Graph Theory and Algorithms

M. Ashraf Iqbal

ii

Copyright c 2010 by M Ashraf Iqbal

All rights reserved.

ISBN . . .

. . . Publications

To my grand daughter Nariman

iv

Contents
1 Introduction 1.1 1.2 1.3 1.4 1.5 1.6 1.7 Why a new book? . . . . . . . . . . . . . . . . . . . . . . . . . What do we emphasize? . . . . . . . . . . . . . . . . . . . . . How the book is organized? . . . . . . . . . . . . . . . . . . . How is the book designed? . . . . . . . . . . . . . . . . . . . . Some salient features of the book . . . . . . . . . . . . . . . . 1 2 3 6 9 9

What tools do we use? . . . . . . . . . . . . . . . . . . . . . . 10 A Possible Sequence . . . . . . . . . . . . . . . . . . . . . . . 10 19 31

2 Basic Definitions in Graph Theory and Algorithms 3 Problems, Models & Graphs: Why Study Graph Theory? 3.1 3.2 3.3

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 Reducing One Problem into Another . . . . . . . . . . . . . . 32 The Satisfiability Problem in Logic Circuits . . . . . . . . . . 34 3.3.1 3.3.2 3.3.3 Reducing a 3-SAT Problem into an Independent Set Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 35 Reducing the 3-CNF Satisfiability Problem into the 3DNF Satisfiability Problem . . . . . . . . . . . . . . . 39 Reducing the 3-CNF Satisfiability Problem into another graph Problem . . . . . . . . . . . . . . . . . . . 41

vi 3.3.4 3.4 3.5 3.6 3.7

Contents Reducing the 2-CNF Satisfiability Problem into a Graph Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 41

An Activity Scheduling Problem . . . . . . . . . . . . . . . . . 47 A Dual Machine Serial Processing Environment . . . . . . . . 48 Sequencing by Hybridization in Computational Biology . . . . 57 Discussion & Problems . . . . . . . . . . . . . . . . . . . . . . 62 77

4 Basics of Graph Theory 4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 A Mutual Friendship Graph . . . . . . . . . . . . . . . . . . . 78 Representation of a Graph . . . . . . . . . . . . . . . . . . . . 79 Complement of a Graph . . . . . . . . . . . . . . . . . . . . . 82 Equal Graphs & Isomorphic Graphs . . . . . . . . . . . . . . . 83 The Degree Sequence . . . . . . . . . . . . . . . . . . . . . . . 86 Walks, Trails, & Paths . . . . . . . . . . . . . . . . . . . . . . 100 Multi-graphs and Pseudo-graphs . . . . . . . . . . . . . . . . . 102 Broad Categories of Graphs & some Special Graphs . . . . . . 102 4.9.1 4.9.2 4.9.3 Tree Graphs . . . . . . . . . . . . . . . . . . . . . . . . 104 Bipartite Graphs . . . . . . . . . . . . . . . . . . . . . 104 Special Graphs . . . . . . . . . . . . . . . . . . . . . . 104

4.10 Integration of Concepts, Properties, and Action Items . . . . . 112 4.11 Self Complementing Graphs . . . . . . . . . . . . . . . . . . . 114 4.11.1 Regular Self Complementing graphs . . . . . . . . . . . 114 4.11.2 Non Regular Self Complementing graphs . . . . . . . . 119 4.11.3 Constructing Self Complementary Graphs . . . . . . . 121 4.11.4 Transforming a SC graph with 4k + 1 vertices into another SC graph with 4k vertices . . . . . . . . . . . . . 124

Contents

vii

4.11.5 Transforming a SC graph with 4k vertices into another SC graph with 4k + 1 vertices . . . . . . . . . . . . . . 129 4.11.6 The Self Complementary problem and Graph Isomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 4.11.7 A SC graph has diameter 2 or 3 - not less than 2 and not more than 3? . . . . . . . . . . . . . . . . . . . . . 132 4.11.8 Bipartite self complementary graphs . . . . . . . . . . 136 4.11.9 Decomposition of a SC graph G . . . . . . . . . . . . . 140 4.11.10 Permutation, Isomorphism, automorphism & Self Complementing Graphs . . . . . . . . . . . . . . . . . . . . 142 5 Basics of Graph Algorithms 5.1 153

Design of Algorithms . . . . . . . . . . . . . . . . . . . . . . . 156 5.1.1 5.1.2 What is Design? . . . . . . . . . . . . . . . . . . . . . 157 The Moore Method . . . . . . . . . . . . . . . . . . . . 157

5.2

The Bucket Algorithm . . . . . . . . . . . . . . . . . . . . . . 157 5.2.1 5.2.2 5.2.3 5.2.4 5.2.5 Understanding the Bucket Algorithm . . . . . . . . . . 159 How does it Work? . . . . . . . . . . . . . . . . . . . . 160 Playing with the Algorithm . . . . . . . . . . . . . . . 161 Solving Other Problems . . . . . . . . . . . . . . . . . 161 The Right Provocation . . . . . . . . . . . . . . . . . . 161

5.3

Finding if a Graph is Connected . . . . . . . . . . . . . . . . . 162 5.3.1 5.3.2 The Number of Connected Components . . . . . . . . 163

Finding a Bridge in a Graph . . . . . . . . . . . . . . . 163

5.4

Finding if a Graph is a Tree . . . . . . . . . . . . . . . . . . . 164 5.4.1 5.4.2 5.4.3 Every Edge in a Tree is a Bridge . . . . . . . . . . . . 164 The Number of Edges in a Graph . . . . . . . . . . . . 165 The Spanning Tree of a Tree . . . . . . . . . . . . . . . 165

viii 5.4.4 5.5

Contents A Comparison . . . . . . . . . . . . . . . . . . . . . . . 165

Finding a Spanning Tree of a Graph . . . . . . . . . . . . . . 166 5.5.1 5.5.2 5.5.3 5.5.4 Cutting Edges . . . . . . . . . . . . . . . . . . . . . . . 166 Growing Edges . . . . . . . . . . . . . . . . . . . . . . 166 Selecting Edges . . . . . . . . . . . . . . . . . . . . . . 166 Integrating Concepts and Discovering Algorithms . . . 167

5.6

Finding a Minimum Spanning Tree of a Weighted Graph . . . 167 5.6.1 5.6.2 5.6.3 5.6.4 Cutting or Growing Edges: A Krushkal’s like greedy algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 167 Selecting Edges: A Prim’s like greedy Algorithm . . . . 171 A Panoramic Picture of various MST Finding Techniques173 The Maximum Spanning Tree Problem . . . . . . . . . 176

5.7

Finding a Path in a Graph . . . . . . . . . . . . . . . . . . . . 177 5.7.1 5.7.2 Cutting Edges . . . . . . . . . . . . . . . . . . . . . . . 177 Selecting Edges . . . . . . . . . . . . . . . . . . . . . . 179

5.8

The Shortest Path Problem . . . . . . . . . . . . . . . . . . . 179 5.8.1 5.8.2 5.8.3 5.8.4 5.8.5 Dijkstra’s (like) Algorithm . . . . . . . . . . . . . . . . 181 Discussion on Dijkstra’s (like) Algorithm . . . . . . . . 181 The Shortest Path Problem Redefined: The k-edge Shortest Path Problem . . . . . . . . . . . . . . . . . . 187 The k-edge Longest Path Problem . . . . . . . . . . . . 190 The Shortest Path Problem in Undirected Graphs with Negative Weights . . . . . . . . . . . . . . . . . . . . . 192

5.9

Graph Traversal Techniques . . . . . . . . . . . . . . . . . . . 195 5.9.1 5.9.2 Traditional Techniques & the Bucket-Algorithm . . . . 196 The Underlying Data Structure . . . . . . . . . . . . . 196

5.10 Some Graph Theoretic Claims . . . . . . . . . . . . . . . . . . 198 5.11 Shortest Path Algorithms . . . . . . . . . . . . . . . . . . . . 204

Contents

ix

5.11.1 Single Source Shortest Path Algorithms with positive edge weights . . . . . . . . . . . . . . . . . . . . . . . . 205 5.11.2 Single Source Shortest Path Algorithms for Directed Acyclic Graphs . . . . . . . . . . . . . . . . . . . . . . 210 5.11.3 Single Source Shortest Path Algorithms for directed graphs with negative edge weights . . . . . . . . . . . . 218 5.11.4 All Pair Shortest Path Algorithms . . . . . . . . . . . . 227 5.11.5 Johnson’s all Pair Shortest Path Algorithm . . . . . . . 233 5.12 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247 6 Network Flows, Connectivity and Matching Problems 6.1 6.2 6.3 6.4 249

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250 Definitions & Prior Knowledge . . . . . . . . . . . . . . . . . . 250 Konig’s Theorem, Menger’s Theorem & Hall’s (Marriage) Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261 Menger’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . 264 6.4.1 6.4.2 6.4.3 6.4.4 6.4.5 6.4.6 Maximum Edge-Disjoint Paths in Directed Graphs . . 264 The Concept of a Minimum Cut in Directed Graphs . . 271 A Proof of Menger’s Theorem and Finding the Minimum Cut in Directed Graphs . . . . . . . . . . . . . . 275 Finding Maximum Vertex-Disjoint Paths & Minimum Vertex Cut in Directed Graphs . . . . . . . . . . . . . 279 Menger’s Theorem for Undirected Graphs . . . . . . . 282 Edge Connectivity and Vertex Connectivity for Undirected Graphs . . . . . . . . . . . . . . . . . . . . . . . 285

6.5 6.6

Konig’s Theorem, Menger’s Theorem & Hall’s (Marriage) Theorem Revisited . . . . . . . . . . . . . . . . . . . . . . . . . . 298 Network Flows . . . . . . . . . . . . . . . . . . . . . . . . . . 306 6.6.1 6.6.2 Finding Maximum Edge-Disjoint Paths in Multi-Graphs306 The Maximum Flow & the Minimum Cut . . . . . . . 310

. 324 6. . . .8. . .5 6. . . . . . . 365 New Problems . . 311 Lower Bounds on Edge Flows and the Max-Cut . . .9 Network flows with lower & upper bounds on flow and the Circulation Problem . . . . . . . . . 348 Category 2 (and 1) network flow Problems . . . . .7. . . .9. .4 6.7. . . . . . . .3 6. . . . .8. . . . . . . . . . . . . .6 6. . .8. . . . .3 6. . . 370 When upper bound is higher than a non zero lower bound372 Is it possible to solve the Circulation Problem for undirected graphs? . . . 345 Finding a Maximum Flow or Finding a Shortest Path? 345 Category 3 network flow Problems . .4 6. . . . . . . . . . . . . . . . . 377 381 7 Eulerian Graphs & the Chinese Postman Problem . . . . .6. . . .5 Introduction . .8. . . . 312 The Matching Problem in Bipartite Graphs . . . . . . . . . . . . . . .9. . . .3 Maximum Matching in Un-weighted bipartite graphs . 366 Finding a feasible flow in a Circulation graph with one special edge . . . . . . . .9. . . . .8. . . . . . . . .2 6. . . . . .7. . . . .x 6. . 345 6. .7 Contents Algorithmic Issues & Complexity Calculations . . . . . . . . . . . . . .8 The Max-Flow Min-Cost Problem . . . . . . . . .1 6. . .4 6. .1 6. . . . . . . 364 6. . . . . . .6. . . . . . . . . . . .9. . . . . . . . . . . . . . . . . . 369 Finding a feasible flow in a network flow graph with one special edge . .2 6. . . . .9. . . . .9. . .3 6. . . . . . . . . 363 6. . . . . . . . . 325 Maximum Matching in Complete (Binary) Weighted Bipartite Graphs . . . . . . . 350 A Panoramic Picture of Similar Problems & Solutions (once again) . . 328 Maximum Weighted Matching in Complete Weighted Bipartite Graphs . . . . . . . . . . .2 6. . . . . . . . . . . . . . . . . . . . . . . 332 6. 364 New concepts . .7 Prior knowledge: . . .9. . . . . .1 6. . . . . .

. . 397 The Chinese Postman Problem for Directed Graphs . . . . . . .2. . . . . . . . . . . . . . 425 A Concept Map . . .1 7.3.7 A Puzzle: . . . . . .2. . . . . . .3. . . . . . . . . . .6 Bipartite Hamiltonian Graphs . . . . . . . . . . . 425 8. . . .5 8. . . . . . .5 8. . .3 8. . . .5 xi A Special Class of Graphs .5 Necessary Conditions for a Connected Graph . . . . . . . . . . . . . . . . . .2. . . . 438 Closure of a Graph: . . . . . . . . . . . . . . . . . . . . . . . . . .1 8. . . . . . . . . .3. . . . Properties & Actions . . . . . . . . 446 Some Theoretical Claims . . . . . . . . .1 Concepts. . . .2. . . . . . . . . 410 423 8 Hamiltonian Graphs 8. . . . . . . . . . . . . . . . 439 8. . . 464 . . . . . .1 8. . . . . . . . 456 463 9 Strongly Connected Directed Graphs and Tournaments 9. . . . . 424 A Loose Sufficient Condition for a Hamiltonian Graph 425 Sufficient Condition for a Connected Graph . .3. . 394 Eulerian Walk and the Chinese Postman Problem . . 436 Bondy & Chvatal’s Theorem: . . . . . .3 8. . 436 Summary . . . . . . . . . . . . . . . . . . . . 382 Eulerian Circuits and Graphs . .2 8. . . . 424 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429 Actually Finding a Hamiltonian Cycle in the Puzzle: . . . . . . 428 8. 438 Ore’s Theorem: . . .2 Introduction . . . . . .4 7. .Contents 7. . . . . . . . . . . . . . . .4 8. . . . . .3 7. . . . .2 8. . . . . . . . . . . 429 Basic Intuition .2. . . . .4 8. . . . . . . .3. . . . . . . .3 Hamiltonian Graphs . . . . . . . .1 8. . 447 A Categorization of Hamiltonian Graphs . . .2 7. . . . . . . . . . . 424 Necessary Conditions for a Hamiltonian Graph . . . . . .4 8. . . . . . 388 Eulerian Trails and Related Problems . . . . .3. . . . . . . .6 8. . 424 Prior Knowledge .3.

485 Tournaments . 498 A Hamiltonian Cycle in a Strong Tournament . . . . . . .4 9. . 498 9. . . . . . . . . . . . . . . . . . . 518 . . .3 9. . . . . . . . . . . . . . . . .2 9. .6 A Panoramic Picture and a Concept Map . . . . . . 471 Directed Acyclic Graphs (DAG’s) . . 503 Unilaterally Connected Directed Graphs: . .xii 9. . .5. .3 9. . . . . . . 511 9. . . . . . . . .1 9. . . . 498 A Hamiltonian Path in a Tournament . .6. . . . . . 476 Strongly Connected Components . . . . . . .2 9. . . . . .1 Unilaterally orient-able Un-directed Graphs . .5. . . . .5. . . . . .5 Contents Strongly Connected Directed Graphs . . . . . . .

6 1.2 1.7 Why a new book? What do we emphasize? How the book is organized? How is the book designed? Some salient features of the book What tools do we use? A Possible Sequence .1 1.4 1.Chapter 1 Introduction 1.5 1.3 1.

For example graph. and a learner needs to become familiar with these concepts. visualize. In fact we try our best to design constructive proofs which not only prove a connection but also provide an algorithm to find something useful. edge weight.2 Introduction 1. Algorithms (actions which transform an input into an output) Let us start with a few concepts used in our subject. For example we use constructive techniques to find as well as prove that the maximum number of edge-disjoint paths are equal to the size of the minimum cut in a graph. For example if every edge is a bridge edge in a connected graph G then the graph G is a tree. thus reinforcing understanding at various levels. An implication connects two or more concepts. Many books tilting towards graph theory do not emphasize graph algorithms. Unlike these books we operate on all the three layers. We need to prove a theorem like this one in a formal but convincing manner. connected graph. & edge cut are all concepts. There are basically three layers on which such a course operates: 1. vertex. Implications & theorems (connections between concepts) 3. tree.1 Why a new book? There are a number of excellent books available on topics covering an introductory course on Graph Theory & Graph Algorithms . For example an algorithm is to be designed to find maximum edge-disjoint paths as well as the minimum edge cut in a given input graph. edge. and connect these concepts. bridge edge. We shall discuss here why there was a need for another book designed with a different pedagogical structure and suitable for students and practitioners working in diverse fields who intend to use graph theory in their respective fields of study. Similarly a number of graph algorithm books ignore the theory part of this course namely graph theorems and their proofs. Definitions (concepts) 2. . edge-disjoint paths. He or she should be able to feel. Before we start making comparison between different pedagogical approaches and talk about the merit of our approach we shall first describe the basic structure of the said course.almost each with a different approach. path. An algorithm converts an input into a useful output. Menger’s theorem connects maximum edge-disjoint paths to the minimum edge cut in a graph.

2.1).What do we emphasize? 3 Most of the other textbooks use a historical perspective of how and when a graph algorithm was discovered. We provide platforms where a learner is provoked to transform one proof technique (or an algorithm) into another proof technique (or an algorithm). We think that not only the similarities but also the differences should be highlighted between two almost similar concepts or algorithms. We expect our readers to first think of a non graph problem in terms of a graph.2 What do we emphasize? The single most prominent feature which distinguishes this book from other books in this field is an emphasis on transformations (or reductions). We think it is important from a learning perspective to integrate as well as differentiate concepts and techniques especially the ones which solve nearly identical problems. Similarly. Transforming one concept into another and transforming one graph into another is the driving force behind this exercise (see Fig. We then encourage our readers to transform that non-graph problem into a graph problem.2. 1. . We also encourage them to transform a graph problem into another graph problem.2). 1. (see Fig. We think of this course on graph theory & algorithms as a course on (intelligent) transformations. The historical perspective encourages one to teach these two algorithms in isolation without making any connections between the two algorithms. We thus encourage our readers to transform one algorithm (Prim’s algorithm) into another (Dijksta’s algorithm). For example Dijkstra designed a greedy algorithm to find shortest path between two vertices in a weighted graph. it is possible and desirable to transform one proof into another. For example we encourage our readers to transform the techniques used to prove that Dijkstra’s algorithm really finds shortest paths into techniques to prove that Prim’s algorithm really finds a minimum spanning tree in a weighted graph. 1. Similarly we encourage them to transform one theorem into another theorem. For example Prim’s algorithm can find a minimum spanning tree or a maximum spanning tree (after a minor modification) and can handle positive or negative edge weights while Dijkstra’s algorithm is unable to do so. Similarly Prim designed a greedy algorithm to find a minimum spanning tree in a weighted graph.

com/worldcup/matches/index. (http://www.2.1: World Cup 2010 round of 16 is represented by a graph.fifa.4 Introduction Figure 1.html) .

.k)} j 3 g k 3 2 6 a f 2 d 1 e 5 3 1 3 5 2 2 b e 4 h 3 c 6 7 3 4 c f 2 d 1 5 3 1 i 3 k 3 6 3 5 a 2 2 b g 2 4 h 6 7 3 4 c j 3 i 3 k 4 2 b Figure 1.2.What do we emphasize? 5 j 3 g 2 3 f 2 1 e 3 a 2 d 3 1 2 b e 4 h 6 9 i 3 k 3 c f 2 d 1 3 3 5 g 2 4 j 3 i 6 9 a 2 2 3 1 b e 4 3 k 3 c f 2 d 1 3 6 3 5 g 4 2 j 3 i 6 9 a 2 2 3 1 b 4 3 k 3 c h h ut sb ce s tan path d i s e st st rte ort h o e sh s he of th dt fin rack We se t lo j 3 g 2 6 a f 2 d 1 e 5 3 1 3 5 2 4 h 3 6 7 i 3 w(a.k) = min{w(a. j)+w(j. w(a.k).2: A visual depiction of how an algorithm transforms a graph into another graph while trying to find the shortest paths (Chapter 5).

All the proofs used here are constructive . 1. For example.1). and computer science. molecular biology. We also show how one problem (and its solution). We show .3. Chapter 6 is the longest chapter in the book. The fourth chapter focuses on basics of graph theory with an introduction to several concepts and properties related to graphs.3 How the book is organized? The book is organized into eight chapters other than this introduction (see Fig. and the circulation problem. network flows. minimum cost flows. We discuss concepts related to graph connectivity. for example. The non-graph problems come from diverse fields like operations research. We start with a stupid algorithm known as the Bucket Algorithm which provides us with a platform where we encourage our learners to modify and mould the Bucket Algorithm to design various useful graph algorithms.we not only prove that an Eulerian Circuit exists in a graph with even degree but we also find that circuit using an algorithm. that is.3. This chapter consists of multiple & diverse topics as described before but a conscious effort has been made to make sure that the number of important milestones or bottlenecks in learning remains very small. civil engineering. The fifth chapter handles some important graph algorithms. the circulation problem is transformed into another problem like the minimum cost-maximum flow problem. digital logic. matching problems. The third chapter describes how a (non-graph) problem is transformed into a graph problem and encourages the readers to think in graph theoretic terms. We also provide other tools like a colorful visual puzzle which differentiates between and integrates various shortest path algorithms (see Fig. Instead we encourage our learners to devise cruder versions of an algorithm which are relatively easy to discover and appreciate intuitively. distributed computing. The second chapter provides standard definitions.6 Introduction 1. if we need to find multiple edge-disjoint paths in a graph then we should reverse the direction of an already found path before finding another edge-disjoint path. Here we effectively start using our emphasis on transformations. We actively demonstrate how a theorem (and its corresponding proof) like Menger’s theorem is transformed into Hall’s theorem or the Konig’s theorem. We purposely do not use a historical perspective and avoid describing an algorithm in its published (or polished) form. all the above mentioned topics depend upon one central and crucial idea. 1. In Chapter 7 we discuss necessary and sufficient conditions for Eulerian graphs and the Chinese Postman problem.2).

The directed acyclic graph illustrates a possible teaching strategy.1: Organization of our book consisting of nine chapters. .3.How the book is organized? 7 Figure 1.

In case of tournament graphs we come back to the problem of finding a Hamiltonian path and a Hamiltonian . unilaterally connected graphs. In Chapter 6. Hamiltonian directed graphs are discussed again in Chapter 8.3.8 Introduction that the Circulation problem (already discussed in Chapter 6) was in fact the Chinese Postman problem in its general form. we transform one problem into another and the corresponding algorithm is also transformed in the process. We discuss sufficient conditions for an undirected as well as a directed graph to be Hamiltonian. and tournaments in this chapter as well. Figure 1. We address strongly connected graphs and components. Both problems and their respective solutions in fact converge into a single problem and a single technique to solve it. We describe Hamiltonian graphs in Chapter 7.2: In shortest path algorithms (described in Chapter 5) we transform one algorithm into another solving the same problem.

Identify a potential bottleneck in learning a specific concept.How is the book designed? cycle provided the tournament is strongly connected.5 Some salient features of the book 1.4 How is the book designed? We have followed a three step design strategy while writing each section and sub-section of this book: 1. its usefulness and limitations 4. Instead of describing a concept in its finished & sophisticated form we first describe its cruder version which is easy to discover and appreciate 3. 9 1. Pakistan). A number of research projects undertaken by students registered in a course titled “Problems of Learning & Teaching ” highlighting various problems of learning in the field of graph theory & algorithms have immensely helped us in the design of various sections). 2. Emphasis on prior knowledge. algorithms or elaborate proofs . Identify its nature using theories of learning & pedagogy (this is done through a statistical analysis of student feedback obtained in the last six years of teaching this course at LUMS. 1. Several hundred colored diagrams play a central role in the design of the book. and theorems reinforce learning and understanding 5. 2. algorithms. Starting with simple and easy to use building blocks which are used by a learner to construct more sophisticated concepts. Some text is added where needed in order to supplement diagrams unlike other books where diagrams supplement text. 3. An integrated approach where concepts. Discovery based learning is practised by first asking provoking questions before actually describing an algorithm or a theorem. Lahore. Remove the bottleneck by introducing a number of bridging concepts and by drawing suitable number of colored diagrams taking advantage of the role of visualization in learning graph theory.

. A lot of technology tools are nowadays available for synchronizing class room video with multimedia slides and lecture notes (synchronizing annotations for educational multimedia). We then encourage our learners to appreciate by themselves the repercussions of that error and in the process find an alternate path to solve the problem correctly. 1. concept maps. It has also been used to teach a similar course at the Virtual University of Pakistan. 1.com. 1.. We use a number of tools from the science of learning. The book has been used several times to teach a graduate level course on Graph Theory & Algorithms at LUMS. We always make comparisons highlighting similarities as well as differences between concepts.synote. Constructive proofs with the help of algorithms 8.6.1. Some of the VU lectures along with Power Point slides are available on Synote. 1.7 A Possible Sequence A possible sequence of lectures covering this book is given below in a typical 3-credit course in a semester system at a higher under-graduate or early graduate level.10 Introduction 6. algorithms as well as theorems 7.3). At the VU.g. 10. 9. Complexity of learning under our control because of a conscious effort to keep it under limits. Komal Syed and I tried to discuss and debate various issues concerning graph theory & algorithms using this book (see Fig. and can be viewed after getting an account on www. we have used a 3-person drama format instead of a single person monologue.6 What tools do we use? We occasionally use concept map as a tool for better learning which enable a learner to explicitly make connections between concepts as shown in Fig. Yasser Hashmi. e. We sometimes encourage our readers to make errors as we think that making an error is a step towards meaningful learning.6.

We use several visualization tools in order to enable the learner to visualize the working and the subsequent time complexity of an algorithm as shown in Fig.6.2.6.1: A concept map showing a number of relevant concepts and a number of theorems which relate different concepts (taken from Chapter 6). 1.A Possible Sequence 11 Figure 1. .

This helps a learner to visualize the time complexity of an algorithm. .6.2: Visualizing how an algorithm changes a graph and what price is paid in terms of number steps performed. Each edge in the pink graph exactly corresponds to one step in the algorithm.12 Introduction Figure 1.

.6.A Possible Sequence 13 Figure 1.3: A 3-person drama format for teaching Graph Theory & Algorithms at the Virtual University of Pakistan.

A learner has the facility of expressing and sharing his or her feedback regarding the understanding of a concept in this software platform.14 Introduction Figure 1.6.synote.4: Virtual University video lectures along with Power Point slides are available on www. .com.

Finding a minimum (maximum) spanning (MST) tree. Algorithmic Issues & Complexity Calculations. the Shortest Path Problem. Properties. Equal Graphs & Isomorphic Graphs. & Hall’s (Marriage) Theorem Re-visited. (3 Lectures) Network Flows: Finding Maximum Edge-Disjoint Paths in Multi-Graphs. Integration of Concepts. Menger’s Theorem. Multi-graphs and Pseudo-graphs. Menger’s Theorem for Un-directed Graphs. the Maximum Flow & the Minimum Cut. Menger’s Theorem. Finding if a Graph is a Tree. some Graph Theoretic Claims. Single source & all pair shortest paths. Complement of a Graph. Matching Problems & Network Flows Definitions. the Satisfiability Problem (in Logic Circuits). Chapter 6 (9 Lectures) Connectivity. Konig’s Theorem. Lower Bounds on Edge Flows: Min-Flow & the Max-Cut. Trails & Paths. the Bucket Algorithm. the Degree Sequence. Finding if a Graph is Connected. & Warm-up Exercises. the Matching Problem in Bi- . Models & Graphs: Why Study Graph Theory 15 Reducing one Problem into another. Finding Maximum Vertex-Disjoint Paths. Graph Traversal Techniques. Broad Categories of Graphs. Finding Maximum Edge-Disjoint Paths. Edge Connectivity & Vertex Connectivity. Representation of a Graph. Finding a Path in a Graph. Menger’s Theorem. & Hall’s (Marriage) Theorem. Walks. Sequencing by Hybridization (in Computational Biology). Finding a Spanning Tree of a Graph. Prior Knowledge. An Activity Scheduling Problem (in Civil Works). discussion & problems (from Operations Research) Chapter 4 (3 Lectures) Basics of Graph Theory A Mutual Friendship Graph. a Dual Machine Serial Processing Environment (from Distributed Computing). Konig’s Theorem. a Proof of Menger’s Theorem & Finding the Min-Cut. The Concept of a Minimum Cut. and Action Items Chapter 5 (4 Lectures) Basics of Graph Algorithms Design of Algorithms.A Possible Sequence Chapter 3 (1 Lectures) Problems. Lower as well as Upper Bounds on Edge Flows.

Eulerian Circuits & Graphs. discussion & Conclusions Chapter 9 (5 Lectures) Strongly Connected Directed Graphs and Tournaments Concepts. Finding Maximum Flow at Minimum Cost (Category 1). Maximum Matching in Unweighted Bipartite graphs. Strongly Connected Directed Graphs. Tournaments.16 Introduction partite Graphs. finding a Feasible Flow Under Lower & Upper Bounds. Properties. Connecting the Network Flow Problem with the Circulation Problem. Maximum Matching in Complete (Binary) Weighted Bipartite Graphs. & Actions. Unilaterally Connected Directed Graphs. Hamiltonian Graphs. a Panoramic Picture of Similar Problems & Solutions (once again). Directed Acyclic Graphs (DAG’s). (3 Lectures) Chapter 7 (3 Lectures) Eulerian Graphs & the Chinese Postman Problem A Special Class of Graphs. (3 Lectures) The Max Flow Min Cost Problem: Finding a Maximum Flow or Finding a Shortest Path? A Panoramic Picture of Similar Problems & Solutions. . Negative Weight Cycles & Improvement in Cost of Flow. The Chinese Postman Problem for Directed Graphs Chapter 8 (4 Lectures) Hamiltonian Graphs Prior Knowledge. Maximum Weighted Matching in Complete Weighted Bipartite Graphs. Eulerian Walk & the Chinese Postman Problem. Problems & Claims. Eulerian Trails & Related Problems. the Circulation Problem. Bipartite Hamiltonian Graphs. Finding a Minimum Cost Feasible Flow. Strongly Connected Components.

Matching Problems.A Possible Sequence A Possible Program of Study in one semester Topic Introduction Definitions Problems. & Graphs Basics of Graph Theory Basics of Graph Algorithms Connectivity. & Network flows Chinese Postman Problem & Eulerian Graphs Hamiltonian Graphs Strongly Connected Graphs & Tournaments 17 Chapter Chapter Chapter Chapter Chapter Chapter Chapter Chapter Chapter Chapter No 1 2 3 4 5 6 7 8 9 Hours 0 0 2 6 8 18 6 8 10 . Models.

Turkey. 2005. 7. Harika Masood. 6. Vol.” proceedings of the IADIS International Conference on Cognition and Exploratory Learning in Digital Age (CELDA 2005). Arlington. 409-418. Ed. The Magic of Dynamic Programming. Proceedings of the International Conference on Engineering Education. Sara Tahir. 2005 2. Int. Yasser Hashmi. VA. 2006. Nooman Nadeem. ”Teaching Science and Mathematics: Discovery Based Learning”.” proceedings of the Second International Conference on Pedagogies and Learning -2005. Int. VA. Ed. Valencia Spain. 409-418. ”Problems of Learning & Teaching.2. The Magic of Dynamic Programming. 3. Arlington. Network for Eng.” proceedings of the Third International Conference on the Teaching of Mathematics at the undergraduate level.18 Acknowledgements Introduction 1. ”Should We Teach Algorithms?” IJECE. Istanbul. Yasser Hashmi. pp. Ed. Australia. Network for Eng. No. (iNEER). (iNEER). pages 78 to 87. Alvi Atif. and Res. 5. 134-140. ”Meaningful Learning of Graph Algorithms. SummerFall 2003. 4. . Int. pp. Innovations 2004: World Innovations in Engineering Education and Research. ”Bridging over Problems of Learning in Finding Strongly Connected Components. pp. (iNEER).2. VA. 2004. 2003. Innovations 2005: World Innovations in Engineering Education and Research. Alvi Atif. and Res. and Res. Network for Eng. Arlington.

Chapter 2 Basic Definitions in Graph Theory and Algorithms .

It is a set of ordered pairs in case of a directed graph and a set of un-ordered pairs in case of an un − directed graph. v} in graph G. (Please see Chapter 4 for more details) Adjacent Vertices A vertex u is said to be adjacent to vertex v if there is an edge {u. (Please see Chapter 4 and 5 for more details) Multi Graphs We assume that multiple edges (known as parallel edges) between the same two vertices are not allowed in a graph. Thus a graph consists of vertices and edges. A weighted graph is one where there may be a weight associated with each edge of the graph. We may represent the set of vertices by V (G) and the set of edges may be represented by E(G). Similarly an edge coming out of a vertex and terminating at the same vertex (known as a self loop) is not allowed. (Please see Chapter 4 for more details) Regular graph An un-directed graph G is regular if the degree of each vertex is the same. Please note that the set E is a set of pairs of connected vertices. (Please see Chapter 4 for more details) Degree of a vertex Number of edges connected to a vertex x is known as the degree of vertex x in an un-directed graph G. (Please see Chapter 4 for more details) .some of the pair of vertices may be connected by directed or un-directed links which are known as edges. In M ultiGraphs we allow parallel edges as well as self loops. Usually the size of the vertex set V is represented by p while the size of the edge set E is represented by q.20 Basic Definitions in Graph Theory and Algorithms Graph A graph G is made up of a (non empty) set of objects called vertices or nodes .

(Please see Chapter 4 for more details) .(Please see Chapter 4 for more details) Bipartite Graph A graph G is bipartite if the vertex set V (G) can be divided into two subsets (or partites) A and B such that every edge in G connects a vertex in partite A and a vertex in partite B.21 Star graph A graph of p vertices in which one vertex has degree equal to p−1 while every other vertex has a degree equal to 1. v} in a graph G there is an edge {u. Please note that two equal graphs are always isomorphic but two isomorphic graphs may not be equal. (Please see Chapter 4 for more details) Isomorphic Graphs Two graphs G and H are isomorphic if graph G becomes equal to graph H by some relabeling of vertices of graph H. (Please see Chapter 4 for more details) Complete Graph An un-directed graph G is complete if there is an edge between every pair of vertices of graph G. It is known as a star graph because it looks like a star with rays of lighy coming out? (Please see Chapter 4 for more details) Chain or Cycle graph A connected graph G such that the degree of each vertex is exactly 2. (Please see Chapter 4 for more details) Equal Graphs If for every edge {u. v} in graph H and vice versa then the two graphs G and H are equal.

it is known as a trivial permutation. The identity permutation is always an automorphism . (Please see Chapter 4 for more details) Automorphism If a permutation p of a graph G creates a graph H which is equal to graph G then that permutation is known as an automorphism of graph G. v} in G.22 Basic Definitions in Graph Theory and Algorithms Complement c(G) of an un-directed graph G The complement c(G) of graph G is a graph with as many vertices as in G and an edge {u. v} in c(G) if and only if there is no edge {u. (Please see Chapter 4 for more details) Walk You can walk on the edges of graph G edges starting from vertex u and ending at vertex v traversing different edges and vertices. (Please see Chapter 4 for more details) Transpose of a directed graph G T ranspose T (G) of a directed graph G is obtained by reversing the direction of each edge in the directed graph G. (Please see Chapter 4 for more details) Complementing Permutation If a permutation p of vertices of graph G creates a graph H which is the complement of graph G and at the same time H is isomorhic to G then the permutation p is known as the complementing permutation of graph G and graph G will be a self complementary graph. We may be more interested in the non trivial permutations of graph G? It may be possible for a certain category of graphs that the only automorphism is the trivial (identity) permutation? Please note that if p is an automorphism of graph G then the permutation p2 is always an automorphism of graph G. (Please see Chapter 9 for more details) Self Complementing graph G A graph G is self complementing (SC) if graph G and its complement are isomorphic to each other. In a walk you .

that is the direction of that edge. (Please see Chapter 4 and 7 for more details) Path If neither an edge nor a vertex is repeated in a walk starting from a vertex u and ending at vertex v then the walk is known as a path. (Please see Chapter 4 and 7 for more details) Trail and Circuit If no edge is repeated in a walk from a vertex u to a vertex v then the walk is known as a trail. (Please see Chapter 4 and 5 for more details) .23 can traverse an edge more than once. A path is always a trail (or a walk) but it is not the other way round. (Please see Chapter 4 and 5 for more details) Shortest path Among all paths between vertex u and vertex v. Please note that it is possible to traverse a vertex more than once but an edge should not be traversed more than once in a trail or in a circuit.that is when vertex u and v are the same then it is known as a circuit. A walk is open if vertex u and v are different. It is known as a u − v path. Please note that in an un-directed graph you can traverse an edge in both directions but in a directed graph you can traverse an edge in only one direction . Thus a trail is always a walk but it is not the other way round. the one with minimum length is known as the shortest path between vertex u and vertex v. It is a closed walk if vertex u and v are the same. In a weighted graph G the minimum length is measured in terms of sum of weights of all edges in the u − v path. (Please see Chapter 4 for more details) Cycle If a path is closed that means you come back to the vertex from where you have started then that path is known as a cycle. A closed trail . A cycle is a circuit but a circuit may not be a cycle as no vertex should be repeated in a cycle. In an unweighted graph G the minimum length is measured in terms of number of edges encountered in the u − v path.

In other words there are no cycles in a . The Reachable Relation matrix of an undirected connected graph will contain all 1’s. (Please see Chapter 9 for more details) A Unilaterally Connected Directed Graph A directed graph D is unilaterally connected if there is directed path from vertex u to vertex v or a path from vertex v to u for every pair (u. (Please see Chapter 9 for more details) Square of a directed graph The square of a graph D is another graph (known as) D2 in which there is an edge from vertex u to vertex v provided there is a two edge path from u to v or there is an edge (that is a one edge path) from u to v in D. (Please see Chapter 4 and 8 for more details) A Strongly Connected Directed Graph A directed graph D is strongly connected if there is a directed path from vertex u to vertex v and a path from vertex v to u for every pair (u. and it is zero otherwise. (Please see Chapter 4 and 5 for more details) Reachable relation or Transitive closure of a graph The Reachable Relation (or the transitive closure) of a directed graph D is another directed graph in which there is an edge from vertex u to vertex v provided v is reachable from u in D. (Please see Chapter 9 for more details) Directed Acylic Graph A directed graph D is directedacyclic (or a DAG) if there is directed path from vertex u to vertex v then there is no path from vertex v to u for every pair (u.24 Basic Definitions in Graph Theory and Algorithms A Connected Graph An un-directed graph G is connected if there is a path between every pair of vertices of that graph. v) in directed graph D. v) in directed graph D. The Reachable Relation graph of D can be represented by an adjacency matrix A in which A(u. v) = 1 provided there is a directed path from u to v in D. v) in directed graph D.

v) in directed graph D. (Please see Chapter 4 and 5 for more details) Cyclic graph G A graph G is cyclic if it contains one or more cycles. It is also known as a tree graph provided graph G is connected. In other words a connected graph G is a tree if every edge of G is a bridge edge. 5 and 9 for more details) Acyclic graph or a Tree An un-directed graph G is acyclic if it does not contain a cycle. v} is a bridge edge if its removal disconnects an un-directed graph G. (Please see Chapter 4 and 5 for more details) Bridge edge or Cut edge An edge {u. (Please see Chapter 4. It is known as a tournament graph as some (actual) tournaments (like f ootball league are played in the form of a tournament graph? (Please see Chapter 9 for more details) Disconnected graph An un-directed graph is disconnected if for any pair of vertices u and v there is no u − v path. (Please see Chapter 5 and 9 for more details) A Tournament Graph A directed graph D is a tournament if there is a directed edge from vertex u to vertex v exclusive OR a directed edge from vertex v to u for every pair (u.25 directed acyclic graph. 5 and 9 for more details) A Forest A (disconnected)graph G with no cycles. (Please see Chapter 4 and 5 for more details) . (Please see Chapter 4. In other words if we put directions on edges in a completely connected un-directed graph G then the un-directed graph G transforms into a tournament (directed) graph D. Please note that a tree is a connected graph with no cycles while a forest may be a disconnected graph.

We have already witnessed the subset sum problem (in previous courses) in which we have to select integers (out of a set of integers) such that the sum of the selected integers is equal to a given constant.26 Basic Definitions in Graph Theory and Algorithms Spanning tree of a graph G A spanning tree (known as ST ) of a connected graph G contains all vertices of G and some edges of G. In other words if you remove all non bridge edges in graph G then you get a spanning tree of G. and it is a tree. A graph which contains a Eulerian circuit is known as a Eulerian graph. what is the smallest subset T of S such that the union of all these sub sets in T covers all elements of U . (Please see Chapter 6 for more details) . A graph which contains a Hamiltonian cycle is known as a Hamiltonian graph. (Please see Chapter 4 for more details) Hamiltonian Cycle It is a cycle in a graph G which traverses each vertex of G exactly once. (Please see Chapter 5 for more details) A Binary Tree A binary tree is a tree such that the degree of each vertex is not more than three. (Please see Chapter 8 for more details) Eulerian Circuit A circuit in a graph G such that every edge of graph G is traversed exactly once. (Please see Chapter 4 for more details) A Path graph A path graph is a tree provided it has two vertices with degree one while all other vertices has degree exactly equal to two. (Please see Chapter 7 for more details) Set Cover Given a set of subsets S of the U niversal Set U .

no two edges in the subset share a common vertex. In other words a matching is a set of non-adjacent edges in a graph G. (Please see Chapter 6 for more details) Maximal Matching This is a matching in which more edges cannot be added in the existing matching to increase the size of this matching. (Please see Chapter 6 for more details) Maximum Matching This is a matching in a graph with as many edges as possible? Please note that maximum matching is always maximal. What is the smallest subset of vertices of the graph that covers all edges? (Please see Chapter 6 for more details) Independent (Vertex) Set It is the largest subset S of vertices of a graph such that no pair of vertices in S has an edge in between. as compared to the number of vertices in a graph? (Please see Chapter 6 for more details) . It may be possible to increase the size of the matching by first discarding the initial matching edges. Is there a connection between the vertex cover and the independent set? (Please see Chapter 3 and 6 for more details) Matching (Independent Edge Set) It is a subset of edges in graph G such that no two edges in the subset has a common vertex in G. (Please see Chapter 4. 5 and 6 for more details) Edge Cover The Universal set is the set of all vertices in a graph. which covers all vertices? How small (or big) can the size of the edge cover become.27 Vertex Cover The Universal set U is the set of all edges in a graph. The edges in this subset are also known as independent edges.Thus it is a matching of maximum size. that is. Then the edge cover is the smallest subset of edges.

the term complete matching is used for it. (Please see Chapter 6 for more details) Bridge edge or Cut edge An edge {u. (Please see Chapter 4. (Please see Chapter 6 for more details) Minimum Edge-Cut or MinCut A set of (minimum number of) edges which if removed will disconnect a special vertex s from another special vertex t in a graph G. In some literature. Every perfect matching is both maximum and hence maximal. (Please see Chapter 6 for more details) . (Please see Chapter 6 for more details) Vertex Connectivity It is the minimum number of vertices which if removed will disconnect an un-directed connected graph G. 5 and 6 for more details) Cut vertex A vertex u is a cut vertex if its removal disconnects an un-directed graph G. (Please see Chapter 6 for more details) Minimum Vertex-Cut A set of (minimum number) of vertices which if removed will disconnect a special vertex s from another special vertex t in a graph G. (Please see Chapter 6 for more details) Edge Connectivity It is the minimum number of edges which if removed will disconnect an undirected connected graph G.28 Basic Definitions in Graph Theory and Algorithms Perfect Matching A P erf ect Matching is a matching which covers all vertices of the graph. every vertex of the graph is incident to exactly one edge of the matching. That is. v} is a bridge edge if its removal disconnects an un-directed graph G.

29

Network Flow problems in a Network Flow graph D
It is a directed graph D with two special vertices. One is a source vertex s and the other is a sink vertex t. The source vertex may produce flow while the sink vertex sinks flow. Every edge may have an associated lower bound on flow, an upper bound on flow, and a cost function associated with flow. We may like to find a f easible flow from vertex s to vertex t. We may like to find the maximum or minimum feasible flow or maximum f low at minimum cost.

Feasible Flow in a network Flow Graph
A f easible flow in a network flow graph from a source vertex s to a sink vertex t is one in which flow through every vertex (other than the source and the sink vertices) is conserved (that is inflow is equal to out flow) and flow through every edge is within the prescribed upper as well as lower bounds. (Please see Chapter 6 for more details)

The Circulation Problem
We consider a network flow graph in which the incoming flow in every vertex should be equal to the outgoing flow in every vertex. The lower bound on flow through every edge is exactly 1. There is a uniform cost of flow through every edge. We need to find a minimum cost f easible f low in the network flow graph (also known as the circulation graph). The problem is solvable in polynomial time for directed graphs. We assume that the Circulation graph is connected if it is un-directed and strongly connected if it is directed. (Please see Chapter 6 and 7 for more details)

The Chinese Postman Problem
We are given a graph D which is strongly connected if directed and connected if un-directed. We need to find a closed walk in this graph such that the total distance covered is minimum in terms of the number of edges of the graph. The problem is solvable in polynomial time for directed as well as un-directed graphs. Please note that the Chinese Postman Problem transforms into the Circulation Problem (and vice versa) for directed graphs. (Please see Chapter 6 for more details)

30

Basic Definitions in Graph Theory and Algorithms

Chapter 3 Problems, Models & Graphs: Why Study Graph Theory?

3.1 3.2 3.3 3.4 3.5 3.6 3.7

Introduction Reducing One Problem into Another The Satisfiability Problem in Logic Circuits An Activity Scheduling Problem A Dual Machine Serial Processing Environment Sequencing by Hybridization in Computational Biology Discussion & Problems

32

Problems, Models & Graphs: Why Study Graph Theory?

3.1

Introduction

When we face a real life problem then one possibility is to solve it right from scratch. A more desirable option is to better understand the problem in terms of reducing it into one of the known problems in computer science. There may be good chances that you are able to transform (or reduce) your unknown problem into one of the solvable problems in graph theory. Even if you end up reducing your problem into one of the hard (or unsolvable) problems – you certainly get a better insight. We shall discuss a number of diverse problems in this chapter; the problems are taken from digital logic, civil works, distributed computing, molecular biology, and the field of operations research. Almost all these problems do not seem to have a any relationship with graphs, yet each of them is transformed into a graph problem. As demonstrated in this chapter, it is very much possible (and in fact desirable) to reduce an unknown problem into multiple known problems instead of just one known graph problem. It may happen that out of these multiple known problems, one of the problems may be relatively simple to solve while the other known problem may be a hard one. Once your problem is transformed into a known problem, the complexity of solving your unknown problem will depend upon the complexity of your transformation as well as that of solving the known problem (see Fig. 3.1.1). The real challenge is thus to find an intelligent transformation into a simpler problem.

3.2

Reducing One Problem into Another

A reduction is a transformation of one problem into another problem. Sometimes we reduce a known hard problem into the new problem; thus showing that the new problem is as hard as some of the known hard problems. Usually reductions are used in such a negative context especially in the field of complexity theory. It is, however, possible to make reductions play a positive role: In such cases we transform a problem into one of the solvable problems, thus showing that a new problem is indeed solvable. Transforming one problem into another requires that each instance of the new problem should be transformed into instances of the old problem; we then solve the old problem using a known algorithm and then again transform its results to obtain the final solution of the new problem. We show such a transformation in Fig. 3.1.1. It is important to note the total time complexity of solving

Reducing One Problem into Another
Inputs

33

Outputs

A Known Algorithm for an Old Problem

Algorithm B Algorithm A

Inputs Outputs

A New Problem

Figure 3.1.1: Reducing a new problem into an old problem. The inputs for the new problem should be transformed into the inputs of the old problem. Similarly the outputs should also be transformed. a new problem by reducing it into an old problem will be time complexity of the known algorithm (needed to solve the old problem) plus the time complexities of Algorithm A and that of Algorithm B. We shall categorize a reduction into the following four categories: 1. Reduction of a hard problem (a problem for which) a polynomial time algorithm is not yet designed) into another hard problem. In complexity theory such problems are known as NP-complete problems; 2. Reduction of a problem (with an existing) polynomial time algorithm into another such problem. In complexity theory such problems are known as P problem; 3. Reduction of a P problem into an NP-complete problem; Does it mean that a P problem has become an NP-complete problem? 4. Reduction of an NP-complete problem into a P problem: Does it mean that an NP-complete problem has become a P problem? Why?

34

Problems, Models & Graphs: Why Study Graph Theory?

We shall provide (or discuss) at least one example from each category in this chapter.

3.3

The Satisfiability Problem in Logic Circuits

Satisfiability in logic circuits is the problem of finding if we can assign 0 or 1 to the input variables so as to make the output of the logic circuit equal to 1. If no assignment of input variables can make the output 1 then we claim that the logic formula (or the circuit) is not satisfiable. In the 3CNF Satisfiability problem (or the 3-SAT Problem), we are given a boolean expression in 3-conjunctive normal form; in simple words it is the AND output of clauses of OR gates with exactly three inputs (see Fig. 3.3.1). Given such an expression, the 3-Satisfiability problem is to find if it is possible to assign binary values (0 or 1) to its inputs that will make the output equal to 1. In the 3-DNF Satisfiability Problem, we are given a Boolean expression in disjunctive normal form (DNF); in terms of logic circuits it is the OR output of clauses of AND gates with exactly 3 inputs (see Fig. 3.3.2). Again we need to assign input variable such that the output of the circuit is 1.

a

b c

a b c

a b c

a b c

Figure 3.3.1: A logic circuit consisting of OR gates (each with 3 inputs) and one AND gate. The output is 1 for the selected inputs shown in orange color. A lot many of you might have played with logic circuits; have explored assignments of inputs for which the output of the logic circuit is 1 – but many

The Satisfiability Problem in Logic Circuits

35

of you may not have realized (or truly appreciated) that the 3-CNF problem is one of the NP-complete problems while the 3-DNF is a solvable problem (in polynomial time). Remember in the 3-CNF problem one has to select inputs such that the output of each OR gate is high so as to satisfy the output. In case of 3-DNF problem the output of the circuit will be high provided the output of any AND gate is high. Surprisingly applying DeMorgans laws it is possible to convert a 3-CNF Boolean expression in terms of a 3-DNF – giving a false impression that we can reduce an NP-complete problem into a P problem. We shall discuss this issue (in detail) in the coming paragraphs.
OutPut

a

b

c

a

b c

a b c

a b c

Figure 3.3.2: A logic circuit consisting of AND gates (each with 3 inputs) and one OR gate. The output is 1 for the selected inputs shown in orange color. Problem Set 3.1. Problem 3.1.1. Design an efficient algorithm to solve the 3-DNF Satisfiability problem. Show that it is indeed possible to solve this problem in polynomial time (in fact in linear time). A 3-DNF circuit is shown in Fig. 3.3.2. Problem 3.1.2. Try to use a similar technique to solve the 3-CNF Satisfiability problem. It will not work. The purpose of this problem is to appreciate the inherent hardness of this problem. A 3-CNF circuit is shown in Fig. 3.3.1.

3.3.1

Reducing a 3-SAT Problem into an Independent Set Problem

Let us start with an example from Category 1. We shall reduce the 3-SAT problem into the Independent Set Problem in graphs. Thus in Fig. 3.3.3, the

36

Problems, Models & Graphs: Why Study Graph Theory?

New Problem is the 3-CNF Satisfiability problem: we are given a Boolean expression in 3-conjunctive normal form; in simple words it is the AND output of clauses of OR gates with exactly three inputs (see Fig. 3.3.1). Given such an expression, the 3-Satisfiability problem is to find if it is possible to assign 0 and 1 to its inputs that will make the output equal to 1. The figure also shows a combination of inputs (shown in orange color) for which the output of this logic circuit is 1. Please note that at least one input from each OR gate should be 1 in order to pass the test for Satisfiability.
Inputs

Old Problem The Independent Set Problem in a Triangular Graph

Outputs

Algorithm B Algorithm A

Inputs Outputs

New Problem The 3-Satisfiability Problem

Figure 3.3.3: Reducing the 3-Satisfiability problem into Independent Set problem in a graph. The old problem in Fig. 3.3.3 is the Independent Set problem in a special graph consisting of k triangles with edges connecting certain vertices within different triangles as shown in Fig. 3.3.4. The problem is to find if it is possible to select one vertex from each triangle such that no two selected vertices have an edge in common (they should not be adjacent). In other words we need to find if the size of the Independent Set (a set of vertices with no common edges) in this graph is equal to the number of triangles in the graph. It is obvious that for the size of the Independent Set to be equal to k we have to select exactly one vertex from each triangle such that no two selected vertices are adjacent.

The Satisfiability Problem in Logic Circuits

37

Figure 3.3.4: A graph consisting of k triangles where some vertices from different triangles are adjacent. We need to find if the size of the independent Set is equal to k.

Algorithm 1: The 3-SAT Problem: Find for what inputs, the output is high for the given logic circuit? Input : A Boolean formula (or circuit) in 3-CNF. Output: Yes or No; If Yes then a combination of inputs for which the output is high. 1. We consider this as a new problem: You need to transform it into a known problem in graph theory?

Algorithm 2: The Independent Set Problem: Find if the size of the independent set in a given graph is k. Input : A graph consisting of k triangles. Output: Yes/No; If Yes then select a vertex from each triangle such that no two selected vertices is adjacent. 1. We consider this as a known problem in graph theory

38

Problems, Models & Graphs: Why Study Graph Theory?

OutPut AND

a ¬b c

a b ¬c

¬a ¬b c

¬a ¬b ¬c

c a

¬c

c

¬c

¬b

a

b

¬a

¬b

¬a

¬b

Figure 3.3.5: The 3-Satisfiability problem (top) is reduced into the Independent Set problem in the graph shown in the bottom. Orange vertices in the bottom graph provides a solution to the Independent Set problem in the bottom graph while their orange counterparts (input variables to the logic circuit) provides a satisfiable solution to the 3-SAT problem in the top circuit.

The Satisfiability Problem in Logic Circuits

39

It is obvious that a possible solution of the Independent Set problem in the graph provides a combination of inputs for which the output will be high in the logic circuit. Thus if the size of the Independent Set in the graph is equal to the number of OR gates in the logic circuit then it is possible to find a combinations of inputs which will make the output high. In fact the reduction goes in both directions; the solution of any one problem implies a possible solution for the other. Unfortunately both these problems belong to the class known as NP-Complete problems. In the next section we shall talk about a Category 2 reduction. Problem Set 3.2. Problem 3.2.1. We have already talked about the 3-Sat problem and its reduction to independent set problem as modeled in Fig. 3.3.3, and depicted in Fig. 3.3.5. Please design Algorithm A and Algorithm B as shown in Fig. 3.3.3. Problem 3.2.2. Try to use a similar reduction to reduce the 2-SAT problem into the independent set problem in a graph. Discuss why this may or may not be possible. Problem 3.2.3. Try to use a similar reduction to reduce the 3-DNF problem into the independent set problem in a graph. Discuss why this may or may not be possible.

3.3.2

Reducing the 3-CNF Satisfiability Problem into the 3-DNF Satisfiability Problem

We know that it is possible to convert a 3-CNF Boolean formula into a 3DNF formula. Such a conversion is done after drawing the truth table or the K-Map of the 3-CNF expression as shown in Fig. 3.3.6. We also know that the 3-CNF Satisfiability problem is NP-Complete while it is possible to solve the 3-DNF Satisfiability problem in polynomial time. Is there a contradiction somewhere? Design the corresponding conversion algorithms (Algorithm No. 1 & 2 of Fig. 3.1.1) and you will be able to resolve this contradiction your self.

40

Problems, Models & Graphs: Why Study Graph Theory?

OutPut AND

Dr aw

K

Ma p

OR

OR

OR

OR

ab
a ¬b c a b ¬c OutPut OR ¬a ¬b c ¬a ¬b ¬c

c

0 1
0 0

1
0

00 01 11 10

1
0

1

1

AND

AND

AND

AND

aw Dr

it rc u Ci

a ¬b ¬c

a ¬b c

¬a ¬b ¬c

¬a b c

Figure 3.3.6: The 3-CNF Satisfiability problem (top) is reduced into the 3DNF-Satisfiability Problem as shown in the bottom. The truth table, also shown on the right, helps us in this reduction.

The Satisfiability Problem in Logic Circuits

41

3.3.3

Reducing the 3-CNF Satisfiability Problem into another graph Problem

Here we intend to discuss a reduction of the 3-SAT problem into the Clique problem in graphs. It is left as an exciting exercise for the reader?

3.3.4

Reducing the 2-CNF Satisfiability Problem into a Graph Problem

We have already talked about reducing a 2-CNF Satisfiability problem into an Independent Set problem in graphs. In fact this problem can be reduced to another graph problem which is solvable in polynomial time. We shall describe this Category 2 reduction briefly in this section. We shall partially justify this reduction and leave the rest of the details as an interesting problem for the reader. The 2-SAT problem is similar to the 3-SAT problem except that now each of the OR gates have two inputs rather than three. We still need to find a combination of inputs for which the output of the circuit is 1. The corresponding graph problem deals with directed graphs; we need to find if any two given vertices, x and y, belong to a single strongly connected component; in other words we need to check if there is a directed path from vertex x to vertex y and from vertex y back to vertex x. This problem and its possible solutions are discussed in detail in Chapter 9. Path finding algorithms are described in Chapter 5. Given the Boolean expression or the logic circuit, we construct a directed graph D according to the following rules: 1. For every variable x in the Boolean expression we create two vertices with labels x and ¬x in the directed graph D. 2. For every OR gate in the Boolean expression with inputs x and y, we add two directed edges: One directed edge from vertex ¬x to vertex y and another directed edge from vertex ¬y to vertex x. Once we have a constructed directed graph D, we claim that the given Boolean expression (or the logic circuit) is not satisfiable (that means for any combination of input variables) if and only if any vertex x and its complement vertex ¬x in graph D belongs to the same strongly connected component.

42

Problems, Models & Graphs: Why Study Graph Theory?

¬a b ¬a ¬b c

OR

¬c

OR AND OR
Convert

a

b

a
¬c

¬b

¬a

a

OR

c

Figure 3.3.7: The 2-CNF Satisfiability problem is reduced into a graph problem. If any vertex x and its complement in the graph (shown in the right diagram) belongs to the same strongly connected component then the Boolean expression corresponding to the circuit (shown in the left diagram) is not satisfiable. We apply Algorithm 3 for every vertex x (and its complement ¬x) in graph D, and if the answer comes out to be NO in each case then the given Boolean expression is satisfiable otherwise not. If the Boolean expression is satisfiable then we have to find a combination of input variables for which the output of the logic circuit is 1. This requires a deeper understanding and appreciation of different concepts involved in this reduction. We shall provide some hints in this regard and leave the rest as a problem for imagination of the reader. Algorithm 3: Find if two given vertices belong to the same strongly connected component in a directed graph D. Input : A directed graph D, and two vertices x and y. Output: Yes/No; 1. Check if there is a path from vertex x to y and from y to x in the directed graph D. If yes then vertices x and y belong to the same strongly connected component otherwise not.

A 2-CNF expression consisting of a single OR/AND combination is shown in the left diagram of Fig. 3.3.8. It is transformed into a directed graph shown

The Satisfiability Problem in Logic Circuits

43

a
0
Tr Lo uth gi Ta c C ble irc o ui f t

a→ b ¬a OR b ¬b → b ¬a
0 1 1 0 1 1 1 0 1 1 1 0
Im Trut pli h T ca a tio ble n Gr of ap h

0 1 1

¬a b

AND OR

For what values of input the output is 1

a
¬b

b

¬a

If a = 1 then b=1 for Output to be 1

Figure 3.3.8: A 2-CNF expression with only one OR/AND gate is reduced into a directed graph. The directed graph is in fact an implication graph with a truth table shown at the top. Note that the implication graph tells us that if a = 1 then b should also be 1 otherwise the output of the logic circuit will be zero.

If a = 0 then Output =1

¬a b

OR AND
For what values of input the output is 1

a

b

¬a ¬b

OR
If a = 1 then b=1 then ¬a =1

¬b

¬a

Figure 3.3.9: The 2-CNF Satisfiability problem is reduced into a graph problem. Note that if a = 1 then we face a contradiction in the implication graph. On the other hand if a = 0 then there is no such contradiction and the output of the logic circuit will be 1 for any value of b.

We show various 2-CNF logic circuits for varying number of OR gates and the corresponding implication graphs in Fig. The implication tells us that in order to make the output 1 we shall make b = 1 if a = 1. The Satisfiability problem in this logic circuit is transformed into a graph problem in the Implication graph shown in the bottom diagram. As you should appreciate this directed graph is not just a directed graph – this is in fact an implication graph.10. it is possible to reach from vertex a to vertex ¬a. It should now be possible to design an efficient graph algorithm which operates on the implication graph but which finds the input combination for which the logic circuit output is 1.11.3. This implies that the logic circuit is not satisfiable. The Satisfiability problem is also transformed into an independent set problem where each OR gate in the logic diagram now corresponds to two adjacent vertices (known as an adjacent pair) instead of a triangular graph. and in case it is then what inputs should be applied so that the output is 1 for each logic circuit shown in this figure. Models & Graphs: Why Study Graph Theory? in the right diagram of this figure. 3. A directed edge in this graph from vertex x to vertex y implies that if the vertex x (or meaningfully) variable x is 1 then y is 1.3. We show the same 2-CNF logic circuit in Fig. It will be interesting to find if the circuit is satisfiable (as we add more OR gates). If a = 1 then b should be 1 so that the output of the top OR gate becomes 1. This transformation works both ways: if the circuit is satisfiable then we can select an independent set . Again the implication graph is telling us how it is possible to make the Boolean expression satisfiable. But if a = 0 then the implication does not dictate any thing – it means that the Boolean expression is satisfiable for any value of b.3.3. it is also possible to reach from vertex ¬a to a as shown by the closed path shown in red color.10. Please note that in the bottom diagram of Fig. 3. If it is possible to select one vertex from each adjacent pair such that no two selected vertices are adjacent then we claim that the logic circuit is satisfiable. 3.44 Problems. The truth table of this implication is exactly the same as that of the Boolean expression as shown in this figure. But if b = 1 then according to the implication ¬a should be 1 so as to make the output of the second OR gate 1 – but that is a contradiction. This makes sense because if both inputs of the OR gate are 0 then the expression will not be satisfiable. 3.9. The implication graph for a Boolean expression consisting of two OR gates are shown in Fig.

10: The 2-CNF Satisfiability problem is reduced into a graph problem.3. .The Satisfiability Problem in Logic Circuits 45 a ¬a b AND OR Convert b ¬b ¬a ¬a b OR AND Convert a b ¬a ¬b ¬a b OR ¬b ¬c ¬a OR a AND Convert b ¬a ¬b c OR ¬b ¬a a ¬a b ¬a ¬b c OR c OR ¬c OR AND OR a b a ¬c ¬b ¬a a OR Convert c Figure 3. vertex a and ¬a do not belong to the same strongly connected component – it means the output of the corresponding logic circuits is satisfiable. In the bottom diagram this is not possible. In the top three graphs (shown on the right side).

The same problem is reduced into a graph problem in an implication graph shown in the bottom diagram.3. It is now obvious that the independent set problem in the graph (which was derived from a 2-CNF logic circuit) can be transformed into a path finding problem in an Implication graph. Given a general graph how can you determine that this graph in fact represents a 2-CNF logic circuit? If it does represent such a logic circuit then we can solve this problem after reducing it into an implication graph? ¬a b ¬a ¬b c OR Transform ¬a b OR AND ¬a ¬b a ¬c OR a c a OR a ¬c Tr ¬c a b ¬b ¬a c Figure 3. Models & Graphs: Why Study Graph Theory? with size equal to the number of OR gates in the logic circuit. an sf o Tr an sf or m rm .11: The 2-CNF Satisfiability problem is reduced into an independent set problem shown in the top diagram. This implies that the independent set problem (in some special graphs) can be reduced into an implication graph problem. Describe an efficient algorithm which finds an independent set in such a graph using an Implication graph.46 Problems.

Now for each vertex x in Fig.htm). we have described a number of tasks. the tasks they represent have no pre-requisites) as shown in Fig. In the table below. We need to find out the minimum possible time to complete a house according to the activities described in the table above.e.2. 3. 3. 3. We represent each activity in the table using a vertex. . How will you model this problem in graph theoretic terms and then solve it? We start with an example of constructing a house as defined in Table 3. their IDs.An Activity Scheduling Problem 47 3. the lower limit will be the length of the longest path in the graph. If we are unable to schedule any activity in parallel then the total time needed to construct the house will be the sum total of (the duration of) all activities. The problem is then transformed into a graph problem as shown in Fig. 3.4. We also need to identify those activities which are critical: that means increasing the time duration of these activities will certainly increase the total completion time. with the duration of the activity as a weight on the vertex.1. For each task vertex v. A longest path in this graph (Fig.waa-inc.2. we assign all the incoming edges to x the same weight as the weight of vertex x in Fig. 3. their duration (in days) and what tasks must be completed before they can begin (prerequisites) (http://www.1.4 An Activity Scheduling Problem The problem of activity scheduling is described in simple words.4. So that is an upper limit on time to complete the job. we add a directed edge to all the vertices such that the task they represent are dependent on the completion of task represented by v.4.3) corresponds to minimum amount of time needed to build the house? Why? The underlying assumption is that we should be able to schedule as many activities in parallel as possible by the pre-requisite relationships.4. On the other hand there are certain activities which if delayed do not necessarily increase the total completion time.1.com/projex/PERT/aoa.4. Then there are activities which can run in parallel while some are strictly sequential. Then we add a special vertex s and connect it to all such vertices that have no incoming edges (i.

48 Task ID 1 2 3 4 5 6 7 8 9 10 Problems.1). Each program module in the chain structure may have a very different computation requirement. This kind of computing has a typical serial or chain like structure as shown in Fig. one module may run faster on one processor while the same module may take much longer time on a different machine in a typical distributed . 8 Table 3. Input : Tasks. 9 3 7. 9 2 6 5. their duration (in days) and what tasks must be completed before they can begin (prerequisites). A new problem: You need to transform it into a known problem in graph theory? The challenge is to transform one problem (Algorithm 4) into a text book problem (Algorithm 5) 3.1: We have tabulated a number of tasks. their IDs. Duration. Models & Graphs: Why Study Graph Theory? Task Description Clear land Lay foundations Build walls Electrical wiring Plastering Landscaping Gardening Interior work Roof Handover Duration 14 28 42 21 21 20 10 35 50 00 Prerequisites none 1 2 3 4. 3.5 A Dual Machine Serial Processing Environment Chain like program graphs is common in many digital signal processing applications. Output: Minimum Time to complete the job. In such applications each packet or frame of data may be processed through various transforms in a fixed sequence.5. 1. and Pre-requisites (Table 3.1. Algorithm 4: Find Minimum Time to Complete the Job.

A Dual Machine Serial Processing Environment 49 Algorithm 5: Find Longest Path between two given vertices in a directed acyclic graph D Input : A weighted directed acyclic graph D and two vertices Output: The Longest Path between two given vertices in D 1.4. see Chapter 5 6 20 7 10 1 14 2 28 42 3 4 21 50 9 21 5 8 35 10 00 Figure 3. A text book problem.2: Transformation to a new graph .4.1: Graph representing a specific instance of the activity scheduling problem 10 6 20 21 21 35 00 7 00 14 1 28 2 42 3 4 5 8 10 50 9 21 35 Figure 3.

In such a distributed heterogeneous computing environment we should somehow take advantage of the diverse and special characteristics of each machine as assigning all modules of the program to one machine may not be an optimal solution.3: Finding the longest path corresponds to the optimal scheduling of activities heterogeneous computing environment consisting of two processors as shown in the bottom diagram of Fig. For example the execution cost of module 2 on processor A is 20 while it is 90 on processor B. the total time of computation is the sum of total execution times plus the total communication times.1. A straight forward greedy solution would be to assign a module on a processor where it is least costly but then if two modules with a lot of communication traffic in between are assigned to different machines then it will again degrade the over all performance. The execution cost of each module on either processor is indicated below each module. Let us concentrate on a sub-problem in order to appreciate the intricacies of the problem. 3. Models & Graphs: Why Study Graph Theory? 10 6 20 21 21 35 00 7 00 14 1 28 2 42 3 4 5 8 10 50 9 21 35 Figure 3. The communication cost between module 2 and 3 is 50 provided the two modules are placed on different machines. .5. A module can be processed on either processor A or on processor B but only one processor is active at any time. it is zero if the two modules are assigned to the same machine. The top diagram of the same figure shows a chain structured modular program consisting of four modules.50 Problems.4.

2. while there will be no cost of communication as the two modules are executed on the same machine.5. The execution cost of each module on either processor is indicated below each module. For example the execution cost of module 2 on processor A is 20 while it is 90 on processor B.A Dual Machine Serial Processing Environment 51 Communication Costs 10 Execution Costs Processor B 50 Processor A 70 90 20 50 30 40 20 0 20 80 Processor Processor Figure 3. while there will be no Cost of Communication as the two modules are executed on the same machine.1: (Bottom) Processor A and processor B connected with a high speed communication link. The communication cost between module 2 and 3 is 50 provided the two modules are placed on different machines. There are basically four possibilities for the first pair of adjacent modules to be assigned onto the dual processor system: 1. (Top) A chain structured modular program consisting of four modules (or nodes). The two machines have different capabilities. Module 1 is executed on machine A while module 2 is processed on . Module 1 as well as 2 is processed on processor A: Then Cost of Execution will be 70+20 = 90. 3. it is zero if the two modules are assigned to the same machine. Module 1 as well as 2 is processed on processor B: Then Cost of Execution will be 50+90 = 140.

52 Problems. This is also sequential processing but this time on machine B.5. Problem 3.3. module 2 on processor A. For example assign module 1 on processor B. Find the total cost if all modules are assigned to machine B.3. Problem 3.3. This is standard sequential processing on machine A. The total time of computation is the sum of total execution times plus the total communication times. . while there will be a Communication Cost = 10. Problem Set 3.6 and Algorithm 2. and so on. 2. cost of execution of a module on a processor. Problem 3. How to model this problem in graph theoretic terms? Problem 3. Total Cost = 80. it is not a distributed assignment. 4. Now calculate the Total Cost of this assignment after taking into account the communication costs. Please note that the cost consists of two parts.3. Hints are provided in the following figures. The challenge is to find the optimal solution for the entire problem efficiently (without enumerating all possibilities as we did for the sub-problem). Cost of Execution will be 70 + 90 = 160. and the cost of communication between two adjacent modules provided the two modules are assigned to different machines. How about if we (initially) ignore communication costs and assign a module on a machine where it is least costly. Find the total cost if all modules are assigned to machine A. Cost of Execution will be 50 + 20 = 70.3.4.2. Models & Graphs: Why Study Graph Theory? B. How about if instead of a chain structure. Total Cost = 170. Module 1 is executed on machine B while module 2 is processed on A. Problem 3. The last option provides the optimal solution in terms of minimum completion time for the sub-problem. There is no communication cost if the two (adjacent) modules are assigned to the same machine. Problem 3.3. while there will be a Communication Cost = 10.3.6.8.3. we design Algorithm 2. and 3.1. Find the Optimal Assignment by hit and trial (the size of this problem is small enough)? It should be less than (or equal to) the ones found in 1.

(Bottom) The possibility of some modules assigned to processor A while other modules to processor B is shown with a zigzag path between the two dummy vertices. .A Dual Machine Serial Processing Environment 53 90 50 30 20 0 u4 00 Path B: All Modules on Machine B 70 Path A: All Modules on Machine A 00 0 20 40 80 90 50 10+90 30 50+30 20 0 u4 00 20+20 70 10+20 50+40 20+80 00 0 20 40 80 Figure 3. A shortest path corresponds to an optimal assignment.2: (top) We show the possibility of all modules assigned to either processor A or to processor B.5. This is equivalent to traversing a path which passes though the top vertices or the bottom vertices in a graph with two dummy vertices.

(bottom) The module assignment corresponding to path X in the top diagram. Models & Graphs: Why Study Graph Theory? 90 50 10+90 30 50+30 20 0 u4 00 20+20 70 10+20 50+40 20+80 00 0 20 40 80 50 70 10 90 20 50 30 0 40 20 20 0 80 Figure 3.5. Path X corresponds to some modules assigned to processor A while others are assigned to processor B.54 Problems. .3: (top) Path B corresponds to all modules assigned to processor B while path A corresponds to all modules assigned to processor B.

Cut B corresponds to all modules assigned to processor B.4: (top) A cut in this graph (which disconnects vertex A from vertex B) corresponds to an assignment of modules. Cut A corresponds to all modules assigned to processor A.5.A Dual Machine Serial Processing Environment 55 50 B 70 10 90 50 30 0 20 40 20 0 80 A 20 B B 50 10 70 20 90 50 B 30 0 40 20 20 0 80 A CUT A 70 20 40 80 50 70 10 90 20 50 30 40 20 20 0 80 Cut X 50 90 30 20 CUT B Figure 3. Cut X corresponds to the module assignment as shown in the bottom diagram. .

1. see Chapter 6 . A text book problem. A new problem: You need to transform it into a known problem in graph theory? Algorithm 7: Find Shortest Path between two given vertices in a directed acyclic graph Input : A directed acyclic graph D and two given vertices Output: A Shortest Path between two given vertices in D 1. Models & Graphs: Why Study Graph Theory? Algorithm 6: Find Minimum Cost Assignment Input : Execution & Communication costs of each module on Processors A & B Output: Minimum Time to complete the job & the corresponding module assignment.56 Problems. A new problem: You need to transform it into another known problem in graph theory? Algorithm 9: Find Minimum Cut which disconnects vertex s from vertex t in a graph G Input : An un-directed graph G and two vertices (s and t) Output: A Minimum Cut (which disconnect s from t) 1. See Chapter 5 Algorithm 8: Find Minimum Cost Assignment Input : Execution & Communication Costs of each module on Processor A & B Output: Minimum Time to complete the job & the corresponding module assignment. 1. A text book problem.

Fig. and Fig. GAG. . C. Then we can find a sequence s as shown in Fig. For example if the (unknown) sequence s = CAT GAGT then a set of all substrings of length 3 that s contains (known as Spectrum(s. T GA. CAT A.6. ACAC. an array (also known as the Gene Chip) tells us about all sub-strings of a fixed length that the DNA sequence contains. G. AT G. 3.6 Sequencing by Hybridization in Computational Biology The Sequencing by Hybridization technique depends on the hybridization of target DNA fragment against a very large array of relatively short probes. ACAT. 3)) will be equal to {CAT. CACA. The challenge is to tranform this problem into a know problem in graph algorithms.1.3. 4) = {AT AC. it does not provide us about the order of the strings or their position in the DNA fragment. ACAA. With a given unknown DNA sequence (of four letters A. In fact there are multiple strings possible (s1 = ACACAACAT ACAG & s2 = ACAT ACAACACAG) with the same spectrum as illustrated in these diagrams. 3. 3. given its spectrum. CAAC.6. AACA}. T ACA. The following figures will provide hints to make multiple transformations.6. 3.2. & T ). ACAG. The elements in the spectrum may not appear in the same order and the challenge is to find the string s. AGT }. Another challenge is to design Algorithm 8 and then use the text book Algorithm 9 in order to solve the Minimum Cost Assugnment Problem. The sequencing by hybridization problem in molecular biology (or simply the spectrum problem) can be transformed either into a Hamiltonian path problem or an Euler path problem in directed graphs depending upon whether we map every element of the spectrum into a node or an edge of a directed graph respectively.Sequencing by Hybridization in Computational Biology 57 The challenge is to design Algorithm 6 and then use the text book Algorithm 7 in order to solve the Minimum Cost Assignment Problem. Assume that Spectrum(s.

58 Problems. see Chapter 8 Algorithm 12: Find an Euler Trail in G Input : A graph G Output: An Euler path in G 1. 1. l) of an unknown s. Output: Sequence s. Models & Graphs: Why Study Graph Theory? Algorithm 10: Find a (Correct) Sequence s given its Spectrum(s. A text book problem. l) Input : Spectrum(s. A text book problem. A new problem: You need to transform it into a known problem in graph theory? Algorithm 11: Find a Hamiltonian Path in G Input : A graph G Output: A Hamiltonian Path in G 1. see Chapter 7 Transform the sequencing problem into Hamiltonian path and then into an Eulerian path problems (see the figures below) .

ACAG. ACAA.6. CATA. CACA. 4) = {AT AC. CAAC. CACA. ACAG and ACAT .4) = {ATAC. ACAC. CAT A. ACAC. AACA} ACAC ATAC ATAC CATA ACAT ACAG TACA CACA ATAC TACA ATAC TACA CATA TACA ACAG ACAA ACAT CAAC AACA ATAC CATA ACAT ACAG TACA Figure 3. That is why there are four directed edges emanating from the vertex T ACA. .1: We start with a Spectrum(s. the edges are going to vertices labeled with ACAA. ACAT. ACAA. AACA} and show a directed graph in which every element of the spectrum was mapped onto a vertex.Sequencing by Hybridization in Computational Biology 59 Spectrum(s. CAAC. TACA. ACAT. ACAG. ACAC. A directed edge exists from one vertex to another provided the last three letters of the first vertex match with the first three letters of the second vertex. T ACA.

Models & Graphs: Why Study Graph Theory? Start ACAC ATAC ACAC ACACA CACA ACATAC ATAC ACAC CACA ACATA CATA TACA ACAG ACAA ACACAA CATA TACA ACATACA ACAG ACAA ACAT ACACAACAT ACAT CAAC AACA ACACAAC ACAT Start CAAC AACA ACACAACA Figure 3.1. Both these diagrams show various stages in the reconstruction of a sequence.2: (Left) A Hamiltonian path is shown in bold in the graph of Fig. 3. .6. (Right) Another Hamiltonian path is shown in the same graph. this Hamiltonian path corresponds to the sequence s1 = ACACAACAT ACAG.6.60 Problems. this Hamiltonian path corresponds to a sequence s2 = ACAT ACAACACAG.

ACAC. ACAA. TACA. T ACA. CAAC. the start vertex carries the first three letters of the edge while the end vertex is labeled with the last three letters of the edge. An edge also specifies its end vertices. CAAC.4) = {ATAC. ACAT. ACAA. AACA} and this time we map every element of the spectrum into an edge of a directed graph.Sequencing by Hybridization in Computational Biology 61 Spectrum(s. CACA. 4) = {AT AC. CATA. CAT A. ACAC. For example the directed edge AACA emanates from CAAC and terminates at AACA. ACAG. AACA} ? ? ? ATAC ? ACAG ACAC ACA CACA ACAT AACA ACAA CAC CATA ? ? CAG ACAT TACA TAC AAC CAAC CAA CAT ATA ATAC TAC ACA ACAT CAT ATAC CATA CAT CATA ATA TAC CAT CATA ATA ACA ACAT ATAC ATA Figure 3.3: We start with the same Spectrum(s. ACAG. .6. CACA. ACAT.

Does your problem have an inherent graphical structure? If yes then the transformation may be a lot simple.62 Problems. we can only rely on heuristics.1. It is certainly not possible to design an exact algorithm to do this transformation. Experience can help you a lot in moving forward.6.6. Note that most of these questions may not have a black and white answer. Please note that in both the Euler paths shown in this diagram there is a vertex which is repeated several times while every edge is traversed exactly once. 3. (Bottom) Another Euler path is shown in the same graph.7 Discussion & Problems How can we exploit graph theory in order to solve an unknown problem? In other words how can we reduce the unknown problem into a graph problem? Finding the right reduction or transformation is not easy. you should also try to find why or why not. We shall provide you below a certain sequence of questions to guide your search towards a suitable transformation. Models & Graphs: Why Study Graph Theory? AACA ACA ACAC CAC CACA ACA ACAA CAA CAAC AAC ACA ACAT ACAG TACA ATAC CATA CAG ACA TAC ATA CAT ACA ACAT CAT CATA ATA ATAC TAC TACA ACA ACAA CAA CAAC CAG ACAG ACA CACA CAC ACAC ACA AACA AAC Figure 3. this Euler path corresponds to the sequence s1 = ACACAACAT ACAG. 1. this Euler path corresponds to the sequence s2 = ACAT ACAACACAG. 3. For example in the first example .4: (Top) A Euler path in the graph of Fig.

4. In the first example we need to minimize the total completion time of building a house. 3.Discussion & Problems 63 we have activities and a pre-requisite relationship among themselves. Do components of your problem have a sequence (or an order) which can be transformed into a directed graph? The first example certainly implies a directed graph. Thus this example is also reduced to a directed graph problem. If you decide to map each element of the Spectrum onto a vertex then an edge in this graph would represent what (and on what basis)? Will this be a directed edge or undirected? 5. The last example is not an optimization problem. this naturally leads to a directed graph. The second example is perhaps more interesting. in its first transformation it is transformed into a directed graph. it is rather impossible to capture its essence without directions. In the second example we need to minimize the sum total of execution and communication costs of a modular program. The activities have weights. these weights can be shifted to vertices if each vertex corresponds to an activity in your graph. only a certain sequence (or order) of letters can correspond to a correct sequence s. we need to find an unknown sequence s for which the Spectrum(s. it is much easier to make a special case reduction. Strangely this is equivalent to finding a longest path in a directed acyclic graph as shown in Fig. 4. This problem is reduced to finding a shortest path in a directed graph. The last example has an implicit directed structure.1. 3. Is your unknown problem an optimization problem or otherwise? If it is an optimization problem where you need to maximize (or minimize) some parameter then you need to search for a graph optimization problem. 2. For example in the second example it is much easier to visualize the special case (in terms of a . Do you map each element of your unknown problem into a vertex or an edge of a graph? Example 3 is a good illustration of this decision. This problem is reduced to a decision problem (not an optimization problem) in graph theory whereby we decide if it is possible to find a Hamiltonian Path (or an Euler trail) in a directed graph. l) is given. it is also reduced into a graph problem applicable to an un-directed graph. Sometimes it is difficult to make a general transformation.

The problem is to find out which reduction is correct and which one is false. If it is not possible to find such a Perfect Matching then we should perhaps maximize the number of women each of whom is married to man whom she knows.1. 3.7.4. This reduction depends upon the hypothesis that each vertex-disjoint path between vertex s and t in graph G corresponds to a matching edge in the bipartite graph. Problem Set 3.) This problem can be modeled by a graph consisting of each man and woman as a vertex. we assume that we have a collection of men and an equal number of women. Such a graph is known as a bipartite graph.4. 3. It is a special graph in the sense that an edge between two vertices belonging to the same part (men or women) is not allowed. The graph is shown below. (Please note that here we have not accounted for the amount of liking between a man and a woman – we just claim that a certain woman knows or likes or does not like a certain man. The problem is to find out if it is possible to marry each woman to a man she likes or knows. each women knows (or likes) some of the men.5. Please note that the graph consists of two parts – an A part consisting of men. This reduction depends upon the hypothesis that each edge-disjoint path between vertex s and t in graph G corresponds to a matching edge in the bipartite graph. Please note that the reduction (which ever is correct) not only tells you if a . Models & Graphs: Why Study Graph Theory? graph problem) when all modules are assigned either to processor A or processor B as shown in the top diagram of Fig. One is shown in Fig. The Marriage Problem is a well-known problem in mathematics as well as in any middle class conservative society. Problem 3. In one case you have to design an informal proof and in the other case you have to design a counter example.64 Problems. we reduce the Marriage Problem into a problem of finding maximum edge-disjoint paths between vertex s and vertex t in a graph G. The quantitative aspects of this problem will be discussed in the coming problems. 3. and a B part consisting of women.2. and the other is shown in Fig.2. We show two reductions of this problem.1. This special case more or less resembles the activity scheduling problem discussed in the first example. In the first reduction. In the second reduction the problem is reduced to finding maximum vertex-disjoint paths between the same two vertices in the graph G.7. A liking between a man and a woman is represented by an edge between the two corresponding vertices.

7.Discussion & Problems 65 a1 a2 A a3 a4 b1 b2 B b3 b4 A Matched Edge a1 a2 A a3 a4 b1 b2 B b3 b4 Bipartite Graph C Corresponds Transforms Maximum Matching in Graph C Transforms Graph G with nodes s & t A Path from s to t Maximum Edge-Disjoint Paths in Graph G a1 a2 b1 b2 B b3 b4 a1 a2 a3 a4 b1 b2 s A a3 a4 t Find Max EdgeDisjoint Paths s A t b3 b4 Figure 3.1: We add vertex s and vertex t to a bipartite graph and then try to reduce a maximum bipartite matching problem into another problem where we maximize edge-disjoint paths between vertex s and vertex t. .

Models & Graphs: Why Study Graph Theory? a1 a2 A a3 a4 b1 b2 B b3 b4 A Matched Edge Corresponds a1 a2 A a3 a4 b1 b2 B b3 b4 Bipartite Graph C Transforms Maximum Matching in Graph C Transforms Graph G with nodes s & t A Path from s to t Maximum Vertex-Disjoint Paths in Graph G a1 a2 b1 b2 B b3 b4 a1 a2 a3 a4 b1 b2 s A a3 a4 t Find Max NodeDisjoint Paths s A t b3 b4 Figure 3.7.66 Problems. .2: We add vertex s and vertex t to a bipartite graph and then try to reduce a maximum bipartite matching problem into another problem where we maximize vertex-disjoint paths between vertex s and vertex t.

The permanent P er(C) of a matrix C is computed like the determinant of C except that the signs of all permutations are positive. Here we start with the adjacency matrix C of the bipartite graph C.2. 3. We need to find if an edge (y. Once we know that a perfect matching exists in the given bipartite graph C. 2. Problem 3. z) is part of a perfect matching in the bipartite graph. the problem is how to find one? Assume that the only operation that we can perform is to find a Permanent of the matrix after or before removing an edge of the bipartite graph. The said edge will be part of a perfect matching if and only if P er(Cyz) is non zero. On the other hand if it is non zero then there will be as many perfect matchings as the value of P er(C).Discussion & Problems 67 perfect matching is possible. the reduction maximizes the number of marriages taking place. The problem of finding a perfect matching is thus reduced to finding the permanent of a matrix. 1.4.7. We consider the same Marriage Problem as described before. Thus if P er(C) is zero then no perfect matching exists. An alternate way of finding a perfect matching in a bipartite graph is demonstrated in Fig. We need to find if a Perfect Matching exists in a given bipartite graph C.4. it also tells us which woman to marry whom. We reduce the problem into finding the permanent of the matrix C. On the . In case a perfect matching is not possible. How can you find a Perfect Matching in this graph? Discuss briefly. If a perfect matching does not exist then we need to maximize the number of marriages as before. For example the P er(C) of the matrix shown in the figure below will be computed as follows: P er(C) = a11 a22 a33 + a11 a23 a32 + a12 a21 a33 + a12 a23 a31 + a13 a22 a31 + a13 a21 a32 P er(C) = 0 + 1 + 0 + 1 + 1 + 0 = 3 As you can see each individual term in the expression for Permanent is a permutation – there is a one to one correspondence between each non zero term in this expression and a perfect matching in the bipartite graph. We also need to identify which woman is marrying whom.

Models & Graphs: Why Study Graph Theory? 1 1 2 2 3 3 Pe r fe ct M at ch in g if Pe r(C )n ot ze ro Figure 3.68 Problems. We take the permanent of the adjacency matrix and claim that a perfect matching in the bipartite graph exists if and only if the value of the permanent is non zero.7.3: We show a bipartite graph and its adjacency matrix shown in the top diagrams. .

7.4: If a perfect matching exists in a bipartite graph then it is possible to find if an edge is part of that perfect matching. Find Pe M ct rfe atc hin g .Discussion & Problems 69 Does this edge part of Perfect Matching? 1 0 1 Pa Ed 1 1 1 1 1 0 Adjacency Matrix of C a1 Bipartite Graph C b1 b2 B b3 ge 1 0 1 1 1 1 1 1 0 rt of a2 A a3 0 1 1 1 Pe rf e ct M at c hi ng Ex is ts Find Permanent of C not zero If Permanent is not zero Figure 3.

70

Problems, Models & Graphs: Why Study Graph Theory? basis of this observation, design an efficient algorithm to find all edges belonging to the perfect matching.

3. It is much more complex to find the permanent of a matrix as compared to finding the determinant of a matrix – thus we are tempted to check if the determinant is as helpful as the permanent of a bipartite graph? How about if we find that the determinant of an adjacency matrix of a bipartite graph is non zero? Under such conditions, can we make a claim that a perfect matching exists in the graph? How about if the determinant is zero – can we claim that a perfect matching does not exist in the bipartite graph? Remember that the determinant of a completely connected bipartite graph is zero in spite of the fact that every permutation of the vertices of the bipartite graph is a perfect matching.

4. We know that the determinant of a completely connected bipartite graph is zero, why? How about if we put random weight on each edge of the bipartite graph? The probability of the value of the determinant becoming zero will depend upon the randomness of the weights assigned to the edges of the bipartite graph. So hopefully the determinant of a bipartite graph will not be zero unless there is not a single perfect matching in the bipartite graph. How to find the edges of a perfect matching in this graph? How about using the previous technique?

Problem 3.4.3. Consider the (minimum weight) assignment problem. We are given a complete balanced weighted bipartite graph, and we need to find a minimum weight perfect matching in this graph – that means a perfect matching in which the sum of weights of all (matched) edges is minimum. For the time being assume that the minimum weight perfect matching is unique. We shall try to relax this condition later.We use the same reduction in this problem as was used in the last problem with one slight modification. All edge weights in the adjacency matrix C are raised to the power of 2, and then we take the permanent of matrix C. The highest power of 2 which divides the value of the permanent is the weight of the minimum weight perfect matching.

Discussion & Problems

71

P er(C) = 28+1+8 + 28+8+8 + 28+8+8 + 28+8+2 + 23+1+2 + 23+8+8 = 217 + 224 + 224 + 218 + 26 + 219 = 26 (1 + 217 + 224 + 224 + 218 + 219 ) Raising all weights to a power of two and then finding the permanent of the weighted matrix gives us a powerful reduction in which each individual term in the permanent is in fact the weight of a perfect matching raised to power of 2. The value of the permanent will be divisible (with remainder equal to zero) by the weight of minimum weight perfect matching raised to the power of 2 provided we have a unique minimum weight perfect matching in the bipartite graph. 1. Design an efficient algorithm to find edges belonging to the minimum weight perfect matching. The only reduction that you can use is to find permanent of a matrix. Of course you can raise a number to the power of 2 or any other number of your choice. 2. What complications can arise in finding the value of minimum weight perfect matching provided such a matching is not unique? 3. Assume that now there is a possibility that the minimum weight perfect matching is not unique. You need not find the value of the minimum weight perfect matching – just make an efficient check if it is unique or not?

72

Problems, Models & Graphs: Why Study Graph Theory?

All other edge weights are 8

a1
3 2 1

b1 b2 B b3

8 8

8 8 8 8

8 2 8 2 2 2

8 2 1 2 8 2

3 2 8 2 8 2

a2 A a3

Transforms

Weighted Bipartite Graph C

Weighted Adjacency Matrix of C
Find

a1
3 2 1

b1 b2 B b3
The weight of minimum weight perfect matching
The highest power of 2 which divides the permanent will be

a2 A a3

1+2+3=6

6

217 + 224+ 224+ 218+ 26 + 219

Figure 3.7.5: A minimum weight assignment problem is reduced to evaluating the permanent of a matrix. Problem 3.4.4. Consider the assignment problem in which we have a complete weighted bipartite graph and we need to find a minimum weight perfect matching. In Chapter 5 we shall reduce this problem into a well known shortest path finding problem in any directed or un-directed graph. Shortest path finding algorithms are relatively simple and are discussed in Chapter 4. Here in this problem we shall discuss a (bizarre) reduction in which we reduce the shortest path problem into the assignment problem. We start with a black box which accepts a complete weighted bipartite graph as input and outputs the minimum weight perfect matching. The output also includes edges (or vertices) belonging to minimum weight perfect matching. We show a weighted graph G in Fig. 3.7.6 where we need to find a shortest path from vertex a to vertex d. The weighted adjacency matrix of this graph is also shown in the top right diagram of this figure. Using this pp matrix we

Discussion & Problems

73

a
8

2 4

b

a a b c
0 2 5 8

b
2 0 3 4

c
5 3 0 6

d
8 4 6 0

3 5

d

6

c

d

Given a weighted graph G find a shortest path between a to d

Adjacency Matrix of weighted graph G

a
5

2

b
3

0

b
3

b a b c
Transform

Transforms

c
5 3 0

d
8 4 6

2 0 3

c
6

0 4

c
8

d

Figure 3.7.6: A weighted graph G is shown in the top left diagram. We need to find a shortest path between vertex a and vertex d. The adjacency matrix of a weighted graph G is shown in the top right diagram. The shortest path problem is converted into a minimum weight perfect matching problem as shown in the bottom diagrams.

74

Problems, Models & Graphs: Why Study Graph Theory?

generate a weighted (complete) bipartite graph of size (p−1)(p−1) by deleting the column corresponding to vertex a and the row corresponding vertex d as shown in the bottom left diagram. We claim that a minimum weight perfect matching in this bipartite graph gives us a shortest path between vertex a and vertex d as shown in Fig. 3.7.7.
a
8 5 2 4 3

b

0 2 5

2 0 3 4

5 3 0 6

8 4 6 0
Transforms

2 0 3

5 3 0

8 4 6

d

Given a weighted graph G find a shortest path from a to d

Adjacency Matrix of Graph G

a a
8 5 2 4 3
Transform

2 5

a
5

Transform

6

c

8

2

b b

0 3

b
3
Transform

b
3

0

b
3

c c
6

d

0 4

c
8

c
6

0 4

c
8

6

A Shortest Path from vertex a to d

d
Find a Minimum Weight Perfect Matching

d
A Complete Weighted Bipartite graph

Figure 3.7.7: The shortest path problem is converted into a minimum weight perfect matching problem as shown in the top diagrams. We find a minimum weight perfect matching in the complete bipartite graph shown in the bottom right diagram. The minimum weight perfect matching (shown in the bottom middle diagram) provides a shortest path between vertex a and vertex d as shown in the bottom left diagram. Consider Fig. 3.1.1, here the New Problem is finding shortest path between two vertices in a graph G while the Old Problem is the minimum weight

Discussion & Problems assignment problem.

75

1. Design an efficient algorithm to transform this problem (of finding a shortest path between two vertices in a weighted graph) into a complete bipartite graph as shown in Fig. 3.7.7. In terms of Fig. 3.1.1, design Algorithm No. 1. 2. Design an efficient algorithm to transform the minimum weight perfect matching into a shortest path between the two given vertices; in terms of Fig. 3.1.1, you will be designing Algorithm No. 2.

76

Problems, Models & Graphs: Why Study Graph Theory?

Chapter 4 Basics of Graph Theory

4.1 4.2 4.3 4.4 4.5 4.6 4.7 4.8 4.9

Introduction A Mutual Friendship Graph Representation of a Graph Complement of a Graph Equal Graphs & Isomorphic Graphs The Degree Sequence Walks, Trails, & Paths Multi-graphs and Pseudo-graphs Broad Categories of Graphs & some Special Graphs

4.10 Integration of Concepts, Properties, and Action Items 4.11 Self Complementing Graphs

78

Basics of Graph Theory

4.1

Introduction

We shall discuss a number of basic definitions in this chapter. We will be solving a couple of puzzles; hopefully the students will realize that in such problems a stage comes when common sense alone is not sufficient to solve the puzzle; we need tools and techniques of graph theory. We shall discuss some necessary conditions for a degree sequence to be graphical. We shall then talk about graph isomorphism and then come back to a discussion of necessary and sufficient conditions for a degree sequence to be graphical. The concept of graph connectedness will also be provided; we shall define a walk, a trail and a path in a connected graph. We shall also talk about some special graphs at the end of this chapter.

4.2

A Mutual Friendship Graph

Assume that we have 5 persons in a hall. Let us name these people as A, B, C, D, & E. Some of these people are friendly with each other while others are not; we assume that their friendship is symmetric; that means if person x knows y then it means that y knows x (for every pair (x, y)). We ask each person with whom he or she is friendly with and we get an answer like A is friendly with B only while B is friendly with A, D, and C. This so called mutual friendship is represented by the top left diagram in Figure 4.2.1. Please note that the double sided arrows emphasize the mutual friendship between two persons; as discussed before we are talking about a symmetric relationship. With each person in this diagram we show the number of people with whom he (or she) is friendly with; thus A is friendly with just one person while B is friendly with three persons. This number in fact represents the amount of popularity a person enjoys; Persons B and D are the most popular persons while persons A and E are the least popular. If we note down this popularity number, and then sort this sequence we get the so called friendship sequence; this sequence is also shown in the top left diagram. The top left diagram in Fig. 4.2.1 is transformed into a graph shown in the top right corner of the same figure. Here each person is represented by a node or a vertex; a double arrow edge is represented by an un-directed edge; the relationship of friendship in the top left diagram is now transformed into an adjacency relationship in the graph shown in the top right diagram. Now instead of a friendship sequence we have a degree sequence; this sequence is

Representation of a Graph

79

exactly the same as the friendship sequence and is shown in the top right corner as well.

Figure 4.2.1: We show five persons with a symmetric friendship relationship indicated by lines with double sided arrows in the top left diagram. The relationship is transformed into an un-directed graph as shown in the top right diagram. The friendship sequence as well as the degree sequence is also indicated along with the respective diagrams. The adjacency matrix representing the graph is shown in the bottom diagram.

4.3

Representation of a Graph

The graph shown in the top right corner in Fig. 4.2.1 is modeled by an adjacency matrix data structure as shown in the bottom diagram of Fig. 4.2.1. In this adjacency matrix a 1 represents an adjacency relationship while a

Is the husband telling the truth assuming the guests told the truth? How about if he answers 3 (he showed no discrimination). Salma’s husband Aslam (known as A) received the guest when they have arrived while Salma was in the kitchen preparing food. he answers zero (he does not shake hands with females). It will be interesting to explore the (special) structure of the friendship sequence (or the degree sequence). the smallest number can not be less than zero. The sum of all numbers in the sequence will always be an even number. The odd numbers in the sequence will appear even number of times. .1.2. In a group of five persons (as shown in Fig. for example the degree of vertex A is 1 while the degree of vertex B is 3. she gets the following answer: B says 3. Why? 3. The degree of each vertex is also shown in the last column of the bottom diagram. C says 2 and D says 1. Find out the number of times the husband actually shook hands and more importantly with whom he shook hands and whom he has ignored.1. Aslam shook hands with some of the guests.80 Basics of Graph Theory zero (or an absence of 1) represents no adjacency relationship between the corresponding nodes in the graph. Why? Problem Set 4. The sum of all one’s in a row corresponds to the degree of a vertex. Salma invites three of her office colleagues for dinner. there is neither any room for negative friendship with some one which means animosity. Why? 2. This is because in our understanding of friendship there is no room for self friendship.1) the largest number can not be larger than four. what does that mean? Under such conditions the sequence will neither be a friendship sequence. Now Salma asks the same question to her husband. At least one number will be repeated in the sequence. from now onwards we shall use the term “the sequence”. Obviously Salma was curious about whom Aslam shook hands and with whom he did not. When we sort the numbers in the last column then we get the degree sequence or the friend-ship sequence.1. Problem 4. In addition to these we have other interesting limitations as discussed below. what does that mean? 1. 4. Please note that if these conditions are not met then the sequence will not be graphical. She can not ask this question directly so she simple asks every colleague about the number of hands she had shaken.

No one shook hands with one self or with his or her spouse. Salma & Aslam has invited three married couples. Several handshakes took place when the guests have arrived. Surprisingly each person gave a different answer but Aslam does not know which answer belongs to whom. It is not yet obvious where Salma is and where Aslam is? Each bigger circle in these diagrams contains a husband and wife pair. This time Aslam asked each person including his wife to write on a slip of paper how many hands he or she has shaken. The diagram on the right shows the next stage. (b) If answer to (a) is no then is it possible to seat the people around a table so that every two neighbors are acquainted as far as possible? (c) Determine if you can seat these people around a table so that every two neighbors are not acquainted D H G H G .2.1: We show some intermediate stages (and hints) in the solution of Problem 4. The diagram on the left shows the graph in the making.1.3. the motivation behind this seating strategy is to create a lively atmosphere.1.2. The acquaintance graph of a group of eight people is given in the following diagram. Now Aslam is curious about the following: (a) Did a female shake hands with a male? (b) Did Salma shake hands with a male friend? (c) How many times Salma shake hands? (d) How many times Aslam shake hands? (e) Assume that Salma has not shaken hands with any male friend then did Aslam shake hands with a female? Under such conditions was there a husband who refused to shake hands with a female? A B A B C C D F E F E Figure 4. Problem 4. (a) Determine if you can seat these people around a round table so that every two neighbors are acquainted.1.3.Representation of a Graph 81 Problem 4.

4 Complement of a Graph Complement of a graph G is another graph H with the same number of vertices such that there is an edge uv in H if and only if there is no edge uv in G. Now try to visualize a graph E having .3. we should be able to draw an actual graph with the same degree sequence. 4. The first problem set (some parts of this at least) also exploits some of the necessary conditions.1 on the left most graph and the middle graph respectively Try to visualize a graph C having same number of vertices as in G or H and having an edge uv provided the edge uv exists either in G or in H. the motivation behind this seating strategy is to encourage strangers to become acquainted with each other. we shall need some of its (graph isomorphism) results for formulating necessary and sufficient conditions for a graphical sequence. Determine if you can seat these people around a round table so that every two neighbors are acquainted.2: The acquaintance graph of a group of eight people is given.82 Basics of Graph Theory as far as possible.4. it will be useful and informative if in each problem you try to actually draw the underlying graph. We have already discussed a number of necessary conditions for a sequence to be graphical. We show a graph G and its complement H in Fig. 4. F H A B G E D C Figure 4. We shall explore necessary and sufficient conditions for a sequence to be graphical but before we do that let us first discuss some more basic stuff (in general) and graph isomorphism (in particular).

All these graphs have the same number of vertices. Thus these two graphs are not equal. 4.4. Note that vertices with the same degree in two different graphs are colored similarly.5.5. If now you draw the adjacency matrices of the graphs shown on the extreme left and middle left in Fig.2.1 after identical labeling then the two matrices comes out to be different as shown in Fig. the degree of each vertex is also indicated along with each vertex.1. 65433210 A B 76544321 A B 77777777 A B C D C D C D F E H Figure 4. If you closely look at the two middle graphs you realize that they are different graphs meaning that they are neither equal nor isomorphic. The graph on the extreme left and the graph on the extreme right are in fact equal.Equal Graphs & Isomorphic Graphs 83 same number of vertices as in G or H and having an edge uv provided the edge uv exists in G and in H. if you draw an adjacency matrix of each graph (after identical labeling) then you will find that the two adjacency matrices are exactly the same. H H G 4.1: A Graph is shown in the left diagram and its complement is shown in the middle. and same number of edges. A completely connected graph is shown in the right diagram. The degree sequences of both these graphs are also indicated.5.5 We show 4 graphs in Fig. in fact they all have the same degree sequence as shown in the same figure. they may still be isomorphic because we fail to find a visible G F E Equal Graphs & Isomorphic Graphs G F E . 4. This is because in one graph all the four degree vertices are connected in the form of a triangle while in the other graph this is not so. 4. So there is a good possibility that they are all equal to each other.

There is still a possibility that the two graphs are isomorphic to each other as the two graphs have a number of similar qualities and no obvious differences.5.84 Basics of Graph Theory Figure 4.1: We show four graphs each having the same degree sequence.5. Out of these graphs at least two are equal graphs. . This is evident from the adjacency matrices of the two graphs shown in the bottom diagram. and at least two are isomorphic but not equal. Please note that vertices with the same degree are drawn in the same color. Figure 4. at least two are unequal graphs (why?). at least two are non isomorphic graphs (why?).2: Graph G shown in the top left corner is not equal to the graph H shown in the top right corner.

Let f be an isomorphism from the left graph (G) to the right graph (H).1? These two graphs may still be isomorphic? What does that mean? Two graphs G and H are isomorphic provided they can be drawn with identical graph drawings. It is quite obvious now that the drawing in the top left diagram is the same as the drawing shown in the top right corner.5. 4.5.3. If. Then the two graphs are isomorphic provided there is an edge between f (u1 ) and f (u2 ) in the right graph if and only if there was an edge between u1 and u2 in the left graph for every two vertices in the two graphs as shown in the bottom diagram of this figure. 4.Equal Graphs & Isomorphic Graphs 85 difference like the one that we have found for the two middle graphs of Fig. 4.2 is redrawn here in the top right corner after rotating it by an angle of 180 degrees.2 by an angle of 180 degrees then we get a drawing which is exactly the same as the graph G.3: Graph H shown in the top right diagram of Fig. this is shown in Fig.5. Figure 4. 4.5. .5. for example. we rotate graph H of Fig.

and the same degree sequence. Basics of Graph Theory Problem 4. Find a visible difference incase the two graphs are not isomorphic.3.2.5.4. which two are isomorphic. which are isomorphic but not equal and which are not isomorphic (and also not equal).86 Problem Set 4. There is a possibility that the degree sequence may not be graphical? Under such conditions .4: Eight graphs with same number of vertices. Problem 4. 4.2. Find which two graphs are equal (and isomorphic).5. Find an isomorphic function from a graph G to graph H in case graphs G and H are isomorphic to each other. We show eight graphs with the same degree sequence in Fig.2. Figure 4.1.2.6 The Degree Sequence Assume that we are given a degree sequence of a graph and we need to find the corresponding graph provided the sequence is graphical. and which two are not isomorphic. 4. Draw as many graphs as possible such that no two of them should be isomorphic to each other and each graph should have a degree sequence 4443322. Problem 4. Draw as many graphs as possible such that no two of them should be isomorphic to each other and each graph should have a degree sequence 332222.2. Find which two of them are equal.

4. The resulting sequence will be 443322 as shown in the top middle diagram of Fig. this transformation makes no sense unless we visualize these operations as if they are performed on a graph. Before moving forward find a sequence which satisfies all necessary conditions that we have discussed (earlier in this chapter) but it is impossible to draw a graph corresponding to this sequence. Now when we perform the first operation of removing the maximum degree from the degree sequence.6.The Degree Sequence 87 we claim that it is not possible to draw a graph for that sequence (why it is not possible?). Remove the maximum degree (which is 4) from the degree sequence SG (thus reducing the length of the sequence from seven to six). How about the degree sequence 543211? Let us assume we are given a degree sequence SG equal to 4443322. the new graph H will have number of vertices one less than G.1 2. Assume that we have a graph G in which the highest degree vertex is known as v and its degree is u. 4. We apply the following procedure on this degree sequence SG and convert it into a new degree sequence SH which is equal to 332222. The new sequence SH will become 332222 shown in the same diagram. The maximum degree here is 4 and the total number of integers in the degree sequence is 7. Subtract 1 from the first 4 (because maximum degree was 4) integers of the remaining sequence as shown in the top right diagram of Fig.6. Input : Original degree sequence SG (Example 4443322) Output: New degree sequence SH (Example 332222) 1. Also assume that (the highest degree) vertex v is connected to the first u vertices (after v) in the degree sequence SG of G as shown in the bottom left diagram of Fig. Please note that removing the vertex v from G means that all .6. So we need to study necessary and sufficient conditions for a sequence to be graphical. Algorithm 13: Convert degree sequence SG into SH . 4.1.1. Algorithm 13 transforms a degree sequence into another degree sequence. that amounts to removing the vertex v from the graph G. we have already studied some necessary condition at the start of this chapter but those conditions were not sufficient (for a sequence to be graphical).

6.1: We are given a degree sequence equal to 4443322 shown in the top left diagram.6. We claim (Havel-Hakimi) that the original sequence SG (corresponding to a graph G) is graphical if and only if the new sequence SH (corresponding to graph H) is graphical. The new sequence will be shorter by 1 as compared to the original sequence as shown in the top right diagram. Claim 4. In order to prove this necessary and sufficient condition we have to make and prove two claims as follows: Claim 4. Figure 4.2. If the original sequence SG is graphical then the new sequence SH is graphical. The maximum degree here is 4 and the total number of integers in the degree sequence is 7. You may have realized that the new graph will have a degree sequence equal to SH . We remove the maximum degree from the sequence and subtract 1 from the first 4 integers of the remaining sequence as shown in the top diagrams.6. If the new sequence SH is graphical then the original sequence SG is graphical.88 Basics of Graph Theory edges emanating from G will also be removed and that amounts to operation number 2. It is interesting to note that the above claims not only provides us a necessary and sufficient condition for a sequence to be graphical. they also provide us means to draw a graph corresponding to a graphical sequence.1. This meaningful interpretation is possible provided we have the crucial assumption: the highest degree vertex v in G is connected to the first u vertices (after v) in the degree sequence SG of G. This is equivalent to reducing the degree by one of all adjacent vertices of v in G. So before proving the above claims let first do the more interesting exercise of finding a .

you can draw the corresponding graph (output “Yes”. and then find the actual graph corresponding to this sequence. a stage comes when it is very much possible to draw the corresponding graph as shown in Fig. 1. 2. If you get a number less than zero in the new sequence then the original sequence was not graphical (output “No”. and eventually into a 7 vertex graph as shown in the middle left diagram of Fig. 4. Either by carefully looking at the latest new sequence. and subtracting 1 from the first remaining u degrees of the degree sequence). In this figure the seven digit degree sequence is converted into a six digit degree sequence and then ultimately into a four digit degree sequence as shown in the top diagrams. this has been illustrated by the bottom diagram of the same figure. It is then converted into a 5 vertex graph.2. 4.The Degree Sequence 89 graph. Each iteration in the above procedure makes the degree sequence smaller in size. The four digit degree sequence can be recognized to be a graphical sequence. and apply the following steps which outputs whether the sequence SG is graphical or not. and terminate) or if it is not possible then repeat step (1) on the latest new degree sequence (but sort it if it is not already sorted).6.6. The resulting graph shown in the bottom left diagram is different from the middle left graph shown in the same figure. The two graphs are neither equal nor isomorphic (why?). We start with a p length degree sequence. It is very much possible to have two different graphs corresponding to the same degree sequence. The corresponding 4 vertex graph is shown in the middle right diagram. and terminate) otherwise there are two possibilities. . verify that it is graphical.2. Let us start with the same degree sequence. transform the sequence into a new sequence using Algorithm 13 (by removing the first vertex v from the degree sequence with a degree equal to u.

We get a four vertex graph which is then converted into a five vertex graph and ultimately into a seven vertex graph as shown in the middle and bottom left diagrams.1. This four vertex sequence is graphical as shown by the middle and bottom right diagrams. Let us now take up Claim 4. and then ultimately into a four vertex sequence as shown in the top right corner. This seven vertex sequence is converted into a six vertex sequence. it says that if a new sequence SH is graphical then the original sequence SG is also graphical.6.90 Basics of Graph Theory Figure 4. Let us design an algorithm to construct a graph G corresponding to an original sequence SG provided the new sequence SH and its corresponding graph H is given.6. . For example if 332222 is graphical then 4443322 is graphical. It is very much possible to get two different graphs from the same degree sequence. Can you design a formal proof for this claim? Is your proof based on induction or is it proof by contradiction? Discuss briefly.2: A seven vertex degree sequence is shown in the top left corner.

2.6. Problem 4. (2) Original degree sequence SG is also given (Example: 4443322). this is certainly a serious deficiency of this algorithm? Before correcting this deficiency let us look at it a bit more closely in the next part. Add a new vertex v in the given graph H. If the assumption is not true then this algorithm will not provide correct results.1 then you should be able to prove this claim also.3. 4.3. Output: Original Graph G corresponding to the degree sequence SG . We are also given the new sequence SH and we need to show that it is graphical. Input : (1) New degree sequence SH (Example 332222).6. and its graph H is given. If you can prove Claim 4. Note down the highest degree in the degree sequence SG .2. it says that if the original sequence SG is graphical then the new sequence SH is graphical. It essentially means that now we are given a graph G corresponding to a degree sequence SG . just move backwards in Fig. this means that we need to draw the corresponding graph H. Add an edge between the vertex v and the appropriate u vertices in the degree sequence SH (why?).3. Problem 4. Does Algorithm 14 perform the intended function correctly? Problem 4. Let us now take up Claim 4. The said algorithm provides a correct solution provided we have the crucial assumption: the highest degree vertex v in G is connected to the first u vertices (after v) in the degree sequence SG of G.6. let it be u.6.The Degree Sequence Algorithm 14: Find a graph G corresponding to an original degree sequence SG provided the new sequence SH and its corresponding graph H is given. (the degree of v should be u (why?)) 3.3. 91 1.1. 4.2. Problem Set 4.3. Is it possible to apply .2? Can you design an algorithm similar to (or almost a mirror image of) Algorithm 14. Does Algorithm 15 perform the intended function correctly? This is an important question because the answer may be no.3. Assume that we are given a degree sequence 4443322 and a graph G as shown in the left diagram of Fig.

Now repeat the above part with the same degree sequences but this time G is as shown in the right diagram of Fig. The following algorithm performs this transformation. What will be the result of our algorithm? Why our algorithm fails this time? Discuss briefly. Suppose that we are given a graph F and its degree sequence SF . and this vertex is not connected with the first u vertices in the degree sequence (after v1 ) of this graph. Let this vertex be v and let its degree be equal to u.3. Input : (1) Original degree sequence SG (Example 4443322). Please note that there are three vertices in this graph with a highest degree equal to four. the maximum degree vertex v1 has a degree u. 1 (Note that graph H corresponds to degree sequence SH ?) Algorithm 15 in order to draw a graph with a degree sequence equal to 332222? Please note that there are three vertices with a highest degree equal to four. Locate the vertex of highest degree in G.92 Basics of Graph Theory Algorithm 15: Find a graph H corresponding to a new degree sequence SH provided the original sequence SG and its corresponding graph G are given. 1. 4. Problem 4. We claim that such a transformation is always possible. Output: Graph H corresponding to the degree sequence SH . 3. (2) New degree sequence SH is also given (Example: 332222). none of these vertices are connected to the first four vertices in the degree sequence of this graph. 2. We need to convert the graph F into another graph G with the same degree sequence but in G the vertex v1 (having highest degree equal to u) is connected to the first u vertices in the degree sequence of this graph. Remove all edges emanating from the vertex v.6.3.4. Remove the vertex v from the resulting graph and we obtain H (why?). two of these vertices are not connected to the first four vertices in the degree sequence of this graph. Why? Once it has been established by Algorithm 16 that a given graph F can . and its graph G is given.

In the graph H. and this vertex is not connected with the first u vertices in the degree sequence (after v1 ) of this graph. vk ) & (vj . graph H is now transformed into graph G? . 1. and this vertex is connected to the first u vertices in the degree sequence of this graph. vj ) & (vk .The Degree Sequence 93 Figure 4. the maximum degree vertex v1 has a degree u. vn ). the maximum degree vertex v1 has degree u. locate a vertex vn such that vj is connected to vn while vk is not connected to vn . vn ). Output: A graph G with the same degree sequence. (Why are you guaranteed to find such a vertex vn ? 3. remove edges (v1 .6. (If you can not find such vertices then graph F is already transformed into G.3: Two graphs with the same degree sequence 4443322. Please note that the two graphs are not isomorphic. From graph H. there is no need to do any thing else) 2. Locate vertex vk and vertex vj in the graph F such that v1 is connected to vertex vk and not connected to vertex vj while the degree of vj is larger than that of vk . Algorithm 16: Convert graph F with a degree sequence SF into a graph G with the same degree sequence (but with an important difference?) Input : A graph F and its degree sequence SF . and insert edges (v1 .

Let us summarize our recent findings: 1. 3. The . 4. It is possible to convert this graph G into another graph with the same degree sequence by using Algorithm 16 as shown in the top right diagram of Fig.6. Given a graph G in which the maximum degree vertex is not connected to the first u vertices in the degree sequence of this graph (where u is the degree of the highest degree vertex) it is not possible to apply Algorithm 15 in order to transform G into another graph H. Please note that the three algorithms perform a dual purpose.4. We show the position of our claims and the respective proofs. 2. What does appropriate mean? Carefully read your text book (page 17) and then make your decision appropriately.6.6. Algorithm 16 has also made it possible to design a constructive proof for Claim Number 2.6. Given a graph G in which the maximum degree vertex is connected to the first u vertices in the degree sequence of this graph (where u is the degree of the highest degree vertex) it is possible to apply Algorithm 15 in order to transform G into another graph H. Problem Set 4.94 Basics of Graph Theory always be transformed into another graph G. Concentrate on the third line of Algorithm 14: “Add an edge between the vertex v and the appropriate u vertices in the degree sequence SH (why?)”. 4.5. Under such conditions we first transform G into another graph G using Algorithm 16 and then Algorithm 15 will correctly transform G into H. What is that constructive proof? We show our strategy in handling the necessary & sufficient conditions for a degree sequence to be graphical in Fig. As shown in this figure the highest degree vertex is connected to a vertex of lowest degree while it is not connected to a vertex of a relatively higher degree.6. We are given a graph G with a degree sequence 432221 as shown in the top left diagram of Fig.4.2.1. Problem 4. 4. Problem 4. they help us in transforming one graph into another but more importantly they provide crucial insight in designing constructive proofs for the two claims. it gives rises to a number of important conclusions.4. Remember we have discussed this problem in the class but have not resolved it completely.

. All these three vertices (with a degree of 4) are connected to a vertex of minimum degree.The Degree Sequence 95 Figure 4.6. v1 is connected to two vertices with degree 4 and two vertices with degree three. Here there is at least one vertex (v1 ) with a (highest degree equal to four) which is connected to vertices with higher degrees only. it is not connected to a vertex of degree 2. The graph in the right diagram has the same degree sequence but there is an important difference. There are three vertices in this graph with a highest degree equal to four. the minimum degree in this graph is 2. The graph in the left diagram is converted into a graph shown in the right diagram by deleting two edges and by inserting two edges (shown in bold).4: A seven vertex graph is shown in the left diagram with a degree sequence shown in the top of the graph.

the minimum degree in this graph is 2. All these three vertices (with a degree of 4) are connected to a vertex of minimum degree. .96 Basics of Graph Theory Figure 4.6.5: A seven vertex graph is shown in the left diagram with a degree sequence shown in the top of the graph. There are three vertices in this graph with a highest degree equal to four.

i.3.4. j. Problem 4. The new graph with the same degree sequence and the problem completely resolved is shown in the bottom diagram. The figure down below may be helpful for your imagination. j. f. Do we know how to draw the graph for SH provided we have a graph for SG ? Discuss briefly. f − 1.The Degree Sequence 97 problem (that the highest degree vertex is connected to (relatively) lower degree vertices) is still not completely resolved as is evident from the middle left diagram of the same figure. i. Is it possible to do some thing in this specific problem so that the problem is resolved just by applying the said algorithm only once? Can you generalize your findings? How much can you save in time in the worst case analysis? Figure 4. As you may have noticed in this specific problem we have to apply the said algorithm twice to obtain the desired results. k and assume that a = 5. d. h. k. c. d − 1. We again apply the same algorithm to resolve the rest of the problem as shown in the middle right diagram. Let a degree sequence consisting of 11 numbers be SG = a.6: A graph with a given degree sequence shown in the top left diagram is converted into another graph with the same degree sequence shown in the bottom diagram. b. g. h.6. e − 1. g. . c − 1. Let another sequence be SH = b − 1. e.

We are given a degree sequence SG . If the degree sequence is graphical then you are supposed to draw a graph G corresponding to this sequence such that the . the highest degree vertex v (with a degree equal to u) will always be connected to the first u vertices in the degree sequence SG ? Problem 4. Do we know how to draw the graph for SH provided we have a graph for SG ? Discuss briefly.4. Let another sequence be SH = b. h. d. In this graph G the highest degree vertex v (with a degree equal to u) will always be connected to the first u vertices in the degree sequence SG .6. i. Both these graphs have the same degree sequence which is 54444411111. b. If the answer to the above problem is no then how can we do some thing to make sure that in the final graph. f. Discuss briefly.7: Two graphs are shown in this figure. h − 1. The graph shown in the left diagram is a connected graph while the graph shown in the right diagram is a disconnected graph.4.7. We check if the degree sequence is graphical. i − 1.4. Please verify if this is always right or wrong in general. g. d. j − 1. j.6. c. Let a degree sequence consisting of 11 numbers be SG = a. k and assume that a = 5. e. k − 1. if it is graphical then we find and draw the actual graph G corresponding to this sequence. e.5. Figure 4. The figure down below may be helpful for your imagination. Problem 4. c. g − 1.98 Basics of Graph Theory Problem 4. We are given a degree sequence SG .4. Problem 4. We check if the degree sequence is graphical.4. f.

4. Either prove or find a counter example.4. Is it always possible to convert it into another graph with the same degree sequence but now the highest degree vertex v (with a degree equal to u) is connected to the last u vertices in the degree sequence SG .The Degree Sequence 99 highest degree vertex v (with a degree equal to u) should be connected to the last u vertices in the degree sequence SG . Problem 4.10. Discuss briefly if this is always possible and how will you do it? Problem 4.8.4. . What are necessary & sufficient conditions for a degree sequence SG to be graphical such that in the resulting graph G the highest degree vertex v (with a degree equal to u) in the graph should be connected to the last u vertices in the degree sequence SG . Problem 4. We are given a graph G (and its degree sequence SG ) in which the highest degree vertex v (with a degree equal to u) is connected to the first u vertices in the degree sequence SG . Is there a possibility that a degree sequence SG is graphical but it is impossible to draw a graph corresponding to this sequence such that the highest degree vertex v (with a degree equal to u) in the graph should be connected to the last u vertices in the degree sequence SG .9.

Remember in a trail it is not allowed to traverse an edge more than once – but it is allowed to traverse a vertex several times. .7. The path is shortest in terms of number of edges in between the two terminal vertices. The graph shown in the left diagram is a connected graph while the graph shown in the right diagram is a disconnected graph. We also show a path between the same two vertices in this graph.8: Two graphs are shown in this figure. You may have realized that in a path it is not allowed to traverse an edge or a vertex more than once.6. & Paths We show a walk from vertex a to vertex d in a graph shown in Fig. A walk or a trail can always be converted into a path as shown in this figure. 4. A shortest path between the same two vertices is also shown. Both these graphs have the same degree sequence which is 54444411111. 4. We also show a trail between the same two vertices in the same graph. Trails. Please note that in a walk it is possible to traverse an edge (and therefore a vertex) several times.100 Basics of Graph Theory Figure 4.7 Walks.1.

Walks. A path is shown from vertex a to vertex d in the top right diagram. the path consists of six edges. the trail consists of six edges. no edge is repeated.7. The shortest path from vertex a to d is shown in the bottom right diagram. . In this trail only a vertex is repeated. vertices as well as edges are repeated in the walk. The trail in the top middle diagram is converted into a path from vertex a to d as shown in the bottom middle diagram. Trails. & Paths 101 Figure 4.1: A walk from vertex a to vertex d consisting of eight edges as shown in the top left diagram. A trail from vertex a to d is shown in the top middle diagram. The walk in the top left corner (from vertex a to d) is converted into a four edge path as shown in the bottom left diagram.

How about graphs which are not simple? 4. 4. & (e) Star graphs. Under special graphs we consider: (a) Completely connected graphs.1. (2) Bipartite graphs.e. (b) Regular graphs. i.8. . graphs containing cycles. This graph can be represented by an adjacency matrix also shown in the same Fig. it consists of no self loops or parallel edges. & (3) Cyclic graphs.8 Multi-graphs and Pseudo-graphs We show a simple graph.8. Figure 4. It can be represented by an adjacency matrix. (c) Cycle graphs.1: A simple graph containing no self loops and no parallel edges. (d) Line graphs.9 Broad Categories of Graphs & some Special Graphs We shall talk about very broad categories of graphs and then some special graphs. We categorize graphs into three major categories: (1) Acyclic graphs (or trees). We also show graphs with self loops and parallel edges.102 Basics of Graph Theory 4..

an acyclic graph (or a tree).9.Broad Categories of Graphs & some Special Graphs 103 Figure 4. shown in the left diagrams. is also a bipartite graph.1: A cyclic graph which is neither bipartite nor acyclic is shown in the right diagrams. A bipartite graph which is not acyclic is shown in the middle diagrams. .

4. .104 Basics of Graph Theory 4. 4.9. Please note that partite A is an independent set while partite B is also an independent set but neither of the two is a maximum independent set.9. In other words every edge in a bipartite graph connects a vertex from set A to a vertex in Set B. The vertex set V (G) of a bipartite graph G can be partitioned into two disjoint sets A and B whereas both A as well as B are independent sets.9. 4. 4.9.1 Tree Graphs A connected graph G is a tree provided it does not contain any cycles. A graph is a line graph (or a chain graph) if the degree of every vertex is 2 except for two vertices where the degree is 1. it is bipartite because it does not contain any odd cycles.3. A bipartite graph is shown in the left diagram of Fig. 4.1 is not bipartite as it contains an odd cycle. The graph shown in the right diagram of Fig.3 Special Graphs A graph G is k-regular if the degree of every vertex is exactly equal to k. The minimum vertex cover of this graph is also shown in this diagram. 4.9. it is bipartite because it does not contain any cycle at all.1. If G does not contain a cycle then G is not only bipartite it is also a tree.9.3. It may contain even cycles or no cycles at all. Another bipartite graph is shown in the middle diagram of Fig. A graph G having p vertices is completely connected if the degree of every vertex is p − 1 (please note that the degree of a vertex in a simple graph can not be more than p − 1). 4. The A partite as well as the B partite are shown in the middle diagram where the bipartite graph is drawn with a different orientation to highlight the two parts. A tree graph is shown in Fig.2 Bipartite Graphs A graph G is bipartite provided it does not contain odd cycles. The maximum independent set in this graph is shown in the right diagram of this figure.1.9. 4.2. A line graph consisting of two vertices is a special case where the degree of both the vertices is 1. We show another bipartite graph in Fig.9. A graph is a cycle graph if the degree of every vertex is exactly two.1 and 4. We show a number of non isomorphic trees with p larger than 1 and smaller than 6 in Fig.9.9.

9. This graph is in fact a bipartite graph as shown in the middle diagram consisting of an A partite and a B partite.9. The minimum (sized) vertex cover and the maximum (sized) independent set in graph G are shown in the right diagram. .2: We show a number of non isomorphic trees with p larger than 1 and smaller than 6.3: A graph G is shown in the left diagram. Figure 4.Broad Categories of Graphs & some Special Graphs 105 Figure 4.

9. The degree of every vertex is 2 except for two vertices where the degree is 1 in a line graph.9. A graph G is referred to as a forest if it contains a collection or set of trees. A cycle graph is also a 2-regular graph as shown in this figure. A start graph consisting of two vertices is a special case. being acyclic. A line graph is not regular unless it consists of a special case of a connected graph of two vertices. is always bipartite.106 Basics of Graph Theory A graph G is a star graph if the degree of one vertex is p − 1 while the degree of every other vertex is 1. It is obvious from Fig.4 that a completely connected graph is (p − 1)regular while a k-regular graph may not be completely connected. Figure 4. The degree of every vertex is 1 except for one vertex where the degree is p − 1 in a star graph. A line graph. The degree of every vertex is exactly 2 in a cycle graph. 4. . A star graph is not regular unless it has a size equal to 2 when it becomes a line graph which is 1-regular. The degree of every node is the same for every vertex in a regular graph.4: The degree of every vertex in a completely connected graph is p − 1.

is always bipartite where one partite consists of size 1 while the other of size p − 1. These new graphs are regular and not bipartite.5. We show a number of k-regular graphs in Fig. They are rather 4-partite graphs meaning that the vertex sets of each of these graphs can be partitioned into four disjoint sets of vertices (or partites).9. A 3-regular graph is shown in the middle diagrams. being acyclic. All these regular graphs are bipartite as shown in the bottom diagrams.9. It is interesting to compare these graphs with the ones shown in Fig.4. We show a number of k-regular graphs in Fig. A curious reader might .7.9.5.9.5: A 2-regular graph shown in the right diagrams. 4. 4. 4. The four partites are indicated in different colors in the bottom diagrams of this figure.Broad Categories of Graphs & some Special Graphs 107 Also a star graph. Please note that all these graphs are bipartite as shown in the bottom diagrams of this figure. any edge in these graphs connects a vertex in one partite to any vertex in one of the other partites.9. A number of cycle graphs are shown in Fig. Figure 4. 4. A 4-regular graph is shown in the right diagrams.

Another 4-regular graph is shown in the right diagrams. . A 4-regular graph is shown in the middle diagrams. Figure 4. What may not be obvious is that an odd cycle graph will always be a 3-partite graph as shown in this figure.108 Basics of Graph Theory appreciate the fact that a cycle graph is bipartite if it consists of even number of vertices. All these graphs are 3-partite as shown in the bottom diagrams.9.6: A 6-regular graph shown in the right diagrams. while it is not bipartite if the cycle graph consists of odd number of vertices.

9. Problem 4. we may obtain a graph which is not isomorphic to graph B. The middle graph shows graph B which is derived from graph A after deleting or subtracting a number of edges from graph A.9. Similarly graph C is derived from graph A by deleting another set of edges from B. It is a 7-regular graph consisting of eight vertices.7: Cycle graphs of different sizes are shown. graph B is a 6-regular graph.1.9 above shows the same graph A as shown in Fig 4. Problem Set 4.5.8. Graph E is obtained by deleting a vertex from graph A. Similarly draw all possible non isomorphic graphs of 8 vertices which are 5regular.9.5. Similarly . By deleting a different set of edges from graph A. Draw all possible graphs consisting of 8 vertices which are not isomorphic to B but which are 6-regular. The Figure 4. The figure below shows a completely connected graph A in the left diagram.Broad Categories of Graphs & some Special Graphs 109 Figure 4. Problem 4. Please note that graph C is again a 5-regular graph consisting of 8 vertices.5.2.

Figure 4.8: Graphs B and C are derived from graph A by deleting certain edges.110 Basics of Graph Theory Figure 4.9.9: Graphs E and F are derived from graph A by deleting certain vertices.9. .

5. Try to match the graphs that you have obtained with the ones shown in the figure below.4.3. Problem 4.9. By carefully selecting and deleting certain edges of graph C shown in Fig.Broad Categories of Graphs & some Special Graphs 111 graph F is obtained by deleting another vertex of graph E. Figure 4. (a) Find which of these graphs is a . How many such non-isomorphic graphs we shall be able to obtain? The figure below shows all non-isomorphic graphs consisting of 8 vertices which are 4regular.8. Problem 4. Please note that both E and F are completely connected and regular graphs. by deleting certain edges transform it into a 5-regular graph. (a) Is it possible to get a graph which is isomorphic to graph F by deleting any two vertices of graph A? Discuss briefly. 4. Look at the graphs consisting of 8 vertices and are 4regular as shown in the figure above. (c) By deleting another set of edges transform it into a 4-regular graph. we can obtain a 4-regular graph of eight vertices.5.10: Shows all non-isomorphic graphs consisting of 8 vertices which are 4-regular. (b) Now concentrate on graph E.9. (d) Draw all non isomorphic graphs which are 4-regular consisting of 7 vertices. it is a 6-regular graph.

and Action Items We have talked about various concepts in graph theory in this chapter. Draw all non-isomorphic graphs consisting of 8 vertices which are 5-regular. and are regular consisting of eight vertices. a cycle graph. Connected graphs can also belong to certain categories like a line graph. We have mainly confined ourselves to simple graphs in which there are no parallel edges and no self loops. Figure 4. Un-directed graphs can be further classified into connected graphs and disconnected graphs.10 Integration of Concepts. Problem 4. We have further limited our study to un-directed graphs in this chapter. we shall study directed graphs in detail in Chapter 8. (c) Draw all nonisomorphic graphs consisting of 12 vertices and are 6-regular. The graph G shown in the figure below is derived from graph A by subtracting certain edges from graph A. Note that graph G is 5-regular.112 Basics of Graph Theory bipartite graph? (b) Draw all non-isomorphic bipartite graphs which are connected. (d) Draw all non-isomorphic graphs consisting of eight vertices which are 3-regular. This category of graphs can easily be represented by an adjacency matrix or an adjacency list data structure. Properties.5.9. 4. Out of these graphs indicate which ones are bipartite graphs.11: Graphs G is derived from graph A by deleting certain edges.5. a tree .

Integration of Concepts. or a path between two vertices in a graph. We have also talked about a walk. It will be interesting if we integrate a couple of concepts with a number of properties. a trail. . a cyclic graph and a completely connected graph. We have also talked about certain properties of graphs. and Action Items 113 graph. Properties. like a Hamiltonian path or an Eulerian path in a graph.

We shall conclude with a number of interesting problems. Degree sequence of a graph 5. Some of the necessary conditions of SC graphs will be discussed next. We shall discuss these graphs in some detail as they provide us a platform to connect a number of key concepts in graph theory & algorithms. Complement c(G) of a graph G 2.11.1 Regular Self Complementing graphs We show a number of non trivial (a graph of one vertex is a trivial example of such graphs) self complementing graphs in the following figure.11 Self Complementing Graphs A self complementing (SC) graph is a graph G whose complement c(G) is isomorphic to itself. Deleting and inserting a vertex in a given graph 6. That means that both these graphs are self complementing. The problem of how to construct a SC graph of fixed vertices will also be discussed. 4. Graph isomorphism between two graphs G & H 3. We shall show how one SC graph can be transformed into another SC graph with less or more vertices? Some of the interesting properties of these graphs will also be elaborated. We shall first show some SC graphs in order to give you a feel of such graphs. Both of them are also Hamiltonian? Necessary Conditions . The following concepts are used in a meaningful manner to advance our discussion on self complementing graphs. 1. It can be seen that the left graph is the complement of the right graph as well as isomorphic to the right graph. Regular & non regular graphs 4. These graphs are also regular as the degree of each vertex is 2. Eccentricity of a vertex 7.114 Basics of Graph Theory 4. Radius & Diameter of a graph This section is organized as follows.

Self Complementing Graphs 115 Figure 4. When k = 2. p will become 9.that means the degree of every vertex in a regular self complementing graph should be p−1 . The complement c(G) of graph G is obtained by deleting all edges of graph G from the corresponding completely connected graph.1: A graph and its complement which is isomorphic to the original graph. Can you draw such a graph if p is even? 4 Can you draw such a graph if p is any odd number? It will be essential to answer the above questions before moving forward? If you have tried to draw such graphs then you will soon realize that a nregular graph (where n = p−1 ) is possible if and only if p = 4k + 1 where 4 k is an integer equal to or larger than one.11. Let us try to explore graphs with 9 vertices with the degree of each vertex equal to 4. If c(G) is isomorphic to G then it should have the same number of edges as G . Does that mean that p or p − 1 in an SC graph should be divisible by 4 4? Are these necessary or sufficient conditions for such a regular graph to be self complementing? It will be rewarding if you draw a couple of regular graphs with the property that the degree of each vertex is p−1 . Does this mean that all regular graphs are self complementing? But a cycle graph of three or six vertices is not self complementing? There must be a class of regular graphs which will be self complementing? The following figure shows the same graph along with the corresponding completely connected graph of five vertices. We have already drawn such a graph with p = 5 when k = 1. Will all such graphs be isomorphic to each other? Will such graphs (or at . Then the number of edges in the graph will be exactly 2 p(p−1) .

Are any two of them isomorphic to each other? Is one complement of the other? Are they self complementing? Please try to answer these questions before moving forward. It will be exciting if you draw the complement of the left graph? You will soon realize that the top right graph is the complement of the top left graph while the bottom graph is isomorphic to the top right graph.2: A self complementing graph of 5 vertices (left diagram) and a completely connected graph of 5 vertices. What does this example tell us? .11.116 Basics of Graph Theory Figure 4. least some of them) self complementing? If you take the complement of any such graph it will certainly have the same degree sequence and same number of edges? Again it will be rewarding if you draw some of these graphs before arriving at a conclusion? We show three such graphs in the figure below.

Top right diagram shows the complement c(G) of graph G. .Self Complementing Graphs 117 1 9 2 9 1 2 8 3 8 3 7 6 5 4 7 4 6 5 1 6 5 2 9 7 3 8 4 Figure 4.11. The bottom graph is equal to c(G).3: Top left diagram shows a graph G.

One such self complementing graph is shown below. 1 7 4 9 1 2 5 6 8 3 2 8 3 9 7 6 5 4 Figure 4.11.11.4: A regular self complementing graph of 9 vertices. How about imagining non isomorphic regular graphs having p = 13 and degree of each vertex equal to 6. A very few of them indeed transform into themselves when you take the complement. One such graph is shown in the figure below. . If we take the complement of one such graph it will transform either into one of such graphs or into itself. There are (367860) six-regular non isomorphic graphs with 13 vertices. 1 13 12 2 3 12 10 1 3 5 11 4 8 7 10 6 8 7 5 6 11 2 13 9 9 4 Figure 4.118 Basics of Graph Theory If you actually draw all non isomorphic graphs with p = 9 and degree of each vertex equal to 4 you will realize that there are 16 such graphs possible.5: A regular self complementing graph of 13 vertices. Most of them transform into one another if you take the complement of one such graph as shown above.

2 Non Regular Self Complementing graphs One important question that should agitate you is every self complementing graph a regular graph? Is it possible to have a non regular graph to be self complementing? How about a line graph of four vertices? Its degree sequence is 2211. The degree sequence of the complement of this line graph can be obtained by subtracting 2211 from 3333. The degree sequence of a completely connected graph of four vertices is 3333. Now we have to verify if these degree sequences are really graphical. and if they are then do they really belong to graphs which are self complementary? Let us start with the sequence 32221. 1 2 3 4 1 2 3 4 Figure 4. & 32221.6: A self complementary line graph of 4 vertices. The other two belongs to non regular graphs and are of interest to us. Necessary Conditions Is it possible to have a non regular graph of five vertices which is self complementary? Let us try out various degree sequences with the above necessary condition? There are only three choices possible under the conditions laid out before: 22222. This provides us a necessary condition for a non regular graph to be self complementary? The degree sequence of c(G) should be such that if we add it (after sorting it in increasing order) with the degree sequence of G (having p vertices) then we should get the degree sequence of a completely connected graph of p vertices.11. it comes out to be the same as the four vertex line graph is self complementary. Can you draw its graph and check out if the graph is self complementary? The sequence is graphical but is itself complementing? . 33211.11.Self Complementing Graphs 119 4. The first degree sequence corresponds to a cycle graph which was regular and we have already seen it.

we can safely say that a SC graph of 4k vertices will always be non regular? Why? . In fact we have earlier claimed that a regular SC graph can have only 4k + 1 vertices. we have seen a non regular SC graph of 4 vertices in the last diagram. Thus this condition along with others (for example the number of odd degrees in a degree sequence should be even for a degree sequence to graphical) may narrow down our search for SC graphs but we always have to verify if a given graph G is a SC graph? These simple conditions tells us that an SC graph (which may or may not be regular) can only have p equal to 4k or 4k + 1 vertices where k can be equal to or larger than 1.7: A 5 vertex non-regular self complementing graph. this indeed comes out to be graphical as well as self complementary as shown below: 1 2 3 4 5 1 2 3 4 5 Figure 4.11. it is not sufficient. Please note that the condition (the degree sequence of c(G) should be such that if we add it (after sorting it in increasing order) with the degree sequence of G (having p vertices) then we should get the degree sequence of a completely connected graph of p vertices) is a necessary condition. we have also seen a non regular SC graph of 5 vertices. We have already seen a regular SC graph of 4 + 1 = 5 vertices.120 Basics of Graph Theory Now let us check the other sequence which is 33221.

Some of the possible degree sequences are listed here: 66661111.3 Constructing Self Complementary Graphs Let us consider graphs where p = 8 and assume that its degree sequence satisfies the stated necessary conditions for a SC graph. 66443311.8: Here vertex G is a graph while c(G) is complement of graph G. and G may contain more than one vertex. 44443333. If G contains a single vertex then this graph is certainly SC graph. 55443322.11.we shall now discuss ways of constructing large SC graphs bypassing this tiring process? The new method will help us further in making meaningful connection between relevant concepts.Self Complementing Graphs 121 4. The resulting super graph is a self complementing graph. 55552222. An edge between G and c(G) means that every vertex G is connected to every vertex of c(G). If we connect graph G and its complement in the following configuration then we claim that the new graph will be a SC graph.11. Let us actually construct such graph and see how it looks? Assume that G . But if G contains more than 1 vertex then why this composite graph is self complementary? Figure 4. Please note that a line between G and c(G) means that each vertex of G is connected to every vertex of c(G). It will be interesting if you try to check if a graph corresponding to these degree sequences is indeed self complementary? This will certainly be a tiring process . Given any graph G we can always find its complement c(G).

Another way to construct a SC graph is shown below . The number of vertices in the SC graph will still be 4p where G has p vertices. Then c(G) will be its complement . How the above configuration would look like when it is actually drawn? If you look at the diagram below you will realize that the graph is indeed self complementary. Another copy of G will consist of a line graph of two vertices .11. 7 5 8 6 1 6 5 2 1 3 2 4 7 3 7 4 Figure 4.122 Basics of Graph Theory is a line graph of two vertices . The degree sequence of this graph is 44443333 it is one of the sequences that we have predicted earlier in our discussion? We can also start with G equal to two isolated vertices .then c(G) will simply consist of two isolated vertices. . Why? The number of vertices in the resulting SC graph will be 4p + 1 where p is the number of vertices in G.consisting of vertex 7 and vertex 8.and that will be two isolated vertices 3 & 4. Using these building blocks you can also construct another SC graph? If you have the patience of constructing such a graph you will realize that its degree sequence will be 55552222.then c(G) will be a line graph of two vertices.here also G is any graph and is connected with its neighboring graph in the same fashion? Here x is just one vertex and we claim as before that the resulting graph will be self complementary.9: Here graph G is a line graph of two vertices 1 & 2.

What would happen if the self complementary graph H in the above step is regular? How about if it is non regular? 5.Self Complementing Graphs 123 G c(G) G ) c(G c(G) G x c(G) x G Figure 4. Please find the degree sequence of each of the following graph? Can you figure out how these SC graphs are constructed? . 1. Will the above configuration result into a regular or a non regular SC graph? The following diagrams will help you answer these interesting questions. Is it possible that we insert graph G in the place of x in the above configuration? Please note that G is any graph? 3. An edge between G and c(G) means that every vertex G is connected to every vertex of c(G). How about if we insert a self complementary graph H in the place of x as well as G in the above configuration? 4.10: Here vertex G is a graph while vertex c(G) is the complement of graph G.11. There a number of interesting possibilities in the above configuration. Vertex x may be a single vertex graph. How about if instead of single x vertex we have a graph H connected to its neighbors? In order to make the whole configuration SC should it be any graph H or a special graph H? What special property it should possess? As you can discover yourself H should be a self complementary graph for the above configuration to be self complementary? 2. The resulting super graph is a self complementing graph.

4 Transforming a SC graph with 4k + 1 vertices into another SC graph with 4k vertices We know that a SC graph should have either 4k or 4k + 1 vertices? This immediately leads us to conjecture that if we have a SC graph of 4k + 1 vertices then by deleting one vertex we can convert it into another SC graph having 4k vertices? Is it straight forward or we need to devise an intelligent algorithm to do so? Similarly if we are given a SC graph having 4k vertices then is it possible to construct a SC graph having 4k + 1 vertices by inserting a new vertex and connecting it to some of the vertices of the original graph? Again do we need some thinking to do so or is it a trivial problem? In order to answer these questions let us start with a regular graph which is SC.11: A line graph of 4 vertices is used as a building block in the left diagram.11. 4.124 Basics of Graph Theory Figure 4. By deleting any vertex from this graph is it possible to convert it into another SC graph having 4k vertices? Let us start with a simpler problem of a regular SC graph having 9 vertices as shown below. It will contain 4k + 1 vertices. Can we delete any vertex and the resulting graph would stay a SC graph of 8 vertices? Please note that initially the degree sequence is .11. A cycle graph of 5 vertices is used as a building block in the right diagram.

It becomes 44443333 after deleting that vertex. 125 What about if we delete any vertex from the 25 vertex graph shown earlier. Will the new graph be also self complementary? Why? How about if we have a non regular SC graph of 4k + 1 vertices? Can we delete any vertex to make it another SC graph with 4k vertices? How about this graph? It has a degree sequence 332211.Self Complementing Graphs 444444444. It is no longer a trivial problem and we should devise an algorithm to solve the problem? .

We remove vertex 2 from the same SC graph in the bottom diagram. .126 Basics of Graph Theory Figure 4.12: We remove vertex 1 in the top diagram from a SC graph in the top diagram.11.

13: We remove a vertex from the 25 vertex SC graph.11.Self Complementing Graphs 127 Figure 4. Is the resulting graph a SC graph? .

after deleting one vertex the new degree sequence should possess certain properties if it represents a self complementing graph? For example it can be 66443311 or 77443300.14: A SC graph having 5 vertices which is not regular.11. Can we remove any vertex and still it remains a SC graph? Given a degree sequence 774444411 of a self complementing graph .15: A SC graph having 9 vertices. Here we show another interesting graph which is not regular but is self complementing. The problem is to decide which vertex should be removed? . etc.128 Basics of Graph Theory Figure 4. Figure 4. This graph may also provide insights needed to design an intelligent algorithm to solve the above problem. Its degree sequence is 774444411.11.

although it has not been drawn in a suitable manner in the top diagram. The resulting graph is SC and is also regular .11. Now start with the same original graph (with a degree sequence equal to 44443333) while the new graph should have a degree sequence equal to 444444444. Again we shall show you a number of graphs and then provoke you to design an efficient algorithm to solve the problem.Self Complementing Graphs 129 Should we always remove the vertex with degree equal to 4? Why? Should that vertex be connected with lower degrees or higher degrees? 4. The graph is redrawn in the bottom diagram to emphasize that it is indeed a beautiful regular graph? .16: We insert a vertex x in a self complementing graph. We show the same graph with a degree sequence 44443333 being converted into a new SC graph with degree sequence 555543333.5 Transforming a SC graph with 4k vertices into another SC graph with 4k + 1 vertices Now let us look into the problem of inserting a vertex in a SC graph consisting of 4k vertices such that the new graph having 4k + 1 vertices is also selfcomplementing. Please note that we can also put an extra constraint that the resulting SC graph should be regular. Figure 4.11.

.17: After inserting vertex x we have a regular SC graph.11.130 Basics of Graph Theory Figure 4.

Self Complementing Graphs 131 So the problem is where to connect the inserted vertex and what should be its degree? Perhaps another example may provide you a solution? Please check the degree sequence of the graph before after inserting a new vertex? Figure 4.then can we solve the graph isomorphism problem? How? .18: Inserting a vertex x in a SC graph? 4.6 The Self Complementary problem and Graph Isomorphism If we know how to find if G and H are isomorphic then we can always check if G is a self complementary graph? How? Take the complement of G . But suppose we know how to check if G is a self complementary graph .11.this is c(G) and now check if G and c(G) are isomorphic.11.

Can we prove that for any graph G the following configuration will give us a SC graph with a diameter not more than 3.19: We need to check if graph G is isomorphic to graph H? Look at the figure above. We need to check if a given graph G is isomorphic to another given graph H.not less than 2 and not more than 3? Instead of proving the above statement let us first do a simple one.11.11.20: We need to check if graph G is isomorphic to graph H? 4. .7 A SC graph has diameter 2 or 3 .132 Basics of Graph Theory Figure 4. We claim that if the top (or the bottom) graph is self complementary then graph G is isomorphic to graph H? Why? We can also use the following configuration to check if G and H are isomorphic also by substituting H in an appropriate place in this diagram? Figure 4.11.

and the radius and diameter of a graph.11. The diameter and radius for the bottom line graph will be respectively 4 and 2. Similarly we show another five vertex graph where the .Self Complementing Graphs 133 Figure 4.11. Let us see how the eccentricities of various vertices look like in a line graph of eight and five vertices as shown below.21: What is the diameter of this super graph? Again can we prove that for any graph G the following configuration will give a SC graph with diameter not more than 2? Figure 4. The diameter in the top graph will be 7 while the radius will be 4 in the top diagram.22: What is the diameter of this super graph? We shall now attack the more general problem? But before that let us refresh our knowledge regarding some old concepts: the eccentricity of a vertex.

The eccentricity of each vertex is indicated in the line graph as well as in its complement.11.11.134 Basics of Graph Theory Figure 4. eccentricity of each vertex is 1. The complement of this graph is also shown in the bottom diagram. The five vertex line graph and its complement are shown in the bottom diagram. Another five vertex graph G and its complement is shown below. Its diameter as well radius will be infinite. It is a completely connected graph . Please note that the diameter in the line graph was 4 while it has reduced to 2 in the complement of the line graph. You can well imagine that if the diameter in a graph G is more than 3 then . Thus its diameter is 2 while its radius is 1. The diameter as well as the radius in the complement graph is 2.23: The eccentricities of different vertices in a line graph. Here the eccentricity of each vertex in G is 2 except for vertex 1. Figure 4. The diameter in the complement of graph G is infinite while its radius is also infinite.the diameter as well as the radius is 1 in this graph.24: A completely connected graph G and its complement.

11.26: A line graph and its complement. .11. Figure 4.25: A star graph and its complement.Self Complementing Graphs 135 Figure 4.

Now assume that the bipartite complement of bipartite graph B is isomorphic to bipartite graph B.the resulting graph will be a bipartite complement of bipartite graph BP .not more not less? 4. Please note that it is not a Category B SC graph? Why? Can you draw another bipartite graph which is a Category A SC graph? Can you prove that such graphs are not possible if any partite contains more than two vertices? We show a category B SC graph BP in the figure below. we first form a completely connected general graph consisting of all BP vertices .11.we then remove those edges which are already present in BP . Similarly if the radius or the diameter in a graph G is 1 then it cannot be isomorphic to its complement? Why? Does that mean that a SC graph can have diameter 2 or 3 .136 Basics of Graph Theory the diameter in its complement is reduced to 2. Here we first form a completely connected bipartite graph consisting of as many vertices in the A as well as B partites of bipartite graph BP .this will give us the complement of graph BP . Please note that bipartite complement of a bipartite graph is different from this complement. Alternatively in order to find complement of graph BP . That means a graph G with a diameter more than 3 cannot be self complementary. Let us call these category A SC bipartite graphs.thus all the A (and B) partite vertices in BP will become completely connected in the complement of BP . Let us call these category B SC bipartite graphs.not more than 3 and not less than 2? How about the radius of a SC graph? Should it be always 2 . We assume that while taking the complement of graph BP we simply consider it a general graph . Please note that this not a Category A SC bipartite graph? Why? . We then remove existing edges in BP from this completely connected bipartite graph . The following graph is a Category A SC graph.8 Bipartite self complementary graphs Definition: Assume that the complement of a bipartite graph BP is isomorphic to bipartite graph BP .

The corresponding completely connected graph is shown in the top right diagram.Self Complementing Graphs 137 Figure 4.27: A bipartite graph BP shown in the top left diagram. The complement of graph BP is shown in the bottom diagram.11.28: A bipartite graph and its bipartite complement.11. . Figure 4.

Figure 4.138 Basics of Graph Theory The same graph BP is drawn below in a different shape.11. The bottom diagrams once again shows the two bipartite graphs shown in the standard form. As you can see this is a regular SC bipartite graph. While that of the completely connected bipartite graph is 44444444. The degree sequence of the above graph is 22222222. The corresponding completely connected bipartite graph and the complement of bipartite graph BP is also shown here. Top middle diagram shows the corresponding completely connected bipartite graph. The degree sequence of a corresponding completely connected bipartite graph is also 44444444.29: Top left diagram shows a bipartite graph BP. Here we show another Category B SC graph with red vertices in one partite . The following bipartite graph is a non regular self complementing graph with a degree sequence 33222211. Top right diagram shows the bipartite complement of the bipartite graph.

11. and green vertices in another partite. The number of edges in a SC bipartite graph will be mn/2. Thus either m or n should be even. .11. That is an important necessary condition for a bipartite graph to be self complementary.we shall simply refer them as SC bipartite graphs.31: A regular SC graph having 12 vertices. How many edges are there in a complete SC bipartite graph with m vertices in one partite and n vertices in another partite? That is equal to mn. Figure 4. From now on wards we shall consider only category B self complementary bipartite graphs .30: A SC bipartite graph which is not regular.Self Complementing Graphs 139 Figure 4.

we call this the graph G(A). bipartite graph BP .140 Basics of Graph Theory Figure 4. Thus even if we remove vertex x the resulting graph G − x will still be a SC graph as shown below . We put vertices in G − x which are adjacent to x in set A while the rest of the vertices in G − x goes in the set B. In addition to that there will be a graph inside G consisting of vertices belonging to the set A . The original graph can thus be decomposed into edge-disjoint graphs .11.32: A SC graph where the size of two partites is different. graph G(B) and the vertex x connected to all . Please note that the following graph is a SC graph with or without the vertex x.although vertex x provides a useful function of pinpointing certain vertices needed for the decomposition of the graph? Vertices in G − x which are adjacent to x belong to one partite while the rest of the vertices in G − x belong to the other partite. 4.9 Decomposition of a SC graph G Here we shall talk about the decomposition of a SC graph G into a number of edge-disjoint graphs .one of them is a SC bipartite graph? Initially we shall talk informally and provide some insight and then we shall discuss it more formally.graph G(A).11. There will be a bipartite graph having edges going between the set A and the set B. Similarly there is a graph G(B) consisting of vertices in the set B.

1. The bipartite graph BP (A.11.Self Complementing Graphs vertices in graph A. It can be observed that the following properties should hold for the original graph G to be SC.there is some relationship between these two graphs? . The graph (x. 141 Figure 4.33: A self complementing graph is decomposed into a number of edge-disjoint graphs. The two partites are balanced that is the size of A and B is the same. A) 3. 2. B) will also be a SC bipartite graph. The graph G(A) and graph G(B) .

The vertex 9 is mapped onto itself. Figure 4. This graph H will always be isomorphic to graph G but may not be equal to graph G (as the adjacency matrix may be different). For example consider the permutation p1 equal to (1234)(5678)(9) as shown in Fig. Isomorphism. 4. 2 onto 3. 3 onto 4.10 Permutation. This permutation permutes the vertices of G and produces a new graph which is not equal but isomorphic to the original graph G. 4. automorphism & Self Complementing Graphs Any permutation of vertices of any graph G may create a different graph H.11.11.142 Basics of Graph Theory Let us consider another example to confirm/enhance our observations before formally discussing the decomposition of a SC graph. and 4 back to 1.11. This is a so called circular permutation in which we map 1 onto 2.34: A SC graph decomposed into a number of edge-disjoint graphs. In order to show that this new graph is isomorphic to the original graph G we have to find an isomorphic function .35.

In this case this permutation p2 will be (4321)(8765)(9). Figure 4. We may be more interested in the non trivial permutations of graph G? It may be possible for a certain category of graphs that the only automorphism is the trivial identity permutation? Please note that if p is an automorphism of graph G then the permutation p2 is always an automorphism of graph G If a permutation p of vertices of graph G creates a graph H which is the complement of graph G then the permutation p is known as the complementing . The two graphs are not equal but they are isomorphic thus the permutation p1 is not an automorphism of graph G.11.Self Complementing Graphs 143 (or a permutation) which maps the vertices of the new graph back onto the vertices of graph G such that adjacency as well as non adjacency is preserved in the two graphs. Please note that graph G is equal to the bottom graph If a permutation p of a graph G creates a graph H which is equal to graph G (that means the adjacency matrix will exactly be the same) then that permutation is known as an automorphism of graph G.35: Graph G (top left diagram) and another graph (top right diagram) where the vertices of G are permuted according to the given permutation p1 .it is known as a trivial permutation. The identity permutation is always an automorphism .

144 Basics of Graph Theory Figure 4.36: Graph G (left diagram) and another graph (right diagram) where the vertices of G are permuted according to the given permutation p.11. Any complementing permutation of a graph G is certainly not an automorphism of graph G but the permutation p2 is always an automorphism of graph G if p is a complementing permutation. permutation of graph G and graph G is a SC graph. The two graphs are not only isomorphic but also equal thus the permutation p is an automorphism of graph G but it is not a complementing permutation of graph G. Whenever we claim that G is self complementing then we have to find out the self complementing permutation? . Thus every self complementing graph G has a complementing permutation p associated with that graph G.

This new graph is equal to graph G .thus p2 (G) is an automorphism of graph G.11.Self Complementing Graphs 145 Figure 4. . This new graph is also the complement of graph G . The bottom diagram shows another graph where the vertices of p(G) are permuted once again according to the same permutation p.that means the permutation p is a complementing permutation of graph G but it is not an automorphism of graph G.37: Graph G (top left diagram) and another graph (top right diagram) where the vertices of G are permuted according to the permutation p.

two of them are shown in the figures below. 4. But in the top permutation one of the partites is mapped onto the other partite . The permutation in Fig.38: Bipartite graph BP and a complementing permutation. The permutation in Fig. According to this permutation vertices of one partite are permuted onto vertices of the other partite in the complement of BP .146 Basics of Graph Theory Let us now consider certain self complementary bipartite graphs and the related complementary permutations .in the lower permutation it is mapped onto the same partite? .39 is also a self complementing permutation.38 is a self complementing permutation. Figure 4. 4.11.11.11.

In the decomposition or the synthesis effort to be explained in the coming examples we are more interested in those bipartite graphs which are similar to bipartite graphs similar to the one shown in Fig.39: Bipartite graph BP and its complement. Why? In the coming figures you find certain clues of synthesizing self complementing graphs using simple building blocks? .11. According to this complementing permutation vertices of one partite are permuted onto vertices of the same partite.Self Complementing Graphs 147 Figure 4. 38. The complementing permutation p = (1A3A)(2A)(4A)(1B3B)(2B)(4B).

148 Basics of Graph Theory Figure 4.40: A successful attempt to synthesize a self complementing graph.11. . p(G) and p2 (G) in that order. The bottom diagram shows graph G.

.41: A successful attempt to synthesize a self complementing graph. p(G) and p2 (G) in that order. The bottom diagram shows graph G.11.Self Complementing Graphs 149 Figure 4.

42: A successful attempt to synthesize a self complementing graph.150 Basics of Graph Theory Figure 4. p(G) and p2 (G) in that order.11. . The bottom diagram shows graph G.

. The bottom diagram shows graph G.43: A failed attempt to synthesize a self complementing graph from individual components.Self Complementing Graphs 151 Figure 4. p(G) and p2 (G) in that order.11.

152 Basics of Graph Theory .

Chapter 5 Basics of Graph Algorithms 5.5 5.4 5.6 5.7 5.9 Design of Algorithms The Bucket Algorithm Finding if a Graph is Connected Finding if a Graph is a Tree Finding a Spanning Tree of a Graph Finding a Minimum Spanning Tree of a Weighted Graph Finding a Path in a Graph The Shortest Path Problem Graph Traversal Techniques 5.3 5.1 5.8 5.10 Some Graph Theoretic Claims 5.2 5.12 Discussion .11 Shortest Path Algorithms 5.

We have also studied a number of properties linking different concepts. . Again this building block can be used to design a number of sophisticated shortest path algorithms based on dynamic programming. We also provide a number of powerful learning tools to understand and design various algorithms. For example. and its title was Should We Teach Algorithms. The initial part of this chapter is based on one of our papers published in IJECE.154 Basics of Graph Algorithms Introduction We have studied a number of concepts related to graph theory in the last chapter. Another four line procedure known as 2-edge Shortest Path Algorithm which finds shortest paths of length 2 from a given vertex to every other vertex in a weighted directed graph. Summer-Fall 2003. how can we check if an edge is a bridge edge.1). how can we check if a graph is cyclic. and Yasser Hashmi and its title was Shortest Path Algorithms . and ofcourse computer science. how can we find an actual cycle in a cyclic graph (see Concept Map 5. In this chapter. Vol. 2. The recursiuon tree and the colored puzzle are some of these visual aids which facilitate a learner or a designer in his or her path towards discovery. engineering. mathematics. biology. For example. 2. The paper was co-authored by Sara Tahir. social sciences. We have also studied that a graph is cyclic if it contains a cycle. we talked about a bridge edge and a cycle in a graph. For example an edge is a bridge edge provided its removal disconnects a graph. A four line procedure known as the Bucket Algorithm which can be molded into a number of useful and powerful graph algorithms based on greedy strategies. The later part of this chapter is based on one of our CS department research reports which was coauthored by Komal Syed.2. No. The original research papers were about encouraging students to discover and learn (graph algorithms) by themselves with minimal help provided by an instructor in the form of provocative questions.Making and Breaking Connections. Almost all of these algorithms are based on the following easy to understand and friendly to use buliding blocks: 1. we provide a detailed study of a number of graph algorithms that have applications in diverse fields like chemistry. In this chapter we shall study a number of graph algorithms.

Properties.1. Action items and Graph Algorithms.155 Concept Map 5. A panoramic picture of some Concepts. .

is essentially a creative effort containing all the ingredients of a thriller: adventure. Though there are rules of thumb that can be followed to help an individual design an algorithm. on how we deal with the analysis phase. dealing with analysis. can sow the seeds that could blossom into the genius that produces efficient yet astonishingly simple algorithms. is to equip the students with the necessary tools and techniques. it appears to be correct. challenge. following our approach and providing proper guidance. The second one. we believe that the instructor. timely hints. Despite the fact that one cannot guarantee that a student could become an efficient algorithm designer. The first objective. and above all the confidence required in solving a non-textbook problem.1 Design of Algorithms Teaching the standard course “Analysis & Design of Algorithms” at an undergraduate level in a typical Computer Science program essentially has two objectives. and stimulating questions posed by the instructor. to a large extent. there is no precise algorithm available that can be used to design new algorithms. excitement. concerned with the design of algorithms. one can always encourage students to redesign an algorithm right from scratch. students to create algorithms themselves using some very fundamental concepts. This second objective. Our experience of teaching algorithms indicates that creativity in algorithm design depends. rather incite. The objective is that students should experience the thrill and excitement of discovery even during the initial phases of understanding existing algorithms. It is now . The study of the methods and rules of discovery and invention is a field in its own right. There is no guarantee that one who critiques literature can learn to write beautiful poetry. we should not formally teach anything. and suspense. Instead we should encourage. We stress that while we are familiarizing students with existing algorithms. is to familiarize students with existing algorithms. which is perhaps far more important. Similarly the ability to understand and analyze algorithms does not guarantee that one could become an efficient algorithm designer. Polya [10] remembers the time when he was a student himself: he was always perturbed by the question: ”Yes the solution seems to work. but how is it possible to invent such a solution? How could I invent or discover such things by myself ?” We feel that with the availability of some pre-requisite knowledge.156 Basics of Graph Algorithms 5.

[6]. In the words of Miller [9].The Bucket Algorithm important to find a good working definition of design (of algorithms). He also does not allow the use of any source material.1 What is Design? According to the Webster’s dictionary [1]. and other courses.2 The Bucket Algorithm We start with a simple algorithm which we call the Bucket Algorithm (the bucket symbolizes a friendly container where a child puts every new toy or every new discovery) consisting of just four lines of pseudo code: We shall show how this primitive procedure can be used to reinvent a number . allowing his students to find the answers in their own ways”.2 The Moore Method R. L. ”Design is to conceive and plan out in the mind”. and have advanced or modified the Moore Method in a number of ways [3]. but in analysis.1. mostly quiet. solution. and then actively guides the students in their path of discovery. leading to an effective or efficient outcome. does not allow collective effort on the part of the students inside or outside of class. Moore was a professor of mathematics at the University of Texas. directed set of decisions that are introduced. We. occasionally asking a question. ”Design is the thought process comprising the creation of an entity”. made and deployed. 157 5. It is interesting to note that our approach is similar in some aspects with the so-called Moore Method of teaching and learning. ”What was so special about his mode of teaching was that he did not lecture. 5. or technology”. game theory. starts with something (very simple). In the words of Hale [5]. The teacher. encourage lively discussions inside as well as outside the classroom. algebra. in our model.1. Taylor [18]. he did not profess. The last definition suits our discovery based learning approach in which a teacher formulates a directed set of questions and hints in order to help his/her students design algorithms. on the other hand. Many professors still use his teaching style not only in his subject of specialization (topology). Rine [13] defines design as ”A systematic. while characterizing (his version of) the Moore method of teaching. He sat in the back of the room. 5.

1: Two pictures of what the Bucket B will look like in the initial stages of the Bucket Algorithm. j g h k f d e i g j i h k f d e a b c a b c Figure 5.158 Basics of Graph Algorithms Algorithm 17: The Bucket-Algorithm input : A Graph G output: A Bucket B 1 Put any vertex x of Graph G in the Bucket B. .2. 4 Put v in B . 2 while there are edges coming out of the Bucket B do 3 Select an edge connecting vertex u in B to v not in B.

6. Conduct a breadth first search in a graph.1 Understanding the Bucket Algorithm The Bucket Algorithm is simple and straightforward. Find a path between two vertices in a graph provided a path exists. Rediscover Prim’s Algorithm.2. Find the number of connected components of a graph. Find if a graph is a tree. Solve the single-source shortest-paths problem: rediscover Dijkstra’s Algorithm. Find if a given graph is connected. Find a minimum-spanning tree of a graph. . We start with something simple but potentially very powerful. and the procedures followed in order to arrive at innovative solutions.The Bucket Algorithm 159 of existing powerful algorithms in graph theory ([16]. Conduct a depth first search in a graph. 4. They would learn the ways and means of devising their own algorithms. [4]). [15]. the students should develop a keen desire and ability to understand the motives behind. 8. Find a bridge in a graph. Find a spanning tree of a graph. [14]. [17]. Simple. 7. 9. 11. [2]. 2. 10. [11]. Specifically the Bucket Algorithm would be used to solve the following problems: 1. because it is easy to understand and at the same time flexible enough to handle a variety of different problems. 5. Rediscover Kruskal’s Algorithm. With some encouragement from the instructor. 5. 3. It is just a 4-line algorithm with a simple while loop (with no conditional statements or recursive calls). 12.

k}. d}} to choose from in Step 3 as we iterate through the while loop.2 How does it Work? We identify a Graph G and a Bucket B (See Figure 5. Next we choose any edge joining vertex a to any other node. The third is the set of edges (the “branches coming out of the Bucket B” in Step 2) connecting vertices inside the bucket to vertices outside the bucket (see the middle diagram of Figure 5. {i. Figure 5. b}. See Figure 5. Step 1 instructs us to put any node. {g.1). The second is the set of edges connecting vertices outside the bucket with each other: {{f. This set of edges is equal to {{a. The first is the set of edges connecting vertices inside the bucket with each other: {{a.2. {b. c}. say node b. {b. Notice that there are two types of vertices: those inside the bucket represented by the set {a.2. f }. An edge belonging to this last set of edges is called a cross edge and is of most interest to us. e}.1 for a picture of what the Bucket B will look like at this stage. i. we can implement numerous algorithms. Now we have a set of nodes {a. {b. {a. {i.2. say node a. {a. These two different kinds of vertices give rise to three different kinds of edges.2.160 j g h k f d e a b e c f d i g h k a b c j i Basics of Graph Algorithms j g h k f d e a b c i Figure 5. .2). f }}. k}. i}. {d. j. c}. of the Graph G into the bucket. {d.2: Two pictures of the Bucket B after different iterations through the Bucket-Algorithm.2(middle diagram) shows the bucket B after different iterations through the Bucket Algorithm. d. {h. and those outside the bucket. 5. f }. k}}. g}. b. k}. Depending on the constraint we place on the selection of cross edges in Step 3. {b. d}. c. h}. g. {c. the set {f. k}.2.2. j}. b} in the bucket giving rise to a set of edges {{a. e}}. h. in the graph (since all other nodes are currently outside of the bucket) and put b in the bucket.

After the students are confident that they understand the idea behind the Bucket Algorithm. “Half way home to solving a problem is a clear understanding of the problem”.5 The Right Provocation It is well known that a real understanding of the problem is a necessary condition to solve any problem. It may be advisable to include it at a later stage. During this activity the instructor should ask thought provoking questions such that the students focus on multiple facets of the algorithm that would later help in designing new algorithms.The Bucket Algorithm 161 5.2. the first question is ”Do I really understand the problem?” Then comes the role of the teacher in terms of how he/she states a problem and provokes (or guides) his/her students to solve it in a specified manner. 5. Out of a sequence of six questions posed by Skiena [16] in order to guide one to discover the right algorithm. the instructor can start asking them to modify it to solve more complex problems. and when would it be false? Does it make any difference if we have a different starting vertex? Note that there are situations when it really makes a difference. What is the worst-case complexity of this algorithm? It is recommended that the instructor not involve the underlying data structure at this stage in order to tackle the issue of complexity.3 Playing with the Algorithm We strongly encourage our readers to play around with the Bucket Algorithm to get comfortable with it. It will be useful at this stage if the students are asked to derive the time complexity of the Bucket Algorithm.2. Would all the vertices of the graph move into the bucket after the completion of the algorithm? When would this scenario be true.4 Solving Other Problems The above questions would induce a deeper understanding amongst students about how the Bucket Algorithm works under different conditions and give some hints while solving more complex problems. For . v} we select in Step 3 that we discover the new vertex v. Such questions could be: Under what conditions would there be no edges coming out of the Bucket? Note that this condition should be met otherwise the algorithm would never terminate. Students must come to realize the importance of cross edges: it is because of this cross edge {u.2. According to David [12]. 5.

5.3. Why we should do this and how should we do this are both equally important for designing. after the Bucket Algorithm has been applied to a graph G.3 Finding if a Graph is Connected Assuming that the students know what a connected graph is.2.1). Please see Concept Map 5. Brighter students would have been able to identify at that stage that some nodes will be left outside the Bucket B when a graph is not connected since cross edges do not exist connecting them to nodes inside the bucket (Figure 5. Please see Concept Map 5. . Notice while students were becoming familiar with the Bucket Algorithm. he/she cannot expect his/her students to discover the said algorithm just after understanding the sorting problem.162 Basics of Graph Algorithms example if a teacher is talking about Quick Sort. The understanding of the previous state of an abstract system and the (usefulness of the) final system state after the application of a so called fundamental operation [7] (for example the partitioning procedure in Quick Sort) is crucial in problem solving in computer science as in elsewhere.2. discovering (and even understanding) the said sorting algorithm. Not all students may be able to identify this property of the Bucket Algorithm. there are still any nodes left outside the bucket then the graph is not connected. The instructor in this case will have to make an extra effort to guide such students. i. Once all students have understood the solution (having arrived at it on their own with well-timed prodding from the instructor) the instructor should start the discussion regarding cost calculation. the instructor should ask the students: “Can you modify the Bucket Algorithm such that you may be able to determine whether a given graph G is connected?” The emphasis should be on using the existing techniques with minimum modification. the instructor asked when there would be nodes left outside the bucket.e. the complexity of the modified algorithm. The answer is simple: if. however. If.. all nodes come inside the bucket. then graph G is connected. The teacher should first make the students appreciate the need for partitioning the array into halves such that all numbers in the first half are smaller than each number in the second half.

3. it is the job of the instructor to at least identify them for those students who cannot visualize the solution immediately.1: A graph G that is not connected. Applying the Bucket Algorithm once on a graph with more than one connected component would tell us that the graph is not connected as all the vertices of the graph do not end up in the bucket.3. The number of times we have to apply the Bucket-Algorithm depends upon the number of connected components. 5. nodes i and j will be left outside the Bucket B. The vertices that do end up in the bucket belong to a single connected component. This could be solved if we remove the given edge and then check the number of connected components in the resulting graph.Finding if a Graph is Connected j g h i k h 163 g j i k f d e a b c f d e a b c Figure 5. Applying the algorithm again with a new bucket would give us a new connected component. 5. once the Bucket Algorithm terminates. What would be the resulting time complexity of this algorithm? The . and this would determine the worst-case time complexity.1 The Number of Connected Components Once we understand how to find if an un-directed graph is connected the above problem becomes simple and very little imagination is needed to answer the above question.3. The first problem is to check if a given edge is a bridge. and so on and so forth.2 Finding a Bridge in a Graph A cut edge or bridge is one whose removal produces a graph with more connected components than the original. There are essentially two different problems here.

1 Every Edge in a Tree is a Bridge We know that a tree having n vertices consists of bare minimum number of edges. Thus every edge in a tree is a bridge.3. Solving a problem from different angles and then making a comparison is the single most important exercise for a student studying algorithms (Rawlins [11]).4. This in not only true for this problem but is true for a majority of problems.164 Basics of Graph Algorithms second problem is to find or locate a bridge in a given graph. which makes it a connected graph.4 Finding if a Graph is a Tree The algorithms that solve this problem depend on how we define a tree. 5. This implies that removing any edge would disconnect a tree.2: A cut edge or bridge is one whose removal produces a graph with more connected components than the original graph. Once the first problem is solved it should be a simple matter to handle it. We already know how to check if a given edge is a bridge in a graph. How many times the Bucket-Algorithm is applied and what is the resulting worst-case complexity of the algorithm? g h j g i k h j i k c b e f d e a f d a b c Figure 5. It highlights the fact that looking at various definitions or properties (which come from a study of graph theory) is sometimes extremely useful and it provides the seed for designing a number of very powerful algorithms. 5. The problem is thus reduced to repeatedly applying the algorithm designed to .

How complex is this problem? Is it possible to count the number of edges while we are checking if the given graph is connected? Would that perhaps reduce the complexity? 5. 5. So the problem is reduced to counting the number of edges.4..e. a connected graph is a tree provided the number of edges in the graph is exactly equal to one minus the number of vertices in the graph. all of these definitions are equivalent implies. p − 1. This definition or property can be used to design an algorithm to check if a given graph is a tree.4. i. Thus the spanning tree of a tree would be exactly the same tree.3 The Spanning Tree of a Tree We know that a tree has the minimum number of edges required to connect a given number of vertices. A spanning tree of a given graph also satisfies this property. The number of times we would have to do this and finding the resulting complexity is an interesting exercise by itself.4 A Comparison A comparison of all these algorithms would be extremely beneficial to the students if they are encouraged to work it out independently. In fact.Finding if a Graph is a Tree 165 check if an edge is a bridge. We know how to find if a given graph is connected using the Bucket Algorithm. Once they have the answers it would again be stimulating for them to compare their findings with their colleagues within the classroom. as it is a tree. 5. For example. The catch is that the graph should be connected otherwise the definition would not apply (why?).2 The Number of Edges in a Graph We can define a tree in a number of ways. Only when the students have gained confidence that they understand the basic problem and can find an efficient solution should we move to more complex problems such as finding whether a given graph is a forest. Encouraging and initiating interesting discussions and even heated debates is one of the most important responsibilities of a teacher: (s)he must simply coordinate and make sure that the interaction is moving in the right direction. .4.

5. Or we can start with no edges and start growing edges until we get a tree.1 Cutting Edges If we remove all redundant edges from a given graph and just keep edges essential to keep it connected the remaining graph would be a spanning tree of the given graph. we might have noticed that every time we discover a new vertex it is because of a cross edge (step 3).5. The resulting complexity would change dramatically depending upon the approach used. It is also possible to identify some of the so-called cross edges (edges which are coming out of the Bucket). and that the number of such cross edges would be exactly equal to p − 1.5. This idea would give birth to an algorithm: Remove all edges that do not disconnect the given graph. 5.166 Basics of Graph Algorithms 5. How many times we use the Bucket-Algorithm eventually decide the overall worst-case complexity. What would be the worst-case time complexity of this algorithm? 5. This approach is opposite to the one discussed above: instead of pruning we are growing edges.2 Growing Edges We start with no edges at all but with p isolated vertices.5 Finding a Spanning Tree of a Graph The algorithm that we design to solve this problem depends on how we visualize the development of a spanning tree. If we just keep a record of all such edges we might get the spanning tree of the given . In the second approach we should be careful not to create cycles in the graph. In each case the Bucket Algorithm helps us.5. We add edges out of the edge pool of the graph such that the resulting graph remains a tree. In the earlier approach we should be careful and should not disconnect the graph. We can start with the original graph and start with pruning or removing edges until the graph becomes a tree. Each approach has its merits and demerits and the comparison itself is very enlightening especially because each approach has more advanced applications. which would constitute the spanning tree.3 Selecting Edges While running the Bucket Algorithm.

What if we find all distinct spanning trees of a given graph using any approach and then select the one with minimum weight? Why is this approach. thus giving rise to different spanning trees.2) which integrates different concepts. which looks at all possible solutions and then selects the one of our choice.2. While cutting edges we select the edge of maximum weight (provided it does not disconnect the graph). How efficient would this be if compared with the algorithms described earlier? 5. This would give rise to an algorithm very similar to Krushkal’s. will provoke the learner to discover a number of interesting techniques. See Figure. which if refined will lead to a number of important algorithms.4 Integrating Concepts and Discovering Algorithms We show a concept map (Concept Map 5. properties and the Bucket Algorithm. having first sorted the edges in descending order of weight. we can grow edges starting from the edge of minimum weight (making sure no cycle is created).5. a number of systematic questions that if asked.6. Similarly. not feasible (although it is correct)? 5.6 Finding a Minimum Spanning Tree of a Weighted Graph There could be many non-isomorphic spanning trees possible for a given graph: each approach that we have described for finding a spanning tree of a graph was flexible and there was a lot of maneuvering possible within it. It is very much possible to have multiple non isomorphic minimum spanning trees of a weighted graph but the weight . 5. It is very much possible to discover most of the algorithms (that we have presented in this chapter) in class once we have become comfortable using and manipulating the Bucket Algorithm. We show in Concept Map 5.Finding a Minimum Spanning Tree of a Weighted Graph 167 graph.6.3. 5.1 Cutting or Growing Edges: A Krushkal’s like greedy algorithm Each algorithm used to find a spanning tree in the previous section could be used with proper modification to find a minimum-spanning tree of a connected and weighted graph.

6. Another minimum spanning tree is shown in the right diagram. The Concept Map and an iterative sequence (ascending order) of asking questions help student discover or understand a number of useful graph algorithms. . A minimum spanning tree of G is shown in the middle diagram.168 Basics of Graph Algorithms Concept Map 5.1: A weighted graph G shown in the left diagram.2. j g 2 4 f 2 e 1 3 a 4 h 3 i 6 9 2 3 k 5 c f 2 1 e 3 d 4 g 2 4 j 3 i 6 9 2 3 b 2 e 3 k a 5 c f 2 1 3 d 4 g 2 4 j 3 i 6 9 2 3 b 2 3 k a 5 c h h d 3 b 2 Figure 5.

3: Pictures of the graph after cutting edges of maximum weight in a Krushkal’s like algorithm. X c .6.2: A weighted graph G shown in the left diagram.Finding a Minimum Spanning Tree of a Weighted Graph 169 j g 2 4 f 2 e 1 3 a 4 h 3 i 6 9 2 3 k 5 c f 2 1 e 3 d 4 g 2 4 j 3 i 6 9 2 3 b 2 e 3 k a 5 c f 2 1 3 d 4 g 2 4 j 3 i 6 9 2 3 b 2 3 k a 5 c h h d 3 b 2 Figure 5.6. j g 2 4 f 2 d 3 e 1 3 a 4 h 3 i 6 3 k 4 5 c 2 b e 1 2 f 2 d 3 g 2 4 j j 3 i g 3 k a 2 3 b e 5 c 2 2 1 d 3 4 3 f a 2 4 h 3 i 3 k h 6 9 5 2 b 2 j g 2 4 f 2 d e 1 3 a 4 h 3 i 3 k 4 f 2 d e 1 3 g 2 4 j 3 i 3 k a c 2 b 2 h c 2 X3 2 b Figure 5. Another maximum spanning tree is shown in the right diagram. A maximum spanning tree of G is shown in the middle diagram. The minimum spanning tree is also shown in the bottom right corner.

.6.4: Pictures of the graph after growing edges of minimum weight in a Krushkal’s like algorithm.170 Basics of Graph Algorithms j g 2 4 3 f 2 d 3 e 1 a 4 h 3 i 6 9 3 k 5 c 2 b e 1 2 f 4 3 2 d g 2 4 j 3 i 6 9 2 2 3 b e 1 3 k a 5 c f 2 d 4 3 g 2 4 j 3 i 6 9 2 2 3 b 3 k a 5 c h h j g 2 4 3 f 2 d 3 e 1 a 4 h 3 i 6 9 2 2 b e 1 3 k 5 c 3 f 2 d 4 g 2 4 j 3 i 6 9 2 2 3 b e 1 3 k a 5 c f 4 3 2 d g 2 4 j 3 i 6 9 2 2 3 b 3 k a 5 c h h Figure 5.

unlike some textbooks.2 Selecting Edges: A Prim’s like greedy Algorithm While forming a spanning tree we can select any cross edge. while there are edges coming out of the Bucket B do Out of all the edges coming out of B. We show here how the simple Bucket Algorithm can be reduced into a minimum spanning tree finding algorithm which resembles Prim’s Algorithm. Without reading proofs given in the textbook they should come up with something of their own making. It would be useful if the students were asked to prove that this greedy approach would actually find a minimum spanning tree. It is important that the minimum spanning tree problem is an optimization problem in which we intend to minimize the sum of weights of all edges in the spanning tree.Finding a Minimum Spanning Tree of a Weighted Graph 171 of each tree would be the same (why?). put this edge in M ST . select one with minimum weight connecting u in B to v not in B.6. We are lucky this time: a so-called greedy approach is working optimally and is in fact optimizing the global sum also.5). In order to form a minimum spanning tree. although greedy approaches are relatively efficient (being based on local conditions only). Put v in B. They should also be asked to derive the time complexity . It follows that among all cross edges that we may select we should pick the one of minimum weight. Please note that here we are not using any fancy data structure since the objective is not to have a complicated design. Using this simple technique the Bucket-Algorithm can easily be modified to find a minimum spanning tree of a weighted directed or undirected graph (Figure 5. Algorithm 18: Find Minimum Spanning Tree of a graph G input : A weighted Graph G output: Minimum Spanning Tree (MST) of Graph G 1 2 3 4 Put any vertex x of Graph G in a Bucket B. However. they are not always optimal. we should try to include edges of less weight thus excluding those of higher weight. In order to minimize the global sum.6. A lively discussion can be initiated to find the merits and demerits of individual work. we are trying to minimize a local quantity. 5.

.172 Basics of Graph Algorithms j 3 j i 6 3 3 g 2 4 3 4 g 2 4 i 6 3 h h k a 2 9 5 4 3 k a 2 9 5 f 2 c d 3 2 f 2 c 2 b e 1 d 3 b e 1 j g 2 4 4 3 j i g 3 2 4 3 i 6 3 h 6 h k a 9 2 5 4 k a 9 2 5 f 2 3 c 2 f 2 3 c 2 d e 1 3 b e d 1 3 b Figure 5.6.5: Pictures of the Bucket B while finding a Minimum Spanning Tree of a weighted graph in a Prim’s like algorithm.

the two approaches look identical.6.6. We show three different techniques in action on the same weighted graph in Fig. We should understand both the differences as well as the similarities. In the top diagrams we select an edge of minimum weight out of all edges coming out of the single bucket in which we have our minimum spanning tree growing. In the middle diagrams we can visualize that each vertex is initially in a separate bucket. This technique (based .6. there is something else. However the time complexity of Prim’s algorithm (as stated in most textbooks) is better. the process of selecting the desired edge out of all edges coming out of the single bucket makes sure that no cycle is generated. something magical which cuts down the time complexity for not so obvious reasons. It would be useful if they compare this approach with Prim’s algorithm. The middle diagrams show a minimum spanning forest growing up in various buckets using a Kruskal’s like algorithm. What is that magic? How and why it is working? Can this magic be used elsewhere and under what conditions? We shall discuss it later in this chapter. In the top algorithm there was no explicit need to check that this condition is true. How about if we start Prim’s algorithm from every vertex? Thus we may be able to avoid an expensive check (that the resulting graph should not be cyclic) each time we insert a new edge (as in Prim’s algorithm) and at the same time we can exploit the inherent parallelism which was lacking in Prim’s algorithm? We initially put every vertex in a separate bucket. Why? The reason is in fact more exciting because Prim’s algorithm is not just greedy. 5. In fact. 5.3 A Panoramic Picture of various MST Finding Techniques It will be useful from a learning perspective if we make a comparison between various minimum spanning tree finding techniques.Finding a Minimum Spanning Tree of a Weighted Graph 173 of this approach. we select an edge out of all edges (coming out of all buckets) in the graph and thus this technique grows various spanning trees in different buckets. In this algorithm (middle diagram) it is essential to check that the new edge that is selected should not form a cycle with the previously found minimum spanning tree. The top diagrams in this figure show how we grow a minimum spanning tree in a single bucket (this is a Prim’s like algorithm). for each bucket we select the edge of minimum weight coming out and thus grow various minimum spanning trees in different buckets.

174 Basics of Graph Algorithms c 2 7 5 4 12 6 b 1 8 c 2 7 5 4 12 6 b 1 8 c 2 7 5 4 12 6 b 1 8 c 2 7 5 4 12 6 b 1 8 g 14 a 3 9 f 11 g 14 a 3 9 f 11 g 14 a 3 9 f 11 g 14 a 3 9 f 11 d e d e Prim’s like Algorithm d e d e c 2 7 5 4 12 6 b 1 8 c 2 7 5 4 12 6 b 1 8 c 2 7 5 4 12 6 b 1 8 c 2 7 5 4 12 6 b 1 8 g 14 a 3 9 f 11 g 14 a 3 9 f 11 g 14 a 3 9 f 11 g 14 a 3 9 f 11 d e d e Krushkal’s like Algorithm d e d e c 2 7 5 4 12 6 b 1 8 c 2 7 5 4 12 6 b 1 8 c 2 7 5 12 6 b 1 8 c 2 7 5 4 12 6 b 1 8 g 14 a 3 9 f 11 g 14 a 3 9 f 11 g 14 a 4 3+43 9 f 11 g 14 a 3 9 f 11 d e d e Boruvka’s Algorithm d e d e Figure 5.6: Pictures of different Buckets while finding a Minimum Spanning Tree of a weighted graph using different techniques including the Boruvka’s Algorithm. .6.

In case of Prim’s algorithm we need not be worried about the formation of a cycle. 2. This is because of the fact that we consider edges coming out of the bucket only. We try to exploit the inherent parallelism in this scheme where each processor is running a Prim’s like algorithm on each vertex. But we do need a cycle test in case the edge weights are not unique. If all edge weights in a graph are unique then we do not need a cycle test in Boruvka’s algorithm. 3. It will also . In case of Kruskal’s algorithm either we are growing a minimum spanning tree (by inserting edges starting from low cost ones) or we are cutting edges starting from high weight edges.6. The big question is do we need to make an expensive cycle check each time we insert an edge? It will be interesting to find under what conditions we need a cycle test and when we do not need such a test? This algorithm terminates when the number of edges selected becomes exactly equal to one less than the number of vertices in the graph. the way we move forward makes sure that no cycle is formed. we grow a minimum spanning tree in each bucket just like Prim’s algorithm. It will be useful to formally prove that in case of Prim’s algorithm we do not need a cycle test and still no cycles will be formed. Why? Do we need a cycle test in Kruskal’s algorithm in case all edge weights are unique? We do not need a cycle test in Prim’s algorithm in spite of the fact that edge weights are not unique? Why? 5. The algorithm terminates when all edges have been considered. 5. if. 4. Prim’s algorithm terminates when all vertices come inside the bucket. In the first case we have to make sure that no cycle is created and in the second case we should be worried about disconnecting the graph. however we consider all edges incident on a vertex inside a bucket then we do need to be worried about the formation of a cycle.6. It is interesting to note the following about the three minimum spanning tree finding algorithms: 1.Finding a Minimum Spanning Tree of a Weighted Graph 175 on the Boruvka’s algorithm) is shown in action (on the same weighted graph) in the bottom diagrams of Fig. In case of Boruvka’s algorithm.

and find the maximum spanning tree while showing the contents of the Bucket B after each step. 5. Do you think there will be any complication in finding a maximum spanning tree if the edge weights are negative (or if they are positive). Problem 5.4 The Maximum Spanning Tree Problem All the minimum spanning tree algorithms described here can easily be modified in order to find the maximum spanning tree of a weighted graph G. Put v in B. We shall discuss that the shortest path problem becomes complicated if all edge weights are not positive. Apply the above algorithm to the following graph G consisting of positive as well as negative edge weights.1. Apply this algorithm to the graph shown in Fig.1.6.7. A curious reader may have realized by now that positive or negative edge weights in a weighted graph does not create any problem while evaluating the minimum or maximum spanning tree of a graph. while there are edges coming out of the Bucket B do Out of all the edges coming out of B. We can make a small change in one of the Minimum Spanning Tree Algorithms in order to convert it into a Maximum Spanning Tree Algorithm as follows. Algorithm 19: Find Maximum Spanning Tree of a graph G input : A weighted Graph G output: A Maximum Spanning Tree (MaxST) of Graph G 1 2 3 4 Put any vertex x of Graph G in a Bucket B. Do we face a similar problem while finding a minimum spanning tree of a graph having negative edge weights? . A small change in the algorithm can do the job. select one with maximum weight connecting u in B to v not in B. put this edge in M axST . Problem Set 5. We shall repeat this problem while finding the shortest path (or the longest path) in a graph having negative edge weights.2.1.6. We shall address these and other related issues in a problem set. Problem 5.1.176 Basics of Graph Algorithms be interesting to prove that in case of Boruvka’s algorithm no cycles will be formed even if we do not make a cycle test provided all edge weights are unique. 5.

6. a time would come when we would reach our destination.1 Cutting Edges If we remove all redundant edges from a given graph and just keep the edges essential to keep the two vertices connected. What is wrong with this approach? If there are cycles in the graph it is possible that we never reach our destination.7: A Graph G having negative edge weights for problem set. Now instead of checking whether the graph is connected or not.1). What would be the worst-case time complexity of this algorithm? Note that we have used a similar technique to find a spanning tree of a graph.Finding a Path in a Graph 177 Figure 5. 5. It would be useful to pinpoint the similarities as well as the differences. . 5. 5. the remaining graph would be a “straight forward” path between the two vertices in the given graph (Fig. since we might have to do a lot of backtracking.7. What if there are no cycles in the graph – what if we first make a spanning tree of the graph? Even now it would be difficult to find a path. we better check if the two given vertices belong to a single connected component. If we keep moving along the edges connecting one vertex to another within the graph.7.7 Finding a Path in a Graph It is possible to find a path between two vertices provided the graph is connected.

.178 Basics of Graph Algorithms j g h i k g j i h k g j i h k f d e a b c f d e a b c f d e a b c j g h i k g j i h k g h i k f d e a b c f d a b c f d a b c Figure 5.1: Pictures of the graph after cutting edges in order to find a path between vertex a and vertex c.7.

and thus there still exist many diversions. We assume that we are finding shortest paths from a single vertex to all other vertices.The Shortest Path Problem 179 5. that a minimum spanning tree of a weighted graph does not always provide shortest distances from a given vertex. With this additional information would it be easier to find a path from the given vertex. It will still be interesting to investigate how a shortest path algorithm resembles and at the same time differs from a minimum spanning tree algorithm. Do we need a different algorithm from the one used to find a shortest path in a graph with uniform edge weights? Why? If somehow we remove edges of higher weights from the graph without disconnecting the two given vertices.8 The Shortest Path Problem If all edges in the graph were to have the same weight.2 Selecting Edges Does the problem become simpler if we first find the spanning tree of the given graph? Now if we start moving from the given vertex to the destination vertex. so first we should solve this problem (which is simpler) before attacking a more complex one. be a shortest path? If not then what should be done to achieve our objective? Note that it is easier to find a shortest path in a graph with uniform edge weights. 5.1.8. but again we may start our journey in the wrong direction and would have to backtrack. 5. . found using the algorithms of the previous section.7. would the problem become simpler? What if we first find a minimum spanning tree of a graph and then move backwards from the destination to the source vertex in order to find shortest paths as described earlier? It is obvious from Figure 5. Suppose we apply the Bucket Algorithm starting with the given vertex: the spanning tree thus formed would originate from the given vertex since the given vertex would be the root. to the destination vertex? The answer is still “no” because a parent may have multiple children.2). would it be less confusing? Perhaps. would the path. Now assume that the edge weights are different. we would eventually reach the root without any confusion (Fig. Students should experience this confusion and the resultant backtracking. We shall come back to this problem after discussing graph traversal techniques in a later section.7. However if we start from the destination vertex and keep selecting the parent vertex. now the root. We also keep a record of the parent of every vertex in the spanning tree.

2: Various stages in finding a path from every vertex to vertex a in a graph. .7.180 Basics of Graph Algorithms j g h i k a d e j g h i k a d e a f g h i j k c j e d b g h i k c b e c f d g b e j c f d g j i h k a b c f i h k a b c f a f d e b Figure 5.

5. It should also be kept in mind that the minimum spanning tree can easily be modified to find a maximum spanning tree of a graph while it is not possible to do so in case of finding a longest path in a graph G (with positive edge weights).2 Discussion on Dijkstra’s (like) Algorithm It was interesting to compare the described shortest path algorithm with the corresponding minimum spanning tree algorithm – but at the same time . The step by step working of the two algorithms is shown in the figure below (Fig.8. It is obvious that initially the two algorithms produce similar results but then they depart ultimately producing different results. it will become obvious that the two algorithms are derived from a common ancestor – the Bucket Algorithm.8. There is indeed a delicate difference between the two – this difference should make us understand why minimum spanning tree algorithm fails to find a shortest path spanning tree and why a shortest path spanning tree algorithm fails to find a minimum spanning tree of a weighted graph G. 5. The shortest distances (as shown in red color) from vertex a provided by the two spanning trees are different as indicated in these diagrams.1 Dijkstra’s (like) Algorithm We produce both the algorithms side by side.8.1: The middle diagram shows a minimum spanning tree (MST) of a weighted graph G shown in the left diagram. The right diagram shows a shortest path spanning tree (SST) of the same graph.The Shortest Path Problem j g 2 4 f 2 e 1 3 a 2 d 3 b 9 4 h 6 17 j 3 i 3 k 5 c 2 4 7 f 2 6 1 e 3 d 5 11 g 2 4 h 13 a 0 3 2 b 2 2 6 e 9 6 15 j 3 k 5 3 c 4 f 2 3 d 1 5 7 g 23 4 2 4 h 9 a 0 3 6 9 2 b 2 181 3 20 i 3 i 12 3 k 5 c 4 9 2 Minimum Spanning Tree Weight of the tree = 26 Shortest Path Spanning Tree Weight of the tree = 31 Figure 5. 5. The weight of the minimum spanning tree is also different from that of the shortest path spanning tree.2).8.

and Dist(i) of every other vertex i equal to ∞ . a vertex x output: Shortest Distance Dist(i) of every vertex i from x 1 2 3 4 Put vertex x in Bucket B. Put edge (u. select one with minimum weight connecting u in B to v not in B. while there are edges coming out of the Bucket B do Out of all the edges coming out of B. while there are edges coming out of the Bucket B do Out of all the edges coming out of B. Initialize Dist(x) = 0. select the edge for which Dist(u) + w(u. v) in SST . Put vertex v in B and edge (u.182 Basics of Graph Algorithms Algorithm 20: Find Minimum Spanning Tree of a graph G input : A weighted Graph G output: A Minimum Spanning Tree M ST 1 2 3 4 5 Put any vertex x of Graph G in a Bucket B. v). Algorithm 21: Find shortest distance of every vertex from a given vertex x in G & also the shortest path spanning tree of G from x input : A weighted Graph G. v) in M ST . Put vertex v in B. Dist(v) = Dist(u) + w(u. v) is minimum where vertex u is in B and vertex v is outside the Bucket B. .

Out of all such vertices (at a distance of one edge from x).3 where some edge weights are negative. In the first iteration of the while loop. The cost of this action is equivalent to making p − 2 comparisons. and which are indirectly connected to x through vertex i. The cost of this action is equivalent to making p − 2 + p − 3 = 2p − 5 steps. It is obvious from this figure that Dijsktra’s algorithm fails to produce correct results. The vertex j goes in the Bucket B and the corresponding edge goes in shortest path spanning tree. we select a vertex i which is at a shortest distance from x. 5. Remember that this algorithm works on a greedy strategy – it makes decisions on the basis of local conditions – but produces optimal results on a global basis provided all edge weights in a graph are positive. however.The Shortest Path Problem 183 it will be a learning experience if we look into the working of the shortest path algorithm under conditions when it fails to provide optimal results.8. But before we do that let us discuss some of the salient features of this algorithm. We show another weighted graph in the right diagram of this figure. 2. i) goes in the shortest path spanning tree. In the second iteration of the while loop.4. we consider vertices which are directly connected to x. Remember vertex i is the last vertex which went into the Bucket B. be violated if some edge weights in G are negative. We show a directed graph in the Figure 5. the edge (x. we consider vertices at a distance of one edge. In order to assert that Dijkstra’s like algorithm is able to find correct results let us first present a scenario where the above mentioned algorithm fails to find shortest paths. This graph is different from the one shown in the left (and middle) diagram in the following ways: . 1.8. and put it in Bucket B. we select a vertex j which is at a shortest distance from vertex x. Out of all these vertices (some of which are at a distance of one edge and some at a distance of two edges). This action is tantamount to a claim that we have found a shortest path from vertex x to vertex i. The same weighted graph and its (correct) shortest path spanning tree is shown in Fig. If all edge weights are positive then this claim will be right – it will. Again this action means that we have found a shortest path of vertex j from vertex x.

.8.184 Basics of Graph Algorithms Figure 5.2: Pictures of the Bucket B while finding a Minimum Spanning Tree (left diagrams) as compared to the pictures while finding a Shortest Path Spanning Tree (right) from vertex a.

The final answer as shown in the bottom right diagram is incorrect. .The Shortest Path Problem 185 j g 2 4 f 2 e -5 3 a -4 h 3 i 6 9 2 3 k 3 c f 2 e -5 4 3 g 2 -4 j 3 i 6 0 a 2 9 3 k 3 c b 2 2 e 3 f 2 -5 4 3 g 2 j -4 h 3 i 6 0 a 2 9 3 k 3 c b 2 2 h d -4 b 2 d d -4 -4 Vertex a at a shortest distance of zero edge Vertex b at a shortest distance of one edge Vertex f at a shortest distance of (at most) two edges 5 j g 2 4 3 f 2 e -5 3 -4 h 0 a 2 d -4 b 2 2 e 6 9 3 i 3 k 3 c 4 3 f 2 -5 d 3 5 -4 4 0 a 2 b 2 2 0 e 9 g 2 -4 h 6 j 3 i 3 k 3 c 3 f 4 2 -5 d 3 5 -4 4 j 7 g 2 9 -4 h 0 a 2 b 2 2 6 9 3 i 8 3 k 7 3 c 4 Vertex c at a shortest distance of (at most) three edges Vertex d at a shortest distance of (at most) four edges The (incorrect) Shortest Path Spanning Tree Figure 5.8.3: Pictures of the Bucket B while finding a shortest path spanning tree using a Dijsktra’s like algorithm.

The right diagram shows a graph where shortest distances of red vertices from vertex a keep on decreasing if we move in a cyclic path as shown in red. The magnitude of this shortest path is finite.186 Basics of Graph Algorithms 1.4).8. In other words the shortest distance between any two vertices does not reduce if we move in a cyclic path. 5. . 5.8. This happens in any graph where there are negative weight cycles. 5 j 7 g 2 4 3 f 2 0 e -5 d 9 3 5 -4 -4 h 0 a 2 b 2 2 0 e 6 9 3 i j 8 3 k 7 3 c 4 3 f 2 3 7 4 g 2 9 5 3 i 6 0 a 2 9 8 3 k 6 4 3 c 3 2 1 e f 2 -5 g 2 -12 j 3 i 6 0 a 9 2 -4 b 2 3 k 3 c -4 h h 3 d d -5 5 -4 b Dijkstra’s shortest path spanning tree (incorrect) (Correct) shortest path spanning tree Shortest distances keep on decreasing in a negative weight cycle Figure 5.4) the above observations are not true. The middle diagram shows the correct shortest path spanning tree for the same graph. The above observation is true for the left or the middle graph in the figure below (Fig.8. The shortest distance path from vertex a to any other vertex is a simple path – no edge or vertex is repeated in this path.4: The left diagram shows an (incorrect) shortest path spanning tree produced by a Dijkstra’s like algorithm. In the graph shown in the right diagram (Fig. Please note that there are negative weight edges in this graph but no negative weight cycles: a negative weight cycle is a cycle in a graph where the net sum of the weight of the edges in the cycle is less than zero. The shortest distances of the red vertices keep on decreasing if we move in a cyclic path as shown in red color with the result that the shortest distance of such vertices ultimately approaches minus infinity. 2.

this happens when k = 5. Once we have the 1-edge shortest paths – we can convert them into 2-edge shortest paths and then into 3-edge shortest paths according to the algorithm described below. Let us explore the consequences of the presence of negative weight cycles in a graph like the one shown in the right diagram of Fig.5. This graph is reproduced in Fig.5. If there are no parallel edges in the graph then these paths will also be the shortest paths from vertex x to every vertex adjacent to x. The weighted adjacency matrix of the given directed graph provides all one edge paths from vertex x. It is evident from the figure below – as we move from k-edge to (k + 1)-edge shortest paths there is some improvement in the length of a shortest path.8. input : A weighted directed Graph G.8. it is denoted by Distk+1 (i) 1 2 3 for i = 1 to p do for j = 1 to p do Distk+1 (i) = min{Distk (i).8. 5. now we claim that we have found the shortest distance of every vertex from x.3 The Shortest Path Problem Redefined: The kedge Shortest Path Problem In order to overcome the above mentioned complication we redefine the shortest path problem as follows.6 where we find different edge path from vertex . A k-edge shortest path (between two given vertices) requires that the path should be shortest but it should not consist of more than k edges. if no improvement takes place in any path then we stop.8. now if we increase k to 6 there is no improvement in any shortest distance. 5. a vertex x. For the graph shown in the Fig.4.8. Algorithm 22: Find (k+1)-edge shortest distance of every vertex from a given vertex x in a weighted directed graph G. w(i. Such one edge paths are shown in the top left diagram in Figure 5. k-edge shortest distance of every vertex i from vertex x denoted by Distk (i) output: (k + 1)-edge Shortest distance of vertex i from vertex x. j) + Distk (j)}. 5.The Shortest Path Problem 187 5. Instead of finding the shortest path between two vertices we intend to find k-edge shortest paths from a given vertex to every other vertex in a directed graph containing negative edges – the graph may even contain negative weight cycles.

.5: We show k-edge shortest paths from vertex a to every vertex in graph G. When k changes from 1 to 5.188 Basics of Graph Algorithms j g 2 4 3 f 2 e -5 3 0 a -4 h j 3 i 6 9 2 3 k 9 3 c b 2 2 e 3 f 2 d -5 5 7 4 3 0 a g 2 -4 h 3 i 12 6 9 3 k 9 3 c 4 b 2 2 0 e 3 f 2 -5 3 d 5 7 4 g 2 9 -4 j 3 i 6 0 a 2 -4 b 12 3 k 9 7 3 c 4 2 1 h 2 -4 d -4 Vertices at one edge shortest distances 5 3 i 6 0 a 3 2 0 -5 e Vertices at (at most) four edge shortest distances: Not Stable 2 d 5 -4 b 2 1 0 e 9 Vertices at (at most) two edge shortest distances: Not Stable j 5 3 i 6 0 a 3 2 -5 2 d 5 -4 b 2 1 0 e 9 8 3 k 6 3 c 3 7 4 3 f 2 Vertices at (at most) three edge shortest distances: Not Stable j g 2 9 5 3 i 6 0 a 3 2 d 5 -4 b 2 1 -5 9 8 3 k 6 3 c 3 j 7 4 3 f g 2 9 -4 h 10 3 k 7 4 3 c 3 3 f 7 g 2 9 -4 h -4 h Vertices at (at most) five edge shortest distances: Not Stable Vertices at (at most) six edge shortest distances: STABLE Figure 5. But when k increases beyond 5 then there is no change in shortest paths. shortest path distances of some vertices change.8.

8. and this situation never becomes stable. . When k changes from 1 to 6. But when k goes beyond six then we are caught in a negative cycle – the shortest path of three vertices (belonging to a negative weight cycle and shown in red color) keeps going down.6: We show k-edge shortest paths from vertex a to every other vertex in graph G. the shortest paths of some vertices change.The Shortest Path Problem j g 2 4 3 f 2 e -5 3 0 a 2 d -4 b 2 2 e 9 -12 h 6 j 3 i 3 k 9 3 c 3 f 2 d -5 5 7 4 3 0 a 2 -4 b 2 2 0 e Vertices at (at most) three edge shortest distances 9 g 2 -12 h 6 j i 12 3 k 9 3 c 4 3 f 2 -5 3 d 5 7 4 g 2 9 -12 h 0 a 2 -4 b 6 9 3 i 189 3 12 3 k 7 3 c 4 2 1 Vertices at one edge shortest distances -3 3 -12 2 9 3 2 0 -5 e Vertices at (at most) four edge shortest distances 2 d 5 -4 b 2 1 0 e h 0 a 6 9 i Vertices at (at most) two edge shortest distances -3 3 i 6 0 a 3 2 -5 2 d 5 -4 b 2 1 9 0 3 k 6 3 c 3 3 f 7 4 j 7 4 3 f g j 10 3 k 7 4 3 c 3 3 f 7 g 2 9 j -3 g 2 6 0 a 3 2 0 -5 e -12 h -12 h 3 i 6 0 3 k 9 6 3 c 3 2 d 5 -4 b 2 1 Vertices at (at most) five edge shortest distances Vertices at (at most) six edge shortest distances: Still Not Stable j -6 7 4 3 f 2 0 -5 e g 2 6 0 a 3 -12 h 3 i 6 j -6 0 3 k 9 6 4 3 c 3 b 2 1 3 f 2 0 -5 e 3 7 g 2 6 0 a -12 h 3 i 6 j -3 3 k 9 6 4 3 c 3 b 2 1 3 f 2 0 -5 e 3 7 g 2 -12 -∞ 3 6 0 a 2 i -∞ 3 k 9 6 3 c 3 b 2 1 -∞ h 2 d 5 -4 2 d 5 -4 d 5 -4 Vertices at (at most) seven edge shortest distances: Still Not Stable Vertices at (at most) eight edge shortest distances: Still Not Stable Vertices at (at most) infinite edge shortest distances Figure 5.

the same weighted graph is reproduced below in the left diagram of Fig. it may become stable after some time as shown in Fig. 5. . Remember if we increase k beyond 5 while finding k-edge shortest paths then no (shortest) distance (with respect to vertex a) changes with k. It means that k would be less than or equal to p − 1 in case there are no negative weight cycles in the graph. 5. Problem Set 5. This situation never becomes stable because of the presence of a negative weight cycle comprising three vertices shown in red color. As soon as all the k-edge shortest paths become stable with increase in k. As we increase k the longest path of certain vertices increases.8.8. Find k-edge shortest paths with respect to vertex a while k changes from 1 to 10.8. An Important Conclusion If there are no negative weight cycles in a directed graph consisting of some negative weight edges then we can use our k-edge shortest path technique to find shortest paths. This is quite expected in view of our prior experience: if there are negative weight cycles in a graph then the shortest path (of certain vertices) keep increasing with k.5. We now change the magnitude of the weight associated with the edge (f. we claim that we have found the shortest paths. Thus k keeps increasing and the distances keep going down. Problem 5. Find how and at what value of k.2. 5.1. however. Please note that the resulting shortest paths will be simple paths.190 Basics of Graph Algorithms x. the shortest paths becomes stable. g) from 4 to 40 as shown in the middle diagram of this figure.6. When k increases the shortest path distances go down. there are positive weight cycles in the graph then we shall be caught in an infinite loop and the longest distances (of at least some vertices) will keep increasing with increase in k.7 where we operate on the same graph of Fig. 5. it is just like maximum spanning tree versus minimum spanning tree problem.8.4 The k-edge Longest Path Problem The k-edge longest path problem is quite similar to the k-edge shortest path problem.8. Please recall the weighted graph of Fig. If. 5. no edge or vertex is repeated in these paths.2. The shortest distances of the rest of the vertices become stable as soon as k approaches 5.8.

8.8.The Shortest Path Problem j g 2 4 3 f 2 e -5 3 0 a 2 d -4 b 2 2 e 9 -12 h 6 j 3 i 3 k 9 3 c 3 f 2 d -5 5 7 4 3 0 a 2 -4 b 2 2 0 e Three edge (at most) longest paths 9 g 2 -12 h 6 j i 12 3 k 9 3 c 4 3 f 2 -5 3 d 5 2 -4 b 2 2 7 4 g 2 18 -12 h 0 a 6 9 3 i 191 3 12 3 k 9 3 c 4 One edge longest paths Two edge (at most) longest paths 6 j 3 i 6 0 a 2 -4 b 7 4 3 f g 2 -12 h 18 3 12 3 k 9 9 3 c 4 2 2 0 e 3 f 2 -5 7 4 g 2 6 j -12 h 18 3 d 5 3 i 6 0 a 2 -4 b 12 3 k 9 9 3 c 4 2 2 2 -5 0 e d 5 Four edge (at most) longest paths Not Stable Five edge (at most) longest paths Stable Figure 5. j g 2 4 f 2 d e -5 -4 3 0 a 2 b 2 2 e 9 -4 h 6 3 i 3 k 3 c f 2 -5 40 3 0 a 2 d -4 b 2 e 9 g 2 -4 h 6 j 3 i 3 k 3 c f 2 d -5 -40 4 3 0 a 2 b 2 9 g 2 -4 h 6 j 3 i 3 k 3 c Figure 5. But when k goes beyond four then the longest distance of any vertex does not change.7: We show k-edge longest paths from vertex a to every other vertex in graph G. the situation becomes stable and we claim that we have found longest paths in this graph. When k changes from 1 to 5 the longest path of some vertices improves.8: A weighted graph G for a problem set .

4. Problem 5.8. Repeat the above problem for the graph shown in the right diagram of Fig.8. We show another weighted graph as shown in the right diagram of Figure 5.2. 5.2.8.8.8.8.2.2. Problem 5.192 Basics of Graph Algorithms Problem 5. We already know that if there . Find k-edge longest paths for each vertex with respect to vertex a in the graph shown in the left diagram of Fig. 5.3. if we further increase k then there is no change in the shortest distance of any vertex with respect to vertex a. Comment on the claim that the longest path problem is a hard problem while the shortest path problem is a solvable problem.5 The Shortest Path Problem in Undirected Graphs with Negative Weights We consider the shortest path problem in undirected graphs with negative edge weights but no negative weight cycles. Is it possible to use the above algorithm to find if an undirected graph contains negative weight cycles? Problem 5. Find the value of k at which the shortest distances become stable.2.8.8. Find if the longest paths become stable with increasing value of k. Please note that there is a big negative weight equal to -40 associated with the edge (d. the shortest distances become stable at a finite value of k? Why? What do you think are necessary and sufficient conditions for the shortest distances to become stable for a finite value of k in a weighted graph? Problem 5. 5. i. vary k from 1 to 11.2. 5. b). Problem 5.9.6. Find k-edge shortest distances of all vertices with respect to vertex a. We need to explore if shortest distances of all vertices with respect to vertex a become stable when k increases while finding k-edge shortest paths.2. In spite of a large negative weight in the graph shown in the right diagram of Fig.7.8.2.8. Problem 5.e.2. What do you think are necessary and sufficient conditions for the longest distances to become stable for a finite value of k in a weighted graph? Problem 5.5. Describe an efficient algorithm (based on finding k-edge shortest distances) to find if a directed graph contains negative weight cycles..

The undirected graph is converted into a directed graph as shown in the right diagram.8.8.9.8. When we apply our k-edge shortest path algorithm to the directed graph (shown in the right diagram of Fig.9) we run into a complication. Each such shortest path starts from vertex x and terminates . With negative edge weights (but no negative weight cycles) we expect that our kedge shortest path technique will work with the hope that the k-edge shortest paths will stabilize for a finite value of k and we shall get the optimal answer.9. 0 0 5 a 5 b a 5 5 a 5 6 -3 6 -3 2 2 c b b 6 d -3 6 -3 2 c d 4 2 c 2 d Figure 5. 5. The correct shortest paths & distances with respect to vertex a are indicated in the middle diagram. The working of the k-edge shortest path technique on this graph is shown in Fig.8. 5. The undirected graph is converted into a directed graph as shown in the right diagram of this figure. 5. Let us represent a k-edge shortest path of any vertex j from vertex x by SP athk (j). we have created a two-edge negative weight cycle in the directed graph. The shortest distances with respect to vertex a are indicated in the middle diagram. It is quite evident that k-edge shortest distances do not stabilize in this graph with increase in k. and we get incorrect answers for shortest distances.8.10. Remember that the so called technique was designed for directed graphs and in order to apply it to undirected graphs we have to first convert the undirected graph into a directed one. The challenge is how to handle this complication? Please understand that we are limiting our focus on a directed graph which is derived from an undirected graph as shown in Fig.9: We show an undirected graph with negative edge weights but no negative weight cycles in the left diagram. We show an undirected graph in left diagram of Fig.The Shortest Path Problem 193 are no negative weights then a simple greedy strategy (like that of Dijskrta’s Algorithm) will solve the shortest path problem in undirected graphs. 5.

10: Shows k-edge shortest paths & distances with respect to vertex a with k changing from one to five.8. .194 Basics of Graph Algorithms 0 5 5 0 5 5 0 5 -1 a 5 6 -3 b 6 a 5 -3 b 6 a 5 -3 -3 b 6 3 2 2 6 6 -3 2 2 d 6 c ∞ d 6 2 2 c 2 d 4 c 2 One edge shortest paths Two edge (at most) shortest paths Three edge (at most) shortest paths 0 5 -1 5 5 -7 a 6 b -3 6 a 5 b -3 6 -3 2 2 6 -3 2 2 d 4 c -4 d -2 c -4 Four edge (at most) shortest paths Five edge (at most) shortest paths Figure 5. The shortest distances do not become stable for any finite value of k.

3. Problem 5.1.3.9 and verify that it provides correct results for values of k in the range of 1 to 5. Apply the above Algorithm to the graph shown in Fig. We must make every . Will it be possible to provide a warning that the given graph contains negative weight cycles? What modifications are needed in our existing algorithms to solve this problem? 5. Algorithm 23: Find (k + 1)-edge shortest path & distance of every vertex from a given vertex x in a weighted directed graph G. 5. Its weight is represented by Distk (j). Problem 5. output: (k + 1)-edge Shortest path of vertex i from vertex x. its weight is equal to Distk+1 (i) 1 We need to find SP athk+1 (i): We consider k-edge shortest path for every vertex j in the graph except for j where yj = i. The last vertex in a shortest path SP athk (j) is vertex yj before terminating at j. Out of the remaining paths we claim that Distk+1 (i) = min{Distk (i). k-edge shortest path of every vertex j from vertex x denoted by SP athk (j). it is denoted by SP athk+1 (i).9.8. Now we are in a position to find (k + 1)-edge shortest path of a vertex i from vertex x in a graph using the following algorithm.8. 5. a vertex x. w(i.8. Problem Set 5.3.9 Graph Traversal Techniques It is possible to traverse a graph in a haphazard manner.3. Problem 5. Check if k-edge shortest paths stabilize for a finite value of k. Consider an undirected graph containing negative weight cycles.3.Graph Traversal Techniques 195 at vertex j. 5.11 for values of k in the range from 1 to 8. assume that the last vertex in this shortest path is yj before terminating at j. j) + Distk (j)} . Apply the above Algorithm to the graph shown in Fig. Efficiency demands that we do not visit the same vertex again and again. input : A weighted directed Graph G. Please note that this graph contains negative weight cycles.2. Compare your results with the correct shortest paths given in the middle diagram of Fig.

11: We show an un-directed graph (left diagram) with a negative weight cycle.9. 5. in most of the current textbooks. Baase [2] uses JAVA to describe algorithms and this may be one reason why the book is relatively difficult to read even if students have prior knowledge of the language. Cormen [17] and Skiena [16] use a pseudo programming language and operate at a slightly higher level. You might have noticed that the Bucket-Algorithm is essentially a graph traversal algorithm. The undirected graph is converted into a directed graph as shown in the right diagram.196 Basics of Graph Algorithms a b a b d c d c Figure 5.1 Traditional Techniques & the Bucket-Algorithm It is interesting to note that the Breadth as well as Depth First Searches are two different implementations of the Bucket-Algorithm (Fig.9. 5. you would only find some very specific techniques like the Breadth and Depth First Search traversal algorithms – more specialized and less flexible than our Bucket Algorithm.8. While introducing this algorithm we purposely did not disclose the implementation details ignoring the underlying data structure required to program the algorithm.2 The Underlying Data Structure We know that we use a cross edge to discover a new vertex in the BucketAlgorithm (step 3). However. move in a systematic manner to ensure that we do not miss out any vertex belonging to the same connected component [3]. Some of these cross edges come from vertices that en- . The Bucket Algorithm is simple because it is more abstract and flexible. The objective was to highlight the basic idea and initially suppress the programming details. 5.9.1).

9.Graph Traversal Techniques 197 j g h i k a d e BFS S DF j g h i k a d e BFS j g h i k a d e b c f c b f c b f j g h i k a d e b c DFS j g h i k a d e b e c f d g j i h k a b c f f Figure 5.1: Pictures of the Bucket B while performing a BFS (DFS) traversal in a graph starting from a vertex a shown in the top (bottom) diagram. .

j g h i k a d e b e c f d g h j i k a b e c f d g h j i k a b c f Figure 5. or a combination of the two. Claim 5.1 below.1.10.198 Basics of Graph Algorithms tered the bucket earlier. The non spanning tree edges are shown in pink color by thin lines. Please see the Figure 5. others from vertices that are new comers in the bucket.10. In the first iteration of the BFS traversal. 5.9. Using a Last in First Out (LIFO or a stack) or a First In First Out (FIFO or a queue) data structure to store the already discovered vertices would make all the difference: a stack implementation would convert the Bucket-Algorithm into a Depth First Search while a queue would transform it into a Breadth First Search. in the k th iteration of the BFS traversal. . The way we decide which vertex to choose would convert the BucketAlgorithm into a Breadth First Search.10. all vertices at a distance of exactly k edges from the starting vertex will be discovered. BFS spanning tree of a graph G is a minimum distance spanning tree in terms of number of edges between the starting vertex and any other vertex in graph G. all vertices at a distance of one edge from the starting vertex are selected (or goes in the Bucket). Depth First Search.10 Some Graph Theoretic Claims We make the following claims about the BFS traversal in an un-directed and connected graph G: Claim 5.2: A Graph G shown in the left diagram. its BFS spanning tree is shown in the middle while a DFS spanning tree of this graph is shown in the right diagram.2.

vertices at a distance of one edge from vertex a goes in the bucket. the middle diagram shows that vertices at a distance of two edges moves in the bucket in the 2nd iteration.1: Pictures of the Bucket B while performing a BFS traversal in a graph starting from a vertex a shown in the top diagrams. and so on. The distance of each vertex with respect to vertex a is indicated in red color along with each vertex in the bottom diagrams.10. The bottom left diagram shows that in the first iteration of the BFS algorithm. .Some Graph Theoretic Claims 199 j g h i k a d e b e c f d g j i h k a b e c f d g j i h k a b c f a 0 a 0 f 1 a 0 f 1 k 1 g 2 2 b 1 k 1 f 1 b 1 b 1 k 1 g 2 2 d e 3 c 2 i 2 d c 2 i 2 3 3 j h Figure 5.

If there is an edge in G other than the ones in the BFS spanning tree of G connecting vertex x and vertex y in the BFS spanning tree such that Dist(x) − Dist(y) is zero then there will be an odd cycle in graph G. h makes a cycle. If there is an edge in G other than the ones in the BFS spanning tree of G then this edge will connect two vertices x and y in the BF S spanning tree such that Dist(x) − Dist(y) is either zero or one whereas Dist(x) is the distance of vertex x from the starting vertex in the BFS spanning tree. Please see Fig. If there is no edge in G other than the ones in the BFS spanning tree of G then G is a tree provided G is a connected graph. Claim 5.4.10. Claim 5.10.2: BFS spanning tree of a graph G is shown. As we know the distance of vertex h from vertex a.3.10.2. If there is an edge in G other than the ones in the BFS spanning tree of G connecting vertex x and vertex y in the BFS spanning .10. The edge between vertex h and vertex j creates an odd cycle in the graph as both these vertices are at the same distance from vertex a. We also know a path between vertex j and vertex a.6.10. Please see Fig. we also know a path between vertex h and vertex a as shown in green color in the top right diagram.10. these two paths along with edge j. It will be an odd cycle as vertices h and j are at the same distance from vertex a. 5.5.10. 5.2 Claim 5. (Remember Dist(x) is the distance of vertex x from the starting vertex in the BFS spanning tree). j g h k f d e a b h c g 2 2 3 a 0 i f 1 b 1 k 1 d e 3 c 2 i 2 3 j Figure 5.200 Basics of Graph Algorithms Claim 5.

0 j g h k f d e a b 3 a i f 1 b 1 k 1 c g 2 2 d e 3 c 2 i 2 3 j h Figure 5. When we grow a BFS spanning tree in a Bucket starting from a given vertex.8.10. Claim 5. Similarly if there is an edge of G connecting vertex x to vertex y in the BFS spanning tree of G such that Dist(x) − Dist(y) is exactly equal to 0 then graph G is not a bipartite graph.10. Claim 5.10. Please see the Fig.4 below. Claim 5.3 below. the given vertex goes in the bucket first. If a graph G is a tree then there will be a unique path between every pair of vertices of G.Some Graph Theoretic Claims 201 tree such that Dist(x) − Dist(y) is one then there will be an even cycle in graph G. The edge between vertex h and vertex i creates an even cycle in the graph G as vertex h is at a distance one larger than the distance of i from a.10. Please see Fig.10. If every edge of G connects a vertex x to vertex y in the BFS spanning tree of G such that Dist(x)−Dist(y) is exactly equal to 1 then graph G is a bipartite graph.10.9. Similarly if there is a unique path between every pair if vertices in a connected graph G then G is a tree. Claim 5.10.7. 5. (Remember Dist(x) is the distance of vertex x from the starting vertex in the BFS spanning tree).10. In the first . 5.3: BFS spanning tree of a graph G. A graph G is bipartite if and only if it does not contain any odd cycles.

BFS spanning tree of another graph G which is a not bipartite is shown in the bottom diagrams.202 Basics of Graph Algorithms j g h k f d e a b 3 a 0 i f 1 b 1 f a k 1 d c 2 i 2 g c 3 j b k e h c g 2 2 d e 3 h i j f a 0 j g h k f d e a b h c g 2 2 3 i f 1 b 1 a k 1 d g c e 3 3 j b k e h d c 2 i 2 i j Figure 5.10.4: BFS spanning tree of a graph G which is a bipartite graph is shown in the top diagrams. .

10.Some Graph Theoretic Claims 203 iteration of this algorithm.11. At any point in time there will be k vertices and k − 1 edges in the Bucket. Initially there will be only one vertex and no spanning edge as shown in the top left corner.10.5: We show the contents of the Bucket at different times while we make a BFS traversal of a graph. Finally there will be 10 spanning edges and 11 vertices in the Bucket. j g h k f a b e No spanning edge & One vertex in the Bucket j i g h k f a b e One spanning edge & Two vertices in the Bucket j i g h k f a b i c c c d e d d Two spanning edges & 3 vertices in the Bucket j g h k f a b e f c d i g j i h k a b e 4 spanning edges & five vertices in the Bucket j g h k f a b i c c d e d 3 spanning edges & four vertices in the Bucket 10 spanning edges & 11 vertices in the Bucket Figure 5.10. Please see the Fig.5. 5. . one more vertex and an edge go in Bucket. If G is a connected graph and if the number of edges is one less than the number of vertices in G then G is a tree. Claim 5. we discover a new vertex because of an edge going out of the Bucket.

Modifying the Bucket algorithm to find a shortest path spanning tree (SST) or shortest distances in a weighted graph from a given vertex. We shall again modify the Bucket algorithm to achieve our objective. Finding k-edge shortest or longest paths in a weighted graph. Modifying the Bucket algorithm to find a minimum spanning tree (MST). faster all pair. 4. . Design of single source shortest path algorithms based on our prior knowledge of finding k-edge shortest paths from a given vertex. 3. We intend to do the following in this section: 1. This is about single source shortest path algorithms assuming that all edge weights are positive in the given weighted graph. Here we assume that there may be negative edge weights in the given weighted directed graph. Design of all pair shortest path algorithms including the slow all pair. We shall further modify the Bucket algorithm to achieve our objectives. 3.12. 5. Design of a shortest path algorithm for directed acyclic graphs with negative as well positive edge weights. We also assume that the directed graph may have cycles. The interesting thing about this style of design is that we shall be using a single building block (sometimes the Bucket Algorithm and some times a 2-edge Shortest path Algorithm) to design or describe an algorithm. If G is acyclic & number of edges in G is one less than the number of vertices then G is a tree.11 Shortest Path Algorithms We have already done the following in earlier sections of this chapter. Floyd-Warshall and Johnsons shortest path algorithms. 2. 1. Analyze the existing minimum spanning tree & shortest path finding algorithms and improve their efficiency as far as possible. On the basis of that precious prior knowledge we can design interesting shortest path algorithms. 2.204 Basics of Graph Algorithms Claim 5.10.

We shall study this algorithm using different tools before improving its time complexity. We introduce a visual tool known as ”opening up the graph” as shown in Fig. The visual tool demonstrates at what stage and when a vertex enters the Bucket while the shortest path algorithm moves forward.Shortest Path Algorithms 205 5. Dist(k) = Dist(j) + w(j. Put edge (j.11.11. Put vertex k in B. Food for thought: The graph shown in Fig. Initialize Dist(a) = 0. But we also know that this very algorithm provides correct results when all edge weights are positive.11.1. k). 5. and Dist(k) of every other vertex k from vertex a equal to ∞.1 Single Source Shortest Path Algorithms with positive edge weights We copy here the modified Bucket algorithm designed in the last section to find shortest paths from a given start vertex. We apply the shortest path algorithm and find the shortest distances from vertex a in this graph where a = 1 also shown in Fig. while there are edges coming out of the Bucket B do Select the edge for which Dist(j) + w(j.1. As soon as a vertex enters the bucket its color changes from blue to orange and then its distance from the start vertex cannot change . That is why the shortest paths found by the crude Dijkstra’s algorithm are not correct. . 5. We have already witenessed that this algorithm does not always provide correct results for graphs with negative weight edges.11. Algorithm 24: (Crude-Dijkstra): Find shortest distance of every vertex from a given vertex a in G & also the shortest path spanning tree (SST ) of G from vertex a input : A weighted Graph G. 5. k) in SST . The shortest distances from vertex 1 as found by this algorithm are also indicated in the bottom diagram.it is fixed and finalized.1 has negative edge weights. a vertex a output: Shortest distance Dist(k) of every vertex k from vertex a 1 2 3 4 5 Put vertex a in Bucket B. k) is minimum where vertex j is in B and vertex k is outside the Bucket.

11. .206 Basics of Graph Algorithms Figure 5.1: We use a new tool ”Opening up the graph” as shown here. It shows at what stage (edge distance) a vertex enters the Bucket while executing crude shortest path algorithm. It also shows how and when the distance of a vertex changes from the start vertex.it is fixed and finalized. As soon as a vertex enters the bucket its color changes from blue to orange and then its distance from the start vertex cannot change .

this will automatically cut down the time complexity of this single source shortest path algorithm to p2 . In fact there are two complications with this algorithm: 1. Now we consider the edges coming out of the Bucket from (only) vertex a instead of edges coming out of the Bucket from all vertices in the Bucket. however. 2. As this loop runs as many times as p so the overall time complexity will be p3 of this algorithm.3. Surprisingly this can be done with a slight modification in the crude algorithm as shown below. The outcome of this step is that the number of edges (which really matters) coming out of the bucket B from vertex a is limited by p and not p2 . The shortest paths.2: The number of edges coming out of the Bucket B at any time will be proportional to p2 under worst case conditions. could not be found . It is interesting to note how the shortest distances are provided in the output in this new algorithm. Somehow we should reduce the number of relevant edges coming out of the Bucket to as small as p . This requires as many comparison steps to move forward in the while loop. The number of edges coming out of the Bucket B at any time is proportional to p2 as illustrated in Fig.Shortest Path Algorithms 207 Figure 5. Although it provides correct results for positive edge weights. The issue of negative edge weights will be handled in the next sub-section. 5. Its time complexity is too high.2.11. What we essentially do is to change the graph itself at each step as shown in Fig.11. 5.11. We shall reduce the time complexity in the following refined version of shortest path algorithm. It does not provide correct results for negative edge weights.

Algorithm 26: Find weight of MST of a weighted graph G input : Adjacency matrix of a weighted Graph G output: Modified adjacency matrix of graph G in which the row corresponding to vertex a gives weight of MST 1 2 3 4 5 6 Put any vertex a in Bucket B. j) + w(j. Put vertex j in B. It is important to note that we may be able to make similar modifications in the crude minimum spanning tree algorithm to make it more efficient. k) We can recover shortest paths by adding a parent table as shown in Fig. a vertex a output: Modified adjacency matrix of graph G in which the row corresponding to vertex a gives shortest distances of every vertex from vertex a. Put vertex j in B for every edge coming out of B from vertex j do assign min{w(a. w(a. while there are edges coming out of the Bucket B from vertex a do Select the edge for which w(a. j) is minimum where vertex j is outside the Bucket B. . Again it will be interesting to understand that this algorithm will provide the weight of the minimum spanning tree of a weighted graph . k) without an extra effort using this refined algorithm. k)} to w(a. for every edge coming out of B from vertex j do assign min{w(a. The refined version of that algorithm is given below . w(j. 1 2 3 4 5 Put vertex a in Bucket B while there are edges coming out of the Bucket B from vertex a do Select the edge for which w(a. k). k)} to w(a. k).not the minimum spanning tree itself as shown in the diagrams below.its time complexity also reduces from O(p3 ) to O(p2 ). j) is minimum where vertex j is outside the Bucket B.208 Basics of Graph Algorithms Algorithm 25: (Refined-Dijkstra):Find shortest distance of every vertex from a given vertex a in weighted graph G input : A weighted Graph G.

3: We modify the graph as we move forward in the shortest path algorithm. j)+w(j.k). .Shortest Path Algorithms 209 j 3 g 2 3 f 2 1 e 3 a 2 d 3 1 2 b e 4 h 6 9 i 3 k 3 c f 2 d 1 3 3 5 g 2 4 j 3 i 6 9 a 2 2 3 1 b e 4 3 k 3 c f 2 d 1 3 6 3 5 g 4 2 j 3 i 6 9 a 2 2 3 1 b 4 3 k 3 c h h ut sb ce s tan path is t d test r es o rt h o sh he s the of t d fin rack We se t lo j 3 g 2 6 a f 2 d 1 e 5 3 1 3 5 2 4 h 3 6 7 i 3 w(a.11.k) = min{w(a.k)} j 3 g k 3 2 6 a f 2 d 1 e 5 3 1 3 5 2 2 b e 4 h 3 c 6 7 3 4 c f 2 d 1 5 3 1 i 3 k 3 6 3 5 a 2 2 b g 2 4 h 6 7 3 4 c j 3 i 3 k 4 2 b Figure 5. w(a.

We have provided a hint in the same figure.2 Single Source Shortest Path Algorithms for Directed Acyclic Graphs We know that by definition a directed acyclic graph contains no cycles.11. If we can recover shortest paths then we should also be able to recover a minimum spanning tree using the refined minimum spanning tree algorithm. You can yourself make suitable modifications in the algorithm according to these modifications.6 & 5.210 Basics of Graph Algorithms Figure 5.11. We can find shortest paths in such graphs in the presence of negative edge weights and we can also find longest paths in the presence of positive edge weights.7. 5. We can even solve the Hamiltonian Path problem in this very restricted class . See how by increasing memory or space requirements we can reduce the time complexity of an algorithm.11. Food for thought: Given a parent table and shortest distances as shown in Fig. 5.11.4: We modify the graph as we move forward in the minimum spanning tree algorithm.11. 5.7 determining the actual shortest paths in a directed graph is an interesting problem.

Shortest distances with respect to vertex a are indicated in the bottom left graph.11. .5: The original graph is shown in the top diagram.Shortest Path Algorithms 211 Figure 5. Weight of the minimum spanning tree is indicated in the bottom right graph. Please note that the shortest paths and minimum spanning tree are not easily available here.

6: We need to add a parent table array in order to find shortest paths in addition to shortest distances.11.7: How to recover shortest paths with the help of a parent table? .212 Basics of Graph Algorithms Figure 5. Figure 5.11.

it will also decide its time complexity.Shortest Path Algorithms 213 of graphs.8: A directed acyclic graph D shown in the top diagram. In the bottom diagram we arrange its vertices such that all edges move from left to right. So coming back to the important question: which vertex (and on what basis) . The crucial question is which will be the next vertex to go in the bucket and on what basis.9. Figure 5. The Bucket Algorithm (described earlier) can easily be modified to create an algorithm which can find shortest paths from any given vertex in a very efficient manner. This observation is explained in detail in Chapter 8 while discussing directed acyclic graphs. If we need to find shortest paths from a source vertex a then we should put that vertex in the bucket first. The intuition of this algorithm comes from the observation that we can always arrange the vertices of a DAG such that all edges in the DAG move from left to right as shown in the figure below.11.11. 5. The numbers inside each vertex is the start time and finish times obtained during a depth first search of the directed graph. Please see the concept map in Fig. In both these algorithms we made certain comparisons to select the next entrant into the bucket. Recall how we select the next candidate vertex which enters into the bucket in case of Dijkstra like algorithm or in case of Prim’s like algorithm. An answer to this question will not only determine the character of this algorithm .

5. We need no comparisons or extra steps to make this decision.9: A concept map depicting which vertex should next enter the bucket in different algorithms.11. .214 Basics of Graph Algorithms Figure 5. It will be an interesting challenge to derive the time complexity of this elegant shortest path algorithm.10. It is the next left vertex in the new arrangement of the vertices of the graph. should enter the bucket after putting the start vertex in the bucket? The interesting observation is that after arranging vertices of a DAG (such that all edges move from left to right) the next vertex to enter the bucket has already been decided.11. The corresponding algorithm is described below. Its working is shown in Fig.

Do you think Dijkstra’s like algorithm will also find correct shortest paths in a DAG? And at what cost? 2.Shortest Path Algorithms 215 Algorithm 27: Find shortest distance of every vertex from a given vertex a in a directed acyclic graph D input : A directed acyclic and weighted Graph D. 1 2 3 4 5 6 Arrange vertices of the DAG such that all edges move from left to right Put given vertex a in Bucket B while there is an edge going out of the Bucket B do Select the next right vertex j. 11. Put vertex j in B for every edge coming out of B from vertex j do assign min{w(a. Can we still apply this algorithm without any modification to find shortest paths? How about if we need to find shortest paths from a vertex other than the source vertex in a DAG? 3. a vertex a output: Modified graph D in which the weighted edges coming out of vertex a provides shortest distances from this vertex.10.11. w(a.11. k) Food for Thought 1. k). Consider the directed acyclic graph shown in Fig. How about if there are negative weight edges in the graph? Would this algorithm still provide correct results? 4. k)} to w(a. j) + w(j. . The graph is already drawn such that all edges are going from left to right. We need to find shortest paths from vertex a in this graph. How about if we need to find longest paths instead of shortest paths from a given vertex in this or any other directed acyclic graph? Which vertex will next end up in the bucket? What changes are needed in this algorithm to find longest paths? Please see figure 5. If we had applied Dijkstra’s algorithm to this graph then it would have selected vertex c (instead of vertex b) as the edge joining vertex a with vertex c having the minimum weight. Assume that we have a DAG with multiple source vertices. The next vertex which goes in the bucket is vertex b in spite of the fact that the edge joining vertex a with vertex b is the heaviest edge in this graph.

216 Basics of Graph Algorithms 5. The vertices of the DAG have been arranged such that all edges move from left to right.10: We need to find shortest distances from vertex a in this graph. Can we use a similar algorithm to find a Hamiltonian Path in a directed acyclic graph provided it exists? Figure 5. . Weight of every edge not indicated in the diagram is equal to 1. Now the next vertex to move in the bucket will always be the next left vertex.11.

Weight of every edge not indicated in the diagram is equal to 1.11. .Shortest Path Algorithms 217 Figure 5.11: We need to find longest distances from vertex a. Now the next vertex to move in the bucket will always be the next left vertex. The vertices of the DAG have been arranged such that all edges move from left to right.

i-edge shortest distance of every vertex k from vertex a denoted by Disti (k) output: (i + 1)-edge Shortest distance of vertex k from vertex a. Algorithm 28: (Bellman-Ford Building Block-A): Find (i + 1)-edge shortest distance of every vertex k from a given vertex a in a weighted directed graph G. Then we need to revert back to algorithms discussed earlier. Remember we have already reduced the time complexity of our modified shortest path algorithm from p3 to p2 . Now is the time to get rid of the bucket as it is hindering our way to handle negative edge weights.11. If a graph is cyclic then we cannot use the simplicity and elegance of this algorithm. input : A weighted directed Graph G. We assume that all distances are to be measured and minimized with respect to a start vertex a. We shall follow this terminology throughout this section. k) + Disti (j)} A 2-edge shortest path algorithm Let us make this building block consistent with our earlier policy of modifying the graph as we move forward in the algorithm. it is denoted by Disti+1 (k) 1 2 3 for each (terminating) vertex k do for each (intermediate) vertex j do Disti+1 (k) = min{Disti (k). a vertex a. w(j. It has been copied below. That was Dijkstra’s like algorithm which can handle positive edge weights in a graph with cycles but does not provide us with correct results if there are negative edge weights in the graph. The terminating vertex is k and the intermediate vertex is indicated by vertex j.3 Single Source Shortest Path Algorithms for directed graphs with negative edge weights We have discussed directed acyclic graphs and the ease with which we can find shortest paths in such graphs in the last section. This will make it friendlier to use as a building block.218 Basics of Graph Algorithms 5. In this building block we just convert the one edge distances into 2-edge shortest distances as given in the following algorithm. k) represents the initial one edge path of vertex k from . Let us recall the (i + 1)-edge shortest path algorithm described earlier in this chapter. Here w(a.

input : A weighted directed Graph G. w(a.1) w(a. k). w(a. k) = min{w(a.Shortest Path Algorithms 219 vertex a. The 2-edge shortest distance w(a. j) + w(j. a vertex a. j) + w(j. k)} to w(a. k) . k). 1-edge shortest distance of every vertex k from vertex a provided by the row corresponding to vertex a in the adjacency matrix G output: Modified graph G in which row corresponding to vertex a provides 2-edge shortest distance of every vertex k from vertex a 1 2 3 for each (terminating) vertex k of graph G do for each (intermediate) vertex j of graph G do Find a 2-edge path from vertex a to k passing through j. k) from vertex a to vertex k will then be given by the following recursive equation: (5. and assign min{w(a. k)} Algorithm 29: (Bellman-Ford Building Block-B): Find 2-edge shortest distance of every vertex k from a given vertex a in a weighted directed graph G.

12).12: Execution of the 2-edge shortest path algorithm from start vertex a. The 2-edge shortest distances are indicated in the modified graph shown in the bottom right corner.k)} .w(a. Here a = 1.k). w(a. 5.220 0 1 ∞ 2 ∞ 3 ∞ 4 ∞ 5 ∞ 6 Basics of Graph Algorithms ∞ 7 0-edge apart 1-edge apart 1 2 1 3 4 5 1 6 7 1 2-edge apart 1 2 3 4 5 6 7 1 1 7 2 2 ∞ 1 1 3 1 1 0 2 0 2 2 7 1 6 4 3 6 3 2 4 3 5 4 Input Graph 4 5 ∞ 1 2-edge Shortest paths? Figure 5. Edge weights not shown are equal to 1. Please see the accompanying figure (Fig. We vary j and k for the entire vertex range and find the 2-edge shortest distances of every vertex from vertex a.j)+w(j.11.11.k)=min{w(a.

The end result remains the same.11. Please note that it is immaterial whether you execute the j-loop first and the k-loop later or vice versa. Here intermediate vertex j is fixed at 4 and the destination vertex k is varied first.Shortest Path Algorithms 221 Figure 5. 5. This switching of two loops is an exciting idea in the development of shortest path algorithms . .13: We execute the k-loop first and then the j-loop in the 2-edge shortest path algorithm.it also becomes a powerful tool for the algorithm designer (Fig.13).11.

5. While vertex (13) in the top row of the recursion tree corresponds to the two edge shortest distance from vertex 1 to vertex 3 in the original graph.14. As you can understand this 2-edge shortest path algorithm becomes a building block for so many shortest path algorithms. k).11. The number of edges in the recursion tree corresponds exactly to the number of steps performed by the two edge shortest distance algorithm. A vertex (13) in the bottom row of the recursion tree corresponds to one edge distance of vertex 3 from vertex 1 in original graph. w(a. Please note that the number of edges in the recursion tree is exactly equal to the number of steps performed by the algorithm.222 Basics of Graph Algorithms Figure 5.14: Execution of the 2-edge shortest path algorithm is illustrated. k) = min{w(a.11. The Recursion Tree We illustrate the 2-edge shortest path algorithm by one additional tool known as the recursion tree (although strictly speaking it is a DAG) as shown in Fig. k)} is also shown in the bottom diagram. j) + w(j. . A recursion tree corresponding to the equation w(a.

w(a. This outer loop executes Algorithm 29 p − 1 number of times and minimizes the value of w(a.15. j) + w(j. k)}. Algorithm 30: (Bellman-Ford1): Find shortest distance of every vertex k from a given vertex a in G input : Adjacency matrix of weighted Graph G and a start vertex a output: Modified adjacency matrix G in which row corresponding to vertex a provides shortest distance of every vertex k from vertex a in G 1 2 3 4 for i=1 to p-1 do for each (intermediate) vertex j of graph G do for each (terminating) vertex k of graph G do Find a two edge path from vertex a to vertex k in G passing through an intermediate vertex j. k)} to w(a. w(a. k) according the following equation: w(a. k) := min{w(a. The other is the intermediate vertex j loop represented by orange color in the puzzle.Shortest Path Algorithms Bellman-Ford like Shortest Path Algorithm 223 In order to find a shortest distance in a weighted graph with negative edge weights we just have to run the 2-edge shortest path building block sufficient number of times as shown in the following algorithm. See without this blue box the remaining two boxes (brown and orange) represents the 2-edge shortest path algorithm known as Algorithm 29. k). k) This algorithm resembles the so called Bellman-Ford algorithm with a time complexity of p3 as we have three nested for loops. The outer most loop in the Bellman-Ford Algorithm (known as Algorithm 30) is represented by blue color in Fig. One is the terminating vertex k loop represented by brown color in the colored puzzle. 5. Please note that the nesting of f or loops in these algorithm can be nicely . The Colored Puzzle The colored puzzle highlights the fact that Algorithm 29 can be used to become a building block for Bellmam-Ford Algorithm. Here we introduce another tool to study an algorithm. Fig. Algorithm 29 has two f or loops. j) + w(j. and assign min{w(a. k).11.

It is interesting to note that each such algorithm appears in pairs . The outer most loop in the algorithm is represented by blue color. The positioning of the j and k loops in the Bellman-Ford Algorithm can be interchanged without affecting the outcome of this algorithm as shown in Fig.11.15: A colored puzzle depicting the positioning of the different for loops in the shortest distance finding algorithm. . however. cannot be changed without adversely affecting the performance of the algorithm. 5. The position of the blue loop.thanks to the colored puzzle which provokes a learner to note this interesting property.224 Basics of Graph Algorithms captured by nested colored boxes in the colored puzzle.11.15.16 along with the corresponding colored puzzles. 5.11. It is possible to run the intermediate vertex j-loop (orange color) first and then the destination vertex k-loop (brown color) or vice versa. for i =1 to p-1 for each vertex k for each vertex j for i =1 to p-1 for each vertex j for each vertex k Figure 5. The recursion tree corresponding to the Bellman-Ford Algorithm is shown in Fig. Other shortest path algorithms can also be represented by this puzzle with the addition of the start vertex a loop shown by green color in coming figures.

The complexity of algorithm is equal to the number of edges in this recursion tree .11.16: The recursion tree corresponding to Bellman-Ford like algorithm. It is also possible to rephrase this algorithm so that we do not need any .which is equal to p3 .Shortest Path Algorithms 225 Figure 5. How about a sparse graph and if we represent it using an adjacency list representation? It may be a good idea to see if the time complexity can be reduced to at least O(pq) which will be less than O(p3 ) for a sufficiently sparse graph. Time Complexity of Bellman-Ford shortest path algorithm We have witnessed that the time complexity of Bellman-Ford1 shortest path algorithm is O(p3 ) with an adjacency matrix data structure. The colored puzzle corresponding to this algorithm is again shown in the bottom diagram.

k) k a Dist(k) = 6 Dist(j) = 4 j w(j.17: How and when the (shortest) distance Dist(k).k) = 3 k j w(j.k) = 1 k New Dist(k) = 6 New Dist(k) = 5 Figure 5. 1 2 3 for i=1 to p-1 do for every directed edge (j. . with respect to vertex a changes when we consider the directed edge (j. k)} Dist(k) = min{ Dist(k).11. The rephrased algorithm is shown below.226 Basics of Graph Algorithms extra intelligence (in its implementation) to make it a O(pq) algorithm. output: Distance array Dist(k) storing minimum distances of every vertex k from vertex a. k). Its working is explained in the figure below. Line 3 is the basic building block of this algorithm. k) in graph D do Dist(k) = min{Dist(k). Dist(j) + w(j. In this version of Bellman-Ford the time complexity will always be pq with an adjacency list data structure and without an extra intelligence in its implementation. Dist(j) + w(j. Algorithm 31: (Bellman-Ford2): Find shortest distance of every vertex k from a given vertex a in G input : Directed and weighted graph D. k)} a Dist(k) Dist(j) j a Dist(k) = 6 Dist(j) = 4 w(j. Distance Dist(a) of vertex a from itself is zero and from any other vertex it is infinite.

Shortest Path Algorithms

227

5.11.4

All Pair Shortest Path Algorithms

We shall describe now three all pair shortest path algorithms. They will all use a single building block which is the 2-edge shortest path algorithm described earlier. The slow all pair shortest path algorithm has a O(p4 ) complexity, faster all pair shortest path algorithm has a O(p3 log2 p) complexity while Floyd-Warshall all pair shortest path algorithm has O(p3 ) time complexity under worst case conditions. The Slow All Pair shortest Path Algorithm The Bellman-Ford like algorithm finds shortest distances from a fixed start vertex in a graph. If we run this algorithm for every vertex in the graph then we end up with the so called slow all pair shortest path algorithm with a time complexity of O(p4 ). This algorithm is depicted pictorially by the colored puzzle shown in the left diagram of the figure below. Algorithm 32: Slow All Pair: Find shortest distance of every vertex from every vertex in G input : Adjacency matrix of weighted Graph G output: Shortest distance of every vertex from every vertex in G
1 2 3 4 5

for each (start) vertex a of graph G do for i=1 to p-1 do for each (intermediate) vertex j of graph G do for each (terminating) vertex k of graph G do Find a two edge path from vertex a to vertex k in G passing through an intermediate vertex j, and assign min{w(a, k), w(a, j) + w(j, k)} to w(a, k)

The Faster All Pair shortest Path Algorithm It is interesting to note that if we switch the blue p-loop with the green a-loop in the colored puzzle as shown in the diagram below then we end up with another all pair shortest path algorithm, known as faster all pair shortest path algorithm. It will be interesting to derive its time complexity and compare it with that of slow all pair shortest path algorithm. The working of this algorithm is shown in the diagram below.

228

Basics of Graph Algorithms

for each vertex a for i =1 to p-1 for each vertex j for each vertex k

for i =1 to p-1 for each vertex a for each vertex j for each vertex k

w(a,k) = min{w(a,k), w(a, j)+w(j, k)}
Figure 5.11.18: The start vertex loop is represented by green color. The intermediate vertex j loop is represented by orange color and the terminating vertex k loop is represented by brown color. There is another p-loop - it is represented by blue color. The diagram shows the effect of switching between the blue and the green loops. The left diagram depicts the slow all pair shortest path algorithm while the right diagram represents the faster all pair shortest path algorithm.

Algorithm 33: Faster All Pair: Find shortest distance of every vertex from every vertex in G input : Adjacency matrix of weighted Graph G output: Shortest distance of every vertex from every vertex in G
1 2 3 4 5

for i=1 to p-1 do for each (start) vertex a of graph G do for each (intermediate) vertex j of graph G do for each (terminating) vertex k of graph G do Find a two edge path from vertex a to vertex k in G passing through an intermediate vertex j, and assign min{w(a, k), w(a, j) + w(j, k)} to w(a, k)

Shortest Path Algorithms

229

Figure 5.11.19: Working of the faster all pair shortest path algorithm is shown. Please note that from two edge all pair shortest distances we jump to 4-edge and then to 8-edge shortest distances.

230

Basics of Graph Algorithms

Figure 5.11.20: The recursion tree and the colored puzzle corresponding to faster all pair shortest path algorithm. Here it is also obvious that from 2-edge shortest distances we directly jump to 4-edge shortest distances.

Shortest Path Algorithms Food for thought

231

1. Is the time complexity of this algorithm any better than that of slow all pair shortest path algorithm? If yes then why? 2. You may have noticed that in this algorithm (FAster All Pair) the number of edges in shortest paths jumps in the powers of 2 as shown in Figure 5.11.20. What does that mean? The outer most loop (the blue colored loop in the colored puzzle) should run from 1 to p or from 1 to log2 p? How will it affect the time complexity of this so called faster all pair shortest path algorithm. 3. In the slow all pair shortest path algorithm the blue loop should run from 1 to p and not from 1 to log2 p? Why? All Pair (Floyd-Warshall) Shortest Path Algorithm Consider the faster all pair shortest path algorithm. We know it can be represented by the colored puzzle as shown in Fig. 5.11.21. The recursive equation used as a building block is also indicated in this figure. There are essentially four for loops in this algorithm represented by four nested rectangles in the multi colored puzzle. If the intermediate vertex j loop (also known as the orange loop in the colored puzzle) becomes the outermost loop while the source vertex a loop (green) and destination vertex k loop (brown) are inner loops (in any order) then (surprisingly) we end up with one of the most efficient all pair shortest path algorithms as shown in Fig. 5.11.22. Please note that now we have only three loops and there is in fact no need to have the fourth loop - the so called blue loop in the colored puzzle. The time complexity will now be O(p3 ) as there are only three loops. Its time complexity is as good (or as bad) as that of Bellman-Ford algorithm (which is a single source shortest path algorithm) for non sparse graphs. The basic building block of this algorithm is the same recursive equation that we used in other algorithms. See Fig. 5.11.22 & 5.11.23.

232

Basics of Graph Algorithms

w(a,k) = min{w(a,k), w(a, j)+w(j, k)}
for i =1 to log2p for each vertex a for each vertex j for each vertex k for i =1 to log2p for each vertex a for each vertex k for each vertex j

Figure 5.11.21: The colored puzzles corresponding to faster all pair shortest path algorithm.

Algorithm 34: (Floyd-Warshall): Find shortest distance of every vertex from every vertex in G input : Adjacency matrix of weighted Graph G output: Shortest distance of every vertex from every vertex in G
1 2 3 4

for each (intermediate) vertex j of graph G do for each (start) vertex a of graph G do for each (terminating) vertex k of graph G do Find a two edge path from vertex a to vertex k in G passing through an intermediate vertex j, and assign min{w(a, k), w(a, j) + w(j, k)} to w(a, k)

Shortest Path Algorithms
for i =1 to for each vertex j p-1 for each vertex k for each vertex a for each vertex j for each vertex a for each vertex k

233

Figure 5.11.22: We show the possibility of having the intermediate vertex j loop become the outer most loop. There is no fourth loop in this diagram and surprisingly there is no need for it. Once the j-loop becomes the outer most loop then it does not matter if the order of the green loop and that of the brown loop is interchanged. Food for Thought: What is the trick or intuition behind this algorithm? Is this not surprising that without a fourth loop we can design an all pair shortest path algorithm? A spectrum of single source and all pair shortest path algorithms We show a spectrum of shortest path algorithms in Fig. 5.11.24. We also show connections or links between different algorithms. Note that the Bucket algorithm is the ancestor of most of these algorithms. This panorama of shortest path algorithms is also depicted by the colored puzzle shown in Fig. 5.11.25. Please appreciate the fact that for the entire galaxy of such algorithms we use a single building block - the 2-edge shortest path algorithm.

5.11.5

Johnson’s all Pair Shortest Path Algorithm

We have studied in the last section that all pair shortest path algorithm (Floyd-Warshall) has a worst case time complexity of O(p3 ). This is the best performance seen so far for an all pair shortest path algorithm. In this subsection we shall study another all pair shortest path algorithm which works faster than O(p3 ) for graphs which are sufficiently sparse. Required Prior Knowledge: First we shall talk about the prior knowledge required to understand this

234

Basics of Graph Algorithms

Figure 5.11.23: The recursion tree corresponding to Floyd-Warshall all pair shortest path algorithm. The number of edges in the recursion tree is equal to the number of steps performed by the said algorithm - and this is equal to O(p3 ).

Shortest Path Algorithms

235

Figure 5.11.24: We show a concept map of various single source and all pair shortest path algorithms.

236

Basics of Graph Algorithms

w(a,k) = min{w(a,k), w(a, j)+w(j, k)}

2-edge shortest path algorithm

Bellman-Ford (single source)

Slow All Pair

Faster All Pair

Floyd-Warshall (all pair)

Figure 5.11.25: We show the galaxy of single pair and all pair shortest path algorithms. Each algorithm can be represented by a different color arrangement in the rectangular puzzle. It is interesting to note that each shortest path algorithm has at least one dual with the same performance and output.

Figure 5.11.26: It will be interesting to see if any of these colored arrangements represents one of the already discussed shortest path algorithms.

If all edge weights are positive then we can apply Dijkstra’s algorithm p times to find all pair shortest paths in O(p2 logp + pq). 2. The improved time complexity is O(qlogp).it innovatively combines two shortest path algorithms (Dijkstra + Bellman-Ford) and creates an all pair shortest path algorithm such that the overall time complexity becomes better than that of best known shortest path algorithm . Once all edge weights are made positive we can use . If that becomes clear then it is almost trivial to appreciate the innovation behind this algorithm. So we need to do something more? Something very innovative? Johnsons’Algorithm = Bellman-Ford + Innovation + Dijkstra Johnson’s algorithm first uses Bellman-Ford to check if there are any negative weight cycles. We also know that its time complexity is O(p2 ) if we use an adjacency matrix as a data structure. If we convert this O(pq) Bellman-Ford algorithm into an all pair shortest path algorithm then its time complexity would become O(p2 q) which is worse than O(p3 ) for Floyd-Warshall.Floyd-Warshall (p3 ) under certain conditions. We know that Dijkstra’s shortest path algorithm finds correct shortest paths from a single vertex provided all edge weights are positive. The time complexity of this algorithm is O(pq) provided we use an adjacency list as a data structure to represent the input graph. Its time complexity can be improved with an adjacency list data structure provided we use a minimum heap to locate the next vertex which goes in the bucket. Again this is an improvement over O(p3 ) provided we have a sufficiently sparse graph. This can further be improved to O(plogp + q) if we use a Fibonacci heap to implement the minimum priority queue. 1. If there are no negative cycles then this algorithm somehow uses results of Bellman-Ford algorithm to convert negative edge weights into positive without disturbing the relative path lengths. This time complexity is better than O(p3 ) (Floyd-Warshall) in sufficiently spare graphs. We know that Bellman-Ford algorithm can find shortest paths from a single source vertex in time O(p3 ) even if there are negative edge weights in a directed graph. But if there are negative edge weights then we shall get incorrect results. This requires time proportional to pq.Shortest Path Algorithms 237 algorithm. It is interesting to note that this is not entirely a new algorithm .

output: Distance array Dist(k) storing minimum distances of every vertex k from vertex a. All distances are measured with respect to vertex a. If it does then it means we have negative weight cycles in the directed graph reachable from vertex a as shown in the figure below.11. . Food for Thought 1. Consider the Bellman-Ford algorithm described earlier and reproduced here. 5. We initialize Dist(a) = 0 and set Dist(x) of every other vertex x from vertex a equal to infinity. Does that mean there are no negative weight cycles and should we stop without further iterating? See Fig.11. If there are negative weight cycles then complications arise as already discussed. We then run this loop one more time to check if any distance changes. Under such conditions this algorithm should at least inform us that in the given graph there are negative weight cycles. If it does not then there are no negative cycles reachable from vertex a in the directed graph. Please see Fig.28. Algorithm 35: Find shortest distances of every vertex k from vertex a in D input : Directed and weighted graph D. Checking Negative Weight Cycles in a directed graph: The above algorithm finds shortest paths correctly in case there are no negative weight cycles in the directed graph. Dist(j) + w(j. 5. Distance Dist(a) of vertex a from itself is zero. 1 2 3 for i = 1 to p − 1 do for every directed edge (j. k)}. k) in the graph D do Dist(k) = min{Dist(k).238 Basics of Graph Algorithms the improved Dijkstra’s algorithm to find all pair shortest paths in time O(p2 logp + pq).27 . It was already discussed that its worst case time complexity will be O(pq). What modification is needed in this algorithm for this extra intelligence? The required modification is simple and elegant? We run the outer most loop p − 1 times and store distance of each vertex x from the start vertex a. How about if the distance of no vertex changes after an initial iteration i when i < p − 1.

29.11. 5. 2. Does that mean we have to apply Bellman-Ford algorithm at each vertex to find if there are any negative weight cycles in the graph? But that will be very costly? Figure 5.28: After finding 3-edge shortest paths in this graph from vertex a there will be no change in distance calculations.11. Does that mean that we should stop here and declare that there are no negative weight cycles in this graph? Applying Bellman-Ford Algorithm once to determine Negative Weight Cycles The problem is how can we apply Bellman-Ford algorithm just once from a (special) vertex and check if there are any negative weight cycles in the di- .27: If we find i-edge distances in this graph from vertex a then we observe that some distances will change when i goes from 4 to 5. This does not confirm that there is a negative weight cycle.11. How about if we apply Bellman-Ford at a vertex which is not reachable to a negative cycle. But when a distance of a vertex from vertex a changes when i goes from 5 to 6 then that is a confirmation that there is indeed a negative weight cycle in this graph.Shortest Path Algorithms 239 Figure 5. See Fig.

30.11. 2. Whether there are any negative weight cycles in the graph.240 Basics of Graph Algorithms Figure 5. This information will further be used to convert negative edge weights into positive edge weights. Thus we cannot add an arbitrary positive number in each edge weight as it will disturb the relative path lengths in the new graph as shown in the following figure.31.11. If there are no negative weight cycles then what are the shortest distances of each vertex from the newly added vertex x. Converting negative weight edges into positive weight edges Please note that we need to simultaneously fulfill the following two objectives: . If the graph does not contain any negative weight cycles but contains negative weight edges then we should somehow try to make the negative weight edges positive and then apply Dijsksta’s algorithm from each vertex as already planned. rected graph? The answer to this problem is given in elegant transformation shown in Fig. If.11. 5. 5. we apply this algorithm from vertex a then it is possible to verify that indeed there are negative weight cycles in this graph. The application of Bellman-Ford algorithm in the transformed graph just once (from the newly added vertex x) provides us the following information: 1.29: If we apply Bellman-Ford algorithm in this graph to find shortest distances from vertex e then we should never be able to confirm that there will be negative weight cycles in this graph. however. If there are any negative weight cycle then the algorithm should not move forward and should terminate. How about adding a big positive number in each edge weight such that every (edge) weight becomes positive? See what complication would arise if we do so as shown in Fig.

Figure 5. All edges coming from vertex x have a zero weight.30: The transformation in this diagram allows us to apply Bellman-Ford algorithm just once and verify if there are any negative weight cycles in the graph.11.Shortest Path Algorithms 241 Figure 5.31: We can always convert negative edge weights by adding a positive number in each weight so that we can apply a more efficient algorithm to find shortest paths in a graph with positive edge weights.11. . But adding a constant in each edge weight disturbs the relative weights of different paths and leads to a wrong answer. We apply Bellman-Ford algorithm to this transformed graph and find shortest distances from vertex x.

k)old is negative then w(j. 5. In the top diagram we try to initially label k such that . The other problem (converting all negative edge weights into positive) remains to be solved. 5. Consider a directed edge (j. k)new = w(j.32. In fact the new path length from a vertex a to a vertex b will be the old path length plus Label(a) − Label(b). Before finding a systematic scheme of providing labels to each vertex let us try to work out a simple example.11.33. If we associate an arbitrary number with each vertex of the graph and change edge weights according to the formula described then we claim that relative path lengths between any two vertices will be the same in the new graph as compared to the old graph. See Fig. All relative distances between two vertices should remain the same in the modified graph with positive edge weights. 2. k)old + Label(j) − Label(k) becomes positive. namely the relative distances between two vertices remain the same in the new graph as compared to the old graph. So now the problem is reduced to finding an appropriate number Label(j) to be associated with a vertex j and another number Label(k) to be associated with vertex k (assuming that there is an edge from j to k with a weight w(j. It requires that vertices should not be labeled arbitrariliy but with some intelligence as described below. and Label(k) are all negative.This would require that wjk (old) + Label(j)−Label(k) is greater than or equal to zero. Now the edge weight w(j. It means that wjk (old)+ Label(j) is not less than Label)k). First we shall devise a simple scheme to alter edge weights such that relative distances between any two vertices do not change in the given graph. Then we shall modify this scheme so as to convert all negative edge weights into positive edge weights thus fulfilling both the above objectives. k)) such that if w(j.242 Basics of Graph Algorithms 1. 5.11. In other words Label(k) should be at least equal or more negative than wjk (old)+Label(j) assuming that wjk (old). Label(j). Putting arbitrary labels solves only one problem.11. How about associating any arbitrary number Label(j) with a vertex j and another arbitrary number Label(k) with vertex k. k) is changed according to a formula where wnew = wold +Label(j)−Label(k). What numbers are desirable and should be associated with two adjacent vertices j and k are indicated in Fig.30. Here vertex x is a vertex added to the given directed graph just like the one shown in Fig. k) with a weight equal to w. All negative edge weights should be converted into positive weights.

11.32: How relative distances remain the same if we associate an arbitrary number with a vertex and then add a number in the edge weight of each edge according to a fixed formula. .Shortest Path Algorithms 243 b 11 11-6=5 -6 a -2 -2+3-4+7=4 c 7 -4 -3 d 3 e f 0 11 b -6 11-6=5 -4 a -3 -2 -2+3-4+7=4 c b 0 11+0+3 11-6+0+4=9 -6-3+4 7 -4 d -4 -1 3 e -2 f -5 a -2+0+1 c 7-5+4 -2+3-4+7+0+4=8 d -1 3-1+2 e -2 -4-2+5 f -5 Figure 5.

11.33 where the label of a vertex is updated and now loook at Fig. It should become quite evident now that if Label(j) is the shortest distance of vertex j from vertex x and Label(k) is the shortest distance of vertex k from vertex x then the required inequality would be satisfied and the edge weight for edge (j. If you are still undecided then read the following paragraphs and look at the coming figure.11. 5. 5. See Fig. Check if there are any negative weight cycles in the graph D using Bellman-Ford algorithm applied to the graph after inserting a source vertex x as shown in Fig. Please see once again Fig.34 where the shortest distance of vertex j is updated. Please note that the new label of k (equal to -80) will make sure that all edges incident to vertex k have positive edge weights.11.35. k) becomes positive. k) also becomes positive. 5. If there are no negative weight cycles in graph D then move to step 2 otherwise terminate.11. k) will eventually become positive if it was initially negative. Summary So now we are in a position to describe Johnson’s algorithm in the following meaningful manner as applied on a directed graph D.244 Basics of Graph Algorithms weight of edge (j1 . Now coming back to the general question: What systematic scheme should be applied which guarantee the allocation of desirable numbers to vertices of a graph such that all edge weights become positive while relative distances between any two vertices remain the same? After looking at the last figure you must have some idea of what is going on or what should be done. Let us recall the basic building block of Bellman-Ford algorithm and how the shortest distance of a vertex k from a source vertex x is updated. In the bottom diagram we intend to re-label k such that weight of edge (j3 . See Fig. Now draw another copy of graph D with all edge . In step 1 we have already found shortest distance Dist(k) of each vertex k from newly inserted vertex x. 2. 5.34 for a demonstration of this updating. 5.35. That shortest distance Dist(k) become Label(k) of vertex k.11. The interesting thing is that we need not spend extra time in finding these labels as they have already been found while checking if the given graph has negative weight cycles. 1.

and j3 are already labelled while we need to label k so as to make edge weights positive.11. k) becomes positive. . In the bottom diagram we intend to re-label k such that weight of edge (j3 .Shortest Path Algorithms 245 Figure 5.33: We assume that vertices j1 . k) also becomes positive. Please note that the new label of k (equal to -80) will make sure that all edges incident to vertex k have positive edge weights. j2 . In the top diagram we try to initially label k such that weight of edge (j1 .

34: While executing Bellman-Ford algorithm we use a basic building block as shown above. Dist(j) + w(j. x Transform 0 b 8 -2 -4 -4 1 -2 2 b 0 -2 a -2+2 8-2 1+0 2+4 -4+4 a c -4 c d 0 3 e 0 d 0 3+0 e 0 Figure 5.246 Basics of Graph Algorithms Dist(k) = min{ Dist(k). . The bottom diagrams show how and when the shortest distance of vertex k from a source vertex x is updated.k) k x Dist(k) = -70 j w = -40 k New Dist(k) = -60 j Dist(j) = -20 w = -40 k New Dist(k) = -70 Figure 5.11. k)} x Dist(k) Dist(j) j x Dist(k) = -10 Dist(j) = -20 w(j.11.35: How to determine which labels to associate with each vertex.

The teacher should help.” . At times it is almost impossible to solve a given problem while it is easy to solve a related problem (the shortest path problem is solvable while the longest path problem is unsolvable). If the teacher helps too much. To learn means to cause your mind to function in a different way: new memories are created and/or new connections are forged. In this chapter we have demonstrated how a teacher can help students discover a number of graph algorithms with some initial help.11. 3. nothing is left to the students. According to Hale [5]: “There are different kinds of learning. but I refer here to the intellectual kind. According to Polya [10]: If the student is left alone with his problem without any help or with insufficient help. It is extremely useful to find why a certain technique works under certain conditions and why it fails in others (greedy methods provide optimal solutions in finding the shortest path but fail to find the longest path). but not too much and not too little. All negative edges become positive as shown in Fig. so that the student shall have a reasonable share of the work. he may make no progress at all.12 Discussion The most important task of a teacher should enable the students to discover and acquire experience of independent work. We have also shown that making comparisons between various techniques and solutions provides a deep insight which itself is very useful in solving otherwise difficult problems [8]. starting with something seemingly simplistic yet capable of being transformed into a number of powerful algorithms with minor modifications. The theory of NP-Completeness connects all problems that are NP-Complete: it is also possible to find a useful relationship among solvable problems and this is what we have attempted to do in this chapter.35. Now apply Dijksta’s shortest path algorithm at every vertex in the modified graph (with positive edge weights) to find all pair shortest distances and paths. We have shown that by asking thought provoking questions it becomes possible for the teacher to guide the students while solving complex problems. 5.Discussion 247 weights modified. 5. Similarly it is difficult to solve a problem in its original form while it is easier to solve it while placing certain restrictions (the graph isomorphism problem is solvable for trees but is difficult to solve in general).

Maud. J. S. . M. Baase. We also wish to thank S. Ikram. Mahkari. S. K. H. T.248 Basics of Graph Algorithms These relationships provide the algorithm designer a perspective that proves invaluable when solving new problems and analyzing old one’s. Alvi. Khan for providing motivation as well as inspiration for this project. Acknowledgement We are thankful to the Department of Computer Science. Skiena. Mian. Jadoon for their help and encouragement. Fahd. Lahore University of Management Sciences for providing support for this research. A. We wish to specially thank R.

7 6.9 Introduction Definitions & Prior Knowledge Konig’s Theorem.4 6.3 6.6 6.5 6.1 6. Menger’s Theorem & Hall’s (Marriage) Theorem Revisited Network Flows The Matching Problem in Bipartite Graphs The Max-Flow Min-Cost Problem Network flows with lower & upper bounds on flow and the Circulation Problem .2 6.Chapter 6 Network Flows.8 6. Menger’s Theorem & Hall’s (Marriage) Theorem Menger’s Theorem Konig’s Theorem. Connectivity and Matching Problems 6.

will be to integrate concepts so that the enabled learner is able to appreciate the bigger picture where one relationship implies another and one theorem can be used to prove the other. We have witnessed the subset sum problem in which we have to select integers (out of a set of integers) such that the sum of the selected integers is equal to a given constant. Connectivity and Matching Problems 6. We shall also discuss Konig’s Theorem which relates the size of the vertex cover to the size of maximum matching in a bipartite graph. and last but not the least the Circulation problem.1 Introduction We shall first address the problem of vertex or edge connectivity in general graphs. The Marriage (Hall’s) theorem provides necessary and sufficient conditions for a bipartite graph to have a perfect matching.250 Network Flows.2 Definitions & Prior Knowledge Set Cover: Given a set of subsets S of a Universal Set U . That is why we first provide a unified picture and then go deeper in order to analyze each area in detail. We shall also discuss the matching problem in bipartite graphs. What is the smallest subset of vertices of the graph that covers all edges in the graph? Independent (Vertex) Set: What is the largest subset S of vertices of a graph such that no pair of vertices in S has an edge in between? Is there a . 6. We shall also be discussing the network flow problem. we shall be designing algorithms to solve a number of related problems. Our intention (and desire) in this chapter. In addition to making formal proofs for a number of theorems. what is the smallest subset T of S such that the union of all these sub sets in T covers all elements of U (it means that every element in U is contained in any of the subsets in T and the set T has the smallest possible size). which if removed will disconnect a special node from another special node in a graph. Specifically we shall be describing Menger’s Theorem which relates maximum number of vertex-disjoint (or edge-disjoint) paths with minimum number of vertices (or edges). We use a single building block in this entire chapter for designing almost every algorithm. the maximum flow at minimum cost problem. Vertex Cover: The Universal set U is the set of all edges in a graph.

What is the smallest subset of edges. (A line graph is a tree where the degree of each node is not more than 2) Maximum Matching: It is a matching in a graph with maximum possible size? How bad can a maximal matching become as compared to maximum matching? (It means how small the size of the maximal matching can become as compared to the size of the maximum matching) Edge Cover: The Universal set is the set of all vertices in a graph. Every edge connecting a leaf vertex u and a non leaf vertex v will always be part of the edge cover. i. Each leaf vertex is at a distance of 3 from the root vertex.Definitions & Prior Knowledge connection between the vertex cover and the independent set problem? 251 v u Figure 6. Any independent set will include leaf vertex u. These edges are also known as independent edges. while any vertex cover will include non leaf vertex v. where the maximum degree of each vertex is one. Maximal Matching: This is a matching in which more edges cannot be added to increase the size of this matching. Find a simple algorithm to find a maximal matching in a line graph. that is. as compared to the number of vertices .e. The size of a matching is equal to the size of set A. The root vertex is shown at the top of the diagram.2. no two edges in the set A share a common vertex.1: A perfect binary tree is shown. Matching (Independent Edge Set): A sub-graph of G. Note that every non leaf vertex is a cut vertex and every edge is a cut edge or a bridge in a tree graph. which covers all vertices? How small (or big) can the size of the edge cover become. Perhaps a more meaningful definition will be a set A of non-adjacent edges. no edge has a common end point.

It has a single vertex with degree equal to 2.2 for the perfect binary . Let us call it the root vertex. 6. We know that in a tree every edge is a cut edge or bridge while every (non leaf) vertex is a cut vertex. A Perfect Binary Tree: We know that a tree is a connected graph where each edge is a bridge edge.1) has the following features: 1.2. the term complete matching is used for perfect matching.2.1 has 8 leaf vertices. That is.1. It has exactly 2h − 1 non leaf vertices including the root vertex.2.1. 6. We shall observe later in this chapter that there is indeed a relationship between graph connectivity and number of paths in a graph (Menger’s Theorem). Every perfect matching is both maximum and hence maximal. 3.2. 6. In a perfect binary tree graph G. The perfect binary tree shown in Fig. It has exactly 2h leaf vertices. 6. It has exactly 2h+1 − 1 vertices where h is the path length (in terms of number of edges) between the root vertex and any leaf vertex in the perfect binary tree.2. Connectivity and Matching Problems A Perfect Matching is a matching which covers all vertices of the graph. The perfect binary tree shown in Fig. a leaf vertex u will always part of an independent set (why?). While a non leaf vertex v which is adjacent to a leaf vertex will always be part of a vertex cover (why?). We are given a perfect binary tree G (see Fig. A perfect binary tree (shown in Fig. We also know that there is always a unique path between every two vertices in a tree. Using this logic all leaf vertices will be part of the independent set and will not be part of the vertex cover.252 in a graph? Network Flows. 6. 6. There are 7 non leaf vertices in tree shown in Fig.2. The root vertex is shown as the top most vertex in Fig. A binary tree is a tree where the degree of each vertex is less than or equal to 3. Problem Set 6.2. 2.1. every vertex of the graph is incident to exactly one edge of the matching. In some literature. 4. As described before each leaf vertex in a perfect binary tree has a path length equal to h from the root vertex.1) and we intend to solve a number of problems related to connectivity and matching. All non leaf vertices other than the root vertex has a degree exactly equal to 3.1 has h equal to 3 and it has 15 vertices. Here we define a perfect binary tree. 6. Independent set and vertex cover vertices are shown in Fig.

2: The top left diagram shows the vertices belonging to the vertex cover highlighted by double circled vertices.Definitions & Prior Knowledge 253 u Vertex Cover u Independent Set u u v u u u v u u Maximum Matching u Edge Cover u u v u u u v u Figure 6. You may have also noticed some relationship between the size of the vertex cover and the size of the independent set? . The bottom left diagram highlights. The bottom right diagram shows the edges belonging to the edge cover. You may have noticed that the size of the maximum matching is equal to the size of the vertex cover in this graph. the maximum matching edges in the given tree. The maximum matching in this graph is not a perfect matching. as all vertices here are not matched as shown by black circled vertices. in brown bold.2. The rest of the vertices belong to the independent set in this diagram as shown in the top right corner. these edges are shown in bold in green.

Check Fig. How can we find an independent (vertex) set in a perfect binary tree? Problem 6. meaning a maximum matching in a perfect binary tree? Problem 6.2.2. It may thus be possible to find a maximum matching given a vertex cover in a tree graph.2. Please note that a perfect binary tree graph is a fairly restricted structure and our existing prior knowledge of graph theory and algorithms is sufficient to solve these problems.5. How can we efficiently find an edge cover in a perfect binary tree? Problem 6. 6.1. Based upon the above observations it is possible to design efficient algorithms to solve the following problems in a tree graph.2.6. The bottom left diagram of Fig.2. what conditions are necessary and which are sufficient? Problem 6. How can we efficiently find a vertex cover in a perfect binary tree? Problem 6. A perfect matching may exist for the tree graph if the two halves have the same size. It also indicates matched and unmatched vertices and edges. We shall observe later that in a bipartite graph (such as a tree) the size of a maximum matching is equal to the size of vertex cover (Konig’s Theorem).1.4.1.2 to see if a near perfect matching exists in a perfect binary tree.1) will always be part of the edge cover.1. 6. If the two halves are not of the same size.3.2.1.7. however a near perfect matching may exist.1.2.1. A near perfect matching requires that all vertices belonging to the smaller half are matched to vertices in the larger half in a bipartite graph. In a perfect binary tree graph an edge connecting a leaf vertex u with a non leaf vertex v (see Fig. 6. Does a perfect matching exist in a perfect binary tree? Discuss briefly. Connectivity and Matching Problems tree graph of Fig. 6. 6. Problem 6.1. For a perfect matching to exist in a tree graph.2 shows a maximum matching in the perfect binary tree graph of Fig. How can we find an independent (edge) set. .1. 6.2. We know that a tree is a (restricted) bipartite graph consisting of two halves.1.254 Network Flows. Problem 6. then a perfect matching cannot exist. Edges belonging to the edge cover are shown in bold in the bottom right diagram of Fig.

. A concept map showing a number of relevant concepts and a number of theorems which relate different concepts.Definitions & Prior Knowledge 255 Concept Map 6.1.

while a tree graph T has λ(T ) = 1. where both s and t belonged to G.2. Thus λ(s. Vertex Connectivity κ(G) of a Graph G is the minimum number of vertices which if removed will disconnect the graph G (see Disconnected Graph). the resulting graph will be a disconnected graph. will disconnect it so that all paths between vertices s and t are destroyed. The edge connectivity for a completely connected graph (having p vertices) is equal to p−1. t) is the minimum number of edges which if removed will destroy all paths between vertices s and t in G. So our new requirement is that when G is disconnected (by removing certain edges) then vertices s and t should belong to different connected components of G. Connectivity and Matching Problems Figure 6. vertices s and t of G would now belong to different connected components. t) as the minimum number vertices which if removed from G. .1.256 Network Flows. A disconnected graph G has λ(G) = 0. What are the necessary and sufficient conditions for a near perfect matching to exist in a tree graph? Edge Connectivity λ(G) of a Graph G is the minimum number of edges which if removed will disconnect the graph G (see Disconnected Graph).8.3: There are several paths between vertices s and t in this graph. By removing certain edges (or vertices) it is possible to remove all these paths between the two vertices. Problem 6. Sometimes we not only want the graph G to be disconnected but also want to make sure that a special vertex s is separated from another special vertex t. Similarly we define κ(s.

Problem 6. Out of all paths that you have listed in Problem 6.2.2.2. while vertex-disjoint paths do not share any vertex except the terminal vertices. Draw more than one edge-disjoint paths between the two vertices such that at least one path should be of length six.2.1. it may be broken up into more connected components or we may require that G would be disconnected such that two special vertices s and t of G should lie in separate connected components.2. Out of all paths that you have listed in Problem 6. 6. a}.4.5. Problem 6. What is the maximum number of such paths? Problem 6. We intend to explore different paths between these two vertices and see how we can destroy these paths in G.2. . We are given a graph G with two special vertices s and t as shown in Fig. Problem 6.2 short list the ones that are vertex-disjoint. then the graph G is disconnected but vertex s is still accessible to vertex t.1 and Problem 6. What is the maximum number of such paths? Problem 6.2.2.2.2.3. Draw all 4-edge paths between vertex s and t.4.2. Draw a 4-edge path between the two vertices such that not more than one additional edge-disjoint path is possible between the two vertices.2.2. Problem Set 6.2. 6. Problem 6.and 6.6.2. There is only one bridge edge in this graph. if we remove this edge {s. Vertex-disjoint paths are also edge-disjoint but it may not be true the other way round. We have to remove a lot more than one vertex in order to disconnect G so as to destroy all paths between vertex s and t. There is no cut vertex in this graph. See Fig.Definitions & Prior Knowledge 257 Edge-disjoint paths do not share any edge.3.4.1 and Problem 6. See Fig. Disconnected Graph: A graph G may simply be disconnected into two or more connected components. Draw all 5.2 shortlist the ones that are edge-disjoint.edge paths between the two vertices. that is. 6.

Connectivity and Matching Problems Four Edge and Six Edge Paths a g a g b s c e h t s b e h t f i s c f i d Ca l wn ra e d is? n b th ike j b c d d j e g h i i f j i f j t t t t t t t Figure 6.4: 4-edge and 6-edge paths from vertex s to vertex t in graph G.258 Network Flows. s a g Can be drawn like this? b b s c f i e h t g e i h i c d f j i f j d j t t t t t t t Figure 6.5: Some edge disjoint paths.2.2. these paths are not vertex disjoint? .

The right diagram of Fig. A minimal subset of these edges shown by a cut is sufficient to do the job. You can remove all edges belonging to the . 6. In fact it is not necessary to remove all the edges in the single 6-edge path as shown in Fig. Problem 6. Draw a 6-edge path between the two vertices such that no additional edge-disjoint (or vertex-disjoint) path is possible between the two vertices. then all the paths between vertex s and t are destroyed. but only because we have selected one wrong path initially. that other possible paths have disappeared.7 shows a 6-edge path (shown in bold) between vertices s and t in the same graph.2.7. Problem 6.6: A 6-edge path and a couple of 4-edge disjoint paths.2. Why? What are its ramifications in designing an algorithm to find all possible edge-disjoint paths in a graph? Problem 6.10.8. Can you find another subset of three edges which when removed will disconnect s and t? What is λ(s. We know that there existed more than one path between the two vertices. Note that if we remove all edges in this 6-edge path.2. The left diagram of Fig. 6. Do these three edges corresponds to a minimum number of edges which if removed will destroy all paths between vertices s and t.9. it will no longer be possible to find an extra edge-disjoint path between the two vertices.2. Once these paths are selected. List down these edges.2.2. 6.2.7 shows two edge disjoint paths of length 4 (shown in bold) between the vertices s and t.7 in order to destroy connectivity between s and t.2. t) for this graph? Problem 6.Definitions & Prior Knowledge 259 Figure 6.

Now we need to find three edge-disjoint paths between the same vertices. 6. paths and see for yourself that the connectivity between s and t is destroyed. What is a minimal subset of these edges which will do the same job of destroying the connectivity between s and t? One such subset is shown in the right diagram of Fig. The right diagram shows two 4 edge disjoint paths (shown in bold) between vertices s and t. the diagram also shows. The left diagram of Fig. two vertex-disjoint paths between s and t. in bold. which if removed will destroy the connectivity between vertices s and t.8 shows a minimum sized subset of edges.2.7: The left diagram shows a 6-edge path (shown in bold) between vertices s and t in the graph. The right diagram shows a minimum sized subset of vertices which if removed will disconnect s and t from each other.2. if removed will disconnect s from t. e and f .2. Connectivity and Matching Problems g a A Cut g b e h b e h s c f i t s c f i t d j d j Figure 6. This subset consists of two vertices.2. Does this mean that there will be three edge-disjoint paths in this graph? Find these three paths.2. Be careful as this is the same cut shown in the left diagram of Fig. 6. The blue (green) cut cuts those edges which.11.7. . This subset consists of four edges unlike three in the last part? Problem 6.260 a A Cut Network Flows.7 where there existed a single path of six edges (with no room for additional paths). Is this a coincidence that the (maximum) number of vertex-disjoint paths is exactly equal to the (minimum) number of vertices which if removed will disconnect s and t? The left diagram shows an edge cut of size three. 6.

Konig’s Theorem tells us that the cardinality of a maximum matching in a bipartite graph is equal to the (minimum) size of the vertex cover (see Concept Map 6.8 is an illustration of Menger’s Theorem. Menger’s Theorem & Hall’s (Marriage) Theorem Before discussing connectivity.8: The left diagram shows a minimum sized subset of edges (shown in bold) which.2. matching. if removed. 6. Menger’s Theorem & Hall’s (Marriage) Theorem a g h a 261 An Edge Cut g e h b e f b s c d i j t s c d f A Vertex Cut t i j Figure 6.3 Konig’s Theorem. The vertex cover of this bipartite graph is also indicated in this figure. It states that the maximum number of vertex-disjoint paths between vertices s and t are equal to minimum number of vertices which. if removed from the graph will disconnect s and t (see Concept Map 6. The left diagram of Fig. 6. The right diagram of Fig. and the above theorems in detail in subsequent sections.3. 6. will destroy the connectivity between vertices s and t. it also shows a maximum matching with the maximum matching edges shown in bold.1).1 shows the same bipartite graph with two dummy nodes designated as s and t.Konig’s Theorem. we shall first provide a panoramic picture of how the different concepts are interrelated.1 shows a bipartite graph with partite A and partite B.1. The vertex cover in the bipartite graph (shown .1).3. The right diagram shows a minimum sized subset of vertices (shown by double circles) which if removed will disconnect s and t from each other.2. All vertices in the first partition of the bipartite graph are connected to s while all vertices in the second partition are connected to the dummy vertex t as shown in the right diagram of Fig. How one theorem implies another and can be derived from each other. The right diagram of Fig. 6. 6.3.

Under such conditions the size of the maximum matching and that of the vertex cover will be equal to the size of remaining set A which is equal to three.3. N ({a1 }. Note that all vertices in the set A are not matched to all vertices in the set B although the size of A is equal to size of B in the given bipartite graph. {a4 }) = {b2 . while in the remaining two the said condition is violated.262 Network Flows.1 would be equal to the size of partition A (and B). {a3 }. Similarly each matched edge (in the left diagram) will correspond to a vertex-disjoint path between vertex s and t in the right diagram (why?). b3 .1) implies that the (maximum) number of vertex-disjoint paths (and κ(s. It is interesting to note that if we remove vertex a3 from A and vertex b3 from B then the condition |N (S)| ≥ |S| is true for every S ⊆ A. {a3 }. A relationship between (minimum) vertex cover and maximum matching in a bipartite graph (Konig’s Theorem) is transformed into a relationship between κ(s. b4 } N ({a1 }. {a2 }) = {b1 .3. For a better understanding of the Marriage Theorem a few neighborhood subsets are indicated below for the bipartite graph of Fig. You may have realized yourself that this necessary and sufficient condition (applicable to the bipartite graph shown on the left of Fig. b3 . Had a perfect matching existed then the size of the vertex cover would have been equal to the size of A (or B). b2 } N ({a2 }. Connectivity and Matching Problems in the left diagram) becomes the minimum sized subset of vertices which. 6.1. 6. b2 . if removed will destroy all paths between vertex s and t in the right diagram (why?).3. {a2 }. {a3 }) = {b2 } N ({a2 }. b4 } Out of the five neighborhood subsets N (S). {a3 }) = {b1 . t) and (maximum) number of vertex-disjoint paths between vertex s and t in the right diagram of Fig. {a4 }) = {b1 . only three satisfy the condition that |N (S)| ≥ |S|. t)) in the graph shown on the right side of Fig. 6.3. In other words a perfect matching does not exist in this bipartite graph. What should be the size of the neighborhood N (S) for every subset S of A which will guarantee a perfect matching (or a vertex cover equal to the size of A) is the subject of Marriage (Hall’s) Theorem.1 (Menger’s Theorem). 6. It states that a bipartite graph (of equal halves) has a perfect matching if and only if |N (S)| ≥ |S| for every S ⊆ A. b2 } N ({a1 }. {a2 }. .

Konig’s Theorem. The top right diagram shows the same bipartite graph with two dummy vertices s and t. .3. and vertices belonging to the vertex cover highlighted by orange circles. Menger’s Theorem & Hall’s (Marriage) Theorem 263 Every Matching Edge a1 a2 A a3 a4 b3 b4 b1 b2 B Corresponds A Path from s to t a1 a2 b1 b2 B b3 b4 b1 b2 B b3 b4 s A a3 a4 a1 a2 t a1 a2 A a3 a4 b1 b2 B b3 b4 Corresponds s A a3 a4 t Every Node of Vertex Cover Maximum Matching Minimum Vertex Cover Konig’s Theorem A Node of Vertex Cut Maximum Paths Transforms Minimum Vertex Cut Menger’s Theorem Figure 6.1: The top left diagram shows a bipartite graph with maximum matching edges shown in bold.

Then according to Menger’s Theorem.4. Given an undirected graph how can we find its connectivity: the minimum number of edges (belonging to Edge Cutset) or minimum number of vertices (belonging to Vertex Cutset) which if removed will disconnect the graph into two or more connected components? The graph theoretic aspects of the above problems and a proof for Menger’s Theorem will be presented after a better appreciation of the issues involved. t).264 Network Flows.1 Maximum Edge-Disjoint Paths in Directed Graphs Given a directed graph D and two specific vertices s and t. Menger’s Theorem equates the maximum number of edge disjoint (vertex disjoint) paths between vertices s and t to minimum number of edges (vertices) which if removed will disconnect s from t. (see Concept map 6. equivalent to finding a minimum sized edge set which if . how can we efficiently find the maximum number of edge-disjoint paths from vertex s to vertex t? Let U be the minimum sized subset of the edge set of D. Connectivity of a directed graph D also poses similar problems.4 Menger’s Theorem Connectivity of an un-directed graph is expressed in terms of minimum number of edges or minimum number of vertices which if removed will disconnect the graph G. the maximum number of edge-disjoint paths from vertex s to t will be equal to λ(s. Thus the problem of finding maximum number of paths from s to t is. 6. Given a directed graph D containing two special vertices s and t. We shall start this section with directed graphs and then generalize our results for undirected graphs. We shall also address the following algorithmic problems: 1. such that D − U does not contain any directed path from vertex s to t. in fact. how can we find the maximum edge-disjoint (vertex-disjoint) paths from s to t in a directed graph D? 2. Given a directed graph D how can we find the minimum number of edges (vertices) which if removed will destroy all paths from vertex s to t in D? 3.1). The size of U will obviously be represented by λ(s. Connectivity and Matching Problems 6. t).

Menger’s Theorem 265 Concept Map 6.2. . A concept map showing a number of concepts related to vertex connectivity and edge connectivity and some important relationships that we shall explore in this section.

If you are successful in finding a path then keep a record of this path. or (s → d → t).1. Pick the initial path more intelligently . But before finding another edge-disjoint path we should first remove the edges of P1 (why?) and then find another path P2 . It may be (s → a → c → t).4. We are in dire need of some innovation? There are potentially two ways to fix this problem: 1. we shall be able to find the two edge-disjoint paths. it has now become impossible to find another path in this graph. we shall land in a difficult situation as depicted in Fig. we can find a directed path P1 . and vertices s & t.4.and do something else with these edges.1. What is that something else? Before we actually try something new consider the following: . If path P1 = (s → a → d → t) (as shown by bold lines in Fig. we can move forward in Algorithm 36. Let us look at what are different possibilities for P1 . one being (s → a → c → t). How about devising the following common sense algorithm to find the maximum number of edge-disjoint paths from vertex s to t in a directed graph D? Algorithm 36: Find Maximum edge-disjoint paths from s to t in D.4. 1 2 Find a directed path P from vertex s to vertex t in D.1) then we are blocked. from vertex s to t in D. go to step 1 . 6. (s → a → d → t). 6.but this may not be an efficient solution. Let us apply this simple algorithm to solve the problem in the graph of Fig. input : Directed graph D. Thus an initial wrong choice will make things hard for us. Remove all edges in the path P and. otherwise exit the algorithm. Once we have selected an initial path P1 . Do not delete the edges of the chosen path . and the other (s → b → d → t or s → d → t). Using any path finding algorithm. why? 2. If we are lucky. 6. (s → b → d → t). Connectivity and Matching Problems removed from D will disconnect t from s. d) we can move from ‘a’ to ‘d’ but not from ‘d’ to ‘a’. output: (Maximum) edge-disjoint paths from s to t in D. It cannot be (s → d → a → c → t) or (s → b → d → a → c → t) because on directed edge (a.266 Network Flows. But if we are unlucky.

Is there a class of directed or undirected graphs where an initial wrong selection would not create any complication? How is the class of graphs (where an initial wrong selection really matters) structurally different from the other class (where an initial wrong selection does not matter)? 4. We are given a directed acyclic graph D with a source vertex (a vertex with zero in-degree) s and a sink vertex (a vertex with zero out-degree) t. Is it possible to convert one class of graphs into another? If this is possible then after the conversion we can use our (stupid) Algorithm 36 to find out maximum edge-disjoint paths in a graph? 5. 2. How about if the directed graph D is cyclic but the in-degree as well as the out-degree of every node are equal? You should consider both cases: when vertices s and t are not part of any cycle and when they are also part of some cycles.4. . Algorithm 36 fails to find the maximum number of edge-disjoint paths in a graph but it still finds maximal number of edge-disjoint paths in a graph.Menger’s Theorem a c a c 267 s b d t s b d t Figure 6. The in-degree is equal to the out-degree for every vertex i of D other than the vertices s and t. (where an initial wrong choice will make things hard) a general phenomenon in almost all graphs or is this a problem in a certain class of graphs? 3.4. See Fig. 6. 1.1: If we initially select a path as shown by bold red lines (left diagram) and remove it from the graph (right diagram) then it becomes impossible to find another path in this graph.2 before arriving at a conclusion. Is this complication. Show that if you apply Algorithm 36 to this graph then you will be able to find the maximum number of edgedisjoint paths without any complication.

y). We shall later show in this section that the number of pseudo edge-disjoint paths between two vertices in a graph is exactly equal to the number of edgedisjoint paths. The status of such an edge. y). . In edge-disjoint paths an edge (x. y) cannot be shared by more than one path. can be traversed by a new path from y to x and not from x to y. Finding maximum number of pseudo edge-disjoint paths is a problem which can be solved using the exciting process of discovery based learning. Please note that rule number 1 is common in pseudo edge-disjoint as well as edge-disjoint paths. not already in use by a path. which is traversed by two paths in opposite directions will be “not used by any path” or “unused”. It is rule number 2. which makes the two categories different. 2.4. in fact. Finding edge-disjoint paths in a general directed graph may be hard to find. This diagram is provided by Khawaja Fahd. in case of pseudo edge-disjoint paths we follow the following rules of the game: 1. A directed edge (x. The status of such an edge when traversed by a single path from x to y will be “used”. A directed edge (x. let us try to solve a simpler problem (finding pseudo edge-disjoint paths) and then use our newly found experience and confidence to solve the harder problem (finding edge-disjoint paths). given a set of pseudo edge-disjoint paths in a graph it is possible to find an equal sized corresponding set of edge disjoint paths in the graph.2: A directed graph D in which the in-degree of every vertex is equal to the out-degree of every vertex including vertices s and t. can be traversed (or used) by a path from x to y and not from y to x. expecting an innovation to achieve this objective may be unrealistic. already traversed by a path from x to y.268 Network Flows. Connectivity and Matching Problems a c a c s b d t s b d t Figure 6.

Reverse the direction of edge e and then go back to step 2.4. If you are successful in finding a path P then keep a record of the number of paths found so far otherwise exit with graph F as output. Algorithm 37: Find Maximum pseudo edge-disjoint paths from s to t in directed graph D.3. Initially no edge of D. Algorithm 37 is based on this innovative idea. In this process (Algorithm 37) we convert graph . for every edge e in path P do set Status(e) = used if initially it was unused otherwise set Status(e) = unused. is occupied by any path and thus the Status of every edge is unused. 1 2 3 4 Copy Graph D into F . output: Maximum pseudo edge-disjoint paths from s to t in D and a Status(e) for every directed edge e of D. Note that Algorithm 37 (unlike Algorithm 36) allows an edge to be used by two different edge disjoint paths moving in opposite directions. We first find maximum pseudo edge-disjoint paths from vertex s to vertex t in this graph. 6. and then go back to step number 2. It also outputs the Status(e) of every edge e of D. In step number 4 (instead of removing every edge in the path as was done in Algorithm 36) we reverse the direction of every edge in the path found in step number 1. As soon as an unused edge is occupied by a path its Status is changed from unused to used in step number 4. Example 6. Status(e) = unused for every directed edge e in D. Find a directed path P from vertex s to vertex t in F . Let us execute this simple four line algorithm on the graph of Fig. input : Directed graph D. and vertices s & t.4. Let us carefully look at this new algorithm which claims to find the maximum number of pseudo edge-disjoint paths from s to t in a directed graph D.1. We apply the above algorithms on the graph D shown in the same figure. In step number 2 we find a path from vertex s to t in D using any path finding algorithm.Menger’s Theorem 269 So the current problem is how to maximize the number of pseudo edgedisjoint paths in D? Instead of deleting all edges in a (recently found) path (as is done in step number 2 of Algorithm 36) we reverse the direction of each edge in the path.

Figure 6. Minimum Cut will consist of edges (x. output: Edges of D that belong to the Minimum Cut. y) of D Where x belongs to P and y belongs to V (D) − P . Connectivity and Matching Problems Algorithm 38: Find Minimum Cut in a graph D. input : Directed graph D.4. and vertices s & t.270 Network Flows.3: Original Graph D . 1 2 3 Convert D into F using Algorithm 37. Let P be the set of vertices (of F ) reachable from vertex s in F .

6. T ) will be a subset of E(D) and is known as a cut in the graph D (a cut signifies edges which if removed from a graph will destroy all paths from vertex s to . Then (S. The top left diagram shows the directed graph D with a selected path P1 from vertex s to vertex t shown in bold.4.6. We execute Algorithm 37 once again on a directed graph D shown in Fig. The most useful result of this algorithm is the Status of every edge e in D after its termination. Example 6.4. 6.4. Again Algorithm 37 provides the maximum number of edge-disjoint paths present in this graph although the paths P1 . P2 & P3 found by this algorithm are not edge-disjoint as shown in Fig. Using this set P we identify edges of D which belong to the minimum cut. they both belong to a (desirable) class? You can apply (the very naive) Algorithm 36 to this class of graphs and you will find maximum sized subset of edge-disjoint paths without any difficulty. The top right diagram shows the edges of another path P2 shown in bold. The original graph D is shown in the bottom right diagram with used edges shown in bold. 6.1) minus the unused edges as shown in the bottom of Fig.4. Before jumping to any conclusions and in order to gain more confidence let us solve another example.2.Menger’s Theorem 271 D into F .2) minus the unused edges shown in Fig. The paths P1 (green) and P2 (blue) that we have found during the execution of this algorithm are not edge-disjoint paths (see Fig. 6. 6. yet this algorithm tells us the maximum number of possible edge-disjoint paths in D (why and how?). 6.4 shows the original graph D with used edges only.5. Let us also consider the graph D (of Example 6. The bottom left diagram shows the edges of the path P3 shown in bold. this new graph has some very desirable properties which we shall elaborate later.4. Let S be the set of vertices of D containing s but not t and let the subset T = V (D) − S. The diagram on the bottom right of Fig.2 The Concept of a Minimum Cut in Directed Graphs Let us define the concept of a cut more formally. The Minimum Cut is found in Fig.4. T ) is a set of directed edges from a vertex in S to a vertex in T .4.4. After deleting unused edges (from D) we end up with this new graph. Graph F helps us identify the set P using Algorithm 38.6.4.4.4).4. 6. Both these graphs share something in common. 6. The set (S. 6. Let us again consider graph D (of Example 6.4.

.4: Graph D is the input to Algorithm 37. This Algorithm outputs maximum pseudo edge-disjoint paths and the Graph F . Connectivity and Matching Problems a c a Used c Used s b Graph D d t s b t Used d Initial Graph F a c Used a Used c Used n U s b d t s Used t Used U se d b d a U n c a c s Un Used t d s b d t d se U b Un Used Final Graph F Graph D without Unused Edges Figure 6.272 Network Flows.4.

The bottom left diagram shows the final graph F where the Min Cut is found and then it is applied to the original graph D on the bottom right diagram.5: The top left diagram shows the Original Graph D.4. .Menger’s Theorem 273 a c a c s b Graph D d t s b Graph F d t a c a c s b Graph F d t s b d t Graph D Figure 6. the top right diagram shows the final graph F with a couple of paths found and reversed.

6: The top left diagram shows a directed graph D with a selected path P1 from vertex s to vertex t shown in bold. The top right diagram shows the edges of another path P2 shown in bold. Connectivity and Matching Problems a g e f h a g e f h b b s c d a i j g e f h t s c d a i j g e f h t b b s c d i j t s c d i j t Figure 6. The bottom left diagram shows the edges of the path P3 shown in bold. . The original graph D is shown in the bottom right diagram with used edges shown in bold.274 Network Flows.4.

c). edge (a. . TA ) is equal to U (which is the minimum sized subset of edges which. 6. d) in the Cut C does not belong to the set (SC .1. While in Cut C. (SA . TC ) = {(s. We shall provide a constructive proof. t}. The theorem states that the maximum number of edge-disjoint paths from vertex s to t is equal to the minimum number of edges of D which if removed will destroy all paths from s to t in D. (a. d) is physically cut by Cut B as well as Cut C but it is only included in the Cut B and not in Cut C. Please note that it is the same directed graph for which we have found maximum number of edge-disjoint paths as shown in Fig.7. TB = {c. 6. d}. d) is not contributing to a path from s to t in D and is therefore not required to be removed. Looking at this problem from a different angle: In Cut B you have to remove all four edges present in this cut in order to destroy all paths from vertex s to t in D. TA = {c.Menger’s Theorem 275 vertex t). According to Menger’s Theorem the maximum number of edge-disjoint paths in this graph will equal λ(s. (SB .3 A Proof of Menger’s Theorem and Finding the Minimum Cut in Directed Graphs We are given a directed graph D and two special nodes s and t. (d. We intend to prove Menger’s Theorem in terms of edge connectivity of a directed graph. b}. Out of many possible cuts we show only three cuts in the directed graph shown in the top diagram of Fig.4. (SA . TC ) as shown in the bottom right diagram of Fig. TC = {a. The size of Cut B is larger then the size of Cut A and that of Cut C.7. (b. How can we find the minimum cut systematically? We shall discuss it in a proof of Menger’s Theorem.4. t}. (SC . a).7. (d.4. For Cut A : SA = {s. d)} For Cut C : SC = {s. b. which is 2. d).4. 6. if removed will destroy all paths from s to t in D). t)} It is interesting to note that the edge (a. t) = |(SA . This is because in Cut B vertex ‘a’ belongs to SB while vertex d belongs to TB as shown in the bottom left diagram of Fig. t}. d). Thus the edge (a. c. TA )| = 2.4. d. (a. Thus λ(s. vertex ‘a’ belongs to TC and vertex d belongs to SC . 6. By a little inspection in this graph it can easily be verified that Cut A is indeed one of the minimum cuts. t). TB ) = {(s. 6. c). TA ) = {(a. b. d}. a. While in Cut C you have to remove only two edges present in this cut and vertex s is disconnected from t. a. t)} For Cut B : SB = {s.

. The bottom right diagram shows the subset SC along with the Cut C edges shown in bold. The bottom left diagram shows the subset SB along with the Cut B edges shown in bold. Connectivity and Matching Problems Cut C a c a c Cut A s b Cut B d t Cut A s SA t b d a c a Cut C c s SB Cut B d t s b SC d t b Figure 6.4.7: We are given a directed graph D and we intend to find maximum number of edge-disjoint (directed) paths from vertex s to vertex t.276 Network Flows. The top diagram also shows three cuts.

The maximum number of pseudo edge-disjoint paths from vertex s to t is equal to the minimum number of edges of D which if removed will destroy all paths from s to t in D.4. 6. there can not be more than k pseudo edge-disjoint paths in D.4. TX ) where vertex s belongs to the subset SX while vertex t belongs to TX as defined earlier in this section. We shall make the following claims here: Claim 6. T )|min where the minimum is taken over all cuts (S. the minimum number of vertex-disjoint paths from s to t is equal to the cardinality of minimum sized vertex cut-set which will disconnect t from s in D. We then reverse the direction of each edge of the path just found.4.. Claim 6. We know that number of edge-disjoint paths from vertex s to t in a graph cannot exceed the size of a cut (SX .4).3.2: In step number 2 of Algorithm 37 we find a path from vertex s to t in D using any path finding algorithm.1.2. The maximum number of the two paths will be exactly equal (given one. we should be able to find the other).Menger’s Theorem 277 Using this proof one can also find the maximum number of edge-disjoint paths as well as the minimum cut.4. i. That means the maximum number of edgedisjoint paths from s to t in D will not exceed |(S. Proof for Claim 6. and go back to step number 2 to find a new path (see Fig.1 & 6.4. 6. Claim 6. As we have seen before there may be different cuts possible with different sizes in the same graph as shown in Fig.4.7. may consist of unused edges (not used by any path so far) or used edges (occupied by a previous path in the . T ) in D.e. This proof technique can easily be adapted for undirected graphs and can also be used to prove the vertex form of Menger’s Theorem which states that. Please note that each (pseudo) path found in step number 2. If we try to maximize the number of paths from s to t they will only be limited by the bottleneck in the graph which will be the minimum cut.4. Algorithm 37 finds k pseudo edge-disjoint paths from s to t in D before it terminates. The number of real (not pseudo) edge-disjoint paths from s to t in D can not exceed the number of (pseudo) edge-disjoint paths from s to t in D. the minimum cut. Menger’s Theorem tells us that the two quantities (maximum number of paths and size of the minimum cut) are exactly equal.

The Cut (P. 6. 6. what we need to show now is that there will actually be a cut whose size will be equal to the maximum number of paths from s to t in D. It is obvious that if there is a (pseudo) path left (in addition to the one’s that are already found) then Algorithm 37 will find it before its exits. We define subset Q to be equal to V (D) − P .8: The bottom left diagram of Fig. Let us assume that while executing this algorithm we have reached a stage where we are no longer able to find a new path and the Algorithm 37 terminates.6.278 Network Flows. a c 1 a 2 2 1 2 c 2 Pseudo Paths 1 s b P d t s b a c Edge-Disjoint Paths t Min Cut d s unused unused t b unused d Figure 6. We have seen that the maximum number of pseudo edge-disjoint paths cannot exceed the size of a minimum cut.5 is duplicated in the top left corner.4. We keep on doing this until we are no longer able to find a path from s to t in D and then the algorithm terminates. Connectivity and Matching Problems opposite direction) as shown in the top right corner of Fig.4. Although no (additional) path from vertex s to t exists anymore. The maximum number of pseudo edge-disjoint paths is shown in bold in the top right diagram. but it still may be possible to reach a number of vertices in the modified graph D from the vertex s. it shows the number of vertices (enclosed in the shaded area) which are reachable from s in D after the termination of Algorithm 37. Let P be a nonempty set (a subset of V (D)) containing vertex s and all other vertices which are reachable from s. Q) will consist of edges such that each edge which is part of this cut will be occupied by a unique .4. The bottom diagram shows the modified graph D with used edges shown in bold. It is important to note that the vertex t will belong to Q otherwise it will be in P and then we can find an extra path from vertex s to vertex t in D.

The out-degree of s will be equal to the in-degree of t in D. The above three properties guarantee that the modified graph D has as many (real) edge-disjoint paths from s to t as the number of pseudo edgedisjoint paths in the original graph D. y) which is part of the cut but not part of any path from s to t in D. It will be a directed acyclic graph. In fact it is possible to find the (real) edge-disjoint paths from the modified graph D (as shown in the bottom of Fig. Proof for Claim 6.4. This situation is depicted in Fig. The in-degree of every node other than the vertices s and t will be equal to its out-degree. Thus each edge of the Cut (P.5.5.6.3: We have seen that the Algorithm 37 not only provides maximum number of pseudo edge-disjoint paths (see Fig.4.) but also the status of every edge is provided by it.4. The set of vertices which are reachable from s (known as P ) are shown shaded in this diagram.4.Menger’s Theorem 279 path from vertex s to t in D. Under such conditions vertex y will also be reachable from vertex s which contradicts our initial assumption. Q) consists of edges all of which are already occupied by existing paths as shown in the bottom right corner of Fig. 6. Can the number of paths be smaller than the size of this cut? If the number of paths is smaller than the size of the cut then you can identify at least one edge (x. Q) is in fact the Minimum Cut equal to the maximum number of edge-disjoint paths from vertex s to t in D.5. 6. 6. When Algorithm 37 is unable to find an extra path it terminates as shown in the bottom left corner of Fig. If we remove all edges of D which have a Status = unused from D then the resulting graph D would have the following properties: 1. 6. The Cut (P. The Cut (P.4. The number of paths cannot be larger than the size of this cut as discussed before.4.4 & Fig.4) using (the very simple) Algorithm 36. 6. Q) will be part of exactly one pseudo path from vertex s to t in D. 3. 2.4. 6. 6.4.4 Finding Maximum Vertex-Disjoint Paths & Minimum Vertex Cut in Directed Graphs We have shown earlier that the maximum number of edge-disjoint paths (from vertex s to vertex t) in a directed graph is equal to the minimum .

There are basically two issues that we would like to tackle: 1. a s b Graph H c t d s b c t d Graph H after removing vertex a c s d Graph H after removing vertex a & b c t s b Graph H after removing vertex a & d t Figure 6. The vertex form of Menger’s Theorem equates maximum number of vertexdisjoint paths (from vertex s to vertex t) to minimum number of vertices which if removed will destroy all paths from vertex s to t in a directed graph. Connectivity and Matching Problems number of edges (which if removed will destroy all paths from vertex s to t).280 Network Flows.4. Please note that removal of certain vertices does not destroy all paths from vertex s to vertex t in this graph (see the top right and bottom left diagrams) while the removal of some other vertices does indeed destroy the s − t connectivity in this graph. 2. The Fig. A proof that maximum number of vertex-disjoint paths (from vertex s to vertex t) is equal to the minimum number of vertices which if removed will destroy all paths from vertex s to t in a directed graph. An efficient algorithm to find maximum number of vertex-disjoint paths from vertex s to t and a minimum cut in terms of vertex cut-set.9: Various vertices are removed to show that the graph does not necessarily become disconnected by their removal and only specific ones make the graph disconnected. It also shows how this graph would look like if we remove a number of vertices. 6.4.9 shown below shows a directed graph H in the top left diagram. .

10: Each vertex of the directed graph of shown in the left diagram is split up into two vertices.Menger’s Theorem 281 Interestingly both these problems can be resolved using our prior experience provided we make a couple of transformations on a given directed graph H (left diagram) and convert it into another directed graph D (right diagram) as shown in Fig. x) in H insert a directed edge from s to x1 in D.10. each directed edge (x. x2 ). Once the directed graph H is transformed into D it has now become possible to appreciate the following: . Thus for every edge in H. In addition to these edges we have edges of the form (x1 . Thus for the new graph D we have V (D) = 2V (H) − 2 and E(D) = E(H) + V (H) − 2. there will be a corresponding edge in D as shown in brown color in the right diagram of Fig.4. Algorithm 39: Transform directed graph H into directed graph D.4. 6. input : Directed graph H with special vertices s & t. For every edge (s. y1 ) in D.10 using Algorithm 39.4. Similarly for every edge (x. a c a1 a2 c1 c2 s b Graph H d t s b1 b2 Graph D t d1 d2 Figure 6. 6. we call these edges internal edges. t) in H insert a directed edge from x2 to t in D. Let us call these edges external edges. y) in H is transformed into a directed edge (x2 . Identify internal and external edges of the graph D. as shown in the right diagram. External edges are shown in brown color while internal edges are shown in bold orange color. We split each vertex x (excluding vertex s and t) of directed graph H into x1 and x2 in D. output: Directed graph D with internal & external edges marked 1 2 Convert H into D by splitting all the vertices (except s and t) as described. We insert an extra edge between x1 and x2 in D for every vertex x in H except for s and t as shown here.

4.5 Menger’s Theorem for Undirected Graphs We shall consider Menger’s Theorem in terms of edge connectivity (Menger’s Theorem in terms of vertex connectivity can also be derived from in a similar .11. The corresponding path in D passes through external as well as internal edges as shown in Fig. The above observations provide us enough insight to find maximum edgedisjoint paths in the directed graph D which will correspond to maximum vertex-disjoint paths in H. A Minimum (edge) Cut in graph D will correspond to a Minimum (vertex) Cut in graph H provided the Minimum (edge) Cut in D passes through internal edges only.11 (top left corner). may pass through some of the external edges as shown in the middle right diagram in Fig.10 is shown in Fig. We can construct (an almost) similar constructive proof that the maximum number of vertex-disjoint paths from s to t in H is exactly equal to the minimum number of vertices which if removed from H will destroy all paths from s to t in H. There is just one problem to be resolved.282 Network Flows. We can use Algorithm 39 to convert graph H into graph D and then use Algorithm 37 without any change to find the maximum edge-disjoint paths and the Minimum Cut. Connectivity and Matching Problems 1.4. 6. 6. 3. There is a path from vertex s to t in D corresponding to any path from s to t in H. In order to do so we make a one line modification in our earlier approach in terms of Algorithm 40 as described below. 6. 6. 6.11 (thereby forcing the entire cut to pass through internal edges alone).4. Every path going from vertex s to t in H has to pass through a number of k intermediate nodes (nodes other than s and t). the working of this approach on the directed graph D of Fig.4. the Minimum Cut as found by the original Algorithm 37.11. It is important to appreciate that the minimum cut should (be forced to) pass through internal edges only. and thus any number of edge-disjoint paths in D will correspond to the same number of vertexdisjoint paths in the directed graph H. We first add the external edges of graph D in graph F and then find the appropriate cut as shown in the bottom diagrams of Fig.4. If it passes through (some of the) external edges then the minimum cut will not correspond to a minimum vertex cut. the corresponding path in D will pass through k internal edges. 2.4. 6.

4. Now when we reverse selected edges of this path then it is no longer possible to find a path from vertex s to t as shown in the middle right diagram.11: The top left diagram shows a directed graph with a selected c a path P1 from vertex s to vertex t shown in pink. The top right diagram Graph H shows selected edges of the path P1 reversed. This is the final Graph F .Menger’s Theorem 283 a1 a2 c1 c2 a1 a2 c1 c2 s t s t b1 b2 d1 d2 b1 b2 d1 d2 Graph D and initial Graph F Graph F a1 a2 c1 c2 a1 a2 c1 c2 s t s t b1 b2 d1 d2 b1 b2 d1 d2 Final Graph F: Cut passing through an external edge a1 a2 c1 c2 a1 a2 c1 c2 s t s t b1 b2 d1 d2 b1 b2 d1 d2 Graph F plus External Edges of D: Cut passing through internal edges only Graph D Figure 6. to which we add the external edges of Graph D and find the Minimum Cut. The reason for the step in the bottom right diagram is to ensure the Minimum Cut passes through internal edges only. . We again find a path (blue) as shown in the middle left diagram. The minimum (vertex) cut is indicated by a bold line.

We can in fact use Algorithm 37 without any changes to find maximum number of edge-disjoint paths from s to t as shown in Fig.4. Let P be the set of vertices (of F ) reachable from vertex s in F . The undirected graph G of Fig. 6.4. input : Directed graph D with special vertices s & t and internal & external edges marked output: Internal Edges of D that belong to the Minimum Cut. y} of G is thus transformed into two directed edges (x. Minimum Cut will consist of edges (x. Connectivity and Matching Problems Algorithm 40: Find a Minimum Cut in a directed graph D passing through internal edges only. It is reasonable to assume that any directed path from s to t in D will consume either (x. 6. y) and (y. It states that the maximum number of edge-disjoint paths between two specific nodes in an un-directed graph G is equal to minimum number of edges needed to destroy all paths between the two specific vertices.13. 6. fashion).13.4. The path between s and t in G (Fig. Each undirected edge {x.12 along with two vertices s and t and a path from s to t shown by bold lines. For every un-directed path between vertex s and t in G there is a corresponding directed path from s to t (or from t to s) in D.4. An un-directed graph G is shown in Fig. 6. 6.13. x) but not both. The maximum number of directed edge-disjoint paths from s to t in D will be equal to the maximum number of un-directed edge-disjoint paths between s and t in G. Menger’s Theorem states that the maximum .12 is transformed into a directed graph D as shown in Fig.13.4. Add all the external edges of Graph D to F . Maximum Vertex-disjoint Paths & Minimum Vertex Cut in un-directed graphs Now considering Menger’s Theorem in terms of vertex connectivity of a graph where s and t are already given. x) in D. 6. 6. The minimum cut corresponding to minimum number of edges required for destroying all paths between s and t can also be found using similar techniques as shown in the bottom left diagram of Fig.4.284 Network Flows.4. y) of D Where x belongs to P and y belongs to V (D) − P .12) is also shown in bold in the top left diagram of Fig. 1 2 3 4 Convert D into F using Algorithm 37. y) or (y.

4.4. Vertex Connectivity κ(G) of a Graph G is the minimum number of vertices which if removed will . Thus each un-directed edge {x. |V (D)| = 2|V (G)| − 2 and |E(D)| = 2|E(G)|. An un-directed graph G is shown at the top of Fig.12: This is Graph G (left diagram). We then apply the standard algorithm for finding edgedisjoint paths from vertex s to vertex t as shown in the top-right diagram of Fig.14 and reverse the edges.Menger’s Theorem 285 a c a c s b d t s b d t Figure 6. y) and (y.4. 6.4.14 is transformed into a directed graph D as shown in the middle diagram of Fig. 6. for every vertex x there is a corresponding two vertices {x1 .15 the external edges of Graph D of Fig. 6.4.14.14.4.6 Edge Connectivity and Vertex Connectivity for Undirected Graphs Edge Connectivity λ(G) of an undirected graph G is the minimum number of edges which if removed will disconnect graph G. Then all the vertices except for s and t are split into two as shown in the bottom diagram of Fig. Thus the total number of vertices and edges are almost doubled. The un-directed graph G of Fig. y} of G is transformed into two directed edges (x. 6.4.14 are added where they do not exist in graph F (not shown in the Figure) and the minimum cut is found and applied to Graph G. x) in D. if we initially select a path as shown by bold red lines and then remove its edges from G then it becomes impossible to find another path in the remaining un-directed graph (right diagram).4. 6. 6. 6. This way in the bottom-left diagram of Fig. The only difference from the previous procedure is that we apply Algorithm 39 to find the Minimum Cut for vertexdisjoint paths. number of vertex-disjoint paths between two specific nodes in an un-directed graph G is equal to minimum number of vertices needed to destroy all paths between the two specific vertices. 6.4.14 along with two vertices s and t. x2 }.

286 Network Flows. Connectivity and Matching Problems a c a c s b Graph D d t s b d t Initial Graph F a c a c Min Cut s b d t s b d t a c a c s b d t s b Graph D d t Final Graph F Figure 6.13: The top left diagram shows an un-directed graph with a selected path P1 from vertex s to vertex t shown in bold. . The minimum cut is indicated by a shaded region in the bottom diagrams.4. We again find a path as shown in the middle left diagram. The top right diagram shows the edges of the path P1 reversed. Now when we reverse the edges of this path then it is no longer possible to find a path from vertex s to t as shown in the middle right diagram.

Then each vertex other than vertex s and t is split and the graph G is transformed into a directed graph D shown in the bottom diagram. Each edge of this graph (is split to indicate directed edges going both ways) is replaced by two directed edges as shown in the middle diagram. .14: The top diagram is an undirected graph G.4.Menger’s Theorem a c 287 s b d t Un-directed Graph H a1 a a2 Directed Graph D before vertex splitting s1 c1 c c2 s s2 t1 b1 b b2 d1 d d2 t t2 a1 a2 c1 c2 s1 s s2 t1 t t2 Directed Graph D after vertex splitting b1 b2 d1 d2 Figure 6.

288 Network Flows. Hence we apply Algorithm 39 and find the minimum cut for the vertex-disjoint paths in the bottom-left diagram and that cut is applied to original Graph G of Fig.4.14. The middle diagram shows another path P2 (blue) found and its edges are reversed in the middle-right diagram. where path P1 (pink) is found and the edges of the path are reversed as shown in the topright diagram. .15: The top-left diagram shows the initial graph F . Connectivity and Matching Problems a1 a2 c1 c2 a1 a2 c1 c2 s1 s2 t1 t2 s1 s2 t1 t2 b1 b2 d1 d2 b1 b2 d1 d2 Graph F a1 a2 c1 c2 a1 a2 c1 c2 s1 s2 t1 t2 s1 s2 t1 t2 b1 b2 d1 d2 b1 b2 d1 d2 Final Graph F a1 a2 c1 c2 a s1 s2 t1 t2 c s b Graph H t d b1 b2 d1 d2 Final Graph F with External Edges Figure 6. as can be seen no more paths can be found. 6.4.

Is it possible to arbitrarily select s and t in the graphs below. find λ(s.4. For each such pair (we call it an s−t pair) we find the maximum number of edge-disjoint paths from vertex s to vertex t in the undirected graph G. Edge Connectivity a p b f g i m c e h j n d Vertex Connectivity k Figure 6. (We actually do it by finding maximum number of edge-disjoint paths from vertex s to t in a graph G). The top graph shows its edge connectivity.Menger’s Theorem 289 disconnect graph G. Our prior knowledge and expertise tells us that given an undirected graph G and two already selected special vertices s and t we can efficiently find λ(s. there could be O(p2 ) distinct pairs of vertices. The bottom diagram shows the vertex connectivity of the bottom graph.4. What is then the way out? Perhaps we should consider all possible (s − t) pairs in the graph G? As the graph G consists of p vertices. The diagrams in Fig. t) = λ(G)? The problem may not be that simple as it is evident from the diagram below (Fig. 6. We know that the maximum number of edge disjoint . λ(s. t) is equal to the edge connectivity of graph G. that is. The problem is to efficiently find both edge connectivity and the vertex connectivity of an undirected graph G.17).4.16 show two undirected graphs. 6. There is a possibility that a wrong choice for vertex s and vertex t may give you an incorrect result as shown in the following diagrams. t) which is equal to minimum number of edges which if removed will disconnect vertex s from vertex t in G.16: The top diagram shows the edge connectivity of the top graph and the bottom one shows the vertex connectivity of the bottom graph. t) and claim that λ(s.

290 Network Flows.19. paths is equal to λ(s. The next logical step should be to find ways to reduce this complexity. t). Edge Connectivity of an undirected Graph We have already hinted before that in order to efficiently compute edge connectivity of an undirected graph. You may have realized that there is no need to consider all O(p2 ) distinct pairs of vertices. . we may fix s arbitrarily but then compute λ(s. 6.17: Various Min-cuts in the same graph. Connectivity and Matching Problems 8 8 8 t t 2 1 7 6 2 7 6 2 7 6 3 4 5 1 3 4 5 s 1 4 3 5 s The Min-Cut from s to t is 2 s t The Min-Cut from s to t is 2 The Min-Cut from s to t is 1 Figure 6.4. This intelligent observation will certainly cut down the time complexity of our earlier technique as described in the algorithm given below. It will be a useful experience to derive the overall time complexity of this simple algorithm. Out of all results we select the minimum. We thus have to apply the maximum edge-disjoint paths finding algorithm for all possible p − 1 pairs where s is fixed while t takes on every possible value as shown in Fig. t) for every different value of t in the undirected graph as shown in the figure below. only O(p) pairs will be sufficient for calculating edge connectivity as shown in Fig. Out of all the O(p2 ) values for λ(s. We select a different vertex t in each of the p − 1 graphs and in each such graph we find the maximum number of edge-disjoint paths from vertex s to vertex t in time equal to O(pq). We repeat this process as many number of times as the number of graphs. The problem of efficiently finding vertex connectivity is slightly more complex as we shall explain in the coming section.4. t) (because of O(p2 ) distinct pairs of s−t vertices) we select the one with the minimum value and that will be the edge connectivity λ(G) of the graph. 6.4.18. We select a vertex s and keep it fixed throughout the working of the algorithm.

4. Ouput the minimum value of all Max-Paths. In the j th copy of graph G select vertex j as t where 2 ≤ j ≤ p. input : Undirected Graph G with vertices numbered from 1 to p. Find Maximum Number (Max-Paths) of edge-disjoint paths from s to t in each copy of graph G. .2.6.18: When does vertex s moves from one place to another? Algorithm 41: Find edge connectivity of an un-directed graph G. 1 2 3 4 Construct p − 1 copies of graph G numbered from 2 to p. or 8 will provide the optimal answer If s is fixed at 1 then t=5. In every copy of graph G select vertex 1 as s.7.3. or 4 will provide the optimal answer Figure 6.7. output: Edge Connectivity of graph G.6. or 8 will provide the optimal answer If s is fixed at 8 then t=1.Menger’s Theorem 291 2 3 4 5 6 8 7 Edge Connectivity 1 8 8 s 6 2 7 8 7 6 2 7 6 s 1 2 3 4 5 1 3 4 5 1 4 3 5 s If s is fixed at 2 then t=5.

3.19: When is vertex s fixed and when does vertex t change its position? The total time complexity will thus be O(p2 q). We now know how to find the magnitude of λ(G). Connectivity and Matching Problems 8 8 8 t t 2 1 7 6 2 7 6 2 7 6 3 4 5 1 3 4 5 s 1 4 3 5 s No. Problem Set 6.292 Network Flows.3.4. In the j th copy of graph G we select vertex j as t where 2 ≤ j ≤ p. Assume that instead of finding edge connectivity (which is an optimization problem) we intend to solve the corresponding decision problem: Is the edge connectivity of graph G less than or equal to k where k is an arbitrary number and is always less than p (why?)? Of course you can use the edge-connectivity finding algorithm to solve this problem but then it will be an overkill.3. In fact the problem becomes a search problem in a finite search space.2.1. Try to derive the time complexity of this algorithm. of Max-Paths from s to t is 2 s t No. So let us design an efficient algorithm to solve this problem. You have the option of making a linear search or a more efficient binary search in order to find the edge connectivity of a graph. We make p − 1 copies of graph D numbered 2 to p and then fix vertex s and t in each copy. Problem 6. of Max-Paths from s to t is 1 Figure 6. of Max-Paths from s to t is 2 No. Once we have solved the decision problem it becomes almost trivial to solve the corresponding optimization problem of finding the (minimum) edge connectivity of a graph. Let us now try to make the previous algorithm more efficient. In every . How about finding the actual edges belonging to the minimum sized set of edges which if removed will disconnect graph G? Problem 6.

Now instead of finding Max-Paths or the Min-Cut in each graph we find k edge-disjoint paths in each copy of graph G starting from k equal to 1.Menger’s Theorem 293 8 8 8 7 2 1 6 2 7 6 2 7 6 t t 3 5 1 3 4 5 s 1 4 3 5 s 4 s t Copies of graph G with s fixed at 1 and a different t 8 8 8 6 t 7 2 1 6 2 7 6 2 7 t 3 5 1 3 4 5 s 1 4 3 5 s 4 s t Find only One (Edge-Disjoint) Path from s to t in each graph 8 8 8 6 t 7 2 1 6 2 7 6 2 7 t 3 5 1 3 4 5 s 1 4 3 5 s 4 s t Find another (Edge-Disjoint) Path from s to t in each graph (if possible) Figure 6. If we can find k edge-disjoint paths in all copies then we move forward otherwise we exit out with the edge connectivity of the graph G.4.20: We make p − 1 copies of graph D numbered 2 to p and then fix vertex s and t in each copy. .

Connectivity and Matching Problems Algorithm 42: Find if the edge connectivity λ(G) of a graph G less than or equal to k? input : An un-directed graph G output: Yes/No 1 ? copy of graph G we select vertex 1 as s. t) for every possible pair of vertices in the graph G and then select the minimum (out of all O(p2 ) possible values). Increment k and go to step 3. If we can find k edge-disjoint paths in all copies then we move forward otherwise we exit out with the edge connectivity of the graph G. Vertex Connectivity of an undirected Graph We know how to find κ(s. If you can not find k edge-disjoint paths in any one of p − 1 graphs then exit with edge connectivity λ(G) = k − 1. Algorithm 43: Find edge connectivity λ(G) of graph G. Try to derive the time complexity of this algorithm. Now instead of finding Max-Paths or the Min-Cut in each graph we find k edge-disjoint paths in each copy of graph G starting from k equal to 1. In the j th copy of graph G select vertex j as t where 2 ≤ j ≤ p. that is the minimum number of vertices which if removed will destroy all paths between vertex s and t in a given undirected graph G. we shall get the vertex connectivity κ(G) for graph G. Let k = 1. How about if we use the same complexity cutting strategy used in finding the edge connectivity. t).294 Network Flows. 1 2 3 4 Construct p − 1 copies of graph G numbered from 2 to p. The modified Algorithm 43 is described below. Find k edge-disjoint path from vertex s to vertex t in each of the p − 1 graphs. output: Edge Connectivity λ(G) of graph G. select s arbitrarily but allow t to have all possible values? Please concentrate on the following algorithm (Algorithm 44) . If we find a κ(s. In every copy of graph G select vertex 1 as s. input : Un-directed Graph G with vertices numbered from 1 to p.

How this intelligent strategy cuts down the . Output the minimum value of all κ(s. The output of the above algorithm will not be equal to the vertex connectivity of the given graph but it will still give us an upper bound on vertex connectivity. In the j th copy of graph G select vertex j as t where 2 ≤ j ≤ p.22. The outcome of the above algorithm when applied to this graph (Fig. input : Undirected Graph G with vertices numbered from 1 to p. a p b f g s i m e h s j n d k Figure 6. Any other choice for vertex s will always provide us the correct answer in this graph. 6. Assume that the algorithm (Algorithm 44) is applied to the graph shown below. That means we should change s and again apply this algorithm but we need not do this repetition more than the latest output of our algorithm. if we select s among the vertices which belong to the minimum sized vertex set (which if removed will disconnect graph G) then our algorithm will provide an incorrect estimate for vertex connectivity.4. The vertex connectivity κ(G) of this graph is also indicated here.4. output: Vertex Connectivity κ(G) of graph G? 1 295 2 3 Construct p − 1 copies of graph G numbered from 2 to p. t)’s . 6.4. In every copy of graph G select vertex 1 as s.21) is shown in Fig.Menger’s Theorem Algorithm 44: Find vertex connectivity κ(G) of graph G. You can easily generalize these observations.21: Original Graph G The figure shows that if s is initially selected as g (or h) in this graph then the output of this algorithm will be 3 which is a wrong answer for vertex connectivity. t) in every graph. Find κ(s.

. Connectivity and Matching Problems time complexity will be interesting to explore.296 Network Flows.

m.22: The arbitrary pair s and t are used to find the Min Cut for vertex-disjoint paths between each pair of s and t and then the minimum value of those p cuts is chosen to find Vertex Connectivity of the graph as a whole.n or p and t = a.t) will not provide the Optimal answer Figure 6.j.4.d.t) will not provide the Optimal answer p If s is fixed at i.t) will provide the Optimal answer a m n k b c d If s = g and t = e then κ(s.d.k.c. or f then κ(s.b.t) will not provide the Optimal answer f e p t b c d s g s h i j k s g t h i j k m n If s = g and t = b then κ(s. .e.c.k.b.e.m.j.Menger’s Theorem 297 a Min-Cut p g h i j t k m n b c a f e Min-Cut p s g h i j k m n s b c d f e t d If s is fixed at a. or f and t = i.n or p then κ(s.t) will not provide the Optimal answer f e p s g s h i j s g s h i j k m n t t a p f e a m n b c d If s = g and t = h then κ(s.t) will provide the Optimal answer a b c f e d If s = g and t = c then κ(s.

The minimum cut in terms of number of edges will be equal to the minimum cut in terms of minimum number of vertices. 2. We shall. The following observations will help us in proving the remaining two theorems. 1.298 Network Flows.5 Konig’s Theorem.5. namely Konig’s and Hall’s Theorem. Thus maximizing number of paths in D is equivalent to finding maximum matching in B. therefore. . Let us now start with an (undirected) bipartite graph B. 3. If you cannot find an additional path from vertex s to vertex t in graph D then it means that you can not improve the size of the existing matching in the bipartite graph.1. Connectivity and Matching Problems 6. Every edge-disjoint path from vertex s to t in the graph D is in fact a vertex-disjoint path between the same two vertices. Menger’s Theorem & Hall’s (Marriage) Theorem Revisited We have earlier proved Menger’s Theorem which states that the maximum number of edge-disjoint (vertex-disjoint) paths from vertex s to t in a directed graph D is equal to the minimum number of edges (vertices) of G which need to be removed in order to disconnect vertex t from s. We transform this graph into a directed graph D after adding vertices s and t as shown in the left diagrams in Fig. This theorem is applicable to any graph while the other two theorems (Hall’s Theorem and Konig’s Theorem) are applicable to bipartite graphs only. The directed graph D with maximum number of edge-disjoint paths is transformed into a directed graph F after the direction of each path in D is reversed as shown in the top left diagram of Fig. 6. Thus the maximum number of paths between vertex s and t in G will be equal to the size of the maximum matching in the bipartite graph. There will be a path between vertex s and t in D corresponding to every matching edge in the bipartite graph B.5. We find maximum edge-disjoint paths from s to t in the graph D as shown in the bottom right diagram of this figure. 6.2. Each path from vertex s to vertex t in this diagram corresponds to a matching edge in the bipartite graph B as shown in the top right diagram of this figure. refer to these paths as paths only and not vertex-disjoint or edgedisjoint paths.

1: The top left diagram shows a bipartite graph B. We find maximum number of edgedisjoint (or vertex-disjoint) paths in graph D as shown in the bottom right diagram. Maximizing number of paths in D is equivalent to finding maximum matching in B. .5. The bipartite graph is transformed into a directed graph D with additional vertices s and t as shown in the bottom left diagram.Konig’s Theorem. there is a matching edge in the bipartite graph as shown in the top right diagram. Menger’s Theorem & Hall’s (Marriage) Theorem Revisited 299 a1 a2 A a3 a4 b1 b2 B b3 b4 A a1 a2 a3 a4 b1 b2 B b3 b4 Bipartite Graph B Transforms Maximum Matching in Graph B Transforms Directed Graph D with nodes s & t Maximum Paths in Graph D a1 a2 b1 b2 B b3 Find Max-Paths a1 a2 b1 b2 s A a3 a4 t s A a3 a4 b3 b4 t b4 Figure 6. Corresponding to each edge-disjoint path from vertex s to vertex t in the directed graph D.

The bottom left diagram shows the set Y containing (minimum number of) vertices which if removed will destroy all paths from vertex s to t in the directed graph.300 Network Flows.2: The top left diagram shows directed graph D where it is no longer possible to find an additional path from vertex s to t.5. . these vertices form the set P . However it is still possible to reach some vertices from s. The set X contains those vertices which are common in the set P and the partite A. Connectivity and Matching Problems P a1 a2 b1 b2 B b3 b4 a1 a2 b1 b2 s A a3 a4 t s a3 A a4 b4 b3 t Graph F: Nodes Reachable from s are in P Partition P in Graph D P a1 X N(X) b1 b2 a1 a2 b1 b2 a2 s a3 A a4 b4 Set X: Vertices common in A & P t b3 s a3 Min-Cut A-X t b3 b4 a4 Neighborhood of X & the Min-Cut a1 X Y b1 b2 a1 a2 b1 b2 B b3 b4 a2 s a3 a4 b3 b4 t A a3 a4 Set Y= Minimum Vertex Cut in D Set Y= Minimum Vertex Cover in B Figure 6. The set Y is in fact the vertex cover in the bipartite graph shown in the bottom right diagram. The middle right diagram shows the minimum cut.

There are two major cases to be considered as shown in Fig. then reverse the direction of each path.2. 9. 6. 6.4.5. 6. The size of the minimum cut in terms of vertex-disconnecting set will be A − X + N (X) . If the set P contains a vertex x belonging to A then obviously vertex x is not matched to a vertex in the B in the maximum matching of the bipartite graph.5. The same set Y is in fact the vertex cover in the bipartite graph in the bottom right diagram. we find maximum edge-disjoint paths in D. 5.5. Menger’s Theorem & Hall’s (Marriage) Theorem Revisited 301 4. the resulting directed graph F is shown in the top right diagram of this figure. 6. The minimum cut in terms of edge-disconnecting set will be edges from s to A − X and from N (X) to t as shown in the middle right diagram of Fig.1. If the set P does not contain any vertex of A then it implies that the number of paths from s to t is equal to the degree of node s in the graph. The minimum sized vertex-disconnecting set in the graph D will correspond to the vertex cover in the bipartite graph B. Each matching edge in the bipartite graph corresponds to a .5. It also implies that every vertex belonging to A is matched to a vertex in B in the corresponding bipartite graph. 8. Problem 6.2.5.2. 6. The size of the minimum cut will thus be equal to the size of A − X and N (X). Let X represents vertices which are common between P and the partition A as shown in the middle left diagram of the same figure. It is first transformed into a directed graph D as described earlier.Konig’s Theorem. Under such conditions the size of the maximum matching will be (at least one) less than the size of partition A.4 below shows a bipartite graph B in the top left diagram. This set (A-X +N(X)) is denoted by Y in the bottom left diagram of Fig. 6. The neighborhood N (X) of X is shown in the middle right diagram of this figure.4. Obviously P will not contain t (why?). Problem Set 6. 7.3. Let P represent the vertices which are reachable from vertex s once we have found maximum number of paths and it is no longer possible to find an additional path from s to t in D as shown in the top left diagram of Fig. The Fig.

Connectivity and Matching Problems All vertices of A not matched in the Bipartite graph Implies Number of Paths less than size of A a1 a2 A a3 a4 b1 b2 B b3 b4 a1 a2 b1 b2 B b3 b4 s A a3 a4 t Case 1: From s you can reach a vertex of A Partite but not to a vertex of B Partite in the graph F a1 a2 A a3 a4 b1 b2 B b3 b4 a1 a2 P b1 b2 B s A a3 a4 b3 b4 t Case 2: From s you can reach a vertex of A Partite as well to a vertex of B Partite in graph F Figure 6.5.3: Hints for a proof of Hall’s Theorem .302 Network Flows.

For each bipartite graph (with a maximum matching indicated) draw the corresponding graph F .5.5. It is interesting to note that all vertices belonging to P . The figure shown above (Fig. are part of a directed cycle containing vertex s. Menger’s Theorem & Hall’s (Marriage) Theorem Revisited 303 directed path from vertex t to s in F . and N (X). Find the minimum cut in terms of edges as well as vertices of F . . It is quite obvious that now it is no longer possible to find an additional path in the graph F .4.2. there is a directed path from vertex t to vertex s corresponding to each matching edge in B. A − X. 6. Find the sets X.Konig’s Theorem.4: Problem 6. X. and A − X in the graph F . N (X). Find the set P.4) shows two different maximum matchings of the same bipartite graph B in the bottom diagrams. Also find the minimum cut in terms of vertices as well as edges. It further implies that we have found the maximum matching in the bipartite graph B. All vertices which are still reachable from vertex s belong to the set P which is also indicated in the right diagram. a1 a2 A a3 a4 b1 b2 B b3 b4 Transforms a1 P a2 b1 b2 s a3 a4 b3 b4 t a1 a2 A a3 a4 b1 b2 B b3 b4 A a1 a2 a3 a4 b1 b2 B b3 b4 Figure 6. You know that in graph F .

Can we prove this observation in general or is this localized to this graph only? What are the implications of this observation? Is this something to do with the proof of Hall’s Theorem? a1 a2 a3 a4 b1 b2 a1 a2 b1 b2 s b3 b4 a3 a4 b3 b4 t Figure 6.4. known as the set P is also indicated in each graph F . We show a bipartite graph with two different maximum matchings in Fig.6.5.7) shows a bipartite graph B with maximum matching edges shown in different colors.5. It is obvious that the maximum matching in this bipartite graph is not a perfect matching – all vertices of partition A are not matched. The right diagram shows the same bipartite graph with two extra vertices s and t added to it. Is this a coincidence in this bipartite graph or will it always be true.5. The figure shown below (Fig. The left diagram of Fig. Note that the bipartite graph B in this problem is transformed into an un-directed graph G instead of a directed graph D.5. there is a path from . 6. 6. Under such conditions vertex s will be part of a directed cycle.4. Prove or give a counter example. It is obvious that corresponding to every matching edge in this bipartite graph. It is quite evident from these diagrams that for different maximum matchings in the same bipartite graph we get the same set P .4. The corresponding graph F showing reversed paths is shown on the right side of each bipartite graph. Connectivity and Matching Problems Problem 6. 6.3.5 shows a bipartite graph with maximum matching edges indicated in different colors. Problem 6. The set of vertices reachable from vertex s. such that the number of vertices of partition A in the cycle will always be one larger than the number of vertices of partition B belonging to the same cycle.304 Network Flows.5: Problem 6.5.4.

Menger’s Theorem & Hall’s (Marriage) Theorem Revisited 305 a1 a2 a3 a4 b1 b2 a1 a2 b1 b2 s b3 b4 a3 a4 b3 b4 t a1 a2 a3 a4 b1 b2 a1 a2 b1 b2 s b3 b4 a3 a4 b3 b4 t Figure 6.Konig’s Theorem.6: .5.

6. 6.7) into an undirected graph G as shown in the figure above. Assume that we convert a bipartite graph (any bipartite graph – not necessarily the one shown in Fig. Connectivity and Matching Problems between vertex s and vertex t in the graph G (Is this one to one correspondence between a matching edge in the bipartite graph and an edge-disjoint path in graph G a general phenomena or is it restricted to this bipartite graph? We shall address this issue in the next problem).6 6. 6. Do you think that corresponding to every edge-disjoint path between vertex s and t in this undirected graph.306 Network Flows. there will be a corresponding matching edge in the bipartite graph B? Either prove or give a counter example.5.6.7: Problem 6.1 Network Flows Finding Maximum Edge-Disjoint Paths in MultiGraphs We consider here the problem of finding maximum number of edge-disjoint paths (and a minimum cut in terms of edges) from a vertex s to a vertex . We find maximum edge-disjoint (or vertexdisjoint) paths between vertex s and vertex t in graph G. a1 a2 a3 a4 b1 b2 a1 a2 b1 b2 s b3 b4 a3 a4 Graph G t b3 b4 Bipartite Graph B Figure 6.4. Find the minimum edge cut & minimum vertex cut in graph G and show that it is equal to the size of the vertex cover in this bipartite graph B.5.

For example if P = (s → a → d → t) as shown in the top left diagram of Fig. We reproduce an earlier algorithm below for a ready reference. If you are successful in finding a path P then keep a record of the number of paths found so far otherwise exit with graph F as output.Network Flows 307 t in a directed multi-graph. 6.1 then we have not one but three edgedisjoint paths passing through the same vertices. output: Maximum pseudo edge-disjoint paths from s to t in D and a Status(e) for every directed edge e of D. 1 2 3 4 Copy Graph D into F . for every edge e in path P do set Status(e) = used if initially it was unused otherwise set Status(e) = unused. input : Directed graph D. all paths parallel to P (that means passing through the same set of vertices) can be discovered right away.6. Status(e) = unused for every directed edge e in D. How this will cut down the time complexity of our modified approach? We have asked you to derive the time complexity of this algorithm (with or without modification) when applied to Multi-graphs. This may be possible without damaging the original character of our algorithm. We designed this algorithm to find maximum edge-disjoint paths in a directed graph. Find a directed path P from vertex s to vertex t in F . and vertices s & t. We can use our earlier algorithms to solve this problem. Whenever we find a path P from vertex s to a vertex t according to step number 2 of this algorithm we find the vertices through which P passes. It will be useful if we spend some time on the selection of a suitable data structure to represent . Reverse the direction of edge e and then go back to step 2. We assume that there are no self loops but parallel edges are allowed in the directed multi-graph. Instead of finding these paths sequentially (strictly according to this algorithm) we should be able to do it in one go as shown in the same figure. Algorithm 45: Find Maximum pseudo edge-disjoint paths from s to t in directed graph D. calculating the resulting time complexity would be interesting? We can certainly make adjustments in order to increase the efficiency of our earlier approach.

1: The top left diagram shows a multi-directed graph with three selected paths from vertex s to vertex t shown in bold. The top right diagram shows the edges of these paths reversed. .308 Network Flows.6. Connectivity and Matching Problems a c a c s b Graph D a c d t s b d t Initial Graph F a c s b d t s b d t a c a c s b a d c t s b d t a c s b d t s b Graph D d t Final Graph F Figure 6. The minimum cut is also indicated in the bottom diagrams. The bottom right diagram shows a stage when it is no longer possible to find an additional path from s to t in this graph. We then find another path and reverse the direction of its edges.

The right diagram shows how it can be represented by a simple weighted graph. Please note that it is an un-weighted graph.6. How can we exploit this data structure in order to use Algorithm 45 (efficiently) to solve the edge connectivity problem? When you derive the time complexity of Algorithm 45 (or its modified version) you may realize that the complexity expression may depend upon the graph edge weights in addition to the size of the problem (that means number of vertices and edges). the weight of each edge (x. That is in fact a more serious problem (than having a less efficient algorithm). 6.Network Flows 309 a multi-graph.2 reproduces the multi-directed graph of Fig. The right diagram shows how it can be represented by a weighted graph.1.6.1. and can conveniently be represented by a weighted adjacency list or adjacency matrix data structure. Why has this happened? How can we over come this shortcoming? a c 3 a 4 c 2 3 s b d t s 1 3 t d 4 b 1 Figure 6. Note that this graph (shown in the right diagram) is a simple graph with no parallel edges. 6.6.2: The left diagram shows the un-weighted multi-directed graph of Fig. . 6. it implies that your algorithm is no longer an algorithm but a technique? Remember that in an algorithm the complexity depends upon the number of vertices and edges of a graph and does not depend upon the edge weights in the graph while in a technique it may depend upon edge weights as well. The weight of each edge (x. The left diagram of Fig. y) in this weighted graph corresponds to the number of edges from vertex x to vertex y in the un-weighted multi-directed graph shown in the left diagram.6. y) in this weighted graph corresponds to the number of edges from vertex x to vertex y in the un-weighted multi-directed shown in the left diagram.

Then the flow taking place on the cut from S to T will be donated by (S. There are a number of following related problems and issues with network flows: 1. Connectivity and Matching Problems 6.3. 2. v) does not exceed the capacity of the edge. the weight of the edge is known as the capacity of the edge. that is.310 Network Flows. The maximum flow in a . T ) and some flow in the direction from t to s which is represented by f (T. Let us assume that (S. the network flow f (N ) is the difference between the two. v) taking place in any edge (u. The net flow coming in or going out of a vertex other than s to t is zero. T ) − f (T.2 The Maximum Flow & the Minimum Cut We are given a network which is essentially a weighted directed graph with two special vertices (a source vertex) s and (a sink vertex) t as shown in Fig. Each directed edge is weighted with a positive integer. v) is a directed edge then the capacity of this edge is denoted by c(u. whereas in the case of s and t the flow going out of s is absorbed by the flow going into t. 3. 6. We assume that flow can originate from the source vertex and can be consumed by the sink vertex. v). Thus the network flow f (N ) is given by the equation: f (N ) = f (S. The net flow coimng out of the source verex s is greater than zero while the netflow sinking in the sink vertex t is also finite.6. c(u. S) .6. v). The network flow f (N ) taking place in a network graph N is equal to the net flow coming out of the source vertex s or the net flow consumed in the sink t. The network flow problem is to find the maximum flow which can take place from the source vertex to the sink vertex such that the flow f (u. Hence every cut will have some flow in the direction from s to t represented by f (S. which means that whatever flow goes into an intermediate vertex is equal to the flow coming out of that vertex and no new flow is generated by the vertex itself. It means that the incoming flow through an intermediate vertex is exactly equal to the outgoing flow through that vertex. T ) is a cut in the network such that the source vertex s belongs to set S while the sink vertex t belongs to set T . S). If (u. T ) while the flow taking place from T to Swill be represented by (T < S). All vertices other than the source or the sink (known as intermediate vertices) can neither generate any flow nor consume any flow.

6. This relationship is described by the famous MinCut-MaxFlow Theorem. It also shows the maximum flow and the minimum cut. Here we apply our earlier techniques of finding maximum edge-disjoint paths in a .6. 5.6. the curious reader may have realized that the algorithm for finding maximum edge-disjoint paths for multi-graphs can be used to find maximum flow as well the minimum cut in the network. The right diagram shows the actual flow taking place in an edge divided by the capacity of that edge. 6. The Fig. Given a network it is possible to efficiently find the maximum flow and the minimum cut in the network.4 shows a network graph with upper bounds on edge flow. T )} where the minimum is taken over all cuts (S. and minimizing the flow in the opposite direction (that is f (T. In every network the value of the maximum flow is equal to the capacity of a minimum cut. 4. one serious problem regarding (the complexity of) this algorithm. a 3 4 c 2 3 3/3 a 1/4 c 1/2 s 1 3 t d 4 s 1/1 1/3 2/3 t d 4/4 b 1 b 1/1 Figure 6. S)). T )). T ) in N . The flow f (N ) in the network is bounded by the expression: f (N ) ≤ min{c(S. 6.3 Algorithmic Issues & Complexity Calculations Although we have not formally described an algorithm to find a maximum flow in a network in the last section. There is however. we have briefly talked about this issue before and shall try to settle it now. 6.3: The left diagram shows a network flow graph with the capacity of each directed edge shown.Network Flows 311 network is achieved through maximizing the flow in one direction (that is f (S.

every cut cuts a number of edges which if removed will destroy all paths from vertex s to vertex t. 6.5 shows various stages of the working of our algorithm while it is trying to find a maximum flow in the network.6. Please note that if the upper bound on flow in every edge is 3000 instead of 3 – even then we have to apply the same number of steps to find the maximum flow. The Fig. Here the time complexity has become dependent not only on the size of the problem but the magnitude of the numbers involved. Unfortunately the said proof and the resulting time complexity calculations are beyond the scope of this book – we. We find another shortest path from s to 2 to t in the second step – again a flow of 3 units is possible in this path. The maximum flow and the minimum cut are found after applying BFS twice in this graph. We apply our earlier technique of finding any path (not necessarily shortest) from vertex s to vertex t in this network. This time we take a longer path from vertex s to vertex t.6. The algorithm finally converges and we get the correct answer but after passing through a number of iterations proportional to k where k is an upper bound on flow in the network. it can be proved that application of BFS (a minor change in our algorithm) ensures (something really big) that the time complexity of the resulting algorithm will not depend upon the magnitudes of the upper bounds on flow – it will only depend upon the size of the problem.6. Fig. therefore rely on our earlier techniques. In fact. we use breadth first search – thus ensuring that we find a shortest path in terms of number of edges. In the shortest path we move from vertex s to 1 and then to t – thus ensuring a network flow of 3 units from the source to the sink as shown in the top right diagram of this figure. In other words the time complexity does not depend upon the magnitude of the upper bound on flow at least in this example. The Min-Cut passes through minimum number of edges while the . The path goes from s to 1 to 2 and then to t.312 Network Flows. Connectivity and Matching Problems directed graph with just one important difference – whenever we find a path from vertex s to vertex t. We reverse its edges and then take another path – this time it goes from vertex s to 2 to 1 to t.6. We also show a number of cuts in this graph. Let us first find out what extra price we have to pay if we do not use BFS. 6.4 Lower Bounds on Edge Flows and the Max-Cut We show a directed graph with two special vertices known as s and t in the Fig.5 below shows the same network graph. 6. 6.6 below.

4: While finding a path from vertex s to vertex t we do use BFS.6. . we show the resulting repercussions.Network Flows 313 3 3 1 3/3 3/3 0/1 s 3 t 3 s 0/3 t 0/3 3/3 3/3 0/1 3/3 3/3 0/1 s 3/3 t 3/3 s 3/3 t 3/3 Figure 6.

Connectivity and Matching Problems 1 3 3 1 2 1 3 s 3 t 3 s 3 1 t 2 2 2 2 2 2 1 1 2 s 1 t s 1 t 2 2 2 2 2 2 1 2 2 3 1 2 s 2 1 t 2 s 2 1 t 3 2 2 3 1 3 3/3 1 3/3 s 1 t 3 s 0/1 t 3 2 Min-Cut 3/3 2 3/3 Figure 6. it shows the resulting repercussions.6. .5: While finding a path from vertex s to vertex t we do not use BFS.314 Network Flows.

6.Network Flows 315 Max-Cut passes through maximum number of edges. Food for thought If we intend to find a Max-Cut in a network flow graph then why not remove all edges of the graph .6. 6. The (size of the) Max-Cut specifies the minimum number of paths from vertex s to t where each path passes through every edge of the graph at least once (see the right diagram of Fig.6: The left diagram shows various cuts which disconnect graph and the right diagram shows the Min-Cut for this graph.7).6). The size of the Min-Cut in this graph is only two and thus there are only two edge-disjoint paths in this graph (between s and t) which are also indicated in the right diagram shown below (Fig.6. The size of the MinCut specifies the maximum number of edge-disjoint paths between s and t in a graph (according to Menger’s Theorem). 6.the source vertex will certainly be separated from the sink vertex and the size of the cut will be maximum? Please think about this strategy and find what is wrong with this otherwise simple approach? We now address the problem of how to find Max-Cut and minimum number .6. Min-Cut Figure 6. 6. Remember the (size of the) Min-Cut specifies the maximum number of (edge-disjoint) paths where each edge is traversed at most once. The Max-Cut is indicated in the diagram above and is replicated again in Fig. As it is clear from its name Max-Cut passes through maximum number of edges which if removed will destroy all paths between vertex s and t in the graph.7. Interestingly the Max-Cut in a graph ia also important. Please note that we have already described an efficient algorithm to find a Min-Cut or maximum edge-disjoint paths in a graph.

6. 6. Connectivity and Matching Problems Max-Cut Max-Cut Figure 6. It may have become obvious now that if we can solve the problem of finding a Max-Cut and Min-Flow in a network graph with lower bounds on edge . In other words we need to find the maximum cut in the network graph D that will disconnect vertex t from vertex s in D. Instead of devising an entirely new algorithm let us explore if we can solve the problem using existing or modified algorithms.6.8 shows a network graph D with the maximum capacity of each edge indicated.7: The left diagram shows the Max-Cut for the graph and the right diagram shows the minimum number of paths where each path passes through every edge of the graph at least once. The problem is to find a minimum flow in this graph such that the flow taking place through any edge does not go below the lower bound of that edge. where each edge is traversed at least once. 6. of paths between vertex s and vertex t in a given directed acyclic graph.6. The intellectual exercise of reducing one problem into another is always an exciting venture – especially when the similarity between the two problems is not so obvious.9. The Fig.316 Network Flows.9 except that each weight associated with an edge in this network graph signifies not the upper bound but a lower bound on the flow that can take place through that edge.6. The maximum flow and the corresponding minimum cut are shown in the right diagram of the same figure. The minimum flow and the corresponding maximum cut are indicated in the right diagram of Fig. We have earlier discussed the network flow problem where each edge has an upper bound on the amount of the flow that can take place. and we need to maximize the total flow taking place in the entire network.6. We show the same network graph with the same edge capacities in Fig.

6. 4 3 2 3 1 1 A Network Flow Graph D with Lower Bounds on Edge Flow 4 7/3 4/4 4/2 Max-Cut 3/3 1/1 3/1 Size of the Minimum Flow and Maximum Cut in the Network = 11 s 3 t s 3/3 7/4 t Figure 6. .9: The left diagram shows the lower bound for the capacity of each edge and the right diagram shows the Minimum Flow and Max-Cut for the same graph.8: The left diagram shows the capacity of each edge and the right diagram shows the corresponding Maximum Flow and Min-Cut for the graph.6.Network Flows 317 4 3 2 3 1 1 3/3 1/4 1/2 Min-Cut 1/1 1/3 1/1 Size of the Maximum Flow and Minimum Cut in the Network = 5 2/3 4/4 s 3 t 4 s t A Network Graph D with Upper Bounds on Edge Flow Figure 6.

input : A weighted directed network graph D with vertices s and t. y) . The weight of each edge signifies an actual and acceptable flow taking place through that edge. The input as well as the output networks of this algorithm is shown in Fig.6. output: A weighted network graph D. The weight of any edge signifies the lower bound of flow that can take place through that edge.10 is not a minimum flow taking place from vertex s to t in the network.10. we assume that the lower bound on flow through each edge is equal to 1. 1 2 3 4 for every directed edge (x. A zero flow through every edge is a possible answer. Then we try to increase and maximize the flow from vertex s to vertex t in the network. Algorithm 46: Find an acceptable (not necessarily minimum) flow from vertex s to vertex t in a given network graph D. Algorithm 46 finds an acceptable flow through the network D. We should be careful about one thing – the conservation of the flow taking place in the network. Please note that an acceptable (or legal flow) shown in the right diagram of Fig. b) in path P do Push an additional flow in the entire path P equal to the lower bound on edge (x. 6. Remember when we have (only) an upper bound on flow that can take place through any edge in a network then we start with a small acceptable flow – so small that it can take place through every edge without violating any bounds. y) in graph D do Find a path P from vertex s to vertex t in graph D passing through edge (x. 6.6. It requires that the total flow coming towards a vertex should be exactly equal to the total flow going out of that vertex (except for vertices s and t in the network). Now when we have lower bounds on flow that can take place through any edge then we should start with a large acceptable flow – large enough that the lower bound (or limit) on flow through any edge is not violated. Connectivity and Matching Problems capacities then we can also solve the problem of finding minimum number of paths between vertex s and vertex t in a given directed graph where each edge is traversed at least once.318 Network Flows. y) for every edge (a.

x) in graph F with a weight equal to w −c(x. y) in D with a weight equal to w. It means that all vertices as well as directed edges of D are copied in directed graph F . For every edge (x. 6.11. y). The . We copy graph D into graph F without any edge weights.6. We start with an acceptable flow in the network graph D as shown in the top left diagram of Fig. where c(x.6. The resulting graph F is shown in the top right diagram of Fig.10: The left diagram shows the input graph to Algorithm 46 and the right diagram shows one of the possible output for the algorithm. x). The weight of each edge in this diagram shows how much actual flow is taking place in this network without violating any lower bound on flow taking place in any edge. We find the maximum flow in graph F taking place from vertex t to vertex s as shown in the middle left diagram.11. now we need to minimize it. The Min-Cut in graph F from vertex t to vertex s is shown in the middle right diagram. 6. After we have found an acceptable flow from vertex s to vertex t in the network graph D. 6.Network Flows 319 1 1 1 3 2 2 1 2 s 1 1 1 1 t 1 s 2 2 t 5 Given Directed Graph D with Lower Bounds on Edge Flow An Acceptable Flow: the weight w of an edge shows how much actual flow is taking place Figure 6. The maximum flow taking place through any edge (y. x) in this graph is represented by m(y. Instead of minimizing this flow directly we do so indirectly – by pushing an opposite flow taking place from vertex t to vertex s in D as shown in the Fig. The network graph D is converted into a directed graph F as follows: 1. y) is the lower bound on flow that can take place in the edge (x.6. y) in the graph D.11 below.6. there is an additional directed edge (y. 2.

11: The diagram shows the entire process of finding the MaxCut for the graph by minimizing the flow(middle left diagram) after the acceptable flow (left-right diagram) which eventually leads to finding the Max-Cut for the graph (in the bottom-right diagram).6. . Connectivity and Matching Problems x 3 2 y 2 2 x 1 y 1 s 2 2 2 1 5 t s 1 1 0 4 1 t An Acceptable Flow in D: the weight w of an edge shows how much actual flow is taking place through this edge Directed Graph F: the weight of a black edge shows an upper bound on edge flow x 1/2 1/1 y 1/1 1/2 1/1 1/1 Min-Cut s 1/1 1/1 0 2/4 1/1 t s 1/1 1/ 1 0 2/4 1/1 t Maximum Flow from t to s in Directed Graph F = 3: Size of the Min-Cut from t to s in Graph F = 3 x 2 1 y 1 Max-Cut s 1 1 1 t 3 s t 1 The weight of an edge is equal to the Acceptable Flow in D minus the Maximum Flow in F in that edge. Minimum Flow from s to t in D = 4 Size of the Max-Cut from s to t in Graph D = 4 Figure 6.320 Network Flows.

y) − m(y. Here the weight of an edge (x. y) is equal to w(x. the size of the two cuts are however different as is evident from Fig.Network Flows 321 corresponding minimum flow in the graph D is shown in the bottom left diagram. .6.11. The Max-Cut in the graph D from vertex s to vertex t is shown in the bottom right diagram. 6. Please note that the Max-Cut in D from s to t cuts the same edges as the Min-Cut in graph F from vertex t to vertex s. x).

How can you design an efficient algorithm which will output either yes or no. We are given a network graph D with special vertices s and t. Musharraf designed the following intelligent algorithm to find an acceptable flow provided it exists in the network graph of the above problem: (a) Initially ignore the upper bound on each edge and find a minimum flow in the network (from vertex s to vertex t) keeping into account the lower bounds on flow through each edge.6.5. 6. How about if the lower bound is a positive number (it may be different for different edges) while the upper bound on flow is the same for all edges? Repeat the earlier problem for this new situation.5. Network Flows. We need to find minimum number of paths from vertex s to vertex t such that each vertex of D is traversed at least once by any of the s − t paths.4.5. Such graphs (where an acceptable flow) is possible have a special structure as we shall study in coming chapters. We are given a directed network graph D with special vertices s and t. (c) If an acceptable flow is possible then find a minimum flow in this network.5. Problem 6.6. We need to make sure that a same fixed amount of flow should take place through every directed edge in that graph. (b) If an acceptable flow is possible and is provided to you then find a maximum flow in this network. Connectivity and Matching Problems Problem 6. (a) Design an efficient algorithm to find if an acceptable flow is possible in this network.322 Problem Set 6.5. Once we find an acceptable flow – it can always be maximized or minimized.2. Problem 6. (b) An acceptable flow exists if and only if the flow through any edge does not exceed its upper bound.5. Now Musharraf insists that his algorithm will find an optimal answer for the first problem in this problem set.3. Problem 6. . we assume that the lower bound as well as upper bound is the same for each edge although the two bounds may be different from each other. Kashif finds the following counter example for the above problem as shown in the figure below (Fig. prove that this algorithm always finds the correct result or find a counter example. Either prove that Musharraf is right or find a counter example.1.12. Problem 6.12. Problem 6. 6.6.5. Each edge in this network has a lower as well as an upper bound on edge flow. We are given a network graph D with special vertices s and t as in Fig.5. yes in case an acceptable flow is possible and no in case it is not possible? A special case of this problem is when the fixed flow through every edge is exactly one.

5.6. 6. however. violates the upper limit. We are given a network graph D with special vertices s and t.6 (also shown below Fig.13). shown in red color.5. Problem 6.4. The flow.8. We need to find if there is a feasible flow in this network graph? Can we design an efficient algorithm to solve this problem? If the in-degree of any node (other than vertex s and t) is more than two times the out-degree of that node in this graph then it will not be possible to find a feasible flow in the graph (why?).12: Upper/Lower limit for flow in each edge is shown. classify used and unused edges and then redraw the directed graph . Please recall the directed graph D shown in top left diagram of Fig.6.13: Problem 6.7. We apply Algorithm 45 on this graph.Network Flows 323 6/6 6 6/1 7 s 6/4 t 4 6/1 3 Figure 6. 6. If.6. The lower limit for flow is 1 while the upper limit is 2 for each edge of this graph. this condition is false in a network graph (that means the in-degree is at most two times the out-degree of any vertex other than vertex s and t) then a feasible flow will exist in the network graph? Can you counter this argument? Can you now design an efficient algorithm to find a feasible flow in the graph? a g e f h a g e f h b b s c d i j t s c d i j t Graph D with Used and Unused edges Graph D without Unused edges Figure 6.

We claim that in any network graph D without the unused edges. & (b) We can use (the so called stupid) Algorithm 36 to find maximum edge-disjoint paths in this graph.324 Network Flows. Can you think of a directed graph where property (a) is applicable but (b) is not? Can you think of a graph where (b) is applicable but (a) is not? Problem 6.5. either prove this or give a counter example.10. Connectivity and Matching Problems D without the unused edges.9.7 The Matching Problem in Bipartite Graphs We have already talked about the Matching problem in different contexts.6. The two graphs are reproduced in the figure below for ready reference. A directed network graph D without unused edges has two important properties: (a) every edge is part of a directed path from vertex s to vertex t in D. it is possible to push a flow of exactly one unit in every directed edge. Consider a directed graph D where the in-degree of every vertex (other than s and t) is equal to the corresponding out-degree. Here we shall address the matching problem in un-weighted and then weighted .5. (a) Is it possible to push a flow of exactly one unit through every directed edge in this graph? (b) If it is possible then does it mean that every edge in this graph will be part of a directed path from vertex s to vertex t in this graph? (c) If (a) is possible then can we find maximum edge-disjoint paths from s to t in this graph using the stupid Algorithm 36? Vertex 1 is s and 7 is t Any vertex is s and also t Figure 6.14: 6. Problem 6.

We convert undirected edges of the bipartite graph (top left diagram) into directed edges as shown in the top right corner.The Matching Problem in Bipartite Graphs 325 bipartite graphs. An edge between a boy and a girl shows a degree of compatibility. We need to find maximum pairs of boys and girls such that the boy and girl .7. this completes the transformation from a bipartite graph of the top left corner into the directed graph shown in the top right diagram. The algorithm described above works well for finding a maximum matching in an unweighted bipartite graph. We are given a bipartite graph having a partition B (consisting of a number of boys) and a partition G (consisting of a number of girls).7. please note that the direction of each such edge is from the partition A to partition B. We add directed edges from vertex s to every vertex in partite A. The matching problem is converted into a connectivity problem by adding two dummy vertices s and t as shown in the top right corner of Fig. present (in the coming sub-section) a slightly different version of this algorithm to find maximum matching in an unweighted bipartite graph? Problem Set 6.7. therefore. Once we find the maximum number of paths in the directed graph (see the bottom left diagram of Fig. Unfortunately it is difficult to adapt this algorithm to find a maximum weighted matching in a weighted bipartite graph. The problem of finding maximum matching in a bipartite graph is thus transformed into the problem of finding maximum number of edge-disjoint (or vertex-disjoint) paths in the directed graph. We shall devise efficient algorithms to find maximum matching in un-weighted bipartite graphs and weighted maximum matching in a weighted bipartite graph. 6. 6.7.1. 6. As explained before each matched edge in the bipartite graph (top left corner) corresponds to a path in the directed graph (top right corner). We. We also add directed edges from every vertex of the partition B to vertex t. in fact any two matched edges in the bipartite graph will correspond to two edge-disjoint directed paths from vertex s to vertex t in the corresponding directed graph.6. 6.1.1 Maximum Matching in Un-weighted bipartite graphs Given a maximal matching.1) we can find the corresponding matched edges in the bipartite graph as shown in the bottom right diagram of the same figure. how can we find a maximum matching in a bipartite graph? A maximal matching in a bipartite graph is shown by bold lines in the top left corner of Fig.

. Connectivity and Matching Problems a1 a2 A a3 b1 b2 B b3 a1 s a2 a3 b1 b2 b3 t Directed Graph F a1 s a2 a3 b1 b2 b3 a1 t s a2 a3 b1 b2 b3 t a1 s a2 a3 b1 b2 b3 a1 t A a3 a2 b1 b2 B b3 Figure 6. The middle right diagram shows maximum number of pseudo edge-disjoint paths from vertex s to t. The top right diagram shows a path from vertex s to vertex t corresponding to each matching edge in the bipartite graph. The bottom left diagram shows maximum number of edge-disjoint paths in the directed graph while the bottom right diagram shows the corresponding maximum matching in the bipartite graph.326 Network Flows.7. This diagram is converted into the middle left diagram by reversing every edge in each path.1: The top left diagram shows a maximal matching in a bipartite graph.

This problem is also known as the Marriage Problem.The Matching Problem in Bipartite Graphs 327 in every pair are compatible. Assume that we intend to solve the decision problem in which we intend to find a yes/no answer corresponding to the question: Is it possible to marry all boys? (Or is it possible to marry all girls?) .7.3.7. a partition G showing. Assume that we are required to solve the marriage problem: namely we intend to marry maximum number of girls. How will you model this problem in terms of a known graph problem? Under what conditions it will be possible? Problem 6. it can be modeled by an un-weighted bipartite graph as shown below (Fig. Assume that we are required to solve the marriage problem: namely we intend to marry maximum number of boy. each girl can marry a single boy.2.1. each boy can marry a single girl and each girl can marry a single boy. Compatibility between a boy and a girl is indicated by an edge between the corresponding vertices.2). The problem can be solved by techniques similar to the ones that we have just studied. How will you model this problem in terms of a known graph problem? Under what conditions it will be possible? Problem 6.6. 6. Problem 6.6. a set of girls. 2 3 Figure 6.2: A bipartite graph showing a partition B consisting of a set of boys.6.

Can we still claim that the maximum matching in the bipartite graph is equal to the maximum number of edge-disjoint paths between vertex s and vertex t in the undirected graph D. The maximum matching in a bipartite graph is equal to the maximum number of edge-disjoint paths in a directed graph D. Discuss with the help of an example.6. Problem 6.7.1. We need to find maximum weighted matching in this weighted bipartite graph. Assume that a boy can marry four girls. thus converting it into an undirected graph. How will you model this problem in terms of a known graph problem? Under what conditions will it be possible? Problem 6.2 Maximum Matching in Complete (Binary) Weighted Bipartite Graphs We are given an unweighted balanced bipartite graph as shown in left diagram of Fig.6. Connectivity and Matching Problems Problem 6. How about if we remove all directions from the directed graph D. 6. Problem 6.6.2. Assume that a boy can marry four girls and each girl can marry a single boy. 6. How will you model this problem in terms of a known graph problem? Under what conditions will it be possible? Problem 6. we intend to maximize the number of marriages taking place.7.6. We are given a bipartite graph with edge weights equal to either zero or one.4.5. How will you model this problem in terms of a known graph problem? Under what conditions will it be possible? Problem 6.7. Discuss how you will solve this problem efficiently using similar techniques.7.7. In other words the problem of finding maximum matching in a bipartite graph is transformed into the problem of finding maximum number of edge-disjoint (or vertex-disjoint) paths in a directed graph D. We intend to maximize the number of married boys and girls subject to the condition that a boy can marry four girls while a girl can marry a single boy.7.328 Network Flows. we intend to maximize the number of boys who are married. we intend to maximize the number of girls who are married.3.6. Assume that a boy can marry four girls. Problem Set 6. A maximum matching in this graph is also shown in this .

A maximum weighted matching in the right graph of this figure will be a maximum matching in the original bipartite graph. As shown in Fig. It can handle (with or without minor modification) the more general problem of finding maximum weighted matching in a bipartite graph. We now describe a (not very . The bipartite graph is converted into a completely connected binary weighted bipartite graph as shown in the right diagram. 6.7. The results of this algorithms can be used to find maximum matching in a bipartite graph.see the right diagram of this figure.7. It can be found using our expertise gained in the last section. An edge in the original bipartite graph has a weight of 1 in the complete bipartite graph while every other edge has a weight equal to zero. The algorithm can also be used to find maximum weighted in a complete bipartite graph with non binary weights as described in the next section. We describe in the following paragraphs a useful algorithm to solve the maximum weighted matching problem in a complete bipartite graph with binary weights. We convert the bipartite graph into a weighted directed graph D after adding vertices s and t according to the rules described previously.The Matching Problem in Bipartite Graphs 329 diagram.3 we convert this bipartite graph into a complete weighted bipartite graph .3: A balanced bipartite graph G with maximum matching of size 3 is shown in the left diagram. We start with a complete weighted bipartite graph G of size k (it means there will be k vertices in A as well as k vertices in B). Edge weights not shown in the right diagram are equal to zero. a1 a2 A a3 b3 b1 a1 1 1 b1 1 b2 B Convert into completely connected bipartite graph a2 A a3 1 b2 B b3 1 Weight of Maximum Matching = 3 Weight of existing edge is 1 and weight of new edge is 0 Figure 6. Thus the complete bipartite graph has binary weights. Here we shall describe another algorithm to find the maximum mathing this new algorithm may not be as efficient as the one described before but it has the added advantage of being flexible.

If you can not find such a path then terminate.330 Network Flows. y) with a negative sign and go back to Step 2 Problem Set 6. Whereas in the case of Algorithm 47 for a complete weighted bipartite graph. How is Algorithm 47 applicable to find maximum matching in a general weighted graph D with weights greater than 1? Discuss the possible modification that need to made to the general graph D before we can use it as input for the Algorithm. y) Multiply the weight of edge (x. Problem 6. for each edge (x. What is the time complexity of Algorithm 47 and the previous algorithm described for unweighted graphs in the last section? Note that we are applying a simple path finding algorithm for un-weighted edges in the previous algorithm of the last section and its time complexity was O(p + q).8.2. 1 2 3 4 5 Transform the bipartite graph G into a directed graph D after adding vertices s and t to G according to the rules already defined. Would we be able to get the maximum matching in a bipartite graph if the graph is neither complete nor balanced by applying Algorithm 47? Problem 6. input : A complete balanced (binary) weighted bipartite graph. Find a longest path P from vertex s to t with a weight equal to or larger than zero. Problem 6. Algorithm 47: Find a maximum weighted matching in a complete balanced bipartite graph G.8. y) of path P in graph D do Reverse the direction of edge (x.1. Connectivity and Matching Problems efficient but useful algorithm) which works on the directed graph D and outputs the maximum weighted matching in the bipartite graph.8. the modifications should be such that the maximum matching in the new bipartite is equal to the maximum matching in the original graph D. the Bellman-Ford algorithm is applied and its time complexity is O(p3 ) and again it is applied p number of times on the given network graph. also it is applied p (number of vertices) times on the graph until all the paths have been found.3. output: A maximum weighted matching in G. .8.

by adding dummy vertices s and t and add zero weight edges from s to A partite and B partite to t.7.4: We convert a bipartite graph into a maximum edge-disjoint path problem. .The Matching Problem in Bipartite Graphs 331 a1 1 b1 1 a1 1 b1 1 a2 A a3 1 b2 B b3 s a2 a3 1 b2 b3 t 1 1 Directed Graph F a1 1 b1 1 a1 1 b1 -1 s a2 a3 1 b2 b3 t s a2 a3 1 b2 b3 t 1 1 Weight of the Longest Path = +1 a1 1 b1 -1 a1 1 b1 1 s a2 a3 1 b2 b3 t s a2 a3 -1 b2 b3 t 1 -1 Weight of the Longest Path = +1-1+1=+1 a1 1 b1 1 a1 1 b1 1 s a2 a3 -1 b2 b3 t A a2 a3 1 b2 B b3 -1 1 Weight of the Longest Path = 0 Figure 6. we then find the longest path in terms of weight and reverse the edges of that path.

b2 . that we describe.5: A complete balanced bipartite graph G is shown with positive edge weights. . assume that vertices in the A partite are numbered as a1 . 6.3 Maximum Weighted Matching in Complete Weighted Bipartite Graphs Problem: We need to find a maximum. We define layers in the graph G: layer 1 contains a1 and b1 . 1. while vertices in the B partite are numbered as b1 . The weight of the maximum matching is equal to 15 + 8 + 16 = 39. Output: A maximum weighted perfect matching.5. The graph consisting of first x layers contains first x vertices from both the partitions A & B.7. layer x contains vertices ax and bx . . See Fig. Graph consisting of First 2 Layers a1 6 13 15 b1 B a1 6 13 15 b1 A a2 10 4 8 16 17 9 b2 a2 10 4 8 17 b2 16 a3 b3 a3 9 b3 Weight of Maximum Matching is 15+16+8 = 39 Graph consisting of first two layers is shown in circle Figure 6. Input: A weighted complete & balanced bipartite graph G with positive edge weights. bk . and all edges between these vertices as shown in Fig. it will be interesting to make a .7. perfect matching in a weighted complete and balanced bipartite graph G. Maximum weighted matching is shown in red color. The partition A as well as the partition B is indicated in the left diagram. All edge weights are positive. weighted. The algorithms. Connectivity and Matching Problems 6. . .332 Network Flows. a3 . The first two layers of bipartite graph G are shown in the right diagram. ak . are similar in some respects and different in others. a2 .7. A Comparison of Two Algorithms: We shall describe two algorithms to solve the above problem. b3 .

6. 48 requires us to find the maximum weighted matching in the first x layers. 2: Given a maximum weighted matching of size x in G it finds the maximum weighted matching of size x + 1 in G. 6. 49 efficiently we need to design an efficient procedure which performs the following function: Procedure No. In order to implement this algorithm efficiently we need to design an efficient procedure which performs the following function: Procedure No.7.6.The Matching Problem in Bipartite Graphs comparison between the two. Algorithm 49: Find maximum weighted matching in a complete bipartite graph G input : A weighted complete bipartite graph G output: maximum weighted matching 1 2 for x = 1 to k do Find maximum weighted matching of size x in bipartite graph G. 333 Algorithm 48: Find maximum weighted matching in a complete bipartite graph G input : A weighted complete bipartite graph G output: maximum weighted matching 1 2 for x = 1 to k do Find maximum weighted matching in the first x layers of bipartite graph G. Similarly in order to implement Algorithm No. Before discussing the details of the two procedures we first need to transform the weighted bipartite undirected graph G into a directed graph D with . Do not move forward before understanding Fig. 1: Given a maximum weighted matching in the first x layers of G it finds the maximum weighted matching in the first x + 1 layers of G. Algorithm No. It is obvious from these diagrams that the intermediate results may be different but the end results are same for the two algorithms. The working of the two algorithms is shown in Fig.7. 6.

7.6: The left diagrams show the working of Algorithm No. 49 when applied to the same bipartite graph G. . Connectivity and Matching Problems Maximum Matching in the first layer: Value = 13 Maximum Matching of size one: Value = 17 a1 6 13 15 b1 Different a1 6 13 15 b1 a2 10 4 8 16 9 17 b2 a2 10 4 8 16 9 17 b2 a3 b3 a3 b3 Maximum Matching in the first two layers: Value = 6+15 = 23 Maximum Matching of size 2: Value = 15+16=31 a1 6 13 15 b1 a1 6 13 15 b1 a2 10 4 8 16 9 17 b2 Different a2 8 4 b2 16 a3 b3 a3 9 17 b3 Maximum Matching in all layers: Value = 8+16+15 = 39 Maximum Matching of size 3: Value = 8+15+16=39 a1 6 13 15 b1 a1 6 13 15 b1 a2 10 4 8 16 9 17 b2 Same a2 10 4 8 16 9 17 b2 a3 b3 a3 b3 Figure 6. 48 when applied to a bipartite graph G. The right diagrams show the working of Algorithm No. Please note that intermediate results may be different but the end result is the same for the two algorithms.334 Network Flows.

Algorithm 50: Procedure 1: Transform maximum weighted matching in the first x layers into a maximum weighted matching in the first x+1 layers in a bipartite graph input : Maximum weighted matching in the first x layers of bipartite G output: Maximum weighted matching in the first x + 1 layers of bipartite G 1 2 3 Transform the weighted bipartite graph into a directed graph D. 2. b) in bipartite graph G add a direction going from vertex b to a in the directed graph D. For every unmatched edge (a. b) in G add a direction going from vertex a to b in the directed graph D.7. We shall now discuss the details of Procedure No.The Matching Problem in Bipartite Graphs 335 additional vertices s and t according to the following rules as shown in the diagram below.1. Claim 6. 4. Now add another vertex t such that there is a directed edge from every unmatched vertex in B to vertex t with a weight equal to 0. Add vertex s to G such that in the resulting graph D there is a directed edge from vertex s to every unmatched vertex in A with a weight equal to 0. 3. The weight of the longest path from vertex s to t in D is equal to the gain in the weight of maximum matching when the size of matching is increased from x to x + 1 in the bipartite graph G. Find a longest path from vertex s to t passing through the first x + 1 layers of D and now reverse the edges in this path. . 1 & 2. 1. Note that vertex a belongs to A while b belongs to B in the bipartite graph G. The sign of every weight w for every matched edge in G is changed to a minus sign in the directed graph D. Every directed edge from B partite to A partite in D corresponds to a maximum weighted matching in the first x + 1 layers of G. All other edge weights retain their original signs in D as there are in G. For every matched edge (a.

7: The diagrams show the working of Procedure No. Maximum weighted matching in the first two layers in a bipartite graph G is converted into a maximum weighted matching in the first three layers of the same bipartite graph. 1.7. .336 Network Flows. Connectivity and Matching Problems Maximum Matching in First 2 layers a1 6 Transform -10 5 b1 -4 8 17 a1 6 10 5 b1 4 8 17 a2 s 16 b2 10 t b3 a2 16 b2 10 a3 9 a3 9 b3 a4 b4 a1 6 -10 5 b1 -4 8 17 a4 b4 a2 s 16 b2 10 25 t b3 a3 a1 6 10 5 9 b1 a4 a1 b2 10 17 -6 10 5 b4 a2 16 4 8 b1 4 8 a3 a2 b3 s -16 b2 10 9 t 25 a3 a4 b4 Transform -17 9 b3 Maximum Matching in First 3 layers a4 b4 Figure 6.

Find a longest path from vertex s to t in D and now reverse the edges in this path. 2.The Matching Problem in Bipartite Graphs 337 Algorithm 51: Procedure 2: Transform maximum weighted matching of size x into a maximum weighted matching of size x + 1 in a bipartite graph input : Maximum weighted matching of size x in bipartite graph G output: Maximum weighted matching of size x + 1 in bipartite graph G 1 2 3 Transform the weighted bipartite graph into a directed graph D. .8: The diagrams show the working of Procedure No. Maximum Matching of size One a1 6 13 15 b1 a1 6 13 15 b1 s a2 10 4 8 16 -17 9 b2 t a2 10 4 8 16 9 17 b2 Transform a3 b3 b3 a1 6 13 15 a3 b1 14 Convert s a2 10 4 8 16 9 -17 b2 t a1 6 13 15 b1 a1 13 6 Transform -15 a3 b1 b3 Find Longest path a2 8 4 b2 16 a3 9 17 s a2 10 4 8 -16 17 9 b2 t b3 a3 Maximum Matching of size Two b3 Reverse Longest path Figure 6. Every directed edge from B partite to A partite in D corresponds to a maximum weighted matching of size x + 1 in G.7. Maximum weighted matching of size 1 in a bipartite graph G is converted into a maximum weighted matching of size 2 in the same bipartite graph.

9: The diagrams illustrates that every path from s to t in the directed graph D passing through the first three layers corresponds to a matching in the first three layers of the bipartite graph G.338 Network Flows. Connectivity and Matching Problems a1 6 10 5 b1 a2 l= tia Ini 14 4 8 17 b2 10 16 a3 6 9 b3 b4 a1 6 -10 5 b1 16 a4 8 17 -4 8 a1 6 10 5 b1 4 8 17 4 a2 16 b2 10 s 0 t 9 Increase by 9 a2 16 b2 10 a3 9 b3 a3 a4 6 9 b3 b4 -10 6 5 a1 b1 -4 8 17 16 a1 8 10 6 5 b1 4 8 17 a2 16 b2 10 12 s 0 t 21 Increase by 21 a2 16 b2 10 a3 9 b3 a3 a4 6 9 b3 b4 -10 6 5 a1 b1 16 -4 8 17 12 a1 6 10 5 b1 4 8 17 8 a2 16 b2 10 a2 t Increase by 25 16 b2 10 s 0 25 a3 9 b3 a3 a4 9 b3 b4 a4 b4 Figure 6. The longest path corresponds to a perfect matching of maximum weight. .7.

7. 6.7. 6 a1 6 -10 5 b1 16 -4 8 17 12 6 a1 6 -10 5 b1 16 -4 8 17 8 8 a2 16 b2 10 4 a2 16 b2 10 s 0 t 25 s 0 t 9 a3 9 b3 a3 9 b3 a4 b4 a1 6 -10 5 a4 b1 -4 8 17 b4 a2 s 16 b2 10 t b3 e rs ve Re Ed s ge a3 -9 a4 b4 Figure 6.7. 6.7. Assume that we have already found a maximum matching in the first x layers of graph G.9. Path 1 is the longest path and has a weight equal to y and Path 2 is a relatively shorter path and has a weight equal to z. If we select Path 1 (as dictated by Procedure No. 6. If on the other hand we reverse edges in Path 2 then a cycle with a net positive value at most equal to y − z will be formed as shown in Fig. We transform the bipartite graph into a directed graph also shown in Fig.7.10. we now need to extend this maximum matching in the first x + 1 layers of G as shown in Fig. Then the reversed edges of Path 1 plus edges belonging to Path 2 will form a cycle in the graph with a net negative value at most equal to y − z.9.The Matching Problem in Bipartite Graphs 339 Claim 6. Suppose we have found two paths (one longest and one relatively shorter) from vertex s to vertex t. 1) then we need to reverse edges belonging to Path 1.10: The top left diagram shows the longest path from s to t in D while the right diagram shows a relatively shorter path. .2. If we reverse edges of the shorter path then a cycle of net positive value will be formed.

Problem Set 6. Please comment. We are working under the assumption that there are no positive weight cycles in the directed graph D. Please note that there are 2k + 2 vertices in this directed graph and your answer should be independent of edge weights in the graph. 6. Assume that we are given a complete balanced weighted bipartite graph with positive edge weights. The left diagram of Fig. 2). 2).4.9.13 shows a directed graph in which there is only one edge going from a vertex b to a vertex a. Connectivity and Matching Problems Claim 6. Problem 6. .7.1. Assume that we have found a 3-edge longest path from vertex s to vertex t in a directed graph consisting of the first x + 1 layers of graph D (see right diagram of Fig. be present in the directed graph D. Design an efficient algorithm to solve this problem. Problem 6. Describe an efficient algorithm to find a maximum matching in the very first layer of the bipartite graph.1 or Procedure No.13).3.7. We shall never encounter a situation where there will be a positive weight cycle in D (provided we follow steps given in Procedure No. Problem 6.7. In Procedure No. 6.9.9. Negative weight cycles will.9. Carefully derive the time complexity of this algorithm. Please note that there are 2k + 2 vertices in this directed graph and your answer should be independent of edge weights in the graph.2. Discuss briefly if your algorithm is a greedy algorithm or does it use dynamic programming. we need to find a longest path from vertex s to vertex t. Someone thinks that the longest path problem is NP-Complete.13 shows a directed graph in which there are x edges going from a b vertex to an a vertex. Problem 6.5. The right diagram of Fig.7. however. 6. Problem 6.340 Network Flows.9. We need to find what will be the maximum length (in terms of number of edges) of the longest path passing through the first x + 1 layers of the graph. Carefully derive the time complexity of this algorithm in terms of the size of the problem.9. 1 (and Procedure No. We need to find what will be the maximum length (in terms of the number of edges) of the longest path passing through the first two layers of the graph.3. Now we need to find a 5-edge longest path from s to t in the same graph consisting of first x + 1 layers of graph D.

The Matching Problem in Bipartite Graphs 341 Finding Maximum Matching in the First layer a1 0 6 10 5 b1 4 8 17 10 a2 s 16 b2 10 t b3 a1 10 a3 9 b1 a4 b4 a1 6 Finding Maximum Matching in First 2 layers 10 5 b1 4 -5 a2 a1 6 0 -10 5 0 b1 4 8 17 5 -5 b2 a1 6 -10 5 b1 4 8 17 5 a2 s 16 b2 10 4 a2 t s 16 b2 10 4 t b3 a1 6 10 5 b1 4 8 17 a3 9 b3 a3 9 a2 a4 b4 a4 b4 16 b2 10 a3 9 b3 6 a1 6 -10 5 b1 16 -4 8 17 8 Finding Maximum Matching in the First 3 layers 6 a1 6 -10 5 b1 16 -4 8 17 12 4 a2 16 b2 10 6 a1 6 -10 5 b1 -4 8 17 16 8 a2 16 b2 10 s 0 t 9 8 s a2 16 0 t 25 a3 9 b3 s b2 12 10 0 a3 t 9 b3 a4 b4 a3 9 21 b3 a4 b4 a4 b4 Figure 6.7. . 48 on a bipartite graph which is already been transformed into a directed graph D.11: The diagrams show the working of Algorithm No.

49 on a bipartite graph which is already been transformed into a directed graph D.342 Network Flows. .12: The diagrams show the working of Algorithm No. Connectivity and Matching Problems Finding Maximum Matching of size one a1 6 13 15 b1 15 a1 6 13 15 b1 s a2 10 4 8 16 9 17 b2 8 t a2 10 4 8 16 9 17 b2 a1 6 13 15 17 a3 b3 a3 b3 b1 a2 Finding Maximum Matching of size 2 8 13 6 8 15 4 b2 16 0 a1 6 13 15 b1 13 0 a1 b1 8 14 a3 9 17 b3 s -1 a2 10 4 8 16 9 -17 b2 16 t s -1 a2 10 4 8 16 9 -17 b2 16 t a1 6 13 15 b1 0 a3 b3 0 a3 b3 a2 10 4 8 16 9 17 b2 a3 Finding Maximum Matching of size 3 b3 -7 a1 6 13 -15 b1 10 -4 a1 6 13 -15 b1 10 s -5 a2 10 4 8 17 b2 8 t a1 -4 6 13 -15 b1 10 s -5 a2 10 4 8 17 b2 8 t -16 9 9 -16 12 9 a3 b3 s a2 -5 10 4 8 17 b2 8 t a3 b3 -16 12 9 a3 b3 Figure 6.7.

49 (see Fig. Carefully derive the time complexity of this algorithm. Assume that the time complexity of Procedure No.9.9. 1 is bounded by O(x · k 2 ).7.11.14. The right diagram shows another directed graph D where there are x edges going from a vertex b to a vertex a. Problem 6.9. 2.9.11).9. Problem 6.7. Find the actual longest path in each of the directed graphs shown in Fig. Problem 6.7. 2 and then the time complexity of Algorithm No.7. 48. Carefully derive the time complexity of Procedure No.9.8.10. Is it possible to use an existing textbook algorithm (without any modification) in order to solve the previous problem? Discuss briefly. 6. Carefully derive the time complexity of Procedure No.The Matching Problem in Bipartite Graphs Only 1 edge is reversed 343 a1 6 -10 5 b1 4 8 17 First layer Only x edges are reversed a1 6 -10 5 b1 First x Layers 8 17 -4 a2 s 16 b2 10 a2 t s 16 b2 10 t b3 a3 9 b3 a3 9 ak bk ak bk Figure 6. 6. Problem 6. What would be the maximum value of the length (in terms of number of edges) of the longest path from vertex s to vertex t in D? Discuss briefly. Consider the working of Procedure No. Problem 6. Problem 6. 48 (see Fig. Problem 6. .13: The left diagram shows a directed graph D in which there is only one edge going from a vertex b to a vertex a.12. Assume that we are given a complete balanced weighted bipartite graph with positive edge weights. Now derive the time complexity of Algorithm No.6.9.9.12). 6.7. Describe an efficient algorithm to find a maximum matching of size 1 in this bipartite graph. 1 and then the time complexity of Algorithm No.

48 or 49 to solve this problem? How can you design a better algorithm to solve this problem? Discuss briefly.15. Problem 6.7.13. Discuss if we can use Algorithm No. Connectivity and Matching Problems b1 a1 6 13 15 b1 15 s a2 10 4 8 16 -17 9 b2 t s a2 10 -17 8 16 4 9 b2 8 t ak bk ak bk a1 6 13 15 b1 a1 6 13 15 b1 s a2 10 -4 8 16 17 9 b2 t s a2 10 -4 8 16 17 -9 b2 t ak bk ak bk Figure 6. Discuss briefly. Discuss if we can use Algorithm No.9.48 and 49 may be the same in terms of the Big O notation yet one algorithm is considerably faster than the other. Suppose we need to find a best possible matching of size x in a given complete balanced weighted bipartite graph G of size k and x is much smaller than k. Suppose we need to find a best possible matching for the first x members of partite B in a given complete balanced weighted bipartite graph G.344 a1 6 13 15 Network Flows. 48 or 49 to solve this problem? How can you design a better algorithm to solve this problem? Discuss briefly. Discuss how you will solve this problem and carefully derive the time complexity of your algorithm in terms of x and k.14. Suppose we need to find a best possible matching for the first x members of partite A in a given complete balanced weighted bipartite graph G. Problem 6.9. Problem 6. .9. Problem 6.9.14: We need to find longest path from vertex s to t in the directed graphs shown here.16. Although the worst case time complexity of Algorithm No.

8.8. The problem is to efficiently find maximum flow at minimum cost in a given network graph. 6. 1/8 2/8 2/8 a 2/0 b 1/4 3/5 8/2 1/0 2/0 a 1/4 6/9 1/4 1/5 3/5 b 8/2 1/0 1/2 2/0 s 1/0 6/9 t 2/0 s 2/0 t c 1/9 d c 1/9 d The Max-Flow Min-Cost Problem (Category 1) Maximum Flow (3 units) at Minimum Cost (8+5+4+2) Figure 6.2 Finding a Maximum Flow or Finding a Shortest Path? Before answering the above question it is important to revise the relevant prior knowledge. We also need to make sure that the maximum flow is taking place at minimum cost. 6.The Max-Flow Min-Cost Problem 345 6. and also remind ourselves of some of the powerful techniques which will be useful in solving this problem.8 6.8. it is indicated with each edge in the figure below. We need to push the maximum flow in this network graph (starting from vertex s and ending at vertex t) at a minimum cost. The capacity/cost of an edge is shown along with each edge. A possible solution of this network graph is shown in the right diagram of the same figure.1: We show a directed graph D (left diagram) having two special vertices s and t. We assume that the capacity as well cost per unit flow through every edge is a positive integer.8.1 The Max-Flow Min-Cost Problem Introduction We consider the Maximum Flow at Minimum Cost problem in a network graph as shown in Fig. we are pushing a maximum flow of 3 units at a total cost of 19. The maximum flow at minimum cost is indicated by colored lines in the right diagram.1. We need to find maximum flow coming out of vertex s and being absorbed by vertex t. It is interesting to note .

8. If we start minimizing the cost by finding shortest paths. Before we solve this general problem we shall try to reflect on similar problems we have already solved and what are some of the special cases of this general problem which can be resolved using our prior knowledge? We shall then extend or modify specialized solutions to solve this general problem. then how will we be able to tackle the problem of maximizing the flow? Different edges have different capacities. 2. the sum of edge costs in all edge-disjoint (shortest) paths should be as small as possible. So let us list down the general problem (once again) and its special cases: Category 1: Given a network flow graph as shown in Fig. i. 6. Category 3: The network flow graph is derived from a complete balanced and weighted bipartite graph. Category 2: Given a network flow graph as shown in Fig. finding shortest paths without considering capacities will create complications. Please note that this problem is equivalent to finding the maximum edge-disjoint paths from vertex s to vertex t at minimum cost. and this somehow looks like finding a shortest path from vertex s to t (we are aware of a number of simple and efficient algorithms to solve shortest path problems). 6.1 how can we find maximum flow at minimum cost from a source vertex s to a sink vertex t? We assume that edge capacities are integers while per unit cost of flow through any edge may be any non-negative real number. Connectivity and Matching Problems that this challenging problem (which we call a Category 1 problem) has two requirements: 1.8.e. We know how to solve the above two problems in isolation but how can we fulfill the two requirements simultaneously? If we start finding flows without looking at costs then we may end up with a maximum flow but at higher cost. Here we add a source vertex s to the A and .2 how can we find the maximum flow at minimum cost from a source vertex s to a sink vertex t? We assume that the capacity of each edge is exactly 1 while per unit cost of flow through any edge may be a real number. We need to maximize the flow and we already know how to do it (but we should also remember our shortcomings and limitations).346 Network Flows. It is certainly a very exciting mixture of two important problems. We need to minimize the cost of (unit) flow starting from vertex s and ending at vertex t.

2: A Category 2 problem is equivalent to finding maximum edgedisjoint paths at minimum cost as shown in the top diagram.The Max-Flow Min-Cost Problem 347 a 1/0 1/9 b 1/0 1/5 1/2 s 1/0 1/4 1/9 t 1/0 Equivalent to c 1/3 d The Problem of finding Max edge-disjoint paths at Min Cost Category 2 The Max-Flow Min-Cost Problem Binary Costs a 1/0 1/0 1/0 1/0 1/1 b 1/0 1/1 1/0 1/1 s c 1/0 d Special case of Category 2 The Max-Flow Min-Cost Problem When costs are binary Figure 6. A special case of this category is shown in the bottom diagram An i nt p r o er es t bl e i n g m t Reduced to .8.

It is possible to recognize that finding maximum flow (or maximum edge-disjoint paths) at minimum cost in the right diagram is equivalent to finding a minimum cost perfect matching in the bipartite graph shown in the left diagram. Here the problem is to find maximum flow at minimum cost in the network flow graph shown in the right diagram. We need to recall our expertise of finding maximum cost perfect matching in a weighted bipartite graph discussed in earlier sections. All edge capacities are 1. 6.3. Both problems have their applications in graph theory and elsewhere. costs not shown are equal to 1 Category 3: The network flow graph is derived from a bipartite graph.8. A curious reader might have noticed that while solving the maximum cost perfect matching problem in the bipartite graph shown in Fig. Connectivity and Matching Problems a sink vertex t to B as shown in Fig. 6. 6. All edge weights associated with source vertex s and sink vertex t are zero 6. We assume that the capacity of each edge is one in the network flow graph.8.3 we were essentially solving the maximum flow at maximum cost problem as shown in a network flow graph in Fig. Similarly while we were solving the minimum weighted perfect matching .348 Network Flows. a1 6 13 15 b1 a1 6 13 15 b1 a2 4 8 16 17 9 b2 s B a2 10 4 8 16 17 9 b2 t A 10 a3 b3 a3 b3 A Bipartite graph with a weight or cost on each edge.8.8. A related problem in a bipartite graph would be to find a maximum cost perfect matching .that would require us to find maximum flow (or maximum edge-disjoint paths) at maximum cost.4. Figure 6.3: In a Category 3 problem a network flow graph is derived from a weighted bipartite graph by inserting a source vertex and a sink vertex.8.3 Category 3 network flow Problems We shall now try attacking these problems starting from Category 3.

The step by step working of this algorithm is shown in Fig.8.5. 6. The algorithm terminates when it is no longer possible to find a path from the source to the sink as shown in the bottom right diagram of Fig. We have in fact solved the Category 3 network flow problem without explicitly saying so as our primary objective was to find a perfect matching of maximum (or minimum) cost.8. problem in a bipartite graph we are essentially solving the minimum cost maximum flow problem. The problem is how to recover maximum flow at minimum cost from this (final) graph F . Maximum Flow at Maximum Cost in Category 3 problems Please recall Algorithm 47 which was designed to find a maximum weighted matching in a complete bipartite graph. 6. We reverse the direction of each edge in path P and also multiply weight of each edge in the path by negative 1.4: A maximum weighted perfect matching in bipartite graph (left diagram) corresponds to a maximum flow at maximum cost in the right diagram. . What we essentially do here is to find a longest path P from vertex s to vertex t in the given graph.3. 6. We apply this algorithm to the network flow graph shown in the right diagram of Fig. It is slightly modified as shown below (Algorithm 52).5.The Max-Flow Min-Cost Problem 349 a1 6 13 15 b1 a1 6 13 15 b1 a2 4 8 16 9 17 b2 s B a2 10 8 4 16 b2 t A 10 a3 17 9 b3 a3 b3 Maximum Weighted Perfect Matching in a Bipartite Graph Category 3: Maximum Flow at Maximum Cost Figure 6.8.8. We then again find a longest path from the source vertex to the sink vertex in the modified graph. As we shall show later this algorithm is powerful enough to handle Category 2 and 1 network flow problems.

Find a flow of one unit through a longest path P (in terms of edge costs) from vertex s to t in F . y) with a negative sign and go back to Step 2 The solution to this problem is again consistent to our earlier approach of deleting unused edges from the network flow graph. The unused edges in final graph F are those which have positive weights as shown in Fig. The only change that we need to do is to replace longest path in line 2 of this algorithm by shortest path as shown in the following algorithm (Algorithm 53). But in order to do that we need . output: Maximum Flow at minimum cost from s to t in D 1 2 3 4 5 Copy Graph D into F . One would like to compare this answer with the one obtained while finding a maximum weighted perfect matching in a bipartite graph shown in Fig. We shall first show that Algorithm 53 can be used solve these problems without any modification. Maximum Flow at Minimum Cost in Category 3 Problems The above algorithm can easily be adopted to find maximum flow at minimum cost or in other words maximum edge-disjoint paths at minimum cost in a Category 3 network flow graph. for each edge (x.6. We intend to solve the maximum flow at minimum cost problem using multiple algorithms in order to provide a better insight to the problem and its possible solutions. 6. y) Multiply the weight of edge (x. and vertices s & t. 6.6.8. y) of path P in graph F do Reverse the direction of edge (x.4 Category 2 (and 1) network flow Problems Here we shall consider Category 2 (and Category 1) network flow problems. Connectivity and Matching Problems Algorithm 52: Find a Maximum Flow at maximum cost from vertex s to vertex t in a directed network graph D belonging to Category 3 input : Directed & Weighted graph D.350 Network Flows. 6.8. By deleting these edges it is possible to find maximum flow at maximum cost in a Category 3 problem as shown in Fig.8. If you are successful in finding a flow then keep a record of the path found otherwise exit with graph F as output.4. 6.8.

6.5: The step by step working of Algorithm 52 is shown here on the network graph of Fig. Please note that in this graph it is no longer possible to find another path from the source vertex to the sink vertex.8. .The Max-Flow Min-Cost Problem 351 Find a longest path from s to t a1 6 13 15 Reverse the edges of the longest path a1 6 13 15 b1 b1 s a2 10 4 8 16 17 9 b2 t s a2 10 4 8 16 9 -17 b2 t a3 b3 a3 b3 Find another longest path from s to t a1 6 13 15 Reverse the edges of the longest path a1 6 13 -15 b1 b1 s a2 10 4 8 16 -17 9 b2 t s a2 10 4 8 17 9 b2 -16 t a3 b3 a3 b3 Find another longest path from s to t a1 6 13 -15 Reverse the edges of the longest path a1 6 13 -15 b1 b1 s a2 8 10 4 b2 -16 t s a2 10 -8 4 b2 -16 t a3 17 9 b3 a3 17 9 b3 Final graph F: No more paths left from the source to the sink vertex Figure 6.8.3. The final graph F is shown in the bottom right corner.

352 a1 6 13 -15 Network Flows. Find a flow of one unit through a shortest path P (in terms of edge costs) from vertex s to t in F . If you are successful in finding a flow then keep a record of the path found otherwise exit with graph F as output. y) with a negative sign and go back to Step 2 . and vertices s & t.6: Final graph F taken from the last figure is shown in the top left diagram. Algorithm 53: Find a Maximum Flow at minimum cost from vertex s to vertex t in a directed network graph D belonging to Category 3 input : Directed & Weighted graph D. y) of path P in graph F do Reverse the direction of edge (x. We can recover maximum flow at maximum cost by removing edges with positive weights in this final graph F . Connectivity and Matching Problems b1 a1 -15 b1 s a2 10 -8 4 b2 -16 t Remove +ive edges from graph F s a2 -16 b2 t a3 17 9 -8 b3 a3 b3 Final graph F: No more paths left from the source to the sink vertex Maximum Flow at Maximum Cost in the apposite direction Re gra mov ph e +i F f ve rom ed gra ges i ph n D a1 15 b1 s a2 8 16 b2 t a3 b3 Maximum Flow at Maximum Cost in the normal direction Figure 6. y) Multiply the weight of edge (x. output: Maximum Flow at minimum cost from source vertex s to sink vertex t in D 1 2 3 4 5 Copy Graph D into F .8. for each edge (x.

Now we claim that as directed graph D does not contain any negative weight cycle then graph F will also not contain any negative weight cycle.2. Now we claim that as directed graph D does not contain any negative edges therefore graph F will not contain any negative weight cycle. The value of net weight in this cycle is -2.8.8. By redirecting the flow in the negative weight cycle it is possible to reduce the cost of flow by an amount exactly equal to 2 as shown in the bottom diagram of this figure. We copy this graph in graph F . We also multiply the weight of each edge in path P by negative one in graph F . We now reverse the direction of each edge in path P in graph F . We are given a directed & weighted graph D with no negative weight cycles but it may contain negative weight edges. We now reverse the direction of each edge in path P in graph F . We find a shortest path P from source vertex s to the sink vertex t in directed graph F . It contains a source vertex s and a sink vertex t. Given a network flow graph D. It contains a source vertex s and a sink vertex t.8. We then find a negative weight cycle in this new graph F highlighted by orange color.8. We reverse the direction of the edges used by the flow and multiply the cost of these edges by negative 1 as shown in graph F (see top right diagram).7. Claim 6. Please note that this flow of one unit is not taking place on a shortest path from vertex s to vertex t. Negative Weight Cycles & Improvement in Cost of Flow We show a flow of one unit in the network D shown in the top left diagram of Fig.8.The Max-Flow Min-Cost Problem 353 to do some serious graph theoretic work in terms of claims and some hints for their proofs. If there is a negative weight cycle in graph F with a net weight equal to -k (graph F is derived from a network graph D after reversing the edges in the direction of the flow and multiplying the costs in . We copy this graph in graph F .8. If on the other hand there are no negative weight cycles then it is not possible to reduce the cost of existing flow. Claim 6.1.3. We also multiply the weight of each edge in path P by minus one in graph F . Such a scenario is shown in Fig. The graph D may contain cycles but it does not contain any negative weight edge. 6. 6. We find a flow of one unit from vertex s to vertex t. We find a shortest path P from source vertex s to the sink vertex t in directed graph F . Claim 6. We are given a directed & weighted graph D.

354 Network Flows. . Connectivity and Matching Problems 2/8 a 2/0 b 1/0 2/0 1/4 3/5 8/2 2/8 a 3/5 6/9 1/0 1/4 b 1/0 8/2 s 1/0 6/9 t 2/0 s t 1/0 c 1/9 d c 1/-9 d One edge-disjoint path or a flow of 1 unit in a Category 1 Problem An Improvement is possible as there is a negative cycle 2/8 Reverse the edges in the path shown earlier There is a negative cycle of value = -2 a 2/0 b 1/0 1/4 3/5 8/2 s 1/0 6/9 t 2/0 c 1/9 d Cost can be improved by 2 units by redirecting the flow in the direction of negative cycle Figure 6.7: We show a flow of one unit in the network shown in the top left diagram. If there is a negative weight cycle (after reversing the edges in the direction of the flow) then the cost of flow can be further reduced.8.

Given a graph D and a finite flow taking place from vertex s to vertex t.8. Claim 6. . If (after reversing the edges in the direction of the flow) there are no negative weight cycles then the cost of flow can not be further reduced.9. We know the path taken by each unit of flow from vertex s to vertex t in D. we keep a record of the length of each such path and call these paths existing paths in D. If there is no negative weight cycle in graph F (derived from a network graph D after reversing the edges in the direction of the flow and multiplying the costs in these edges by negative 1) then it is not possible to reduce the cost of the existing flow by re-adjusting the flow in any direction. Claim 6.8. If there is no improvement possible in the cost of existing flow then there will be no negative weight cycle in graph F (derived from a network graph D after reversing the edges in the direction of the flow and multiplying the costs in these edges by negative 1).5.The Max-Flow Min-Cost Problem 355 these edges by minus 1) then it is possible to reduce the cost of the existing flow by k by re-adjusting the flow in the direction of the negative cycle. 6.8.8: We show a flow of two units in the network shown in the top left diagram. 2/8 2/8 a 2/0 b 1/0 1/4 3/5 8/2 1/0 1/0 a 6/9 1/-4 1/-5 2/5 b 8/2 1/0 s 1/0 6/9 t 2/0 s 1/0 t 2/0 c 1/9 d c 1/9 d A flow of 2 units in a Category 1 Problem Reverse the edges in the paths shown earlier 2/8 1/0 1/0 2/8 a 6/9 1/-4 1/-5 2/5 b 8/2 1/0 1/0 1/0 a 1/-5 6/9 1/-4 2/5 b 8/2 1/0 s 1/0 t s 1/0 t 1/0 c 1/9 d The cycle is of net positive weight en t a em l e f o r v 1/0 p ro si b ni ts o Imis pos 2 u N t of os ow n C ed f l i x fi c 1/9 d The cycle is of net positive weight Figure 6.8. Please see Fig.4.

9: We show two existing paths from s to t in graph D (top left diagram). . The paths are reversed in graph F as shown in the top right diagram. Connectivity and Matching Problems 2/8 2/8 a 2/0 b 1/0 1/4 3/5 8/2 1/0 1/0 a 6/9 1/-4 1/-5 2/5 b 8/2 1/0 s 1/0 6/9 t 2/0 s 1/0 t 2/0 c 1/9 d Convert c 1/9 d A flow of 2 units in graph D The red path has length equal to 4.8. the green path has length equal to 5 Graph F: we have reversed the edges in the direction of the flow in graph D and have multiplied the edge costs by -1 If there are no improvement possible in existing flow in graph D then there will be no negative cycles in F Figure 6.356 Network Flows.

output: Maximum Flow from s to t in D at Minimum Cost 1 2 3 4 Copy graph D into a graph F .The Max-Flow Min-Cost Problem Finding Maximum Flow at Minimum Cost 357 In view of the above claims it is obvious that Algorithm 53 can be used to find maximum flow at minimum cost in Category 2 as well as Category 1 network flow problems. Problem 6. 6. the removal of a negative weight cycle will certainly reduce the cost of flow by at least one unit. now go to step 3. First we find maximum flow.8.10. We now describe an alternate algorithm (Algorithm 54) to find a maximum flow at minimum cost for Category 1 problems. Find Maximum Flow from vertex s to t in F (ignoring costs). The step by step detailed working of the above algorithm on a Category 1 problem is shown in Fig. Our last three claims support the argument that the following algorithm would be able to correctly solve maximum flow at minimum cost problem in Category 1 network flow graphs. If there are any negative weight cycles in the graph then we have to remove every negative weight cycle. We have demonstrated the working of an algorithm (to find maximum flow at minimum cost in Category 3 problems) in Fig.1. It will be interesting to derive the time complexity of that algorithm for both these categories. Reverse the edges in the path of every flow.10. Algorithm 54: Find Maximum Flow at Minimum Cost from s to t in a directed graph D in a Category 1 problem input : Directed & Weighted graph D.8.10. 6. The process is repeated until there are no more negative weight cycles left in the graph. multiply the cost of these edges with minus one. we then reverse the edges in the direction of the flow.11. Find if there is a negative weight cycle in graph F.8. and vertices s & t. The process of finding maximum flow at minimum cost may be accelerated if we some how find a negative weight cycle of higher value and remove it as shown in Fig. multiply the costs corresponding to these edges with negative 1. .4. Remove the negative weight cycle by adjusting the flow accordingly. Problem Set 6. If you find one then go to step 4 else output the maximum flow at minimum cost and exit. but it may give rise to another negative weight cycle. 6.

Connectivity and Matching Problems 2/8 2/ a 2/0 b 1/0 1/4 3/5 8/2 2/ a 6/ 1/ 3/ b 8/ 1/ s 1/0 6/9 t 2/0 s 1/ t 2/ c 1/9 d c 1/ d Integer Capacity Integer Cost in graph D 2 2 2/ Forget the Costs in D for the time being 2 2/ a 6/ 1/ 3/ b 8/ 2 1/ 2/0 2/8 a 6/ 1/4 3/5 b 8/2 1/0 s 1/ t 2/ s 1/0 t 2/0 c 1/ d c 1/9 d Find the Max Flow in D 2 0/ 2 0/ Find the Cost of Max flow 2×8+2×1+9×1=27 2/-8 0/8 a 6/ 1/ 3/ b 7/ 0/ 0/0 a 1/4 6/9 3/5 b 7/2 0/0 1/-2 0/0 s 0/ t 0/ s 0/0 t c 0/ d c 0/9 1/-9 d Reverse the Flow in graph F 2/-8 0/8 Reverse the Costs also in F 2/-8 0/8 0/0 a 1/4 3/5 6/9 b 7/2 0/0 1/-2 0/0 0/0 a 3/5 6/9 1/4 b 7/2 0/0 1/-2 0/0 s 0/0 t s 0/0 t c 1/-9 0/9 d c 1/-9 0/9 d There is a negative cost cycle of value = -8 in F There is a negative cost cycle of value = -6 in F Figure 6. We then remove negative weight cycles and subsequently reduce cost.358 Network Flows. We first find maximum flow ignoring costs (cost comes out to be = 27). .8.10: We show different steps in finding maximum flow at minimum cost in a Category 1 problem.

11: We show different steps in finding maximum flow at minimum cost in a Category 1 problem. .8. A maximum flow of 3 units can pushed in this network at an optimal cost of 8+4+5+2 = 19. The cost is thus reduced from 27 to 19 for the same amount of flow.The Max-Flow Min-Cost Problem 2/-8 359 0/8 2/-8 0/8 0/0 a 3/5 6/9 1/4 b 7/2 0/0 1/-2 0/0 0/0 a 3/5 6/9 1/4 b 7/2 0/0 1/-2 0/0 s 0/0 t s 0/0 t c 1/-9 0/9 d c 1/-9 0/9 d There is a negative cost cycle of value = 6 in F 2/-8 0/8 Circulate a unit flow in the direction of the negative cycle and the cost will decrease by 6 1/-8 1/8 0/0 1/-2 0/0 0/0 0/0 a 3/5 6/9 1/4 b 7/2 a 0/4 6/9 1/-4 3/5 b 8/2 0/0 s 0/0 t s 0/0 t 0/0 c 1/-9 0/9 d c 0/9 1/-9 d Reverse the direction of the edges in the Cycle in graph F 1/-8 1/8 0/0 Make corresponding adjustments in graph F 1/-8 1/8 a 0/4 6/9 1/-4 3/5 b 8/2 0/0 0/0 a 0/4 6/9 1/-4 3/5 b 8/2 0/0 s 0/0 t 0/0 s 0/0 t 0/0 c 1/-9 0/9 d c 0/9 1/-9 d A newly formed negative cycle Circulate a unit flow in this cycle 1/-8 1/8 0/0 Reverse the direction of the edges in the Cycle in graph F 1/8 2/8 a 0/4 6/9 1/-4 1/-5 2/5 b 0/0 2/0 a 1/4 6/9 1/4 1/5 3/5 b 8/2 1/0 1/2 2/0 s 0/0 7/2 1/-2 0/0 t s 1/0 t c 1/9 d c 1/9 d Make adjustments in graph F Now No more Negative Cycles Maximum Flow at Minimum Cost (8+5+4+2) in the original Graph D Figure 6.

. Connectivity and Matching Problems 2/-8 0/8 2/-8 0/8 0/0 a 1/4 3/5 6/9 b 7/2 0/0 1/-2 0/0 0/0 a 1/4 3/5 6/9 b 7/2 0/0 1/-2 0/0 s 0/0 t s 0/0 t c 1/-9 0/9 d c 1/-9 0/9 d There is a negative cost cycle of value = 8 in F Circulate a flow of 1 unit 1/-8 1/8 0/0 Reverse the direction of the edges in the cycle 1/8 2/8 a 0/4 6/9 1/-4 1/-5 2/5 b 0/0 2/0 a 1/4 6/9 1/4 1/5 3/5 b 8/2 1/0 1/2 2/0 s 0/0 7/2 1/-2 0/0 t s 2/0 t c d 1/9 c 1/9 d Make corresponding adjustments in graph F Maximum Flow at Minimum Cost in the original Graph G Figure 6.360 Network Flows.8.12: Removing a negative weight cycle of a relatively higher value may speed up the process of finding a maximum flow at minimum cost.

13: We show different steps in finding maximum flow at minimum cost in a Category 2 problem by removing negative weight cycles.The Max-Flow Min-Cost Problem 361 a 1/0 1/2 b 1/0 1/8 a b 1/9 s 1/0 1/7 t 1/0 s t c d c 1/5 d Category 2 Category 2: Forget Capacities & ignore Costs a b 0 a 2 b 0 -8 -9 -7 s d t c s 0 t 0 c 5 d Find max edge-disjoint paths Reverse the edges in the paths and multiply edge costs in the paths by -1 a 0 2 b 0 0 a -2 b 0 -9 9 -8 s 0 -7 0 t s 0 8 7 0 t c 5 There are -ive cost cycles d 2 c -5 Reverse the edges in the cycles d a 0 Max Edge-Disjoint Paths at Min Cost b 0 8 9 7 s 0 t 0 c 5 d Figure 6.8. Please note that in this category all edge capacities are equal to one .

14. We show an interesting special case of Category 4 in Fig. Is it possible to solve problems belonging to this category using Algorithm 53 even if the cost of a unit flow in an edge is a real number? Discuss briefly. 6. You may design a better algorithm if you want and if you can? a 4/0 ∞/8 ∞/3 ∞/7 b 5/0 s 3/0 t d 2/0 c ∞/6 Figure 6.2. The step by step working of this algorithm is demonstrated in Fig. . Problem 6. In this special category of graphs the sum of capacities of all edges coming out of s is equal to the sum of capacities of all edges going into vertex t.10.3.8. Problem 6. Discuss briefly. We need to find maximum flow at minimum cost using an efficient algorithm. Algorithm 54 can be used to find maximum flow at minimum cost in Category 1 problems. The cost of every edge coming out of s and going into t is zero.10. Is this possible to use this technique (without any appreciable change) to find the minimum cost for a fixed flowthe amount of fixed flow may not be the maximum flow in the network.8. The capacity of every edge in the D − s − t graph is infinite while the cost is a positive integer indicated in the diagram. The sum of capacities of all edges coming out of source s is exactly equal to the sum of capacities of edges going in sink t while the costs of these edges are zero.8. 6. Connectivity and Matching Problems Describe this algorithm (known as Algorithm 53) in your own words and carefully derive its time complexity.10.14: We show a special network graph D. Here edge capacities in graph D − s − t are all infinite.362 Network Flows. Carefully derive the time complexity of this algorithm. Find if you can use any earlier techniques (Algorithm 53) to solve this interesting problem.

Design a (very) efficient algorithm to find a maximum flow at minimum cost in this special category.8. applying Algorithm 54 seems to be overkill. In this category all edge capacities are equal to 1.8.12.15: We show that a Category 3 problem (where edges adjacent to s and t have non-zero cost) is as difficult (or as easy) to solve as a Category 3 problem where the s and t edges have zero costs (top diagram).8.4. . Some of these problems were addressed in the last problem set.6. 6. 6. we expect to solve it using a simple and a very efficient algorithm. The bottom diagram shows an interesting variation of Category 3 problem where we need to minimize cost not for a maximum flow but for a fixed flow in the network. On the other extreme Category 2 & 3 problems can be solved using very efficient algorithms. thus maximum flow here corresponds to maximum edge-disjoint paths as shown in Fig. We have applied Algorithm 54 to solve a Category 2 problem as shown in Fig.The Max-Flow Min-Cost Problem 363 Problem 6. Describe an efficient algorithm to handle Category 2 problems. Now consider a special case of this category where costs are expressed by binary numbers. We have not yet devised an exact algorithm to find maximum flow (even if we ignore costs) in such problems. 6.6. Because of its restricted nature. this special case is shown in the bottom diagram of Fig. 6.8. a1 1/0 1/7 b1 1/4 1/5 1/0 s 1/0 t b2 1/0 Source/sink edge costs are integers a1 1/3 1/7 b1 1/4 1/5 1/4 s 1/2 t b2 1/7 a2 1/3 a2 1/3 Equivalent Category 3 The Max flow Min-Cost Problem Category 3 The Max flow Min-Cost Problem Figure 6.8. In between these two extremes there is an exciting range of problems to be explored.10.5 A Panoramic Picture of Similar Problems & Solutions (once again) Please note that Category 1 problems are the most general.

9.9. Given a network flow graph with zero lower bound and a finite upper bound on flow in each edge.1).9.9 Network flows with lower & upper bounds on flow and the Circulation Problem We shall discuss the Circulation Problem and the related problems of finding feasible flow in network flow graphs. The right diagram shows the maximum flow and the minimum cut in the network shown in the left diagram. how can we find a maximum flow from a source vertex to a sink vertex? The maximum flow as well as the minimum cut is indicated in the graph shown below (Fig. It is interesting to note that this section relies on old concepts like network flows.364 Network Flows.1 Prior knowledge: 1. etc. We are given a network flow graph with zero lower bound. maximum flow and minimum cut. 6. 2. We provide a systematic and step by step treatment of a number of theorems and algorithms. We start this section with the required prior knowledge which is essential to understand the theory and the practice described here. a finite . We shall introduce one new concept and that is of a Circulation graph. The lower bound on flow through every edge is zero. maximum flow at minimum cost.1: Left diagram shows a network flow graph with upper bound on flow in each edge. Understanding of these transformations is a must for appreciating the new knowledge described here. We shall also list down the specific problems that we shall address here. These theorems as well as algorithms depend upon a number of powerful transformations. 6. a 3 4 c 2 3 3/3 a 4/1 c 2/1 s 1 3 t d 4 s 1/1 3/1 3/2 t d 4/4 b 1 b 1/1 Figure 6. Connectivity and Matching Problems 6. Please note that the flow is conserved at every vertex except the source and the sink vertex.

In a Circulation Graph the law of conservation of flow should hold for every vertex. 6.2: A network flow graph with upper bound on flow in each edge as well as per unit cost of flow through that edge.2 New concepts In a network flow graph we assume that there is a single source vertex and a single sink vertex. In a network flow graph.9.Network flows with lower & upper bounds on flow and the Circulation Problem 365 upper bound on flow in each edge.9. For example in the graph. and now we need to find a maximum (or a fixed) flow at minimum cost from a source vertex to a sink vertex in the network. the source vertex has the capability to produce an infinite flow while the sink vertex has . The problem is to find maximum flow at minimum cost.9.that means the actual flow entering a vertex is equal to the actual flow coming out of it. Please note that the flow is conserved at every vertex except the source and the sink vertex. shown in Fig. We are also given a cost per unit flow in each edge of the network graph. For the rest of the vertices the law of conservation of flow holds . The lower bound on flow through every edge is zero.1. it is possible to push a flow of 2 units from the source vertex to the sink vertex. 6. For example in the following network flow graph (Fig. in contrast. 3. we should also be able to solve a slightly different problem in which we can find in terms of yes or no if a fixed flow of k units can be pushed in a network flow graph from a source vertex and taken back into a sink vertex from the network. Using our prior knowledge.2) the minimum cost of a flow of one unit from the source to the sink vertex is 4.9. 2/8 a 2/0 b 1/4 3/5 8/2 1/0 s 1/0 6/9 t 2/0 c 1/9 d Figure 6. 6.

6. There may also be per-unit cost associated with each edge.3: A Circulation graph with lower as well as upper bound on flow in each edge.9. 2. For example a feasible flow of 5 is possible in the circulation graph shown below (Fig. 6. A feasible flow in the graph is also indicated. Usually this problem is known as the minimum cost Circulation Problem or the Circulation Problem.7 a 5 3.3 New Problems 1. We need to know how to find a feasible flow in a Circulation Graph where every edge has a lower bound as well as an upper bound on flow. We need to find a feasible flow in a network graph with a source and a sink vertex. Please note that if the lower bound on flow in each edge is zero then a zero flow will always be a feasible flow from the .9.7 b Figure 6. 6.5 t 5. and perunit cost on flow through that edge. Thus in a circulation graph there is no source and no sink vertex.5). Every edge has an associated lower bound.9 s 4. an upper bound.366 Network Flows.9. Every edge has a lower bound as well as an upper bound on flow (see Fig.9. Every edge in a circulation graph may have a nonzero lower bound and an upper bound on flow through this edge. 6. Please see the solution of the Circulation problem in the following diagram (Fig.4). We need to find a minimum cost feasible flow in a Circulation graph.3). Please note that the flow is conserved at every vertex without any exception. 3. Connectivity and Matching Problems the capability to sink an infinite flow. 2.9.

6. the lower bound on flow in any edge is nonzero then finding a feasible flow is not a trivial problem as the zero flow is not a correct answer. The cost of a unit flow through every edge is equal to one in this graph. Every edge has an associated lower bound. We need to find a maximum or a fixed flow at minimum cost in a network graph with a source and a sink vertex. 4.Network flows with lower & upper bounds on flow and the Circulation Problem 2 367 a e d b 4 a 2 b e 3 3 a 2 b e 2 c d 3 c d 2 c Figure 6. source to the sink.6). however. A feasible flow in the circulation graph is indicated in the middle diagram. and per unit cost on flow through this edge.9. Please note that the flow is conserved at every vertex without any exception. If. . an upper bound.4: A Circulation graph (left diagram) with lower bound equal to 1 and upper bound equal to five on flow in each edge. Please see the solution of this problem in the diagram below (Fig.9. A minimum cost feasible flow is indicated in the right diagram.

taking place at minimum cost. an upper bound.6/3 t Figure 6.3.2.5/3 a 1.that is 4 units .2. per unit cost on flow. .9.7/3 x 3.4/2 t 2.6: Left graph is a network flow with a flow of 4 units taking place from the source to the sink vertex. Connectivity and Matching Problems w 1.5: A network flow graph with lower bound.4. upper bound. The right diagram shows the same amount of flow . 1.5/3 s 0.9/2 2. and actual flow taking place in that edge. Also note that the feasible flow taking place may not be a maximum flow (or a minimum flow) from the source to the sink vertex.3. Each edge has an associated lower bound.5/2 a 1.4.8/2 z Figure 6.15/1 s 2. Please note that the flow is conserved at every vertex except the source and the sink vertex.5/5 3.368 Network Flows.6/2 t s 2.15/2 1. A flow of 4 units is a feasible flow but it is taking place at a higher cost in the left diagram. and actual flow in each edge is indicated in the respective order.6/4 y 2.9.

The edge (s.9. 6.7) while a feasible flow exists in the same graph having a source and sink vertex as shown below in the left diagram of Fig. 6.2 a 2 0. 2.5 t 0.9.9 2. Here all edges except one have a lower bound equal to zero. 6. here a feasible flow does not exist because of obvious reasons. It will be interesting to formally prove this claim. Right graph is a circulation graph. Thus it may be possible that a feasible flow does not exist in a Circulation graph (see the right diagram of Fig.8. We transform the left circulation graph into the right network flow graph and then claim that a feasible flow exists in the circulation graph if and only if we can push a specified flow (equal to 2 in this case) from the source vertex x and pull the same amount of flow from vertex y in the network flow graph.9.4 Finding a feasible flow in a Circulation graph with one special edge Consider the graph shown in the left diagram in Fig. . 6.9.5 t b 0 0.7.1 s 0.9. a) has a lower bound equal to the upper bound on flow and they are both equal to 2. here a feasible flow exists.9 s 0.1 b Figure 6.2 a No Feasible Flow 0. Obviously a feasible flow does not exist in this example because of obvious reasons.Network flows with lower & upper bounds on flow and the Circulation Problem 369 Before finding feasible flow in a Network graph and in a Circulation graph we should again note that in a Circulation graph the flow is conserved at every vertex while in a network flow graph it is not conserved at the source as well as the sink vertex.7: Left graph is a network flow graph with a source and sink vertex.

1 Figure 6.9. It will be interesting to formally prove this claim. Here the source vertex x is trying to push a flow of 2 units in the network graph while the sink vertex y is trying to pull 2 units from the network graph.9. 6.370 Network Flows.5 Finding a feasible flow in a network flow graph with one special edge Consider the network flow graph shown in the left diagram with a designated source and a sink vertex (Fig.9 y 2 a s 0.9 s 0. a) has a lower bound equal to upper bound equal to 2.5 t b 0.9.9). The edge (s. Once this relationship is established then we can use earlier results to find a feasible flow provided it exists as shown in Fig. We transform this network flow graph into a circulation graph as shown in the middle diagram.10.2 a 0. 6. Connectivity and Matching Problems x 2 2. Please note that except for one edge all edges have lower bounds equal to zero.1 t b 0. The Circulation graph is converted into a network flow graph shown in the right diagram.5 0. 6.9. We claim that a feasible flow exists in the network flow graph (left diagram) if and only if a feasible flow exists in the circulation graph shown in the middle.8: Left graph is a Circulation graph with associated lower bound and upper bound on flow in each edge. .

x 2 2.∞ 0.1 0.9 y t 2 a s 0.5 t b 0.9. First it is converted into a Circulation graph (top right diagram).2 a 0.9: Left diagram is a network flow graph with associated lower bound and upper bound on flow in each edge.2 a 2 0.1 2.9 s 0.9 s 0 0.Network flows with lower & upper bounds on flow and the Circulation Problem 371 x 2 2. .9 y t 2 a s 0.1 0.5 t b 0.9 s 0.5 2 b 0. We find a feasible flow in the Circulation graph and then convert it into a feasible flow in the original network flow graph shown in the bottom diagram.5 0.2 a 0.10: Top left diagram is a network flow graph with associated lower bound and upper bound on flow in each edge. First it is converted into a Circulation graph (middle diagram) and then we find a feasible flow in the circulation graph.9.2 a 0.1 t b 0.∞ b 0.9 2.∞ b 0.1 t Figure 6.1 0.1 Figure 6.1 s 0.

6 When upper bound is higher than a non zero lower bound We again consider a circulation graph (shown in the left diagram) where every edge has a lower bound equal to zero except for one special edge .9 a 0. It will be interesting to formally prove this claim. 6. 2 2.7 s 0. This edge is split into two edges as shown in the middle graph.9 s 0.5 y a 0. The edge (s. a) has a lower bound equal to 2 and an upper bound equal to 7.9 0.but now the upper bound is higher than a non zero lower bound on flow through this edge.7 s 0.9.9. .11: Left graph is a Circulation graph with associated lower bound and upper bound on flow in each edge.7 t b 0.372 Network Flows.11. The Circulation graph is converted into a network flow graph shown in the right diagram.7 a 0.9.9 2. We transform this circulation graph (left diagram) into another circulation graph shown in the middle diagram and claim that a feasible flow in the left circulation graph exists if and only if a feasible flow exists in the middle circulation graph. Please note that except for one edge all edges have lower bounds equal to zero.5 t b 0. Once this relationship is established then we can use earlier results to find a feasible flow provided it exists as shown in Fig.5 a 0.7 x 2.7 t Figure 6.5 t s 0.5 2 b 0.2 0.5 2 b 0. Here the source vertex x is trying to push a flow of 2 units in the network graph while the sink vertex y is trying to pull 2 units from the network graph. Connectivity and Matching Problems 6.

9 0. . Using earlier transformations we convert the circulation graph shown in the left diagram into a network flow graph shown in the middle diagram.12: Left graph is a Circulation graph with associated lower bound and upper bound on flow in each edge.6 1 a 0.1 t b 0. Here the source vertex x is trying to push a flow of number of units in the network graph while the sink vertex y is trying to pull the same amount of units from the network graph.2 4 5 x 2 3 2 1 x Figure 6. The Circulation graph is converted into a network flow graph shown in the middle diagram.1 t b 0. The middle network flow diagram is in turn transformed into another simplified network flow diagram shown in the right diagram. We claim that a feasible flow in the circulation graph exists if and only if we can push a specified amount of flow from vertex x in the middle or the right diagram.5 t b 5.7 y 5 0.6 2 a 3.9.2 s 0. y 2 2.5 3 a 4 0.5 s 4.Network flows with lower & upper bounds on flow and the Circulation Problem Finding feasible flow in a general Circulation graph 373 Here we consider a circulation graph where each edge may have a nonzero lower bound and a different upper bound on flow through this edge.7 s 0.

.9 3+2 s 4. The middle diagram shows an intermediate stage.1/1 0. This is per unit flow cost for each edge .9 a 0.5/3 a 3.13: Left graph is a Circulation graph with associated lower bound and upper bound on flow in each edge.2/0 t b s 4+1 4. We need to find a feasible flow of minimum cost in this circulation.9.5 t b 5+0 5.7 0.7 2 1 x Figure 6. How to solve the minimum cost Circulation Problem Consider the following circulation problem where the lower bound for each edge is 1 while the upper bound on flow through each edge is 3.in this example it is equal to 1.5 t b 5. Connectivity and Matching Problems y 1 2 2. In addition to lower and upper bounds we have a cost associated with each edge.7 s 0. The right diagram shows a feasible flow in the circulation graph.374 Network Flows.6/2 2+3 2.7 a 3.

6. The Circulation graph is converted into a network flow graph shown in the right diagram.9.14: The left diagram is a Circulation graph with associated lower bound and upper bound on flow in each edge. .9. If we are successful then it means that a feasible flow exists in the circulation.15. If we just want to find a feasible flow (not the minimum cost feasible flow) then we know what to do. Such a feasible flow is shown in the left diagram of Fig. We try to push 2 units of flow from vertex x. Here the source vertex x is trying to push a flow of 2 units in the network graph while the sink vertex y is trying to pull the same amount of units from the network graph.Network flows with lower & upper bounds on flow and the Circulation Problem 375 x Find a Feasible Flow 1 1 Push 2 units of Flow -1 +1 -1 +1 1 Upper Bound is 3 while Lower bound is 1 Upper Bound is 2 while Lower bound is 0 for black edges 1 y Figure 6. The cost of per-unit flow through each edge is equal to 1. We convert the circulation (left diagram) into a network flow graph as shown in the right diagram. But we know that this feasible flow may not be the minimum cost feasible flow.

This flow of 2 units at minimum cost is then translated into a feasible flow at minimum cost as shown in the right diagram of Fig.14.15. The right diagram shows that it is possible to push the same amount of flow from vertex x at a lower cost in fact at a minimum cost.376 Network Flows. 6.15: The left diagram shows that it is possible to push 2 units of flow from the source vertex x at some cost.9.9.9. . This provides us a solution to the (minimum cost) Circulation Problem.9. 6. In order to find a feasible flow at minimum cost in the circulation graph of Fig. Connectivity and Matching Problems x 1 1 Push 2 units of Flow at any Cost x 1 1 Push 2 units of Flow at Minimum Cost -1 +1 -1 +1 -1 -1 +1 1 Upper Bound is 2 while Lower bound is 0 for black edges 1 Upper Bound is 2 while Lower bound is 0 for black edges 1 +1 1 y y Figure 6. 6. we should push 2 units of flow from vertex x at a minimum cost as shown in the right diagram of Fig.16.

But we know that this problem is a hard problem? Assume that we need to find minimum flow from vertex s to vertex t in a network flow graph which is un-directed. We know that minimum flow from vertex s to vertex t is equal to max cut in a network .7 Is it possible to solve the Circulation Problem for un-directed graphs? We need to find a (or size of) maximum cut in an undirected network flow graph having vertices s and t.9. The cut should separate vertex s from vertex t.9.16: The left diagram shows the source vertex x pushing two units of flow at minimum cost.Network flows with lower & upper bounds on flow and the Circulation Problem 377 x 1 1 Push 2 units of Flow at Minimum Cost Minimum Cost Feasible Flow -1 +1 -1 +1 1 Upper Bound is 2 while Lower bound is 0 for black edges 1 y Upper Bound is 3 while Lower bound is 1 Figure 6. Here the flow in each edge is 1 except for the red bold edges where the flow is 2 units. The right diagram shows the resulting solution of the minimum cost Circulation Problem. We call this Problem 2. We call this Problem 1. Our Strategy of Solving the Circulation & other related problems Our strategy of solving the Circulation and other related problems are summarized in the following four figures which are self explainatory. 6.

19: Finding feasible flow in a network with non zero lower bounds.18: Finding a minimum cost feasible flow in a Circulation. Find Feasible Flow at Min Cost Circulation Problem? (non zero lower bounds) Network Flow at Min Cost (Zero lower bounds) Figure 6.17: Finding a feasible flow in a Circulation with non zero lower bounds. Find Feasible Flow Network Flow Problem? (non zero lower bounds) Find Feasible Flow Circulation Problem? (non zero lower bounds) Network Flow Problem (Zero lower bounds) Figure 6.9. Connectivity and Matching Problems Find Feasible Flow Circulation Problem? (non zero lower bounds) Network Flow Problem (Zero lower bounds) Figure 6. .378 Network Flows.9.9.

3 5.4 t 2. In order to find minimum flow we need to make it a circulation as shown above and then find a minimum cost circulation (Problem 3) in this graph .3 1.7 x 3.5 1.3 s 0.∞ y 2.9 2.9 2. The minimum flow is 5 and thus the max cut will also have same size.3 w t 2.8 z y 2.but instead a negative result.7 x 3.9. Thus we moved in the following fashion to find the max cut in a network flow graph? In case of un-directed graphs we have the same sort of strategy .20: The network flow problem is first converted into a circulation problem.6 5. .8 z Find Maximum Flow at Minimum Cost The Circulation Problem Figure 6. The top figure shows that if we need to find max cut in this directed graph (Problem 1) then we should find the min flow from vertex s to vertex t (Problem 2).4 s 0.that will give us a minimum flow from s to t in the network flow graph as shown below.5 1.Network flows with lower & upper bounds on flow and the Circulation Problem 379 flow graph? Suppose we can solve the circulation problem (Problem 3) at minimum cost in an un-directed graph? Then we can use the solution of Problem 3 in order to solve Problem 2? We can then use this solution to find a solution to Problem 1? w 1.6 0.

3/2 x 3.9.380 Network Flows.4/2 t 2.7/5 w 1.8/4 z Figure 6. It essentially means that we cannot find a minimum cost circulation in an un-directed graph while we can solve this problem in a directed graph? .9.21: Finding minimum flow at minimum cost in a network flow problem.6/2 y 2.3/3 1.9/2 2.22: Solving one problem solves another.5/3 s 0. The maximum cut can now be found. Connectivity and Matching Problems 5. Solve Problem 3 Solve Problem 2 Solve Problem 1 Figure 6.

Chapter 7 Eulerian Graphs & the Chinese Postman Problem 7.2 7.5 A Special Class of Graphs Eulerian Circuits and Graphs Eulerian Trails and Related Problems Eulerian Walk and the Chinese Postman Problem The Chinese Postman Problem for Directed Graphs .4 7.3 7.1 7.

We have claimed earlier that there is a class of graphs where even an unintelligent algorithm (like Algorithm 36) can efficiently find the maximum edge-disjoint paths. Thus we start with a panoramic picture. 7. Remove all edges in the path P and go to step 1. We end this chapter with a detailed study of the Chinese Postman problem for both directed as well as undirected graphs. We then come back to our categories of graphs and look at these in the light of our newly acquired experience about Eulerian graphs. We make a number of inter-related claims about such graphs and then show how the proof of one claim can lead to the proof of another. We shall consider connected undirected graphs. In this graph vertex a is a source vertex having only out-degree while the vertex d is a sink vertex . output: (Maximum) edge-disjoint paths from s to t in D 1 2 Find a directed path P from vertex s to vertex t in D.382 Eulerian Graphs & the Chinese Postman Problem Introduction We shall start with a special category of graphs which was earlier discovered in the last chapter.1. Algorithm 55: Find Maximum edge-disjoint paths from s to t in D input : Directed graph D. and vertices s & t. move in depth with one category. Consider the directed graph D shown in Fig. Please recall Algorithm 36 which is reproduced below. Eulerian graphs belong to one of these categories which we discuss in detail. If you are successful in finding a path then keep a record of this path otherwise exit the algorithm. We shall study this and similar classes of graphs in detail in this chapter. 7.1. Exploiting our prior knowledge about this category we define a number of new categories of graphs which are to some extent similar and at the same time different from the graphs we have seen previously.1 A Special Class of Graphs We have studied the problem of finding the maximum number of edge-disjoint paths from a source vertex to a sink vertex in a graph in the last chapter. and then come back to the panoramic picture with new tools and techniques. for a directed graph we shall assume that the underlying undirected graph is connected.

A Special Class of Graphs 383 having only in-degree. for the rest of the vertices of D.1. etc) in order to execute the first step of this algorithm. In this special class of graphs the in-degree of vertex d is always equal to the out-degree of vertex a (why?). 2. In fact all class A directed graphs possess the following properties: 1. The right diagram shows the same graph with one trail and one path from vertex a to vertex d in the given graph. Please note that each edge of this graph is covered either by the trail or by the path as shown in this diagram. We can find maximum edge-disjoint paths from vertex a to d in this class of directed graphs using Algorithm 55. z e f y d a b x c Figure 7. 7. the in-degree is equal to the corresponding out-degree. Please note that we can use any traversal algorithm (Breadth First Search. The maximum number of edge-disjoint paths from the source vertex to the sink vertex is equal to the out-degree of the source vertex or the in-degree of the sink vertex. Depth First Search. Every edge of the graph is covered by one of the edge-disjoint trails (or paths) from the source vertex to the sink vertex.1: A directed graph D with two special nodes a and d. We can use Algorithm 55 to find maximum number of edge-disjoint paths from vertex a to d in this graph (there is no need to use the more sophisticated Algorithm 37).2. . The maximum (number of) edge-disjoint paths in the directed graph D are shown in the left diagram of Fig.1. All directed graphs fulfilling the above properties are known as Class A graphs. The indegree is equal to the out-degree for every node in this graph except a and d.

Class E deals with undirected graphs where the degree of every vertex is even except for two special vertices a and d where the degree is an odd number. Class F deals with a more general category of undirected graphs in which . the in-degree of every vertex is equal to the corresponding out-degree. For every other node in this graph the in-degree is equal to the corresponding out-degree. For the rest of the vertices in D the in-degree is equal to the respective out-degree. We shall prove these properties later in this chapter. There are two edge-disjoint trails in this graph such that each edge of the graph is covered exactly once by either of the two trails as shown in the right diagram. Thus the vertex a resembles a source vertex while vertex d resembles a sink vertex.2: We show a directed graph of Class A: Vertex a has only outdegree while vertex d has only in-degree. Again we shall discuss the proofs later. while the vertex d has an in-degree larger than the out-degree. z z e f y d e f y d a b x c a b x c Figure 7. Class C deals with directed graphs in which there are no special vertices. however. There are two edgedisjoint paths from vertex a to vertex d in the graph D as shown in the left diagram. We also assume that vertex a has an out-degree larger than the in-degree. an important difference this time: Vertex a has an out-degree but also an indegree while the vertex d has an in-degree but also an out-degree. There is.384 Eulerian Graphs & the Chinese Postman Problem 3. Let us now define a Class B category of graphs: In this category of directed graphs we again have two special nodes a and d.1. Class C may also include undirected graphs where the degree of every vertex is even. The maximum number of edge-disjoint paths from the source vertex to the sink vertex can be found by Algorithm 55. How is this class B different from class A and in what respect are they similar? Try to answer this question before moving forward.

1. . A concept map showing various classes of some special graphs.A Special Class of Graphs 385 Concept Map 7.

4. The out-degree of vertex f is 3 while the in-degree is equal to 1. the rest of the p − 2k vertices have an even degree. here are a few questions that you should be able to answer by yourself: 1.1. The degree of each node in this undirected graph is even.3. We show a Class B directed graph in Fig. Can we also find trails in this graph such that each edge of this graph is covered exactly once by one of the trails? We show a Class C directed graph in the left diagram of Fig. the two paths are shown in Fig. How other properties of a graph change (or do not change) after such a transformation is interesting to explore. the out-degree of vertex x is 1 while its in-degree is equal to 3. if we now remove edges of this cycle from the original graph then the new graph will also belong to the same class (why?).1.1. Both the graphs shown in this diagram can be partitioned into edge disjoint cycles (or circuits) shown in different colors.386 Eulerian Graphs & the Chinese Postman Problem every graph G (having p vertices) has 2k nodes with an odd degree. for the rest of the nodes.4. As you should discover yourself Class C directed and undirected graphs have some special and interesting properties.3. Before moving forward. It is interesting to note that a directed graph (where the in-degree of each node is equal to the corresponding out-degree) belongs to the same class as an undirected graph where the degree of each vertex is even (why?). 7. 7. What are the necessary and sufficient conditions for a (connected) graph to be cyclic? . it will become a Class A or Class B graph. the in-degree is equal to the corresponding out-degree. We can use Algorithm 55 to find the two edge-disjoint paths from vertex f to vertex x in D. Similarly if we add a path between vertex a and d (these are the only two vertices having an odd degree) in a Class E graph then it will be transformed into a Class C category. Remember that the number of vertices having odd degree in a graph can not be odd (why?). We can use any traversal algorithm to find a cycle in such graphs. There is no special vertex in this graph. If however we remove (all edges in) a path from any vertex x to a vertex y in a Class C graph then the new graph will not be a Class C graph any more. One such property is shown in Fig. the in-degree of every vertex is equal to the corresponding out-degree. 7. An undirected graph belonging to the same category is shown in the left diagram of this figure. 7.1.

3: A directed graph of Class B: The out-degree of node f is larger than the in-degree while the in-degree of x is larger than its out-degree. For every other node in this graph the in-degree is equal to the corresponding out-degree. An un-directed graph where the degree of each vertex is even is shown in the right diagram. Both these graphs belong to our Class C category. . The edge set of these graphs can be partitioned into edge-disjoint cycles (shown by different colors) which if combined together will create a circuit consisting of all edges of the graph.1. z z e f y d e f y d a b x c a b x c Figure 7.A Special Class of Graphs 387 Figure 7.1.4: Every node in the directed graph has an in-degree equal to the corresponding out-degree as shown in the left diagram.

A graph G is Eulerian if and only if every edge of G lies on an odd number of cycles. Let us consider a special case of Class C graphs where the degree of each vertex is not only even it is exactly two. What about if each vertex in a graph has a degree equal to at least two? Remember we are considering connected graphs only as mentioned earlier in this section. 3. 2. Let us do . Can we make the above claim if the degree of every vertex is even? 5. We shall start with proving Claim Number 4 and then work backwards in order to prove earlier claims.388 Eulerian Graphs & the Chinese Postman Problem 2. A graph G is Eulerian if and only if the edge set of G can be decomposed into edge-disjoint cycles. Can we claim that in such a graph every vertex will lie on some cycle? 4. What can you say about other properties of this special graph? 3. then we should be able to prove that the degree of each vertex of G is even (you should be able to do it easily). The degree of each vertex in a graph G is even if and only if the edge set of G can be decomposed into edge-disjoint cycles. Try to visualize such a graph. A graph G is Eulerian if and only if the degree of each vertex is even. So we assume that the degree of each vertex is even and now we should be able to prove that the graph G can be decomposed into edge-disjoint cycles. 7. Now we make a number of claims (it will become evident from these claims that an Eulerian graph belongs to our Class C category): 1. Assume that a graph G can be decomposed into edge-disjoint cycles. 4. Please remember that in a circuit it is possible to visit a vertex several times but we are allowed to traverse an edge only once.2 Eulerian Circuits and Graphs A graph G is Eulerian provided it contains an Eulerian circuit. a circuit which contains every edge of G.

1: An Eulerian graph containing an Eulerian circuit is shown in the left diagram.2.Eulerian Circuits and Graphs 389 it now. Remember in a circuit a vertex may be repeated but an edge cannot be repeated. If the graph is Eulerian then there will be an Eulerian Circuit inside that graph as shown in left diagram of Fig.1. . Thus the cycles forming the Eulerian circuit will be edge-disjoint as shown in the left diagram of Fig. 7. so the graph will be Eulerian.1. 7. Let us now tackle Claim Number 3. If this circuit is a cycle then the proof is complete otherwise it will consist of several cycles. The circuit can be decomposed into edge-disjoint cycles as shown in the right diagram. z z e f y d e f y d a b x c a b x c Figure 7. It is possible to build logic on this observation in order to design a formal proof. If the edge set of a graph can be partitioned into edge-disjoint cycles then we can always combine these cycles to create a circuit which will cover every edge of the graph exactly once. As the degree of each vertex in G is even we can find a cycle C in G using any traversal algorithm. Combining the two we can prove that a Graph is Eulerian if and only if the degree of each node in the graph is even. We have earlier proved in Claim Number 4 that a graph can be partitioned into edge-disjoint cycles if and only if the degree of each vertex in the graph is even. If we remove all the edges belonging to C from G then in the resulting graph the degree of each vertex will again be even (why?) but this new graph will have fewer edges as compared to the original graph G. This part of the proof is done and let us attempt the other part: if a graph is Eulerian then (we shall be able to prove that) it can be partitioned into edge-disjoint cycles.2. In Claim Number 3 we have proved that a graph G is Eulerian if and only if the edge set of G can be decomposed into edge-disjoint cycles.2.

in fact there will always be an odd number of ways out (why?). Let us now concentrate on Claim Number 1. thus the degree of node m as well as that of node n will be even. We start from a (in Fig. Proving that edge (m. as the degree of every vertex in G is even so if you can enter a vertex then you can leave it as well. It means that the degree of node m as well as that of n will be even. As the graph G is Eulerian thus the degree of each node will be even. First assume that every edge (m.390 Eulerian Graphs & the Chinese Postman Problem Figure 7.2 will be equal to .2. 7. Now assume that the graph G is Eulerian.3.2) and arrive at the adjacent vertex e. n) in graph G is part of an odd number of cycles. Consider vertex a which is adjacent to vertex b in this graph. 7. The total number of trails in the graph of Fig.2: An Eulerian graph is shown.2. 7. This means that if you draw a tree of all possible trails then the out-degree of every node in this tree will be an odd number as shown in Fig.2.2. we have to prove that every edge (m. Please note that if there is a counter clockwise cycle from vertex f to vertex y to vertex z and back to f then there is clockwise cycle from vertex f to z to y and back to vertex f . 7. n) in G will be part of an odd number of cycles as shown in Fig. Please remember that in a trail we may repeat vertices but we cannot repeat edges while in a path neither vertices nor edges can be repeated. This may not be very obvious so we shall prove this after first finding the number of distinct trails between the two vertices. n) will be part of an odd number of cycles is equivalent to proving that there are an odd number of paths between vertex m and vertex n (why?). We intend to prove that the edge ab is part of an odd number of cycles in this graph. Proving that there is an odd number of paths between vertex m and vertex n is in fact equivalent to proving that there is an odd number of trails from vertex m to vertex n. thus G will be an Eulerian graph.2.2.

We show here (middle diagram) all possible trails starting from vertex a and ending at vertex b.2.2. We show all possible paths between vertex a and vertex b in the bottom diagram.3: Consider the Eulerian graph shown in Fig.2.Eulerian Circuits and Graphs 391 z z e f y d e a b z f x c f y d a y d a b z x c e e f z b y e f y d a b x c a b x c y b z x b x f b f b d c b d c b b c x b d x b c x b d x b z e f y d a b x c Figure 7. 7. .

There are a number of algorithmic issues apart from the above (theoretical) claims and their respective proofs: 1. How can you efficiently check if a graph G is Eulerian using different (necessary & sufficient) conditions for a graph to be Eulerian? 2. Every cycle in the graph contributes to two trails as shown in Fig. How can you efficiently find an Eulerian circuit in an Eulerian graph? 3.2. How can you efficiently find edge-disjoint cycles in an Eulerian graph? There are a number of exciting theoretical problems which you should attempt before moving forward: 1.2. How can you generalize (or modify) the four claims in case of directed graphs? .2. 7.3 which will be an odd number (why?).2 and Fig.3. in multi-graphs we allow parallel edges and self loops? (Hint: Can you convert a multi-graph into a simple graph?) 2. Thus the total number of paths between vertex m and vertex n will be an odd number (why?).392 Eulerian Graphs & the Chinese Postman Problem the number of leaf vertices in the tree of Fig. 7. 7. How can you generalize the four claims that we have made for multigraphs.

2.Eulerian Circuits and Graphs 393 Problem Set 7. Problem 7.1. e f y e f y a Graph G b x a b Graph H x c c Figure 7. Problem 7. Remove all edges in the Cycle C from graph G. . Prove the claim or give a counter example. Please read the following algorithm which is primarily designed to find edge-disjoint cycles in a graph G where the degree of every vertex is even.4: Problem 7. Algorithm 56: Find Cycles in an un-directed graph G where degree of every vertex is even input : Undirected graph G where degree of every vertex is even output: Edge-disjoint cycles in graph G 1 2 Find a Cycle C in graph G. Apply the above algorithm on the graph G as shown in Fig.1.2. Someone claims that the algorithm outputs maximum number of edge-disjoint cycles in a graph G.1. and specify each cycle output by the algorithm. If some edges are still left in G then go to step 1 otherwise exit.3.4.4.1. Someone claims that the algorithm outputs maximal number of edge-disjoint cycles in a graph G. and keep a record of it. Some one claims that the algorithm outputs edge disjoint cycles.1. 7.1. Discuss why or why not. Problem 7. Prove the claim or give a counter example.2.

How about if we have a directed graph in which the indegree of every vertex is equal to its out-degree. How would the algorithm behave in this type of graph? What will be the output of the algorithm? How will the output be different in this case? Discuss briefly.1.5. . How will the solution of earlier problems be affected? 7.7.1.6.1.3 Eulerian Trails and Related Problems Let us now consider Class E undirected graphs where the degree of every vertex is even except for two special vertices f and x where the degree is an odd number (will the degree of the two odd vertices be the same? Why?). Algorithm 57: Find graph G given a set of edge-disjoint cycles of G input : Edge-disjoint cycles in an unknown graph G output: Un-directed graph G (degree of every vertex in G will be even) Algorithm 58: Find an Eulerian circuit in graph G given a set of edge-disjoint cycles of G input : Edge-disjoint cycles in an unknown graph G output: An Eulerian circuit in G Algorithm 59: Find a set of edge-disjoint cycles of G given an Eulerian circuit in a graph G input : An Eulerian circuit in G output: Edge-disjoint cycles in an unknown graph G Problem 7.394 Eulerian Graphs & the Chinese Postman Problem Problem 7. Problem 7. Assume that we apply the above algorithm to graph H which is different from graph G. It will be useful to design the following algorithms for an undirected graph G.

We can use Algorithm 56 to find out efficiently one such set of edge-disjoint cycles.Eulerian Trails and Related Problems 395 e f y a b x e f y c e f y a b x a b x c e f y e c f y a b x a b x c c Figure 7.5: A graph (with every vertex having even degree) can be split up into edge-disjoint cycles.2. Several sets of edge-disjoint cycles are shown here. some having a larger size than others. Note that every edge is part of an edge-disjoint cycle. .

2.6: We have seen that if the degree of every vertex is even then the graph can be split into edge-disjoint cycles again shown in the top diagram. There will be an Eulerian circuit in the reconstructed graph. . The original graph will have all nodes with even degree.396 Eulerian Graphs & the Chinese Postman Problem b x e f e f f y c f a y b a b b x x x c e f y a b x c Figure 7. The reconstruction algorithm will help you reconcile with this claim. On the other hand if a graph can be split into edge-disjoint cycles or if the edge-disjoint cycles of a graph are given then we can reconstruct the original graph as shown in the middle diagram.

3. This class comprises of undirected graphs having 2k nodes with an odd degree. Its counterpart in directed graphs is shown in the right diagram of Fig.2 and 7.3.3. the degree of every vertex is not even then the problem is to traverse each edge at least once (not exactly once) and making sure . for the rest of the vertices the in-degree is equal to the corresponding out-degree. Again we should appreciate that we can use Algorithm 55 to find maximum number of edge-disjoint paths from a source to a sink vertex? We have earlier presented Class B and Class A category directed graphs in Fig.1. If.3. 7. The right diagram shows the same graph where we add an extra edge between two odd vertices converting this graph into a Class C category where the degree of each vertex is even.1. 7. a trail in which every edge of the graph is covered (exactly once).3. It will then become possible to find an Eulerian circuit in the resulting graph which is also shown in the right diagram of Fig.1. In both these graphs it is possible to find an Eulerian trail from vertex f to vertex x. Such a graph is shown in the left diagram of Fig. with odd vertices shown in bold. 7. 7. 7.3. however. This observation should lead you to design a formal proof for the above claim.2 for comparison with graphs in Fig. the rest of the p − 2k vertices have an even degree.1.3. 7. The claim for such a graph G is that the edge set of G can be partitioned into k trails where each trail is connecting two odd vertices.4 Eulerian Walk and the Chinese Postman Problem We know that if the degree of every vertex is even in an undirected graph G then we can find an Eulerian Circuit in G by traversing each edge exactly once.3. These claims can be generalized with some interesting modifications for Class B directed graphs and then proved using our newly acquired experience of Eulerian Graphs? We now present one last claim which is applicable to Class F undirected graphs. How can we prove this and how can we find an Eulerian trail? Perhaps we can find a constructive proof which will solve both the problems. 7. here the outdegree of one special vertex is larger than its in-degree by one.Eulerian Walk and the Chinese Postman Problem 397 Such a graph is shown in the left diagram of Fig. We have made certain claims about Class A graphs earlier in this chapter.3. 7. while it is the other way round for the other special vertex. These graphs are reproduced in Fig.1.

2: A directed graph of Class B category is shown in the left diagram. A directed graph of Class A category is shown in the right diagram. vertex a has only out-degree while vertex d has only in-degree. the degree of vertex f and x is odd while the degree of every other vertex is even. A directed graph D with two special nodes f and x is shown in the right diagram. the in-degree is equal to the out-degree for every node in this graph except f and x. the out-degree of vertex f is one larger than its in-degree and it is the other way round for vertex x in this directed graph.3.398 Eulerian Graphs & the Chinese Postman Problem z z e f y d e f y d a b x c a b x c Figure 7. this is a Class E un-directed graph. it is the other way round for vertex x.3. . the out degree of vertex f is three while its in-degree is 1. the in-degree is equal to the out-degree for every node in this graph except a and d.1: An un-directed graph shown in the left diagram. z z e f f y d e f y d a b x x c a b x c Figure 7.

By adding two edges between odd vertices we can convert this graph into a Class C category where the degree of every vertex is even.Eulerian Walk and the Chinese Postman Problem 399 z 3 3 4 z 4 e f y d e f y d 3 3 4 4 a b x c a b x c Degree of 4 vertices is odd Degrees of odd vertices is made even by inserting edges (shown in red color) between odd vertices z z e f y d e f y d a b x c a b x c We get an Eulerian Circuit We remove the dotted edges and get two trails Figure 7.3. b. . y and x is odd while the degree of every other vertex is even.3: An undirected graph of Class F shown in the left diagram. the degree of vertex f .

7.3. We find a path from one odd vertex to another odd vertex and duplicate every edge encountered in that path.4. The resulting graph will be Eulerian as shown in the right diagram of this figure. 1. in graph G we shall be traversing an edge twice when in graph H we shall be traversing a duplicated edge. we need to minimize the total sum of edge weights in a closed walk which covers every edge of the graph at least once. This problem is faced by any post man delivering letters in houses along lanes or a sweeper who is sweeping roads.400 Eulerian Graphs & the Chinese Postman Problem that the number of edges. This problem is formally defined as the Chinese Postman Problem: We need to find a shortest closed walk in a graph G which passes through every edge of G at least once.4. in a weighted graph. We show a simple technique of converting a graph with odd vertices into a graph with all vertices even in Fig. we show graphs where there are two vertices of odd degree. We need to convert a given graph G into an Eulerian graph H by duplicating some existing edges of G. are minimized. he would certainly like to traverse each lane at least once while making sure that the traversals of the same lane should be minimized.2. In other words the Eulerian graph H should be of minimum size in terms of number of edges or in terms of sum of edge weights of H. In the next figure. There are basically two problems that we intend to solve simultaneously but first we need an understanding of these problems in isolation. Each resulting Eulerian graph H will . All these possible paths gives rise to different number of duplicated edges required to convert a graph into an Eulerian graph. In the left diagram of this figure we have an un-weighted graph while the right diagram shows a weighted graph. Please note that the degree of the two terminal odd vertices in the path will become even while the degree of an even vertex in the middle of the path will stay even. Indeed there could be paths of different edge lengths between the two odd vertices in a graph as shown in Fig. 2. traversed more than once. Such a walk (walk because some edges will be traversed more than once) is also known as an Eulerian walk in the graph G. Now an Eulerian circuit in H will correspond to a closed Eulerian Walk in G. As you can well imagine. 7. We need to make sure that the Eulerian trail that we have created in the first part is shortest in terms of number of edges involved or in terms of edge weights.

For a weighted graph we need to convert G into an Eulerian graph by duplicating certain edges such that the sum total of edge weighted corresponding to duplicated edges is minimized.Eulerian Walk and the Chinese Postman Problem 401 Two vertices with odd degree (red color) in an unweighted graph Two vertices with odd degree (red color) in a weighted graph Figure 7. we need to select the one with the minimum number. We show different Eulerian graphs corresponding to a weighted graph G in Fig.4.1: An undirected graph and un-weighted graph (left diagram). shown in the right diagram. consists of more edges than the graph shown in the middle diagram. have different number of edges. . The Chinese Postman Problem for an un-weighted graph can thus be rephrased: We need to convert a graph G (having some odd vertices) into an Eulerian graph H (having all vertices with even degree) by duplicating minimum number of existing edges of graph G. The right diagram shows a weighted graph.4. 7.4. Both graphs have two vertices with an odd degree. It is interesting to note that optimal graph.

402 Eulerian Graphs & the Chinese Postman Problem 1 3 2 G 2 Transform G into H 1 4 H 2 4 4 4 2 3 3 4 3 Find a path between the two odd degree vertices in graph G After Duplicating the edges in the path the odd vertices in G become even and even vertices remain even Figure 7.3: It is possible to convert a graph having two odd vertices into different Eulerian graphs with varying number of edges. The duplicated edges are shown in red color in the graph H. An Eulerian Circuit in H is equal to an Eulerian Walk in G.2: Shows how can we convert a graph G having two odd vertices into a graph H where the degree of each vertex is even. . The Eulerian circuit in each graph is also indicated.4. 1 2 1 2 1 2 4 3 4 3 4 3 The duplicated edges are shown in red. the graph becomes Eulerian Optimal Solution Only one edge (red) is duplicated to create an Eulerian Circuit Figure 7. the graph becomes Eulerian The duplicated edges are shown in red.4.

The resulting Eulerian graph is shown in the right diagram of the same figure. 7. Do we have to enumerate all possible sets of pairs of odd vertices in the graph G? How to do that systematically and estimate what is the total number of possibilities? Is a brute force approach suitable or should we find a way to solve this problem in a more efficient manner? Given 2k items (corresponding to 2k odd vertices) numbered from 1 to 2k. A minimum weight . Let us start with an arbitrary selection of vertices in the three pairs as shown in Fig.6.5? How about if we apply the same trick of finding a shortest path between vertices belonging to different pairs of odd vertices? (Remember the total number of odd vertices in any graph will always be even). The Eulerian circuit in each graph is also indicated.9. It is possible to put weights on edges of the completely connected graph as shown in Fig. What about if a graph G consists of more than 2 odd vertices as shown in Fig. Finding shortest paths for a fixed set of pairs of odd vertices certainly helps to reduce the cost of making the degree of each vertex even .but a different set of pair of odd vertices may help us in further reducing this cost as shown in Fig.7.Eulerian Walk and the Chinese Postman Problem 1 9 6 4 6 2 403 1 9 2 2 1 9 2 2 2 4 6 4 4 3 7 4 7 3 4 7 3 The duplicated edges shown in red Cost is 6+7=13 The duplicated edge shown in red Cost is 9 Optimal Solution Two edges are duplicated Cost is 2+4 = 6 Figure 7.4.4. 7. Thus the problem is reduced to finding the set of pairs of odd vertices which minimizes the cost of duplicating the edges. The total number of edges duplicated is also indicated in this diagram.8. It is quite obvious now that a shortest path between the two odd vertices provides us an optimal solution to the Chinese Postman Problem given that the graph has only two odd vertices.4: It is possible to convert a graph having two odd vertices into different Eulerian graphs with varying number of edges. 7.4.4. let us draw a completely connected graph consisting of 2k = 6 vertices (of G) as shown in Fig. 7. 7.4. A perfect matching in this graph provides the desired set of all pairs of odd vertices.4.

4. 2 & 4. and 5 & 6 2 4 c b 3 1 a 6 e 5 d Add an extra edge in each path All vertices become even now Cost=Extra Edges=3+3+1=7 4 3 e Figure 7.6: We find shortest paths between a fixed set of odd vertices in graph G. The resulting Eulerian graph H is shown in the right diagram.5: A graph G consisting of 6 odd vertices.4.404 Eulerian Graphs & the Chinese Postman Problem Figure 7. 2 c b 1 a 6 5 d Find shortest paths between odd vertices 1 & 3. .

. perfect matching in the completely connected weighted graph will provide an efficient solution to our problem as shown in the following figures.Eulerian Walk and the Chinese Postman Problem 405 2 c b 1 a 6 5 d Find shortest paths between odd vertices 1 & 2. and 5 & 6 2 4 c b 3 1 a 6 e 5 d Add an extra edge in each path All vertices become even now 4 3 e Cost=Extra Edges=2+1+1=4 Figure 7.4. 3 & 4. The resulting Eulerian graph H is shown in the right diagram.7: We find shortest paths between a different set of odd vertices in graph G.

. we duplicate edges along the shortest path between vertex x and vertex y in G. The weight of an edge between two vertices in graph K is equal to the weight of the shortest path between the corresponding odd vertices in graph G. 2. For every such pair (x.8: A perfect matching in a completely connected graph K of odd vertices of graph G.6) A Perfect Matching (1.4). Create a completely connected graph K.2). y).4).406 1 Eulerian Graphs & the Chinese Postman Problem 2 3 4 3 6 2 1 2 3 4 1 2 1 3 2 2 1 1 2 3 4 1 2 2 6 3 2 3 6 2 1 3 2 3 1 2 5 4 5 2 4 5 2 4 A Perfect Matching (1. (5. Identify odd vertices in graph G. A minimum cost perfect matching in the completely connected graph K provides us the desired pairs of vertices (x. (5. An efficient solution of the Chinese Postman Problem in an un-directed graph G is thus reduced to the following steps: 1. 3.4). (2.5).6) Figure 7. (3.3).6) A Perfect Matching (1.4. every odd vertex of G is a vertex of this completely connected graph K. y) in graph K. (3. (2.

(5. (3.Eulerian Walk and the Chinese Postman Problem 407 2 c b a 6 5 d e 5 a 4 c 2 4 b 1 3 1 3 6 e d (1. (1.3).4.6) 1 2 2 1 3 2 3 6 1 1 3 6 1 3 5 4 5 4 Figure 7.4).6) .2). (2.4).9: The weight of a shortest path between two odd vertices in G corresponds to the weight of an edge between the two corresponding odd vertices in the graph K. (5.

2). (2.2).6). (2.5) 4 1 2 1 2 1 2 1 1 6 2 3 6 1 1 3 3 6 1 1 3 5 3 (1. (2.5).2).6).6) 4 5 (1. (3. (2. (2. (4.5) 4 5 (1.6) 1 2 1 2 1 2 3 6 3 3 6 3 2 2 3 6 1 3 3 1 5 (1. (2.4).3).5).5) 4 Figure 7. (2.3).3). (3.4).6) 4 5 (1.6) 4 5 (1. (2.4). (3. (5.3). (2. (3.4).10: We enumerate all distinct perfect matching in a completely connected graph consisting of 6 odd vertices of G.4).6) 1 2 3 6 4 2 3 4 5 (1. (4. (3. (2.5). (3.3). (5.4) 3 2 5 3 (1. (3. (4.6).5).3).6). (3.4).5) 4 (1. (4. (5.5) 4 1 2 1 2 1 2 2 6 4 1 3 6 4 1 5 (1. (3.6) 2 1 4 5 2 4 5 3 (1.6) 4 5 3 (1.6). (4.4.6).5). (2. Please note that looking at all enumerations will be very costly.5). (2.4) 4 1 2 1 2 1 2 1 6 2 3 6 1 3 3 6 1 2 1 4 5 (1.6). (4. how do we find a better solution? .4).6) 4 5 (1.408 1 Eulerian Graphs & the Chinese Postman Problem 2 2 1 2 2 1 2 2 6 3 6 3 6 1 3 1 5 (1.5).

. Without enumerating all perfect matchings in a graph we can still efficiently find the minimum cost perfect matching in a graph.11: The minimum weight perfect matching in a completely connected graph K among all odd vertices in G corresponds to the minimum number of edges of G which if duplicated will convert graph G into an Eulerian graph H.Eulerian Walk and the Chinese Postman Problem 409 2 c 2 3 2 4 1 2 1 2 3 1 2 1 1 3 2 4 b 1 a 6 5 d e 3 Minimum Weight Perfect Matching in a completely connected weighted graph Add extra edges corresponding to Minimum Weight Perfect Matching Cost=Extra Edges=1+1+1=3 Figure 7.4.

5. 1.1. this graph is converted into a directed graph as shown in the top right diagram of the same figure. The undirected graph G has even as well as odd vertices . The out-degree is equal to the corresponding in-degree. 2. We show a directed graph G where there is only one vertex with positive ∆ as shown in Fig. We show an undirected graph G in the top left corner of Fig. There are four such vertices in graph D. all shown in orange. There are three such vertices in graph D. Two such vertices exist in the directed graph D and are shown in green color. Here the problem is simple: Finding a path from each of the three orange vertices to the only red vertex and duplicating edges in these paths will convert the graph G into an Eulerian graph.5. If a directed graph G have all vertices where the in-degree is equal to the corresponding out-degree then the graph G is Eulerian. hence ∆ is negative. some vertices have out-degree larger or smaller than the in-degree then we need to convert graph G into an Eulerian graph H by duplicating minimum number of existing edges of G. the directed graph D has three types of vertices as described below. Finding shortest paths from each orange vertex (with ∆ negative) to the only red vertex (with ∆ positive) will provide us the optimal solution to . The out-degree is less than the corresponding in-degree. however.1.5. however there are vertices with +ive and/or -ive ∆ then we face the challenge of solving the Chinese Postman Problem. If. 7. In contrast. The out-degree is larger than the corresponding in-degree. By putting directions on edges of G.that is why graph G is not Eulerian. 7. If. The difference ∆ (equal to out-degree minus in-degree) for any vertex is zero for such vertices.2. all shown in red.410 Eulerian Graphs & the Chinese Postman Problem 7. hence ∆ is positive. A solution to this problem for an undirected graph G is not helpful even if the directed graph is derived from the same undirected graph as shown in the bottom left and right diagrams of Fig. If all vertices in a directed graph D have ∆ equal to zero then graph D is Eulerian.5 The Chinese Postman Problem for Directed Graphs We discuss here the Chinese Postman Problem for a directed graph. 7. 3.

It is converted into a directed graph D as shown in the top right diagram. The Eulerian walk in G is not an Eulerian walk in the directed graph D. It is obvious that graph G is not Eulerian. After duplicating certain edges in graph G it is transformed into an Eulerian graph as shown in the bottom left diagram.1: We again show an undirected graph G in the top left diagram. -.5. & zero ∆ 8 9 4 1 6 1 7 9 8 7 4 6 2 3 5 2 3 An Eulerian Walk for the un-directed G: Not applicable for directed graph D 5 We duplicate certain edges as shown in red color. Figure 7.The Chinese Postman Problem for Directed Graphs 411 8 9 4 1 6 1 1- 8 7 1+ 1- 9 4 2+ 7 1- 6 1- 2 3 Graph G: All six odd vertices are shown in red while even vertices are shown in green 5 2 0 1+ 5 0 3 A Directed graph D derived from the un-directed graph G: Difference ∆ between out-degree and In-degree for each vertex is indicated: There are vertices with +. . The resulting multi-graph H has now become Eulerian. An Eulerian walk in G is shown in the bottom right diagram.

.5.2: We show a directed graph G where there is only one vertex (shown in red) where the out-degree is larger than the in-degree.412 Eulerian Graphs & the Chinese Postman Problem 8 0 0 8 7 1- 9 9 7 1 0 3+ 4 6 0 1 4 6 2 1- 5 3 0 1- 2 3 5 The Difference ∆ between out-degree and In-degree is indicated with each vertex: There is only one vertex with positive ∆ shown in red in this graph G We find a shortest path from each orange vertex (where ∆ is negative) to a vertex with positive ∆ shown in red color 0 8 2 2 0 9 7 0 0 5 5 4 1 0 4 6 0 3 7 2 0 5 3 0 0 The lengths of different shortest paths between orange vertices and the red vertex in graph G Edges in each shortest path are duplicated: The resulting graph H has now become Eulerian Figure 7. All other vertices have either out-degree equal to in-degree (shown in green) or less than in-degree (shown in orange).

5. This has been illustrated in Fig. that single vertex has ∆ exactly equal to +3. duplicating edges on shortest paths ensures that the number of extra edges (added) is minimized. the problem is to minimize the cost also. 7. Vertex 4 has ∆ equal to −3 while vertex 10 has ∆ equal to −1. Finding a maximum flow efficiently (in polynomial time) in this (special) network graph N is by itself an interesting problem.5. The solution is simple: we create a weighted bipartite graph B where the weight of an edge from an orange vertex x to a red vertex y in B signifies the weight of the shortest path between the orange vertex x to red vertex y in graph G. Please note that there should be three paths coming out of vertex 4 in order to increase its ∆ from −3 to zero. Further adding vertices s and t and edge capacities (middle right diagram) the problem is converted into a flow problem. 7.The Chinese Postman Problem for Directed Graphs 413 the Chinese Postman problem as shown in the same figure. vertex 9 has ∆ equal to +2. Please note that all the three shortest paths originating from vertices with ∆ equal to minus 1 are terminating at a single vertex. all shown in red color. An efficient solution of the Chinese Postman Problem in a directed graph G is thus reduced to the following steps: . 7. Finding a maximum flow with minimum cost in a network graph N will ensure that the graph becomes Eulerian and the number of edges (which are) duplicated are minimized at the same time. this will make the new directed graph Eulerian. 7. there will be an edge from every orange vertex to every red vertex in B signifying that there will be a path from every orange vertex to every red vertex in graph G.3. Please note that this will be a complete bipartite graph. The maximum flow in this graph (as shown in the bottom diagram) will provide the required information. Similarly there are three vertices with ∆ positive. We show a directed graph in the top diagram of Fig.5.3. In order to make ∆ of every vertex zero (thus converting the graph G into an Eulerian graph) we should find out which path originating from an orange vertex should terminate at which red vertex.5. There are two vertices (4 & 10) with ∆ negative shown in orange color. The graph G will become Eulerian but the cost in terms of number of edges of G that are duplicated may be high. We create a bipartite graph B consisting of an A partite (consisting of all orange vertices) and a B partite (consisting of red vertices) as shown in the middle left diagram of Fig.3). Similarly two paths should be terminating at vertex 9 to convert its ∆ from +2 to zero (please see Fig.4.

5. 6.3: A maximum flow in a network graph N (middle right diagram) helps us find how to convert a directed graph into an Eulerian graph. & 9 6 10 2 0 1- 1+ 5 0 1 4 6 10 9 3 ∞ 3 0 1 ∞ 1 4 ∞ The path requirement can be fulfilled by finding maximum flow in this network graph: edge capacities are indicated with each edge s 1 6 10 ∞ ∞ ∞ 1 t 2 ∞ ∞ 3 1 1 9 4 ∞ s 1 6 1 t Maximum flow tells us which path coming from 4 &10 should terminate at which vertex (1. 6. or 9) to make the graph Eulerian 10 ∞ ∞ ∞ 2 9 Figure 7.414 Eulerian Graphs & the Chinese Postman Problem 8 2+ 0 9 4 3- 7 0 1 1+ The Number of paths coming out of 4 & 10 should be exactly equal to the number of paths going in 1. .

4. It will be interesting to derive the time complexity of the above algorithm and compare it with that of the algorithm used to solve the Chinese Postman Problem for undirected graphs. Corresponding to every such (x.The Chinese Postman Problem for Directed Graphs 1. the capacity of these edges will be infinite.5.5.2. The cost of any other edge in N will be the corresponding weight in the bipartite graph. Problem 7. it means that there should be a path between every pair of vertices in graph D. 7. The weight of an edge from an orange vertex x to a red vertex y in B signifies the weight of the shortest path between the orange vertex x to red vertex y in graph G. Convert bipartite graph B into a network flow graph after adding vertices s & t.2. Here there are four vertices with ∆ negative (shown in orange color) while . Create a weighted bipartite graph B consisting of orange vertices as an A partite and red vertices as the B partite. A minimum cost maximum flow in the network graph N will provide us the desired pairs of orange/red vertices in graph G. while y belongs to partite B of red vertices). The examples of directed graphs that we have considered in this section have one thing in common: If we sum ∆ of all vertices in a graph it comes out to be zero.2.1. Identify orange and red vertices in G. Show that a necessary & sufficient condition for a directed graph D to have a solution to the Chinese Postman Problem is that graph D should be strongly connected. 415 2. duplicate edges in the shortest path from vertex x to vertex y in graph G. The capacity of an edge from s to an orange vertex x is equal to ∆(x). y) pair (x belongs to partite A consisting of orange vertices. Problem 7. 3. Problem Set 7.2. The costs of these edges will be zero. We show a directed graph in the top left diagram of Fig.3. The capacity of an edge from a red vertex y to t will be equal to ∆(y). Problem 7.2. Is this a coincidence or will every directed graph possess this property? Discuss briefly.

4: We show capacity/cost associated with each edge in the middle network graph. .5.416 Eulerian Graphs & the Chinese Postman Problem 8 2+ of s t t hs C o t Pa nd s Fi orte Sh 0 9 4 3- 7 0 1 1+ 6 10 1- 1+ 1 3 2 0 5 0 4 5 3 3 6 1 6 4 Ma Mi ximu n im m um Flo Co w st 0 10 ∞/3 ∞/5 1 1/0 9 s 8 2+ 0 3/0 4 ∞/3 6 1/0 1/0 t 10 ∞/1 ∞/6 ∞/4 2/0 9 4 3- 7 0 9 1+ al tim ph Op Gr a An rian le Eu 1 1+ 6 10 1- 2 0 5 0 3 0 Figure 7. A maximum flow at minimum cost in this network graph helps us find how we can convert a directed graph into an Eulerian graph by duplicating minimum number of existing edges.

We describe how we can convert a directed graph G into an Eulerian graph H by duplicating minimum number of edges of G. Note that vertex 4 is duplicated in the Bipartite graph? The Difference ∆ between outdegree and In-degree is indicated with each vertex for the graph G 0 1 6 2 7 7 4 3 5 3 2 2 1 3 4 2 6 3 3 0 8 9 4 0 7 0 6 4 0 1 4 0 6 0 7 4 8 9 2 0 5 3 0 We find a minimum cost perfect matching in the complete bipartite graph We have duplicated edges in each path corresponding to the minimum cost perfect matching: The ∆ for each vertex becomes 0: The new graph H has now become Eulerian Figure 7.5: We show a directed graph G where there are three vertices.5. where the out-degree is less than the corresponding in-degree. . Green vertices have out-degree equal to the in-degree.The Chinese Postman Problem for Directed Graphs 8 1+ 1- 417 9 4 7 1 16 2 7 7 4 3 5 3 2 2 1 3 4 2 6 3 3 1 1- 2+ 6 4 6 1- 7 4 4 2 0 5 0 8 9 3 1+ We find a shortest path from every vertex with negative ∆ to each vertex with a positive ∆. Vertices where the out-degree is larger than the in-degree are shown in red color. shown in orange color.

2. Problem 7. Problem 7.6. We need to convert these graphs into Eulerian graphs by duplicating minimum number of existing edges of these graphs.418 Eulerian Graphs & the Chinese Postman Problem there are three red vertices with ∆ positive.6: We show two directed graphs which are not Eulerian. We show two directed graphs in Fig.4.5. Solve the Chinese Postman problem for both these graphs using an efficient algorithm.2.5. We demonstrate an alternate scheme to solve the Chinese Postman problem. 7. 8 2+ 0 8 7 1+ 3+ 0 9 4 3- 9 4 3- 7 1+ 1 1+ 6 0 0 1 10 1- 6 0 10 2 0 1- 5 0 2 0 5 0 3 0 3 0 There are three vertices where out-degree is larger than in-degree. two vertices where in-degree is larger than out-degree 3 1 6 4 5 4 7 4 6 7 7 10 6 8 9 10 9 5 9 Figure 7. Derive the time complexity of the algorithm (that we have described in the text) which can be used to solve the Chinese Postman prob- .5. two vertices where in-degree is larger than out-degree There are two vertices where out-degree is larger than in-degree. Describe an algorithm behind this demonstration and calculate the time complexity of this algorithm in comparison with that of the algorithm discussed in the text (based on minimum cost maximum flow algorithm).

5.8.6. please note that the in-degree of every vertex in this diagram is exactly equal to its corresponding out-degree.7.The Chinese Postman Problem for Directed Graphs 419 lem for directed graphs. if that is not possible then we should allow him (the Chinese Postman) to traverse the red edges at least once and come back after traversing a minimum number of edges in this graph.7. Problem 7. This graph is not Eulerian. We show a directed graph in the left diagram of Fig. Figure 7. the graph shown in the left diagram is Eulerian while the one shown in the right diagram is neither Eulerian nor it is strongly connected. The Chinese Postman is supposed to start delivering his mail from a specific . We show a directed graph G in the left diagram of Fig. The Chinese Postman is supposed to deliver mail in red streets only. Problem 7. We need to find an Eulerian circuit consisting of edges corresponding to red streets alone. Problem 7. 7. thus the graph is Eulerian. show the detailed working of this algorithm on this graph. 7. Please describe an efficient algorithm to solve this problem.2. The Chinese Postman is supposed to distribute mail on streets shown in red color only.5.5. it is neither strongly connected yet it may be possible for the Chinese Postman to distribute mail in (only red) streets while traversing minimum number of extra graph edges.8. Compare its value with the time complexity of the algorithm used for undirected graphs.7.7: We show two directed graphs. We show another directed graph in the right diagram of Fig. 7.2.2.5.

5. The right diagram of this figure shows a transformation of graph G into another graph where certain edges of G are duplicated. 7. If this is not possible then he should traverse each edge at least once and also minimize the number of edges which are traversed more than once. Does the right diagram of Fig.5. .8: We show a directed graph G in the left diagram.8 provide an optimal solution to our problem for the graph shown in the left diagram? Discuss briefly? 5 7 4 1 End here 5 6 7 4 3 1 End here 6 3 10 10 2 8 Start here 9 2 8 Start here 9 Figure 7. Show that an Euler trail is possible from vertex 8 to vertex 3 in this graph. Design an efficient algorithm to efficiently solve the above problem.420 Eulerian Graphs & the Chinese Postman Problem point and end his delivery job at another point as shown in the figure. the green vertices have out-degree equal to the corresponding out-degree. The right diagram shows a transformation of graph G into another graph where it is possible to find an Eulerian Path from vertex 8 to vertex 3. the red vertices have in-degree smaller than the corresponding out-degree while the orange vertices have in-degree larger than the out-degree. The postman is supposed to traverse every edge of this graph exactly once.

5.9: Eu An O ler p ian tim Gr a l ap h n Fi d M um im in o Fl w 0 t 0 .The Chinese Postman Problem for Directed Graphs 421 8 2+ 0 8 7 0 2+ 0 9 4 3- 9 4 7 0 1 1+ 6 10 2 0 1- 1+ 3- 1 1+ 6 10 1- 1+ 5 0 0 2 3 0 5 0 3 0 0 8 2+ 9 7 4 30 0 1+ 1 s 0 6 0 1+ 10 1- 2 0 5 0 3 0 Figure 7.

5.11: .422 Eulerian Graphs & the Chinese Postman Problem a b a +1 b s t +2 s t -1 d c F Eu ind l e r Op ian tim Gr al ap h d c -2 ∞/3 2/0 a ∞/2 1/0 a +1 b c x 1/0 y t ∞/1 ∞/2 2/0 s +2 t -1 s a 2 s 1 1 d 1 c 1 3 b 1 d c -2 1 i M um n im ow Fl t Figure 7.5.10: 0 8 2+ 0 9 7 4 31+ 0 0 8 2+ 9 7 0 0 1+ 1 s 0 6 0 0 0 t 1+ 1 s 0 4 31+ 6 0 10 0 0 t 1- 2 0 10 1- 5 0 2 0 5 0 3 0 3 0 Figure 7.

2 8.Chapter 8 Hamiltonian Graphs 8.3 8.6 Introduction Prior Knowledge Hamiltonian Graphs Bipartite Hamiltonian Graphs Some Theoretical Claims A Categorization of Hamiltonian Graphs .5 8.4 8.1 8.

This means that no vertex should be a cut vertex and no edge is a . 8. It will be interesting to find sufficient conditions for a graph to be connected.1 Prior Knowledge Necessary Conditions for a Connected Graph Let us start with a familiar and simple example of a connected graph. Having edges less than this limit will be a sufficient condition for a graph to be disconnected.424 Hamiltonian Graphs 8. we are unaware of such a characterization so we shall be talking about necessary and sufficient conditions separately. we can easily differentiate between necessary and sufficient conditions.2 Necessary Conditions for a Hamiltonian Graph A graph can not be Hamiltonian unless it is connected. contain a Hamiltonian Path). 8. But then there are graphs where the number of edges is more than this critical number but still these graphs are not connected. We begin by using constructive proof techniques for certain sufficient conditions.2 8. In case of Eulerian Graphs where we have a neat classification. Thus a Hamiltonian graph should be more than a tree graph. It should be connected and there should be a cycle spanning all vertices. That means this (q ≥ p − 1) will be a necessary condition for a connected graph.2. although a loose necessary condition. however. the same techniques can be used to find a Hamiltonian Cycle in a graph where a Hamiltonian Cycle is guaranteed (by the sufficient conditions) to exist. we shall address the problem of finding a Hamiltonian cycle and a Hamiltonian Path in a graph.1 Introduction In this chapter. but we shall discuss this issue later in this topic. In Hamiltonian Graphs however. A tree graph is connected but it can not be Hamiltonian as it has no cycles (it can. there should be at least p − 1 edges otherwise the graph will be guaranteed to be disconnected. We shall also be considering certain conditions (necessary and then sufficient) for a graph to be Hamiltonian.2. Students of this topic sometimes get confused between necessary and sufficient conditions. For a graph (with p vertices) to be connected. It is thus a necessary condition. Such a graph is known as a tree where each vertex is a cut vertex and each edge is a bridge edge.

It will be useful if a learner solves this problem by himself or herself. 2 It can also be shown that if the maximum degree of any vertex in a graph is equal to (p−2) then you can always draw a disconnected graph.3 A Loose Sufficient Condition for a Hamiltonian Graph Let us think of a familiar graph where a Hamiltonian Cycle is guaranteed to exist and we can find this cycle using a trivial algorithm. We can certainly find sharper necessary conditions as are explained in most other text books. (How?) This 2 implies that the condition (minimum degree ≥ (p−1) ) is certainly a sharp 2 condition for connectedness. and Hamiltonian (see Concept Map . This loose sufficient condition will be used in later pages to derive sharper sufficient conditions. It is interesting to note that complete connectedness is not a sufficient condition for a graph to be Eulerian. Again these will be necessary conditions as there are graphs which satisfy these conditions but are not Hamiltonian. 8.1).4 Sufficient Condition for a Connected Graph Before finding a tighter sufficient condition for a Hamiltonian graph. Any permutation of the vertices of this graph will give you a Hamiltonian Cycle (see Fig. On the other hand if a graph contains a bridge edge (or a cut vertex) then it is certainly not Hamiltonian.2.2. The minimum degree of any vertex in a graph should be at least (p−1) for a graph to be connected.2. The other side of the spectrum is a completely connected graph where every vertex is connected to every other vertex. 8. 8.Prior Knowledge 425 bridge edge in a Hamiltonian graph. Obviously a completely connected condition would be just too loose for a graph to be just connected. although it is perhaps an overkill and thus not a sharp sufficient condition.2. let us solve a much familiar and also simpler problem of finding a sufficient condition for a graph to be connected. 8. A tree graph is just connected and belongs to one side of the extreme of connected graphs.5 A Concept Map We show a concept map indicating some necessary and sufficient conditions for a graph to be connected. Thus the completely connected property is a sufficient condition for a graph to be Hamiltonian. Eulerian.

Two Hamiltonian Cycles are shown in the completely connected graph. We shall 2 . in fact we can draw a graph with the above property which is clearly not Hamiltonian. Subsequently we can design efficient algorithms to actually find an Euler Cycle in a graph provided the graph satisfies the given conditions. each cycle corresponds to a different permutation of vertices of the graph. Summary It will be interesting to summarize that what is possible (or not possible) and at what cost. In case of connected or Hamiltonian graphs it is not possible to find necessary and sufficient conditions. In case of connected property again it is not possible to devise necessary and sufficient conditions. The graph may still not be Hamiltonian.1: A completely connected graph of 6 vertices. A simple Breadth First Search (or any traversal algorithm) can solve this problem efficiently. If the 2 minimum degree in a graph is equal to or more than (p−1) then certainly the 2 graph will be connected. In case of Eulerian graphs we can find a nice characterization that is necessary and sufficient for a graph to be Eulerian. And a minimum degree slightly larger than (p−1) is a sufficient condition for a graph to be Hamiltonian. We can still draw a disconnected graph if the minimum degree of a node in the graph is less than or equal to (p−2) . Thus for a general graph which does not satisfy the sufficient conditions for a Hamiltonian Cycle it is extremely hard to design an algorithm which can find a Hamiltonian Cycle (provided such a cycle exists) or out puts in negative if a Hamiltonian Cycle does not exist. however it is possible to design efficient algorithms which can determine if a given graph is connected or not.2.426 Hamiltonian Graphs 1). u5 u4 u5 u4 u6 u3 u6 u3 u1 u2 u1 u2 Figure 8. If the number of edges is less than p − 1 then the graph will certainly be disconnected. Interestingly a minimum degree equal to (p−1) is a sufficient condition for 2 graph to contain a Hamiltonian Path.

. A concept map of certain necessary and sufficient conditions for a graph to be connected.1. and Hamiltonian. Eulerian.Prior Knowledge 427 Concept Map 8.

Deciding when a specific tool can be used and where it can not be used is certainly a valuable learning experience. the experience that you will gain will provide you powerful tools that we shall use in this chapter in solving various problems.428 Hamiltonian Graphs now provide constructive proof techniques to prove a number of sufficient conditions for a graph to be Hamiltonian. You just need some prior knowledge in addition to some common sense. and as it is obvious it is a function of the order of the graph. Problem 8. Let us start with simpler problems. Suppose G is a line graph then G2 has a Hamiltonian Cycle.3 Hamiltonian Graphs We shall make a drastically different start.1. The diameter of a line graph is proportional to the size of the graph but still G2 is Hamiltonian. Please see Fig. Please note that the diameter of a line graph is p − 1.3. You must solve this problem before moving forward. 8. Problem Set 8. Suppose G is a star graph then G2 has a Hamiltonian Cycle.3.1. Problem 8. we shall make you solve a number of related but simple problems.1. We shall also encourage you to solve a puzzle. For each of the problem you are supposed to design a formal proof. It will be interesting to explore if this is a more general result: G2 of any tree is Hamiltonian? Either show it or find a counter example. No graph (other than a line graph) can have a diameter as large as p − 1. 8.1. If the diameter of a graph G is k then Gk will be a completely connected graph and that is why G2 of a star graph is Hamiltonian. Please note that the diameter of a star graph is equal to 2 irrespective of the order of the graph. It is important for the learner to gain confidence: You can discover and create new knowledge. . Instead of simply describing the algorithms. 8. The k th power Gk of a graph G is a graph with same number of vertices as in G in which two vertices are adjacent if and only if they are at most d distance apart from each other in G.2.2. Let us now try to solve a problem which belongs to the so called critical activity section. For each of the above two problems it will be useful if the learner actually draws such graphs and then discovers the answer himself or herself.1 and Fig.

3.3. instead of resisting a Hamiltonian Cycle. Now the degree of one terminal vertex will be larger and the other smaller. you are adamant to form one as soon as the first opportunity arises then what will be the minimum degree of the two terminal vertices which will guarantee a Hamiltonian cycle? Now come back to the previous problem where we try to resist a Hamiltonian Cycle as far as possible. But let us relax the condition that the degree of the two terminal vertices should be the same. You may like to solve the puzzle for the graph shown in Fig. If. You are supposed to design an efficient algorithm which will output the actual Hamiltonian Cycle in G.Hamiltonian Graphs 429 8. You are allowed to insert new edges between u1 and any of the intermediate nodes. Again this will be a sufficient condition for the modified graph to be Hamiltonian. That will be a sufficient condition for this graph to be Hamiltonian. . Again find what will be the minimum sum of the two degrees (of the terminal vertices) when it will be impossible to resist a Hamiltonian Cycle.2 8. You can not insert an edge between the two terminal nodes otherwise a Hamiltonian Cycle will immediately be formed (we already have a Hamiltonian Path between the two terminal vertices). The problem is to find out (by an actual trial and error drawing) the minimum degree of the two terminal vertices when it will be impossible to resist a Hamiltonian Cycle.3.1. While inserting an edge from u1 to an intermediate vertex (and then an edge from up to an intermediate vertex) try your best that a Hamiltonian Cycle is not formed.1 A Puzzle: We are given a line graph with p nodes. While inserting edges you should keep the degree of u1 and up exactly the same.2 Actually Finding a Hamiltonian Cycle in the Puzzle: Let us look at the same problem from an algorithmic point of view: You are given a graph G which contains a Hamiltonian path between two non adjacent vertices u and v. where 2 ≥ i ≥ p − 1 as shown in Fig. 8. Assume that the sum of the degrees of u and v is equal to or larger than p. the starting and end node of the line graph are designated as u1 and up respectively. Similarly you can insert edges between up and any of the intermediate nodes. 8. The intermediate nodes are labeled as ui . Also assume that the Hamiltonian Path between u and v is already provided as an input.3.

A Hamiltonian Path between vertex u and v.1: A line graph with six nodes. Algorithm 60: Find a Hamiltonian Cycle in a graph G where deg(u)+ deg(v) ≥ p for a pair of non adjacent vertices u & v in G input : A Graph G.. We should add extra edges in this graph but do our best to avoid a Hamiltonian Cycle.3..430 Hamiltonian Graphs u1 up u1 up u2 u3 u4 u5 u2 u3 u4 u5 A graph where the degree of every vertex is 2 except for two terminal vertices An edge between the terminal vertices will instantly create a Hamiltonian Cycle u1 up u1 up u2 u3 u4 u5 Then u2 u3 u4 u5 Suppose we add an edge as shown. Now we should not add an edge which will now create a Hamiltonian Cycle Figure 8. . output: A Hamiltonian Cycle in G 1 Use the knowledge and expertise that you have gained while solving the puzzle.

.3.2: Without the red dotted edge there is no Hamiltonian Cycle in each of these graphs.Hamiltonian Graphs 431 u v u v u v u v Figure 8.

3. In the top graph the degree sum of the two special vertices is equal to the number of nodes in the graph.3: We are given three graphs where a Hamiltonian path exists between two special vertices. in the bottom graphs the degree sum of the two special vertices is less than the number of nodes in the graph.432 Hamiltonian Graphs u5 u4 3 u6 u3 3 u1 u2 u4 u5 u4 2 3 u6 u4 2 3 u6 u3 u5 u2 u1 u2 u1 u3 Figure 8. . Please note that a Hamiltonian Cycle exists in each of these graphs. The Hamiltonian path between the two vertices is also indicated by shaded lines in each graph.

Is this not strange that a graph contains a Hamiltonian Cycle but your algorithm can not find it? Let us now describe the details of the algorithm which solves the problem outlined above.Hamiltonian Graphs 433 Apply your algorithm on the graphs given in Fig. a Hamiltonian path between u and v is also given. Find a vertex x such that vertex 1 is adjacent to vertex x + 1 while vertex p is adjacent with vertex x in G.4. .5). 8.3. p − 2. Interestingly there may be graphs where the condition (that deg(u) + deg(v) ≥ p for a pair of non adjacent vertices u & v in G) is not met yet our algorithm will be able to find a Hamiltonian cycle as shown in the same figure. 8. It is interesting to note that our algorithm also has a serious short coming: a graph may have a Hamiltonian Cycle but we can not find it (as shown in Fig. . The Hamiltonian Cycle will be 1. output: A Hamiltonian Cycle in G 1 2 3 Let us index the vertices of the Hamiltonian path like 1. 3. . 2. . 8. .3. . Algorithm 61: Find a Hamiltonian Cycle in an un-directed graph G where deg(u) + deg(v) ≥ p for a pair of non adjacent vertices u & v in G. The index of u will be 1 while that of v will be p.3. p. please note that each of the graphs does indeed contain a Hamiltonian Cycle (which can easily be found by hit and trial method). . x + 1. p − 1.3 in order to find a Hamiltonian Cycle in each graph. input : A Hamiltonian Path between vertices u and v where deg(u) + deg(v) ≥ p in an un-directed graph G. . . . 1. p. Let us see where and why your algorithm fails and where it does find a Hamiltonian Cycle. The working of the algorithm is illustrated in Fig. 2. x.

3.4: Intermediate stages of how Algorithm 61 works to find a Hamiltonian Cycle in a graph G where the condition (that deg(u) + deg(v) ≥ p for a pair of non adjacent vertices u & v in G) is met and where there is a Hamiltonian path between vertices u and v in G.434 Hamiltonian Graphs u 2 1 v 3 6 (1) 4 5 2 1 ) (2 2 1 3 6 3 6 x x+1 4 5 (3) x x+1 4 5 Figure 8. .

but a Hamiltonian Cycle exists in this graph as shown by blue lines.Hamiltonian Graphs 435 Degree = 2 u5 u 3 2 u5 u u6 u3 v u6 u3 v 2 u1 u2 u1 u2 A Hamiltonian Cycle exists as shown by blue lines yet our algorithm will not be able to find it The Condition is not met yet our algorithm will be able to find the Hamiltonian Cycle shown in blue Figure 8.3.5: A graph G shown (left diagram) where the condition (that deg(u) + deg(v) ≥ p for a pair of non adjacent vertices u & v in G) is not met and the Algorithm 61 will NOT be able to find a Hamiltonian cycle. . A graph G shown where the condition (that deg(u) + deg(v) ≥ p for a pair of non adjacent vertices u & v in G) is not met yet the Algorithm 61 will be able to find a Hamiltonian cycle as shown in the right diagram.

Then. Then. If this hypothetical Hamiltonian Cycle does not pass through original graph edges then insert Extra edges as shown in (red) in this figure to create a cycle consisting of some original and some Extra edges. p and then going back to 1. passing through 2. 8.7.3. that there will not be a Hamiltonian Cycle in the graph G (without that extra edge) as shown in the left diagrams of the same figure. This intuition provides a powerful technique which can be used in most of Hamiltonian finding algorithms in Hamiltonian graphs. G+uv is Hamiltonian means that there is a Hamiltonian Cycle in the graph G+uv.3. It provides a couple of very powerful theoretical results and useful algorithmic tools which can be used to find a Hamiltonian cycle in a certain category of graphs. This essentially means that there will be a Hamiltonian Path starting from vertex 1 and terminating at vertex p in the original graph G. Now visualize a hypothetical Hamiltonian Cycle in this graph starting from 1. Consider a graph G where the degree sum of any pair of vertices in G is equal to or larger than p as shown in Fig. .6.4 Bondy & Chvatal’s Theorem: Now let us come back to a text book theorem which says that a graph G is Hamiltonian if and only if G + uv is Hamiltonian provided vertices u and v are non adjacent and deg(u) + deg(v) ≥ p. we claim.3. Before attempting to prove this theorem let us look at the problem more closely and try to relate it to the previous knowledge that we have acquired.3. 8.6. . Assume that a graph G has two vertices where the degree sum of these two vertices is more than p as shown in the figure below.3. that there will be a Hamiltonian Cycle in graph G (without that extra edge) as shown in the right diagrams of the Fig. There are two possibilities: . 4. .436 Hamiltonian Graphs 8. 3.3 Basic Intuition Let us carefully look at the ramifications of Algorithm 61. Now assume that after inserting that extra edge between the two vertices the resulting graph does not have a Hamiltonian Cycle. we claim. 8. We insert an edge between these two vertices and assume that in the resulting graph there is a Hamiltonian Cycle as shown in the top left diagram of Fig. 8. You know that we can always delete Extra edges one by one (using the techniques that we have recently acquired) forcing the Hamiltonian Cycle to divert through only graph edges.

u1 up u1 up u2 u3 u4 u5 u2 u3 u4 u5 Consider a graph where the degree sum of every pair of vertices is equal to or larger than p We force a Hamiltonian Cycle in this graph: If there no graph edges available then we insert Extra edges as shown in red Figure 8.7: We make a number of claims in this diagram: A graph G is Hamiltonian if and only if the graph G plus an edge between two non adjacent vertices of G is Hamiltonian provided certain conditions are met. We do not know if a HAM cycle exists in graph G u1 up u1 up u2 u3 u4 u5 u2 u3 u4 u5 We insert a direct edge from u1 to up Assume that now a HAM Cycle does Not exist in the resulting graph Then Then We insert a direct edge from u1 to up Assume that now a HAM Cycle does exist in the resulting graph u1 up u1 up u2 u3 u4 u5 u2 u3 u4 u5 Now we remove the direct edge from u1 to up A HAM Cycle will NOT exist in graph G? Now we remove the direct edge from u1 to up A HAM Cycle will still exist in graph G? Figure 8.3.Hamiltonian Graphs 437 u1 up u2 u3 u4 u5 A graph G where the degree of u1 and up is equal to or more than p.3. .6: We make a number of claims in this diagram: A graph G is Hamiltonian if and only if the graph G plus an edge between two non adjacent vertices of G is Hamiltonian provided certain conditions are met.

3. As the edge uv is not part of the Hamiltonian Cycle so it is possible to remove this edge and still a Hamiltonian Cycle will exist in G. This means the possibility. . This also means that if we remove the edge uv even then there will be a Hamiltonian Cycle in G provided there was a Hamiltonian Cycle in G + uv. Under such conditions we can map this problem to the last problem that we have discussed except that the actual Hamiltonian Path between u and v is not provided here. 2. 8. Under such conditions a Hamiltonian Path between u and v may not exist.8. Again notice the figure of eight in this diagram. Again it will be interesting to relate this algorithmic problem with the last such problem that we have discussed. 8. Such a possibility is shown in the right diagram of Fig. The Hamiltonian Cycle passes through the edge uv.3.8. 8.6 Closure of a Graph: A closure c(G) of a graph G of order p is a graph obtained from G by recursively joining pairs of non adjacent vertices u and v whose degree sum is more than or equal to p until no such pair remains in G. The edge uv does not make any difference.3. If there is a graph G in which there are two adjacent nodes u and v such that deg(u) + deg(v) ≥ p then G is Hamiltonian if and only if G + uv is Hamiltonian. 8. Such a possibility is indicated in the left diagram of Fig.438 Hamiltonian Graphs 1. It will be instructive to actually draw such a graph. Then there will certainly be a Hamiltonian Path between the vertex u and vertex v. that the graph G is not Hamiltonian while the graph G + uv is Hamiltonian.3.5 Summary Let us look at what we have really understood so far. does not exist. its inclusion does not convert a non Hamiltonian graph into Hamiltonian and its removal does not convert a Hamiltonian graph into a non Hamiltonian graph provided other conditions are also met. It can now be proved that a graph G is Hamiltonian if and only if its closure c(G) is Hamiltonian. The Hamiltonian Cycle does not pass through the edge uv. How about if we are supposed to design an algorithm to find a Hamiltonian Cycle in the graph G provided the above stated conditions are true.

3. 2. Extra edges inserted between very pair of non adjacent vertices u and v where deg(u) + (v) ≥ p. We have also stated that for a completely connected graph. But remember the Hamiltonian Cycle that we have found belonged to the closure of G. The Hamiltonian Cycle in c(G) will pass through some of the original edges of the graph. How to find a Hamiltonian Cycle which passes through E(G) alone is an interesting problem. These edges were not part of G. You can verify that a Hamiltonian path does not exist between these two vertices in the right diagram. finding a Hamiltonian Cycle is a trivial problem.7 Ore’s Theorem: The usefulness of the above result becomes obvious when in a graph G every pair of non adjacent vertices satisfies the above mentioned condition. degu + v ≥ p.Hamiltonian Graphs u5 u4 u5 u4 439 u6 u3 u6 u3 u1 u2 u1 u2 Figure 8. that means E(G). the same Hamiltonian Cycle may not exist in the actual graph G? This is because of the fact that c(G) consists of two types of edges: 1.3. We have already stated that complete connectedness is a very loose sufficient condition for a graph to be Hamiltonian. It can also be used constructively . But it may also pass through some of the extra edges belonging to the second type. The right diagram shows a graph where the Hamiltonian Cycle does not pass through the edge between vertex u1 and u2 . that is. 8.8: A graph containing a Hamiltonian Cycle which passes through the edge between vertex u1 and u2 is shown in the left diagram. We call these edges extra edges. then the closure c(G) will be obviously be a completely connected graph. Edges which were originally present in G. The following algorithm can be used to find a Hamiltonian Cycle in graph G.

In the second graph H the degree sum of each pair of non adjacent vertices is not p but the closure of H is a completely connected graph. Discuss briefly where it is possible to use Algorithm 62 without any modification to find a Hamiltonian Cycle. This sufficient condition was originally discovered by Ore. Also pinpoint the graph where it may not be easy to find a Hamiltonian Cycle using the knowledge that we have acquired until now. Algorithm 62: Find a Hamiltonian Cycle in a graph G where deg(u)+ deg(v) ≥ p for very pair of non adjacent vertices u & v in G. input : An un-directed Graph G where deg(u) + deg(v) ≥ p for very pair of non adjacent vertices u & v in G. In the third graph J.2. Problem Set 8. We find a Hamiltonian Cycle in graph H. thus if there is a Hamiltonian Path between u and v in H then we can find a new Hamiltonian cycle in H which does not pass through the edge uv (use Algorithm 61).2. in the first graph G the degree sum of each pair of non adjacent vertices should be at least p. and last but not the least discuss where Algorithm 62 have to be modified in order to find a Hamiltonian Cycle.this closure is our starting graph H.1. Problem 8. output: A Hamiltonian Cycle in G 1 2 3 4 We find a closure of graph G known as c(G) . Draw three different graphs.440 Hamiltonian Graphs to prove that if deg(u) + deg(u)v ≥ p for every non adjacent pair of vertices in a graph G then G is Hamiltonian. We know that deg(u) + deg(v) ≥ p. . the degree sum of each pair of non adjacent vertices is not equal to p. This cycle may pass through some of the graph G original edges and may also pass through some of the edges not present in G (but are present in H) known as extra edges. This is certainly a tighter sufficient condition for a graph G to be Hamiltonian as compared to the complete connectedness of a graph. while there are extra edges in the Hamiltonian Cycle in graph H do Remove an extra edge uv from the cycle and from graph H thereby creating a Hamiltonian Path between two vertices u and v in the new graph H.

3. Note that the Hamiltonian Cycles passes entirely through extra edges. the graph is shown in the left diagram. however. The closure of this graph is a completely connected graph as shown in the right diagram of the same figure. interestingly the Hamiltonian Cycle passes entirely through extra edges.10.3. We remove extra edges one by one and each time find a new Hamiltonian Cycle.Hamiltonian Graphs 441 Dirac’s Theorem: A graph G where the degree of each node is more than or equal to p/2 is Hamiltonian. u5 u4 u5 u4 u6 u3 u6 u3 u1 u2 u1 u2 Figure 8. The Hamiltonian Cycle is also indicated in this diagram. 8. The closure of this graph is.3. a completely connected graph as shown in the top right diagram. Note that the closure is a completely connected graph containing a Hamiltonian Cycle also shown in the right diagram.12. We apply Algorithm 62 to this graph and show a number of intermediate results in Fig. become important in the next example. The closure of this graph thus contains a Hamiltonian . the degree sum of a pair of non adjacent vertices is less than 6 in this graph. Example 2: We show a graph G in the top left diagram of Fig.3. A formal proof of this theorem can easily be derived from the above discussion.9. 8. A constructive proof can also be designed on similar lines. however. The closure of this graph is shown in the right diagram. 8. the extra edges removed are indicated by dotted lines while a Hamiltonian Cycle is shown by bold lines in Fig. 8.9: A graph G where the degree sum of each pair of non adjacent vertices is equal to p. Why? The order will.10. the degree of some of the nodes is also indicated in this diagram. Please note that the order in which extra edges are removed is not very important in this example. Example 1: We show a graph G where the degree of each node is equal to p/2 in the left diagram of Fig.3.

. We repeat this process until all extra edges are removed.442 Remove One of the Extra edges shown in orange Hamiltonian Graphs Find Hamiltonian cycle (blue color) in the remaining graph u5 u4 u5 u4 u6 u3 u6 u3 u1 u2 w th e gr ap h ra u1 u2 u1 u4 R Remove One more Extra edge Find Hamiltonian cycle (blue color) in the remaining graph ed u1 u4 u6 u3 u6 u3 u5 u2 w th e gr ap h u5 ra u2 u4 u1 R Remove One more Extra edge Find Hamiltonian cycle (blue color) in the remaining graph ed u4 u1 u6 u3 u6 u3 u5 u2 u5 u2 u6 u4 R ed u3 u1 w th e gr ap Remove One more Extra edge ra Find Hamiltonian cycle (blue color) in the remaining graph h u3 u1 u6 u4 u5 u2 u5 u2 Figure 8.10: We start with the closure of the graph shown in the last figure. We apply Algorithm 62 to this graph.3. All extra edges which are removed are shown by dotted lines. remove an extra edge and find a new Hamiltonian Cycle.

Another Extra Edge can be removed u3 u1 u3 u1 u5 u4 u6 u4 u6 u4 u6 u3 u5 u2 u5 u2 u1 u2 Redraw the graph and remove the last Extra Edge Hamiltonian Cycle. Another Extra Edge can NOT be removed Redraw the graph and remove another Extra Edge Hamiltonian Cycle shown in blue. Another Extra Edge can be removed Redraw the graph and remove another Extra Edge u4 u1 u4 u1 u1 u4 u6 u3 u6 u3 u6 u3 u5 u2 u5 u2 u5 u2 Hamiltonian Cycle shown in blue. .Hamiltonian Graphs 443 u5 u4 u5 u4 u1 u4 u6 u3 u6 u3 u6 u3 u1 u2 u1 u2 u5 u2 Extra edges are shown in red.3. shown in blue. passes through No Extra Edge Figure 8. At the end it passes through the graph edges as shown in the bottom right diagram. Initially the Hamiltonian Cycle passes through all the extra edges as shown in the top left diagram.11: It is possible to remove more than one extra edge (in a single step) from the graph. We remove one Extra Edge Hamiltonian Cycle shown in blue.

The closure is complete u5 u4 u5 u4 u5 u4 u6 u1 u6 u1 u6 u1 u3 u2 u3 u2 u3 u2 The Hamiltonian Cycle does not pass through any Extra Edge Find a Hamiltonian Cycle in the remaining graph as shown in blue color Delete an Extra Edge u5 u4 u5 u4 u6 ? u1 u2 u3 u6 u1 u3 u2 Redraw the graph and try to remove the last Extra Edge through which the Hamiltonian Cycle passes NOT Possible? Find a different Hamiltonian Cycle in the remaining graph as shown in blue color. This will increase the degree? Add Extra Edges where the degree sum condition is recently satisfied. It passes through one Extra Edge. remove other Extra Edges Figure 8. .444 Hamiltonian Graphs 4 u5 u4 2 u5 u4 u5 u4 3 u6 u3 u6 u3 u6 u3 u1 u2 4 u1 u2 u1 u2 We start with a graph where degree sum of a few pairs of vertices is not p Add Extra Edges where the degree sum condition is satisfied. one should be very careful about the order in which they are removed otherwise one may fall into a trap as shown in the bottom diagrams.3.12: While removing Extra edges.

13: There is no need to make the closure of a graph complete.3. the closure is not yet complete 1 6 2 1 6 2 5 4 1 6 3 5 4 3 2 It is possible to add just one Extra Edge so as to get a Hamiltonian Cycle 5 4 3 Figure 8.Hamiltonian Graphs 445 A given graph where degree sum condition for every pair of vertices is satisfied We add Extra Edges so that a Hamiltonian Cycle is formed. . we can add just enough extra edges such that a Hamiltonian Cycle becomes possible in the graph.

The closure of this graph is still a completely connected graph as shown in the bottom left diagram.3. 4 u5 u4 2 3 u5 u4 4 4 u6 u3 u6 u3 u1 u2 4 u1 u2 u5 u4 u5 u4 u6 u3 u6 u3 u1 u2 u1 u2 Figure 8.9? It is essential for you to answer these questions before moving forward.14: The top left diagram shows a graph in which the degree sum of every pair of non adjacent vertices is not equal to (or larger than) the number of nodes in the graph. 8. The bottom right diagram shows that a difficult situation arises if we remove extra edges in the wrong order. while inserting edges you should keep in mind that in a bipartite graph.446 Hamiltonian Graphs Cycle. 8.3.12. If we now remove the extra edges in the wrong order then we may end up with a difficult situation as depicted in the bottom right diagram of Fig. 8.3.1 once again. Let us solve puzzle No. 8. Why this order has suddenly become important and why it was not important in the graph of Fig.3. Be careful this time. an edge uv exists provided .4 Bipartite Hamiltonian Graphs We are now in a position to design or discover similar sufficient conditions for a bipartite graph to be Hamiltonian (It will also be exciting to discover some of the necessary conditions for a bipartite graph to be Hamiltonian).

2 and Fig. 8. The problem is to find out (by an actual trial and error drawing) the minimum degree of the two terminal vertices when it will be impossible to resist a Hamiltonian Cycle. 8. That will be a sufficient condition for this graph to be Hamiltonian.1. 8.5 Some Theoretical Claims We shall describe a number of special Hamiltonian graphs and then present a number of graph theoretical claims. You should also keep the degree of u1 and up exactly the same while inserting new edges.4. Hamiltonian Connected: (p+1)-Closure is complete: A Hamiltonian Path passes through every pair of vertices.Some Theoretical Claims 447 vertex u and vertex v belong to different partites. It is always Moorish Connected. While inserting an edge from u1 to an intermediate vertex (and then an edge from up to again an intermediate vertex) try your best that a Hamiltonian Cycle is not formed. You can not insert an edge between the two terminal nodes otherwise a Hamiltonian Cycle will immediately be formed (we already have a Hamiltonian Path between the two terminal vertices).1: We show a bipartite graph consisting of eight vertices containing a Hamiltonian Path between two end vertices. Please see Fig. Fig. Hamiltonian Graphs: p-closure is complete in G: A Hamiltonian Cycle exists in G. 8.3 u1 u2 u3 u4 u5 u6 u7 u8 Figure 8.4. If we insert an edge between the two end vertices then a Hamiltonian Cycle will immediately be formed. .4.4.

Notice the bi-colored vertices.1. an edge uv exists provided u and v belong to different partite. 8. similar color vertices belong to one partite while all the other color vertices are in the second partite.4. Edges not allowed in a bipartite graph are indicated in both the diagrams in the form of dotted lines.448 Hamiltonian Graphs u1 up u2 u7 u3 u4 u5 u6 u1 up u2 u7 u3 u4 u5 u6 Figure 8. .2: Here we show a Hamiltonian Path between two vertices in the bipartite graph shown in Fig.4. In a bipartite graph.

. 10. You can easily think of sufficient condition for a bipartite graph to be Hamiltonian.3: A bipartite graph of sizes 8.Some Theoretical Claims 449 u v u v u v u v Figure 8. and 12.4.

1: We show a graph G. It may or may not be Hamiltonian Connected. . we add and connect a vertex x to vertices u and v of G as shown in the top diagrams. Claim No. Now if p + 1 closure of H (having p + 1 vertices) is complete then there will be a HAM path in G with end points u and v in G. If the closure of G+x is complete then there will be a HAM cycle in G + x and a HAM path in G between vertex u and vertex v of G. u k k G k k Then there will be a HAM cycle passing through vertex u & v provided u and v are adjacent in G v k Figure 8.5. we add a vertex x to G and connect it to vertex u and vertex v of G. If vertex u and v are adjacent in graph G then there will be a HAM cycle in G passing through u and then v. 1: Given a graph G (having p vertices).450 Hamiltonian Graphs Moorish Connected: A Hamiltonian Path passes through every pair of adjacent vertices. the new graph is known as graph H as shown in the figure below. k k G Degree of every vertex larger than or equal to k k k k u 2 k+1 k k k x k+1 v If p+1 closure is complete in graph G+x then there will be a HAM cycle in the graph G+x.

Degree sum of other pairs becomes 6 1 2 2 4 5 1 5 4 5 1 5 4 x 4 2 x 2 5 5 2 4 5 2 2 5 5 5 5 4 4 3 2 3 All vertices in graph G+x are now completely connected 3 All vertices in G are completely connected There is a HAM path between vertex 4 and 5 Figure 8. We know before hand that a HAM path exists between these vertices but we also know that there is a possibility that a HAM path exists in a graph but we are unable to confirm or find it. we add and connect a vertex x to vertices 4 and 5 of G as shown in the top diagram. 3 4 5 1 2 4 2 1 4 2 1 3 4 x 2 5 3 2 2 x 2 5 4 x 2 5 5 4 4 4 3 3 2 3 2 3 Connect vertices where degree sum is equal to 6 Connect x to 4 and 5. It follows that such a graph G is Hamiltonian. It is somewhat surprising to note that a p closure of graph G is not complete yet a p + 1 closure of graph G + x is complete. 1: We show a graph G in Fig.5. Such a graph G (where a HAM path exists between every pair of vertices in G) is known as Hamiltonian Connected. 8. Degree sum of 4 and 5 becomes 6 Connect 4 & 5. 2: Assume that in a graph G. .Some Theoretical Claims 451 Example No.5. the p + 1 closure is complete then there will be a HAM path between every pair u.2 and we need to check if our newly acquired knowledge can really confirm if there is a HAM path between two vertices 4 & 5 in graph G.2: We show a graph G. v of G. Claim No. v will be part of some HAM cycle. The closure of G + x is complete and there will be a HAM cycle in G + x and a HAM path in G between vertex 4 and 5 of G. in fact it is more than that: every edge u. We are lucky in this example as we do find a HAM path between the two given vertices.

4: Assume that in a graph G. Claim No.4 with end vertices 3 and 5. 5: Given a graph G (having p vertices) with three vertices u. 2: We show a chain graph G in Fig. k+1 k k+1 G Degree of every vertex larger than or equal to k k k p k+1 x k k+1 AH AM pat h in k k+1 G If p+1 closure of G+x is complete then there will be a HAM cycle in G+x Figure 8. Then we connect vertex x to every vertex in G. and w such that u is adjacent to v and v is adjacent to w. We add a vertex x to G and connect it to vertex 3 and 5 as shown in the top left diagram. 8. thus confirming that a HAM path do exist in this graph.3. v. Now if p + 1 closure of H (having p + 1 vertices) is complete then there will be a HAM path in G. now the p + 1 closure of G + x becomes complete as shown in the rest of the diagrams. If the closure of G + x is complete then there will be a HAM path in G. we add a new vertex x to G and connect it to every vertex of G. . Claim No. We add a vertex x and a vertex y to G such that vertex x is connected to vertices u and v of G. the new graph is known as the graph H as shown in Fig. Example No. The p + 1 closure of graph G + x (having p + 1 vertices) is not complete.5.452 Hamiltonian Graphs Claim No. 3: Given a graph G (having p vertices). We know that a HAM path exists between vertex 3 and 5 in this graph.5. thus we can not confirm that a HAM path exists between vertex 3 and 5 in G.3: We show a graph G.5. we add and connect a vertex x to every vertex of G as shown in the top diagrams. 8. the p − 1 closure is complete then there will be a HAM path in G.

4: The p + 1 closure of the graph G + x (having p + 1 vertices) is not complete as shown in the top left diagram. 3 Now we can connect 2 and 5 and 3 & 4 as the degree sum is 6 Figure 8. .Some Theoretical Claims 453 1 2 2 4 2 5 1 3 3 4 2 5 1 3 4 4 2 5 2 3 4 2 2 x 2 2 x 2 2 x 3 A HAM path exists between 3 and 5 but the 6 closure of this graph is not complete 3 We add vertex x to G but now connect x it to every vertex of graph G 3 We connect 2 and 4 as the degree sum is 6 1 5 5 4 5 5 1 3 5 4 3 5 5 5 2 5 x 2 3 x 3 The closure is a completely connected graph. The p + 1 closure of graph G + x is complete and thus a HAM path exists in the graph G as shown in the rest of the diagrams.5.

Claim No. 7: In a regular graph G where the degree of every vertex is p/2. 6: Assume that in a graph G. k v G k k k There will be a HAM cycle passing through u then v and then w in G provided u is adjacent to v and v is adjacent to w. 8.5.5.454 Hamiltonian Graphs and vertex y is connected to vertex v and w of G. and so on. the new graph G + x + y is known as graph H as shown in the figure below. v} and {v.5: If the p + 2 closure of graph G + x + y (having p + 2 vertices) is complete then there will be a HAM cycle in G passing from vertex u to vertex v and then to vertex w of G.5). Claim No. Now if p + 2 closure of H (having p+2 vertices) is complete then there will be a HAM cycle in G which will be passing from vertex u to v and then to vertex w (see Fig. u k G Degree of every vertex larger than or equal to k k 2 k+1 x k k+2 k v k y k k 2 k+1 k w k u k If p+2 closure is complete in G+x+y then there will be a HAM cycle passing from u to x to v to y and then to w. w} in G. a HAM path exists between vertex u and vertex v of G if and only if p + 1 . the p + 2 closure is complete then some HAM cycle will pass through every pair of adjacent edges {u. w Figure 8.

1. u p/2 p/2 u 2 p/2 p/2 G Degree of every vertex is exactly p/2 p/2 p/2 G v p/2 p/2 x p/2 v p/2 There will be a HAM path passing through vertex u & v in G If and only if p+1 closure is complete in graph G+x Figure 8. Look at the graphs in Fig.5. If some of the graphs do not satisfy sufficient conditions for Hamiltonian graphs then there is a possibility that a Hamiltonian Cycle exists but our expertise is not able to find it. . where the degree of each vertex is p/2.Some Theoretical Claims closure of G + x is complete (please see Fig.6).2.3.3. 8: In a regular graph G where the degree of each vertex is p/2.6 in this regard). 9: A regular graph G. We add a vertex x to G and connect x to vertex u and v in G as shown in the right diagram. 8. Problem 8.3. 455 Claim No. Problem Set 8.5. How can you modify the algorithm (which was earlier used to find a Hamiltonian Cycle in a Hamiltonian graph) which makes sure that the resulting Hamiltonian Cycle passes through a given edge.7.5. Claim No. 8. You can use our past experience of finding a Hamiltonian cycle in these graphs. Problem 8. the degree of every vertex is p/2. 8. every edge {u.5. v} will be part of a HAM cycle. is Hamiltonian Connected (that means there is a HAM Path between every pair of vertices in G) if and only if p + 1 closure of G + x is complete for every pair u and v of G (please see Fig.6: We show a regular graph G.

7. We call it category E graphs. We call it category A graphs. Draw a graph which is Hamiltonian (that means it satisfies the sufficient conditions) but in which there is a Hamiltonian path between every pair of vertices. Category B: There is a Hamiltonian path between every pair of non adjacent vertices in G and G is Hamiltonian (that means it satisfies the sufficient conditions) and it is not Hamiltonian Connected (B1). We call it category D graphs.3. We call it category B graphs.3.4.6. . check if there is a Hamiltonian cycle and also check if our expertise can find it or can not find it. Problem 8. Problem 8.6 A Categorization of Hamiltonian Graphs Category A: There is not a Hamiltonian path between every pair of non adjacent vertices in G and G is Hamiltonian (that means it satisfies the sufficient conditions). Design an efficient algorithm to find a Hamiltonian path in these graphs between any two vertices.456 Hamiltonian Graphs 1.5. For every graph in Fig.3.3. Draw a graph which is not Hamiltonian (that means it does not satisfy the sufficient conditions) but in which there is a Hamiltonian path between every pair of vertices. Also consider the option when it is Hamiltonian Connected (B2). Design an efficient algorithm to find a Hamiltonian path in these graphs between two given (not any two but two special) vertices. Problem 8.3. 8. Problem 8.3. 8. 3. Draw a graph which is Hamiltonian (that means it satisfies the sufficient conditions) but in which there is not a Hamiltonian path between every pair of non adjacent vertices. 2. Problem 8. Draw a graph which is Hamiltonian (that means it satisfies the sufficient conditions) but in which there is a Hamiltonian path between every pair of non adjacent vertices. We call it category C graphs.5.7. Draw a graph which is not Hamiltonian (that means it does not satisfy the sufficient conditions) but in which there is a Hamiltonian path between every pair of non adjacent vertices.

5.7: You may find some graphs here which you were supposed to draw in earlier problems.A Categorization of Hamiltonian Graphs 457 Figure 8. .

8: A panoramic hierarchy of graphs. etc. some containing a Hamiltonian Cycle.5.458 Hamiltonian Graphs u5 u4 u6 u3 u1 u2 Figure 8. . some containing a Hamiltonian Path.

3.A Categorization of Hamiltonian Graphs 459 Category C: G is Hamiltonian (that means it satisfies the sufficient conditions) & Hamiltonian Connected (C1).1).3.3.12. 8. We need to determine k such that a Hamiltonian Path (or a cycle) passes through every two vertices u and v in G (see the bottom left diagram in Fig. You can relax this condition later and cater to the condition when the degree of every vertex may be larger than p/2.3.1). Initially you can assume that G is a regular graph and thus the degree of every vertex is the same. Problem 8. Problem 8. Problem 8. Assume that we are given a graph G with number of vertices equal to p where p is even.3.3.13. Problem 8. Problem 8. Problem 8. A graph G where the minimum degree is k is given as shown below.14.3.16.9.10.3. v and w in G (see the bottom right diagram in Fig. Try to draw a connected graph G where G3 is not Hamiltonian Connected.8. 8.6. We need to find (that is design an efficient algorithm to solve this .6. Try to draw a graph G where G2 is not Hamiltonian while G3 is Hamiltonian. Category D: There is a Hamiltonian path between every pair of non adjacent vertices in G and G is not Hamiltonian (that means it does not satisfy the sufficient conditions).6. We need to determine k such that a Hamiltonian Path (or a cycle) passes through three vertices u. We need to determine conditions under which there will be a Hamiltonian Path in G (see the top diagrams in Fig.15. Problem 8. Prove that G3 of a graph G is always Hamiltonian. 8. and the degree of every vertex is larger than or equal to p/2.3. Problem 8. Also consider the option when G is not Hamiltonian Connected (C2).1). Try to draw a graph G where G2 is not Hamiltonian while G3 is Hamiltonian Connected.11. Category E: G is Hamiltonian Connected but is not Hamiltonian. Problem 8. Try to draw a graph G where G2 is not Hamiltonian while G3 is not Hamiltonian Connected.

1: Figures for the last problem set .460 Hamiltonian Graphs k k k k+1 k+1 G k k p k+1 x k+1 k+1 u 2 k+1 k k x 2 u k+1 k v k+1 x k+1 2 k y w v k+1 k Figure 8.6.

and (b) when they are not adjacent. Consider two options separately: (a) when the two given vertices are adjacent.2 may provide some food for thought. 8. . If there is really a Hamiltonian path then our algorithm should be able to output this path (we should provide a proof for this).A Categorization of Hamiltonian Graphs 461 problem) if there is a Hamiltonian path between two given vertices.6. The following graphs in Fig.

.462 Hamiltonian Graphs Figure 8.6.2: We show here a number of regular graphs where the degree of each vertex is exactly p/2 and p is even.

2 9.Chapter 9 Strongly Connected Directed Graphs and Tournaments 9.3 9.1 9.6 Concepts.4 9. Properties & Actions Strongly Connected Directed Graphs Directed Acyclic Graphs (DAG’s) Strongly Connected Components Tournaments Unilaterally Connected Directed Graphs: .5 9.

Some of the properties of graphs. A completely connected graph is always Hamiltonian. 9. and how they are connected. An undirected graph is known as strongly orient-able if it has the potential of becoming a strongly connected directed graph. Graphs are also judged by different properties that they possess. some action items (or so called transformations). directed graphs can also be classified into different categories. We shall start with different categories (shown as concepts in Concept map 1) of directed graphs. Each category of directed graph possesses some specific properties. On the other extreme we have seen the completely connected graph. and a classification of un-directed and directed graphs are shown in a concept map (see Concept map 1). It will be interesting to examine how different properties of a graph change (or do not change) if we transform a graph into a new directed graph using . This is how we shall study directed graphs in this chapter. These actions on a graph G transform G into another graph H. and if not then what are its strongly connected components. When we make a transformation on a graph we create a graph with new properties. We shall be talking about strongly connected directed graphs. directed acyclic graphs. Properties & Actions We have earlier classified an un-directed graph on the basis of connectedness.1 Concepts.464 Strongly Connected Directed Graphs and Tournaments Introduction We shall talk about some specific features of directed graphs in this chapter. For example a line graph which is acyclic and is not Hamiltonian. We had an un-directed graph which was not connected. unilaterally connected directed graphs and tournaments. As with un-directed graphs. We have also talked about certain actions on graphs like the square of a graph or the complement of a graph. We shall discuss and prove (again using constructive proof strategies) necessary and sufficient conditions for an undirected graph to be strongly orient-able. We shall also describe efficient algorithms to find if a directed graph is strongly connected. We have seen in the last chapter that if we take the square of a line graph or a star graph then the resulting graph is Hamiltonian. We have also discussed a connected graph which was not a tree that means a graph which was cyclic. and then we had a just barely connected graph known as a tree.

some action items or transformations.Concepts. Please note that the action items are represented by elliptical objects while the properties are shown by square boxes. and a classification of undirected and directed graphs. . A concept map showing some of the properties of graphs. Properties & Actions 465 Concept Map 9.1.

This means that if we intelligently put directions in a completely connected un-directed graph then the resulting directed graph will be strongly connected (or unilaterally connected). If the underlying undirected graph of D is connected then D is known as a weekly connected directed graph. In a strongly connected directed graph there is a path from u to v and a path from v to u. When we put directions in an un-directed (or bi-directed) graph G then it becomes a directed graph D.466 Strongly Connected Directed Graphs and Tournaments some of the action items shown in Concept map 1. That means a strongly connected graph is always a unilaterally connected graph but not vise versa. The other is unilaterally connected but not strongly connected. Not all un-directed graphs can be converted into a strongly connected directed graph. A directed graph D is unilaterally connected provided for every pair of vertices u and v. Those which can be converted into a strongly connected graph are known as strongly orient-able un-directed graphs. A directed graph formed by putting arbitrary directions in a completely connected undirected graph has a special name that is a tournament. It is also possible to put directions in a completely connected graph in such a manner that the resulting directed graph is acyclic.1 shows three different directed graphs. Please try to pinpoint which one is which. Completely connected un-directed graphs are both unilaterally orient-able and strongly orient-able. Un-directed graphs which can be converted into a unilaterally connected directed graph are known as unilaterally orient-able undirected graphs. A directed graph where you can reach any vertex from any other vertex is known as a strongly connected graph. A directed graph having no cycles is called an directed acyclic graph or a DAG. Both strongly connected as well as unilaterally connected directed graphs are always weakly connected but again it may not be true the other way round. If we remove directions from a directed graph D then the resulting graph will become undirected and is known as the underlying undirected graph. One of them is strongly connected. Fig. there is a path from u to v or a path from v to u.1. Concepts As you can see in Concept map 1. One of these graphs is acyclic while the rest are cyclic. and is therefore not strongly connected. It is obvious that a tournament directed graph may be acyclic . directed graphs can also be classified on the basis of connectedness. The third is neither unilaterally connected nor strongly connected. 9.

Concepts. Problem 9. The resulting graph should be acyclic and unilaterally connected. as we shall discuss in detail in this chapter. After putting appropriate directions draw the directed graph D. or may contain a cycle.1.1. in fact it may contain a Hamiltonian Cycle.1.1: Three different directed graphs. We have not asked you to draw a graph which should be acyclic and strongly connected. Problem 9. Why? . You are supposed to put directions in these undirected edges in order to fulfill some objective as defined below. Properties & Actions 467 2 1 3 4 6 5 2 1 3 6 1 2 3 4 5 4 6 5 Figure 9. 9.1. One of them is strongly connected.1. A graph consisting of six vertices is shown in Fig.4.2.1. Problem 9. Only two edges in this graph are directed while the rest are left un-directed or bi-directed. Problem Set 9.1. Problem 9.3.1.2. The resulting graph should be cyclic and not unilaterally connected. The resulting graph should be acyclic and not unilaterally connected.

The resulting directed graph should be cyclic but not strongly connected. Problem 9. Problem 9.1. Problem 9. The resulting graph should be unilaterally connected. Will the resulting graph always be unilaterally connected? Discuss briefly. A directed graph in which if there is a path from u to v then there is a path from v to u for every pair of vertices u and v in D.2.2: A graph in which only two edges are directed while the rest are left un-directed.2.8.1.2.1. You are supposed to make it a directed graph by putting directions on un-directed edges in order to fulfill certain objectives Problem Set 9.2.5. 3 4 .6.468 Strongly Connected Directed Graphs and Tournaments Problem 9.7.1. The directed graph should be weekly connected. Problem 9.3. The resulting directed graph should be cyclic and strongly connected. Draw a directed graph D in order to fulfill the following objectives: Problem 9. 2 1 6 5 Figure 9. A directed graph in which every node lies on a directed cycle but the graph is not strongly connected. Problem 9.1.1. The directed graph may not be weekly connected. and D is not strong. not strongly connected but cyclic. The resulting graph should be unilaterally connected but not strongly connected.2. A directed graph in which every node lies on a directed cycle but the graph is not strongly connected.

and it is zero otherwise. or acyclic? Discuss briefly. Properties & Actions 469 Problem 9. This definition is almost the same as given in the last chapter for un-directed graphs. A directed graph D is transitive if and only if there is an edge from u to v provided . The transpose of a directed graph D is another directed graph E in which there is an edge from u to v if and only if there is an edge from v to u in D. Will the resulting graph be strongly connected. unilaterally connected. if there is a path from u to v then there is no path from v to u. A Hamiltonian Path or a spanning path spans every vertex only once but is not closed so you can not come back from where you have started. An Eulerian Trail is a closed trail which spans all edges exactly once.2. The Reachable Relation graph of D can be represented by an adjacency matrix A in which A(u. The square of a graph D is another graph (known as) D2 in which there is an edge from vertex u to vertex v provided there is a two edge path from u to v or there is an edge (that is a one edge path) from u to v in D.Concepts. The Reachable Relation (or the transitive closure) of a directed graph D is another directed graph in which there is an edge from vertex u to vertex v provided v is reachable from u in D. We know that a spanning cycle passes through every vertex exactly once. v) = 1 provided there is a directed path from u to v in D. As we know in a walk a node as well as an edge may be traversed several times. A directed graph D is n − cyclic if it contains a directed cycle consisting of n nodes where n ≤ p. If n = p then the graph D is Hamiltonian.4. A closed spanning walk in a directed graph D passes through every node of D and comes back from where it has started. Properties A directed graph D is Hamiltonian provided there is a spanning cycle in the graph. We shall introduce more action items as the need arises. Action Items Let us now discuss some of the action items or so called transformations. A weekly connected directed graph in which for every pair of vertices u and v. An Eulerian Trail may pass through a vertex several times but it should pass through every edge exactly once. Similarly the Reachable Relation matrix of a strongly connected directed graph will also contain all 1’s. The Reachable Relation matrix of an undirected connected graph will contain all 1’s.

3. 2 1 3 4 6 5 Figure 9.3. answer the following: 9.3. . Problem 9.1. Is this graph Hamiltonian? Discuss briefly.3.3. Is D2 Hamiltonian? Discuss briefly. Find a closed spanning walk in this graph.3.3.4.3.3. Is D or D2 transitive? Discuss briefly.3: A directed graph D for the problem set.2.5. Problem Set 9. Problem 9.8. For the directed graph D shown in Fig. try to Problem 9. Is the transpose also strongly connected? Why? Problem 9.7. Draw the transpose of the graph. Problem 9.1.1. Find a Reachable Relation Matrix for this graph.3. Is this graph strongly connected? Why? Problem 9. Problem 9.470 Strongly Connected Directed Graphs and Tournaments there is an edge from u to any vertex w and an edge from w to v for every pair of vertices u and v in D. Draw D2 . Problem 9.3.6.

that is.Strongly Connected Directed Graphs 471 9. Why it can not be an acyclic graph? Because otherwise it will not be possible to reach from v to u if it is possible to reach from u to v. When a graph is strongly connected then it is possible to reach every vertex from every vertex in that graph. the Reachable Relationship of a directed graph. a trail passes through every edge exactly once but may pass through a vertex several times. As every vertex is visited at least once and we come back from where we have started thus every Eulerian graph . Is it the other way round also.2). and thus such a graph will be strongly connected. Let me now pick one action item. These properties are: 1. A strongly connected graph may not be Hamiltonian but it perhaps should always contain a cycle of length n where n < p. Closed spanning walk 4. Let us now focus on Eulerian directed graphs. Keep in mind a number of properties a directed graph can possess.2 Strongly Connected Directed Graphs Let us focus on a strongly connected graph (we consider it as the current concept). this implies that every vertex is reachable on this cycle. This directly implies that the Reachable Relation matrix of a strongly connected graph will contain all 1’s (see Concept map 9. Transitive. that is every strongly connected graph is Hamiltonian? I think you should be able to find a counter example quite easily. Thus a strongly connected graph should never be acyclic. Eulerian Trail 3. Again it is not the other way round that is every cyclic graph may not be strongly connected (we have already witnessed this in a problem set). Hamiltonian Cycle 2. Does it also imply that if the Reachable Relation matrix of a graph contains all 1’s then the graph is strongly connected? I think it is quite obvious that the answer is yes. In such graphs it is possible to find a Eulerian trail. n-cyclic 5. Please do it before moving forward. A directed graph which is Hamiltonian contains a Hamiltonian Cycle.

go to vertex v. that will in fact be a closed walk. however. When we merge U and W we shall get a closed spanning walk in the graph D. This will result in another closed walk W which when merged in U gives us a closed walk with more vertices that in U . be not a closed spanning walk. As U does not span all vertices thus there will be a vertex u in U which will be connected to a vertex v of D which is not part of U (why?). it means that you can always come back to a vertex from where you have started. We are now in a position to design algorithms to solve a number of related problems.2. Let us summarize of what we have gained so far.1). 9. let us look at a relatively more general graph D which contains a closed spanning walk.2.1: A closed walk U is indicated in a directed graph D.472 Strongly Connected Directed Graphs and Tournaments will be strongly connected but not vise versa. it will still be possible to come back to vertex u. Let us call this a closed walk U . It uses the Reachable Relation of a graph as an intermediate or a so called bridging concept. Algorithm 63 answers if a given directed graph is strongly connected (or not) in a polynomial time algorithm. We know that it is possible to traverse a node or even an edge several times in a walk. 1 2 7 1 2 u v 7 6 U 3 8 6 W 3 8 5 4 9 5 4 9 Figure 9. Another closed walk W is also shown in the same directed graph. it may. A graph D is strongly connected if and only if we can find a closed spanning walk in D or D is strongly connected if and only if the Reachable Relationship of D contains all 1’s. We can start from u. . We repeat this process until the closed walk becomes a closed spanning walk (see Fig. Once we have the Reachable Relation we have some information about the nature of the directed graph. After looking at Hamiltonian or Eulerian graphs which are quite restrictive in nature. We also know that in a strongly connected graph it is possible to move from any vertex to any other vertex.

a couple of properties and an action item which is the Reachable Relation. Eulerian.Strongly Connected Directed Graphs 473 Algorithm 63: Check if directed graph D is strongly connected input : Directed graph D output: Graph D is Strongly Connected or not 1 Transform the Directed Graph D into a Reachable Relation graph by applying any traversal algorithm on every vertex of D.2. A map showing the concept of a strongly connected graph. . Concept Map 9. If the Reachable Relation matrix contains all 1’s then the graph D is strongly connected otherwise not. and strongly connected directed graph is exactly the same. It is interesting to note that the Reachable Relation graph of a Hamiltonian.

Note that this algorithm is directly related to the (constructive) proof which proves that a directed graph D is strongly connected if and only if D contains . It will be interesting to derive the time complexity of this algorithm and to make it more efficient if possible. Convert it into a cycle. Khawaja Fahd says that we should just check the diagonal boxes in the Reachable Matrix and there is no need to check any other box. A closed walk can always be converted into a cycle. Algorithm 65: Find a closed spanning walk in a strong graph D input : A strongly connected directed graph D output: A closed spanning walk in D 1 2 Find a closed walk u in D. We shall address this problem once we acquire the required knowledge. Algorithm 64 finds a cycle in a directed graph D. In this algorithm first we find a closed walk in graph D using any traversal algorithm. If D is strongly connected then it will certainly contain a cycle. Algorithm 64: Find a cycle in a directed graph D input : A directed graph D which is not acyclic output: A cycle in D 1 2 Find a closed walk u in D. The given directed graph D may not be strongly connected but it should contain a cycle. Increase its size until it becomes a closed spanning walk as previously explained in this section. What do you think? It is interesting to note that it is possible to design a much more efficient algorithm to solve the same problem but then we need some more concepts and a deeper insight into the problem.474 Strongly Connected Directed Graphs and Tournaments If the relation matrix contains all 1’s then the graph D is strongly connected otherwise it is not.

2. Is a strongly connected graph D transitive? Should it never be transitive? If it is transitive then will it be a special strongly connected graph? If a strongly connected graph D is not transitive then will D2 be transitive? Under what conditions D2 be transitive? We know that in a strongly connected graph D it is possible to reach any vertex from any vertex in D. We next consider a class of un-directed graphs (known as strongly orient-able) where if we put a direction intelligently on each edge then the resulting directed graph will be strongly connected. We should remember that the desired directed graph should not only be cyclic. One of the necessary conditions for a graph to be Hamiltonian is that there should be no cut vertex. every directed edge should be part of a directed cycle otherwise it will be impossible to ensure reach ability from any vertex to every other vertex. Will it be still possible to reach any vertex from any vertex in this graph? If yes then the transpose of D will always be strongly connected provided D was strongly connected. The proof was earlier presented informally in this section. Let us discuss strongly connected graphs in terms of one more property (transitive) and one more transformation (transpose). It is interesting (or shocking) to note that the underlying un-directed graph of a strongly connected graph may not be strongly orient-able (why?). Is this also a necessary condition for a graph G to be strongly orient-able? The following algorithm converts an un-directed graph with no bridge edges into a strongly connected directed graph by putting appropriate direction on each edge (see Fig. It will again be interesting to derive the time complexity of this algorithm and to make it more efficient if possible. It should not be acyclic otherwise it will be impossible to create a directed cyclic graph out of it which is strongly connected. that is. 9.1). Now assume that we reverse the direction of each edge in graph D. It is obvious that the un-directed G should be connected. This implies that a graph G should not contain any bridge edge for G to be strongly orient-able.Strongly Connected Directed Graphs 475 a closed spanning walk. Strongly orient-able Un-directed Graphs We need to discover necessary and sufficient conditions for an undirected graph G to be transformed into a strongly connected graph. The algorithm is quite similar to the algorithm . we have taken the transpose of D. On the other hand there are un-directed graphs where it is impossible to convert them into strongly connected directed graphs.

an Eulerian Trail. You should either prove this conjecture or find a counter example.3 Directed Acyclic Graphs (DAG’s) In a strongly connected directed graph every vertex is reachable from every other vertex. That will in fact complete the proof of a theorem which states that a connected graph G is strongly orient-able if and only if G contains no bridge edges. Can a DAG be unilaterally connected? Can a DAG be not unilaterally connected? Should a DAG be always transitive or always not transitive? These are some of the questions . or a closed walk can not exist in a DAG. We have yet to prove that if there is a bridge edge in a graph G then it is impossible to convert G into a strongly connected graph D. As opposed to a strongly connected graph. As it should be evident. We have seen in the last chapter that G2 of a line graph is Hamiltonian and is therefore strongly orient-able. It will be worthwhile to conjecture that if G is connected but not strongly orient-able then G2 is certainly strongly orient-able. a DAG is quite the apposite.476 Strongly Connected Directed Graphs and Tournaments used to find a closed spanning walk in a strongly connected directed graph. Increase the size of the spanning walk until it becomes a closed spanning walk as previously explained. here if a vertex v is reachable from u then the vertex u is not reachable from v for every pair of vertices u and v in a DAG. a Hamiltonian Cycle. We have also seen that G2 of a star graph is also Hamiltonian. Put appropriate directions. 9. a vertex u is reachable from v and the vertex v is reachable from u for every vertex u and v in D. In other words there is no cycle in a DAG and that is why the name: directed acyclic graph. Algorithm 66: Convert an un-directed G into a strongly connected D input : An un-directed graph G output: A strongly connected directed graph D 1 2 3 Find a closed spanning walk u in G. The algorithm can also be used as a constructive proof to prove that if there are no bridge edges in an un-directed graph G then it is possible to convert G into a strongly connected directed graph D. an n-Cycle. Thus a DAG can never be strongly connected. that is.

the direction of each edge is reversed only. 9. Similarly the sink of the original DAG will be transformed into a source in the transpose of the DAG. A tournament DAG is always transitive (why?). How about if we take the square of a DAG? Will the resulting graph also a DAG? How the roles of different vertices change or remain the same in the square graph? Will the resulting graph be transitive? If we draw a Reachable Relation matrix A for a DAG. A source node is defined as a node with no in degree while a sink node has no out degree. Answer these questions before moving forward. Will the transpose of a DAG always a DAG? Why? It will obviously contain the same number of vertices and exactly the same number of edges. For example every DAG will have at least one source node and at least one sink node in addition to other ordinary (sometimes called intermediate nodes). Looking from the other way round you can start from any undirected graph G and put a direction on each edge such that the resulting directed graph becomes a DAG (in fact it can be converted into several different DAG’s). as this DAG is derived from a completely connected un-directed graph so it is known as a tournament DAG. a sink vertex.4. On the other hand if an un-directed graph is completely connected then we have to be careful in selecting the direction of an edge in order to transform G into a DAG.3.4). It is interesting to note that the underlying undirected graph of a DAG will be an undirected graph. and a number of intermediate vertices. If the undirected graph is a tree then we may put an arbitrary direction on each edge and the resulting directed graph will guaranteed to be a DAG. Try your luck and design an algorithm which can transform an un-directed completely connected graph into a DAG (see Fig. But then a DAG consists of a source vertex. how the matrix will tell us that it represents the reachable relation for a directed acyclic graph? . the transpose of a DAG. What happens to these vertices when we take a transform of a DAG? A source node has no in-degree but when we reverse directions it will have no out-degree so it will be transformed into a sink as shown in Concept map 9. We shall shortly provide an efficient algorithm to solve this problem.Directed Acyclic Graphs (DAG’s) 477 which should trouble your mind. Perhaps you may like to find a non tournament DAG which is also transitive? Let us look at a DAG along with an action item. An intermediate node has both in degree as well as out degree. There are certain nodes in a DAG which have some special features not noticed in other directed graphs.

For the directed acyclic graph D shown in Fig.1.2. Start a DFS from a vertex other than u and again locate a vertex with the maximum finishing time.2). do the following: Problem 9. DFS spanning tree consisting of tree edges 2. identify tree edges.3). it should still tell us somehow that it represents a directed acyclic graph? Let us now introduce a new action item. The source vertex will always be on the extreme left while the sink will be on the extreme right on the horizontal line (see Fig.4. Back edges 5. . you must have become familiar with Depth First Search of a graph while studying algorithms.4. This visualization of a DAG may encourage us in devising a scheme to convert an un-directed graph into an acyclic directed graph.3. The output of the DFS are: 1. the vertices of the DAG can be so arranged on a horizontal line such that all directed edges go from left to right. Perform a DFS from any vertex u in D. 9. We know that a directed graph is acyclic if and only if a DFS of the directed graph does not produce a back edge.3. Forward edges 3. forward edges.3. the directed edges should always go only from left to right. In this so called topological sort. Problem Set 9. and cross edges if any. It is interesting to note that the vertices of a cyclic graph can not be so arranged on a horizontal line such that all edges go in only one direction.1.4. We also know that if a directed graph is a DAG then the vertex with maximum finishing time will always be (one of) the source vertices of the DAG. Cross edges 4.478 Strongly Connected Directed Graphs and Tournaments If we take the transpose of a DAG then obviously the Reachable Relation matrix for the transpose will be different from that of the original graph. More importantly it is beneficial to sort the vertices of the DAG on the basis of decreasing finishing times.4. Find the vertex of maximum finishing time. Finishing times of all vertices (please see Concept map 9. Problem 9. back edges. 9. Problem 9.

Directed Acyclic Graphs (DAG’s) 479 Concept Map 9. a couple of properties and two action items which are Reachable Relation and transpose of a DAG.3. . A concept map showing the concept of a directed acyclic graph.

2.480 Strongly Connected Directed Graphs and Tournaments Problem 9.5.4. Repeat the above three steps (1. Design an efficient algorithm to solve the following: Problem 9.5.5. Find if a DAG has a node u such that you can reach u from every other vertex of the DAG. Problem 9.7. Convert a connected graph into a DAG.4. Problem 9. 2.5. Problem 9. Discuss the outcome? 2 1 3 4 6 5 Figure 9.1.4.3. We are now in a position to address a number of algorithmic issues.5.1: A directed graph D for Problem Set.5.5. reverse the direction of the edge between vertex 2 and vertex 4.3. Find the source and the sink node in a DAG.5. Let us try to address the following problems: Problem Set 9.5. . Find if a given directed graph (not necessarily a DAG) has a node u such that you can reach every vertex from u. Find if a given directed graph is acyclic. Convert a completely connected un-directed graph into a DAG.6. Problem 9. Problem 9. Problem 9. Find if a given DAG has a node u such that you can reach every other vertex of the DAG from u. & 3) on the same graph with one change.

if yes then the source and the sink vertices 1 2 3 Find a Reachable Relation R for the given graph D. You should try to solve these problems yourself before moving forward. It is not possible to find the sink node directly. Sink vertex: You find out.Directed Acyclic Graphs (DAG’s) Algorithm 67: Check if D is a DAG. Algorithm 68: Check if D is a DAG. this will be a source vertex. You might have noticed that Algorithm 68 is much more costly than Algorithm 67. Algorithm 67 used DFS as a bridge to solve a couple of problems. Look at Concept map 9. input : A directed graph D output: Yes/No. Using a Reachable Relation Matrix we can easily find if there is a node u in D from where it is possible to reach every vertex of D. If all diagonal entries in R are zero then D is a DAG. How about if instead of a DAG we have an ordinary directed graph which may contain cycles? There . find the source and the sink. Now we can start a new BFS from this very vertex.3 for further hints. Algorithm 67 can also be slightly modified to locate a vertex u in a special DAG from where it is possible to reach every other vertex. Source vertex will be the one with maximum finishing time. First we can find the vertex with maximum finishing time. Source: Vertex with maximum reachability in R?. if all vertices of D are spanned by the BFS then the DAG possesses the desired characteristics (Does it mean that D is strongly connected?). But then the Reachable Relation provides much more information about a graph. if yes then the source and the sink vertices 1 2 3 481 Apply DFS on graph D. input : A directed graph D output: Yes/No. Perhaps you should take the transpose of D and then again perform a DFS now on the transpose of D. In fact it can answer this question even if the graph D is not acyclic. This is because the Reachable Relation Matrix R is expensive as compared to a single DFS. If no back edge then D is a DAG. Sink: Vertex with zero reachabiliy?. The source vertex can be found directly by identifying the vertex with maximum finishing time. find the source and the sink.

Please note that if the undirected graph G is not completely connected then it will be hard to find a Hamiltonian Path in G (as is done in Algorithm 69. We can easily design an efficient algorithm to check if such a path exists. Perhaps in an ordinary graph. will the maximum finishing time be able to locate this u in D? It is important for you to verify or contradict this conjecture at this very stage before moving forward.482 Strongly Connected Directed Graphs and Tournaments may still be a node u in D from where it is possible to reach every where.3. now the vertices of the DAG can be so arranged on a horizontal line such that the source vertex will always be on the extreme left while the sink will be on the extreme right on the horizontal line (see Fig. Arrange the Hamiltonian Path as a Horizontal line. The first step should be to do a topological sort on the vertices of the given directed graph. This node u is not a conventional source vertex as it may have an in-degree. All edges of the graph will go from left to right in this arrangement otherwise it will not be a directed acyclic graph. Algorithm 69: Convert a completely connected graph into a DAG input : An undirected completely connected graph G output: A DAG D (with same number of nodes and edges) 1 2 3 4 Find a Hamiltonian Path in G. Put back the remaining edges of G with directions going from left to right only.4). and if it does then the algorithm outputs the actual Hamiltonian Path. we should first put directions arbitrarily and then remove cycles by reversing back edges? Or put edge directions such that no cycle is created in the first place? Or why not convert the given un-directed graph into a completely connected un-directed graph and then use Algorithm 69. For a Hamiltonian path to exist in this graph there should be a directed edge .2). 9.4. But then we have to remove the extra edges that we have inserted into the original un-directed graph? So what? It is interesting to note that finding a Hamiltonian Path in a DAG is not a hard problem as it is in other graphs. Algorithm 69 can be used to put appropriate directions on a completely connected un-directed graph in order to convert it into a DAG (see Fig. 9. It is left as an exercise for the learner to design an algorithm in order to convert an ordinary connected graph into a DAG. How about if we perform a DFS in this directed graph. Put directions on this line going from left to right only.

Directed Acyclic Graphs (DAG’s) 483 (from left to right) between every two consecutive vertices. Algorithm 70: Find a Hamiltonian path in a DAG input : A directed acyclic graph D output: A Hamiltonian path in D provided it exists 1 2 Perform a topological sort on vertices of D. As one can see all edges of this graph are going from left to right. A directed edge between two consecutive vertices (8/9 & 1/6) is missing. thus it is not possible to find a directed path passing through all the topologically sorted vertices. a Hamiltonian path in D will essentially be a path between topological sorted vertices in D (in the topological sorted order). The vertices of D are topologically sorted on the basis of decreasing finishing times as shown in the bottom diagram. We start a Depth First Search in this graph starting from a vertex designated as a start vertex in the top diagram. .. A path passing through all topological sorted vertices means a Hamiltonian path exists otherwise not. Fig.2 shows a directed acyclic graph D.3. 9. The resulting start / finishing times are indicated along with each vertex in this diagram. It implies that a Hamiltonian path does not exist in this directed acyclic graph. this observation confirms that D is a directed acyclic graph.

. The start/finishing times of each vertex is also indicated here. The directed graph D is acyclic thus all edges move from left to right in the bottom diagram. Vertices of D are topologically sorted on the basis of decreasing finishing times as shown in the bottom diagram.3.2: A depth first search is conducted on the directed acyclic graph D starting from a vertex labeled as start node as shown in the top diagram. A direct edge between two consecutive vertices is missing in the bottom diagram thus a Hamiltonian path is not possible in this graph.484 Strongly Connected Directed Graphs and Tournaments 3/4 8/9 2/5 7/10 1/6 Start Node 11/12 11/12 7/10 8/9 1/6 2/5 3/4 Figure 9.

In such a graph every vertex u in D may not be reachable to every vertex v in D but if u is reachable to v then v is reachable to u for every pair of vertices u.Strongly Connected Components 485 9. The two sub-graphs in Fig. An independently connected graph D which is not strongly connected is shown in the top of Fig.1. v in D. it is in fact an efficient way of determining vertices belonging to a strongly connected .4. In the bottom diagram of Fig. we add an edge between component E and F . An independently connected graph may not be strongly connected but a strongly connected directed graph is always independently connected (see Concept map 9.4 Strongly Connected Components We know that in a strongly connected graph D. 9. 9.4). in the top diagram there is no edge between component E and F . both of these sub-graphs are strongly connected while graph D as a whole is not strongly connected. The problem that we intend to address in this section is to understand how we can efficiently find the following: 1. v in D.1 and Concept map 9. the resulting graph is no longer independently connected. It is quite evident from this figure that graph D consists of two sub-graphs E and F .4). 9. it can not enter into another strongly connected component.4.4. a vertex u is reachable to vertex v and the vertex v is reachable to vertex u for every pair of vertices u. Let us now consider a (related) class of directed graphs which we call independently connected directed graphs. Nodes belonging to a strongly connected component. It is now natural to define a new concept before moving forward: a strongly connected component in a directed graph D is a maximal set of vertices of a directed graph in which vertices u and v are reachable from each other for every pair of vertices u and v in that maximal set.1. It is important to appreciate that in an independently connected directed graph there are no edges among strongly connected components (see top diagram of Fig. The number of strongly connected components in a directed graph.1 are in fact two strongly connected components. Thus if we start a BFS from any vertex belonging to any strongly connected component in an independently connected graph then the search is confined to that strongly connected component. & 2.4. The concept of independently connected graphs is introduced as it is much simpler to solve the above problem in such graphs. 9.

4. In the bottom diagram there is an edge from E to F and the graph D is neither strongly connected nor independently connected.4. 9. 9.486 Strongly Connected Directed Graphs and Tournaments 4 7 1 3 5 E 2 D F 6 4 7 1 3 5 E D 2 6 F Figure 9.5) in D shown in the bottom of Fig. it may also enter another strongly connected component.1) Thus the problem of determining nodes belonging to a strongly connected component becomes harder if a directed graph is neither strongly connected nor independently connected. In the top diagram. note the directed edge from component E to component F in the directed graph D shown in the bottom of Fig.1) thereby converting it into an independently connected graph then the problem of determining nodes belonging to a strongly connected component is much simplified.1 If you start a BFS in this graph then the search may not be confined to only one strongly connected component. there is no edge between the two components and graph D is independently connected. For example if you start a BFS from vertex 1 belonging to component E then the search will traverse vertices belonging to component E as well as F (see Fig. The . 9.4. component (it is as simple as determining nodes belonging to a connected component in an un-directed graph). In a directed graph which is neither strongly connected nor independently connected there are edges between different strongly connected components. If some how we can remove edges connecting different strongly connected components in a directed graph (like the edge (3.4.1: Sub-graph E and sub-graph F are both strongly connected components.

4. it is not reachable to vertices belonging to other strongly connected components. 9.4. If you carefully look at the resulting Reachable matrix you will notice that the relationship it depicts on this graph is reflexive.2. In simple words it means that in an independently connected graph a vertex is reachable only to vertices belonging to the same strongly connected component. 9. The Reachable Relation Matrix for the independently connected graph of Fig. Once we have the Reachable matrix of an independently connected graph. The corresponding equivalence classes (that the relationship generates) are in fact the strongly connected components E and F of the independently connected graph D.Strongly Connected Components question is how to locate and then remove these edges? 487 It is important at this learning stage to apply the Reachable transformation to a number of independently connected directed graphs and draw the corresponding Reachable Matrices. .1.4. it becomes trivial to pinpoint vertices belonging to a strongly connected component. We shall certainly encourage you to draw the Reachable matrix for the independently connected graph shown at the top of Fig. it is an equivalence relationship on the vertices of the independently connected graph. symmetric as well as transitive. 9.1 is shown by the bold 1’s in any of the table shown in Fig.

although vertex 1 belongs to strongly connected component E (please see Reachable Relation Matrix A in the table shown in Fig.2 is again equivalent to removing the edge (3.488 Strongly Connected Directed Graphs and Tournaments The picture is not so rosy for a directed graph which is not independently connected. you will realize the inherent complication. The question is again how to do it? Let us again consider an independently connected directed D and assume that we take a transpose of D. The problem of determining strongly connected components in a directed graph is thus reduced to recovering the bold 1’s (or throwing away the small 1’s) in the Reachable Relation Matrices of Fig.4.4. It is quite obvious in the tables that the reachability within a strongly connected component is not affected by the transpose transformation (see the bold 1’s in the two tables in Fig. however.4. 9.4.5) in the bottom diagram of Fig.4. you can reach all vertices belonging to component E as well as component F from vertex 1. The Reachable Relation Matrix A of D and the Reachable Relation Matrix B of the transpose of D are also shown in this table. 9. You might have noticed that removing those small ones in the table shown in Fig. 9. 9. hence it is not an equivalence relationship. we take the transpose of a directed graph which is not independently connected then things will be different. The reachability across strongly connected components is certainly affected by taking the transpose of a directed graph. 9. 9. For example if you draw the Reachable matrix of the graph shown in the bottom of Fig.2.2).4.4.4.1 thereby converting it again into an independently connected directed graph. .2). It is interesting to note that the graph D is different from its transpose but the Reachable relation of the two graphs will be exactly the same as demonstrated by bold 1’s in the two tables of Table shown in Fig. We show the same directed graph D as depicted in the bottom of Fig.4. 9. A careful look at the two tables will provide the required answer. this is indicated by the small 1’s in the two tables given in the figure. 9.2. If.1 as well as its transpose on the top of Table in Fig.2. 9.1.4.2. It is thus no longer trivial to determine vertices in a strongly connected component in such directed graphs? If somehow we can convert it back into a symmetric relation without disturbing the connectivity inside any strongly connected component and without changing the number of strongly connected components then the problem will become much simpler? This means that some how we should be able to remove the small 1’s in the Reachable Relation Matrix A of table in Fig. The Reachable relation is not symmetric. 9.

4. It is very different from the transpose of D in Fig.Strongly Connected Components 489 A problem For each of the graphs shown in Fig.4.2 except that the edge (3. 3. Please note that: 1. Algorithm 71 is a straight forward algorithm to find all strongly connected components in a directed graph as already illustrated in the Table shown in Fig. 9. Again it is quite different from graph D in Fig.3).3.2 except the connectivity inside strongly connected component E in D is altered. that is (3. 9.3) is reversed and another edge (6. find the Reachable Relation Matrix and compare this matrix with the ones given in Table in Fig. Graph I is the same as the transpose of D in Fig. Graph K is almost the same as graph I except that the edge (2. 9. Graph J is almost the same as graph D in Fig. Find Transpose T (D) of D.5). Algorithm 71: Find strongly connected components of graph D input : A directed graph D output: Strongly connected components of graph D 1 2 3 4 Find Reachable Relation Matrix A of D.4.4.4.2 except for one common edge.2 (this comparison will certainly be a useful learning experience).4.4. 9.4.5) in D is reversed. 9. 9.4. Graph H is the same as graph D in Fig. 9.4.2 except that the edge (5. Find Reachable Relation Matrix B of T (D). It may not be very efficient (still a polynomial time algorithm) but certainly a good start for a new learner in this field. 9.2 except for one similar edge and that is (5. 9. We shall describe hints for designing a more efficient algorithm to solve a similar problem in .6) in I is reversed in K thus creating a new cycle in the directed graph.2). Take AND of Matrix A & B.2) is added.2. 2. (please see Fig. It uses two transformations: Reachable Relation as well as the Transpose as intermediate building blocks of this algorithm.

4.2). This looks very similar but there are interesting differences between the two algorithms. When you insert edges you will soon realize that in order to achieve the above mentioned objective you should be careful not to create a cycle otherwise all strongly connected components within that cycle will collapse into one strongly connected component and the above condition will be violated.2. In the top left diagram of Fig. Again consider an independently connected directed graph D consisting of K strongly connected components. Graph I has two strongly connected components in spite of the extra edges (2.490 Strongly Connected Directed Graphs and Tournaments the coming paragraphs. Before discussing this algorithm (Algorithm 72). Find Transpose T of D.6) while Graph K is reduced into one strongly connected component because of a cycle created by the edge (6. Please recall graph I & K in Fig. This further implies that there will be at least one source and one sink strongly connected component.3. While in the top right diagram of the same figure. Let us for the time being restrict ourselves to find one strongly connected component of a directed graph.4. let us describe another algorithm (Algorithm 73) which will also output a strongly connected component of F . 9.. Call BFS on T starting from u. . find vertex u of maximum finishing time. Output all vertices traversed by the BFS. the component E was a source while the component F was a sink. This immediately suggests that strongly connected components in a directed graph are always connected in the form of a directed acyclic graph (see Concept map 9. A comparison of the two will certainly be an exciting learning experience.2). 9. Algorithm 72: Find a strongly connected components of graph D input : A directed graph D output: One Strongly Connected Component of D 1 2 3 4 Call DFS on D. You are supposed to insert extra edges between the k strongly connected components such that the new directed graph should still have as many as k strongly connected components. We know there are no edges between strongly connected components in an independently connected graph. component F was a source and E was a sink strongly connected component.

A map showing the concept of a strongly connected graph.4. and strongly connected directed graph is exactly the same. . Algorithm 73: Find a strongly connected component of graph D input : A directed graph D output: One Strongly Connected component of D 1 2 3 4 Find Transpose T of D.Strongly Connected Components 491 Concept Map 9. Call BFS on D starting from u. It is interesting to note that the Reachable Relation graph of a Hamiltonian. Output all vertices traversed by the BFS. a couple of properties and an action item which is the Reachable Relation. Call DFS on T : find vertex u of maximum finishing time. Eulerian.

492 Strongly Connected Directed Graphs and Tournaments E 1 4 Graph D 5 7 4 Transpose of D 3 5 7 F 6 3 2 1 F 6 2 E 1 1 2 3 4 5 6 7 1 1 1 1 2 1 1 1 1 3 1 1 1 1 4 1 1 1 1 5 1 1 1 6 1 1 1 1 7 1 1 1 1 1 1 2 3 4 5 6 7 1 1 1 1 1 1 1 2 1 1 1 1 1 1 1 3 1 1 1 1 1 1 1 4 1 1 1 1 1 5 6 7 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Reachable Relation Matrix A of graph D Reachable Relation B of Transpose of graph D Figure 9.4.2: The Reachable Relation Matrix A for D is shown on the left. . Reachability within a strongly connected component is shown by a bold 1’s. A similar matrix B for the Transpose of D is shown on the right.

This vertex will belong to the source component of the transpose of D. If we start a traversal from this node in the graph F then the traversal will not only traverse vertices belonging to the source strongly connected component. Vertex u will belong to the source strongly connected component of directed graph D. But how to determine if a directed graph possesses this property? How about if the sink strongly connected component consists of all vertices of the directed graph? We can determine this in O(p + q) instead of O(p3 ) .5)? Please recall the algorithm in which we have tried to solve the problem of determining if a given directed graph is strongly connected. We therefore take the transpose of D and then start a traversal from vertex u. The source strongly connected component in D will become a sink in the transpose of D. the output will thus be all vertices belonging to the source component in D as shown in Concept map 9. this time a traversal will be contained in the sink component of the transpose of D. The output of this algorithm will thus be the source or the sink strongly connected component of directed graph D (see Concept map 9.3: Directed graphs for the problem set As you may have noticed that Algorithm 72 locates a vertex u of maximum finishing time after doing a DFS on D. What does that mean in terms of our newly acquired knowledge of strongly connected components? A strongly connected graph is just one strongly connected component.4.Strongly Connected Components Graph H 4 E 1 2 3 5 6 F 1 2 7 E 3 5 6 F 4 Graph I 493 7 4 E 1 2 3 Graph J 7 E 5 6 F 1 4 Graph K 7 3 2 5 6 F Figure 9. it will also enter and traverse vertices belonging to other strongly connected components.5. Algorithm 73 first takes the transpose of D and then locates the vertex u of maximum finishing time.

494 Strongly Connected Directed Graphs and Tournaments Concept Map 9. Transpose of F also consists of a (different) DAG of strongly connected components.5. The source strongly connected component belonging to F is the same as the sink strongly connected component of the transpose of F . A directed graph F consists of a DAG of its strongly connected components. .

4. . Please note that all edges in the bottom left diagram are not directed.6. Fig. Directions are added in this undirected graph in order to convert it into an acyclic directed graph shown in the bottom right. What we need to do is to slightly modify the last line of Algorithm 72 or 73. The bottom of this figure shows two directed graph derived from the top graph by putting a directions on edges of the un-directed complete graph (such a directed graph is known as a tournament).4 (top) shows a completely connected undirected graph. Different directions are added in the completely connected un-directed graph to convert it into a strongly connected graph shown in the bottom left. Problem 9. Problem Set 9. Please solve the following: 5 1 2 3 4 1 5 5 1 2 3 4 1 5 1 2 5 3 4 1 5 1 Figure 9. if BFS traverses all vertices of the directed graph D then D is strongly connected.1. It should read.4 determine the corresponding strongly connected components and the underlying DAG connecting the strongly connected components.4.4. It is interesting to note that the directed graph (bottom right) is a directed acyclic graph as all edges are going from left to right (see Algorithm 69).6. For the two bottom graphs in Fig.Strongly Connected Components 495 for Algorithm 63. 9.4: A completely connected undirected graph of five vertices is shown in the top diagram. The bottom left diagram of this figure is one strongly connected directed graph or component. 9. the graph still remains a strongly connected graph. You may put any directions on these edges (indicated by thin lines).

7.4. What are their respective sizes in terms of number of nodes? Please remember that each node in a DAG corresponds to a strongly connected component.5 show three non-isomorphic directed graphs (tournaments) derived by putting directions on edges of the completely connected un-directed graph of Fig. .3.7. How many non-isomorphic DAG’s are possible corresponding to all possible tournaments of 5 vertices.4 into a directed graph such that the resulting directed graph (which will be another tournament graph) is neither strongly connected nor acyclic. please see Fig.4.4. 9. 9. Is it possible to convert the top un-directed graph Fig. Problem 9. 9.496 Strongly Connected Directed Graphs and Tournaments Problem 9. it shows two tournaments (top and middle) with the same DAG and one with a different DAG (bottom). Carefully examine each of non-isomorphic tournaments that you have drawn in the last problem set. Find and draw all non isomorphic directed graphs derived from the completely connected un-directed graph shown in the figure. Group all tournaments which correspond to the same DAG.4. Fig.4.6. Convert the completely connected un-directed graph in Fig. Problem 9.4.6. 9. The DAG of strongly connected components of the bottom graph is not isomorphic to any of the two. Problem 9. Problem Set 9.4 into a directed graph consisting of two strongly connected components such that one strongly connected component should consist of two nodes and the other will consist of the remaining three nodes? Discuss briefly how this is possible or why it is not possible. Draw all these DAGs.1.6. Find its strongly connected components and determine the DAG connecting these components. Is it possible to draw a tournament of 5 nodes which consists of 4 strongly connected components? Discuss briefly. Note that the top two directed graphs are not isomorphic to each other yet the DAG’s of strongly connected components are isomorphic to each other. The corresponding DAG of strongly connected components of each tournament graph is drawn in front of it.7. Problem 9.5.2.2.5.6. and the respective DAG of strongly connected components of each tournament.4. Draw the corresponding DAG of strongly connected components for each of the tournament graph that you have drawn. 9. Problem 9.

you may put any directions on these edges.Strongly Connected Components 497 1 2 3 4 5 1 2.3. Note that all edges are not directed as shown by thin lines in the bottom graph. the strongly connected components and the corresponding DAG will still remain the same.2 3. 2 3 4 5 1 2 3 4 5 1. .4 5 5 Figure 9.4. The corresponding DAG of strongly connected components of each tournament graph is drawn in front of it.4 5 1 2 3 4 5 1.4. 9.4.5: We show here three non-isomorphic directed graphs derived by putting directions on each edge of a completely connected graph of Fig.

3.5.7. We can take any node u in T which is not part of P . a tournament is always unilaterally connected. you might have noticed that a tournament can be strongly connected while another tournament may be a directed acyclic graph again shown in Fig. 9.5. There are several non-isomorphic tournaments possible which are not acyclic (Fig.498 Strongly Connected Directed Graphs and Tournaments Problem 9. 9.5.5. Derive an exact expression for the number of non isomorphic DAGs of strongly connected components of tournaments of size p. Similarly a tournament is transitive if and only if it is acyclic (why?).7.1 are in fact all tournaments. The directed graphs that you have studied in Fig.1).5.5. Let us start with some very simple properties. Is the DAG of strongly connected components of a tournament also a tournament? Why? Problem 9.1 Tournaments A Panoramic Picture and a Concept Map A directed graph obtained by putting directions on edges in a completely connected un-directed graph is known as a tournament (see Fig. 9. While playing with tournaments in previous problems. we can design an efficient algorithm to find that path. We shall study these and many other interesting properties of tournaments in this section.4.2) but there is always a unique acyclic tournament for a fixed number of vertices (why?). Assume that we have a tournament of p nodes and we have already found a path P of length k in this tournament.5 9. Is it possible to draw a DAG of 4 strongly connected components corresponding to a tournament of size 4? Problem 9. We know . The goal is to find a Hamiltonian path in the tournament which will be a simple path of length p − 1.7. Within these two extremes there is a lot of variety of possible tournaments consisting of various strongly connected components (see Concept map 9. 9.1 . 9. Why? Only because it is a tournament? Recall how a tournament is constructed from a completely connected un-directed graph and also recall the definition of a unilaterally connected directed graph? 9.6).2 A Hamiltonian Path in a Tournament Every tournament has a Hamiltonian path.5.

it possesses some very special properties as shown in this concept map.Tournaments 499 Concept Map 9. Such a directed graph is known as a Tournament. An un-directed and completely connected graph can be converted into a directed graph by putting directions on each edge in an arbitrary fashion. .6.

5. .500 Strongly Connected Directed Graphs and Tournaments We start with a completely connected un-directed graph Which is usually drawn in this fashion 1 5 2 Put directions on each edge 1 2 3 4 1 5 4 3 1 5 2 1 2 3 4 1 5 4 3 A directed acyclic tournament We get a Tournament: This is a directed acyclic graph 1 2 3 4 1 5 1 2 3 4 1 5 A strongly connected tournament A strongly connected tournament Figure 9.1: Different Tournament graphs constructed from a completely connected un-directed graph.

Tournaments 501 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 11 2 3 4 5 1 2 3 4 5 1 2 3 4 5 5 1 2 3 4 5 1 2 3 4 5 Figure 9.2: Different Tournaments and the associared Strongly Connected Components.5. .

Any exercise to put directions so that the path length does not increase will fail as shown in the bottom diagrams.5.3: A directed path P of length k is shown in a tournament. Consider any vertex u in T which is not included in P. The extended Hamiltonian path is shown in bold in the bottom diagrams. Vertex u will be connected to every vertex in P. .502 Strongly Connected Directed Graphs and Tournaments 1 2 3 4 1 k u 1 2 3 4 1 k 1 1 2 1 3 4 1 k u u Figure 9. you are at liberty to put any directions in the top diagram you will always end up with an extended path P of length k+1.

9. Using Algorithm 74 it is possible to find a Hamiltonian path in a tournament.3 (bottom). let us look at a related problem: Assume that we are given a Hamiltonian cycle in a tournament T of size k.5. Algorithm 74: Extend path length in P from k to k + 1 input : A Tournament T . and a cycle C of length k in T output: A cycle of length k + 1 in T Before providing an answer to the above problem. Before proving that general result. See if you can find the Basic Idea and Time Complexity of Algorithm 75 as shown below. thus it may not be strongly connected. We will now show that T + u .Tournaments 503 that u is connected to all nodes in T (why?) as shown in Fig. Algorithm 75: Extend cycle length in T from k to k + 1 input : A strong Tournament T. you may put any directions on these edges.3 A Hamiltonian Cycle in a Strong Tournament We know that a tournament can be acyclic. How about if we use a similar approach used earlier for extending the path length.5. 9. and a simple path P of length k in T output: A simple path P of length k + 1 in T 1 2 Consider any node u in T not already in the path P .3 (top). you are bound to get an increase in the path length of P from k to k + 1. We now prove that if a tournament is strongly connected then it will always be Hamiltonian. it will not contain a Hamiltonian cycle. 9.5. we first prove a simpler hypothesis: if there is a cycle C of length k in a tournament then it is possible to find a cycle of length k +1 in that tournament. The algorithm also serves as a constructive proof that there is always a Hamiltonian path in any tournament. Include u in P and output the new path P . Try your luck and check if it is possible. Also assume that now we add another vertex u to T such that T +u is strongly connected and is also a tournament. It will be an interesting (but futile) exercise to put directions such that the path length does not increase by one as is demonstrated in Fig.

5. It means that if T is Hamiltonian then T + u will also be Hamiltonian provided T + u is a strong tournament. This type of approach will be very similar to Algorithm 74 where we extend the path length incrementally and finally output the Hamiltonian path. Is it feasible or not? Carefully look at Fig.4. In fact this special case is simple to solve as there is only one way to convert a directed acyclic tournament graph (we already . A new vertex u is added to T such that T + u is a tournament and is also strongly connected.4).5.4: A Hamiltonian cycle in a tournament T of size k is shown. and so on until we find a Hamiltonian cycle. 9. As T + u is a tournament thus u will be connected to every vertex of T. how about if T is a directed acyclic tournament but T +u is a strongly connected.504 Strongly Connected Directed Graphs and Tournaments is Hamiltonian. Thus it will always be possible to find a Hamiltonian cycle in T + u as shown in Fig. Then how can we show that T + u will be Hamiltonian? Let us look at the extreme. that is certainly not a hard problem? Then we should locate a vertex u such that the cycle length increases from 3 to 4. it is possible that a tournament T is not strongly connected but T + u is a strongly connected tournament. How the above hypothesis will help us in finding a Hamiltonian cycle in a strongly connected tournament or in proving that a strongly connected tournament is Hamiltonian? How about finding a cycle of length 3 in a strong tournament. As you can see there is a serious complication here.5 before making up your mind.5.5. the vertex u will have a finite in-degree and a finite out-degree as T + u is strongly connected (see Fig. 3 4 3 4 2 u 5 2 u 5 1 k 1 k Figure 9. The vertex u will have at least one incoming edge and one outgoing edge (right). It is thus always possible to find a Hamiltonian cycle in T + u as shown in bold in the left diagram. 9. 9.

A cycle of length 5 is shown in bottom left diagram. both these vertices are outside the cycle C and are double circled. Please note that vertex 4 has all incoming edges from vertices belonging to the cycle C while vertex 5 has all out going edges to vertices belonging to the cycle C. in fact it is more than that. If it is right then the tournament will have a cycle of every possible length until we have a Hamiltonian cycle. but there is a possibility when T is neither strongly connected nor acyclic. How to cater to that class of graph.Tournaments 505 know that it will have a Hamiltonian path inside it) into a strongly connected graph is to insert the extra node u is to connect it with the source node of the DAG and then connect the sink with node u. Let us come back to the problem of designing Algorithm 75 or proving that if a cycle C of length k exists in a strong tournament then a cycle of length k + 1 also exists in the strong tournament.5. A cycle of length 4 is shown in the bottom right. . Before proving this hypothesis let us look at its repercussions. it will be interesting for you to explore? 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 Figure 9. Consider vertices 4 and 5 in this figure.5: A directed cycle C of length 3 is shown (enclosed in a dotted circle) in a strongly connected tournament of size 5 (top). We have already considered the possibility when T is strongly connected. thus a strongly connected tournament is Hamiltonian. we have also considered the possibility when T is a directed acyclic graph. this will certainly result in a Hamiltonian graph (why?).

5.5. There are essentially two possibilities.5). Now it is possible to find a cycle of length k + 2 (see the bottom left of Fig. 2. This also provides an efficient algorithm to find a cycle from length 3 to all the way to a Hamiltonian cycle of length p − 1.4. Note that if (a) or (b) is not true then we shall end up with the first possibility.5. (b) all edges from v to every vertex of the cycle C are out going.5. Assume that we have found a cycle C of length k in a strongly connected tournament. 9. You can find a vertex u such that it has at least one incoming and one out going edge connecting u to vertices already in the given cycle C. You are already familiar with the concept of a cut vertex in an un-directed graph (a vertex v in an undirected connected graph G is a cut vertex provided G − v is not connected).6 ). This completes the constructive proof that you can extend the cycle length in a strong tournament of p nodes from 3 to p−1.5and also Fig. 9. 9. This scenario is depicted in Fig. Now we intend to find a cycle of length k + 1 in the same tournament.5. 1.5. and (c) there is an edge from u to v. 9. You can certainly find a cycle of length K +1 in this situation by extending the cycle length of C from k to k + 1. The Concept of a Rip Vertex Let us define a new term before designing an alternate strategy to find a Hamiltonian cycle in a strong tournament. Let us now come to the proof of the above hypothesis. If (c) is not true then the tournament T will not be strongly connected. it is possible to convert this cycle into a cycle of length k + 1 by a simple manipulation (see the bottom right of Fig. 9. This possibility is quite similar to the one illustrated in Fig.506 Strongly Connected Directed Graphs and Tournaments a cycle of length k is possible in a strongly connected tournament where k is equal to or more than 3. If you can not find a vertex u with the above property then you will certainly find a vertex u and a vertex v such that (a) all edges to u from every vertex of C are incoming towards u. Corresponding to a cut vertex in a connected un-directed graph there is a counter part concept in a (strongly) connected directed graph: a vertex v is a rip vertex (in a strongly connected directed graph D) provided D − v is not .

.5.6: Extending a 3-cycle into a 5-cycle in a strongly connected tournament. It can be extended into a 5-cycle directly 1 2 3 4 5 Figure 9.Tournaments 507 1 2 3 4 5 We find a 3-cycle in the graph The 3-cycle is extended into a 4-cycle 1 2 3 4 5 1 2 3 4 5 The 4-cycle is extended into a 5-cycle which is a HAM Cycle 1 2 3 4 5 We start with a different 3-cycle in the same graph This 3-cycle can not be extended into a 4-cycle.

Draw a strong tournament in which there are at least 3 rip vertices in a tournament of size 5 or 6 . It will be interesting to derive and compare the time complexities of the two algorithms. Problem 9. In the later algorithm we remove non rip vertices from a tournament one by one until it is possible to find a Hamiltonian cycle in the reduced size tournament.8. Problem 9.5. Prove that there always exists at least one non rip vertex in a strong tournament.3. We can do this step of removing a non rip vertex recursively until the original tournament is reduced to such a small sized tournament where it will be trivial to find a Hamiltonian cycle. the Hamiltonian cycle will grow incrementally (with the graph) as shown in Fig. How can you efficiently find a non rip vertex in a strong tournament? .5.8.5. we then insert the removed vertices one by one and Hamiltonian cycle also increases incrementally with the size of the graph. In the earlier algorithm we grow a cycle within the original tournament from a small size to p. It will be a learning experience to compare the working of the two algorithms that we have described to find a Hamiltonian cycle in a strong tournament. We first locate and then remove a non rip vertex u.8. 9. It is important for you to answer the following in order to meaningfully understand the last algorithm that we have described to find a Hamiltonian cycle in a strong tournament.8. Now assume that we are given a strongly connected tournament T of size p. The resulting graph T − u will be a tournament (why?) and a strong tournament (because u was not a rip vertex) of size p − 1. Now we can start inserting back the removed vertices one by one in the last removed first inserted order.8. Problem 9. Problem Set 9. until you find the Hamiltonian cycle in the original tournament T .8.4.508 Strongly Connected Directed Graphs and Tournaments (strongly) connected.2. 9. Indicate which vertices are rip vertices and which vertices are not rip vertices in Fig. Problem 9.8. Draw a strong tournament in which no vertex is a rip vertex.5.1. On the basis of these ideas it is possible to construct an alternate algorithm to find a Hamiltonian cycle in a strong tournament. Problem 9.

We show four different tournaments in the top diagram.7: Finding a rip vertex in a Tournament. In the other diagrams we check if a given vertex is a rip vertex any any of the top tournaments.Tournaments 509 5 5 5 5 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 5 1 2 3 4 1 2 3 4 1 2 3 4 5 1 2 3 4 1 is a rip vertex 1 is a rip vertex Not a rip vertex Not a rip vertex 5 1 2 3 4 1 2 3 4 1 2 5 3 4 1 2 3 4 Not a rip vertex 2 is a rip vertex Not a rip vertex 2 is a rip vertex 5 1 2 3 4 1 2 5 3 4 1 2 3 4 1 2 3 4 Not a rip vertex Not a rip vertex 3 is a rip vertex 3 is a rip vertex 5 1 2 3 4 1 2 3 4 1 2 3 4 5 1 2 3 4 4 is a rip vertex Not a rip vertex 4 is a rip vertex Not a rip vertex Figure 9.5. .

.8: Extending a 3-cycle into a 5-cycle in a strongly connected tournament.510 Strongly Connected Directed Graphs and Tournaments 1 2 3 4 5 2 3 4 5 Identify 1 as a non-rip vertex Remove the non-rip vertex 2 3 4 5 2 4 5 Identify 3 as a non-rip vertex Remove the non-rip vertex 2 4 5 2 3 4 5 Find a 3 cycle in the last graph Insert back the last vertex removed 2 3 4 5 1 2 3 4 5 Find a 4 cycle in the last graph Insert back the last vertex removed Find a 5 cycle in the graph which is the HAM Cycle 1 2 3 4 5 Figure 9.5.

we remove a rip vertex then what will be the complication in reconstructing the Hamiltonian cycle in T ? Problem 9. shown in Table 9. j] = 1 for every pair of vertices i. If instead of removing a non rip vertex in a strong tournament. j in D then D will be more than unilaterally connected. Why we insist that a non rip vertex should be removed instead of a rip vertex? Problem 9.6. A variety of non-isomorphic directed acyclic graphs are possible which will produce such a Reachable Relation matrix. j] and A[j. It will be useful at this stage to draw a couple of such .1 corresponds to a unilaterally connected graph which is also directed acyclic (why?) assuming that all other entries in the matrix are equal to zero. i] are equal to 1 for every pair of vertices i. j in D if j > i. Each of the matrices.6 Unilaterally Connected Directed Graphs: A directed graph D is unilaterally connected provided for every pair of vertices u and v.8.9 9.8. Thus a directed graph D will be unilaterally connected if A[i. It is clear from the definition of such directed graphs that either A[i. What will be the overall time complexity of finding a Hamiltonian cycle in a strong tournament using the above algorithm.9.8. Problem 9. j] = 1 or A[j. 9. Let us first address the problem of how a Reachable Relation Matrix A of a unilaterally connected graph would look like. We have also seen that a tournament graph is also unilaterally connected. i] = 1 for every pair of vertices i.1. Can you visualize a tournament graph where there is no rip vertex? Please see Fig.7. That means a strongly connected graph is always a unilaterally connected graph but not vise versa. there is a path from u to v or a path from v to u. If instead of one both A[i.5. 9. In this section we shall discuss various properties of unilaterally connected directed graphs.8. In a strongly connected directed graph there is a path from u to v and a path from v to u. in fact it will be strongly connected. j in D.8.Unilaterally Connected Directed Graphs: 511 Problem 9.6.6. This means that the Reachable Relation Matrix A for a unilaterally connected directed graph will have all 1’s either on upper side of the diagonal (or on the lower side of the diagonal) as shown in Table.

5.512 Strongly Connected Directed Graphs and Tournaments 1 2 3 4 1 5 2 3 4 1 5 1 1 3 4 1 5 1 2 4 1 5 1 1 2 3 5 1 2 3 4 1 A tournament where the in-degree as well as out-degree of every vertex is equal Figure 9.9: The effect of deleting a vertex in a Special Tournament .

An un-directed connected graph which can be converted into a unilaterally connected directed graph is known as a unilaterally orient-able graph.Unilaterally Connected Directed Graphs: 513 Concept Map 9. A map showing the concept of a unilaterally connected directed graph. a couple of properties and an action item which is the Reachable Relation. .7. A tree graph is unilaterally orient-able provided it is a path graph.

5 1 1 5 3 1 4 5 2 3 4 1 5 2 Figure 9.1: The Reachable Relation Matrix A for a unilaterally connected directed graph should have all 1’s either on the upper side of the diagonal (left) or on the lower side as shown in the right table. The right diagram shows such a graph with maximum number of edges.2: A unilaterally connected directed graph D having minimum number of edges with all 1’s in the upper side of the diagonal and 0’s else where as shown in the left diagram. .514 Strongly Connected Directed Graphs and Tournaments Figure 9.6.6. Both theses graphs are directed acyclic.

as well as cyclic containing a Hamiltonian path (as before) as shown in the middle diagram of Fig. Assume that in addition to all 1’s on the upper side of the diagonal there are some additional 1’s as shown (in bold) in the top diagram of Fig.3. We shall soon prove that a unilaterally connected directed acyclic graph D will always contain a Hamiltonian path inside D.6.Unilaterally Connected Directed Graphs: 515 graphs with the above property and the assumption that all other entries in the matrix are zero.6. We are dealing with a directed acyclic graph D.6. this would imply a Hamiltonian path in D (it will . What is quite unexpected is that the Hamiltonian path may disappear because of the extra 1’s that we have added in the Reachable Relation matrix as shown in the bottom diagram of Fig. 9. 9. 9. In fact we shall prove that a directed graph is unilaterally connected if and only if it contains an open spanning walk.2. 9. we show in Fig.3. If there are cycles in the unilaterally connected graph then there may or may not be a Hamiltonian path inside the graph (see Fig.6.6.3. This means there will be a direct edge between every two consecutive vertices (again from left to right). But before that let us try to imagine what will happen if there are more 1’s then are absolutely essential for a graph to be unilaterally connected. v} in D either u is reachable to v or v is reachable to u. We have observed that a unilaterally connected graph contains a Hamiltonian path if it is acyclic (see Fig. one having minimum number of edges (left) and the other having maximum number of edges as shown on the right of this diagram. assume that u is on the left of v in linear ordering then there must be a direct edge from u to v otherwise the condition stated above will be violated or there will be an edge going from right to left.2). 9. What we have not observed until now is that whether a unilaterally connected directed graph contains a Hamiltonian path or not it will certainly contain an open spanning walk. thus for every consecutive vertex pair {u. 9. There will not be any edge going from right to left because otherwise graph D will be cyclic.6. thus if we do a topological sort on vertices of D then the ordering of the vertices of the graph D along a horizontal line is such that all directed edges will be going from left to right as shown in Fig. Let us summarize our observations before making a number of formal proofs.2. 9. such a unilaterally connected graphs.6. Let us first find a constructive proof that a unilaterally connected directed acyclic graph will always contain a Hamiltonian path inside it. Under such conditions one would expect that the resulting directed graph will be unilateral.3). We are dealing with a unilaterally connected directed graph D.

The extra 1’s (shown in bold) will make the unilaterally connected directed graph cyclic but there is still a Hamiltonian path inside the graph as shown in the middle diagram. .3: A Reachable Relation matrix A for a unilaterally connected directed graph with some additional 1’s in the matrix (top). what is quite unexpected is that the Hamiltonian path may disappear from the resulting directed graph because of the extra 1’s as shown in the bottom diagram.6.516 Strongly Connected Directed Graphs and Tournaments 1 1 2 3 4 5 2 1 3 1 4 1 5 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 3 5 1 4 5 1 2 3 1 4 5 Figure 9.

6.6. If.3) and so it may not have a Hamiltonian path. thus there will be an open spanning .6.Unilaterally Connected Directed Graphs: essentially be a path between topological sorted vertices in D). D contains cycles then we can always find strongly connected components of D and we know that the graph interconnecting its strongly connected components will be a directed acyclic graph. Let us first handle the hypothesis that if D is unilaterally connected then there will be an open spanning walk inside D. Note that inside each strongly connected component there is a closed walk.4. 517 1 2 3 4 5 Figure 9. 9. We can again make a topological sort on strongly connected components of D and argue that there will be an edge going from left to right between every two consecutive strongly connected components of D. Also note that the directed acyclic graph connecting all strongly connecting components should be a line graph otherwise D will not be a unilaterally oriented directed graph.4: A unilaterally connected directed graph D having 5 strongly connected components. now we claim that a directed graph D is unilaterally connected if and only if it contains an open spanning walk. however. 9. thus there will be (not only a spanning walk) but a spanning path inside D. Let us now consider the possibility when a unilaterally connected directed graph is not directed acyclic (see Fig. We also know that there is a closed walk spanning all vertices inside each strongly connected component. We have already observed that if D is a unilaterally connected directed acyclic graph then there will be a Hamiltonian path inside D. Thus there will be a Hamiltonian path passing through all strongly connected components of D as shown in Fig.

If you recall this is a sufficient condition for a graph to be strongly orient-able. Let us consider the second possibility when every edge of G is a non bridge edge.5. We also know that a directed acyclic unilaterally connected graph D always contain a Hamiltonian path.518 walk inside D. There are essentially three possibilities: 1. Consider each connected component as a single (super) vertex. Some edges are non bridge edges while some are bridges. Every edge of G is a non-bridge edge. A tree of ordinary vertices can be unilaterally . Now let us find necessary and sufficient conditions for a general connected graph G (G is no longer restricted to be a tree and may contain cycles) to be unilaterally orient-able. We have already considered the first possibility when G is a tree and every edge of G is a bridge edge. 2. Thus when we put directions on edges in a tree graph in order to make it unilaterally connected then the resulting directed graph should also have a Hamiltonian path inside it. This would require that there should be an un-directed Hamiltonian path inside the tree graph G in the first place otherwise it would have been impossible to convert G into D with a directed Hamiltonian path inside it.6. Strongly Connected Directed Graphs and Tournaments 9. Every edge of G is a bridge edge.6. If G is strongly orient-able then it will also be unilaterally orient-able. 9. If we put back the bridge edges then these (super) vertices will be connected in the form of a tree (why?). Thus a tree graph is unilaterally orient-able if and only if it is a line graph (or a path graph). The only tree graph which contains a Hamiltonian path is in fact a line graph. 3. Now we shall consider the third possibility when some edges of G are bridge edges while some other edges are non bridges as shown in Fig. If the graph G contains K bridge edges and if we remove all these K edges then G will be decomposed into k + 1 connected components.1 Unilaterally orient-able Un-directed Graphs We know that if we put directions on edges in an un-directed acyclic connected graph G (that means a tree) then no matter what is our direction scheme the resulting graph D would be a directed acyclic graph.

.5: An un-directed connected graph G with 4 bridge edges shown in bold in the top diagram. The connected components are always connected in the form of a tree as shown in the bottom diagram.6.Unilaterally Connected Directed Graphs: 519 3 1 2 4 5 3 1 2 4 5 Figure 9. If we remove all bridge edges then the un-directed graph is disconnected into 5 connected components shown in shaded circles.

no edge inside the component will be a bridge edge (why?). A tree of super vertices can be unilaterally orient-able provided: (1) The tree is a line graph and (2) We can find a closed spanning walk inside each connected component (why?).520 Strongly Connected Directed Graphs and Tournaments orient-able if and only if the tree is a line graph. You may prove the converse as an exercise? . So once again concentrate on a single connected component. Thus each connected component is strongly orient-able (and will subsequently become a strongly connected component in the directed graph). This completes the proof that if a graph G is unilaterally orient-able then it will contain an open spanning walk. there will be a closed spanning walk inside each component.

How to teach a class by the modified moore method. 2003. The teaching of problem solving.htm.com/dphils/dphil1. [8] M. Rawlins. [7] M. Lecture. [3] Donald R. national computer conference. Foulds. 2003. 1992. In 2003 International Conference on Engineering Education. Computer Algorithms: Introduction to Design and Analysis. http://www.edu/∼mhale/teach/index. 1992. [9] William R. 2000. Iqbal and Atif Alvi. Math. How to Solve it. Narosa Publishing House. New York. Teaching computer science. Amer. 82:466–470. [5] Margie Hale. 102:317–321. [4] L. http://www. [11] G. The magic of dynamic programming. A. Miller. Monthly. [12] David Rine. Technical report. 1975. .tcdc. [10] G. Chalice. R. Graph Theory Applications. Halmos. 1995. Polya. Monthly. R.htm#top. 1980. A. Addison-Wesley. Lahore University of Management Sciences. Princeton University Press. Amer. Iqbal. [6] P. Department of Computer Science.stetson. Math. [2] Sara Baase and Allen Van Gelder. A New Aspect of Mathematical Method.Bibliography [1] Websters dictionary. Compared to What? Computer Science Press. 1988.

Leiserson and Clifford Stein. The Algorithm Design Manual. University of Houston. Prentice Hall International. C. WCB/McGraw-Hill. Rivest T. Ross and Charles R. B. Taylor. [17] R. george mason university. MIT Press. Rosen. 1972. moore. 1992.522 Bibliography [13] David Rine. l. [18] D. Skiena. 149. 1997. 1972. Discrete Mathematics. Creative teaching: Heritage of r. [14] Kenneth H. QA 29 M6T7. 2002. Springer-Verlag. Introduction to Algorithms. Discrete Mathematics and its Applications. [15] Kenneth A. Cambridge MA. 2001. Lectures in patterns . [16] Steven S. Wright. New York. 2003.directed design. Cormen. .

Sign up to vote on this title
UsefulNot useful

Master Your Semester with Scribd & The New York Times

Special offer for students: Only $4.99/month.

Master Your Semester with a Special Offer from Scribd & The New York Times

Cancel anytime.