Graph Algorithms, 2nd Edition

Shimon Evens Graph Algorithms, published in 1979, was a seminal introductory book
on algorithms read by everyone engaged in the eld. This thoroughly revised second
edition, witha forewordbyRichardM. Karpandnotes byAndrewV. Goldberg, continues
the exceptional presentation from the rst edition and explains algorithms in formal but
simple language with a direct and intuitive presentation.
The material covered by the book begins with basic material, including graphs and
shortest paths, trees, depth-rst search, and breadth-rst search. The main part of the
book is devoted to network ows and applications of network ows. The book ends with
two chapters on planar graphs and on testing graph planarity.
S HI MON E VE N (19352004) was a pioneering researcher on graph algorithms and
cryptography. He was a highly inuential educator who played a major role in establish-
ing computer science education in Israel at the Weizmann Institute and the Technion.
He served as a source of professional inspiration and as a role model for generations
of students and researchers. He is the author of Algorithmic Combinatorics (1973) and
Graph Algorithms (1979).
Graph Algorithms
2nd Edition
SHIMON EVEN
Edited by
GUY EVEN
Tel-Aviv University
CAMBRI DGE UNI VE RS I T Y P RE S S
Cambridge, New York, Melbourne, Madrid, Cape Town,
Singapore, So Paulo, Delhi, Tokyo, Mexico City
Cambridge University Press
32 Avenue of the Americas, New York, NY 10013-2473, USA
www.cambridge.org
Information on this title: www.cambridge.org/9780521736534
Shimon Even 1979
Shimon Even and Guy Even 2012
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First edition published 1979 by Computer Science Press
Second edition published 2012
Printed in the United States of America
A catalog record for this publication is available from the British Library.
Library of Congress Cataloging in Publication Data
ISBN 978-0-521-51718-8 Hardback
ISBN 978-0-521-73653-4 Paperback
Cambridge University Press has no responsibility for the persistence or accuracy of URLs
for external or third-party Internet Web sites referred to in this publication and does not
guarantee that any content on such Web sites is, or will remain, accurate or appropriate.
Contents
Foreword by Richard M. Karp page vii
Preface to the Second Edition ix
Preface to the First Edition xi
1 Paths in Graphs 1
1.1 Introduction to Graph Theory 1
1.2 Computer Representation of Graphs 3
1.3 Euler Graphs 6
1.4 De Bruijn Sequences 9
1.5 Shortest-Path Algorithms 11
1.6 Problems 22
2 Trees 29
2.1 Tree Denitions 29
2.2 Minimum Spanning Tree 31
2.3 Cayleys Theorem 34
2.4 Directed Tree Denitions 37
2.5 The Innity Lemma 39
2.6 Problems 42
3 Depth-First Search 46
3.1 DFS of Undirected Graphs 46
3.2 Algorithm for Nonseparable Components 52
3.3 DFS on Directed Graphs 57
3.4 Strongly Connected Components of a Digraph 58
3.5 Problems 62
4 Ordered Trees 65
4.1 Uniquely Decipherable Codes 65
v
vi Contents
4.2 Positional Trees and Huffmans Optimization Problem 69
4.3 Application of the Huffman Tree to Sort-by-Merge
Techniques 75
4.4 Catalan Numbers 77
4.5 Problems 82
5 Flow in Networks 85
5.1 Introduction 85
5.2 The Algorithm of Ford and Fulkerson 87
5.3 The Dinitz Algorithm 94
5.4 Networks with Upper and Lower Bounds 102
5.5 Problems 109
5.6 Notes by Andrew Goldberg 115
6 Applications of Network Flow Techniques 117
6.1 Zero-One Network Flow 117
6.2 Vertex Connectivity of Graphs 121
6.3 Connectivity of Digraphs and Edge Connectivity 129
6.4 Maximum Matching in Bipartite Graphs 135
6.5 Two Problems on PERT Digraphs 137
6.6 Problems 141
7 Planar Graphs 146
7.1 Bridges and Kuratowskis Theorem 146
7.2 Equivalence 157
7.3 Eulers Theorem 158
7.4 Duality 159
7.5 Problems 164
8 Testing Graph Planarity 168
8.1 Introduction 168
8.2 The Path Addition Algorithm of Hopcroft and Tarjan 169
8.3 Computing an st-Numbering 177
8.4 The Vertex Addition Algorithm of Lempel, Even, and
Cederbaum 179
8.5 Problems 185
Index 187
Foreword
In Appreciation of Shimon Even
Shimon was a great computer scientist who inspired generations of Israeli
stutents and young researchers, including many future leaders of theoretical
computer science.
He was a master at creating combinatorial algorithms, constructions, and
proofs. He always sought the simplest and most lucid solutions. Because he
never allowedhimself touse a knowntheoremunless he understoodits proof, his
discoveries were often based on original methods. His lectures were legendary
for their clarity.
Shimon was devoted to his family, generous to his colleagues, and freely
available to the students in his classes.
He expressed his views forcefully and with complete honesty. He expected
honesty in return, and reserved his disapproval for those who tried to obfuscate
or mislead.
Shimon had an unending supply of interesting anecdotes, and would laugh
uproariously at good jokes, including his own.
In sum, he was a great and unforgettable man and a great scientist, and his
name has a permanent place in the annals of theoretical computer science.
Richard M. Karp
Berkeley, April 2011
vii
Preface to the Second Edition
My father, Shimon Even, died on May 1, 2004. In the year prior to his illness,
he began revising this book. He used to tell me with great satisfaction whenever
he completed the revision of a chapter. To his surprise, he often discovered
that, after twenty-ve years, he preferred to present the material differently (the
rst edition was published in 1979). Unfortunately, he only managed to revise
Chapters 1, 2, 3, and 5. These revised chapters appear in this edition. However,
since the material in Chapters 9 and 10 on NP-completeness is well covered in
a few other books, we decided to omit these chapters from the second edition.
Therefore, the second edition contains only the rst eight chapters.
As I was reading the manuscript for the second edition, my fathers deep voice
resonated clearly in my mind. Not only his voice, but also his passion for teach-
ing, for elegant explanations, and, most importantly, for distilling the essence.
As an exceptional teacher, he used his voice and his physique to reinforce his
arguments. His smile revealed how happy he was to have the opportunity to
tell newcomers about this wonderful topic. One cannot overvalue the power of
such enthusiasm. Luckily, this enthusiasm is conveyed in this book.
Many people tell me (with a smile) about being introduced to the topic of
algorithms through this book. I believe the source of their smiles is its outstand-
ing balance between clarity and preciseness. When one writes mathematical
text, it is very easy to get carried away with the desire to be precise. The written
letter is long lasting, and being aware that ones text leaves certain gaps requires
boldness. For my father this task was trivial. The audience he had in mind con-
sisted simply of himself. He wrote as he would have wanted the material to be
presented to him. This meant that he elaborated where he needed to, and he did
not hesitate to skim over details he felt comfortable with. As a child, I recall
seeing him prepare for class by reading a chapter from his book. I asked him:
ix
x Preface to the Second Edition
Why are you reading your own book? Presumably, you know what is there.
True, he replied, but I dont remember!
This second edition would have never been completed without Oded
Goldreich. Oded Goldreich began by convincing me to prepare the second edi-
tion. Then he put me in touch with Lauren Cowles from Cambridge University
Press. Finally, he continuously encouraged me to complete this project. It took
almost seven years! There is no good excuse for it. We all knowhowsuch a task
can be pushed aside by more urgent and demanding tasks. Apparently, it
took me some time to realize how important this task was, and that it could not
be completed without a coordinated effort. Only after I recruited LotemKaplan
to do the typesetting of the unrevised chapters and complete the missing gure
and index terms did this project begin to progress seriously. I am truly grateful
to Oded for his insistence, to Lotem for her assistance, and to Lauren for her
kind patience.
Finally, I wish to thank Richard M. Karp, an old friend of my fathers, for his
foreword. I also wish to thank Andrew Goldberg, the expert in network ow
algorithms, for the notes he contributed in Chapter 5. These notes outline the
major developments in the algorithms for maximumow that have taken place
since the rst edition of this book was published.
Guy Even
Tel-Aviv, March 2011
Preface to the First Edition
Graph theory has long been recognized as one of the more useful mathematical
subjects for the computer science student to master. The approachthat is natural
to computer science is the algorithmic one; our interest is not so much in the
existence proofs or enumerationtechniques as it is in nding efcient algorithms
for solving relevant problems or, alternatively, in showing evidence that no such
algorithmexists. Although algorithmic graph theory was started by Euler, if not
earlier, its development in the last ten years has been dramatic and revolutionary.
Much of the material in Chapters 3, 5, 6, 8, 9, and 10 is less than ten years old.
This book is meant to be a textbook for an upper-level undergraduate, or a
graduate course. It is the result of my experience in teaching such a course
numerous times, since 1967, at Harvard, the Weizmann Institute of Science,
Tel-Aviv University, the University of California at Berkeley, and the Tech-
nion. There is more than enough material for a one-semester course; I am sure
that most teachers will have to omit parts of the book. If the course is for
undergraduates, Chapters 1 to 5 provide enough material, and even then, the
teacher may choose to omit a few sections, such as 2.6, 2.7, 3.3, and 3.4.
1
Chapter 7 consists of classical nonalgorithmic studies of planar graphs, which
are necessary in order to understand the tests of planarity, described in Chapter
8; it may be assigned as a preparatory reading assignment. The mathematical
background needed for understanding Chapters 1 to 8 includes some knowl-
edge of set theory, combinatorics, and algebra, which the computer science
student usually masters during his freshman year through courses on discrete
mathematics and on linear algebra. However, the student also needs to know
something about data structures and programming techniques, or he may not
The rst edition was published in 1979 (G.E.).
1
Sections 2.6 and 2.7 were removed from the second edition by Shimon Even.
xi
xii Preface to the First Edition
appreciate the algorithmic side or may miss the complexity considerations. It
is my experience that after two courses in programming, students have the nec-
essary knowledge. However, in order to follow Chapters 9 and 10,
2
additional
background is necessary, namely, in theory of computation. Specically, the
student should know about Turing machines and Churchs thesis.
The book is self-contained. No previous knowledge is needed beyond the
general background just described. No comments such as the rest of the proof
is left to the reader or this is beyond the scope of this book are ever made.
Some unproved results are mentioned, with a reference, but are not used later
in the book.
At the end of each chapter, there are a few problems teachers can use for
homework assignments. The teacher is advised to use them discriminately,
since some of them may be too hard for his students.
I would like to thank some of my past colleagues for our joint work and for the
inuence they have had on my work, and therefore on this book: I. Cederbaum,
M. R. Garey, J. E. Hopcroft, R. M. Karp, A. Lempel, A. Pnuely, A. Shamir,
and R. E. Tarjan. Also, I would like to thank some of my former Ph.D. students
for all that I have learned from them: O. Kariv, A. Itai, Y. Perl, M. Rodeh,
and Y. Shiloach. Finally, I would like to thank E. Horowitz for his continuing
encouragement.
S.E., Techinion, Haifa, Israel
2
Chapters 9 and 10 are not included in the second edition.
1
Paths in Graphs
1.1 Introduction to Graph Theory
A graph G(V, E) is a structure consisting of a set of vertices V = {v
1
, v
2
, . . .}
and a set of edges E = {e
1
, e
2
, . . .}; each edge e has two endpoints, which are
vertices, and they are not necessarily distinct.
Unless otherwise stated, both V and E are assumed to be nite. In this case
we say that G is nite.
For example, consider the graph in Figure 1.1. Here, V = {v
1
, v
2
, v
3
, v
4
, v
5
},
E = {e
1
, e
2
, e
3
, e
4
, e
5
}. The endpoints of e
2
are v
1
and v
2
. Alternatively, we say
that e
2
is incident on v
1
and v
2
. The edges e
4
and e
5
have the same endpoints
and are therefore called parallel. Both endpoints of e
1
are the same v
1
; such
an edge is called a self-loop.
The degree of a vertex v, d(v), is the number of times v is used as an endpoint.
Clearly, a self-loop uses its endpoint twice. Thus, in our example, d(v
1
) = 4,
d(v
2
) =3. Also, a vertexwhose degree is zerois called isolated. In our example,
v
3
is isolated since d(v
3
) = 0.
Lemma 1.1 In a nite graph the number of vertices of odd degree is even.
Proof: Let |V| and |E| be the number of vertices and edges, respectively. It is
easy to see that
|V|

i=1
d(v
i
) = 2 |E|,
since each edge contributes two to the left-hand side: one to the degree of each
of its two endpoints if they are distinct; and two to the degree of its endpoint if
the edge is a self-loop. For the left-hand side to sum up to an even number, the
number of odd terms must be even.
1
2 1 Paths in Graphs
e
3
v
1
e
1
e
2
v
2
v
3
v
4
e
5
e
4
v
5
Figure 1.1: Example of a graph.
The notation u
e
v means that the edge e is incident on vertices u and v. In
this case we also say that e connects vertices u and v, or that vertices u and v
are adjacent.
A path, P, is a sequence of vertices and edges, interweaved in the following
way: P starts with a vertex, say v
0
, followed by an edge e
1
incident to v
0
,
followed by the other endpoint v
1
of e
1
, and so on. We write
P : v
0
e
1
v
1
e
2
v
2

If P is nite, it ends with a vertex, say v
l
. We call v
0
the start-vertex of P and
v
l
the end-vertex of P. The number of edge appearances in P, l, is called the
length of P. If l = 0, then P is said to be empty, but it has a start-vertex and
an end-vertex, which are identical. (We shall not use the term path unless a
start-vertex exists.)
In a path, edges and vertices may appear more than once, unless otherwise
stated. If no vertex appears more than once, and therefore no edge can appear
more than once, the path is called simple.
A circuit, C, is a nite path in which the start and end vertices are identical.
However, an empty path is not considered a circuit. By denition, the start and
end vertices of a circuit are the same, and if there is no other repetitionof a vertex,
the circuit is called simple. However, a circuit of length two, a
e
b
e
a, where
the same edge, e, appears twice, is not considered simple. (For a longer circuit,
it is superuous to state that if it is simple, then no edge appears more than
once.) A self-loop is a simple circuit of length one.
If for every two vertices u and v of a graph G, there is a (nite) path that starts
in u and ends in v, then G is said to be connected.
A digraph or directed graph G(V, E) is dened similarly to a graph, except
that the pair of endpoints of every edge is now ordered. If the ordered pair of
1.2 Computer Representation of Graphs 3
endpoints of a (directed) edge e is (u, v), we write
u
e
v.
The vertex u is called the start-vertex of e; and the vertex, v, the end-vertex of
e. The edge e is said to be directed fromuto v. Edges with the same start-vertex
and the same end-vertex are called parallel. If u,=v, u
e
1
v and v
e
2
u, then
e
1
and e
2
are antiparallel. An edge u u is called a self-loop.
The out-degree d
out
(v) of vertex v is the number of (directed) edges having
v as their start-vertex; in-degree d
in
(v) is similarly dened. Clearly, for every
nite digraph G(V, E),

vV
d
in
(v) =

vV
d
out
(v).
A directed path is similar to a path in an undirected graph; if the sequence of
edges is e
1
, e
2
, then for every i 1, the end-vertex of e
i
is the start-vertex of
e
i+1
. The directed path is simple if no vertex appears on it more than once. A
nite directed path Cis a directed circuit if the start-vertex and end-vertex of C
are the same. If C consists of one edge, it is a self-loop. As stated, the start and
end vertices of C are identical, but if there is no other repetition of a vertex, C
is simple. A digraph is said to be strongly connected if, for every ordered pair
of vertices (u, v) there is a directed path which starts at u and ends in v.
1.2 Computer Representation of Graphs
To understand the time and space complexities of graph algorithms one needs
to know how graphs are represented in the computers memory. In this section
two of the most common methods of graph representation are briey described.
Let us consider graphs and digraphs that have no parallel edges. For such
graphs, the specication of the two endpoints is sufcient to specify the edge;
for digraphs, the specication of the start-vertex and the end-vertexis sufcient.
Thus, we can represent such a graph or digraph of nvertices by an nnmatrix
M, where M
ij
=1 if there is an edge connecting vertex v
i
to v
j
, and M
ij
=0, if
not. Clearly, in the case of (undirected) graphs, M
ij
= 1 implies that M
ji
=1;
or in other words, Mis symmetric. But in the case of digraphs, any nnmatrix
of zeros and ones is possible. This matrix is called the adjacency matrix.
Given the adjacency matrix Mof a graph, one can compute d(v
i
) by counting
the number of ones in the i-th row, except that a one on the main diagonal
represents a self-loop and contributes two to the count. For a digraph, the number
4 1 Paths in Graphs
of ones in the i-th row is equal to d
out
(v
i
), and the number of ones in the j-th
column is equal to d
in
(v
j
).
The adjacency matrix is not an efcient representation of the graphif the graph
is sparse; namely, the number of edges is signicantly smaller than n
2
. In these
cases, it is more efcient to use the incidence lists representation, described
later. We use this representation, which also allows parallel edges, in this book
unless stated otherwise.
A vertex array is used. For each vertex v, it lists vs incident edges and a
pointer indicating the current edge. The incidence list may simply be an array
or may be a linked list. Initially, the pointer points to the rst edge on the list.
Also, we use an edge array, which tells us for each edge its two endpoints (or
start-vertex and end-vertex, in the case of a digraph).
Assume we want an algorithm TRACE(s, P), such that given a nite graph
G(V, E) and a start-vertex s V traces a maximal path P that starts at s and does
not use any edge more than once. Note that by maximal we do not mean that
the resulting path, P, will be the longest possible; we only mean that P cannot
be extended, that is, there are no unused incident edges at the end-vertex.
We can trace a path starting at s by taking the rst edge e
1
on the incidence list
of s, marking e
1
as used in the edge array, and looking up its other endpoint
v
1
(which is s if e
1
is a self-loop). Next, use the vertex array to nd the pointer
to the current edge on the list of v
1
. Scan the incidence list of v
1
for the rst
unused edge, take it, and so on. If the scanning hits the last edge and it is used,
TRACE(s, P) halts. A PASCAL-like description of TRACE(s, P) is presented
in Algorithm 1.1. Here is a list of the data structures it uses:
(i) A vertex table such that, for every v V, it includes the following:
A list of the edges incident on v, which ends with NIL
A pointer N(v) to the current item in this list. Initially, N(v) points to
the rst edge on the list (or to NIL, if the list is empty).
(ii) An edge table such that every e E consists of the following:
The two endpoints of e
A ag that indicates whether e is used or unused. Initially, all edges are
unused.
(iii) An array (or linked list) P of edges that is initially empty, and when
TRACE(s, P) halts, will contain the resulting path.
Notice that in each application of the while loop of TRACE (lines 210 in
Algorithm 1.1), either N(v) is moved to the next item on the incidence list of
v (line 4), or lines 610 are applied, but not both. The number of times line
1.2 Computer Representation of Graphs 5
Procedure TRACE(s, P)
1 v s
2 while N(v) points to an edge (and not to NIL) do
3 if N(v) points to a used edge do
4 change N(v) to point to the next item on the list
5 else do
6 e N(v)
7 change the ag of e to used
8 add e to the end of P
9 use the edge table to nd the second endpoint of e, say u
10 v u
Algorithm 1.1: The TRACE algorithm.
4 is applied is clearly O(|E|). The number of times lines 610 are applied is
also O(|E|), since the ag of an unused edge changes to used, and each of these
lines takes time bounded by a constant to run. Thus, the time complexity of
TRACE is O(|E| ).
1
(In fact, if the length of the resulting P is L then the time
complexity is O(l); this follows from the fact that each edge that joins P can
cause a waste of computing time only twice: once when it joins P and, at
most, once again by its appearance on the incidence list of the adjacent vertex.)
If one uses the adjacency matrix representation, in the worst case, the tracing
algorithmtakes time (and space) complexity (|V |
2
).
2
And if |E| << |V |
2
, as is
the case for sparse graphs, the complexity is reduced by using the incidence-list
representation. Since in most applications, the graphs are sparse, we prefer to
use the incidence-list representation.
Note that in our discussions of complexity, we assume that the word length of
our computer is sufcient to store the names of our atomic components: vertices
and edges. If one does not make this assumption, then one may have to allow
(log(|E| +|V|)) bits to represent the atomic components, and to multiply the
complexities by this factor.
1
f(x) is O(g(x)) if there are two constants k
1
and k
2
, such that for every x, f(x)k
1

g(x)+k
2
.
2
f(x) is (g(x)) if there are two constants k
3
and k
4
, such that for every x, f(x)k
3

g(x)+k
4
.
6 1 Paths in Graphs
1.3 Euler Graphs
An Euler path of a nite undirected graph G(V, E) is a path such that every
edge of G appears on it once. Therefore, the length of an Euler path is |E|. If G
has an Euler path, then it is called an Euler graph.
Theorem 1.1 A nite (undirected) connected graph is an Euler graph if and
only if exactly two vertices are of odd degree or all vertices are of even degree.
In the latter case, every Euler path of the graph is a circuit, and in the former
case, none is.
As an immediate conclusion of Theorem 1.1 we observe that none of the
graphs in Figure 1.2 is an Euler graph because both have four vertices of odd
degree. The graph shown in Figure 1.2(a) is the famous Knigsberg bridge
problemsolved by Euler in 1736. The graph shown in Figure 1.2(b) is a common
misleading puzzle of the type draw without lifting your pen from the paper.
Proof: It is clear that if a graph has an Euler path that is not a circuit, then
the start-vertex and the end-vertex of the path are of odd degree, while all the
other vertices are of even degree. Also, if a graph has an Euler circuit, then all
vertices are of even degree.
Assume now that G is a nite graph with exactly two vertices of odd degree,
a and b. We nowdescribe an algorithm(A), which will nd an Euler path from
a to b.
First, trace a maximal path P, as in the previous section, starting at vertex a.
Since G is nite, the algorithm halts, producing a path. But as soon as the path
emanates froma, one of the edges incident to a is used, and as residual degree
becomes even. Thus, every time a is reentered, there is an unused edge to leave
(a) (b)
Figure 1.2: Non-Eulerian graphs.
1.3 Euler Graphs 7
a by. This proves that P cannot end in a. Similarly, if vertex v V \ {a, b}, then
P cannot end in v. It follows that P ends in b.
If P contains all the edges of G, we are done. If not, we make the following
observations:

The residual degree of every vertex is even.

There is an unused edge incident on some vertex v that is on P. To see that

this must be so, let u
e
w be an unused edge. If either u or w is on P, we
are done. If not, since G is connected, there is a path Q from a to u. There
must be unused edges on Q. Going from a on Q, the rst unused edge we
encounter ts the bill.
Now, trace a maximal path P

in the residual graph, which consists of the

set V of vertices and all edges of E that are not in P. Start P

at v. Since all
vertices of the residual graph are of even degree, P

ends in v (and is therefore

a circuit). Next, combine P and P

to form one path from a to b as follows:

Follow P until it enters v. Now, incorporate P

, and then follow the remainder

of P.
Repeat, incorporating additional circuits into the present path as long as there
are unused edges. Since the graph is nite, this process will terminate, yielding
an Euler path.
If all vertices of the graph are of even degree, the rst traced path can start at
any vertex, and will be a circuit. The remainder of the algorithm is similar to
to this process.
In the case of digraphs, a directed Euler path is a directed path in which every
edge appears once. A directed Euler circuit is a directed Euler path for which
the start and end vertices are identical. In addition, digraph is called Euler if it
has a directed Euler path (or circuit).
The underlying (undirected) graph of a digraph is the graph resulting from
the digraph if the direction of the edges is ignored. Thus, the underlying graph
of the digraph in Figure 1.3(a) is shown in Figure 1.3(b).
(b) (a)
Figure 1.3: A digraph and its underlying graph.
8 1 Paths in Graphs
Theorem 1.2 A nite digraph is an Euler digraph if and only if its underlying
graph is connected and one of the following two conditions holds:
(i) There is one vertex asuch that d
out
(a) =d
in
(a)+1, and another vertex b
such that d
out
(b) +1 =d
in
(b), while for every other vertex v, d
out
(v) =
d
in
(v).
(ii) For every vertex v, d
out
(v) =d
in
(v).
In the former case, every directed Euler path starts at a and ends in b. In the
latter, every directed Euler path is a directed Euler circuit.
The proof of Theorem1.2 is along the same lines as the proof of Theorem1.1,
and is therefore not repeated here.
Let us make now a few comments about the complexity of the algorithm /
for nding an Euler path, as described in the proof of Theorem1.1. Our purpose
is to show that the time complexity of the algorithm is O(|E|).
Assume G(V, E) is presented in the incidence lists data structure. The main
path P and the detour P

will be represented by linked lists, where each item

on the list represents an edge.
In the vertex table, we add for each vertex v the following two items:
(i) A ag that indicates whether v is already on the main path P or the detour
P

. Initially, this ag is unvisited.

(ii) For every visited vertex v, there is a pointer E(v) to the location on the path
of the edge through which v was rst encountered. Initially, for every v,
E(v) =NIL.
We shall also use a list L of visited vertices. Each vertex enters L once, when
its ag is changed from unvisited to visited.
/ starts by running TRACE(a, P), updating the vertices ags, and E(v) for
each newly visited vertex v. Next, the following loop is applied:
If L is empty, / halts. If not, take a vertex v from L, and remove v from L.
Use TRACE(v, P

) to produce P

. Look up edge E(v), recording the location

of the edge e it is linked to. Change this link to point to the rst edge on P

.
Now, let the last edge of P

point to e.
Note that when TRACE(v, P

) terminates, v has no unused incident edges.

This explains why we can remove v from L.
Now that P

has been incorporated into P, the loop is repeated.

It is not hard to see that both the time and space complexities of /are O(|E|).
1.4 De Bruijn Sequences 9
1.4 De Bruijn Sequences
Let = {0, 1, ..., 1} be an alphabet of letters. Clearly, there are L =
n
different words of length nover . Ade Bruijn sequence is a (circular) sequence
a
0
a
1
a
L1
over such that for every word wof length n over there exists
a (unique) 0 j <L such that
a
j
a
j+1
a
j+n1
=w,
where the computation of the indexes is modulo L.
The most important case is that of = 2. Binary de Bruijn sequences are of
great importance in coding theory and can be generated by shift registers. (See
Golomb, 1967, on the subject.) In this section we discuss the existence of de
Bruijn sequences for every and every n.
For that purpose let us dene the de Bruijn digraph G
,n
(V, E) as follows:
(i) V =
n1
; i.e., the set of all
n1
words of length n1 over .
(ii) E =
n
.
(iii) The directed edge b
1
b
2
b
n
starts at vertex b
1
b
2
b
n1
and ends in
vertex b
2
b
3
b
n
.
Digraphs G
2,3
and G
2,4
are shown in Figure 1.4. G
3,2
is shown in Figure 1.5.
Observe that if w
1
, w
2

n
, then w
2
can follow w
1
in a de Bruijn sequence
only if in G
,n
the edge w
2
starts at the vertex in which w
1
ends. It follows
that there is a de Bruijn sequence for and n if and only if there is a directed
Euler circuit in G
,n
.
For example, consider the directed Euler circuit of G
2,3
, which consists of the
following sequence of directed edges:
000, 001, 011, 111, 110, 101, 010, 100.
The corresponding de Bruijn sequence, 00011101, follows by reading the rst
letter of each word (edge) in the circuit.
Theorem 1.3 For every and n, G
,n
has a directed Euler circuit.
Proof: To use Theorem 1.2 we have to show that the underlying undirected
graph of G
,n
is connected and that for every vertex v, d
out
(v) =d
in
(v).
Let us show that G
,n
is strongly connected. This implies that its underlying
undirected graph is connected.
10 1 Paths in Graphs
00
01
11
10
000
001
011
111
110
100
010
101
000
001
011
111
110
100
010
101
0000
0001
0011 1100 0101 1010
1011 1101
1001
0010 0100
1000
1111
0111 1110
0110
G
2,3
G
2,4
Figure 1.4: Examples of de Bruijn digraphs for = 2.
0
1 2
00
11
21
01 20
22
02 10
12
Figure 1.5: G
3,2
Let b
1
b
2
b
n1
and c
1
c
2
c
n1
be any two vertices. The directed path
b
1
b
2
b
n1
c
1
, b
2
b
3
b
n1
c
1
c
2
, . . . , b
n1
c
1
c
n1
is of length n 1, it starts at vertex b
1
b
2
b
n1
and ends in vertex
c
1
c
2
c
n1
, showing that G
,n
is strongly connected. (Observe that this
directed path is not necessarily simple; it may use vertices and edges more
than once.)
Now, observe that for each vertex v =b
1
b
2
b
n1
, every outgoing edge is of
the formb
1
b
2
b
n1
c, where c can be any of the letters. Thus, d
out
(v) =.
1.5 Shortest-Path Algorithms 11
Asimilar statement holds for d
in
(v). Thus, the condition that d
out
(v) =d
in
(v)
holds, uniformly.
Corollary 1.1 For every and n there exists a de Bruijn sequence.
1.5 Shortest-Path Algorithms
In general, the shortest-path problem is concerned with nding shortest paths
between vertices. Many interesting problems arise, and their variety depends on
the type of graph in our application and the exact question we want to answer.
Some of the characteristics that may help in dening the exact problem are as
follows:
(i) The graph is nite or innite.
(ii) The graph is undirected or directed.
(iii) The edges are all of length one, or all lengths are nonnegative, or negative
lengths are allowed.
(iv) We may be interested in shortest paths from a given vertex to another, or
from a given vertex to all the other vertices, or from each vertex to all the
other vertices.
(v) We may be interested in nding just one path, or in all paths, or in counting
the number of shortest paths.
Clearly, this section will deal only with a few of all possible problems. We
attempt to describe the most important techniques.
It is assumed that we are given a graph (or digraph) G(V, E) and a length
function l : E R. Namely, if e is an edge, then l(e) is the length of e. The
length of a path is the sum of the lengths of edges on it. The distance from
vertex u to vertex v, d(u, v), is the length of a shortest path fromu to v if there
are paths fromu to v, and is innite if no such path exists.
In most algorithms described in this section, the scenario is as follows: We
assume that there is a designated vertex s V, called the source. We denote by
(v) the value d(s, v). The algorithm will assign a label (v) to some (or all)
vertices v. Thus prove that when the algorithm halts, (v) =(v).
1.5.1 Breadth-First Search
Let us start with the case that G is nite and undirected, l(e) = 1 for every
edge e, and s V is a designated source. Our goal is to compute (v) for every
vertex v.
12 1 Paths in Graphs
Procedure BFS(G, s;)
1 empty Q and put s in Q
2 (s) 0
3 T {s}
4 while Q,= do
5 remove the rst vertex, u, from Q
6 for every edge u v, if v / T, do
7 T T {v}
8 (v) (u) +1
9 put v in Q
Algorithm 1.2: The BFS algorithm.
The natural and simple algorithm that follows was rst suggested by
Moore 1957, and was later called Breadth-First Search, or BFS. Intuitively, all
it does is the following:
We start by assigning (v) for every v V. We then proceed in waves.
In wave i 0 a set W(i) of vertices is assigned a nite label (v) i, and
no vertex with a nite label will ever be relabeled. In wave 0, (s) 0, and
therefore W(0) = {s}. As long as W(i) ,= , i is incremented, for every edge
u v, such that (u) =i1 and (v) =, assign (v) i and put v W(i).
In the Pascal-like algorithm, described in Algorithm 1.2, we use a slightly
different presentation; our reasons for doing so are discussed later. In this pre-
sentation we do not use the sets of waves, nor is there a running index i. Instead,
we use a queue Qof vertices and a set T of vertices. The rst vertextobe removed
from Q is s, and for every edge incident on s, the adjacent vertex, if new,
3
is labeled 1, joins T, and is put on Q. (These are the vertices of W(1).) Next,
vertices (that would have been in W(1)) are removed from Q, and their new
neighbors
4
are labeled 2, join T, are put in Q, and so on.
Let us say that a vertex v is accessible from u if there is a path from u to v.
Theorem 1.4 Algorithm BFS assigns every vertex v, which is accessible from
s, a label (v) and (v) =(v).
3
Not in T.
4
Adjacent vertices.
1.5 Shortest-Path Algorithms 13
Proof: First, let us prove that if v is accessible from s, then it gets a (nite)
label (v), and (v) (v). This is proved by induction on the value of (v).
The basis is established by Line 2, where s is labeled 0.
Now, assume the claim holds for vertices whose distance from s is less than
i, and let us prove it for vertex v for which (v) =i.
There is a path P from s to v of length i. The vertex u that precedes v on
P, satises (u) =i 1.
5
Thus, by the inductive hypothesis, u is labeled, and
(u) i 1. When u is removed from Q, the edge between u and v on P is
examined. If at that time v is not yet labeled, it gets a label (v) i, proving
the claim. If v is already labeled, by the fact that waves exit Q according to
nondecreasing order, (v) cannot be higher than i.
It is easy to see that if a vertex v gets a label (v), then there is a path of
length (v) froms to v. Such a path can be traced back fromv to s by the edges
through which vertices have been labeled. This proves that (v) (v). Putting
the two inequalities together, the theorem follows.
The foregoing discussion holds for digraphs as well. The only change one
has to make in Algorithm 1.2 is to replace u v by u v.
In the case of innite graphs (digraphs), one can still use a modication of
BFS to solve the problem, provided the following conditions hold:
(i) Since it is impossible to store the entire input graph, there should be an
algorithm that provides the data of the vertex table of a specied vertex,
when such is requested. Clearly, the degrees (out-degrees) of the vertices
must be nite.
(ii) There is a designated-target vertex t, which is accessible froms, and all we
want to do is to nd (t) (and (v) for every v, such that (v) <(t)).
All one needs to change in Algorithm 1.2 is Line 4, as follows: while Q,=
and t / T do. Note that the description of BFS as in Algorithm 1.2 does not
require the prelabeling of all vertices by , which is an impossible task if G
has innitely many vertices.
Let us discuss the time complexity of BFS for nite graphs (digraphs). We
can represent T by an array of length |V| in which the ith position is 0 if v
i
/ T
and one if v
i
T. Looking up whether v T or changing the ith position takes
constant time, but the performance of Line 3 takes (|V|) time. In addition,
note that each edge is examined at most twice once from each of its endpoints,
5
It is easy to prove that every subpath of a shortest path is itself a shortest path between its two
ends. This is sometimes referred to as the principle of dynamic programming.
14 1 Paths in Graphs
and when the edge is examined it may cause a computation that takes constant
time. Thus, the time complexity of BFS is O(|V| + |E|).
Finally, let us comment on how one can use the data generated by BFS to
trace a shortest path from s to a vertex v T. This can be done as described
in the last paragraph of the proof of Theorem 1.4. To nd the edge by which
a vertex has been labeled, one can add to the vertex table an item that carries
this information; initially, the value of this item is NIL for every vertex, and
when a vertex joins T, the item is updated. This shows that the time complexity
remains essentially unchanged.
1.5.2 Dijkstras Algorithm
In this subsection, we shall rst assume that we are given a nite digraph
G(V, E); there is a source vertex s, and every (directed) edge e has a nonnegative
length l(e) 0. Our task is to compute (v) for every vertex v. Later, we shall
discuss other cases in which similar algorithms apply.
If all edge lengths are positive integers, it may seem that by replacing each
edge e by a directed path that goes through l(e) new edges, all of length 1, and
l(e)1 new(intermediate) vertices, the problemis mapped tothe case described
in the previous subsection, and BFS can be used to solve it. Although, this is
a valid statement, from a computer-science point of view, it is a bad idea. The
reason is that it takes logl(e) bits to represent l(e), and the foregoing suggested
transformation introduces l(e) 1 new vertices and edges. This blows up the
length of the input data exponentially. The Dijkstra Algorithm (Dijkstra 1959,
vol.1) avoids this blowup and keeps the complexity bounded polynomially in
terms of the length of the input data.
The Dijkstra algorithm is presented in Algorithm 1.3 in Pascal-like style.
Two sets of vertices are used: T is the set of temporarily labeled vertices; that
is, vertices for which has been assigned but its value is still subject to change.
P is the set of permanently labeled vertices. Vertices neither in T, nor in P,
have not been labeled yet. A vertex v, for which (v) is minimum, among the
vertices in T, is chosen in Line 5. This vertex moves from T to P, and every
edge e outgoing from v is examined. If the end-vertex u of e is in T, then its
label is lowered in Line 10, if e enables such an improvement. If u is new, it
gets a label in Line 12, and joins T in Line 13.
Lemma 1.2 When procedure DIJKSTRAis applied to any nite digraph Gand
start-vertex s, it halts.
1.5 Shortest-Path Algorithms 15
Procedure DIJKSTRA(G, s, l;)
1 (s) 0
2 T {s}
3 P
4 while T ,= do
5 choose a vertex v T for which (v) is minimum
6 T T \ {v}
7 P P {v}
8 for every v
e
u do
9 if u T then do
10 (u) min{(u), (v) +l(e)}
11 else, if u / P then do
12 (u) (v) +l(e)
13 T T {u}
Algorithm 1.3: The Dijkstra algorithm.
Proof: Every vertex enters T at most once, and every vertex chosen in Line 5
leaves T. The performance of Line 5 takes at most O(|V|) time, and the sum
total time to perform Lines 813, for all chosen vertices, is O(|E|), since no
edge is examined more than once. After performing Line 5 |V| times, T must
be empty, and the procedure halts.
It follows that the time complexity of Dijkstras algorithm is O(|V|
2
+ |E|).
We shall return to this issue.
Lemma 1.3 During the computation of Dijkstras algorithm, every vertex
accessible from s gets a label.
Proof: Let T be a directed path from s to v. If v does not get a label, let u be
the rst unlabeled vertex along T when the algorithm halts. Consider the edge
w
e
u on T. Since w has been labeled, eventually, it is chosen to leave T,
and by the edge e, u gets labeled. A contradiction.
Lemma 1.4 At any time during the computation of Dijkstras algorithm, if a
vertex v is labeled (v), then there is a directed path from s to v whose length
is (v).
16 1 Paths in Graphs
Proof: By induction on the time during which an assignment or reassignment
of a label takes place. The rst assignment is in Line 1, and indeed, there is a
path of length 0 froms to itself, that is, the empty path.
Now, look at an assignment, or reassignment, that occurs at time , and
assume all previous assignments satisfy the claim. Assume vertex u is assigned
a label at time (for the rst time) in Line 12. Thus, the label of v at time
was assigned earlier. By the inductive hypothesis, there is a directed path
from vertex s to v of length (v), and this path, appended with e, is a path of
length (u) from s to u. A similar argument holds in case of a reassignment,
per Line 10.
Note that Lemma 1.4 implies that for every labeled vertex v (and at any time),
(v) (v). (1.1)
Theorem 1.5 When Dijkstras algorithm halts, for every vertex v accessible
from s,
(v) =(v).
Proof: If v is accessible from s, then by Lemma 1.3, v has an assigned label
(v). By Equation 1.1, (v) (v). It remains to show that (v) (v).
The proof is by induction on the order in which vertices join the set P.
Let T be a shortest directed path from s to v. Let be the time when vertex
v is chosen to join P. At that time, s is in P and v is not. Thus, there must be
an edge u
e
w on T such that u P and w / P. Consider the following
sequence of claims:

Since the length of edges is nonnegative, the subpath of T, from w to v, is

of a nonnegative length. Thus,
(v) (w).

Since every subpath of a shortest path is shortest from its start-vertex to its
end-vertex, the subpath of T from s to w is shortest, and its length consists
of the length of a shortest path froms to u plus l(e), that is,
(w) =(u) +l(e).

Since uhas joined P before time , by the inductive hypothesis, (u) (u).
Thus,
(u) +l(e) (u) +l(e).
1.5 Shortest-Path Algorithms 17

Since u has joined P before time , all its outgoing edges, including e, have
been examined, as per Lines 813. Thus, at time , (w) has been assigned,
and
(u) +l(e) (w).

However, at time , v has been chosen to join P, not w. Thus,

(w) (v).
These foregoing ve relations imply that, at time ,
(v) (v).

As we have seen, a simple implementation of the Dijkstra algorithmis of time

complexityO(|V|
2
+|E|). For sparse graphs the bulk of the computationis in the
(|V|) applications of Line 5, where many of themmay take (|V|) time each.
If one uses a heap to store the vertices of T (see, e.g., Cormen et al. [4]), then
the complexity is reduced to O(|E| log|V|). If one uses Fibonacci heaps (see,
e.g., [5] or [4]), then the complexity is reduced further to O(|V| log|V| + |E|).
In case of undirected graphs, the Dijkstra algorithm is applicable; the only
change one should make is to replace v
e
u by v
e
u in Line 8. Alter-
natively, transform the given graph to a digraph by replacing each undirected
edge with a pair of antiparallel directed edges of the same length, and apply
Dijkstras algorithm as is.
If there is a target vertex t, and one does not want to continue the computation
after t has been put in P, then one should replace Line 4 with while T ,= and
t / P do. This change makes the algorithm applicable for innite graphs,
provided t is accessible from s, the out-degrees of the vertices are nite, and
the number of vertices v, for which (v) (t), is nite.
Note that Dijkstras algorithm may not be applicable if there are edges of
negative length. This is the case even if the graph is directed and there are no
directed circuits whose length is negative.
1.5.3 The Ford Algorithm
In this subsection we assume that the given digraph, G(V, E), is nite, every
(directed) edge e has a length l(e), which may be negative. We are also given
a source vertex s. Our task is to compute (v) for every vertex v.
18 1 Paths in Graphs
Procedure gen-FORD(G, s, l;)
1 for every v V do
2 (v)
3 (s) 0
4 while there is an edge u
e
v such that (u) is nite and (v) >
(u) +l(e) do
5 (v) (u) +l(e)
Algorithm 1.4: The generic Ford algorithm.
Let us call a directed circuit negative if its length is negative. Notice that if
there is a negative circuit C, and it is accessible from s, then the distance from
s to the vertices on C is not dened; for every real number r, one may take a
path from s to one of Cs vertices and go around C sufciently many times,
to build up a path of length less than r. But, if there are no negative circuits
accessible froms, then either v is not accessible froms, and then (v) =, or
only simple paths froms to v need to be considered. The number of such paths
is nite, and therefore, (v) is well dened. Thus, we conclude that () is well
dened.
The generic Ford algorithm[6, 7],
6
described in Algorithm1.4, computes for
every vertex v, a value (v). As we shall see, if there are no negative circuits
accessible froms, the procedure will terminate, and upon termination, for every
vertex v, (v) =(v).
Lemma 1.5 While running gen-FORD, if (v) is nite then there is a directed
path from s to v whose length is (v).
The proof is similar to that of Lemma 1.4.
Lemma 1.5 holds even if there are negative circuits. However, if there are
no such circuits, the path traced in the proof cannot return to a vertex visited
earlier. For if it does, then by going around the directed circuit, a vertex has
improved its own label. This implies that the sum of the edge lengths of the
circuit is negative. Therefore, we have:
6
Sometimes called the Bellman-Ford algorithm.
1.5 Shortest-Path Algorithms 19
Lemma 1.6 If the digraph has no accessible negative circuits and if, while
running, gen-FORD (v) is nite, then there is a simple directed path from s
to v whose length is (v).
Under the conditions of Lemma 1.6, since each new assignment of ()
corresponds to a new simple directed path from s, and since the number of
simple directed paths (froms) is nite, we conclude that under these conditions
procedure gen-FORD must terminate, and is therefore, an algorithm.
Lemma 1.7 For a digraph with no accessible negative circuits, upon termina-
tion of the Ford algorithm, (v) =(v) for every vertex v.
Proof: If v is not accessible from s, then both (v) and (v) are equal to
and the claim holds.
If v is accessible froms, by Lemma 1.5, (v) (v). It remains to be shown
that (v) (v).
Let P be a shortest path froms to v, where
P : s =u
0
e
1
u
1
e
2
u
2
u
l1
e
l
u
l
=v.
We prove, by induction on 0 i l, that (u
i
) (u
i
). The claim clearly
holds for u
0
, by Line 3 and the fact that labels never increase.
Let i be the least value for which (u
i
) > (u
i
). Since every section of a
shortest path is shortest between its two endpoints, the length of the subpath
of P, from u
0
to u
i
is (u
i
) and (u
i
) =(u
i1
) +l(e
i
). By the choice of i,
(u
i1
) (u
i1
). It follows that
(u
i
) >(u
i
) =(u
i1
) +l(e
i
) (u
i1
) +l(e
i
).
Thus, the algorithm should not have terminated.
In spite of the fact that gen-FORDis a valid algorithm, the lack of determinism
in the choice of the order in which the edges are observed in Line 4 may be
abused to cause the algorithmto take exponential time. (See Johnson [8].) There
is a simple remedy: Order the edges of the digraph any order will do and
perform Line 4 by scanning the edges in this order. Once the scan is complete,
repeat it until there is no improvement in a complete scan. This procedure,
adv-FORD, is described in Algorithm 1.5, where it is assumed that
E = {e
1
, e
2
, . . . , e
m
}.
20 1 Paths in Graphs
Procedure adv-FORD(G(V, E), s, l;)
1 for every v V do
2 (v)
3 (s) 0
4 repeat
5 Flag False
6 for every 1 i mdo
7 let u
e
i
v
8 if (u) is nite and (v) >(u) +l(e
i
) then do
9 (v) (u) +l(e
i
)
10 Flag True
11 until Flag = False
Algorithm 1.5: The advanced Ford algorithm.
Theorem1.6 If the digraph has no accessible negative circuits, then procedure
adv-FORD terminates in O(|V| |E|) time, and when it terminates, (v) =(v)
for every vertex v.
Proof: Let us prove by induction on kthat for everyvertexv, if there is a shortest
path from s to v, which consists of k edges, then after the k-th application of
the loop (Lines 4-11), or if procedure adv-FORD halts earlier, (v) =(v).
For k = 0, the only applicable vertex is s, and Line 3 establishes the claim.
Assume now that the claim holds for 0 k j and show that it holds for
k =j +1.
If procedure adv-FORDterminates before the j +1-st application of the loop,
the claimfollows fromLemma 1.7. Let v be a vertex such that there is a shortest
path, P, from s to v that consists of j +1 edges. (If there is also a shortest path
from s to v which consists of less edges, then there is nothing to prove.) Let
u
e
v be the last edge in P. Since the subpath of P from s to u is a shortest
path to u, and since it consists of j edges, by the inductive hypothesis, after the
j-th application of the loop, (u) =(u). In the j +1-st application of the loop,
when e is checked, (v) gets the value (v), if it has not had that value already.
That proves the claim.
If v is accessible from s, and since there are no accessible negative circuits,
a shortest path froms to v is simple and consists |V| 1 edges, or fewer. Thus,
1.5 Shortest-Path Algorithms 21
during the |V|-th application of the loop no vertex improves its label, and the
procedure halts. Since the time complexity of the loop is O(|E|), the whole
procedure takes O(|V| |E|) time.
A simple conclusion of the proof of Theorem 1.6 is that, if adv-FORD does
not halt after the V-th application of the loop, then there must be an accessible
negative circuit. The algorithm can easily be modied to detect the existence
of negative circuits in digraphs in time O(|V| |E|).
1.5.4 The Floyd Algorithm
As in the previous subsection, assume we are given a nite digraph G(V, E) and
a length function l : E!.
7
We shall assume that there are no negative circuits
in G. Our aim is to compute a complete distance table; that is, to compute the
distance (u, v), from vertex u to vertex v, for every (ordered) pair of u and v.
We shall assume that there are no parallel edges; if there is more than one
edge from u to v, then we can remove them all, except for one of the shortest.
Self-loops are also superuous, but we shall allow them, and as we shall see,
the algorithm can be used to check if there are negative circuits, including the
case of negative self-loops.
If all edges were of nonnegative lengths, we could have used Dijkstras algo-
rithm from every vertex, and the complexity, in the simple implementation,
would have been O(|V|
3
). If there are negative edges, however, this cannot
be done. If we use Fords algorithm from every vertex, the complexity is
O(|V|
2
|E|), and for dense graphs, that can take (|V|
4
) time.
Floyds algorithm, [9], presented bellow, achieves the goal in time complexity
O(|V|
3
).
Let us assume that V = {1, 2, . . . , n} and for every 0 k n let
k
be an
nn matrix.
8

k
(i, j) stands for the length of a shortest path from vertex i to
vertex j, among all paths which do not go through vertices k+1, k+2, . . . , n
as intermediate vertices.
The Floyd algorithm is described in Algorithm 1.6.
It is easy to see that the time complexity of the Floyd algorithmis O(n
3
); there
are napplications of the outer loop (Lines 911), and in each of its applications,
there are n
2
applications of Line 11.
The proof of validity is also easy, by induction on k. A shortest path from i
to j, among paths that do not go through vertices higher than k, either does not
7
Rdenotes the set of real numbers.
8
We will show that only one matrix is necessary, but for didactic reasons, let us start with n+1
matrices.
22 1 Paths in Graphs
Procedure FLOYD(G(V, E), l;
n
)
1 for every 1 i n do
2 if there is a self-loop i
e
i and l(e) <0 then do
3
0
(i, i) l(e)
4 else
0
(i, i) 0
5 for every 1 i, j n such that i ,=j do
6 if there is an edge i
e
j then do
7
0
(i, j) l(e)
8 else
0
(i, j)
9 for every k, starting with k = 1 and ending with k =n do
10 for every 1 i, j n do
11
k
(i, j) min{
k1
(i, j),
k1
(i, k) +
k1
(k, j)}
Algorithm 1.6: The Floyd algorithm.
go through vertex k, and is therefore equal to
k1
(i, j), or does go through k
and is therefore equal to
k1
(i, k) +
k1
(k, j).
However, the space complexity of the algorithm, as stated in Algorithm 1.6,
is (n
3
), since there are n matrices of size nn each. It is easy to see that
there is no need to keep previous matrices. Two matrices sufce: The previous
one and the one being computed. In fact, one matrix will do, where some
of the entries are as in
k1
; and some, as in
k
. (Observe that a shortest path
from i to k, or from k to j, never needs to go through k, since there are no
negative circuits.) Thus, in fact, the space complexity can be reduced to O(n
2
),
by dropping the superscript indexes of .
Finally, one can use Floyds algorithmto check whether there are negative cir-
cuits in the digraph. Simply apply the algorithm, and in the end, check whether
there is an i for which (i, i) <0. If so, there is a negative circuit.
1.6 Problems
Problem 1.1 Prove that if a connected (undirected) nite graph has exactly 2k
vertices of odd degree, then the set of edges can be partitioned into k paths such
that every edge is used exactly once. Is the condition of connectivity necessary
or can it be replaced by a weaker condition?
1.6 Problems 23
Figure 1.6: A graph for Problem 1.4.
Problem 1.2 Let G(V, E) be an undirected nite circular Euler graph; that is,
Gis connected and for every v V, d(v) is even. A vertex s is called universal
if every application of TRACE(s, G), no matter how the edges are ordered in
the incidence lists, produces an Euler circuit.
Prove that s is universal if and only if s appears in every simple circuit
of G.
9
Problem1.3 Let G(V, E) be a nite digraph such that for every v V, d
in
(v) =
d
out
(v). Also, assume that the exits from v are labeled 1, 2, . . . , d
out
(v).
Consider a tour in G, which starts at a given vertex s. Every time a vertex
v is visited, the next exit is chosen to leave, starting with exit number 1 and
continuing cyclically. However, the tour stops if s is reached and all its exits
have been taken.
Prove that the tour stops and that every edge has been used at most once.
10
Problem1.4 AHamilton path (circuit) is a simple path (circuit) in which every
vertex of the graph appears exactly once.
Prove that the graph shown in Figure 1.6 has no Hamilton path or circuit.
Problem 1.5 Prove that in every completely connected digraph (a digraph in
which every two vertices are connected by exactly one directed edge in one of
the two possible directions), there is always a directed Hamilton path. (Hint:
Prove by induction on the number of vertices.)
Problem 1.6 Prove that a directed Hamilton circuit of the de Bruijn digraph,
G
,n
, corresponds to a directed Euler circuit of G
,n1
. Is it true that G
,n
always has a directed Hamilton circuit?
9 See[10].
10
More about such tours and nding Euler circuits can be found in [11].
24 1 Paths in Graphs
... ...
a
c
b
Figure 1.7: A switch for Problem 1.7.
Problem1.7 In the following, assume that G(V, E) is a nite undirected graph,
with no parallel edges and no self-loops.
(i) Describe an algorithm which attempts to nd a Hamilton circuit in G by
working with a partial simple path. If the path cannot be extended in either
direction, then try to close it into a simple circuit by the edge between its
endpoints, if it exists, or by a switch, as suggested by Figure 1.7, where
edges a and b are added and c is deleted. Once a circuit is formed, look
for an edge from one of its vertices to a new vertex, and open the circuit
to a now longer simple path, and so on.
(ii) Prove that if for every two vertices u and v, d(u) +d(v) n, where
n = |V|, then the algorithm never fails to produce a Hamilton circuit.
(iii) Deduce Diracs Theorem [12]: If for every vertex v, d(v)
n
2
, then G
has a Hamilton circuit.
Problem 1.8 Describe an algorithm for nding the number of shortest paths
from s to t, after the BFS algorithm has been performed.
Problem 1.9 A digraph is called acyclic if there are no directed circuits.
11
Let
G(V, E) be a nite acyclic digraph. A bijection f : V {1, 2, . . . , n}, where
n = |V|, is called a topological sorting if for every edge u v, f(u) <f(v).
Consider the procedure described in Algorithm 1.7. A queue Qof vertices is
used, which is initially empty.
Prove that this is an algorithm, that it computes a topological sorting and that
its time complexity is O(|V| + |E|).
Problem 1.10 Show that the Dijkstra algorithm is not applicable if there are
negative edges, even if the digraph is acyclic.
11
Sometimes called DAG, for directed acyclic graph.
1.6 Problems 25
Procedure TOPO.SORT(G;f)
1 for every v V compute d
in
(v)
2 for every v V do
3 if d
in
(v) = 0 then put v in Q
4 i 1
5 while Q,= do
6 remove the rst vertex u from Q
7 f(u) i
8 i i +1
9 for every edge u v do
10 d
in
(v) d
in
(v) 1
11 if d
in
(v) = 0 then put v in Q
Algorithm 1.7: Topological sorting.
Problem1.11 In the Dijkstra algorithm, assume the sequence of vertices which
join P, in this order, is s =v
1
, v
2
, . . .. Prove that the sequence (v
1
), (v
2
), . . .
is nondecreasing.
Problem1.12 Assume G(V, E) is a nite digraph, l : E!a length function,
and assume the length of every directed circuit is positive. Also, assume s V
is the source, V

is the set of vertices accessible from, s and : V

!is the
distance function.
We want to compute the function : V

:, where (v) is the number of

shortest paths froms to v.
(i) Let H(V

, E

) be a subgraph of G, where E

is the set of edges u

e
v
such that (v) =(u) +l(e). Prove that H is acyclic.
(ii) Show how a modication of the topological sorting algorithm, applied to
H, can compute . What is the complexity of this algorithm?
Problem 1.13 Prove that a connected undirected graph G is orientable (by
giving each edge some direction) into a strongly connected digraph if and only
if each edge of G is in some simple circuit in G.
Problem 1.14 Prove that if the digraph G(V, E), with edge lengths l : E R,
has a negative circuit, then there is a vertex i on this circuit such that in the
matrix computed by Floyds algorithm, (i, i) <0.
26 1 Paths in Graphs
Procedure WARSHALL(G(V, E);T)
1 for every 1 i, j n do
2 if there is an edge i j in G then do
3 T(i, j) 1
4 else T(i, j) 0
5 for every k, starting with k = 1 and ending with k =n do
6 for every 1 i, j n do
7 T(i, j) max{T(i, j), T(i, k) T(k, j)}
Algorithm 1.8: The Warshall algorithm.
Problem1.15 The transitive closure of a digraphG(V, E) is a digraphT(V, E
T
)
such that there is an edge uv in E
T
if and only if there is a nonempty directed
path from u to v in G.
Show how BFS can be used to construct T in time O(|V| |E|).
Problem 1.16 This problem is about Warshalls algorithm [13] for the
computation of the transitive closure.
Given a digraph G(V, E), where V = {1, 2, . . . , n}, we want to compute an
nn matrix T, such that
T(i, j) =

1 if there is a nonempty directed path in G from i to j

0 otherwise
Warshalls algorithm is described in Algorithm 1.8.
(i) What is the complexity of Warshalls algorithm? Compare it with the
repeated BFS above.
(ii) Prove the validity of Warshalls algorithm. (Hint of one possible proof:
Consider a (simple) path fromi to j and the order in which the intermediate
vertices on it are processed in the loop of Lines 57.)
(iii) Show that there is a close relationship between Floyds algorithm and
Warshalls: (i, j) is nite if and only if T(i, j) = 1.
Problem 1.17 This problem is about Dantzigs algorithm [14] for computing
all distances in a nite digraph, like Floyds algorithm.
1.6 Problems 27
Procedure Dantzig(G(V, E), l;
n
)
1 for every 1 i, j, k n do
2
k
(i, j)
3
1
(1, 1) l(1, 1)
4 for every k, starting with k = 2 and ending with k =n do
5 for every 1 i <k do
6
k
(i, k) min
1j<k
{
k1
(i, j) +l(j, k)}
7
k
(k, i) min
1j<k
{l(k, j) +
k1
(j, i)}
8 for every 1 i, j <k do
9
k
(i, j) min{
k1
(i, j),
k
(i, k) +
k
(k, j)}
Algorithm 1.9: The Dantzig algorithm.
For 1 i, j k n, let
k
(i, j) denote the length of a shortest path fromi to
j among paths that do not use vertices higher than k. The algorithm computes

k
(i, j) for all i, j and k.
Set l(i, j) as follows:
l(i, j) =

l(e) if i ,=j and there is an edge i

e
j
if i ,=j and there is no edge i j
min{0, l(e)} if i =j and there is an edge i
e
i
0 if i =j and there is no edge i i
The Dantzig algorithm is described in Algorithm 1.9.
(i) Show that Dantzigs algorithm is valid.
(ii) How can negative circuits be detected?
(iii) What is the time complexity of this algorithm?
(iv) What is the space complexity of the algorithm? Can it be reduced?
Problem1.18 Describe an algorithmwhose input is a nite, strongly connected
digraph G(V, E), and it determines whether there is a directed circuit whose
length (in terms of the number of edges in it) is odd. The algorithm should be
of time complexity O(|E|). (Hint: Does the fact that for some pair of vertices,
a and b, there are two directed paths froma to b, one of odd length and one of
even length, imply the existence of an odd circuit?)
28 1 Paths in Graphs
Problem 1.19 Let an undirected connected nite graph G(V, E) be the road
map of a country, where each edge represents a road, and each vertex represents
an intersection. Let h : E !
+
be a function such that h(e) species the
maximum height of vehicles allowed on the road represented by e. Also, let
c V be a specied vertex (such as the capital of the country).
Write an algorithmthat, when given the data above, computes for each vertex
v the maximum height of vehicles which can travel from c to v.
Make sure the time complexity of your algorithm is O(|V|
2
).
(Hint: A proper modication of Dijkstras algorithm can solve the problem.)
Bibliography
[1] Golomb, S. W., Shift Register Sequences, Holden-Day, 1967.
[2] Moore, E. F., The Shortest Path Through a Maze, Proc. Iternat. Symp. Switching
Th., 1957, Part II, Harvard Univ. Press, 1959, pp. 285292.
[3] Dijkstra, E. W., ANote on Two Problems in Connexion with Graphs, Numerische
Math., Vol. 1,1959, pp. 269271.
[4] Cormen, T. H., C. E. Leiserson, and R. L. Rivest, Introduction to Algorithms, The
MIT Press, 1990.
[5] Fredman, M. L., and R. E. Tarjan, Fibonacci Heaps and Their Uses in Improved
Network Optimization Algorithms, J. of the ACM, Vol. 34, No. 3, 1987, pp. 596
615.
[6] Ford, L. R., Jr., Network Flow Theory, The Rand Corporation, P-923, August,
1956.
[7] Ford, L. R., Jr. and D. R. Fulkerson, D. R., Flows in Networks, Princeton University
Press, 1962, Chap. III, Sec. 5.
[8] Johnson, D. B., A Note on Dijkstras Shortest Path Algorithm, J. of the ACM,
Vol. 20, No. 3, 1973, pp. 385388.
[9] Floyd, R. W., Algorithm 97: Shortest Path, Comm. of the ACM, Vol. 5, 1962,
p. 345.
[10] Ore, O., A Problem Regarding the Tracing of Graphs, Elem. Math., Vol. 6, 1951,
pp. 4953.
[11] Bhatt, S., D. Greenberg, S. Even, and R. Tayar, Traversing Directed Eulerian
Mazes, J. of Graph Algorithms and Applications, Vol. 6, No. 2, 2002, pp. 157173.
[12] Dirac, G. A., Connectivity Theorems for Graphs, Quart. J. of Math., Ser. (2),
Vol. 3, 1952, pp. 171174.
[13] Warshall, S., ATheoremon Boolean Matrices, J. of the ACM, Vol. 9, No. 1, 1962,
pp. 1112.
[14] Dantzig, G. B., All Shortest Routes in a Graph, Oper. Res. House, Stanford
University Tech. Rep. 663, November 1966.
2
Trees
2.1 Tree Denitions
Let G(V, E) be an undirected, nite, or innite graph. We say that G is circuit-
free if there are no simple circuits in G. G is called a tree if it is connected and
circuit-free.
Theorem 2.1 The following four conditions are equivalent:
(a) G is a tree.
(b) G is circuit-free, but if any new edge is added to G, a simple circuit is
formed.
(c) G has no self-loops, and for every two vertices, there is a unique simple
path connecting them.
(d) G is connected, but if any edge is deleted, the connectivity of G is
interrupted.
Proof: We shall prove that (a) (b) (c) (d) (a).
(a) (b): We assume that G is connected and circuit-free. Let a
e
b be
a new edge. If a = b then e forms a self-loop, and therefore a simple circuit
exists. If a ,=b, there is a simple path in G(without e) connecting a and b. The
addition of e creates a simple circuit.
(b) (c): We assume that G is circuit-free and that no edge can be added to
G without creating a simple circuit. Clearly, G has no self-loops.
Let a and b be any two vertices of G. If there is no path connecting them,
then we can add an edge between a and b without creating a simple circuit.
Thus, G must be connected.
Moreover, if there are two simple paths, P and P

, connecting a and b, then

there must be a simple circuit in G. To see that, assume that
P : a =v
0
e
1
v
1
e
2
v
2
v
l1
e
l
v
l
=b,
29
30 2 Trees
and
P

: a =v

0
e

1
v

1
e

2
v

2
v

1
e

=b.
Since both paths are simple, one cannot be the beginning of the other. Let i be
the rst index for which e
i
,=e

i
. That is, the two paths split at v
i1
=v

i1
. Let
v be the rst vertex on P, after the split, which is also on P

. Thus, for some

j, k i, v =v
j
=v

k
. The subpath of P, between v
i1
and v
j
, and the subpath
of P

, between v

i1
and v
k
, form a simple circuit.
(c) (d): We assume that Ghas no self-loops and that for every two vertices,
there is a unique simple path connecting them. Thus, G is connected.
Assume now that we delete an edge a
e
b from G. Since G has no self-
loops, a ,=b. If there is still a (simple) path in (V, E\ {e}) connecting a and b,
then in (V, E), there are two simple paths connecting a and b. A contradiction.
(d) (a): We assume that G is connected and that no edge can be deleted
without interrupting the connectivity.
If G has a simple circuit, any edge on this circuit can be deleted without
interrupting the connectivity. Thus, G is circuit-free.
There are two, more common ways to dene a nite tree. These are presented
in the following theorem:
Theorem 2.2 Let G(V, E) be a nite graph and n = |V|. The following three
conditions are equivalent:
(a) G is a tree.
(b) G is circuit-free and |E| =n1.
(c) G is connected and |E| =n1.
Proof: For n =1 the theorem is trivial. Assume n2. We shall prove that (a)
(b) (c) (a).
(a) (b): Let us prove, by induction on n, that if G(V, E) is a tree, then
|E| = n1. This statement clearly holds for n = 1. Assume that it is true for
all n <m, and let G be a tree with m vertices.
Delete from G any edge e. By condition (d) of Theorem 2.1, the resulting
graph is not connected any more, and has two disjoint connected components.
Each of these components is circuit-free and is therefore a tree. By the inductive
hypothesis, each component has one edge less than its number of vertices. Thus,
together they have m2 edges. Add e back, and the number of edges is m1.
(b) (c): We assume that G is circuit-free and has n1 edges. Let us rst
show that G has at least two vertices of degree 1.
2.2 Minimum Spanning Tree 31
Choose any edge u
e
v; there is at least one edge, since the number of
edges is n1 and n2. Also, u ,=v since Gis circuit-free. As in the TRACE
procedure, walk on new edges, starting from u. Since the number of edges is
nite, this walk must end, say, at vertex t
1
. Also, the traced path is simple,
or a simple circuit would have been formed. Thus, d(t
1
) = 1. A similar walk,
starting from v, yields another vertex, t
2
, and d(t
2
) = 1, as well. Thus, there
are two vertices of degree 1.
Now, the proof that G is connected proceeds by induction on the number of
vertices, n. Obviously, the statement holds for n = 1. Assume that it holds for
n=m1, and let G be a circuit-free graph with mvertices and m1 edges.
Eliminate from G a vertex v of degree 1 and its incident edge. The resulting
graph is still circuit-free and has m1 vertices and m2 edges. Thus, by the
inductive hypothesis it is connected. Therefore, G is connected, as well.
(c) (a): Assume that G is connected and has n1 edges. As long as G
has simple circuits, we can eliminate edges (without eliminating vertices) and
maintain the connectivity. When this process terminates, the resulting graph is
a tree, and, by (a) (b), it has n1 edges. This, however, is the number of
edges we started with. Thus, no edge has been eliminated, and therefore G is
circuit-free.

Let us call a vertex whose degree is 1, a leaf. A corollary of Theorem2.2 and

the statement proved in the (b) (c) part of its proof is the following:
Corollary 2.1 A nite tree with more than one vertex has at least two leaves.
2.2 Minimum Spanning Tree
A graph G

(V

, E

) is called a subgraph of a graph G(V, E), if V

V and
E

E.
1
Assume G(V, E) is a nite, connected (undirected) graph and each edge e E
has a known length l(e) >0. Assume that we want to nd a connected subgraph
G

(V, E

), whose total length,

eE
l(e), is minimum; or, in other words, we
want to remove fromG a subset of edges whose total length is maximum, and
yet, the resulting subgraph remains connected. Such a subgraph is a tree. For
G

is assumed to be connected, and since its total length is minimum, none of

its edges can be removed without destroying its connectivity. By Theorem 2.1
(see part (d)) G

is a tree. A subgraph of G that contains all of its vertices and

1
Note that an arbitrary choice of V

V and E

E may not yield a subgraph, simply

because it may not be a graph; that is, some of the endpoints of edges in E

may not be in V

.
32 2 Trees
Procedure PRIM(G, l;T)
1 for every v V do
2 (v)
3 choose a vertex s V
4 (s) 0
5 (s)
6 TEMP V
7 T
8 while TEMP ,= do
9 choose a vertex v TEMP for which (v) is minimum
10 TEMP TEMP \ {v}
11 T T (v)
12 for every v
e
u do
13 if u TEMP and (u) >l(e) then do
14 (u) l(e)
15 (u) {e}
Algorithm 2.1: The Prim algorithm.
is a tree is called a spanning tree of G. Thus, our problem is that of nding a
minimum (length) spanning tree (MST) of G.
There are several known algorithms for constructing an MST. We describe
here the Primalgorithm[1]. Additional algorithms are discussed inthe problems
section in the end of the chapter.
The Prim algorithm is described in Algorithm 2.1.
2
In the following discussion, we shall use T to denote a subgraph of G, which
is a tree, and the set of edges in this tree. TEMP denotes the set of vertices, not
yet in T.
The idea of the algorithm is to grow T, by starting with a tree with one
vertex, s, and in each step adding a new leaf, v, to T, until all vertices have
joined it. Thus, v is chosen in such a way that the edge connecting it to T (the
single edge in the set (v)) is of minimum length among edges which connect
a vertex of T to a vertex of TEMP.
2
The structure of the algorithm is similar to that of Dijkstra, but historically, Prims algorithm
precedes it.
2.2 Minimum Spanning Tree 33
When a new vertex, v, joins T, all its incident edges, v
e
u are checked. If
u is not a vertex of T yet (this is equivalent to the fact that u TEMP), and if
l(e) is shorter than the presently known shortest edge (if any) that connects u
to a vertex of T, then the value of (u) is updated to be l(e), and the edge e is
recorded in (the set) (u).
Since G is connected, when PRIM halts, the set of edges in T constitutes a
spanning tree. The complexity of the algorithm is similar to that of Dijkstras
algorithm; see Section 1.5.2.
Before we discuss the validity of the algorithm, let us dene the concept of a
cut.
Let S V and

S =V \S. The cut (S;

S) is the set of edges with one endpoint

in S and the other in

S.
Theorem2.3 The tree T produced by Prims algorithmis a MinimumSpanning
Tree of G.
Proof: Let us prove that for every T produced while the algorithm is running,
there is an MST, T
opt
, of G, such that T is a subgraph of T
opt
. This implies that
the nal T is an MST. The proof is by induction on , the number of vertices
in T.
Initially, = 1 and there are no edges in T. Thus, the claim holds trivially.
Assume now that the claim holds for <|V|, and let us prove it for +1.
By the inductive hypothesis, there is an MST, T
opt
, such that T is a subgraph
of T
opt
. If the newly added edge, e, is such that e T
opt
, then the claim holds
for T {e} as well, and we are done.
If e / T
opt
, consider the graph H(V, T
opt
{e}). By Theorem 2.1 part (b),
there is a simple circuit Cin H. Also, let Sbe the set of vertices in T and consider
the cut (S;

S) of G. There are at least two edges of C in (S;

S). One of them is

e, and assume another is e

. Now consider the graph H

(V, (T
opt
{e}) \ {e

}).
Since an edge of a simple circuit has been removed, H

remains connected, and

the number of edges is |V| 1. By Theorem 2.2 part (c), H

is a spanning tree.
By the choice of v, (v) is minimum(among values of for vertices in TEMP),
and the edge e (v) satises l(e) =(v) l(e

). Thus, the total length of H

is less than or equal to that of T

opt
, and is therefore an MST.

The analogous problem for digraphs, namely, that of nding a subset of the
edges E

whose total length is minimum among those for which (V, E

) is
a strongly connected subgraph, is much harder. In fact, even the case where
l(e) = 1 for all edges is hard. This is discussed in Chapter 10.
34 2 Trees
2 3
1
2 3
1
2 3
1
Figure 2.1: The spanning trees on three named vertices.
2.3 Cayleys Theorem
In a later section we shall consider the question of the number of spanning trees
of a given graph. Here we consider the more restricted, and yet interesting,
problem of the number of trees one can construct on a given set of vertices,
V = {1, 2, . . . , n}.
For n=2, there is only one tree that one can construct, consisting of an edge
between the two vertices. For n = 3, there are three possible trees, as shown
in Figure 2.1. The reader can verify, by exhausting all cases, that for n = 4 the
number of trees is 16. The following theorem is due to Cayley [3]. Warning:
We shall use the integers 1, 2, . . . , n in three different meanings. As names of
vertices, as integers, and as letters of an alphabet.
Theorem 2.4 The number of spanning trees for n distinct vertices is n
n2
.
The remainder of this section describes a proof due to Prfer [4]. (For a survey
of various proofs see Moon [5].)
Assume V = {1, 2, . . . , n}. Let us display a bijection between the set of
the spanning trees and the n
n2
words of length n 2 over the alphabet
{1, 2, . . . , n}. The mapping fromthe set of spanning trees to the corresponding
set of words is dened by an algorithm which is described in Algorithm 2.2.
For example, assume that n = 6 and T is as shown in Figure 2.2. We now
apply TREEtoWORD. We start with the given T and an empty template for
a word w of 4 letters. This is depicted in the rst line of Figure 2.3. T has 3
leaves, and vertex 2 has the least name. Therefore, vertex 2 and its incident
edge, 2 4, are removed. The rst letter in the word w is 4, as shown in the
second line of Figure 2.3. The next leaf to be removed is 3, and a
2
= 1, and
so on. After four steps, the tree consists of two vertices (4 and 6) and an edge
between them, while w= 4164.
By Corollary 2.1, Line (2) of TREEtoWORD algorithm can always be per-
formed. Note that when a leaf is removed, the remaining graph is still a tree. It
follows that for every tree of n vertices, a word wof length n2 is produced.
Since algorithm TREEtoWORD is deterministic, it denes a mapping f from
2.3 Cayleys Theorem 35
Procedure TREEtoWORD(T(V, E);w=a
1
a
2
a
n2
)
1 for every i starting with 1 and ending with n2 do
2 choose a leaf v whose name is minimum among the leaves of T
3 let v
e
u be its incident edge
4 a
i
u
5 remove e and v from T
Algorithm 2.2: The TREEtoWORD algorithm. Mapping a spanning
tree T to a word w.
2
5
4 6 1 3
Figure 2.2: T: An example of a spanning tree with six vertices.
trees to words. It remains to be shown that no word is produced by two different
trees, and that for every word w there is a tree T such that f(T) =w.
Notice that TREEtoWORD is insensitive to the nature of the set of vertices,
V, of T, as long as the names of the vertices are distinct and an order is dened
on these names. We shall assume that V is a set of integers, not necessarily
{1, 2, . . . , n}.
Lemma 2.1 If in T a vertex v has a degree d(v), then in w=f(T) the letter v
appears d(v) 1 times.
Proof: When each of the edges incident on v are removed, except the last one,
one writes the letter v in w. Thus, v appears d(v) 1 times in w. The last edge
may not be removed at all, if v is one of the two vertices which remain when
TREEtoWORD halts. And if vs last edge is removed, then v is the removed
leaf, and its neighbor, not v, is written in w.
Lemma 2.2 For every word w = a
1
a
2
a
n2
over an alphabet V, where
|V| =n, there is a unique tree T whose vertices are V and f(T) =w.
36 2 Trees
2
4 6 1 3
5
4 6 1 3 4 5
4 6 1 4 5 1
6 4 6 4 5 1
6 4 6 4 1 4
Figure 2.3: Applying TREEtoWORD to the the example tree T.
Proof: By induction on n2. For n=2, wis the empty word and if V = {u, v},
there is only one tree whose set of vertices is V. TREEtoWORD does nothing
(leaving T intact, and w empty).
Now assume the claim holds for n 1 and let us prove it for n. Let w =
a
1
a
2
a
n2
be a word over an alphabet V, where |V| =n.
Since the alphabet has n letters, while there are n2 appearances of letters
in w, there are at least two letters missing fromw. Let v be the least letter that
does not appear in w. Now consider the word w

= a
2
a
3
a
n2
, with the
alphabet V

= V \ {v}. By the inductive hypothesis, there is a unique tree T

whose set of vertices is V

, such that f(T

) =w

. Notice that a
1
appears in w,
and therefore a
1
,=v, and thus a
1
V

.
By Lemma 2.1, the set of leaves of T

is equal to the set of elements of V

that do not appear as letters in w

. Thus, v is less than every leaf of T

, with the
possible exception of a
1
.
Now dene T to be the tree of n vertices that results from T

by adding to it
the vertex v and an edge connecting v to a
1
. Clearly T is a tree, and a
1
is not a
leaf of T, even if it has been a leaf of T

. It follows that v is the least leaf of T.

Thus, if one applies TREEtoWORD to T, the rst letter assigned to the word
TREEtoWORD builds is indeed a
1
, v and its incident edge are removed, and
the tree on which TREEtoWORD continues is T

. By the inductive hypothesis,

T

is the unique tree that produces w

. We conclude that f(T) =w.

2.4 Directed Tree Denitions 37
Since v is the least letter missing fromw, and since a
1
is the rst letter of w,
every tree that produces w must have an edge v a
1
. Thus, T is the unique
tree that produces w.

2.4 Directed Tree Denitions

A digraph G(V, E) is said to have a root r if r V and every vertex v V is
reachable from r; that is, there is a directed path that starts in r and ends in v.
A digraph (nite or innite) is called a directed tree if it has a root and its
underlying undirected graph is a tree.
Theorem 2.5 Assume G is a digraph. The following ve conditions are
equivalent:
(a) G is a directed tree.
(b) G has a root from which there is a unique directed path to every vertex.
(c) Ghas a root r for which d
in
(r) =0, and for every other vertex v, d
in
(v)=1.
(d) G has a root and the deletion of any edge (but no vertices) interrupts this
condition.
(e) The underlying undirected graph of G is connected and G has one vertex
r for which d
in
(r) = 0, while for every other vertex v, d
in
(v) = 1.
Proof: We prove that (a) (b) (c) (d) (e) (a).
(a) (b): We assume that G has a root, say r, and its underlying undirected
graph G

is a tree. Thus, by Theorem 2.1 part (c), there is a unique simple path
fromr to every vertex in G

. Also, G

is circuit-free. Thus, a directed path from

r to a vertex v, in G, must be simple and unique.
(b) (c): Here, we assume that G has a root, say r, and a unique directed
path from it to every vertex v. First, let us show that d
in
(r) = 0.
Assume there is an edge u
e
r. There is a directed path from r to u, and
it can be continued, via e, back to r. Thus, in addition to the empty path from
r to itself (containing no edges), there is one more path, in contradiction to the
assumption of the path uniqueness.
Now, we have to show that if v ,=r, then d
in
(v) =1. Clearly, d
in
(v) >0, for
it must be reachable from r. If d
in
(v) > 1, then there are at least two edges,
say, v
1
e
1
v and v
2
e
2
v. Since there is a directed path, P
1
, fromr to v
1
, and
a directed path, P
2
, from r to v
2
, by adding e
1
to P
1
and e
2
to P
2
, we get two
different directed paths from r to v, contradicting the uniqueness assumption.
(Note that the two paths are different, even if v
1
=v
2
.)
38 2 Trees
(c) (d): This proof is trivial, for the deletion of any edge u
e
v will make
v unreachable from r.
(d) (e): We assume that G has a root, say r, and the deletion of any edge
interrupts this condition.
Clearly, d
in
(r) = 0, for any edge entering r can be deleted without inter-
rupting the condition that r is a root. Also, if v ,=r, then d
in
(v) >0, since v is
reachable fromr. If d
in
(v) >1, let v
1
e
1
v and v
2
e
2
v be two edges entering
v. Let P be a simple directed path fromr to v. At least one of the edges e
1
and
e
2
is not used in P. This edge can be deleted without interrupting the fact that
r is a root. Thus, d
in
(v) =1.
(e) (a): We assume that the underlying undirected graph of G, G

, is
connected, d
in
(r) = 0, and for v ,=r, d
in
(v) = 1. First let us prove that r is a
root of G.
Let P

be a simple (undirected) path connecting r and v in G

. This must
correspond to a directed path, P, fromr to v in G, for if any of the edges points
in the wrong direction, it would imply either that d
in
(r) >0 or that for some
u, d
in
(u) >1.
Now, assume G

has a simple circuit, C

. In G, all corresponding edges must

be directed in the same circular direction, or there would be a vertex v for which
d
in
(v) 2. Thus, r is not one of the vertices of C, the directed circuit of G
which corresponds to C

. Now, let P be a shortest directed path from r to a

vertex of C, say it is u. Then, d
in
(u) must be at least 2. A contradiction. Thus,
G

is circuit-free, and is therefore an undirected tree.

In case of nite digraphs one more useful denition of a directed tree is

possible:
Theorem 2.6 A nite digraph G(V, E) is a directed tree if and only if its
underlying undirected graph, G

, is circuit-free, one of its vertices, r, satises

d
in
(r) = 0, and for all other vertices v, d
in
(v) = 1.
Proof: The only if part follows directly fromthe denition of a directed tree
and Theorem 2.5 part (c).
To prove the if part we rst observe that the number of edges is |V| 1.
Thus, by Theorem 2.2, (b) (c), G

is connected. Thus, by Theorem 2.5, (e)

(a), G is a directed tree.

2.5 The Innity Lemma 39

...
Figure 2.4: An example of an innite arbitrated digraph with no root.
Let us say that a digraph is arbitrated (Berge [6] calls it quasi strongly con-
nected) if for every two vertices, v
1
and v
2
, there is a vertex v called an arbiter
of v
1
and v
2
, such that there are directed paths from v to v
1
and from v to v
2
.
There are innite digraphs which are arbitrated but do not have a root. For exam-
ple, see the digraph of Figure 2.4. However, for nite digraphs the following
theorem holds:
Theorem 2.7 If a nite digraph is arbitrated then it has a root.
Proof: Let G(V, E) be a nite arbitrated digraph, where V = {1, 2, . . . , n}.
Let us prove, by induction, that every set {1, 2, . . . , m}, where mn, has an
arbiter; i.e., a vertex a
m
such that every 1 i m is reachable from a
m
. By
denition, a
2
exists. Assume a
m1
exists. Let a
m
, be the arbiter of a
m1
and
m. Since a
m1
is reachable from a
m
, and every 1 i m1 is reachable
from a
m1
, every 1 i m1 is also reachable from a
m
.
Thus, for nite digraphs, the condition that it has a root, as in Theorem 2.5
parts (a), (b), (c), and (d), can be replaced by it being arbitrated.
2.5 The Innity Lemma
The following is known as Knigs Innity Lemma [7]:
Theorem 2.8 If G is an innite digraph with a root r, and every vertex has a
nite out-degree, then G has an innite directed path starting in r.
Before we present our proof, let us point out the necessity of the niteness of
the out-degrees of the vertices. For, if we allow a single vertex to be of innite
out-degree, the conclusion does not follow. Consider the digraph of Figure 2.5.
The root r is connected to vertices v
1
1
, v
2
1
, v
3
1
, . . ., where v
k
1
is the second vertex
on a directed path of length k. It is clear this directed tree is innite, and yet
it has no innite path. Furthermore, the replacement of the condition of nite
degrees by the condition that for every k, the tree has a path of length k, does
not work either, as the same example shows.
40 2 Trees
r
... v
1
1
v
2
1
v
3
1
v
3
2
v
3
3
v
2
2
Figure 2.5: An example of an innite digraph with no innite path.
Proof:
3
First, let us restrict our attention to a directed tree T which is an innite
subgraph of G. Ts root is r. All vertices of distance 1 away fromr in Gare also
of distance 1 away from r in T. In general, if a vertex v is of distance away
fromr in G, it is also of distance away fromr in T; all the edges entering v in
G are now dropped, except one which connects a vertex of distance 1 to v.
It is sufcient to showthat in T there is an innite directed path fromr. Clearly,
since T is a subgraph of G, all its vertices are of nite out-degrees too.
In T, r has innitely many descendants (vertices reachable fromr). Since r is
of nite out-degree, at least one of its sons (the vertices reachable via one edge),
say r
1
, must have innitely many descendants. One of r
1
s sons has innitely
many descendants, too, and so we continue to construct an innite directed path
r, r
1
, r
2
, . . . .
Despite the seeming simplicity of the theorem, it is useful. For example,
imagine we conduct a search on a directed tree of nite degrees (but it is not
known that there is a bound on the degrees). If it is known that the tree has no
innite directed path, then the theorem assures us that the tree is nite and our
search will terminate.
An interesting application of Theorem 2.8 was made by Wang [8]. Consider
the problem of tiling the plane with square tiles, all of the same size (Wang
calls the tiles dominoes). There is a nite number of tile families. The sides
of the tiles are labeled by letters of an alphabet, and all tiles of one family have
the same labels and are indistinguishable. Tiles may not be rotated or reected,
and the labels are specied for their north side, south side, and so on. There
3
We use a naive approach to set theory and thus ignore the applications of the axiom of choice.
2.5 The Innity Lemma 41
a
b
c
c
a
a
a
d
a
d
b
c
b
a
b a
Figure 2.6: A set of families of tiles.
c
a
a
a
a
b
c
a
d
a
d
b
d
a
d
b
a
b
c
a
c
b
a
b
Figure 2.7: A torus constructed from the tile families depicted in Figure 2.6.
is an innite supply of tiles of each family. Tiles may abut one another if the
joining sides have the same labels. For example, if the tile families are as shown
in Figure 2.6, then we can construct the torus, as shown in Figure 2.7. Now,
by repeating this torus innitely many times horizontally and vertically, we can
tile the whole plane.
Wang proved that if it is possible to tile the upper right quadrant of the plane
with a given nite set of t tile families, then it is possible to tile the whole plane.
The reader should realize that a southwest shift of the upper-right tiled quadrant
cannot be used to cover the whole plane. In fact, if the number of tile families is
not restricted to be nite, one can nd sets of families for which the upper-right
quadrant is tileable, whereas the whole plane is not. (See Problem 2.14.)
Consider the following directed tree T: The root r is connected to vertices,
each representing a tile family, that is, a 1 1 square, tiled with the tile of that
family. Thus, the out-degree of r is t. For every k 1 and every legitimate way
of tiling a (2k +1) (2k +1) square, there is a vertex in T; its father is the
vertex which represents the tiling of a (2k1) (2k1) square, identical to
the center part of the tiled square represented by the son.
Now, if the upper-right quadrant is tilable, then T has innitely many vertices.
However, a vertex representing a certain tiling of a (2k1) (2k1) square
has at most t
8k
sons. Since the out-degree of each vertex is nite (although it
may not be bounded), Theorem2.8 implies that there is an innite directed path
in T. Such a path describes a way to tile the whole plane.
42 2 Trees
2.6 Problems
Problem 2.1 Let T
1
(V, E
1
) and T
2
(V, E
2
) be two spanning trees of G(V, E).
Prove that for every E
1
\E
2
there is a E
2
\E
1
such that each of the sets
(E
1
\ {}) {}
(E
2
\ {}) {}
denes a spanning tree.
Problem 2.2 Let G(V, E) be a nite connected graph and l : E !. Describe
an algorithm for nding a maximum length spanning tree of G; explain why it
is valid and analyze its complexity.
Problem2.3 Algorithm2.3 describes the Kruskal [2] algorithmfor computing
an MST of a nite connected undirected graph G(V, E) with a length function
l : E !. The set of edges in the resulting tree is T. Prove that the resulting T
is indeed an MST and analyze the complexity of the algorithm.
Problem 2.4 Describe an algorithm that is similar to Kruskals, but instead of
adding edges, fromlight to heavy, as long as they do not create a simple circuit,
it deletes edges, from heavy to light, as long as connectivity is maintained.
Prove its validity and analyze its complexity.
Problem 2.5 Let G(V, E) be a nite undirected graph, with a length function
l : E !. Also, T is an MST of G.
Procedure KRUSKAL(G(V, E), l;T)
1 sort the set E into E = {e
1
, e
2
, . . . , e
|E|
}, so that if i <j, then l(e
i
)
l(e
j
)
2 T
3 for i, starting with i = 1 and ending with i = |E| do
4 if (V, T {e
i
}) is circuit-free then
5 T T {e
i
}
Algorithm 2.3: The Kruskal algorithm.
2.6 Problems 43
A new edge, e, is added, of length l(e). The following is a sketch of an
algorithmfor mending the tree to be an MST of the new graph: Add e to T. Let
e

be an edge of maximum length in the simple circuit that forms. If e

,= e,
then remove e

. Otherwise, leave T intact.

Prove the validity of this algorithm and compare its time complexity with
computing an MST anew.
Problem 2.6 Let G(V, E) be a nite undirected graph, with a length function
l : E !. Also, T is an MST of G.
An edge e E is removed from G to form G

. The following is a sketch of

an algorithm for mending the tree to be an MST of G

: If e / T, do nothing.
Otherwise, let S and

S be the two sets of vertices in the two connected compo-
nents of T \ {e}. If in G

, the cut (S;

S) = then G

has no MST. Otherwise, let

e

be of minimum length in (S;

S). The new MST is (T \ {e}) {e

}.
Prove the validity of this algorithm and compare its time complexity with
computing an MST anew.
Problem2.7 Compute the number of trees that can be built on n, given labeled
vertices with unlabeled edges, in such a way that one specied vertex is of
degree k.
Problem 2.8 Let V = {1, 2, . . . , n} and for each 1 i n d(i) is a positive
integer. Prove that if
n

i=1
d(i) = 2n2,
then there exists a tree T(V, E) in which for every i, the degree of i is d(i). How
many such trees are there if the edges have no names?
Problem 2.9 What is the number of trees that one can build with n labeled
vertices and m=n1 labeled edges? Prove that the number of trees that can
be built with m2 labeled edges (and no labels on the vertices) is (m+1)
m2
.
Explain why the condition that m2 is necessary.
4
Problem 2.10 A digraph which has no directed circuits is called a DAG
(directed acyclic graph). One wants an algorithm that checks whether a given
nite DAG, G(V, E), has a root.
One way to do this is rst to check that there is only one vertex r, such that
d
in
(r) = 0, and then check if all vertices are reachable from r. Explain why
this is valid, and prove that the time complexity of such an algorithmis O(|E|).
4
This problem was inspired by an unpublished report of S. Golomb and A. Lempel, and a
comment made by A. Pnueli.
44 2 Trees
Problem 2.11 Given a nite digraph G(V, E). Describe an algorithm which
runs in time O(|V|) and checks whether G is a directed tree.
Problem2.12 Prove that, if Gis an innite undirected connected graph whose
vertices are of nite degrees, then every vertex of G is the start vertex of some
simple innite path.
Problem 2.13 Show that, if rotation or ipping of tiles is allowed, then the
question of tiling the plane becomes trivial.
Problem 2.14 Consider the following (innite) set of tile families: For every
ordered pair of positive integers (i, j), there is a tile family with a label i 1 in
the West, i in the East, j 1 in the South, and j in the North. Prove that one can
tile the upper-right quadrant with this set of families, but not the whole plane.
Problem 2.15 Let T be an undirected tree with n vertices. We want to invest
O(n) time, labeling the graph in a way that will allow one to nd a (minimum)
path between any two vertices that are of distance apart, in time O(). Describe
both a preparatory algorithmand an algorithmfor nding the path once the two
vertices are specied.
Problem 2.16 A clique is a simple undirected graph such that for every two
vertices there is an edge connecting them. Let G(V, E) be a clique and T(V, E

)
a spanning tree of G. Prove that the complement of T (= (V, E

)) is either
connected or consists of one isolated vertex, while the remaining vertices form
a clique.
Problem2.17 Let G(V, E) be an undirectednite connectedgraph, where each
edge e has a given length l(e) >0 and s is a designated vertex. Also, let (v)
denote the distance froms to v.
(a) Explain why every vertex v ,= s, has an incident edge u
e
v, such that
(v) =(u) +l(e).
(b) Show that if one such edge is chosen for each vertex v ,=s, then the set of
these edges forms a spanning tree of G.
(c) Show that such a tree is not necessarily an MST.
Bibliography
[1] Prim, R.C., Shortest Connection Networks and Some Generalizations, Bell
System Tech. J., Vol. 36, 1957, pp. 13891401.
[2] Kruskal, J.B., On the Shortest Spanning Subtree of a Graph and the Traveling
Salesman Problem, Proc. of the Amer. Math. Society, Vol. 7, 1956, pp. 4850.
2.6 Problems 45
[3] Cayley, A., A Theorem on Trees, Quart. J. Math., Vol. 23, pp. 376378. Also in
Collected Papers, Vol. 13, Cambridge, 1897, pp. 2628.
[4] Prfer, H., Neuer Beweise eines Satzes ber Permutationen, Arch. Math. Phys.,
Vol. 27, 1918, pp. 742744.
[5] Moon, J.W., Various Proofs of Cayleys Formula for Counting Trees, A Seminar
on Graph Theory, F. Harary (ed.), Holt, Rinehart and Winston, 1967, pp. 7078.
[6] Berge, C., and A. Ghouila-Houri, Programming, Games and Transportation
Networks, Wiley, 1965, Sec. 7.4.
[7] Knig, D., Theorie der endlichen und unendlichen Graphen, Liepzig, 1936.
Reprinted by Chelsea, 1950.
[8] Wang, H., Proving Theorems by Pattern Recognition, Bell System Tech. J.,
Vol. 40, 1961, pp. 141.
3
Depth-First Search
3.1 DFS of Undirected Graphs
The depth-rst search (DFS) technique is a method of scanning a nite,
undirected graph. Since the publication of the papers of Hopcroft and Tar-
jan [4, 6], DFS has been widely recognized as a powerful technique for solving
various graph problems. However, the algorithmhas been known since the nine-
teenth century as a technique for threading mazes. See, for example, Lucas
report of Trmauxs work [5]. Another algorithm, which was suggested later
by Tarry [7], is just as good for threading mazes, and in fact, DFS is a special
case of it. But the additional structure of DFS is what makes the technique so
useful.
3.1.1 Trmauxs Algorithm
Assume one is given a nite, connected graph G(V, E), which we will also refer
to as the maze. Starting in one of the vertices, one wants to walk along the
edges, from vertex to vertex, visit all vertices, and halt. We seek an algorithm
that will guarantee that the whole graph will be scanned without wandering
too long in the maze, and that the procedure will allow one to recognize when
the task is done. However, before one starts walking in the maze, one does not
knowanything about its structure, and therefore, no preplanning is possible. So,
decisions about where to go next must be made one by one as one goes along.
We will use markers, which will be placed in the maze to help one to
recognize that one has returned to a place visited earlier and to make later
decisions on where to go next. Let us mark the passages, namely the connections
of the edges to vertices. If the graph is presented by incidence lists, then we
can think of each of an edges two appearances in the incidence lists of its two
endpoints as its two passages. It sufces to use two types of markers: F for the
rst passage used to enter the vertex, and E for any other passage used to leave
46
3.1 DFS of Undirected Graphs 47
Procedure TRMAUX(G, s)
1 v s
2 while there is an unmarked passage in v or v has a passage marked F do
3 if there is an unmarked passage to edge v
e
u, then do
4 mark the passage of e at v by E
5 if u has no marked passages then do
6 mark the passage of e at u by F
7 v u
8 else mark the passage of e at u by E
9 else (there is a passage in v marked F) do
10 use the passage marked F to move to the neighboring vertex u
11 v u
Algorithm 3.1: The Trmaux algorithm.
s
a
b
c
E
E
E
F
E
E
E
F E
E
Figure 3.1: An example of running Trmauxs algorithm.
the vertex. No marker is ever erased or changed. There is no use of memory
other than the markers on the passages and the stage of the algorithmone is has
reached; in other words, the moves are controlled by a nite state automaton.
As we shall prove later, the algorithmdescribed in Algorithm3.1 will terminate
in the original starting vertex s, after scanning each edge once in each direction.
Let us demonstrate the algorithmon the graph shown in Figure 3.1. The initial
value of v, the place where we are situated, or the center of activity, is s. All
passages are unlabeled. We choose one, mark it Eand traverse the edge. Its other
endpoint is a (u =a). None of its passages are marked, therefore we mark the
48 3 Depth-First Search
passage through which a has been entered by F, the new center of activity is
a (v = a), and we are back in Line 2. Since a has two unmarked passages,
assume that we choose the one leading to b. The passage at a is marked E and
the one at b is marked F since b is new, etc. The complete excursion is shown
in Figure 3.1 by the dashed line.
Lemma 3.1 Trmauxs algorithm never allows an edge to be traversed twice
in the same direction.
Proof:
1
If a passage is used as to exit a vertex and enter an edge), then either it
is being marked E in the process, and thus the edge is never traversed again in
this direction, or the passage is already marked F. It remains to be shown that
no passage marked F is ever reused for entering the edge.
Let u
e
v be the rst edge to be traversed twice in the same direction, from
u to v. The passage of e, at u, must be labeled F. Since s has no passages
marked F, u ,=s. Vertex u has been left d(u) +1 times; once through each of
the passages marked E and twice through e. Thus, u must have been entered
d(u) +1 times and some edge been used twice to enter u, before e is used for
the second time. A contradiction.
An immediate corollary of Lemma 3.1 is that the process described by Tr-
mauxs algorithm will always terminate. Clearly, it can only terminate in s,
since every other visited vertex has an F passage. Therefore, all we need to
prove is that, upon termination, the whole graph has been scanned.
Lemma 3.2 Upon termination of Trmauxs algorithm, every edge of the graph
has been traversed once in each direction.
Proof: Let us state the proposition differently: For every vertex, all its incident
edges have been traversed in both directions.
First, consider the start vertex s. Since the algorithm has terminated, all the
incident edges of s have been traversed from s outward. Thus, s has been left
d(s) times, and since we end up in s, it has also been entered d(s) times.
However, by Lemma 3.1 no edge is traversed more than once in the same
direction. Therefore, every edge incident to s has been traversed once in each
direction. Let S be the set of vertices for which the statement that each of their
incident edges has been traversed once in each direction holds. Since s S,
S ,= . Assume V ,= S. By the connectivity of the graph there must be edges
1
The following terminology is used in the proof. A passage from a node uto an edge v
e
u is
used as an exit if the algorithm exits v via the passage. In such a case, the passage is used to
enter the edge. (G.E.)
3.1 DFS of Undirected Graphs 49
connecting vertices of S with V \ S. All these edges have been traversed once
in each direction. Let e be the rst edge to be traversed from a vertex v S to
u V \ S. Clearly, the passage of e, at u, is marked F. Since this passage has
been entered, all other passages of u must have been marked E. Thus, each of
us incident edges has been traversed outward. The search has not started in u
and has not ended in u. Therefore, u has been entered d(u) times, and each of
its incident edges has been traversed inward. A contradiction, since u belongs
in S.
Observe that the loop (Lines 210) is applied at most once for every passage,
and the number of computational steps in each application of the loop is bounded
by a constant. Thus, the time complexity is O(|E|).
3.1.2 The Hopcroft-Tarjan Version of DFS
The Hopcroft and Tarjan version of DFS is essentially the same as Trmauxs,
except that they number the vertices from 1 to n(= |V|) in the order in which
they are discovered. This is not necessary, as we have seen, for scanning the
graph, but the numbering is useful in applying the algorithmfor more advanced
tasks. Let us denote the number assigned to vertex v by k(v). Also, instead of
marking passages, they mark edges as used, and instead of using the F mark to
indicate the edge through which the vertex was discovered and through which
it is left for the last time, let us record for each vertex v other than s the vertex
f(v) from which v has been discovered. Then f(v) is called the father of v; this
name will be justied later. DFS is described in Algorithm 3.2.
Since this algorithm is just a simple variation of the previous one, our proof
that the whole (connected) graph will be scanned, each edge once in each
direction, still applies. Here, in Line 11, if k(u) ,=0, then u is not a new vertex,
and v, the center of activity, does not change. This is equivalent to the scanning
of e fromv to u and back to v, as was done in Trmauxs version. Also, moving
our center of activity fromv to f(v) (Line 16) corresponds to traversing the edge
v f(v), in this direction. Thus, the whole algorithm is of time complexity
O(|E|), namely, linear in the size of the graph.
Now that we have applied DFS to a nite and connected G(V, E), let us
consider the set of edges E

, consisting of all edges of the form f(v) v

through which new vertices have been discovered. Also, direct each such edge
fromf(v) to v. The digraph (V, E

) is called the DFS tree. This name is justied

by the following Lemma:
Lemma 3.3 The digraph (V, E

) dened above is a directed tree with root s.

50 3 Depth-First Search
Procedure DFS(G(V, E), s;k(), f())
1 for every e E mark e new
2 for every u V do
3 k(u) 0
4 f(u) NIL
5 v s
6 k(s) 1
7 i 2
8 while v has a new incident edge or f(v) ,= NIL do
9 if v has a new incident edge v
e
u then do
10 mark e old
11 if k(u) = 0 (u is a new vertex) then do
12 f(u) v
13 k(u) i
14 i i +1
15 v u
16 else v f(v)
Algorithm 3.2: DFS.
Proof: Since f(s) = NIL, d
in
(s) = 0. For every other vertex v, d
in
(v) = 1.
Now one can nd a directed path
s =v
0
e
1
v
1
e
2
v
2
v
l
=v
from s to any vertex v, where for every 1 i l, e
i
is the directed edge from
f(v
i
) to v
i
, by starting fromv and tracing the path backwards. Since for every i,
f(v
i
) was discovered before v
i
, the directed path is simple. Also, every vertex,
other than s, has a dened father, and therefore the only vertex in which this
backwards search can end is s.
By Theorem 2.5, Part (c), (V, E

) is a directed tree.
Clearly, if one ignores the edge directions, (V, E

) is a spanning tree of G.
In a directed tree, vertex u is called an ancestor of v, and v is called a
descendant of u if there is a directed path from u to v.
The following very useful lemma is due to Hopcroft and Tarjan [4, 6]:
3.1 DFS of Undirected Graphs 51
Lemma 3.4 Let (V, E

) be the undirected version of a DFS tree, T, of G(V, E).

If an edge a
e
b is in E \ E

, then either a is an ancestor of b or a is a

descendant of b in T.
Proof: Without loss of generality, assume that k(a) <k(b). In the DFS algo-
rithm, the center of activity (v in the algorithm) moves only along the edges of
the tree (V, E

). If b is not a descendant of a, and since a is discovered before

b, the center of activity must rst move froma to some ancestor of a before it
moves up to b. However, we backtrack from a (in Line 16 of Algorithm 3.2)
only when all as incident edges are old. This means that e is old. Thus, b
is already discovered. A contradiction.
Let us call the edges of the DFS tree tree edges, and all other edges of the
graph, back edges. The justication for this name is in Lemma 3.4; every nontree
edge connects some vertex back to one of its ancestors.
Consider, as an example, the graph shown in Figure 3.2. We apply DFS to this
graph, starting with s =c, and assume that we discover vertices d, e, f, g, b, a,
in this order. The resulting vertex numbers, tree edges, and back edges are shown
in Figure 3.3, where the tree edges are shown as solid lines and are directed
fromlowto high, and the back edges are shown as dashed lines and are directed
from high to low. In both cases the direction of the edge indicates the direction
in which the edge was scanned rst. For tree edges this is the dened direction,
and for back edges we can prove it as follows: Assume v
e
u is a back edge,
and u is an ancestor of v. The edge e could not have been scanned rst fromu,
for if v was new at that time, then e would have been a tree edge, and if v has
already been discovered (after u), then the center of activity could have been in
u only if we have backtracked from v, and this means that e has already been
scanned from v.
a
b
c
d
e
f
g
e
1
e
2
e
3
e
5
e
6
e
8
e
9
e
4
e
7
Figure 3.2: A graph to which DFS is applied.
52 3 Depth-First Search
c
1
d
2
e
3
b
6
f
4
a
7
g
5
e
6
e
7
e
5
e
3
e
2
e
1
e
8
e
9
e
4
Figure 3.3: The DFS tree edges and back edges.
3.2 Algorithm for Nonseparable Components
A connected graph G(V, E) is said to have a separation vertex v (sometimes
also called an articulation point) if there exist vertices a and b, distinct from
v, such that every path connecting a and b passes through v. In this case, we
also say that v separates a fromb. A graph that has a separation vertex is called
separable, and one that has none is called nonseparable.
Let V

V. The induced subgraph G

(V

, E

) is called a nonseparable com-

ponent if G

is nonseparable, and if for every V

V the induced
subgraph G

(V

, E

) is separable. For example, in the graph shown in

Figure 3.2, the subsets {a, b}, {b, c, d} and {d, e, f, g} induce the nonsepara-
ble components of the graph. If a graph G(V, E) contains no separation vertex,
then clearly the whole G is a nonseparable component. However, if v is a
separation vertex, then V \ {v} can be partitioned into {V
1
, V
2
, . . . , V
k
} such
that V
1
V
2
V
k
= V \ {v}, and if i ,= j, then V
i
V
j
= . Also, two
vertices a and b are in the same V
i
if and only if there is a path connecting
them that does not include v. Thus, no nonseparable component can contain
vertices from more than one V
i
. We can next consider each of the subgraphs
induced by V
i
{v} and continue to partition them into smaller parts, if they
are separable. Eventually, we end up with nonseparable parts. This shows that
3.2 Algorithm for Nonseparable Components 53
no two nonseparable components can share more than one vertex because each
such vertex is a separating vertex. Also, every simple circuit must lie entirely in
one nonseparable component. Now, let us discuss how DFS can help to detect
separating vertices.
Let the lowpoint of v, L(v), be the least number, k(u), of a vertex u that can
be reached fromv by a, possible empty, directed path consisting of tree edges,
followed by at most one back edge. Clearly L(v) k(v), for we can use the
empty path fromv to itself. Also, if a nonempty path is used, then its last edge
is a back edge, for a directed path of tree edges leads to vertices higher than v.
For example, in the graph of Figure 3.2, with the DFS as shown in Figure 3.3,
the lowpoints are as follows: L(a) = 7, L(b) = L(c) = L(d) = 1 and L(e) =
L(f) =L(g) =2.
Lemma 3.5 Let G be a graph whose vertices have been numbered by DFS. If
uv is a tree edge, k(u) >1, and L(v) k(u), then uis a separating vertex
of G.
Proof: Let S be the set of vertices on the path from the root r (k(r) = 1)
to u, including r but not including u, and let T be the set of vertices on the
subtree rooted at v, including v (i.e., all descendants of v, including v itself). By
Lemma 3.4, there cannot be any edge connecting a vertex of T with any vertex
of V \ (S {u} T). Also, if there is any edge connecting a vertex t T with
a vertex s S, then the edge t s is a back edge and clearly k(s) < k(u).
Now, L(v) k(s), since one can take the tree edges from v to t, followed by
t s. Thus, L(v) <k(u), contradicting the hypothesis. Thus, u separates the
S vertices from the T vertices and is, therefore, a separating vertex.
Lemma 3.6 Let G(V, E) be a graph whose vertices have been numbered by
DFS. If u is a separating vertex, and k(u) > 1, then there exists a tree edge
u v such that L(v) k(u).
Proof: Since u is a separating vertex, there is a partition of V \ {u} into
V
1
, V
2
, . . . , V
m
such that m 2, and if i ,= j, then all paths from a ver-
tex of V
i
to a vertex of V
j
, pass through u. Since k(u) >1, the the search has
not started in u. Let us assume that it starts in r, and w.l.o.g. r V
1
. The center
of activity of the DFS must pass through u. Let u v be the rst tree edge
for which v / V
1
. W.l.o.g. assume v V
2
. Since there are no edges connecting
vertices of V
2
with vertices of V \ (V
2
{u}), L(v) k(u).
54 3 Depth-First Search
Lemma 3.7 Let G(V, E) be a graph whose vertices have been numbered by
DFS, starting with r (k(r) = 1). Vertex r is a separating vertex if and only if
there are at least two tree edges out of r.
Proof: Assume r is a separating vertex. Let V
1
, V
2
, . . . , V
m
be a partition
of V \ {r} such that m2, and if i ,=j, then all paths from a vertex of V
i
to a
vertex of V
j
, pass through u. Therefore, no path in the tree starting with r v,
v V
i
, can lead to a vertex of V
j
where j ,=i. Thus, there are at least two tree
edges out of r.
Now, assume r v
1
and r v
2
are two tree edges out of r. Let T be the
set of vertices in the subtree rooted at v
1
. By Lemma 3.4, there are no edges
connecting vertices of T with vertices of V \ (T {r}). Thus, r separates T
from the rest of the graph, which is not empty, since it includes at least the
vertex v
2
.
Let C
1
, C
2
, . . . , C
m
be the nonseparable components of a connected graph
G(V, E), and let s
1
, s
2
, . . . , s
p
be its separating vertices. Let us dene

G(

V,

E),
the superstructure of G(V, E), as follows:

V = {s
1
, s
2
, . . . , s
p
} {C
1
, C
2
, . . . , C
m
},

E = {s
i
C
j
| s
i
is a vertex of C
j
in G}.
By the observations we made in the beginning of the section,

G is a tree. By
Corollary 2.1, if m>1, then there must be at least two leaves in

G. However,
the degree of a separating vertex s
i
in

G is greater than 1. Thus, there are at
least two leaf components in G, each containing only one separating vertex.
By Lemma 3.2, the whole graph will be explored by the DFS. Now, assume
the search starts in a vertex r that is not a separating vertex. Even if it is in one of
the leaf components, eventually, we will enter another leaf component C, say,
via its separating vertex s and an edge s v. By Lemma 3.6, L(v) k(s),
and if L(v) is known when we backtrack fromv to s, then by using Lemma 3.5,
we can detect that s is a separating vertex. Also, as far as the component C is
concerned, from the time C is entered via s v until it is entirely explored,
we can think of the algorithm as running on C alone, with s as the starting
vertex. Thus, by Lemma 3.7, there is only one tree edge froms to other vertices
of C, and all other vertices of Care descendants of v and are therefore explored
after v is discovered and before the backtrack from v to s. This suggests the
use of a stack (pushdown store) for producing the vertices of the component.
We store the vertices in the stack in the order that they are discovered. If on
backtracking from v to f(v), we discover that f(v) is a separating vertex, we
3.2 Algorithm for Nonseparable Components 55
read off all vertices fromthe top of the stack down to and including v. All these
vertices, plus f(v), (which is not removed at this point from the stack even if it
is the next on top) constitute a component. This, in effect, removes the leaf C
fromthe tree

G, and if its adjacent separating vertex s has nowd(s) = 1, then we
may assume that it is removed too. The new superstructure is again a tree, and
the same process will repeat itself, detecting and trimming one leaf at a time
until only one component is left when the DFS terminates.
If the search starts in a separating vertex r, then all but the components con-
taining r are detected and produced as before. All components that contain r,
except the last one, are detected by Lemma 3.7: Each time we backtrack into r,
on r v, if r still has additional unexplored incident edges, then we conclude
that r is a separating vertex, and the vertices on the stack above, including v,
plus r, constitute a component.
Finally, when the search is about to end, since we are going to backtrack
to r, and r has no new incident edges, all vertices on the stack form the last
component, although no separating vertex is discovered at this point.
The remaining problem is that of computing L(v) in time; that is, its value
should be known by the time we backtrack from v. If v is a leaf of the DFS
tree, then L(v) is the least element in the following set: {k(u) | u =v or v
uis a back edge}. Let us assign L(v) =k(v) immediately when v is discovered,
and as each back edge v uis explored, let us assign L(v) =min{L(v), k(u)}.
Clearly, by the time we backtrack fromv, all the back edges have been explored,
and L(v) has the right value. If v is not a leaf of the DFStree, then L(v) is the least
element in the following set: {k(u) | u =v or v u is a back edge} {L(u) |
v uis a tree edge}. When we backtrack fromv, we have already backtracked
from all its sons earlier, and therefore already know their lowpoint. Thus, in
addition to what we do for a tree leaf, it sufces to do the following: When we
backtrack from u to v =f(u), we assign L(v) = min{L(v), L(u)}.
Putting together the ideas described above leads to the algorithmrepresented
in Algorithm 3.3. In this representation, L() is the lowpoint function, and
S is a stack of vertices. The remaining variables are as in DFS. For the given
undirected, connected, and nite graphG(V, E), |V| >1, the algorithmproduces
the set of separating vertices and a list of its nonseparable components, both of
which are assumed to be initially empty. Note that just before the last backtrack
into s, Lines 2627 produce the last nonseparable component, which may be
the entire V if G has no separating vertices.
Although this algorithm is more complicated then the original DFS, its time
complexity is still O(|E|). This follows easily from the fact that each edge is
still scanned exactly once in each direction. The number of operations per edge
56 3 Depth-First Search
Procedure NONSEPARABLE(G(V, E), s;set of separating vertices,
list of nonseparable components)
1 for every e E mark e new
2 for every u V do
3 k(u) 0
4 f(u) NIL
5 v s
6 k(s) 1
7 i 2
8 vacate S
9 push s into S
10 while v has a new incident edge or f(v) ,= NIL do
11 if v has a new incident edge v
e
u then do
12 mark e old
13 if k(u) = 0 (u is a new vertex) then do
14 push u into S
15 f(u) v
16 k(u) i
17 L(u) i
18 i i +1
19 v u
20 else (u is old) do
21 L(v) min{L(v), k(u)}
22 else (f(v) is dened)
23 if L(v) k(f(v)) then do
24 if f(v) ,=s or s has a new incident edge, then do
25 add f(v) to the set of separating vertices
26 pop vertices from S down to and including v
27 the set of popped vertices, with f(v), is an element
of the set of nonseparable components
28 else (L(v) <k(f(v))) then do
29 L(f(v)) min{L(f(v)), L(v)}
30 v f(v)
Algorithm 3.3: Using DFS to nd the separating vertices and
nonseparable components.
3.3 DFS on Directed Graphs 57
is bounded by a constant, except when a nonseparable component is produced.
Each vertex is pushed into S once, and popped once. Thus, the total time to
produce the components is O|V|.
3.3 DFS on Directed Graphs
Running DFS on digraphs is similar to runningit on undirected graphs. We start
scanning from a new vertex and will scan all vertices (and edges) that can be
reached from it via directed paths. A new scan is started from a new vertex, as
long as the previous search has left unscanned vertices. Thus, the fact that the
whole graph is scanned is trivial.
Assuming the given nite digraph is G(V, E), where |V| =n. Again, we assign
a number k(u) to every vertex u, where k : V {1, 2, . . . , n} is a bijection. If
a vertex u is rst discovered by scanning an edge v u, then we assign
f(u) =v. Here, too, v denotes the center of activity.
The algorithm is described in Algorithm 3.4. It is easy to see that the time
complexity is O(|V| + |E|).
Let us call a vertex v ripe if all its outgoing edges are old and the center of
activity is at v. When v is ripe, and if f(v) ,= NIL, then we backtrack to f(v).
Otherwise, we return to Line 6 in Algorithm 3.4.
Let us denote by T
u
the directed subtree whose root is u, and all vertices that
are discovered from the time u is discovered to the time u is ripe are in it.
2
The
edges of T
u
are of the formf(w)
e
w, where both f(w) and ware in T
u
, and
w has been discovered via e.
Lemma 3.8 If vertex wis newwhen u is discovered, and if at that time there is
a directed path from u to w such that all its intermediate vertices (and edges)
are new, then w is in T
u
.
Proof: By contradiction. Assume that wis as in the premise of the Lemma, but
wis not in T
u
. Let P be a directed path fromuto wsuch that all its intermediate
vertices are new when u is discovered, and let b be the rst vertex on P that
does not belong to T
u
. Let a
e
b be the edge on P that enters b.
Since a T
u
, and it eventually becomes ripe, e must have been investigated,
as in Line 11 of Algorithm 3.4, and b has been discovered and belongs to T
u
.
A contradiction.
2
Note that u is not necessarily a root of a search in the sense that it may not have been picked in
Line 7, and may belong to some T
v
for v =u.
58 3 Depth-First Search
Procedure Directed-DFS(G(V, E), s;k(), f())
1 for every e E mark e new
2 for every u V do
3 k(u) 0
4 f(u) NIL
5 i 1
6 while there is a vertex u for which k(u) = 0 do
7 let v be such a vertex
8 k(v) i
9 i i +1
10 while v has a new outgoing edge or f(v) ,= NIL do
11 if v has a new outgoing edge v
e
w, then do
12 mark e old
13 if k(w) = 0 (u is a new vertex) then do
14 f(w) v
15 k(w) i
16 i i +1
17 v w
18 else v f(v)
Algorithm 3.4: DFS on a directed graph.
3.4 Strongly Connected Components of a Digraph
Let G(V, E) be a nite digraph. Let us dene the relation , a subset of V V,
in the following way: For x, y V, x y if there is a directed path from x to y
and also a directed path from y to x.
The relation is easily seen to be reexive, symmetric, and transitive. Thus,
it is an equivalence relation. An equivalence class of this relation is called a
strongly connected component or, in short, a strong component, and if there is
only one equivalence class, then G is said to be strongly connected.
The super-structure of G,

G(

V,

E) is a digraph constructed as follows:

V is the set of strong components of V.

E = {X
e
Y | there exists an edge x
e
y in E such that x X and y Y}.
3.4 Strongly Connected Components of a Digraph 59
Observe that

G is a DAG (directed acyclic graph). A strong component C is
called a source if there are no edges that enter C in

G. A sink component is
similarly dened.
Given a digraph G(V, E), our purpose is to describe an efcient algorithmfor
nding its strong components. The rst linear-time algorithmthat achieved this
goal was presented by Tarjan [6]. However, we present an algorithm attributed
to Kosaraju and Sharir [1], which is simpler to explain.
3
The algorithm consists of three phases:
Phase 1: Run a DFS-A on G, in which vertices are numbered h : V
{1, 2, . . . , n} in the order in which they become ripe. This is described
in Algorithm 3.5.
Phase 2: Reverse the direction of all edges of G(V, E) to obtain G
R
(V, E
R
).
Phase 3: Run a DFS-B on G
R
, where each time a new search begins, it is
started in the new vertex v for which h(v) is maximum; the set of
vertices found in a search is declared to be a strong component of G.
This is described in Algorithm 3.6.
Observe that the strong components of G
R
are identical to those of G. Also,
the algorithm consists of three phases, each of time complexity O(|V| + |E|).
Thus, the whole algorithm is of time complexity O(|V| + |E|).
Lemma 3.9 The function h(), produced by DFS-A, satises the following
condition: For every u V,
h(u) = max
wT
u
{h(w)} .
Proof: Follows from the fact that u is the last vertex in T
u
to become ripe.
Lemma 3.10 Let C be a strong component of G, and let u be the vertex of
C for which h(u) is maximum. It follows that u is the rst vertex of C to be
discovered in DFS-A.
Proof: If a is the rst vertex of C to be discovered in DFS-A, then by
Lemma 3.8, u T
a
. By Lemma 3.9, h(a) h(u). Since h is a bijection,
and h(u) is maximum in C, it follows that a =u.
Corollary 3.1 If the premise of Lemma 3.10 holds, then all vertices of C are
in T
u
.
3
The algorithm of Tarjan uses DFS once, while that of Kosaraju and Sharir uses two runs of DFS.
60 3 Depth-First Search
Procedure DFS-A(G(V, E);h())
1 for every e E mark e new
2 for every u V do
3 f(u) NIL
4 mark u new
5 i 1
6 while there are new vertices do
7 let v be such a vertex
8 mark v old
9 while v has new outgoing edges or f(v) ,=NIL do
10 if v has new outgoing edges, then do
11 let v
e
w be a new edge
12 mark e old
13 if w is new then do
14 mark w old
15 f(w) v
16 v w
17 else (f(v) ,=NIL) then do
18 h(v) i
19 i i +1
20 v f(v)
21 h(v) i
22 i i +1
Algorithm 3.5: DFS-A.
Lemma 3.11 Assume DFS-A was applied to G and vertex v was assigned the
highest value of h(v). The strong component C to which v belongs is a source
component.
Proof: By contradiction. Assume there is an edge a
e
b, such that a / C
and b C.
Let r be the root, chosen in Line 7, of the search in which a is discovered.
Then r / C, for otherwise, there would be a directed path fromC, which starts
in r, to a, to b, and thus, a belongs to C, contradicting the assumption that
a / C. We conclude that r ,=v.
3.4 Strongly Connected Components of a Digraph 61
Procedure DFS-B(G
R
(V, E
R
), h();strong components of G)
1 for every e E
R
mark e new
2 for every u V do
3 f(u) NIL
4 mark u new
5 while there are new vertices do
6 let v be the new vertex for which h(v) is maximum
7 S {v}
8 mark v old
9 while v has new outgoing edges or f(v) ,=NIL do
10 if v has new outgoing edges, then do
11 let v
e
w be a new edge
12 mark e old
13 if w is new then do
14 mark w old
15 f(w) v
16 S S{w}
17 v w
18 else (f(v) ,=NIL) then do
19 v f(v)
20 print: The set S is a strong component
Algorithm 3.6: DFS-B.
Observe that the last vertexto serve as a root of a search is the vertex for which
h() is maximum. Thus, v is a root of a search. Also, T
r
is constructed before v
is discovered. But when r is chosen to be a root of a search, all vertices on the
path fromr to a, to b, to v, are new. Thus, by Lemma 3.8, v T
r
, contradicting
the fact that v is a root of a new search.
Since the component C, containing the vertex v for which h(v) is maximum,
is a source component of G, it is a sink component of G
R
. It follows that in
DFS-B, since the root of the rst search is v, all vertices of C, and none else,
are discovered in this search, and indeed, this set S is declared to be a strong
component.
62 3 Depth-First Search
Procedure TARRY(G, s)
1 v s
2 while there is an unmarked passage in v or v has a passage marked F do
3 if there is an unmarked passage to edge v
e
u then do
4 mark the passage of e at v by E
5 if u has no marked passages then do
6 mark the passage of e at u by F
7 v u
8 else (there is a passage in v marked F) do
9 use the passage marked F to move to the neighboring vertex u
10 v u
Algorithm 3.7: The Tarry algorithm.
Now, assume we remove from G all vertices of the rst declared component
and their incident (outgoing) edges to forma directed subgraphG

. The remain-
ing directed subforest, of the original DFS-A forest of G, and the values of h()
for the remaining vertices are legitimate in the sense that one can run DFS-A
on G

to yield exactly this forest and assignments of h(). It follows that the
next declared strong component is valid too. By induction, all declared strong
components are valid as well.
3.5 Problems
Problem 3.1 Tarrys algorithm [7] is like Trmauxs, with the following
change. Upon reaching an old vertex u, one moves the center of activity to
that vertex anyway, instead of insisting on marking the passage through which
one has just reached u by E and not moving the center of activity to u. Tarrys
algorithmis described in Algorithm3.7. Prove that Tarrys algorithmterminates
after all edges of G have been traversed, once in each direction.
Problem 3.2 Consider the set of edges which upon termination of Tarrys
algorithm(see Problem 3.1) have one endpoint marked E and the other marked
F. Also, assume these edges are now directed from E to F.
(i) Prove that this set of edges form a directed spanning tree of G, with root s.
(ii) Does a statement like that of Lemma 3.4hold in this case? Prove or disprove.
3.5 Problems 63
Problem 3.3 Fraenkel [2, 3] showed that the number of edge traversals can
sometimes be reduced, in comparison to the Tarry algorithm(see Problem3.1),
if the use of a two-way counter is allowed. Each time a new vertex is entered,
the counter is incremented. When it is realized that all incident edges of a vertex
have been traversed at least in one direction, the counter is decremented. If the
counter reaches the start value, the search is terminated. (One can return to s
via the passages marked F.)
Write an algorithm that realizes this idea. (Hint: An additional mark that
temporarily marks the passages used to reenter a vertex is used.) Prove the
validity of your algorithm. Showthat for some graphs the algorithmwill traverse
each edge exactly once for others; the savings depends on the choice of passages.
Yet there are graphs for which the algorithm can save only one traversal, even
if we do not insist on returning to s.
Problem 3.4 Assume that G is drawn in the plane in such a way that no two
edges cross. Show how Trmauxs algorithm can be modied in such a way
that the whole scanning path never crosses itself.
Problem 3.5 In an undirected graph G, a set of vertices K is called a clique if
every two vertices of K are connected by an edge. Prove that, in the spanning
(directed) tree resulting from running DFS on a nite and connected G, all
vertices of a clique appear on one directed path. Do they necessarily appear
consecutively on the path? Justify your answer.
Problem 3.6 Prove or disprove the following claim: If C is a simple circuit in
an undirected, nite, and connected graph G to which DFS is applied, then all
vertices of C appear on one directed path of the (directed) DFS tree. Does you
answer change if C is an induced circuit? (A circuit C is an induced circuit in
G if G does not contain an edge between two nonadjacent vertices in C.)
Problem 3.7 Prove that if C is a directed circuit of a nite digraph to which
DFS is applied and v is a vertex on C for which k(v) is minimum, then v is a
root of a subtree of the resulting directed forest, and all vertices of C are in this
subtree.
Problem3.8 An edge e of a connected undirected graph Gis called a bridge if
the deletion of e destroys Gs connectivity. Describe an algorithm to compute
all bridges of a given nite G. (Hint: There are two ways to solve the problem.
In the rst solution, one modies the graph and uses the algorithm for nding
the separating vertices. In the second, one modies the algorithm and applies
the modied algorithm to G.)
64 3 Depth-First Search
Problem 3.9 (This problem was suggested by Silvio Micali.) Let G be a
connected graph.

Prove that a vertex u ,= s is a separating vertex of G if and only if, upon

termination of DFS on G, there is a tree edge u v for which there is no
back edge x y such that x is a descendant of v and y is a proper ancestor
of u.

Describe an algorithm, of time complexity O(|E|), which detects the sepa-

rating vertices of G without numbering the vertices and, therefore, without
the use of lowpoint. (Hint: For every back edge x y, mark all tree edges
on the path from y to x; proceed from x to y until an already marked edge
is encountered or the edge tree outgoing from y is reached. The latter edge
is not marked. The back edges are considered by rescanning G again, while
remembering which edges are tree edges and which are back edges, only that
in the second scan the back edges are processed rst. When all this is over,
v ,=s is a separating vertex if and only if there is an unmarked tree edge out
of v. Also note that a third scan can be used to produce the nonseparable
components.)
Problem3.10 (This problemwas suggested by Alessandro Tescari.) Showthat
the algorithm for nonseparable components can be simplied by adding a new
vertex r and a new edge r s, and starting the search at r. Do we still need
Lemma 3.7?
Bibliography
[1] J.E. Hopcroft, A.V. Aho, and J.D. Ullman. Data Structures and Algorithms.
Addison-Wesley, 1983.
[2] A.S. Fraenkel. Economic traversal of labyrinths. Math. Mag., 43(1):125130,
1970.
[3] A.S. Fraenkel. Economic traversal of labyrinths (correction). Math. Mag., 44,
1971.
[4] J. Hopcroft and R. Tarjan. Algorithm 447: Efcient algorithms for graph
manipulation. Comm. of the ACM, 16:372378, 1973.
[5] E. Lucas. Rcreations Mathmatiques. Paris, 1882.
[6] R. Tarjan. Depth-rst search and linear graph algorithms. SIAMJ. on Computing,
1:146160, 1972.
[7] G. Tarry. Le problme des labyrinthes. Nouvelles Annales de Math., 14:187,
1895.
4
Ordered Trees
4.1 Uniquely Decipherable Codes
Let = {0, 1, . . . , 1}. We call an alphabet and its elements are called
letters; the number of letters in is . (Except for this numerical use of ,
the numerical value of the letters is ignored; they are just meaningless
characters. We use the numerals just because they are convenient characters.)
Anite sequence a
1
a
2
a
l
, where a
i
is a letter, is called a word whose length
is l. We denote the length of a word w by l(w). A set of (nonempty and
distinct) words is called a code. For example, the code {102, 21, 00} consists of
three code-words: one code-word of length 3 and two code-words of length 2;
the alphabet is {0, 1, 2} and consists of three letters. Such an alphabet is called
ternary.
Let c
1
, c
2
, . . . , c
k
be code-words. The message c
1
c
2
c
k
is the word resulting
from the concatenation of the code-word c
1
with c
2
, and so on. For example, if
c
1
= 00, c
2
= 21, and c
3
= 00, then c
1
c
2
c
3
= 002100.
A code C over (i.e., the code-words of C consist of letters in ) is said to
be uniquely decipherable (UD) if every message constructed from code-words
of C can be broken down into code-words of C in only one way. For example,
the code {01, 0, 10} is not UD because the message 010 can be parsed in two
ways: 0, 10 and 01, 0.
Our rst goal is to describe a test for deciding whether a given code Cis UD.
This test is an improvement on a test of Sardinas and Patterson [1] and can be
found in Gallagers book [2].
If s, p, and w are words, and ps =w, then p is is called a prex of w and s
is called a sufx of w. We say that a word w is nonempty if l(w) >0.
A nonempty word t is called a tail if there exist two messages c
1
c
2
c
m
and
c

1
c

2
c

n
with the following properties:
65
66 4 Ordered Trees
(1) c
i
, 1 i m, and c

j
, 1 j n are code-words, and c
1
,=c

1
;
(2) t is a sufx of c

n
;
(3) c
1
c
2
c
m
t =c

1
c

2
c

n
.
Lemma 4.1 A code C is UD if and only if no tail is a code-word.
Proof: If a code-word c is a tail then, by denition, there exist two messages
c
1
c
2
c
m
and c

1
c

2
c

n
that satisfy c
1
c
2
c
m
c =c

1
c

2
c

n
, while c
1
,=c

1
.
Thus, there are two different ways to parse this message, and C is not UD.
If C is not UD, then there exist messages that can be parsed in more than
one way. Let be such an ambiguous message, whose length is minimum: =
c
1
c
2
c
k
=c

1
c

2
c

n
; i.e. all the c
i
-s and c
j
-s are code-words and c
1
,= c

1
.
Now, without loss of generality we can assume that c
k
is a sufx of c

n
(or
change sides). Thus, c
k
is a tail.
The following algorithmgenerates all the tails. If a code is a tail, the algorithm
terminates with a negative answer.
Algorithm for UD:
(1) For every two code-words, c
i
and c
j
(i ,=j), do the following:
a. If c
i
=c
j
, halt; C is not UD.
b. If for some word s, either c
i
s =c
j
or c
i
=c
j
s, put s in the set of tails.
(2) For every tail t and every code-word c do the following:
a. If t =c, halt; C is not UD.
b. If some word s, either ts =c or cs =t, put s in the set of tails.
(3) Halt; C is UD.
Clearly, in Step (1), the words declared to be tails are indeed tails. In step
(2), since t is already known to be a tail, there exist code-words c
1
, c
2
, . . . , c
m
and c

1
, c

2
, . . . , c

n
such that c
1
c
2
c
m
t = c

1
c

2
c

n
. Now, if ts = c, then
c
1
c
2
c
m
c = c

1
c

2
c

n
s, and therefore s is a tail; and if cs = t, then
c
1
c
2
c
m
cs =c

1
c

2
c

n
and s is a tail.
Next, if the algorithmhalts in (3), we want to showthat all the tails have been
produced. Once this is established, it is easy to see that the conclusion that C
is UD follows; each tail has been checked, in Line (2a.), whether it is equal to
a code-word, and no such equality has been found. By lemma 4.1, the code C
is UD.
For every t let m(t) = c
1
c
2
c
m
be a shortest message such that
c
1
c
2
c
m
t = c

1
c

2
c

n
and t is a sufx of c

n
. We prove by induction on
4.1 Uniquely Decipherable Codes 67
the length of m(t) that t is produced. If m(t) =1, then t is produced by ((1)b.),
since m=n = 1.
Now assume that all tails p for which m(p) < m(t) have been produced.
Since t is a sufx of c

n
, let p denote the word such that pt = c

n
. Therefore,
c
1
c
2
c
m
=c

1
c

2
c

n1
p, and p is a tail.
If p =c
m
, then c
m
t =c

n
, and t is produced in Step (1).
If p is a sufx of c
m
, then p is a tail. Also, m(p) is shorter then m(t). By
the inductive hypothesis, p has been produced. When Step (2b) is applied to
pt =c

n
(with p as a tail and c

n
as a code-word), the tail t is produced.
If c
m
is a sufx of p, then c
m
t is a sufx of c

n
, and therefore, c
m
t is a
tail. Also, m(c
m
t) =c
1
c
2
c
m1
and is shorter than m(t). By the inductive
hypothesis, c
m
t has been produced. When Step (2b) is applied to the tail c
m
t
and code-word c
m
, the tail t is produced.
This proves that the algorithm halts with the right answer.
Let the code consist of nwords, and l be the maximumlength of a code-word.
Step (1) takes at most O(n
2
l) elementary operations. The number of tails is
at most O(n l). Thus, Step (2) takes at most O(n
2
l
2
) elementary operations.
Therefore, the whole algorithmis of time complexity O(n
2
l
2
). Other algorithms
of the same complexity can be found in [3] and [4]; these tests are extendible
to test for additional properties [5, 6, 7].
Theorem 4.1 Let C = {c
1
, c
2
, . . . , c
n
} be a UD code over an alphabet of
letters. If l
i
=l(c
i
), i = 1, 2, . . . , n, then
n

i=1

l
i
1. (4.1)
The left-hand side of 4.1 is called the characteristic sum of C; clearly, it
characterizes the vector (l
1
, l
2
, . . . , l
n
), rather than C. The inequality 4.1is called
the characteristic sumcondition. The theoremwas rst proved by McMillan [8].
The following proof is due to Karush [9].
Proof: Let e be a positive integer
_
n

i=1

l
i
_
e
=
n

i
1
=1
n

i
2
=1

n

i
e
=1

(l
i
1
+l
i
2
++l
i
e
)
.
There is a unique term, on the right-hand side, for each of the n
e
messages of
e code-words. Let us denote by N(e, j) the number of messages of e code-words
68 4 Ordered Trees
whose length is j. It follows that
n

i
1
=1
n

i
2
=1

n

i
e
=1

(l
i
1
+l
i
2
++l
i
e
)
=
e

j=e
N(e, j)
j
,
where

l is the maximum length of a code-word. Since C is UD, no two
messages can be equal. Thus, N(e, j)
j
. We now have
e

j=e
N(e, j)
j

j=e

j

j
e

l.
We conclude that for all e 1,
_
n

i=1

l
i
_
e
e

l.
This implies 4.1.
A code C is said to be prex if no code-word is a prex of another. For
example, the code {00, 10, 11, 100, 110} is not prex, since 10 is a prex of
100; the code {00, 10, 11, 010, 011} is a prex. A prex code has no tails and
is therefore UD. In fact, it is very easy to parse the messages: Reading the
messages from left to right, as soon as we read a code-word, we know that
it is the rst code-word of the message, since it cannot be the beginning of
another code-word. Therefore, in most applications, prex codes are used. The
following theorem, due to Kraft [10], in a sense shows us that we do not need
nonprex codes.
Theorem 4.2 If the vector of integers, (l
1
, l
2
, . . . , l
n
), satises
n

i=1

l
i
1, (4.2)
then there exists a prex code C= {c
1
, c
2
, . . . , c
n
} over the alphabet of letters
such that l
i
=l(c
i
).
Proof: Let
1
<
2
< <
m
be integers such that each l
i
is equal to one of
the
j
-s and each
j
is equal to at least one of the l
i
-s. Let k
j
be the number of
4.2 Positional Trees and Huffmans Optimization Problem 69
l
j
-s that are equal to
j
. We have to show that there exists a prex code C such
that the number of code-words of length
j
is k
j
.
Clearly, 4.2 implies that
m

j=1
k
j

j
1 (4.3)
We prove by induction on r that for every 1 r m there exists a prex
code C
r
such that, for every 1 j r, the number of its code-words of
length
j
is k
j
.
First assume that r = 1. Inequality 4.3 implies that k
1

1
1, or k
1

1
.
Since there are

1
distinct words of length
1
, we can assign any k
1
of them
to constitute C
1
.
Now, assume C
r
exists. If r <m then 4.3 implies that
r+1

j=1
k
j

j
1.
Multiplying both sides by

r+1
yields
r+1

j=1
k
j

r
+1
j

r
+1
,
which is equivalent to
k
r+1

r+1

j=1
k
j

r
+1
j
. (4.4)
The number of distinct words of length
r+1
, a prex of which of length
j
is a code-word in C
r
, equals k
j

r
+1
j
. Thus, 4.4 implies that among the

r+1
words of length
r+1
, there are at least k
r+1
words, none of which has a
prex in C
r
. The enlarged set of code-words is C
r+1
.
This proof suggests an algorithm for the construction of a code with a given
vector of code-word length. We return later to the question of prex code
construction, but rst we introduce positional trees.
4.2 Positional Trees and Huffmans Optimization Problem
A positional -tree (or when is known, a positional tree) is a directed tree
with the following property: Each edge out of a vertex v is associated with one
70 4 Ordered Trees
00
0
01
011 100 101
10
1

0
0 0 1
1 1 0
1
Figure 4.1: A positional 2-tree.
of the letters of the alphabet = {0, 1, . . . , 1}; different edges, out of v, are
associated with different letters. It follows that the number of edges out of a
vertex is at most , but may be less; in fact, a leaf has none.
We associate with each vertex v the word consisting of the sequence of letters
associated with the edges on the path fromthe root r to v. For example, consider
the binary tree (positional 2-tree) of Figure 4.1, where the associated word is
written in each vertex. ( denotes the empty word.)
Clearly, the set of words associated with the leaves of a positional tree is a pre-
x code. Also, every prex code canbe described by a positional tree inthis way.
The level of a vertex v of a tree is the length of the directed path fromthe root
to v; it is equal to the length of the word associated with v.
Our next goal is to describe a construction of an optimumcode in a sense that
we discuss shortly. It is described here as a communication problem, as it was
viewed by Huffman [11], who solved it. In the next section, we shall describe
one more application of this optimization technique.
Assume that words over a source alphabet of n letters have to be transmitted
over a channel that can transfer one letter of the alphabet = {0, 1, . . . , 1} at
a time, and <n. We want to construct a code over with n code-words and
associate a code-word with each source letter. A word over the source alphabet
is translated into a message over the code, by concatenating the code-words that
correspond to the source letters in the same order as they appear in the source
word. This message can now be transmitted through the channel. Clearly, the
code must be UD.
4.2 Positional Trees and Huffmans Optimization Problem 71
Assume further that the source letters have given probabilities p
1
, p
2
, . . . p
n
of
appearance and that the choice of the next letter in the source word is indepen-
dent of its previous letters. If the vector of code-word lengths is (l
1
, l
2
, . . . , l
n
),
then the average code-word length, l, is given by

l =
n

i=1
p
i
l
i
. (4.5)
We want to nd a code for which

l is minimumin order to minimize the expected

length of the message.
Since the code must be UD, by Theorem4.1, the vector of code-word lengths
must satisfy the characteristic sum condition. This implies, by Theorem 4.2,
that a prex code with the same vector of code-word lengths exists. Therefore,
in seeking an optimumcode, for which

l is minimum, we may restrict our search

to prex codes. In fact, all we have to do is nd a vector of code-word lengths
for which

l is minimum, among the vectors that satisfy the characteristic sum
condition.
First, let us assume that p
1
p
2
p
n
. This is easily achieved by sorting
the probabilities. We rst demonstrate Huffmans construction for the binary
case ( =2). Assume the probabilities are 0.6, 0.2, 0.05, 0.05, 0.03, 0.03, 0.03,
0.01. We write this list as our top row (see Figure 4.2). We add the last (and
therefore least) two numbers, and insert the sum in the proper place to maintain
the nonincreasingorder. We repeat this operation until we get a vector with only
two probabilities. Now, we assign each probability a word-length 1, and start
working our way back up by assigning each of the probabilities of the previous
step its length in the present step, if it is not one of the last two, and each of the
two last probabilities of the previous step is assigned a length that is larger by
one than the length assigned to their sum in the present step.
Once the vector of code-word lengths is found, a prex code can be assigned
to it by the technique of the proof of Theorem4.2. (An efcient implementation
is discussed in Problem 4.6.) Alternatively, the back-up procedure can produce
a prex code directly. Instead of assigning lengths to the last two probabilities,
we assign the two words of length one: 0 and 1. As we back up from a present
step in which a word is already assigned to each probability, to the previous
step, the rule is as follows: All but the last two probabilities of the previous
step are assigned the same words as in the present step. The last two probabil-
ities are assigned c0 and c1, where c is the word assigned to their sum in the
present step.
72 4 Ordered Trees
0.6
1 2 4 4 5 5 5 5
1 2 4 4 4 5 5
1 2 4 4 4 4
1 2 3 4 4
1 2 3 3
1 2
1 1
2
0.2 0.05 0.05 0.03 0.03 0.03 0.01
0.6 0.2 0.05 0.05 0.04 0.03 0.03
0.6 0.2 0.06 0.05 0.05 0.04
0.6 0.2 0.09 0.06 0.05
0.6 0.2 0.11 0.09
0.6 0.2 0.2
0.6 0.4
Figure 4.2: A demonstration of Huffmans construction for the binary case.
In the general case, when 2, we add in each step the last d probabilities
of the present vector of probabilities; if n is the number of probabilities of this
vector then d is given by
1 <d and n d mod (1) (4.6)
After the rst step, the length of the vector, n

, satises n

1 mod (1),
and will be equal to one, mod ( 1), from there on. The reason for this
rule is that we should end up with exactly probabilities, each to be assigned
length 1. Now, 1 mod (1), and since in each ordinary step the number
of probabilities is reduced by 1, we want n 1 mod (1). In case this
condition is not satised by the given n, we correct it in the rst step as is
done by our rule. Our next goal is to prove that this indeed leads to an optimum
assignment of a vector of code-word lengths.
Lemma 4.2 If C= {c
1
, c
2
, . . . , c
n
} is an optimum prex code for the probabili-
ties p
1
, p
2
, . . . , p
n
, then p
i
>p
j
implies that l(c
i
) l(c
j
).
Proof: Assume l(c
i
) >l(c
j
). Make the following switch: Assign c
i
to proba-
bility p
j
, and c
j
to p
i
; all other assignments remain unchanged. Let

l denote the
average code-word length of the new assignment, while

l denotes the previous
4.2 Positional Trees and Huffmans Optimization Problem 73
one. By (4.5), we have

l l = [p
i
l(c
i
) +p
j
l(c
j
)] [p
i
l(c
j
) +p
j
l(c
i
)]
= (p
i
p
j
)(l(c
i
) l(c
j
)) >0
contradicting the assumption that

l is minimum.
Lemma 4.3 There exists an optimum prex code for the probabilities p
1

p
2
p
n
such that the positional tree that represents it has the following
properties:
(1) All the internal vertices of the tree, except possibly one internal vertex v,
have exactly sons.
(2) Vertex v has 1 < sons, where n mod (1).
(3) Vertex v of ((1)) is on the lowest level that contains internal vertices, and
its sons are assigned to p
n+1
, p
n+2
, . . . , p
n
.
Proof: Let T be a positional tree representing an optimumprex code. If there
exists an internal vertex u that is not on the lowest level of T containing internal
vertices and it has less than sons, then we can perform the following change
in T: Remove one of the leaves of T from its lowest level and assign to the
probability a new son of u. The resulting tree, and therefore its corresponding
prex code, has a smaller average code-word length. A contradiction. Thus, we
conclude that no such internal vertex u exists.
If there are internal vertices on the lowest level of internal vertices that have
less than sons, choose one of them, say v. Now eliminate sons from v and
attach their probabilities to new sons of the others, so that their number of sons
is . Clearly, such a change does not change the average length, and the tree
remains optimum. If, before lling in all the missing sons, v has no more sons,
we can use v as a leaf and assign to it one of the probabilities from the lowest
level, thus creating a new tree that is better than T. A contradiction. Thus, we
never run out of sons of v to be transferred to other lacking internal vertices
on the same level. Also, when this process ends, v is the only lacking internal
vertex (proving (1)), and its number of remaining sons must be greater than
one, or its son can be removed and its probability attached to v. This proves
that the number of sons of v, , satises 1 < .
If vs sons are removed, the new tree has n

=n+1 leaves and is full

(i.e., every internal vertex has exactly sons). In such a tree, the number of
leaves, n

, satises n

1, mod (1). This is easily proved by induction on

74 4 Ordered Trees
the number of internal vertices. Thus, n+1 1 mod (1), and therefore
n mod (1), proving (2).
We have already shown that v is on the lowest level of T that contains internal
vertices, and the number of its sons is . By Lemma 4.2, we know that the least
probabilities are assigned to leaves of the lowest level of T. If they are not
sons of v, we can exchange sons of v with sons of other internal vertices on
this level, to bring all the least probabilities to v without changing the average
length.
For a given alphabet size and probabilities p
1
p
2
p
n
, let

(p
1
, p
2
, . . . , p
n
) be the set of all -ary positional trees with nleaves, assigned
with the probabilities p
1
, p
2
, . . . , p
n
in such a way that p
nd+1
, p
nd+2
, . . . , p
n
(see Equation 4.6) are assigned, in this order, to the rst d sons of a vertex v,
which has no other sons. By Lemma 4.3,

(p
1
, p
2
, . . . , p
n
) contains at least
one optimum tree. Thus, we may restrict our search for an optimum tree to

(p
1
, p
2
, . . . , p
n
).
Lemma 4.4 There is a one to one correspondence between

(p
1
, p
2
, . . . , p
n
)
and the set of -ary positional trees, with n d + 1 leaves assigned with
p
1
, p
2
, . . . , p
nd
, p

, where p

=
n

i=nd+1
p
i
. The average word-length

l of the
prex code, represented by a tree T of

(p
1
, p
2
, . . . , p
n
), and the average code-
word-length

l

of the prex code represented by the tree T

, which corresponds
to T, satisfy

l =

l

+p

. (4.7)
Proof: The tree T

which corresponds to T is achieved as follows: Let v be the

father of the leaves assigned p
nd+1
, p
nd+2
, . . . , p
n
. Remove all the sons of
v and assign p

to it.
It is easy to see that two different trees T
1
and T
2
in

(p
1
, p
2
, . . . , p
n
) will
yield two different trees T

1
and T

2
, and that every -ary tree T

with nd+1
leaves assigned p
1
, p
2
, . . . , p
nd
, p

, is the image of some T; establishing the

correspondence.
Let l
i
denote the level of the leaf assignedp
i
inT. Clearly l
nd+1
=l
nd+2
=
=l
n
. Thus,

l =
nd

i=1
p
i
l
i
+l
n

i=nd+1
p
i
=
nd

i=1
p
i
l
i
+l
n
p

=
nd

i=1
p
i
l
i
+(l
n
1) p

+p

=

l

+p

.
4.3 Application of the Huffman Tree to Sort-by-Merge Techniques 75
Lemma 4.4 suggests a recursive approach to nd an optimum T. For

l to be
minimum,

l

must be minimum. Thus, let us rst nd an optimumT

, and then
nd T by attaching d sons to the vertex of T

assigned p

; these d sons are

assigned p
nd+1
, p
nd+2
, . . . , p
n
. This is exactly what is done in Huffmans
procedure, thus proving its validity.
It is easy to implement Huffmans algorithmin time complexity O(n
2
). First,
we sort the probabilities, and after each addition, the resulting probability is
inserted in a proper place. Each such insertion takes at most O(n) steps, and
the number of insertions is ,(n)/(1)|. Thus, the whole forward process
is of time complexity O(n
2
). The back up process is O(n) if pointers are left
in the forward process to indicate the probabilities of which it is composed.
However, the time complexity can be reducedtoO(nlogn). One wayof doing
it is the following: First sort the probabilities. This can be done in O(nlogn)
steps [14]. The sorted probabilities are put on a queue S
1
in a non-increasing
order from left to right. A second queue, S
2
, initially empty, is used too. In the
general step, we repeatedly take the least probability of the two (or one, if one
of the queues is empty) appearing at the right hand side ends of the two queues,
and add up d of them. The result, p

, is inserted at the left hand side end of

S
2
. The process ends when after adding d probabilities both queues are empty.
This adding process and the back up are O(n). Thus, the whole algorithm is
O(nlogn).
The construction of an optimumprex code, when the cost of the letters are not
equal is discussed in Reference [12]; the case of alphabetic prex codes, where
the words must maintain lexicographically the order of the given probabilities,
is discussed in Reference [13]. These references give additional references to
previous work.
4.3 Application of the Huffman Tree to Sort-by-Merge Techniques
Assume that we have n items, and there is an order dened between them. For
ease of presentation, let us assume that the items are the integers 1, 2, . . . , n,
and the order is less than. Assume that we want to organize the numbers in
nondecreasing order, where initially they are put in L lists, A
1
, A
2
, . . . , A
L
. Each
A
i
is assumed to be ordered already. Our method of building larger lists from
smaller ones is as follows: Let B
1
, B
2
, . . . , B
m
be any m existing lists. We read
the rst, and therefore least, number in each of the lists, take the least number
among them away from its list, and put it as the rst number of the merged list.
The list from which we took the rst number is now shorter by one. We repeat
this operation on the same m lists until they merge into one. Clearly, some of
76 4 Ordered Trees
1
5
4
3
2
5
2
4
3
1
3
4
1
Figure 4.3: An example of a sort-by-merge described by a positional tree.
the lists become empty before others, but since this depends on the structure of
the lists, we only know that the general step of nding the least number among
mnumbers (or less) and its transfer to a new list is repeated b
1
+b
2
+ +b
m
times, where b
i
is the number of numbers in B
i
.
The number m is dictated by our equipment or decided upon in some other
way. However, we shall assume that its value is xed and predetermined. In
fact, in most cases m=2.
The whole procedure can be described by a positional tree. Consider the
example shown in Figure 4.3, where m= 2. First, we merge the list 3 with
1, 4. Next, we merge 2, 5 with 1, 3, 4. The original lists, A
1
, A
2
, . . . , A
L
,
correspond to the leaves of the tree. The number of transfers can be computed
as follows: Let a
i
be the number of numbers in A
i
, and l
i
be the level of the
list A
i
in the tree. The number of elementary merge operations is then
L

i=1
a
i
l
i
. (4.8)
Burge [15] observed that the attempt to nd a positional m-ary that mini-
mizes (4.8) is similar to that of the minimum average word-length problem
solved by Huffman. The fact that the Huffman construction is in terms of prob-
abilities does not matter, since the fact that p
1
+p
2
+ +p
L
= 1 is never
used in the construction or its validity proof. Let us demonstrate the implied
procedure by the following example:
4.4 Catalan Numbers 77
9 8 8 7 6 5 5 4 6 6 3 3
1 2 2 2 2 2 2 2 2 2 2 2
10 9 8 8 7 6 6 6 5 5
1 1 2 2 2 2 2 2 2 2
22 10 9 8 8 7 6
1 1 1 2 2 2 2
29 22 10 9
1 1 1 1
Figure 4.4: An example of sort-by-merge for L = 12 and m= 4.
29 22 10
70
9
8 8 7 6 6 6 5 5 4 3 3
Figure 4.5: The positional tree describing the example in Figure 4.4.
Assume L = 12 and m = 4; the b
i
s are given in nonincreasing order:
9, 8, 8, 7, 6, 6, 6, 5, 5, 4, 3, 3. Since L 0(mod 3), according to (4.6), d =3.
Thus, in the rst step we merge the last three lists to form a list of length 10,
which is now put in the rst place (see Figure 4.4). From there on, we merge
each time the four lists of least length. The whole merge procedure is described
in the tree shown in Figure 4.5.
4.4 Catalan Numbers
The set of well-formed sequences of parentheses is dened by the following
recursive denition:
(1) The empty sequence is well formed.
(2) If A and B are well-formed sequences, so is AB (the concatenation of A
and B).
(3) If A is well formed, so is (A).
78 4 Ordered Trees
(4) There are no other well-formed sequences.
For example, (()(())) is well formed; (()))(() is not.
Lemma 4.5 A sequence of (left and right) parentheses is well formed if and
only if it contains an even number of parentheses, half of which are left, and the
other half, right. Also as we read the sequence from left to right, the number of
right parentheses never exceeds the number of left parentheses.
Proof: First let us prove the only if part. Since the construction of every well-
formed sequence starts with no parentheses (the empty sequence), and each time
we add on parentheses (Step 3) there is one left and one right, it is clear that
there are nleft parentheses and nright parentheses. Now, assume that for every
well-formed sequence of m left and m right parentheses where m < n, it is
true that as we read it from left to right the number of right parentheses never
exceeds the number of left parentheses. If the last step in the construction of
our sequence was (2), then since Ais a well-formed sequence, as we read from
left to right, as long as we still read A, the condition is satised. When we are
between Aand B, the count of left and right parentheses equalizes. From there
on, the balance of left and right is safe, since B is well formed and contains less
than n parentheses. If the last step in the construction of our sequence was (3),
since A satises the condition, so does (A).
Now, we shall prove the if part, again by induction on the number of paren-
theses. (Here, as before, the basis of the induction is trivial.) Assume that the
statement holds for all sequences of mleft and mright parentheses, if m<n,
and we are given a sequence of n left and n right parentheses that satises
the condition. Clearly, if after reading 2m symbols of it from left to right, the
number of left and right parentheses is equal, and if m< n, then this subse-
quence, A, by the inductive hypothesis, is well formed. Now, the remainder
of our sequence, B, must satisfy the condition, too, and again by the inductive
hypothesis is well formed. Thus, by Step (2), AB is well formed. If there is no
such nonempty subsequence A, which leaves a nonempty B, then as we read
from left to right, the number of right parentheses, after reading one symbol
and before reading the whole sequence, is strictly less then the number of left
parentheses. Thus, if we delete the rst symbol, which is a (, and the last,
which is a ), the remainder sequence, A, still satises the condition, and
by the inductive hypothesis, is well formed. By Step (3) our sequence is well
formed too.
4.4 Catalan Numbers 79
We shall now show a one-to-one correspondence between the non-well-
formed sequences of n left and n right parentheses, and all sequences of n1
left parentheses and n+1 right parentheses.
Let p
1
p
2
p
2n
be a sequence of n left and n right parentheses that is not
well formed. By Lemma 4.5, there is a prex of it that contains more right
parentheses than left parenthesis. Let j be the least integer such that the number
of right parentheses exceeds the number of left parentheses in the subsequence
p
1
p
2
p
j
. Clearly, the number of right parentheses is then one larger than the
number of left parentheses, or j is not the least index to satisfy the condition.
Now, invert all p
i
s for i >j fromleft parentheses to right parentheses, and from
right parentheses to left parentheses. Clearly, the number of left parentheses is
now n1, and the number of right parentheses is now n+1.
Conversely, given any sequence p
1
p
2
p
2n
of n 1 left parentheses and
n+1 right parentheses, let j be the rst index such that p
1
p
2
p
j
contains one
right parenthesis more than left parentheses. If we nowinvert all the parentheses
in the section p
j+1
p
j+2
p
2n
from left to right and from right to left, we get
a sequence of n left and n right parentheses which is not well formed. This
transformation is the inverse of the one in the previous paragraph. Thus, the
one-to-one correspondence is established.
The number of sequences of n1 left and n+1 right parentheses is
_
2n
n1
_
,
for we can choose the places for the left parentheses, and the remaining places
will have right parentheses. Thus, the number of well-formed sequences of
length n is
_
2n
n
_

_
2n
n1
_
=
1
1 +n
_
2n
n
_
. (4.9)
These numbers are called Catalan numbers.
An ordered tree is a directed tree such that for each internal vertex there is
a dened order of its sons. Clearly, every positional tree is ordered, but the
converse does not hold: In the case of ordered trees there are no predetermined
potential sons; only the order of the sons counts, not their position, and there
is no limit on the number of sons.
An ordered forest is a sequence of ordered trees. We usually draw a forest
with all the roots on one horizontal line. The sons of a vertex are drawn from
left to right in their given order. For example, the forest shown in Figure 4.6
consists of three ordered trees whose roots are A, B, and C.
80 4 Ordered Trees
B A C ((( )( ))) (( )( )( )) ((( )( ))( ))
(( )( )) (( )( ))
( ) ( ) ( ) ( )
( ) ( ) ( ) ( )
Figure 4.6: An example of three ordered trees.
There is a natural correspondence between well-formed sequences of n pairs
of parentheses and ordered forests of n vertices. Let us label each leaf with
the sequence (). Every vertex whose sons are labeled w
1
, w
2
, . . . , w
s
is labeled
with the concatenation (w
1
w
2
w
s
); clearly, the order of the labels is in the
order of the sons. Finally, once the roots are labeled x
1
, x
2
, . . . , x
r
, the sequence
corresponding to the forest is the concatenation x
1
x
2
x
n
. For example, the
sequence correspondingto the forest of Figure 4.6 is ((()())())(()()())((()())). The
inverse transformation clearly exists, and thus the one-to-one correspondence
is established. Therefore, the number of ordered forests of n vertices is given
by (4.9).
We now describe a one-to-one correspondence between ordered forests and
positional binary trees. The leftmost root of the forest is the root of the binary
tree. The leftmost son of the vertex in the forest is the left son of the vertex in
the binary tree. The next brother on the right, or, in the case of a root, the next
root on the right is the right son in the binary tree. For example, see Figure 4.7,
where an ordered forest and its corresponding binary tree are drawn. Again, it
is clear that this is a one-to-one correspondence, and therefore the number of
positional binary trees with n vertices is given by (4.9).
There is yet another combinatorial enumeration that is directly related to
these.
A stack is a storage device that can be described as follows. Suppose that n
cars travel on a narrow one-way street where no passing is possible. This leads
into a narrow two-way street on which the cars can park or back up to enter
another narrowone-way street (see Figure 4.8). Our problemis to nd howmay
4.4 Catalan Numbers 81
A
D E
J K
J
K
E
D
A
F
B
G
H L
M
I
C
F G H
B C
I
L M
Figure 4.7: An ordered forest and its corresponding binary tree.
Figure 4.8: An example of a stack of cars.
permutations of the cars can be realized from input to output if we assume that
the cars enter in the natural order.
The order of operations in the stack is fully described by the sequence of
drive-in and drive-out operations. There is no need to specify which car drives
in, for it must be the rst one on the leading-in present queue; also, the only one
that can drive out is the top one in the stack. If we denote a drive-in operation
by (, and a drive-out operation by ), the whole procedure is described by a
well-formed sequence of n pairs of parentheses.
The sequence must be well formed, by Lemma 4.5, since the number of drive-
out operations can never exceed the number of drive-in operations. Also, every
well-formed sequence of n pairs of parentheses denes a realizable sequence
of operations, since, again by Lemma 4.5, a drive-out is never instructed when
82 4 Ordered Trees
the stack is empty. Also, different sequences yield different permutations. Thus,
the number of permutations on n cars realizable by a stack is given by (4.9).
Let us now consider the problem of nding the number of full binary trees.
Denote the number of leaves of a binary tree T by L(T), and the number of
internal vertices by I(T). It is easy to prove, by induction on the number of
leaves, that L(T) =I(T) +1. Also, if all leaves of T are removed, the resulting
tree of I(T) vertices is a positional binary tree T

. Clearly, different T-s will

yield different T

-s, and one can reconstruct T fromT

by attaching two leaves

to each leaf of T

, and one leaf (son) to each vertex that in T

has only one son.

Thus, the number of full binary trees of n vertices is equal to the number of
positional binary trees of (n1)/2 vertices. By (4.9) this number is
2
n+1
_
n1
n1
2
_
.
4.5 Problems
Problem 4.1 Prove the following theorem: A position between two letters in
a message m over a UD code C is a separation between two code-words if
and only if both the prex and the sufx of mup to this position are messages
over C.
Problem 4.2 Use the result of Problem 4.1 to construct an efcient algorithm
for parsing a message m over a UD code C in order to nd the words which
compose m. (Hint: scan m from left to right, and mark all the positions such
that the prex m up to them is a message. Repeat from right to left to mark
positions corresponding to sufxes that are messages.)
Problem 4.3 Test the following codes for UD:
(1) {00, 10, 11, 100, 110},
(2) {1, 10, 001, 0010, 00000, 100001},
(3) {1, 00, 101, 010}.
Problem 4.4 Construct a prex binary code, with minimum average
code-word length, which consists of ten words whose probabilities are
0.2, 0.18, 0.12, 0.1, 0.1, 0.08, 0.06, 0.06, 0.06, 0.04. Repeat the construction for
= 3 and 4.
Problem 4.5 Prove that, if a prex code corresponds to a full positional tree,
then its characteristic sum is equal to 1.
4.5 Problems 83
Problem 4.6 Prove that, if the word-length vector (l
1
, l
2
, . . . , l
n
) satises the
characteristic sum condition, and if l
1
l
2
l
n
, then there exists a
positional tree with n leaves whose levels are the given l
i
s, and the order of
the leaves from left to right is as in the vector.
Problem 4.7 A code is called exhaustive if every word over the alphabet is the
beginning of some message over the code. Prove the following:
(1) If a code is prex, and its characteristic sumis 1, then the code is exhaustive.
(2) If a code is UD and exhaustive, then it is prex and its characteristic sum
is 1.
Problem 4.8 Construct the ordered forest, the positional binary tree and the
permutation through a stack that corresponds to the following well-formed
sequence of ten pairs of parentheses:
(()(()()))((()()())).
Problem 4.9 A direct method for computing the number of positional binary
trees of n vertices through the use of a generating function goes as follows: Let
b
n
be the number of trees of n vertices. Dene b
0
= 1 and dene the function
B(x) =b
0
+b
1
x +b
2
x
2
+ .
(1) Prove that b
n
=b
0
b
n1
+ +b
n1
b
0
.
(2) Prove that xB
2
(x) B(x) +1 = 0.
(3) Use the formula
(1 +a)
1
2
=1 +
1
2
1!
a+
1
2
(
1
2
1)
2!
a
2
+
1
2
(
1
2
1)(
1
2
2)
3!
a
3
+
to prove that
b
n
=
1
n+1
_
2n
n
_
.
Bibliography
[1] Sardinas, A. A., and Patterson, G. W., A Necessary and Sufcient Condition for
the Unique Decomposition of Coded Messages, IRE Convention Record, Part 8,
1953, pp. 104108.
[2] Gallager, R. G., Information Theory and Reliable Communication, John Wiley,
1968. Problem 3.4, page 512.
84 4 Ordered Trees
[3] Levenshtein, V. I., Certain Properties of Code Systems, Dokl. Akad. Nauk,
SSSR, Vol. 140, No. 6, Oct. 1961, pp. 12741277. English translation: Soviet
Physics, Doklady, Vol. 6, April 1962, pp. 858860.
[4] Even, S., Test for Unique Decipherability, IEEE Trans. on Infor. Th., Vol. IT-9,
No. 2, April 1963, pp. 109112.
[5] Levenshtein, V. I., Self-Adaptive Automata for Coding Messages, Dokl. Akad.
Nauk, SSSR, Vol 140, Dec. 1961, pp. 13201323. English translation: Soviet
Physics, Doklady, Vol. 6, June 1962, pp. 10421045.
[6] Markov, Al. A., On Alphabet Coding, Dokl. Akad. Nauk, SSSR, Vol. 139, July
1961, pp. 560561. English translation: Soviet Pysics, Doklady, Vol. 6, Jan.
1962, pp. 553554.
[7] Even, S., Test for Synchronizability of Finite Automata and Variable Length
Codes, IEEE Trans. on Infor. Th., Vol. IT-10, No. 3, July 1964, pp. 185189.
[8] McMillan, B., Two Inequalities Implied by Unique Decipherability, IRE Tran.
on Infor. Th., Vol. IT-2, 1956, pp. 115116.
[9] Karush, J., A Simple Proof of an Inequality of McMillan, IRE Tran. on Infor.
Th., Vol. IT-7, 1961, pp. 118.
[10] Kraft, L. G., A Device for Quantizing, Grouping and Coding Amplitude
Modulated Pulses, M.S. Thesis, Dept. of E.E., M.I.T.
[11] Huffman, D. A., AMethodfor the Constructionof MinimumRedundancyCodes,
Proc. IRE, Vol. 40, No. 10, 1952, pp. 10981101.
[12] Perl, Y., Garey, M. R., and Even, S., Efcient Generation of Optimal Prex Code:
Equiprobable Words Using Unequal Cost Letters, J.ACM, Vol. 22, No. 2, April
1975, pp. 202214.
[13] Itai, A., Optimal Alphabetic Trees, SIAMJ. Comput., Vol. 5, No. 1, March 1976,
pp. 918.
[14] Knuth, D. E., The Art of Computer Programming, Vol. 3: Sorting and Searching,
Addison-Wesley, 1973.
[15] Burge, W. H., Sorting, Trees, and Measures of Order, Infor. and Control, Vol. 1,
1958, pp. 181197.
5
Flow in Networks
5.1 Introduction
A network N(G, s, t, c) consists of the following data:

A nite digraph G(V, E) with no self-loops and no parallel edges.

1

Two vertices s and t are specied; s is called the source; and t, the sink.
2

The capacity function, c : E R

+
. The positive real number, c(e), is called
the capacity of edge e.
For every vertex v V, let (v) denote the set of edges that enter v in G.
Similarly, let (v) denote the set of edges that emanate fromv.
A ow function, f : E R, is an assignment of a real number f(e) to each
edge e such that the following two conditions hold:
The Edge Rule. For every edge e E, 0 f(e) c(e).
The Vertex Rule. For every vertex v V \ {s, t},

e(v)
f(e) =

e(v)
f(e).
The total ow F, of f in N, is dened by
F =

e(t)
f(e)

e(t)
f(e) . (5.1)
Namely, F is the net sum of ow into the sink.
1
Self-loops are useless in this context, and parallel edges can be replaced with one edge whose
capacity is the sum of the capacities of the parallel edges.
2
The source is not necessarily a graphical source; i.e., it may have incoming edges. Similarly,
the sink is not necessarily a graphical sink.
85
86 5 Flow in Networks
In the next two sections, we shall discuss methods for computing a ow
function f for which F is maximum.
Given a set S V, let

S =V \S. In the following, we shall discuss sets S, such
that s S and t

S. Also, (S;

S) denotes the set of edges which are directed

from a vertex in S to a vertex in

S; this set of edges is called a forward cut. The
set (

S;S) is similarly dened, and is called a backward cut. The union of (S;

S)
and (

S;S) is called the cut dened by S.

By denition, the total owF is measured at the sink. Our purpose is to show
that F can be measured at any cut.
Lemma 5.1 For every {s} S (V \ {t}) and every ow function f of total
ow F, the following holds:
F =

e(S;

S)
f(e)

e(

S;S)
f(e) . (5.2)
Proof: By the vertex rule, for every v (V \ {s, t}),
0 =

e(v)
f(e)

e(v)
f(e) . (5.3)
Also, consider again Equation 5.1:
F =

e(t)
f(e)

e(t)
f(e) .
Now, add the equations, as in Equation 5.3, for every v (

S \ {t}), as well as
Equation 5.1. The aggregate equation has F on the l.h.s.
3
In order to see what happens on the r.h.s., consider an edge x
e
y. If both
x and y belong to S then f(e) does not appear on the r.h.s. of the aggregate
equation at all, in agreement with Equation 5.2. If both x and y belong to

S
then f(e) appears twice on the r.h.s. of the aggregate equation; once positively,
in the equation for y, and once negatively, in the equation for x. Thus, it is
canceled out in the summation, again in agreement with Equation 5.2. If x S
and y

S then f(e) appears on the r.h.s. of the aggregate equation, as part of
Equation 5.3 for y, positively, and in no equation for other vertices included
in the summation. In this case e (S;

S), and again we have agreement with

Equation 5.2. Finally, if x

S and y S, f(e) appears negatively on the r.h.s.
3
Left-hand side.
5.2 The Algorithm of Ford and Fulkerson 87
of the aggregate equation, as part of Equation 5.3 for x, and again this agrees
with Equation 5.2, since e (

S;S).
Let us denote by c(S) the capacity of the cut determinedby S, which is dened
as follows:
c(S) =

e(S;

S)
c(e) . (5.4)
Lemma 5.2 For every ow function f, with total ow F, and every {s} S
(V \ {t}), the following inequality holds:
F c(S) . (5.5)
Proof: By Lemma 5.1,
F =

e(S;

S)
f(e)

e(

S;S)
f(e) .
By the edge rule, for every edge e, 0 f(e) c(e). Thus,
F

e(S;

S)
c(e) 0 .
Finally, by Equation 5.4, F c(S).
The following is a very important corollary of Lemma 5.2. It allows us to
detect that a given total ow F is maximum, and the capacity of a given cut,
dened by S, is minimum.
Corollary 5.1 If F and S satisfy Equation 5.5 by equality, then F is maximum
and the cut dened by S is of minimum capacity.
5.2 The Algorithm of Ford and Fulkerson
Ford and Fulkerson [7] suggested the use of augmentingpaths to change a given
ow function f in order to increase the total ow. A procedure for nding an
augmenting path, if one exists, is used. If an augmenting path is found, then it is
used to increase the total ow. If no augmenting path is found, then the present
ow is declared maximum and the process terminates.
If the direction of the edges is ignored, an augmenting path P is a simple path
froms to t. It is used to add a quantity >0 to the total ow by pushing along
88 5 Flow in Networks
P additional units of ow. To do this, if an edge e of P is directed in the
direction froms to t, c(e) f(e) + must hold. But if e of P is directed in the
opposite direction, we must be able to reduce its ow from f(e) to f(e) .
Thus, f(e) must hold.
In attempt to nd an augmenting path for a given ow, a labeling procedure is
used. We rst label s. Then, as long as there is an unlabeled vertex v for which
an augmenting path from s to v is found, v is labeled. If t is labeled, then an
augmenting path has been found and the labeling process is terminated.
Every vertex v is assigned a label (v), which is one of the following:

(s) =; only vertex s gets this label, and it is the only label s gets.

(v) =NIL; in this case, we say that v has no label.

(v) = (e, ), where e is the edge through which v has been assigned its label,
= + if e has been used forwardly and = if e has been used backwardly.
An edge u
e
v is said to be useful from u to v if (u) ,=NIL, (v) =NIL
and one of the following conditions holds:

u
e
v and f(e) <c(e). In this case, we say that e is forwardly useful.

v
e
u and f(e) >0. In this case, we say that e is backwardly useful.
Algorithm 5.1 describes the labeling procedure, named LABEL. The input
of LABEL consists of the network N and the present ow function f. () is
dened for all vertices. The procedure may modify these labels. Thus, () is
both an input and output. The output of LABEL includes a function (), which
is dened on a subset of E, and its value is a positive real number.
Procedure LABEL(N, f, , )
1 while there is an edge u
e
v which is useful from u to v and
(t) =NIL do
2 if e is forwardly useful then do
3 (v) (e, +)
4 (e) c(e) f(e)
5 if e is backwardly useful then do
6 (v) (e, )
7 (e) f(e)
Algorithm 5.1: The labeling procedure.
5.2 The Algorithm of Ford and Fulkerson 89
Procedure AUGMENT(G, , , f)
1 empty Q
2
3 v t (v is called the center of activity)
4 while v ,=s do
5 put (v) into Q
6 let (v) = (e, )
7 min{, (e)}
8 let e be u
e
v
9 v u
10 while Q is not empty do
11 remove the rst label, (e, ), from Q
12 if = + then do
13 f(e) f(e) +
14 else ( = ) do
15 f(e) f(e)
Algorithm 5.2: The augmenting procedure.
If (t) ,= NIL when LABEL is terminated, then an augmenting path exists.
The augmenting path is found in the AUGMENT procedure and used to change
f and thus, increment F. This is presented in Algorithm 5.2.
The input to AUGMENTconsists of G(V, E), () and (). The owfunction
f is both an input and an output. A queue Q of vertex labels is used, as well as
a real these variables are internal.
The Ford and Fulkerson procedure is described in Algorithm 5.3 and uses
LABEL and AUGMENT as subroutines. Initially, f is any legitimate ow in
the network N, and in the absence of better ideas, one can use f(e) = 0 for
every edge e. Thus, the input is N and f is both an input and an output.
Note that, if for an edge e (s), initially f(e) =0, then f(e) is never changed.
The same is true for an edge e (t).
As an example, consider the networks shown in Figure 5.1. Next to each edge
e we write c(e), f(e), in this order. We assume a zero initial ow everywhere.
A rst wave of label propagation might be as follows: s is labeled, e
2
is used to
label c, e
6
is used to label d, e
4
is used to label a, e
3
used to label b and nally,
90 5 Flow in Networks
Procedure FORD-FULKERSON(N, f)
1 for every v V do
2 (v) NIL
3 (s)
4 call LABEL(N, f, , )
5 while (t) ,=NIL do
6 call AUGMENT(G, , , f)
7 for every v V \ {s} do
8 (v) NIL
9 call LABEL(N, f, , )
10 print The present ow f is maximum
Algorithm 5.3: The Ford-Fulkerson procedure.
s
a b
c d
t
15,0
12,0
7,0
4,0
10,0
5,0
10,0
3,0
e
1
e
3
e
6
e
4
e
7
e
8
e
2
e
5
Figure 5.1: An example of a network.
e
7
is used to label t. The path is
s
e
2
c
e
6
d
e
4
a
e
3
b
e
7
t ,
= 4, and the new ow is shown in Figure 5.2. The second augmenting path
may be
s
e
1
a
e
3
b
e
5
c
e
6
d
e
8
t ,
= 3 and the new ow is shown in Figure 5.3. The third augmenting path
may be
s
e
1
a
e
3
b
e
7
t ,
=3 and the newowis shown in Figure 5.4. Up to nowonly forward labeling
5.2 The Algorithm of Ford and Fulkerson 91
s
a b
c d
t
15,0
12,4
7,4
4,4
10,4
5,4
10,0
3,0
e
1
e
3
e
6
e
4
e
5
e
7
e
8
e
2
Figure 5.2: The example after the rst augmenting path.
s
a b
c d
t
15,3
12,7
7,4
4,4
10,7
5,4
10,3
3,3
e
1
e
3
e
6
e
4
e
5
e
7
e
8
e
2
Figure 5.3: The example after the second augmenting path.
s
a b
c d
t
15,6
12,10
7,7
4,4
10,7
10,3
3,3
5,4
e
1
e
3
e
6
e
4
e
5
e
7
e
8
e
2
Figure 5.4: The example after the third augmenting path.
has been used. In the next application of LABEL, we can still label vertex a
via e
1
; and vertex b, via e
3
, both of which are forward labeling, but no further
forward labeling exists. However, e
4
is useful backwardly, and through it one
92 5 Flow in Networks
s
a b
c d
t
12,10
7,7
10,7
10,7
5,0
3,3
15,10
4,4
e
1
e
3
e
6
e
4
e
5
e
7
e
8
e
2
Figure 5.5: The example after the fourth augmenting path.
can label d. Now one can use e
8
to label t. The augmenting path is
s
e
1
a
e
4
d
e
8
t ,
= 4, and the new ow is shown in Figure 5.5. Now, the total ow, F, is equal
14. The next application of LABEL does not reach t. The set of labeled vertices
S is {s, a, b} and the forward cut, (S;

S) consists of the edges e

2
, e
5
, and e
7
.
All of these edges are saturated; that is, f(e) =c(e). The backward cut, (

S;S),
consists of one edge, e
4
, and its ow is 0. Thus, F =c(S), and by Corollary 5.1,
F is maximum and c(S) is minimum.
It is easy to see that the ow produced by the algorithm remains legitimate
throughout. The denitions of () and guaranty that forwardly used edges
will not be overowed, that is, f(e) c(e), and that backward edges will not
be underowed, that is, f(e) 0. Also, since units of ow are are pushed
froms to t on a path, the incoming ow will remain equal to the outgoing ow
in every vertex v V \ {s, t}.
Assuming the Ford and Fulkerson procedure halts, the last labeling process
has not reached t. As above, let S be the set of vertices labeled in the last
application of the labeling process. If e (S;

S), then f(e) =c(e), since e is not

forwardly useful. If e (

S, S), then f(e) = 0, since e is not backwardly useful.

By Lemma 5.1,
F =

e(S;

S)
f(e)

e(

S;S)
f(e) =

e(S;

S)
c(e)

e(

S;S)
0 =c(S) .
By Corollary 5.1, F is maximum and c(S) is minimum.
The question of whether the Ford and Fulkerson procedure will always halt
remains to be discussed. Note rst a very important property of the procedure:
5.2 The Algorithm of Ford and Fulkerson 93
If the initial ow is integral, for example, zero everywhere, and if all capacities
are integers, then the algorithm never introduces fractions. The algorithm adds
and subtracts, but it never divides. Also, if t is labeled, the resulting augmenting
path is used to increase the total owby at least one unit. Since there is an upper
bound on the total ow (any cut), the process must terminate.
Ford and Fulkerson showed that their procedure may fail if the capacities are
allowed to be irrational numbers. Their counterexample ([7, p. 21]) displays an
innite sequence of owaugmentations. The owconverges (in innitely many
steps) to a value which is one-fourth of the maximum total ow. We shall not
present their example here; it is fairly complex, and as the reader will shortly
discover, it is not as important any more.
One could have arguedthat, for all practical purposes, we may assume that the
algorithm is sure to halt. This follows from the fact that our computations are
usually through a xed radix (decimal, binary, etc.) number representation with
a bound on the number of digits used; in other words, all gures are multiples
of a xed quantumand the termination proof works here as it does for integers.
However, a simple example shows the weakness of this argument. Consider the
network shown in Figure 5.6.
Assume that M is a very large integer. If the algorithm starts with f(e) = 0
for every e, and alternately uses
s a b t
and
s b a t
a
b
s t
M M
M M
1
Figure 5.6: An example which demonstrating the possibility that the
Ford and Fulkerson procedure may be inefcient.
94 5 Flow in Networks
as augmenting paths, it will use 2Maugmentations before F =2Mis achieved.
This is exponential time in terms of the length of the input data, since it takes
only ,log
2
(M+1)| bits to represent M.
Edmonds and Karp [4] were rst to overcome this problem. They showed
that if Breadth-First Search (BFS) is used in the labeling algorithm and thus,
every augmentingpath is a shortest one, the algorithmterminates in O(|V| |E|
2
)
time, regardless of the capacities. (Here, of course, we assume that our com-
puter can handle, in one step, any real number.) In the next section we shall
present the more advanced work of Dinitz (see [2]); his algorithm has time
complexity O(|V|
2
|E|). A signicantly more efcient algorithm was pre-
sented by Goldberg and Tarjan [9], but this algorithm is not described in this
book.
The existence of the algorithmof Edmonds and Karp, or that of Dinitz, assures
that if one proceeds according to a proper strategy in the labeling procedure,
the algorithmis guaranteed to halt (in polynomial time). When it does, the total
ow is maximum, and the cut indicated is minimum, thus providing the max
ow min-cut theorem:
Theorem 5.1 Every network has a maximum total ow which is equal to the
capacity of a cut for which the capacity is minimum.
5.3 The Dinitz Algorithm
As in the Ford and Fulkerson algorithm, the Dinitz algorithm
4
starts with some
legitimate ow function f and improves it. When no improvement is possible
the algorithm halts, and the total ow, F, is maximum.
Given a network N(G(V, E), s, t, c) and a ow function f, let us dene the
secondary network N

(G

(V, E

), s, t, c

) as follows:

For every e Esuch that f(e) <c(e), e E

. Also, c

(e) =c(e)f(e). (Such

an edge is called forwardly useful and c

(e) is called its residual capacity.)

For every e E, a
e
b such that f(e) >0, there is an edge b
e

a in E

.
Also, c

(e

) =f(e). (Such an edge is called backwardly useful.)

Note that if 0 < f(e) < c(e), then e gives rise to two antiparallel edges in
G

, since it is useful both in the forward and the backward directions. Thus,
|E| |E

| 2|E|.
4
In [3], Yem Dinitz tells the story of his algorithm, and the differences between his version and
the one presented here (G.E.).
5.3 The Dinitz Algorithm 95
Procedure LAYER(N

, V
i
, vertex labels, T)
1 T
2 while there is an edge u v in N

, u V
i
and v is not labeled, do
3 T T {v}
4 label v
Algorithm 5.4: Finding the next layer in the BFS of N

.
Before we go into a detailed description of the Dinitz algorithm, here is an
abstract of its structure:
The Dinitz algorithm proceeds in phases. In each phase the current f is used
to produce the corresponding secondary network N

. A layered network, N

,
is then produced by a BFS on N

, starting from the source s. If the sink t is

not reached, the whole algorithm halts, and the current ow is maximum. If t
is reached, a maximal (or blocking) ow, f

, is found in N

. The ow f is then
augmented, by using f

, and a new phase is launched.

We nowpresent a detailed description of the Dinitz algorithmand later discuss
its validity and time complexity.
The construction of N

, fromN

, uses the BFS layers. The production of the

layers is described in the subroutine LAYERS, which in turn uses a subroutine
LAYER, described in Algorithm 5.4. The input to LAYER is the secondary
network N

, the previous layer (a set of vertices) V

i
, and the vertex labels. The
set of labeled verces can change when LAYERis run (since additional vertices
may get labeled), and is therefore also an output. The set of vertices that may
become the next layer is provided by the output T.
The LAYERSsubroutine is described in Algorithm5.5. The input to LAYERS
consist of N

. The output is l, and if t is reached, the layers are specied in

V
0
, V
1
, . . . V
l
. Obviously, l n1.
If t has been reached, the layered network N

(G

(V

, E

), s, t, c

) is
constructed as follows:

=
l
i=0
V
i
,

i
= {u
e

v | u V
i
, v V
i+1
, e

},

=
l1
i=0
E

i
,

for every e

(e

) =c

(e

).
96 5 Flow in Networks
Procedure LAYERS(N

;V
0
, V
1
, . . . V
n1
, l)
Local variables: T a set of vertices, vertex labels.
1 label every v V unlabeled
2 label s
3 V
0
{s}
4 i 0
5 repeat
6 call LAYER(N

, V
i
, vertex labels

, T)
7 i i +1
8 V
i
T
9 until t T or T =
10 if t T then do
11 l i
12 V
l
{t}
13 else (T =) declare: f is a maximum ow.
Algorithm 5.5: Finding the BFS Layers of N

.
s t
a b
c d
1,1
1,1
1,1
1,0
1,0
1,0
1,0
V
0
V
1
V
2
V
3
Figure 5.7: An example of a maximal ow in a layered network that
is not maximum.
Next, we construct a maximal, or blocking ow in N

. A maximal ow f

is a legitimate ow in N

such that no augmenting path of length l exists in

the presence of f

. A maximal ow may not be maximum, as demonstrated in

Figure 5.7. The total ow is 1. This ow is not maximum, since there is a ow
5.3 The Dinitz Algorithm 97
Procedure FIND-PATH(N

, blocking-labels, S)
1 empty S
2 push (NIL, s) into S
3 while S is not empty and t is not the right element in the top pair of S do
4 let (x, v) be the top pair on S
5 if there is an unblocked edge v
e

, then do
6 push (e

, v

) into S
7 else do
8 if x ,=NIL then label x blocked
9 pop (x, v) from S
Algorithm 5.6: Finding an augmenting path of length l in N

.
for which the total is 2. Yet it is maximal, since on every directed path of length
3 from s to t, there is at least one saturated edge.
It is easier to nd maximal ow than maximum ow, since the ow in edges
never has to decrease. In other words, no backward labeling of vertices is
necessary.
The procedure MAXIMALto nd a maximal owin a given layered network
N

starts with f

(e

) = 0 for every e

. It uses two subroutines: FIND-

PATH and INCREASE-FLOW. Initially, all edges are marked unblocked.
When an edge becomes saturated, its label changes to blocked. Also, if we
conclude that the ow in u
e

v cannot increase since all paths fromv to t are

blocked, we change the label of e

to blocked.
FIND-PATH is described in Algorithm 5.6. Its input is N

. The blocking-
labels are both an input and an output, and the stackSis anoutput. The procedure
uses a DFS strategy. As long as there is an unblocked edge to continue on, we
keep going, and store the sequence of the edges and their end-points in a stack
S. Thus, S is a stack of pairs, such as (e, v), where v is the vertex that e enters.
If t is reached, the sequence of edges stored in S is a directed augmenting path
from s to t of length l. If there are no unblocked edges out of the vertex we
have reached, we pop the top edge (and its endpoint) from S, change its label
to blocked and backtrack to the previous vertex on the path. If we are stuck
at s, the search terminates with no edges in S.
The procedure INCREASE-FLOW is described in Algorithm 5.7. The pro-
cedure uses two local variables: S

is a stack of vertices, and is a real number.

98 5 Flow in Networks
Procedure INCREASE-FLOW(N

, S, f

, blocking-labels)
1 empty S

2
3 while the pair on top of S is not (NIL, s) do
4 pop the top pair (e, v) from S
5 min{, c

(e) f

(e)}
6 push e into S

7 while S

is not empty do
8 pop the top edge e from S

9 f

(e) f

(e) +
10 if f

(e) =c

(e) then mark e blocked

Algorithm 5.7: Increasing the ow in the edges of the augmenting
path in N

.
N

and S are the input, while f

and the blocking-labels are both an input and

an output. The augmenting path stored in S is scanned twice. During the rst
scan (Lines 36), the minimum residual capacity of the edges on the augment-
ing path is computed and recorded in , and the augmenting path is restored
S

. In the second scan (Lines 710), the ow in every edge of the augmenting
path is increased by , and if an edge becomes saturated, its label is changed
from unblocked to blocked. Clearly, at least one edge on the path becomes
saturated.
The procedure MAXIMAL described in Algorithm 5.8 nds a maximal ow,
f

, in a given layered network N

. The stack S and the blocking-labels are

internal variables. Finally, we are ready to present the Dinitz algorithm, as
described in Algorithm 5.9. Note that the input to DINITZ is the network N,
and the output is a maximum ow f, in N. All other variables used in the
procedure, such as N

, N

and all their components, including l, and f

, are
internal variables.
Lemma 5.3 If procedure DINITZ halts, the resulting f is a legitimate ow
in N.
Proof: First, observe that in each phase the ow f

is legitimate in N

. The
edge rule is observed since is chosen in INCREASE-FLOW in a way that
5.3 The Dinitz Algorithm 99
Procedure MAXIMAL(N

, f

)
1 for every e E

do
2 label e unblocked
3 f

(e) 0
4 run FIND-PATH(N

, blocking-labels, S)
5 while S is not empty then do
6 run INCREASE-FLOW(N

, S, f

, blocking-labels)
7 run FIND-PATH(N

, blocking-labels, S)
Algorithm 5.8: Constructing a maximal ow in a layered network N

.
Procedure DINITZ(N;f)
1 for every e E do
2 f(e) = 0
3 N

N
4 run LAYERS(N

;V
0
, V
1
, . . . , V
n1
, l)
5 while f has not been declared to be maximum do
6 construct N

7 run MAXIMAL(N

;f

)
8 for every e

do
9 if e

corresponds to a forward edge e E then do

10 f(e) f(e) +f

(e

)
11 else (u
e

v N

, but v
e
u N) do
12 f(e) f(e) f

(e

)
13 construct N

from (N, f)
14 run LAYERS(N

;V
0
, V
1
, . . . , V
n1
, l)
Algorithm 5.9: The Dinitz algorithm.
ensures that no edge is overowed. The vertex rule is observed, since the ow
is built up by augmenting paths.
By the denition of c

and thus, c

, when f is changed by adding (subtracting)

f

to (from) the previous f, the edge rule is maintained in the edges of G. Also,
100 5 Flow in Networks
since f is the superposition of two ows, each of which observes the vertex rule,
so does their sum.
Lemma 5.4 If procedure DINITZ halts, the resulting f is a maximum ow
in N.
Proof: The proof here is very similar to the one in the Ford and Fulkerson
algorithm. Consider the last phase, in which vertex t is not reached in procedure
LAYERS (see Line 13). Let S be the union of V
0
, V
1
, . . . , V
i
, where V
i
is the
last (nonempty) layer before T = is encountered. Nowconsider the cuts (S;

S)
and (

S;S) in G. Every edge u

e
v in (S;

S) is saturated, or else e is useful from

uV
i
to v, and T is not empty. Also, for every edge u
e
v in (

S;S), f(e) =0,

or e is useful fromv V
i
to u, and again, T is not empty. The remainder of the
argument is identical to that of the Ford-Fulkerson case.
Lemma 5.5 In each phase, except the last, the ow f

is maximal in N

.
Proof: Initially, all edges of N

are unblocked (see Algorithm 5.8). An edge

u
e
v is marked blocked in two cases:

e is saturated. See Line 10 of Algorithm 5.7.

There is no unblocked edge out of v. See Line 8 of Table 5.6.

In either case, when an edge is blocked, it is already known that no additional
augmenting path through it is possible. Thus, whenall edges out of s are blocked,
the ow f

is maximal. This event is detected by procedure FIND-PATH (see

Algorithm 5.6) by the emptiness of S.
Lemma 5.6 The running time of each phase is O(|V| |E|).
Proof: The time it takes to construct N

is O(|E|), since it is essentially a BFS

on N

.
Note that in Algorithm5.6 we traverse edges (see Lines 56). If we backtrack
to a vertex other than s (Lines 79), an edge is marked blocked.
The number of consecutive edge traversals between two blockings of edges
is bounded by l for the following reasons:

If l traversals are performed consecutively without any backtracking, then

t has been reached and an augmenting path is stored in S. Procedure
INCREASE-FLOW is called and at least one edge becomes saturated and
is marked blocked.
5.3 The Dinitz Algorithm 101

It is possible that we have backtracked into a vertex other than s. In that case,
in less than l consecutive traversals, either another backtrack occurs or t is
reached, and as in the previous case, at least one edge is marked blocked.
Since the number of edges in N

is O(|E|), and since every edge can be

blocked at most once, the total number of edge traversals is O(|V| |E|).
It is not hard to see that the time of all other operations in proce-
dure MAXIMAL is bounded by a constant times the number of edge
traversals.
Note that the length of a layered network N

was denoted by l. Since we

need to compare lengths of the layered networks of consecutive phases, let us
denote by l
k
the length of the layered network of the k-th phase, k 1.
Lemma 5.7 If the (k+1)-st phase is not the last, then l
k+1
>l
k
.
Proof: There is a path P of length l
k+1
in the layered network of the (k+1)-st
phase, which starts with s and ends with t:
P : s =v
0
e
1
v
1
e
2
v
2
v
l
k+1
1
e
l
k+1
v
l
k+1
=t .
First, let us assume that all the vertices of Pappear in the k-th layered network.
Let V
j
be the j-th layer of the k-th layered network. We claim that if v
a
V
b
,
then a b. This is proved by induction on a. For a = 0 (v
0
=s), the claim is
obviously true. Now, assume v
a+1
V
c
.
By the inductive hypothesis, (ab), and if b+1 c, then a+1 b+1 c,
proving the inductive step. However, if b+1 <c, then the edge e
a+1
has not
been used in the k-th phase since it is not even in the k-th layered network,
where all edges are between adjacent layers. If e
a+1
has not been used in the
k-th layered network and is useful from v
a
to v
a+1
in the beginning of phase
k +1, then it has been useful from v
a
to v
a+1
in the beginning of phase k, as
well. Thus, v
a+1
cannot belong to V
c
since by the algorithm it belongs to a
previous layer of the k-th network.
Now, in particular, t =v
l
k+1
V
k
. Therefore, l
k+1
l
k
. Also, equality cannot
hold because in this case, the entire P is in the k-th layered network, and if all
its edges are still useful in the beginning of phase k+1, then the nal ow f

of phase k was not maximal. This proves the claim of the lemma in the case
that all vertices of P are in the k-th layered network.
If not all the vertices of P appear in the kth layered network, then let v
a
e
a+1

v
a+1
be the rst edge of P such that for some b, v
a
V
b
, but v
a+1
is not in the
102 5 Flow in Networks
k-th layered network. Thus, e
a+1
was not used in phase k. Since e
a+1
is useful
in the beginning of phase k+1, it was also useful in the beginning of phase k.
The only possible reason for v
a+1
not to belong to V
b+1
is that b+1 =l
k
and
v
a+1
,=t. Since t =v
l
k+1
, it follows that a+1 <l
k+1
. By the argument of the
previous paragraph, a b. Thus, l
k
=b+1 a+1 <l
k+1
.
Corollary 5.2 The number of phases is bounded by |V|.
Proof: Clearly, l
1
1. By Lemma 5.7, for the k-th phase, if it is not the last,
l
k
k. Since k l
k
|V| 1, the number of phases, including the last, is
bounded by |V|.
Theorem 5.2 The Dinitz algorithm terminates in time O(|V|
2
|E|) and yields a
maximum ow.
Proof: By Corollary 5.2, the number of phases is bounded by |V|. By
Lemma 5.6, each phase requires O(|V| |E|) time. Thus, the whole algo-
rithm takes O(|V|
2
|E|) time to terminate. By Lemma 5.4, the resulting ow
is maximum in N.
Theorem 5.3 (The max-ow min-cut theorem) Every nite network has a
maximum ow and a minimum cut, and their values are equal.
Proof: By Theorem 5.2 there is a max-ow, and by the proof of Lemma 5.4,
there is a cut of the same value. By Corollary 5.1, this cut is of minimum
value.
5.4 Networks with Upper and Lower Bounds
In the previous sections, we have assumed that the owin each edge is bounded
from above by the capacity of the edge, but that the lower bound on the ow in
every edge is zero. The signicance of this assumption is that the assignment
of f(e) = 0, for every edge e, denes a legitimate ow, and the algorithm for
improving the ow can be started with this zero ow.
In this section, in addition to the upper bound c(e) on the owin e, we assume
that the owis also boundedfrombelowby b(e). Thus, the edge rule is changed
as follows: The ow f(e), in every edge e, must satisfy
b(e) f(e) c(e). (5.6)
The vertex rule remains unchanged.
Thus, the problem of nding a maximum ow in a given network
N(G(V, E), s, t, b, c) is divided into two subproblems. First, check whether N
5.4 Networks with Upper and Lower Bounds 103
s t a
0,1 2,3
Figure 5.8: An example of a network which has no legitimate ow.
has legitimate ows, and if the answer is positive, nd one. Second, increase
this initial ow and nd a maximum ow.
A simple example of a network that has no legitimate ow is shown in
Figure 5.8. Here, next to each edge e we write b(e), c(e).
The following method for testing whether a given network Nhas a legitimate
ow function is due to Ford and Fulkerson [7]. It reduces the problem to one
of determining a maximum ow in an auxiliary network

N(

G(

V,

E), s,

t, c), in
which all lower bounds are zero. Here, s is called the auxiliary source, and

t
is called the auxiliary sink. In the auxiliary ow problem, s and t are not the
source and the sink anymore and must satisfy the vertex rule. We shall show
that the original N has legitimate ows if and only if the maximum ow in

N
satises a condition to be specied shortly.

N is dened as follows:
(i)

V =V { s,

t}, where s and

t are two new vertices.
(ii)

E = E o T {e

, e

}, where o, T and {e

, e

} are sets of new edges

dened as follows:
o = { s v | v V},
T = {v

t | v V},
and s
e

t and t
e

s.
(iii) c :

ER
0
{} is dened separately for each of the four sets of edges:
For e E,
c(e) =c(e) b(e). (5.7)
For s

v, c() =

e(v)
b(e).
For v

t, c() =

e(v)
b(e).
c(e

) =and c(e

) =.
Let us demonstrate this construction on the network shown in Figure 5.9. The
auxiliary network is shown in Figure 5.10. The upper bounds are shown next
to the edges to which they apply.
Now we can use the Ford and Fulkerson or the Dinitz algorithm to nd a
maximumow in the auxiliary network. It is left to the reader to verify that the
104 5 Flow in Networks
s t
w x
y z
5,7
1,3 3,5
0,10 2,6
2,4
1,3
2,8
Figure 5.9: An example of a network for which we want to determine
whether it has a legitimate ow.
s t
w x
y z
1 2 4 4 0
1 0 3 5 5 2
2
2
2 2
2
6
10 4
5
s

Figure 5.10: The auxiliary network of the example network.

maximum total ow in the auxiliary network of our example is 16. We return
to this example shortly.
Theorem 5.4 The original network N has a legitimate ow if and only if the
maximum ow of the auxiliary network

N saturates all the edges emanating
from s; that is, all edges of o.
Clearly, if all edges of o are saturated, then so are all edges of T . This follows
fromthe fact that every edge e E contributes b(e) to the capacity of one edge
5.4 Networks with Upper and Lower Bounds 105
in o and of one edge in T . Thus, the sum of capacities of edges of o is equal
to the sum of capacities of edges of T .
Proof: First, we prove the sufciency of the condition. Assume that a maximum
owfunction

fof

Nsaturates a11 edges of o. Dene the following owfunction
f, for N: For every e E,
f(e) =b(e) +

f(e). (5.8)
Since

f satises the edge rule in

N, it follows that
0

f(e) c(e). (5.9)
Thus,
b(e) b(e) +

f(e) b(e) + c(e).
By Equations 5.8 and 5.7, one gets
b(e) f(e) c(e) , (5.10)
and f satises the edge rule.
Now, let v V \ {s, t}. We remind the reader that (v) is the set of edges that
enter v in G, and (v) is the set of edges that emanate fromv in G. Let s

v
be the new edge that enters v in

G, and v

t be the new edge that emanates

from v in

G. Since

f satises the vertex rule in

N, we have

e(v)

f(e) +

f() =

e(v)

f(e) +

f() . (5.11)
By the assumption, and are saturated. Thus,

e(v)

f(e) + c() =

e(v)

f(e) + c() .
By the denitions of c() and c(), it follows that

e(v)

f(e) +

e(v)
b(e) =

e(v)

f(e) +

e(v)
b(e) .
By Equation 5.8 and merging the sums on each side of the equation, one gets

e(v)
f(e) =

e(v)
f(e) , (5.12)
106 5 Flow in Networks
and the vertex rule is proved. Thus, f is legitimate in N.
Next, we prove the necessity of the condition. Assume there is a legitimate
ow f in N. Let us show that there is a maximum ow

f in

N for which all
edges of o are saturated.
The steps of the previous proof are reversible, with minor modications. For
an edge e E, use Equation 5.8 to dene

f(e). For an edge o, dene

f() = c(), and for T dene

f() = c(). Finally, if F 0 is the total ow
according to f, dene

f(e

) =F and

f(e

) = 0, and if F <0, dene

f(e

) = 0
and

f(e

) = F.
To show that

f observes the edge rule, all we need to do is show it for e E;
obviously, in the remaining edges,

f is dened in a way which satises the edge
rule. Since f is legitimate in N, it satises Equation 5.10. By Equations 5.8 and
5.7, one concludes that Equation 5.9 holds.
To show that

f observes the vertex rule, rst consider a vertex v V \ {s, t}.
Since f observes the vertex rule in N, Equation 5.12 holds. By reversing the
steps of the proof above, one gets Equation 5.11, which implies that v satises
the vertex rule.
Finally, for vertex s,

f satises the vertex rule for the following reasons. In N,
s may not be balanced. If F >0, then the total outgoing ow is F units greater
than the total incoming ow. The total outgoing ow in

Nin edges of (s) and
in is as it is in N, and the total incoming ow in

N in edges of (s) and in
is also as in N. However,

f(e

) causes the incoming and outgoing total ows

to balance. Similar arguments show that the total ow in t is balanced as well,
and that both s and t are balanced when F <0.
Let us demonstrate the technique for establishing whether the network has a
legitimate ow, andnding one inthe case the answer is positive, onour example
(Figure 5.9). First, we apply the Dinitz algorithm to the auxiliary network of
Figure 5.10 and end up with the ow, as in Figure 5.11. The maximum ow
saturates all edges that emanate from s, and we conclude that the original
network has a legitimate ow. We use Equation 5.8 to dene a legitimate ow
in the original network; this is shown in Figure 5.12. (Next to each edge e we
write b(e), c(e), and f(e), in this order.)
Once a legitimate ow f has been found in N, we turn to the question of
optimizing it. First, let us consider the question of maximizing the total ow.
One canuse the FordandFulkersonalgorithm, except that the backwardlabeling
must be redened as follows. An edge v
e
uis backwardly useful for labeling
vertex v, if
5.4 Networks with Upper and Lower Bounds 107
s t
w x
y z
1, 1 0, 0 3, 3 5, 5 5, 5 2, 2
2, 0
2, 2 2, 0
2, 2
10, 2
2, 1
4, 0
6, 2
5, 5 1, 1 2, 2 4, 4 0, 0 4, 4

t
, 5
, 0
Figure 5.11: Amaximumowin the auxiliary network of the example
network.
s t
y z
w x
1,3,3
5,7,5
3,5,3
2,4,2
1,3,2
0,10,2
2,6,2
2,8,4
Figure 5.12: A legitimate ow in the original example network.
(i) u is labeled and v is not,
(ii) f(e) >b(e).
The label that v gets is (e, ). In this case, Line 7 of the procedure LABEL (see
Algorithm 5.1) is changed to
(e) f(e) b(e).
108 5 Flow in Networks
With this exception, the algorithm is exactly as described in Section 5.2. The
proof that the ow is maximum when the algorithm terminates is similar, but
we need to redene the capacity of a cut determined by S (see Equation 5.4) as
follows:
c(S) =

e(S;

S)
c(e)

e(

S;S)
b(e) .
It is easy to prove that a statement just like Lemma 5.2 still holds; namely,
F c(S) .
Now, the set of labeled vertices S, when the algorithm terminates, satises this
inequality by equality. Thus, the ow is maximum, and the indicated cut is
minimum.
Clearly, the Dinitz algorithm can also be used. It is left as an exercise for the
reader to make the necessary changes.
In certain applications, what we want is a minimum ow, that is, a legitimate
ow function f for which the total ow F is minimum. Consider a network
N(G(V, E), b, c), where function b : E R species the lower bounds on the
ow in the edges, and c : E R species the upper bounds on the ow in the
edges. Let s, t V be any two vertices. Denote by f : E R a legitimate ow
function of the network (G(V, E), s, t, b, c), where s plays the role of the source
and t plays the role of the sink. Also, let F
s,t
be the corresponding total ow.
Clearly, f is legitimate in (G(V, E), t, s, b, c) as well, where the roles of the
source and the sink are reversed. Also,
F
s,t
= F
t,s
.
It follows that f is a minimum ow in (G(V, E), s, t, b, c) if and only if f is a
maximum ow in (G(V, E), t, s, b, c). Thus, our techniques solve the problem
of minimizing the total ow by simply exchanging the roles of s and t.
For a set {s} S V \ {t}, let us dene c(S), (the value of the cut (S;

S) for
minimum ow purposes) as follows:
c(S) =

e(S;

S)
b(e)

e(

S;S)
c(e) .
Clearly, c(S) = c(

S). Now, assume f is a maximumowin (G(V, E), t, s, b, c).

By the max-ow min-cut theorem (Theorem 5.3), which also applies to net-
works with lower and upper bounds, provided they have legitimate ows, there
5.5 Problems 109
is a cut dened by some {s} S V \ {t}, such that
F
t,s
=c(

S) .
Thus,
F
s,t
= c(S) ,
and
F
s,t
=c(S) .
One can prove and use a lemma similar to Lemma 5.2, with the inequality
reversed. Thus, c(S) is maximum. It follows that for networks with lower and
upper bounds, which have legitimate ows, a min-owmax-cut theoremholds.
5.5 Problems
Problem 5.1 Find a maximum ow in the network shown below.
s
a b
c d
e f
t
3
2
9
6
2
7
5
7
1
2
4
6
The number next to each edge is its capacity. Show a minimum cut. How
many minimum cuts can you nd?
Problem 5.2 In the following network x
1
, x
2
, x
3
, are all sources (of the same
commodity). The supply available at x
1
is 5, at x
2
is 10, and at x
3
is 5. The
110 5 Flow in Networks
a
b c
d
7
9
4
3
3
7
3
2 5
8
8
1 6
8
1
2
4
2
7
x
1
x
2
x
3
y
1
y
2
y
3
vertices y
1
, y
2
, y
3
, are all sinks. The demand required at y
1
is 5, at y
2
is 10,
and at y
3
is 5. Find out whether all requirements can be met simultaneously.
(Hint: One way of solving this type of problem is to reduce it to the familiar
one-source one-sink format. Introduce auxiliary source s and sink t. Connect s
to x
i
through a directed edge of capacity equal to x
i
s supply. Connect each y
i
to t through a directed edge of capacity equal to y
i
s demand. Find a maximum
ow in the resulting network, and observe whether all demands are met.)
Problem 5.3 In the following network, in addition to the capacities of the
edges, each vertex other than s and t has an upper bound on the ow that may
s t
5
4
5
a
5
b
c
5
d
e
4
f
3
3
2
9
6
7
2
7
5
3
4
2
6
1
5.5 Problems 111
ow through it. These vertex capacities are written below the vertex labels.
Find a maximum ow for this network. (Hint: One way of solving this type
of problem is to replace each vertex v by two vertices v

and v

with an edge
v

e
v

, where c(e) is the upper bound on the ow through v. All edges that
previously entered v now enter v

, and all edges that previously emanated from

v now emanate from v

.)
Problem 5.4
(1) Describe an alternative labeling procedure, like that of Ford and Fulkerson,
for maximizing the ow, except that the labeling starts at t, and if it reaches
s, an augmenting path is found.
(2) Demonstrate your algorithm on the following network:
s t
a
c
b
d
4
5
6
4
4
4
2
7
6
(3) Describe a method of locating an edge with the property that increasing
its capacity increases the maximum ow in the network. (Hint: One way
of doing this is to use both source-to-sink and sink-to-source labelings.)
Demonstrate your method on the network above.
(4) Does an edge like this always exist? Prove your claim.
Problem 5.5 In a network N(G(V, E), s, t, c), there are two sets, {s} S
1

V \ {t} and {s} S
2
V \ {t}, and each of them denes a minimum cut. Prove
that each of S
1
S
2
and S
1
S
2
denes a minimum cut as well.
Problem 5.6 Let f be a maximum ow in N(G(V, E), s, t, c), where all edge
capacities are positive integers, andfor everye E, f(e) is a nonnegative integer.
112 5 Flow in Networks
The capacity of u
e
v is reduced by one unit. It is necessary to nd a
maximum ow in the new network.
Describe an algorithm to achieve this goal whose time complexity is O(|E|).
(Hint: Without loss of generality, assume f(e) = c(e). If there is a directed
circuit via e that carries ow, reduce the ow on every edge of the circuit by
one unit. If not, rst locate a directed path from s to t via e that carries ow,
and reduce the ow in every edge of the path by one unit, and then look for an
augmenting path from s to t in the new network.)
Problem 5.7 Let G(V, E) be a nite directed graph without parallel edges.
Describe an algorithmthat is a modication of the algorithmof Ford and Fulk-
erson for nding a maximum ow in a given network N(G, s, t, c), except that,
in each stage, one looks for an augmenting path for which , the value by
which the ow is increased, is maximum. The time complexity of nding each
such path should be O(|V|
2
). (Comment: This algorithm can be shown to be
polynomial; see [4].)
Problem 5.8 In the following network, the capacities of the edges are written
next to them. Use the Dinitz algorithmto nd its maximumowwhen the initial
ow is zero everywhere. How many phases are there in which t is reached?
s t
2
2 2 2 2
2 2 2 2
2
2
2
1 1
Problem 5.9 A ow in a network is said to have circuits if there is at least one
directed circuit such that on all its edges the ow is positive. Such a circular
ow is superuous since it contributes nothing to the total ow.
5.5 Problems 113

Show that if we start with zero ow everywhere and use the Dinitz algorithm
to nd a maximum ow, we may end up with a ow that has circuits. (Hint:
Consider the graph depicted in Figure 5.13.)
s t
Figure 5.13: Hint for Problem 5.9.

Describe an O(|E|
2
) algorithm to remove all circular ow from a given ow
function.
Problem 5.10 Let N(G(V, E), c) be a network where G is a nite directed
graph, and c is a capacity function on the edges. For every x, y V, let F
xy
denote the maximum ow in case x is the source and y is the sink.
Prove that for every three vertices u, v, w V, F
uw
min{F
uv
, F
vw
}.
Problem 5.11 Prove that in a network with a lower bound b(e) 0 for every
edge e, but no upper bound (c(e) = ), there is a legitimate ow if and only
if for every edge e, for which b(e) >0, either e is in a directed circuit or e is
in a directed path from s to t or from t to s. Show that in such a network, a
legitimate ow can be found in time O(|V| |E|).
Problem5.12 Find a minimumowfroms to t for the network of Problem5.1,
where the numbers next to the edges are now assumed to be lower bounds, and
there are no upper bounds.
Problem 5.13 The two networks shown below have both lower and upper
bounds on the ow through the edges. Which of the two networks has no
legitimate ow? Find both a maximum ow and a minimum ow if a legit-
imate ow exists. If no legitimate ow exists, display a set of vertices that
114 5 Flow in Networks
includes neither the source nor the sink and is required to produce ow or to
absorb it.
5
s t
3,5
2,6
9,12
3,4
1,2
4,6
2,4
3,6
s t
3,5
2,6
5,10
3,4
1,2
4,6
2,4
3,6
(a)
(b)
Problem 5.14 Prove that a network with lower and upper bounds on the ow
in the edges has no legitimate ow if and only if there exists a set of vertices
which includes the neither source nor the sink and is required to produce ow
or to absorb it.
5
A set of vertices AV \{s, t} is required to absorb ow if

e(

A;A)
b(e) >

e(A;

A)
c(e) .
5.6 Notes by Andrew Goldberg 115
5.6 Notes by Andrew Goldberg
The augmenting path algorithm is due to Ford and Fulkerson [7, 6]. Dinitz [2]
developed the blocking ow algorithm, which runs in polynomial time. The
observation that augmenting along a shortest path leads to a polynomial-time
algorithmhas been made independentlyby Edmonds and Karp [4]. Amore ef-
cient, O(|V|
3
), variant of the blocking ow algorithm based on preows is due
to Karzanov [12]. Sleator and Tarjan [14] used the dynamic tree data structure
to improve this bound to O(|V||E| log|V|). This has been further improved to
O(|V||E| log(|V|
2
/|E|)) by Goldberg and Tarjan [10].
Goldberg and Tarjan [9] developed the push-relabel method that uses pre-
ows and the push operation similar to Karzanovs algorithm. However, instead
of building a layered network, the algorithm uses the relabel operation for
ne-grains updates of vertex distances. The push-relabel method leads to the
best currently known, strongly polynomial bound of King et al. [13]. This
bounds comes very close, but does not quite achieve, O(|V||E|). The push-
relabel algorithm is also highly practical when used in combination with
efciency-enhancing heuristics [1].
Karzanov [11] and, independently, Even and Tarjan [5] have shown that
on unit capacity networks, the blocking ow algorithm algorithm runs in
O(min((|E|
1/2
, |V|
2/3
)|E|) time. This leaves a polynomial gap between the
general and the unit-capacity cases.
When talking about shortest augmenting paths, one has to assign lengths
for residual edges. All polynomial-time algorithms mentioned above use the
unit length function. Edmonds and Karp [4] observed that one can use
other length functions. However, for a long time nobody was able to use
this observation to improve the time bounds. Goldberg and Rao [8] use the
binary length function (zero for high- and one for low-capacity edges) to get
an O(min((|E|
1/2
, |V|
2/3
)|E| log
|V|
2
|E|
logU) bound for networks with integral
capacities in the range [1, U]. This closes the gap between the general and the
unit-capacity case.
Bibliography
[1] B. V. Cherkassky and A. V. Goldberg. On Implementing Push-Relabel Method
for the Maximum Flow Problem. Algorithmica, 19:390410, 1997.
[2] E. A. Dinic. Algorithmfor Solution of a Problemof Maximum Flowin Networks
with Power Estimation. Soviet Math. Dokl., 11:12771280, 1970.
116 5 Flow in Networks
[3] YemDinitz. Dinitz Algorithm: The original version and Evens version. In Oded
Goldreich, Arnold L. Rosenberg, and Alan L. Selman, editors, Essays in Memory
of Shimon Even, Vol. 3895 of Lecture Notes in Computer Science, pp. 218240.
Springer, 2006.
[4] J. Edmonds and R. M. Karp. Theoretical Improvements in Algorithmic Efciency
for Network Flow Problems. J. Assoc. Comput. Mach., 19:248264, 1972.
[5] S. Even and R. E. Tarjan. Network Flow and Testing Graph Connectivity. SIAM
J. Comput., 4:507518, 1975.
[6] L. R. Ford, Jr. and D. R. Fulkerson. Maximal Flow Through a Network.
Canadian Journal of Math., 8:399404, 1956.
[7] L. R. Ford, Jr. and D. R. Fulkerson. Flows in Networks. Princeton University Press,
1962.
[8] A. V. Goldberg and S. Rao. Beyond the Flow Decomposition Barrier. J. Assoc.
Comput. Mach., 45:753782, 1998.
[9] A. V. Goldberg and R. E. Tarjan. A New Approach to the Maximum Flow
Problem. J. Assoc. Comput. Mach., 35:921940, 1988.
[10] A. V. Goldberg and R. E. Tarjan. Finding Minimum-Cost Circulations by
Successive Approximation. Math. of OR, 15:430466, 1990.
[11] A. V. Karzanov. O nakhozhdenii maksimalnogo potoka v setyakh spetsialnogo
vida i nekotorykh prilozheniyakh. In Matematicheskie Voprosy Upravleniya
Proizvodstvom, volume 5. Moscow State University Press, 1973. In Russian; title
translation: On Finding Maximum Flows in Networks with Special Structure and
Some Applications.
[12] A. V. Karzanov. Determining the Maximal Flow in a Network by the Method of
Preows. Soviet Math. Dok., 15:434437, 1974.
[13] V. King, S. Rao, and R. Tarjan. A Faster Deterministic Maximum Flow
Algorithm. J. Algorithms, 17:447474, 1994.
[14] D. D. Sleator and R. E. Tarjan. A Data Structure for Dynamic Trees. J. Comput.
System Sci., 26:362391, 1983.
6
Applications of Network Flow Techniques
6.1 Zero-One Network Flow
Several combinatorial problems can be solved through network owtechniques.
In the networks we get, the capacity of all the edges is one. To get better
algorithms with lower time complexities, we need to study these network ow
problems. We follow here the work of Even and Tarjan [1].
Consider a maximum ow problem where for every edge e of G(V, E),
c(e) = 1.
The rst observation is that in the Dinitz algorithm for maximal ow in a
layered network, each time we nd a path, all the edges on it become blocked;
in case the last edge leads to a dead end, we backtrack on this edge, and it
becomes blocked. Thus, the total number of edge traversals is bounded by |E|,
and the whole phase is of time complexity O(|E|). Since the number of phases
is bounded by |V|, the Dinitz algorithm for maximum ow is of complexity
O(|V| |E|).
Our rst goal is to prove a better bound yet: O(|E|
3/2
). However, we need to
prepare a few results beforehand.
Let G(V, E) be a 0-1 network in which c(e) = 1 for all e E with some
integral legal ow function f. Dene

G(V,

E) as follows:
(i) If u
e
v in G, and f(e) = 0, then e

E.
(ii) If u
e
v in G, and f(e) = 1, then u
e

v is in

G. Clearly, e

is a new
edge that corresponds to e.
Thus, |E| = |

E|. Clearly, the useful edges of the layered network that is con-
structed for G with present ow f, with their direction of usefulness, are all
edges of

G.
117
118 6 Applications of Network Flow Techniques
Let us denote by (SI

S)
G
, where s S, t / S, and

S =V S, the set of edges
that emanate from a vertex of S and enter a vertex of

S, and let c(S, G) be the
capacity of the corresponding cut in G. Also, let Mbe the total maximum ow
in G, while F is the total present ow.
Lemma 6.1

M=MF.
Proof: Let S be a subset of V such that s S and t / S. The denition of

G
implies that
c(S,

G) = |(S;

S)

G
| =

e(S;

S)
G
(1 f(e)) +

e(

S;S)
G
f(e).
However,
F =

e(S;

S)
G
f(e)

e(

S;S)
G
f(e).
Thus,
c(S,

G) = |(S;

S)
G
| F =c(S, G) F.
This implies that the minimum cut of G corresponds to the minimum cut
of

G; that is, is dened by the same S. By the max-ow min-cut theorem
(Theorem 5.1), the capacity of a minimum cut of

G is

M (the maximum total
ow in

G). Thus, the lemma follows.
Lemma 6.2 The length of the layered network for the 0-1 network dened by
G(V, E) (with a given s and t) and zero ow everywhere is at most |E|/M.
Proof: We remind the reader that V
i
is the set of vertices of the i-th layer of
the layered network, and E
i
is the set of edges fromV
i1
to V
i
. Since f(e) =0
for every e E, the useful directions are all forward. Thus, every E
i
is equal to
(S;

S)
G
, where S =V
0
V
1
. . . V
i1
. Thus, by Lemma 5.1,
M|E
i
|. (6.1)
Summing up (6.1), for every i = 1, 2, . . . , l where l is the length of the layered
network, we get l M|E|, or
l |E|/M.

Theorem 6.1 For 0-1 networks, Dinitzs algorithm is of time complexity

O(|E|
3/2
).
6.1 Zero-One Network Flow 119
Proof: If M |E|
1/2
, then the number of phases is bounded by |E|
1/2
, and
the result follows. Otherwise, consider the phase during which the total ow
reaches M |E|
1/2
. The total ow F in G(V, E) when the layered network for
this phase is constructed satises
F <M |E|
1/2
.
This layered network is identical with the one constructed for

G, with zero ow
everywhere. Thus, by Lemma 6.1;

M=MF >|E|
1/2
.
By Lemma 6.2, the length l of this layered network satises
l |E|/

M<|E|/|E|
1/2
= |E|
1/2
.
Therefore, the number of phases up to this point is at most |E|
1/2
1, and since
the number of additional phases to completionis at most |E|
1/2
, the total number
of phases is at most 2|E|
1/2
.
A 0-1 network is of type 1 if it has no parallel edges. For such a network we
can prove another upper bound on the time complexity. First, we prove a lemma
similar to Lemma 6.2.
Lemma 6.3 Let G(V, E) dene a 0-1 network of type 1, with maximum total
ow M from s to t. The length l of the rst layered network, when the ow is
zero everywhere, is at most 2|V|/M
1/2
.
Proof: Let V
i
be the set of vertices of the i-th layer. Since there are no parallel
edges, the set of edges, E
i+1
, from V
i
to V
i+1
in the layered network satises
|E
i+1
| |V
i
| |V
i+1
| for every i = 0, 1, . . . , l 1. Since each |E
i
| is the capacity
of a cut, we get that
M|V
i
| |V
i+1
|.
Thus, either V
i
M
1/2
or |V
i+1
| M
1/2
. Clearly,
|V|

i=0
|V
i
|
_
l +1
2
_
M
1/2
.
Thus,
|V|
M
1/2

_
l +1
2
_

2
,
120 6 Applications of Network Flow Techniques
and
2|V|
M
1/2
,
and the lemma follows.
Theorem6.2 For 0-1 networks of type 1, Dinitzs algorithmhas time complexity
O(|V|
2/3
|E|).
Proof: If M |V|
2/3
, the result follows immediately. Let F be the total ow
when the layered network, for the phase during which the total ow reaches the
value M|V|
2/3
, is constructed. This layered network is identical with the rst
layered network for

Gwith zero oweverywhere.

Gmay not be of type 1 since
it may have parallel edges, but it can have at most two parallel edges from one
vertex to another; if e
1
and e
2
are antiparallel in G, f(e
1
) = 0 and f(e
2
) = 1,
then in

Gthere are two parallel edges: e
1
and e
2
. A result similar to Lemma 6.3
yields that
l <2
3/2
|V|/

M
1/2
.
Since

M=MF >M (M |V|
2/3
) = |V|
2/3
, we get
l <
2
3/2
|V|
|V|
1/3
= 2
3/2
|V|
2/3
.
Thus, the number of phases up to this point is O(|V|
2/3
). Since the number of
phases from here to completion is at most |V|
2/3
, the total number of phases is
O(|V|
2/3
).
In certain applications, the networks that arise satisfy the condition that for
each vertex other than s or t, there is either only one edge emanating from it or
only one edge entering it. Such 0-1 networks are called type 2.
Lemma 6.4 Let the 0-1 network dened by G(V, E) be of type 2, with maximum
total owMfrom s to t. The length l of the rst layered network, when the ow
is zero everywhere, is at most (|V| 2)/M+1.
Proof: The structure of G implies that a max-ow in G can be decomposed
into vertex-disjoint directed paths from s to t; that is, no two of these paths
share any vertices, except their common start-vertex s and end-vertex t. (The
ow may imply some directed circuits that are vertex-disjoint from each other
and from the paths above, except possibly at s or t. These circuits are of no
interest to us). The number of these paths is equal to M. Let be the length of a
6.2 Vertex Connectivity of Graphs 121
shortest of these paths. Thus, each of the paths uses at least 1 intermediate
vertices. We have
M (1) |V| 2,
which implies (|V| 2)/M + 1. However, l . Thus, the lemma
follows.
Lemma 6.5 If the 0-1 network dened by Gis of type 2, and if the present ow
function is f, then the corresponding

G also denes a type 2, 0-1 network.
Proof: Clearly

G denes a 0-1 network. What remains to be shown is that in

G, for every vertex v, there is either one emanating edge or only one entering
edge. If there is no ow through v (per f), then, in

G, v has exactly the same
incident edges, and the condition continues to hold. If the ow going through v
is 1, (clearly, it cannot be more) assume that it enters via e
1
and leaves via e
2
.
In

G, neither of these two edges appears, but two edges e

1
and e

2
are added,
which have directions opposite to e
1
and e
2
, respectively. The other edges of G
that are incident to v remain intact in

G. Thus, the numbers of incoming edges
and outgoing edges of v remain the same. Since G is of type 2, so is

G.
Theorem 6.3 For a 0-1 network of type 2, Dinitzs algorithm is of time
complexity O(|V|
1/2
|E|).
Proof: If M |V|
1/2
, then the number of phases is bounded by |V|
1/2
, and
the result follows. Otherwise, consider the phase during which the total ow
reaches the value M |V|
1/2
. Therefore, the layered network for this phase is
constructed when F < M |V|
1/2
. This layered network is identical with the
rst for

G, with zero oweverywhere. Also, by Lemma 6.5,

Gis of type 2. Thus,
by Lemma 6.4, the length l of the layered network is at most (|V| 2)/

M+1.
Now,

M=MF >M (M |V|
1/2
) = |V|
1/2
. Thus,
l
|V| 2
|V|
1/2
+1 =O(|V|
1/2
).
Therefore, the number of phases up to this one is at most O(|V|
1/2
). Since the
number of phases to completion is at most |V|
1/2
more, the total number of
phases is at most O(|V|
1/2
).
6.2 Vertex Connectivity of Graphs
Intuitively, the connectivity of a graph is the minimum number of elements
whose removal from the graph disconnect it. There are four cases. We may
122 6 Applications of Network Flow Techniques
discuss undirected graphs or digraphs; we may discuss the elimination of edges
or vertices. We start with the problem of determining the vertex-connectivity
of an undirected graph. The other cases, which are simpler, are discussed in the
next section.
Let G(V, E) be a nite undirected graph, with no self-loops and no parallel
edges. A set of vertices, S, is called an (a, b) vertex separator if {a, b} V \S,
and every path connecting a and b passes through at least one vertex of S.
Clearly, if a and b are connected by an edge, no (a, b) vertex separator exists.
Let a, b mean that there is no such edge. In this case, let N(a, b) be the
least cardinality of an (a, b) vertex separator. Also, let p(a, b) be the maximum
number of pairwise vertex-disjoint paths connecting a and b in G; clearly, all
these paths share the two end-vertices, but no other vertex appears on more than
one of them.
Theorem 6.4 If a , b then N(a, b) =p(a, b).
This is one of the variations of Mengers theorem [2]. It is not only rem-
iniscent of the max-ow min-cut theorem, but can be proved by it. Dantzig
and Fulkerson [3] pointed out how this can be done, and we shall follow their
approach.
Proof: Construct a digraph

G(

V,

E) as follows. For every v V put two vertices
v

and v

in

V with an edge v

e
v
v

. For every edge u

e
v in G, put two
edges u

and v

in

G. Dene now a network, with digraph

G,
source a

, sink b

, unit capacities for all the edges of the e

v
type (let us call
them internal edges), and innite capacity for all the edges of the e

and e

type (called external edges). For example, in Figure 6.1(b) the network for G,
as shown in Figure 6.1(a), is demonstrated.
We now claim that p(a, b) is equal to the total maximum ow F (froma

to
b

) in the corresponding network. First, assume we have p(a, b) vertex disjoint

paths froma to b in G. Each such path, a
e
1
v
1
e
2
v
2
e
3
. . .
e
l1
v
l1
e
l
b,
indicates a direct path in

G:
a

1
v

1
e
v
1
v

1
e

2
v

2
e
v
2
v

2
e

3
. . .
e

l1
v

l1
e
v
l1
v

l1
e

l
b

These directed paths are vertex-disjoint, and each can be used to ow one
unit froma

to b

. Thus,
F p(a, b).
6.2 Vertex Connectivity of Graphs 123
a
(a) (b)
b

b b
1
1
1
1

a a
Figure 6.1: Construction in the proof of Theorem 6.4
Next, assume f is a ow function which achieves a maximum total ow F in
the network. We may assume that f(e) is either zero or one, for every e

E.
This follows from the fact that one can use the Ford and Fulkerson algorithm,
or the Dinitz algorithm, in which the ow is always integral. Also, the edges
with innite capacity enter a vertex with a single outgoing edge whose capacity
is one and which must satisfy the conversation rule (C2), or they emanate from
a vertex with only one incoming edge of unit capacity which is subject to C2;
thus, the ow through them is actually bounded from above by one. (We have
assigned them innite capacity for convenience reasons, which will become
clear shortly.) Therefore, the total ow F can be decomposed to paths, each
describing the way that one unit reaches b

from a

. These paths are vertex-

disjoint, since the ow through v

or v

, if v / {a, b}, is bounded by one. Each

indicates a path in G. These F paths in G are vertex-disjoint too. Thus,
F p(a, b).
We conclude that F =p(a, b).
By the max-owmin-cut theorem, F is equal to the capacity c(S) of some cut
dened by some S

V, such that a

S and b

/ S. Since
c(S) =

e(S;

S)
c(e),
the set (S;

S) consists of internal edges only. Now, every directed path from

a

to b

in

G uses at least one edge of (S;

S). Thus, every path from a to b

in G uses at least one vertex v such that e
v
(S;

S). Therefore, the set R =

{v | v V and e

v
(S;

S)} is an (a, b) vertex separator. Clearly |R| = c(S).

124 6 Applications of Network Flow Techniques
Thus, we have an (a, b) vertex separator whose cardinality is F. Proving that
N(a, b) F p(a, b).
Finally, it is easy to see that N(a, b) p(a, b), since every path from a
to b uses at least one vertex of the separator, and no two paths can use the
same one.
The algorithm suggested in the proof, for nding N(a, b), when the Dinitz
algorithm is used to solve the network problem, is of time complexity
O(|V|
1/2
|E|). This results from the following considerations. The number of
vertices in

G is 2|V|; the number of edges is |V| +2|E|. Assuming |E| |V|, we
have |

V| =O(|V|) and |

E| =O(|E|). Since we can assign unit capacity to all the

edges without changing the maximum total ow, the network is of type 2. By
Theorem 6.3, the algorithm is of time complexity O(|V|
1/2
|E|). We can even
nd a minimum (a, b) vertex separator as follows: Once the ow is maximum,
change the capacity of the external edges back to and apply the construction
of the layered network. The set of vertices which appear also in this layered
network, S, denes a minimum cut which consists of internal edges only. Let
R be the vertices of G which correspond to the internal edges in (S;

S). R is a
minimum(a, b) vertex separator in G. This addition work is of time complexity
O(|E|).
The vertex connectivity, c, of an undirected graph G(V, E) is dened as
follows:
(i) If G is completely connected, (i.e., every two vertices are connected by an
edge), then c = |V| 1.
(ii) If G is not completely connected, then
c = min
a, b
N(a, b).
Lemma 6.6 If G is not completely connected, then
min
a, b
p(a, b) = min
a,b
p(a, b);
namely, the smallest value of p(a, b) occurs also for some two vertices a and
b that are not connected by an edge.
Proof: Let a, b be a pair of vertices such that a
e
b and p(a, b) is minimum
over all pairs of vertices of the graph. Let G

be the graph obtained from G by

dropping e. Clearly, the number of vertex disjoint paths connecting a and b in
6.2 Vertex Connectivity of Graphs 125
G

, p

(a, b), satises

p

(a, b) =p(a, b) 1.
Also, since a, b in G

, then by Theorem6.4, there is an (a, b) vertex separator

R in G

such that p

(a, b) = |R|.
If |R| = |V| 2, then p(a, b) = |V| 1, and p(a, b) cannot be the least of all
{p(u, v) | u, v V}, since for any u, v, p(u, v) |V| 2. Hence, |R| <|V| 2.
Therefore, there must be some vertex v V (R{a, b}). Now, without loss of
generality, we may assume that R is also an (a, v) vertex separator (or exchange
a and b). Thus, a, v in G and R{b} is an (a, v) vertex separator in G. We
now have
p(a, v) |R| +1 =p(a, b),
and the lemma follows.
Theorem 6.5 c = min
a,b
p(a, b).
Proof: If G is completely connected, then for every two vertices a and b,
p(a, b) = |V| 1, and the theorem holds. If G is not completely connected,
then, by denition,
c = min
a, b
N(a, b).
By Theorem 6.4, min
a, b
N(a, b) = min
a, b
p(a, b). Now by Lemma 6.6,
min
a, b
p(a, b) = min
a,b
p(a, b).
We can use the intermediate result,
c = min
a, b
p(a, b),
to compute the vertex connectivity of G with time complexity O(|V|
5/2
|E|).
However, a slightly better bound can be obtained.
Lemma 6.7 c 2|E|/|V|.
Proof: The vertex (or edge) connectivity of a graph cannot exceed the degree
of any vertex. Thus,
c min
v
d(v).
Also,

v
d(v) = 2 |E|.
Thus, min
v
d(v) 2 |E|/|V|, and the lemma follows.
126 6 Applications of Network Flow Techniques
Procedure VERTEX-CONNECTIVITY(V, E)
1 Order the vertices v
1
, v
2
, . . . , v
|V|
in such a way that v
1
, v for some v.
2
3 i 1
4 while i do
5 for every v such that v
i
, v do
6 min{, N(v
i
, v)}
7 return .
Algorithm 6.1: Constructing the vertex connectivity of a graph G=
(V, E) that is not completely connected.
A procedure to nd the vertex connectivity c of a graph G that is not
completely connected is listed in Algorithm 6.1.
Theorem6.6 The procedure VERTEX-CONNECTIVITYterminates with =c.
Proof: Clearly, after the rst computation of N(v
1
, v) for some v
1
, v,
satises
c |V| 2. (6.2)
From there on, can only decrease, but (6.2) still holds. Thus, for some
k |V| 1, the procedure will terminate. When it does, k +1 c +1.
By denition, c equal to the cardinality of a minimum vertex separator R of
G. Thus, at least one of the vertices v
1
, v
2
, . . . , v
k
is not in R, say v
i
. R separates
the remaining vertices into at least two sets, such that each path from a vertex
of one set to a vertex of another passes through at least on vertex of R. Thus,
there exists a vertex v such that N(v
i
, v) |R| =c, and therefore c.
Clearly, the time complexity of this procedure is O(c |V|
3/2
|E|). By
Lemma 6.7, this is bounded by O(|V
1/2
|E|
2
).
If c =0, then Gis not connected. We can use DFS (or BFS) to test whether this
is the case in O(|E|) time. If c =1, then Gis separable, and as we sawin Section
3.2, this can be tested also by DFS in O(|E|) time. This algorithm determines
also whether c 2, that is, whether it is nonseparable. Before we discuss testing
6.2 Vertex Connectivity of Graphs 127
for a given k, whether c k, let us consider the following interesting theorem
about equivalent conditions for G to be nonseparable.
1
Theorem6.7 Let G(V, E) be an undirected graph with |V| >2 and no isolated
vertices.
2
The following six statements are equivalent.
1. G is nonseparable.
2. For every two vertices x and y there exists a simple circuit which goes
through both.
3. For every two edges e
1
and e
2
there exists a simple circuit which goes through
both.
4. For every two vertices x and y and an edge e there exists a simple path from
x to y which goes through e.
5. For every three vertices x, y and z there exists a simple path from x to z
which goes through y.
6. For every three vertices x, y and z there exists a simple path from x to z
which avoids
Proof: First we prove that (1) is equivalent to (2).
(1) (2): Since G is nonseparable, c 2. By Theorem 6.5, for every two
vertices x and y p(x, y) 2; thus, there is a simple circuit that goes through x
and y.
(2) (1): There cannot exist a separation vertex in G, since every two vertices
lie on some common simple circuit.
Next, let us show that (1) and (3) are equivalent.
(1) (3): From G construct G

as follows. Remove the edges u

1
e
1
v
1
and u
2
e
2
v
2
(without removing any vertices). Add two new vertices, x and
y, and four new edges: u
1
x v
1
, u
2
y v
2
. Clearly, none of the
old vertices become separation vertices by this change. Also, x cannot be a
separation vertex, or either u
1
or v
1
are separation vertices in G. (Here, |V| >2
is used.) Thus, G

is nonseparable. Hence, by the equivalence of (1) and (2),

G

satises (2). Therefore, there exists a simple circuit in G

that goes through

x and y. This circuit indicates a circuit through e
1
and e
2
in G.
(3) (1): Let x and y be any two vertices. Since G has no isolated vertices,
there is an edge e
1
incident to x and an edge e
2
incident to y. (If e
1
=e
2
, choose
any other edge to replace e
2
; the replacement need not even be incident to y; the
replacement exists since there is at least one other vertex, and it is not isolated.)
1
Many authors use the term biconnected to mean nonseparable. I prefer to call a graph
biconnected if c = 2.
2
Namely, for every v V, d(v) >0. Ghas been assumed to have no self-loops.
128 6 Applications of Network Flow Techniques
By (3) there is a simple circuit through e
1
and e
2
, and therefore a circuit through
x and y. Thus, (2) holds, and (1) follows.
Now, let us prove that (3) (4) (5) (6) ((3)).
(3) (4): Since (3) holds, the graph G is nonseparable. Add a new edge
x
e

y, if such does no exist already in G. Clearly, the new graph G

, is still
nonseparable. By the equivalence of (1) and (3), G

satisfy statement (3). Thus,

there is a simple circuit which goes through e and e

in G

. Therefore, there is
a simple path in G from x to y through e.
(4) (5): Let e be an edge incident to vertex y; such an edge exists, since
there are no isolated vertices in G. By (4), there is a simple path from x to z
through e. Thus, this path goes through y.
(5) (6): Let p be a simple path which goes from x to y through z; such a
path exists, since (5) holds for every three vertices. The rst part of p, from x
to z does not pass through y.
(6) (1): If (6) holds, then there cannot be any separation vertex in G.
Let us now return to the problem of testing the vertex connectivity of a given
graph G; that is, testing whether c is greater than or equal to a given positive
integer k. We have already seen that for k=1 and 2, there is an O(|E|) algorithm.
Hopcroft and Tarjan [4] showed that k =3 can also be tested in linear time, but
their algorithmis quite complicated and does not seem to generalize for higher
values of k. Let us present a method suggested by Kleitman [5] and improved
by Even [6].
Let L = {v
1
, v
2
, . . . , v
l
} be a subset of V, where l k. Dene

G as follows:

G
includes all the vertices and edges of G. In addition, it includes a new vertex
s connected by an edge to each of the vertices of L;

G is called the auxiliary
graph.
Lemma 6.8 Let u V L. If p(v
i
, u) k in G, for every v
i
L, then, in

G,
p(s, u) k.
Proof: Assume not. Then p(s, u) < k. By Theorem 6.4, there exists a (s, u)
vertex separator S in

Gsuch that |S| <k. Let R be the set of vertices such that all
paths in

Gfroms to v R pass through at least one vertex of S. Clearly, v
i
/ R,
since v
i
is connected by an edge to s. However, since l k >|S|, there exists
some 1 i l such that v
i
/ S. All paths from v
i
to u go through vertices of
S. Thus, p(v
i
, u) |S| <k, contradicting the assumption.
Let V = {v
1
, v
2
, . . . , v
n
}. Let j be the least integer such that for some i < j,
p(v
i
, v
j
) <k in G.
6.3 Connectivity of Digraphs and Edge Connectivity 129
Lemma 6.9 Let j be as dened above and

G be the auxiliary graph for L =
{v
1
, v
2
, . . . , v
j1
}. In

G, p(s, v
j
) <k
Proof: Consider a minimum (v
i
, v
j
) vertex separator S. By Theorem 6.4,
|S| <k. Let R be the set of all vertices v V such that all the paths from v
i
to v in Gpass through vertices of S. Clearly, v
j
R. If for some j

<j, v
j
R,
then p(i, j

) |S| <k, and the choice of j is wrong. Thus, v

j
is the least vertex
in R (i.e. the vertex for which the subscript is minimum). Hence, L R = .
Thus, in

G, S is an (s, v
j
) vertex separator, and p(s, v
j
) <k.
The following algorithm determines whether the vertex connectivity of a
given undirected graph G(V, E), where V = {v
1
, v
2
, . . . , v
n
}, is at least k.
1. For every i and j such that 1 i < j k, check whether p(v
i
, v
j
) k. If
for some i and j this test fails then halt; Gs connectivity is less than k.
2. For every j such that k+1 j n, form

G(with L = {v
1
, v
2
, . . . , v
j1
}), and
check whether, in

G, p(s, v
j
) k. If for some j this test fails, then halt; Gs
connectivity is less than k.
3. Halt; the connectivity of G is at least k.
The proof for the algorithms validity is as follows: If Gs connectivity is at
least k, then clearly no failure will be detectedinStep (6.2). Also, by Lemma 6.8,
no failure will occur in Step (6.2), and the algorithmwill halt in Step (6.2) with
the right conclusion. If Gs connectivity is less than k, and it is not detected
directly in Step (6.2) then, by Lemma 6.9, it will be detected in Step (6.2).
Step (6.2) takes O(k
3
|E|) steps, since we have to solve k(k 1)/2 ow
problems. In each we have to nd k augmenting paths, and each path takes
O(|E|) steps to nd.
Step (6.2) takes O(k |V| |E|) steps, since we have to solve |V| k ow
problems, again for each up to total ow k.
Thus, if k |V|
1/2
then the time complexity of the algorithmis O(k |V| |E|).
The readers who are familiar with the interesting result of Gomory and Hu [7]
for nding all |V| (|V| 1) total ows, for all source-sink pairs in an undirected
network by solving only |V| 1 network ow problems, should notice that this
technique is of no help in our problem. The reason for that is that even if G is
undirected, the network we get for vertex connectivity testing is directed.
6.3 Connectivity of Digraphs and Edge Connectivity
First, let us consider the problem of vertex-connectivity of a digraph G(V, E).
The denition of an (a, b) vertex separator is the same as in the undirected case,
130 6 Applications of Network Flow Techniques
except that now all we are looking at are directed paths from a to b; i.e., an
(a, b) vertex separator is a set of vertices S such that {a, b} S = and every
directed path from a to b passes through at least one vertex of S. Accordingly,
N(a, b) and p(a, b) are dened. The theorem analogous to Theorem 6.4, still
holds, except that the algorithm is even simpler: For every edge u
e
v in G
there is only one edge u

in

G.
The vertex connectivity, c, of a digraph G(V, E) is dened as follows:
(i) If G is completely connected, (i.e., for every two vertices a and b, there
are edges a b and b a), then c = |V| 1.
(ii) If G is not completely connected, then
c = min
a, b
N(a, b).
The lemma analogous to Lemma 6.6 still holds, and the proof goes along
the same lines. Also, the theorem analogous to Theorem 6.5 holds, and the
complexity it yields is the same. If G has no parallel edges, a statement like
Lemma 6.7 holds, and the procedure and the proof of its validity (Theorem6.6)
extend to the directed case, except that for each v
i
, we compute both N(v
i
, v)
and N(v, v
i
).
The algorithmfor testing k connectivity extends also to the directed case, and
again all we need to change is that whenever p(a, b) was computed, we now
have to compute both p(a, b) and p(b, a).
Let us nowconsider the case of edge connectivity both in graphs and digraphs.
Let G(V, E) be an undirected graph. Aset of edges, T, is called an (a, b) edge
separator if every path from a to b passes through at least one edge of T. Let
M(a, b) be the least cardinality of an (a, b) edge separator. Let p(a, b) be now
the maximum number of edge disjoint paths which connect a with b.
Theorem 6.8 M(a, b) =p(a, b).
The proof is similar to that of Theorem 6.4, only simpler. There is no need
to split vertices. Thus, in

G,

V = V. We still represent each edge u v of
G by two edges u
e

v and v
e

u in

G. There is no loss of generality in
assuming that the ow function in

Gsatises the condition that either f(e

) =0
or f(e

) =0; for if f(e

) =f(e

) =1 then replacing both by 0 does not change

the total ow. The rest of the proof raises no difculties.
The edge connectivity, c, of a graph G is dened by c =min
a,b
M(a, b). By
Theorem 6.8 and its proof, we can nd c by the network ow technique. The
6.3 Connectivity of Digraphs and Edge Connectivity 131
networks we get are of type 1. Both Theorem6.1 and Theorem6.2 apply. Thus,
each network ow problem is solvable by Dinitzs algorithm with complexity
O(min{|E|
3/2
, |V|
2/3
|E|}).
Let T be a minimum edge separator in G; that is, |T| =c. Let v be any vertex
of G. For every vertex v

, on the other side of T, M(v, v

) =c. Thus, in order

to determine c, we can use
c = min
v

V{v}
M(v, v

).
We need to solve at most |V| 1 network owproblems. Thus, the complexity
of the algorithm is O(|V| |E| min{|E|
1/2
, |V|
2/3
}).
In the case of edge connectivity of digraphs, we need to consider directed
paths. The denition of an (a, b) edge separator is accordingly a set of edges,
T, such that every directed path from a to b uses at least one edge of T. The
denition of p(a, b) again uses directed paths, and the proof of the statement
analogous to Theorem6.8 is the easiest of all, since

Gis nowGwith no changes.
In the denition of c, the edge connectivity, we need the following change:
c = min{M(a, b) | (a, b) V V},
namely, we need to consider all ordered pairs of vertices.
The networks we get are still of type 1 andthe complexity of each is still O(|E|
min{|E|
1/2
, |V|
2/3
}). The approach of testing for one vertex v, both M(v, v

) and
M(v

, v) for all v

V {v} still works, to yield the same complexity: O(|V|

|E| min{|E|
1/2
, |V|
2/3
}). However, the same result, with an improvement only
in the constant coefcient follows fromthe following interesting observation of
Schnorr [8], which applies both to the directed and undirected edge connectivity
problems.
Lemma 6.10 Let v
1
, v
2
, . . . , v
n
be a circular ordering (i.e. v
n+1
= v
1
) of the
vertices of a digraph G. The edge connectivity, c, of G satises
c = min
1in
M(v
i
, v
i+1
).
Proof: Let T be a minimumedge separator in G. That means that there are two
vertices a and b such that T is an (a, b) edge separator. Dene
L = {v | there is a directed path from a to v which avoids T}.
R = {v | there is no directed path from a to v which avoids T}.
132 6 Applications of Network Flow Techniques
Clearly, L R = V and L R = . Let l L and r R. T is an (l, r) edge
separator; for if it is not, then r belongs in L. Therefore, M(l, r) |T|. Since
T is a minimum edge separator, M(l, r) = |T|. Now neither L nor R are empty,
since they contain a and b, respectively.
Consider now the circular ordering of V. There must be an i, 1 i n, such
that v
i
L and v
i+1
R. Hence, the result.
In the case of graphs and digraphs we can test for k connectivity, easily, in
time complexity O(k |V| |E|). Instead of running each network ow problem
to completion, we terminate it when the total ow reaches k. Each augmenting
path takes O(|E|) time and there are |V| ow problems. As we can see, testing
for k edge connectivity is much easier than for k vertex connectivity. The reason
is that vertices cannot participate in the separating set which consists of edges.
We can also use this approach to determine the edge connectivity, c, in time
O(c |V| |E|). We run all the |V| network ow problems in parallel, one aug-
menting path for each network in turn. When no augmenting path exists in any
of the |V| problems, we terminate. The cost increase is only in space, since we
need to store all |V| problems simultaneously. One can use binary search on c to
avoid this increase in space requirements, but in this case the time complexity
is O(c |V| |E| logc).
We conclude our discussion of edge connectivity with a very powerful
theorem of Edmonds [9]. The proof presented here is due to Lovsz [10].
Theorem 6.9 Let a be a vertex of a digraph G(V, E) and min
vV{a}
M(a, v) = k. There are k edge-disjoint directed spanning trees of G rooted
at a.
Proof: The theoremtrivially holds for k=1. We prove the theorembyinduction
on k. Let us denote by
G
(S) the number of edges in (S;

S) in G. If H is a
subgraph of G then GH is the digraph resulting from the deletion of all the
edges of H from G.
Clearly, the condition that min
vV{a}
M(v, a) k is equivalent to the
statement that, for every S V, S ,=V and a S,
G
(S) k.
Let F(V

, E

) be a subgraph of G such that

(i) F is a directed tree rooted at a (which is not necessarily spanning);
(ii) For every S V, S ,=V and a S,
GF
(S) k1.
If F is spanning directed tree then we get the result immediately; by the
inductive hypothesis there are k1 edge-disjoint spanning trees rooted at a in
GF, and F is one more.
6.3 Connectivity of Digraphs and Edge Connectivity 133
The crux of the proof is to show that if F is not spanning then an edge of the
set (V

;

V

) can be added to F, to increase its number of vertices by one and

still satisfy both (i) and (ii).
Consider the following three conditions on a subset of vertices, S:
(1) a S,
(2) SV

,=V,
(3)
GF
(S) =k1.
Let us show that if no such S exists, then one can add any edge e (V

;

V

)
to F. Clearly, F+e satises (i). Now, if (ii) does not hold, then there exists an S
such that S ,=V, a S, and
G(F+e)
(S) <k1. It follows that
GF
(S)<k.
Now, by (ii),
GF
(S) k1. Thus,
GF
(S) =k1, and S satises condi-
tion ((3)). Let uand v be vertices such that u
e
v. Since
G(F+e)
(S) <k1
and
GF
(S) = k 1, v / S. Also, v / V

. Thus, S V

,= V, satisfying
condition ((2)). Therefore, S satises all three conditions; A contradiction.
Now, let A be a maximal
3
set of vertices that satises ((1)), ((2)), and ((3)).
Since the edges of F all enter vertices of V

GF
(AV

) =
G
(AV

) k.
By condition ((3)),

GF
(AV

) >
GF
(A).
The inequality implies that there exists an edge x
e
y that belongs to (A
V

;AV

) and does not belong to (A;

A) in GF. Hence, x

AV

and
y

A

V

. Clearly, F+e satises (i). It remains to be shown that it satises (ii).

Let S V, S ,=V and a S. If e / (S;

S), then

G(F+e)
(S) =
GF
(S) k1.
Assume e (S;

S). It is not hard to prove that for every two subsets of V, S,

and A,

GF
(SA) +
GF
(SA)
GF
(S) +
GF
(A),
by considering the sets of edges connecting S A, S

A,

S A, and

S

A.
Now,
GF
(A) =k1 and
GF
(SA) k1. Therefore,

GF
(SA)
GF
(S).
3
Namely, no larger set that contains Ahas all three properties.
134 6 Applications of Network Flow Techniques
Since x Sand x / A, SA; namely, SAis larger than A. Also, y

S, y

A,
and y

V

. Thus, (SA)V

,=V. By the maximality of A,

GF
(SA) k.
This implies that
GF
(S)k; therefore
G(F+e)
(S)k1, proving(ii).
The proof provides an algorithm for nding k edge-disjoint directed trees
rooted at a. We look for a tree F such that min
vV{a}
M(a, v) k 1 in
GF, by adding to F one edge at a time. For each candidate edge e, we have
to check whether min
vV{a}
M(a, v) k 1 in G (F +e). This can be
done by solving |V| 1 network ow problems, each of complexity O(k |E|).
Thus, the test for each candidate edge is O(k |V| |E|). No edge needs be
considered more than once in the construction of F, yielding the time complexity
O(k |V| |E|
2
). Since we repeat the construction k times, the whole algorithm
is of time complexity O(k
2
|V| |E|
2
).
The following theorem was conjectured by Y. Shiloach and proved by Even,
Garey, and Tarjan [11].
Theorem 6.10 Let G(V, E) be a digraph whose edge connectivity is at least k.
For every two vertices, uand v, and every l, 0 l k, there are l directed paths
from u to v and kl directed paths from v to u, which are all edge disjoint.
Proof: Construct an auxiliary graph

Gby adding to Ga newvertex a, l parallel
edges from a to u, and k l parallel edges from a to v. Let us rst show that
in

G min
wV
M(a, w) =k.
If min
wV
M(a, w) <k, then there exists a set S such that S V{a}, aS,
and |(S;

S)| <k in

G. Clearly, S ,= {a} for |({a};V)| =k. Let x S {a}. Thus,
M(x, y) < k for every y

S, and the same also holds in G, since G is a sub-
graph of

G. This contradicts the assumption that in G, the edge connectivity is
at least k.
Now, by Theorem 6.9, there are k edge-disjoint directed spanning trees, in

G, rooted at a. Exactly one edge out of a appears in each tree. Thus, each
of the trees that uses an edge a u contains a directed path from u to v,
and each of the trees that uses an edge a v contains a directed path from
v to u. All these paths are, clearly, edge disjoint.
Corollary 6.1 If the edge connectivity of a digraph is at least 2, then for every
two vertices u and v there exists a directed circuit that goes through u and v
in which no edge appears more than once.
Proof: Use k = 2, l = 1 in Theorem 6.10.
It is interesting to note that no such easy result exists in the case of vertex
connectivity and a simple directed circuit through given two vertices. In [11],
6.4 Maximum Matching in Bipartite Graphs 135
a digraph with vertex connectivity 5 is shown such that for every two of its
vertices there is no simple directed circuit that passes through both. The author
does not know whether any vertex connectivity will guarantee the existence of
a simple directed circuit through any two vertices.
6.4 Maximum Matching in Bipartite Graphs
A set of edges, M, of a graph G(V, E) with no self-loops, is called a matching
if every vertex is incident to at most one edge of M. The problem of nding a
maximummatching was rst solved in polynomial time by Edmonds [12]. The
best known result of Even and Kariv [13] is O(|V|
2.5
). These algorithms are too
complicated to be included here, and they do not use network ow techniques.
An easier problemis to nd a maximummatching in a bipartite graph, that is,
a graph in which V =XY, XY =, and each edge has one end-vertex in X
and one in Y. This problemis also known as the marriage problem. We present
here its solution via network owand showthat its complexity is O(|V|
1/2
|E|).
This result was rst achieved by Hopcroft and Karp [14].
Let us construct a network N(G). Its digraph

G(

V,

E) is dened as follows:

V = {s, t} V,

E = {s x|x X} {y t|y Y} {x y|x y in G}.

Let c(s x) =c(y t) = 1 for every x X and y Y. For every edge
x
e
y let c(e) =. (This innite capacity is dened in order to simplify our
proof of Theorem 6.12. Actually, since there is only one edge entering x, with
unit capacity, the ow in x y is bounded by 1.) The source is s and the
sink is t. For example, consider the bipartite graph G shown in Figure 6.2(a).
Its corresponding network N(G) is shown in Figure 6.2(b).
Theorem 6.11 The number of edges in a maximum matching of a bipartite
graph G is equal to the maximum ow, F, in its corresponding network, N(G).
Proof: Let M be a maximum matching. For each edge x y of M, use the
directed path s x y t to ow one unit froms to t. Clearly, all these
paths are vertex-disjoint. Thus, F |M|.
Let f be a ow function on N(G), which is integral. (There is no loss of
generality here. As we saw, in Chapter 5, every network with integral capacities
has a maximum integral ow.) All the directed paths connecting s and t are of
the form s x y t. If such a path is used to ow (one unit) from s
136 6 Applications of Network Flow Techniques
x
1
x
3
x
4
x
5
x
2
y
1
y
2
(a) (b)
y
3
y
4

1
1
1
1
1
1
1
1
1

s t
x
1
x
3
x
4
x
5
x
2
y
1
y
2
y
3
y
4
Figure 6.2: Construction of N(G).
,1
,1
1,1
1,1
1,1
1,1
1,1
,1
,0
,0
,0
,1
,0
1,1
1,1
1,1
1,1
s t
x
1
x
3
x
4
x
5
x
2
y
1
y
2
y
3
y
4
x
1
x
3
x
4
x
5
x
2
y
1
y
2
y
3
y
4
(a) (b)
Figure 6.3: Maximum ow found by Dinitzs algorithm and its
corresponding matching.
to t, then no other edge x y

or x

y can carry ow, since there is only

one edge s x and its capacity is one, and the same is true for y t. Thus,
the set of edges x y, for which f(x y) = 1, constitutes a matching in
G. Thus, |M| F.
The proof indicates how the network ow solution can yield a maximum
matching. for our example, a maximum ow, found by Dinitzs algorithm is
shown in Figure 3(a) and its corresponding matching is shown in Figure 3(b).
The algorithmdescribed in the proof constructs a type 2 network; therefore, by
Theorem 6.3, its running time is O(|V|
1/2
|E|).
6.5 Two Problems on PERT Digraphs 137
Next, let us showthat one can also use the max-owmin-cut theoremto prove
a theorem of Halls [15]. For every AX, let (A) denote the set of vertices
(all in Y) that are connected by an edge to a vertex of A. A matching M, is
called complete if |M| = |X|.
Theorem 6.12 A bipartite graph G has a complete matching M if and only if
for every AX, |(A)| |A|.
Proof: Clearly, if G has a complete matching M, then each x has a unique
mate in Y. Thus, for every AX, |(A)| |A|.
Assume now that G does not have a complete matching. Let S be the set of
labeled vertices (in the Ford and Fulkerson algorithm or the Dinitz algorithm)
upon termination. Clearly, the maximumtotal owis equal to |M|, but |M| <|X|.
Let A=XS. Since all the edges of the type x y are of innite capacity,
(A) S. Also, no vertex of Y \(A) is labeled, since there is no edge from a
labeled vertex to it. We have
(S;

S) = ({s};XA) ((A);{t}).
Since (S;

S) = |M| <|X|, we get

|XA| + |(A)| <|X|,
which implies |(A)| <|A|.
6.5 Two Problems on PERT Digraphs
The ProgramEvaluation and ReviewTechnique, commonly abbreviated PERT,
is a model for project management. APERT digraph is a nite digraph G(V, E)
with the following properties:
(i) There is a vertex s, called the start-vertex, and a vertex t(,=s), called the
termination vertex.
(ii) G has no directed circuits.
(iii) Every vertex v V \ {s, t} is on some directed path from s to t.
A PERT digraph has the following interpretation: Every edge represents a
process. Recall that (v) denotes the set of edges that enter v; (v) denotes
the set of edges that emanate from v. All the processes which are represented
by edges of (s) can be started right away. For every vertex v, the process
represented by edges of (v) can be started when all the processes represented
by edges of (v) are completed.
138 6 Applications of Network Flow Techniques
Our rst problem deals with the question of how soon can the whole project
can be completed; that is, what is the shortest time, from the moment the pro-
cesses represented by (s) are started, until all the processes represented by
(t) are completed. We assume that the resources for running the processes are
unlimited. For this problemto be well dened, let us assume that each e E has
an assigned length l(e), which species the time it takes to execute the process
represented by e. The minimumcompletion time can be found by the following
algorithm:
1. Assign s the label 0 ((s) 0). All other vertices are unlabeled.
2. Find a vertex v such that v is unlabeled and all edges of (v) emanate from
labeled vertices. Assign
(v) max
e(v)
{(u) +l(e)}.
3. If v = t, halt; (t) is the minimum completion time. Otherwise, go to
Step (2).
In Step (2), the existence of a vertex v such that all the edges of (v) emanate
from labeled vertices is guaranteed by conditions (i) and (iii): If no unlabeled
vertex satises the condition, then for every unlabeled vertex v, there is an
incoming edge which emanates from another unlabeled vertex. By repeatedly
tracing back these edges, one nds a directed circuit. Thus, if no such vertex is
found, then we conclude that either (i) or (ii) does not hold.
It is easy to prove by induction on the order of labeling, that (v) is the
minimum time in which all processes represented by the edges of (v) can be
completed.
The time complexity of the algorithm can be kept down to O(|E|) as follows:
For each vertex v we keep count of its incoming edges fromunlabeled vertices;
this count is initially set to d
in
(v); each time a vertex u gets labeled, we use
the list (u) to decrease the count for all v such that uv, accordingly; once
the count of a vertex v reaches 0, it enters a queue of vertices to be labeled.
Once the algorithm terminates, by going back from t to s, via the edge that
determined the label of the vertex, we can trace a longest path froms to t. Such
a path is called critical.
4
Clearly, there may be more than one critical path. If
one wants to shorten the completion time, (t), then on each critical path at
least one edge length must be shortened.
4
The whole process is sometimes called the critical path method (CPM).
6.5 Two Problems on PERT Digraphs 139
s
t
Figure 6.4: Example of a PERT digraph.
Next, we shall consider another problem concerning PERT digraphs, where
there is no reference to edge lengths. Assume that each of the processes, repre-
sented by the edges, uses one processor for its execution. The question is: How
many processors do we need to be sure that no execution will ever be delayed
because of a shortage of processors? We want to avoid such a delay without
relying on the values of l(e)s either because they are unknown or because they
vary from time to time.
Let us solve a minimum ow problem in the network whose digraph is G,
source s, sink t, lower bound b(e) = 1 for all e E and no upper bound (i.e.,
c(e) = for all e E). Condition (6.5) assures the existence of a legal ow
(see Problem 6.5).
For example, consider the PERT digraph of Figure 6.4. The minimum ow
(which in this case is unique) is shown in Figure 6.5(a), where a maximum cut
is shown too.
A set of edges is called concurrent if for no two edges in the set there is a
directed path that passes through both. Now, let T be the set of vertices that
are labeled in the last attempt to nd an augmenting path from t to s. Clearly,
t T and s / T. The set of edges (

T;T) is a maximum cut; there are no edges

in (T;

T), for there is no upper bound on the ow in the edges, and any such
edge would enable to continue the labeling of vertices. Thus, the set (

T;T) is
concurrent.
If S is a set of concurrent edges, then the number of processors required is,
at least |S|. This can be seen by assigning the edges of S a very large length;
and all the others, a short length. Since no directed path leads from one edge
of S to another, they all will be operative simultaneously. This implies that the
number of processors required is at least |(

T;T)|.
140 6 Applications of Network Flow Techniques
s t
s t
3
3
3
3
2
1
2
(a)
2
1
1
1
1
2
2
2
2
2
1
1
1
1
1
1
(b)
Figure 6.5: (a) The minimum ow in the PERT; (b) decompositions
of the ow into the PERT into F directed paths.
However, the owcan be decomposed into F directed paths froms to t, where
F is the minimum total ow, such that every edge is on at least one such (since
f(e) 1 for every e E). This is demonstratedfor our example in Figure 6.5(b).
We can now assign to each processor all the edges of one such path. Each such
processor executes the processes represented by the edges of the path in the
order in which they appear on the path. If one process is assigned to more than
one processor, then one of them executes while the others are idle. It follows
that whenever a process that corresponds to u v is executable (because all
the processes which correspond to (u) have been executed), the processor to
which this process is assigned is available for its execution. Thus, F processors
are sufcient for our purpose.
Since F = |(

T;T)|, by the min-owmax-cut theorem, the number of processors

thus assigned is minimum.
The complexity of this procedure is as follows: We can nd a legal initial ow
in time O(|V| |E|), by tracing for each edge a directed path froms to t via this
edge, and ow through it one unit. This path is found by starting fromthe edge,
and going forward and backwardfromit until s and t are reached. Next, we solve
6.6 Problems 141
a maximum ow problem, fromt to s. Thus, the complexity of the whole pro-
cedure is dominated by the complexity of solving one maximumow problem.
6.6 Problems
Problem 6.1 Let G(V, E) be an acyclic nite digraph.
5
We wish to nd a
minimum number of directed vertex-disjoint paths that cover all the vertices;
that is, every vertex is on exactly one path. The paths may start anywhere and
end anywhere, and their lengths are not restricted in any way. A path may be
of zero length; that is, it may consist of one vertex.
1. Describe an algorithm for achieving this goal, and make it as efcient as
possible. (Hint. Form a network as follows:
V

= {s, t} {x
1
, x
2
, . . . , x
|V|
} {y
1
, y
2
, . . . , y
|V|
}.
E

= {s x
i
| 1 i |V|} {y
i
t | 1 i |V|}
{x
i
y
j
| v
i
v
j
in G}.
The capacity of all edges is 1.
Showthat the minimumnumber of paths that cover V in Gis equal to |V| F,
where F is the maximum total ow of the network.)
2. Is the condition that Gis acyclic essential for the validity of your algorithm?
Explain.
3. Give the best upper bound you can on the time complexity of your algorithm.
Problem 6.2 This problem is similar to Problem 6.1, except that the paths are
not required to be vertex- (or edge-) disjoint.
1. Describe an algorithm for nding a minimum number of covering paths.
(Hint. Form a network as follows:
V

= {s, t} {x
1
, x
2
, . . . , x
|V|
} {y
1
, y
2
, . . . , y
|V|
}.
E

= {s x
i
| 1 i |V|} {y
i
t | 1 i |V|}
{x
i
y
i
| 1 i |V|} {y
i
x
j
| v
i
v
j
in G}.
The lower bound of each x
i
y
i
edge is 1.
The lower bound of all other edges is 0.
The upper bound of all the edges is . Find a minimum ow from s to t.)
5
A digraph is called acyclic if it has no directed circuits.
142 6 Applications of Network Flow Techniques
2. Is the condition that Gis acyclic essential for the validity of your algorithm?
Explain.
3. Give the best upper bound you can on the time complexity of your algorithm.
(Hint. O(|V| |E|) is achievable.)
4. Two vertices u and v are called concurrent if no directed path exists from
u to v or fromv to u. A set of concurrent vertices is such that every two in
the set are concurrent. Prove that the minimum number of paths that cover
the vertices of G is equal to the maximum number of concurrent vertices.
(This is Dilworths Theorem [16].)
Problem 6.3 1. Let G(X, Y, E) be a nite bipartite graph. Describe an efcient
algorithm for nding a minimum set of edges such that each vertex is an
end-vertex of at least one of the edges in the set.
2. Discuss the time complexity of your algorithm.
3. Prove that the size of a minimum set of edges which cover the ver-
tices (as in (1)) is equal to the maximum size of an independent set of
vertices of G.
6
Problem 6.4 1. Prove that if G(X, Y, E) is a complete bipartite graph (i.e., for
every two vertices x X and y Y, there is an edge x y) then the vertex
connectivity of G is
c(G) = min{|X|, |Y|}.
2. Prove that for every k, there exists a graph G such that c(G) k and G has
no Hamilton path. (See Problem 1.4.)
Problem 6.5 Let M be a matching of a bipartite graph. Prove that there exists
a maximummatching M

such that every vertex matched in Mis matched also

in M

.
Problem 6.6 Let G(V, E) be a nite acyclic digraph with exactly one vertex
s for which d
in
(s) = 0 and exactly one vertex t for which d
out
= 0. We say
that the edge a b is greater than the edge c d if and only if there is a
directed path in G fromb to c. A set of edges is called a slice if no edge in it is
greater than another and it is maximal; no other set of edges with this property
contains it. Prove that the following three conditions on a set of edges, P, are
equivalent:
6
A set of vertices of a graph is called independent if there is no edge between two vertices of the
set.
6.6 Problems 143
1. P is a slice.
2. P is an (s, t) edge separator in which no edge is greater than any other.
3. P = (S;

S) for some {s} S V {t} such that (

S;S) =.
Problem 6.7 (The problem of a system of distinct representatives [SDR]). Let
S
1
, S
2
, . . . , S
m
be nite sets. A set {e
1
, e
2
, . . . , e
m
} is called an SDR if for every
1 i m, e
i
S
i
.
1. Describe an efcient algorithm for nding an SDR, if one exists. (Hint.
Dene a bipartite graph and solve a matching problem.)
2. Prove that an SDR exists if and only if the union of any 1 k m of the
sets contains at least k elements.
Problem 6.8 Let
1
and
2
be two partitions of a set of m elements, each
containing exactly r disjoint subsets. We want to nd a set of r elements such
that each of the subsets of
1
and
2
is represented.
1. Describe an efcient algorithm to determine whether there is such a set of r
representatives.
2. State a necessary and sufcient condition for the existence of such a set,
similar to Theorem 6.12.
Problem6.9 Let G(V, E) be a completely connecteddigraph (see Problem1.5);
it is called classiable if V can be partitioned into two nonempty classes, Aand
B, such that all the edges connecting between them are directed from A to B.
Let V = {v
1
, v
2
, . . . , v
n
} where the vertices satisfy
d
out
(v
1
) d
out
(v
2
) . . . d
out
(v
n
).
Prove that G is classiable if and only if there exists a k <n such that
k

i=1
d
out
(v
i
) =
_
k
2
_
.
Problem 6.10 Let S be a set of people such that |S| 4. We assume that
acquaintance is a mutual relationship. Prove that if, in every subset of four
people, there is one who knows all the others, then there is someone in S who
knows everybody.
Problem 6.11 In the acyclic digraph shown in Figure 6.6, there are both AND
vertices (designated by
_
) and OR vertices (designated by
_
). As in a PERT
network, the edges represent processes, and the edge length is the time the
144 6 Applications of Network Flow Techniques
t s
4
5
6
3
3
5
4
6
2
6
2
5
4
3
3
5
6
1
Figure 6.6: Graph for Problem 6.11.
process requires. The processes represented by the edges that emanate from
an AND (OR) vertex can be started when all (at least one of) the incoming
processes are (is) completed. Describe an algorithm for nding the minimum
time fromstart (s) to reach termination (which depends on the type of t). Apply
your algorithmon the given network. What is the complexityof your algorithm?
Problem 6.12 Consider Problem 6.11 on digraphs that are not necessarily
acyclic. Show how to modify the algorithm to solve the problem or conclude
that there is no solution.
Problem 6.13 In a school are n boys and n girls. Each boy knows exactly k
girls (1 k n), and each girl knows exactly k boys. Prove that if knowing"
is mutual, then all the boys and girls can participate in one dance, where every
pair of dancers (a boy and a girl) know each other. Also show that it is always
true that k consecutive dances can be organized so that everyone will dance
once with everyone he or she knows.
Problem6.14 Prove or disprove the following claim: If the vertex connectivity
of a digraph is at least 2, then for every three vertices x, y, and z, there exists a
simple directed path from x to z via y.
Problem 6.15 Let G be a nite undirected graph whose edge connectivity is
at least 2. For each one of the following claims, determine whether it is true or
false. Justify your answer.
1. For every three vertices x, y, and z there exists a path in which no edge
appears more than once, from x to z via y.
2. For every three vertices x, y, and z there exists a path in which no edge
appears more than once, from x to z, which avoids y.
6.6 Problems 145
Bibliography
[1] Even, S., Tarjan, R. E., Network Flow and Testing Connectivity, SIAM J. on
comput., Vol. 4, 1975, pp. 507518.
[2] Menger, K., Zur Aligemeinen Kurventheorie, Fund. Math., Vol. 10, 1927,
pp. 96115.
[3] Dantzig, G. B., Fulkerson, D. R., On the Max-Flow Min-Cut Theorem of
Networks, Linear Inequalities and Related Systems, Annals of Math. Study 38,
Princeton University Press, 1956, pp. 215221.
[4] Hopcroft, J., and Tarjan, R. E., Dividing a Graph into Triconnected Components,
SIAM J. on Comput., Vol. 2, 1973, pp. 135158.
[5] Kleitman, D. J., Methods for Investigating Connectivity of Large Graphs, IEEE
Trans. on Circuit Theory, CT-16, 1969, pp. 232233.
[6] Even, S., Algorithm for Determining whether the Connectivity of a Graph is at
Least k, Siam J. on Comput., Vol. 4, 1977, pp. 393396.
[7] Gomory, R. E., and Hu, T. C., Multi-Terminal Network Flows, J. of SIAM,
Vol. 9, 1961, pp. 551570.
[8] Schnorr, C. P., Multiterminal Network Flow and Connectivity in Unsymmetrical
Networks, Department of Applied Math, University of Frankfurt, October 1977.
[9] Edmonds, J., Edge-Disjoint Branchings, in Combinatorial Algorithms, Courant
Inst. Sci. Symp. 9, R. Rustin, Ed., Algorithmics Press Inc., 1973, pp. 9196.
[10] Lovsz, L., On Two Minimax Theorems in Graph Theory, Journal of Combina-
torial Theory, Series B, Vol. 21:2, 1976, pp. 96103.
[11] Even, S., Garey, M. R., and Tarjan, R. E., A Note on Connectivity and Circuits
in Directed Graphs, unpublished manuscript (1977).
[12] Edmonds, J., Paths, Trees, and Flowers, Canadian J. of Math., Vol. 17, 1965,
pp. 449467.
[13] Even, S., and Kariv, O., An O(n
2.5
) Algorithmfor MaximumMatching in General
Graphs, 16th Annual Symposium on Foundations of Computer Science, IEEE,
1975, pp. 100112.
[14] Hopcroft, J., and Karp, R. M., An O(n
5/2
) Algorithm for Maximum Matching
in Bipartite Graphs, SIAM J. on Comput., 1975, pp. 225231.
[15] Hall, P., On Representation of Subsets, J. London Math. Soc. Vol. 10, 1935,
pp. 2630.
[16] Dilworth, R. P., A Decomposition Theorem for Partially Ordered Sets, Ann.
Math., Vol. 51, 1950, pp. 161166.
7
Planar Graphs
7.1 Bridges and Kuratowskis Theorem
Consider a graph drawn in the plane in such a way that each vertex is repre-
sented by a point; each edge is represented by a continuous line connecting
the two points that represent its end vertices, and no two lines, which represent
edges, share any points, except in their ends. Such a drawing is called a plane
graph. If a graph G has a representation in the plane that is a plane graph then
it is said to be planar.
In this chapter, we shall discuss some of the classical work concerning planar
graphs. The question of efciently testing whether a given nite graph is planar
is discussed in the next chapter.
Let S be a set of vertices of a nonseparable graph G(V, E). Consider the
partition of the set V S into classes, such that two vertices are in the same
class if and only if there is a path connecting them that does not use any vertex
of S. Each such class K denes a component as follows: The component is a
subgraph H(V

, E

), where V

K. In addition, V

includes all the vertices of S

that are connected by an edge to a vertex of K, in G. Also, E

contains all edges

of G that have at least one end-vertex in K. An edge u
e
v, where both u and
v are in S, denes a singular component ({u, v}, {e}). Clearly, two components
share no edges, and the only vertices they can share are elements of S. The
vertices of a component that are elements of S are called its attachments.
In our study we usually use a set S, which is the set of vertices of a simple
circuit C. In this case, we call the components bridges; the edges of C are not
considered bridges.
For example, consider the plane graph over the vertices a, . . . , k shown in
Figure 7.1. The edges of the plane graph are 1, . . . , 18. Let C be the outside
boundary:
146
7.1 Bridges and Kuratowskis Theorem 147
c k
d
b
i
j
a
g
f
e
h
6
3
4
8
10
9
7
12
11
14
15
16
18
17
5
2
1
13
Figure 7.1: Example of a plane graph.
C=a
1
b
2
c
3
. . .
6
g
7
a.
The circuit C partitions V \ C into two classes: {i, j, h} and {k}. The set of
bridges is
({e, g}, {8}),
({i, j, h, a, e, g}, {9, 10, 11, 12, 13, 14}),
({a, e}, {15}), and
({k, b, c, d}, {16, 17, 18}).
The rst and third bridges are singular.
Lemma 7.1 Let B be a bridge, and a
1
, a
2
, a
3
, three of its attachments.
There exists a vertex v, not an attachment, for which there are three vertex-
disjoint paths in B: P
1
(v, a
1
), P
2
(v, a
2
) and P
3
(v, a
3
). (P(a, b) denotes a path
connecting a to b).
Proof: Let a
1
e
1
v
1
, a
2
e
2
v
2
, and a
3
e
3
v
3
be edges of B. If any of the
v
i
s (i = 1, 2, 3) is an attachment, then the corresponding edge is a singular
component and is not part of B. Thus, v
i
K, where K is the class that denes
B. Hence, there is a simple path P

(v
1
, v
2
) that uses vertices of Konly; if v
1
=v
2
,
this path is empty. Also, there is a simple path P

(v
3
, v
1
) that uses vertices of K
only. Let v be the rst vertexof P

(v
3
, v
1
) that is also on P

. Now, let P
1
(v, a
1
) be
the part of P

that leads fromv to v

1
, concatenated with v
1
a
1
; let P
2
(v, a
2
)
be the part of P

that leads fromv to v

2
, concatenated with v
2
a
2
; let P
3
(v, a
3
)
be the part of P

that leads fromv to v

3
, concatenated with v
3
a
3
. It is easy
to see that these paths are disjoint.
148 7 Planar Graphs
Let C be a simple circuit of a nonseparable graph G, and B
1
, B
2
, . . . , B
k
be
the bridges with respect to C. We say that B
i
and B
j
interlace if at least one of
the following conditions holds:
1. There are two attachments of B
i
, a and b, and two attachments of B
j
, c and
d, such that all four are distinct and appear on C in the order a, c, b, d.
2. There are three attachments common to B
i
and B
j
.
For each bridge B
i
, consider the subgraph C+B
i
. If any of these graphs is not
planar, then clearly Gis not planar. Now, assume all these subgraphs are planar.
In every plane realization of G, C outlines a contour that divides the plane into
two disjoint parts: its inside and outside. Each bridge must lie entirely in one
of these parts. Clearly, if two bridges interlace they cannot be on the same side
of C. Thus, in every plane realization of G, the set of bridges is partitioned into
two sets: those drawn inside C and those drawn outside it. No two bridges in
the same set interlace.
Lemma 7.2 If B
1
, B
2
, . . . , B
m
is set of bridges of a nonseparable graph Gwith
respect to a simple circuit C and the following two conditions are satised:
1. for every 1 i n, C+B
i
is planar, and
2. no two bridges interlace,
then C+B
1
+B
2
+. . . +B
m
has a plane realization in which all these bridges
are inside C.
Proof: We shall only outline the proof. As we go clockwise around C there
must be a bridge B
i
such that we encounter all its attachments in some order:
a
1
, a
2
, . . . , a
t
, and no attachment of any other bridge appears between a
1
and a
t
on C. Such a bridge can be found by starting from any attachment a of B
1
and
going around, say, clockwise. If before encountering all the attachments of B
1
we encounter an attachment of another bridge B
i
, then all B
i
s attachments are
between consecutive attachments of B
1
that may also belong to B
i
. We repeat
the same process on B
i
, and so on. Since the number of bridges is nite, and
those discarded will not interfere" with the new ones, the process will yield
the desired bridge.
This observation allows a proof by induction. First B
i
is drawn, and since no
other bridge uses any of the vertices of C between a
1
and a
t
, we can take C

to be the circuit that describes the part of C from a

t
to a
1
, clockwise, and the
boundary of B
i
froma
1
to a
t
to forma simple circuit C

, whose inside is so far

empty. The remaining bridges are also bridges of C

and, clearly, satisfy (7.2)

and (7.2) with respect to C

.
7.1 Bridges and Kuratowskis Theorem 149
Corollary 7.1 Let Gbe a nonseparable graph; and C, a simple circuit in G. In
this case G is planar if and only if the bridges B
1
, B
2
, . . . , B
k
of G, with respect
to C, satisfy the following conditions:
1. For every 1 i k, C+B
i
is planar.
2. The set of bridges can be partitioned into two subsets, such that no two
bridges in the same subset interlace.
Let us introduce the coloring of graphs. Here, we consider only 2-coloring
of vertices. A graph G(V, E) is said to be 2-colorable if V can be partitioned
into V
1
and V
2
in such a way that there is no edge in G with two end vertices
in V
1
(V
2
). (Obviously, a 2-colorable graph is bipartite, and vice versa. It is
customary to use the termbipartite" if the partition is given, and 2-colorable"
if one exists.) The following theorem is due to Knig [1].
Theorem 7.1 A graph G is 2-colorable if and only if it has no odd length
circuits.
Proof: It is easy to see that if a graph has an odd length circuit, then it is not
2-colorable. In order to prove the converse, we may assume that Gis connected,
for if each component is 2-colorable, then the whole graph is 2-colorable.
Let v be any vertex. Let us perform BFS (see Section 1.5.1) starting from v.
There cannot be an edge u win Gif u and wbelong to the same layer, that
is, are the same distance fromv. For if such an edge exists then we can display
an odd length circuit as follows: Let P
1
(v, u) be a shortest path from v to u.
P
2
(v, w) is dened similarly and is of the same length. Let x be the last vertex
that is common to P
1
and P
2
. The part of P
1
fromx to u, and the part of P
2
from
x to ware of equal length, and together with u w, they form an odd length
simple circuit.
Now, we can color all the vertices of even distance fromv with one color and
all the vertices of odd distance fromv with a second color.
We can use the concept of 2-colorability to decide whether the bridges
B
1
, B
2
, . . . , B
k
, with respect to a simple circuit C, can be partitioned into two
pairwise noninterlacing subsets, as follows: Construct a graph G

whose ver-
tices are the bridges B
1
, B
2
, . . . , B
k
, and two vertices are connected by an edge
if and only if the corresponding bridges in G interlace. Now test whether the
graph G

is 2-colorable, by giving one vertex color 1 and using some search

technique, such as DFS or BFS to color alternate vertices with different colors
150 7 Planar Graphs
a
1
b
1
b
2
a
2
a
3
b
3
a
(b) (a)
e b
d c
e
1
e
5
e
2
e
3
e
4
Figure 7.2: The graphs of Kuratowski: (a) K
3,3
and (b) K
5
.
out of {1, 2}. If no contradiction arises, a coloring is obtained and thus, a parti-
tion. If a contradiction occurs, there is no 2-coloring, and therefore no partition
of the bridges; in this case, the graph is nonplanar, by Corollary 7.1.
Let us now introduce the graphs of Kuratowski [2]: K
3,3
and K
5
. They are
shown in Figure 7.2(a) and (b), respectively.
K
5
is completely connected graph of ve vertices, or a clique of ve vertices.
K
3,3
is a completely connected bipartite graph with three vertices on each side.
Lemma 7.3 Neither K
3,3
nor K
5
is planar.
Proof: First, let us consider K
5
, and its circuit C: a b c d
e a. Clearly, there are ve bridges, all singular, corresponding to the edges
e
1
, e
2
, e
3
, e
4
, and e
5
. Let us construct G

as in the preceding discussion. It is

shown in Figure 7.3. For example, the bridge e
1
interlaces with e
5
and e
2
, and
so on. Since G

contains an odd circuit, by Theorem 7.1, it is not 2-colorable,

and the set of bridges of K
5
with respect to C is not partitionable. Thus, by
Corollary 7.1, K
5
is not planar.
In the case of K
3,3
, take C: a
1
b
1
a
2
b
2
a
3
b
3
a
1
.
The bridges a
1
b
2
, a
2
b
3
and a
3
b
1
form a triangle in the
corresponding G

.
Before we take on Kuratowskis theorem, we need a fewmore denitions and
a lemma.
Let G(V, E) be a nite nonseparable plane graph with |V| >2. A face of G is
a maximal part of the plane such that, for every two points x and y in it, there is
a continuous line fromx and y that does not share any point with the realization
of G. The contour of each face is a simple circuit of G; if any of these circuits
is not simple, then G is separable. Each of these circuits is called a window.
7.1 Bridges and Kuratowskis Theorem 151
e
1
e
5
e
2
e
4
e
3
Figure 7.3: G

for the proof of Lemma 7.3.

One of the faces is of innite area. It is called the external face, and its window
is the external window. It is not hard to show that for every window W, there
exists another plane realization, G

, of the same graph, which has the same set

of windows, but in G

, the window W is external. First, draw the graph on a

sphere, maintaining the windows; this can be achieved by projecting each point
of the plane vertically up to the surface on a sphere whose center is in the plane
and its intersecting circle with the plane encircles G. Next, place the sphere on
a plane that is tangent to it, in such a way that a point in the face whose contour
is W is the north pole, that is, furthest from the plane. Project each point P
of the sphere (other than the north pole) to the plane by a straight line that
starts from the north pole and goes through P. The graph is now drawn in the
plane, and W is the external window.
Lemma 7.4 Let G(V, E) be a 2-vertex connected graph with a separating pair
a, b. Let H
1
, H
2
, . . . , H
m
be the components with respect to {a, b}. G is planar
if and only if for every 1 i m, H
i
+ (a
e
b) is planar.
By H
i
+(a
e
b) we mean the subgraph obtained by adding the edge a
e
b
to H
i
.
1
Proof: In each H
i
, there is a path from a to b, or G is not 2-connected. Also
m>1. Thus, for each H
i
we can nd a path P from a to b in one of the other
components. If Gis planar, so is H
i
+P, and therefore H
i
+(a
e
b) is planar.
Now assume that each H
i
+ (a
e
b) is planar. For each of these realization
we can assume that e is on the external window. Thus, a planar realization of
G exists, as demonstrated in Figure 7.4 in the case of m= 3.
1
Since Gis 2-vertex connected, it follows that a, bH
i
. By denition, the component H
i
does not contain e, even if e E. Thus, we add e to H
i
regardless of whether e E.
152 7 Planar Graphs
a
b
b
b
a
a
a
b
H
1
H
1
H
2
H
3
H
2
H
3
Figure 7.4: A demonstration of a planar realization of G, with m=3,
for the proof of Lemma 7.4.
Two graphs are said to be homeomorphic if both can be obtained from the
same graph by subdividing edges, that is, an edge is replaced by a simple path
whose intermediate vertices are all new.
2
Clearly, if two graphs are homeomor-
phic, then either both are planar or both are not. We are now ready to state
Kuratowskis Theorem [2].
Theorem 7.2 A graph G is non-planar if and only if there is a subgraph of G
which is homeomorphic to either K
3,3
or K
5
.
Proof: If G has a subgraph H that is homeomorphic to either K
3,3
or K
5
, then
by Lemma 7.3, H is nonplanar, and therefore G is nonplanar. The converse is
much harder to prove. We prove it by contradiction.
Let G be a graph that is nonplanar and which does not contain a subgraph
that is homeomorphic to one of Kuratowskis graphs; in addition, let us assume
that among such graphs, G has the minimum number of edges.
First, let us show that the vertex connectivity of G is at least 3. Clearly, G is
connected and nonseparable, or the number of its edges is not minimum. Assume
that it has a separating pair {a, b}, and let H
1
, H
2
, . . . , H
m
be the components
2
Two graphs G
1
(V
1
, E
1
) and G
2
(V
2
, E
2
) are said to be isomorphic if there are bijections
f : V
1
V
2
and g: E
1
E
2
such that f(u)
g(e)
f(v) in G
2
for every edge u
e
v in
G
1
. Clearly, G
1
is planar if and only if G
2
is. Thus, we are not interested in the particular
names of the vertices or edges, and we distinguish between graphs only up to isomorphism.
7.1 Bridges and Kuratowskis Theorem 153
with respect to {a, b}, where m>1. By Lemma 7.4, there exists an 1 i m
for which H
i
+(a b) is nonplanar. Clearly H
i
+(a b) does not contain
a subgraph that is homeomorphic to one of Kuratowskis graphs either. This
contradicts the assumption that G has the minimum number of edges.
We now omit an edge a
0
e
0
b
0
from G. The resulting graph, G
0
, is planar.
Since the connectivity of G is at least 3, G
0
is nonseparable. By Theorem 6.7,
there is a simple circuit in G
0
that goes through a
0
and b
0
. Let

G
0
be a plane
realization of G
0
and Cbe a simple circuit that goes througha
0
and b
0
, such that
C encircles the maximum number of faces of all such circuits in all the plane
realizations of G
0
. Note that the selection of

G
0
and C is done simultaneously,
and not successively. Assuming u and v are vertices on C, let C[u, v] be the part
of C going fromu to v, clockwise. C(u, v) is dened similarly, but the vertices
u and v are excluded.
Consider now the external bridges of C in

G
0
. If such a bridge, B, has two
attachments either on C[a
0
, b
0
] or C[b
0
, a
0
], then C is not maximum. To see
this, assume that B has two attachments, a and b, in C[a
0
, b
0
]. There is a simple
path P(a, b) connecting a to b via the edges and vertices of B, which is disjoint
fromC, and is therefore exterior to C. FormC

by adding to P the path C[b, a].

C

goes through a
0
and b
0
, and the interior of C is either completely included
in the interior of C

or in the exterior (see Figure 7.5). In the rst case, C

has
a larger interior than C, in

G
0
. In the later case, we have to nd another plane
realization that is similar to G
0
, but the exterior of C

is now the interior. In

either case, the maximality of the choice of

G
0
and Cis contradicted. Thus, each
P(a,b)
P(a,b)
a
0
a
b
(b) (a)
b
0
a
0
a
b
b
0
Figure 7.5: (a) The interior of the circuit Cis contained in the interior
of C

; (b) the interior of the circuit C is contained in the exterior

of C

.
154 7 Planar Graphs
external bridge has at most one attachment in C[a
0
, b
0
] and C[b
0
, a
0
]. Since an
external bridge must have at least two attachments, it follows that neither a
0
nor b
0
can be an attachment of an external bridge.
It follows that each external bridge has at most two attachments. Since the
vertex connectivity is at least 3, we conclude that all external bridges are singular
(i.e., consist of a single edge).
Finally, there is at least one such external singular bridge; otherwise, one
could draw the edge e
0
outside C, to yield a planar realization of G. Every
external singular bridge interlaces with a
0
e
0
b
0
since the attachments are on
C(a
0
, b
0
) and C(b
0
, a
0
).
Similarly, there must be an internal bridge, B

, which prevents the drawing of

e
0
inside, and which cannot be transferred outside; that is, B

interlaces with an
external singular bridge, say, a
1
e
1
b
1
. The situation is schematically shown
in Figure 7.6. We divide the argument to two cases according to whether B

has any attachment other than a

0
, b
0
, a
1
, b
1
.
Case 1: B

has an attachment a
2
other than a
0
, b
0
, a
1
, b
1
. Without loss of
generality we may assume that a
2
is on C(a
1
, a
0
). Since B

prevents the drawing

of e
0
, it must have an attachment on C(a
0
, b
0
). Since B

interlaces e
1
, it must
have an attachment on C(b
1
, a
1
).
Case 1.1: B

has an attachment b
2
on C(b
1
, b
0
). In B

, there is a path P
connecting a
2
with b
2
. The situation is shown in Figure 7.7. The subgraph of
a
1
b
1
b
0
a
0
e
1
B
Figure 7.6: The bridge B

in the proof of Theorem 7.2.

7.1 Bridges and Kuratowskis Theorem 155
a
1
a
2
a
0
e
1
b
1
P
b
2
b
0
e
0
Figure 7.7: Case 1.1 in the proof of Theorem 7.2.
G shown in Figure 7.7 is homeomorphic to K
3,3
, where a
1
, a
0
, and b
2
play the
role of the upper vertices of Figure 7.2(a), and a
2
, b
1
and b
0
play the role of
the lower vertices.
Case 1.2: B

has no attachments on C(b

1
, b
0
). Thus, B

has one attachment b

2
on C(a
0
, b
1
); that is, it may be b
1
but not a
0
. Also, B

has an attachment b

2
on
C[b
0
, a
1
). By Lemma 7.1, there exists a vertex v and three vertex-disjoint paths
in B

: P
1
(v, a
2
), P
2
(v, b

2
), and P
3
(v, b

2
). The situation is shown in Figure 7.8.
If we erase from the subgraph of G, shown in Figure 7.8 the edges in the path
C[b
1
, b
0
] and all its intermediate vertices, the resulting subgraph is homeomor-
phic to K
3,3
: Vertices a
2
, b

2
, and b

2
play the role of the upper vertices; and
a
0
, a
1
, and v, the lower vertices.
Case 2: B

has no attachments other than a

0
, b
0
, a
1
, b
1
. In this case all four
must be attachments; for, if a
0
or b
0
are not, then B

and e
1
do not interlace; if
a
1
or b
1
are not, then B

does not prevent the drawing of e

0
.
Case 2.1: There is a vertex v, in B

, from which there are four disjoint

paths in B

: P
1
(v, a
0
), P
2
(v, b
0
), P
3
(v, a
1
), and P
4
(v, b
1
). This case is shown
in Figure 7.9, and the shown subgraph is clearly homeomorphic to K
5
.
Case 2.2: No vertex as in Case 2.1 exists. Let P
0
(a
0
, b
0
) and P
1
(a
1
, b
1
) be two
simple paths in B

. Let c
1
be the rst vertex on P
1
that is common with P
0
, and
let c
2
be the last on P
1
that is common with P
0
. We use only the rst part, A,
156 7 Planar Graphs
e
1
a
0
b
2
b
2
b
1
a
2
a
1
b
0
v
e
0
P
2
P
1
P
3
Figure 7.8: Case 1.2 in the proof of Theorem 7.2.
e
1
a
0
b
1
a
1
b
0
e
0
v
Figure 7.9: Case 2.1 in the proof of Theorem 7.2.
of P
1
, connecting a
1
and c
1
; and the last part, B, connecting c
2
with b
1
. The
pertaining subgraph of G is now shown in Figure 7.10, and is homeomorphic
to K
3,3
after the edges in C[a
0
, b
1
] and C[b
0
, a
1
] and all their intermediate
vertices are erased: Vertices a
0
, b
1
, and c
1
play the role of the upper vertices;
and b
0
, a
1
, and c
2
, the lower. (If c
1
is closer to a
0
than c
2
, then we erase C[a
1
, a
0
]
and C[b
1
, b
0
], instead, and the upper vertices are a
0
, a
1
and c
2
.)
Kuratowskis theorem provides a necessary and sufcient condition for a
graph to be planar. However, it does not yield an efcient algorithmfor planarity
7.2 Equivalence 157
e
1
a
0
b
1
a
1
b
0
e
0
c
1
c
2
A
B
P
0
Figure 7.10: Case 2.2. in the proof of Theorem 7.2.
testing. The obvious procedure, that of trying for all subsets of ve vertices to
see whether there are ten vertex disjoint paths connecting all pairs, or for every
pairs of three and three vertices whether there are nine paths, suffers from two
shortcomings. First there are
_
|V|
_
choices of ve-sets and 1/2
_
|V|
3
_

_
|V|3
3
_
choices of three, and three vertices; this alone is O(|V|
6
). But what is worse,
we have no efcient way to look for the disjoint paths; this problem may, in
fact, be exponential.
Fortunately, there are O(|E|) tests, as we shall see in the next chapter, for
testing whether a given graph is planar.
7.2 Equivalence
Let

G
1
and

G
2
be two plane realizations of the graph G. We say that

G
1
and

G
2
are equivalent if every window of one of them is also a window in the other.
G may be 2-connected and have nonequivalent plane realization; for example,
see Figure 7.11
Let us restrict our discussion to planar nite graphs with no parallel edges
and no self-loops. Our aim is to show that if the vertex connectivity of G, c(G),
is at least three then the plane realization of G is unique up to equivalence.
Lemma 7.5 A planar nonseparable graph G is 2-connected if there is a plane
realization of it,

G, and one of its windows has more than one bridge.
Proof: If C is a window of

G with more than one bridge, then all Cs bridges
are external. Therefore, no two bridges interlace. As in the rst paragraph of the
158 7 Planar Graphs
1
2
4
6 7
5
3
8
1
2 3
4
5
6
8
7
Figure 7.11: Two nonequivalent plane realizations of the same graph.
proof of Lemma 7.2, there exists a bridge B whose attachments can be ordered
a
1
, a
2
, . . . a
t
, and no attachments of any other bridge appear on C(a
1
, a
t
). It
is easy to see that {a
1
, a
t
} is a separating pair; it separates the vertices of B
and C(a
1
, a
t
) fromthe set of vertices of all other bridges and C(a
t
, a
1
), where
neither set can be empty since G has no parallel edges.
Theorem7.3 If Gis a plane graph with no parallel edges and no self-loops and
if its vertex connectivity, c(G), is at least 3, then every two plane realizations
of G are equivalent.
Proof: Assume that Ghas two plane realizations

G
1
and

G
2
that are not equiv-
alent. Without loss of generality we may assume that there is a windowCin

G
1
that is not a windowin

G
2
. Therefore, Chas at least two bridges; one interior and
one exterior. By Lemma 7.5, and since C is a window in

G
1
, G is 2-connected.
A contradiction, since c(G) 3.
7.3 Eulers Theorem
The following theorem is due to Euler.
Theorem 7.4 Let G(V, E) be a nonempty connected plane graph. The number
of faces, f, satises
|V| +f |E| = 2. (7.1)
7.4 Duality 159
Proof: By induction on |E|. If |E| =0, then Gconsists of one vertex and there is
one face, and Equation 7.1 holds. Assume that the theorem holds for all graphs
with m = |E|. Let G(V, E) be a connected plane graph with m+1 edges. If
G contains a circuit, then we can remove one of its edges. The resulting plane
graph is connected and has medges and, therefore, by the inductive hypothesis,
satises Equation 7.1. Adding back the edge increases the number of faces by
one and the number of edges by one, and thus Equation 7.1 is maintained. If G
contains no circuits, then it is a tree. By Corollary 2.1, it has at least two leaves.
Removing a leaf and its incident edge yields a connected graph with one less
edge and one less vertex, which satises Equation 7.1. Therefore, G satises
Equation 7.1 too.
The theorem implies that all connected plane graphs with |V| vertices and |E|
edges have the same number of faces. One can draw many conclusions from
the theorem. Some of them are the following:
Corollary 7.2 If G(V, E) is a connected plane graph with no parallel edges,
no self-loops and |V| >2, then
|E| 3|V| 6. (7.2)
Proof: Since there are no parallel edges, every window consists of at least
three edges. Each edge appears on the windows of two faces, or twice on the
window of one face. Thus, 3 f 2 |E|. By Equation 7.1, |E| = |V| +f 2.
Thus, |E| |V| +2/3|E| 2, and (7.2) follows.
Corollary 7.3 Every connected plane graph with no parallel edges and no
self-loops has at least one vertex whose degree is 5 or less.
Proof: Assume the contrary; that is, the degree of everyvertexis at least 6. Thus,
2 |E| 6 |V|; note that each edge is counted in each of its two end-vertices.
This contradicts (7.2).
7.4 Duality
Let G(V, E) be a nite undirected and connected graph. A set KE is called a
cutset if it is a minimal separating set of edges; that is, the removal of K from
G interrupts its connectivity, but no proper subset of K does it. It is easy to
see that a cutset separates Ginto two connected components: Consider rst the
removal of K{e}, where e K. Gremains connected. Nowremove e. Clearly,
G breaks into two components.
160 7 Planar Graphs
x
6
x
1
x
5
x
3
x
4 x
2
y
3
y
2
y
4
y
5
y
1
y
6
z
5
z
2
z
6
z
1
z
4
z
3
Figure 7.12: (a) A connected graph G
1
; (b) G
2
, a dual of G
1
; (c) G
3
,
also a dual of G
1
.
The graph G
2
(V
2
, E
2
) is said to be the dual of a connected graph G
1
(V
1
, E
1
)
if there is a 1 1 correspondence f : E
1
E
2
, such that a set of edges S forms
a simple circuit in G
1
if and only if f(S) (the corresponding set of edges in
G
2
) forms a cutset in G
2
. Consider the graph G
1
shown in Figure 7.12 (a). G
2
shown in Figure 7.12(b) is a dual of G
1
, but so is G
3
, shown in Figure 7.12(c),
as the reader can verify by considering all (six) simple circuits of G
1
and all
cutsets of G
2
, or G
3
.
Acontraction of an edge x
e
yof a graph G(V, E) is the following operation:
Delete the edge e, and merge x with y. The new contracted graph, G

, has one
less edge and one less vertex, if x ,=y. Clearly, if Gis connected, so is G

. Also,
graph G

is a contraction of G if by repeated contractions we can construct G

from G.
Lemma 7.6 If a connected graph G
1
has a dual and G

1
is a connected subgraph
of G
1
, then G

1
has a dual.
Proof: We can get G

1
from G
1
by a sequence of two kinds of deletions:
7.4 Duality 161
1. A deletion of an edge e of the present graph, which is not in G

1
, and whose
deletion does not interrupt the connectivity of the present graph.
2. Adeletion of a leaf of the present graph, which is not a vertex of G

1
, together
with its incident edge.
We want to showthat each of the resulting graphs, starting with G
1
and ending
with G

1
, has a dual.
Let Gbe one of these graphs, except the last, and its dual be G
d
. First consider
a deletion of type (7.4), of an edge e. Construct f(e) in G
d
, to get G
dc
. If C
is a simple circuit in Ge, then clearly it cannot use e, and therefore it is a
circuit in G too. The set of edges of C is denoted by S. Thus, f(S) is a cutset of
G
d
, and it does not include f(e). Thus, the end vertices of f(e) are in the same
component of G
d
with respect to f(S). It follows that f(S) is a cutset of G
dc
too. If K is a cutset of G
dc
, then it is a cutset of G
d
too. Thus, f
1
(K) form a
simple circuit C

in G. However, f(e) is not in K, and therefore e is not in C

.
Hence, C

is a simple circuit of Ge.

Next, consider a deletion of type (7.4) of a leaf v and its incident edge e.
Clearly, e, plays no role in a circuit. Thus, f(e) cannot be a part of a cutset in G
d
.
Hence, f(e) is a self-loop. The deletion of v and e form G, and the contraction
of f(e) in G
d
(which effectively, only deletes f(e) from G
d
), does not change
the sets of simple circuits in G and cutsets in G
d
, and the correspondence is
maintained.
Lemma 7.7 Let Gbe a connected graph and e
1
, e
2
be two of its edges, neither
of which is a self loop. If for every cutset, either both edges are in it or both are
not, then e
1
and e
2
are parallel edges.
Proof: If e
1
and e
2
are not parallel, then there exists a spanning tree which
includes both. (Such a tree can be found by contracting both edges and nding
a spanning tree of the contracted graph.) The edge e
1
separates the tree into
two connected components whose sets of vertices are S and

S. The set of edges
between S and

S in G is a cutset that includes e
1
and does not include e
2
. A
contradiction.
Lemma 7.8 Let G be a connected graph with a dual G
d
, and let f be the
1 1 correspondence of their edges. If u
e
1
x
1
e
2
x
2
e
3
. . . x
l1
e
l
v is a
simple path or circuit in G such that x
1
, x
2
, . . . , x
l1
are all of degree two, then
f(e
1
), f(e
2
), . . . , f(e
l
) are parallel edges in G
d
.
162 7 Planar Graphs
Proof: Every circuit of G that contains one e
i
, 1 i l, contains all the rest.
Thus, in G
d
, if one edge, f(e
i
), is in a cutset then all the rest are. By Lemma. 7.7,
they are parallel edges.
Lemma 7.9 If a connected graph G
1
has a dual, and G
2
is homeomorphic to
G
1
, then G
2
has a dual.
Proof: If an edge x
e
y of G
1
is replaced in G
2
by a path x
e
1
v
1
e
2
v
2
. . .
e
l
y, then in the dual f(e) is replaced by parallel edges:
f(e
1
), f(e
2
), . . . , f(e
l
). If a path of G
1
is replaced in G
2
by an edge e, then the
edges of the dual that correspond to the edges of the path are all parallel (by
Lemma 7.8) and can be replaced by a single edge f(e). It is easy to see that
every circuit which uses an edge e, when it is replaced by a path, will use all
the edges of the path instead, while in the dual, every cutset that uses f(e) will
use all the parallel edges which replace it. Thus, the correspondence of circuits
and cutsets is maintained.
Theorem 7.5 A (connected) graph has a dual if and only if it is planar.
Proof: Assume G
1
(V
1
, E
1
) is a planar graph and

G
1
is a plane realization of
it. Choose a point p
i
in each face F
i
of

G
1
. Let V
2
be the set of these points.
Since G
1
is connected, the boundary of every face is a circuit (not necessarily
simple) which we call its window. Let W
i
be the window of F
i
, and assume
it consists of l edges. We can nd l curved lines, all starting in p
i
, but no two
share any other point, such that each of the lines ends on one of the l edges
of W
i
, one line per edge. If a separating edge
3
e appears on W
i
, then clearly
it appears twice. In this case there will be two lines from p
i
hitting e from
both directions. These two lines can be made to end in the same point on e,
thus creating a closed curve that goes through p
i
and crosses e. If e is not a
separating edge, then it appears on two windows, say, W
i
and W
j
. In this case,
we can make the line fromp
i
to e meet e at the same point as does the line from
p
j
to e, to forma line connecting p
i
with p
j
, which crosses e. None of the set of
lines thus formed crosses another, and we have one per edge of

G
1
. Now dene

G
2
(V
2
, E
2
) as follows: The set of lines connecting the p
i
s is a representation
of E
2
. The 1 1 correspondence f : E
1
E
2
is dened as follows: f(e) is the
edge of

G
2
which crosses e. Clearly,

G
2
is a plane graph that is a realization
of a graph G
2
(V
2
, E
2
). It remains to show that there is a 1 1 correspondence
of the simple circuits of G
1
to the cutsets of G
2
. The construction described
3
An edge whose removal from G
1
interrupts its connectivity.
7.4 Duality 163
Figure 7.13: For the proof of Theorem 7.5.
above is demonstrated in Figure 7.13, where

G
1
is shown in solid lines, and

G
2
is shown in dashed lines.
Let C be a simple circuit of G
1
. Clearly, in

G
1
, C describes a simple closed
curve in the plane. There must be at least one vertex of

G
2
inside this circuit,
since at least one edge of

G
2
crosses the circuit, and it crosses exactly once.
The same argument also applies to the outside. This implies that f(S), where S
is the set of edges of C, forms a separating set of G
2
. Let us postpone the proof
of the minimality of f(S) for a little while.
Now let K be a cutset of G
2
. Let T and

T be the sets of vertices of the two
components of G
2
formed by the deletion of K. The set of faces of

G
1
that
correspond to the vertices of T, forms a continuous region, but not the whole
plane. The minimality of K implies that the boundary of this region is a circuit
in G
1
, whose edges correspond to K. Thus, we have shown that every simple
circuit of G
1
corresponds to a separating set of G
2
, and every cutset of G
2
corresponds to a circuit of G
1
.
Now let us handle the minimality. If S is the set of edges of a simple circuit C
of G
1
and f(S) is not a cutset of G
2
, then there is a proper subset of f(S) which
is a cutset, say K. Therefore, f
1
(K) is a circuit of G
1
. However, f
1
(K) S. A
contradiction to the assumption that C is simple. The proof that if K is a cutset
then f
1
(K) is a simple circuit is similar.
This completes the proof that the connected planar graph has a dual. We turn
to the proof that a nonplanar graph has no dual. Here, we follow the proof of
Parson [3].
First, let us show that neither K
3,3
nor K
5
have a dual. In K
3,3
, the shortest
circuit is of length four, and for every two edges there exists a simple circuit
that one but does not use the other. Thus, in its dual no cutset consists of less
164 7 Planar Graphs
than four edges and there are no parallel edges. Therefore, the degree of each
vertex is at least 4 and there must be at least ve vertices. The number of edges
is therefore at least (5 4/2) = 10, while K
3,3
has 9 edges. Thus, K
3,3
has no
dual. In the case of K
5
, it has ten edges, ten simple circuits of length three and no
shorter circuits, and for every two edges there is a simple circuit that uses one
but not the other. Thus, the dual must have ten edges, ten cutsets of three edges,
no cutset of lesser size and therefore the degree of every vertex is at least three.
Also, there are no parallel edges in the dual. If the dual has ve vertices, then it is
K
5
itself (ten edges and no parallel edges), but K
5
has no cutsets of three edges. If
the dual has seven vertices or more, then it has at least eleven edges (,7 3/2|).
Thus, the dual must have six vertices. Since it has ten clusters of three edges,
there is one that separates S from

S, where neither consists of a single vertex.
If |S| = 2, then it contains a vertex whose degree is 2. If |S| = |

S| = 3, then the
maximum number of edges in the dual is nine. Thus, K
5
has no dual.
Now, if G is a nonplanar graph with a dual, then by Kuratowskis theorem
(Theorem7.2) it contains a subgraph G

that is homeomorphic to either K

3,3
or
K
5
. By Lemma 7.6, G

also has a dual. By Lemma 7.9, either K

3,3
or K
5
has a
dual. A contradiction. Thus, no nonplanar graph has a dual.
Theorem7.5provides another necessary and sufcient condition for planarity,
in addition to Kuratowskis theorem. However, neither has been shown to be
useful for testing planarity.
There are many facts about duality that we have not discussed. Among them
are the following:
1. If G
d
is a dual of G, then G is a dual of G
d
.
2. A 3-connected planar graph has a unique dual.
The interested reader can nd additional information and references in the
books of Harary [4], Ore [5], and Wilson [6].
7.5 Problems
Problem 7.1 The purpose of this problem is to prove a variation of Kura-
towskis theorem.
1. Prove that if G is a connected graph and v
1
, v
2
, v
3
are three ver-
tices, then there exists a vertex v and three (vertex) disjoint paths
P
1
(v, v
1
), P
2
(v, v
2
), P
3
(v, v
3
), one of which may be empty.
2. Prove that if G is a connected graph and S is a set of four vertices, then
either there is a vertex v with four disjoint paths to the members of S or there
7.5 Problems 165
are two vertices u and v such that two members of S are connected to u by
paths, two to v, and u is connected to v; all ve paths are vertex-disjoint.
3. Show that if a graph is contractible to K
3,3
, then it has a subgraph
homeomorphic to K
3,3
.
4. Show that if a graph is contractible to K
5
, then either it has a subgraph that
is homeomorphic to K
5
or it has a subgraph that is contractible to K
3,3
.
5. Prove the theorem: A graph is nonplanar if and only if it contains a subgraph
that is contractible to K
3,3
or K
5
.
Problem 7.2 Show a graph that is nonplanar but not contractible to either K
3,3
or K
5
. Does it contradict the result of Problem 7.1(7.1)?
Problem7.3 Use Kuratowskis theoremto prove that the Petersengraph, shown
in Figure 7.14, is nonplanar.
Figure 7.14: Peterson graph.
Problem7.4 Is the graph depicted in Figure 7.15 planar? Justify your answer.
Figure 7.15: Is this graph planar?
166 7 Planar Graphs
Problem 7.5 A plane graph is called triangular if each of its windows is a
triangle. Let N
i
be the number of vertices whose degree is i.
1. Prove that every plane graph with three or more vertices that has no self-
loops and no parallel edges can be made triangular by adding new edges
without creating any self-loops or parallel edges. (This process is called
triangulation.)
2. Let G be a triangular plane graph as in part 1, with |V| >3. Prove that
N
1
=N
2
= 0.
3. Let G be a triangular plane graph as in part 1. Prove that
12 = 3 N
3
+2 N
4
+N
5
N
7
2 N
8
3 N
9
4 N
10
. . .
4. Prove that in a graph, in part 1, if there are no vertices of degree 3 or 4, then
there are at least tweleve vertices of degree 5.
5. Prove that if G is a planar graph with no self-loops or parallel edges, and
|V| >3, then it has at least 4 vertices with degrees less than 6, and if N
3
=
N
4
= 0, then N
5
12.
6. Prove that if the vertex connectivity, c(G), of a graph G(V, E) is at least ve,
then |V| 12.
7. Prove that if G is planar, then c(G) <6.
Problem 7.6 Prove that if G
1
(V
1
, E
1
) and G
2
(V
2
, E
2
) are homeomorphic, then
|E
1
| |V
1
| = |E
2
| |V
2
|.
Problem 7.7 Show a triangular plane graph without parallel edges that is not
Hamiltonian.
Problem7.8 Let Gbe a plane graph. The plane graph G

, which is constructed
by the procedure in the rst part of the proof of Theorem 7.5, is called its
geometric dual.
Prove that if G

is the geometric dual of G, then G is the geometric dual

of G

.
7.5 Problems 167
Bibliography
[1] Knig, D., Theorie der endlichen und unendlichen Graphen. Leipzig, 1936.
Reprinted Chelsea, 1950.
[2] Kuratowski, K., Sur le problme des Courbes Gauches en Topologie, Fund. Math.,
Vol. 15, 1930, pp. 217283.
[3] Parson, T. D., On Planar Graphs, Am. Math. Monthly, Vol. 78, No. 2, 1971,
pp. 176178.
[4] Harary, F., Graph Theory, Addison Wesley, 1969.
[5] Ore, O., The Four-Color Problem, Academic Press, 1976.
[6] Wilson, R. J., Introduction to Graph Theory, Longman, 1972.
8
Testing Graph Planarity
8.1 Introduction
There are two known planarity testing algorithms that have been shown to be
realizable in a way that achieves linear time (O(|V|)). The idea in both is to
follow the decisions to be made during the planar construction of the graph,
piece by piece, as to the relative location of the various pieces. The construction
is not carried out explicitly because there are difculties, such as the crowding
of elements into a relatively small portion of the area allocated, which, as yet,
we do not knowto avoid. Also, an explicit drawing of the graph is not necessary,
as we shall see, to decide whether such a drawing is possible. We shall imagine
that such a realization is being carried out, but will only decide where the various
pieces are laid relative to each other, and not their exact shape. Such decisions
may change later to make a place for later additions of pieces. In both cases, it
was shown that the algorithmterminates within (O(|V|)) steps, and if it fails to
nd a realization, then none exists.
The rst algorithmstarts by nding a simple circuit and adding toit one simple
path at a time. Each such new path connects two old vertices via new edges
and vertices. (Whole pieces are sometimes ipped over, around some line).
Thus, we call it the path addition algorithm. The basic ideas were suggested
by various authors, such as Auslander and Parter [1] and Goldstein [2], but the
algorithm in its present form, both from the graph-theoretic point of view, and
complexity point of view, is the contribution of Hopcroft and Tarjan [3]. They
were rst to show that planarity testing can be done in linear time.
The second algorithm adds one vertex in each step. Previously drawn edges
incident to this vertex are connected to it, and new edges incident to it are
drawn and their other endpoints are left unconnected. (Here, too, sometimes
whole pieces have to be ipped around or permuted). The algorithm is due to
Lempel, Even, and Cederbaum [4]. It consists of two parts. The rst part was
168
8.2 The Path Addition Algorithm of Hopcroft and Tarjan 169
shown to be linearly realizable by Even and Tarjan [5]; the second part was
shown to be linearly realizable by Booth and Leuker [6]. We call this algorithm
the vertex addition algorithm.
Each of these algorithms can be divided into its graph-theoretic part and
its data structures and their manipulation. The algorithms are fairly complex
and a complete description and proof would require a much more elaborate
exposition. Thus, since this is a book on graphs, and not on programming, I
have chosen to describe in full the graph-theoretic aspects of both algorithms
and to only briey describe the details of the data manipulation techniques.
An attempt is made to convince the reader that the algorithms work, but in
order to see the details which make it linear one will have to refer to the papers
mentioned above.
Throughout this chapter, for reasons explained in Chapter 7, we shall assume
that G(V, E) is a nite undirected graph with no parallel edges and no self-
loops. We shall also assume that G is nonseparable. The rst thing that we can
do is check whether |E| 3 |V| 6. By Corollary 7.1, if this condition does
not hold, then G is nonplanar. Thus, we can restrict our algorithm to the cases
where |E| =O(|V|).
8.2 The Path Addition Algorithm of Hopcroft and Tarjan
The algorithmstarts with a DFS of G. We assume that Gis nonseparable. Thus,
we drop from the DFS the steps for testing nonseparability. However, we still
need the lowpoint function, to be denoted now L1(v). In addition, we need the
second lowpoint function, L2(v), to be dened as follows: Let S(v) be the set
of values k(u) of vertices u reachable from descendants of v by a single back
edge. Clearly, L1(v) = min{{k(v)} S(v)}.
Dene
L2(v) = min{{k(v)} [S(v) {L1(v)}]}.
Let us now rewrite the DFS in order to compute these functions (we denote
an assignment statement, x gets the value of y by x :=y):
1. Mark all the edges unused. For every v V, let k(v) := 0. Let i := 0 and
v :=s (the vertex s is where we choose to start the DFS).
2. i :=i +1, k(v) :=i, L1(v) :=i, L2(v) :=i.
3. If v has no unused incident edges, go to Step (8.2).
4. Choose an unused incident edge v
e
u. Mark e used. If k(u) ,= 0 then
do the following:
If k(u) <L1(v), then L2(v) :=L1(v), L1(v) :=k(u).
170 8 Testing Graph Planarity
If k(u) >L1(v), then L2(v) := min{L2(v), k(u)}.
Go to Step (8.2).
Otherwise (k(u) = 0), let f(u) :=v, v :=u and go to Step (8.2).
5. If k(v) =1, halt.
6. (k(v) >1; we backtrack).
If L1(v) < L1(f(v)) then L2(f(v)) := min{L2(v), L1(f(v))}, L1(f(v))
:=L1(v).
If L1(v) =L1(f(v)), then L2(f(v)) := min{L2(v), L2(f(v))}.
Otherwise (L1(v) >L1(f(v))), L2(f(v)) := min{L1(v), L2(f(v))}.
v :=f(v).
Go to Step (8.2).
Fromnowon we refer to the vertices by their k(v) number; that is, we change
the name of v to k(v).
Let A(v) be the adjacency list of v; that is, the list of edges incident fromv.
We remind the reader that after the DFS each of the edges is directed; the tree
edges are directed from low to high, and the back edges are directed from high
to low. Thus, each edge appears once in the adjacency lists. Now, we want to
reorder the adjacency lists, but rst, we must dene an order on the edges. Let
the value (e) of an edge u
e
v be dened as follows:
(e) =

2 v if u
e
v is a back edge.
2 L1(v) if u
e
v is a tree edge and L2(v) u.
2 L1(v) +1 if u
e
v is a tree edge and L2(v) <u.
Next, we order the edges in each adjacency list to be in nondecreasing order
with respect to . [This can be done on O(|V|) time as follows. First, compute for
each edge its value. Prepare 2 |V| +1 buckets, numbered 1, 2, . . . , 2 |V| +1.
Put each e into the (e) bucket. Now, empty the buckets in order, rst bucket
number 1, then 2 and so on, putting the edges taken out into the proper new
adjacency lists, in the order that they are taken out of the buckets.]
The new adjacency lists are now used, in a second run of a DFS algorithm, to
generate the paths, one by one. In this second DFS, vertices are not renumbered,
and there is no need to recompute f(v), L1(v) or L2(v). The tree remains the
same, although the vertices may not be visited in the same order. The paths
nding algorithm is as follows:
1. Mark all edges unused and let v := 1.
2. Start the circuit C; its rst vertex is 1.
8.2 The Path Addition Algorithm of Hopcroft and Tarjan 171
3. Let the rst unused edge in A(v) be v
e
u.
If u ,= 1, then mark e as used, add u to C, update v := u, and repeat
Step (3). Otherwise, (u = 1), C is closed. Output C.
4. If v =1, halt.
5. If every edge in A(v) is used, then update v :=f(v) and go to Step (4).
6. Start a path P with v as its rst vertex.
7. Let v
e
u be the rst unused edge in A(v).
If e is a back edge (u <v), terminate P with u, output P, and go to (5).
8. Otherwise, (e is a tree edge). Add u to P, update v :=u, and go to Step (7).
Lemma 8.1 The paths nding algorithm nds rst a circuit C that consists of
a path from 1 (the root) to some vertex v, via tree edges, and a back edge from
v to 1.
Proof: Let 1 u be the rst edge of the tree to be traced (in the rst appli-
cation of Step (3)). We assume that G is nonseparable, and |V| > 2. Thus, by
Lemma 3.7, this edge is the only tree edge out of 1, and u=2. Also, 2 has some
descendants, other than itself. Clearly, 2 3 is a tree edge. By Lemma 3.5,
L1(3) <2, that is, L1(3) =1. Thus, L1(2) =1. The reordering of the adjacency
lists assures that the rst path to be chosen out of 1 will lead back to 1 as
claimed.
Lemma 8.2 Each generated path P is simple, and it contains exactly two ver-
tices in common with previously generated paths; they are the rst vertex, f,
and the last, l.
Proof: The edge scanning during the paths nding algorithm is in a DFS
manner, in accord with the structure of the tree (but not necessarily in the same
scanning order of vertices). Thus, a path starts from some (old) vertex f, goes
along tree edges, via intermediate vertices which are all new, and ends with
a back edge which leads to l. Since back edges always lead to ancestors, l is
old. Also, by the reordering of the adjacency lists and the assumption that G is
nonseparable, l must be lower than f. Thus, the path is simple.
Let S
v
denote the set of descendants of v, including v itself.
Lemma 8.3 Let f and l be the rst and last vertices of a generated path P and
f v be its rst edge.
1. If v ,=l then L1(v) =l.
2. l is the lowest vertex reachable from S
f
via a back edge that has not been
used in any path yet.
172 8 Testing Graph Planarity
Proof: Let us partition S
f
{f} into two parts, and , as follows. Adescendant
ubelongs to if and only if when the construction of P begins, we have already
backtracked from u. Clearly, f and all the back edges out of have been
scanned already. Let u be the lowest vertex reachable via an unused back edge
from. Clearly, the rst remaining (unused) edge of A(f) is the beginning of a
directed path to u, which is either an unused back edge fromf to u or a path of
tree edges, via vertices of , followed by a single back edge to u. Thus, u =l,
and the lemma follows.
Lemma 8.4 Let P
1
and P
2
be two generated paths whose rst and last ver-
tices are f
1
, l
1
and f
2
, l
2
, respectively. If P
1
is generated before P
2
and f
2
is a
descendant of f
1
then l
1
l
2
.
Proof: The lemma follows immediately from Lemma 8.3.
So far, we have not made any use of L2(v). However, Lemma 8.5 relies on it.
Lemma 8.5 Let P
1
: f
e
1
v
1
. . . l and P
2
: f
e
2
v
2
. . . l be
two generated paths, where P
1
is generated before P
2
. If v
1
,=l and L2(v
1
) <f,
then v
2
,=l and L2(v2) <f.
Proof: By the denition of (e), (e
1
) = 2 l +1. If v
2
= l or L2(v2) f,
then (e
2
) =2 l, and e
2
should appear in A(f) before e
1
. A contradiction.
Let Cbe 1 v
1
v
2
. . . v
n
l. Clearly 1 <v
1
<v
2
<. . . <v
n
.
Consider now the bridges of G with respect to C.
Lemma 8.6 Let B be a nonsingular bridge of G with respect to C, whose
highest attachment is v
i
. There exists a tree edge v
i
e
u that belongs to B,
and all other edges of B with endpoints on C are back edges.
Proof: The lemma follows from the fact that the paths nding algorithm is
a DFS. First C is found. We then backtrack from a vertex v
j
only if all its
descendants have been scanned. No internal part of B can be scanned before
we backtrack into v
i
. There must be a tree edge v
i
u, where u belongs to
B, for the following reasons. If all the edges of B, incident to v
i
are back edges,
they all must come fromdescendants or go to ancestors of v
i
(see Lemma 3.4).
An edge from v
i
to one of its ancestors (which must be on C) is a singular
bridge and is not part of B. An edge from a descendant w of v
i
implies that w
cannot be in B, for it has been scanned already, and we have observed that no
internal part of B can be scanned before we backtrack into v
i
. If any other edge
v
k
x of B is also a tree edge, then, by denition of a bridge, there is a path
8.2 The Path Addition Algorithm of Hopcroft and Tarjan 173
connecting u and x that is vertex-disjoint from C. Along this path there is at
least one edge that contradicts Lemma 3.4.
Corollary 8.1 Once a bridge B is entered, it is completely traced before
it is left.
Proof: By Lemma 8.6, there is only one edge through which B is entered.
Since eventually the whole graph is scanned, and no edge is scanned twice in
the same direction, the corollary follows.
Assuming that C and the bridges explored from vertices higher than v
i
have
already been explored and drawn in the plane. Lemma 8.7 provides a test for
whether the next generated path could be drawn inside; the answer is negative,
even if the path itself can be placed inside, but it is already clear that the whole
bridge to which it belongs cannot be placed there.
Lemma 8.7 Let v
i
u
1
u
2
. . . u
l
(=v
j
) be the rst path, P, of
the bridge B to be generated by the paths nding algorithm. P can be drawn
inside (outside) C if there is no back edge w v
k
drawn inside (outside)
for which j < k < i. If there is such an edge, then B cannot be drawn inside
(outside).
Proof: The sufciency is immediate. If there is no back edge drawn inside C,
that enters the path, v
j
v
j+1
v
i
in one of its internal vertices,
then there cannot be any inside edge incident to these vertices, since bridges to
be explored fromvertices lower than v
i
have not been scanned yet. Thus, P can
be drawn inside if it is placed sufciently close to C.
Now, assume there is a back edge wv
k
, drawn inside, for which j <k <i.
Let P

be the directed path from v

p
to v
k
whose last edge is the back edge
wv
k
. Clearly, v
p
is on C and p i; P

is not necessarily generated in one

piece by the path-nding algorithm, if it is not the rst path to be generated in
the bridge B

to which it belongs.
Case 1: p > i. The bridges B and B

interlace by part (i) of the denition of

interlacement. Thus, B cannot be drawn inside.
Case 2: p =i. Let P

be the rst path of B

to be generated, P

: v
i
x
1

x
2
. . . v
q
. By Lemma 8.4, q j, since B

is explored before B.
Case 2.1: q<j. Since v
i
and v
j
are attachments of B, v
k
and v
q
are attachments
of B

and q < j < k < i, the two bridges interlace. Thus, B cannot be drawn
inside.
174 8 Testing Graph Planarity
Case 2.2: q = j. P

cannot consist of a single edge, for in this case, it is a

singular bridge and v
k
is not one of its attachments. Also, L2(x
1
) <v
k
. Thus,
L2(x
1
) <v
i
. By Lemma 8.5, u
1
,=v
j
and L
2
(u
1
) <v
i
. This implies that B and
B

interlace by either part (i) or part (ii) of the denition of interlacement, and
B cannot be drawn inside.

The algorithm assumes that the rst path of the new bridge B is drawn inside
C. Now, we use the results of Corollary 7.1, Theorem 7.1 and the discussion
that follows it, to decide whether the part of the graph explored so far is planar,
assuming that C+B is planar. By Lemma 8.7, we nd which previous bridges
interlace with B. The part of the graph explored so far is planar if and only if
the set of its bridges can be partitioned into two sets such that no two bridges in
the same set interlace. If the answer is negative, the algorithm halts, declaring
that graph nonplanar. If the answer is positive, we still have to check whether
C+B is planar.
Let the rst path of B be P: v
i
u
1
u
2
. . . v
j
. We now have a
circuit C

consisting of C[v
j
, v
i
] and P. The rest of Cis an outside path P

, with
respect to C

, and it consists of C[v

i
, 1] and C[1, v
j
]. The graph B+C[v
j
, v
i
]
may have bridges with respect to C

, but none of them has all its attachments

on C[v
j
, v
i
], for such a bridge is also a bridge of Gwith respect to C, and is not
a part of B.
Thus, no bridge of C

, with attachments of C(v

j
, v
i
) may be drawn outside
C

, since it interlaces with P

. We conclude that C+B is planar if and only

if B+C[v
j
, v
i
] is planar and its bridges can be partitioned into an inside set
and an outside set so that the outside set contains no bridge with attachments
in C(v
j
, v
i
). The planarity of B+C[v
j
, v
i
] is tested by applying the algorithm
recursively, using the established vertex numbering L
1
, L
2
, f functions and order-
ing of the adjacency lists. If B+C[v
j
, v
i
] is found to be nonplanar, clearly G
is nonplanar. If it is found to be planar, we check whether all its bridges with
attachments in C(v
j
, v
i
) can be placed inside. If so, C+B is planar; if not,
G is nonplanar.
Hopcroft and Tarjan devised a simple tool for deciding whether the set of
bridges can be partitionedproperly. It consists of three stacks
1
,
2
, and
3
. Of
these
1
contains in nondecreasing order (the greatest entry on top) the vertices
of C(1, v
i
) (where v
i
is the last vertex of C into which we have backtraced),
which are attachments of bridges placed inside C in the present partition of the
bridges. Avertex may appear on
1
several times if there are several back edges
into it from inside bridges.
2
is similarly dened for outside bridges. Both
1
8.2 The Path Addition Algorithm of Hopcroft and Tarjan 175
and
2
are doubly linked lists, in order to enable switching over of complete
sections from one of them to the other, investing, per section, time bounded by
some constant.
3
consists of pairs of pointers to entries in
1
and
2
. Its task
will be described shortly.
Let S be a maximal set of bridges, explored up to now, such that a decision
for any one of these bridges as to the side it is in, implies a decision for all the
rest. (S corresponds to a connected component of the graph of bridges, as in the
discussion following Theorem 7.1.) Let the set of entries in
1
and
2
, which
correspond to back edges from the bridges of S, be called a block.
Lemma 8.8 Let K be a block, whose highest element is v
h
and lowest element
is v
l
. If v
p
is an entry in
1
or
2
, then v
p
belongs to K if v
l
<v
p
<v
h
.
Proof: We prove the lemma by induction on the order in which the bridges are
explored. At rst, both
1
and
2
are empty and the lemma is vacuously true.
The lemma trivially holds after one bridge is explored.
Assume the lemma is true up to the exploration of the rst path P of the
new bridge B, where P : v
i
. . . v
j
. If there is no vertex v
k
on
1
or

2
such that v
j
<v
k
<v
i
then clearly the attachments of B (in C(1, v
i
)) form
a mew block (assuming C+B is planar), and the lemma holds. However, if
there are vertices of
1
or
2
in between v
j
and v
i
, then by Lemma 8.7, the
bridges they are attachments of all interlace with B. Thus, the old blocks which
these attachments belong to, must now be merged into one new block with
the attachments of B (in C(1, v
i
)). Now, let v
l
be the lowest vertex of the new
block and v
h
be the highest. Clearly, v
l
was the lowest vertex of some old block
whose highest vertexwas v

h
, and v

h
>v
j
. Thus, by the inductive hypothesis, no
attachment of another block could be between v
l
and v

h
and therefore cannot
be in this region after the merger. Also, all the attachments in between v
j
and
v
h
are in the new block since they are attachments of bridges which interlace
with B. Thus, all the entries of
1
or
2
which are in between v
l
and v
h
belong
to the new block.
Corollary 8.2 The entries of one block appear consecutively on
1
(
2
).
Thus, when we consider the rst path P : v
i
. . . v
j
of the new bridge
B in order to decide whether it can be drawn inside, we check the top entries
t
1
and t
2
of
1
and
2
, respectively.
If v
j
t
1
and v
j
t
2
, then no merger is necessary; the attachments of B (in
C(1, v
i
)) are entered as a block (if C+B is found to be planar) on top of
1
.
If v
j
<t
1
and v
j
t
2
, then we rst join all the blocks for which their highest
entry is higher than v
j
. To this end there is no need to check
2
, since v
j
t
2
;
176 8 Testing Graph Planarity
however, several blocks still may merge. Next, switch the sections of the new
block, that is, the section on
1
exchanges places with the section on
2
. Finally,
place the attachments of B in nondecreasing order on top of
1
; these entries
join the new block.
If v
j
t
1
and v
j
<t
2
, then again we join all blocks whose highest element is
greater than v
j
; only the sections on
2
need to be checked. The attachments
of B join the same block are placed on top of
1
.
If v
j
<t
1
and v
j
<t
2
, then all blocks whose highest element is greater than v
j
are joined into one new block. As we join the blocks, we examine them one by
one. If the highest entry in the section on
1
is higher than v
j
, then we switch the
sections. If it is still higher, then we halt and declare the graph nonplanar. If all
these switches succeed, then the merger is completed by adding the attachments
of B on top of
1
.
In order to handle the sections switching, the examination of their tops and
mergers efciently, we use a third stack,
3
. It consists of pairs of pointers, one
pair (x, y) per block; x points to the lowest entry of the section in
1
; and y, to
the lowest entry of the section in
2
. (If the section is empty, then the pointers
value is 0). Two adjacent blocks are joined by simply discarding the pair of the
top one. When several blocks have to be joined together upon the drawing of
the rst path of a new bridge, only the pair of the lowest of these blocks need
remain, except when one of its entries is 0. In this case, the lowest nonzero entry
of the pairs above it on the same side, if any such entry exists, takes its place.
When we enter a recursive step, a special end of stack marker E is placed on
top of
2
, and the three stacks are used as in the main algorithm. If the recursive
step ends successfully, we rst attempt to switch sections for each of the blocks
with a nonempty section on
2
, above the top most E. If we fail to expose E,
then C+B is nonplanar and we halt. Otherwise, all the blocks created during
the recursion are joined to the one which includes v
j
(the end vertex of the rst
path of B). The exposed E, on top of
2
, is removed, and we continue with the
previous level of recursion. When we backtrack into a vertex v
i
, all occurrences
of v
i
are removed from the top of
1
and
2
, together with pairs of pointers of

3
which point to removed entries on both
1
and
2
. (Technically, instead of
pointing to an occurrence of v
i
, we point to 0 and pairs (0, 0) are removed).
Theorem 8.1 The complexity of the path addition algorithm is O(|V|).
Proof: As in the closing remarks of Section 8.1, we can assume |E| =O(|V|).
The DFSand the reordering of the adjacency lists have been shown to be O(|V|).
Each edge in the paths nding algorithm is used again, once in each direction.
The total number of entries in the stacks
1
and
2
is bounded by the number of
8.3 Computing an st-Numbering 177
back edges (|E| |V| +1), and is therefore O(|V|). After each section switching
the number of blocks is reduced by one; thus the total work invested in section
switchings is O(|V|).
8.3 Computing an st-Numbering
In this section, we dene an st-numbering and describe a linear time algorithm
to compute it. Thus numbering is necessary for the vertex addition algorithm,
for testing planarity, of Lempel, Even, and Cederbaum.
Given any edge s t of a nonseparable graph G(V, E), a 1-1 function
g : V {1, 2, . . . , |V|} is called an st-numbering if the following conditions are
satised:
1. g(s) = 1.
2. g(t) = |V| (=n).
3. For every v V {s, t} there are adjacent vertices u and wsuch that g(u) <
g(v) <g(w).
Lempel, Even, and Cederbaum showed that, for every nonseparable graph
and every edge s t, there exists an st-numbering. The algorithm described
here, following the work of Even and Tarjan [5], achieves this goal in linear
time.
The algorithm starts with a DFS whose rst vertex is t and rst edge is
t s. (i.e., k(t) = 1 and k(s) = 2). This DFS computes for each vertex v, its
DFS number, k(v), its father, f(v), and its lowpoint L(v) and distinguishes tree
edges fromback edges. This information is used in the paths-nding algorithm
to be described next, which is different from the one used in the path addition
algorithm.
Initially, s, t, and the edge connecting them are marked old, and all the
other edges and vertices are marked new. The path-nding algorithm starts
from a given vertex v and nds a path from it. This path may be directed from
v or into v.
1. If there is a new back edge v
e
w (in this case k(w) < k(v)), then do
the following:
Mark e old.
The path is v
e
w.
Halt.
2. If there is a new tree edge v
e
w(in this case k(w) >k(v)), then do the
following:
178 8 Testing Graph Planarity
Trace a path whose rst edge is e, and fromthere it follows a path that dened
L(w), that is, it goes up the tree and ends with a back edge into a vertex u
such that k(u) =L(w). All vertices and edges on the path are marked old.
Halt.
3. If there is a new back edge w
e
v (in this case k(w) > k(v)), then do
the following:
Start the path with e (going backwards on it) and continue backwards via
tree edges until you encounter an old vertex. All vertices and edges on the
path are marked old. Halt.
4. (All edges incident to v are old). The path produced is empty. Halt.
Lemma 8.9 If the path nding algorithm is always applied from an old
vertex v ,=t, then all the ancestors of an old vertex are old too.
Proof: By induction on the number of applications of the path nding algo-
rithm. Clearly, before the rst application, the only ancestor of s is t, and it is
old. Assuming the statement is true up to the present application, it is easy to
see that if any of the four steps is applicable, the statement continues to hold
after its application.
Corollary 8.3 If G is nonseparable and under the condition of Lemma 8.9,
each application of the path-ndingalgorithmfroman old vertex v produces
a path, through new vertices and edges, to another old vertex, or in case
all edges incident to v are old, it returns the empty path.
Proof: The only case which requires a discussion is when case (2) of the path
nding algorithm is applied. Since G is nonseparable, by Lemma 3.5, L(w) <
k(v). Thus, the path ends bellow v, in one of its ancestors. By Lemma 8.9,
this ancestor is old.
We are now ready to present the algorithmwhich produces an st-numbering.
It uses a stack S that initially contains only t and s, s on top of t.
1. i 1.
2. Let v be the top vertex on S. Remove v from S. If v =t then g(t) i and
halt.
3. (v ,= t) Apply the path nding algorithm to v. If the path is empty, then
g(v) i, i i +1 and go to Step (2).
4. (The path is not empty) Let the path be v u
1
u
2
. . . u
l
w.
Put u
l
, u
l1
, . . . , u
2
, u
1
, v on S in this order (v comes out on top), and go to
Step (2).
8.4 The Vertex Addition Algorithm of Lempel, Even, and Cederbaum 179
Theorem 8.2 The algorithm above computes an st-numbering for every
nonseparable graph G(V, E).
Proof: First, we make a few observations about the algorithm:
1. No vertex ever appears in two or more places on S at the same time.
2. Once a vertex v is placed on S, nothing under v receives a number until v
does.
3. A vertex is permanently removed from S only after all its incident edges
become old.
Next, we want to show that each vertex v is placed on S before t is removed.
Since t and s are placed on S initially, the statement needs to be proved for
v ,= s, t only. Since G is nonseparable, there exists a simple path from s to v
which does not pass through t (see Theorem 6.7 part (??)). Let this path be
s =u
1
u
2
u
l
= v. Let m be the rst index such that u
m
is not
placed on S. Since u
m1
is placed on S, t can be removed only after u
m1
(fact (ii)), and u
m1
is removed only after all its incident edges are old (fact
(ii)). Thus, u
m
must be placed on S before t is removed.
It remains to be shown that the algorithm computes an st-numbering.
Since each vertex is placed on S and eventually is removed, each vertex v
gets a number g(v). Clearly, g(s) = 1, for it is the rst to be removed. After
each assignment i is incremented. Thus, g(t) = |V|. Every other vertex v is
placed on S, for the rst time, as an intermediate vertex on a path. Thus, there is
an adjacent vertex stored below it, and an adjacent vertex stored above it. The
one above it (by fact (ii)) gets a lower number and the one below it, a higher
number.
It is easy to see that the whole algorithm is of time complexity O(|E|): First,
the DFS is O(|E|). The total time spent on path nding is also O(|E|) since
no edge is used more than once. The total number of operations in the main
algorithm is bounded also by O(|E|) because the number of stack insertions is
exactly |E| +1.
8.4 The Vertex Addition Algorithm of Lempel, Even, and Cederbaum
In this section, we assume that G(V, E) is a nonseparable graph whose vertices
are st-numbered. Fromnowon, we shall refer to the vertices by their st-number.
Thus, V = {1, 2, . . . , n}. Also, the edges are now directed from low to high.
A (graphical) source of a digraph is a vertex v such that d
in
(v) =0; a (graph-
ical) sink is a vertex v such that d
out
(v) = 0. (Do not confuse with the source
180 8 Testing Graph Planarity
2
(b) (a)
3
2
1
3
4
5 6
1 4
6 5
6 5
4
Figure 8.1: (a) An example of a digraph; (b) B
3
of the digraph from (a).
and sink of a network. The source of a network is not necessarily a (graphical)
source, etc.) Clearly, vertex 1 is a source of G and vertex n is a sink. Further-
more, due to the st-numbering, no other vertex is either a source or a sink. Let
V
k
= {1, 2, . . . , k}. Then, G
k
(V
k
, E
k
) is the digraph induced by V
k
, that is, E
k
consists of all the edges of G whose endpoints are both in V
k
.
If Gis planar, let

Gbe a plane realization of G. It contains a plane realization
of G
k
. The following simple lemma reveals the reason for the st-numbering.
Lemma 8.10 If

G
k
is a plane realization of G
k
contained in a plane digraph

G, then all the edges and vertices of

G

G
k
are drawn in one face of

G
k
.
Proof: Assume that a face F of

G
k
contains vertices of V V
k
. Since all the
vertices on Fs window are lower than the vertices in F, the highest vertex in F
must be a sink. Since G has only one sink, only one such face is possible.
Let B
k
be the following digraph. G
k
is a subgraph of B
k
. In addition, B
k
contains all the edges of G which emanate from vertices of V
k
and enter in
G, vertices of V V
k
. These edges are called virtual edges, and the leaves
they enter in B
k
are called virtual vertices. These vertices are labeled as their
counterparts in G, but they are kept separate; that is, there may be several virtual
vertices with the same label, each with exactly one entering edge. For example,
consider the digraph shown in Figure 8.1(a). B
3
of this digraph is shown in
Figure 8.1(b).
By Lemma 8.10, we can assume that if G is planar, then there exists a plane
realization of B
k
in which all the virtual edges are drawn in the outside face.
Furthermore, since 1
e
n is always an edge in G, if k < n then vertex 1 is
on the outside window and one of the virtual edges is e. In this case, we can
8.4 The Vertex Addition Algorithm of Lempel, Even, and Cederbaum 181
1
4 4 5 5 6 6
2
3
Figure 8.2: A bush form of B
3
, of our example.
draw B
k
in the following form: Vertex 1 is drawn at the bottom level. All the
virtual vertices appear on one horizontal line. The remaining vertices of G
k
are
drawn in such a way that vertices with higher names are drawn higher. Such a
realization is called a bush form. A bush form of B
3
, of our example, is shown
in Figure 8.2.
In fact, Lemma 8.10 implies that, if Gis planar, then there exists a bush form
of B
k
such that all the virtual vertices with labels k +1 appear next to each
other on the horizontal line.
The algorithm proceeds by successively drawing B
1
, B
2
, . . . , B
n1
and G.
If in the realization of B
k
all the virtual vertices labeled k+1 are next to each
other, then it is easy to draw B
k+1
: One joins all the virtual vertices labeled
k+1 into one vertex and pulls it down from the horizontal line. Now all the
edges of G that emanate from k +1 are added, and their other endpoints are
labeled properly and placed in an arbitrary order on the horizontal line, in the
space evacuated by the former virtual vertices labeled k+1.
However, a difculty arises. Indeed, the discussion up to now guarantees that
if G is planar, then there exists a sequence of bush forms such that each one is
grown from the previous one. But since we do not have a plane realization
of G, we may put the virtual vertices, out of k +1, in a wrong order. It is
necessary to show that this does not matter; namely, by simple transformations
it will be possible later to correct the mistake.
Lemma 8.11 Assume v is a separation vertex of B
k
. If v >1, then exactly one
component of B
k
, with respect to v, contains vertices lower than v.
182 8 Testing Graph Planarity
Note that here we ignore the direction of the edges, and the lemma is actually
concerned with the undirected underlying graph of B
k
.
Proof: The st-numbering implies that for every vertex u there exists a path
from 1 to u such that all the vertices on the path are less than u. Thus, if u <v
then there is a path from 1 to u that does not pass through v. Therefore, 1 and
u are in the same component.
Lemma 8.11 implies that a separation vertex v of B
k
is the lowest vertex in
each of the components, except the one which contains 1 (in case v >1). Each
of these components is a sub-bush; that is, it has the same structure as a bush
form, except that its lowest vertex is v rather than 1. These subbushes can be
permuted around v in any of the p! permutations, in case the number of these
subbushes is p. In addition, each of the subbushes can be ipped over. These
transformations maintain the bush form. It is our purpose to show that if

B
1
k
and

B
2
k
are bush forms of a planar B
k
, then through a sequence of permutations
and ippings, one can change

B
1
k
into a

B
3
k
such that the virtual vertices of B
k
appear in

B
2
k
and

B
3
k
in the same order.
For efciency reasons, to be become clear to those readers who will study the
implementation through P Q-trees, we assume that when a subbush is ipped,
smaller subbushes of other separation vertices of the component are not ipped
by this action. For example, consider the bush formshown in Figure 8.3(a). The
bush form of Figure 8.3(b) is achieved by permuting about 1, Figure 8.3(c) by
ipping about 1 and Figure 8.3(d) by ipping about 2.
Lemma 8.12 Let H be a maximal nonseparable component of B
k
and
y
1
, y
2
, . . . , y
m
be the vertices of H that are also endpoints of edges of B
k
H.
In every bush form

B
k
, all the ys are on the outside window of H and in the
same order, except that the orientation may be reversed.
Proof: Since

B
k
is a bush form, all the ys are on the outside face of H. Assume
there are two bush forms

B
1
k
and

B
2
k
in which the realizations of H are

H
1
and

H
2
, respectively. If the ys do not appear in the same order on the outside
windows of

H
1
and

H
2
, then there are two ys, y
i
and y
j
which are next to each
other in

H
1
but not in

H
2
(see Fig. 8.4). Therefore, in

H
2
, there are two other
ys, y
k
and y
l
which interpose between y
i
and y
j
on the two paths between
them on the outside window of

H
2
. However, from

H
1
we see that there are
two paths, P
1
[y
i
, y
j
] and P
2
[y
k
, y
l
], which are completely disjoint. These two
paths cannot exist simultaneously in

H
2
. A contradiction.
8.4 The Vertex Addition Algorithm of Lempel, Even, and Cederbaum 183
7
2
3
1
4
5
6
9 8 9 8 9
2
3
1
5
7
6
4
9
5
8
4
9
3
2
1
9 8
6
7
9
(a) (b)
(c) (d)
8 9 9 8 8 9 8 9
9 8
9 8 9
5
8
4
9
7
6
3
2
1
8 9
Figure 8.3: (a) Abush form; (b) a bush formobtained by permuting (a)
about vertex 1; (c) a bush formobtained by ipping (a) about vertex 1;
(d) a bush form obtained by ipping (c) about vertex 2.
y
i
y
i
y
k
y
j
y
l
y
j

H
1

H
2
Figure 8.4: Proof of Lemma 8.12.
Theorem 8.3 If

B
1
k
and

B
2
k
are bush forms of the same B
k
, then there exists a
sequence of permutations and ippings that transforms

B
1
k
into

B
3
k
, such that
in

B
2
k
and

B
3
k
, the virtual vertices appear in the same order.
Proof: By induction on the size of bush or subbush forms.
1
Clearly, if each
of the two (sub-)bushes consists of only one vertex and one virtual vertex,
then the statement is trivial. Let v be the lowest vertex in the (sub-)bushes
1
Size the number of vertices.
184 8 Testing Graph Planarity

B
1
and

B
2
of the same B. If v is a separation vertex, then the components of
B appear as subbushes in

B
1
and

B
2
. If they are not in the same order, it is
possible, by permuting them in

B
1
, to put them in the same order as in

B
2
. By
the inductive hypothesis, there is a sequence of permutations and ippings that
will change each subbush of

B
1
to have the same order of its virtual vertice as
in its counterpart in

B
2
, and therefore the theorem follows.
If v is not a separating vertex then let Hbe the maximal nonseparable compo-
nent of B which contains v. In

B
1
(

B
2
) there is a planar realization

H
1
(

H
2
) of H.
The vertices y
1
, y
2
, . . . , y
m
of H, which are also endpoints of edges of BH,
by Lemma 8.12, must appear on the outside windowof Hin the same order, up
to orientation. If the orientation of the ys in

H
1
is opposite to that of

H
2
, ip
the (sub-)bush

B
1
about v. Now, each of the ys is the lowest vertex of some
subbush of B, and these subbushes appear in (the new)

B
1
and

B
2
in the same
order. By the inductive hypothesis, each of these subbushes can be transformed
by a sequence of permutations and ippings to have its virtual vertices in the
same order as its counterpart in

B
2
.
Corollary 8.4 If G is planar and

B
k
is a bush form of B
k
, then there exists a
sequence of permutations and ippings which transforms

B
k
into a

B

k
in which
all the virtual vertices labeled k+1 appear together on the horizontal line.
It remains to be shown how one decides which permutation or ipping to
apply, how to represent the necessary information, without actually drawing
bush forms, and how to do it all efciently. Lempel, Even, and Cederbaum
described a method that uses a representation of the pertinent information by
proper expressions. However, a better representation was suggested by Booth
and Leuker. They invented a data structure, called PQ-trees, through which the
algorithmcan be run in linear time. PQ-trees were used to solve other problems
of interest; see [6].
I shall not describe PQ-trees in detail. The description in [6] is long (30 pages)
although not hard to follow. It involves ideas of data structure manipulation, but
almost no graph theory. The following is a brief and incomplete description.
A PQ-tree is a directed ordered tree, with three types of vertices: P-vertices,
Q-vertices and leaves. Each P-vertex or Q-vertex, has at least one son. The
sons of a P-vertex, which in our application represents a separating vertex v of
B
k
, may be permuted into any new order. Each sons and its subtree, represents
a subbush. A Q-vertex represents a maximal nonseparable component, and its
sons, which represent the ys, may not be permuted, but their order can be
reversed. The leaves represent the virtual vertices.
8.5 Problems 185
The attempt to gather all the leaves labeled k +1 into an unbroken run, is
done from sons to fathers, starting with the leaves labeled k +1. Through a
technique of template matching, vertices are modied while the k +1 labeled
leaves are bunched together. Only the smallest subtree containing all the k+1
labeled leaves is scanned. All these leaves, if successfully gathered, are merged
into one P-vertex, and its sons represent the virtual edges out of k +1. This
procedure is repeated until k+1 =n.
8.5 Problems
Problem 8.1 Demonstrate the path-addition algorithm on the Peterson graph
(see Problem 7.3). Show the data for all the steps: the DFS for numbering
the vertices, dening the tree and computing L1 and L2. The function on
the edges. The sorting of the adjacency lists. Use the path-nding algorithm
in the new DFS to decompose the graph into C and a sequence of paths. Use

1
,
2
,
3
and end-of-stack markers to carry out all the recursive steps up to
planarity decision.
Problem8.2 Repeat the path-addition planarity test, as in Problem8.1, for the
graph depicted in Figure 8.5.
Figure 8.5: A graph for Problem 8.2.
Problem 8.3 Demonstrate the vertex-addition planarity test on the Peterson
graph. Show the steps for the DFS, the st-numbering, and the sequence of bush
forms.
Problem 8.4 Repeat the vertex-addition planarity test for the graph of
Problem 8.2.
Problem8.5 Showthat if a graph is nonplanar, then a subgraph homeomorphic
to one of the Kuratowskis graphs can be found in O(|V|
2
). (Hints: Only O(|V|)
edges need to be considered. Delete edges if their deletion does not make the
graph planar. What is left?)
186 8 Testing Graph Planarity
Bibliography
[1] Auslander, L., and Parter, S. V., On Embedding Graphs in the Plane, J. Math. and
Mech., Vol. 10, No. 3, May 1961, pp. 517523.
[2] Goldstein, A. J., An Efcient and Constructive Algorithm for Testing Whether a
Graph Can Be Embedded in a Plane, Graph and Combinatorics Conf., Contract
No. NONR 1858-(21), Ofce of Naval Research Logistics Proj., Dept. of Math.,
Princeton Univ., May 1618, 1963, 2 pp.
[3] Hopcroft, J., and Tarjan, R., Efcient Planarity Testing, JACM, Vol. 21, No. 4,
Oct. 1974, pp. 549568.
[4] Lempel, A., Even, S., and Cederbaum, I., An Algorithm for Planarity Test-
ing of Graphs, Theory of Graphs, International Symposium, Rome, July, 1966.
P. Rosenstiehl, Ed., Gordon and Breach, N.Y. 1967, pp. 215232.
[5] Even, S., and Tarjan, R. E., Computing and st-numbering, Th. Comp. Sci., Vol. 2,
1976, pp. 339344.
[6] Booth, K. S., and Lueker, G. S., Testing for the Consecutive Ones Property, Interval
Graphs, and Graph Planarity Using PQ-tree Algorithms, J. of Comp. and Sys.
Sciences, Vol. 13, 1976, pp. 335379.
Index
acyclic digraph, 141
adjacency list, 170
adjacency matrix, 3
attachments, 146
augment procedure, 89
augmenting path, 87
backward cut, 86
BFS, see breadth-rst search
biconnected, 127
bipartite graph, 135
Breadth-rst search, 11
bridge, 63
bridges, 146
interlace, 148
capacity, 85
capacity function, 85
capacity of a cut, 87
residual capacity, 94
circuit, 2
simple, 2
classiable, 143
clique, 44, 63
code, 65
word length, 65
alphabet, 65
characteristic sum, 67
characteristic sum condition, 67
exhaustive, 83
letters, 65
message, 65
prex, 65, 68
sufx, 65
tail, 65
uniquely decipherable, 65
word, 65
coloring, 149
component, 146
nonseparable, 52
singular component, 146
strong, 58
superstructure, 54
concurrent set of edges, 139
connectivity
vertex, 121, 124
contraction, 160
critical path, 138
cut, 33, 86
cutset, 159
Dantzig algorithm, 27
De Bruijn sequence, 9
depth-rst search, 46
directed, 57
lowpoint, 53
number, 49
tree, 49
back edges, 51
tree edges, 51
DFS, see depth-rst search
digraph, see graph
Dijkstra algorithm, 14
Dinitz algorithm, 94
187
188 Index
edge, 1
antiparallel, 3
edge array, 4
parallel, 3
self-loop, 3
edge rule, 102
edge separator, 130
Euler, 6
ow
absorb, 114
blocking, 95
maximal, 95
ow function, 85
Floyd algorithm, 21
Ford algorithm, 17
Ford-Fulkerson algorithm, 87
forward cut, 86
geometric dual, 166
graph, 1
2-colorable, 149
acyclic, 24
bipartite, 149
circuit-free, 29
connected, 2
directed (digraph), 2
arbitrated, 38
dual, 160
Euler graph, 6
planar, 146
sparse, 4
strongly connected, 3, 58
triangular, 166
triangulation, 166
underlying graph, 7
homeomorphic, 152
incidence list, 4
independent set, 142
isomorphic, 152
label
labeling procedure, 88
leaf, 31
marriage problem, 135
matching, 135
complete, 137
max-ow min-cut theorem, 94
Mengers Theorem, 122
network, 85
0-1, 117
0-1 type 1, 119
0-1 type 2, 120
auxiliary network, 103
layered network, 95
secondary network, 94
nonseparable, 126
path, 2
directed, 3
directed Euler path, 7
Euler path, 6
Hamilton, 23
shortest, 11
simple, 2
PERT digraph, 137
Petersen graph, 165
planar
PQ-trees, 184
block, 175
bush form, 181
equivalent, 157
face, 150
external, 151
path addition, 168
second lowpoint function, 169
sink, 179
source, 179
st-numbering, 177
vetex addition, 169
virtual edges, 180
virtual vertices, 180
window, 150
plane graph, 146
Prim algorithm, 32
root, 37
separating edge, 162
separation vertex, 52
sink, 85
auxiliary sink, 103
slice, 142
source, 85
auxiliary source, 103
spanning tree, 32
subgraph, 31
Index 189
topological sorting, 24
total ow, 85
Trmauxs algorithm, 46
tree, 29
directed, 37
useful
backwardly useful, 88
forwardly useful, 88
useful edge, 88
vertex, 1
degree, 1
in-degree, 3
out-degree, 3
sink, 58
source, 58
vertex array, 4
vertex separator, 122, 130
Warshalls Algorithm, 26
Warshalls algorithm, 26