Graph Theory: Penn State Math 485 Lecture Notes: Licensed Under A
Graph Theory: Penn State Math 485 Lecture Notes: Licensed Under A
Notes
Version 1.0.6
Christopher Grin
2011-2012
Licensed under a Creative Commons Attribution-Noncommercial-Share Alike 3.0 United States License
With Contributions By:
Suraj Shekhar
Contents
List of Figures v
Using These Notes xi
Chapter 1. Preface and Introduction to Graph Theory 1
1. Some History of Graph Theory and Its Branches 1
2. A Little Note on Network Science 2
Chapter 2. Some Denitions and Theorems 3
1. Graphs, Multi-Graphs, Simple Graphs 3
2. Directed Graphs 8
3. Elementary Graph Properties: Degrees and Degree Sequences 9
4. Subgraphs 15
5. Graph Complement, Cliques and Independent Sets 16
Chapter 3. More Denitions and Theorems 19
1. Paths, Walks, and Cycles 19
2. More Graph Properties: Diameter, Radius, Circumference, Girth 21
3. More on Trails and Cycles 22
4. Graph Components 23
5. Introduction to Centrality 28
6. Bipartite Graphs 29
7. Acyclic Graphs and Trees 31
Chapter 4. Some Algebraic Graph Theory 39
1. Isomorphism and Automorphism 39
2. Fields and Matrices 45
3. Special Matrices and Vectors 47
4. Matrix Representations of Graphs 47
5. Determinants, Eigenvalue and Eigenvectors 50
6. Properties of the Eigenvalues of the Adjacency Matrix 53
Chapter 5. Applications of Algebraic Graph Theory: Eigenvector Centrality and
Page-Rank 57
1. Basis of R
n
57
2. Eigenvector Centrality 59
3. Markov Chains and Random Walks 62
4. Page Rank 66
Chapter 6. Trees, Algorithms and Matroids 69
iii
1. Two Tree Search Algorithms 69
2. Prims Spanning Tree Algorithm 71
3. Computational Complexity of Prims Algorithm 75
4. Kruskals Algorithm 77
5. Shortest Path Problem in a Positively Weighted Graph 79
6. Greedy Algorithms and Matroids 83
Chapter 7. A Brief Introduction to Linear Programming 87
1. Linear Programming: Notation 87
2. Intuitive Solutions of Linear Programming Problems 88
3. Some Basic Facts about Linear Programming Problems 91
4. Solving Linear Programming Problems with a Computer 94
5. Kurush-Kuhn-Tucker (KKT) Conditions 96
6. Duality 99
Chapter 8. An Introduction to Network Flows and Combinatorial Optimization 105
1. The Maximum Flow Problem 105
2. The Dual of the Flow Maximization Problem 106
3. The Max-Flow / Min-Cut Theorem 108
4. An Algorithm for Finding Optimal Flow 111
5. Applications of the Max Flow / Min Cut Theorem 115
Chapter 9. A Short Introduction to Random Graphs 119
1. Bernoulli Random Graphs 119
2. First Order Graph Language and 0 1 properties 122
3. Erd os-Renyi Random Graphs 123
Chapter 10. Some More Algebraic Graph Theory 129
1. Vector Spaces and Linear Transformation 129
2. Linear Span and Basis 131
3. Vector Spaces of a Graph 132
4. Cycle Space 133
5. Cut Space 136
6. The Relation of Cycle Space to Cut Space 139
Bibliography 141
iv
List of Figures
2.1 It is easier for explanation to represent a graph by a diagram in which vertices
are represented by points (or squares, circles, triangles etc.) and edges are
represented by lines connecting vertices. 4
2.2 A self-loop is an edge in a graph G that contains exactly one vertex. That is, an
edge that is a one element subset of the vertex set. Self-loops are illustrated by
loops at the vertex in question. 5
2.3 The city of K onigsburg is built on a river and consists of four islands, which
can be reached by means of seven bridges. The question Euler was interested
in answering is: Is it possible to go from island to island traversing each
bridge only once? (Picture courtesy of Wikipedia and Wikimedia Commons:
http://en.wikipedia.org/wiki/File:Konigsberg_bridges.png) 5
2.4 Representing each island as a dot and each bridge as a line or curve connecting
the dots simplies the visual representation of the seven Konigsburg Bridges. 5
2.5 During World War II two of the seven original Konigsburg bridges were
destroyed. Later two more were made into modern highways (but they are still
bridges). Is it now possible to go from island to island traversing each bridge
only once? (Picture courtesy of Wikipedia and Wikimedia Commons: http:
//en.wikipedia.org/wiki/File:Konigsberg_bridges_presentstatus.png) 6
2.6 A multigraph is a graph in which a pair of nodes can have more than one edge
connecting them. When this occurs, the for a graph G = (V, E), the element E
is a collection or multiset rather than a set. This is because there are duplicate
elements (edges) in the structure. 7
2.7 (a) A directed graph. (b) A directed graph with a self-loop. In a directed graph,
edges are directed; that is they are ordered pairs of elements drawn from the
vertex set. The ordering of the pair gives the direction of the edge. 8
2.8 The graph above has a degree sequence d = (4, 3, 2, 2, 1). These are the degrees
of the vertices in the graph arranged in increasing order. 10
2.9 We construct a new graph G
= 6, 7, 8, 9, 10. The
subgraph shown in (b) is a spanning sub-graph and is induced by edge subset E
=
1, 6 , 2, 9 , 3, 7 , 4, 10 , 5, 8 , 6, 7 , 6, 10 , 7, 8 , 8, 9 , 9, 10. 16
2.13 A clique is a set of vertices in a graph that induce a complete graph as a
subgraph and so that no larger set of vertices has this property. The graph in
this gure has 3 cliques. 17
2.14 A graph and its complement with cliques in one illustrated and independent sets
in the other illustrated. 17
2.15 A covering is a set of vertices so that ever edge has at least one endpoint inside
the covering set. 18
3.1 A walk (a), cycle (b), Eulerian trail (c) and Hamiltonian path (d) are illustrated. 20
3.2 We illustrate the 6-cycle and 4-path. 21
3.3 The diameter of this graph is 2, the radius is 1. Its girth is 3 and its
circumference is 4. 22
3.4 We can create a new walk from an existing walk by removing closed sub-walks
from the walk. 23
3.5 We show how to decompose an (Eulerian) tour into an edge disjoint set of cycles,
thus illustrating Theorem 3.26. 24
3.6 A connected graph (a), a disconnected graph (b) and a connected digraph that
is not strongly connected (c). 24
3.7 We illustrate a vertex cut and a cut vertex (a singleton vertex cut) and an edge
cut and a cut edge (a singleton edge cut). Cuts are sets of vertices or edges
whose removal from a graph creates a new graph with more components than
the original graph. 25
3.8 If e lies on a cycle, then we can repair path w by going the long way around the
cycle to reach v
n+1
from v
1
. 26
3.9 Path graph with four vertices. 28
3.10 The graph for which you will compute centralities. 29
3.11 A bipartite graph has two classes of vertices and edges in the graph only exists
between elements of dierent classes. 30
3.12 Illustration of the main argument in the proof that a graph is bipartite if and
only if all cycles have even length. 31
3.13 A tree is shown. Imagining the tree upside down illustrates the tree like nature
of the graph structure. 32
3.14 The Petersen Graph is shown on the left while a spanning tree is shown on the
right in red. 32
vi
3.15 The proof of 4 = 5 requires us to assume the existence of two paths in graph
T connecting vertex v to vertex v
vV
deg(v)
Proof. Consider two vertices v
1
and v
2
in V . If e = v
1
, v
2
then a +1 is contributed to
vV
deg(v) for both v
1
and v
2
. Thus every non-self-loop edge contributes +2 to the vertex
degree sum. On the other hand, if e = v
1
is a self-loop, then this edge contributes +2
to the degree of v
1
. Therefore, each edge contributes exactly +2 to the vertex degree sum.
Equation 2.4 follows immediately.
Corollary 2.38. Let G = (V, E). Then there are an even number of vertices in V with
odd degree.
Exercise 5. Prove Corollary 2.38.
Definition 2.39 (Graphic Sequence). Let d = (d
1
, . . . , d
n
) be a tuple in Z
n
with d
1
d
2
d
n
. Then d is graphic if there exists a graph G with degree sequence d.
Corollary 2.40. If d is graphic, then the sum of its elements is even.
Exercise 6. Prove Corollary 2.40.
Lemma 2.41. Let d = (d
1
, . . . , d
n
) be a graphic degree sequence. Then there exists a
graph G = (V, E) with degree sequence d so that if V = v
1
, . . . , v
n
then:
(1) deg(v
i
) = d
i
for i = 1, . . . , n and
(2) v
1
is adjacent to vertices v
2
, . . . , v
d
1
+1
.
Proof. The fact that d is graphic means there is at least one graph whose degree
sequence is equal to d. From among all those graphs, chose G = (V, E) to maximize
(2.5) r = [N(v
1
) v
2
, . . . , v
d
1
+1
[
Recall that N(v
1
) is the neighborhood of v
1
. Thus maximizing Expression 2.5 implies we are
attempting to make sure that as many vertices in the set v
2
, . . . , v
d
1
+1
are adjacent to v
1
as possible.
If r = d
1
, then the theorem is proved since v
1
is adjacent to v
2
, . . . , v
d
1
+1
. Therefore well
proceed by contradiction and assume r < d
1
. We know the following things:
(1) Since deg(v
1
) = d
1
there must be a vertex v
t
with t > d
1
+ 1 so that v
t
is adjacent
to v
1
.
(2) Moreover, there is a vertex v
s
with 2 s d
1
+ 1 that is not adjacent to v
1
.
(3) By the ordering of V , deg(v
s
) deg(v
t
); that is d
s
d
t
.
11
(4) Therefore, there is some vertex v
k
V so that v
s
is adjacent to v
k
but v
t
is not
because v
t
is adjacent to v
1
and v
s
is not and the degree of v
s
is at least as large as
the degree of v
t
.
Let us create a new graph G
= (V, E
:
r
= [N
G
(v
1
) v
2
, . . . , v
d
1
+1
[
and we have r
> r. This contradicts our initial choice of G and proves the theorem.
Theorem 2.42 (Havel-Hakimi Theorem). A degree sequence d = (d
1
, . . . , d
n
) is graphic
if and only if the sequence (d
2
1, . . . , d
d
1
+1
1, d
d
1
+2
, . . . , d
n
) is graphic.
Proof. () Suppose that d = (d
1
, . . . , d
n
) is graphic. Then by Lemma 2.41 there is a
graph G with degree sequence d so that:
(1) deg(v
i
) = d
i
for i = 1, . . . , n and
(2) v
1
is adjacent to vertices v
2
, . . . , v
d
1
+1
.
If we remove vertex v
1
and all edges containing v
1
from this graph G to obtain G
then in G
for all i = 2, . . . d
1
+1 the degree of v
i
is d
i
1 while for j = d
1
+2, . . . , n the degree of v
j
is
d
j
because v
1
is not adjacent to v
d
1
+2
, . . . , v
n
by choice of G. Thus G
from G by
12
adding a vertex v
1
to G and creating an edge from v
1
to each vertex v
2
through v
d
1
+1
. It
is clear that the degree of v
1
is d
1
, while the degrees of all other vertices v
i
must be d
i
and
thus d = (d
1
, . . . , d
n
) is graphic because it is the degree sequence of G
i=0
d
i
k(k + 1) +
n1
i=k+1
mink + 1, d
i
Exercise 8 (Independent Project). There are several proofs of Theorem 2.45, some
short. Investigate them and reconstruct an annotated proof of the result. In addition
investigate Bergs approach using ows [Ber73].
Remark 2.46. There has been a lot of interest recently in degree sequences of graphs,
particularly as a result of the work in Network Science on so-called scale-free networks. This
has led to a great deal of investigation into properties of graphs with specic kinds of degree
sequences. For the brave, it is worth looking at [MR95], [ACL01], [BR03] and [Lu01] for
interesting mathematical results in this case. To nd out why all this investigation started,
see [BAJB00].
3.1. Types of Graphs from Degree Sequences.
Definition 2.47 (Complete Graph). Let G = (V, E) be a graph with [V [ = n with
n 1. If the degree sequence of G is (n 1, n 1, . . . , n 1) then G is called a complete
graph on n vertices and is denoted K
n
. In a complete graph on n vertices each vertex is
connected to every other vertex by an edge.
Lemma 2.48. Let K
n
= (V, E) be the complete graph on n vertices. Then:
[E[ =
n(n 1)
2
13
Corollary 2.49. Let G = (V, E) be a graph and let [V [ = n. Then:
0 [E[
_
n
2
_
Exercise 9. Prove Lemma 2.48 and Corollary 2.49. [Hint: Use Equation 2.4.]
Definition 2.50 (Regular Graph). Let G = (V, E) be a graph with [V [ = n. If the
degree sequence of G is (k, k, . . . , k) with k n 1 then G is called a k-regular graph on n
vertices.
Example 2.51. We illustrate one complete graph and two (non-complete) regular graphs
in Figure 2.10. Obviously every complete graph is a regular graph. Every Platonic solid is
(a) K
4
(b) Petersen Graph (c) Dodecahedron
Figure 2.10. The complete graph, the Petersen Graph and the Dodecahedron.
All Platonic solids are three-dimensional representations of regular graphs, but not
all regular graphs are Platonic solids. These gures were generated with Maple.
also a regular graph, but not every regular graph is a Platonic solid. In Figure 2.10(c) we
show a attened dodecahedron, one of the ve platonic solids from classical geometry. The
Peteron Graph (Figure 2.10(b)) is a 3-regular graph that is used in many graph theoretic
examples.
3.2. Digraphs.
Definition 2.52 (In-Degree, Out-Degree). Let G = (V, E) be a digraph. The in-degree
of a vertex v in G is the total number of edges in E with destination v. The out-degree of
v is the total number of edges in E with source v. We will denote the in-degree of v by
deg
in
(v) and the out-degree by deg
out
(v).
Theorem 2.53. Let G = (V, E) be a digraph. Then the following holds:
(2.7) [E[ =
vV
deg
in
(v) =
vV
deg
out
(v)
Exercise 10. Prove Theorem 2.53.
14
4. Subgraphs
Definition 2.54 (Subgraph). Let G = (V, E). A graph H = (V
, E
) is a subgraph of G
if V
V and E
V or E
E.
Example 2.55. We illustrate the notion of a sub-graph in Figure 2.11. Here we illustrate
a sub-graph of the Petersen Graph. The sub-graph contains vertices 6, 7, 8, 9 and 10 and
the edges connecting them.
(a) Petersen Graph (b) Highlighted Subgraph (c) Extracted Subgraph
Figure 2.11. The Petersen Graph is shown (a) with a sub-graph highlighted (b)
and that sub-graph displayed on its own (c). A sub-graph of a graph is another
graph whose vertices and edges are sub-collections of those of the original graph.
Definition 2.56 (Spanning Subgraph). Let G = (V, E) be a graph and H = (V
, E
) be
a subgraph of G. The subgraph H is a spanning subgraph of G if V
= V .
Definition 2.57 (Edge Induced Subgraph). Let G = (V, E) be a graph. If E
E. The
subgraph of G induced by E
is the graph H = (V
, E
) where v V
E.
The subgraph of G induced by V
is the graph H = (V
, E
) where v
1
, v
2
E
if and only
if v
1
and v
2
are both in V
.
Remark 2.59. For directed graphs, all sub-graph denitions are modied in the obvious
way. Edges become directed as one would expect.
Example 2.60. Using the Petersen Graph we illustrate a subgraph induced by a vertex
subset and a spanning subgraph. In Figure 2.12(a) we illustrate the subgraph induced by
the vertex subset V
= 1, 6 , 2, 9 , 3, 7 , 4, 10 , 5, 8 , 6, 7 , 6, 10 , 7, 8 , 8, 9 , 9, 10
15
(a) Highlighted Subgraph (b) Spanning Subgraph
Figure 2.12. The subgraph (a) is induced by the vertex subset V
= 6, 7, 8, 9, 10.
The subgraph shown in (b) is a spanning sub-graph and is induced by edge subset
E
= 1, 6 , 2, 9 , 3, 7 , 4, 10 , 5, 8 , 6, 7 , 6, 10 , 7, 8 , 8, 9 , 9, 10.
5. Graph Complement, Cliques and Independent Sets
Definition 2.61 (Clique). Let G = (V, E) be a graph. A clique is a set S V of
vertices so that:
(1) The subgraph induced by S is a complete graph (or in general graphs, every pair of
vertices in S is connected by at least one edge in E) and
(2) If S
) so that:
e = v
1
, v
2
E
v
1
, v
2
, E
16
(a) Cliques (b) Independent Set
Figure 2.13. A clique is a set of vertices in a graph that induce a complete graph
as a subgraph and so that no larger set of vertices has this property. The graph in
this gure has 3 cliques.
Example 2.67. In Figure 2.14, the graph from Figure 2.13 is illustrated (in a dierent
spatial conguration) with its cliques. The complement of the graph is also illustrated.
Notice that in the complement, every clique is now an independent set.
Figure 2.14. A graph and its complement with cliques in one illustrated and in-
dependent sets in the other illustrated.
Definition 2.68 (Relative Complement). If G = (V, E) is a graph and H = (V, E
) is a
spanning sub-graph, then the relative complement of H in G is the graph H
= (V, E
) with:
e = v
1
, v
2
E
v
1
, v
2
E and v
1
, v
2
, E
E. If v I,
then clearly v
is in I,
but this does not aect that fact that V I must be a cover since for every edge e E at
least one element is in V I.
() Now suppose that V I is a vertex covering. Choose any two vertices v and v
in I.
The fact that V I is a vertex covering implies that v, v
, E
) with:
(1) V
= v
1
, . . . , v
n+1
(2) E
= e
1
, . . . , e
n
= (v
m
, e
m
, . . . , e
m+k1
, v
m+k
)
is a closed sub-walk of w. The walk w
= (v
1
, e
1
, v
2
, . . . , v
m+k
, e
m+k
, v
m+k+1
, . . . , v
n
, e
n
, v
n+1
)
22
that is shorter than the original walk. This is illustrated in Figure 3.4. [GY05] calls this a
walk reduction, though this notation is not standard.
1 2 3
4
5
6
7 8 1 2 3 7 8
w
w
Figure 3.4. We can create a new walk from an existing walk by removing closed
sub-walks from the walk.
Lemma 3.25. Let G = (V, E) be a graph and suppose that t is a non-trivial tour (closed
trail) in G. Then t contains a cycle.
Proof. The fact that t is closed implies that it contains at least one pair of repeated
vertices. Therefore a closed sub-walk of t must exist since t is itself has these repeated
vertices. Let c be a minimal (length) closed sub-walk of t. We will show that c must be a
cycle. By way of contradiction, suppose that c is not a cycle. Then since it is closed it must
contain a repeated vertex (that is not its rst vertex). If we applied our observation from
Remark 3.24 we could produce a smaller closed walk c
,
then t is still a tour and has length at most N. We can now apply the induction hypothesis
to see that this new tour t
E. Then the
graph G
. We write this as G
= GE
.
Definition 3.37 (Vertex Deletion Graph). Let G = (V, E) and let V
V . Then the
graph G
. We write this as G
= GV
.
Definition 3.38 (Vertex Cut and Cut Vertex). Let G = (V, E) be a graph. A set
V
E
is a edge cut if the graph G
of G is minimal
if when we remove any edge from E
to form E
= Ge
is connected if and only if e lies on a cycle in G.
25
Proof. () Recall a graph G is connected if and only if for every pair of vertices v
1
and v
n+1
there is a walk w from v
1
to v
n+1
with:
w = (v
1
, e
1
, v
2
, . . . , v
n
, e
n
, v
n+1
)
Let G
we
will have the path:
c
= (u
1
, f
1
, . . . , u
m
)
The fact that e is in the walk w implies there are vertices v
i
and v
i+1
so that e = v
i
, v
i+1
(with v
i
= u
1
and v
i+1
= u
m
). In deleting e from G we remove the sub-walk (v
i
, e, v
i+1
) from
w. But we can create a new walk with structure:
w
= (v
1
, e
1
, . . . , v
i
, f
1
, u
2
, . . . , u
m1
, f
m1
, u
m
, . . . , e
n
, v
n+1
)
This is illustrated in Figure 3.8.
u
1
v
1
v
2
v
i
v
n+1
v
i+1
e
u
m
Figure 3.8. If e lies on a cycle, then we can repair path w by going the long way
around the cycle to reach v
n+1
from v
1
.
() Suppose G
is connected,
there is a walk from v
1
to v
n+1
. Applying Remark 3.24, we can reduce this walk to a path p
with:
p = (v
1
, e
1
, . . . , v
n
, e
n
, v
n+1
)
Since p is a path, there are no repeated vertices in p. We can construct a cycle c containing
e in G as:
p = (v
1
, e
1
, . . . , v
n
, e
n
, v
n+1
, e, v
1
)
since e = v
1
, v
n+1
= v
n+1
, v
1
. Thus, e lies on a cycle in G. This completes the proof.
Corollary 3.43. Let G = (V, E) be a connected graph and let e E. The edge e is a
cut edge if and only if e does not lie on a cycle in G.
Exercise 20. Prove Corollary 3.43.
26
Remark 3.44. The next result is taken from Extremal Graph Theory, the study of
extremes or bounds in properties of graphs. There are a number of results in Extremal
Graph Theory that are of interest. See [Bol04] for a complete introduction.
Theorem 3.45. If G = (V, E) is a graph with n vertices and k components, then:
[E[
(n k + 1)(n k)
2
Proof. Assume that each component of G has n
i
vertices in it with
k
i=1
n
i
= n.
Applying Lemma 2.48 we know that Component i has at most n
i
(n
i
1)/2 edges; that is,
each component is a complete graph on n
i
vertices. This is the largest number of edges that
can occur under these assumptions.
Consider the case where k 1 of the components has exactly 1 vertex and the remaining
component has n (k 1) vertices. Then the total number of edges in this case is:
(n (k 1))(n (k 1) 1)
2
=
(n k + 1)(n k)
2
edges. It now suces to show that this case has the greatest number of vertices of all cases
where the k components are each complete graphs.
Consider the case when component i is K
r
and component j is K
s
with r, s 2 and
suppose r s. Then the total number of edges in these two components is:
r(r 1) +s(s 1)
2
=
r
2
+s
2
r s
2
Now, suppose we move one vertex in component j to component i. Then component i is
now K
r+1
and component j is now K
s1
. Applying Lemma 2.48, the number of edges in this
case is:
(r + 1)(r) + (s 1)(s 2)
2
=
r
2
+r +s
2
3s + 2
2
Observe that since r s, substituting s for r we have:
r
2
+r +s
2
3s + 2 r
2
+s
2
2s + 2
By a similar argument:
r
2
+s
2
2s r
2
+s
2
r s
Thus we conclude that:
r
2
+r +s
2
3s + 2
2
r
2
+s
2
2s + 2
2
r
2
+s
2
2s
2
r
2
+s
2
r s
2
Repeating this argument over and over shows that in a k component graph with n vertices,
the largest number of edges must occur in the case when there is one component with
n(k1) vertex and k1 components with exactly 1 vertex. This completes the proof.
Corollary 3.46. Any graph with n vertices and more than (n 1)(n 2)/2 edges is
connected.
Exercise 21. Prove Corollary 3.46.
27
5. Introduction to Centrality
Remark 3.47. There are many situations in which wed like to measure the importance
of a vertex in a graph. The problem of measuring this quantity is usually called determining
a vertexs centrality.
Definition 3.48 (Degree Centrality). Let G = (V, E) be a graph. The degree centrality
of a vertex is just its degree or for a centrality in the set [0, 1], we may dene the degree
centrality of vertex v
i
as deg(v
i
)/2[E[
Exercise 22. Show that if we require the degree centralities of a graph to be in the
interval [0, 1], then the sum of the centralities equals 1.
Remark 3.49. Degree centrality is only the simplest measurement of centrality. There
are many other measures of this quantity we discuss one more and then continue our discus-
sion of this topic in Chapter 5.
Definition 3.50 (Geodesic Centrality). Let G = (V, E) be a graph. The geodesic
centrality (sometimes called the betweeness) of a vertex v V is the fraction of times
v occurs on any shortest path connecting any other pair of vertices s, t V . Put more
formally, let
st
be the total number of shortest paths connecting vertex s with vertex t. Let
st
(v) be the number of these shortest paths containing v. The geodesic centrality of v is:
(3.4) C
B
(v) =
s=t=v
st
(v)
st
These values can be normalized so that they fall within [0, 1] by dividing each C
B
(v) by the
sum of all C
B
(v).
Example 3.51. Consider the graph with 4 vertices shown below. The degrees of the
graph are (2, 3, 3, 2), which is the unnormalized degree centrality. The normalized degree
centrality of the vertices is:
(1) v
1
:
1
5
(2) v
2
:
3
10
(3) v
3
:
3
10
(4) v
4
:
1
5
1 2 3 4
Figure 3.9. Path graph with four vertices.
To compute the normalized Geodesic centrality, we must compute the fraction of times
a vertex appears in a shortest path. This is shown in the table below in Table 1: In the
vertex pair (1, 2) there is exactly one shortest path connecting 1 to 2. Since 1 and 2 are
the end points, they are not counted. Vertices 3 and 4 do not appear in this shortest path,
so they each receive a zero. For (1, 4) there are two shortest paths (one through 2 and the
other through 3) therefore 1/2 of the shortest paths contain vertex 2 and 1/2 of the shortest
28
Vertex Pair 1 2 3 4
(1,2) - - 0 0
(1,3) - 0 - 0
(1,4) -
1
2
1
2
-
(2,3) 0 - - 0
(2,4) 0 - 0 -
(3,4) 0 0 - -
SUM 0
1
2
1
2
0
Table 1. A table showing the intermediate computations for geodesic centrality.
paths contain vertex 3. The remainder of the table is lled out in exactly the same way. The
normalized geodesic centrality is:
(1) v
1
: 0
(2) v
2
:
1
2
(3) v
3
:
1
2
(4) v
4
: 0
In this case, we see that the centrality measures are similar in their ordering, but dierent
in their values.
Exercise 23. Compute the geodesic centrality and the degree centrality for the graph
shown in Figure 3.10. Compare your results.
1
2 3
4 5
Figure 3.10. The graph for which you will compute centralities.
Remark 3.52. Its clear from this analysis that cut vertices should have high geodesic
centrality if they connect two large components of a graph. Thus, by some measures, cut
vertices are very important elements of graphs.
6. Bipartite Graphs
Definition 3.53. A graph G = (V, E) is bipartite if V = V
1
V
2
and V
1
V
2
= and
if e = E, then e = v
1
, v
2
with v
1
V
1
and v
2
V
2
. This denition is valid for non-simple
graphs as well.
Remark 3.54. In a bipartite graph, we can think of the vertices as belonging to one
of two classes (either V
1
or V
2
) and edges only exist between elements of the two classes,
29
not between elements in the same class. We can also dene n-partite graphs in which the
vertices are in any of n classes and there are only edges between classes, not within classes.
Example 3.55. Figure 3.11 shows a bipartite graph in which V
1
= 1, 2, 3 and V
2
=
4, 5, 6, 7. Notice that there are only edges connecting vertices in V
1
and vertices in V
2
.
There are not edges connecting elements in V
1
to other elements in V
1
or elements in V
2
to
other elements in V
2
.
Figure 3.11. A bipartite graph has two classes of vertices and edges in the graph
only exists between elements of dierent classes.
Definition 3.56 (Path Concatenation). Let p
1
= (v
1
, e
1
, v
2
, . . . , v
n
, e
n
, v
n+1
) and let
p
2
= (v
n+1
, e
n+1
, v
n+2
, . . . , v
n+m
, e
n+m
, v
n+m+1
). Then the concatenation of path p
1
with
path p
2
is the path:
p = (v
1
, e
1
, v
2
, . . . , v
n
, e
n
, v
n+1
, e
n+1
, v
n+2
, . . . , v
n+m
, e
n+m
, v
n+m+1
)
Remark 3.57. Path concatenation is illustrated in the proof of Theorem 3.58.
Theorem 3.58. A graph G = (V, E) is bipartite if and only if every cycle in G has even
length.
Proof. () Suppose G is bipartite. Every cycle begins and ends at the same vertex
and therefore in the same partition, either V
1
or V
2
. Starting at a vertex v
1
V
1
we must
take a walk of length 2 to return to V
1
. The same is true if we start at a vertex in V
2
. Thus
every cycle must contain an even number of edges in order to return to either V
1
or V
2
.
() Suppose that every cycle in G has even length. Without loss of generality, assume
G is connected. We will create a partition of V so that V = V
1
V
2
and and V
1
V
2
=
and there is no edge between vertices if they are in the same class.
Choose an arbitrary vertex v V and dene:
V
1
= v
V : d
G
(v, v
) 0 mod 2 (3.5)
V
2
= v
V : d
G
(v, v
) 1 mod 2 (3.6)
Clearly V
1
and V
2
constitute a partition of V . Choose u
1
, u
2
V
1
and suppose e = u
1
, u
2
E. The distance from v to u
1
is even, so there is a path p
1
with an even number of edges
beginning at v and ending at u
1
. Likewise the distance from v to u
2
is even, so there is a
path p
2
beginning at u
2
and ending at v with an even number of edges. If we concatenate
30
paths p
1
and the length 1 path q = (u
1
, u
1
, u
2
, u
2
) and path p
2
we obtain a cycle in G that
has odd length. Therefore, there can be no edge connecting two vertices in V
1
.
Choose u
1
, u
2
V
2
and suppose that e = u
1
, u
2
E. Using the same argument, there
is a path p
1
of odd length from v to u
1
and a path p
2
of odd length from u
2
to v. If we
concatenate paths p
1
and the length 1 path q = (u
1
, u
1
, u
2
, u
2
) and path p
2
we again obtain
a cycle in G that has odd length. Therefore, there can be no edge connecting two vertices
in V
2
. These arguments are illustrated in Figure 3.12
u
1
u
2
v
Both even or odd length paths
e
Figure 3.12. Illustration of the main argument in the proof that a graph is bipartite
if and only if all cycles have even length.
In the case when G has more than one component, execute the process described above
for each component to obtain partitions V
1
, V
2
, V
3
, V
4
, . . . , V
2n
. Create a bipartition U
1
and
U
2
of V with:
U
1
=
n
_
k=1
V
2k1
(3.7)
U
2
=
n
_
k=1
V
2k
(3.8)
Clearly there can be no edge connecting a vertex in U
1
with a vertex in U
2
. This completes
the proof.
7. Acyclic Graphs and Trees
Definition 3.59 (Acyclic Graph). A graph that contains no cycles is called acyclic.
Definition 3.60 (Forests and Trees). Let G = (V, E) be an acyclic graph. If G has
more than one component, then G is called a forest. If G has one component, then G is
called a tree.
Example 3.61. A randomly generated tree with 10 vertices is shown in Figure 3.13.
Note that a tree (if drawn upside down) can be made to look exactly like a real tree growing
up from the ground.
Remark 3.62. We can dene directed trees and directed forests as acyclic directed
graphs. Generally speaking, we require the underlying graphs to be acyclic rather than just
having no directed cycles. See Chapter 4 of [Gri11a] (http://www.personal.psu.edu/
cxg286/Math486.pdf) for the use of directed trees in Game Theory. For the remainder of
this chapter we will deal undirected trees, but results will apply to directed trees unless
otherwise noted.
31
Figure 3.13. A tree is shown. Imagining the tree upside down illustrates the tree
like nature of the graph structure.
Definition 3.63 (Spanning Forest). Let G = (V, E) be a graph. If F = (V
, E
) is an
acyclic subgraph of G such that V = V
) of G.
32
Proof. We proceed by induction on the number of vertices in G. If [V [ = 1, then G
is itself a (degenerate) tree and thus has a spanning tree. Now, suppose that the statement
is true for all graphs G with [V [ n. Consider a graph G with n + 1 vertices. Choose an
arbitrary vertex v
n+1
and remove it and all edges of the formv, v
n+1
from G to form G
with
vertex set V
= v
1
, . . . , v
n
. The graph G
).
Exercise 24. Prove Corollary 3.66.
Definition 3.67 (Leaf). Let T = (V, E). If v V and deg(v) = 1, then v is called a
leaf of T.
Lemma 3.68. Every tree with one edge has at least two leaves.
Proof. Let:
w = (v
1
, e
1
, v
2
, . . . , v
n
, e
n
, v
n+1
)
be a path of maximal length in T. Consider vertex v
n+1
. If deg(v
n+1
) > 1, then there are
two possibilities: (i) there is an edge e
n+1
and a vertex v
n+2
with v
n+2
not in the sequence
w. In this case, we can extend w to w
dened as:
w
= (v
1
, e
1
, v
2
, . . . , v
n
, e
n
, v
n+1
, e
n+1
, v
n+2
)
which contradicts our assumption that w was maximal in length. (ii) there is an edge e
n+1
and a vertex v
n+2
and for some k 1, . . . , n, v
n+2
= v
k
; i.e., v
n+2
is in the sequence w. In
this case, there is a closed sub-walk:
w
= (v
k
, e
k
, v
k+1
, . . . , v
n+1
, e
n+1
, v
n+2
)
Since w is a path, there are no other repeated vertices in the sequence w
and thus w
is a
cycle in T, contradicting our assumption that T was a tree. The reasoning above holds for
vertex v
1
as well, thus the two end points of every maximal path in a tree must be leaves.
This completes the proof.
Corollary 3.69. Let G = (V, E) be a graph. If each vertex in V has degree at least 2,
then G contains a cycle.
Exercise 25. Prove Corollary 3.69.
Lemma 3.70. Let T = (V, E) be a tree with [V [ = n. Then [E[ = n 1.
33
Proof. We will proceed by induction. For the case when n = 1, this statement must be
true. Now, suppose that the statement is true [V [ n. We will show that when [V [ = n+1,
then [E[ = n, assuming that T = (V, E) is a tree. By Lemma 3.68 we know that if T is a tree,
then it contains one component and at least two leaves. Therefore, choose a vertex v V
that is a leaf in T. There is some edge e = v
= (V
, E
)
with: V
= V v and E
must:
(1) have one component since v was connected to only one other vertex v
V and T
had only one component and
(2) by acyclic since T itself was acyclic and we have not introduced new edges to create
a cycle.
Therefore T
from T it follows
that T had exactly n edges and our originally assumed n + 1 vertices. The required result
follows immediately from induction.
Corollary 3.71. If G = (V, E) is a forest with n vertices, then G has n c(G) edges.
(Recall c(G) is the number of components in G).
Exercise 26. Prove Corollary 3.71.
Theorem 3.72. A graph G = (V, E) is connected if and only if it has a spanning tree.
Exercise 27. Prove Theorem 3.72.
Theorem 3.73. Let T = (V, E) be a graph with [V [ = n. Then the following are
equivalent:
(1) T is a tree.
(2) T is acyclic and has exactly n 1 edges.
(3) T is connected and has exactly n 1 edges.
(4) T is connected and every edge is a cut-edge.
(5) Any two vertices of T are connected by exactly one path.
(6) T is acyclic and the addition of any new edge creates exactly one cycle in the resulting
graph.
Proof. (1 = 2) Assume T is a tree. Then by denition T is acyclic and the fact that
it has n 1 edges follows from Lemma 3.70.
(2 = 3) Since T is acyclic, it must be a forest and by Corollary 3.71 [E[ = n c(T).
Since we assumed that T has n 1 edges, we must have n c(T) = n 1 and thus the
number of components of T is 1 and thus T must be connected.
(3 = 4) The fact that T is connected is assumed from 3. Suppose we consider the
graph T
= (V, E
) where E
contains n vertices and must still be acyclic (that is a forest) and therefore n2 = nc(T
).
Thus c(T
.
v
v
w
w
cannot exist.
(5 = 6) The fact that any pair of vertices is connected in T implies T is connected
(i.e., has one component). Now suppose that T has a cycle (like the one illustrated in Figure
3.15). Then it is easy to see there are (at least) two paths connecting w and w
contradicting
our assumption. Therefore, T is acyclic. The fact that adding an edge creates exactly one
cycle can be seen in the following way: Consider two vertices v and v
, v
)
in T connecting v and v
, v
, v, v
, v)
so at least one cycle is created. To see that this cycle is unique, note that if there is another
cycle present then it must contain the edge v, v
, v
, v, v
, v)
where there is at least one vertex w
i
not present in the set u
1
, . . . , u
n
(otherwise the two
cycles are identical). We now see there must be two disjoint paths connecting v and v
,
namely:
(v, v, u
1
, u
1
, . . . , u
n
, u
n
, v
, v
)
and
(v, v, w
1
, w
1
, . . . , w
n
, w
n
, v
, v
)
this contradicts our assumption on T. Thus the created cycle is unique.
(6 = 1) It suces to show that T has a single component. Suppose not, there are
at least two components of T. Chose two vertices v and v
i=1
d
i
= 2n 2
(2) n 2, d
i
> 0 for i = 1, . . . , n and Equation 3.9 holds.
Proof. () See Exercise 28.
() Now suppose that Equation 3.9 holds. If n = 1, then 2n 2 = 0 and thus d
1
= 0
and d = (d
1
). This is the degree sequence of a degenerate tree with one vertex. We now
proceed by induction to establish the remainder of the theorem. If n = 2, 2n 2 = 2 and if
d
1
, d
2
> 0 then then d
1
= d
2
= 1 by necessity. This is the degree sequence for a tree with two
vertices joined by a single edge and thus it is a tree-graphic degree sequence. Now assume
the statement holds for all integers up to some n. We will show it is true for n+1. Consider
a degree sequence (d
1
, . . . , d
n+1
) such that:
n+1
i=1
d
i
= 2(n + 1) 2 = 2n
We assume that the degrees are ordered (largest rst) and positive. Therefore, d
1
2
(because otherwise d
1
+ + d
n+1
n + 1) and that d
1
n (note in the case that d
1
= n
we must have d
2
= d
3
= = d
n+1
= 1). Moreover, if d
1
= d
2
= = d
n1
= 2,
then d
n
= d
n+1
= 1. Since d
i
> 0 for i = 1, . . . , n + 1 from the previous two facts we
see that for any positive value of d
1
, we must have d
n
= d
n+1
= 1 in order to ensure that
d
1
+d
2
+ +d
n+1
= 2n. Consider the sequence of degrees:
d
= (d
1
1, d
2
, . . . , d
n
)
Since d
n+1
= 1, we can see that (d
1
1)+d
2
+ +d
n
= 2n2. Thus, a permutation of d
to
correct the order leads to a tree-graphic sequence by the induction hypothesis. Let T
be the
tree that results from this tree-graphic degree sequence and let v
1
be the vertex with degree
d
1
1 in T
and it must be acyclic since we have not connected two vertices already in T
of G composed
of only cycles C
1
, . . . , C
n
will have m components with 1 m n. Each component is
composed of at most n edge disjoint cycles and therefore applying the induction hypothesis,
each has a tour. Denote the components K
1
, . . . , K
m
. The cycle C
n+1
shares one vertex in
common with at least one of these components (and perhaps all of them). Without loss of
generality, assume that K
1
is a component sharing a vertex in common with C
n+1
(if not,
reorder the components to make this true). Begin following the tour around K
1
until we
encounter the vertex v
1
that component K
1
and C
n+1
share. At this point, break the tour
of K
1
and begin traversing C
n+1
until (i) we return to v
1
or (ii) we encounter a vertex v
2
that is shared by another component (say K
2
). In case (i), we complete the tour of K
1
and
necessarily we must have completed a tour of the entire graph since it is connected. In case
(ii) we follow the tour of K
2
until we return to v
2
and then continue following C
n+1
until
either case (i) occurs or case (ii) occurs again. In either case, we apply the same logic as
before. Since there are a nite number of components, this process will eventually terminate
with case (i), we complete the tour of K
1
and thus we will have constructed a tour of the
entire graph.
This theorem is illustrated in Figure 3.16. This completes the proof.
37
1
2
3
4
5
1
6
4
5
2
3
7
1
2
5
1
6
7
= +
1
2
3
4
4
5
2
3
Figure 3.16. We illustrate an Eulerian graph and note that each vertex has even
degree. We also show how to decompose this Eulerian graphs edge set into the
union of edge-disjoint cycles, thus illustrating Theorem 3.77. Following the tour
construction procedure (starting at Vertex 5), will give the illustrated Eulerian tour.
Exercise 29. Show by example that Theorem 3.77 does not necessarily hold if we are
only interested in Eulerian trails.
38
CHAPTER 4
Some Algebraic Graph Theory
1. Isomorphism and Automorphism
Definition 4.1 (Injective Mapping). Let S and T be sets. A function f : S T is
injective (sometimes one-to-one) if for all s
1
, s
2
S: f(s
1
) = f(s
2
) s
1
= s
2
.
Definition 4.2 (Surjective Mapping). Let S and T be sets. A function f : S T is
surjective (sometimes onto) if for all t T there exists an s S such that f(s) = t.
Definition 4.3 (Bijective Mapping). Let S and T be sets. A function f : S T is
bijective if f is both injective and surjective.
Definition 4.4 (Isomorphism). Let G = (V, E) and let G
= (V
, E
). The graphs G
and G
are isomorphic, we
write G
= G
.
Exercise 30. Prove that
= is an equivalence relation. [Hint: Recall an equivalence
relation is a binary relation dened on a set S so that (i) for all s S, s s (reexiveness);
(ii) for all s, t S, s t t s (symmetry) and (iii) for all r, s, t T r s and s t
implies r t (transitivity). Here the set is the set of all graphs.]
Definition 4.5. Let G = (V, E) be a graph. Then the set H : H
= G is called the
isomorphism type (or isomorphism class) of G.
Theorem 4.6. Suppose G = (V, E) and G
= (V
, E
with
f : V V
the graph isomorphism between the graphs. Further suppose that the degree
sequence of G is d and the degree sequence of G
is d
. Then:
(1) [V [ = [V
[ and [E[ = [E
[,
(2) For all v V , deg(v) = deg(f(v))
(3) d = d
,
(4) For all v V , ecc(v) = ecc(f(v))
(5) (G) = (G
),
(9) rad(G) = rad(G
)
(10) The girth of G is equal to the girth of G
,
(11) The circumference of G is equal to the circumference of G
.
39
Remark 4.7. The proof of Theorem 4.6 is long and should be clear from the denition
of isomorphism. Isomorphism is really just a way of renaming vertices; we assume that the
vertices in graph G are named from the set V , while the vertices in the set G
are named
from the set V
. If the graphs are identical except for the names we give the vertices (and
thus the names that appear in the edges), then the graphs are isomorphic and all structural
properties are preserved as a result of this. We should not that the converse of Theorem 4.6
does not hold. We illustrate this in Exampleex:DegreeIsom.
Example 4.8. Given two graphs G and G
it is necessary that d = d
when G
= G
,
1
2 3
4
5 6
1
2 3
4
5 6
G
G
0
6
=
Figure 4.1. Two graphs that have identical degree sequences, but are not isomorphic.
but these graphs cannot be isomorphic, since they have dierent numbers of components.
The same is true with the other graph properties. The equality between a property of G
and that same property for G
, but
not sucient. We will not encounter any property of a graph that provides such a necessary
and sucient condition.(See Remark 4.12).
Theorem 4.9. Suppose G = (V, E) and G
= (V
, E
with
f : V V
= f(H) is a subgraph of G
= (V
, E
are
isomorphic.
Definition 4.11 (Subgraph Isomorphism). Given two graphs G = (V, E) and H =
(V
, E
i=1
x
i
y
i
Definition 4.46. Two vectors x and y are orthogonal if x y = 0. (Here 0 is the zero
in the eld over which the vectors are dened.)
Definition 4.47 (Matrix Multiplication). If A R
mn
and B R
np
, then C = AB
is the matrix product of A and B and
(4.7) C
ij
= A
i
B
j
Note, A
i
R
1n
(an n-dimensional vector) and B
j
R
n1
(another n-dimensional vector),
thus making the dot product meaningful.
Example 4.48.
(4.8)
_
1 2
3 4
_ _
5 6
7 8
_
=
_
1(5) + 2(7) 1(6) + 2(8)
3(5) + 4(7) 3(6) + 4(8)
_
=
_
19 22
43 50
_
Definition 4.49 (Matrix Transpose). If A R
mn
is a mn matrix, then the transpose
of A dented A
T
is an mn matrix dened as:
(4.9) A
T
ij
= A
ji
Example 4.50.
(4.10)
_
1 2
3 4
_
T
=
_
1 3
2 4
_
The matrix transpose is a particularly useful operation and makes it easy to transform
column vectors into row vectors, which enables multiplication. For example, suppose x is
an n 1 column vector (i.e., x is a vector in R
n
) and suppose y is an n 1 column vector.
Then:
(4.11) x y = x
T
y
46
Exercise 40. Let A, B R
mn
. Prove by example that AB ,= BA; that is, matrix
multiplication is not commutative. [Hint: Almost any pair of matrices you pick (that can be
multiplied) will not commute.]
Exercise 41. Let A R
mn
and let, B R
np
. Use the denitions of matrix multipli-
cation and transpose to prove that:
(4.12) (AB)
T
= B
T
A
T
[Hint: Note that C
ij
= A
i
B
j
, which moves to the (j, i) position. Now gure out what is
in the (j, i) position of B
T
A
T
.]
3. Special Matrices and Vectors
Definition 4.51 (Identify Matrix). The n n identify matrix is:
(4.13) I
n
=
_
_
1 0 . . . 0
0 1 . . . 0
.
.
.
.
.
.
.
.
.
0 0 . . . 1
_
_
Definition 4.52 (Zero Matrix). The n n zero matrix an n n consisting entirely of
0.
Exercise 42. Let A R
nn
. Show that AI
n
= I
n
A = A. Hence, I is an identify for
the matrix multiplication operation on square matrices. [Hint: Do the multiplication out
long hand.]
Definition 4.53 (Symmetric Matrix). Let M R
nn
be a matrix. The matrix M is
symmetric if M = M
T
.
Definition 4.54 (Invertible Matrix). Let A R
nn
be a square matrix. If there is a
matrix A
1
such that
(4.14) AA
1
= A
1
A = I
n
then matrix A is said to be invertible (or nonsingular) and A
1
is called its inverse. If A is
not invertible, it is called a singular matrix.
4. Matrix Representations of Graphs
Definition 4.55 (Adjacency Matrix). Let G = (V, E) be a graph and assume that
V = v
1
, . . . , v
n
. The adjacency matrix of G is an n n matrix M dened as:
M
ij
=
_
1 v
i
, v
j
E
0 else
Proposition 4.56. The adjacency matrix of a (simple) graph is symmetric.
Exercise 43. Prove Proposition 4.56.
Theorem 4.57. Let G = (V, E) be a graph with V = v
1
, . . . , v
n
and let M be its
adjacency matrix. For k 0, the (i, j) entry of M
k
is the number of walks of length k from
v
i
to v
j
.
47
1 2
3 4
Figure 4.4. The adjacency matrix of a graph with n vertices is an n n matrix
with a 1 at element (i, j) if and only if there is an edge connecting vertex i to vertex
j; otherwise element (i, j) is a zero.
Proof. We will proceed by induction. By denition, M
0
is the n n identity matrix
and the number of walks of length 0 between v
i
and v
j
is 0 if i ,= j and 1 otherwise, thus
the base case is established.
Now suppose that the (i, j) entry of M
k
is the number of walks of length k from v
i
to v
j
.
We will show this is true for k + 1. We know that:
(4.15) M
k+1
= M
k
M
Consider vertices v
i
and v
j
. The (i, j) element of M
k+1
is:
(4.16) M
k+1
ij
=
_
M
k
i
_
M
j
Let:
(4.17) M
k
i
=
_
r
1
. . . r
n
where r
l
, (l = 1, . . . , n), is the number of walks of length k from v
i
to v
l
by the induction
hypothesis. Let:
(4.18) M
j
=
_
_
b
1
.
.
.
b
n
_
_
where b
l
, (l = 1, . . . , n), is a 1 if and only there is an edge v
l
, v
j
E and 0 otherwise.
Then the (i, j) term of M
k+1
is:
(4.19) M
k+1
ij
= M
k
i
M
j
=
n
l=1
r
l
b
l
This is the total number of walks of length k leading to a vertex v
l
, (l = 1, . . . , n), from
vertex v
i
such that there is also an edge connecting v
l
to v
j
. Thus M
k+1
ij
is the number of
walks of length k + 1 from v
i
to v
j
. The result follows by induction.
Example 4.58. Consider the graph in Figure 4.4. The adjacency matrix for this graph
is:
(4.20) M =
_
_
0 1 1 1
1 0 0 1
1 0 0 1
1 1 1 0
_
_
48
Consider M
2
:
(4.21) M
2
=
_
_
3 1 1 2
1 2 2 1
1 2 2 1
2 1 1 3
_
_
This tells us that there are three distinct walks of length 2 from vertex v
1
to itself. These
walks are obvious:
(1) (v
1
, v
1
, v
2
, v
2
, v
1
, v
2
, v
1
)
(2) (v
1
, v
1
, v
2
, v
3
, v
1
, v
3
, v
1
)
(3) (v
1
, v
1
, v
4
, v
4
, v
1
, v
4
, v
1
)
We also see there is 1 path of length 2 from v
1
to v
2
: (v
1
, v
1
, v
4
, v
4
, v
2
, v
4
, v
2
). We can
verify each of the other numbers of paths in M
2
.
Exercise 44. Devise an inecient test for isomorphism between two graphs G and G
using their adjacency matrix representations. Assume it takes 1 time unit to test whether
two nn matrices are equal. What is the maximum amount of time your algorithm takes to
determine that G ,
= G
and E = e
1
, . . . , e
n
. Then the incidence matrix of G is an mn matrix A with:
(4.22) A
ij
=
_
_
0 if v
i
is not in e
j
1 if v
i
is in e
j
and e
j
is not a self-loop
2 if v
i
is in e
j
and e
j
is a self-loop
Theorem 4.62. Let G = (V, E) be a graph with V = v
1
, . . . , v
m
and E = e
1
, . . . , e
n
with incidence matrix A. The sum of every column in A is 2 and the sum of each row in A
is the degree of the vertex corresponding to that row.
Proof. Consider any column in A; it corresponds to an edge e of G. If the edge is a self-
loop, there is only one vertex adjacent to e and thus only one non-zero entry in this column.
Therefore its sum is 2. Conversely, if e connects two vertices, then there are precisely two
49
vertices adjacent to e and thus two entries in this column that are non-zero both with value
1, thus again the sum of the column is 2.
Now consider any row in A; it corresponds to a vertex v of G. The entries in this row are
1 if there is some edge that is adjacent to v and 2 if there is a self-loop at v. From Denition
2.9, we see that adding these values up yields the degree of the vertex v. This completes the
proof.
Exercise 46. Use Theorem 4.62 to prove Theorem 2.37 a new way.
Definition 4.63. Let G = (V, E) be a digraph with V = v
1
, . . . , v
m
and E =
e
1
, . . . , e
n
. Then the incidence matrix of G is an mn matrix A with:
(4.23) A
ij
=
_
_
0 if v
i
is not in e
j
1 if v
i
is the source of e
j
and e
j
is not a self-loop
1 if v
i
is the destination of e
j
and e
j
is not a self-loop
2 if v
i
is in e
j
and e
j
is a self-loop
Remark 4.64. The adjacency matrices of simple directed graphs (those with no self-
loops) have very useful properties, which we will come to when we study network ows. In
particular, these matrices have the property that every square sub-matrix has a determinant
that is either 1, -1 or 0. This property is called total unimodularity and it is particularly
important in the analysis of network ows.
5. Determinants, Eigenvalue and Eigenvectors
Definition 4.65 (Determinant). Let M R
nn
. The determinant of M is:
(4.24) det(A) =
Sn
sgn()
n
i=1
A
i(i)
Here S
n
represents a permutation over the set 1, . . . , n and (i) represents the value
to which i is mapped under .
Example 4.66. Consider an arbitrary 2 2 matrix:
M =
_
a b
c d
_
There are only two permutations in the set S
2
: the identity permutation (which is even) and
the transposition (1, 2) which is odd. Thus, we have:
det(M) =
a b
c d
= M
11
M
22
M
12
M
21
= ad bc
This is the formula that one would expect from a course in matrices (like Math 220).
Definition 4.67 (Eigenvalue and (Right) Eigenvector). Let M R
nn
. An eigenvalue,
eigenvector pair (, x) is a scalar and n 1 vector such that:
(4.25) Mx = x
50
Remark 4.68. A left eigenvector is dened analogously with x
T
M = x
T
, when x is
considered a column vector. We will deal exclusively with right eigenvectors and hence when
we say eigenvector we mean a right eigenvector.
Definition 4.69 (Characteristic Polynomial). If M R
nn
then its characteristic poly-
nomial is:
(4.26) det (I
n
M)
Remark 4.70. The following theorem is useful for computing eigenvalues of small ma-
trices and denes the characteristic polynomial for a matrix. Its proof is outside the scope
of these notes, but would occur in a Math 436 class. (See Chapter 8.2 of [Lan87].)
Theorem 4.71. A value is an eigenvalue for M R
nn
if and only if it satises the
characteristic equation:
det (I
n
M) = 0
Furthermore, M and M
T
share eigenvalues.
Example 4.72. Consider the matrix:
M =
_
1 0
0 2
_
The characteristic polynomial is computed as:
det (I
n
M) =
1 0
0 2
= ( 1)( 2) 0 = 0
Thus the characteristic polynomial for this matrix is:
(4.27)
2
3 + 2
The roots of this polynomial are
1
= 1 and
2
= 2. Using these eigenvalues, we can compute
eigenvectors:
x
1
=
_
1
0
_
(4.28)
x
2
=
_
0
1
_
(4.29)
and observe that:
(4.30) Mx
1
=
_
1 0
0 2
_ _
1
0
_
= 1
_
1
0
_
=
1
x
1
and
(4.31) Mx
2
=
_
1 0
0 2
_ _
0
1
_
= 2
_
0
1
_
2
x
2
as required. Computation of eigenvalues and eigenvectors is usually accomplished by com-
puter and several algorithms have been developed. Those interested readers should consult
(e.g.) [Dat95].
51
Example 4.73. You can use Matlab to compute the eigenvalues of a matrix using the
eig command. The same command can also return the eigenvalues (as the diagonals of
a matrix) and the corresponding eigenvectors in a second matrix. An example is shown
in Figure 4.5. This command will return the eigenvalues when used as: d = eig(A) and
Figure 4.5. Computing the eigenvalues and eigenvectors of a matrix in Matlab can
be accomplished with the eig command. This command will return the eigenvalues
when used as: d = eig(A) and the eigenvalues and eigenvectors when used as [V
D] = eig(A). The eigenvectors are the columns of the matrix V.
the eigenvalues and eigenvectors when used as [V D] = eig(A). The eigenvectors are the
columns of the matrix V.
Remark 4.74. It is important to remember that eigenvectors are unique up to scale.
That is, if M is a square matrix and (, x) is an eigenvalue eigenvector pair for M, then so
is (, x) for ,= 0. This is because:
(4.32) Mx = x = M(x) = (x)
Definition 4.75 (Degenerate Eigenvalue). An eigenvalue is degenerate if it is a multiple
root of the characteristic polynomial. The multiplicity of the root is the multiplicity of the
eigenvalue.
Example 4.76. Consider the identify matrix I
2
. It has characteristic polynomial (1)
2
,
which has one multiple root 1. Thus = 1 is a degenerate eigenvalue for this matrix.
However, this matrix does have two eigenvectors [1 0]
T
and [0 1]
T
.
52
Remark 4.77. The theory of eigenvalues and eigenvectors of matrices is deep and well
understood. A substantial part of this theory should be covered in Math 436, for those
interested. We will use only a few result in our study of graphs. The following results are
proved in Chapter 8 of [GR01]. Unfortunately, the proofs are well outside the scope of the
class.
Theorem 4.78 (Spectral Theorem for Real Symmetric Matrices). Let M R
nn
be a
symmetric matrix. Then the eigenvalues of M are all real.
6. Properties of the Eigenvalues of the Adjacency Matrix
Lemma 4.79. Let a
n
x
n
+ +a
1
x+a
0
for x = p/q with gcd(p, q) = 1 and a
n
, . . . , a
0
Z.
Then p is divisible by a
0
and q is divisible by a
n
.
Theorem 4.80. Let G = (V, E) be a graph with adjacency matrix M. Then:
(1) Every eigenvalue of M is real and
(2) If is a rational eigenvalue of M, then it is integer.
Exercise 47 (Project). Prove the Spectral Theorem for Real Symmetric Matrics and
then use it to obtain Part 1 of Theorem 4.80. Then prove and apply Lemma 4.79 to prove
Part 2 of Theorem 4.80. You should discuss the proof of the Spectral Theorem for Hermitian
Matrices. [Hint: All these proofs are available in references or online, expand on these sources
in your own words.]
Remark 4.81. Two graphs that are not isomorphic can have the same set of eigenvalues.
This can be illustrated through an example that can be found in Chapter 8 of [GR01]. The
graphs are shown in Figure 4.6. We can see the two graphs are not isomorphic since there
(a) G
1
(b) G
2
Figure 4.6. Two graphs with the same eigenvalues that are not isomorphic are
illustrated.
is no vertex in Graph G
1
that has a degee of 6 unlike Vertex 7 of graph G
2
. The adjacency
53
matrices for the two graphs are:
M
1
=
_
_
0 1 0 0 1 0 0
1 0 1 0 0 0 1
0 1 0 1 0 1 1
0 0 1 0 1 1 1
1 0 0 1 0 1 0
0 0 1 1 1 0 1
0 1 1 1 0 1 0
_
_
M
2
=
_
_
0 1 0 0 0 1 1
1 0 1 0 0 0 1
0 1 0 1 0 0 1
0 0 1 0 1 0 1
0 0 0 1 0 1 1
1 0 0 0 1 0 1
1 1 1 1 1 1 0
_
_
However, (using a computer) one can determine that they share the same set of eigenvalues:
2, 1
7, 1 +
7, 1, 1, 1, 1
Definition 4.82 (Irreducible Matrix). A matrix M R
nn
is irreducible if for each
(i, j) pair, there is some k Z with k > 0 so that M
k
ij
> 0.
Lemma 4.83. If G = (V, E) is a connected graph with adjacency matrix M, then M is
irreducible.
Exercise 48. Prove Lemma 4.83.
Theorem 4.84 (Perron-Frobenius Theorem). If M is an irreducible matrix, then M has
an eigenvalue
0
with the following properties:
(1) The eigenvalue
0
is positive and if is an alternative eigenvalue of M, then
0
[[,
(2) The matrix M has an eigenvectors v
0
corresponding to
0
with only positive entries
when properly scaled,
(3) The eigenvalue
0
is a simple root of the characteristic equation for M and therefore
has a unique (up to scale) eigenvector v
0
.
(4) The eigenvector v
0
is the only eigenvector of M that can have all positive entries
when properly scaled.
Remark 4.85. The Perron-Frobenius theorem is a classical result in Linear Algebra with
several proofs (see [Mey01]). Meyer says of the theorem,
In addition to saying something useful, the PerronFrobenius theory is el-
egant. It is a testament to the fact that beautiful mathematics eventually
tends to be useful, and useful mathematics eventually tends to be beautiful.
54
One should note that you can say more than we state in this version of the Perron-Frobenius
Theorem. See Chapter 8 of [Mey01] for details and a proof.
Corollary 4.86. If G = (V, E) is a connected graph with adjacency matrix M, then
it has a unique largest eigenvalue which corresponds to an eigenvector that is positive when
properly scaled.
Proof. Applying Lemma 4.83 we see that M is irreducible. Further, we know that
there is an eigenvalue
0
of M that is (i) greater than or equal to in absolute value all other
eigenvalues of M and (ii) a simple root. From Theorem 4.80, we know that all eigenvalues
of M are real. But for (i) and (ii) to hold, no other (real) eigenvalue can have value equal
to
0
(otherwise it would not be a simple root). Thus,
0
is the unique largest eigenvalue of
M. This completes the proof.
55
CHAPTER 5
Applications of Algebraic Graph Theory: Eigenvector Centrality
and Page-Rank
Remark 5.1. In this chapter, were going to explore two applications of Algebraic Graph
Theory: Eigenvector Centrality and Page-Rank. The goal is to devise ways for ranking the
vertices of a graph. This topic is actually very important. Google uses Page-Rank to rank
the search results they return. Social scientists have used eigenvector centrality as a way of
determining leaders in organizations. Well rst review a key set of denitions from Linear
Algebra and then discuss eigenvector centrality. We then move on to Markov chains and
Page-Rank.
1. Basis of R
n
Remark 5.2. We will be considering the eld R and vectors dened over it. By this, we
mean n 1 matrices, which are just column vectors. Therefore, by a vector in R
n
we really
mean a matrix x R
n1
.
Definition 5.3. Let x
1
, . . . , x
m
be vectors in R
n
and let
1
, . . . ,
m
R be scalars.
Then
(5.1)
1
x
1
+ +
m
x
m
is a linear combination of the vectors x
1
, . . . , x
m
.
Definition 5.4 (Span). Let A = x
1
, . . . , x
m
be a set of vectors in R
n
, then the
span of A is the set:
(5.2) span(A) = y R
n
[y is a linear combination of vectors in A
Definition 5.5 (Linear Independence). Let x
1
, . . . , x
m
be vectors in R
n
. The vectors
x
1
, . . . , x
m
are linearly dependent if there exists
1
, . . . ,
m
R, not all zero, such that
(5.3)
1
x
1
+ +
m
x
m
= 0
If the set of vectors x
1
, . . . , x
m
is not linearly dependent, then they are linearly independent
and Equation 5.3 holds just in case
i
= 0 for all i = 1, . . . , n.
Example 5.6. In R
3
, consider the vectors:
x
1
=
_
_
1
1
0
_
_
, x
2
=
_
_
1
0
1
_
_
, x
3
=
_
_
0
1
1
_
_
We can show these vectors are linearly independent: Suppose there are values
1
,
2
,
3
R
such that
1
x
1
+
2
x
2
+
3
x
3
= 0
57
Then:
_
_
1
0
_
_
+
_
_
2
0
2
_
_
_
_
0
3
_
_
=
_
_
1
+
2
1
+
3
2
+
3
_
_
=
_
_
0
0
0
_
_
Thus we have the system of linear equations:
1
+
2
= 0
1
+
3
= 0
2
+
3
= 0
Solving this problem yields the unique solution:
1
=
2
=
3
= 0. Thus these vectors are
linearly independent.
Definition 5.7 (Basis). Let A = x
1
, . . . , x
m
be a set of vectors in R
n
. The set A is
called a basis of R
n
if A is a linearly independent set of vectors and every vector in R
n
is
in the span of A. That is, for any vector w R
n
we can nd scalar values
1
, . . . ,
m
such
that
(5.4) w =
m
i=1
i
x
i
Example 5.8. We can show that the vectors:
x
1
=
_
_
1
1
0
_
_
, x
2
=
_
_
1
0
1
_
_
, x
3
=
_
_
0
1
1
_
_
form a basis of R
3
. We already know that the vectors are linearly independent. To show
that R
3
is in their span, chose an arbitrary vector in R
m
: [a, b, c]
T
. Then we hope to nd
coecients
1
,
2
and
3
so that:
1
x
1
+
2
x
2
+
3
x
3
=
_
_
a
b
c
_
_
Expanding this, we must nd
1
,
2
and
3
so that:
_
_
1
0
_
_
+
_
_
2
0
2
_
_
+
_
_
0
3
_
_
=
_
_
a
b
c
_
_
This problem can be solved in terms of a, b, and c to yield:
1
= 1/2 a + 1/2 b 1/2 c
2
= 1/2 b + 1/2 a + 1/2 c
3
= 1/2 c + 1/2 b 1/2 a
which clearly has a solution for all a, b, and c.
Remark 5.9. The following theorem on the size of a basis in R
n
is outside the scope of
this course. A proof can be found in [Lan87].
Theorem 5.10. If A is a basis of R
n
, then A contains precisely n vectors.
58
2. Eigenvector Centrality
Remark 5.11. This approach to justifying eigenvector centrality comes from Leo Spizzirri
[Spi11]. It is reasonably nice, and fairly rigorous. This is not meant to be anymore than
a justication. It is not a proof of correctness. Before proceeding, we require a theorem,
whose proof is outside the scope of the course. (See [GR01], Chapter 8, Page 170.)
Theorem 5.12 (Principle Axis Theorem). Let M R
nn
be a symmetric matrix. Then
R
n
has a basis consisting of eigenvectors of M.
Remark 5.13 (Eigenvector Centrality). We can assign to each vertex of a graph G =
(V, E) a score (called its eigenvector centrality) that will determine its relative importance
in the graph. Here importance it measured in a self-referential way: important vertices
are important precisely because they are adjacent to other important vertices. This self-
referential denition can be resolved in the following way.
Let x
i
be the (unknown) score of vertex v
i
V and let x
i
= (v
i
) with being the
function returning the score of each vertex in V . We may dene x
i
as a pseudo-average of
the scores of its neighbors. That is, we may write:
(5.5) x
i
=
1
vN(v
i
)
(v)
Here will be chosen endogenously during computation.
Recall that M
i
is the i
th
row of the adjacency matrix M and contains a 1 in position j
if and only if v
i
is adjacent to v
j
; that is to say v
j
N(v
i
). Thus we can rewrite Equation
5.5 as:
x
i
=
1
j=1
M
ij
x
j
This leads to n equations, one for vertex in V (or each row of M). Written as a matrix
expression we have:
(5.6) x =
1
Mx = x = Mx
Thus x is an eigenvector of M and is its eigenvalue.
Clearly, there may be several eigenvectors and eigenvalues for M. The question is, which
eigenvalue / eigenvector pair should be chosen? The answer is to choose the eigenvector with
all positive entries corresponding to the largest eigenvalue. We know such an eigenvalue /
eigenvector pair exists and is unique as a result of the Perron-Frobenius Theorem and Lemma
4.83.
Theorem 5.14. Let G = (V, E) be a connected graph with adjacency matrix M R
nn
.
Suppose that
0
is the largest real eigenvalue of M and has corresponding eigenvalue v
0
. If
x R
n1
is a column vector so that x v
0
,= 0, then
(5.7) lim
k
M
k
x
k
0
=
0
v
0
59
Proof. Applying Theorem 5.12 we see that the eigenvectors of M must form a basis for
R
n
. Thus, we can express:
(5.8) x =
0
v
0
+
1
v
1
+ +
n1
v
n1
Multiplying both sides by M
k
yields:
(5.9) M
k
x =
0
M
k
v
0
+
1
M
k
v
1
+ +
n1
M
k
v
n1
=
0
k
0
v
0
+
1
k
1
v
1
+ +
n1
k
n
v
n1
because M
k
v
i
=
k
i
v
i
for any eigenvalue v
i
. Dividing by
k
0
yields:
(5.10)
M
k
x
k
0
=
0
v
0
+
1
k
1
k
0
v
1
+ +
n1
k
n1
k
0
v
n1
Applying the Perron-Frobenius Theorem (and Lemma 4.83) we see that
0
is greater than
the absolute value of any other eigenvalue and thus we have:
(5.11) lim
k
k
i
k
0
= 0
for i ,= 0. Thus:
(5.12) lim
k
M
k
x
k
0
=
0
v
0
Remark 5.15. We can use Theorem 5.14 to justify our denition of eigenvector centrality
as the eigenvector corresponding to the largest eigenvalue. Let x be a vector with a 1 at
index i and 0 everywhere else. This vector corresponds to beginning at vertex v
i
in graph
G with n vertices. If M is the adjacency matrix, then Mx is the i
th
column of M whose
j
th
index tells us the number of walks of length 1 leading from vertex v
j
to vertex v
i
and by
symmetry the number of walks leading from vertex v
i
to vertex v
j
. We can repeat this logic
to see that M
k
x gives us a vector of whose j
th
element is the number of walks of length k
from v
i
to v
j
. Note for the remainder of this discussion, we will exploit the symmetry that
the (i, j) element of M
k
is both the number of walks from i to j and the number of walks
from j to i.
From Theorem 5.14 we know that no matter which vertex we choose in creating x that:
(5.13) lim
k
M
k
x
0
=
0
v
0
Reinterpreting Equation 5.13 we observe that as k , M
k
x will converge to some multiple
of the eigenvector corresponding to the eigenvalue
0
. That is, the eigenvector corresponding
to the largest eigenvalue is a multiple of the number of walks of length k leading from some
initial vertex i, since the Perron-Frobeinus eigenvector is unique (up to a scale).
Now, suppose that we were going to choose one of these (many) walks of length k at
random, we might ask what is the probability of arriving at vertex j given that we started
at vertex i. If we let y = M
k
x then the total number of walks leading from i is:
(5.14) T =
n
j=1
y
j
60
The probability of ending up at vertex j then is the j
th
element of the vector
1
T
M
k
x. Put
more intuitively, if we start at a vertex and begin wandering along the edges of the graph at
random, the probability of ending up at vertex j after k steps is the j
th
element of
1
T
M
k
x.
If we let: [x[
1
be the sum of the components of a vector x, then from Equation 5.13 we
may deduce:
(5.15) lim
k
M
k
x
[M
k
x[
1
=
v
0
[v
0
[
1
Thus, eigenvector centrality tells us a kind of probability of ending up at a specic vertex
j assuming we are allowed to wander along the edges of a graph (from any starting vertex)
for an innite amount of time. Thus, the more central a vertex is, the more likely we will
arrive at it as we move through the edges of the graph. More importantly, we see the
self-referential nature of eigenvector centrality: the more likely we are to arrive at a given
vertex after walking along the edges of a graph, the more likely we are to arrive at one of its
neighbors. Thus, eigenvector centrality is a legitimate measure of vertex importance if one
wishes to measure the chances of ending up at a vertex when wondering around a graph.
We will discuss this type of model further when we investigate random walks on graph.
Example 5.16. Consider the graph shown in Figure 5.1. Recall from Example 4.58 this
1 2
3 4
Figure 5.1. A matrix with 4 vertices and 5 edges. Intuitively, vertices 1 and 4
should have the same eigenvector centrality score as vertices 2 and 3.
graph had adjacency matrix:
M =
_
_
0 1 1 1
1 0 0 1
1 0 0 1
1 1 1 0
_
_
We can use a computer to determine the eigenvalues and eigenvectors of M. The eigenvalues
are:
_
0, 1,
1
2
+
1
2
17,
1
2
1
2
17,
_
while the corresponding oating point approximations of the eigenvalues the columns of the
matrix:
_
_
0.0 1.0 1.0 1.000000001
1.0 0.0 0.7807764064 1.280776407
1.0 0.0 0.7807764069 1.280776408
0.0 1.0 1.0 1.0
_
_
61
The largest eigenvalue is
0
=
1
2
+
1
2
_
1.0
0.7807764064
0.7807764064
1.0
_
_
We can normalize this vector to be:
v
0
=
_
_
0.2807764065
0.2192235937
0.2192235937
0.2807764065
_
_
Illustrating that vertices 1 and 4 have identical (larger) eigenvector centrality scores and
vertices 2 and 3 have identical (smaller) eigenvector centrality scores. By way of comparison,
consider the vector:
x =
_
_
1
0
0
0
_
_
We consider M
k
x/[M
k
x[
1
for various values of k:
M
1
x
[M
1
x[
1
=
_
_
0.0
0.3333333333
0.3333333333
0.3333333333
_
_
M
10
x
[M
10
x[
1
=
_
_
0.2822190823
0.2178181007
0.2178181007
0.2821447163
_
_
M
20
x
[M
20
x[
1
=
_
_
0.2807863651
0.2192136380
0.2192136380
0.2807863590
_
_
M
40
x
[M
40
x[
1
=
_
_
0.2807764069
0.2192235931
0.2192235931
0.2807764069
_
_
Its easy to see that as k , M
k
x/[M
k
x[
1
approaches the normalized eigenvector cen-
trality scores as we expected.
3. Markov Chains and Random Walks
Remark 5.17. Markov Chains are a type of directed graph in which we assign to each
edge a probability of walking along that edge given we imagine ourselves standing in a specic
vertex adjacent to the edge. Our goal is to dene Markov chains, and random walks on a
graph in reference to a Markov chain and show that some of the properties of graphs can
be used to derive interesting properties of Markov chains. Well then discuss another way
of ranking vertices; this one is used (more-or-less) by Google for ranking webpages in their
search.
62
Definition 5.18 (Markov Chain). A discrete time Markov Chain is a tuple / = (G, p)
where G = (V, E) is a directed graph and the set of vertices is usually referred to as the
states, the set of edges are called the transitions and p : E [0, 1] is a probability assignment
function satisfying:
(5.16)
No(v)
p(v, v
) = 1
for all v V . Here, N
o
(v) is the neighborhood reachable by out-edge from v. If there is no
edge (v, v
) E then p(v, v
) = 0.
Remark 5.19. There are continuous time Markov chains, but these are not in the scope
of these notes. When we say Markov chain, we mean discrete time Markov chain.
Example 5.20. A simple Markov chain is shown in Figure 5.2. We can think of a
Markov chain as governing the evolution of state as follows. Think of the states as cities
with airports. If there is an out-edge connecting the current city to another city, then we
can y from our current city to this next city and we do so with some probability. When
we do y (or perhaps dont y and remain at the current location) our state updates to the
next city. In this case, time is treated discretely.
1 2
1
2
1
7
6
7
1
2
Figure 5.2. A Markov chain is a directed graph to which we assign edge proba-
bilities so that the sum of the probabilities of the out-edges at any vertex is always
1.
A walk along the vertices of a Markov chain governed by the probability function is called
a random walk.
Definition 5.21 (Stochastic Matrix). Let / = (G, p) be a Markov chain. Then the
stochastic matrix (or probability transition matrix) of / is:
(5.17) M
ij
= p(v
i
, v
j
)
Example 5.22. The stochastic matrix for the Markov chain in Figure 5.2 is:
M =
_
1
2
1
2
1
7
6
7
_
Thus a stochastic matrix is very much like an adjacency matrix where the 0s and 1s indi-
cating the presence or absence of an edge are replaced by the probabilities associated to the
edges in the Markov chain.
Definition 5.23 (State Probability Vector). If / = (G, p) be a Markov chain with n
states (vertices) then a state probability vector is a vector x R
n1
such that x
1
+x
2
+ +
x
n
= 1 and x
i
0 for i = 1, . . . , n and x
i
represents the probability that we are in state i
(at vertex i).
63
Remark 5.24. The next theorem can be proved in exactly the same way that Theorem
4.57 is proved.
Theorem 5.25. Let / = (G, p) be a Markov chain with n states (vertices). Let x
(0)
R
n1
be an (initial) state probability vector. Then assuming we take a random walk of length
k in / using initial state probability vector x
(0)
, the nal state probability vector is:
(5.18) x
(k)
=
_
M
T
_
k
x
(0)
Remark 5.26. If you prefer to remove the transpose, you can write x
(0)
R
1n
; that is,
x
(0)
is a row vector. Then:
(5.19) x
(k)
= x
(0)
M
k
with x
(k)
R
1n
.
Exercise 49. Prove Theorem 5.25. [Hint: Use the same inductive argument from the
proof of Theorem 4.57.]
Example 5.27. Consider the Markov chain in Figure 5.2. The state vector:
x
(0)
=
_
1
0
_
states that we will start in State 1 with probability 1. From Example 5.22 we know what
M is. Then it is easy to see that:
x
(1)
=
_
M
T
_
k
x
(0)
=
_
1
2
1
2
Which is precisely the state probability vector we would expect after a random walk of length
1 in /.
Definition 5.28 (Stationary Vector). Let / = (G, p) be a Markov chain. Then a
vector x
is stationary for / if
(5.20) x
= M
T
x
j
M
ij
= 1
for all i = 1, . . . , n.
Exercise 50. Prove Lemma 5.30.
Lemma 5.31. / = (G, p) be a Markov chain with n states and with stochastic matrix
M. If G is strongly connected, then M and M
T
are irreducible.
64
Proof. If G is strongly connected, then there is a directed walk from any vertex v
i
to
any other vertex v
j
in V , the vertex set of G. Consider any length k walk connecting v
i
to
v
j
(such a walk exists for some k). Let e
i
be the vector with 1 in its i
th
component and 0
everywhere else. Then (M
T
)
k
e
i
is the nal state probability vector associated with a walk
of length k starting at vertex v
i
. Since there is a walk of length k from v
i
to v
j
, we know
that the j
th
element of this vector must be non-zero. That is:
e
T
j
(M
T
)
k
e
i
> 0
where e
j
is dened just as e
i
is but with the 1 at the j
th
position. Thus, (M
T
)
k
ij
> 0 for
some k for every (i, j) pair and thus M
T
is irreducible. The fact that M is irreducible follows
immediately from the fact that (M
T
)
k
= (M
k
)
T
. This completes the proof.
Theorem 5.32 (Perron-Frobenius Theorem Redux). If M is an irreducible matrix, then
M has an eigenvalue
0
with the following properties:
(1) The eigenvalue
0
is positive and if is an alternative eigenvalue of M, then
0
[[,
(2) The matrix M has an eigenvectors v
0
corresponding to
0
with only positive entries,
(3) The eigenvalue is a simple root of the characteristic equation for M and therefore
has a unique (up to scale) eigenvectors v
0
.
(4) The eigenvector v
0
is the only eigenvector of M that can have all positive entries
when properly scaled.
(5) The following inequalities hold:
min
i
j
M
ij
0
max
i
j
M
ij
Theorem 5.33. Let / = (G, p) be a Markov chain with stochastic matrix M, M
T
, is
irreducible then / has a unique stationary probability distribution.
Proof. From Theorem 4.71 we know that M and M
T
have identical eigenvalues. By
the Perron-Frobenius theorem, M has a largest positive eigenvalue
0
that satises:
min
i
j
M
ij
0
max
i
j
M
ij
By Lemma 5.30, we know that:
min
i
j
M
ij
= max
i
j
M
ij
= 1
Therefore, by the squeezing lemma
0
= 1. The fact that M
T
has exactly one strictly
positive eigenvector v
0
corresponding to
0
= 1 means that:
(5.22) M
T
v
0
= v
0
Thus v
0
is the unique stationary state probability vector for / = (G, p). This completes
the proof.
65
4. Page Rank
Definition 5.34 (Induced Markov Chain). Let G = (V, E) be a graph. Then the induced
Markov chain from G is the one obtained by dening a new directed graph G
= (V, E
) with
each edge v, v
) and (v
, v) in E and dening
the probability function p so that:
(5.23) p(v, v
) =
1
deg
out
G
v
Example 5.35. An induced Markov chain is shown in Figure 5.3. The Markov chain in
1
2 3
4 1
2 3
4
1/3
1/3
1/3
1
1/2
1/2
1/2
1/2
Original Graph Induced Markov Chain
Figure 5.3. An induced Markov chain is constructed from a graph by replacing
every edge with a pair of directed edges (going in opposite directions) and assigning
a probability equal to the out-degree of each vertex to every edge leaving that vertex.
the gure has the stationary state probability vector:
x
=
_
_
3
8
2
8
2
8
1
8
_
_
which is the eigenvector corresponding to the eigenvalue 1 in the matrix M
T
. Arguing as
we did in the proof of Theorem 5.14 and Example 5.16, we could expect that for any state
vector x we would have:
lim
k
_
M
T
_
k
x = x
and we would be correct. When this convergence happens quickly (where we leave quickly
poorly dened) the graph is said to have a fast mixing property.
If we used the stationary probability of a vertex in the induced Markov chain as a measure
of importance, then clearly vertex 1 would be most important followed by vertices 2 and 3
and lastly vertex 4. We can compare this with the eigenvector centrality measure, which
assigns a rank vector of:
x
+
=
_
_
0.3154488065
0.2695944375
0.2695944375
0.1453623195
_
_
Thus eigenvector centrality gives the same ordinal ranking as using the stationary state
probability vector, but there are subtle dierences in the values produced by these two
ranking schemes. This leads us to PageRank [BP98].
66
Remark 5.36. Consider a collection of web pages each with links. We can construct a
directed graph G with the vertex set V consisting of the we web pages and E consisting of
the directed links among the pages. Imagine a random web surfer who will click among these
web pages following links until a dead-end is reached (a page with no outbound links). In
this case, the web surfer will type a new URL in (chosen from the set of web pages available)
and the process will continue.
From this model, we can induce a Markov chain in which we dene a new graph G
with
edge set E
to every other
vertex in V and we then dene:
(5.24) p(v, v
) =
1
deg
out
G
v
exactly as before. In the absence of any further insight, the PageRank algorithm simply
assigns to each web page a score equal to the stationary probability of its state in the
induced Markov chain. For the remainder of this remark, let M be the stochastic matrix of
the induced Markov chain.
In general, however, PageRank assumes that surfers will get bored after some number
of clicks (or new URLs) and will stop (and move to a new page) with some probability
d [0, 1] called the damping factor. This factor is usually estimated. Assuming there are n
web pages, let r R
n1
be the PageRank score for each page. Taking boredom into account
leads to a new expression for rank (similar to Equation 5.5 for Eigenvector centrality):
(5.25) r
i
=
1 d
n
+d
_
n
j=1
M
ji
r
j
_
for i = 1, . . . , n
Here the d term acts like a damping factor on walks through the Markov chain. In essence,
it stalls people as they walk, making it less likely a searcher will keep walking forever. The
original System of Equations 5.25 can be written in matrix form as:
(5.26) r =
_
1 d
n
_
1 +dM
T
r
where 1 is a n 1 vector consisting of all 1s. It is easy to see that when d = 1 r is precisely
the stationary state probability vector for the induced Markov chain. When d ,= 1, r is
usually computed iteratively by starting with an initial value of r
0
i
= 1/n for all i = 1, . . . , n
and computing:
r
(k)
=
_
1 d
n
_
1 +dM
T
r
(k1)
The reason is that for large n, the analytic solution:
(5.27) r =
_
I
n
dM
T
_
1
_
1 d
n
_
1
is not computationally tractable
1
.
1
Note,
_
I
n
dM
T
_
1
computes a matrix inverse, which we reviewed briey in Chapter 4. We should
note that for stochastic matrices, this inverse is guaranteed to exist. For those interested, please consult and
of [Dat95, Lan87, Mey01].
67
Example 5.37. Consider the induced Markov chain in Figure 5.3 and suppose we wish
to compute PageRank on these vertices with d = 0.85 (which is a common assumption). We
might begin with:
r
(0)
=
_
_
1
4
1
4
1
4
1
4
_
_
We would then compute:
r
(1)
=
_
1 d
n
_
1 +dM
T
r
(0)
=
_
_
0.462499999999999967
0.214583333333333320
0.214583333333333320
0.108333333333333337
_
_
We would repeat this again to obtain:
r
(2)
=
_
1 d
n
_
1 +dM
T
r
(1)
=
_
_
0.311979166666666641
0.259739583333333302
0.259739583333333302
0.168541666666666673
_
_
This would continue until the dierence between in the values of r
(k)
and r
(k1)
was small.
The nal solution would be close to the exact solution:
r
=
_
_
0.366735867135100591
0.245927818588310476
0.245927818588310393
0.141408495688278513
_
_
Note this is (again) very close to the stationary probabilities and the eigenvector centralities
we observed earlier. This vector is normalized so that all the entries sum to 1.
Exercise 51. Consider the Markov chain shown below:
1
2 3
4
1/3
1/3
1/3
1/2
1/2
1/2
1/2
1/2
1/2
Suppose this is the induced Markov chain from 4 web pages. Compute the page-rank of
these web pages using d = 0.85.
Exercise 52. Find an expression for r
(2)
in terms of r
(0)
. Explain how the damping
factor occurs and how it decreases the chance of taking long walks through the induced
Markov chain. Can you generalize your expression for r
(2)
to an expression for r
(k)
in terms
of r
(0)
?
68
CHAPTER 6
Trees, Algorithms and Matroids
1. Two Tree Search Algorithms
Remark 6.1. Tree searching is the process of enumerating the vertices of a tree (for the
purpose of searching them). One can consider this process as generating a walk hitting all
the vertices of the tree at least once or as a way to create a sequence of the vertices. In this
section, we will take the latter view, though it will be clear how to create a walk as well.
The rst algorithm, called Breadth First Search, explores the vertices of a tree starting
from a given vertex by exploring all the neighbors of this given vertex, then the neighbors
of the neighbors and so on until all the vertices have been encountered.
Breadth First Search
Input: T = (V, E) a tree, v
0
a starting vertex
Initialize: F = (v
0
) A sequence of vertices to enumerate.
Initialize: F
next
= () The sequence of next vertices to enumerate.
Initialize: w = () The sequence of vertices traversed.
(1) while F ,=
(2) for each v F do
(3) Remove v from F
(4) Append v to w
(5) for each v
V do
(6) if v
, w and v, v
E then
(7) Append v
to F
next
(8) end if
(9) end for
(10) end for
(11) F = F
next
(12) F
next
= ()
(13) end while
Output: w A breadth rst sequence of vertices in T.
Algorithm 1. Breadth First Search
Example 6.2. Figure 6.1 shows the order the vertices are added to w during a breadth
rst search of the tree.
Proposition 6.3. A breadth rst search of a tree T = (V, E) enumerates all vertices in
w.
Proof. We proceed by induction. If T has one vertex, then clearly v
0
in the algorithm
is that vertex. The vertex is added to w in the rst iteration of the while loop at Line 1 and
69
a
b c
d e
1
2 3
4 5
F
next
= (b, c)
Figure 6.1. The breadth rst walk of a tree explores the tree in an ever widening
pattern.
F
next
is the empty set, thus the algorithm terminates. Now suppose that the statement is
true for all trees with at most n vertices. We will show the statement is true for a tree with
n +1 vertices. To see this, construct a new tree T
from
T. Clearly the algorithm must enumerate every vertex in T
in T. At this point, v
would be added to F
next
and it would be added to w in the next execution through the while
loop since F ,= the next time. Thus, every vertex of T must be enumerated in w. This
completes the proof.
Remark 6.4. Breadth First Search can be modied for directed trees in the obvious way.
Necessarily, we need v
0
to be strongly connected to every other vertex in order to ensure
that BFS enumerates every possible vertex.
Remark 6.5. Another algorithm for enumerating the vertices of a tree is the depth rst
search algorithm. This algorithm works by descending into the tree as deeply as possible
(until a leaf is identied) and then working back up. We present Depth First Search as a
recursive algorithm.
Depth First Search
Input: T = (V, E) a tree, v
0
a starting vertex
Initialize: v
now
= v
0
The current vertex.
Initialize: w = (v
0
) The sequence of next vertices to enumerate.
Initialize: v
last
= The sequence of vertices traversed.
(1) Recurse(T, v
now
, v
last
, w)
Output: w A breadth rst sequence of vertices in T.
Recurse
Input: T = (V, E) a tree, v
now
current vertex, v
last
the last vertex, w the sequence
(1) for each v V do
(2) if v
now
, v E and v , w then
(3) Append v to w
(4) Recurse(T, v, v
now
, w)
(5) end if
(6) end for
Algorithm 2. Depth First Search
70
1
2 5
3 4
Figure 6.2. The depth rst walk of a tree explores the tree in an ever deepening
pattern.
Example 6.6. Figure 6.2 shows the order the vertices are added to w during a depth
rst search of the tree.
Proposition 6.7. A depth rst search of a tree T = (V, E) enumerates all vertices in
w.
Exercise 53. Prove proposition 6.7. [Hint: The proof is almost identical to the proof
for Breadth First Search.]
Remark 6.8. We note that breadth and depth rst search can be trivially modied to
search through connected graph structures and construct spanning trees for these graphs.
We also note that BFS and DFS can be modied to function on directed trees (and graphs)
and that all vertices will be enumerated provided that every vertex is reachable by a directed
path from v
0
.
Remark 6.9. In terms of implementation, we note that the recursive implementation of
Depth First Search works on most computing systems provided the graph your are searching
has a longest path of at most some specied value. This is because most operating systems
prevent a recursive algorithm from making any more than a specied number of recursion
calls.
Exercise 54. Rewrite the breadth rst search algorithm to produce a spanning tree of
a connected graph.
2. Prims Spanning Tree Algorithm
Definition 6.10 (Weighted Graph). A weighted graph is a pair (G, w) where G = (V, E)
is a graph and w : E R is a weight function.
Remark 6.11. A weighted digraph is dened analagously. A Markov chain is an example
of a weighted graph where the probabilities play the role of the edge weights.
Example 6.12. Consider the graph shown in Figure 6.3. A weighted graph is simply
a graph with a real number (the weight) assigned to each edge. Weighted graphs arise in
several instances, such as travel planning, communications and military planning.
Remark 6.13. Any graph can be thought of as a weighted graph in which we assign the
weight of 1 to each edge. The distance between two vertices in a graph can then easily be
generalized in a weighted graph. If p = (v
1
, e
1
, v
2
, . . . , v
n
, e
n
, v
n+1
) is a path, then the weight
71
3
4
-1 6
7
3
5
Figure 6.3. A weighted graph is simply a graph with a real number (the weight)
assigned to each edge.
of the path is:
n
i=1
w(e
i
)
Thus in a weighted graph, the distance between two vertices v
1
and v
2
is the weight of the
weight of the least weight path connecting v
1
and v
2
. We study the problem of nding this
distance in Section 5.
Definition 6.14 ((Sub)Graph Weight). Let (G, w) be a weighted graph with G = (V, E).
If H = (V
, E
eE
w(e)
Definition 6.15 (Minimum Spanning Forrest Problem). Let (G, w) be a weighted graph
with G = (V, E). The minimum spanning forest problem for G is to nd a forest F = (V
, E
)
that is a spanning subgraph of G that has the smallest possible weight.
Remark 6.16. If (G, w) is a weighted graph and G is connected, then the minimum
spanning forest problem becomes the minimum spanning tree problem.
Example 6.17. A minimum spanning tree for the weighted graph shown in Figure 6.3 is
shown in Figure 6.4. In the minimum spanning tree problem, we attempt to nd a spanning
3
4
-1 6
7
3
5
Figure 6.4. In the minimum spanning tree problem, we attempt to nd a spanning
subgraph of a graph G that is a tree and has minimal weight (among all spanning
trees).
subgraph of a graph G that is a tree and has minimal weight (among all spanning trees).
72
We will verify that the proposed spanning tree is minimal when we derive algorithms for
constructing a minimum spanning forest.
Remark 6.18. The next algorithm, commonly called Prims Algorithm [Pri57] will
construct a minimum spanning tree for a connected graph.
Prims Algorithm
Input: (G, w) a weighted connected graph with G = (V, E), v
0
a starting vertex
Initialize: E
= v
0
New vertices added to the spanning tree.
(1) while V
,= V
(2) Set X := V V
so (i) v V
; (ii) v
X and:
w(e) = min
uU,u
X
w
_
u, u
_
(4) Set E
= E
e
(5) Set V
= V
, E
that has
minimum weight. Note at Iteration 2, we could have chosen to add either edge 1, 3 or
edge 4, 6 the order doesnt matter, so any tie breaking algorithm will suce. We continue
adding edges until all vertices in the original graph are in the spanning tree.
Theorem 6.20. Let (G, w) be a weighted connected graph. Then Prims algorithm returns
a minimum spanning tree.
Proof. We will show by induction that at each iteration of Prims algorithm, the tree
(V
, E
) is a subtree of a minimum spanning tree T of (G, w). If this is the case, then at the
termination of the algorithm, (V
, E
= v
0
and E
= and
therefore (V
, E
, E
) at
iteration k. Suppose at iteration k +1 we add edge e = v, v
to T
k
to obtain T
k+1
= (U, F)
with U = V
and F = E
= E e. Suppose that T
k+1
is not a subtree of T, then e
is not an edge in T and thus e must generate a cycle in T. On this cycle, there is some edge
e
= u, u
with u V
and u
, V
from T by removing e
) w(T) thus T
= {1}
E
= {}
1
2
5
3
4
6
3
4
-1 6
7
3
5
Iteration 1
E
= {{1, 4}}
V
= {1, 4}
1
2
5
3
4
6
3
4
-1 6
7
3
5
Iteration 2
V
= {1, 3, 4}
E
= {{1, 4}, {1, 3}, {4, 6}, {1, 4}, {4, 5}}
V
= {1, 2, 3, 4, 5, 6}
1
2
5
3
4
6
3
4
-1 6
7
3
5
Iteration 5 (Stop)
Figure 6.5. Prims algorithm constructs a minimum spanning tree by successively
adding edges to an acyclic subgraph until every vertex is inside the spanning tree.
Edges with minimal weight are added at each iteration.
e
e'
e
e'
T
T
= v
0
New vertices added to the spanning tree.
(1) while V
,= V
(2) Set X := V V
(3) Set e :=
(4) Set w
=
(5) for each v V
X
(8) Let j be the row index corresponding to row v
(9) if M
ij
,= 0 and w(v, v
) < w
(10) w
= w(v, v
)
(11) e := v, v
(12) end if
(13) end for
(14) end for
(15) Set E
= E
e
(16) Set V
= V
, E
at each while loop iteration (and thus remove one vertex from X at
each while loop iteration). The for loop beginning at Line 5 will have k iterations and the
for loop starting at Line 7 will have n k iterations. This means that for any iteration of
the while loop, we will perform O(k(n k)) operations. Thus, for the whole algorithm we
will perform:
O
_
n1
k=1
k(n k)
_
= O
_
1
3
n
3
1
6
n
_
1
Algorithm 4 is not optimal. It is intentionally not optimal so that we can compute its complexity in
Theorem 6.27 easily and we do not have to appeal to special data structures. See Remark 6.28 for more on
this.
76
Thus, the running time for Algorithm 4 is O(n
3
) = O([V [
3
).
Remark 6.28. As it turns out, the implementation of Prims algorithm can have a
substantial impact on the running time. There are implementations of Prims algorithm
that run in O([V [
2
), O([E[ log([V [)) and O([E[ +[V [ log([V [)) [CLRS01]. Thus, we cannot
just say, Prims algorithm is an O(g(x)) algorithm, we must know which implementation of
Prims algorithm we are using. Clearly, the implementation in Algorithm 4 is not a very
good one. To learn how to properly implement Prims algorithm, you might consider taking
CSE 465, which covers data structures and algorithms. You can also refer to [CLRS01].
Definition 6.29 (Polynomial Time). For a specic implementation of an algorithm, its
running time is polynomial if there is some polynomial p(x) so that when the running time
of the algorithm is f(x) then f(x) O(p(x)).
Theorem 6.30. There is an implementation of Prims algorithm that is polynomial.
4. Kruskals Algorithm
Remark 6.31. In this section, we will discuss Kruskals algorithm [Kru56], an alterna-
tive way to construct a minimum spanning tree of a weighted graph (G, w). The algorithm
is shown in Algorithm 5.
Kruskals Algorithm
Input: (G, w) a weighted connected graph with G = (V, E) and n = [V [
Initialize: Q = E
Initizlize: V
= V
Initizlize: E
=
Initialize: For all v V dene C(v) := v C(v) is the set of vertices connected to v at each
iteration.
(1) while E
)
(4) for each u C(v): C(u) := C(u) C(v
)
(5) for each u C(v
:= E
e
(7) Q := Q e
(8) else
(9) Q := Q e
(10) GOTO 2
(11) end if
(12) end while
Output: T = (V
, E
. Dijkstras Algorithm
[Dij59] answers this question by growing a spanning tree starting at v so that the unique
path from v to any other vertex v
=
Initialize: For all v V if v ,= v
0
dene d(v
0
, v) := otherwise dene d(v
0
, v) := 0 d(v
0
, v) is
the best distance from v
0
to v.
Initialize: For all v V , p(v) := undened A parent function that will be used to build
the tree.
(1) while Q ,=
(2) Choose v Q so that d(v
0
, v) is minimal
(3) Q := Q v
(4) for each v
N(v)
(5) Dene (v
0
, v
) := d(v
0
, v) +w(v, v
)
(6) if (v
0
, v
) < d(v
0
, v
)
(7) d(v
0
, v
) = (v
0
, v
)
(8) p(v
) = v
(9) end if
(10) end for
(11) end while
(12) Set V
:= V
(13) Set E
=
(14) for each v V
(15) if p(v) ,= undened
(16) E
= E
v, p(v)
(17) end if
(18) end for
Output: T = (V
, E
) and d(v
0
, ) T is a Dijkstra tree, d(v
0
, ) provides the distances.
Algorithm 6. Dijkstras Algorithm (Adapted from Wikipedias Pseudo-code,
http://en.wikipedia.org/wiki/Dijkstra%27s_algorithm)
any vertex v
in (G, w).
Before proceeding note, that for any vertex v
) is xed
permanently after that iteration. Thus, the path from v
0
to v
in T
k
is the same as the path
from v
0
to v
in T
k+1
. Thus, assuming that d(v
0
, v
before reaching
v while the path recorded in T
k+1
reaches v
, v) and w
2
= w(v
) + w
1
> d(v
0
, v
) + w
2
.
By the induction hypothesis, d(v
0
, v
) and d(v
0
, v
and v
when N(v
) was evaluated,
then p(v) = v
) + w
1
<
d(v
0
, v
) +w
2
. On the other hand, if p(v) = v
when N(v
) =
eE
w(e)
is as small as possible and E
I).
Remark 6.50. One can dene a maximum weight problem in precisely the same way, if
we replace the word minimum with maximum and small with large. The following algorithm
called the Greedy Algorithm can be used (in some cases) to solve the minimum weight
problem. The name, Greedy makes a little more sense for maximum weight problems, but
the examples we give are minimum weight problems.
Remark 6.51. Let (G, w) be a weighted graph and consider the weighted hereditary
system with (E, J) with J the collection of edge subsets of E that induce acyclic graphs and
w is just the edge weighting. Kruskals Algorithm is exactly a greedy algorithm. We begin
with the complete set of edges and continue adding them to the forest (acyclic subgraph of
a given weighted graph (G, w)), each time checking to make sure that the added edge does
not induce a cycle (that is, that we have an element of J). We will use this fact to prove
Theorem 6.34.
83
Greedy Algorithm
Input: (E, J, w) a weighted hereditary system
Initizlize: E
=
Initizlize: A = E
(1) while A ,=
(2) Choose e A to minimize w(e)
(3) A := A e
(4) if E
e J
(5) E
:= E
e
(6) end if
(7) end while
Output: E
i=1
w(e
i
)
then w(I
k
) w(J
k
) for all k = 1, . . . , n. We proceed by induction. Since the Greedy
Algorithm selects the element e with smallest weight rst, it is clear that w(I
1
) w(J
1
),
thus we have established the base case. Now assume that the statement is true up through
84
some arbitrary k < n. By denition, we know that [J
k+1
[ > [I
k
[ and therefore by the
augmentation property there is some e J
k+1
with e , I
k
so that I
k
e is an element of J.
It follows that w(e
k+1
) w(e) because otherwise, the Greedy Algorithm would have chosen
e instead of e
k+1
. Furthermore, w(e) w(f
k+1
) since the elements of I and J are listed in
ascending order and e J
k+1
. Thus, w(e
k+1
) w(e) w(f
k+1
) and therefore we conclude
that w(I
k+1
) w(J
k+1
). The result follows by induction at once.
() We will proceed by contrapositive to prove that M is a matroid. Suppose that the
augmentation property is not satised and consider I and J in J with [I[ < [J[ so that there
is no element e J with e , I so that I e is in J. Without loss of generality, assume
that [I[ = [J[ + 1. Let [I[ = p and consider the following weight function:
w(e) =
_
_
p 2 if e I
p 1 if e J I
0 else
After the Greedy algorithm chooses all the elements of I, it cannot decrease the weight of the
independent set because only elements that are not in J will be added. Thus, the total weight
will be p(p + 2) p
2
2p. However, the set J has weight (p + 1)(p + 1) = p
2
2p 2.
Thus any set independent set containing J has weight at most p
2
2p 2. Thus, the
Greedy Algorithm cannot identify a maximal independent set with minimum weight when
the augmentation property is not satised. Thus by contrapositive we have shown that if
the Greedy Algorithm identies a maximal independent set with minimal weight, then M
must be a matroid. This completes the proof.
Theorem 6.55. Let G = (V, E) be a graph. Then the hereditary system M(G) = (E, J)
where J is the collection of subsets that induce acyclic graphs is a matroid.
Proof. From Proposition 6.46, we know that (E, J) is a hereditary system, we must
simply show that it has the augmentation property. To see this, let I and J be two elements
of J with [I[ < [J[. Let H be the subgraph of G induced from the edge sets I J. Let
F be a spanning forest of this subgraph H that contains I. We know from Corollary 3.66
that H has a spanning subgraph and we know that we can construct such a graph using the
technique from the proof of Theorem 3.65.
Since J is acyclic, F has at least as many edges as J and therefore there exists at least
one edge e that is in the forest F but that does not occur in the set I and furthermore, it
must be an element of J (by construction of H). Since e is an edge in F, it follows that
the subgraph induced by the set I e is acyclic and therefore, I e is an element of J.
Thus M(G) has the augmentation property and is a matroid.
Corollary 6.56 (Theorem 6.34). Let (G, w) be a weighted graph. Then Kruskals al-
gorithm returns a minimum spanning tree when G is connected.
Exercise 65. Prove Theorem 6.34.
Remark 6.57. Matroid Theory is a very active and deep area of research in combinatorics
and combinatorial optimization theory. A complete study of this eld is well outside the scope
of this course. The interested reader should consider a text on Matroid Theory like [Oxl92].
85
CHAPTER 7
A Brief Introduction to Linear Programming
Remark 7.1. It turns out the many graph theoretic problems can be expressed as linear
optimization problems. Furthermore, the proofs of some of the most fundamental theorems
of graph theory are greatly simplied by the use of a linear optimization formulation.
Even though it seems like were going to go far o topic, we use this chapter to introduce
Linear Optimization and its fundamental results. We will then use these results to prove key
results in Graph Theory and thereby illustrate the link between the theory of optimization
and the theory of graphs.
1. Linear Programming: Notation
Definition 7.2 (Linear Programming Problem). A linear programming problem is an
optimization problem of the form:
(7.1)
_
_
max z(x
1
, . . . , x
n
) = c
1
x
1
+ +c
n
x
n
s.t. a
11
x
1
+ +a
1n
x
n
b
1
.
.
.
a
m1
x
1
+ +a
mn
x
n
b
m
h
11
x
1
+ +h
n1
x
n
= r
1
.
.
.
h
l1
x
1
+ +h
ln
x
n
= r
l
Remark 7.3. You will recall from your matrices class (Math 220) that matrices can be
used as a short hand way to represent linear equations. Consider the following system of
equations:
(7.2)
_
_
a
11
x
1
+a
12
x
2
+ +a
1n
x
n
= b
1
a
21
x
1
+a
22
x
2
+ +a
2n
x
n
= b
2
.
.
.
a
m1
x
1
+a
m2
x
2
+ +a
mn
x
n
= b
m
Then we can write this in matrix notation as:
(7.3) Ax = b
where A
ij
= a
ij
for i = 1, . . . , m, j = 1, . . . , n and x is a column vector in R
n
with entries
x
j
(j = 1, . . . , n) and b is a column vector in R
m
with entries b
i
(i = 1 . . . , m). Obviously, if
87
we replace the equalities in Expression 7.2 with inequalities, we can also express systems of
inequalities in the form:
(7.4) Ax b
Using this representation, we can write our general linear programming problem using
matrix and vector notation. Expression 7.1 can be written as:
(7.5)
_
_
max z(x) =c
T
x
s.t. Ax b
Hx = r
Definition 7.4. In Problem 7.5, if we restrict some of the decision variables (the x
i
s)
to have only integer (or discrete) values, then the problem becomes a mixed integer linear
programming problem. If all of the variables are restricted to integer values, the problem is
an integer programming problem and if every variable can only take on the values 0 or 1, the
program is called a 0 1 integer programming problem. [WN99] is an excellent reference
for Integer Programming.
2. Intuitive Solutions of Linear Programming Problems
Example 7.5. Consider the problem of a toy company that produces toy planes and toy
boats. The toy company can sell its planes for $10 and its boats for $8 dollars. It costs $3
in raw materials to make a plane and $2 in raw materials to make a boat. A plane requires
3 hours to make and 1 hour to nish while a boat requires 1 hour to make and 2 hours to
nish. The toy company knows it will not sell anymore than 35 planes per week. Further,
given the number of workers, the company cannot spend anymore than 160 hours per week
nishing toys and 120 hours per week making toys. The company wishes to maximize the
prot it makes by choosing how much of each toy to produce.
We can represent the prot maximization problem of the company as a linear program-
ming problem. Let x
1
be the number of planes the company will produce and let x
2
be
the number of boats the company will produce. The prot for each plane is $10 $3 = $7
per plane and the prot for each boat is $8 $2 = $6 per boat. Thus the total prot the
company will make is:
(7.6) z(x
1
, x
2
) = 7x
1
+ 6x
2
The company can spend no more than 120 hours per week making toys and since a plane
takes 3 hours to make and a boat takes 1 hour to make we have:
(7.7) 3x
1
+x
2
120
Likewise, the company can spend no more than 160 hours per week nishing toys and since
it takes 1 hour to nish a plane and 2 hour to nish a boat we have:
(7.8) x
1
+ 2x
2
160
88
Finally, we know that x
1
35, since the company will make no more than 35 planes per
week. Thus the complete linear programming problem is given as:
(7.9)
_
_
max z(x
1
, x
2
) = 7x
1
+ 6x
2
s.t. 3x
1
+x
2
120
x
1
+ 2x
2
160
x
1
35
x
1
0
x
2
0
Remark 7.6. Strictly speaking, the linear programming problem in Example 7.5 is not a
true linear programming problem because we dont want to manufacture a fractional number
of boats or planes and therefore x
1
and x
2
must really be drawn from the integers and not
the real numbers (a requirement for a linear programming problem). This type of problem
is generally called an integer programming problem. However, we will ignore this fact and
assume that we can indeed manufacture a fractional number of boats and planes. If youre
interested in this distinction, you might consider taking Math 484, where we discuss this
issue in depth.
Linear Programs (LPs) with two variables can be solved graphically by plotting the
feasible region along with the level curves of the objective function. We will show that we
can nd a point in the feasible region that maximizes the objective function using the level
curves of the objective function. We illustrate the method rst using the problem from
Example 7.5.
Example 7.7 (Continuation of Example 7.5). Lets continue the example of the Toy
Maker begin in Example 7.5. To solve the linear programming problem graphically, begin
by drawing the feasible region. This is shown in the blue shaded region of Figure 7.1.
After plotting the feasible region, the next step is to plot the level curves of the objective
function. In our problem, the level sets will have the form:
7x
1
+ 6x
2
= c = x
2
=
7
6
x
1
+
c
6
This is a set of parallel lines with slope 7/6 and intercept c/6 where c can be varied as
needed. The level curves for various values of c are parallel lines. In Figure 7.1 they are
shown in colors ranging from red to yellow depending upon the value of c. Larger values of
c are more yellow.
To solve the linear programming problem, follow the level sets along the gradient (shown
as the black arrow) until the last level set (line) intersects the feasible region. If you are
doing this by hand, you can draw a single line of the form 7x
1
+ 6x
2
= c and then simply
draw parallel lines in the direction of the gradient (7, 6). At some point, these lines will fail
to intersect the feasible region. The last line to intersect the feasible region will do so at a
point that maximizes the prot. In this case, the point that maximizes z(x
1
, x
2
) = 7x
1
+6x
2
,
subject to the constraints given, is (x
1
, x
2
) = (16, 72).
Note the point of optimality (x
1
, x
2
) = (16, 72) is at a corner of the feasible region. This
corner is formed by the intersection of the two lines: 3x
1
+x
2
= 120 and x
1
+2x
2
= 160. In
89
x
1
= 35
(7x
1
+ 6x
2
)
x
1
+ 2x
2
= 160
3x
1
+x
2
120
x
1
+ 2x
2
160
x
1
35
x
1
0
x
2
0
3x
1
+x
2
= 120
(x
1
, x
2
) = (16, 72)
3x
1
+x
2
120
x
1
+ 2x
2
160
x
1
35
x
1
0
x
2
0
Figure 7.1. Feasible Region and Level Curves of the Objective Function: The
shaded region in the plot is the feasible region and represents the intersection of
the ve inequalities constraining the values of x
1
and x
2
. On the right, we see the
optimal solution is the last point in the feasible region that intersects a level set
as we move in the direction of increasing prot.
this case, the constraints
3x
1
+x
2
120
x
1
+ 2x
2
160
are both binding, while the other constraints are non-binding. In general, we will see that
when an optimal solution to a linear programming problem exists, it will always be at the
intersection of several binding constraints; that is, it will occur at a corner of a higher-
dimensional polyhedron.
Remark 7.8. It can sometimes happen that a linear programming problem has an innite
number of alternative optimal solutions. We illustrate this in the next example.
Example 7.9. Suppose the toy maker in Example 7.5 nds that it can sell planes for a
prot of $18 each instead of $7 each. The new linear programming problem becomes:
(7.10)
_
_
max z(x
1
, x
2
) = 18x
1
+ 6x
2
s.t. 3x
1
+x
2
120
x
1
+ 2x
2
160
x
1
35
x
1
0
x
2
0
Applying our graphical method for nding optimal solutions to linear programming problems
yields the plot shown in Figure 7.2. The level curves for the function z(x
1
, x
2
) = 18x
1
+6x
2
90
are parallel to one face of the polygon boundary of the feasible region. Hence, as we move
further up and to the right in the direction of the gradient (corresponding to larger and
larger values of z(x
1
, x
2
)) we see that there is not one point on the boundary of the feasible
region that intersects that level set with greatest value, but instead a side of the polygon
boundary described by the line 3x
1
+x
2
= 120 where x
1
[16, 35]. Let:
S = (x
1
, x
2
[3x
1
+x
2
120, x
1
+ 2x
2
160, x
1
35, x
1
, x
2
0
that is, S is the feasible region of the problem. Then for any value of x
1
[16, 35] and any
value x
2
so that 3x
1
+x
2
= 120, we will have z(x
1
, x
2
) z(x
1
, x
2
) for all (x
1
, x
2
) S. Since
there are innitely many values that x
1
and x
2
may take on, we see this problem has an
innite number of alternative optimal solutions.
Every point on this line
is an alternative optimal
solution.
S
Figure 7.2. An example of innitely many alternative optimal solutions in a linear
programming problem. The level curves for z(x
1
, x
2
) = 18x
1
+ 6x
2
are parallel to
one face of the polygon boundary of the feasible region. Moreover, this side contains
the points of greatest value for z(x
1
, x
2
) inside the feasible region. Any combination
of (x
1
, x
2
) on the line 3x
1
+x
2
= 120 for x
1
[16, 35] will provide the largest possible
value z(x
1
, x
2
) can take in the feasible region S.
3. Some Basic Facts about Linear Programming Problems
Definition 7.10 (Canonical Form). A maximization linear programming problem is in
canonical form if it is written as:
(7.11)
_
_
max z(x) =c
T
x
s.t. Ax b
x 0
91
A minimization linear programming problem is in canonical form if it is written as:
(7.12)
_
_
min z(x) =c
T
x
s.t. Ax b
x 0
Definition 7.11 (Standard Form). A linear programming problem is in standard form
if it is written as:
(7.13)
_
_
max z(x) =c
T
x
s.t. Ax = b
x 0
Remark 7.12. The following theorem is outside the scope of the course. You may cover
it in a Math 484 [Gri11b].
Theorem 7.13. Every linear programming problem in canonical form can be put into
standard form.
Exercise 66. Show that a minimization linear programming problem in canonical form
can be rephrased as a maximization linear programming problem in canonical form. [Hint:
Multiply the objective and constraints 1. Dene new matrices.]
Remark 7.14. To illustrate Theorem 7.13, we note that it is relatively easy to convert
any inequality constraint into an equality constraint. Consider the inequality constraint:
(7.14) a
i1
x
1
+a
i2
x
2
+ +a
in
x
n
b
i
We can add a new slack variable s
i
to this constraint to obtain:
a
i1
x
1
+a
i2
x
2
+ +a
in
x
n
+s
i
= b
i
Obviously this slack variable s
i
0. The slack variable then becomes just another variable
whose value we must discover as we solve the linear program for which Expression 7.14 is a
constraint.
We can deal with constraints of the form:
(7.15) a
i1
x
1
+a
i2
x
2
+ +a
in
x
n
b
i
in a similar way. In this case we subtract a surplus variable s
i
to obtain:
a
i1
x
1
+a
i2
x
2
+ +a
in
x
n
s
i
= b
i
Again, we must have s
i
0.
Example 7.15. Consider the linear programming problem:
_
_
max z(x
1
, x
2
) = 2x
1
x
2
s.t. x
1
x
2
1
2x
1
+x
2
6
x
1
, x
2
0
92
This linear programming problem can be put into standard form by using both a slack and
surplus variable.
_
_
max z(x
1
, x
2
) = 2x
1
x
2
s.t. x
1
x
2
+s
1
= 1
2x
1
+x
2
s
2
= 6
x
1
, x
2
, s
1
, s
2
0
Definition 7.16 (Row Rank). Let A R
mn
. The row rank of A is the size of the
largest set of row (vectors) from A that are linearly independent.
Example 7.17. The row rank of the matrix
A =
_
_
1 2 3
4 5 6
7 8 9
_
_
is 2. To see this note that:
_
7 8 9
=
_
1 2 3
+ 2
_
4 5 6
It is also clear that [1 2 3] and [4 5 6] are linearly independent. Thus showing that the row
rank of A is 2.
Remark 7.18. The column rank of a matrix A R
mn
is dened analogously on columns
rather than rows. The following theorem relates the row and column rank. Its proof is
outside the scope of the course.
Theorem 7.19. If A R
mn
is a matrix, then the row rank of A is equal to the column
rank of A. Further, rank(A) minm, n.
Definition 7.20. Suppose that A R
mn
and let m n. Then A has full row rank if
rank(A) = m.
Remark 7.21. We will assume, when dealing with Linear Programming Problems in
standard or canonical form that the matrix A has full row rank and if not, we will adjust it
so this is true. The following theorem tells us what can happen in a Linear Programming
Problem.
Theorem 7.22. Consider any linear programming problem:
P
_
_
max z(x) =c
T
x
s.t. Ax b
x 0
Then there are exactly four possibilities:
(1) There is a unique solution to problem P denoted x
.
(2) There are an innite number of alternative optimal solutions to P.
(3) There is no solution to P because there is no x that satises Ax = b.
(4) There is no solution to P because the problem is unbounded. That is for any x such
that Ax = b there is another x
,= x so that Ax
= b and c
T
x < c
T
x
.
93
4. Solving Linear Programming Problems with a Computer
Remark 7.23. There are a few ways to solve Linear Programming problems. The most
common approach is called the Simplex Algorithm. Unfortunately, we will not have time to
cover the Simplex Algorithm in this class. This is covered in IE 405 and Math 484, for those
interested [Gri11b].
Remark 7.24. Well show how to solve Linear Programs using Matlab. Matlab assumes
that all linear programs are input in the following form:
(7.16)
_
_
min z(x) =c
T
x
s.t. Ax b
Hx = r
x l
x u
Here c R
n1
, so there are n variables in the vector x, A R
mn
, b R
m1
, H R
ln
and r R
l1
. The vectors l and u are lower and upper bounds respectively on the decision
variables in the vector x.
The Matlab command for solving linear programs is linprog and it takes the parameters:
(1) c,
(2) A,
(3) b,
(4) H,
(5) r,
(6) l,
(7) u
If there are no inequality constraints, then we set A = [] and b = [] in Matlab; i.e., A and
b are set as the empty matrices. A similar requirement holds on H and r if there are no
equality constraints. If some decision variables have lower bounds and others dont, the term
-inf can be used to set a lower bound at (in l). Similarly, the term inf can be used if
the upper bound on a variable (in u) is innity. The easiest way to understand how to use
Matlab is to use it on an example.
Example 7.25. Suppose I wish to design a diet consisting of Raman noodles and ice
cream. Im interested in spending as little money as possible but I want to ensure that I eat
at least 1200 calories per day and that I get at least 20 grams of protein per day. Assume that
each serving of Raman costs $1 and contains 100 calories and 2 grams of protein. Assume
that each serving of ice cream costs $1.50 and contains 200 calories and 3 grams of protein.
We can construct a linear programming problem out of this scenario. Let x
1
be the
amount of Raman we consume and let x
2
be the amount of ice cream we consume. Our
objective function is our cost:
(7.17) x
1
+ 1.5x
2
Our constraints describe our protein requirements:
(7.18) 2x
1
+ 3x
2
20
94
and our calorie requirements (expressed in terms of 100s of calories):
(7.19) x
1
+ 2x
2
12
This leads to the following linear programming problem:
(7.20)
_
_
min x
1
+ 1.5x
2
s.t. 2x
1
+ 3x
2
20
x
1
+ 2x
2
12
x
1
, x
2
0
Lets use Matlab to solve this problem. Our original problem is:
_
_
min x
1
+ 1.5x
2
s.t. 2x
1
+ 3x
2
20
x
1
+ 2x
2
12
x
1
, x
2
0
This is not in a form Matlab likes, so we change it by multiplying the constraints by 1
on both sides to obtain:
_
_
min x
1
+ 1.5x
2
s.t. 2x
1
3x
2
20
x
1
2x
2
12
x
1
, x
2
0
Then we have:
c =
_
1
1.5
_
A =
_
2 3
1 2
_
b =
_
20
12
_
H = r = []
l =
_
0
0
_
u = []
The Matlab code to solve this problem is shown in Figure 7.3 The solution Matlab returns
in the x variable is x
1
= 3.7184 and x
2
= 4.1877. It turns out there are actually an innite
number of alternative optimal solutions to this problem. You could draw a picture of this
scenario to see if you can gure out why.
Exercise 67. In previous example, you could also have just used the problem in standard
form with the surplus variables and had A = b = [] and dened H and r instead. Use Matlab
to solve the diet problem in standard form. Compare your results to Example 7.25
95
%%Solve the Diet Linear Programming Problem
c = [1 1.5];
A = [[-2 -3];...
[-1 -2]];
b = [-20 -12];
H = [];
r = [];
l = [0 0];
u = [];
[x obj] = linprog(c,A,b,H,r,l,u);
Figure 7.3. Matlab input for solving the diet problem. Note that we are solving
a minimization problem. Matlab assumes all problems are mnimization problems,
so we dont need to multiply the objective by 1 like we would if we started with a
maximization problem.
5. Kurush-Kuhn-Tucker (KKT) Conditions
Remark 7.26. The single most important thing to learn about Linear Programming (or
optimization in general) is the Kurush-Kuhn-Tucker optimality conditions. These conditions
provide necessary and sucient conditions for a point x R
n
to be an optimal solution to
a Linear Programming problem. We state the Kurush-Kuhn-Tucker theorem, but do not
prove it. A proof can be found in Chapter 8 of [Gri11b].
Theorem 7.27. Consider the linear programming problem:
(7.21) P
_
_
max cx
s.t. Ax b
x 0
with A R
mn
, b R
m
and (row vector) c R
n
. Then x
R
n
if and only if there exists
(row) vectors w
R
m
and v
R
n
and a slack variable vector s
R
m
so that:
Primal Feasibility
_
Ax
+s
= b
x
0
(7.22)
Dual Feasibility
_
_
w
Av
= c
w
0
v
0
(7.23)
Complementary Slackness
_
w
(Ax
b) = 0
v
= 0
(7.24)
Remark 7.28. The vectors w
and v
_
max z(x
1
, x
2
) = 7x
1
+ 6x
2
Dual Variable
s.t. 3x
1
+x
2
120 (w
1
)
x
1
+ 2x
2
160 (w
1
)
x
1
35 (w
3
)
x
1
0 (v
1
)
x
2
0 (v
2
)
In this problem we have:
A =
_
_
3 1
1 2
1 0
_
_
b =
_
_
120
160
35
_
_
c =
_
7 6
_
_
_
3 1
1 2
1 0
_
_
_
x
1
x
2
_
_
_
120
160
35
_
_
_
x
1
x
2
_
_
0
0
_
Dual Feasibility
_
_
_
w
1
w
2
w
3
_
_
3 1
1 2
1 0
_
_
_
v
1
v
2
=
_
7 6
_
w
1
w
2
w
3
_
0 0 0
_
v
1
v
2
_
0 0
Complementary Slackness
_
_
_
w
1
w
2
w
3
_
_
_
_
3 1
1 2
1 0
_
_
_
x
1
x
2
_
_
_
120
160
35
_
_
_
_
120
160
35
_
_
_
_
_
v
1
v
2
_
x
1
x
2
= 0
Recall that at optimality, we had x
1
= 16 and x
2
= 72. The binding constraints in this case
where
3x
1
+x
2
120
x
1
+ 2x
2
160
To see this note that if 3(16) + 72 = 120 and 16 + 2(72) = 160. Then we should be able
to express c = [7 6] (the vector of coecients of the objective function) as a positive
combination of the gradients of the binding constraints:
(7x
1
+ 6x
2
) =
_
7 6
(3x
1
+x
2
) =
_
3 1
(x
1
+ 2x
2
) =
_
1 2
97
That is, we wish to solve the linear equation:
(7.25)
_
w
1
w
2
_
3 1
1 2
_
=
_
7 6
1
, x
2
) = (16, 72)
3x1 +x2 120
x1 + 2x2 160
x1 35
x1 0
x2 0
Figure 7.4. The Gradient Cone: At optimality, the cost vector c is obtuse with
respect to the directions formed by the binding constraints. It is also contained
inside the cone of the gradients of the binding constraints, which we will discuss at
length later.
since x
1
< 35, we know that x
1
35 is not a binding constraint and thus its dual variable
w
3
is also zero. This leads to the conclusion:
_
x
1
x
2
_
=
_
16
72
_
_
w
1
w
2
w
=
_
8/5 11/5 0
_
v
1
v
=
_
0 0
_
min cx
s.t. Ax b
x 0
[Hint: Remember, every minimization problem can be converted to a maximization problem
by multiplying the objective function by 1 and the constraints Ax b are equivalent to
the constraints Ax b.
6. Duality
Remark 7.30. In this section, we show that to each linear programming problem (the
primal problem) we may associate another linear programming problem (the dual linear
programming problem). These two problems are closely related to each other and an analysis
of the dual problem can provide deep insight into the primal problem.
Consider the linear programming problem
(7.27) P
_
_
max c
T
x
s.t. Ax b
x 0
Then the dual problem for Problem P is:
(7.28) D
_
_
min wb
s.t. wA c
w 0
Remark 7.31. Let v be a vector of surplus variables. Then we can transform Problem
D into standard form as:
(7.29) D
S
_
_
min wb
s.t. wAv = c
w 0
v 0
Thus we already see an intimate relationship between duality and the KKT conditions. The
feasible region of the dual problem (in standard form) is precisely the the dual feasibility
constraints of the KKT conditions for the primal problem.
In this formulation, we see that we have assigned a dual variable w
i
(i = 1, . . . , m) to
each constraint in the system of equations Ax b of the primal problem. Likewise dual
variables v can be thought of as corresponding to the constraints in x 0.
Remark 7.32. The proof of the following lemma is outside the scope of the class, but it
establishes an important fact about duality.
99
Lemma 7.33. The dual of the dual problem is the primal problem.
Remark 7.34. Lemma 7.33 shows that the notion of dual and primal can be exchanged
and that it is simply a matter of perspective which problem is the dual problem and which is
the primal problem. Likewise, by transforming problems into canonical form, we can develop
dual problems for any linear programming problem.
The process of developing these formulations can be exceptionally tedious, as it requires
enumeration of all the possible combinations of various linear and variable constraints. The
following table summarizes the process of converting an arbitrary primal problem into its
dual. This table can be found in Chapter 6 of [BJS04].
MINIMIZATION PROBLEM MAXIMIZATION PROBLEM
C
O
N
S
T
R
A
I
N
T
S
V
A
R
I
A
B
L
E
S
C
O
N
S
T
R
A
I
N
T
S
V
A
R
I
A
B
L
E
S
UNRESTRICTED
= UNRESTRICTED
=
0
0
0
0
Table 1. Table of Dual Conversions: To create a dual problem, assign a dual
variable to each constraint of the form Ax b, where represents a binary relation.
Then use the table to determine the appropriate sign of the inequality in the dual
problem as well as the nature of the dual variables.
Example 7.35. Consider the problem of nding the dual problem for the Toy Maker
Problem (Example 7.5) in standard form. The primal problem is:
max 7x
1
+ 6x
2
s.t. 3x
1
+x
2
+s
1
= 120 (w
1
)
x
1
+ 2x
2
+s
2
= 160 (w
2
)
x
1
+s
3
= 35 (w
3
)
x
1
, x
2
, s
1
, s
2
, s
3
0
Here we have placed dual variable names (w
1
, w
2
and w
3
) next to the constraints to which
they correspond.
The primal problem variables in this case are all positive, so using Table 1 we know that
the constraints of the dual problem will be greater-than-or-equal-to constraints. Likewise, we
know that the dual variables will be unrestricted in sign since the primal problem constraints
are all equality constraints.
100
The coecient matrix is:
A =
_
_
3 1 1 0 0
1 2 0 1 0
1 0 0 0 1
_
_
Clearly we have:
c =
_
7 6 0 0 0
b =
_
_
120
160
35
_
_
Since w = [w
1
w
2
w
3
], we know that wA will be:
wA =
_
3w
1
+w
2
+w
3
w
1
+ 2w
2
w
1
w
2
w
3
This vector will be related to c in the constraints of the dual problem. Remember, in this
case, all variables in the primal problem are greater-than-or-equal-to zero. Thus
we see that the constraints of the dual problem are:
3w
1
+w
2
+w
3
7
w
1
+ 2w
2
6
w
1
0
w
2
0
w
3
0
We also have the redundant set of constraints that tell us w is unrestricted because the
primal problem had equality constraints. This will always happen in cases when youve
introduced slack variables into a problem to put it in standard form. This should be clear
from the denition of the dual problem for a maximization problem in canonical form.
Thus the whole dual problem becomes:
(7.30)
min 120w
1
+ 160w
2
+ 35w
3
s.t. 3w
1
+w
2
+w
3
7
w
1
+ 2w
2
6
w
1
0
w
2
0
w
3
0
w unrestricted
Again, note that in reality, the constraints we derived from the wA c part of the dual
problem make the constraints w unrestricted redundant, for in fact w 0 just as we
would expect it to be if wed found the dual of the Toy Maker problem given in canonical
form.
101
Exercise 70. Identify the dual problem for:
max x
1
+x
2
s.t. 2x
1
+x
2
4
x
1
+ 2x
2
6
x
1
, x
2
0
Exercise 71. Use the table or the denition of duality to determine the dual for the
problem:
(7.31)
_
_
min cx
s.t. Ax b
x 0
Remark 7.36. The following theorems are outside the scope of this course, but they can
be useful to know and will help cement your understanding of the true nature of duality.
Theorem 7.37 (Strong Duality Theorem). Consider Problem P and Problem D. Then
(Weak Duality): cx
and w
respectively
and cx
= w
b.
(2) Problem P is unbounded and Problem D is infeasible.
(3) Problem D is unbounded and Problem P is infeasible.
(4) Both problems are infeasible.
Theorem 7.38. Problem D has an optimal solution w
R
m
if and only if there exists
vectors x
R
n
and s
R
m
and a vector of surplus variables v
R
n
such that:
Primal Feasibility
_
w
A c
w
0
(7.32)
Dual Feasibility
_
_
Ax
+s
= b
x
0
s
0
(7.33)
Complementary Slackness
_
(w
Ac) x
= 0
w
= 0
(7.34)
Furthermore, these KKT conditions are equivalent to the KKT conditions for the primal
problem.
Remark 7.39. The nal theorem illustrates the true nature of duality. Two linear pro-
gramming problems are dual if they share KKT conditions. That is, if they share conditions
for optimality.
102
1 4
3
2
$3 / 1
$4 / 3
$1 / 2 $5 / 2
Cost Per Flow Unit / Capacity
Produces 3 units
Consumes 3 units
x
1
x
2 x
3
x
4
$2 / 2
x
5
Figure 7.5. In this problem, it costs a certain amount to ship a commodity along
each edge and each edge has a capacity. The objective is to nd an allocation of
capacity to each edge so that the total cost of shipping three units of this commodity
from Vertex 1 to Vertex 4 is minimized.
Exercise 72. Consider the directed graph shown in Figure 7.5 The amount of ow along
each edge is given by the variables x
1
, . . . , x
5
. The total cost of shipping ow from Vertex 1
to Vertex 5 is
(7.35)
5
i=1
c
i
x
i
where c
i
is the cost associated to the ow in each edge. For each edge, we know that x
i
0
and x
i
u
i
where u
i
is the capacity on each edge. Finally, we must be able to assert
that commodities are neither created nor destroyed (except at Vertex 1, where 3 units of
commodity are created and at Vertex 4 where 3 units of commodity are consumed). Thus
we have constraints of the form:
x
1
+x
2
= 3
x
1
= x
4
+x
5
x
2
+x
5
= x
3
x
3
+x
4
= 3
(1) Put all these constraints together to form a linear programming problem whose
solution yields a minimal cost assignment of ow to the edges.
(2) Use Matlab to nd an optimal ow.
(3) Notice that each equation in the equality constraints represents the balance of ow
into and out of a vertex. Rewrite each equation so that is has the form
ow-out ow-in = ow produced at vertex ow consumed at vertex
(4) Compute the A matrix that results from the equations you just constructed and
compare it to the incidence matrix of the directed graph. [Hint: They should be
the same.]
103
CHAPTER 8
An Introduction to Network Flows and Combinatorial
Optimization
Remark 8.1. For the remainder of this chapter, we will consider directed graphs with no
isolated vertices and no self-loops. That is, we will only consider those graphs whose incident
matrices do not have any zero rows. These graphs will be connected and furthermore will
have two special vertices v
1
and v
m
and we will assume that there is at least one directed
path from v
1
to v
m
.
1. The Maximum Flow Problem
Definition 8.2 (Flow). Let G = (V, E) be a digraph and suppose V = v
1
, . . . , v
m
and E = e
1
, . . . , e
n
. If e
k
= (v
i
, v
j
) is an edge, then a ow on e
k
is a value x
k
R
+
that
determines that amount of some quantity (of a commodity) that will leave v
i
and ow along
e
k
to v
j
.
Definition 8.3 (Vertex Supply and Demand). Let G = (V, E) be a digraph and suppose
V = v
1
, . . . , v
m
. The ow supply for vertex v
i
is a real value b
i
assigned to v
i
that quanties
that amount of ow produced at vertex v
i
. If b
i
< 0, then vertex v
i
consumes ow (rather
than producing it).
Definition 8.4 (Flow Conservation Constraint). Let G = (V, E) be a digraph with
no self-lops and suppose V = v
1
, . . . , v
m
and E = e
1
, . . . , e
n
. Let I(i) be the set of
edges with destination vertex v
i
and O(i) be the set of edges with source v
i
. Then the ow
conservation constraint associated to vertex v
i
is:
(8.1)
kO(i)
x
k
kI(i)
x
k
= b
i
i
Remark 8.5. Equation 8.1 states that the total ow out of vertex v
i
minus the total
ow into v
i
must be equal to the total ow produced at v
i
. Or put more simply, excess ow
is neither created nor destroyed.
Definition 8.6 (Edge Capacity). Let G = (V, E) be a digraph with no self-lops and
suppose V = v
1
, . . . , v
m
and E = e
1
, . . . , e
n
. If e
k
E then its capacity is a positive
real value u
k
that determines the maximum amount of ow the edge may be assigned. We
can think of (G, u) as being a weighed graph where the weights are the edge capacities.
Proposition 8.7. Let G = (V, E) be a digraph with no self-lops and suppose V =
v
1
, . . . , v
n
and let A be the incidence matrix of G (see Denition 4.63). Then Equation
8.1 can be written as:
(8.2) A
i
x = b
i
105
where x is a vector of variables of the form x
k
taken in the order the edges are represented
in A.
Proof. From Denition 4.63 we know that:
(8.3) A
ik
=
_
_
0 if v
i
is not in e
k
1 if v
i
is the source of e
k
1 if v
i
is the destination of e
k
The equivalence between Equation 8.2 and Equation 8.1 follows at once from this fact.
Remark 8.8. For the remainder of this chapter, Let e
1
R
m1
be the vector with a 1
at position 1 and 0 everywhere else. Dene e
m
similarly.
Definition 8.9 (Maximum Flow Problem). Let G = (V, E) be a digraph with no self-
loops and suppose that V = v
1
, . . . , v
m
. Without loss of generality, suppose that there is
no edge connecting v
m
to v
1
. The maximum ow problem for G is the linear programming
problem:
(8.4)
_
_
max f
s.t. (e
m
e
1
) f +Ax = 0
x u
x 0
f unrestricted
Here u is a vector of edge ow capacity values.
Remark 8.10. The constraints (e
m
e
1
) f + Ax = 0 are ow conservation constraints
when we assume that there is an (imaginary) ow backwards from v
m
to v
1
along an edge
(v
m
, v
1
) and that no ow is produced in the graph. That is, we assume all ow is circulating
within the graph. The value f determines the amount of ow that circulates back to vertex
v
1
from v
m
under this assumption. Since all ows are circulating and excess ow is neither
created nor destroyed, the value of f is then the total ow that ows from v
1
to v
m
. By
maximizing f, Problem 8.4 is exactly computing the maximum amount of ow that can go
from vertex v
1
to v
m
under the assumptions that ows are constrained by edge capacities
(x u), ows are non-negative (x 0) and ows are neither created nor destroyed in the
graph.
2. The Dual of the Flow Maximization Problem
Theorem 8.11. The dual linear programming problem for Problem 8.4 is:
(8.5)
_
_
min
n
k=1
u
k
h
k
s.t. w
m
w
1
= 1
w
i
w
j
+h
k
0 e
k
= (v
i
, v
j
) E
h
k
0 (v
i
, v
j
) E
w
i
unrestricted i 1, . . . , m
106
Proof. Consider the constraints of Problem 8.4 and suppose that the imaginary edge
from v
m
to v
1
is edge e
0
. We rst add slack variables to constraints of the form x
k
u
k
to
obtain:
x
k
+s
k
= u
k
k 1, . . . , n
The constraints (other than x 0 and f unrestricted) can be rewritten in matrix form as:
(8.6)
_
_
1 a
11
a
12
a
1n
0 0 0
0 a
12
a
22
a
2n
0 0 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1 a
m1
a
m2
a
mn
0 0 0
0 1 0 0 1 0 0
0 0 1 0 0 1 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 1 0 0 1
_
_
_
_
f
x
1
x
2
.
.
.
x
n
s
1
s
2
.
.
.
s
n
_
_
=
_
_
0
0
.
.
.
0
u
1
u
2
.
.
.
u
n
_
_
or more simply as:
(8.7)
_
e
m
e
1
A 0
0 I
n
I
n
_
_
_
f
x
s
_
_
=
_
0
u
_
where all elements written 0 are zero matrices or vectors of appropriate dimension. This
matrix has 2n + 1 columns and m + n rows. To the rst m rows, we associate the dual
variables w
1
, . . . , w
m
. To the next n rows we associate the dual variables h
1
, . . . , h
n
. Our
dual variable vector is then:
y = [w
1
, . . . , w
m
, h
1
, . . . , h
n
]
Since the constraints in Expression 8.6 are all equality constraints, we know that there dual
variables are unrestricted. Further from Expression 8.6 we know that the objective function
vector is:
c = [1, 0, . . . , 0]
We can now compute our dual constraints. The rst constraint is computed fromy multiplied
by the rst column of the matrix in Expression 8.6 and gives:
(8.8) w
1
+w
m
= 1
Since this dual constraint corresponds to the variable f in the primal problem we know it
will be an equality constraint (f is unrestricted) and that its right hand side will be 1 the
coecient of f in the primal problem. The next n constraints are derived by multiplying y
by the next n columns of the matrix in Expression 8.6 and will have the form:
(8.9) w
i
w
j
+h
k
0
This follows since there will be a 1 in the matrix whenever edge e
k
has as destination vertex
v
j
and a +1 in the matrix whenever edge e
k
has source at vertex v
i
. Clearly there is a 1 in the
k
th
row of the identity matrix below the A matrix in Expression 8.6 thus yielding the +h
k
107
term. These constraints correspond to the variables x
1
, . . . , x
n
, which are all non-negative
and thus the constraints are non-negative. The nal n constraints have form:
(8.10) h
k
0
and are derived by multiplying y by the last n columns of the matrix in Expression 8.6.
These constraints correspond to the variables s
1
, . . . , s
n
which are non-negative and thus the
constraints are non-negative. The objective function of the dual problem is computed by
multiplying y by the right-hand-side of Expression 8.6. This yields:
(8.11)
n
k=1
u
k
h
k
Problem 8.5 follows at once. This completes the proof.
3. The Max-Flow / Min-Cut Theorem
Remark 8.12. Let G = (V, E) be a directed graph and suppose V = v
1
, . . . , v
m
and
E = e
1
, . . . , e
n
. Recall we dened an edge cut in Denition 3.39. Let V
1
be any set
of vertices containing v
1
and not containing v
m
and let V
2
= V v
1
. Immediately we see
v
m
V
2
. Clearly, the edges connecting vertices in V
1
with vertices in V
2
form an edge cut
and moreover any edge cut that divides G into two components, one containing v
1
and the
other containing v
m
corresponds to some sets V
1
and V
2
. Thus, we will refer to all such
edge cuts by these generated sets; that is, (V
1
, V
2
) corresponds to the edge cut dened when
v
1
V
1
and v
m
V
2
and V
1
V
2
= and V
1
V
2
= V . For the remainder of this chapter, a
cut refers to a cut of this type.
Definition 8.13 (Cut Capacity). Let G = (V, E) be a directed graph and suppose V =
v
1
, . . . , v
m
and E = e
1
, . . . , e
n
. Let (V
1
, V
2
) be a cut separating v
1
from v
m
containing
edges e
s
1
, . . . , e
s
l
with sources in V
1
and destinations in V
2
, where s
1
, . . . , s
l
is a subset of the
edge indexes 1, . . . , n. Then the capacity of the cut (V
1
, V
2
) is:
(8.12) C(V
1
, V
2
) =
l
k=1
u
s
k
Lemma 8.14. Let G = (V, E) be a directed graph and suppose V = v
1
, . . . , v
m
and
E = e
1
, . . . , e
n
. The solution to the maximum ow problem is bounded above by the
minimal cut capacity.
Proof. Let (V
1
, V
2
) be the cut with minimal capacity. Consider the following solution
to the dual problem:
(8.13) w
i
=
_
0 v
i
V
1
1 v
i
V
2
and
(8.14) h
k
=
_
1 e
k
= (v
i
, v
j
) and v
i
V
1
and v
j
V
2
0 else
108
It is clear this represents a feasible solution to the dual problem. Thus by the strong duality
theorem (Theorem 7.37) the objective function value:
(8.15)
k
u
k
h
k
is an upper bound for the primal problem. But this is just the capacity of the cut with the
smallest capacity. This completes the proof.
Lemma 8.15. In any optimal solution to Problem 8.4 every directed path from v
1
to v
m
must have at least one edge at capacity.
Proof. Note rst, problem 8.4 is bounded above by the capacity of the minimal cut as
shown in Lemma 8.14 and since the zero ow is a feasible solution, we know from Theorem
7.22 there is at least one optimal solution to Problem 8.4 because the problem can neither
be unbounded nor infeasible.
Consider any optimal solution to Problem 8.5. Then it corresponds to some optimal solu-
tion to the primal problem and these solutions satisfy the Kurush-Kuhn-Tucker conditions.
We show that in this primal solution, along each path from v
1
to v
m
in G at least one edge
must have ow equal to its capacity. To see this note that for any edge that does not carry
its capacity (that is x
k
< u
k
) we must have h
k
= 0 (to ensure complementary slackness).
Suppose this path has vertices (u
1
, v
2
, . . . , u
s
) with v
1
= u
1
and v
m
= u
s
. If there is some
path from v
1
to v
m
that does not carry its capacity, then we have the following requirements:
w
s
> w
1
w
1
w
2
.
.
.
w
s1
w
s
But this implies that w
s
> w
1
w
2
w
s
, which is a contradiction. Therefore every
path from v
1
to v
m
has at least one edge at capacity.
Theorem 8.16. Let G = (V, E) be a directed graph and suppose V = v
1
, . . . , v
m
and
E = e
1
, . . . , e
n
. There is at least one cut (V
1
, V
2
) so that the ow from v
1
to v
m
is equal to
the capacity of the cut (V
1
, V
2
).
Proof. Denote one such solution to Problem 8.4 as (x
, f
.
If E
is not yet an edge cut in the underlying graph of G, then there are some paths from
v
1
to v
m
in the underlying graph of G that are not directed paths from v
1
to v
m
. In each
such path, there is at least one edge directed toward v
m
from v
1
. Choose one edge from each
of these paths directed from v
m
to v
1
to minimize the total cardinality of edges chosen and
add these edges to E
V
2
= v : there is a path from v to v
m
in the underlying graph of GE
= (V
1
, V
2
) is an edge cut in the underlying graph of G.
Proof. See Exercise 73.
Suppose E
= e
s
1
, . . . , e
s
l
.
Claim 2. If there is some edge e
k
with source in V
2
and destination in V
1
, then x
k
= 0.
Proof. If x
k
,= 0, we could reduce this ow to zero and increase the net ow from v
1
to
v
m
by adding this ow to f. If ow cannot reach x
1
along e
k
(illustrated by the middle path
in Figure 8.1, then ow conservation ensures it must be equal to zero.
Claim 3. The total ow from v
1
to v
m
must be equal to the capacity of the edges in E
=
(V
1
, V
2
).
Claim 3 establishes that the ow f
, f
) is a maximum ow in G from v
1
to v
m
, then there
is a cut (V
1
, V
2
) so that the capacity of this cut is equal to f
. Since f
is bounded above by
the capacity of the minimal cut separating v
1
from v
m
, the cut constructed in the proof of
Theorem 8.16 must be a minimal capacity cut. Thus, the maximum ow from v
1
to v
m
is
equal to the capacity of the minimum cut separating v
1
from v
m
.
Exercise 73. Prove Claim 1 in the proof of the Theorem 8.16.
4. An Algorithm for Finding Optimal Flow
Remark 8.18. The proof of the Max Flow / Min Cut theorem we presented is a very
non-standard proof technique. Most techniques are constructive; that is, they specify an
algorithm for generating a maximum ow and then show that this maximum ow must be
equal to the capacity of the minimal cut. In this section, well develop this algorithm and
show that it generates a maximum ow and then (as a result of the Max Flow / Min Cut
theorem this maximum ow must be equal to the capacity of the minimum cut.)
Definition 8.19 (Augment). Let G = (V, E) be a directed graph and suppose V =
v
1
, . . . , v
m
and E = e
1
, . . . , e
n
and let x be a feasible ow in G. Consider a simple path
p = (v
1
, e
1
, . . . , e
l
, v
m
) in the underlying graph of G from v
1
to v
m
. The augment of p is the
quantity:
(8.16) min
k{1,...,l}
_
u
k
x
k
if the edge e
k
is directed toward v
m
x
k
else
Definition 8.20 (Augmenting Path). Let G = (V, E) be a directed graph and suppose
V = v
1
, . . . , v
m
and E = e
1
, . . . , e
n
and let x be a feasible ow in G. A simple path p in
in the underlying graph of G from v
1
to v
m
is an augmenting path if its augment is non-zero.
In this case we say that ow x has an augmenting path.
Example 8.21. An example of augmenting paths is shown in Figure 8.2. An augmenting
u
k
1
2
3
4
2 3
3 1
3
u
k
x
k
1
2
3
4
1 / 2 3 / 3
3 / 3 1 / 1
2 / 3
x
k
1
2
3
4
1 / 2 1.25 / 3
1 / 3
0.75 / 1
0.25 / 3
x
k
Augment = 0.25 No augmenting path (Augment = 0)
1
2
3
4
1 / 2 2 / 3
2 / 3 1 / 1
1 / 3
x
k
Augment = 1
Figure 8.2. Two ows with augmenting paths and one with no augmenting paths
are illustrated.
path is simply an indicator that more ow can be pushed from vertex v
1
to vertex v
m
. For
example, in the ow on the bottom left of Figure 8.2 we could add an additional unit of
111
ow on the edge (v
1
, v
3
). This one unit could ow along edge (v
3
, v
2
) and then along edge
(v
2
, v
4
). Augmenting paths that are augmenting soley because of a backward ow away from
v
1
to v
m
can also be used to increase the net ow from v
1
to v
m
by removing ow along the
backward edge.
Definition 8.22 (Path Augment). If p is an augmenting path in G with augment ,
then by augmenting p by we mean add to the ow in each edge directed from v
1
toward
v
m
and subtract from the ow in each edge directed from v
m
to v
1
.
Example 8.23. If we augment the augmenting paths illustrated in Example 8.21, the
resulting ows are illustrated in Figure 8.3.
u
k
x
k
1
2
3
4
1 / 2 1.25 / 3
1 / 3
0.75 / 1
0.25 / 3
x
k
Augment = 0.25
1
2
3
4
1 / 2 2 / 3
2 / 3 1 / 1
1 / 3
x
k
Augment = 1
1
2
3
4
1 / 2 3 / 3
3 / 3 1 / 1
2 / 3
x
k
1
2
3
4
1.25 / 2 1.25 / 3
1 / 3
1 / 1
0 / 3
x
k
Figure 8.3. The result of augmenting the ows shown in Figure 8.2.
Remark 8.24. The next algorithm, sometimes called the Edmonds-Karp Algorithm will
nd a maximum ow in a network by discovering and removing all augmenting paths.
Maximum Flow Algorithm
Input: (G, u) a weighted directed graph with G = (V, E), V = v
1
, . . . , v
m
, E = e
1
, . . . , e
n
is feasible.
Proof. To see that x
is
optimal.
Corollary 8.32 (Integral Flow Theorem). If the capacities of a network are all integers,
then there exists an integral maximum ow.
Remark 8.33. It is worth noting that the original form of Algorithm 8 did not specify
which augmenting path to nd. This leads to a pathological condition in which the algorithm
occasionally will not terminate. This is detailed in Ford and Fulkersons original paper and
more recently in [Zwi95]. The shortest augmenting path can be found using a breadth
rst search on the underlying graph. This breadth rst search is what leads to the proof of
Theorem 8.30.
Exercise 74. Prove the Integral Flow Theorem.
114
Remark 8.34. The problem of nding a maximum ow (or minimum cut) is still very
much an area of interest for researchers with new results being published as recently as the
late 90s. See [KV08] for details.
5. Applications of the Max Flow / Min Cut Theorem
Theorem 8.35 (Mengers First Theorem). Let G be an (undirected) graph with V =
v
1
, . . . , v
m
. Then the number of edge disjoint paths from v
1
to v
m
is equal to the size of
the smallest edge cut separating v
1
from v
m
.
Exercise 75. Prove Mengers Theorem. [Hint: Enumerate all edge disjoint paths from
v
1
to v
m
and replace them with directed paths from v
1
to v
m
. If any edges remain un-directed,
then give them arbitrary direction. Assign each arc a ow capacity of 1.]
Remark 8.36. The next two proofs of classical theorems are based on the same results
in [KV08].
Theorem 8.37 (Mengers Second Theorem). Let G = (V, E) be a directed graph. Let
v
1
and v
2
be two non-adjacent and distinct vertices in V . The maximum number of vertex
disjoint directed paths from v
1
to v
2
is equal to the minimum number of vertices (excluded v
1
and v
2
) whose deletion will destroy all directed paths from v
1
to v
2
.
Proof. We construct a new graph by replace each vertex v in G by two vertices v
and
v
and an edge (v
, v
, w
, v
), illustrated below.
v
v' v''
u w
u'' w'
Note that each arc of the form (v
, v
be a maximal
matching for G and let C
[ [C
[. To see this
note that if a matching contains every vertex (e.g. it is perfect) then a covering of G can be
constructed by simply choosing (at random) one vertex from each match. The fact that G is
bipartite means that each edge in G is now adjacent to some vertex in the cover constructed
from M
.
Construct a new graph N from G by introducing new vertices s and t so that s is adjacent
to all vertices in V
1
and t is adjacent to all vertices in V
2
. This is illustrated below.
s t
In the remainder of the proof, s will be our source (v
1
) and t will be our sink (v
m
). Consider
a maximal (in cardinality) set P of vertex disjoint paths from s to t (we may think of G
being directed from vertices in V
1
toward vertices in V
2
). Each path p P has the form
(s, e
1
, v
1
, e
2
, v
2
, e
3
, t) with v
1
V
1
and v
2
V
2
. It is easy to see that we can construct a
matching M(P) from P so for path p we introduce the edge e
2
= v
1
, v
2
into our match-
ing M(P). The fact that the paths in P are vertex disjoint implies there is a one-to-one
correspondence between elements in M(P) and elements in P. Thus, [P[ [M
[ since we
assumed that M
was a
minimal vertex cover, it follows that J is not itself a vertex cover of G and thus G J
leaves at least one edge in G. But this edge must connect a vertex in V
1
to a vertex in V
2
because G is bipartite. Thus, N J has a path from s to t, which is a contradiction. Thus,
[C
[ [C
[ = [C
[.
This completes the proof.
Remark 8.44. It is clear that K onigs Theorem does not hold in general. To see this,
consider K
3
(see Figure 8.7). In this case, the general inequality that that the cardinality
of the maximal matching is at most the cardinality of the minimal covering does hold (and
this will always hold), but we do not have equality.
Remark 8.45. Let G = (V, E) be a (bipartite) graph with V = v
1
, . . . , v
m
and E =
e
1
, . . . , e
n
. The minimal vertex covering problem for G can be written as the integer
programming problem:
(8.21)
_
_
min x
1
+ +x
m
s.t. x
i
+x
j
1 v
i
, v
j
E
x
i
0, 1 i = 1, . . . , m
If A is the incidence matrix for G, then this problem can be written in matrix notation as:
(8.22)
_
_
min 1
T
x
s.t. A
T
x 1
x 0, 1
m
where 1 is a vector consisting of only ones of appropriate length.
Exercise 76. Consider the relaxation of Problem 8.21:
(8.23)
_
_
min x
1
+ +x
m
s.t. x
i
+x
j
1 v
i
, v
j
E
x
i
1 i = 1, . . . , m
x
i
0 i = 1, . . . , m
Compute the dual problem for Problem 8.21. [Hint: Use slack and surplus variables to
construct a problem with only equality constraints (except for x 0). Compute the dual
problem. You will obtain an objective function that looks like max w
1
+ +w
n
+u
1
+u
2
+
+u
m
and some constraints.]
Exercise 77. Use the dual problem you constructed in the Exercise 76 along with a
restriction to 01 variables to show that to show that for K
3
the cardinality of the maximum
matching cannot be greater than 1 while the cardinality of the minimum covering can be 2.
Thus, illustrate that strong duality does not hold for integer programming problems.
117
Exercise 78. When G is a bipartite graph, argue that there is an integer solution of the
dual problem you found in Exercise 76 (even if you dont force it to have 0 1 variables).
Finally prove that under the assumptions weve given in this chapter (i.e., that no isolated
vertex exists in a bipartite graph G) the cardinality of a minimal vertex covering is equal to
the cardinality of a maximal matching. [Hint: Argue an optimal integer solution exists to
the relaxation of both problems. The result follows from strong duality.]
Remark 8.46. There are many other applications of Linear Programming (and Integer
Programming) to the study of graphs. Please consult [BJS04, KV08, PS98, WN99] for
details.
118
CHAPTER 9
A Short Introduction to Random Graphs
The study of random graphs presupposes a little knowledge of probability theory. For
those readers not familiar with probability at all, Chapter 1 of [Gri11a] contains a small
introduction to discrete probability spaces, which should provide sucient background for the
material on Random Graphs. The material can be obtained from http://www.personal.
psu.edu/cxg286/Math486.pdf. We provide one denition that is not contained in Chapter
1 of [Gri11a] that is critical for the rest of this chapter.
Definition 9.1 (Bernoulli Random Variable). A Bernoulli Random Variable is a random
variable underlying a probability space with exactly 2 outcomes 0 and 1 and for which the
probability of 1 occurring is p.
Remark 9.2. It is easy to see that if X is a Bernoulli random variable, then (the
expected value of X) E(X) = p = Pr(X = 1). Furthermore, given a set of n Bernoulli
Random variables, the expected number of these events that will be 1 is np.
Remark 9.3. Random graphs are misnamed, as it makes one think of a graph that is
somehow random. In reality, the term random graph usually refers to a family of graphs,
which serves as a discrete probability space on which we derive a probability measure that
assigns to each graph some probability that it occurs.
1. Bernoulli Random Graphs
Definition 9.4 (Bernoulli Random Graph). Let n Z
+
and let p (0, 1) Then ((n, p)
is Bernoulli Family of Random Graphs, which is the discrete probability space so that:
(1) The sample space is the set of all graphs with n vertices
(2) The probability that any two vertices chosen at random has an edge between them
is p and this probability is independent of all other edges.
(3) Therefore, for any given graph G = (V, E) in ((n, p) with [E[ = m, the probability
assigned to G is:
p
m
(1 p)
(
n
2
)m
Remark 9.5. Even though they are called Bernoulli Random Graphs, they were invented
by E. N. Gilbert [Gil59]. Gilbert went on to look at other (more useful) classes of random
graphs that are useful for modeling more realistic phenomena in communications [Gil61].
Example 9.6. We illustrate three graphs generated randomly where n = 10 and p = 0.5.
That means that any two edges have a 50% chance of having an edge between them. The rst
two graphs each have 21 edges and therefore the probability of these graphs is 0.5
2
1 0.5
2
4,
while the probability of the third graph is 0.5
2
4 0.5
2
1 because it has 24 edges. Of course
these two values are identical.
119
(a) (b) (c)
Figure 9.1. Three random graphs in the same random graph family (
_
10,
1
2
_
.
The rst two graphs, which have 21 edges, have probability 0.5
2
10.5
2
4. The third
graph, which has 24 edges, has probability 0.5
2
4 0.5
2
1.
Theorem 9.7. Let 2 k n. Then the probability that a graph G ((n, p) has a set
of k independent vertices is at most:
(9.1)
_
n
k
_
(1 p)
(
k
2
)
Proof. For any set of k vertices, the probability that they are independent is simply
the probability that none of the pairs of vertices are connected by an edge. There are
_
k
2
_
such pairs and each has a probability of (1 p) of not being connected by an edge. There
are
_
n
k
_
subsets of the n vertices each containing k elements. Therefore, the probability that
any of these sets is an independent set is:
_
n
k
_
(1 p)
(
k
2
)
Thus we have proved:
(9.2) Pr((G) k)
_
n
k
_
(1 p)
(
k
2
)
where is the independence number of the graph G.
Remark 9.8. Observe that Equation 9.2 is a weak bound in that if we choose k much
smaller than n and x p, then the bound exceeds 1. This is because, while it is true that
each edges existence in the set U in the proof is independent, not all the
_
n
k
_
possible sets
U are independent (they will share vertices). Thus we are over estimating the probability
when we sum them in the proof.
Exercise 79. Find a limiting expression for the probability that G ((n, p) has a clique
of size k with 2 k n. [Hint: Remember, a clique is the exact opposite of an independent
set.]
120
Definition 9.9 (Almost Sure Properties). Let P be a statement about a graph G. (For
example, P might be the statement G is a tree.) A property P is said to hold almost
surely (abbreviated a.s) for graphs in the family ((n, p) if:
lim
n
Pr (P holds for an arbitrary graph G ((n, p)) = 1
Remark 9.10. This notion of almost surely is a funny one, but theres any easy way to
get used to it. As n grows large, we are really absorbing more and more of the graphs that
are possible (that is structures G = (V, E) with [V [ some nite integer). If a property holds
with probability 1 as n goes to innity, it means that for almost every graph that property
must hold because no matter how large an n we choose, there are always more graphs with
more than n vertices than there are with fewer than n vertices. Thus, almost surely should
be interpreted over the set of all graphs with a nite (but unbounded) number of vertices.
Lemma 9.11. Let p (0, 1) and let G be a graph in ((n, p). Then almost surely every
pair of vertices v, u in G is connected by a path of length 2.
Proof. Let G = (V, E) and consider w V v, u. The probability that both edges
v, w and u, w are in G is p
2
. Thus the probability that at least one of these edges is
absent is 1 p
2
. Over all n 2 possible choices for w, the probability that this occurs each
time, therefore is:
_
1 p
2
_
n2
This quantity approaches 0 as n approaches innity and thus, a.s. there is a path of length
2 connecting u and v.
Theorem 9.12. A graph G ((n, p) is almost surely connected.
Proof. This proof is immediate from the previous lemma.
Theorem 9.13. Let p (0, 1) and let G be a graph in ((n, p) and let H = (V, E) be an
arbitrary xed graph. Then the property G has H as a subgraph holds a.s.
Proof. Suppose that H contains m vertices. Then partition the n vertices available in
graphs in ((n, p) into m sets each with size k = n/m|. If there are vertices remaining,
they may be added to an m + 1
st
partition and ignored. Suppose that H contains s edges.
Well order the vertices in each partition and choose the i
th
element from each of these m
partitions to be the vertices of H. (We will see that i will run from 1 to k.) With s edges,
the probability that the edges of the graph H are present is exactly p
s
and therefore the
probability these edges are not present is 1p
s
. Each of the k vertex m-tuples is independent
of all the others because (by ordering the partitions) we do not repeat any pair of vertices
that might form an edge. Thus, the probability that H is not spanned by any of these k
m-tuples of vertices is:
(1 p
s
)
k
= (1 p
s
)
n/m
Since p (0, 1) we have:
(9.3) lim
n
(1 p
s
)
n/m
= 0
Thus, the probability that H is a subgraph approaches 1 (1 = 10) as n approaches innity.
This completes the proof.
121
Remark 9.14. Our use of independence in the proof of Theorem 9.13 is a little subtle.
Clearly, when we are dealing with the m-tuple of vertices, the existence of the s edges
between each of those m vertices is independent by the denition of the Bernoulli Random
Graph Family. By ordering the elements in the m partitions of the vertices and choosing
only m-tuples that never repeat vertices, we can be sure that we never repeat edges and
thus the n/m| trials we take to nd H as a subgraph are independent. This is how we can
obtain Equation 9.3. Compare this to the proof of Theorem 9.7, in which we only seek a
bound.
Exercise 80. Let G be a graph in ((n, p) for p (0, 1). Show that the property G is
not a forest holds a.s.
2. First Order Graph Language and 0 1 properties
Remark 9.15. Theorems like Theorem 9.13 are called 0 1 properties. In turns out
that a great deal of these types of properties exist for simple graphs, as we prove.
Definition 9.16 (First Order Graph Language and the First Order Theory of Graphs).
The rst order graph language consists of a single relation (verb) E in which we write x and
y are connected by an edge as x, y E; here x and y are variables (vertex placeholders).
Remark 9.17. Sentences in rst order graph language are formed from the connectives
and (), or (), implies ( = ), not () and the quantiers there exists () and for all ().
Definition 9.18 (First Order Theory of Graphs). In the rst order theory of graphs, we
assume that
(1) for all x and y, x, y E y, x E and
(2) x, x , E for any x.
We also assume that there is an equality operation so that we can determine whether to
variables are equal.
Example 9.19. A very simple statement in rst order graph language is There is a
path of length 3. In this case, the statement would read:
(9.4) v
1
v
2
v
3
v
4
(v
1
, v
2
E v
2
, v
3
E v
3
, v
4
E
v
1
,= v
2
v
1
,= v
3
v
1
,= v
4
v
2
,= v
3
v
2
,= v
4
v
3
,= v
4
)
This simply says, there exists four distinct vertices v
1
, v
2
, v
3
and v
4
and v
1
and v
2
are
adjacent, v
2
and v
3
are adjacent, and v
3
and v
4
are adjacent. Other statements become
harder to write, but many graph properties can be expressed in rst order graph language.
Theorem 9.20. For every property T written in rst order graph language either T holds
a.s. or T fails to holds a.s.
Remark 9.21. The interested reader should see [Bol01] (Chapter 2) or (for the logicians
point of view) see [Mar00] (Chapter 2) for a proof of this theorem. This tells us that any
property that can be expressed in rst order graph language is no particularly interesting in
that either it will hold for almost every graph or fail to hold for almost every graph.
122
3. Erd os-Renyi Random Graphs
Definition 9.22 (Erdos-Renyi Family of Random Graphs). Let n Z
+
and let m be
an integer between 0 and
_
n
2
_
(see Corollary 2.49). Then ((n, m) is Erdos-Renyi Family of
Random Graphs, which is the discrete probability space so that:
(1) The sample space is the set of all graphs with n vertices and m edges and
(2) The probability function assigns a uniform probability:
p(G) =
__
n
2
_
m
_
1
to each graph in the sample space.
Remark 9.23. Erd os-Renyi graphs were developed by P. Erdos and A. Renyi in 1959
while studying the interactions of probability theory and discrete mathematics [ER60,
ER59]. Before giving an example of an Erd os-Renyi Family of Random Graphs, we re-
quire three lemmas, which will be useful.
Lemma 9.24. Consider the Path Graph P
n
; that is the graph on n vertices that is itself
a path. The isomorphism type of P
n
contains exactly n!/2 distinct graphs.
Proof. Consider, without loss of generality the example graph shown in Figure 9.2 For
Figure 9.2. A path graph with 4 vertices has exactly 4!/2 = 12 isomorphic graphs
obtained by rearranging the order in which the vertices are connected.
n vertices, we could arrange these vertices in any of n! ways. However, (by way of example)
the graph in which the vertices are arranged in order 1 to n with vertex i adjacent to vertex
i 1 and vertex i + 1 (save for vertices 1 and n, which are only adjacent to vertices 2 and
n 1 respectively) has an identical edge set to the graph with the vertices arranged in the
order n to 1. Thus, to obtain the size of the isomorphism type of P
n
, we must divide n! by 2
to remove duplicates. Thus the size of the isomorphism type of P
n
is n!/2. This completes
the proof.
Lemma 9.25. There are exactly n +1 distinct graphs in the isomorphism type of S
n
(the
star graph on n vertices).
Proof. By way of example, consider the graph S
3
shown in Figure 9.3. We can choose
any of the 4 vertices to be the center of the star and obtain a dierent edge relation. For S
n
,
which contains n + 1 vertices, there are n + 1 possible vertices that could act as the center
of the star. Thus there are n + 1 distinct graphs in the isomorphism class. This completes
the proof.
Remark 9.26. In the previous two lemmas, we were interested in the number of distinct
graphs in the isomorphism class. Here distinct means that the edge sets are dierent. In the
case of the star graph shown in Figure 9.3 this means that the graph in which vertex 1 is
123
Figure 9.3. There are 4 graphs in the isomorphism class of S
3
, one for each possible
center of the star.
the center is distinct from the star graph with vertex 4 at the center because the edge sets
in these two instances would be:
E = 0, 1, 0, 2, 0, 3 E
= 3, 0, 3, 1, 3, 2
Note this is distinct from the number of automorphisms, which might be quite a bit higher
because two distinct automorphism might create the same edge set. (For example, in the
gure if we map the vertex 1 to 3 and leave all other vertices unchanged, this is an automor-
phism, but the edge structure has not changed.)
Lemma 9.27. The number of distinct graphs in the isomorphism class of K
n
is 1.
Exercise 81. Prove Lemma 9.27.
Example 9.28. This example appears in Chapter 11.3 of [GY05]. We just explain it
a little more completely. Consider the random graph family ((5, 3). This family of graphs
contains
__
5
2
_
3
_
=
_
10
3
_
= 120
Some of these graphs are isomorphic however. In fact there are exactly 4 isomorphism
classes contained in ((5, 3), which we illustrate through exhaustive enumeration. Consider
the graph shown in Figure 9.4(a): This graph consists of an isolated vertex (in the gure
Vertex 5) and a copy of P
4
. There are 5 ways that the isolated vertex can be chosen and
from Lemma 9.24 we know there are 4!/2 = 12 elements in the isomorphism class of P
4
.
Thus there are 60 graphs isomorphic to the on in Figure 9.4(a) in ((5, 3).
Another type of graph we might consider is shown in Figure 9.4(b). This graph has an
isolated vertex and a copy of the star graph S
3
. Again, there are 5 distinct ways to choose
a vertex as the isolated vertex and by Lemma 9.25 there are 4 distinct graphs in the S
3
isomorphism class. Thus there are 20 distinct graphs isomorphic to the on in Figure 9.4(b)
in ((5, 3).
124
1 2 3 4
5
(a)
1
2
3
4 5
(b)
1
2
3
4
5
(c)
1
2
3
4
5
(d)
Figure 9.4. The 4 isomorphism types in the random graph family ((5, 3). We
show that there are 60 graphs isomorphic to this rst graph (a) inside ((5, 3), 20
graphs isomorphic to the second graph (b) inside ((5, 3), 10 graphs isomorphic to
the third graph (c) inside ((5, 3) and 30 graphs isomorphic to the fourth graph (d)
inside ((5, 3).
The third graph type in ((5, 3) is shown in Figure 9.4(c). Here there are two isolated
vertices and one copy of K
3
. By Lemma 9.27 we know there is only one distinct element in
the isomorphism class of K
3
, but there are
_
5
2
_
= 10 ways to choose the two isolated vertices.
Thus there are 10 distinct graphs isomorphic to the on in Figure 9.4(c) in ((5, 3).
In Figure 9.4(d) we have the nal graph type that appears in ((5, 3). We can tell this
because we have a single copy of K
2
which has only one distinct element in its isomorphism
class by Lemma 9.27 and a copy of S
2
, which by Lemma 9.25 has 3 elements in its isomor-
phism class. There are
_
5
2
_
= 10 to pick the two vertices that form K
2
thus there are 30
distinct graphs isomorphic to the on in Figure 9.4(d) in ((5, 3).
Since 60 + 20 + 10 + 30 = 120 we know we have enumerated all the distinct graphs in
((5, 3). This yields some interesting properties. For example, it we let X be the random
variable that returns the number of isolated vertices assuming we draw a graph (sample) at
random from the family ((5, 3), then we see that:
(9.5) E(X) = (1)
60
120
+ (1)
20
120
+ (2)
10
120
+ (0)
30
120
=
100
120
=
5
6
We might instead dene Y be to the random variable that is 1 if and only if there is a copy
of K
3
(as a subgraph) of a graph chosen at random from ((5, 3) and 0 else. Then Y is a
Bernoulli Random Variable (see Remark 9.2.) In this case, its easy to see that:
(9.6) Pr(Y = 1) =
10
120
=
1
12
since only the graphs isomorphic to the graph in Figure 9.4(c) contain a copy of K
3
.
Remark 9.29. Given a graph G = (V, E) (possibly drawn from a random graph family)
we say that G has a copy of K
n
if there is a subgraph H of G that is isomorphic to K
n
.
Figure 9.4 obviously contains a copy of K
3
and every graph with at least one edge contains
a copy of K
2
.
Theorem 9.30. The expected number of copies of K
s
in a graph chosen at random from
((m, n) is:
(9.7)
_
n
s
__
m
_
s
2
_
__
_
n
2
_
_
s
2
_
_
1
125
Proof. Dene a random variable Y to be 1 just in case there the vertex set v
1
, . . . , v
s
induces a complete sub-graph in a randomly chosen graph from ((m, n) and 0 otherwise.
We now ask the question, what is the probability that Y = 1 (or equivalently what is E(Y )).
If we let this probability be p then, from Remark 9.2, the expected number of copies of K
s
is simply:
_
n
s
_
p
since there are
_
n
s
_
ways to pick these s vertices and by Lemma 9.27 the isomorphism class
of K
n
contains only one element. To compute p, consider the following. There are out of all
the graphs in ((m, n), we must choose one in which k =
_
s
2
_
edges are proscribed (because
they link the elements of v
1
, . . . , v
s
). We now may select m k edges from a possible
collection of
_
n
2
_
k edges. Thus, there are:
__
n
2
_
k
mk
_
ways this can be done. This means the probability of choosing any one of these graphs is:
(9.8) p =
__
n
2
_
k
mk
___
n
2
_
m
_
1
since there are:
__
n
2
_
m
_
graphs in the family ((m, n) from Denition 9.22. Simplifying Equation 9.8 we obtain:
(9.9) p =
__
n
2
_
k
mk
__
n2
m
_
1
=
__
n
2
_
k
_
!
__
n
2
_
k (mk)
_
!(mk)!
m!
__
n
2
_
m
_
!
_
n
2
_
!
=
__
n
2
_
k
_
!
(mk)!
m!
_
n
2
_
!
=
__
n
2
_
k
_
!
_
n
2
_
!
m!
(mk)!
=
__
n
2
_
k
_
!k!
_
n
2
_
!
m!
k!(mk)!
=
__
n
2
_
k
_
1
_
m
k
_
Thus we see that the expected number of copies of K
s
in a graph chosen at random from
((m, n) is:
(9.10)
_
n
s
__
m
k
___
n
2
_
k
_
1
=
_
n
s
__
m
_
s
s
_
__
_
n
2
_
_
s
s
_
_
1
This completes the proof.
Lemma 9.31. There are exactly (n1)!/2 distinct graphs in the isomorphism type of C
n
(the cycle graph on n vertices).
Proof. We will build a cycle iteratively using the n vertices allotted. Choose a vertex
at random from the n vertices and denote it v
1
. This vertex will be joined by an edge to
exactly one vertex v
2
. There are n 1 possible choices for v
2
. Now, v
2
will be joined by an
edge to exactly one (new) vertex v
3
for which we have n 2 possible choices. We can now
see that the number of ways of making a such a cycle is (n1)(n2) (1). However, The
cycle that rst creates an edge from v
1
to v
2
and then v
2
to v
3
and so on is precisely the
126
same as the cycle that rst creates and edge from v
1
to v
n
and then from v
n
to v
n1
and so
on. Thus, the value (n1)(n2) (1) double counts each cycle once. Therefore, the total
number of distinct graphs in the isomorphism class of C
n
is (n 1)!/2.
Theorem 9.32. The expected number of copies of C
k
in a random graph chosen from
the family ((m, n) is:
(9.11)
(k 1)!
2
_
n
k
__
m
k
___
n
2
_
k
_
1
Proof. We follow the proof of Theorem 9.30. Dene a random variable Y to be 1 just
in case there the vertex set v
1
, . . . , v
s
induces a cycle C
k
in a randomly chosen graph from
((m, n) and 0 otherwise. We now ask the question, what is the probability that Y = 1
(or equivalently what is E(Y )). If we let this probability be p then, from Remark 9.2 and
Lemma 9.31, the expected number of distinct copies of C
k
is simply:
(n 1)!
2
_
n
k
_
p
since there are
_
n
k
_
ways to choose the k vertices in C
k
and (n 1)!/2 distinct isomorphic
copies of each cycle C
k
on k vertices. Further, since C
k
has k edges, we have already derived
the value of p in the proof of of Theorem 9.30. We have:
p =
__
n
2
_
k
_
1
_
m
k
_
Thus we conclude that the expected number of copies of C
k
in a random graph chosen from
the family ((m, n) is
(k 1)!
2
_
n
k
__
m
k
___
n
2
_
k
_
1
This completes the proof.
Exercise 82. Let H be a graph on k vertices whose isomorphism class contains exactly
s distinct elements. State and prove a theorem on the expected number of copies of H in a
random graph chosen from the family ((m, n). [Hint: Theorem 9.32 is an example of such
a general theorem. Extrapolate from this.]
Remark 9.33. The spaces ((n,
1
2
) and ((n, m) are closely related to each other. Consider
((n,
1
2
) restricted to only those graphs with exactly m edges. The probability of any one of
these graphs in ((n,
1
2
) is:
(9.12)
_
1
2
_
m
_
1
2
_
(
n
2
)m
That is, they all have equal probability in ((n,
1
2
). But, if we were to compute the conditional
probability of any of these graphs in ((n,
1
2
) given that we require a graph to have m edges,
then their probabilities all reduce to precisely the probability one expects in the model
((n, m) by the properties of conditional probability.
127
CHAPTER 10
Some More Algebraic Graph Theory
I have never actually taught this chapter. It is suitable for a very advanced class or as a substitute for
an earlier chapter.
129
dened. In general, all you need to dene a vector space is a eld (the scalars) a group (the
vectors) and a multiplication operation (scalar-vector multiplication) that connects the two
and that satises all the properties listed in Denition 10.2.
Remark 10.6. Note that dot products are not necessarily dened over general vector
spaces. They are, however, dened over vector spaces in which vectors are n-tuples (or
column / row vectors) with elements drawn from the underlying eld as in R
n
.
Definition 10.7 (Subspace). If 1 = (F, +, , 0, 1, V, +, ) is a vector space and U V
with | = (F, +, , 0, 1, U, +, ) also a vector space then, | is called a subspace of 1. Note
that U must be closed under + and .
Example 10.8. If we consider R
3
as a vector space over the reals (as usual) then it has
as a subspace several copies of R
2
. The easiest is to consider the subset of vectors:
U = (x, y, 0) : x, y R
Clearly U is closed under the addition and scalar multiplication of the original vector space.
Definition 10.9 (Linear Transformation). Let 1 and | be two vector spaces (over some
appropriately dened elds). A linear transformation is a function f : 1 | such that:
(1) f(v
1
+v
2
) = f(v
1
) +f(v
2
) for all vectors v
1
and v
2
in vector space 1 and
(2) f(sv) = sf(v) for all vectors v and scalars s in vector space 1.
Example 10.10. Consider the vector spaces R
2
and R
3
and the matrix:
M =
_
_
1 2
3 4
5 6
_
_
Then the function f : R
2
R
3
with f(x) = Mx is a linear transformation. To see this,
consider
x
1
=
_
x
11
x
12
_
x
2
=
_
x
21
x
22
_
Then we have:
M(x
1
+x
2
) =
_
_
1 2
3 4
5 6
_
_
__
x
11
x
12
_
+
_
x
21
x
22
__
=
_
_
1 2
3 4
5 6
_
_
_
x
11
+x
21
x
12
+x
22
_
=
_
_
(x
11
+x
21
) + 2(x
12
+x
22
)
3(x
11
+x
21
) + 4(x
12
+x
22
)
5(x
11
+x
21
) + 6(x
12
+x
22
)
_
_
=
_
_
(x
11
+ 2x
12
) + (x
21
+ 2x
22
)
(3x
11
+ 4x
12
) + (3x
12
+ 4x
22
)
(5x
11
+ 6x
12
) + (5x
12
+ 6x
22
)
_
_
=
_
_
x
11
+ 2x
12
3x
11
+ 4x
12
5x
11
+ 6x
12
_
_
+
_
_
x
21
+ 2x
22
3x
12
+ 4x
22
5x
12
+ 6x
22
_
_
=
_
_
1 2
3 4
5 6
_
_
_
x
11
x
12
_
+
_
_
1 2
3 4
5 6
_
_
_
x
21
x
22
_
= Mx
1
+Mx
2
A similar argument will show that M(cx) = cMx for all vectors x R
2
and all scalars
c R.
130
Remark 10.11. It is (relatively) easy to generalize the previous example to see that
(left) multiplication of a matrix M R
mn
by a vector x R
n1
constitutes a linear
transformation from R
n
to R
m
.
2. Linear Span and Basis
Remark 10.12. This section is a review of the material covered in Chapter 5.1. We now
generalize it to an arbitrary eld and vector space.
Definition 10.13. Let x
1
, . . . , x
m
be vectors in a vector space 1 and let
1
, . . . ,
m
be
scalars. Then
(10.1)
1
x
1
+ +
m
x
m
is a linear combination of the vectors x
1
, . . . , x
m
.
Definition 10.14 (Span). Let A = x
1
, . . . , x
m
be a set of vectors in a vector space
1, then the span of A is the set:
(10.2) span(A) = y 1 : y is a linear combination of vectors in A
Definition 10.15 (Linear Independence). Let x
1
, . . . , x
m
be vectors in vector space 1.
The vectors x
1
, . . . , x
m
are linearly dependent if there exists scalars
1
, . . . ,
m
, not all zero,
such that
(10.3)
1
x
1
+ +
m
x
m
= 0
If the set of vectors x
1
, . . . , x
m
is not linearly dependent, then they are linearly independent
and Equation 10.3 holds just in case
i
= 0 for all i = 1, . . . , n.
Remark 10.16. It is worthwhile to note that the zero vector 0 makes any set of vectors
a linearly dependent set.
Exercise 83. Prove the remark above.
Definition 10.17 (Basis). Let A = x
1
, . . . , x
m
be a set of vectors in vector space
1. The set A is called a basis of 1 if A is a linearly independent set of vectors and every
vector in 1 is in the span of A. That is, for any vector w 1 we can nd scalar values
1
, . . . ,
m
F such that
(10.4) w =
m
i=1
i
x
i
Remark 10.18. The following statements on the size of a bases in vectors spaces are
outside the scope of this course. Proof can be found in [Lan87].
Theorem 10.19. Every basis of a vector space 1 has precisely the same cardinality.
Definition 10.20 (Dimension). The cardinality of any basis of a vector space 1 is called
the dimension of the vector space.
Remark 10.21. Theorem 10.19 ensures that the dimension of a vector space is uniquely
specied.
131
3. Vector Spaces of a Graph
Definition 10.22 (Galois Field with 2 Elements). The Galois Field with 2 elements
(denoted GF2) is the eld (0, 1, +, , 0, 1) where:
(1) 0 + 0 = 0,
(2) 0 + 1 = 1,
(3) 1 + 1 = 0,
(4) 0 0 = 0,
(5) 0 1 = 0, and
(6) 1 1 = 1
Remark 10.23. The eld GF2 is the rst of an innite number of nite elds that have
several practical applications. The interested reader should consult [LP97] for an excellent
introduction to applied abstract algebra.
Definition 10.24 (Symmetric Dierence). Let S
1
and S
2
be any two sets. The symmet-
ric dierence of S
1
and S
2
is:
(10.5) S
1
+S
2
= (S
1
S
2
) (S
2
S
1
)
That is:
S
1
+S
2
= s S
1
S
2
: s is in exactly one of S
1
or S
2
with:
i
=
_
1 e
i
S
0 else
Thus B = e
1
, . . . , e
m
spans c. To see that B is linearly independent, note that:
=
1
e
1
+
2
e
2
+ +
m
e
m
is only possible if
1
=
2
= =
m
= 0 since the elements of B are mutually disjoint.
The fact that the dimension of c is [E[ follows at once from Denition 10.20.
Exercise 85. State and prove a similar Theorem to Theorem 10.31 for 1.
Definition 10.32 (Characteristic Vector). Let G = (V, E) Let S be a vector in c. The
characteristic vector of S is the m-dimensional vector of values drawn from GF2 so that:
S =
1
e
1
+
2
e
2
+ +
m
e
m
Remark 10.33. Obviously for any subset of V we may likewise dene a characteristic
vector based on Exercise 85. This means that c can really be considered to be the vector
space GF2
n
the set of m-tuples of elements of GF2 for a graph G = (V, E) with [E[ = n.
Likewise, if [V [ = m, then 1 is equivalent to the vector space GF2
m
. The following theorem
is now straightforward to prove.
Theorem 10.34. Let G = (V, E) and let M be the incidence matrix of G. Then M is
a linear transformation from c to 1 when elements of c are treated as their characteristic
vectors and elements of 1 are treated as their characteristic vectors.
Remark 10.35. Before proving Theorem 10.34 it is important to note that all operations
are performed over GF2 and not the real numbers.
Proof of Theorem 10.34. The fact that multiplication by M is a linear transform
from c to 1 is a result of the fact that matrix multiplication is always a linear transform
from one vector space over a eld to another vector space over the same eld (see Remark
10.11).
Exercise 86. Consider a graph G = (V, E) and its incidence matrix M. Let x be the
characteristic vector for a standard basis vector in c (a vector corresponding to the one
element edge sets of c). What is the result (in 1) of the transformation Mx?
4. Cycle Space
Definition 10.36 (Cycle Space). Let G = (V, E) be a graph. The cycle space of G is
an element of 2
E
denoted ( and is the smallest set (of sets) containing the , all cycles in G
(where each cycle is treated as a set of edges) and all unions of edge disjoint cycles in G.
133
Example 10.37. We show an example of the cycle space. The cycle space of a graph can
be thought of as all the cycles contained in that graph along with the subgraphs consisting
of cycles that only share vertices but no edges.
G
C
Figure 10.1. The cycle space of a graph can be thought of as all the cycles con-
tained in that graph along with the subgraphs consisting of cycles that only share
vertices but no edges. This is illustrated in this gure.
Theorem 10.38. Let G = (V, E) be a graph. The cycle space ( is a subspace of c.
Proof. It suces to prove that ( is closed under + since it is trivially closed under
vector-scalar multiplication. Consider two elements C
1
and C
2
(.
(1) If c
1
and c
2
are the edge disjoint cycles (or the empty set) then C
1
+ C
2
is simply
the union of edge disjoint cycles and clearly C
1
+C
2
(.
(2) If c
1
is a union of edge disjoint cycles and C
2
is the union of edge disjoint cycles (or
the empty set) and C
1
and C
2
share no edges in common, then it is again easy to
see that C
1
+C
2
is simply the union of edge disjoint cycles and in (.
(3) If C
1
and C
2
are unions of edge disjoint cycles but share edges in common, then C
1
and C
2
must share two or more cycles in common. In constructing C
1
+ C
2
these
common cycles will be removed (through symmetric dierencing) and the result will
be a union of edge disjoint cycles, which is clearly in (.
Thus, ( is closed under + and it must be a vector space.
Definition 10.39 (Fundamental Cycle). Let G = (V, E) and let F = (V, E
) be a
spanning forest (a spanning subgraph of G that is also a forest). A cycle is fundamental
(with respect to F and G) if it is a cycle created by adding an edge e in E to F that is not
in E
.
Example 10.40. We illustrate the creation of a fundamental cycle in the graph in Figure
10.2.
Definition 10.41 (Fundamental System of Edge Cycles). Let G = (V, E) and let F =
(V, E
) be a spanning
tree of G. Then the fundamental system of cycles of G and T forms a basis for (.
Proof. Recall rst that by Theorem 3.65 such a spanning tree T exists when G is
connected. Consider any fundamental cycle C. This cycle is constructed by adding exactly
one edge to T and nding the cycle that results. Thus no fundamental cycle can be expressed
as the sum of any other fundamental cycles because they will all be missing (at least) one
edge. As a result, the fundamental system of cycles must be linearly independent.
Choose any element C of ( and let e
1
, . . . , e
r
be the edges of C that do not appear in
E
i
=
_
1 if the non-tree edge of C
i
is found in C
0 else
Thus, the fundamental set of cycles is linearly independent and spans ( and so it is a basis.
This completes the proof.
Remark 10.43. The dimension of the space ( for a graph G = (V, E) is called the
cyclomatic number of G and is, of course, equal to the size of (any) set of fundamental cycles
generated by a spanning tree of G.
Exercise 88. The statement of Theorem 10.42 is stated and (more or less) proved
following [GY05]. Diestel [Die10] has a dierent way of dening the basis that does not
require Gto be connected. Note that fundamental systems of cycles were dened for arbitrary
graphs (rather than just connected ones). Is there any reason we couldnt just replace the
135
spanning tree T with a spanning forrest (invoking Corollary 3.66) and state and prove a
more general result?
5. Cut Space
Definition 10.44 (Cut Space). Let G = (V, E) be a graph. The cut space of G is an
element of 2
E
denoted (
and is the smallest set (of sets) containing the , all minimal edge
cuts in G and all unions of edge disjoint minimal edge cuts in G.
Example 10.45. We show an example of the cut space. The cut space of a graph can
be thought of as all the minimal cuts contained in that graph along with the subgraphs
consisting of minimal cuts that only share vertices but no edges.
G
C
Figure 10.3. The cut space of a graph can be thought of as all the minimal cuts
contained in that graph along with the subgraphs consisting of minimal cuts that
only share vertices but no edges. This is illustrated in this gure.
Theorem 10.46. Let G = (V, E) be a graph. The cut space (
is a subspace of c.
Proof. The proof is almost identical to the proof of Theorem 10.38, except we are
dealing with elements made up of cuts, rather than cycles.
Exercise 89. Using the proof of Theorem 10.38 construct a complete proof for Theorem
10.46.
Definition 10.47 (Partition Cut). Let G = (V, E) be a graph and suppose that V =
V
1
V
2
with V
1
V
2
= . That is, V
1
and V
2
constitute a two element partition of V . Then
136
the partition cut generated by V
1
and V
2
is the set of edges connecting elements in V
1
with
elements in V
2
. The partition cut is sometimes denotes V
1
, V
2
.
Lemma 10.48. Let G = (V, E) be a connected graph and suppose that V
1
and V
2
partition
V . Then the partition cut V
1
, V
2
is a minimal edge cut.
Exercise 90. Prove Lemma 10.48. [Hint: The graph G is connected, so we may assume
that the subgraphs induced by V
1
and V
2
are connected since V
1
, V
2
consists only of edges in
which one end resides in V
1
and the other in V
2
. Suppose you remove one edge from V
1
, V
2
.
The graph that results from removing these edges would still have a link connecting one
vertex in V
1
with one vertex in V
2
. The rest of the proof should be easy using the denition
of connectivity.]
Definition 10.49 (Fundamental Edge Cut). Let G = (V, E) and let F = (V, E
) be
a spanning forest (a spanning subgraph of G that is also a forest). Let V
1
and V
2
be the
vertices of the two new components formed when an edge e is removed from the forest F.
Then the partition cut V
1
, V
2
is a fundamental edge cut.
Proposition 10.50. Every fundamental edge cut is minimal.
Proof. Apply Lemma 10.48 to the specic component of G in which the edge removed
from the spanning forest resides.
Example 10.51. We illustrate the creation of a fundamental edge-cut in the graph in
Figure 10.4.
Fundamental Edge Cut
Spanning Tree
Tree Edge to Remove
Figure 10.4. A fundamental edge cut of a graph G (with respect to a spanning
forest F) is a partition cut created from partitioning the vertices based on a cut in
a spanning tree and then constructing the resulting partition cut.
Definition 10.52 (Fundamental System of Edge Cuts). Let G = (V, E) and let F =
(V, E
) be a spanning
tree of G. Every minimal edge cut of G contains at least one element of T.
Proof. Let C be any minimal edge cut of G. We rst show that the graph GC has
2 components. If G C has 1 component, then C is not an edge cut, contradicting our
assumption. If GC has three components, then choose two of them. There must be edges
connecting vertices in these two components in C because G was connected. Restoring any
137
one of these edges will result in a graph that still has more than 1 component (since we
assumed there were more than two components) and thus E
is not minimal.
Let V
1
and V
2
be the vertices of the two components that result from the removal of C
from G. The fact that T is a spanning tree means that there is some (exactly one) edge
e E
connecting a vertex in V
1
to a vertex in V
2
. This edge must be in C, otherwise, V
1
is
connected to V
2
and C is not a cut. This completes the proof.
Theorem 10.54. Let G = (V, E) be a connected graph. A set E
G is an edge cut if
and only if every spanning tree of G contains at least one edge in E
.
Exercise 92. Prove Theorem 10.54. [Hint: Use Theorem 3.72 or Lemma 10.53 or both.]
Theorem 10.55. Let G = (V, E) be a connected graph and let T = (V, E
) be a spanning
tree of G. Then the fundamental system of edge cuts of G and T forms a basis for (
.
Proof. To see that the fundamental system of edge cuts is linearly independent, note
that each system of edge cuts is uniquely dened by the removal of one edge from the spanning
tree T and therefore this edge is an element of that cut. If V
1
and V
2
is the resulting partition,
then there is exactly one edge in the spanning tree in the cut V
1
, V
2
(otherwise there would
be more than one path connecting two vertices in the spanning tree) and thus this partition
cut is uniquely dened by that edge. Thus no fundamental edge cut can be expressed as the
sum of any other fundamental cycles because they will all be missing (at least) one edge. As
a result, the fundamental system of edge cuts must be linearly independent.
Choose any element C of (
and let e
1
, . . . , e
r
be the edges of C that appear in E
.
Further dene the fundamental edge cut C
i
to be the one that arises as a result of removing
e
i
from T.
It is easy to see that no edge in the set e
1
, . . . , e
r
appears in C +C
1
+ +C
r
because
each edge appears once in C and once in one of the C
i
s and therefore, it will not appear in
C +C
1
+ +C
r
. But this means that every edge in C +C
1
+ +C
r
is an edge that does
not appear in T. Every element in (
i
=
_
1 if the edge of C has edge e
i
found in E
0 else
Thus, the fundamental set of edge cuts is linearly independent and spans (
and so it is a
basis. This completes the proof.
Definition 10.56 (Cycle Rank / Edge-Cut Rank). Let G = (V, E) be a graph. The
edge-cut rank of G is the number of edges in a spanning forest of G. The cycle rank (or
Betti number) of G, denoted (G), is the number of edges in the relative complement of
that spanning forest in G (see Denition 2.68).
138
Proposition 10.57. Let G = (V, E) be a graph. The Betti number of G is:
(G) = [E[ [V [ c(G)
While the edge cut rank of G is [V [ c(G).
Exercise 93. Prove Proposition 10.57. [Hint: Use Corollary 3.71.]
Corollary 10.58. Let G = (V, E) be a graph. The dimension of ( is (G), while the
dimension of the edge cut space is the edge cut rank, which is [V [ c(G).
Proof. This follows immediately from the proofs of Theorems 10.42 and 10.55.
6. The Relation of Cycle Space to Cut Space
Lemma 10.59. Let G = (V, E) be a connected graph. Any cycle in G has an even number
of edges in common with any minimal edge cut of G.
Proof. Let C be any cycle and suppose that E
and the theorem is proved. Therefore, assume that C contains some vertices from
V
1
and some from V
2
. Consider the process of walking along C beginning at some vertex in
V
1
. At some point we must cross into V
2
along an edge in E
. Similarly,
if K is in the edge cut space (
are orthogonal
subspaces.
Proof. Choose any elements C ( and / (
.
These subspaces are orthogonal complements if and only if ( (
= .
Proof. The orthogonality of ( and (
= , then ( (
) = dim(() + dim((
) dim(( (
)
From Corollary 10.58, we know that:
dim(() = [E[ [V [ c(G)
dim((
) = [V [ c(G)
Thus:
dim(( (
) = [E[
if and only if
dim(( (
) = 0
The latter can only happen when ( (
= c and (
and (