Introduction Recursive Programming

Introduction to
Recursive
Programming
Introduction to
Recursive
Programming
Manuel Rubio-Sánchez
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742
© 2018 by Taylor & Francis Group, LLC

CRC Press is an imprint of Taylor & Francis Group, an Informa business
No claim to original U.S. Government works
Printed on acid-free paper

Version Date: 20170817
International Standard Book Number-13: 978-1-4987-3528-5 (Paperback)

International Standard Book Number-13: 978-1-138-10521-8 (Hardback)
This book contains information obtained from authentic and highly regarded sources. Reasonable
efforts have been made to publish reliable data and information, but the author and publisher cannot
assume responsibility for the validity of all materials or the consequences of their use. The authors and
publishers have attempted to trace the copyright holders of all material reproduced in this publication
and apologize to copyright holders if permission to publish in this form has not been obtained. If any
copyright material has not been acknowledged please write and let us know so we may rectify in any
future reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or
hereafter invented, including photocopying, microfilming, and recording, or in any information
storage or retrieval system, without written permission from the publishers.
For permission to photocopy or use material electronically from this work, please access
www.copyright.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc.
(CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization
that provides licenses and registration for a variety of users. For organizations that have been granted
a photocopy license by the CCC, a separate system of payment has been arranged.
Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and
are used only for identification and explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
To the future generations
Contents
PREFACE xv
LIST OF FIGURES xxi
LIST OF TABLES xxxi
LIST OF LISTINGS xxxiii
Chapter 1 Basic Concepts of Recursive Programming 1
1.1 RECOGNIZING RECURSION 1

1.2 PROBLEM DECOMPOSITION 7
1.3 RECURSIVE CODE 14
1.4 INDUCTION 20
1.4.1 Mathematical proofs by induction 20
1.4.2 Recursive leap of faith 22
1.4.3 Imperative vs. declarative programming 25
1.5 RECURSION VS. ITERATION 25
1.6 TYPES OF RECURSION 27
1.6.1 Linear recursion 27
1.6.2 Tail recursion 27
1.6.3 Multiple recursion 28
1.6.4 Mutual recursion 28
1.6.5 Nested recursion 29
1.7 EXERCISES 29
Chapter 2 Methodology for Recursive Thinking 31
2.1 TEMPLATE FOR DESIGNING RECURSIVE ALGO-

RITHMS 31
vii
viii Contents
2.2 SIZE OF THE PROBLEM 32

2.3 BASE CASES 34
2.4 PROBLEM DECOMPOSITION 37
2.5 RECURSIVE CASES, INDUCTION, AND DIAGRAMS 41
2.5.1 Thinking recursively through diagrams 41
2.5.2 Concrete instances 45
2.5.3 Alternative notations 47
2.5.4 Procedures 47
2.5.5 Several subproblems 49
2.6 TESTING 52
2.7 EXERCISES 55
Chapter 3 Runtime Analysis of Recursive Algorithms 57
3.1 MATHEMATICAL PRELIMINARIES 57

3.1.1 Powers and logarithms 58
3.1.2 Binomial coefficients 58
3.1.3 Limits and L’Hopital’s rule 59
3.1.4 Sums and products 60
3.1.5 Floors and ceilings 66
3.1.6 Trigonometry 66
3.1.7 Vectors and matrices 67
3.2 COMPUTATIONAL TIME COMPLEXITY 70
3.2.1 Order of growth of functions 71
3.2.2 Asymptotic notation 73
3.3 RECURRENCE RELATIONS 76
3.3.1 Expansion method 80
3.3.2 General method for solving difference equations 89
3.4 EXERCISES 101
Chapter 4 Linear Recursion I: Basic Algorithms 105
4.1 ARITHMETIC OPERATIONS 106

Contents ix
4.1.1 Power function 106

4.1.2 Slow addition 110
4.1.3 Double sum 113
4.2 BASE CONVERSION 115
4.2.1 Binary representation of a nonnegative integer 115
4.2.2 Decimal to base b conversion 117
4.3 STRINGS 119
4.3.1 Reversing a string 119
4.3.2 Is a string a palindrome? 120
4.4 ADDITIONAL PROBLEMS 121
4.4.1 Selection sort 121
4.4.2 Horner’s method for evaluating polynomials 124
4.4.3 A row of Pascal’s triangle 125
4.4.4 Ladder of resistors 127
4.5 EXERCISES 129
Chapter 5 Linear Recursion II: Tail Recursion 133
5.1 BOOLEAN FUNCTIONS 134

5.1.1 Does a nonnegative integer contain a partic-
ular digit? 134
5.1.2 Equal strings? 136
5.2 SEARCHING ALGORITHMS FOR LISTS 139
5.2.1 Linear search 139
5.2.2 Binary search in a sorted list 142
5.3 BINARY SEARCH TREES 143
5.3.1 Searching for an item 144
5.3.2 Inserting an item 147
5.4 PARTITIONING SCHEMES 148
5.4.1 Basic partitioning scheme 149
5.4.2 Hoare’s partitioning method 150
5.5 THE QUICKSELECT ALGORITHM 155
5.6 BISECTION ALGORITHM FOR ROOT FINDING 157
x Contents
5.7 THE WOODCUTTER PROBLEM 158

5.8 EUCLID’S ALGORITHM 164
5.9 EXERCISES 167
Chapter 6 Multiple Recursion I: Divide and Conquer 171
6.1 IS A LIST SORTED IN ASCENDING ORDER? 172

6.2 SORTING 173
6.2.1 The merge sort algorithm 174
6.2.2 The quicksort algorithm 177
6.3 MAJORITY ELEMENT IN A LIST 180
6.4 FAST INTEGER MULTIPLICATION 183
6.5 MATRIX MULTIPLICATION 186
6.5.1 Divide and conquer matrix multiplication 187
6.5.2 Strassen’s matrix multiplication algorithm 190
6.6 THE TROMINO TILING PROBLEM 191
6.7 THE SKYLINE PROBLEM 196
6.8 EXERCISES 203
Chapter 7 Multiple Recursion II: Puzzles, Fractals, and

More. . . 205
7.1 SWAMP TRAVERSAL 205

7.2 TOWERS OF HANOI 209
7.3 TREE TRAVERSALS 213
7.3.1 Inorder traversal 215
7.3.2 Preorder and postorder traversals 216
7.4 LONGEST PALINDROME SUBSTRING 217
7.5 FRACTALS 220
7.5.1 Koch snowflake 220
7.5.2 Sierpiński’s carpet 224
7.6 EXERCISES 226
Contents xi
Chapter 8 Counting Problems 235
8.1 PERMUTATIONS 236

8.2 VARIATIONS WITH REPETITION 238
8.3 COMBINATIONS 240
8.4 STAIRCASE CLIMBING 242
8.5 MANHATTAN PATHS 244
8.6 CONVEX POLYGON TRIANGULATIONS 245
8.7 CIRCLE PYRAMIDS 248
8.8 EXERCISES 250
Chapter 9 Mutual Recursion 253
9.1 PARITY OF A NUMBER 254

9.2 MULTIPLAYER GAMES 255
9.3 RABBIT POPULATION GROWTH 256
9.3.1 Adult and baby rabbit pairs 257
9.3.2 Rabbit family tree 258
9.4 WATER TREATMENT PLANTS PUZZLE 263
9.4.1 Water flow between cities 263
9.4.2 Water discharge at each city 265
9.5 CYCLIC TOWERS OF HANOI 268
9.6 GRAMMARS AND RECURSIVE DESCENT PARSERS 273
9.6.1 Tokenization of the input string 274
9.6.2 Recursive descent parser 279
9.7 EXERCISES 288
Chapter 10 Program Execution 291
10.1 CONTROL FLOW BETWEEN SUBROUTINES 292

10.2 RECURSION TREES 297
10.2.1 Runtime analysis 303
10.3 THE PROGRAM STACK 305
xii Contents
10.3.1 Stack frames 306

10.3.2 Stack traces 309
10.3.3 Computational space complexity 310
10.3.4 Maximum recursion depth and stack over-
flow errors 312
10.3.5 Recursion as an alternative to a stack data
structure 313
10.4 MEMOIZATION AND DYNAMIC PROGRAMMING 317
10.4.1 Memoization 317
10.4.2 Dependency graph and dynamic programming 322
10.5 EXERCISES 325
Chapter 11 Tail Recursion Revisited and Nested Recursion 333
11.1 TAIL RECURSION VS. ITERATION 333

11.2 TAIL RECURSION BY THINKING ITERATIVELY 337
11.2.1 Factorial 337
11.3 NESTED RECURSION 342
11.3.1 The Ackermann function 342
11.3.2 The McCarthy 91 function 342
11.3.3 The digital root 343
11.4 TAIL AND NESTED RECURSION THROUGH FUNC-
TION GENERALIZATION 344
11.4.1 Factorial 345
11.5 EXERCISES 350
Chapter 12 Multiple Recursion III: Backtracking 353
12.1 INTRODUCTION 354

12.1.1 Partial and complete solutions 354
12.1.2 Recursive structure 356
12.2 GENERATING COMBINATORIAL ENTITIES 358
Contents xiii
12.2.1 Subsets 359

12.2.2 Permutations 364
12.3 THE N -QUEENS PROBLEM 368
12.3.1 Finding every solution 370
12.3.2 Finding one solution 372
12.4 SUBSET SUM PROBLEM 372
12.5 PATH THROUGH A MAZE 377
12.6 THE SUDOKU PUZZLE 384
12.7 0-1 KNAPSACK PROBLEM 388
12.7.1 Standard backtracking algorithm 389
12.7.2 Branch and bound algorithm 393
12.8 EXERCISES 397
FURTHER READING 403
Index 407
Preface
Recursion is one of the most fundamental concepts in computer science

and a key programming technique that, similarly to iteration, allows
computations to be carried out repeatedly. It accomplishes this by em-
ploying methods that invoke themselves, where the central idea consists
of designing a solution to a problem by relying on solutions to smaller
instances of the same problem. Most importantly, recursion is a powerful
problem-solving approach that enables programmers to develop concise,
intuitive, and elegant algorithms.
Despite the importance of recursion for algorithm design, most pro-
gramming books do not cover the topic in detail. They usually devote
just a single chapter or a brief section, which is often insufficient for
assimilating the concepts needed to master the topic. Exceptions in-
clude Recursion via Pascal, by J. S. Rohl (Cambridge University Press,
1984); Thinking Recursively with Java, by E. S. Roberts (Wiley, 2006);
and Practicing Recursion in Java, by I. Pevac (CreateSpace Indepen-
dent Publishing Platform, 2016), which focus exclusively on recursion.
The current book provides a comprehensive treatment of the topic, but
differs from the previous texts in several ways.
Numerous computer programming professors and researchers in the
field of computer science education agree that recursion is difficult for
novice students. With this in mind, the book incorporates several ele-
ments in order to foster its pedagogical effectiveness. Firstly, it contains
a larger collection of simple problems in order to provide a solid foun-
dation of the core concepts, before diving into more complex material.
In addition, one of the book’s main assets is the use of a step-by-step
methodology, together with specially designed diagrams, for guiding and
illustrating the process of developing recursive algorithms. The book also
contains specific chapters on combinatorial problems and mutual recur-
sion. These topics can broaden students’ understanding of recursion by
forcing them to apply the learned concepts differently, or in a more so-
phisticated manner. Lastly, introductory programming courses usually
focus on the imperative programming paradigm, where students primar-
xv
xvi Preface
ily learn iteration, understanding and controlling how programs work. In

contrast, recursion requires adopting a completely different way of think-
ing, where the emphasis should be on what programs compute. In this
regard, several studies encourage instructors to avoid, or postpone, cov-
ering how recursive programs work (i.e., control flow, recursion trees, the
program stack, or the relationship between iteration and tail recursion)
when introducing recursion, since the concepts and abilities learned for
iteration may actually hamper the acquisition of skills related to recur-
sion and declarative programming. Therefore, topics related to iteration
and program execution are covered towards the end of the book, when
the reader should have mastered the design of recursive algorithms from
a purely declarative perspective.
The book also includes a richer chapter on the theoretical analysis
of the computational cost of recursive programs. On the one hand, it
contains a broad treatment of mathematical recurrence relations, which
constitute the fundamental tools for analyzing the runtime or recursive
algorithms. On the other hand, it includes a section on mathematical
preliminaries that reviews concepts and properties that are not only
needed for solving recurrence relations, but also for understanding the
statements and solutions to the computational problems in the book.
In this regard, the text also offers the possibility to learn some basic
mathematics along the way. The reader is encouraged to embrace this
material, since it is essential in many fields of computer science.
The code examples are written in Python 3, which is arguably today’s
most popular introductory programming language in top universities. In
particular, they were tested on Spyder (Scientific PYthon Development
EnviRonment). The reader should be aware that the purpose of the book
is not to teach Python, but to transmit skills associated with recursive
thinking for problem solving. Thus, aspects such as code simplicity and
legibility have been prioritized over efficiency. In this regard, the code
does not contain advanced Python features. Therefore, students with
background in other programming languages such as C++ or Java should
be able to understand the code without effort. Of course, the methods
in the book can be implemented in several ways, and readers are encour-
aged to write more efficient versions, include more sophisticated Python
constructs, or design alternative algorithms. Lastly, the book provides
recursive variants of iterative algorithms that usually accompany other
well-known recursive algorithms. For instance, it contains recursive ver-
sions of Hoare’s partition method used in the quicksort algorithm, or of
the merging method within the merge sort algorithm.
Preface xvii
The book proposes numerous exercises at the end of the chapters,

whose fully worked-out solutions are included in an instructor’s manual
available at the book’s official website (see www.crcpress.com). Many of
them are related to the problems analyzed in the main text, which make
them appropriate candidates for exams and assignments.
The code in the text will also be available for download at the book’s
website. In addition, I will maintain a complementary website related
to the book: https://sites.google.com/view/recursiveprogrammingintro/.
Readers are more than welcome to send me comments, suggestions for
improvements, alternative (clearer or more efficient) code, versions in
other programming languages, or detected errata. Please send emails to:
recursion.book@gmail.com.
INTENDED AUDIENCE
The main goal of the book is to teach students how to think and program
recursively, by analyzing a wide variety of computational problems. It is
intended mainly for undergraduate students in computer science or re-
lated technical disciplines that cover programming and algorithms (e.g.,
bioinformatics, engineering, mathematics, physics, etc.). The book could
also be useful for amateur programmers, students of massive open online
courses, or more experienced professionals who would like to refresh the
material, or learn it in a different or clearer way.
Students should have some basic programming experience in order
to understand the code in the book. The reader should be familiar with
notions introduced in a first programming course such as expressions,
variables, conditional and loop constructs, methods, parameters, or el-
ementary data structures such as arrays or lists. These concepts are
not explained in the book. Also, the code in the book is in accordance
with the procedural programming paradigm, and does not use object
oriented programming features. Regarding Python, a basic background
can be helpful, but is not strictly necessary. Lastly, the student should
be competent in high school mathematics.
Computer science professors can also benefit from the book, not just
as a handbook with a large collection and variety of problems, but also
by adopting the methodology and diagrams described to build recursive
solutions. Furthermore, professors may employ its structure to organize
their classes. The book could be used as a required textbook in introduc-
tory (CS1/2) programming courses, and in more advanced classes on the
design and analysis of algorithms (for example, it covers topics such as
xviii Preface
divide and conquer, or backtracking). Additionally, since the book pro-

vides a solid foundation of recursion, it can be used as a complementary
text in courses related to data structures, or as an introduction to func-
tional programming. However, the reader should be aware that the book
does not cover data structures or functional programming concepts.
BOOK CONTENT AND ORGANIZATION

The first chapter assumes that the reader does not have any previous
background on recursion, and introduces fundamental concepts, nota-
tion, and the first coded examples.
The second chapter presents a methodology for developing recursive
algorithms, as well as diagrams designed to help thinking recursively,
which illustrate the original problem and its decomposition into smaller
instances of the same problem. It is one of the most important chapters
since the methodology and recursive diagrams will be used throughout
the rest of the book. Readers are encouraged to read the chapter, re-
gardless of their previous background on recursion.
Chapter 3 reviews essential mathematical fundamentals and nota-
tion. Moreover, it describes methods for solving recurrence relations,
which are the main mathematical tools for theoretically analyzing the
computational cost of recursive algorithms. The chapter can be skipped
when covering recursion in an introductory course. However, it is in-
cluded early in the book in order to provide a context for characteriz-
ing and comparing different algorithms regarding their efficiency, which
would be essential in a more advanced course on design and analysis of
algorithms.
The fourth chapter covers “linear recursion.” This type of recursion
leads to the simplest recursive algorithms, where the solutions to com-
putational problems are obtained by considering the solution to a single
smaller instance of the problem. Although the proposed problems can
also be solved easily through iteration, they are ideal candidates for in-
troducing fundamental recursive concepts, as well as examples of how to
use the methodology and recursive diagrams.
The fifth chapter covers a particular type of linear recursion called
“tail recursion,” where the last action performed by a method is a re-
cursive call, invoking itself. Tail recursion is special due to its relation-
ship with iteration. This connection will nevertheless be postponed until
Chapter 11. Instead, this chapter focuses on solutions from a purely
declarative approach, relying exclusively on recursive concepts.
Preface xix
The advantages of recursion over iteration are mainly due to the use
of “multiple recursion,” where methods invoke themselves several times,
and the algorithms are based on combining several solutions to smaller
instances of the same problem. Chapter 6 introduces multiple recursion
through methods based on the eminent “divide and conquer” algorithm
design paradigm. While some examples can be used in an introductory
programming course, the chapter is especially appropriate in a more ad-
vanced class on algorithms. Alternatively, Chapter 7 contains challenging
problems, related to puzzles and fractal images, which can also be solved
through multiple recursion, but are not considered to follow the divide
and conquer approach.
Recursion is used extensively in combinatorics, which is a branch
of mathematics related to counting that has applications in advanced
analysis of algorithms. Chapter 8 proposes using recursion for solving
combinatorial counting problems, which are usually not covered in pro-
gramming texts. This unique chapter will force the reader to apply the
acquired recursive thinking skills to a different family of problems. Lastly,
although some examples are challenging, many of the solutions will have
appeared in earlier chapters. Thus, some examples can be used in an
introductory programming course.
Chapter 9 introduces “mutual recursion,” where several methods in-
voke themselves indirectly. The solutions are more sophisticated since it
is necessary to think about several problems simultaneously. Neverthe-
less, this type of recursion involves applying the same essential concepts
covered in earlier chapters.
Chapter 10 covers how recursive programs work from a low-level
point of view. It includes aspects such as tracing and debugging, the
program stack, or recursion trees. In addition, it contains a brief intro-
duction to memoization and dynamic programming, which is another
important algorithm design paradigm.
Tail-recursive algorithms can not only be transformed to iterative
versions; some are also designed by thinking iteratively. Chapter 11 ex-
amines the connection between iteration and tail recursion in detail. In
addition, it provides a brief introduction to “nested recursion,” and in-
cludes a strategy for designing simple tail-recursive functions that are
usually defined by thinking iteratively, but through a purely declarative
approach. These last two topics are curiosities regarding recursion, and
should be skipped in introductory courses.
The last chapter presents backtracking, which is another major al-
gorithm design technique that is used for searching for solutions to com-
xx Preface
putational problems in large discrete state spaces. The strategy is usu-

ally applied for solving constraint satisfaction and discrete optimization
problems. For example, the chapter will cover classical problems such as
the N-queens puzzle, finding a path through a maze, solving sudokus, or
the 0-1 knapsack problem.
POSSIBLE COURSE ROAD MAPS

It is possible to cover only a subset of the chapters. The road map for
introductory programming courses could be Chapters 1, 2, 4, 5, and 10.
The instructor should decide whether to include examples from Chap-
ters 6–9, and whether to cover the first section of Chapter 11.
If students have previously acquired skills to develop linear-recursive
methods, a more advanced course on algorithm analysis and design could
cover Chapters 2, 3, 5, 6, 7, 9, 11, and 12. Thus, Chapters 1, 4, and 10
could be proposed as readings for refreshing the material. In both of
these suggested road maps Chapter 8 is optional. Finally, it is important
to cover Chapters 10 and 11 after the previous ones.
ACKNOWLEDGEMENTS
The content of this book has been used to teach computer programming
courses at Universidad Rey Juan Carlos, in Madrid (Spain). I am grate-
ful to the students for their feedback and suggestions. I would also like
to thank Ángel Velázquez and the members of the LITE (Laboratory
of Information Technologies in Education) research group for providing
useful insights regarding the content of the book. I would also like to
express my gratitude to Luís Fernández, computer science professor at
Universidad Politécnica de Madrid, for his advice and experience related
to teaching recursion. A special thanks to Gert Lanckriet and members
of the Computer Audition Laboratory at University of California, San
Diego.
Manuel Rubio-Sánchez
July, 2017
List of Figures
1.1 Examples of recursive entities. 2

1.2 Recursive decomposition of lists and trees. 6
1.3 Family tree representing the descendants of a per-
son, which are its children plus the descendants of
its children. 7
1.4 Recursive problem solving. 8
1.5 Decompositions of the sum of the first positive integers. 9
1.6 Decompositions of the sum of the elements in a list,
denoted as a, of (n = 9) numbers. 12
1.7 Functions that compute the sum of the first n nat-
ural numbers in several programming languages. 15
1.8 Data structures similar to lists, and parameters nec-
essary for defining sublists. 18
1.9 Thought experiment in a classroom, where an in-
structor asks a student to add the first 100 positive
integers. S(n) represents the sum of the first n pos-
itive integers. 23
2.1 General template for designing recursive algorithms. 32

2.2 Additional diagrams that illustrate the decomposi-
tion of the sum of the first positive integers when
the problem size is decreased by a unit. 38
2.3 When thinking recursively we generally do not need
to decompose a problem into every instance of
smaller size. 39
2.4 Decompositions of the Fibonacci function. 40
xxi
xxii List of Figures
2.5 General diagram for thinking about recursive cases

(when a problem is decomposed into a single self-
similar subproblem). 42
2.6 Diagram showing a decomposition of the sum of the
first n positive integers S(n) that uses two subprob-
lems of half the size as the original. 45
2.7 Diagram showing a decomposition of the problem
consisting of printing the digits of a nonnegative in-
teger on the console, in reversed order, and verti-
cally. A particular (n = 2743) and a general m-digit
(n = dm−1 ⋯ d1 d0 ) case are shown in (a) and (b),
respectively. 49
2.8 General diagram for thinking about recursive cases,
when a problem is decomposed into several (N ) self-
similar subproblems. 50
2.9 Alternative diagram showing a divide and conquer
decomposition, and the recursive thought process,
for the problem of finding the largest value in a list.
The thick and thin arrows point to the solutions of
the problem and subproblems, respectively. 50
2.10 Diagram showing a decomposition of the sum of the
first n positive integers S(n) that uses two subprob-
lems of (roughly) half the size as the original, when
n is odd. 54
3.1 Graphical mnemonic for determining the quadratic

formula for the sum of the first n positive integers (S(n)). 63
3.2 Right triangle used for showing trigonometric definitions. 67
3.3 Geometric interpretation of vector addition and sub-
traction. 69
3.4 Rotation matrix (counterclockwise). 70
3.5 The highest-order term determines the order of
growth of a function. For T (n) = 0.5n2 + 2000n +
50000 the order is quadratic, since the term 0.5n2
clearly dominates the lower-order terms (even added
up) for large values of n. 71
List of Figures xxiii
3.6 Orders of growth typically used in computational

complexity. 72
3.7 Graphical illustrations of asymptotic notation defi-
nitions for computational complexity. 74
3.8 Sequence of operations carried by the function in
Listing 1.1 in the base case. 77
3.9 Sequence of operations carried by the function in
Listing 1.1 in the recursive case. 78
3.10 Summary of the expansion method. 81
3.11 An algorithm processes the shaded bits of numbers
from 1 to 2n − 1 (for n = 4). 102
4.1 Conversion of 142 into its representation (1032) in

base 5. 118
4.2 Pascal’s triangle. 126
4.3 Decomposition and recovery of a row of Pascal’s triangle. 126
4.4 Ladder of resistors problem. 127
4.5 Equivalence of circuits with resistors. 127
4.6 Decomposition of the ladder of resistors problem,
and derivation of the recursive case through induction. 128
4.7 The product of two nonnegative integers n and m
can be represented as the number of unit squares
that form an n × m rectangle. 130
4.8 Step of the insertion sort algorithm. 132
5.1 Decomposition related to the binary search algorithm. 141

5.2 Binary search tree that stores information about a
birthday calendar. 145
5.3 Binary search tree in Figure 5.2 and (5.1), where
each node is a list of four elements: name (string),
birthday (string), left subtree (list), and right sub-
tree (list). 145
5.4 Decomposition associated with several algorithms
related to binary search trees. 146
5.5 Partitioning of a list used in the quicksort and quick-
select algorithms. 148
xxiv List of Figures
5.6 Example of Hoare’s partition method. 151

5.7 Decomposition of Hoare’s partitioning problem. 153
5.8 Base case and problem decomposition used by the
quickselect algorithm. 155
5.9 Steps of the bisection algorithm. 158
5.10 Base case of the bisection algorithm (b − a ≤ 2ǫ). 159
5.11 Instance of the woodcutter problem. 160
5.12 Decomposition of the woodcutter problem. 162
5.13 Steps of the binary search algorithm related to an
instance of the woodcutter problem, for w = 10. 164
5.14 Steps in the counting sort algorithm. 167
5.15 Main idea behind Newton’s method. 169
6.1 Merge sort algorithm. 174

6.2 Decomposition of the quicksort algorithm. 178
6.3 Types of trominoes ignoring rotations and reflections. 192
6.4 Tromino tiling problem. 192
6.5 Decomposition of the tromino tiling problem. 192
6.6 L trominoes considering rotations. 195
6.7 The skyline problem. 197
6.8 Base case for the skyline problem with one building. 198
6.9 Recursive case for the skyline problem. 199
6.10 Possible situations when merging skylines that
change at the same location x. 200
6.11 Possible situations when merging skylines and x1 < x2 . 202
7.1 Swamp traversal problem. 206

7.2 Decomposition of the swamp traversal problem. 207
7.3 The towers of Hanoi puzzle. 209
7.4 Solution to the towers of Hanoi puzzle for n = 2 disks. 210
7.5 Solution to the towers of Hanoi puzzle for n = 4 disks. 212
7.6 Decomposition of the towers of Hanoi problem. 213
7.7 Output of Listing 7.3, which represents the solution
to the towers of Hanoi puzzle for n = 4 disks. 214
List of Figures xxv
7.8 Concrete example of the decomposition of the in-

order traversal problem. 215
7.9 Decomposition of the problem of finding the longest
palindrome substring. 218
7.10 Koch curve fractal. 221
7.11 Koch snowflake fractal. 222
7.12 Koch curve decomposition. 223
7.13 New endpoints of shorter segments when applying
an iteration related to the Koch curve. 224
7.14 Sierpiński’s carpet after 0, 1, 2, and 3 iterations. 226
7.15 Sierpiński’s carpet decomposition. 227
7.16 Simulation of tick marks on an English ruler. 229
7.17 Rules for the “sideways” variant of the towers of
Hanoi puzzle. 230
7.18 Sierpiński’s triangle after 0, 1, 2, and 3 iterations. 231
7.19 Hilbert curves of orders 1–6. 232
8.1 Possible permutations (lists) of the first four positive

integers. 236
8.2 Decomposition of the problem that counts the num-
ber of possible permutations of n different elements,
denoted as f (n). 237
8.3 Decomposition of the possible permutations of the
first four positive integers. 238
8.4 Example of a k-element variation with repetition of
n items. 239
ber of k-element variations with repetition of n items. 240
ber of k-element combinations of n items. 241
8.7 A possible way to climb a staircase by taking leaps
of one or two steps. 242
ber of ways to climb a staircase by taking leaps of
one or two steps. 243
xxvi List of Figures
8.9 Manhattan paths problem. 244

8.10 Decomposition of the Manhattan paths problem. 245
8.11 Two possible triangulations of the same convex poly-
gon containing seven vertices. 246
8.12 Six (n−2) possible triangles associated with an edge
of an octagon (n = 8). 246
8.13 Decomposition of the convex polygon triangulation
problem for fixed triangles. 247
8.14 Total number of triangulations related to an octagon. 247
8.15 Valid and invalid circle pyramids. 248
8.16 Pyramids of n = 4 circles on the bottom row,
grouped according to how many circles appear in
the row immediately above it. 249
8.17 Decomposition of the circle pyramids problem for
subproblems of a fixed size. 250
8.18 Two-element variations with repetition of the four
items in {a,b,c,d}. 250
8.19 Tiling of a 2 × 10 rectangle with 1 × 2 or 2 × 1 domino tiles. 251
8.20 Five different binary trees that contain three nodes. 252
8.21 Pyramids of n = 4 circles on the bottom row,
grouped according to their height. 252
9.1 Calls of mutually recursive methods. 254

9.2 Illustration of the rabbit population growth rules. 257
9.3 Rabbit family tree after seven months. 259
9.4 Concrete example of the decomposition of the rabbit
population growth problem into self-similar subproblems. 260
9.5 Concrete decompositions of the problems that lead
to two mutually recursive methods for solving the
rabbit population growth problem. 261
9.6 Water treatment plants puzzle. 263
9.7 Decomposition of the water treatment plants puzzle
when modeling the water flow between cities. 264
9.8 Three problems for the mutually recursive solution
of the water treatment plants puzzle. 266
List of Figures xxvii
9.9 Decompositions of the three problems for the mutu-

ally recursive solution of the water treatment plants
puzzle. 267
9.10 The cyclic towers of Hanoi puzzle. 268
9.11 Illustration of two different problems comprised in
the cyclic towers of Hanoi puzzle. 269
9.12 Three operations used to implement the recursive
methods for solving the cyclic towers of Hanoi puzzle. 270
9.13 Operations for moving n disks clockwise. 271
9.14 Operations for moving n disks counterclockwise. 272
9.15 Decomposition of a mathematical formula into cat-
egories such as expressions, terms, factors, or numbers. 280
9.16 Method calls of the mutually recursive functions
that implement the recursive descent parser asso-
ciated with the calculator program. 282
9.17 Possible decompositions of an expression. 284
9.18 Rules of the Spin-out brainteaser challenge.
®
289
10.1 Sequence of method calls and returns for the code

in Listing 10.1. 293
10.2 Sequence of calls and returns for sum_first_naturals(4). 294
10.3 Sequence of calls and returns for the procedure
mystery_method_1(’Word’). 296
10.4 Sequence of calls and returns for the procedure
mystery_method_2(’Word’). 298
10.5 Recursion trees for sum_first_naturals(4). 299
10.6 Activation tree for sum_first_naturals(4). 299
10.7 Activation tree for gcd1(20,24). 300
10.8 Recursion (a) and activation (b) trees for fibonacci(6). 301
10.9 Activation tree for the mutually recursive functions
in Listing 9.1. 302
10.10 Recursion tree for the mutually recursive functions
in (1.17) and (1.18). 303
10.11 Activation tree for the nested recursive function in (1.19). 304
xxviii List of Figures
10.12 Recursion tree associated with the recurrence

T (n) = 2T (n/2) + n2 . 305
10.13 The stack and queue data structures. 306
10.14 Evolution of stack frames when running the code
in Listing 10.1, where cos, ∣ ⋅ ∣, and ⟨⋅, ⋅⟩ repre-
sent the methods cosine, norm_Euclidean, and
dot_product, respectively. 307
10.15 Program stack and data at step 5 in Figure 10.14. 308
10.16 Evolution of stack frames for sum_first_naturals(4). 309
10.17 Evolution of stack frames for fibonacci(5). 311
10.18 File system tree example. 313
10.19 State of a stack data structure when running the
iterative code in Listing 10.3, for the files and folders
in Figure 10.18. 315
10.20 Evolution of stack frames when running the code in
Listing 10.4, for the files and folders in Figure 10.18. 317
10.21 Overlapping subproblems when computing Fi-
bonacci numbers through F (n) = F (n − 1) + F (n − 2). 319
10.22 Dependency graph for F (n) = F (n − 1) + F (n − 2). 322
10.23 Dependency graph for Listing 10.6, which solves the
longest palindrome substring problem. 323
10.24 Matrix L after running Listing 10.7 with s = 'bcaac'. 325
10.25 Alternative binary search tree that stores informa-
tion about a birthday calendar. 326
10.26 Transformation of a line segment into five smaller
ones for a Koch curve variant. 330
10.27 “Koch square” variant for n = 4. 330
11.1 State of the program stack when running the base

case relative to a call to gcd1(20,24). 334
11.2 Similarities between the iterative and tail-recursive
codes that compute the factorial function. 339
11.3 McCarthy 91 function. 343
12.1 One solution to the four-queens puzzle. 354

List of Figures xxix
12.2 Partial solutions within complete solutions that are

coded as lists or matrices. 355
12.3 Recursion tree of a backtracking algorithm that
finds one solution to the four-queens puzzle. 356
12.4 Binary recursion tree of an algorithm that generates
all of the subsets of three items. 359
12.5 Alternative binary recursion tree of an algorithm
that generates all of the subsets of three items. 362
12.6 Recursion tree of an algorithm that generates all of
the permutations of three items. 364
12.7 Pruning a recursion tree as soon as a partial solution
is not valid. 367
12.8 Indexing principal and secondary diagonals on a ma-
trix or chessboard. 369
12.9 Recursion tree of the procedure that solves the sub-
set sum problem for S = {2, 6, 3, 5} and x = 8. 376
12.10 Problem of finding a path through a maze, and solu-
tion through backtracking when searching in a par-
ticular order. 378
12.11 Different paths through a maze, depending on the
search order. 379
12.12 Decomposition of the problem of finding a path
through a maze. 380
12.13 An instance of the sudoku puzzle and its solution. 384
12.14 Recursive cases for the sudoku solver. 385
12.15 Recursion tree of a backtracking algorithm for the
0-1 knapsack problem. 390
12.16 Recursion tree of a branch and bound algorithm for
the 0-1 knapsack problem. 395
12.17 Alternative recursion tree of an algorithm that gen-
erates all of the subsets of three items. 398
12.18 One solution to the four-rooks puzzle. 398
12.19 A 3 × 3 magic square. 399
12.20 Chess knight moves. 399
xxx List of Figures
12.21 An instance and solution to the traveling salesman

problem. 400
12.22 Two ways to represent a solution for the tug of war
problem. 401
List of Tables
3.1 Concrete values of common functions used in com-

putational complexity. 72
11.1 Program state related to the iterative factorial func-

tion when computing 4!. 338
11.2 Program state related to the iterative base conver-
sion function in Listing 11.5, when obtaining the
base-5 representation of 142, which is 1032. 340
xxxi
List of Listings
1.1 Python code for adding the first n natural numbers. 16

1.2 Alternative Python code for adding the first n nat-
ural numbers. 16
1.3 Python code for computing the n-th Fibonacci number. 17
1.4 Alternative Python code for computing the n-th Fi-
bonacci number. 17
1.5 Recursive functions for adding the elements in a list
a, where the only input to the recursive function is
the list. 19
1.6 Alternative recursive functions for adding the ele-
ments in a sublist of a list a. The boundaries of the
sublist are specified by two input parameters that
mark lower and upper indices in the list. 21
2.1 Misconceptions regarding base cases through the
factorial function. 35
2.2 Code for computing the sum of the digits of a non-
negative integer. 47
2.3 Code for printing the digits of a nonnegative integer
vertically, and in reversed order. 49
2.4 Code for computing the maximum value in a list,
through a divide and conquer approach. 51
2.5 Erroneous Python code for determining if a nonneg-
ative integer n is even. 52
2.6 Correct Python code for determining if a nonnega-
tive integer n is even. 52
2.7 Erroneous Python code for adding the first n posi-
tive numbers, which produces infinite recursions for
most values of n. 53
xxxiii
xxxiv List of Listings
2.8 Incomplete Python code for adding the first n posi-

tive numbers. 54
2.9 Python code for adding the first n positive numbers
based on using two subproblems of (roughly) half
the size as the original. 55
3.1 Measuring execution times through Python’s time
module. 70
3.2 Solving a system of linear equations, Ax = b, in Python. 94
4.1 Power function in linear time for nonnegative exponents. 107
4.2 Power function in linear time. 108
4.3 Power function in logarithmic time for nonnegative
exponents. 109
4.4 Inefficient implementation of the power function
that runs in linear time. 109
4.5 Slow addition of two nonnegative integers. 111
4.6 Quicker slow addition of two nonnegative integers. 112
4.7 Alternative quicker slow addition of two nonnegative
integers. 113
4.8 Recursive functions that compute the double sum in (4.3). 115
4.9 Binary representation of a nonnegative integer. 117
4.10 Conversion of a nonnegative integer n into its rep-
resentation in base b. 118
4.11 Conversion of a nonnegative integer n into its rep-
resentation in base b. 120
4.12 Function that determines if a string is a palindrome. 121
4.13 Recursive selection sort algorithm. 123
4.14 Recursive variant of the selection sort algorithm. 123
4.15 Horner’s method for evaluating a polynomial. 125
4.16 Function that generates the n-th row of Pascal’s triangle. 126
4.17 Function that solves the ladder of resistors problem. 129
5.1 Linear-recursive Boolean function that determines if
a nonnegative integer contains a digit. 135
5.2 Tail-recursive Boolean function that determines if a
nonnegative integer contains a digit. 136
List of Listings xxxv
5.3 Linear-recursive function that determines if two

strings are identical. 138
5.4 Tail-recursive function that determines if two strings
are identical. 138
5.5 Tail-recursive linear search of an element in a list. 139
5.6 Linear-recursive linear search of an element in a list. 140
5.7 Alternative tail-recursive linear search of an element
in a list. 141
5.8 Binary search of an element in a list. 142
5.9 Algorithm for searching an item with a particular
key in a binary search tree. 146
5.10 Procedure for inserting an item with a particular key
in a binary search tree. 147
5.11 Auxiliary methods for partitioning a list. 150
5.12 Hoare’s iterative partitioning algorithm. 152
5.13 Alternative recursive version of Hoare’s partitioning
scheme. 154
5.14 Tail-recursive quickselect algorithm. 156
5.15 Bisection algorithm. 159
5.16 Function that computes the amount of wood col-
lected when cutting trees at height h. 160
5.17 Binary search algorithm for the woodcutter problem. 163
5.18 Euclid’s algorithm for computing the greatest com-
mon divisor of two nonnegative integers. 165
6.1 Function that determines whether a list is sorted in
ascending order. 172
6.2 Merge sort method. 175
6.3 Method for merging two sorted lists. 177
6.4 Variant of the quicksort algorithm. 179
6.5 In-place quicksort algorithm. 179
6.6 Code for counting the number of times an element
appears in a list. 182
6.7 Code for solving the majority element problem. 183
xxxvi List of Listings
6.8 Karatsuba’s fast algorithm for multiplying two non-

negative integers. 185
6.9 Divide and conquer matrix multiplication. 188
6.10 Alternative divide and conquer matrix multiplication. 189
6.11 Auxiliary functions for drawing trominoes. 193
6.12 Recursive method for drawing trominoes. 194
6.13 Code for calling the trominoes method. 195
6.14 Main recursive method for computing skylines. 198
6.15 Recursive method for merging skylines. 201
7.1 Function that determines whether there exists a
path through a swamp. 207
7.2 Alternative function that determines whether there
exists a path through a swamp. 208
7.3 Towers of Hanoi procedure. 214
7.4 Inorder traversal of a binary tree. 216
7.5 Preorder and postorder traversals of a binary tree. 216
7.6 Code for finding the longest palindrome substring
(or sublist). 218
7.7 Alternative code for finding the longest palindrome
substring (or sublist). 220
7.8 Code for generating Koch curves and the Koch snowflake. 225
7.9 Code for generating Sierpiński’s carpet. 228
9.1 Mutually recursive functions for determining the
parity of a nonnegative integer n. 255
9.2 Mutually recursive procedures implementing Alice
and Bob’s strategies when playing a game. 256
9.3 Mutually recursive functions for counting the popu-
lation of baby and adult rabbits after n months. 258
9.4 Alternative mutually recursive functions for count-
ing the population of rabbits after n months. 262
9.5 Function based on multiple recursion for solving the
water treatment plants puzzle. 265
9.6 Mutually recursive procedures for the cyclic towers
of Hanoi puzzle. 273
List of Listings xxxvii
9.7 Mutually recursive functions for tokenizing a math-

ematical expression. 275
9.8 Function that checks whether a string represents a
number. 276
9.9 Function that parses a mathematical expression of
additive terms. 283
9.10 Function that parses a term of multiplicative factors. 285
9.11 Function that parses a term of multiplicative factors,
where the first one is a parenthesized expression. 286
9.12 Function that parses a mathematical factor. 287
9.13 Function that parses a parenthesized expression. 287
9.14 Basic code for executing the calculator program. 287
10.1 Methods for computing the cosine of the angle be-
tween two vectors. 292
10.2 Similar recursive methods. What do they do? 295
10.3 Iterative algorithm for finding a file in a file system. 314
10.4 Recursive algorithm for finding a file in a file system. 316
10.5 Recursive algorithm for computing Fibonacci num-
bers in linear time, by using memoization. 320
10.6 Memoized version of Listing 7.7. 321
10.7 Code based on dynamic programming that com-
putes the longest palindrome substring within a
string s. 324
10.8 Methods that, supposedly, add and count the digits
of a nonnegative integer. Are they correct? 327
10.9 Erroneous code for computing the number of times
that two adjacent elements in a list are identical. 328
10.10 Code for computing the smallest prime factor of a
number n, which is greater than or equal to m. 328
10.11 Erroneous code for computing the floor of a logarithm. 328
10.12 Erroneous code for determining if a list contains an
element that is larger than the sum of all of the rest. 329
10.13 Erroneous code for finding the location of the “peak
element.” 330
xxxviii List of Listings
10.14 Code for generating a Koch fractal based on the

transformation in Figure 10.26. 331
10.15 Code for computing the length of the longest palin-
drome subsequence of a list. 332
11.1 Iterative version of Euclid’s method (gcd1). 335
11.2 Iterative version of the bisection method. 336
11.3 Iterative factorial function. 338
11.4 Tail-recursive factorial function and wrapper method. 338
11.5 Iterative conversion of a nonnegative integer n into
its representation in base b. 340
11.6 Tail-recursive base conversion function and wrapper
method. 341
11.7 The Ackermann function implemented in Python. 341
11.8 Nested-recursive method for finding the digital root
of a nonnegative integer. 344
12.1 Code for printing all of the subsets of the elements
in a list. 360
12.2 Alternative code for printing all of the subsets of the
elements in a list. 363
12.3 Code for printing all of the permutations of the ele-
ments in a list. 366
12.4 Alternative code for printing all of the permutations
of the elements in a list. 368
12.5 Code for finding all of the solutions to the n-queens
puzzle. 371
12.6 Code for finding one solution to the n-queens puzzle. 373
12.7 Backtracking code for solving the subset sum problem. 375
12.8 Backtracking code for finding a path through a maze. 381
12.9 Auxiliary code related to the backtracking methods
for finding a path through a maze. 383
12.10 Code for solving a sudoku puzzle. 386
12.11 Auxiliary code for solving a sudoku puzzle. 387
12.12 Backtracking code for solving the 0-1 knapsack problem. 391
12.13 Auxiliary code related to the 0-1 knapsack problem. 393
List of Listings xxxix
12.14 Branch and bound code for solving the 0-1 knapsack
problem. 396
12.15 Auxiliary code for the branch and bound algorithm
related to the 0-1 knapsack problem. 397
CHAPTER 1
Basic Concepts of
Recursive Programming
To iterate is human, to recurse divine.

— Laurence Peter Deutsch
ECURSION is a broad concept that is used in diverse disciplines

R such as mathematics, bioinformatics, or linguistics, and is even
present in art or in nature. In the context of computer programming,
recursion should be understood as a powerful problem-solving strategy
that allows us to design simple, succinct, and elegant algorithms for
solving computational problems. This chapter presents key terms and
notation, and introduces fundamental concepts related to recursive pro-
gramming and thinking that will be further developed throughout the
book.
1.1 RECOGNIZING RECURSION

An entity or concept is said to be recursive when simpler or smaller
self-similar instances form part of its constituents. Nature provides nu-
merous examples where we can observe this property (see Figure 1.1).
For instance, a branch of a tree can be understood as a stem, plus a
set of smaller branches that emanate from it, which in turn contain
other smaller branches, and so on, until reaching a bud, leaf, or flower.
Blood vessels or rivers exhibit similar branching patterns, where the
larger structure appears to contain instances of itself at smaller scales.
Another related recursive example is a romanesco broccoli, where it is
1
2 Introduction to Recursive Programming
Tree branches Branching rivers
Romanesco broccoli Spiral Droste effect
Sierpiński’s triangle Matryoshka dolls
Figure 1.1 Examples of recursive entities.

Basic Concepts of Recursive Programming 3
apparent that the individual florets resemble the entire plant. Other ex-
amples include mountain ranges, clouds, or animal skin patterns.
Recursion also appears in art. A well-known example is the Droste
effect, which consists of a picture appearing within itself. In theory the
process could be repeated indefinitely, but naturally stops in practice
when the smallest picture to be drawn is sufficiently small (for example,
if it occupies a single pixel in a digital image). A computer-generated
fractal is another type of recursive image. For instance, Sierpiński’s tri-
angle is composed of three smaller identical triangles that are subse-
quently decomposed into yet smaller ones. Assuming that the process is
infinitely repeated, each small triangle will exhibit the same structure as
the original’s. Lastly, a classical example used to illustrate the concept
of recursion is a collection of matryoshka dolls. In this craftwork each
doll has a different size and can fit inside a larger one. Note that the
recursive object is not a single hollow doll, but a full nested collection.
Thus, when thinking recursively, a collection of dolls can be described
as a single (largest) doll that contains a smaller collection of dolls.
While the recursive entities in the previous examples were clearly
tangible, recursion also appears in a wide variety of abstract concepts.
In this regard, recursion can be understood as the process of defining
concepts by using the definition itself. Many mathematical formulas and
definitions can be expressed this way. Clear explicit examples include
sequences for which the n-th term is defined through some formula or
procedure involving earlier terms. Consider the following recursive defi-
nition:
sn = sn−1 + sn−2 . (1.1)
The formula states that a term in a sequence (sn ) is simply the sum of the
two previous terms (sn−1 and sn−2 ). We can immediately observe that the
formula is recursive, since the entity it defines, s, appears on both sides
of the equation. Thus, the elements of the sequence are clearly defined
in terms of themselves. Furthermore, note that the recursive formula in
(1.1) does not describe a particular sequence, but an entire family of
sequences in which a term is the sum of the two previous ones. In order
to characterize a specific sequence we need to provide more information.
In this case, it is enough to indicate any two terms in the sequence.
Typically, the first two terms are used to define this type of sequence.
For instance, if s1 = s2 = 1 the sequence is:
1, 1, 2, 3, 5, 8, 13, 21, 34, 55, . . .

which is the well-known Fibonacci sequence. Lastly, sequences may also

be defined starting at term s0 .
The sequence s can be understood as a function that receives a posi-
tive integer n as an argument, and returns the n-th term in the sequence.
In this regard, the Fibonacci function, in this case simply denoted as F ,
can be defined as:
⎧
⎪ 1 if n = 1,
⎪
⎪
⎪
F (n) = ⎨1 if n = 2, (1.2)
⎪
⎩F (n − 1) + F (n − 2) if n > 2.
⎪
⎪
⎪
Throughout the book we will use this notation in order to describe func-
tions, where the definitions include two types of expressions or cases.
The base cases correspond to scenarios where the function’s output
can be obtained trivially, without requiring values of the function on
additional arguments. For Fibonacci numbers the base cases are, by def-
inition, F (1) = 1, and F (2) = 1. The recursive cases include more
complex recursive expressions that typically involve the defined function
applied to smaller input arguments. The Fibonacci function has one re-
cursive case: F (n) = F (n − 1) + F (n − 2), for n > 2. The base cases are
necessary in order to provide concrete values for the function’s terms
in the recursive cases. Lastly, a recursive definition may contain several
base and recursive cases.
Another function that can be expressed recursively is the factorial of
some nonnegative integer n:
n! = 1 × 2 × ⋯ × (n − 1) × n.
In this case, it is not immediately obvious whether the function can be
expressed recursively, since there is not an explicit factorial on the right-
hand side of the definition. However, since (n − 1)! = 1 × 2 × ⋯ × (n − 1),
we can rewrite the formula as the recursive expression n! = (n − 1)! × n.
Lastly, by convention 0! = 1, which follows from plugging in the value
n = 1 in the recursive formula. Thus, the factorial function can be defined
recursively as:
⎧
⎪
⎪1 if n = 0,
n! = ⎨ (1.3)
⎪(n − 1)! × n if n > 0.
⎪
⎩
Similarly, consider the problem of calculating the sum of the first n
positive integers. The associated function S(n) can be obviously defined
as:
S(n) = 1 + 2 + ⋯ + (n − 1) + n. (1.4)
Again, we do not observe a term involving S on the right-hand side of

the definition. However, we can group the n − 1 smallest terms in order
to form S(n − 1) = 1 + 2 + ⋯ + (n − 1), which leads to the following
recursive definition:
⎧
⎪
⎪1 if n = 1,
S(n) = ⎨ (1.5)
⎩S(n − 1) + n if n > 1.
⎪
⎪
Note that S(n−1) is a self-similar subproblem to S(n), but is simpler,

since it needs fewer operations in order to calculate its result. Thus, we
say that the subproblem has a smaller size. In addition, we say we have
decomposed the original problem (S(n)) into a smaller one, in order
to form the recursive definition. Lastly, S(n − 1) is a smaller instance
of the original problem.
Another mathematical entity for which how it can be expressed re-
cursively may not seem immediately obvious is a nonnegative integer.
These numbers can be decomposed and defined recursively in several
ways, by considering smaller numbers. For instance, a nonnegative inte-
ger n can be expressed as its predecessor plus a unit:
⎧
⎪
⎪0 if n = 0,
n=⎨
⎩predecessor(n) + 1 if n > 0.
⎪
⎪
Note that n appears on both sides of the equals sign in the recursive
case. In addition, if we consider that the predecessor function necessarily
returns a nonnegative integer, then it cannot be applied to 0. Thus, the
definition is completed with a trivial base case for n = 0.
Another way to think of (nonnegative) integers consists of consider-
ing them as ordered collections of digits. For example, the number 5342
can be the concatenation of the following pairs of smaller numbers:
(5, 342), (53, 42), (534, 2).
In practice, the simplest way to decompose these integers consists of

considering the least significant digit individually, together with the rest
of the number. Therefore, an integer can be defined as follows:
⎧
⎪
⎪n if n < 10,
n=⎨
⎩(n//10) × 10 + (n%10) if n ≥ 10,
⎪
⎪
where // and % represent the quotient and remainder of an integer divi-

sion, respectively, which corresponds to Python notation. For example,
Lists Trees
Figure 1.2 Recursive decomposition of lists and trees.
if n = 5342 then the quotient is (n//10) = 534, while the remainder is

(n%10) = 2, which represents the least significant digit of n. Clearly, the
number n can be recovered by multiplying the quotient by 10 and adding
the remainder. Finally, the base case considers numbers with only one
digit that naturally cannot be decomposed any further.
Recursive expressions are abundant in mathematics. For instance,
they are often used in order to describe properties of functions. The
following recursive expression indicates that the derivative of a sum of
functions is the sum of their derivatives:
[f (x) + g(x)]′ = [f (x)]′ + [g(x)]′.
In this case the recursive entity is the derivative function, denoted as [⋅]′ ,
but not the functions f (x) and g(x). Observe that the formula explicitly
indicates the decomposition that takes place, where some initial function
(which is the input argument to the derivative function) is broken up
into the sum of the functions f (x) and g(x).
Data structures can also be understood as recursive entities. Fig-
ure 1.2 shows how lists and trees can be decomposed recursively. On the
one hand, a list can consist of a single element plus another list (this
is the usual definition of a list as an abstract data type), or it can be
subdivided into several lists (in this broader context a list is any collec-
tion of data elements that are arranged linearly in an ordered sequence,
as in lists, arrays, tuples, etc.). On the other hand, a tree consists of a
parent node and a set (or list) of subtrees, whose root node is a child of
the original parent node. The recursive definitions of data structures are
completed by considering empty (base) cases. For instance, a list that
contains only one element would consist of that element plus an empty
list. Lastly, observe that in these diagrams the darker boxes represent a
Figure 1.3 Family tree representing the descendants of a person, which

are its children plus the descendants of its children.
full recursive entity, while the lighter ones indicate smaller self-similar
instances.
Recursion can even be used to define words in dictionaries. This
might seem impossible since we are told in school that the description of
a word in a dictionary should not contain the word itself. However, many
concepts can be defined correctly this way. Consider the term “descen-
dant” of a specific ancestor. Notice that it can be defined perfectly as:
someone who is either a child of the ancestor, or a descendant of any of
the ancestor’s children. In this case, we can identify a recursive structure
where the set of descendants can be organized in order to form a (family)
tree, as shown in Figure 1.3. The darker box contains all of the descen-
dants of a common ancestor appearing at the root of the tree, while the
lighter boxes encompass the descendants of the ancestor’s children.
1.2 PROBLEM DECOMPOSITION

In general, when programming and thinking recursively, our main task
will consist of providing our own recursive definitions of entities, con-
cepts, functions, problems, etc. While the first step usually involves es-
tablishing the base cases, the main challenge consists of describing the
recursive cases. In this regard, it is important to understand the concepts
of: (a) problem decomposition, and (b) induction, which we will cover
briefly in this chapter.
The book will focus on developing recursive algorithms for solving
computational problems. These can be understood as questions that
computers could possibly answer, and are defined through statements
that describe relationships between a collection of known input values or
parameters, and a set of output values, results, or solutions. For example,
Problem
Decompose
Self-similar problems Different problems
Solve and combine
Solution
Figure 1.4 Recursive problem solving.
“given some positive integer n, calculate the sum of the first n positive
integers” is the statement of a computational problem with one input
parameter (n), and one output value defined as 1 + 2 + ⋯ + (n − 1) + n.
An instance of a problem is a specific collection of valid input values
that will allow us to compute a solution to the problem. In contrast,
an algorithm is a logical procedure that describes a step-by-step set
of computations needed in order to obtain the outputs, given the ini-
tial inputs. Thus, an algorithm determines how to solve a problem. It is
worth mentioning that computational problems can be solved by differ-
ent algorithms. The goal of this book is to explain how to design and
implement recursive algorithms and programs, where a key step involves
decomposing a computational problem.
Decomposition is an important concept in computer science and
plays a major role not only in recursive programming, but also in gen-
eral problem solving. The idea consists of breaking up complex problems
into smaller, simpler ones that are easier to express, compute, code, or
solve. Subsequently, the solutions to the subproblems can be processed
in order to obtain a solution to the original complex problem.
In the context of recursive problem solving and programming, de-
composition involves breaking up a computational problem into several
subproblems, some of which are self-similar to the original, as illustrated
in Figure 1.4. Note that obtaining the solution to a problem may re-
Original problem: S(n) S(n) = S(n − 1) + n
n−1
n
(a) (b)
S(n) = 3S( n2 ) + S( n2 − 1) S(n) = 3S( n−1

2
) + S( n+1
2
)
for n even for n odd
n
2 n−1
2
(c) (d)
Figure 1.5 Decompositions of the sum of the first positive integers.
quire solving additional different problems that are not self-similar to

the original one. We will tackle several of these problems throughout the
book, but in this introductory chapter we will examine examples where
the original problems will only be decomposed into self-similar ones.
For the first example we will reexamine the problem of computing
the sum of the first n positive integers, denoted as S(n), which can be
formally expressed as in (1.4). There are several ways to break down
the problem into smaller subproblems and form a recursive definition
of S(n). Firstly, it only depends on the input parameter n, which also
specifies the size of the problem. In this example, the base case is asso-
ciated with the smallest positive integer n = 1, where clearly S(1) = 1 is
the smallest instance of the problem. Furthermore, we need to get closer
to the problem’s base case when considering subproblems. Therefore, we
have to think of how we can reduce the input parameter n.
A first possibility consists of decreasing n by a single unit. In that

case, the goal would be to define S(n) somehow by using the subproblem
S(n − 1). The corresponding recursive solution, derived in Section 1.1
algebraically, is given in (1.5). We can also obtain the recursive case by
analyzing a graphical description of the problem. For instance, the goal
could consist of counting the total number of blocks in a “triangular”
structure that contains n blocks in its first level, n − 1 in its second, and
so on (the n-th level would therefore have a single block), as shown in
Figure 1.5(a) for n = 8. In order to decompose the problem recursively
we need to find self-similar problems. In this case, it is not hard to
find smaller similar triangular shapes inside the original. For example,
Figure 1.5(b) shows a triangular structure of height n − 1 that contains
all of the blocks of the original, except for the n blocks on the first level.
Since this smaller triangular shape contains exactly S(n − 1) blocks, it
follows that S(n) = S(n − 1) + n.
Another option that involves using a problem that adds n − 1 terms
consists of considering the sum 2+ ⋯ +(n−1)+n. However, it is important
to note that it is not a self-similar problem to S(n). Clearly, it is not the
sum of the first positive integers. Instead, it is a special case of a more
general problem consisting of adding all of the integers from some initial
value m up to another larger n: m+(m+1)+ ⋯ +(n−1)+n, for m ≤ n. The
difference between both problems can also be understood graphically.
Regarding the illustrations in Figure 1.5, this general problem would
define right trapezoid structures instead of triangular shapes. Finally, it
is possible to compute the sum of the first n positive integers by using this
more general problem, since we could simply set m = 1. Nevertheless, its
recursive definition is slightly more complex since it requires two input
parameters (m and n) instead of one.
Other possibilities concern decreasing n by larger quantities. For
example, the input value n can be divided by two in order to ob-
tain the decomposition illustrated in Figures 1.5(c) and (d). When n
is even we can fit three triangular structures of height n/2 inside the
larger one corresponding to S(n). Since the remaining blocks also form
a triangular shape of height n/2 − 1 the recursive formula can be ex-
pressed as S(n) = 3S(n/2) + S(n/2 − 1). Alternatively, when n is odd
it is possible to fit three triangular structures of height (n − 1)/2,
and one of size (n + 1)/2. Thus, in that case the recursive formula is
S(n) = 3S((n − 1)/2) + S((n + 1)/2). The final recursive function is:
⎧
⎪
⎪
⎪
1 if n = 1,
⎪
⎪
⎪
⎪3 if n = 2,
S(n) = ⎨
⎪
(1.6)
⎪
⎪3S ( n2 ) + S ( n2 − 1)
⎪
if n > 2 and n is even,
⎪
⎪
⎪
⎩3S ( 2 ) + S ( 2 ) if n > 2 and n is odd.
n−1 n+1
It is important to note that the definition needs an additional base case

for n = 2. Without it we would have S(2) = 3S(1)+S(0) due to the recur-
sive case when n is even. However, S(0) is not defined since, according
to the statement of the problem, the input to S must be a positive in-
teger. The new base case is therefore necessary in order to avoid using
the recursive formula for n = 2.
When dividing the size of a problem the resulting subproblems are
considerably smaller than the original, and can therefore be solved much
faster. Roughly speaking, if the number of subproblems to be solved is
small and it is possible to combine their solutions efficiently, this strat-
egy can lead to substantially faster algorithms for solving the original
problem. However, in this particular example, code for (1.6) is not nec-
essarily more efficient than the one implementing (1.5). Intuitively, this
is because (1.6) requires solving two subproblems (with different argu-
ments), while the decomposition in (1.5) only involves one subproblem.
Chapter 3 covers how to analyze the runtime cost of these recursive
algorithms.
The previous idea for adding the first n positive integers broke up
the problem into two subproblems of smaller size, where the new input
parameters approach the specified base cases. In general, we can decom-
pose problems into any number of simpler subproblems, as long as the
new parameters get closer to the values specified in the base cases. For
instance, consider the following alternative recursive definition of the
Fibonacci function (it is equivalent to (1.2)):
⎧
⎪
⎪
⎪ 1 if n = 1,
⎪
⎪
⎪
⎪1
F (n) = ⎨ n−2
if n = 2,
⎪
(1.7)
⎪
⎪
⎪
⎪
⎪ 1 + ∑ F (i) if n > 2,
⎪
⎩ i=1
i=1 F (i) is the sum F (1)+F (2)+ ⋯ +F (n−2) (see Section 3.1.4).
where ∑n−2
In this example, for some value n that determines the size of the problem,
Original problem: s(a) = a[0] + a[1] + ⋯ + a[n − 1]
5 1 12 4 2 5 6 3 7
0 1 2 3 4 5 6 7 8
(a)
s(a) = s(a[0 ∶ n − 1]) + a[n − 1]
5 1 12 4 2 5 6 3 7
0 1 2 3 4 5 6 7 8
(b)
s(a) = a[0] + s(a[1 ∶ n])
5 1 12 4 2 5 6 3 7
0 1 2 3 4 5 6 7 8
(c)
s(a) = s(a[0 ∶ n//2]) + s(a[n//2 ∶ n])
5 1 12 4 2 5 6 3 7
0 1 2 3 4 5 6 7 8
(d)
Figure 1.6 Decompositions of the sum of the elements in a list, denoted

as a, of (n = 9) numbers.
the recursive case relies on decomposing the original problem into every
smaller problem of size 1 up to n − 2.
For the last example of problem decomposition we will use lists that
allow us to access individual elements by using numerical indices (in
many programming languages this data structure is called an “array,”
while in Python it is simply a “list”). The problem consists of adding
the elements of a list, denoted as a, containing n numbers. Formally, the

problem can be expressed as:
n−1
s(a) = ∑ a[i] = a[0] + a[1] + ⋯ + a[n − 1], (1.8)
i=0
where a[i] is the (i + 1)-th element in the list, since the first one is not
indexed by 1, but by 0. Figure 1.6(a) shows a diagram representing a
particular instance of 9 elements.
Regarding notation, in this book we will assume that a sublist of
some list a is a collection of contiguous elements of a, unless explicitly
stated otherwise. In contrast, in a subsequence of some initial sequence
s its elements appear in the same order as in s, but they are not required
to be contiguous in s. In other words, a subsequence can be obtained
from an original sequence s by deleting some elements of s, and without
modifying the order of the remaining elements.
The problem can be decomposed by decreasing its size by a single
unit. On the one hand, the list can be broken down into the sublist
containing the first n − 1 elements (a[0 ∶ n − 1], where a[i ∶ j] denotes the
sublist from a[i] to a[j − 1], following Python’s notation) and a single
value corresponding to the last number on the list (a[n − 1]), as shown
in Figure 1.6(b). In that case, the problem can be defined recursively
through:
⎧
⎪
⎪0 if n = 0,
s(a) = ⎨
⎪s(a[0 ∶ n − 1]) + a[n − 1] if n > 0.
(1.9)
⎪
⎩
In the recursive case the subproblem is naturally applied to the sublist
of size n − 1. The base case considers the trivial situation when the list
is empty, which does not require any addition. Another possible base
case can be s(a) = a[0] when n = 1. However, it would be redundant in
this decomposition, and therefore not necessary. Note that if n = 1 the
function adds a[0] and the result of applying the function on an empty
list, which is 0. Thus, it can be omitted for the sake of conciseness.
On the other hand, we can also interpret that the original list is its
first element a[0], together with the smaller list a[1 ∶ n], as illustrated
in Figure 1.6(c). In this case, the problem can be expressed recursively
through:
⎧
⎪
⎪0 if n = 0,
s(a) = ⎨
⎪a[0] + s(a[1 ∶ n]) if n > 0.
(1.10)
⎪
⎩
Although both decompositions are very similar, the code for each one
can be quite different depending on the programming language used.
Section 1.3 will show several ways of coding algorithms for solving the
problem according to these decompositions.
Another way to break up the problem consists of considering each
half of the list separately, as shown in Figure 1.6(d). This results in two
subproblems of roughly half the size of the original’s. The decomposition
produces the following recursive definition:
⎧
⎪
⎪
⎪
0 if n = 0,
⎪
s(a) = ⎨a[0]
⎪
if n = 1, (1.11)
⎪
⎪
⎪
⎩s(a[0 ∶ n//2]) + s(a[n//2 ∶ n]) if n > 1.
Unlike the previous definitions, this decomposition requires a base case
when n = 1. Without it the function would never return a concrete value
for a nonempty list. Observe that the definition would not add or return
any element of the list. For a nonempty list the recursive case would
be applied repeatedly, but the process would never halt. This situation
is denoted as infinite recursion. For instance, if the list contained a
single element the recursive case would add the value associated with an
empty list (0), to the result of the same initial problem. In other words,
we would try to calculate s(a) = 0 + s(a) = 0 + s(a) = 0 + s(a) . . ., which
would be repeated indefinitely. The obvious issue in this scenario is that
the original problem s(a) is not decomposed into smaller and simpler
ones when n = 1.
1.3 RECURSIVE CODE

In order to use recursion when designing algorithms it is crucial to learn
how to decompose problems into smaller self-similar ones, and define
recursive methods by relying on induction (see Section 1.4). Once these
are specified it is fairly straightforward to convert the definitions into
code, especially when working with basic data types such as integers,
real numbers, characters, or Boolean values. Consider the function in
(1.5) that adds the first n positive integers (i.e., natural numbers). In
Python the related function can be coded as shown in Listing 1.1. The
analogy between (1.5) and the Python function is evident. As in many of
the examples covered throughout the book, a simple if statement is the
only control flow structure needed in order to code the function. Addi-
tionally, the name of the function appears within its body, implementing
a recursive call. Thus, we say that the function calls or invokes itself,
and is therefore recursive (there exist recursive functions that do not call
themselves directly within their body, as explained in Chapter 9).
C, Java:
1 int sum_first_naturals(int n)
2 {
3 if (n==1)
4 return 1;
5 else
6 return sum_first_naturals(n-1) + n;
7 }
Pascal:
1 function sum_first_naturals(n: integer ): integer ;
2 begin
3 if n=1 then
4 sum_first_naturals := 1
5 else
6 sum_first_naturals := sum_first_naturals(n-1) + n;
7 end;
MATLAB® :
1 function result = sum_first_naturals(n)
2 if n==1
3 result = 1;
4 else
5 result = sum_first_naturals(n-1) + n;
6 end
Scala:
1 def sum_first_naturals(n: Int): Int = {
2 if (n==1)
3 return 1
4 else
5 return sum_first_naturals(n-1) + n
6 }
Haskell:
1 sum_first_naturals 1 = 1
2 sum_first_naturals n = sum_first_naturals (n - 1) + n
Figure 1.7 Functions that compute the sum of the first n natural numbers
in several programming languages.
Listing 1.1 Python code for adding the first n natural numbers.
1 def sum_first_naturals(n):
2 if n == 1:
3 return 1 # Base case
4 else:
5 return sum_first_naturals(n - 1) + n # Recursive case
Coding the function in other programming languages is also straight-

forward. Figure 1.7 shows equivalent code in several programming lan-
guages. Again, the resemblance between the codes and the function def-
inition is apparent. Although the code in the book will be in Python,
translating it to other programming languages should be fairly straight-
forward.
An important detail about the function in (1.5) and the associated
codes is that they do not check if n > 0. This type of condition on an
input parameter, which is is specified in the statement of a problem or
definition of a function, is known as a precondition. Programmers can
assume that the preconditions always hold, and therefore do not have to
develop code in order to detect or handle them.
Listing 1.2 Alternative Python code for adding the first n natural num-
bers.
1 def sum_first_naturals_2(n):
2 if n == 1:
3 return 1
4 elif n == 2:
5 return 3
6 elif n % 2:
7 return (3 * sum_first_naturals_2((n - 1) / 2)
8 + sum_first_naturals_2((n + 1) / 2))
9 else:
10 return (3 * sum_first_naturals_2(n / 2)
11 + sum_first_naturals_2(n / 2 - 1))
Listing 1.2 shows the recursive code associated with (1.6). The func-
tion uses a cascaded if statement in order to differentiate between the
two base cases (lines 2–5) and the two recursive cases (lines 6–11), which
each make two recursive calls to the defined Python function.
It is also straightforward to code a function that computes the n-th
Fibonacci number by relying on the standard definition in (1.2). List-
Listing 1.3 Python code for computing the n-th Fibonacci number.
1 def fibonacci(n):
2 if n == 1 or n == 2:
3 return 1
4 else:
5 return fibonacci(n - 1) + fibonacci(n - 2)
ing 1.3 shows the corresponding code, where both base cases are consid-
ered in the Boolean expression of the if statement.
Listing 1.4 Alternative Python code for computing the n-th Fibonacci
number.
1 def fibonacci_alt(n):
2 if n == 1 or n == 2:
3 return 1
4 else:
5 aux = 1
6 for i in range(1, n - 1):
7 aux += fibonacci_alt(i)
8 return aux
Implementing the Fibonacci function defined in (1.7) requires more

work. While the base cases are identical, the summation in the recursive
case entails using a loop or another function in order to compute and
add the values F (1), F (2),. . . , F (n − 2). Listing 1.4 shows a possible
solution that uses a for loop. The result of the summation can be stored
in an auxiliary accumulator variable aux that can be initialized to 1 (line
5), taking into account the extra unit term in the recursive case. The
for loop simply adds the terms F (i), for i = 1, . . . , n − 2, to the auxiliary
variable (lines 6–7). Finally, the function returns the computed Fibonacci
number stored in aux.
When data types are more complex there may be more variations
among the codes of different programming languages due to low-level
details. For instance, when working with a data structure similar to a
list it is usually necessary to access its length, since recursive algorithms
will use that value in order to define sublists. Figure 1.8 illustrates three
combinations of lists (or similar data structures) together with param-
eters that are necessary in order to use sublists in recursive programs.
In (a) the data structure (denoted as a) allows us to recover its length
a 5 -1 3 2 4 -3 length 6
(a)
a 5 -1 3 2 4 -3 size 6
(b)
a 7 0 5 -1 3 2 4 -3 9 2 ∗ ⋯ ∗
lower 2 upper 7
(c)
Data structures similar to lists, and parameters necessary for

Figure 1.8
defining sublists.
without using any extra variables or parameters. For instance, it could

be a property of the list, or it could be recovered through some method.
In Python, the length of a list can be obtained through the function len.
However, it is not possible to access the length of a standard array in
other programming languages such as C or Pascal. If the length of the
list cannot be recovered directly from the data structure an additional
parameter, denoted as size (which contains the length of the list) is
needed in order to store and access it, as shown in (b). This approach
can be used when working with arrays of a fixed size or partially filled ar-
rays. In these cases a more efficient alternative consists of using starting
and finishing indices for determining the limits of a sublist, as shown in
(c). Note that in this scenario the fixed size of the data structure may be
a large constant (enough for the requirements of the application), but the
true length of the lists and sublists may be much smaller. The graphic
in (c) shows a list where only the first 10 elements are relevant (the rest
should therefore be ignored). Furthermore, a sublist of six elements is
defined through the lower and upper index variables, which delimit its
boundaries. Note that the elements at these indices are included in the
sublist.
Listing 1.5Recursive functions for adding the elements in a list a, where

the only input to the recursive function is the list.
1 # Decomposition: s(a) => s(a[0:n-1]), a[n-1]
2 def sum_list_length_1(a):
3 if len(a) == 0:
4 return 0
5 else:
6 return (sum_list_length_1(a[0:len(a) - 1])
7 + a[len(a) - 1])
8
9
10 # Decomposition: s(a) => a[0], s(a[1:n])
12 if len(a) == 0:
13 return 0
14 else:
15 return a[0] + sum_list_length_2(a[1:len(a)])
16
17
18 # Decomposition: s(a) => s(a[0:n//2]) , s(a[n//2:n])
20 if len(a) == 0:
21 return 0
22 elif len(a) == 1:
23 return a[0]
24 else:
25 middle = len(a) // 2
26 return (sum_list_length_3(a[0:middle ])
27 + sum_list_length_3(a[middle :len(a)]))
28
29
30 # Some list:
31 a = [5, -1, 3, 2, 4, -3]
32
33 # Function calls:
34 print (sum_list_length_1(a))
The constructs and syntax of Python allow us to focus on algorithms

at a high level, avoiding the need to understand low-level mechanisms
such as parameter passing. Nevertheless, its flexibility permits a wide va-
riety of coding possibilities. Listing 1.5 shows three solutions to the prob-
lem of adding the elements in a list, corresponding to the three decompo-
sition strategies in Figure 1.6, where the only input parameter is the list,
which is the scenario in Figure 1.8(a). Functions sum_list_length_1,

sum_list_length_2, and sum_list_length_3 implement the recursive
definitions in (1.9), (1.10), and (1.11), respectively. The last lines of the
example declare a list v and print the sum of its elements using the
three functions. Note that the number of elements in the list n is re-
covered through len. Finally, recall that a[lower,upper] is the sublist
of a from index lower to upper−1, while a[lower:] is equivalent to
a[lower:len(a)]. If the size of the list cannot be obtained from the
list directly it can be passed to the functions through an additional pa-
rameter, as shown in Figure 1.8(b). The corresponding code is similar
to Listing 1.5 and is proposed as an exercise at the end of the chapter.
Alternatively, Listing 1.6 shows the corresponding functions when
using large lists and two parameters (lower and upper) in order to
determine sublists within it, as illustrated in Figure 1.8(c). Observe the
analogy between these functions and the ones in Listing 1.5. In this case,
empty lists occur when lower is greater than upper. Also, the sublist
contains a single element when both indices are equal (recall that both
parameters indicate positions of elements that belong to the sublist).
1.4 INDUCTION
Induction is another concept that plays a fundamental role when design-
ing recursive code. The term has different meanings depending on the
field and topic where it is used. In the context of recursive programming
it is related to mathematical proofs by induction. The key idea is that
programmers must assume that the recursive code they are trying to
implement already works for simpler and smaller problems, even if they
have not yet written a line of code! This notion is also referred to as the
recursive “leap of faith.” This section reviews these crucial concepts.
1.4.1 Mathematical proofs by induction

In mathematics, proofs by induction constitute an important tool for
showing that some statement is true. The simplest proofs involve for-
mulas that depend on some positive (or nonnegative) integer n. In these
cases, the proofs verify that the formulas are indeed correct for every
possible value of n. The approach involves two steps:
a) Base case (basis). Verify that the formula is valid for the smallest
value of n, say n0 .
Listing 1.6Alternative recursive functions for adding the elements in a

sublist of a list a. The boundaries of the sublist are specified by two
input parameters that mark lower and upper indices in the list.
1 # Decomposition: s(a) => s(a[0:n-1]), a[n-1]
2 def sum_list_limits_1(a, lower, upper ):
3 if lower > upper :
4 return 0
5 else:
6 return a[upper] + sum_list_limits_1(a, lower , upper - 1)
7
8
9 # Decomposition: s(a) => a[0], s(a[1:n])
12 return 0
13 else:
14 return a[lower] + sum_list_limits_2(a, lower + 1, upper)
15
16
17 # Decomposition: s(a) => s(a[0:n//2]) , s(a[n//2:n])
20 return 0
21 elif lower == upper:
22 return a[lower] # or a[upper ]
23 else:
24 middle = (upper + lower ) // 2
25 return (sum_list_limits_3(a, lower , middle )
26 + sum_list_limits_3(a, middle + 1, upper))
27
28
29 # Some list:
30 a = [5, -1, 3, 2, 4, -3]
31
33 print (sum_list_limits_1(a, 0, len(a) - 1))
b) Inductive step. Firstly, assume the formula is true for some general
value of n. This assumption is referred to as the induction hy-
pothesis. Subsequently, by relying on the assumption, show that
if the formula holds for some value n, then it will also be true for
n + 1.
If it is possible to prove both steps, then, by induction, it follows that

the statement holds for all n ≥ n0 . The statement would be true for n0 ,
and then for n0 + 1, n0 + 2, and so on, by applying the inductive step
repeatedly.
Again, consider the sum of the first n positive numbers (see (1.4)).
We will try to show if the following identity (which is the induction
hypothesis) involving a quadratic polynomial holds:
n
n(n + 1)
S(n) = ∑ i = . (1.12)
i=1 2
The base case is trivially true, since S(1) = 1(2)/2 = 1. For the induction
step we need to show whether
n+1
(n + 1)(n + 2)
S(n + 1) = ∑ i = (1.13)
i=1 2
holds, by assuming that (1.12) is true. Firstly, S(n + 1) can be expressed

as:
S(n + 1) = ∑ i + (n + 1).
n
i=1
Furthermore, assuming (1.12) holds we can substitute the summation

by the polynomial:
n(n + 1) n2 + n + 2n + 2 (n + 1)(n + 2)
induction
hypothesis
S(n + 1) +n+1=
↓
= = ,
2 2 2
showing that (1.13) is true, which completes the proof.
1.4.2 Recursive leap of faith

Recursive functions typically call themselves in order to solve smaller
subproblems. It is reasonable for beginner programmers to doubt if a re-
cursive function will really work as they code, and to question whether
it is legitimate to call the function being written within its body, since it
has not even been finished! However, not only can functions call them-
selves (in programming languages that support recursion), it is crucial
to assume that they work correctly for subproblems of smaller size. This
assumption, which plays a similar role as the induction hypothesis in
proofs by induction, is referred to as the recursive “leap of faith.” It is
one of the cornerstones of recursive thinking, but also one of the hardest
S(100) = ? S(99) = ?
Instructor Student +100 Classmate
5050 4950
(a)
S(100)=? S(99)=? S(98)=? S(2)=? S(1)=?
+100 +99 +3 +2
5050 4950 4851 3 1

(b)
Figure 1.9 Thought experiment in a classroom, where an instructor asks

a student to add the first 100 positive integers. S(n) represents the sum
of the first n positive integers.
concepts for novice programmers to wrap their heads around. Naturally,

it is not a mystical leap of faith that should be accepted without un-
derstanding it. The following thought experiment illustrates its role in
recursive thinking.
Consider you are in a large classroom and the instructor asks you
to add up the integers from 1 to 100. Assuming you are unaware of the
formula in (1.12), you would have to carry out the 99 involved additions.
However, imagine you are allowed to talk to classmates and there is a
“clever” one who claims he can add the first n positive integers, for
n ≤ 99. In that case you could take a much simpler approach. You could
simply ask him to tell you the sum of the first 99 positive integers, which
is 4950, and all you would need to do is add 100 to your classmate’s
response, obtaining 5050. Although it appears that you are cheating or
applying some trick, this strategy (illustrated in Figure 1.9(a)) is a valid
way to think (recursively) and solve the problem. However, it is necessary
to assume that the clever classmate is indeed providing a correct answer
to your question. This assumption corresponds to the recursive leap of
faith.
There are two possibilities regarding your classmate’s ability to re-
spond to your question. Perhaps he is indeed clever, and can somehow
provide the answer without any help (for this problem it is straightfor-
ward to apply (1.12)). However, this would be unlikely for complicated
problems. Instead, a second possibility is that he is not as clever as he

claims to be, and is also using your strategy (i.e., asking another “clever”
classmate) to come up with his answer. In other words, you can inter-
pret that your friend is simply a clone of yourself. In fact, in this last
scenario, represented in Figure 1.9(b), all of the students in the class
would implement your approach. Note that the process is clearly recur-
sive, where each student asks another one to solve a smaller instance
of the problem (decreasing its size by one at each step), until a final
student can simply reply 1 in the trivial base case. From there on, each
student is able to provide an answer to whoever asked him to compute
a full sum of integers by calculating a single addition. This process is
carried out sequentially until you receive the sum of the first 99 positive
integers, and you will be able to respond correctly to the instructor’s
question by adding 100. It is worth mentioning that the approach does
not halt when reaching the base case (this is one of students’ major mis-
conceptions). Note that there is a first stage where students successively
ask questions to other classmates until one of them responds the trivial
1, and a second phase where they calculate an addition and supply the
result to a classmate.
This approach can be formalized precisely by (1.5) and coded as
in Listing 1.1, where using S(n − 1) or sum_first_naturals(n-1) is
equivalent to asking your classmate to solve a subproblem. The recursive
leap of faith consists of assuming that S(n − 1) in the definition, or
sum_first_naturals(n-1) in the code, will indeed return the correct
answer.
Similarly to proofs by induction, we can reason that the entire pro-
cedure has to be correct. Firstly, your response will be correct as long
as your classmate returns the right answer. But this also applies to your
classmate, and to the student he is asking to solve a subproblem. This
reasoning can be applied repeatedly to every student until finally some
student asks another one to add up the integers from 1 to 1. Since the
last student returns 1, which is trivially correct, all of the responses from
the students must therefore be correct.
All of this implies that programmers can – and should – construct
valid recursive cases by assuming that recursive function calls, involv-
ing self-similar subproblems of smaller size, work correctly. Informally,
we must think that we can obtain their solutions “for free.” Therefore,
when breaking up a problem as illustrated in Figure 1.4, while we will
have to solve the different problems resulting from the decomposition,
we will not need to solve the self-similar subproblems, since we can as-
sume that these solutions are already available. Naturally, we will have
to process (modify, extend, combine, etc.) them somehow in order to
construct the recursive cases. Lastly, recursive algorithms are completed
by incorporating base cases that are not only correct, but that also allow
the algorithms to terminate.
1.4.3 Imperative vs. declarative programming

Programming paradigms are general strategies for developing software.
The imperative programming paradigm focuses on how programs work,
where the code explicitly describes the control flow and the instructions
that modify variable values (i.e., the program state). Iterative code fol-
lows this paradigm. In contrast, recursive code relies on the declarative
programming paradigm, which focuses on what programs should per-
form without describing the control flow explicitly, and can be consid-
ered as the opposite of imperative programming. Functional program-
ming follows this paradigm, which prevents side effects, and where the
computation is carried out by evaluating mathematical functions.
Therefore, when designing recursive code, programmers are strongly
encouraged to think of what functions accomplish, instead of how they
carry out a task. Note that this is related to functional abstraction and
the use of software libraries, where programmers do not need to worry
about implementation details, and can focus exclusively on functional-
ity by considering methods as black boxes that work properly. In this
regard, programmers can – and should – consider calling recursive func-
tions from a high-level point of view as invoking black boxes that return
correct results (as in Figure 1.9(a)), instead of thinking about all of the
recursive steps taken until reaching a base case (as in Figure 1.9(b)).
More often than not, focusing on lower-level details involving the se-
quence of function calls will be confusing.
1.5 RECURSION VS. ITERATION

The computational power of computers is mainly due to their ability
to carry out tasks repeatedly, either through iteration or recursion. The
former uses constructs such as while or for loops to implement repeti-
tions, while recursive functions invoke themselves successively, carrying
out tasks repeatedly in each call, until reaching a base case. Iteration
and recursion are equivalent in the sense that they can solve the same
kinds of problems. Every iterative program can be converted into a re-
cursive one, and vice versa. Choosing which one to use may depend on
several factors, such as the computational problem to solve, efficiency,
the language, or the programming paradigm. For instance, iteration is
preferred in imperative languages, while recursion is used extensively in
functional programming, which follows the declarative paradigm.
The examples shown so far in the book can be coded easily through
loops. Thus, the benefit of using recursion may not be clear yet. In
practice, the main advantage of using recursive algorithms over iterative
ones is that for many computational problems they are much simpler
to design and comprehend. A recursive algorithm could resemble more
closely the logical approach we would take to solve a problem. Thus, it
would be more intuitive, elegant, concise, and easier to understand.
In addition, recursive algorithms use the program stack implicitly
to store information, where the operations carried out on it (e.g., push
and pop) are transparent to the programmer. Therefore, they constitute
clear alternatives to iterative algorithms where it is the programmer’s
responsibility to explicitly manage a stack (or similar) data structure.
For instance, when the structure of the problem or the data resembles a
tree, recursive algorithms may be easier to code and comprehend than
iterative versions, since the latter may need to implement breadth- or
depth-first searches, which use queues and stacks, respectively.
In contrast, recursive algorithms are generally not as efficient as iter-
ative versions, and use more memory. These drawbacks are related to the
use of the program stack. In general, every call to a function, whether
it is recursive or not, allocates memory on the program stack and stores
information on it, which entails a higher computational overhead. Thus,
a recursive program cannot only be slower than an iterative version, a
large number of calls could cause a stack overflow runtime error. Further-
more, some recursive definitions may be considerably slow. For instance,
the Fibonacci codes in Listings 1.3 and 1.4 run in exponential time, while
Fibonacci numbers can be computed in (much faster) logarithmic time.
Lastly, recursive algorithms are harder to debug (i.e., analyze a program
step by step in order to detect errors and fix them), especially if the
functions invoke themselves several times, as in Listings 1.3 and 1.4.
Finally, while in some functional programming languages loops are
not allowed, many other languages support both iteration and recursion.
Thus, it is possible to combine both programming styles in order to build
algorithms that are not only powerful, but also clear to understand (e.g.,
backtracking). Listing 1.4 shows a simple example of a recursive function
that contains a loop.
1.6 TYPES OF RECURSION

Recursive algorithms can be categorized according to several criteria.
This last section briefly describes the types of recursive functions and
procedures that we will use and analyze throughout the book. Each type
will be illustrated through recursive functions that, when invoked with
certain specific arguments, can be used for computing Fibonacci numbers
(F (n)). Lastly, recursive algorithms may belong to several categories.
1.6.1 Linear recursion

Linear recursion occurs when methods call themselves only once. There
are two types of linear-recursive methods, but we will use the term “lin-
ear recursion” when referring to methods that process the result of the
recursive call somehow before producing or returning its own output. For
example, the factorial function in (1.3) belongs to this category since it
only carries out a single recursive call, and the output of the subproblem
is multiplied by n in order to generate the function’s result. The func-
tions in (1.5), (1.9), and (1.10) are also clear examples of linear recursion.
The following function provides another example:
⎧
⎪
⎪
⎪1 if n = 1 or n = 2,
f (n) = ⎨
⎪⌊Φ ⋅ f (n − 1) + ⌋ if n > 2,
1 (1.14)
⎪
⎪
⎩ 2
√
where Φ = (1 + 5)/2 ≈ 1.618 is a constant also known as the “golden
ratio,” and where ⌊⋅⌋ denotes the floor function. In this case, F (n) =
f (n) is the n-th Fibonacci number. Chapter 4 covers this type of linear-
recursive algorithms in depth.
1.6.2 Tail recursion

A second type of linear recursion is called “tail recursion.” Methods that
fall into this category also call themselves once, but the recursive call is
the last operation carried out in the recursive case. Therefore, they do
not manipulate the result of the recursive call. For example, consider the
following tail-recursive function:
⎧
⎪
⎪b if n = 1,
f (n, a, b) = ⎨
⎪
(1.15)
⎩f (n − 1, a + b, a) if n > 1.
⎪
Observe that in the recursive case the method simply returns the re-
sult of a recursive call, which does not appear within a mathematical
or logical expression. Therefore, the recursive case is simply specifying

relationships between sets of arguments for which the function returns
the same value. As the algorithm carries out recursive calls it modifies
the arguments until it is possible to compute the solution easily in a
base case. In this example Fibonacci numbers can be recovered through
F (n) = f (n, 1, 1). This type of recursive algorithm will be covered in
Chapter 5.
1.6.3 Multiple recursion

Multiple recursion occurs when a method calls itself several times in some
recursive case (see Chapter 7). If the method invokes itself twice it is also
called “binary recursion.” Examples seen so far include the functions in
(1.2), (1.6), (1.7), and (1.11). The following function uses multiple recur-
sion in order to provide an alternative definition of Fibonacci numbers
(F (n) = f (n)):
⎧
⎪
⎪
⎪
1 if n = 1 or n = 2,
⎪
⎪
⎪
f (n) = ⎨[f ( n2 + 1)] − [f ( n2 − 1)]
2 2
⎪
if n > 2 and n even, (1.16)
⎪
⎪
⎪
⎪
⎪
⎩[f ( 2 )] + [f ( 2 )]
n+1 2 n−1 2
if n > 2 and n odd.
This type of recursion appears in algorithms based on the “divide and

conquer” design strategy, covered in Chapter 6.
1.6.4 Mutual recursion

A set of methods are said to be mutually recursive when they can call
each other in a cyclical order. For example, a function f may call a
second function g, which in turn can call a third function h, which can
end up calling the initial function f . This type of recursion is also called
“indirect” since a method might not invoke itself directly within its body,
but through a cyclical chain of calls. For example, consider the two
following functions:
⎧
⎪
⎪0 if n = 1,
A(n) = ⎨
⎪
(1.17)
⎩A(n − 1) + B(n − 1) if n > 1,
⎪
⎧
⎪
⎪1 if n = 1,
B(n) = ⎨
⎪
(1.18)
⎩A(n − 1) if n > 1.
⎪
It is clear that A is recursive since it calls itself, but the recursion in B is

achieved indirectly by calling A, which in turn calls B. Thus, recursion
arises since invoking B can produce other calls to B (on simpler instances
of the problem). These functions can also be used to generate Fibonacci
numbers. In particular, F (n) = B(n) + A(n). Mutual recursion will be
covered in Chapter 9.
1.6.5 Nested recursion

Nested recursion occurs when an argument of a recursive function is
defined through another recursive call. Consider the following function:
⎧
⎪
⎪1 + s if n = 1 or n = 2,
f (n, s) = ⎨
⎪
(1.19)
⎩f (n − 1, s + f (n − 2, 0)) if n > 2.
⎪
The second parameter of the recursive call is an expression that involves
yet another recursive function call. In this case, Fibonacci numbers can
be recovered through F (n) = f (n, 0). This type of recursion is rare, but
appears in some problems and contexts related to tail recursion. Finally,
an overview of nested recursion is covered in Chapter 11.
1.7 EXERCISES
Exercise 1.1 — What does the following function calculate?
⎧
⎪
⎪1 if n = 0,
f (n) = ⎨
⎪
⎩f (n − 1) × n if n > 0.
⎪
Exercise 1.2 — Consider a sequence defined by the following recursive

definition: sn = sn−1 + 3. Calculate its first four terms, by considering
that: (a) s0 = 0, and (b) s0 = 4. Provide a nonrecursive definition of sn
in both cases.
Exercise 1.3 — Consider a sequence defined by the following recursive

definition: sn = sn−1 + sn−2 . If we are given the initial values of s1 and s2
we have enough information to build the entire sequence. Show that it is
also possible to construct the sequence when given two arbitrary values
si and sj , where i < j. Finally, find the elements of the sequence between
s1 = 1 and s5 = 17.
Exercise 1.4 — The set “descendants of a person” can be defined re-

cursively as the children of that person, together with the descendants
of those children. Provide a mathematical description of the concept

using set notation. In particular, define a function D(p), where D de-
notes descendants, and the argument p refers to a particular person.
Also, consider that you can use the function C(p), which returns the set
containing the children of a person p.
Exercise 1.5 — Let F (n) = F (n − 1) + F (n − 2) represent a recursive

function, where n is a positive integer, and with arbitrary initial values
for F (1) and F (2). Show that it can be written as F (n) = F (2) +
∑i=1 F (i), for n ≥ 2.
n−2
Exercise 1.6 — Implement the factorial function, defined in (1.3).
Exercise 1.7 — When viewed as an abstract data type, a list can be

either empty, or it can consist of a single element, denoted as the “head”
of the list, plus another list (which may be empty), referred to as the
“tail” of the list. Let a represent a list in Python. There are several ways
to check whether it is empty. For example, the condition a==[] returns
True if the list is empty, and False otherwise. In addition, the head of
the list is simply a[0], while the tail can be specified as a[1:]. Write
an alternative version of the function sum_list_length_2 in Listing 1.5
that avoids using len by using the previous elements.
Exercise 1.8 — Implement three functions that calculate the sum of

elements of a list of numbers using the three decompositions in Fig-
ure 1.6. The functions will receive two input parameters: a list, and its
size (i.e., length), according to the scenario in Figure1.8(b). In addition,
indicate an example calling the functions and printing their results.
Exercise 1.9 — Show by mathematical induction the following identity

related to a geometric series, where n is a positive integer, and x ≠ 1 is
some real number:
n−1
xn − 1
∑
i
x−1
x = .
i=0
Exercise 1.10 — Code the five recursive functions defined in Sec-

tion 1.6 that compute Fibonacci numbers F (n). Since some require
several parameters besides n, or do not compute a Fibonacci number
directly, implement additional “wrapper” functions that only receive the
parameter n, and call the coded functions in order to produce Fibonacci
numbers. Test that they produce correct outputs for n = 1, . . . , 10.
CHAPTER 2
Methodology for
Recursive Thinking
Science is much more than a body of knowledge. It is a way of

thinking.
— Carl Sagan
HEN thinking declaratively it is possible to design recursive al-

W gorithms for solving a wide variety of computational problems by
following a systematic approach. This chapter describes a general tem-
plate for deriving recursive solutions from a declarative perspective that
unravels the process of recursive thinking into a sequence of steps. In ad-
dition, the chapter introduces several useful diagrams that programmers
can utilize when designing recursive cases. These are beneficial since they
force us to think about problem decomposition and induction, the two
most important concepts underlying recursive (declarative) thinking.
Finally, for some tasks (e.g., debugging) or certain complex prob-
lems it may be useful to think about the lower-level details regarding
the sequence of operations carried out by recursive procedures. We will
nevertheless postpone covering these lower-level aspects until Chapter
10. Instead, this chapter focuses exclusively on higher-level declarative
thinking.
2.1 TEMPLATE FOR DESIGNING RECURSIVE ALGORITHMS

The recursive functions introduced in Chapter 1, as well as many of the
algorithms that will appear in the rest of the book, can be designed by
following the methodology described in Figure 2.1. This template em-
31
1. Determine the size of the problem.
2. Define bases cases.
3. Decompose the computational problem into self-similar

subproblems of smaller size, and possibly additional different
problems.
4. Define recursive cases by relying on induction and

diagrams.
5. Test the code.
Figure 2.1 General template for designing recursive algorithms.
phasizes declarative thinking due to its fourth step, since using induction
implies focusing on what an algorithm performs, rather than on how it
solves the problem. The following sections explain each step and mention
common pitfalls and misunderstandings.
2.2 SIZE OF THE PROBLEM

Naturally, the first step when tackling any computational problem is
to understand it. Given the problem statement, the inputs, outputs,
and relationships between them should be clearly identified in order to
proceed. Regardless of the type of algorithm we intend to design, it is
crucial to determine the size of the problem (also denoted as input size).
It can be thought of as a mathematical expression involving quantities
related to the input parameters that determine the problem’s complexity
in terms of the number of operations that algorithms need in order to
solve it. When using recursion it is especially relevant since the base and
recursive cases clearly depend on it.
In many problems the size only depends on one factor, which also
happens to be one of the input parameters. For instance, the size of cal-
culating the n-th Fibonacci number, the factorial of n, or the sum of the
first n positive integers (see Chapter 1) is precisely the input parameter
n. For these problems it is apparent that solving smaller instances (e.g.,
for n − 1 or n/2) requires less operations. In other cases the size of the
Methodology for Recursive Thinking 33
problem might not be specified explicitly by the inputs, but the pro-
gramming language will allow us to retrieve the particular size through
its syntax and constructs. For example, when working with lists, their
length often determines the size of the problem, which can be recovered
through the function len.
In other problems the size may be expressed as a function of the input
parameters. For instance, consider the problem of computing the sum of
the digits of some positive integer n. Although the input parameter is
n, the size of the problem is not n, but rather the number of digits of n,
which determines the number of operations to carry out. Formally this
quantity can be expressed as ⌊log10 (n)⌋ + 1.
The size of a problem may also depend on several input parameters.
Consider the problem of adding the elements of a sublist within a list
(see Listing 1.6), determined by “lower” and “upper” indices. Solving the
problem requires adding n−m+1 elements, where n and m are the upper
and lower indices, respectively. Thus, it needs to compute n−m additions.
The small (unit) difference between both expressions is irrelevant. It is
worth mentioning that it is not necessary to know exactly how many
operations will be carried out in order to solve a problem. Instead, it is
enough to understand when a base case is reached, and how to reduce the
size of the problem in order to decompose it into smaller subproblems.
For this last problem, the size is decreased by reducing the difference
between m and n.
Moreover, the size of a problem does not necessarily need to specify
the number of operations that algorithms must perform. Consider the
problem of adding the elements of a square n × n-dimensional matrix.
Since it contains n2 elements, the sum requires n2 − 1 additions, which is
a function of n. However, in this case the size of the problem is simply n,
and not n2 . Note that smaller subproblems arise by decreasing n, not n2 .
If the matrix were n × m-dimensional, the problem size would depend on
both parameters n and m. In particular, the size of the problem would
be nm, where we could obtain subproblems of smaller size by decreasing
n, m, or both.
The size is generally a property of a problem, not of a particular
algorithm that solves it. Thus, it is not the actual number of computa-
tions carried out by a specific algorithm, since problems can be solved
by different algorithms whose runtime may vary considerably. Consider
the searching problem that consists of determining whether or not some
number appears in a sorted list of n numbers. In the worst case (when
the list does not contain the number and the result is False) it can be
solved naively in n steps, or more efficiently in ⌊log2 (n)⌋ + 1 operations

by the “binary search” algorithm. Regardless of the algorithm used, the
function that describes the runtime only depends on n. Thus, the size of
the problem is n.
Nevertheless, for some problems the size can be defined in several
ways. Moreover, the choice of the size affects the rest of the decisions in
the recursive design process, and ultimately leads to different algorithms
for solving the problem. For example, consider the problem of adding two
nonnegative integers a and b by only using unit increments or decrements
(see Section 4.1.2). The size of the problem can be a+b, a, b, or min(a, b),
and the final algorithm will be different depending on which problem size
we choose to work with.
2.3 BASE CASES

Base cases are instances of problems that can be solved without using
recursion; or more precisely, without carrying out recursive calls. The
most common type of base case is associated with the smallest instances
of a problem, for which the result can be determined trivially, and some-
times may not even require carrying out computations. For instance, the
value of the first and second Fibonacci numbers is simply 1, which fol-
lows directly from the definition. The “sum” of the first positive integer
is evidently 1. Similarly, the “sum” of the digits of a nonnegative number
n that contains only one digit (i.e., n < 10) is obviously that digit (n).
Observe that these functions simply return a value in the base cases, and
do not perform any addition or other mathematical operation.
Some methods may require several base cases. For example, (1.2)
defines a base case for n = 1 and another for n = 2. Since the associated
value (1) is the same, the Python function in Listing 1.3 can use a single
logical expression ((n==1) or (n==2)) in order to take into account
both cases. Other methods need to use a cascaded if structure in order
to separate each base case. For instance, the function in (1.6) defines a
base case for n = 1 and another for n = 2 since it returns different values
for both inputs. Logically, the associated code in Listing 1.2 separates
both cases.
One of the advantages of recursive programming is that the resulting
code is often concise, which helps in understanding it. In this regard, con-
sider the factorial function defined in (1.3). One could argue that if base
cases correspond to trivial instances of a problem, then the definition of
the function could include the case 1! = 1 as well. However, 1! can be
Listing 2.1 Misconceptions regarding base cases through the factorial

function.
1 def factorial(n):
2 if n == 0:
3 return 1
4 else:
5 return factorial(n - 1) * n
6
7
8 def factorial_redundant(n):
9 if n == 0 or n == 1:
10 return 1
11 else:
12 return factorial_redundant(n - 1) * n
13
14
15 def factorial_missing_base_case(n):
16 if n == 1:
17 return 1
18 else:
19 return factorial_missing_base_case(n - 1) * n
20
21
22 def factorial_no_base_case(n):
23 return factorial_no_base_case(n - 1) * n
obtained through the recursive case correctly since 1! = 0! × 1 = 1 × 1 = 1.

Therefore, although 1! is a valid base case (it is obviously correct), in-
cluding it would be redundant, and therefore unnecessary. Moreover,
these dispensable cases could be misleading, since programmers trying
to understand the recursive code could interpret that the functions re-
quire them in order to be correct. Thus, for clarity and conciseness, it is
often recommended to include the least number of base cases that lead
to correct recursive definitions. Lastly, redundant base cases usually do
not have a relevant impact on the efficiency of a recursive program. In
general, although they can reduce the number of recursive function calls
for a given input, the saved computational time is often negligible. For
instance, if 1! is included in the factorial function as a base case the re-
cursive process will only perform one recursive call less, which barely has
any effect on computing time (moreover, the code would need two con-
ditions, and could even be more inefficient if both are evaluated before
reaching a recursive case).
Listing 2.1 uses the factorial function to illustrate this and other
basic pitfalls regarding bases cases. Firstly, the factorial method
provides a perfect implementation of the mathematical function. The
factorial_redundant method is correct since it produces a valid out-
put for any nonnegative integer input argument. However, it contains an
extra base case that is unnecessary.
In addition, we should also strive for generality, which is an im-
portant desirable software feature. More general functions are those
that can operate correctly on a wider range of inputs, and are there-
fore more applicable and useful. The factorial function in (1.3) re-
ceives a nonnegative integer as its input argument, and is defined
adequately since it can be applied correctly not only to every posi-
tive integer, but also to 0. Replacing the base case 0! = 1 by 1! =
1 would lead to a less general function, since it would not be de-
fined for n = 0. The method factorial_missing_base_case imple-
ments this function. If it were called with n = 0 as its input argu-
ment it would fail to produce a valid result since it would run into
an infinite recursion. In particular, factorial_missing_base_case(0)
would call factorial_missing_base_case(-1), which would call
factorial_missing_base_case(-2), and so on. In practice the pro-
cess textcolorredwould halt producing a runtime error message (e.g.,
“stack overflow,” or “maximum recursion depth exceeded”) after per-
forming a large number of function calls (see Chapter 10). Finally,
factorial_no_base_case always generates an infinite recursion since
it does not contain any base case, and can therefore never halt.
For many problems, there are instances for which we can provide a
result without using recursion, even if their size is not small. For example,
consider the problem of determining whether some element appears in
a list. The size of the problem is the number of elements in the list.
A first base case occurs when the list is empty (this is the smallest
instance of the problem, of zero size), where the result is obviously False.
In addition, the algorithm can check if the element appears in some
position (e.g., the first, middle, or last) of the list, and return a true
value immediately if it does, even if the list is very large (Chapter 5
covers this type of searching problems). In both cases there is no need
to carry out a recursive call.
2.4 PROBLEM DECOMPOSITION

The next step in the recursive design methodology consists of identifying
self-similar problems smaller than the original (and possibly other dif-
ferent problems), as illustrated in Figure 1.4. These smaller subproblems
will be used in the fourth step of the template in order to establish the
recursive cases. The process involves decreasing the size of the problem
in order to consider smaller instances that are closer to the base cases.
Therefore, it is essential to correctly establish the size of the problem
and its base cases before proceeding with this step.
Many recursive solutions reduce the size of a problem by decreasing
it a single unit, or by diving it by two, as shown graphically in Figures 1.5
and 1.6. Representing problems through diagrams is useful and highly
recommended since it may help us to recognize self-similar subproblems
(i.e., recursion) visually, and can also facilitate designing recursive cases.
Figure 2.2 shows additional diagrams that illustrate the decomposition of
the sum of the first positive integers when the problem size is decreased
by a unit. Firstly, since the recursive method is a mathematical function
that can be written as a formula, we can simply expand the expression in
order to identify the original problem and the corresponding subproblem,
as shown in (a). In (b) the diagram also contains a formula, but uses
boxes that enclose the formulas instead of using curly braces. In this
case, the darker outer box represents a full problem, while the lighter
box contained in it symbolizes a subproblem. In (c) the diagram uses
a similar scheme, where a problem consists of adding the small circles
arranged in a triangular pattern. This illustration is analogous to that of
Figure 1.5(b). Finally, in (d) the numbers to be added are depicted inside
rectangles, which are concatenated in order to represent the problem and
subproblem.
These apparently simple diagrams are appropriate since they allow
us to identify the subproblem within the original one. However, it is cru-
cial to observe that the subproblem is not decomposed into yet smaller
ones. For instance, the sum of the positive integers up to n − 2 does not
appear explicitly as a subproblem in any of the diagrams. In other words,
the diagrams avoid illustrating subproblems in a nested manner, or in
a tree structure. This is one of the main pitfalls and sources of misun-
derstandings when thinking recursively. Figures 2.3(a) and (c) illustrate
a proper decomposition of the sum of the first positive integers (S(n)).
Instead, while the nested diagram in (b), or the tree representing the
sequence of recursive calls in (d), represent the problem correctly, they
Subproblem
³¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ·¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹µ
1 + 2 + 3 + ⋯ + (n − 2) + (n − 1) + n
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
Original problem
(a)
1 + 2 + 3 + ⋯ + (n − 2) + (n − 1) + n
Subproblem
Original problem
(b)
1
Subproblem
2
3 Original
problem
⋮ ⋮
n−2 ⋯
n−1 ⋯
n ⋯
(c)
1 2 3 ⋯ n−2 n−1 Subproblem
1 2 3 ⋯ n−2 n−1 n Original problem

(d)
Additional diagrams that illustrate the decomposition of the

Figure 2.2
sum of the first positive integers when the problem size is decreased by
a unit.
S(n − 1) S(n) ✔
(a)
S(1) S(2) S(3) ⋯ S(n − 1) S(n) ✘
(b)
S(n)
S(n − 1)
S(n) ⋮
✔ ✘
S(3)
S(n − 1)
(c)
S(2)
S(1)
(d)
Figure 2.3When thinking recursively we generally do not need to decom-

pose a problem into every instance of smaller size.
incorporate additional subproblems (S(1), S(2),. . . , S(n − 2)) that do

not appear in the recursive definition in (1.5). Therefore, these subprob-
lems should be omitted from the diagrams, and therefore also from the
recursive thinking process, in order to avoid confusion. The nested and
tree-structured diagrams will nevertheless be useful for other tasks (see
Chapter 10).
Avoiding nested diagrams does not mean that we should only con-
sider one subproblem. Many recursive definitions require decomposing
F (n − 1) + F (n − 2)
Original problem: F (n)
(a)
1 + F (n − 2) + ⋯ + F (3) + F (2) + F (1)
Original problem: F (n)
(b)
Figure 2.4 Decompositions of the Fibonacci function.
the original problem into numerous subproblems. For instance, Fig-

ure 2.4 shows two diagrams for decomposing a Fibonacci number. In
particular, (a) is associated with the definition in (1.2), while (b) is re-
lated to (1.7). Although the diagram in (b) shows every subproblem
from F (1) until F (n − 2), it is a valid representation since the boxes
that represent subproblems are not nested.
For most problems in this book the size will depend on only one
parameter or factor. However, in some cases it is determined by several.
For instance, problems on matrices usually depend on their height (num-
ber of rows) and width (number of columns). Problems of smaller size
can therefore arise from decreasing one or both of these parameters. In
particular, many recursive algorithms partition matrices in blocks that
correspond to submatrices. These “block matrices” can be interpreted
as the original matrix with vertical and horizontal subdivisions, as illus-
trated in the following example with a 4 × 7 matrix:
⎡ ⎤
⎢ 3 1 2 0 3 4 9 ⎥
⎢ ⎥
⎢ ⎥
] = ⎢ ⎥,
7 2 3 3 1 7 5
A = [
A1,1 A1,2
⎢ ⎥
A2,1 A2,2 ⎢ 0 5 3 2 1 4 8 ⎥
⎢ ⎥
⎣ 6 3 1 5 4 9 0 ⎦
A1,1 = [ ], A1,2 = [ ],
3 1 2 0 3 4 9
7 2 3 3 1 7 5
A2,1 = [ ], A2,2 = [ ].
0 5 3 2 1 4 8
6 3 1 5 4 9 0
In this case, the number of rows and columns of each submatrix results
from performing the integer division of the original height and width of
the matrix by two.
Regarding efficiency, dividing the size of the problem by two, instead
of decreasing it by a single unit, may lead to faster algorithms. There-
fore, this strategy should be considered in general, and especially if a
theoretical analysis shows that it is possible to improve the computa-
tional cost of an algorithm based on reducing the size of the problem by
a unit. However, dividing the size of a problem by two may not lead to
a reasonable recursive algorithm. For instance, it is difficult to obtain a
simple recursive formula for the factorial function n! by considering the
subproblem (n/2)!.
2.5 RECURSIVE CASES, INDUCTION, AND DIAGRAMS

The next step in the template for designing recursive algorithms consists
of defining the recursive cases, which involves figuring out how to build
the full solution to an original problem by using the solutions to the self-
similar subproblems provided by the decomposition stage. As mentioned
in Section 1.4, we can assume that these simpler solutions are readily
available by relying on induction. Thus, the real challenge when thinking
recursively consists of determining how to modify, extend, or combine
the solutions to subproblems in order to obtain the complete solution to
an original problem, which corresponds to the last step in the scheme
presented in Figure 1.4.
2.5.1 Thinking recursively through diagrams

It is often useful to represent the thought process or “mental model” of
deriving recursive cases by means of diagrams. The general illustration
in Figure 2.5 shows a recursive thinking procedure when a problem is
decomposed into a single self-similar subproblem. The process starts by
considering a general instance of the problem, of a certain size, deter-
mined by the input parameters, which are depicted in the top-left box.
Applying the recursive method to those parameters would produce the
results represented in the top-right box. Note that this top row is sim-
ply another way to specify the statement of the problem. In order to
define recursive cases we first choose a particular decomposition that
leads to a smaller instance of the problem. The new, simpler, parame-
PSfrag
Inputs Results
Recursive
method
Problem Input parameters Solution
Derive
Decomposition recursive
cases
Recursive
method
Subproblem Simpler inputs Simpler solution
Induction
Figure 2.5General diagram for thinking about recursive cases (when a

problem is decomposed into a single self-similar subproblem).
ters for the self-similar subproblem are shown in the bottom-left box.
A recursive call to the method with those parameters would therefore
produce the results in the bottom-right box, which are obtained by ap-
plying the statement of the problem to the simpler inputs. Finally, since
we can assume that these results are available by relying on induction,
we derive recursive cases by determining how we can modify or extend
them in order to obtain or arrive at the solution to the original problem
(top-right box).
For the sum of the first n positive integers (S(n)) the diagram would
be:
Inputs Results
S
n ÐÐÐÐÐÐÐÐÐÐ→ 1 + 2 + ⋯ + (n − 1) + n = S(n)
+ n
S
n−1 ÐÐÐÐÐÐÐÐÐÐ→ 1 + 2 + ⋯ + (n − 1) = S(n − 1)
Firstly, we can choose to decompose the problem by reducing its size

by a unit. For this particular decomposition our goal is to figure out
how we can obtain S(n) by modifying or extending the solution to the
subproblem S(n − 1). In this case, it is easy to see that adding n to
S(n − 1) yields S(n). Therefore, the recursive case is defined as S(n) =
S(n − 1) + n.
It is important to understand that we are only required to define

the decomposition, and establish how to obtain the results to the orig-
inal problem given the results of the subproblem. Therefore, we only
have to think about the processes associated with the thicker arrows.
Alternatively, the thinner arrows simply symbolize the effect of solv-
ing particular instances of the problem (on different inputs), which is
completely defined by the statement of the problem. Observe that we
do not have to make any algorithmic design decisions regarding the re-
lationships between inputs and results, since they are specified in the
statement of the problem.
The advantage of the diagram in Figure 2.5 is its generality, since
it can be applied to numerous problems for deriving recursive solutions.
Therefore, the explanations to many of the examples covered throughout
the book will rely on it. Naturally, other graphical representations, such
as the diagrams in Figures 1.5, 1.6, and 2.2 can also be used instead.
These alternative visualizations are especially useful since they not only
show subproblems, but also their relationship with the original problem,
which provides clues about how we can combine, increment, or modify
the solutions to the subproblems in order to solve the original one. In
other words, they visually suggest how we can define the recursive cases.
For instance, it is apparent from the diagrams in Figure 2.2 that the so-
lution to the problem consists of adding n to the result of the subproblem
of size n − 1.
Indeed, more elaborate diagrams, especially tailored for the partic-
ular problem, may help us to find different decompositions that could
eventually lead to more efficient algorithms. For instance, representing
the sum of the first n positive integers through triangular pyramids of
tiled blocks, as shown in Figure 1.5, allows us to derive recursive cases
easily when the size of the subproblems is roughly half of the original
one. Firstly, recall that the problem can be interpreted as counting the
number of blocks in a triangular structure. Furthermore, since the rep-
resentation allows us to superimpose several subproblems (i.e., smaller
triangular structures) simultaneously on top of the original, the visual-
izations indicate how it is possible to “fill” a larger pyramid with smaller
ones. Therefore, the diagrams illustrate how to derive recursive cases by
adding the results (in this case, the number of blocks) of smaller sub-
problems. For example, in Figure 1.5(c) the output of three subproblems
of size n/2 can be added to the result of the subproblem of size n/2 − 1
in order to obtain the total value of the original problem.
Nonetheless, the general diagram in Figure 2.5 can also be used in

order to provide recursive solutions for S(n) when the decomposition
divides the size of the problem by two. In that case we have:
Inputs Results
S
n ÐÐÐÐÐÐÐÐÐÐ→ 1 + 2 + ⋯ + ( n2 ) + ( n2 + 1) + ⋯ + n = S(n)
S
n
2
ÐÐÐÐÐÐÐÐÐÐ→ 1 + 2 + ⋯ + ( n2 ) = S( n2 )
The question is to figure out how (and if) we can obtain S(n) by
modifying or extending S(n/2). A first obvious idea leads to S(n) =
S(n/2) + (n/2 + 1) + ⋯ + n. However, although it is correct, its imple-
mentation requires either a loop, or using a different recursive function,
in order to calculate the sum of the last n/2 terms. Observe that it is
not possible to obtain (n/2 + 1) + ⋯ + n by using a recursive call to the
function under construction (S), since that sum is not an instance of the
problem. Nevertheless, it can be broken up, for example, in the following
way:
(n/2 + 1) + (n/2 + 2) + ⋯ + n
n/2 + n/2 + ⋯ + n/2 = (n/2)2
+ 1 + 2 + ⋯ + n/2 = S(n/2)
This not only simplifies the expression, but it also contains S(n/2), which
we can use in order to obtain a much simpler recursive case:
S(n) = S(n/2) + (n/2 + 1) + ⋯ + n

= S(n/2) + S(n/2) + (n/2)2
= 2S(n/2) + (n/2)2 . (2.1)
The general diagram could be completed in the following way:

Original problem: S(n) S(n) = 2S( n2 ) + ( n2 )2
n
2
n
2
(a) (b)
Figure 2.6 Diagram showing a decomposition of the sum of the first n

positive integers S(n) that uses two subproblems of half the size as the
original.
Inputs Results
S
n ÐÐÐÐÐÐÐÐÐÐ→ 1+2+⋯+ ( n2 ) + ( n2 + 1) + ⋯ + n = S(n)
× 2 + ( n2 )2
S
n
2
ÐÐÐÐÐÐÐÐÐÐ→ 1 + 2 + ⋯ + ( n2 ) = S( n2 )
indicating that S(n) can be obtained by multiplying the result of S(n/2)

times 2, and adding (n/2)2 . Although the decomposition of the problem
is unique (S(n/2)), this example shows that we can use the result of
the subproblem several times in order to arrive at the solution of the
original problem. In this case, we can interpret that the recursive case
uses S(n/2) twice.
This recursive case can also be obtained through the diagrams that
depict S(n) as a triangular pyramid of blocks. In particular, Figure 2.6
illustrates the same reasoning that leads to S(n) = 2S(n/2) + (n/2)2
through these representations. The nature of this particular diagram,
where the solution to the problem consists of adding up blocks, facilitates
obtaining recursive cases.
2.5.2 Concrete instances

In the previous example involving the sum of the first positive integers
it is straightforward to express the result of the recursive method as a
general function of the input parameter n. In other words, it is easy to

obtain the expressions on the right-hand side of the general diagram.
However, in other problems it can be more complicated to describe the
results of the recursive methods (e.g., through some formula), and there-
fore to figure out how to arrive at the solution given the results to the
subproblems. In these cases a useful initial approach consists of analyzing
the problem through concrete instances.
Consider the problem of adding the digits of some nonnegative inte-
ger n, and assume that we choose the decomposition that decreases the
number of digits by a unit, where the least significant digit is dropped
from the original integer. This information provides the elements on the
left-hand side of the general diagram. However, the method’s output on
the right-hand side can be complicated as a function of n. For example,
one possible way to describe it is through the following summation:
⌊log10 n⌋
∑ [(n//10i )%10]. (2.2)
i=0
Although we could use the formula to derive the recursive cases, we will
proceed by analyzing concrete instances of the problem. For example (we
can discard using the method’s name in the results column, and when
labeling the arrows, for the sake of simplicity):
Inputs Results
n = 5342 ÐÐÐÐÐÐÐÐÐ→ 14
+2
n//10 = 534 ÐÐÐÐÐÐÐÐÐ→ 12
or
Inputs Results
n = 879 ÐÐÐÐÐÐÐÐÐ→ 24
+9
n//10 = 87 ÐÐÐÐÐÐÐÐÐ→ 15
It is easy to see in these diagrams that we can obtain the result of the
method by adding the last (underlined) digit of the original number to
the output of the subproblem.
Listing 2.2 Code for computing the sum of the digits of a nonnegative
integer.
1 def add_digits(n):
2 if n < 10:
3 return n
4 else:
5 return add_digits(n // 10) + (n % 10)
2.5.3 Alternative notations

Another approach consists of using some alternative general notation
that facilitates constructing a diagram. Let dm−1 ⋯ d1 d0 represent the
sequence of digits that define some nonnegative integer n in base 10 (in
other words, n = dm−1 ⋅ 10m−1 + ⋯ + d1 ⋅ 10 + d0 , with 0 ≤ di ≤ 9 for all i,
and dm−1 ≠ 0 if m > 1). In that case the general diagram would be:
Inputs Results
n = dm−1 ⋯ d1 d0 ÐÐÐÐÐÐÐÐÐ→ dm−1 + ⋯ + d1 + d0
+ d0
n//10 = dm−1 ⋯ d1 ÐÐÐÐÐÐÐÐÐ→ dm−1 + ⋯ + d1
and the function can be coded as shown in Listing 2.2, where (n%10)
represents d0 . Note that the base case for this problem occurs when n
contains a single digit (i.e., n < 10), for which the result is obviously n.
Lastly, in practice more complicated problems may require analyzing

multiple examples and scenarios (e.g., distinguishing when the input
parameter is odd or even), that lead to different recursive cases.
2.5.4 Procedures
The methods seen so far correspond to functions that return values,
where the results can be defined through formulas. Therefore, they can
be used within expressions in other methods, where they would return
specific values given a set of input arguments. However, there exist meth-
ods, called “procedures” in certain programming languages (e.g., Pas-
cal), which do not return values. Instead, they may alter data structures
passed as inputs to the method, or they can simply print information on

a console. In these cases it is also possible to use the general diagram.
Consider the problem of printing the digits of some nonnegative in-
teger n on the console, in reversed order, and vertically. In other words,
the least significant digit will appear on the first line, the second least
significant digit on the second line, and so on. For instance, if n = 2743
the program will print the following lines on the console:
3¶
4¶
7¶
2¶
Firstly, the size of this problem is the number of digits of n. The base
case occurs when n contains a single digit (n < 10), where the algorithm
would simply print n. As in the previous example, the simplest decompo-
sition considers n//10, where the least significant digit is removed from
the original number. Figure 2.7 shows a possible decomposition diagram
of the problem.
In addition, the general diagram can also be used for this type of
procedure:
Inputs Results
3¶
4¶
n = 2743 ÐÐÐÐÐÐÐÐÐ→ 7¶
2¶
print(3) (before)
4¶
n//10 = 274 ÐÐÐÐÐÐÐÐÐ→ 7¶
2¶
In this example the results are no longer numerical values, but a se-
quence of instructions, which correspond to printed lines on a console.
For this problem it is possible to arrive at the solution by printing the
least significant digit of the input number, and by calling the recursive
method on the remaining digits. However, the order in which these op-
erations are performed is crucial. In particular, the least significant digit
3¶ d0 ¶
4¶ d1 ¶
7¶ ⋮
2¶ dm−1 ¶
(a) (b)
Figure 2.7 Diagram showing a decomposition of the problem consisting

of printing the digits of a nonnegative integer on the console, in reversed
order, and vertically. A particular (n = 2743) and a general m-digit
(n = dm−1 ⋯ d1 d0 ) case are shown in (a) and (b), respectively.
Code for printing the digits of a nonnegative integer vertically,

Listing 2.3
and in reversed order.
1 def print_digits_reversed_vertically(n):
2 if n < 10:
3 print (n)
4 else:
5 print (n % 10)
6 print_digits_reversed_vertically(n // 10)
must be printed before the rest of the digits. The associated code is
shown in Listing 2.3.
2.5.5 Several subproblems

Some algorithms need to decompose a problem into several self-similar
subproblems, where the thought process of deriving recursive cases is
analogous to the one illustrated in Figure 2.5. The diagrams simply
need to include various subproblems and their corresponding solutions,
according to the chosen decomposition of the problem, as shown in Fig-
ure 2.8. Besides extending and modifying the individual solutions to the
subproblems, the recursive cases usually need to combine them as well.
Given a nonempty list of n integers (n ≥ 1), consider the problem of
finding the value of the largest one. For this example, we will decom-
pose the problem by splitting it in two, in order to work with the first
and second halves of the list, as shown in Figure 1.6(d). The following
diagram illustrates the thought process with a concrete example:
PSfrag
Inputs Results
Recursive
method
Problem Input parameters Solution
Derive
Decomposition recursive
cases
Recursive
method
Subproblem #1 Simpler inputs #1 Simpler solution #1
⋮ ⋮
Recursive
method
Subproblem #N Simpler inputs #N Simpler solution #N
Induction
Figure 2.8 General diagram for thinking about recursive cases, when a
problem is decomposed into several (N ) self-similar subproblems.
5 -1 3 2 4 7 2
Figure 2.9 Alternative diagram showing a divide and conquer decompo-

sition, and the recursive thought process, for the problem of finding the
largest value in a list. The thick and thin arrows point to the solutions
of the problem and subproblems, respectively.
Inputs Results
[5, -1, 3 ,2, 4, 7, 2] ÐÐÐÐÐÐÐÐÐ→ 7
max(5, 7)
[5, -1, 3] ÐÐÐÐÐÐÐÐÐ→ 5

[2, 4, 7, 2] ÐÐÐÐÐÐÐÐÐ→ 7
Figure 2.9 shows an alternative diagram that indicates the decomposition

and recursive thought process for the problem. Both cases illustrate that
each recursive call to one of the halves returns the greatest value in
Code for computing the maximum value in a list, through a

Listing 2.4
divide and conquer approach.
1 def max_list_length_DaC(a):
2 if len(a) == 1:
3 return a[0]
4 else:
5 middle = len(a) // 2
6 m1 = max_list_length_DaC(a[0:middle ])
7 m2 = max_list_length_DaC(a[middle :len(a)])
8 return max(m1, m2)
9
10
11 def max_list_limits_DaC(a, lower , upper ):
12 if lower == upper:
13 return a[lower] # or a[upper ]
14 else:
15 middle = (upper + lower ) // 2
16 m1 = max_list_limits_DaC(a, lower , middle )
17 m2 = max_list_limits_DaC(a, middle + 1, upper )
18 return max(m1, m2)
19
20
21 # Some list:
22 v = [5, -1, 3, 2, 4, 7, 2]
23
25 print (max_list_length_DaC(v))
26 print (max_list_limits_DaC(v, 0, len(v) - 1)
the particular half. Therefore, the recursive case can simply return the
maximum value in either half. The recursive function (f ) is defined as
follows:
⎧
⎪
⎪a[0] if n = 1,
f (a) = ⎨
⎪
(2.3)
⎩max (f (a[0 ∶ n//2]), f (a[n//2 ∶ n])) if n > 1.
⎪
Listing 2.4 shows two ways of coding the function. The version that
uses the lower and upper limits is usually faster. Naturally, this problem
also allows recursive solutions based on a single subproblem whose size is
decreased by a single unit. This approach is as efficient as the divide and
conquer strategy. However, in practice it may produce runtime errors for
large lists (see Chapter 10).
Listing 2.5Erroneous Python code for determining if a nonnegative inte-

ger n is even.
1 def is_even_incorrect(n):
2 if n == 0:
3 return True
4 else:
5 return is_even_incorrect(n - 2)
Listing 2.6 Correct Python code for determining if a nonnegative integer

n is even.
1 def is_even_correct(n):
2 if n == 0:
3 return True
4 elif n == 1:
5 return False
6 else:
7 return is_even_correct(n - 2)
2.6 TESTING
Testing is a fundamental stage in any software development process. In
the context of this book its main purpose consists of discovering errors in
the code. Testing therefore consists of running the developed software on
different instances (i.e., inputs) of a problem in order to detect failures.
Novice programmers are strongly encouraged to test their code since the
ability to detect and correct errors (e.g., with a debugger) is a basic
programming skill. In addition, it teaches valuable lessons in order to
avoid pitfalls and code more efficiently and reliably.
Besides checking the basic correctness of the base and recursive cases,
when testing recursive code, programmers should pay special attention to
possible scenarios that lead to infinite recursions. These usually appear
due to missing base cases or by erroneous recursive cases. For example,
consider the function in Listing 2.5 whose goal consists of determining
whether some nonnegative integer n is even. Both base and recursive
cases are correct. Naturally, if a number n is even then so is n − 2, and
the function must return the same Boolean value for both integers. Nev-
ertheless, is_even_incorrect only works for even numbers. Let f (n)
represent is_even_incorrect(n). A call to f (7) produces the following
Erroneous Python code for adding the first n positive numbers,

Listing 2.7
which produces infinite recursions for most values of n.
2 if n == 1:
3 return 1
4 else:
5 return 2 * sum_first_naturals_3(n / 2) + (n / 2)**2
recursive calls:
f (7) → f (5) → f (3) → f (1) → f (−1) → f (−3) → ⋯
which is an infinite recursion since the process does not halt at a base
case. The fact that the function does not contain a base case that re-
turns False provides a warning regarding its correction (not all Boolean
functions need two base cases in order to return True or False). Indeed,
the method can be fixed by adding that base case. Listing 2.6 shows a
function that works for any valid argument (n ≥ 0).
Another example is the function in Listing 2.7 that uses the recursive
case described in (2.1) in order to compute the sum of the first n positive
integers (S(n)). It is incomplete, and generates infinite recursions for
values of n that are not a power of two. Firstly, since Python considers
n to be a real number, n/2 is also a real number in general. Therefore, if
the argument n is an odd integer in any recursive function call then n/2
will have a fractional part, and so will the arguments of the following
recursive calls. Thus, the algorithm would not halt at the base case
n = 1 (which is an integer with no fractional part), and would continue
to make function calls with smaller and smaller arguments. For example,
let f (n) represent sum_first_naturals_3(n). A call to f (6) produces
the following recursive calls:
f (6) → f (3) → f (1.5) → f (0.75) → f (0.375) → f (0.1875) → ⋯
never stopping at the base case. The only situation in which the algo-
rithm works is when the first argument n is a power of two, since each of
the divisions by two produces an even number, eventually reaching n = 2
and afterwards n = 1, where the function can finally return a concrete
value instead of producing another function call.
The method sum_first_naturals_3 does not work properly due
to real-valued arguments. Thus, we could try replacing the real divi-
Original problem: S(n) S(n) = 2S( n−1

2
) + ( n+1
2
)2
n−1
2
n
n+1
2
(a) (b)
Figure 2.10 Diagram showing a decomposition of the sum of the first n

positive integers S(n) that uses two subproblems of (roughly) half the
size as the original, when n is odd.
Listing 2.8 Incomplete Python code for adding the first n positive num-
bers.
2 if n == 1:
3 return 1
4 else:
5 return 2 * sum_first_naturals_4(n // 2) + (n // 2)**2
sions by integer divisions, as shown in Listing 2.8. This forces the ar-
guments to be integers, which prevents infinite recursions. Nevertheless,
the function still does not work properly for arguments that are not
powers of two. The issue with sum_first_naturals_4 is that it is in-
complete. In particular, although the recursive case is correct, it only
applies when n is even. Figure 2.10 shows how to derive the recursive
case (S(n) = 2S((n − 1)/2) + ((n + 1)/2)2 ) for the problem and decom-
position when n is odd. The complete function is therefore:
⎧
⎪
⎪
⎪
⎪
1 if n = 1,
⎪
S(n) = ⎨2S ( n2 ) + ( n2 )
2
⎪
⎪
⎪
⎪
⎪
⎩2S ( 2 ) + ( 2 )
n+1 2
n−1
if n > 1 and n is odd,
and the corresponding code is shown in Listing 2.9. With the new re-
cursive case, every argument to function sum_first_naturals_5 will
also be an integer (given an initial integer input). Finally, replacing the
Python code for adding the first n positive numbers based on

Listing 2.9
using two subproblems of (roughly) half the size as the original.
2 if n == 1:
3 return 1
4 elif n % 2 == 0:
5 return 2 * sum_first_naturals_5(n / 2) + (n / 2)**2
6 else:
7 return (2 * sum_first_naturals_5((n - 1) / 2)
8 + ((n + 1) / 2)**2)
real divisions (/) by integer divisions (//) would also lead to a correct
algorithm.
2.7 EXERCISES
Exercise 2.1 — Let n be some positive integer. Consider the problem
of determining the number of bits set to 1 in the binary representation
of n (i.e., n expressed in base 2). For example, for n = 2510 = 110012
(the subscript indicates the base in which a number is expressed), the
result is three bits set to 1. Indicate the size of the problem and provide
a mathematical expression for such size.
Exercise 2.2 — Consider the function that adds the first n positive
integers in (1.5). Define a more general function that it is also applicable
to all nonnegative integers. In other words, modify the function consid-
ering that it can also receive n = 0 as an input argument. Finally, code
the function.
Exercise 2.3 — Use similar diagrams as in Figures 2.6 and 2.10 in order
to derive recursive definitions of the sum of the first n positive integers
(S(n)), where the recursive case will add the result of four subproblems
of (roughly) half the size as the original. Finally, define and code the full
recursive function.
Exercise 2.4 — Consider the problem of printing the digits of some

nonnegative integer n on the console vertically, in “normal” order, where
the most significant digit must appear on the first line, the second most
significant digit on the second line, and so on. For instance, if n = 2743
the program must print the following lines on the console:
2¶
7¶
4¶
3¶
Indicate the size of the problem and its base case. Draw a diagram for
a general nonnegative input integer n = dm−1 ⋯ d1 d0 , where m is the
number of digits of n, in order to illustrate the decomposition of the
problem, and how to recover the solution to the problem given the result
of a subproblem. Finally, derive the recursive case and code the method.
Exercise 2.5 — Define a general diagram when using a divide and

conquer approach for the problem of calculating the largest value in a
list a of n elements. Use appropriate general notation instead of concrete
examples.
Exercise 2.6 — Define a recursive function that calculates the largest

value in a list a of n elements, where the decomposition simply reduces
the size of the problem by a unit.
CHAPTER 3
Runtime Analysis of
Recursive Algorithms
The faster you go, the shorter you are.

— Albert Einstein
lgorithm analysis is the field that studies how to theoretically

A estimate the resources that algorithms need in order to solve com-
putational problems. This chapter focuses on analyzing the runtime,
also denoted as “computational time complexity,” of recursive algorithms
that solve problems whose size depends on a single factor (which occurs
in the majority of the problems covered in the book). This will provide
a context that will enable us to characterize and compare different al-
gorithms regarding their efficiency. In particular, the chapter describes
two methods for solving recurrence relations, which are recursive mathe-
matical functions that describe the computational cost of recursive algo-
rithms. These tools are used in order to transform a recurrence relation
into an equivalent nonrecursive function that is easier to understand
and compare. In addition, the chapter provides an overview of essential
mathematical fundamentals and notation used in the analysis of algo-
rithms. Lastly, the analysis of memory (storage) or space complexity of
recursive algorithms will be covered in Chapter 10.
3.1 MATHEMATICAL PRELIMINARIES

This section presents a brief introduction to basic mathematical defini-
tions and properties that appear in the analysis of the computational
57
complexity of algorithms, and in several problems covered throughout

the book.
3.1.1 Powers and logarithms

The following list reviews essential properties of powers and logarithms:
• b1 = a • b0 = 1
• bx by = bx+y • (bx )y = bxy = (by )x
• b−x = 1/bx • (ab)x = ax bx
• logb b = 1 • logb 1 = 0
• logb (xy) = logb (x) + logb (y) • logb (x/y) = logb (x) − logb (y)
• logb (xy ) = y logb x • logb x = loga x/ loga b
• logb a = 1/ loga b • xlogb y = y logb x
• logb (bx ) = x • blogb a = a
where a, b, x, and y are arbitrary real numbers, with the exceptions that:
(1) the base of a logarithm must be positive and different than 1, (2) a
logarithm is only defined for positive numbers, and (3) the denominator
in a fraction cannot be 0. For example, logb x = loga x/ loga b is only valid
for a > 0 with a ≠ 1, b > 0 with b ≠ 1, and x > 0.
Logarithms and powers of positive numbers are monotonically in-
creasing functions. Therefore, if x ≤ y then logb x ≤ logb x, and bx ≤ by
(for valid values of x, y, and b).
3.1.2 Binomial coefficients

A binomial coefficient, denoted as (m
n
), is an integer that appears in the
polynomial expansion of the binomial power (1 + x)n . It can be defined
⎧
as:
⎪
⎪1 if m = 0 or n = m,
( ) = ⎨ n!
n
⎪ m!(n−m)! otherwise,
(3.1)
m ⎪
⎩
Runtime Analysis of Recursive Algorithms 59
where n and m are integers that satisfy n ≥ m ≥ 0. In addition, a binomial

coefficient can be defined recursively through:
⎧
⎪
⎪1 if m = 0 or n = m
( ) = ⎨ n−1
n
⎪
(3.2)
⎩(m−1) + ( m ) otherwise.
⎪
n−1
m
Binomial coefficients play an important role in combinatorics. In par-

ticular, (m
n
) determines the number of ways it is possible to choose m
elements from a set of n distinct ones, where the order of selection does
not matter.
3.1.3 Limits and L’Hopital’s rule

The computational cost of different algorithms can be compared through
limits involving a quotient of functions. Firstly,
∞
= ∞,
k
= 0, and
∞ k
where k is a constant, and where ∞ should be understood as the result
of a limit. In addition, since functions that measure computational cost
are generally increasing (with the exception of constants), the limit when
their input approaches infinity will also be infinity. For example:
lim logb n = ∞, or lim na = ∞,

n→∞ n→∞
for a valid base b for the logarithm, and for a > 0. Therefore, the following
indeterminate form appears frequently:
f (n) ∞
lim
∞
= .
n→∞ g(n)
It is usually solved by simplifying the fraction f (n)/g(n) until it is pos-

sible to obtain a result that is not an indeterminate form. A well-known
approach for simplifying the fraction is to use L’Hopital’s rule:
f (n) f ′ (n)
lim = lim ′ (3.3)
n→∞ g (n)
,
n→∞ g(n)
where f ′ (n) and g ′ (n) are the derivatives of f (n) and g(n), respectively.
Formally, L’Hopital’s rule is only valid when the limit on the right-hand
side of (3.3) exists, which is usually the case.
3.1.4 Sums and products

It is important to be familiarized with sums since they not only appear
in function definitions and formulas that we may need to code, but also
when analyzing the efficiency of both iterative and recursive algorithms.
A sum, or summation, is simply the addition of a collection of mathemat-
ical entities (integers, real numbers, vectors, matrices, functions, etc.).
When a sum involves adding some function f (i), evaluated at consecu-
tive integer values from an initial integer index m up to a final index n,
it can be expressed compactly by using “sigma notation”:
∑ f (i) = f (m) + f (m + 1) + ⋯ + f (n − 1) + f (n).

n
(3.4)
i=m
Thus, the result of the sum is simply the addition of terms that arise
from substituting every occurrence of i in function f (i) with integers
from m up to n. For instance:
4
∑ ki = k ⋅ 0 + k ⋅ 1 + k ⋅ 2 + k ⋅ 3 + k ⋅ 4 ,
2 2 2 2 2 2
i=0
where f (i) = ki2 .

It is important to understand that the result of the sum does not
depend on the index variable. For example, the sum in (3.4) logically
depends on the function f , and on the particular limits m and n, but
not on any specific value of i. For instance, notice that i does not appear
on the right-hand side of (3.4). Therefore, the integer index is simply a
“dummy” variable that allows us to indicate all of the terms that need
to be added with a single expression. In this regard, we can choose any
name for the index variable, say j or k:
∑ f (i) = ∑ f (j) = ∑ f (k).

n n n
i=m j=m k=m
Furthermore, we can express the same sum in different ways by perform-

ing a change of variable:
∑ f (i) = ∑ f (i + 1) = ∑ f (n − i + m).
n n−1 n
i=m i=m−1 i=m
In the second sum the limits (and parameter of f ) appear shifted. The
third sum simply adds the terms in “reverse” order (when i = m it adds
f (n), while when i = n the term added is f (m)). Finally, if the lower
limit m is greater than the upper limit n, then the sum evaluates to 0,
by convention.
3.1.4.1 Basic properties of sums

The following basic properties are useful for simplifying and working with
sums, and can be derived easily from the addition and multiplication
properties of the type of elements to be added. Firstly,
n
∑ 1 = 1 + 1 + ⋯ + 1 + 1 = n.
i=1 ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
n times
Notice that f (i) is a constant (1) that does not depend on the index
variable i. Similarly,
n
∑ k = k + k + ⋯ + k + k = kn.
i=1 ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
n times
The previous example also illustrates that when a multiplicative term in

f does not depend on the index variable it can be “pulled out” of the
sum:
∑ k = ∑(k ⋅ 1) = ∑ k ⋅ 1 = k ⋅ (∑ 1) = k ∑ 1 = kn.
n n n n n
(3.5)
i=1 i=1 i=1 i=1 i=1
In this case the constant term k, which we can consider is multiplied
by 1, does not depend on i, and can be pulled out of the sum, where
it appears multiplying it (regarding notation, (3.5) shows that it is not
necessary to use parenthesis inside a sum if f does not contain additive
terms, nor if some term multiplies a sum). In general, the property can
be expressed as:
∑ kf (i) = k ∑ f (i),
n n
i=m i=m
which follows from the distributive law of multiplication over addition,
where the expression has been simplified by extracting the common fac-
tor k from all of the terms being added. Naturally, the factor k can be
the product of several terms that do not depend on the index, and may
contain the upper and lower limits, as shown in the following example:
n n
∑ amn i = amn ∑ i ,
2 3 2 3
i=m i=m
where a is some constant.

Finally, sums where the function f contains several additive terms
can be broken up into simpler individual sums. Formally,
∑ (f1 (i) + f2 (i)) = ∑ f1 (i) + ∑ f2 (i).

n n n
i=m i=m i=m

3.1.4.2 Arithmetic series

An arithmetic series is the sum of the terms in a sequence si , for
i = 0, 1, 2, . . ., in which each term is equal to the previous one plus a par-
ticular constant d (which could be negative). In other words, si = si−1 +d,
for i > 0, which is a recursive definition. These terms can also be de-
scribed nonrecursively as follows:
si = si−1 + d = si−2 + 2d = ⋯ = s0 + id. (3.6)
While technically an arithmetic series is an infinite sum:

∞
∑ si ,
i=0
when analyzing the runtime of algorithms the quantities of interest only

add the terms of finite sequences:
n
∑ si ,
i=m
also referred to as partial sums.

The sum of the first n positive integers (S(n)) is used frequently
when analyzing the efficiency of recursive and iterative algorithms, and
can be understood as a partial sum (of n elements) of the arithmetic
series with s0 = 1, and d = 1:
S(n) = ∑ si = ∑ (s0 + id) = ∑ (1 + i)

n−1 n−1 n−1
i=0 i=0 i=0

n−1 n−1 n−1 n
= ∑ 1 + ∑ i = n + ∑ i = ∑ i.
i=0 i=0 i=0 i=1
As shown in Section 1.4.1, its result can be expressed as a quadratic

polynomial (see (1.12)). A simple way to derive the formula consists of
adding two sums S(n), one expressed with terms in increasing order,
and the other in decreasing order:
S(n) = 1 + 2 + ⋯ + (n − 1) + n
+
S(n) = n + (n − 1) + ⋯ + 2 + 1
2S(n) = (n + 1) + (n + 1) + ⋯ + (n + 1) + (n + 1)
S(n)
2S(n) = n(n + 1)
n
n(n+1)
⇒ S(n) = 2
S(n)
n+1
Graphical mnemonic for determining the quadratic formula for

Figure 3.1
the sum of the first n positive integers (S(n)).
It follows from the result that 2S(n) = n(n + 1), since there are n terms
(each equal to n + 1) on the right-hand side of the identity. Finally,
dividing by 2 yields:
n
n(n + 1)
S(n) = ∑ i = .
i=1 2
The idea can also be understood through the graphical mnemonic in

Figure 3.1, where S(n) is the area of a triangular pyramid of square
(1 × 1) blocks. Two of these triangular structures can be joined together
in order to form a rectangle of area n × (n + 1). Therefore, it follows that
the area (S(n)) of each triangular structure is n(n + 1)/2.
A similar formula can be obtained for the general partial sum of n
terms of an arithmetic sequence (si = si−1 + d, for some initial s0 ):
(s0 + sn−1 ),
n−1
∑ si =
n
i=0 2
which is the average between the first and last elements of the sequence,
multiplied by the number of elements in the sequence.
3.1.4.3 Geometric series

A geometric series is the sum of the terms in a sequence si , for i =
0, 1, 2, . . ., in which each term is equal to the previous one times a par-
ticular constant r. In other words, si = r ⋅ si−1 , for i > 0. These terms can
also be described nonrecursively as follows:
si = r ⋅ si−1 = r2 ⋅ si−2 = ⋯ = ri ⋅ s0 . (3.7)

The partial sum of a geometric series can be obtained through the

following formula:
n n n
rm − rn+1
∑ si = ∑ r ⋅ s0 = s0 ∑ r = s0
i i
(3.8)
1−r
,
i=m i=m i=m
for r ≠ 1. In practice the last equality can be derived easily through

the following approach. First, let S represent the value of the sum. Sub-
sequently, create another sum by multiplying S by r. Finally, subtract
both sums (where most terms cancel out), and solve for S. The process
(ignoring s0 ) can be illustrated as follows:
S = rm + rm+1 + rm+2 + . . . + rn−1 + rn

-
rS = + rm+1 + rm+2 + . . . + rn−1 + rn + rn+1
S − rS = rm − rn+1
Therefore, solving for S yields:

n
rm − rn+1 rn+1 − rm
S = ∑ ri = (3.9)
1−r r−1
= .
i=m
Finally, a geometric series converges to a constant value if ∣r∣ < 1:

∞
1
∑ ar = a
i
1−r
.
i=0
3.1.4.4 Differentiation
Another useful sum is:
n
S = ∑ iri = 1r1 + 2r2 + 3r3 + ⋯ + nrn , (3.10)
i=1
which can be interpreted as a hybrid between arithmetic and geometric

series. It is possible to derive a formula for the sum as follows. Firstly,
consider a general partial sum of a geometric series and its corresponding
simplified formula:
rn+1 − 1
T = 1 + r + r2 + . . . + rn =
r−1
.
We then differentiate with respect to r:

nrn+1 − (n + 1)rn + 1
= 1 + 2r + 3r2 + . . . + nrn−1 =
dT
(3.11)
(r − 1)2
,
dr
and multiply by r in order to obtain a formula for (3.10):

n
nrn+1 − (n + 1)rn + 1
S = ∑ iri = r + 2r2 + . . . + nrn = r (3.12)
(r − 1)2
.
i=1
We can use these formulas in order to simplify similar sums. For

example, let
N +1
S = ∑ i2i−1 = 1 ⋅ 1 + 2 ⋅ 2 + 3 ⋅ 4 + 4 ⋅ 8 + ⋯ + (N + 1) ⋅ 2N .
i=1
It is a special case of (3.11) for r = 2 and n = N + 1. Substituting, we

obtain:
(N + 1)2N +2 − (N + 2)2N +1 + 1
S= = 2N +1 (2N + 2 − N − 2) + 1
1
= N 2N +1 + 1.
3.1.4.5 Products
Similarly to the notation used for sums, a product of several terms of
some function f (i), evaluated at consecutive integer values from an ini-
tial index m up to a final index n, can be written as follows:
∏ f (i) = f (m) ⋅ f (m + 1) ⋅ ⋯ ⋅ f (n − 1) ⋅ f (n),

n
(3.13)
i=m
where by convention the product is 1 if m > n. For example, the factorial

function can be expressed as:
n! = ∏ i = 1 ⋅ 2 ⋅ ⋯ ⋅ (n − 1) ⋅ n,
n
i=1
where for mathematical convenience 0! = 1.

Similarly to sums, in products multiplicative terms that do not de-
pend on the index variable can also be pulled out of the product. How-
ever, if the product involves n terms, they have to be raised to the power
of n:
∏ kf (i) = k ∏ f (i).
n n
n
i=1 i=1
In addition, the product of a sum is not the sum of the products in
general:
∏ (f1 (i) + f2 (i)) ≠ ∏ f1 (i) + ∏ f2 (i).
n n n
i=m i=m i=m

Lastly, the logarithm of a product is a sum of logarithms:
log ( ∏ f (i)) = ∑ log f (i).

n n
i=m i=m
3.1.5 Floors and ceilings

The floor of a real number x, denoted as ⌊x⌋, is the largest integer that
is less than or equal to x. Similarly, the ceiling of x, denoted as ⌈x⌉, is
the smallest integer that is greater than or equal to x. Formally, they
can be described as:
⌊x⌋ = max{m ∈ Z ∣ m ≤ x},

⌈x⌉ = min{m ∈ Z ∣ m ≥ x},
where Z represents the set of all integers, and ∣ can be read as “such
that.” The following list includes several basic properties of floors and
ceilings:
• ⌊x⌋ ≤ x • ⌈x⌉ ≥ x
• ⌊x + n⌋ = ⌊x⌋ + n • ⌈x + n⌉ = ⌈x⌉ + n
• ⌊x⌋ + ⌊y⌋ ≤ ⌊x + y⌋ • ⌈x⌉ + ⌈y⌉ ≥ ⌈x + y⌉
• n = ⌊n/2⌋ + ⌈n/2⌉ • n − 2⌊n/2⌋ = 0 ⇔ n is even
• n // 2 = ⌊n/2⌋ • n >> m = ⌊n/2m ⌋

• ⌊log10 p⌋ + 1 = the number of decimal digits of p
• ⌊log2 p⌋ + 1 = the number of binary digits (bits) of p
where x is a real number, n is an integer, m is a nonnegative integer,

and p is a positive integer. In addition, // and >> represent operators
in Python (and also in other programming languages) that compute the
quotient of an integer division, and perform a right bit-shift, respectively.
Finally, ⇔ denotes “if and only if.”
3.1.6 Trigonometry
Consider the right triangle in Figure 3.2. The following list reviews basic
trigonometric definitions and properties:
c
a
Figure 3.2 Right triangle used for showing trigonometric definitions.
• sin(α) = a/c • cos(α) = b/c
• tan(α) = sin(α)/ cos(α) = a/b • cot(α) = cos(α)/ sin(α) = b/a
• sin(0) = 0 • cos(0) = 1
√
• sin(30○ ) = sin(π/6) = 1/2 • cos(30○ ) = cos(π/6) = 3/2
√ √
• sin(45○ ) = sin(π/4) = 2/2 • cos(45○ ) = cos(π/4) = 2/2
√
• sin(60○ ) = sin(π/3) = 3/2 • cos(60○ ) = cos(π/3) = 1/2
• sin(90○ ) = sin(π/2) = 1 • cos(90○ ) = cos(π/2) = 0
• sin(α) = − sin(−α) • cos(α) = cos(−α)
where sin, cos, tan, and cot denote sine, cosine, tangent, and cotan-
gent, respectively. In most programming languages, the arguments of
the trigonometric functions are in radians (one radian is equal to 180/π
degrees).
3.1.7 Vectors and matrices

A matrix is a collection of numbers (when programming they can be
other data types such as characters, Boolean values, etc.) that are ar-
ranged in rows and columns forming a rectangular structure. Formally,
an n × m-dimensional matrix A contains n rows and m columns of num-
bers, for n ≥ 1 and m ≥ 1. Typically, matrices are written within brackets
or parentheses:
⎡ a1,1 a1,2 ⋯ a1,m ⎤
⎢ ⎥
⎢ a ⎥
⎢ 2,1 a2,2 ⋯ a2,m ⎥
A=⎢ ⎥
⎢ ⋮ ⋮ ⋱ ⋮ ⎥
⎢ ⎥
⎢ an,1 an,2 ⋯ an,m ⎥
⎣ ⎦
where ai,j denotes the particular element or entry of matrix A in its i-th
row and j-th column.
If one of the dimensions is equal to one the mathematical entity is
called a vector, while if both dimensions are equal to one the object is
a scalar (i.e., a number). In this book we will use a standard notation
where matrices are represented by capital boldface letters (A), vectors
through boldface lower case letters (a), and scalars by italic lower case
letters or symbols (a).
The transpose of an n × m-dimensional matrix A is the m × n-
dimensional matrix AT whose rows are the columns of A (and therefore
its columns are the rows of A). For example, if:
⎡ 3 1 ⎤
⎢ ⎥
⎢ ⎥
A=[ ], AT = ⎢ 4 8 ⎥ .
3 4 2
⎢ ⎥
then
1 8 5 ⎢ 2 5 ⎥
⎣ ⎦
It is possible to add and multiply matrices (and vectors). The sum
A + B is a matrix whose entries are ai,j + bi,j . In other words, matrices
(and vectors) are added entrywise. Thus, A and B must share the same
dimensions. For example:
⎡ 4 ⎤ ⎡ 2 ⎤ ⎡ 6 ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ −1 ⎥ + ⎢ 3 ⎥ = ⎢ 2 ⎥ ,
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ 2 ⎥ ⎢ −7 ⎥ ⎢ −5 ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦
is a basic vector sum.
Instead, multiplication is more complex. Let a and b be two n-
dimensional (column) vectors. Their dot product, expressed as aT b
(other notations include a⃗⋅b
⃗ , or ⟨a, b⟩), is defined as the sum of their
entrywise products:
n
aT b = ∑ ai bi . (3.14)
i=1
Vectors of n components (i.e., defined in Rn ) also have a geometrical

interpretation. It can be shown that:
aT b = ∣a∣ ⋅ ∣b∣ ⋅ cos(α), (3.15)
where α is the angle between the vectors a and b, and ∣ ⋅ ∣ denotes the
Euclidean norm: √
∣a∣ = a21 + ⋯ + a2n . (3.16)
u+v v v v−u
u u
(a) (b)
Figure 3.3 Geometric interpretation of vector addition and subtraction.
Alternatively, let A and B be n × m and m × p-dimensional matrices,

respectively. The product C = A⋅B is an n×p-dimensional matrix whose
entries are defined as:
m
ci,j = ai,1 ⋅ b1,j + ai,2 ⋅ b2,j + ⋯ + ai,m ⋅ bm,j = ∑ ai,k ⋅ bk,j .
k=1
In this case, the number of columns of A (i.e., m) must be the same as

the number of rows of B. The entry ci,j corresponds to the dot product
between the i-th row of A, and the j-th column of B (which are obviously
both vectors). For example:
⎡ 3 1 ⎤ ⎡ 10 20 11 ⎤
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
A A=⎢ 4 8 ⎥⋅[ ] = ⎢ 20 80 48 ⎥ ,
3 4 2
⎢ ⎥ ⎢ ⎥
T
⎢ 2 5 ⎥ 1 8 5 ⎢ 11 48 29 ⎥
⎣ ⎦ ⎣ ⎦
which is a symmetric matrix, since it is identical to its transpose.
Vectors can also be regarded as points. Geometrically, adding two
vectors u and v can be understood as creating a new vector whose end-
point is the result of “concatenating” u and v, as shown in Figure 3.3(a).
It therefore follows that the vector that begins at the endpoint of u and
ends at the endpoint of v is the vector v − u, as illustrated in (b).
Lastly, a 2-dimensional vector can be rotated counterclockwise α
degrees (or radians) by multiplying it times the following “rotation”
⎡ ⎤
matrix:
⎢ cos(α) − sin(α) ⎥
R=⎢ ⎢
⎥,
⎥ (3.17)
⎢ sin(α) cos(α) ⎥
⎣ ⎦
as shown in Figure 3.4, where u is a column vector (in most mathematical
texts vectors are expressed as column vectors).
Ru ⎡ ⎤
⎢ cos(α) − sin(α) ⎥
R=⎢
⎢
⎥
⎥
⎢ sin(α) cos(α) ⎥
⎣ ⎦
α
u
Figure 3.4 Rotation matrix (counterclockwise).
Listing 3.1 Measuring execution times through Python’s time module.

1 import time
2
3 t = time.time ()
4 # execute some code here
5 elapsed_time = time.time() - t
6 print (elapsed_time)
3.2 COMPUTATIONAL TIME COMPLEXITY

The computational time complexity of an algorithm is a theoretical mea-
sure of how much time, or how many operations, it needs to solve a
problem (Listing 3.1 shows a simple way to measure execution times in
practice, through Python’s time module). It is determined by analyzing
a function, say T , of the input size that quantifies this number of op-
erations for a particular instance. In computer science the efficiency of
algorithms is generally studied by contemplating how T behaves when
the size of the problem is very large. Moreover, the key factor is the
rate at which T grows as the input size tends to infinity. The following
subsections explain these ideas and the mathematical notation typically
used to characterize computational time complexity. Lastly, while the
size of a problem may depend on several factors, in this introductory
book we will analyze the computational time complexity of algorithms
that solve problems whose size is determined by only one factor. Thus,
the runtime cost of the covered algorithms will be determined by a func-
tion T (n) of one parameter, where n usually represents the size of the
problem.
7
x 10
5 0.5n2
4.5
3.5
2.5
2 2000n + 50000
1.5
0.5
0
0 2000 4000 6000 8000 10000
n
Figure 3.5 The highest-order term determines the order of growth of a

function. For T (n) = 0.5n2 + 2000n + 50000 the order is quadratic, since
the term 0.5n2 clearly dominates the lower-order terms (even added up)
for large values of n.
3.2.1 Order of growth of functions

The function T (n) that quantifies the runtime of an algorithm may
contain several additive terms that contribute differently to its value for a
given input. For example, let T (n) = 0.5n2 +2000n+50000. For moderate
values of n all of the terms are relevant regarding the magnitude of
T (n). However, as n increases the leading term (0.5n2 ) affects the growth
of the function significantly more than the other two (even combined).
Therefore, the order of growth of the function is characterized by
its highest-order term. Figure 3.5 shows a plot of 0.5n2 together with
2000n + 50000. For large inputs it is apparent that the quadratic term
dominates the other two, and therefore characterizes the rate of growth
of T (n). The coefficients of the polynomial have been chosen in order
to illustrate that they do not play a significant role in determining the
function’s rate of growth.
Figure 3.6 plots several orders of growth that appear frequently in
computational complexity. They can be sorted as follows when consid-
ering large values of n:
1 < log n < n < n log n < n2 < n3 < 2n < n!,
216 n!
2n
212
n3
28 n2
n log n
24 n
log n
20 1
2 4 6 8 10 12
n
Figure 3.6 Orders of growth typically used in computational complexity.
where informally f (n) < g(n) if:
f (n)
lim = 0.
n→∞ g(n)
The previous orders of growth are called (from left to right) “constant,”
“logarithmic,” “linear,” “n-log-n,” “quadratic,” “cubic,” “exponential,”
and “factorial.”
Since the scale of the Y axis in Figure 3.6 is logarithmic, the differ-
ences between the orders of growth may appear to be much smaller than
Table 3.1Concrete values of common functions used in computational

complexity.
1 log2 n n n log2 n n2 n3 2n n!
1 0 1 0 1 1 2 1
1 1 2 2 4 8 4 2
1 2 4 8 16 64 16 24
1 3 8 24 64 512 256 40320
1 4 16 64 256 4096 65.536 2.09 ⋅ 1013
1 5 32 160 1024 32.768 4.295.967.296 2.63 ⋅ 1035
they actually are (the difference between consecutive tick marks means
that an algorithm is 16 times slower/faster). Table 3.1 shows concrete
values for the functions, where the fast growth rates of the exponen-
tial or factorial functions clearly stand out. Problems that cannot be
solved by algorithms in polynomial time are typically considered to be
intractable, since it would take too long for the methods to terminate
even for problems of moderate size. In contrast, problems that can be
solved in polynomial time are regarded as tractable. However, the line
between tractable and intractable problems can be subtle. If the run-
time of an algorithm is characterized by a polynomial order with a large
degree, in practice it could take too long to obtain a solution, or for its
intermediate results to be useful.
3.2.2 Asymptotic notation

An important detail about the orders mentioned so far is the lack of
constant multiplicative terms. Similarly to lower-order terms, these are
omitted since they are less relevant than the actual order of growth when
determining the computational efficiency for large inputs. Moreover, it is
not worth the effort to specify the efficiency of an algorithm with exact
precision, since its runtime depends on numerous factors that include
the computer’s hardware, the programming language, the compiler or
interpreter, and many others. Thus, it is sufficient to assume in prac-
tice that it simply takes a constant amount of time to execute a basic
instruction, where its value is irrelevant.
The theoretical analysis of algorithms and problems relies on a type
of notation called “asymptotic notation,” which allows us to discard
the lower-order terms and multiplicative constants when dealing with
arbitrarily large inputs. In particular, asymptotic notation provides def-
initions of sets that we can use in order to specify “asymptotic bounds.”
Big-O notation defines the following set:
O(g(n)) = {f (n) ∶ ∃ c > 0 and n0 > 0 / 0 ≤ f (n) ≤ c ⋅ g(n), ∀n ≥ n0 }.
If a function f (n) belongs to this set, then g(n) will be an asymptotic

upper bound for f (n). This means that g(n) will be greater than f (n),
but the definition only requires this to be true on an interval from some
point n0 > 0 until infinity, where in addition we can multiply g(n) times
a sufficiently large positive constant c. Figure 3.7(a) illustrates the idea
graphically. If g(n) is an asymptotic upper bound for f (n) then it must
be possible to find a positive constant c such that cg(n) ≥ f (n), but
f (n) ∈ O(g(n))
y c ⋅ g(x)
f (x)
0 n0 x
(a)
f (n) ∈ Ω(g(n))
y
f (x)
c ⋅ g(x)
0 n0 x
(b)
f (n) ∈ Θ(g(n))
y
c2 ⋅ g(x) f (x)
c1 ⋅ g(x)
0 n0 x
(c)
Figure 3.7 Graphical illustrations of asymptotic notation definitions for

computational complexity.
from some positive value n0 until infinity (whatever happens for n < n0
is irrelevant). In order to prove that a function belongs to O(g(n)) it
is sufficient to show the existence of a pair of constants c and n0 that
will satisfy the definition, since they are not unique. For instance, if the
definition is true for some particular c and n0 , then it will also be true for
larger values of c and n0 . In this regard, it is not necessary to provide the
lowest values of c and n0 that satisfy the definition O (this also applies
to the notations mentioned below).
Algorithms are often compared according to their efficiency in the
worst case, which corresponds to an instance of a problem, amongst
all that share the same size, for which the algorithm will require more
resources (time, storage, etc.). Since Big-O notation specifies asymptotic
upper bounds, it can be used in order to provide a guarantee that a
particular algorithm will need at most a certain amount of resources,
even in a worst-case scenario, for large inputs. For example, the running
time for the quicksort algorithm that sorts a list or array of n elements
(see Section 6.2.2) belongs to O(n2 ) in general, since it requires carrying
out on the order of n2 comparisons in the worst case. However, quicksort
can run faster (in n log n time) in the best and average cases.
In contrast, Big-Omega notation defines asymptotic lower
bounds:
Ω(g(n)) = {f (n) ∶ ∃ c > 0 and n0 > 0 / 0 ≤ c ⋅ g(n) ≤ f (n), ∀n ≥ n0 }.
Figure 3.7(b) illustrates the idea graphically. Big-Omega notation is use-

ful for specifying a lower bound on the resources needed to solve a prob-
lem, no matter which algorithm is applied. For instance, it is possible to
prove theoretically that any algorithm capable of sorting a list of n real
numbers will require Ω(n log n) comparisons in the worst case.
Finally, Big-Theta notation defines asymptotic tight bounds:
Θ(g(n)) = {f (n) ∶ ∃ c1 > 0, c2 > 0, and n0 > 0 /
0 ≤ c1 g(n) ≤ f (n) ≤ c2 g(n), ∀n ≥ n0 }.
If f (n) ∈ Θ(g(n)) then f (n) and g(n) will share the same order of
growth. Thus, by choosing two appropriate constants c1 and c2 , g(n)
will be both an upper and lower asymptotic bound of f (n), as shown in
Figure 3.7(c). In other words, f (n) ∈ O(g(n)) and f (n) ∈ Ω(g(n)). For
example, the merge sort algorithm for sorting a list or array of n ele-
ments always requires (in the best and worst case) on the order of n log n
comparisons. Therefore, we say its running time belongs to Θ(n log n).
When specifying an asymptotic bound the constants and lower-order

terms are dropped from the function g(n). For example, although it is
true that 3n2 + 10n ∈ O(5n2 + 20n), it is sufficient to indicate 3n2 +
10n ∈ O(n2 ), since these notations indicate orders of growth. In this
regard, the reader may have noticed the lack of the base of a logarithm
when describing the order of a function. The reason for this is that
the difference between two logarithms of different bases is a constant
multiplicative term. Let ρ represent some asymptotic bound, and a and
b two bases for logarithms. The following identities indicate that all
logarithms share the same order, regardless of their base:
⎛ ⎞
⎜ 1 ⎟
) = ρ⎜ logb g(n)⎟
logb g(n)
ρ (loga g(n)) = ρ ( ⎜ ⎟ = ρ (logb g(n)) .
logb a ⎜ logb a ⎟
⎝² ⎠
constant
Thus, the base of a logarithm is not specified when indicating its order
of growth.
Finally, we can also use limits to determine the order of functions, due
to the following equivalent statements that relate them to the definitions
of asymptotic bounds:
f (n)
f (n) ∈ O(g(n)) ⇔ lim < ∞ (constant or zero),
n→∞ g(n)
f (n)
f (n) ∈ Ω(g(n)) ⇔ lim > 0 (constant > 0, or infinity),
n→∞ g(n)
f (n)
f (n) ∈ Θ(g(n)) ⇔ lim = constant > 0.
n→∞ g(n)
3.3 RECURRENCE RELATIONS

The running time or number of operations carried out by a recursive
algorithm is specified through a recurrence relation (or simply, “recur-
rence”), which is a recursive mathematical function, say T , that describes
its computational cost. Consider the code in Listing 1.1 for adding the
n first positive integers. Firstly, the number of operations that it needs
to perform clearly depends on the input parameter n. Thus, T will be a
function of n.
In the base case (when n = 1) the method performs the basic opera-
tions shown in Figure 3.8. Before carrying out instructions the program
needs to store low-level information (e.g., regarding the parameters, or
def sum_first_naturals(n): a0
if n==1:
return 1
else:
return sum_first_naturals(n-1) + n
def sum_first_naturals(n):
if n==1: a1
return 1
else:
if n==1: a2
return 1
else:
if n==1:
return 1 a3
else:
Figure 3.8 Sequence of operations carried by the function in Listing 1.1

in the base case.
the return address). Say this requires a0 units of computing time, where
a0 is a simple constant. The next basic operation evaluates the condition,
taking a1 units of time. Since the result is True, the next operation is a
“jump” to the third line of the method, which requires a2 units of time.
Finally, the method can return the value 1 in the last step, requiring a3
units of time. In total, the method requires a = a0 + a1 + a2 + a3 units
of time for n = 1. Thus, we can define T (1) = a. The exact value of a
is irrelevant regarding the asymptotic computational complexity of the
method. What is important is that a is a constant quantity that does
not depend on n.
Alternatively, Figure 3.9 shows the operations carried out in the re-
cursive case (when n > 1). Let b = ∑5i=0 bi be the total computing time
needed to carry out the basic operations (store low-level information,
evaluate the condition, jump to the recursive case, subtract a unit from
n, add n to the output of the recursive call, and return the result), which
def sum_first_naturals(n): b0
if n==1:
return 1
else:
if n==1: b1
return 1
else:
if n==1: b2
return 1
else:
if n==1:
return 1
else:
return sum_first_naturals(n-1) + n b3
if n==1:
return 1
else:
return sum_first_naturals(n-1) + n T (n − 1)
if n==1:
return 1
else:
if n==1:
return 1
else:
Figure 3.9 Sequence of operations carried by the function in Listing 1.1

in the recursive case.
is also a constant whose exact value is irrelevant regarding asymptotic

computational complexity. In addition to b, we need to consider the time
required by the recursive call. Since it solves a full problem of size n − 1,
we can define it as T (n − 1). Altogether, the recurrence relation T (n)
can be specified as follows:
⎧
⎪
⎪a if n = 1,
T (n) = ⎨
⎪
(3.18)
⎩T (n − 1) + b if n > 1,
⎪
where, for example, T (3) = T (2) + b = T (1) + b + b = a + 2b.

Although T describes the algorithm’s computational cost correctly,
it is not trivial to figure out its order of growth since it is recursive.
Therefore, the next step of the analysis consists of transforming the
function into an equivalent nonrecursive form. This process is referred
to as “solving” the recurrence relation. In this example, it is not hard to
see that T (n) is a linear function of n:
T (n) = b(n − 1) + a = bn − b + a ∈ Θ(n).
The next sections will cover methods for solving common recurrence
relations.
In addition, in this introductory text we will simplify the recurrence
relations in order to ease the analysis. Consider the code in Listing 2.9.
Its associated runtime cost function can be defined as:
⎧
⎪
⎪
⎪
if n = 1,
⎪
a
T (n) = ⎨T ( n2 ) + b
⎪
if n > 1 and n is even, (3.19)
⎪
⎪
⎪
⎩T (⌊ 2 ⌋) + c if n > 1 and n is odd.
n
This recurrence relation is hard to analyze for two reasons. On the one
hand, it contains more than one recursive case. On the other hand, al-
though it is possible to work with the floor function, it is subject to
technicalities and requires more complex mathematics (e.g., inequali-
ties). Moreover, the extra complexity that stems from separating the
odd and even cases, and dealing with a recurrence of finer detail, is un-
necessary regarding the function’s order of growth. Since the algorithm
is based on solving subproblems of (approximately, in the odd case) half
the size, we can work with the following recurrence relation instead:
⎧
⎪
⎪a if n = 1,
T (n) = ⎨ n
⎪
(3.20)
⎩T ( 2 ) + b if n > 1.
⎪
Using this function is analogous to utilizing (3.19) and assuming that the
size of the problem (n) will be a power of two, which can be arbitrarily
large.
Therefore, in this book we will cover recurrence relations with only
one recursive case, which will not involve the floor or ceiling functions.
This will allow us to determine exact nonrecursive definitions of recur-
rence relations whose order of growth can be characterized by the tight
Θ asymptotic bound.
3.3.1 Expansion method

The expansion method, also known as the “iteration” or “backward sub-
stitution” method, can be used primarily to solve recurrence relations
whose recursive case contains one reference to the recursive function (in
some cases it can be applied to recurrences where the recursive func-
tion appears several times in the recursive definitions). The idea consists
of simplifying the recurrence relation progressively step by step, until
noticing a general pattern at some i-th stage. Subsequently, the function
can take concrete values by considering that the base case is reached
at that i-th step. The following examples illustrate the procedure. Fi-
nally, Section 10.2.1 illustrates a related visual approach called the “tree
method.”
3.3.1.1 Common recurrence relations

Consider the function defined in (3.18). Its recursive case:
T (n) = T (n − 1) + b (3.21)
can be applied repeatedly (to arguments of smaller size) in order to
“expand” the T term of the right-hand side. For example, T (n − 1) =
T (n−2) +b, where all we have done is substitute n with (n−1) in (3.21).
Thus, we arrive at:
T (n) = [T (n − 2) + b] + b = T (n − 2) + 2b,
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
T (n−1)
where the expression inside the square brackets is the expansion of T (n−
1). The idea can be applied again expanding T (n−2), which is T (n−3)+b.
Thus, at a third step we obtain:
T (n) = [T (n − 3) + b] + 2b = T (n − 3) + 3b.
´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹¶
T (n−2)
1. Write down the recursive case of the recurrence relation.
2. Expand the recursive terms (T ) on the right-hand side several

times until detecting a general pattern for the i-th step.
3. Determine the value of i that allows us to reach a base case.
4. Substitute for i in the general pattern.
Figure 3.10 Summary of the expansion method.
After several of these expansions we should be able to detect a general

pattern corresponding to the i-th step. For this function it is:
T (n) = T (n − i) + ib. (3.22)
Finally, for some value of i the process will reach a base case. For function
(3.18) it is defined for T (1). Therefore, the term T (n−i) will correspond
to the base case when n−i = 1, or equivalently, when i = n−1. Substituting
this result in (3.22) allows us to eliminate the variable i from the formula,
and provide a full nonrecursive definition of T (n):
T (n) = T (1) + (n − 1)b = a + (n − 1)b = bn − b + a ∈ Θ(n).
Thus, Listing 1.1 runs in linear time with respect to n.

Figure 3.10 presents a summary of the expansion method that we will
now apply to other common recurrence relations. Consider the function
defined in (3.20). The expansion process is:
T (n) = T (n/2) + b (step 1)

= [T (n/4) + b] + b = T (n/4) + 2b (step 2)
= [T (n/8) + b] + 2b = T (n/8) + 3b (step 3)
= [T (n/16) + b] + 3b = T (n/16) + 4b (step 4)
⋮
where the general pattern for the recurrence relation at step i is:
T (n) = T (n/2i ) + ib. (3.23)

The base case T (1) is reached when n/2i = 1. Thus, it occurs when n = 2i ,
or equivalently, when i = log2 n. By substituting in (3.23) we obtain:
T (n) = T (1) + b log2 n = a + b log2 n ∈ Θ(log n).
Since the order of growth is logarithmic, Listing 2.9 is faster than List-
ing 1.1, whose order is linear. This makes intuitive sense, since the former
decomposes the original problem by dividing its size by two, while the
latter is based on decrementing the size of the problem by a unit. Thus,
Listing 2.9 needs fewer recursive function calls in order to reach the base
case.
A subtle detail about recurrence relations that stem from dividing
the size of a problem by an integer constant k ≥ 2 is that they should not
contain a single base case for n = 0. From a mathematical point of view,
it would never be reached, and we would not be able to find a value for
i in the method’s third step. The pitfall is that the argument of T (n)
must be an integer. Thus, the fraction in T (n/k) actually corresponds
to an integer division. Notice that after reaching T (1) the next expan-
sion would correspond to T (0) instead of T (1/k). Therefore, for these
recurrence relations we should include an additional base case, usually
for n = 1, in order to apply the method correctly.
In the previous examples it was fairly straightforward to detect the
general recursive pattern for the i-th stage of the expansion process. For
the next recurrence relations the step is slightly more complex since it
will involve computing sums like the ones presented in Section 3.1.4.
Consider the following recurrence relation:
⎧
⎪
⎪a if n = 0,
T (n) = ⎨
⎪
(3.24)
⎩T (n − 1) + bn + c if n > 0.
⎪
The expansion process is:
T (n) = T (n − 1) + bn + c
= [T (n − 2) + b(n − 1) + c] + bn + c
= T (n − 2) + 2bn − b + 2c
= [T (n − 3) + b(n − 2) + c] + 2bn − b + 2c
= T (n − 3) + 3bn − b(1 + 2) + 3c
= [T (n − 4) + b(n − 3) + c] + 3bn − b(1 + 2) + 3c

= T (n − 4) + 4bn − b(1 + 2 + 3) + 4c
⋮
= T (n − i) + ibn − b(1 + 2 + 3 + ⋯ + (i − 1)) + ic,
where the square brackets indicate particular expansions of terms involv-
ing T . In addition, the last step contains the general recursive pattern,
which can be written as:
(i − 1)i
T (n) = T (n − i) + ibn − b ∑ j + ic = T (n − i) + ibn − b
i−1
+ ic. (3.25)
j=1 2
Two common misconceptions when using sums in the pattern at the i-th
stage are: (1) using i as the index variable of the sum, and (2) choosing
n as its upper limit. It is important to note that i − 1 is the upper limit
of the sum, which implies that the index of the sum cannot be i.
Finally, the base case T (0) is reached when i = n. Therefore, substi-
tuting in (3.25) yields:
T (n) = bn2 − n(n − 1) + cn + a = n2 + (c + ) n + a ∈ Θ(n2 ),

b b b
2 2 2
which is a quadratic polynomial.
The next recurrence relation appears in divide and conquer algo-
rithms such as merge sort (see Chapter 6):
⎧
⎪
⎪a if n = 1,
T (n) = ⎨
⎪
(3.26)
⎩2T (n/2) + bn + c if n > 1.
⎪
The expansion process is:
T (n) = 2T (n/2) + bn + c
= 2[2T (n/4) + bn/2 + c] + bn + c
= 4T (n/4) + 2bn + 2c + c
= 4[2T (n/8) + bn/4 + c] + 2bn + 2c + c
= 8T (n/8) + 3bn + 4c + 2c + c
= 8[2T (n/16) + bn/8 + c] + 3bn + 4c + 2c + c

= 16T (n/16) + 4bn + 8c + 4c + 2c + c
⋮
= 2i T (n/2i ) + ibn + c(1 + 2 + 4 + ⋯ + 2i−1 ).
Again, square brackets indicate expansions of terms involving T . In this
case they are especially useful since each of the expanded terms needs to
be multiplied by two. Students are advised to use them in general, since
omitting them is a common source of mistakes.
In this case, the recursive pattern contains a partial sum of a geo-
metric series.
T (n) = 2i T (n/2i ) + ibn + c ∑ 2j = 2i T (n/2i ) + ibn + c(2i − 1).

i−1
(3.27)
j=0
Finally, we reach base case T (1) when n/2i = 1, which implies that
i = log2 n. Therefore, substituting in (3.27) yields:
T (n) = nT (1) + bn log2 n + c(n − 1) = bn log2 n + n(a + c) − c ∈ Θ(n log n),
whose highest-order term is bn log2 n.
3.3.1.2 Master theorem

The master theorem is a result that can be used as a quick recipe for
determining the computational time complexity of algorithms based on
the divide and conquer design strategy. In particular, it can be applied
to recurrence relations of the following form:
⎧
⎪
⎪c if n = 1,
T (n) = ⎨
⎪
⎩aT (n/b) + f (n) if n > 1,
⎪
where a ≥ 1, b > 1, c ≥ 0, and f is an asymptotically positive function.

Thus, these algorithms solve a subproblems whose size is equal to the
original’s divided by b. In addition, the further processing or combination
of the subsolutions requires f (n) operations. Depending on the relative
contribution of the terms aT (n/b) and f (n) to the algorithm’s runtime
cost, the master theorem states, in its most general definition, that it is
possible to determine an asymptotic tight bound for T in the following
three cases:
1. If f (n) = O(nlogb a−ǫ ) for some constant ǫ > 0, then:
T (n) ∈ Θ(nlogb a ).
2. If f (n) = Θ (nlogb a (log n)k ), with k ≥ 0, then:
T (n) ∈ Θ (nlogb a (log n)k+1 ) .
3. If f (n) = Ω(nlogb a+ǫ ) with ǫ > 0, and f (n) satisfies the regularity
condition (af (n/b) ≤ df (n) for some constant d < 1, and for all n
sufficiently large), then:
T (n) ∈ Θ(f (n)).
For example, for

⎧
⎪
⎪1 if n = 1,
T (n) = ⎨
⎪
⎩T (n/2) + 2
⎪ if n > 1,
n
it is possible to find an ǫ > 0 such that the order of f (n) = 2n will be

greater than that of nlog2 1+ǫ = nǫ . Indeed, in this example any value ǫ > 0
would be valid since the exponential order of 2n is always greater than
the order of a polynomial. Thus, this recurrence relation falls in the third
case of the master theorem, which implies that T (n) ∈ Θ(2n ).
When f (n) is a polynomial of degree k we can apply the following
simpler version of the master theorem:
⎧
⎪ Θ(nk )
⎪
⎪
a
if <1
⎪
⎪
⎪
bk
⎪
⎪
T (n) ∈ ⎨ Θ(nk log n)
⎪
if a
=1 (3.28)
⎪
⎪
bk
⎪
⎪
⎪
⎪
⎪
⎩ Θ(n b )
a
log a
if bk
>1
This result can be derived through the expansion method. Consider

the following recurrence relation:
⎧
⎪
⎪c if n = 1,
T (n) = ⎨
⎪
(3.29)
⎩aT (n/b) + dn
⎪ if n > 1,
k
where a ≥ 1, b > 1, c ≥ 0, and d ≥ 0. The expansion process is:
T (n) = aT (n/b) + dnk

= a [aT (n/b2 ) + d(n/b)k ] + dnk
= a2 T (n/b2 ) + dnk (1 + a/bk )
= a2 [aT (n/b3 ) + d(n/b2 )k ] + dnk (1 + a/bk )
= a3 T (n/b3 ) + dnk (1 + a/bk + a2 /b2k )
⋮
a j
= ai T ( i ) + dnk ∑ ( k ) .
i−1
n
b j=0 b
The base case is T (1) = c, which is reached when i = logb n. Substituting

yields:
logb n−1 logb n−1
a j a j
T (n) = calogb n + dnk ∑ ( ) = cnlogb a
+ dnk
∑ ( ) ,
j=0 bk j=0 bk
where there are three cases depending on the values of a, b, and k:
1. If a < bk then ∑j=0b (a/bk ) will be a constant term (it cannot

log n−1 j
be infinity). Notice that the infinite sum ∑∞ i=0 r = 1/(1 − r) is a

i
constant (i.e., it does does not diverge) for r < 1. Thus, in this
scenario:
T (n) = cnlogb a + dKnk ,
for some constant K. In addition, since a < bk implies that logb a <
k, the highest-order term is nk . Therefore,
T (n) ∈ Θ(nk ).
2. If a = bk then T (n) = cnk + dnk ∑j=0b 1 = cnk + dnk logb n, which

log n−1
implies that
T (n) ∈ Θ(nk log n).
3. If a > bk , then, solving the geometric series:
logb n
T (n) = cnlogb a + dnk

( a
) −1
bk
( ak )−1
b
nlogb a
cnlogb a + dnk
−1
nk
= K
= (c + d
K
) nlogb a − d k
K
n ,
where K = a/bk − 1 is a constant. Finally, since a > bk implies that

logb a > k, we have:
T (n) ∈ Θ(nlogb a ).
3.3.1.3 Additional examples

A recurrence relation that captures the runtime of Listings 1.2 and 2.4
⎧
could be:
⎪
⎪1 if n = 1,
T (n) = ⎨
⎪
(3.30)
⎪
⎩ 2T (n/2) + 1 if n > 1,
where we have made the assumptions that multiplicative constants are
equal to 1, and that the size of the subproblems is exactly half of the
original problem’s. The recurrence multiplies T (n/2) times two since
the algorithm needs to invoke itself twice in the recursive cases, with
different arguments in each call. In addition, it adds the constant 1 since
the results of the subproblems need to be added and returned by the
method. Lastly, for simplicity, we do not need to specify a base case for
n = 2. These assumptions will not affect the runtime’s order of growth.
The recurrence relation can be solved through the expansion method
as follows:
T (n) = 2T (n/2) + 1
= 2[2T (n/4) + 1] + 1 = 4T (n/4) + 2 + 1
= 4[2T (n/8) + 1] + 2 + 1 = 8T (n/8) + 4 + 2 + 1
= 8[2T (n/16) + 1] + 4 + 2 + 1 = 16T (n/16) + 8 + 4 + 2 + 1
⋮
= 2i T (n/2i ) + ∑ 2j = 2i T (n/2i ) + 2i − 1.
i−1
j=0
The base case T (1) = 1 occurs when n = 2i . Substituting yields:
T (n) = n + n − 1 = 2n − 1 ∈ Θ(n),
where T (n) is a linear function of n. Naturally, this result is in accor-

dance with the master theorem (see (3.28)), since the recurrence relation
is a special case of (3.29), where a = 2, b = 2, and k = 0. Therefore,
T (n) ∈ Θ(2log2 n ) = Θ(n).
Consider the functions in Listing 1.5. The first two methods decom-
pose the problem by reducing its size by a unit. Thus, the associated
recurrence relation could be:
⎧
⎪
⎪1 if n = 1,
T (n) = ⎨
⎪
⎩T (n − 1) + 1 if n > 1,
⎪
which is a special case of (3.21), where T (n) ∈ Θ(n). Instead, the third
method decomposes the original problem in two of half the size, where
the subsolutions do not need to be processed any further. Therefore, the
corresponding recurrence relation can be identical to the one in (3.30).
Finally, we will study the computational time complexity of List-
ings 2.2 and 2.3. The associated recurrence relation (ignoring multiplica-
tive constants) is:
⎧
⎪
⎪1 if n < 10,
T (n) = ⎨
⎪
(3.31)
⎩T (n/10) + 1 if n ≥ 10.
⎪
It is different than the previous recurrences since the base case is defined
on an interval. Nevertheless, by assuming that the input will be a power
of 10, we can use an alternative definition of the recurrence:
⎧
⎪
⎪1 if n = 1,
T (n) = ⎨
⎪
(3.32)
⎩T (n/10) + 1 if n > 1.
⎪
Both functions are equivalent for n = 10p (and p ≥ 0).

The second recurrence is a special case of (3.29), where a = 1, b = 10,
and k = 0. Thus, the master theorem indicates that its complexity is
logarithmic, since T (n) ∈ Θ(nk log n) = Θ(log n).
We can obtain this result by the expansion method as follows:
T (n) = T (n/10) + 1
= [T (n/100) + 1] + 1 = T (n/102 ) + 2
= [T (n/1000) + 1] + 2 = T (n/103 ) + 3
= [T (n/10000) + 1] + 3 = T (n/104 ) + 4
⋮
= T (n/10i ) + i.
The base case is obtained when n/10i = 1, or in other words, when

i = log10 n. Substituting we obtain:
T (n) = T (1) + log10 n = 1 + log10 n ∈ Θ(log n).
3.3.2 General method for solving difference equations

The expansion method is effective at solving recurrence relations in
which the recursive function appears only once in the recursive defini-
tions. This section describes a powerful approach, denoted in this book
as the “general method for difference equations,” which allows us to solve
recurrences that may “call” themselves several times. In particular, the
method can be used to solve recurrences of the following form:
T (n) = −a1 T (n − 1) − ⋯ − ak T (n − k) + P1d1 (n)bn1 + ⋯ + Psds (n)bns , (3.33)

´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶ ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
“T difference” terms polynomial × exponential terms
where ai and bi are constants, and Pidi (n) are polynomials (of n) of
degree di . Terms involving T may appear several times on the right-
hand side of the definition. Moreover, their arguments are necessarily n
minus some integer constant. Thus, these recurrence relations are also
known as “difference equations.” In this book we will call these terms
involving T the “T difference” terms, to emphasize that the arguments
cannot take the form n/b, where b is a constant (if these terms appear it is
necessary to transform the recurrence). In addition, the right-hand side
of the definition may contain several terms that consist of a polynomial
times a power of the input n (i.e., an exponential).
Instead of providing a general procedure directly, the following sub-
sections will explain the method progressively, starting with simple re-
currences, and introducing new elements as they grow in complexity.
3.3.2.1 Homogeneous recurrences and different roots

Homogeneous recurrence relations only contain T difference terms:
T (n) = −a1 T (n − 1) − ⋯ − ak T (n − k).
The first step for solving them consists of passing every term to the
left-hand side of the definition:
T (n) + a1 T (n − 1) + ⋯ + ak T (n − k) = 0.
Subsequently, we define its associated characteristic polynomial by
applying the change xk−z = T (n − z) for z = 0, . . . , k:
xk + a1 xk−1 + ⋯ + ak−1 x + ak .
All we have done is replace T (n) with xk , T (n − 1) with xk−1 , and so
on. The coefficient (ak ) associated with the T difference term that has
the smallest argument will be the polynomial’s constant. The next step
consists of finding the k roots of the characteristic polynomial. If ri is
its i-th root then it can be expressed in factorized form as:
(x − r1 )(x − r2 ) ⋯ (x − rk ).
If all of the roots are different then the nonrecursive expression for T
will be:
T (n) = C1 r1n + ⋯ + Ck rkn . (3.34)
Finally, the values of the constants will depend on the base cases of
T , and will be found by solving a system of k linear equations with
k unknown variables (the constants Ci ). We will need to find k initial
values (base cases, for small values of n) of T in order to construct the
k equations.
The following example illustrates the approach applied to the basic
Fibonacci function defined in (1.2), and coded in Listing 1.3. It can be
rewritten as:
⎧
⎪
⎪
⎪
0 if n = 0,
⎪
T (n) = ⎨1
⎪
if n = 1, (3.35)
⎪
⎪
⎪
⎩T (n − 1) + T (n − 2) if n > 1,
where we have included an extra base case for n = 0 that will be useful
shortly. The first step consists of writing each T difference term on the
left-hand side of the recursive identity:
T (n) − T (n − 1) − T (n − 2) = 0.
It is a homogeneous recurrence since there is a 0 on the right-hand side.

We then form the characteristic polynomial:
x2 − x − 1.
The roots of this quadratic function are:

√ √
1+ 5 1− 5
r1 = , and r2 = ,
2 2
which are different. Therefore, we can express the recurrence nonrecur-
sively as:
√ n √ n
1+ 5 1− 5
T (n) = C1 r1 + C2 r2 = C1 (
n n
) + C2 ( ) . (3.36)
2 2
The last step consists of finding values for the constants Ci by solving a
system of linear equations. Each equation is formed by choosing a small
value for n corresponding to a base case, and plugging it in (3.36). The
simplest equation is obtained for n = 0, which is the reason why we
considered the base case n = 0 in (3.35). For the second equation we can
use n = 1. This provides the following system of linear equations:
⎫
C1 + C2 = 0 = T (0) ⎪
⎪
⎪
⎬.
( 2 ) C1 + ( 2 ) C2 = 1 = T (1) ⎪
√ √
⎪
⎪
⎭
1+ 5 1− 5
The solutions are:

1 1
C1 = √ , and C2 = − √ .
5 5
Therefore, the Fibonacci function can be expressed as:
√ n √ n √ n
1 1+ 5 1 1− 5 1+ 5
T (n) = F (n) = √ ( ) −√ ( ) ∈ Θ( ) , (3.37)
5 2 5 2 2
which is an exponential function. Observe that r1n clearly dominates

the growth of the function. On the one hand, r1 > r2 . On the other
hand, ∣r2 ∣ < 1, which means that r2n will approach 0 as n approaches
infinity. Finally, despite its complex appearance, the result is obviously
an integer.
For the next example consider the mutually recursive functions
defined in (1.17) and (1.18). They must be redefined exclusively in
terms of themselves in order to apply the method. On the one hand,

B(n) = A(n − 1) implies that B(n − 1) = A(n − 2). In addition, A(2) = 1.
Thus, we can redefine A(n) as:
⎧
⎪
⎪
⎪
⎪
0 if n = 1,
A(n) = ⎨1
⎪
if n = 2, (3.38)
⎪
⎪
⎪
⎩A(n − 1) + A(n − 2) if n ≥ 3.
On the other hand, since A(n − 1) = B(n), and A(n) = B(n + 1), we
can substitute in (1.17) in order to obtain B(n + 1) = B(n) + B(n − 1).
Furthermore, since B(2) = 0, we can define B(n) as:
⎧
⎪
⎪
⎪
⎪
1 if n = 1,
B(n) = ⎨0
⎪
if n = 2, (3.39)
⎪
⎪
⎪
⎩B(n − 1) + B(n − 2) if n ≥ 3.
Both of these functions have the form in (3.36). Thus, the only difference
between them is the value of the constants. In particular:
√ n √ n
1 1 1+ 5 1 1 1− 5
A(n) = ( − √ ) ( ) + ( + √ )( ) , (3.40)
2 2 5 2 2 2 5 2
and
√ n √ n
1 3 1+ 5 1 3 1− 5
B(n) = (− + √ ) ( ) + (− − √ ) ( ) , (3.41)
2 2 5 2 2 2 5 2
where it is possible to assume that A(0) = 1, and B(0) = −1. Lastly, it

is not hard to see that A(n) + B(n) are Fibonacci numbers F (n), since
adding (3.40) and (3.41) yields (3.37).
3.3.2.2 Roots with multiplicity greater than one

The previous section showed that if the multiplicity of a root ri is 1,
corresponding to a term (x−ri )1 in the factorization of the characteristic
polynomial, then the nonrecursive version of T (n) will contain the term
Ci rin . In contrast, when the multiplicity m of a root r is greater than
1, resulting from (x − r)m , then T (n) will incorporate m terms of the
form: constant × polynomial × rn . In particular, the polynomials will
be different powers of n, ranging from 1 to nm−1 . For example, a term
(x − 2)4 in the factorization of the characteristic polynomial would lead

to the four following terms in the nonrecursive version of T (n):
C1 2n + C2 n2n + C3 n2 2n + C4 n3 2n ,
for some constants Ci . Therefore, T (n) can be expressed in general as:
T (n) = C1 P1 (n)r1n + ⋯ + Ck Pk (n)rkn , (3.42)
where Pi (n) is a polynomial of the form nc , for some c.
For example, consider the following recurrence:
T (n) = 5T (n − 1) − 9T (n − 2) + 7T (n − 3) − 2T (n − 4),
with T (0) = 0, T (1) = 2, T (2) = 11, and T (3) = 28. It can be expressed
as T (n) − 5T (n − 1) + 9T (n − 2) − 7T (n − 3) + 2T (n − 4) = 0, leading to
following the characteristic polynomial:
x4 − 5x3 + 9x2 − 7x + 2.
It can be factorized (for example, by using Ruffini’s rule) as:
(x − 1)3 (x − 2),
which implies that the recurrence will have the following form:
T (n) = C1 ⋅ 1 ⋅ 1n + C2 ⋅ n ⋅ 1n + C3 ⋅ n2 ⋅ 1n + C4 ⋅ 1 ⋅ 2n
= C1 + C2 n + C3 n2 + C4 2n ,
where there are three terms associated with root r = 1. Finally, the
constants can be recovered by solving the following system of linear
equations:
⎫
+ C4 = 0 = T (0) ⎪ ⎪
⎪
⎪
⎪
C1
⎪
⎪
C1 + C2 + C3 + 2C4 = 2 = T (1) ⎪ ⎪
⎪
⎬.
C1 + 2C2 + 4C3 + 4C4 = 11 = T (2) ⎪ ⎪
⎪
⎪
⎪
⎪
⎪
⎪
C1 + 3C2 + 9C3 + 8C4 = 28 = T (3) ⎪ ⎪
⎭
The solutions are C1 = −1, C2 = −2, C3 = 3, and C4 = 1 (Listing 3.2
shows how to solve the system of linear equations, expressed as Ax = b,
with the NumPy package). Finally,
T (n) = −1 − 2n + 3n2 + 2n ∈ Θ(2n ),
whose order of growth is exponential.
Listing 3.2 Solving a system of linear equations, Ax = b, in Python.

1 import numpy as np
2
3 A = np.array ([[1, 0, 0, 1], [1, 1, 1, 2],
4 [1, 2, 4, 4], [1, 3, 9, 8]])
5 b = np.array ([0, 2, 11, 28])
6 x = np.linalg .solve (A, b)
3.3.2.3 Nonhomogeneous recurrences

A nonhomogeneous recurrence contains nonrecursive terms on its right-
hand side. We will be able to solve these recurrences when such terms
consist of a polynomial times an exponential, as shown in (3.33). The
procedure is exactly the same, but for every term Pidi (n)bni we will need
to include (x − bi )di +1 in the characteristic polynomial, where di is the
degree of the polynomial Pi (n).
Consider the following recurrence relation:
T (n) = 2T (n − 1) − T (n − 2) + 3n + n3n + 3 + n + n2 .
As in the previous examples, the first step consists of moving the T

difference terms to the left-hand side of the recurrence:
T (n) − 2T (n − 1) + T (n − 2) = 3n + n3n + 3 + n + n2 .
For the next step it is useful to express the terms on the right-hand side
as the product of a polynomial times an exponential. Naturally, if a term
only contains an exponential then it can be multiplied by 1, which is a
polynomial. Similarly, we can multiply polynomials times 1n . Therefore,
the recurrence can be written as:
T (n) − 2T (n − 1) + T (n − 2) = 1 ⋅ 3n + n ⋅ 3n + 3 ⋅ 1n + n ⋅ 1n + n2 ⋅ 1n .
Furthermore, even though a particular exponential may appear several

times, we must consider that it multiplies a single polynomial. Thus, the
recurrence should be regarded as:
T (n) − 2T (n − 1) + T (n − 2) = (1 + n) ⋅ 3n + (3 + n + n2 ) ⋅ 1n .
The next step consists of determining the characteristic polynomial.

From the left-hand side we have (x2 − 2x + 1) = (x − 1)2 . From the
term (1 + n) ⋅ 3n we need to include (x − 3)2 , where the 3 is the base
of the exponential, and 2 is the degree of the polynomial (1) plus one.
Similarly, (3 + n + n2 ) ⋅ 1n provides the new term (x − 1)3 . Therefore, the
characteristic polynomial is;
(x − 1)2 (x − 3)2 (x − 1)3 = (x − 1)5 (x − 3)2 ,
and T (n) has the following form:
T (n) = C1 + C2 n + C3 n2 + C4 n3 + C5 n4 + C6 3n + C7 n3n .
The next example illustrates the approach with the following non-
homogeneous recurrence relation:
⎧
⎪
⎪1 if n = 0,
T (n) = ⎨
⎪
⎩2T (n − 1) + n + 2
⎪ if n > 0.
n
In this case, we can write the recursive case as T (n) − 2T (n − 1) =

n ⋅ 1n + 1 ⋅ 2n . The corresponding factorized characteristic polynomial is:
(x − 2) ⋅ (x − 1)2 ⋅ (x − 2) = (x − 1)2 (x − 2)2 ,

´¹¹ ¹ ¹ ¹¸¹ ¹ ¹ ¹ ¶ ´¹¹ ¹ ¹ ¹ ¹ ¹¸ ¹ ¹ ¹ ¹ ¹ ¹¶n ´¹¹ ¹ ¹ ¹¸¹ ¹ ¹ ¹ ¶n
from T (n) − 2T (n − 1) from n ⋅ 1 from 1 ⋅ 2
which implies that the recurrence has the following form:
T (n) = C1 + C2 n + C3 2n + C4 n2n .
Since there are four unknown constants we need the values of T evaluated
at four different inputs. Staring with the base case T (0) = 1, we can
compute T (1), T (2), and T (3) by using T (n) = 2T (n − 1) + n + 2n . In
particular, T (1) = 5, T (2) = 16, and T (3) = 43. With this information
we can build the following system of linear equations:
⎫
+ C3 1 = T (0) ⎪
⎪
⎪
⎪
=
⎪
C1
⎪
⎪
C1 + C2 + 2C3 + 2C4 = 5 = T (1) ⎪ ⎪
⎪
⎬.
C1 + 2C2 + 4C3 + 8C4 = 16 = T (2) ⎪⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
C1 + 3C2 + 8C3 + 24C4 = 43 = T (3) ⎪
⎭
The solutions are C1 = −2, C2 = −1, C3 = 3, and C4 = 1. Therefore, the
nonrecursive expression of T (n) is:
T (n) = −2 − n + 3 ⋅ 2n + n2n ∈ Θ(n2n ).

3.3.2.4 Fractional inputs

The general method for difference equations can only be applied when
the inputs of terms involving T are differences of the form n − b, where b
is some integer constant. Nevertheless, it is possible to solve some recur-
rence relations that contain terms where the inputs to T are fractions
of the form n/b, for some constant b. The key resides in transforming
the fraction into a difference through the change of variable n = bk , and
therefore assuming that n is a power of b.
Consider the recurrence relation in (3.30), whose recursive case is:
T (n) = 2T (n/2) + 1,
and where we can assume that the input is a power of two. Since the
argument in the T term of the right-hand side is n/2 we need to apply
the change of variable n = 2k in order to obtain:
T (2k ) = 2T (2k /2) + 1 = 2T (2k−1 ) + 1.
In this new definition T (2k ) is a function of k. Therefore, we can sub-

stitute it with the term t(k):
t(k) = 2t(k − 1) + 1,
which we can solve through the general method for difference equations.
In particular, the expression can be written as t(k) − 2t(k − 1) = 1 ⋅ 1n .
Therefore, the characteristic polynomial is (x−2)(x−1), and the function
will have the following form:
t(k) = C1 + C2 2k . (3.43)
By undoing the change of variable we obtain a general solution to the

recurrence in terms of the original variable:
T (n) = C1 + C2 n. (3.44)
The last step consists of determining the constants C1 and C2 . This can
be done through either (3.43) or (3.44). For T we can use the base cases
T (1) = 1 and T (2) = 3. The analogous base cases for t are: t(0) = T (20 ) =
T (1) = 1, and t(1) = T (21 ) = T (2) = 3. Either way, the system of linear
equations is:
⎫
C1 + C2 = 1 = T (1) = t(0) ⎪ ⎪
⎪
⎬.
C1 + 2C2 = 3 = T (2) = t(1) ⎪ ⎪
⎪
⎭
The solutions are C1 = −1 and C2 = 2, and the nonrecursive expression

of T (n) is:
T (n) = 2n − 1 ∈ Θ(n).
In the next recurrence relation we can assume that the input is a
power of four. In particular, let:
⎧
⎪
⎪1 if n = 1,
T (n) = ⎨
⎪
⎩2T (n/4) + log2 n if n > 1.
⎪
In this case applying the change of variable n = 4k leads to:

T (4k ) = 2T (4k /4) + log2 4k = 2T (4k−1 ) + k log2 4 = 2T (4k−1 ) + 2k.
The next step consists of performing the substitution t(k) = T (4k ):
t(k) = 2t(k − 1) + 2k.
We will call this operation a “change of function” in order to emphasize
that only the T terms are replaced in the expression (by the t terms).
Therefore, the change does not affect the term 2k. The recurrence can be
rewritten as t(k) − 2t(k − 1) = 2k ⋅ 1k , where the associated characteristic
polynomial is (x − 2)(x − 1)2 . Thus, the recurrence has the form:
t(k) = C1 2k + C2 + C3 k.
In order to undo the change variable we can use k = log4 n, but we need
an expression for 2k . In particular, notice that n = 4k = (22 )k = (2k )2 .
√
Therefore, 2k = n, which leads to:
√
T (n) = C1 n + C2 + C3 log4 n.
Finally, from the base case T (1) = 1 we can compute T (4) = 2T (1) +
log2 4 = 4, and T (16) = 2T (4) + log2 16 = 12. This allows us to compute
the constants Ci by solving the following system of linear equations:
⎫
C1 + C2 T (1) ⎪
⎪
⎪
⎪
⎪
= 1 =
⎪
⎪
2C1 + C2 + C3 = 4 = T (4) ⎬ .
⎪
⎪
⎪
⎪
4C1 + C2 + 2C3 = 12 = T (16) ⎪
⎪
⎪
⎭
The solutions are C1 = 5, C2 = −4, and C3 = −2, and the nonrecursive
expression of T (n) is:
√
T (n) = 5 n − 4 − 2 log4 n ∈ Θ(n1/2 ).
When the argument of T in the recursive expression is a square

root, we can use the change of variable n = 2(2 ) . Consider the following
k
recurrence relation:
√
T (n) = 2T ( n) + log2 n (3.45)
where T (2) = 1 and n = 2(2 ) (note that 2(2 ) ≠ 4k ). This last restriction
k k
on n guarantees that the recursive procedure will terminate at the base

case for n = 2. Applying the change of variable, we have:
T (2(2 ) ) = 2T (2(2 /2)

) + log2 2(2 )
= 2T (2(2 )
) + 2k .
k k k k−1
By using the change of function t(k) = T (2(2 ) ), the recurrence is t(k) =

k
t(k − 1) + 2k , whose characteristic polynomial is (x − 2)(x − 2). Therefore,

the new recurrence will have the form:
t(k) = C1 2k + C2 k2k .
In order to undo the change of variable we can use k = log2 (log2 n), and
2k = log2 n. The recurrence as a function of n is therefore:
T (n) = C1 log2 n + C2 (log2 (log2 n)) log2 n. (3.46)
Finally, we can use the base cases T (2) = 1 and T (4) = 4 in order to
find the constants. In particular, we need to solve the following system
of linear equations:
⎫
= 1 = T (2) ⎪
⎪
⎪
⎬.
C1
2C1 + 2C2 = 4 = T (4) ⎪
⎪
⎪
⎭
The solutions are C1 = C2 = 1. Therefore, the final nonrecursive formula
for T (n) is:
T (n) = log2 n + (log2 (log2 n)) log2 n ∈ Θ((log(log n)) log n). (3.47)
3.3.2.5 Multiple changes of variable or function

The previous recurrence in (3.45) can also be solved by applying two
consecutive changes of variable (and function). Firstly, the change n = 2k
leads to: √
T (2k ) = 2T ( 2k ) + log2 2k = T (2k/2 ) + k.
Replacing T (2k ) with t(k), we obtain the following recurrence:
t(k) = 2t(k/2) + k,
which still cannot be solved through the method since it is not a differ-
ence equation. However, we can apply a new change of variable in order
to transform it into one. With the change k = 2m we obtain:
t(2m ) = 2t(2m−1 ) + 2m ,
which can be written as a difference equation by performing the change

of function u(m) = t(2m ) :
u(m) = 2u(m − 1) + 2m .
Its characteristic polynomial is (x−2)2 , which implies that the recurrence

has the following form:
u(m) = C1 2m + C2 m2m .
Undoing the changes of variable leads to:
t(k) = C1 k + C2 (log2 (k))k,
and finally:
T (n) = C1 log2 n + C2 (log2 (log2 n)) log2 n,
which is identical to (3.46). Thus, the solution is provided in (3.47).

The strategy of using several changes of variables or functions can be
used to solve more complex recurrences, such as the following nonlinear
⎧
one:
⎪
⎪1/3 if n = 1,
T (n) = ⎨
⎪
⎩n [T (n/2)] if n > 1,
⎪
2
where n is a power of two. Firstly, we can apply the change of variable

n = 2k , which leads to:
T (2k ) = 2k [T (2k /2)] = 2k [T (2k−1 )] .

2 2
In addition, with the change of function t(k) = T (2k ) we obtain:
t(k) = 2k [t(k − 1)] ,

2
which we still cannot solve. However, we can take logarithms on both

sides of the definition:
log2 t(k) = log2 (2k [t(k − 1)] ) = k + 2 log2 t(k − 1),
2
and apply the more complex change of function u(k) = log2 t(k), which
leads to:
u(k) = k + 2u(k − 1),
and that we can solve through the method. In particular, its charac-
teristic polynomial is (x − 2)(x − 1)2 , which implies that u(k) has the
following form:
u(k) = C1 2k + C2 + C3 k.
Undoing the changes, we first have:
k
t(k) = 2C1 2 +C2 +C3 k
,
and finally:
T (n) = 2C1 n+C2 +C3 log2 n .
The last step consists of determining the constants. Using the initial
recurrence with T (1) = 1/3, we obtain T (2) = 2[T (1)]2 = 2/9, and
T (4) = 4[T (2)]2 = 4(2/9)2 = 16/81. We can use these values to build
the following system of (nonlinear) equations:
⎫
1/3 = T (1) ⎪
⎪
⎪
⎪
⎪
2C1 +C2 =
⎪
⎪
2/9 = T (2) ⎬ ,
⎪
2 2C1 +C2 +C3
=
⎪
⎪
⎪
= 16/81 = T (4) ⎪
⎪
⎪
⎭
2 4C1 +C2 +2C3
which can be transformed into a system of linear equations by taking

logarithms on both sides of the equations:
⎫
C1 + C2 −log2 3 ⎪
⎪
⎪
⎪
⎪
=
⎪
⎪
2C1 + C2 + C3 = 1 − log2 9 = 1 − 2 log2 3 ⎬ .
⎪
⎪
⎪
⎪
4C1 + C2 + 2C3 = 4 − log2 81 = 4 − 4 log2 3 ⎪ ⎪
⎪
⎭
The solutions are: C1 = 2 − log2 3 = log2 (4/3), C2 = −2, and C3 = −1.
Therefore, T (n) can be expressed as:
T (n) = 2(log2 (4/3))n−2−log2 n = [2log2 (4/3) ] ⋅ 2−2 ⋅ 2log2 (1/n)
n
4 n 1 4 n 1
=[ ] ⋅ ∈ Θ ([ ] ⋅ ) .
3 4n 3 n
3.4 EXERCISES
Exercise 3.1 — Prove the following identity:
a logb n nlogb a
( ) =
bk nk
Exercise 3.2 — By using limits, show that log n! ∈ Θ(n log n). Hint:
when n approaches infinity, n! can be substituted by “Stirling’s approx-
√
imation”:
n n
n! ∼ 2πn ( ) .
e
Exercise 3.3 — Show that n log n ∈ O(n1+a ), where a > 0. Use limits,
and L’Hopital’s rule.
Exercise 3.4 — Let m and n be some integers. Determine:

n
∑ 1.
i=m
Exercise 3.5 — Write the sum of the first n odd integers in sigma
notation, and simplify it (the result should be a well-known polynomial).
Exercise 3.6 — Use the following identity:

n
i(i + 1) n
∑ = ∑ i(n − i + 1)
i=1 2 i=1
to provide an expression for the sum of the first n squared integers

(12 + 22 + ⋯ + n2 ).
Exercise 3.7 — Show that a general partial sum of n terms of an arith-

metic sequence (si = si−1 + d, for some initial s0 ) follows:
(s0 + sn−1 ).
n−1
∑ si =
n
i=0 2
Exercise 3.8 — An algorithm processes some of the bits of the binary

representations of numbers from 1 to 2n − 1, where n is the number
of bits of each number. In particular, the algorithm processes the least
significant bits of each number (from right to left), until it finds a bit
set to 1. Given n, determine, using sums, the total number of bits that
the algorithm processes. For example, for n = 4 the algorithm processes
the 26 shaded bits in Figure 3.11.
0 0 0 1
0 0 1 0
0 0 1 1
0 1 0 0
0 1 0 1
0 1 1 0
0 1 1 1
1 0 0 0
1 0 0 1
1 0 1 0
1 0 1 1
1 1 0 0
1 1 0 1
1 1 1 0
1 1 1 1
Figure 3.11 An algorithm processes the shaded bits of numbers from 1 to

2n − 1 (for n = 4).
Exercise 3.9 — Specify recurrence relations that describe the runtime

of algorithms that implement functions (1.14), (1.15), (1.16), and (1.19).
Exercise 3.10 — Solve the following recurrences, but without comput-

ing multiplicative constants (Ci ). Thus, in this exercise the base cases
are not needed.
a) T (n) = 4T (n − 1) − 5T (n − 2) + 2T (n − 3) + n − 3 + 5n2 ⋅ 2n
b) T (n) = T (n − 1) + 3n − 3 + n3 ⋅ 3n
c) T (n) = 5T (n − 1) − 8T (n − 2) + 4T (n − 3) + 3 + n2 + n2n
Exercise 3.11 — Define a recurrence relation for the runtime of List-

ing 2.6, then solve it through the expansion method and the general
method for difference equations. Finally, determine its order of growth.
Exercise 3.12 — Solve the following recurrence relation through the

expansion method:
⎧
⎪
⎪1 if n = 0,
T (n) = ⎨
⎪
⎩1 + ∑j=0 T (j) if n > 0.
⎪
n−1

expansion method:
⎧
⎪
⎪1 if n = 1,
T (n) = ⎨
⎪
⎩2T (n/4) + log2 n if n > 1.
⎪
Exercise 3.14 — Solve the recurrence relation in (3.18) through the

general method for difference equations.
Exercise 3.15 — Define a recurrence relation for the runtime of List-

ing 2.9, solve it through the master theorem, the expansion method, and
the general method for difference equations. Finally, determine its order
of growth.

master theorem, the expansion method, and the general method for dif-
ference equations. Finally, determine its order of growth.
⎧
⎪
⎪1 if n = 1,
T (n) = ⎨
⎪
⎩3T (n/2) + n if n > 1,
⎪
where n is a power of 2.
Exercise 3.17 — Solve the recurrence relation in (3.32) through the

general method for difference equations.
Exercise 3.18 — Solve the following recurrence relations:
a) T (n) = 2T (n − 1) + 3n − 2, with T (0) = 0.
b) T (n) = T (n/2) + n, with T (1) = 1, and n is a power of 2.
c) T (n) = T (n/α) + n, with T (1) = 0, and where α ≥ 2 is an integer.
d) T (n) = T (n/3) + n2 , with T (1) = 1, and n is a power of 3.
e) T (n) = 3T (n/3) + n2 , with T (1) = 0, and n is a power of 3.
f) T (n) = 2T (n/4) + n, with T (1) = 1, and n is a power of 4.
g) T (n) = T (n/2) + log2 n, with T (1) = 1, and n is a power of 2.
h) T (n) = 4T (n/2) + n, with T (1) = 1, and n is a power of 2.

i) T (n) = 2T (n/2) + n log2 n, with T (1) = 1, and n is a power of 2.
j) T (n) = 23 T (n/2) − 12 T (n/4) − n1 , with T (1) = 1, T (2) = 3/2, and n

is a power of 2.
CHAPTER 4
Linear Recursion I: Basic

Algorithms
Do the difficult things while they are easy and do the great things
while they are small. A journey of a thousand miles must begin
with a single step.
— Lao Tzu
ECURSIVE methods can be categorized according to the number

R of times they invoke themselves in the recursive cases. This chap-
ter examines methods that not only call themselves just once, but also
process the output of the recursive call before producing or returning
their own result. This common type of recursion is known as linear re-
cursion, where the single recursive call is not the last operation carried
out by the method. This chapter will present numerous problems and
respective linear-recursive solutions that we will design by relying on the
concepts and methodology introduced in Chapters 1 and 2. Finally, if the
single recursive call is the method’s last action, we say we are using tail
recursion. This special type of recursion will be covered in Chapter 5.
Linear recursion is the simplest type of recursion, and can be viewed
as an alternative way to implement iterative loops that carry out a cer-
tain number of operations repeatedly. In this regard, although it is fairly
straightforward to solve the problems discussed in this chapter itera-
tively, linear recursion provides the clearest examples of how to think
and program recursively. The chapter will therefore present examples of
how to apply recursive concepts (i.e., problem decomposition and induc-
tion), which the reader should comprehend before tackling more complex
problems and types of recursion.
105
4.1 ARITHMETIC OPERATIONS

This section presents recursive solutions to several elemental arithmetic
computations. We will examine them for illustration purposes, since most
correspond to simple operations that can be implemented through basic
commands or expressions.
4.1.1 Power function

A classical problem used to illustrate recursion is the power function.
The goal is to calculate b to the power of n:
n
bn = ∏ b = b × b × ⋯ b × b,
i=1 ´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
n times
where the base b is a real number, and the exponent n is a nonnegative in-
teger (in Python, powers can be obtained through the ** operator.) The
following subsections examine algorithms that can compute the power
in linear and logarithmic time.
4.1.1.1 Powers in linear time

According to the methodology presented in Figure 2.1, the first step
consists of determining the size of the problem, which is related to the
input parameters. For this problem, only the exponent is relevant for
determining the number of operations that algorithms need to solve it.
Clearly, the runtime of the algorithms increases as n grows, since the
result would require more multiplications. Thus, the exponent n is the
size of the problem. The base case corresponds to trivial instances of
the problem, with minimum size. Since n is a nonnegative integer, the
smallest instance is b0 , which is equal to one.
In the decomposition step we have to consider smaller and simpler
self-similar subproblems that stem from reducing the size of the problem.
In this section we will simply decrement it by a single unit, which leads
to the following diagram:
Inputs Results
f
(b, n) ÐÐÐÐÐÐÐÐÐÐ→ bn
×b
f
(b, n − 1) ÐÐÐÐÐÐÐÐÐÐ→ bn−1
Linear Recursion I: Basic Algorithms 107
Listing 4.1 Power function in linear time for nonnegative exponents.

1 def power_linear(b, n):
2 if n == 0:
3 return 1
4 else:
5 return b * power_linear(b, n - 1)
where f (b, n) = bn is naturally a function of the two parameters b and n.

It is apparent that if we assume to know the solution to f (b, n−1) = bn−1 ,
then all we need to do in the recursive case is multiply it times b in
order to obtain bn . Together with the base case, the function can be
implemented as shown in Listing 4.1.
It is an example of linear recursion since there is only one function
call in the recursive case, and the recursive call is not the last opera-
tion that the algorithm performs in the recursive case. Even though the
function call is the last element in the expression (in line 5), the method
first evaluates it, and then it processes it, multiplying the result times
b. Thus, the function is not tail-recursive. It is also important to note
that the function does not need a base case when n = 1, since it would
be redundant. Finally, the computational complexity of this algorithm
is linear with respect to n. In particular, the runtime can be defined
through:
⎧
⎪
⎪1 if n = 0,
T (n) = ⎨
⎪
(4.1)
⎪
⎩ T (n − 1) + 1 if n > 0,
where T (n) = n + 1 ∈ Θ(n).
If the problem statement allowed negative integer exponents as well,
the size of the problem would be the absolute value of n. In this variant
the base case also occurs when n = 0, and the recursive case is identical
if n > 0. When n is negative the decomposition needs to reduce the size
of the problem, ensuring that the corresponding subproblem gets closer
to the base case. Thus, for this recursive case we need to increment n.
The diagram that illustrates the thought process would be:
Listing 4.2 Power function in linear time.

1 def power_general_linear(b, n):
2 if n == 0:
3 return 1
4 elif n > 0:
5 return b * power_general_linear(b, n - 1)
6 else:
7 return power_general_linear(b, n + 1) / b
Inputs Results
f
/b
f
(b, n + 1) ÐÐÐÐÐÐÐÐÐÐ→ bn+1
Thus, the result of the subproblem must be divided by b in order to

obtain the desired power bn . The recursive case is therefore f (b, n) =
f (b, n + 1)/b, as shown in Listing 4.2.
4.1.1.2 Powers in logarithmic time

It is possible to design more efficient algorithms that calculate powers,
which run in logarithmic time, by considering a subproblem of half the
size in the decomposition stage. Assuming that n is a nonnegative inte-
ger, we need two diagrams in order to account for the parity of n. If n is
even, the recursive design thought process can be visualized as follows:
Inputs Results
f
( )2
f
(b, n/2) ÐÐÐÐÐÐÐÐÐÐ→ bn/2
Thus, the result of the subproblem must be squared in order to obtain bn .

Instead, if n is odd, we can consider using a subproblem of size (n − 1)/2
(which is an integer). The diagram would be:
Listing 4.3 Power function in logarithmic time for nonnegative exponents.

1 def power_logarithmic(b, n):
2 if n == 0:
3 return 1
4 elif n % 2 == 0:
5 return power_logarithmic(b, n // 2)**2
6 else:
7 return b * (power_logarithmic(b, (n - 1) // 2)**2)
Inefficient implementation of the power function that runs in

Listing 4.4
linear time.
1 def power_alt(b, n):
2 if n == 0:
3 return 1
4 elif n % 2 == 0:
5 return power_alt(b, n // 2) * power_alt(b, n // 2)
6 else:
7 return (power_alt(b, (n - 1) // 2)
8 * power_alt(b, (n - 1) // 2) * b)
Inputs Results
f
( )2 and × b
f
(b, (n − 1)/2) ÐÐÐÐÐÐÐÐÐÐ→ b(n−1)/2
In this case, the output of the subproblem also needs to be squared,

but a final multiplication times the base b is also required. With these
recursive cases, and the trivial base case, the recursive function is:
⎧
⎪
⎪
⎪ 1 if n = 0,
⎪
⎪ n/2 2
b = ⎨(b )
n
⎪
⎪
⎪
⎪
⎩(b
⎪ ) ⋅ b if n > 0 and n is odd.
(n−1)/2 2
A proper implementation of the function is shown in Listing 4.3, which

again uses linear recursion. The use of the integer division (//) instead
of a real division / is optional, but included in order to emphasize
that the second input parameter is an integer. In addition, when n is
odd (n-1)//2 is equivalent to n//2. The code uses the former expres-
sion since it resembles the mathematical definition of the function more
closely.
Its runtime can be defined through:
⎧
⎪
⎪1 if n = 0,
T (n) = ⎨
⎪
⎩T (n/2) + 1 if n > 0,
⎪
where T (n) = 2 + log2 n for n > 0. Thus, T (n) ∈ Θ(log n). The superior
performance stems from dividing the size of the problem by two in the
decomposition stage. However, the function must make a single recursive
call in each recursive case. For example, the code in Listing 4.4 does
not run in logarithmic time even though the decomposition divides the
problem size by two. The issue is that it calculates the result of the
same subproblem twice by using two identical recursive calls, which is
obviously unnecessary. The runtime cost for this function is:
⎧
⎪
⎪1 if n = 0,
T (n) = ⎨
⎪
⎩2T (n/2) + 1 if n > 0,
⎪
where T (n) = n + 1 ∈ Θ(n). Notice that it is analogous to the runtime

of the function in Listing 4.1, whose runtime is described through (4.1).
Thus, the extra performance that could have been gained by halving the
size of the input is lost by making two identical recursive calls in the
recursive cases.
4.1.2 Slow addition

This problem consists of adding two nonnegative integers a and b, where
we cannot use the general arithmetic commands such as addition, sub-
traction, multiplication, or division. Instead, we are only allowed to add
or subtract single units from the numbers (a constant amount of times).
This simple problem will help illustrate how the choice of the size can af-
fect the way we decompose a problem, which leads to different recursive
algorithms.
In consonance with the template in Figure 2.1, the first step consists
of determining the size of the problem. Since we are only allowed to add
or subtract 1 from a and b, it is reasonable to think that we will have to
perform these simple operations a + b times until a base case condition is
met (we will see shortly that it can be solved in only a or b operations).
If we choose a + b to be the size of the problem, then a base case occurs
Listing 4.5 Slow addition of two nonnegative integers.

1 def slow_addition(a, b):
2 if a == 0:
3 return b
4 elif b == 0:
5 return a
6 else:
7 return slow_addition(a - 1, b) + 1
when a = b = 0, which corresponds to the instance of smallest size. In

addition, we have to consider other possible base cases where we can
provide a result easily, without the need to carry out a recursive call.
For this problem, when a = 0 the result is obviously b, and when b = 0
the output should be a. Furthermore, with these two base cases it is not
necessary to include the one for a = 0 and b = 0. Lastly, the base cases
guarantee that both a and b will be positive in the recursive case.
We could proceed in the decomposition step by reducing the size of
the problem by a unit. Since our choice for the size of the problem is
a + b, we could subtract a unit from either a or b. The recursive diagram
related to decrementing a would be:
Inputs Results
f
(a, b) ÐÐÐÐÐÐÐÐÐÐ→ a+b
+1
f
(a − 1, b) ÐÐÐÐÐÐÐÐÐÐ→ a−1+b
If the function we are implementing is f , then clearly f (a, b) = f (a−1, b)+

1 defines the recursive case, and it can be expressed mathematically as:
⎧
⎪
⎪
⎪
⎪
b if a = 0,
f (a, b) = ⎨a
⎪
if b = 0, (4.2)
⎪
⎪
⎪
⎩f (a − 1, b) + 1 if a > 0, b > 0.
Listing 4.5 shows an implementation of the function in Python. Had

we chosen a to be the size of the problem the two base cases would
also appear in the definition of the function, and we would have also
decremented the value of a when decomposing the problem. Therefore,
Listing 4.6 Quicker slow addition of two nonnegative integers.

1 def quicker_slow_addition(a, b):
2 if a == 0:
3 return b
4 elif b == 0:
5 return a
6 elif a < b:
7 return quicker_slow_addition(a - 1, b) + 1
8 else:
9 return quicker_slow_addition(a, b - 1) + 1
the resulting function would be identical. Alternatively, choosing b as

the size of the problem would imply replacing the recursive case by
f (a, b) = f (a, b − 1) + 1, since we would have decremented b instead of a.
The function in Listing 4.5 can be slow if a is large. Alternatively, we
can build a more efficient algorithm by considering that the size of the
problem is the smallest input parameter, i.e., min(a, b). In this scenario
the two base cases f (0, b) = b, and f (a, 0) = a, correspond to the smallest
instances of the problem (f (0, 0) would be redundant). For the recursive
cases we must find a decomposition that guarantees reducing the size of
the problem. A first approach consists of decrementing the smallest input
parameter, where the size of the subproblem would be min(a, b) − 1. For
example, if a < b the recursive rule would be f (a, b) = f (a−1, b)+1, while
if a ≥ b we would apply f (a, b) = f (a, b − 1) + 1. Thus, the mathematical
function would be:
⎧
⎪
⎪
⎪
⎪
b if a = 0,
⎪
⎪
⎪a if b = 0 (and a ≠ 0),
f (a, b) = ⎨
⎪
⎪ f (a − 1, b) + 1
⎪
⎪
⎪
if a < b (and a ≠ 0, b ≠ 0),
⎪
⎪
⎩f (a, b − 1) + 1 if b ≤ a (and a ≠ 0, b ≠ 0),
which can be coded as shown in Listing 4.6.

A second idea consists in decrementing both parameters, which also
ensures reducing the size of the problem. In particular, note that min(a−
1, b − 1) = min(a, b) − 1. For this decomposition the recursive diagram
would be:
Listing 4.7 Alternative quicker slow addition of two nonnegative integers.

1 def quicker_slow_addition_alt(a, b):
2 if a == 0:
3 return b
4 elif b == 0:
5 return a
6 else:
7 return quicker_slow_addition_alt(a - 1, b - 1) + 1 + 1
Inputs Results
f
(a, b) ÐÐÐÐÐÐÐÐÐÐ→ a+b
+1 + 1
f
(a − 1, b − 1) ÐÐÐÐÐÐÐÐÐÐ→ a−1+b−1
The recursive rule is therefore: f (a, b) = f (a − 1, b − 1) + 1 + 1, for a > 0

and b > 0. According to the problem statement, this is valid if we assume
that we can perform increments a constant number of times (in this
case, f (a − 1, b − 1) is incremented twice), which does not depend on the
inputs. Listing 4.7 shows the associated code.
4.1.3 Double sum

The next example involves calculating the particular double sum:
f (m, n) = ∑ ∑(i + j).

m n
(4.3)
i=1 j=1
By using the properties of sums it can be simplified to a simple expres-

sion:
n(n + 1) nm(m + 1) mn(n + 1)
f (m, n) = ∑ ∑ (i + j) = ∑ (in + )=
m n m
+
i=1 j=1 i=1 2 2 2
mn(m + n + 2)
= .
2
However, we will develop a recursive solution that uses two recursive
functions (one for each sum).
Firstly, the outer sum is a function of the parameters m and n, whose
size is m. The function returns 0 in its base case when m is smaller than
the lower index (i.e., when m ≤ 0). The general diagram for the recursive
case, assuming that the size of the subproblem is m − 1, is:
Inputs Results
n n n
f
(m, n) ÐÐÐÐÐ→ ∑ (1 + j) + ⋯ + ∑ ((m − 1) + j) + ∑ (m + j)
j=1 j=1 j=1
n
+ ∑ (m + j)
j=1
n n
f
(m − 1, n) ÐÐÐÐÐ→ ∑ (1 + j) + ⋯ + ∑ ((m − 1) + j)
j=1 j=1
Clearly, all we have to do to the result of f (m − 1, n), in order to obtain

f (m, n), is add the inner sum ∑nj=1 (m + j), which is a function g that
depends on n and m. Therefore, we can define f as follows:
⎧
⎪
⎪0 if m ≤ 0,
f (m, n) = ⎨
⎪
⎩f (m − 1, n) + g(n, m) if m > 0.
⎪
We can also define g(n, m) recursively. Firstly, we need to determine its

size, which is n, since it adds n terms. In accordance with other sums, it
returns 0 in the base case, which occurs when n ≤ 0. The general diagram
for its recursive case, again decrementing the size of the problem by a
unit, is:
Inputs Results
g
(n, m) ÐÐÐÐÐ→ (m + 1) + ⋯ + (m + (n − 1)) + (m + n)
+ (m + n)
g
(n − 1, m) ÐÐÐÐÐ→ (m + 1) + ⋯ + (m + (n − 1))
In this case, we need to add (m + n) to the result of the subproblem in

order to obtain g(n, m). Thus, the function is:
⎧
⎪
⎪0 if n ≤ 0,
g(n, m) = ⎨
⎪
⎩g(n − 1, m) + (m + n) if n > 0.
⎪
Finally, Listing 4.8 shows an implementation of both sums.

Listing 4.8 Recursive functions that compute the double sum in (4.3).
1 def inner_sum(n, m):
2 if n <= 0:
3 return 0
4 else:
5 return inner_sum(n - 1, m) + (m + n)
6
7
8 def outer_sum(m, n):
9 if m <= 0:
10 return 0
11 else:
12 return outer_sum(m - 1, n) + inner_sum(n, m)
4.2 BASE CONVERSION

Numbers can be represented in different ways, depending on a particular
base or radix. Typically, we use a base-10 numeral system where a se-
quence of digits, say 142, represents the number 1 ⋅ 102 + 4 ⋅ 101 + 2 ⋅ 100 . In
general, for a particular base b, the value of a number x can be expressed
as a unique sequence of m digits dm−1 ⋯d0 , where:
m−1
x = ∑ di bi , (4.4)
i=0
with 0 ≤ di < b, and dm−1 ≠ 0 (i.e., we omit writing leading zeros). There-
fore, different bases lead to distinct sequences of digits that represent the
same number. Regarding notation, the base can be specified through a
subscript, which we usually omit when it is 10. For example, 14210 = 142,
but 1425 = 1 ⋅ 52 + 4 ⋅ 51 + 2 ⋅ 50 = 25 + 20 + 2 = 47. In this section we will
examine algorithms for converting numbers expressed in some base to
another one.
4.2.1 Binary representation of a nonnegative integer

In this example the goal consists of developing a recursive function that,
given a particular nonnegative integer n (in base 10), returns a new in-
teger that contains the binary representation of n, which is the sequence
of base-2 digits (i.e., bits) associated with the value of n. The output
will also be a number in base 10, but its digits will only be either zero or
one. Thus, we can interpret that the digits actually correspond to bits.
Since we need to create the sequence of “bits,” the size of the problem
is the number of bits in the binary representation of n. Mathematically,
this quantity is ⌊log2 n⌋ + 1 (for n > 0). However, we do not need the
formula in order to design the recursive algorithm. All we require is a
clear definition of the size of the problem that will enable us to define base
cases and decompose the problem. In particular, the smallest instances
of the problem correspond to numbers that contain a single bit, which
are 0 and 1. Thus, the base case occurs when n < 2, where the output is
simply n.
For the recursive case we need to decide how we can reduce the size
of the problem. The simplest way consists of decrementing the number
of bits by a single unit, which is accomplished by performing an integer
division of n by two (this shifts the bits one place to the right, where the
least significant bit is discarded). We can start the analysis with concrete
instances. For example:
Inputs Results
f
n = 18 ÐÐÐÐÐÐÐÐÐÐ→ 10010
× 10 + 0
f
n//2 = 9 ÐÐÐÐÐÐÐÐÐÐ→ 1001
and
Inputs Results
f
n = 19 ÐÐÐÐÐÐÐÐÐÐ→ 10011
× 10 + 1
f
n//2 = 9 ÐÐÐÐÐÐÐÐÐÐ→ 1001
Note that the output of f (18) is ten thousand and ten, since it is ex-
pressed in base 10. The diagrams illustrate that we can obtain the de-
sired result by multiplying the output of the subproblem by 10, and
then adding 1 if n is odd. Therefore, the recursive case seems to be
f (n) = 10f (n//2) + n%2.
We can proceed more rigorously through the following general dia-
gram of the recursive thought process:
Listing 4.9 Binary representation of a nonnegative integer.

1 def decimal_to_binary(n):
2 if n < 2:
3 return n
4 else:
5 return 10 * decimal_to_binary(n // 2) + (n % 2)
Inputs Results
f
n = (bm−1 ⋯ b1 b0 )2 ÐÐÐÐÐÐÐÐÐÐ→ (bm−1 ⋯ b1 b0 )10
× 10 + b0
f
n//2 = (bm−1 ⋯ b1 )2 ÐÐÐÐÐÐÐÐÐÐ→ (bm−1 ⋯ b1 )10
where bi represents a bit. In particular, f (n//2) is the decimal number

that represents the sequence of “bits” of n except for the least signifi-
cant one (b0 ), whose value is n%2 (or n&1, where & denotes the bitwise
AND operator). In order to append the discarded bit we need to mul-
tiply f (n//2) by 10 (which shifts the digits one place to the left, and
appends a zero), and add it to the result. Therefore, the recursive case is
indeed f (n) = 10f (n//2)+n%2, which we obtained by analyzing concrete
instances. The complete function is:
⎧
⎪
⎪n if n < 2,
f (n) = ⎨
⎪
⎩10f (n//2) + n%2 if n ≥ 2.
⎪
Listing 4.9 shows the corresponding code. Finally, the binary represen-
tation of 142 = 14210 is 100011102 = 128 + 8 + 4 + 2.
4.2.2 Decimal to base b conversion

The problem in Section 4.2.1 is a particular case of the more general
problem that consists of converting a decimal number into another in
another base b, where b ≥ 2. It can be tackled in a similar way by using
diagrams. Figure 4.1 shows the steps of a general algorithm that solves
the problem. In particular, it illustrates how to obtain the base-5 rep-
resentation of 142, which is 1032 (i.e., 14210 = 10325 ). The full problem
is shown in (a), where the idea consists of performing successive integer
divisions of n by b until reaching 0. At each step the remainder of the
//5 //5 //5 //5

142 28 5 1 0
%5 %5 %5 %5
2 3 0 1
(a)
//5 //5 //5 //5

142 28 5 1 0
%5 %5 %5 %5
2 3 0 1
(b)
Figure 4.1 Conversion of 142 into its representation (1032) in base 5.
Listing 4.10 Conversion of a nonnegative integer n into its representation

in base b.
1 def decimal_to_base(n, b):
2 if n < b:
3 return n
4 else:
5 return 10 * decimal_to_base(n // b, b) + (n % b)
division by b is also calculated, and will constitute a digit in the new

representation in base b, whose position is controlled by a variable asso-
ciated with a power of 10. The diagram in (b) highlights a subproblem
within the original, where 28 = 1035 .
In order to develop a recursive solution we need to establish the size
of the problem. For this problem it is the number of times it is necessary
to divide n until reaching 0, which is essentially the number of digits of
the input in base b. The base case occurs when n can be represented by
a single digit in base b. In other words, when n < b, where the result is
simply n.
The recursive case requires identifying a subproblem within the orig-
inal, as shown in Figure 4.1(b). Note that the subproblem is self-similar
to the original, but starts at 28, which is 142//5. Thus, the decomposi-
tion consists of performing the integer division n//b, which reduces the
size of the problem by a unit. Subsequently, we must determine how
it is possible to modify the solution to the subproblem (103), in order
to obtain the original (1032). The solution consists of multiplying the
result of the subproblem times 10, and adding n%b (which is 2 in the
example). Thus, the function can be coded as shown in Listing 4.10.
4.3 STRINGS
This section analyzes two problems involving strings, which are essen-
tially sequences of characters, and constitute a fundamental data type
in many programming languages.
4.3.1 Reversing a string

Consider the problem of reversing a string. In particular, we will develop
a function f that receives an input string and returns its reverse. For
example, f ('abcd') ='dcba'.
The size of the problem is the length of the input string. A base case
occurs when the input string is empty, where the function obviously
returns an empty string. In the recursive case we need to discard a
character from the input string in order to reduce the size of the problem.
The first and last characters of the string are clear candidates. Omitting
the first one leads to the following diagram where the input string s is
written as the sequence s0 s1 ⋯ sn−2 sn−1 of characters:
Inputs Results
s0 s1 ⋯ sn−2 sn−1 ÐÐÐÐÐÐÐÐÐ→ sn−1 sn−2 ⋯ s1 s0
+ s0
s1 ⋯ sn−2 sn−1 ÐÐÐÐÐÐÐÐÐ→ sn−1 sn−2 ⋯ s1
Thus, the function simply has to concatenate the first character to the
result of the subproblem associated with s1 ⋯ sn−2 sn−1 (the + symbol
represents string concatenation). Together with the base case, the recur-
sive function in Python is shown in Listing 4.11.
Listing 4.11 Conversion of a nonnegative integer n into its representation

in base b.
1 def reverse_string(s):
2 if s == '':
3 return ''
4 else:
5 return reverse_string(s[1:]) + s[0]
4.3.2 Is a string a palindrome?

The next problem consists of determining whether a string is a palin-
drome, which is a sequence of characters that reads the same forward
as it does backward (e.g., “radar”). Its size is the length of the string,
since it determines the number of operations needed in order to obtain
a true result. There are two base cases for this problem: (a) when the
string is empty, and (b) when it comprises a single character. In both
situations the output is True. Similarly to the function in Listing 2.6,
the second base case is necessary since we will decompose the problem
by reducing its size by two units. In particular, we will consider the sub-
problem that discards the first and last characters of the original string
of length n ≥ 2. In addition, the output of the function for some string
s = s0 s1 ⋯ sn−2 sn−1 of length n is:
(s0 = sn−1 ) ∧ (s1 = sn−2 ) ∧ ⋯ ∧ (s⌊ n2 ⌋−1 = sn−⌊ n2 ⌋ ),
where ∧ denotes logical AND (conjunction), and si represents the char-

acter from s at position i (where the first character of the string is located
at position 0). Observe that the indices of the characters that need to be
compared add up to n − 1. Furthermore, we only need to compare ⌊n/2⌋
pairs of characters. The expression can be written more compactly as:
⌊n/2⌋−1
⋀ (si = sn−i−1 )
i=0
similarly to “sigma” notation. The corresponding recursive diagram is:

Listing 4.12 Function that determines if a string is a palindrome.

1 def is_palindrome(s):
2 n = len(s)
3 if n <= 1:
4 return True
5 else:
6 return (s[0] == s[n - 1]) and is_palindrome(s[1:n - 1])
Inputs Results
⌊n/2⌋−1
f
s0 s1 ⋯ sn−2 sn−1 ÐÐÐÐÐÐÐÐÐÐ→ ⋀ (si = sn−i−1 )
i=0
∧ (s0 = sn−1 )
⌊n/2⌋−1
f
s1 ⋯ sn−2 ÐÐÐÐÐÐÐÐÐÐ→ ⋀ (si = sn−i−1 )
i=1
Thus, the Boolean function can be defined as:

⎧
⎪
⎪true if n < 2,
f (s) = ⎨
⎪
⎩(s0 = sn−1 ) ∧ f (s1..n−2 ) if n ≥ 2,
⎪
where s1..n−2 simply denotes the substring s1 ⋯ sn−2 . Finally, Listing 4.12
shows the corresponding code, whose time complexity is Θ(n).
4.4 ADDITIONAL PROBLEMS

This section contains several classical problems that can be solved ele-
gantly through recursion.
4.4.1 Selection sort

The selection sort algorithm is one of the simplest strategies for sorting
the elements of a list. Given an input list a of n numbers, assume the
method must sort it in ascending order. The algorithm begins by placing
the smallest element of the list in the first position of the output list. In
the second step, the method searches for the minimum element in a1..n−1 ,
and places it in the second position of the output list. This procedure is
carried out n − 1 times until the array is finally sorted. Clearly, at the
end of the k-th step the smallest k numbers are sorted correctly, and
after n − 1 the entire list is completely sorted.
We will now show a linear-recursive version of the algorithm. Firstly,

the size of the problem is the length of the list (n). The base case occurs
when the list contains one element, where the output is the input list
since it is obviously sorted. In addition, the algorithm can also return an
empty list if the input is empty. Thus, the method can return a if n ≤ 1.
For the recursive case we can decompose the problem by reducing
its size by a unit, where the main idea is to discard the smallest element
from the list. This can be done in two ways. Firstly, the element can be
swapped with the first one on the list. This naturally does not change
the result (the sorted list), but allows us to discard the first element of
the list after having swapped the elements. Therefore, if the input list
is [7, 5, 3, 8, 4], it can be modified to [3, 5, 7, 8, 4] by swapping the 3 and
the 7. In that case, the input to the subproblem of size n − 1 would be
[5, 7, 8, 4], and a recursive diagram with this particular example could
be:
Inputs Results
[7, 5, 3, 8, 4] ÐÐÐÐÐÐÐÐÐ→ [3, 4, 5, 7, 8]
concatenate [3]
[5, 7, 8, 4] ÐÐÐÐÐÐÐÐÐ→ [4, 5, 7, 8]
In order to solve the original problem, the recursive rule must concate-
nate the discarded smallest element ([3]) with the output list of the
subproblem (note that the results are sorted lists). Listing 4.13 shows
an implementation of the method that uses the function min to deter-
mine the smallest value in a list, the method index that returns the
location of an element in a list, and the + operator to concatenate lists.
An important detail regarding this algorithm is that it needs to make a
new copy of the input list (in line 5) in order to not alter it when calling
the method. In particular, without this copy the method would swap the
first and smallest elements of the list.
Another possibility for reducing the size of the problem consists of
discarding the smallest element of the list directly, by calling the remove
method. In this case, the recursive diagram would be:
Listing 4.13 Recursive selection sort algorithm.

1 def select_sort_rec(a):
2 if len(a) <= 1:
3 return a
4 else:
5 b = list(a)
6 min_index = b.index(min(b))
7 aux = b[min_index]
8 b[min_index] = b[0]
9 b[0] = aux
10
11 return [aux] + select_sort_rec(b[1:])
Listing 4.14 Recursive variant of the selection sort algorithm.

1 def select_sort_rec_alt(a):
2 if len(a) <= 1:
3 return a
4 else:
5 b = list(a)
6 m = min(b)
7 b.remove (m)
8
9 return [m] + select_sort_rec_alt(b)
Inputs Results
[7, 5, 3, 8, 4] ÐÐÐÐÐÐÐÐÐ→ [3, 4, 5, 7, 8]
concatenate [3]
[7, 5, 8, 4] ÐÐÐÐÐÐÐÐÐ→ [4, 5, 7, 8]
The only difference with respect to the previous diagram is the order of
the elements in the input to the subproblem. This does not affect the
recursive rule, where we must also concatenate the discarded smallest
element ([3]) with the output list of the subproblem. Listing 4.14 shows
a possible implementation of the function, which relies on the remove
method. Again, the fifth line makes a copy of the input list in order to
keep it unaltered when calling the method.
Finally, the runtime cost of these algorithms is characterized by

(3.24), since finding (and removing) the smallest element of a list re-
quires on the order of n operations. Thus, the algorithms run in Θ(n2 ).
4.4.2 Horner’s method for evaluating polynomials

In this problem the goal consists of evaluating a polynomial of degree n:
P (x) = cn xn + cn−1 xn−1 + ⋯ + c1 x + c0 , (4.5)
at some real value x. The sum contains powers of x that are multiplied
by the coefficients ci . A naive algorithm that computes each power in-
dependently would require on the order of n2 multiplications. Instead,
Horner’s method only needs to perform n products. Its clever idea is
based on expressing the polynomial as:
P (x) = c0 + x(c1 + x(c2 + ⋯ + x(cn−1 + cn x))).
The size of the problem is clearly the degree n. Thus, the base case
occurs when n = 0, where the result is obviously c0 . In practice, c will be
a list (or a similar data structure such as an array) of n+1 elements that
represents the polynomial. Therefore, the base case is reached when the
length of c is one.
In order to apply recursion we need to detect a self-similar subprob-
lem of smaller size. The decomposition of the problem, decrementing its
size by a unit, is:
³¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹· ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ µ
subproblem
P (x) = c0 + x(c1 + x(c2 + ⋯ + x(cn−1 + cn x))⋯) .

´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
full problem
By discarding coefficient c0 and the first multiplication times x, the re-

sulting subproblem has exactly the same structure as the original prob-
lem. If p(c, x) is the function that evaluates the polynomial defined by
c at x, the recursive formula would be:
p(c, x) = c0 + x ⋅ p(c1..n , x),
where the list c1..n is simply c without its first element. The full function
⎧
is:
⎪
⎪c0 if n = 1,
p(c, x) = ⎨
⎪
⎩c0 + x ⋅ p(c1..n , x) if n > 1,
⎪
Listing 4.15 Horner’s method for evaluating a polynomial.

1 def horner (c, x):
2 if len(c) == 1:
3 return c[0]
4 else:
5 return c[0] + x * horner (c[1:], x)
and the corresponding code is shown in Listing 4.15.

It is important to note that discarding cn instead of c0 does not lead
to Horner’s method. In that case the recursive formula would be:
p(c, x) = p(c0..n−1 , x) + cn xn ,
which would evaluate the polynomial as in (4.5). If the power xn is

computed in linear or logarithmic time the method would require Θ(n2 )
or Θ(n log n) operations, respectively. However, Horner’s method is more
efficient since it runs in linear time (Θ(n)).
4.4.3 A row of Pascal’s triangle

Pascal’s triangle is the triangular arrangement of binomial coefficients
shown in Figure 4.2, where the triangle in (a) shows the particular bi-
nomial coefficients, and the one in (b) shows the actual integer values
associated with them. The following problem consists of determining a
list containing the binomial coefficients of the n-th row:
( ) ( ) ⋯ ( ) ( ),
n n n n
0 1 n−1 n
where the first and last elements are always 1. The problem can be
viewed from a recursive point of view by considering the definition in
3.2. Graphically, it means that a binomial coefficient is the sum of the
two immediately above it in the previous row of Pascal’s triangle. For
example, (43) = (32) + (33). Therefore, it is possible to define a row of
Pascal’s triangle if we know the values of the previous row. Figure 4.3
shows the relationship between a problem (n-th row) and a subproblem
((n − 1)-th row).
The size of the problem is n. Thus, the base case corresponds to
n = 0, where the output is simply a list containing a 1 ([1]). For n = 1
we cannot apply the recursive rule in 3.2. Thus, initially it appears that
we may need an additional base case for n = 1, where the output is
(00) 1
(10) (11) 1 1
(20) (21) (22) 1 2 1
(30) (31) (32) (33) 1 3 3 1
(40) (41) (42) (43) (44) 1 4 6 4 1
(a) (b)
Figure 4.2 Pascal’s triangle.
1 (n−1
1
) (n−1
2
) ⋯ (n−3
n−1
) (n−2
n−1
) 1
+ + + +
1 (n1 ) (n2 ) ⋯ (n−2
n
) (n−1
n
) 1
Figure 4.3 Decomposition and recovery of a row of Pascal’s triangle.
Listing 4.16 Function that generates the n-th row of Pascal’s triangle.
1 def pascal (n):
2 if n == 0:
3 return [1]
4 else:
5 row = [1]
6 previous_row = pascal (n - 1)
7 for i in range(len(previous_row) - 1):
8 row.append (previous_row[i] + previous_row[i + 1])
9 row.append (1)
10 return row
([1,1]). However, since all of the rows for n > 0 begin and end with a
1, these elements can be incorporated in the recursive case by default.
Thus, in this scenario a special base case for n = 1 would be unnecessary.
In particular, in the recursive case we assume that we know the
solution to the subproblem, which allows us to compute every element
PSfrag
r r r
⋯ A A
r r r r r R
1 2 3
⋯ n−1 n
B B
Figure 4.4 Ladder of resistors problem.

r1 r2 r1 + r2
A B A B
(a)
r1 r1 ⋅r2
r1 +r2
A B A B
r2 (b)
Figure 4.5 Equivalence of circuits with resistors.
of the solution except for the ones at the extremes (these elements can
be appended at some point during the process). Listing 4.16 shows a
possible solution to the problem. The recursive case starts by inserting
a 1 in a list (row) that will contain the result. Line 6 computes the
result of the subproblem of size n − 1, and in lines 7 and 8 the loop
adds consecutive integers of the subsolution (as shown in Figure 4.3),
appending the sum to the result. Finally, a 1 is inserted at the end,
which completes the solution. Exercise 4.14 proposes replacing the loop
by a recursive function.
4.4.4 Ladder of resistors

In the next problem the goal is to simplify the electrical circuit in Fig-
ure 4.4 resembling a ladder that contains several layers of resistors whose
resistance is r. In particular, it is possible to substitute the entire circuit
by an equivalent one composed of a single resistor. Thus, the objective
consists of determining the value of the resistance R (as a function of
the resistance r) that will yield equivalent circuits. Figure 4.5 shows how
to transform a circuit with two resistors with resistances r1 and r2 when
they are connected in series, as shown in (a), or in parallel, as illus-
trated in (b). When they are connected in series the resulting resistance
r r r
⋯ A
r r r r r
1 2 3
⋯ n−1 n
B
(a)
r
A A
R(r, n − 1) r R(r, n)
B B
(b)
Figure 4.6Decomposition of the ladder of resistors problem, and deriva-

tion of the recursive case through induction.
is r1 +r2 . Instead, when they are connected in parallel the new resistance
is r = (r1 ⋅ r2 )/(r1 + r2 ). Alternatively, it can be expressed as:
1 1 1
= + . (4.6)
r r1 r2
These rules can be applied successively to pairs of resistors until we
obtain a circuit that contains a single resistor. However, this process is
tedious. Instead, recursion provides a succinct and elegant solution.
The problem has two input parameters: the resistance (r) and the
number of rungs (n) of the ladder. The size of the problem is clearly
n (r plays a role in the final value, but is not responsible for the run-
time of the algorithm). Let R(r, n) denote the recursive function. The
base case occurs when n = 1, where the initial ladder would only contain
a single resistor. Therefore, in that case R(r, 1) = r. For the recursive
case we need to find a subproblem within the original with exactly the
same structure. Figure 4.6(a) shows the decomposition of the problem
by decrementing its size by a unit. The circuit associated with the sub-
problem can be replaced by a single resistor of resistance R(r, n − 1),
Listing 4.17 Function that solves the ladder of resistors problem.

1 def circuit (n, r):
2 if n == 1:
3 return r
4 else:
5 return 1 / (1 / r + 1 / (circuit (n - 1, r) + r))
as shown in (b), where we can assume that we know its value by using
induction. Finally, it is fairly straightforward to simplify the resulting
circuit that contains only three resistors. Firstly, the left and top resis-
tors are connected in series. Thus, they can be merged to form a resistor
with resistance R(r, n − 1) + r. Finally, this new resistor is connected in
parallel to the right resistor. By applying (4.6), R(r, n) can be defined
through:
1 1 1
= +
R(r, n) r R(r, n − 1) + r
.
The recursive function is therefore:

⎧
⎪r
⎪ if n = 1,
R(r, n) = ⎨
⎪
⎪ 1/ ( r + R(r,n−1)+r ) if n > 1.
⎩
1 1
Listing 4.17 shows the associated code. Finally, as a curious note, it

is possible to show that R(r, n) = rF (2n − 1)/F (2n), where F is the
Fibonacci function.
4.5 EXERCISES
Exercise 4.1 — Listing 2.6 contains a linear-recursive Boolean func-
tion that determines whether a nonnegative integer n is even, where the
decomposition reduces the size of the problem (n) by two units. Define
and code an alternative method based on decrementing the size of the
problem by a single unit.
Exercise 4.2 — Implement a recursive function that computes the

power bn in logarithmic time, for a real base b, and an integer n, which
can be negative.
Exercise 4.3 — Implement a recursive function that computes the n-

th power of a square matrix, which is another square matrix of the same
dimensions. Use the NumPy package and its methods:
Figure 4.7The product of two nonnegative integers n and m can be rep-

resented as the number of unit squares that form an n × m rectangle.
• identity: returns the identity matrix.

• shape: indicates the dimensions of a matrix.
• dot: multiplies matrices.
Finally, compute matrices:
n
[ ]
1 1
1 0
for n = 1, . . . , 10, and print the element in their first row and second
column. What are these numbers?
Exercise 4.4 — Implement recursive functions that compute a “slow”

product of two nonnegative integers m and n. The algorithms are allowed
to add and subtract numbers, but they may not use the multiplication
(*) operator. Use decompositions that decrease one or both of the in-
puts by a single unit. In addition, illustrate the recursive design thought
process through diagrams (for example, by using the general diagram
in Figure 2.5; or through diagrams that represent the product of n and
m as the area of a rectangle of height n and width m, where the result
is the number of square blocks of unit area needed to build an n × m
rectangle, as shown in Figure 4.7).
Exercise 4.5 — Implement a more efficient “slow” product recursive

function by using the decomposition that divides both input parameters
by two. Use rectangular diagrams like the one described in Figure 4.7.
Exercise 4.6 — In Python and other programming languages methods

can also be parameters to other methods. Define and code a general
recursive function that computes the sum:
g(m, n, f ) = ∑ f (i) = f (m) + f (m + 1) + ⋯ + f (n − 1) + f (n),

n
i=m
where m and n are integers, and f is a function. Use it in order to

calculate and print:
n
3
∑i ,
i=1
for n = 0, . . . , 4.
Exercise 4.7 — Define and code a function that computes the number
of digits of a nonnegative integer n.
Exercise 4.8 — Define and code a function that, given a decimal num-
ber n whose digits are either zero or one, returns the number whose bi-
nary representation is precisely the sequence of zeros and ones in n. For
example, if n = 1011010 the function returns 22, since 101102 = 22.
Exercise 4.9 — Derive the recursive case of the “decimal to base b

conversion” problem in Section 4.2.2 by using the general diagram for
thinking about recursive cases in Figure 2.5.
Exercise 4.10 — Recall the problem in 3.8 that processes (binary)

numbers of n bits (from right to left). Derive and code a recursive algo-
rithm that, given a particular number x of n bits, determines the position
of its least significant bit set to 1. In particular, consider that the least-
significant bit is located at position 1. For example, for x = 01110100 the
rightmost bit set to 1 is located at position 3.
Exercise 4.11 — Write a recursive method that uses the function de-
veloped in Exercise 4.10 in order to solve Exercise 3.8 computationally.
Print the solutions for numbers expressed with n = 1, . . . , 5 bits.
Exercise 4.12 — Code a recursive function that returns the number

of vowels in a given string.
Exercise 4.13 — A binomial coefficient (m n

) is a function of two inte-
ger parameters n and m. Define three linear-recursive functions for cal-
culating binomial coefficients by considering three decompositions: one
that decrements n, another that decrements m, and one that decrements
both. Use the definition in (3.1) to develop the recursive cases.
Exercise 4.14 — Replace the loop in Listing 4.16 with a recursive func-
tion. It should receive a row of Pascal’s triangle, and return a list with the
sum of its consecutive terms. For example, if the input is [1, 3, 3, 1]
the result should be [4, 6, 4].
sorted unsorted
1 3 5 5 6 4 2 8 1 9 3 7 8
0 1 2 3 4 5 6 7 8 9 10 11 12
step 5
sorted unsorted
1 3 4 5 5 6 2 8 1 9 3 7 8
0 1 2 3 4 5 6 7 8 9 10 11 12
Figure 4.8 Step of the insertion sort algorithm.
Exercise 4.15 — Consider a nonnegative integer n whose digits always

appear in ascending order from left to right, such as 24667. In other
words, if dm−1 ⋯ d1 d0 represents the sequence of (m) digits of n, then
di ≤ dj for i < j. Given an additional digit 0 ≤ x ≤ 9, write a function
that returns the integer that results from inserting x in n, such that its
digits also appear in ascending order from left to right. For instance, if
n = 24667 and x = 5, the function should return 245667. Avoid redundant
base cases.
Exercise 4.16 — This exercise is similar to Exercise 4.15. Given a

sorted list of numbers in ascending order (a), and a number x, imple-
ment a function that inserts x in a location in a such that the resulting
list will remain sorted.
Exercise 4.17 — The “insertion sort” method sorts a list by repeat-

edly applying the procedure shown in Figure 4.8. At the i-th step the
elements from indices 0 to i − 1 will be sorted, while the rest of the list
will not be (necessarily) sorted. In order to proceed, the method inserts
the element at index i into the sorted sublist in a location where the
resulting list (from index 0 to i) will remain sorted. If this operation is
performed from index (step) 1 (or 0) until n − 1 the list will be sorted
after the last step. Implement a recursive variant of the method that
uses the solution to Exercise 4.16 to sort a list.
CHAPTER 5
Linear Recursion II: Tail

Recursion
You are only as beautiful as your last action.

— Stephen Richards
AIL recursion, also known as “final” recursion, is a form of lin-

T ear recursion in the sense that the recursive cases only invoke the
method once. However, in tail recursion a recursive call is the last ac-
tion performed by the method. For example, the function in (1.15) is
tail-recursive since it does not manipulate the result of the function call
in the recursive case. This implies that it will return a value directly
obtained in a base case. In addition, functions may require more param-
eters in order to store intermediate results that will be used in the base
cases for providing the final output. The tail-recursive methods seen so
far include Listings 2.3, 2.5, and 2.6.
In the previous chapter all of the methods were based on decrement-
ing the size of the problem by one or two units, except for the ones in
Section 4.1.1.2. In contrast, this chapter will present more algorithms
that divide the size of the problem by two. In these cases the recursive
calls will solve a single problem of half the size of the original’s, since the
methods only invoke themselves once. Some authors refer to this strategy
as “divide and conquer.” However, we will only use this term when the
algorithms need to solve several independent subproblems whose size is
a fraction of that of the original’s. These algorithms will be covered in
Chapter 6.
Finally, tail recursion is also special due to its relationship with iter-
ation. Chapter 11 will analyze this connection.
133
5.1 BOOLEAN FUNCTIONS

This introductory section examines how the choice of base cases can lead
to linear or tail-recursive algorithms for some Boolean functions.
5.1.1 Does a nonnegative integer contain a particular digit?

Given a nonnegative integer n, and a digit 0 ≤ d ≤ 9 (which is also an
integer), in this problem the goal consists of determining whether d is
present in n. The recursive method will therefore be a Boolean function
with parameters n and d. Additionally, we will assume that n does not
contain any leading zeros (for example, 358 cannot be written as 0358).
However, we will consider that the number zero contains the digit 0.
5.1.1.1 Linear-recursive algorithm

The first step consists of establishing the size of the problem, which is
the number (m) of digits of n. A base case occurs when n only contains a
single digit (i.e., when n < 10), where the result is simply whether n = d.
Let us assume for the time being that this is the only base case.
For the recursive case, it is cumbersome to express the value of a
digit in terms of n. Therefore, we can use an alternative notation, as
discussed in Section 2.5.3. In particular, let dm−1 ⋯ d1 d0 represent the
sequence of m digits of n in base 10. We can form the following diagram
when decrementing the size of the problem by a unit (i.e., using n//10,
which discards the least significant digit of n):
Inputs Results
(n = dm−1 ⋯ d1 d0 , d) ÐÐÐ→ (dm−1 = d) ∨ ⋯ ∨ (d1 = d) ∨ (d0 = d)
∨ (d0 = d)
(n//10 = dm−1 ⋯ d1 , d) ÐÐÐ→ (dm−1 = d) ∨ ⋯ ∨ (d1 = d)
where ∨ denotes logical OR (disjunction). The diagram clearly illustrates

that if d is a digit of n, then it is either present in n//10 (this is the result
of the recursive call), or it should be equal to d0 .
Other notations are also possible. We can simply choose words or
expressions, even if they are not technically correct logically or mathe-
matically. For instance,
Linear Recursion II: Tail Recursion 135
Listing 5.1Linear-recursive Boolean function that determines if a non-

negative integer contains a digit.
1 def contains_digit(n, d):
2 if n < 10:
3 return n == d
4 else:
5 return (n % 10 == d) or contains_digit(n // 10, d)
Inputs Results
(n, d) ÐÐÐÐÐÐÐÐÐ→ d∈n
∨ (n%10 = d)
(n//10, d) ÐÐÐÐÐÐÐÐÐ→ d ∈ (n//10)
introduces a clear abuse of notation, since the membership symbol ∈

should be used with a set. Thus, d ∈ n actually represents whether d
belongs to the set of digits that form n. Nevertheless, the diagram serves
its purpose, accurately indicating the function’s recursive case.
Listing 5.1 shows the code that implements the function, where n%10
is the least significant digit of n. Observe that the method is linear-
recursive, since the function invokes itself once, and needs to process the
output of the subproblem (i.e., compute the result of the expression
involving the or operator). Additionally, if the condition (n%10==d)
is True then a compiler or interpreter that uses “short-circuit” eval-
uation would automatically return a true value, avoiding the call to
contains_digit(n//10,d). In particular, short-circuit evaluation al-
lows us to avoid unnecessary computations by considering the proper-
ties: True ∨ b = True, and False ∧ b = False, where b is some Boolean
expression. Since these hold regardless of the value of b, there is no need
to evaluate b.
5.1.1.2 Tail-recursive algorithm

Most programming languages support short-circuit evaluation. However,
we can force it by considering an additional base case, which leads to a
tail-recursive algorithm. In particular, for this problem an algorithm can
return True as soon as it finds a digit in n that is equal to d. Therefore,
we can return True if the last digit of n is equal to d (i.e., if n%10 = d).
Listing 5.2Tail-recursive Boolean function that determines if a nonnega-

tive integer contains a digit.
1 def contains_digit_tail(n, d):
2 if n < 10:
3 return n == d
4 elif n % 10 == d:
5 return True
6 else:
7 return contains_digit_tail(n // 10, d)
By considering this base case we can be sure that n%10 = d0 ≠ d in the

recursive case. Therefore, the recursive diagram is now:
Inputs Results
(n = dm−1 ⋯ d1 d0 , d) ÐÐÐ→ (dm−1 = d) ∨ ⋯ ∨ (d1 = d)
do nothing
(n//10 = dm−1 ⋯ d1 , d) ÐÐÐ→ (dm−1 = d) ∨ ⋯ ∨ (d1 = d)
This implies that the result of the subproblem is exactly the output of
the original problem, and the function can simply return the result of
the recursive call, without processing it. Together with the base cases,
this leads to a tail-recursive algorithm that can be coded as shown in
Listing 5.2.
Lastly, it is important to understand that although the chosen de-
composition divides the input by a constant (10), the size of the problem
is only reduced by a unit. The time complexity for the algorithm is log-
arithmic with respect to the input n, but it is linear with respect to the
number of digits (m) of n.
Finally, this problem has analogous counterparts that rely on data
structures such as lists, arrays, etc., since these also represent sequences
of elements. For example, it is very similar to deciding if a string contains
a character, or if an element is present in a list. Although the codes may
be different, the underlying reasoning is essentially identical.
5.1.2 Equal strings?

The following problem consists of determining whether two strings are
equal (of course, in Python we can simply use the == operator). A
Boolean function that solves the problem will therefore have two string
input parameters. If their lengths are different the algorithm can return
False immediately in a base case. Thus, the challenge lies in solving
the problem when they have the same length, which would constitute
its size.
5.1.2.1 Linear-recursive algorithm

The smallest instance therefore corresponds to two empty strings, where
the result is obviously True. In this section we will assume that there
are no additional base cases, which leads to a linear-recursive function.
In particular, the decomposition can decrease the size of the problem
by a unit. This implies discarding one character (located at the same
position) from each input string. Omitting the first one leads to the
following recursive diagram, where s and t are two input strings of length
n:
Inputs Results
n−1
f
(s, t) ÐÐÐÐÐÐÐÐÐÐ→ ⋀ (si = ti)
i=0
∧ (s0 = t0 )
n−1
f
(s1..n−1 , t1..n−1 ) ÐÐÐÐÐÐÐÐÐÐ→ ⋀ (si = ti)
i=1
Clearly, the method has to check that the first characters are the same,
and that the remaining substrings are also identical through a recursive
call. The Boolean function is therefore:
⎧
⎪
⎪
⎪
false if length(s) ≠ length(t),
⎪
f (s, t) = ⎨true
⎪
if n = 0,
⎪
⎪
⎪
⎩(s0 = t0 ) ∧ f (s1..n−1 , t1..n−1 ) if n > 0,
Listing 5.3 shows the corresponding code, which describes a linear-

recursive function.
5.1.2.2 Tail-recursive algorithm

Similarly to the example in Section 5.1.1, we can force a short-circuit
evaluation by considering an additional base case. In particular, an al-
gorithm can return False as soon as it detects two different characters
Listing 5.3 Linear-recursive function that determines if two strings are

identical.
1 def equal_strings(s, t):
2 if len(s) != len(t):
3 return False
4 elif s == '':
5 return True
6 else:
7 return s[0] == t[0] and equal_strings(s[1:], t[1:])
Listing 5.4 Tail-recursive function that determines if two strings are iden-
tical.
1 def equal_strings_tail(s, t):
2 if len(s) != len(t):
3 return False
4 elif s == '':
5 return True
6 elif s[0] != t[0]:
7 return False
8 else:
9 return equal_strings_tail(s[1:], t[1:])
at the same position in the strings. Thus, we can incorporate a base case
that checks if s0 ≠ t0 . In that case, the algorithm would automatically
return False. With this base case we can be sure that s0 = t0 in the
recursive case. Therefore, the new recursive diagram would be:
Inputs Results
f n−1
(s, t) ÐÐÐÐÐÐÐÐÐÐ→ ⋀ (si = ti)
i=1
do nothing
n−1
f
(s1..n−1 , t1..n−1 ) ÐÐÐÐÐÐÐÐÐÐ→ ⋀ (si = ti)
i=1
Again, the result of the subproblem is exactly the output of the original
problem, which leads to a tail-recursive algorithm. Finally, together with
the base cases, the function can be coded as in Listing 5.4.
Listing 5.5 Tail-recursive linear search of an element in a list.

1 def linear_search_tail(a, x):
2 n = len(a)
3 if a == []:
4 return -1
5 elif a[n - 1] == x:
6 return n - 1
7 else:
8 return linear_search_tail(a[:n - 1], x)
5.2 SEARCHING ALGORITHMS FOR LISTS

This section will cover algorithms that solve the classical problem of
finding the position i of an element x within a list a = [a0 , a1 , . . . , an−1 ].
In other words, given a and x, the goal is to find an index i for which
ai = x. If the element x appears several times in the list the methods that
solve the problem can return any valid index that indicates a location
where x is found in the list.
For a list of length n, a method that solves the problem will return
an integer index ranging from 0 until n − 1, when the element is present
in the list. In contrast, if the list does not contain x the algorithms must
return some value outside of that interval. Some Python implementations
simply return a Boolean false value. However, this option is not valid in
many other languages (C, Java, Pascal, etc.), where methods can only
return values of a single data type. Since the indices are necessarily
integers, in these languages the value that indicates that an element is
not contained in the list must also be an integer. Thus, we will use the
integer value −1 to indicate that an element is not present in a list, which
is compatible with other programming languages.
5.2.1 Linear search

Firstly, assume that the list a is not (necessarily) sorted. Any algorithm
would require n comparisons in the worst case, which occurs when the
element does not belong to the list. Note that any method would have
to analyze every single element of the list in order to conclude that the
element is missing. In this situation the algorithms would perform a
“linear search” for an element in a list, which runs in O(n).
The size of the problem is clearly n. If the list is empty the method
can trivially return −1. Another base case can return the location of
Listing 5.6 Linear-recursive linear search of an element in a list.

1 def linear_search_linear(a, x, n):
2 if a == []:
3 return -n - 1
4 elif a[0] == x:
5 return 0
6 else:
7 return 1 + linear_search_linear(a[1:], x, n)
8
9
10 def linear_search_linear_wrapper(a, x):
11 return linear_search_linear(a, x, len(a))
the element if it is found. This base case will depend on the choice of
decomposition. If we reduce the size of the problem by discarding the
last element of the list, then the base case will need to check whether
x is located at the last position. If indeed an−1 = x, then the method
can simply return n − 1. Otherwise, it will need to carry out a recursive
call in order to solve the subproblem of size n − 1, and the final result
will be exactly that of the subproblem’s, which leads to a tail-recursive
solution. Listing 5.5 shows a possible implementation in Python of the
described linear search function.
Finally, it is worth noticing that with the previous decomposition
the locations of the elements in the sublist are the same as those in
the original list. Alternatively, if the decomposition omitted the first
element of the list, all of the indices in the sublist would be decremented
by a unit with respect to those in the original list. For example, if the
list were [2, 8, 5], the 8 is also located at position 1 in [2, 8], but it
appears at location 0 in [8, 5]. This leads to more complex methods
that require an additional parameter. For example, Listing 5.6 shows a
solution that adds a unit to the result of each recursive call. The method
trivially returns 0 if x = a0 , but the base case for an empty list is more
complicated. Note that if that base case is reached the algorithm will
have added n units in the n previous recursive calls. Thus, it needs to
return −n − 1, where n is the length of the initial list (not the particular
length of the input argument since it is 0 in that case) in order to return
−1, indicating that x does not appear in the list. Since n cannot be
obtained from an empty list it has to be passed as an extra parameter
in every function call in order to be recovered in the base case. The code
Listing 5.7 Alternative tail-recursive linear search of an element in a list.

1 def linear_search_tail_alt(a, x, index ):
2 if a == []:
3 return -1
4 elif a[0] == x:
5 return index
6 else:
7 return linear_search_tail_alt(a[1:], x, index + 1)
8
9
10 def linear_search_alt_tail_wrapper(a, x):
11 return linear_search_tail_alt(a, x, 0)
Original problem Original problem
a a
m x < am x > am m
Subproblem Subproblem
(a) (b)
Figure 5.1 Decomposition related to the binary search algorithm.
therefore requires a wrapper method, where the third parameter of the

recursive function is initialized to n.
The last decomposition implies that the algorithm will search for
the element starting at index 0, and will progressively advance until
reaching index n − 1. Another solution consists of explicitly specifying
the index of the element that will be compared to x each time the method
invokes itself. This introduces a new parameter that can be viewed as
an accumulator variable that is incremented with each function call, and
leads to the tail-recursive function in Listing 5.7, which correctly returns
−1 if the list is empty. The function must be called with the accumulator
index parameter set to 0 from a wrapper method. Lastly, observe that
the result is stored in that parameter.
Listing 5.8 Binary search of an element in a list.

1 def binary_search(a, x, lower , upper):
2 if lower > upper : # empty list
3 return -1
4 else:
5 middle = (lower + upper ) // 2
6
7 if x == a[middle ]:
8 return middle
9 elif x < a[middle ]:
10 return binary_search(a, x, lower , middle - 1)
11 else:
12 return binary_search(a, x, middle + 1, upper)
13
14
15 def binary_search_wrapper(a, x):
16 return binary_search(a, x, 0, len(a) - 1)
5.2.2 Binary search in a sorted list

When the input list is sorted it is possible to search for an element by
using a faster approach. Assume that the list a is sorted in ascending
order, where ai ≤ ai+1 , for i = 0, . . . , n − 2. The problem can be solved
in O(log n) running time by “binary search” algorithms. The main idea
behind these methods (there are several variants) consists of dividing the
size of the problem progressively by two. Figure 5.1 illustrates the de-
composition of the problem. Firstly, the algorithms compute the middle
index m. If x = am they can terminate in a base case simply returning
m. Otherwise, since the list is sorted, it is possible to keep searching
for x in only one of the sublists to the right or left of m. In particular,
if x < am then x cannot appear in the sublist [am , . . . , an−1 ], and the
methods can continue to search for x in [a0 , . . . , am−1 ], as shown in (a).
Analogously, if x > am then x can only appear in [am+1 , . . . , an−1 ], as
illustrated in (b). This approach essentially divides the size of the prob-
lem (n) by two. Therefore, the runtime will be characterized by (3.20),
which implies that the methods will run in logarithmic time with respect
to n in the worst case. Lastly, the functions will return −1 (at a base
case) when the list is empty.
Similarly to the linear search functions in Listings 5.6 and 5.7, the
binary search methods need to include additional input parameters in
order to indicate the original indices of the list. For example, consider
searching for an 8 in the list [1,3,3,5,6,8,9]. If we simply pass the sublist

[6, 8, 9] to the function in a recursive call, the 8 is now at index 1, while
it was located at position 5 in the original list. Thus, we need to indicate
that the sublist’s first index should really be 5. In other words, we need
to specify an extra parameter that contains the position in the original
list where the sublist begins. Listing 5.8 shows an implementation of the
method that not only includes this parameter (lower), but also uses an
additional one (upper), which is the index of the original list in which
a sublist ends. Thus, lower and upper simply specify the boundaries
of a sublist within a. The upper parameter is not strictly necessary in
Python since we can use the length of the list in order to determine
the sublist’s upper index (Exercise 5.4 proposes implementing a method
that only uses the lower parameter). However, the code in Listing 5.8
is simpler in the sense that it requires less arithmetic operations, does
not invoke itself with sublists (the recursive calls use the entire list a),
and shows how the method can be implemented in other programming
languages where the length of lists is not directly available. In particular,
if lower is greater than upper then the list is empty and the method
returns −1. Otherwise, the function checks the base case associated with
the condition x = am . If x is not at index m, the method carries out one
of the two recursive calls, on the appropriate sublist, depending on the
condition x < am . Finally, the wrapper method is used to initialize the
lower and upper indices to 0 and n − 1, respectively.
5.3 BINARY SEARCH TREES

A binary search tree is a data structure used for storing data items, each
of which is associated with a certain key (we will consider that it has
to be unique), similarly to the entries in a Python dictionary. It is an
important data structure that allows us to carry out certain operations
efficiently, such as searching, inserting, or removing elements. The data
in each node of the binary tree is a pair (key, item), and the nodes are
arranged according to the values of the keys, which can be compared
through the < operator, or some equivalent function that allows us to
determine if a key is less than another one (i.e., a total preorder binary
relation). In particular, given any node in the tree, every key in its left
subtree must be less than its key, while all of the keys in its right subtree
must be greater than its key. This is also known as the “binary search
tree property,” which implies that every subtree of a binary search tree
is also a binary search tree.
For example, this data structure can be used to store the information
of a birthday calendar. Figure 5.2 shows a binary tree of seven persons
that allows us to retrieve their birthdays according to their names, which
are strings that we can assume are unique (naturally, we can include last
names or tags in order to identify each person uniquely). Keys can be
strings since they can be compared, for example, according to the lexico-
graphic order (in Python we can simply use the < operator). Therefore,
the names in the binary search tree are sorted as they would appear
in a regular dictionary. For instance, consider the root node associated
with “Emma”. Observe that all of the names contained in its left subtree
would appear before “Emma” in a dictionary, and all of the names in the
right subtree would be found after “Emma”. Furthermore, this property
holds for every node of the binary search tree.
In order to implement a binary search tree, we first have to decide
how to code a binary tree. There are several possibilities (one of the most
common approaches consists of using object oriented features in order
to declare a class associated with the nodes of the tree), but in this book
we will simply use lists. In particular, every node of the tree will consist
of a list of four elements: the key, the item, the left subtree, and the
right subtree, where the subtrees are also binary search trees. Thus, the
binary search tree in Figure 5.2 would correspond to the following list:
[‘Emma’, ‘2002/08/23’,
[‘Anna’, ‘1999/12/03’, [], []],
[‘Paul’, ‘2000/01/13’,
[‘Lara’, ‘1987/08/23’, (5.1)
[‘John’, ‘2006/05/08’, [], []],
[‘Luke’, ‘1976/07/31’, [], []]],
[‘Sara’, ‘1995/03/14’, [], []]]]
which is illustrated graphically in Figure 5.3. Observe that the left and
right subtrees are lists, which are empty at the leaf nodes.
5.3.1 Searching for an item

The goal of the next problem is to search for some item with a known
key k in a binary search tree, and retrieve its associated item. We can
assume that the size of the problem is the height of the tree. A trivial
base case occurs when the list that represents the binary tree is empty.
In that case the algorithm can simply return None. There is another
situation where the algorithm can provide a result without carrying out
a recursive call. If the key of the root node is k, the method will have
PSfrag
Emma 2002/08/23
Anna 1999/12/03 Paul 2000/01/13
Lara 1987/08/23 Sara 1995/03/14
John 2006/05/08 Luke 1976/07/31
Figure 5.2 Binary search tree that stores information about a birthday
calendar.
‘Emma’,‘2002/08/23’, ,
‘Anna’,‘1999/12/03’,[ ],[ ] ‘Paul’,‘2000/01/13’, ,
‘Lara’,‘1987/08/23’, , ‘Sara’,‘1995/03/14’,[ ],[ ]
‘John’,‘2006/05/08’,[ ],[ ] ‘Luke’,‘1976/07/31’,[ ],[ ]
Figure 5.3 Binary search tree in Figure 5.2 and (5.1), where each node
is a list of four elements: name (string), birthday (string), left subtree
(list), and right subtree (list).
found the item, and can therefore return it. Thus, in the recursive cases
we can be sure that the root node does not contain the searched item.
In the next step our goal is to find an appropriate decomposition of
the problem that reduces its size. We have already seen that trees are
composed recursively of subtrees. Thus, we could consider searching for
the item in the two subtrees of the binary search tree. This guarantees
reducing the size of the problem by a unit, since it discards the root
node. Nevertheless, it is easy to see that in this problem we can also
avoid searching in an entire subtree. If the key k is less than the key of
the root node (kroot ), then we can be sure that the item we are looking for
will not be in the right subtree, due to the binary tree search property.
kroot xroot kroot xroot
Subproblem Subproblem
k < kroot k > kroot
Figure 5.4Decomposition associated with several algorithms related to

binary search trees.
Algorithm for searching an item with a particular key in a

Listing 5.9
binary search tree.
1 def bst_search(T, key):
2 if T == []:
3 return None
4 elif T[0] == key:
5 return T[1] # return the root item
6 elif key < T[0]:
7 return bst_search(T[2], key) # search in left subtree
8 else:
9 return bst_search(T[3], key) # search in right subtree
Analogously, if k > kroot , the item will not appear in the left subtree.
Figure 5.4 illustrates this idea, where xroot is the item stored in the root
node. Clearly, for this particular problem there are two recursive cases.
If k < kroot the method must keep searching in the left subtree through
a recursive call, while if k > kroot it will search for the item in the right
subtree. Listing 5.9 shows an implementation of the searching algorithm,
where each node of the binary tree is coded as a list of four components,
as described in Section 5.3.
Finally, the height of the binary tree determines the cost of the algo-
rithm in the worst case. If the tree is balanced (i.e., it has approximately
the same number of nodes on the left and right subtrees of nodes that
Procedure for inserting an item with a particular key in a

Listing 5.10
binary search tree.
1 def insert_binary_tree(x, T):
2 if T == []:
3 T.extend ([x[0], x[1], [], []])
4 else:
5 if x[0] < T[0]:
6 if T[2] == []:
7 T[2] = [x[0], x[1], [], []]
8 else:
9 insert_binary_tree(x, T[2])
10 else:
11 if T[3] == []:
12 T[3] = [x[0], x[1], [], []]
13 else:
14 insert_binary_tree(x, T[3])
appear on the same level) it runs in Θ(log n), where n is the number of
nodes in the tree, since it would discard approximately half of the nodes
with each recursive call. However, the tree could have a linear structure,
in which case the algorithm would run in O(n).
5.3.2 Inserting an item

The goal in this problem is to insert an item with a given key k in a
binary search tree, such that the resulting tree also satisfies the binary
search tree property. The method will be a procedure that receives two
parameters: (1) a tuple x containing the key and the item, and (2) the
binary search tree T . In particular, it will extend the list (T ) by adding
a new leaf to the tree.
Again we can consider that the size of the problem is the height of
the tree. The simplest base case occurs when the tree is empty. In that
situation the algorithm must extend the tree in order to incorporate a
single node, which obviously will not have any children. The diagrams in
Figure 5.4 also illustrate the decomposition that we will follow for solving
this problem. Clearly, if the tree is not empty, then the procedure will
insert a node in the left subtree if k is smaller than the key of the root
node. Otherwise, it will insert it in the right subtree. Firstly, assume
that k < kroot . There are two possible scenarios that lead to a base case
and a recursive case. If the root node does not have a left subtree, then
the procedure can simply replace the associated empty list by the node
pivot
6 4 1 7 4 7 3 6 5
4 1 3 4 6 7 7 6 5
≤ pivot > pivot
Figure 5.5Partitioning of a list used in the quicksort and quickselect

algorithms.
containing the key and the item (and two empty subtrees). However,
if the root node does have a nonempty subtree then the method must
insert the new node in that subtree, which naturally leads to a recursive
call. The reasoning regarding the right subtree is analogous. Finally,
Listing 5.10 shows a possible implementation of the procedure.
5.4 PARTITIONING SCHEMES

A well-known problem in computer science consists of partitioning a list
in the following manner:
1. Choose a “pivot” element from the list. Typical choices include the
first or middle element of the list, or simply some element selected
at random.
2. Construct a new list with the same elements as the input list, or
simply permute the input list, where the first elements are smaller
than or equal to the pivot, and the last ones are greater than the
pivot.
Figure 5.5 illustrates the idea through a concrete example. The problem
is relevant since the partitioning is key in other eminent algorithms such
as quickselect (see Section 5.5) or quicksort (see Section 6.2.2). There
are several well-known and efficient algorithms, also denoted as “parti-
tioning schemes,” which solve the problem. The most popular is Hoare’s
partitioning method, developed by Tony Hoare, which we will analyze
shortly. Nevertheless, first we will examine a simpler approach for the

sake of introducing more intuitive recursive algorithms.
5.4.1 Basic partitioning scheme

A simple idea consists of scanning the entire input list in order to build:
(1) a new list that contains the elements that are smaller than or equal
to the pivot, and (2) another list formed by the elements that are greater
than the pivot. Afterwards, they can be concatenated together with the
pivot element, which would obviously be placed in between them.
Since the problems are very similar we will explain only the first
one. In particular, the inputs to the problem are a list a of length n,
and some value x that plays the role of the pivot. Its size is clearly the
length of the list. The base case occurs when the input list is empty,
where algorithms can simply return an empty list. For the recursive case
we can decompose the problem by discarding the first element of the list,
which reduces its size by a unit. In that case, we can form the following
recursive diagrams, depending on whether the first element of the list is
smaller than or equal to the pivot. If it is, a concrete diagram could be:
Inputs Results
([3, 6, 1, 7, 4],5) ÐÐÐÐÐÐÐÐÐ→ [3, 1, 4]
concatenate [3]
([6, 1, 7, 4],5) ÐÐÐÐÐÐÐÐÐ→ [1, 4]
Clearly, the algorithm must concatenate the first element of the list to
the output of the subproblem. In contrast, if the first element is greater
than x, the diagram is:
Inputs Results
([9, 6, 1, 7, 4],5) ÐÐÐÐÐÐÐÐÐ→ [1, 4]
do nothing
([6, 1, 7, 4],5) ÐÐÐÐÐÐÐÐÐ→ [1, 4]
In this case, the algorithm simply has to return the solution (list) to
the subproblem. Listing 5.11 shows the linear-recursive codes that solve
Listing 5.11 Auxiliary methods for partitioning a list.

1 def get_smaller_than_or_equal_to(a, x):
2 if a == []:
3 return []
4 elif a[0] <= x:
5 return [a[0]] + get_smaller_than_or_equal_to(a[1:], x)
6 else:
7 return get_smaller_than_or_equal_to(a[1:], x)
8
9
10 def get_greater_than(a, x):
11 if a == []:
12 return []
13 elif a[0] > x:
14 return [a[0]] + get_greater_than(a[1:], x)
15 else:
16 return get_greater_than(a[1:], x)
both problems (notice that in the recursive cases that perform the con-
catenation the function call is not the last action of the method). The
runtime cost of the methods can be characterized by:
⎧
⎪
⎪1 if n = 0,
T (n) = ⎨
⎪
⎩T (n − 1) + 1 if n > 0,
⎪
which implies that their runtime is linear with respect to n. Lastly,

the functions can be substituted with Python expressions that im-
plement “filters.” In particular, get_smaller_than_or_equal_to(a,x)
and get_greater_than(a,x) return the same lists as the expressions
[y for y in a if y <= x] and [y for y in a if y > x], respec-
tively.
5.4.2 Hoare’s partitioning method

Listing 5.12 contains an iterative version of Hoare’s partitioning algo-
rithm. In particular, it partitions a sublist within the input list (a),
which is specified through lower and upper indices. Firstly, the method
chooses a pivot (in this case, the element in the middle of the sublist)
and swaps it with the first element of the sublist (lines 3–6). Afterwards,
it sets a left index parameter at the second position of the sublist, and
a right index at the last location of the sublist (lines 8 and 9). Subse-
quently the method enters the main loop of the algorithm. Inside it, it
lower pivot index upper
select pivot
6 4 1 7 4 7 3 6 5
and indices
left right
swap pivot
4 4 1 7 6 7 3 6 5
and first element
right
advance left
4 4 1 7 6 7 3 6 5
and right indices
left
swap left
4 4 1 3 6 7 7 6 5
and right elements
right
advance left
4 4 1 3 6 7 7 6 5
and right indices
left
swap pivot
3 4 1 4 6 7 7 6 5
and right element
Figure 5.6 Example of Hoare’s partition method.

Listing 5.12 Hoare’s iterative partitioning algorithm.

1 def partition_Hoare(a, lower , upper):
2 if upper >= 0:
4 pivot = a[middle ]
5 a[middle ] = a[lower]
6 a[lower] = pivot
7
8 left = lower + 1
9 right = upper
10
11 finished = False
12 while not finished :
13 while left <= right and a[left] <= pivot :
14 left = left + 1
15
16 while a[right ] > pivot:
17 right = right - 1
18
19 if left < right:
20 aux = a[left]
21 a[left] = a[right ]
22 a[right] = aux
23
24 finished = left > right
25
26 a[lower] = a[right ]
27 a[right] = pivot
28
29 return right
iteratively increases the left index until it references an element that is

larger than the pivot (lines 13 and 14). Analogously, it decreases the
right index until it references an element that is less than or equal to
the pivot (lines 16 and 17). The algorithm then swaps the elements ref-
erenced by the left and right index (if the left index is smaller than the
right one) in lines 19–22. This process is repeated until the indices cross
(i.e., until the left index is greater than the right one, which is checked
in line 24). At the end of the loop the elements located up to the right
index will be smaller than or equal to the pivot, and the elements from
the left index until the end of the sublist will be greater than the pivot.
Finally, the procedure completes the partition by swapping the element
referenced by the right index with the pivot (lines 26 and 27), and re-
swap
>p ⋯ ≤p ⋯
left right left right
(a)
≤p ⋯ ≤p ⋯
>p ⋯ >p ⋯
≤p ⋯ >p ⋯
(b)
Figure 5.7 Decomposition of Hoare’s partitioning problem.
turns the final location of the pivot. Figure 5.6 shows a concrete example
of the partitioning method.
We will now see a tail-recursive algorithm that can substitute the
main loop in Hoare’s partitioning scheme. Given the input list a, initial
left and right indices, as well as the value of the pivot, the method
should partition the list analogously as the loop in Hoare’s method,
returning the final location of the right index. The size of the problem
is the difference between the left and right indices, since it determines
the number of increment/decrement operations to be performed on the
indices until they cross. The base case occurs precisely when they cross
(when the left index is greater than the right one), where the method
can simply return the right index.
For the recursive case we must reduce the size of the problem by
incrementing the left index and/or decrementing the right one. There
are two different scenarios, illustrated in Figure 5.7, where p denotes
Listing 5.13 Alternative recursive version of Hoare’s partitioning scheme.

1 def partition_Hoare_rec(a, left, right , pivot ):
2 if left > right :
3 return right
4 else:
5 if a[left] > pivot and a[right ] <= pivot:
6 aux = a[left]
7 a[left] = a[right]
8 a[right ] = aux
9 return partition_Hoare_rec(a, left + 1,
10 right - 1, pivot)
11 else:
12 if a[left] <= pivot :
13 left = left + 1
14 if a[right] > pivot :
15 right = right - 1
16 return partition_Hoare_rec(a, left, right , pivot)
17
18
19 def partition_Hoare_wrapper(a, lower , upper):
20 if upper >= 0:
22 pivot = a[middle ]
23 a[middle ] = a[lower]
24 a[lower] = pivot
25
26 right = partition_Hoare_rec(a, lower + 1, upper , pivot)
27
28 a[lower] = a[right ]
29 a[right] = pivot
30
31 return right
the pivot. If aleft > p and aright ≤ p the method first needs to swap
the elements referenced by the indices, and afterwards can perform a
function call by advancing both indices, as shown in (a). This leads to a
first recursive case that reduces the size of the problem by two units. If
the previous condition is False, then the method will not swap elements,
and will advance at least one of the indices (thus, reducing the size of
the problem), as shown in (b). If aleft ≤ p it will increment the left index,
and if aright > p it must decrement the right one. Subsequently, the
method can invoke itself with the updated indices. Listing 5.13 shows
the corresponding code, together with a wrapper method that completes
the partitioning algorithm. Note that it is identical to Listing 5.12, where
ip = k − 1
a
(a)
ip > k − 1 ip < k − 1
a Subproblem a Subproblem
(b) (c)
Figure 5.8 Base case and problem decomposition used by the quickselect
algorithm.
the code associated with the loop has been substituted by a call to the
recursive function. Lastly, the method runs in O(n) since the runtime
of the recursive method can be characterized by the recurrence T (n) =
T (n − 1) + 1 in the worst case.
5.5 THE QUICKSELECT ALGORITHM

The quickselect algorithm, also developed by Tony Hoare, is a selec-
tion method based on Hoare’s partitioning scheme (see Section 5.4.2).
In particular, it finds the k-th smallest number (also called the k-th or-
der statistic) in a list that is not sorted. Assume that an instance of the
problem is defined through a numerical input list a, lower and upper
indices defining a sublist within it, and a positive integer k. The algo-
rithm will therefore search for the k-th smallest number in the specified
sublist. The size of the problem is the length of the sublist. The smallest
instances of the problem correspond to lists that contain a single ele-
ment. Therefore, the base case occurs when the lower and upper indices
are identical, where the method can simply return the element of the
list.
For larger lists the method first applies Hoare’s partition scheme,
which arranges the list in three parts. A first sublist contains the elements
of a that are less than or equal to a chosen “pivot” element from the list.
This list is followed by another one that only contains the pivot. Finally,
a third list contains the elements of a that are greater than the pivot.
Listing 5.14 Tail-recursive quickselect algorithm.

1 def quickselect(a, lower , upper , k):
2 if lower == upper:
3 return a[lower]
4 else:
5 pivot_index = partition_Hoare_wrapper(a, lower , upper)
6
7 if pivot_index == k - 1:
8 return a[pivot_index]
9 elif pivot_index < k - 1:
10 return quickselect(a, pivot_index + 1, upper , k)
11 else:
12 return quickselect(a, lower , pivot_index - 1, k)
Let ip denote the index of the pivot. Since the function that imple-
ments Hoare’s partition scheme returns ip , we can check whether it is
equal to k − 1 (since indices start at location 0, the j-th element appears
at position j −1). If it is, then the pivot will be the k-th smallest element
in a, and the method can terminate at this base case. This scenario is
illustrated in Figure 5.8(a).
In the recursive case the algorithm reduces the size of the problem
by considering either the sublist to the left, or to the right, of the pivot.
If ip > k − 1, then the location of the searched element will be smaller
than ip . Thus, the algorithm can focus on the sublist to the left of the
pivot, as shown in Figure 5.8(b). Instead, if ip < k − 1 then the algo-
rithm will proceed by solving the subproblem related to the list to the
right of the pivot, as illustrated in (c). Finally, Listing 5.14 contains an
implementation of the tail-recursive function.
The runtime cost of the algorithm depends on where the pivot is
located after performing the decomposition. If it is always located at the
middle position of the list, the running time can be characterized by:
⎧
⎪
⎪1 if n ≤ 1,
T (n) = ⎨
⎪
⎩T (n/2) + cn if n > 1,
⎪
since the algorithm needs to solve a subproblem of approximately half

the size of the original’s, while the methods that perform the partition
run in linear time. This is the best case scenario, where T (n) ∈ Θ(n).
However, if the pivot is always located at an extreme of the list, then
the runtime cost is determined by:

⎧
⎪
⎪1 if n ≤ 1,
T (n) = ⎨
⎪
⎩T (n − 1) + cn if n > 1,
⎪
which is a quadratic function (i.e, T (n) ∈ Θ(n2 )). This situation corre-
sponds to the worst case scenario for the algorithm.
5.6 BISECTION ALGORITHM FOR ROOT FINDING

Consider a continuous real-valued function f (x) defined over some inter-
val [a, b]. The bisection algorithm (also known as the interval halving,
binary search, or dichotomy method) is a strategy for finding an approx-
imation ẑ to a root of f (x) in the (a, b) interval. Recall that a root, or
zero, of f (x) is a value r for which f (r) = 0.
The inputs to the problem are the extremes of the interval a and b,
where a ≤ b, the function f , and a small value ǫ > 0 that determines the
accuracy of the approximation. In particular, the output of the function ẑ
will satisfy ∣ẑ −r∣ ≤ ǫ. Thus, the smaller ǫ is, the better the approximation
will be. In addition, another prerequisite over the inputs is that f (a) and
f (b) must have opposite signs. This guarantees that there will exist a
root inside the interval (a, b). Lastly, the approximation ẑ is defined as
the midpoint between a and b. In other words, ẑ = (a + b)/2.
The method works by progressively dividing the interval by two
where the root must exist. Figure 5.9 illustrates a few steps of the pro-
cess. In the initial setting the root is located to the left of the approxima-
tion (r < ẑ). Therefore, in the next step b is updated to ẑ, which halves
the interval. Similarly, at step 1, ẑ < r, and a is assigned the value of ẑ in
the following step. This procedure is applied repeatedly until the interval
is small enough, guaranteeing that ẑ meets the required accuracy.
The size of the problem depends on the length of the interval [a, b],
and the base case occurs when b − a ≤ 2ǫ, where the method returns
ẑ = (a + b)/2. Observe that this condition guarantees ∣ẑ − r∣ ≤ ǫ, as
illustrated in Figure 5.10. In particular, if ẑ is the midpoint of the interval
[a, b], then the distance between ẑ and r cannot be greater than ǫ. In
addition, the method can also return ẑ trivially if f (ẑ) = 0 (or if f (ẑ) is
sufficiently small).
The decomposition of the problem involves dividing its size by two.
This is accomplished by replacing one of the extremes of the interval
[a, b] by its midpoint ẑ. Since the signs of the new extremes must be
f (x) f (x)
r r
a b a b
ẑ ẑ
Initial setting Step 1
f (x) f (x)
r r
a b a b
ẑ ẑ
Step 2 Step 3
Figure 5.9 Steps of the bisection algorithm.
different, ẑ should replace b when f (a) and f (z) have opposite signs.
Otherwise, ẑ will replace a. This leads to the tail-recursive method in
Listing 5.15, which contains two recursive cases. The code uses√ the func-
tion f (x) = x − 2 that allows us to find an approximation of √
2
2, since
it is a root of f (x). Note that the initial interval [0, 4] contains 2. Fi-
nally, the error of the approximation will be less than or equal to 10−10
(the result will be accurate up to nine decimal digits to the right of the
decimal point).
5.7 THE WOODCUTTER PROBLEM

In this problem a woodcutter needs to collect w units of wood from a
forest composed of n trees. He has a special machine that is able to
cut the trees at some height h, where he would gather the wood from
the trees above h. Assume that w is an integer (not greater than all of
the wood in the forest), and that the heights of the n trees, which are
also integers, are specified through a list t. The goal of the problem is
to find the largest value of h that will allow him to collect at least w
units of wood. Figure 5.11 shows an instance of the problem for n = 20
f (x)
∣ẑ − r∣ ≤ ǫ
a r b
ǫ ǫ
ẑ
Figure 5.10 Base case of the bisection algorithm (b − a ≤ 2ǫ).
Listing 5.15 Bisection algorithm.

1 def f(x):
2 return x * x - 2
3
4
5 def bisection(a, b, f, epsilon ):
6 z = (a + b) / 2
7
8 if f(z) == 0 or b - a <= 2 * epsilon :
9 return z
10 elif (f(a) > 0 and f(z) < 0) or (f(a) < 0 and f(z) > 0):
11 return bisection(a, z, f, epsilon )
12 else:
13 return bisection(z, b, f, epsilon )
14
15
16 # Print an approximation of the square root of 2
17 print (bisection(0, 4, f, 10**(-10) ))
trees, where the tallest tree has height H = 12. If the goal is to collect 10
units of wood, then the woodcutter should set the height of the cutting
machine to h = 8, where the total wood collected would be exactly 10
units. If the goal were to collect 7, 8, or 9 units of wood the optimal
height h would also be 8. Even though the woodcutter would obtain
more wood than initially required, h cannot be higher since cutting at
height 9 only provides 6 units of wood.
The problem is interesting from an algorithmic point of view since
it can be solved in several ways. For example, the trees can be initially
sorted in decreasing order by their height, and subsequently processed
from the highest to the lowest until obtaining the optimal height. This
Goal: find height (h) for collecting (10) units of wood

12 H
11
10
9
8 h
7
6
5
4
3
2
1
0
1 ⋯ n
Figure 5.11 Instance of the woodcutter problem.
Listing 5.16Function that computes the amount of wood collected when

cutting trees at height h.
1 def compute_wood(t, h):
2 if t == []:
3 return 0
4 else:
5 if t[0] > h:
6 return t[0] - h + compute_wood(t[1:], h)
7 else:
8 return compute_wood(t[1:], h)
approach would run in O(n log n) if we apply general sorting algorithms

like merge sort or quicksort (see Chapter 6). Since the heights are integers
it is possible to apply linear time sorting algorithms like the counting
sort method (see Exercise 5.3), which would lead to a solution that runs
in O(n+H). Thus, if H were small this approach could be more efficient.
We will now present a “binary search” solution that runs in
O(n log H), which does not rely on sorting the heights of the trees.
Firstly, note that a basic approach consists of starting with h = H − 1,
and then decreasing h unit by unit until obtaining the required amount
of wood.
For simplicity, assume that the method calculates the amount of

wood obtained at each new height independently of the amount gath-
ered for other heights. This implies that computing the amount of wood
collected from n trees requires on the order of n operations. Therefore, in
total, the algorithm for solving the problem would require O(nH) oper-
ations in the worst case (at each of the possible H heights the algorithm
would need to carry out n computations). In particular, the amount of
wood collected at a certain height h can be obtained through the linear-
recursive function in Listing 5.16 (Exercise 11.2 proposes implementing
a tail-recursive version of the function). The method returns zero in the
base case if the list of tree heights is empty. The recursive case processes
the first tree height from the list. If it is greater than h the tree will
be cut, and the function must return the difference between the height
of the tree t0 and h, plus the wood collected from the rest of the trees
(through the recursive call). If the tree’s height is less than or equal to
h, the tree will not be cut, and the method can simply return the total
wood gathered from the remaining trees.
Instead of decreasing h one unit at a time, the following algorithm
uses a strategy similar to a binary search of an element in a list, or to
the bisection method. The idea is to start searching for the height h
in the middle of the interval [0, H], and progressively halving it until
a decision can be taken. Thus, the method can receive the tree heights
and specified wood as parameters, together with lower and upper limits
that indicate where the solution h can be found.
The size of the problem is the difference between the upper and lower
limits (which is initially H). A first base case occurs if they share the
same value. In that case, the output height h will be precisely that value.
Otherwise the algorithm will compute a “middle” height, as the (integer)
average of the lower and upper limits. If the total amount of wood that
would be gathered from cutting the trees at this middle height is equal
to the goal w, then the method can return such middle height in another
base case.
Figure 5.12 shows the decomposition of the problem used by the al-
gorithm. The original problem is shown in the top image, given the lower
and upper indices. After computing the wood that would be collected at
the middle height (hm ), there are two possible scenarios. Firstly, if it is
greater than w we can clearly rule out all of the heights below hm . Thus,
we can continue searching for the optimal height (through a recursive
call) by replacing the lower limit with hm . It is important to note that
hm could still be the solution, since the wood gathered for hm + 1 may
upper
middle
lower
Original problem
upper
middle
lower
Subproblem: wood at middle > w
upper
middle
lower
Subproblem: wood at middle < w
Figure 5.12 Decomposition of the woodcutter problem.

Listing 5.17 Binary search algorithm for the woodcutter problem.

1 def collect_wood(t, wood, lower , upper ):
2 middle_h = (lower + upper) // 2
3 wood_at_middle = compute_wood(t, middle_h )
4
5 if wood_at_middle == wood or lower == upper:
6 return middle_h
7 elif lower == upper - 1:
8 if compute_wood(t, upper ) >= wood:
9 return upper
10 else:
11 return lower
12 elif wood_at_middle > wood:
13 return collect_wood(t, wood, middle_h , upper )
14 else:
15 return collect_wood(t, wood, lower , middle_h - 1)
be less than the required amount w. A second scenario occurs when the
wood collected at hm is less than w. In that case we can discard all of the
heights that are greater than or equal to hm , since the woodcutter would
not obtain enough wood. The associated recursive case would therefore
invoke the method by replacing the upper limit with hm − 1.
Lastly, in the previous decomposition, when the upper limit is just a
unit greater than the lower one, the middle height is equal to the lower
one. In that situation the first recursive case would not reduce the size
of the problem (the limits would not vary). Thus, the algorithm needs
an additional base case in order to work properly. When this situation
occurs the method has to return either the lower or the upper limit.
In particular, if the amount of wood associated with the upper limit is
greater than or equal to w, then the upper limit is the correct result.
Otherwise, it will be the lower one.
Listing 5.17 shows the code related to the mentioned base and recur-
sive cases. Lastly, Figure 5.13 shows the steps it takes in order to solve
the instance in Figure 5.11. Step 1 shows an initial situation where the
lower limit is zero and the upper one is H = max{ti }, for i = 1, . . . , n.
Step 2 and 3 are associated with applying the first and second recursive
cases, respectively. Finally, step 4 applies the additional base case since
the upper limit (8) is just a unit over the lower one (7). The solution is
h = 8, since for that height the woodcutter obtains the requested amount
of wood (10 units).
12 12
8 8
4 4
0 0
Step 1 Step 2
12 12
8 8
4 4
0 0
Step 3 Step 4
Steps of the binary search algorithm related to an instance of

Figure 5.13
the woodcutter problem, for w = 10.
5.8 EUCLID’S ALGORITHM

One of the first algorithms developed throughout history is known as
Euclid’s algorithm, named after the ancient Greek mathematician Eu-
clid, who first described it in his Elements (c. 300 BC). Its purpose is
finding the greatest common divisor (gcd) of two nonnegative integers m
and n (where both cannot be zero), which is the largest positive integer
k that divides both m and n without a remainder (i.e., m/k and n/k are
integers). For example, the greatest common divisor of 20 and 24 is 4.
It can also be understood as the product of the common prime factors
of m and n. For m = 20 = 2 ⋅ 2 ⋅ 5, and n = 24 = 2 ⋅ 2 ⋅ 2 ⋅ 3, the product of
common factors is 2 ⋅ 2 = 4.
There are several versions of the method. The original one corre-
sponds to the following function:
⎧
⎪
⎪
⎪
if m = 0,
⎪
n
gcd1 (m, n) = ⎨gcd1 (n, m)
⎪
if m > n, (5.2)
⎪
⎪
⎪
⎩gcd1 (m, n − m) otherwise,
Listing 5.18Euclid’s algorithm for computing the greatest common divi-

sor of two nonnegative integers.
1 def gcd1(m, n):
2 if m == 0:
3 return n
4 elif m > n:
5 return gcd1(n, m)
6 else:
7 return gcd1(m, n - m)
which can be coded as shown in Listing 5.18. Firstly, it is tail-recursive

since it returns the value of a function call in the recursive cases. In
addition, observe that the method returns the value of a parameter in
the base case, which is common in many tail-recursive functions. In par-
ticular, the greatest common divisor of n > 0 and 0 is clearly n, since
n divides both n and 0 without a remainder. The first recursive case
simply switches the order of the arguments. This guarantees that the
first one will be larger than or equal to the second one when reaching
the second recursive case. In this last recursive case reducing the size
of the problem (which depends on m and n) by a unit, or dividing it
by two, does not work. Instead, the recursive case decreases the second
argument by subtracting m (observe that n − m ≥ 0, since n ≥ m). It
implements the mathematical property:
gcd(m, n) = gcd(m, n − m), (5.3)
which is not immediately obvious. It can be shown to be true as fol-

lows. Assume n ≥ m, and let m = az, and n = bz, where z = gcd(m, n)
represents the product of the common prime factors in n and m, and
a ≤ b. This implies that a and b do not share common prime factors. In
addition, let b = a + c. In that case we can express n and m as:
n = bz = (a + c)z = (a1 ⋯ak + c1 ⋯cl )z, (5.4)

m = (a1 ⋯ak )z, (5.5)
where a1 ,. . . ak and c1 ,. . . cl are the prime factors of a and c, respectively.

The key to the proof is the fact that a and c cannot share prime fac-
tors (i.e., ai ≠ cj ), because if they did they could be pulled out of the
parenthesis in (5.4), implying that a and b share those common prime
factors, which is a contradiction. If a and c do not share prime factors
then z = gcd(az, cz), and we can conclude that:
gcd(az, bz) = z = gcd(az, cz) ⇒ gcd(m, n) = gcd(m, n − m).
In addition, it can also be shown that the algorithm is guaranteed to

reach the base case, since the values of the arguments decrease until one
of them is 0. For m = 20 and n = 24 the method carries out the following
recursive calls:
gcd1 (20, 24)
= gcd1 (20, 4)
= gcd1 (4, 20)
= gcd1 (4, 16)
= gcd1 (4, 12) (5.6)
= gcd1 (4, 8)
= gcd1 (4, 4)
= gcd1 (4, 0)
= gcd1 (0, 4) = 4.
Since tail-recursive functions do not manipulate the result of recursive

calls, all of them return the same value, which is obtained at the base
case (gcd1 (0, 4) = 4). Therefore, tail-recursive functions essentially define
relationships between sets of arguments. For example, in (5.6) the pairs
(20, 24), (20, 4), and so on, are all related to the value 4. Lastly, a more
efficient version that is still used today is:
⎧
⎪
⎪n if m = 0,
gcd2 (m, n) = ⎨
⎪
(5.7)
⎩gcd2 (n%m, m) if m ≠ 0.
⎪
The proof of the property in the recursive case is similar to the one for
(5.3). Assume n ≥ m, and let m = az, and n = bz, where z is the product
of the common prime factors in n and m, and a ≤ b. This implies that
a and b do not share common prime factors. In addition, let b = qa + r,
where q and r are the quotient and remainder of the division b/a. The
key in this proof is that a and r cannot share common prime factors,
since otherwise they would also be prime factors of b, and this is not
possible since a and b do not share common prime factors. This implies
that z = gcd(rz, az). Thus, we can conclude that:
gcd(az, bz) = z = gcd(rz, az) ⇒ gcd(m, n) = gcd(n%m, m).

a 2 2 3 2 0 1 3 2 0 0 4
step 1
b 3 1 4 2 1
0 1 2 3 4
step 2
c 0 0 0 1 2 2 2 2 3 3 4
Figure 5.14 Steps in the counting sort algorithm.
5.9 EXERCISES
Exercise 5.1 — Define and code Boolean linear and tail-recursive func-
tions that determine whether a nonnegative integer n contains an odd
digit.
Exercise 5.2 — It is more pleasant to display polynomials in a format-

ted style rather than through a list. Write a method that receives a list
of length n that represents a polynomial of degree n − 1, and prints it
in a formatted style. For example, given the input list p = [3, −5, 0, 1],
corresponding to x3 −5x1 +3 (i.e., pi is the coefficient associated with the
term xi ), the method should print a line similar to: + 1x^3 - 5x^1 + 3.
Assume that pn−1 ≠ 0, except if the polynomial is the constant 0. Finally,
specify its asymptotic computational cost.
Exercise 5.3 — The “counting sort” algorithm is a method for sorting

a list a of n integers in the range [0, k], where k is usually small. The
method runs in O(n + k) time, which implies that it runs in O(n) time
if k ∈ O(n).
Given a, the method creates a new list b that contains the number
of occurrences of each integer in a, as shown in step 1 of Figure 5.14. For
example, b2 = 4, since the integer 2 appears four times in a. Implement
a tail-recursive procedure that receives lists a and b (initialized with
zeros), and fills list b with the occurrences of the integers in a.
In addition, implement a linear-recursive function that receives the
list of integer occurrences b, and returns a new list (c) that is the sorted
version of a, as shown in step 2 of Figure 5.14.
Finally, implement a function that calls the previous two methods in

order to implement a version of the counting sort algorithm, and specify
its asymptotic computational cost.
Exercise 5.4 — Implement an alternative version of Listing 5.8 that

does not use the upper parameter. Finally, specify its asymptotic com-
putational cost.
Exercise 5.5 — Implement a Boolean “binary search” function in

Python that simply indicates whether an element x appears in a list
a. The function will have only two input parameters: x and a, and will
be based on decomposing the problem into another one of half its size.
Finally, specify its asymptotic computational cost.
Exercise 5.6 — Write a function that searches for the item with the
smallest key in a binary search tree T , defined as a list of four compo-
nents, as described in Section 5.3.
Exercise 5.7 — Let a be a sorted list (in ascending order) of different

integers. The goal of this problem is to efficiently search for an element
that is equal to its position (i.e., index) along the list. In other words,
we want to find an element i for which ai = i. For example, if a =
[−3, −1, 2, 5, 6, 7, 9], then the output of the method will be 2, since a2 =
2. Note that the first element of the list is located at position 0. For
simplicity, assume that a will have at most one element that satisfies
ai = i. If the list does not contain any element that satisfies that property
the function will return the value −1. Finally, specify its asymptotic
computational cost.
Exercise 5.8 — Let a be a list of n integers, arranged so that even

numbers appear before odd numbers (the index of any even number
will be smaller than the index of any odd number), like, for example,
a = [2, −4, 10, 8, 0, 12, 9, 3, −15, 3, 1]. The goal of the following problem
is to develop an efficient recursive algorithm for determining the largest
index associated with the even numbers. In other words, the index i for
which ai is even, but aj will be odd for any j > i. For the example the
method would return i = 5, since a5 = 12 is even, but a6 = 9 is odd (and
so are a7 , a8 , and so on). If a does not contain even numbers the result
will be −1, while if it is composed entirely of even numbers the output
will be n − 1. Finally, specify its asymptotic computational cost.
f (x)
xn−1 r
xn
(xn−1 , f (xn−1 ))
Figure 5.15 Main idea behind Newton’s method.
Exercise 5.9 — Newton’s method is an approach for finding succes-

sively better approximations to the roots of a real-valued function f (x)
that is differentiable. It is based on the following recursive rule:
f (xn−1 )
xn = xn−1 −
f ′ (xn−1 )
,
which is explained geometrically in Figure 5.15. Say we have an initial

approximation xn−1 to the root r of function f (x). The procedure finds
the tangent line to f (x) at xn−1 , and computes a new approximation
(xn ) to r that is the value of x where the line crosses the X-axis (note
that xn is closer to r). Thus, starting with some initial value x0 , ẑ = xn
will be the approximation of the root of f (x) after applying the recursive
rule n steps.
This procedure can be used, for example, to obtain very accurate
√
approximations to the square root of a number, in only a few steps. Say
√
we want to calculate a, which is some unknown value x. In that case we
have the following identities: x = a, where x2 = a, which implies that
x2 − a = 0. Therefore, the root of the function x2 − a will be the square
root of a. In that case, the recursive formula associated with Newton’s
method is:
x2 + a
xn = xn−1 − n−1 . (5.8)
2xn−1
√
Implement linear and tail-recursive functions that receive the value a,
√
an initial positive estimate x0 of a, and a certain number of steps n,
and return the final estimate ẑ = xn of a by applying (5.8) n times.
Finally, specify their asymptotic computational cost.
CHAPTER 6
Multiple Recursion I:
Divide and Conquer
∆ιαίρει και βασίλευε (divide and conquer).

— Philip II of Macedon
HE advantages of recursion over iteration, such as code clarity or

T avoiding managing a stack explicitly (see Section 10.3.5), are mainly
due to the possibility of using multiple recursion. Methods based on this
type of recursion invoke themselves several times in at least one recursive
case. Therefore, these algorithms solve more than one simpler self-similar
problem, and must combine, extend, and/or modify their results in order
to obtain the solution to the original problem.
The book devotes three chapters to multiple recursion, and contains
additional examples throughout it as well. In this chapter we will cover
an important class of algorithms based on multiple recursion that decom-
pose problems by dividing their size by some constant. These algorithms
are said to follow the “divide and conquer” approach, which is one of
the most important algorithm design paradigms. It can be viewed as a
general problem solving strategy that consists of breaking up a problem
into several self-similar subproblems whose size is a fraction of the origi-
nal problem’s size. The approach is therefore closely related to recursion
(iterative algorithms can also follow the divide and conquer strategy),
since it relies on recursive problem decomposition (see Figure 1.4). The
function in Listing 1.2 and the third method of Listing 1.5 constitute
examples of this algorithm design approach.
In some references the term “divide and conquer” is also applied to
algorithms where the decomposition is based on a single subproblem
171
Listing 6.1 Function that determines whether a list is sorted in ascending

order.
1 def is_list_sorted(a):
2 n = len(a)
3 if n <= 1:
4 return True
5 else:
6 return (is_list_sorted(a[0:n // 2])
7 and a[n // 2 - 1] <= a[n // 2]
8 and is_list_sorted(a[n // 2:n]))
of half the size of the original’s (the method would invoke itself only
once). However, many authors consider that the term should only be
used when the solution relies on breaking up a problem into two or
more subproblems. This book adopts this last convention. Thus, the
methods that invoke themselves once, dividing the size of the problem
by two, have been covered in earlier chapters (mostly in Chapter 5).
The following sections describe classical recursive divide and conquer
algorithms.
6.1 IS A LIST SORTED IN ASCENDING ORDER?

In this problem the input is a list a of n items that can be sorted through
the ≤ operator (or some function that allows us to compare elements by
implementing a total order binary relation), and the output is a Boolean
value that can be expressed as:
f (a) = ⋀ (ai ≤ ai+1 ).

n−2
(6.1)
i=0
Thus, the result is True when the elements of the list appear in (non-
strictly) increasing or nondecreasing order.
The size of this problem is the number of elements in the list. If the
list contains one element the result is trivially True. In addition, we can
consider that an empty list is also ordered.
The problem can be solved by a linear-recursive method that reduces
the size of the problem one unit in the decomposition stage. However,
we will present a solution that decomposes the problem by dividing the
input list in two halves, and which leads to the following partition of
Multiple Recursion I: Divide and Conquer 173
(6.1):
f (a) = (a0 ≤ a1 ) ∧ ⋯ ∧ (a⌊n/2⌋−2 ≤ a⌊n/2⌋−1 )

´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
Subproblem 1
∧ (a⌊n/2⌋−1 ≤ a⌊n/2⌋ ) ∧ (a⌊n/2⌋ ≤ a⌊n/2⌋+1 ) ∧ ⋯ ∧ (an−2 ≤ an−1 ) .

´¹¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¹ ¶
Subproblem 2
Clearly, if a list is sorted in ascending order then both halves must also be
sorted in ascending order. Thus, the method can invoke itself twice with
the two corresponding sublists, and perform a logical AND operation
with their result. Finally, the combination of the results of the subprob-
lems needs an additional step. In particular, the last element of the first
sublist (a⌊n/2⌋−1 ) must be less than or equal to the first element of the
second sublist (a⌊n/2⌋ ). This condition also requires another AND oper-
ation in the recursive case. Listing 6.1 shows a possible implementation
of the method, which runs in O(n) time, since its cost is characterized
by:
⎧
⎪
⎪1 if n ≤ 1,
T (n) = ⎨
⎪2T (n/2) + 1 if n > 1.
⎪
⎩
6.2 SORTING
Sorting a general list (or array, sequence, etc.) is one of the most studied
problems in computer science. It can be solved by numerous algorithms
that can be used to introduce essential concepts related to computational
complexity and runtime analysis, algorithm design paradigms, or data
structures. The sorting algorithms covered in this chapter will assume
that the input to the problem is a list a of n real numbers. The out-
put is a rearrangement (or permutation) of the elements that provides
another list a′ where a′i ≤ a′i+1 , for i = 0, . . . , n − 2 (≤ can be thought of
as Boolean function that allows us to determine if an element precedes
another, implementing a total order binary relation). The choice of real
numbers for the type of the elements of the list implies that the algo-
rithms need to carry out comparisons through ≤. It can be shown that
any algorithm that solves this problem requires Ω(n log n) comparisons
(i.e., at least on the order of n log n decisions using ≤). Lastly, there exist
algorithms for sorting lists that do not need to compare elements, which
require Ω(n) operations. However, the elements of the list must satisfy
certain conditions in order to apply them. For example, the “counting
merge_sort
3 8 5 3 5 0 4 9
merge_sort merge_sort
3 3 5 8 0 4 5 9
merge
0 3 3 4 5 5 8 9
Figure 6.1 Merge sort algorithm.
sort” algorithm sorts integers that belong to a small interval [0, k] (see
Exercise 5.3).
6.2.1 The merge sort algorithm

The merge sort algorithm is one of the most extensively used examples
to show the potential of the divide and conquer approach. While many
sorting algorithms run in O(n2 ) time in the worst case (see Section 4.4.1
and Exercise 4.17), the merge sort algorithm runs in Θ(n log n). In other
words, its efficiency is optimal from a computational complexity perspec-
tive.
The size of the problem is the length of the list (n). The smallest
instances occur when n ≤ 1, where the algorithm can simply return the
input list. In the recursive case the algorithm decomposes the problem by
Listing 6.2 Merge sort method.

1 def merge_sort(a):
2 n = len(a)
3 if n <= 1:
4 return a
5 else:
6 a1 = merge_sort(a[0:n // 2])
7 a2 = merge_sort(a[n // 2:n])
8 return merge (a1, a2)
dividing the input list in two halves, giving rise to two different subprob-
lems of (roughly) half the size of the original’s. The following diagram
shows a concrete example of the recursive thought process associated
with the algorithm:
Inputs Results
[7, 3, 4, 8, 4, 6, 4, 6] ÐÐÐÐÐÐÐÐÐ→ [3, 4, 4, 4, 6, 6, 7, 8]
merge
[7, 3, 4, 8] ÐÐÐÐÐÐÐÐÐ→ [3, 4, 7, 8]
[4, 6, 4, 6] ÐÐÐÐÐÐÐÐÐ→ [4, 4, 6, 6]
The recursive case is more complicated than in previous examples, since

the final sorted list cannot be obtained through a simple operation (e.g.,
addition, computing the maximum of two numbers, concatenating two
lists, etc.). In this case, it is necessary to solve a new problem in order
to combine the solutions to the subproblems. In particular, the recursive
case needs to “merge” two sorted lists (the outputs of the subproblems)
in order to produce the final sorted list. Assuming we have implemented
a function merge that solves this problem (see below), the succinct code
in Listing 6.2 implements the merge sort method. Note the readability of
the code and the simplicity of the algorithm, which conforms perfectly
to the divide and conquer strategy.
The runtime cost for this algorithm is:
⎧
⎪
⎪1 if n <= 1,
T (n) = ⎨
⎪
⎩2T (n/2) + f (n) if n > 1,
⎪
where we have also assumed that the partition of the list can be obtained
in a constant number of operations. In other words, we have considered
that it is possible to obtain a[0:n//2] and a[n//2:n] in Θ(1) time. In
addition, f (n) measures the number of operations needed by the merge
method in order to combine two sorted lists of (approximately) n/2 ele-
ments. In this case, due to the master theorem (see (3.28)), if f (n) were
a linear function, the merge sort algorithm would run in the (optimal)
order of Θ(n log n). We will see shortly that indeed it is possible to solve
the merge problem in linear time. In general, when faced with a similar
problem, algorithm designers focus their efforts on developing the most
efficient combination method, which may reduce the order of growth of
the divide and conquer algorithm.
The inputs to the merging problem are two sorted lists a and b,
of lengths na and nb , respectively. The output is another sorted list of
length n = na + nb . We can interpret that the size of the problem is the
number of operations needed by the algorithm until it can return a trivial
answer. In this scenario, the size of the problem is m = min(na , nb ), since
the solution to the problem if one of the lists is empty is obviously the
other list. This constitutes the base cases. For the recursive case we can
use the following diagram with the sublists used in the previous example:
Inputs Results
([3, 4, 7, 8],[4, 4, 6, 6]) ÐÐÐÐÐÐ→ [3, 4, 4, 4, 6, 6, 7, 8]
concatenate [3]
([4, 7, 8],[4, 4, 6, 6]) ÐÐÐÐÐÐ→ [4, 4, 4, 6, 6, 7, 8]
The decomposition reduces the size of the problem by a unit, by dis-

carding the smallest element in either list. Naturally, it must appear at
the initial position of the lists, since they are sorted in ascending order.
In the example, the smallest element (3) is found in the first input list.
Clearly, the recursive case simply needs to concatenate this element with
the result of the subproblem. Listing 6.3 shows the corresponding code.
Its runtime cost is characterized by the following function:
⎧
⎪
⎪1 if m = 0,
T (m) = ⎨
⎪
⎩T (m − 1) + 1 if m > 0,
⎪
assuming that the tail of a list (i.e., x[1:]) can be obtained in constant
time. Since its nonrecursive expression is T (m) = m + 1, the merge func-
Listing 6.3 Method for merging two sorted lists.

1 # lists a and b are sorted in ascending order
2 def merge(a, b):
3 if a == []:
4 return b
5 elif b == []:
6 return a
7 else:
8 if a[0] < b[0]:
9 return [a[0]] + merge (a[1:], b)
10 else:
11 return [b[0]] + merge (a, b[1:])
tion runs in linear time with respect to m, and also n, which implies
that the merge sort algorithm runs in Θ(n log n).
6.2.2 The quicksort algorithm

The quicksort algorithm is another method developed by Tony Hoare. It
is based on the divide and conquer approach and receives its name due
to its remarkable efficiency. Unlike the merge sort algorithm, its running
time in the worst case is O(n2 ). However, in the best and average cases
it runs in Θ(n log n) time, and can be several times quicker than the
merge sort algorithm in practice.
To understand the difference between both methods, observe that the
decomposition in merge sort is easy. In particular, the input list can be
divided in two halves by simply using appropriate index ranges (0:n//2
and n//2:n). However, combining the results of the subproblems requires
solving yet another problem (merge) that is not immediately straight-
forward. In contrast, in the quicksort algorithm the decomposition is
hard, but the combination stage is not only trivial, but may not even be
necessary in some implementations.
Specifically, instead of simply dividing the list in two halves, quick-
sort’s decomposition transforms the input by using a partitioning scheme
like the ones presented in Section 5.4. Figure 6.2 illustrates this type of
decomposition, where the sublists within the original list are separated
by a pivot (the sublist to the right of the pivot contains elements that are
less than or equal to the pivot, and the sublist to the left is composed of
numbers that are greater than the pivot). A concrete recursive diagram
based on the previous decomposition could be:
Pivot
6 4 1 7 4 7 3 6 5
Decompose
3 4 1 4 6 7 7 6 5
Figure 6.2 Decomposition of the quicksort algorithm.
Inputs Results
[6, 4, 1, 7, 4, 7, 3, 6, 5] ÐÐÐÐÐ→ [1, 3, 4, 4, 5, 6, 6, 7, 7]
concatenate
[4] in between
sorted sublists
[4, 1, 3] ÐÐÐÐÐ→ [1, 3, 4]

[6, 7, 7, 6, 5] ÐÐÐÐÐ→ [5, 6, 6, 7, 7]
The diagram clearly shows that sorting the original list simply requires
solving the two subproblems through recursive calls, and concatenating
the sorted sublists while maintaining the pivot in between them. Thus,
after solving the subproblems recursively, the combination of the respec-
tive solutions is straightforward. Lastly, an important detail regarding
this decomposition is that the pivot is removed from the list that con-
tains the elements that are smaller than or equal to it. This is necessary
in order to guarantee that the size of the subproblem is indeed smaller
than the size of the original problem.
Listing 6.4 implements a slower variant of the method that is based
on the basic partitioning schemes described in Section 5.4.1. Firstly, it
checks whether the input corresponds to the base case (which is the
same as in the merge sort method). In the recursive case a common
strategy is to consider that the pivot is the first element of the list.
With this choice the worst case occurs when the input list is already
sorted. Moreover, the algorithm would also perform poorly if the input
Listing 6.4 Variant of the quicksort algorithm.

1 def quicksort_variant(a):
2 n = len(a)
3 if n <= 1:
4 return a
5 else:
6 pivot = a[n // 2]
7 v1 = get_smaller_than_or_equal_to(a, pivot)
8 v1.remove (pivot)
9 v2 = get_greater_than(a, pivot )
10 return (quicksort_variant(v1) + [pivot ]
11 + quicksort_variant(v2))
Listing 6.5 In-place quicksort algorithm.

1 def quick_sort_inplace(a, lower , upper ):
2 if lower < upper :
3 pivot_index = partition_Hoare_wrapper(a, lower , upper)
4
5 quick_sort_inplace(a, lower , pivot_index - 1)
6 quick_sort_inplace(a, pivot_index + 1, upper )
is almost sorted. Since these situations may occur frequently in practice,

the function uses the middle element of the input list instead as the pivot.
Subsequently, the method obtains the two sublists associated with the
decomposition (removing the pivot from the sublist that contains it), and
finally concatenates the results of the subproblems, leaving the pivot in
between them.
The merge sort method in Listing 6.2, and the quicksort function
in Listing 6.4, do not alter the input list and return their result in a
new list, which requires twice as much storage space. These types of
algorithms are denoted as “out-of-place.” Instead, it is possible to imple-
ment variants that overwrite the input list, which do not need to allocate
storage space for a full list of length n. These algorithms are known as
“in-place.” Listing 6.5 shows an in-place variant of the quicksort algo-
rithm that uses lower and upper limits to indicate the boundaries of a
sublist within the original one. The code includes the recursive version
of Hoare’s partitioning method described in Listing 5.13 (naturally, the
iterative version, which is more efficient, can be used as well), which
carries out the decomposition of the problem.
The runtime cost analysis of the algorithm is similar to that of the

quickselect algorithm. The main difference is that, instead of solving one
subproblem, quicksort invokes itself twice. The best case occurs when the
pivot is always located at the middle position of the list. In that case,
the running time can be characterized by:
⎧
⎪
⎪1 if n ≤ 1,
T (n) = ⎨
⎪
⎩2T (n/2) + cn if n > 1,
⎪
where T (n) ∈ Θ(n log n). Instead, the worst case occurs when the pivot
is always located at an extreme of the list. In that case the runtime cost
is determined by:
⎧
⎪
⎪1 if n ≤ 1,
T (n) = ⎨
⎪
⎩T (n − 1) + cn if n > 1,
⎪
which is a quadratic function (T (n) ∈ Θ(n2 )).
6.3 MAJORITY ELEMENT IN A LIST

A list is said to have a “majority element” if more than half of its entries
are identical. In this classical problem the goal consists of determining
whether a list a of length n has a majority element, and, if so, to find
that element. In addition, the problem assumes that the elements can
only be compared in Θ(1) (constant) time by using the equality (==)
operator. In other words, we are not allowed to compare them through
relations such as <, >, ≤, or ≥ (the examples will use integers, but not
those operators).
There are several divide and conquer variants that solve the problem.
We will consider a method that returns a tuple (it could also be a list)
containing three values that indicate:
a) Whether the list contains a majority element (Boolean).
b) The majority element if it exists. Otherwise this value is irrelevant

(the function could simply return None).
c) The number of occurrences of the majority element (0 if it does

not exist).
The size of the problem is n. A first base case can correspond to an
empty list, where the result would be the tuple (False, x, 0). The value
x would be irrelevant. Additionally, if the list contains one element then

the function would return (True, a0 , 1).
The recursive case can rely on decomposing the list in two halves,
and solving the corresponding subproblems. We can form initial recursive
diagrams with concrete examples in order to understand how we can
combine the solutions to the subproblems. For example:
Inputs Results
a = [4, 4, 5, 1, 4, 2, 4, 3] ÐÐÐÐÐÐ→ (False, -, 0)
false result
b = [4, 4, 5, 1] ÐÐÐÐÐÐ→ (False, -, 0)

c = [4, 2, 4, 3] ÐÐÐÐÐÐ→ (False, -, 0)
In this case the result of both subproblems is False, which implies that
the input list cannot contain a majority element (even though there are
exactly n/2 occurrences of the element 4 in the example, it is not enough
to produce a true result). In general, this can be shown as follows. Firstly,
the initial list is divided into a sublist b of length ⌊n/2⌋, and another
sublist c of length ⌈n/2⌉, regardless of whether n is even or odd, since
n = ⌊n/2⌋ + ⌈n/2⌉. If the result for both subproblems is False then an
element can appear at most ⌊⌊n/2⌋/2⌋ times in b, and ⌊⌈n/2⌉/2⌋ times
in c. Adding these quantities yields:
⌊n/2⌋ ⌈n/2⌉ ⌊n/2⌋ + ⌈n/2⌉

⌊ ⌋+⌊ ⌋≤⌊ ⌋=⌊ ⌋≤ .
n n
2 2 2 2 2
Therefore, we can conclude that an element cannot appear more than

n/2 times in the initial list.
Another concrete diagram could be:
Inputs Results
a = [4, 4, 5, 4, 1, 2, 4, 3] ÐÐÐÐÐ→ (False, -, 0)
3 + #(c, 4) > n/2 ?

´¹¹ ¹ ¹ ¹ ¹ ¸ ¹ ¹ ¹ ¹ ¹ ¶
=1
b = [4, 4, 5, 4] ÐÐÐÐÐ→ (True, 4, 3)

c = [1, 2, 4, 3] ÐÐÐÐÐ→ (False, -, 0)
Listing 6.6 Code for counting the number of times an element appears in
a list.
1 def occurrences_in_list(a, x):
2 if a == []:
3 return 0
4 else:
5 return int(a[0] == x) + occurrences_in_list(a[1:], x)
In this case the element 4 appears three times in the first sublist, and
is therefore a majority element since 3 > n/2 = 2. The algorithm must
therefore count the number of occurrences of 4 in the second list (denoted
through #(c, 4)) in order to determine whether it is also a majority
element of the initial list a. This can be computed through a simple
linear-recursive function (see Listing 6.6) that receives an input list a
and an element x. If a is empty the result is obviously 0. Otherwise, the
output can consist of the method applied to the tail of a (a1..n−1 ) and
x, plus a unit only if a0 = x.
Listing 6.7 shows a possible implementation of the function that
solves the majority element problem. Lines 3–6 code the base cases.
Lines 8 and 9 decompose the input list into two halves. Line 11 invokes
the method on the first sublist, and if there exists a majority element
(line 12), then line 13 computes the number of occurrences of the el-
ement in the second sublist. If the total number of occurrences of the
element (in both sublists) is greater than n/2 (line 14), then the method
returns a tuple in line 15 with the values: True, the majority element,
and the number of times it appears in the input list. Lines 17–21 are
analogous, but switch the roles of the sublists. Finally, if the function
has not returned, then the list does not contain a majority element (line
23).
In the recursive cases the method needs to invoke itself twice with
one half of the input list, and also needs to compute the occurrences
of an element on two sublists of length n/2 (approximately). Since this
last auxiliary function runs in linear time, the time complexity of the
method can be characterized by:
⎧
⎪
⎪1 if n ≤ 1,
T (n) = ⎨
⎪
⎩2T (n/2) + cn if n > 1.
⎪
Listing 6.7 Code for solving the majority element problem.

1 def majority_element_in_list(a):
2 n = len(a)
3 if n == 0:
4 return (False , None, 0)
5 elif n == 1:
6 return (True, a[0], 1)
7 else:
8 b = a[0:n // 2]
9 c = a[n // 2:n]
10
11 t = majority_element_in_list(b)
12 if t[0]:
13 occurrences = occurrences_in_list(c, t[1])
14 if t[2] + occurrences > n / 2:
15 return (True, t[1], t[2] + occurrences)
16
17 t = majority_element_in_list(c)
18 if t[0]:
19 occurrences = occurrences_in_list(b, t[1])
20 if t[2] + occurrences > n / 2:
21 return (True, t[1], t[2] + occurrences)
22
23 return (False , None, 0)
Therefore, the order of growth of the algorithm is Θ(n log n) (see (3.28)).
Lastly, this problem can be solved through the Boyer–Moore majority
vote algorithm in linear time.
6.4 FAST INTEGER MULTIPLICATION

The classical algorithm taught in school to multiply two nonnegative n-
digit integers requires n2 digit-times-digit multiplications. In this section
we will analyze Karatsuba’s algorithm, which is a faster approach. The
method can be applied to numbers expressed in any base, but we will
focus on multiplying binary numbers. In particular, let x and y be two
nonnegative integers represented by bx and by bits, respectively. Applying
a divide and conquer approach, each binary number can be partitioned
in two as follows:
x = a ⋅ 2m + b,
(6.2)
y = c ⋅ 2m + d,
where m = min(⌊bx /2⌋, ⌊by /2⌋). For example, for x = 594, and y = 69, the
decomposition is:
x = 10010100102 = 1001010 010 = 74 ⋅ 23 + 2,

´¹¹ ¹ ¹ ¹ ¹ ¹ ¸¹ ¹ ¹ ¹ ¹ ¹ ¶ °
a=74 b=2
10001012 = 1000 101 = 8 ⋅ 23 + 5,

±°
y =
c=8 d=5
where bx = 10, by = 7, m = 3, a = 74, b = 2, c = 8, and d = 5. Thus,

the smaller number (in this case, y) is partitioned in two parts with
(roughly) the same number of bits, and the numbers associated with the
lower-significant parts (b and d) are expressed with the same number of
bits.
In many programming languages it is possible to compute the values
of a, b, c, and d by relying on bit shift operations. In Python they can be
carried out through the << and >> operators. On the one hand, (x << m)
is equivalent to x2m , which shifts the bits of x m times towards the
left, appending m least-significant zeros. On the other hand, (x >> m)
performs ⌊x/2m ⌋, shifting the bits of x m times towards the right, which
discards those m bits. These bitwise operations are therefore useful for
multiplying (or performing an integer division) by a power of two, and
can be used as follows to decompose x and y according to (6.2):
a = x >> m,
b = x − (a << m),
c = y >> m,
d = y − (c << m),
where the parentheses are necessary due to operator precedence rules.

According to the decomposition, the product of x and y can be writ-
ten as:
xy = (a ⋅ 2m + b)(c ⋅ 2m + d) = ac22m + (ad + bc)2m + bd. (6.3)
This initial (naive) approach can break up the original problem (a mul-
tiplication) into four smaller subproblems: ac, ad, bc, and bd, which can
be computed through four recursive calls (we can ignore the cost of
multiplications times powers of two, since these can be implemented
very efficiently as bit shifts). However, it is not more efficient than the
Listing 6.8 Karatsuba’s fast algorithm for multiplying two nonnegative

integers.
1 def number_of_bits(n):
2 if n < 2:
3 return 1
4 else:
5 return 1 + number_of_bits(n >> 1)
6
7
8 def multiply_karatsuba(x, y):
9 if x == 0 or y == 0:
10 return 0
11 elif x == 1:
12 return y
13 elif y == 1:
14 return x
15 else:
16 n_bits_x = number_of_bits(x)
17 n_bits_y = number_of_bits(y)
18
19 m = min(n_bits_x // 2, n_bits_y // 2)
20
21 a = x >> m
22 b = x - (a << m)
23 c = y >> m
24 d = y - (c << m)
25
26 ac = multiply_karatsuba(a, c)
27 bd = multiply_karatsuba(b, d)
28 t = multiply_karatsuba(a + b, c + d) - ac - bd
29
30 return (ac << (2 * m)) + (t << m) + bd
“school” method, since it also requires n2 bit-times-bit multiplications

for two n-bit numbers. The algorithm proposed by Karatsuba is able to
reduce such quantity to approximately n1.585 by rearranging the terms
and including more addition/subtraction operations, which turn out to
be negligible regarding asymptotic computational complexity. In partic-
ular, the product xy can be reformulated as:
xy = ac22m + [(a + b)(c + d) − ac − bd] 2m + bd, (6.4)
which requires only three simpler products: ac, bd, and (a + b)(c + d),
leading to a faster algorithm that only carries out three recursive calls.
Listing 6.8 implements Karatsuba’s method, which contains an aux-

iliary function that computes the number of bits of the binary repre-
sentation of a nonnegative integer n (which is equal to ⌊log2 n⌋ + 1, for
n ≥ 1). The size of the problem can be min(bx , by ). Thus, the base cases
can check if the inputs are equal to zero or one, which lead to trivial
results. The recursive case implements (6.4), using three recursive calls.
Regarding its efficiency, consider that both inputs have the same
number of bits n, which is a power of two (i.e., n = 2k ). We can make
this assumption since it is always possible to append 0 leading bits to the
input integers until the assumption is satisfied. In that case, the running
time of the algorithm is characterized by:
⎧
⎪
⎪1 if n ≤ 1,
T (n) = ⎨
⎪
⎩3T (n/2) + cn + d if n > 1.
⎪
The term cn + d is due to additions, subtractions, shift operations, and
calls to number_of_bits, which are carried out in linear time with re-
spect to n. Applying the master theorem, T (n) ∈ Θ(nlog2 3 ) = Θ(n1.585... ).
Thus, the method is more efficient than the approach related to (6.3),
whose computational cost is described by:
⎧
⎪
⎪1 if n ≤ 1,
T (n) = ⎨
⎪
⎩4T (n/2) + en + f if n > 1,
⎪
where T (n) ∈ Θ(n2 ).
Karatsuba’s algorithm computes less multiplications (which corre-
spond to recursive calls) than the school method, but performs more
additions and subtractions. In practice, the algorithm will be quicker for
large values of n. However, if n is small, the extra operations may make
it run slower than the traditional approach. In any case, the method is
noteworthy since it is able to reduce the number of multiplication opera-
tions, which are considerably more costly than additions or subtractions.
6.5 MATRIX MULTIPLICATION

Two matrices can be multiplied by partitioning them into block matrices
as described in Section 2.4. We will now examine a straightforward di-
vide and conquer recursive method that requires n3 elementary (number)
multiplications in order to compute the product of two n×n matrices (de-
fined through the NumPy package). In addition, we will cover Strassen’s
algorithm, which is able to obtain the result using approximately n2.8
basic scalar multiplications.
6.5.1 Divide and conquer matrix multiplication

Let A and B be p × q and q × r-dimensional matrices, respectively. The
size of the problem depends on the three dimensions p, q, and r. A trivial
base case occurs when p = q = r = 1, where the result is a simple scalar
number. Additionally, some implementations may require considering
situations where a dimension is 0. In those cases the output should be
an empty matrix, as will be addressed shortly.
One fairly straightforward way to decompose the problem consists of
partitioning each matrix into four block matrices (forming a 2 × 2 array
of block matrices). In that case their product can be defined as follows:
AB = [ ] [ 1,1 ]
A1,1 A1,2 B B1,2
A2,1 A2,2 B2,1 B2,2
(6.5)
A B + A1,2 B2,1 A1,1 B1,2 + A1,2 B2,2
= [ 1,1 1,1 ].
A2,1 B1,1 + A2,2 B2,1 A2,1 B1,2 + A2,2 B2,2
Notice that the formula is analogous to multiplying two 2 × 2 matrices.
For example, the top-left block of the result (A1,1 B1,1 + A1,2 B2,1 ) can be
viewed as the product between the first (block) row of A and the first
(block) column of B.
The decomposition involves computing eight simpler matrix prod-
ucts. Thus, the method will invoke itself eight times in the recursive
case. The results of each product need to be added and stacked appro-
priately in order to form the output matrix. Listing 6.9 shows a possi-
ble implementation. The recursive case first defines each of the smaller
block matrices, adds the simpler products, and builds the output matrix
through the methods vstack and hstack. One of the base cases com-
putes a simple product when p = q = r = 1. In addition, the code also
considers the possibility of receiving empty input matrices, since they
appear when partitioning the matrices in the recursive case if one of the
dimensions is equal to one (obviously, a vector cannot be partitioned
into four vectors as described in (6.5)). Thus, if any of the dimensions is
0 a special base case returns an empty matrix of dimensions p × r, which
can be handled appropriately in Python.
The previous method creates 1 × 1 matrices (in a base case) and
progressively stacks them together to form the final p × r matrix. In
addition, note that the dimensions of the input matrices to the methods
are not fixed.
Another more efficient alternative consists of passing the entire ma-
trices A and B in each call, and specifying the blocks that need to be
Listing 6.9 Divide and conquer matrix multiplication.

2
3
4 def matrix_mult(A, B):
5 p = A.shape [0]
6 q = A.shape [1]
7 r = B.shape [1]
8
9 if p == 0 or q == 0 or r == 0:
10 return np.zeros ((p, r))
11 elif p == 1 and q == 1 and r == 1:
12 return np.matrix ([[A[0, 0] * B[0, 0]]])
13 else:
14 A11 = A[0:p // 2, 0:q // 2]
15 A21 = A[p // 2:p, 0:q // 2]
16 A12 = A[0:p // 2, q // 2:q]
17 A22 = A[p // 2:p, q // 2:q]
18
19 B11 = B[0:q // 2, 0:r // 2]
20 B21 = B[q // 2:q, 0:r // 2]
21 B12 = B[0:q // 2, r // 2:r]
22 B22 = B[q // 2:q, r // 2:r]
23
24 C11 = matrix_mult(A11 , B11) + matrix_mult(A12 , B21)
28
29 return np.vstack ([np.hstack ([C11 , C12]),
30 np.hstack ([C21 , C22])])
31
32
33 A = np.matrix ([[2, 3, 1, -3], [4, -2, 1, 2]])
34 B = np.matrix ([[2, 3, 1], [4, -1, -5], [0, -6, 3], [1, -1, 1]])
35 print (matrix_mult(A, B))
multiplied through appropriate limits passed as parameters, similarly to

Listing 1.6. In addition, the result can be stored in a p × r matrix param-
eter C (passed by reference). Listing 6.10 shows a possible implemen-
tation of this alternative solution. The method matrix_mult_limits
always passes the entire matrices A and B in each call, storing the re-
sult in a p × r matrix (its third parameter). Additionally, it specifies the
submatrices that it will actually multiply through the rest of the param-
Listing 6.10 Alternative divide and conquer matrix multiplication.

2
3
4 def add_matrices_limits(A, B, C, lp, up, lr, ur):
5 for i in range (lp, up + 1):
6 for k in range(lr, ur + 1):
7 C[i, k] = A[i, k] + B[i, k]
8
9
10 def matrix_mult_limits(A, B, C, lp, up, lq, uq, lr, ur):
11 mp = (lp + up) // 2
12 mq = (lq + uq) // 2
13 mr = (lr + ur) // 2
14
15 if lp == up and lq == uq and lr == ur:
16 C[mp, mr] = A[mp, mq] * B[mq, mr]
17 elif lp <= up and lq <= uq and lr <= ur:
18
19 M1 = np.zeros ((A.shape [0], B.shape [1]))
20 M2 = np.zeros ((A.shape [0], B.shape [1]))
21
22 matrix_mult_limits(A, B, M1, lp, mp, lq, mq, lr, mr)
23 matrix_mult_limits(A, B, M2, lp, mp, mq + 1, uq, lr, mr)
24 add_matrices_limits(M1, M2, C, lp, mp, lr, mr)
25
26 matrix_mult_limits(A, B, M1, lp, mp, lq, mq, mr + 1, ur)
27 matrix_mult_limits(
28 A, B, M2, lp, mp, mq + 1, uq, mr + 1, ur)
29 add_matrices_limits(M1, M2, C, lp, mp, mr + 1, ur)
30
31 matrix_mult_limits(A, B, M1, mp + 1, up, lq, mq, lr, mr)
33 A, B, M2, mp + 1, up, mq + 1, uq, lr, mr)
34 add_matrices_limits(M1, M2, C, mp + 1, up, lr, mr)
35
37 A, B, M1, mp + 1, up, lq, mq, mr + 1, ur)
39 A, B, M2, mp + 1, up, mq + 1, uq, mr + 1, ur)
40 add_matrices_limits(M1, M2, C, mp + 1, up, mr + 1, ur)
41
42
43 def matrix_mult_limits_wrapper(A, B):
44 C = np.zeros ((A.shape [0], B.shape [1]))
45 matrix_mult_limits(A, B, C, 0, A.shape [0] - 1,
46 0, A.shape [1] - 1, 0, B.shape [1] - 1)
47 return C
eters, which indicate lower and upper limits related to the dimensions
p, q, and r. The base case of matrix_mult_limits occurs when both
of the submatrices correspond to scalar numbers, say ai,j and bj,k . In
that case the method simply stores their product in row i and column
k of C. Lastly, the method is not a function, since it does not return a
matrix. Instead, it is a procedure that modifies the parameter C, where
it stores the result. Finally, the iterative method add_matrices_limits
adds the elements of submatrices passed as the first two matrix input
parameters, and stores the result in its third parameter (the submatrices
are specified through parameter limits).
6.5.2 Strassen’s matrix multiplication algorithm

The most expensive arithmetic operation carried out by the previous
algorithms is the scalar multiplication in the base cases. In particular,
both require pqr of these multiplications, similarly to the straightforward
iterative version that uses three loops. It is nevertheless interesting to
analyze the time complexity assuming that the input matrices are n × n.
In that case the runtime can be specified through the following function:
⎧
⎪
⎪1 if n ≤ 1,
T (n) = ⎨
⎪
(6.6)
⎩8T (n/2) + 4Θ(n ) if n > 1,
⎪ 2
since the methods invoke themselves eight times, and need to perform
four matrix additions whose cost is quadratic with respect to n. There-
fore, according to the master theorem (see (3.28)), T (n) ∈ Θ(nlog2 8 ) =
Θ(n3 ). We will now describe Strassen’s algorithm, which is a well-known
method that can reduce the time complexity to Θ(nlog2 7 ) = Θ(n2.807... ).
The method also decomposes each of the input matrices into four
block matrices as in the standard algorithm. Thus, AB = C can be
expressed as:
[ ] [ 1,1 ] = [ 1,1 ]
A1,1 A1,2 B B1,2 C C1,2
A2,1 A2,2 B2,1 B2,2 C2,1 C2,2
The key to the method is the definition of the following new matrices
that involve one matrix multiplication operation:
M1 = (A1,1 + A2,2 )(B1,1 + B2,2 ),
M2 = (A2,1 + A2,2 )B1,1 ,
M3 = A1,1 (B1,2 − B2,2 ),
M4 = A2,2 (B2,1 − B1,1 ), (6.7)
M5 = (A1,1 + A1,2 )B2,2 ,
M6 = (A2,1 − A1,1 )(B1,1 + B1,2 ),
M7 = (A1,2 − A2,2 )(B2,1 + B2,2 ).
Finally, these matrices can be combined as follows to form the output’s
block matrices:
C1,1 = M1 + M4 − M5 + M7 ,
C1,2 = M3 + M5 ,
(6.8)
C2,1 = M2 + M4 ,
C2,2 = M1 − M2 + M3 + M6 .
The algorithm therefore computes seven products and 18 additions (or
subtractions) in every recursive call. Thus, its runtime cost is described
through:
⎧
⎪
⎪1 if n ≤ 1,
T (n) = ⎨
⎪
(6.9)
⎩7T (n/2) + 18Θ(n ) if n > 1,
⎪ 2
which implies that T (n) = Θ(nlog2 7 ) = Θ(n2.807... ). The algorithm can

be faster than the standard method that runs in Θ(n3 ) for large values
of n. Nevertheless, for small or medium-sized matrices it may be slower
due to the larger multiplicative constants that play a role in practice.
Finally, from a theoretical point of view, the inputs for this algo-
rithm need to be square n × n matrices, where n is a power of two. Nev-
ertheless, in practice efficient implementations split the matrices into
numerous square submatrices, and apply the algorithm repeatedly. A
simpler, but slower, alternative consists of padding (i.e., extending) the
input matrices with zeros, in order for them to be 2k × 2k matrices (see
Exercise 6.6).
6.6 THE TROMINO TILING PROBLEM

A tromino is a polygon formed by connecting three equal-sized squares
by their edges. Without considering rotations and reflections, there are
“I” tromino “L” tromino
Figure 6.3 Types of trominoes ignoring rotations and reflections.
n = 2k
n = 2k
Figure 6.4 Tromino tiling problem.
only two types of trominoes: the “I” and the “L” trominoes, as illustrated
in Figure 6.3. The following problem consists of covering a square n × n
board, where n ≥ 2 is a power of two, which contains a “hole” that
cannot be covered, with L trominoes. Figure 6.4 explains the problem
graphically with an example.
The size of the problem is clearly n. The smallest instances of the
problem correspond to 2×2 boards, whose solutions are trivial. Figure 6.5
illustrates the divide and conquer decomposition used in the recursive
case. The initial board in (a) is divided into four smaller square boards
of size n/2, as shown in (b). However, only one of these smaller boards
will contain the initial hole. Therefore, the other three boards do not
n/2 n/2
(a) (b) (c)
Figure 6.5 Decomposition of the tromino tiling problem.

Listing 6.11 Auxiliary functions for drawing trominoes.

1 def draw_L1 (x, y):
2 plt.plot([x, x + 2], [y + 2, y + 2], 'k-')
3 plt.plot([x, x + 1], [y + 1, y + 1], 'k-')
4 plt.plot([x + 1, x + 2], [y, y], 'k-')
5 plt.plot([x, x], [y + 1, y + 2], 'k-')
6 plt.plot([x + 1, x + 1], [y, y + 1], 'k-')
7 plt.plot([x + 2, x + 2], [y, y + 2], 'k-')
8
9
11 plt.plot([x, x + 2], [y + 2, y + 2], 'k-')
12 plt.plot([x, x + 1], [y, y], 'k-')
13 plt.plot([x + 1, x + 2], [y + 1, y + 1], 'k-')
14 plt.plot([x, x], [y, y + 2], 'k-')
15 plt.plot([x + 1, x + 1], [y, y + 1], 'k-')
16 plt.plot([x + 2, x + 2], [y + 1, y + 2], 'k-')
17
18
20 plt.plot([x, x + 2], [y, y], 'k-')
21 plt.plot([x, x + 1], [y + 1, y + 1], 'k-')
22 plt.plot([x + 1, x + 2], [y + 2, y + 2], 'k-')
23 plt.plot([x, x], [y, y + 1], 'k-')
24 plt.plot([x + 1, x + 1], [y + 1, y + 2], 'k-')
25 plt.plot([x + 2, x + 2], [y, y + 2], 'k-')
26
27
29 plt.plot([x, x + 2], [y, y], 'k-')
30 plt.plot([x, x + 1], [y + 2, y + 2], 'k-')
31 plt.plot([x + 1, x + 2], [y + 1, y + 1], 'k-')
32 plt.plot([x, x], [y, y + 2], 'k-')
33 plt.plot([x + 1, x + 1], [y + 1, y + 2], 'k-')
34 plt.plot([x + 2, x + 2], [y, y + 1], 'k-')
constitute self-similar subproblems. This can be solved by placing a tro-

mino in the center of the board, where its three squares will constitute
holes in the smaller boards, creating valid subproblems, as illustrated in
(c).
We will use the package Matplotlib in order to generate images show-
ing the solutions to the problem. In particular, the trominoes can be
drawn by plotting six line segments. Since there are four possible L tro-
minoes considering rotations (see Figure 6.6), we can use the four auxil-
Listing 6.12 Recursive method for drawing trominoes.

1 def trominoes(x, y, n, p, q):
2 if n == 2:
3 if y == q: # hole in bottom tiles
4 if x == p: # hole in bottom -left tile
5 draw_L1 (x, y)
6 else: # hole in bottom -right tile
7 draw_L2 (x, y)
8 else: # hole in top tiles
9 if x == p: # hole in top -left tile
10 draw_L3 (x, y)
11 else: # hole in top -right tile
12 draw_L4 (x, y)
13
14 else:
15 mid_x = x + n // 2
16 mid_y = y + n // 2
17
18 if q < mid_y : # hole in bottom squares
19
20 if p < mid_x : # hole in bottom -left square
21 draw_L1 (mid_x - 1, mid_y - 1)
22 trominoes(x, y, n // 2, p, q)
23 trominoes(x, mid_y , n // 2, mid_x - 1, mid_y)
24 trominoes(mid_x , y, n // 2, mid_x , mid_y - 1)
25 trominoes(mid_x , mid_y , n // 2, mid_x , mid_y)
26 else: # hole in bottom -right square
27 draw_L2 (mid_x - 1, mid_y - 1)
28 trominoes(x, y, n // 2, mid_x - 1, mid_y - 1)
30 trominoes(mid_x , y, n // 2, p, q)
32
33 else: # hole in top squares
34
35 if p < mid_x : # hole in top -left square
36 draw_L3 (mid_x - 1, mid_y - 1)
38 trominoes(x, mid_y , n // 2, p, q)
41 else: # hole top -right square
42 draw_L4 (mid_x - 1, mid_y - 1)
46 trominoes(mid_x , mid_y , n // 2, p, q)
L1 L2 L3 L4
(x, y) (x, y) (x, y) (x, y)
Figure 6.6 L trominoes considering rotations.
Listing 6.13 Code for calling the trominoes method.

1 import random
2 import matplotlib.pyplot as plt
3 from matplotlib.patches import Rectangle
4
5 # Include tromino methods here
6
7 fig = plt.figure ()
8 fig.patch.set_facecolor('white ')
9 ax = plt.gca()
10 n = 16 # power of 2
11 p = random .choice ([i for i in range (n)])
12 q = random .choice ([i for i in range (n)])
13 ax.add_patch(Rectangle((p, q), 1, 1, facecolor=(0.5, 0.5, 0.5)))
14 trominoes(0, 0, n, p, q)
15 plt.axis('equal ')
16 plt.axis('off')
17 plt.show()
iary functions in Listing 6.11 to draw each one. The functions receive the
coordinates (x, y) of the bottom-left corner corresponding to the square
surrounding the tromino. The command plot([x1 , x2 ],[y1 , y2 ],’k-’)
draws a black line segment with endpoints (x1 , y1 ) and (x2 , y2 ).
Listing 6.12 shows a possible implementation of the recursive
method. The procedure needs to know which problem/subproblem it
should solve. This information is provided by the first three parameters.
The first two indicate the bottom-left coordinates (x, y) of the board,
while the third is the size of the board (n). The last two indicate the
location of the hole (in particular, (p, q) specifies the bottom-left corner
of the 1 × 1 square). In both base and recursive cases the method uses
conditions in order to determine the relative position of the hole, and
draws the appropriate tromino. Finally, the method invokes itself four
times in the recursive case, with different parameters indicating the new
subproblems, together with the new holes on three of them.
Finally, Listing 6.13 shows a fragment of code that can be used to call
the trominoes method. Line 7 creates a figure, line 8 sets its background
color to white, and ax captures the axes of the figure in line 9. After defin-
ing the size of the initial board (line 10), the hole is chosen within it at
random, and then drawn in line 13. When using the Matplotlib package
a rectangle can be formed by calling the method Rectangle. It receives
the coordinates of the bottom-left vertex, together with the width and
height, and other possible arguments. Line 14 calls the main method,
and the last lines are included to avoid scaling factors, to eliminate the
axes in the final image, and to draw it.
6.7 THE SKYLINE PROBLEM

The skyline problem consists of finding the outline of a set of rectangular
buildings against the sky. Figure 6.7 explains the idea. The input to the
problem is a list of n ≥ 1 rectangles that represent buildings, as shown
in (a). The bottom side of the rectangles is always located at level 0.
Thus, each building can be specified by using only three parameters. In
particular, we will use tuples of the form (x1 , x2 , h), where x1 marks the
location of left side of the building, x2 indicates the right side, and h
contains its height. Thus, x1 < x2 , and h > 0. The buildings in the exam-
ple are: [(1,7,7), (18,20,7), (2,9,5), (17,19,2), (12,24,3), (3,8,8), (11,13,5),
(15,21,6)]. Note that it is not necessary to sort them in any way (e.g.,
according to the value of x1 ).
The skyline consists of a curve whose height at any point on the X
axis is the maximum height of the buildings at such point, as illustrated
through the thick dark segments in (b). Since the buildings are rectan-
gular the skyline can be specified through a set of coordinates on the
plane (x, h) that mark the location x when its height changes to h, as
illustrated in (c). The output to the problem will therefore be a list of
these coordinates, which can be coded as tuples as well, and will appear
sorted in ascending order. The output for the example is: [(1,7), (3,8),
(8,5), (9,0), (11,5), (13,3), (15,6), (18,7), (20,6), (21,3), (24,0)].
The size of the problem is the number of buildings n. For the base
case, the smallest instance occurs when there is only one building. If it is
specified through the tuple (x1 , x2 , h) then the output is the list [(x1 , h),
(x2 , 0)], as shown in Figure 6.8.
This problem can be decomposed by reducing the size of the problem
by a unit. However, this leads to an algorithm whose runtime is O(n2 ) in
the worst case. Instead, we will examine a divide and conquer approach
5 10 15 20 25
(a)
5 10 15 20 25
(b)
5 10 15 20 25
(c)
Figure 6.7 The skyline problem.

(x1 , h)
(x1 , x2 , h)
(x2 , 0)
Figure 6.8 Base case for the skyline problem with one building.
Listing 6.14 Main recursive method for computing skylines.

1 def compute_skyline(buildings):
2 n = len(buildings)
3 if n == 1:
4 return ([(buildings[0][0], buildings[0][2]) ,
5 (buildings[0][1], 0)])
6 else:
7 skyline1 = compute_skyline(buildings[0:n // 2])
8 skyline2 = compute_skyline(buildings[n // 2:n])
9 return merge_skylines(skyline1 , skyline2 , 0, 0)
that is able to solve the problem in Θ(n log n) time. The idea, illustrated
in Figure 6.9, is similar to the approach used in the merge sort algorithm.
The decomposition step consists of dividing the input list in two smaller
lists of approximately n/2 buildings. The method then carries out two
recursive calls on those sublists that return two independent skylines.
Assuming that the skylines have been constructed correctly (by apply-
ing induction), the final, and challenging, step consists of merging the
skylines in order to produce a final one. Listing 6.14 shows the associated
divide and conquer method, whose structure is essentially identical to
that of Listing 6.2.
The skyline merging problem is a new computational problem in its
own right. While the majority of solutions in texts are iterative, we will
now examine a linear-recursive method. The inputs are the two input
lists of sorted tuples representing skylines. In addition, since a tuple in-
dicates a change in the height of a skyline, the method needs to access
the previous height before such change. Moreover, since the proposed
algorithm will process the first tuples from the lists (until one list is
empty), but progressively discard them as they are analyzed, these pre-
vious heights will not be contained in the input lists. Thus, the method
will need two additional parameters, say p1 and p2 , in order to store the
previous heights of the skylines. Naturally, both of these parameters will
8
6
4
2
5 10 15 20 25
Original problem
8 8
6 6
4 4
2 2
5 10 15 20 25 5 10 15 20 25
Subproblem 1 Subproblem 2
8 8
6 6
4 4
2 2
5 10 15 20 25 5 10 15 20 25
Skyline 1 Skyline 2
8
6
4
2
5 10 15 20 25
Combination problem
8
6
4
2
5 10 15 20 25
Solution
Figure 6.9 Recursive case for the skyline problem.

p1
(x2 , h2 )
p1
(x2 , h2 ) h2 h2
p2 p2
(x1 , h1 ) h1
(x1 , h1 ) h1
x x
(a) (b)
Possible situations when merging skylines that change at the

Figure 6.10
same location x.
be initialized to zero when calling the method within the main skyline
function (see line 9 of Listing 6.14).
The size of the problem depends on the lengths of the input skyline
lists, say n1 and n2 . We can consider that the base case occurs when one
of the lists is empty, where the method must trivially return the other
list. Listing 6.15 codes the method, where the base cases are described
in lines 2–5.
A key observation for determining an appropriate decomposition is
that the output of the merging function produces a new skyline whose
tuples are sorted in ascending order according to their x value. Thus,
the algorithm will analyze the first tuples of each list, and process the
one with a smaller x value (or both if their x values are the same). Thus,
in the recursive case we need the first tuples of the skylines (x1 , h1 ) and
(x2 , h2 ), together with their previous heights p1 and p2 .
Firstly, let us consider the situation when x1 = x2 = x. Since the
tuples mark changes in the skyline, we may need to include in the solu-
tion the one with larger height. For example, in Figure 6.10(a) the point
(x, h2 ) would be included in the final skyline. Furthermore, the recursive
call will use the tails of both input lists, discarding (x, h1 ) and (x, h2 ),
since the possible changes at x will have been processed correctly. This
is accomplished in lines 19–21. Lastly, there is a situation where a new
tuple is not included in the solution. This occurs when the largest new
height is equal to the largest previous height of the skyline (i.e., when
max(h1 , h2 ) = max(p1 , p2 )), since there would be no change of heights
at x. Figure 6.10(b) illustrates this case, where the point (x, h2 ) would
not be included in the final skyline (see lines 16 and 17).
We will now analyze possible scenarios when the x values of the first
tuples of the skylines are not equal. Without loss of generality, assume
Listing 6.15 Recursive method for merging skylines.

1 def merge_skylines(sky1, sky2, p1, p2):
2 if sky1 == []:
3 return sky2
4 elif sky2 == []:
5 return sky1
6 else:
7 x1 = sky1 [0][0]
8 x2 = sky2 [0][0]
9 h1 = sky1 [0][1]
10 h2 = sky2 [0][1]
11
12 if x1 == x2:
13 h = max(p1, p2)
14 new_h = max(h1, h2)
15 if h == new_h :
16 return merge_skylines(sky1[1:], sky2 [1:],
17 h1, h2)
18 else:
19 return ([(x1, new_h)]
20 + merge_skylines(sky1 [1:], sky2[1:],
21 h1, h2))
22
23 elif x1 < x2:
24 if h1 > p2:
25 return ([(x1, h1)]
26 + merge_skylines(sky1 [1:], sky2,
27 h1, p2))
28 elif p1 > p2:
29 return ([(x1, p2)]
30 + merge_skylines(sky1 [1:], sky2,
31 h1, p2))
32 else:
33 return merge_skylines(sky1[1:], sky2,
34 h1, p2)
35
36 else:
37 if h2 > p1:
38 return ([(x2, h2)]
39 + merge_skylines(sky1, sky2[1:],
40 p1, h2))
41 elif p2 > p1:
42 return ([(x2, p1)]
43 + merge_skylines(sky1, sky2[1:],
44 p1, h2))
45 else:
46 return merge_skylines(sky1, sky2[1:], p1, h2)
h1 h1 h1 p1
p2
p1 p1 = p2 p1 h1
p2 p2
x1 x1 x1 x1
h1 > p2 — Include tuple (x1 , h1 )
p1 p1
p2
h1 h1 = p2
x1 x1
h1 ≤ p2 and p1 > p2 — Include tuple (x1 , p2 )
p2 p2
h1 p1 h1 = p2 p1 = p2
p1 h1 p1 h1
x1 x1 x1 x1
Otherwise — Do not include a tuple
Figure 6.11 Possible situations when merging skylines and x1 < x2 .
x1 < x2 . In that case, the algorithm must decide whether to include the
tuple (x1 , h1 ), or (x1 , p2 ), or none at all, as illustrated in Figure 6.11. If
h1 > p2 then the first skyline is above the second one at location x1 , and
therefore must include the tuple (x1 , h1 ) as part of the merged skyline
(see lines 25–27). If h1 ≤ p2 then the algorithm must check if p1 > p2 . If
the result is True then the method includes the tuple (x1 , p2 ) (see lines
29–31). Notice that when h1 < p2 this produces a new tuple that does
not appear in the input skyline lists. Lastly, in other situations, p2 ≥ h1
and p2 ≥ p1 , which implies that the merged skyline will not change at x1 .
Finally, having processed the first tuple from the first skyline (x1 , h1 ), the
method discards it when invoking itself in the corresponding recursive
cases. In addition, the arguments that specify the previous heights of

the skylines will be h1 and p2 (see lines 27, 31, and 34). The rest of the
code at lines 36–46 is analogous to the code at lines 23–34, and handles
the case when x2 < x1 .
6.8 EXERCISES
Exercise 6.1 — Implement a divide and conquer algorithm that deter-
mines whether a list a contains an element x.
Exercise 6.2 — Let a be a list of n nonnegative integers. Write a func-

tion based on the divide and conquer technique that returns the set of
digits shared amongst all of the elements in a, and specify its asymp-
totic computational cost. For example, for a = [2348, 1349, 7523, 3215],
the solution is {3}. The function should call another one that provides
the set of digits in a nonnegative integer. Code this function as well.
Exercise 6.3 — The maximum sublist problem consists of finding

the sublist of contiguous elements within a list of numbers that
has the largest sum of its components. For example, given the list
[−1, −4, 5, 2, −3, 4, 2, −5], the optimal sublist is [5, 2, −3, 4, 2], whose ele-
ments sum up to 10. Given a nonempty input list of numbers a, imple-
ment a divide and conquer function that returns the sum of the elements
of its maximum sublist.
Exercise 6.4 — Design a fast recursive polynomial multiplication al-

gorithm based on a divide and conquer decomposition analogous to the
one used in Karatsuba’s algorithm (see Section 6.4). Code polynomials
through lists, as described in Exercise 5.2. The method will need to call
functions that add and subtract polynomials. Code these functions as
well.
Exercise 6.5 — Implement a recursive function that receives an n × m

matrix (A) and returns its m × n transpose (AT ). Use the NumPy pack-
age, and the following divide and conquer decomposition that breaks up
the input matrix A into four block matrices by dividing each dimension
by two:
A=[ ].
A1,1 A1,2
A2,1 A2,2
In that case, the transpose of A can be defined as:
AT = [ ].
AT1,1 AT2,1
AT1,2 AT2,2
Exercise 6.6 — Implement Strassen’s matrix multiplication algorithm.

The code should include a wrapper function that will allow multiplying
a p × q matrix times a q × r matrix, by padding the input matrices with
zeros.
Exercise 6.7 — Implement a matrix multiplication algorithm by de-

composing the input matrices as follows:
A ⋅ B = [ A1 A2 ] ⋅ [ ] = [ A1 B1 + A2 B2 ] .
B1
B2
Exercise 6.8 — Implement a matrix multiplication algorithm by de-

composing the input matrices as follows:
A⋅B=[ ] ⋅ [ B1 B2 ] = [ 1 1 ].
A1 A B A1 B2
A2 A2 B1 A2 B2
CHAPTER 12
Multiple Recursion III:

Backtracking
He who would search for pearls must dive below.

— John Dryden
ACKTRACKING is one of the most important algorithm design

B paradigms. It can be regarded as an “intelligent” brute-force strat-
egy that performs exhaustive searches for solutions to constraint satis-
faction and discrete optimization problems. The approach can be used to
solve numerous puzzles and problems, including the eight-queens prob-
lem, finding paths through mazes, the sudoku puzzle, the 0-1 knapsack
optimization problem, and many more.
Backtracking methods generally combine recursion and iteration,
contain several parameters, and are not usually designed by strictly
thinking about problem decomposition and induction. Thus, they appear
to be complex to students who study the material for the first time. For-
tunately, backtracking algorithms often share a similar structure, which
can be exploited in order to ease their design. This structure is captured
in what are know as “backtracking templates,” which vary depending on
coding styles and programming languages. In this book the methods will
share a particular structure where they receive a relatively large num-
ber of parameters, and barely rely on additional methods. However, the
reader should be aware that there are other possibilities. In any case, stu-
dents can master backtracking with relative ease by studying examples,
and by applying very similar algorithms to different problems.
353
q
q
q
q
Figure 12.1 One solution to the four-queens puzzle.
12.1 INTRODUCTION
This section introduces fundamental concepts related to backtracking,
and provides an overview of how it works by examining the simple four-
queens puzzle. Its goal consists of placing four chess queens on a 4 × 4
chessboard so that they do not threaten each other. Since queens can
move horizontally, vertically, and diagonally, two (or more) queens can-
not appear in the same row, column, or diagonal on the board. Fig-
ure 12.1 illustrates one of the two possible solutions to the puzzle. Nat-
urally, the problem can be generalized to placing n queens on an n × n
chessboard (see Section 12.3).
12.1.1 Partial and complete solutions

Backtracking is a general strategy for finding solutions to computational
problems among a finite set of possibilities. These solutions are collec-
tions of discrete items. Thus, technically we say that backtracking is a
strategy that searches for solutions in a “discrete state space.” It is also
a brute-force method in the sense that the search is exhaustive. In other
words, if a solution exists, a backtracking algorithm is guaranteed to find
it.
The methods search for solutions by constructing and updating a
partial solution, which could eventually become a valid complete
solution to the problem. Partial solutions are built by incorporating
candidate elements incrementally, step by step, in successive recursive
calls. Thus, there is an implicit order in which the candidates appear
in the solutions. In this regard, partial solutions can be understood as
“prefixes” of complete solutions. When a solution is a list, a partial
solution is simply a sublist that contains the first elements, as illustrated
in Figure 12.2(a) through the lighter-shaded region. Instead, if it is a two-
Multiple Recursion III: Backtracking 355
0 1 2 3 0 4 8 12
4 5 6 7 1 5 9 13
0 1 2 3 4 5 6
8 9 10 11 2 6 10 14
12 13 14 15 3 7 11 15
(a) (b) (c)
Figure 12.2 Partial solutions within complete solutions that are coded as
lists or matrices.
dimensional matrix, we need to consider some linear arrangement of its

elements. The two most natural choices consist of ordering the entries
of a matrix in row or column-major order, as shown in (b) and (c),
respectively. In those cases a partial solution is the collection of the first
elements of the matrix according to one of these linear arrangements.
The key to using backtracking effectively is the possibility of effi-
ciently analyzing whether a partial solution satisfies the constraints of
the problem. If it does, then it could potentially be expanded into a com-
plete valid solution. However, if it is not possible to create a valid partial
solution by including a new item, the algorithm “backtracks” to another
previous valid partial solution, and continues exploring other possibil-
ities not contemplated yet. Backtracking algorithms therefore perform
exhaustive searches for solutions, but can solve some problems efficiently
by avoiding incrementing partial solutions that do not lead to valid com-
plete solutions.
The first task when tackling a problem through backtracking con-
sists of selecting a particular form or data structure for the solutions.
Typically, they are implemented as lists or matrices. For the n-queens
puzzle a first idea that could come to mind is to use a Boolean n × n
matrix to encode the solutions, where (n) True values mark the squares
where the queens lie on the board. A partial solution would have the
same structure, but it would contain less than n queens (as soon as you
place the n-th queen the solution is either not valid, or it is a complete
solution to the problem). However, although it is possible to use a ma-
trix, this choice of data structure does not lead to an efficient algorithm.
Since backtracking methods perform exhaustive searches, an algorithm
would traverse the entire matrix in row or column-major order, deciding
whether to place a queen in each of the chessboard’s squares. This would
0 1
1 3 0 3
2
2
1 3 1 2 0
2 2
Figure 12.3 Recursion tree of a backtracking algorithm that finds one

solution to the four-queens puzzle.
be time-consuming, since, for example, if a queen is placed in a particular

column, then the algorithm does not need to try to place other queens
on that same column. Instead, the method could simply move on to the
next column.
A more appropriate choice consists of using a simple list of length
n. In that case, the indices of the list could represent the columns of
the chessboard, and the particular value could denote the row where a
queen is placed. In other words, if x is a list that represents a solution,
and there is a queen in column i and row j, then we have xi = j. For
example, if columns are numbered from left to right, and rows from top
to bottom, both starting at 0, the solution in Figure 12.1 corresponds to
the list [1, 3, 0, 2]. By using a list, after placing a (valid) queen in a row
for a particular column (i), the algorithm can continue to try to place
a queen in the next column (i + 1). In addition, it can keep track of the
rows where there are queens, and avoid placing further queens in them.
12.1.2 Recursive structure

We can examine how a backtracking algorithm works by analyzing its
recursion tree. Figure 12.3 illustrates it for a backtracking method that
finds one solution to the four-queens puzzle. The root node represents
the first call to the method, where the partial solution, drawn as a 4 × 4
board (but coded as a list), is empty. Subsequently, the order of the

recursive calls is in accordance with a preorder traversal of the tree. The
first recursive call, associated with the root’s left child, places a queen
(depicted as a shaded square) in the first row of the first column. The
label “0” on the corresponding edge indicates the row where a queen is
placed. Since this does not violate the constraints of the problem it can
continue to place queens in the second column. In particular, it starts by
avoiding placing a queen in the first row, since there is already a queen
on it. Subsequently, the algorithm analyzes whether it can place a queen
on the second row. Clearly it cannot, since the two queens would lie on
the same diagonal. The associated partial solution would not be valid,
and the algorithm does not proceed with expanding it. In other words, it
will not carry out any further recursive calls with that (invalid) partial
solution. This is illustrated in the diagram through dotted edges.
Subsequently, the method proceeds by placing a queen in the third
row, which leads to a valid partial solution. Thus, the algorithm can
continue to try to place a queen in the third column. The first free row
is the second one, but this leads to an invalid partial solution. Similarly,
placing a queen on the fourth row violates the constraints of the prob-
lem. Since there are no more options left, the algorithm “backtracks”
to the node where there is only one queen on the first column and first
row. Afterwards, the method explores all of the possible partial solutions
that could be obtained when placing a queen on the fourth row of the
second column. Since there are no valid solutions for this option either,
the algorithm will end up backtracking to the initial node. Note that
at this stage the method will have carried out an exhaustive search for
solutions in which a queen appears on the first row of the first column.
Subsequently, the algorithm continues exploring possibilities when start-
ing by placing a queen on the second row of the first column. The same
procedure is repeated until eventually the method is able to place four
queens that do not threaten each other. In that case it can stop search-
ing if the goal consists of finding only one solution, or it can continue to
search for all of the solutions.
The recursion tree also illustrates several features of the algorithm.
Firstly, the nodes correspond to recursive method calls where the par-
tial solution (which will be an input parameter) is valid. The small dark
nodes attached through dotted edges simply show how the tree would
have been expanded had its corresponding partial solution been valid.
Also, observe that an item is added to the partial solution at each re-
cursive call. In other words, partial solutions that contain k items will
be associated with nodes of depth k. Therefore, complete solutions are

found when reaching leaf nodes, of depth n. Furthermore, a path from
the root to a node specifies the partial solution, or a complete solution
if the node is a leaf. For example, the solution found by the algorithm
is the list [1, 3, 0, 2], which is the sequence of labels associated with the
edges of the path from the root to the leaf where the solution is found. In
addition, since the values in the list represent different rows, the solution
to the four-queens puzzle is actually a permutation of the rows, which
are the first n nonnegative integers. A tree that explores all possible per-
mutations would have n! leaves, which can be very large, even for small
values of n. However, the figure also illustrates that the recursive tree
can be pruned considerably by not expanding invalid partial solutions,
which is key to obtaining efficient algorithms.
Similar details will appear in other problems and examples in the
chapter. In particular, Section 12.3 will cover the n-queens puzzle in
depth, showing a specific coded algorithm that solves it. Lastly, the
tree shown in Figure 8.18 could correspond to a recursion tree of an
algorithm that searches for two-element variations with repetition of four
items. The lists next to the nodes would be partial or complete solutions,
while the labels next to the edges would indicate the element that is
introduced in a partial solution. In practice, backtracking algorithms
prune the recursion trees in order to gain efficiency by discarding invalid
solutions.
12.2 GENERATING COMBINATORIAL ENTITIES

We have just seen that the solution to the n-queens problem is a per-
mutation (of the first n nonnegative integers). This will also be the case
in numerous problems. In addition, the goal in many other problems
consists of finding subsets of n elements. Therefore, it is important to
understand how to generate all of the permutations, and all of the sub-
sets, of n distinct elements. This section presents specific algorithms in
Python for generating these combinatorial entities. They are particularly
beneficial since they can be used as starting points to build backtracking
algorithms for many problems.
Lastly, these algorithms will not generate lists or other data struc-
tures in order to store all of the subsets or permutations. This would
not be practical due to the large number of possibilities (there are n!
possible permutations, and 2n different subsets, of n distinct elements).
Instead, the methods will use a simple list (the partial solution) that rep-
[-,-,-]
0 1
[0,-,-] [1,-,-]
0 1 0 1
[0,0,-] [0,1,-] [1,0,-] [1,1,-]
0 1 0 1 0 1 0 1
[0,0,0] [0,0,1] [0,1,0] [0,1,1] [1,0,0] [1,0,1] [1,1,0] [1,1,1]
Figure 12.4 Binary recursion tree of an algorithm that generates all of the
subsets of three items.
resents only one of the subsets/permutations, but which will be modified

in order to contain all of them as the algorithms are executed. Thus, the
methods will be able to process every subset/permutation (we will sim-
ply print them or count them), but will not return all of them in a single
data structure.
12.2.1 Subsets
This section presents two strategies for generating all of the subsets of n
(distinct) elements, which will be provided through an input list. In both
methods the recursion tree will be binary, and the solutions (subsets)
will be represented at its leaves. One method uses partial solutions of
(fixed) length n, while in the other their length varies as the procedure
carries out recursive calls.
12.2.1.1 Partial solutions of fixed length

Figure 12.4 shows the binary recursion tree of an algorithm that gen-
erates the eight subsets that it is possible to create with three distinct
elements. Each subset is described through a binary list of n zeros or
ones (naturally, Boolean values can be used as well). For example, given
the list of elements [a, b, c], the list [1,0,1] denotes the subset {a, c}. This
binary list is the partial solution, which is complete when the method
reaches a leaf. The binary digits are therefore the candidate elements
that form it. As the method carries out recursive calls it simply appends
Listing 12.1 Code for printing all of the subsets of the elements in a list.
1 def generate_subsets(i, sol , elements ):
2 # Base case
3 if i == len(elements ):
4 # Print complete solution
5 print_subset_binary(sol , elements )
6 else:
7 # Generate candidate elements
8 for k in range(0, 2):
9
10 # Include candidate in partial solution
11 sol[i] = k
12
13 # Expand partial solution at position i+1
14 generate_subsets(i + 1, sol , elements )
15
16 # Remove candidate from partial solution
17 sol[i] = None # optional
18
19
20 def generate_subsets_wrapper(elements ):
21 sol = [None] * (len(elements ))
22 generate_subsets(0, sol , elements )
23
24
25 def print_subset_binary(sol , elements ):
26 no_elements = True
27 print('{', end='')
28 for i in range (0, len(sol)):
29 if sol[i] == 1:
30 if no_elements:
31 print (elements [i], sep='', end='')
32 no_elements = False
33 else:
34 print (', ', elements [i], sep='', end='')
35 print('}')
a new candidate to the partial solution (the concrete binary value is

shown on the edges of the tree), which is passed as a parameter to the
method, and shown next to the nodes of the recursion tree. Initially that
list has length n, but its entries are meaningless. When the algorithm
reaches a base case it will have generated one of the 2n possible subsets,
at one of the leaves of the recursion tree. In those cases the lists contain
n meaningful elements and will represent complete solutions.
Listing 12.1 shows an algorithm that complies with the recursion

tree in Figure 12.4, and which prints each subset through the method
print_subset_binary. Besides the initial list elements that represents
the set of n items that it will group into subsets, the procedure also
receives an input list sol that represents the partial solutions, and
which has n components initialized to None in the wrapper method
generate_subsets_wrapper. In addition, observe that the method in-
cludes a new candidate in the partial solution, at every level, as it de-
scends through the recursion tree. Therefore, it uses a third parameter i
that indicates the index in the partial solution where a new zero or one
will appear. Furthermore, i also corresponds to the depth of the node
associated with the recursive call responsible for introducing the new
element in the partial solution. Since the recursive backtracking method
generate_subsets begins introducing a zero or a one at index 0, i is
initialized to 0 in generate_subsets_wrapper.
The method generate_subsets checks in line 3 if the partial solu-
tion in sol is really a complete solution, which occurs if i is equal to n.
This would correspond to the base case, where the method simply prints
the solution (line 5). In the recursive case the procedure has to perform
two recursive calls in order to include more candidates in the partial
solution. It accomplishes this through a simple loop where the variable
k takes either the value 0 or 1 (line 8), and where the i-th component
of the partial solution (sol) receives the value of k (line 11). The back-
tracking algorithms included in this book will also use a loop in order to
consider every possible candidate that could (potentially) be included in
the partial solution. The procedure then invokes itself (line 14) with the
modified partial solution, and incrementing i in order to include more
candidates (if the partial solution is not complete).
The code in line 17 is optional, since the method works correctly
without it. Nevertheless, the procedure includes it for illustrative pur-
poses, since similar instructions may be necessary in other backtracking
algorithms. Observe that when a particular node in the recursion tree
finishes executing, the control passes to its parent, where the value of
the partial solution at the i-th position is meaningless. Therefore, we
can express this explicitly through the code in line 17. However, the
code works without it since this particular method only prints a subset
when reaching a base case (where partial solutions are complete), and
because the assignment in line 11 simply overwrites a zero when k = 1.
The reader is encouraged to execute the program step by step with a
[]
0 1
[0] [1]
0 1 0 1
[0,0] [0,1] [1,0] [1,1]
0 1 0 1 0 1 0 1
[0,0,0] [0,0,1] [0,1,0] [0,1,1] [1,0,0] [1,0,1] [1,1,0] [1,1,1]
Figure 12.5 Alternative binary recursion tree of an algorithm that gener-

ates all of the subsets of three items.
debugger, with or without line 17, in order to examine the evolution of

the partial solution.
Finally, a call to generate_subsets_wrapper(['a','b','c']) pro-
duces the following output:
{}
{c}
{b}
{b, c}
{a}
{a, c}
{a, b}
{a, b, c}
12.2.1.2 Partial solutions of variable length

In the previous algorithm the partial solutions always had the same
length (n). Thus, the procedure can be used with lists or arrays that
do not vary in size. In that case the parameter i was useful in order to
specify the indices of the elements to be added to the partial solution.
Instead, another possibility is to use partial solutions of variable size,
where i is no longer needed. Figure 12.5 shows another recursive tree
that has the same structure as the one in Figure 12.4, but the labels
next to the nodes specify partial solutions at every method call where
every element is meaningful.
Listing 12.2 Alternative code for printing all of the subsets of the elements
in a list.
1 def generate_subsets_alt(sol , a):
2 # Base case
3 if len(sol) == len(a):
4 # Print complete solution
5 print_subset_binary(sol , a)
6 else:
9
11 sol = sol + [k]
12
14 generate_subsets_alt(sol , a)
15
17 del sol[-1]
18
19
20 def generate_subsets_alt_wrapper(elements ):
21 sol = []
22 generate_subsets_alt(sol , elements )
Listing 12.2 shows a method that is in accordance with this approach,

which is very similar to the code in Listing 12.1. Firstly, it prints the
solution if it contains n elements (base case). In the recursive case it
uses the same loop, appends a zero or a one at the end of the partial
solution, and subsequently invokes itself. In this case, the code in line 17
is necessary, where the method has to “undo” the change to the partial
solution prior to invoking the method recursively. While the algorithm in
Listing 12.1 overwrites a zero in the partial solution when incorporating
a one, in this case the method needs to remove the zero from the list
in order to include a one properly. Thus, calling the recursive method
implies adding an element to the partial solution (line 14), and the al-
gorithm deletes it (line 17) after returning from the recursive call. The
code on line 17 of Listing 12.1 plays a similar role, but is not necessary
in that method. Regarding efficiency, appending and removing elements
from a data structure dynamically can be time-consuming. Thus, it is
often more efficient to allocate a fixed amount of memory for the partial
solution, and update it without altering its size. In this regard, the code
[−, −, −]
a c
b
[a, −, −] [b, −, −] [c, −, −]
b c a c a b
[a, b, −] [a, c, −] [b, a, −] [b, c, −] [c, a, −] [c, b, −]
c b c a b a
[a, b, c] [a, c, b] [b, a, c] [b, c, a] [c, a, b] [c, b, a]
Figure 12.6 Recursion tree of an algorithm that generates all of the per-
mutations of three items.
in Listing 12.1 runs faster than the method generate_subsets_alt. Fi-

nally, for this alternative algorithm the wrapper method must initialize
the partial solution to an empty list.
12.2.2 Permutations
This section examines two similar algorithms that print all of the pos-
sible permutations of the n distinct elements in a given list. One way
to represent a permutation is through a list of indices from 0 to n − 1,
which reference the locations of the elements in the input list. For ex-
ample, given the list [a, b, c], the partial solution [1, 2, 0] would denote
the permutation [b, c, a]. However, the following algorithms will simply
use partial solutions formed by the items of the input list (and None
values). Figure 12.6 shows the structure of their recursion trees, for the
list [a, b, c]. Observe that the partial solutions have length n, but only
the first items are meaningful. In the first call to the methods the partial
solution is “empty,” where all of its elements are set to None. The pro-
cedures also receive a parameter i, initially set to zero, which specifies
how many candidate elements have been included in the partial solution.
Thus, it indicates the position in the partial solution where the methods
introduce a new candidate, and which is also equivalent to the depth of
the node related to a method call in the recursion tree.
The methods begin by introducing one of the n candidate elements

of the input list at the first position of the partial solution. Since there
are n possibilities, the root node has n children. In the next level of the
tree the methods invoke themselves n − 1 times, since the element that
appears in the first position cannot appear again in the permutation.
Similarly, in the next level the partial solutions contain two elements and
the methods invoke themselves n − 2 times. This is repeated in general
until they reach a base case, at a leaf of the recursion tree, where the
partial solutions are complete permutations.
The next subsections describe both methods, where the main dif-
ference between them resides in how they test the validity of a partial
solution.
12.2.2.1 Validity of partial solutions without additional data structures

Listing 12.3 shows a first method that complies with the recursion tree
in Figure 12.6. Since i indicates the number of elements in the partial
solution sol, the first if statement checks if the method has reached
a base case. If the result is True the procedure simply prints the per-
mutation stored in sol. If the algorithm has not reached a base case
it uses a loop in order to try to include every candidate item of the
initial input list elements in the partial solution. The loop’s counter k,
which serves as an index of elements, therefore receives values from 0 to
n − 1. Subsequently, the algorithm tests whether the k-th candidate has
already been included in the partial solution (line 10). It is important
to note that although the statement can be written in a single line, it
requires examining i values in the worst case. If the k-th candidate is
not in the partial solution the method includes it (line 13), and performs
a recursive call incrementing the value of i in line 16. Lastly, the method
does not require “undoing” the assignment in line 19 after carrying out
the recursive call, since it simply overwrites the value sol[i]. In other
words, it is not necessary to reset sol[i] to None after the recursive
call. Nevertheless, the code includes the instruction so that its recursion
tree conforms exactly to the one shown in Figure 12.6.
The method generate_permutations_wrapper is a wrapper proce-
dure needed to initialize the partial solution with n None values, and call
generate_permutations with i set to zero. Lastly, the code includes a
basic procedure that prints a list, separating its elements by space bar
characters.
Listing 12.3 Code for printing all of the permutations of the elements in
a list.
1 def generate_permutations(i, sol , elements ):
2 # Base case
4 print_permutation(sol) # complete solution
5 else:
7 for k in range(0, len(elements )):
8
9 # Check candidate validity
10 if not elements [k] in sol[0:i]:
11
13 sol[i] = elements [k]
14
16 generate_permutations(i + 1, sol , elements )
17
19 sol[i] = None # not necessary
20
21
22 def generate_permutations_wrapper(elements ):
24 generate_permutations(0, sol , elements )
25
26
27 def print_permutation(sol):
29 print (sol[i], ' ', end='')
30 print ()
Finally, similarly to the rest of the algorithms included in the chapter,

the method uses a loop in order to consider including every possible
candidate in the partial solution. Without considering constraints in
the internal nodes the resulting algorithm would lead to a full n-ary
recursion tree (every internal node would have n children), which would
contain nn leaves. Naturally, we could test whether the partial solutions
at those nn leaves are valid. However, it is much more efficient to examine
the validity of a candidate in an internal node, and discard unnecessary
recursive calls as soon as the algorithm detects that a partial solution will
never be valid. Thus, in line 6 the condition verifies that the candidate is
a c
b
c a c a
b b
c b c a b a
[a, b, c] [a, c, b] [b, a, c] [b, c, a] [c, a, b] [c, b, a]
Figure 12.7 Pruning a recursion tree as soon as a partial solution is not

valid.
valid, according to the constraints of the problem. This condition speeds

up the algorithm dramatically by pruning the recursion tree, as shown
in Figure 12.7. The lighter gray nodes and edges indicate method calls
of the full ternary tree that are not executed when pruning the tree with
the condition in line 6. Finally, the resulting recursion tree (depicted
through dark nodes and edges) is identical to the one in Figure 12.6,
and contains n! leaves. Although it is a huge quantity even for relatively
small values of n, it is much smaller than nn .
12.2.2.2 Validity of partial solutions with additional data structures

Instead of examining the validity of a partial solution by analyzing its
first i values (they have to be different than elements[k]), the code
in Listing 12.4 uses the additional Boolean parameter list available,
of size n. The idea consists of storing whether a particular item of the
input list elements is not yet in the partial solution. The list is there-
fore initialized to n True values in a wrapper method. In this case, the
condition in line 10 is much simpler, and can be evaluated in Θ(1) time.
Subsequently, if the k-th element is indeed available the procedure first
includes it in the partial solution (line 13), marks it as unavailable (line
16), and invokes itself incrementing the parameter i (lines 19 and 20).
In addition, in this case it is necessary to undo the change to available
in line 23. Observe that when the method returns from the recursive call
it can carry out more iterations of the loop, inserting other elements at
position i in the partial solution, but the list available has to be re-
stored to its initial values. In particular, available[k] has to be set to
Alternative code for printing all of the permutations of the

Listing 12.4
elements in a list.
1 def generate_permutations_alt(i, available , sol , elements ):
2 # Base case
4 print_permutation(sol) # complete solution
5 else:
7 for k in range(0, len(elements )):
8
9 # Check candidate validity
10 if available[k]:
11
13 sol[i] = elements [k]
14
15 # k-th candidate no longer available
16 available[k] = False
17
19 generate_permutations_alt(i + 1, available ,
20 sol , elements )
21
22 # k-th candidate available again
23 available[k] = True
24
25
26 def generate_permutations_alt_wrapper (elements ):
27 available = [True] * (len(elements ))
29 generate_permutations_alt(0, available , sol , elements )
True, since the algorithm has to include elements[k] at some point in

a further position in the partial solution. Finally, although the method
requires an extra input parameter, it is more efficient than the algorithm
in Listing 12.3.
12.3 THE N-QUEENS PROBLEM

The n-queens puzzle was previously introduced in Section 12.1. Recall
that it is a constraint satisfaction problem whose solutions correspond
to permutations of the rows where the queens are placed. Therefore, we
can use the code that generates permutations in Listing 12.4 as a start-
Principal diagonals Secondary diagonals
0,0 0,1 0,2 0,3 0,0 0,1 0,2 0,3
1,0 1,1 1,2 1,3 1,0 1,1 1,2 1,3

6 0
2,0 2,1 2,2 2,3 2,0 2,1 2,2 2,3
5 1
3,0 3,1 3,2 3,3 3,0 3,1 3,2 3,3
4 2
0 1 2 3 3 4 5 6
column − row + n − 1 row + column
Indexing principal and secondary diagonals on a matrix or

Figure 12.8
chessboard.
ing point for building the backtracking method that solves the problem.
Although the final algorithm will introduce a few modifications, its struc-
ture will be very similar to the method that generates permutations.
Consider the method generate_permutations_alt and its input
parameters. It permutes the items in the list elements, which can con-
tain arbitrary numbers, characters, or other types of data. In this case
the elements that we have to permute are the rows of the chessboard.
Since they are simply the integers from 0 to n − 1, we will be able to
write the code omitting the list elements. However, we will use the list
sol that represents a partial solution containing rows, the parameter i
that indicates the column where a new queen will be placed, and the
Boolean list available that specifies which rows (candidates) are free
to be incorporated in the partial solutions. We will change the name of
this list to free_rows, since even though there might not be a queen
in a particular row, the row might not be available for inclusion in the
partial solution if there is a conflict related to a diagonal.
In addition to those parameters, we can use two more Boolean lists
in order to indicate whether there are queens on the diagonals of the
chessboard. There are two types of diagonals: (a) principal diagonals,
which are parallel to the main diagonal that runs from the top-left to the
bottom-right corners of the board; and (b) secondary diagonals, which
are perpendicular to the principal diagonals. There are exactly 2n − 1
of each kind, as illustrated in Figure 12.8 for n = 4. Thus, we can use

the Boolean lists free_pdiags and free_sdiags, both of length 2n − 1,
which will contain True values if there are no queens on the principal
and secondary diagonals, respectively. The figure also shows how we can
enumerate the diagonals, where the pair of numbers in a square indicate
its row and column on the board. On the one hand, note that the sum
of the column and row for squares along a specific secondary diagonal is
constant. These sums range from 0 to 2n − 2, and can be used as indices
to free_sdiags. On the other hand, we can subtract the row from the
column in order to enumerate the principal diagonals. However, this
provides values from −(n − 1) to n − 1. Since we need to use nonnegative
indices, we can simply add n − 1 in order to obtain a valid index for a
principal diagonal that will range from 0 to 2n−2. These operations allow
us to quickly determine the diagonals in which a square lies according
to its row and column.
12.3.1 Finding every solution

Listing 12.5 shows a recursive backtracking algorithm that finds all
of the solutions to the n-queens puzzle (for n = 8 there are 92 solu-
tions, although only 12 are truly distinct in the sense that they can-
not be obtained from others through rotations and reflections). In the
base case the method processes a valid permutation if it is a complete
solution (lines 6 and 7). For example, it could print it (see method
print_chessboard in Listing 12.6), or draw queens on a chessboard
image. The loop in line 12 is used to iterate through all of the possible
candidates that could be appended to the partial solution. The vari-
able k therefore represents rows of the chessboard. Subsequently, the if
statement in lines 16 and 17 examines whether the k-th row is a valid
candidate for the i-th column. In particular, it makes sure that the row
and the corresponding diagonals are free (i.e., that they do not contain
a queen). If the result is True the algorithm can proceed by incorporat-
ing the k-th row into the partial solution (line 20). Since this implies
placing a queen on the chessboard, it is necessary to update the Boolean
data structures in order to indicate that a row and two diagonals will no
longer be free (lines 24–26). The method can then call itself (in lines 30
and 31) with the modified lists in order to continue placing queens on
the next column (i+1). Lastly, after the recursive call the method needs
to get ready to test whether it can place a queen in the next row, which
would occur in the next iteration of the loop. Therefore, it is necessary
Listing 12.5 Code for finding all of the solutions to the n-queens puzzle.
1 def nqueens_all_sol(i, free_rows , free_pdiags ,
2 free_sdiags , sol):
3 n = len(sol)
4
5 # Test if the partial solution is a complete solution
6 if i == n:
7 print_chessboard(sol) # process the solution
8 else:
9
10 # Generate all possible candidates that could
11 # be introduced in the partial solution
12 for k in range(0, n):
13
14 # Check if the partial solution with the
15 # k-th candidate would be valid
16 if (free_rows[k] and free_pdiags[i - k + n - 1]
17 and free_sdiags[i + k]):
18
19 # Introduce candidate k in the partial solution
20 sol[i] = k
21
22 # Update data structures , indicating that
23 # candidate k is in the partial solution
24 free_rows[k] = False
25 free_pdiags[i - k + n - 1] = False
26 free_sdiags[i + k] = False
27
28 # Perform a recursive call in order to include
29 # more candidates in the partial solution
30 nqueens_all_sol(i + 1, free_rows , free_pdiags ,
31 free_sdiags , sol)
32
33 # Eliminate candidate k from the partial
34 # solution , and restore the data structures ,
35 # indicating that candidate k is no longer
36 # in the partial solution
37 free_rows[k] = True
38 free_pdiags[i - k + n - 1] = True
39 free_sdiags[i + k] = True
40
41
42 def nqueens_wrapper(n):
43 free_rows = [True] * n
44 free_pdiags = [True] * (2 * n - 1)
45 free_sdiags = [True] * (2 * n - 1)
46 sol = [None] * n
47 nqueens_all_sol(0, free_rows , free_pdiags , free_sdiags , sol)
to restore the values of the Boolean lists, as if the k-th row had not been
included in the partial solution. Finally, the algorithm modifies the lists
specifying that the k-th row, and the diagonals related to the square at
column i and row k, will be free of queens (lines 37–39).
12.3.2 Finding one solution

In many situations we are only interested in finding one solution to a
problem. Moreover, some problems only have one solution (for exam-
ple, the sudoku puzzle). Therefore, we can code backtracking methods
that stop searching for solutions as soon as they find one. In Python
it is possible to insert a return statement after finding a solution. For
example, we can include it after the print_chessboard(sol) instruc-
tion in the nqueens_all_sol method of Listing 12.5. However, in many
programming languages this is not possible.
There are many ways to implement a program that finds and pro-
cesses one solution. Listing 12.6 shows one approach where the recursive
method is a Boolean function, which returns True if it is able to find
a solution. It is very similar to the code that finds all of the solutions.
On the one hand, the method nqueens_one_sol_wrapper calls the re-
cursive backtracking function with the if statement, and only prints
the solution if the function is able to find it. Note that the list sol
is mutable (it can be understood as a parameter passed by reference)
and will end up storing the solution after the method call if it exists.
Naturally, another option is to print the solution in the base case of
the recursive function before returning. On the other hand, the function
nqueens_one_sol first defines a Boolean variable sol_found, initialized
to False, which indicates if the method has found a solution. The for
loop is replaced by a while loop in order to stop iterating as soon as a
solution is found. The body of the loop is identical to the one in the pro-
cedure nqueens_all_sol, but the result of the recursive call is assigned
to the variable sol_found. Finally, the function can return the value in
sol_found after executing the while loop.
12.4 SUBSET SUM PROBLEM

Given some set S of n positive integers, and a particular integer x, the
goal of this problem is to determine whether the sum of some subset of
elements of S is equal to x. Formally, if si denotes the i-th integer in S,
the idea consists of checking whether there exists a subset T ⊆ S such
Listing 12.6 Code for finding one solution to the n-queens puzzle.
1 def nqueens_one_sol(i, free_rows , free_pdiags ,
3 n = len(sol)
4 sol_found = False
5
6 if i == n:
7 return True
8 else:
9 k = 0
10 while not sol_found and k < n:
11 if (free_rows[k] and free_pdiags[i - k + n - 1]
12 and free_sdiags[i + k]):
13
14 sol[i] = k
15
16 free_rows[k] = False
17 free_pdiags[i - k + n - 1] = False
18 free_sdiags[i + k] = False
19
20 sol_found = nqueens_one_sol(i + 1, free_rows ,
21 free_pdiags ,
22 free_sdiags , sol)
23
24 free_rows[k] = True
25 free_pdiags[i - k + n - 1] = True
26 free_sdiags[i + k] = True
27
28 k = k + 1
29
30 return sol_found
31
32
33 def nqueens_one_sol_wrapper(n):
34 free_rows = [True] * n
35 free_pdiags = [True] * (2 * n - 1)
36 free_sdiags = [True] * (2 * n - 1)
37 sol = [None] * n
38
39 if nqueens_one_sol(0, free_rows , free_pdiags ,
41 print_chessboard(sol)
42
43
44 def print_chessboard(sol):
46 print (sol[i], ' ', end='')
47 print ()
that:
∑ si = x. (12.1)
si ∈T
In particular, we will design a backtracking algorithm that prints every

subset T of S that satisfies (12.1). For example, if S = {1, 2, 3, 5, 6, 7, 9}
and x = 13, the method should print the following five subsets on the
console: {6, 7}, {2, 5, 6}, {1, 5, 7}, {1, 3, 9}, and {1, 2, 3, 7}.
A naive brute-force solution to the problem consists of generating
every possible subset T of S, and testing whether (12.1) is True. This
exhaustive search can be carried out through a procedure similar to
generate_subsets in Listing 12.1. The set S could be represented by
the input list elements, and the method would contain an additional
parameter that stores the value of x. Furthermore, the procedure would
have to verify (12.1) in the base case before printing a subset. In other
words, it would have to check (12.1) in each leaf of the recursion tree.
However, it is possible to speed up the search by using a backtracking
approach. Listing 12.7 shows a particular implementation where a partial
solution (sol) represents the subset T . In order to avoid calculating the
sum of the elements in the partial solution in every recursive call, it uses
an additional accumulator parameter psum that contains the sum, and
which is obviously initialized to zero in the wrapper method. The key
observation that leads to the algorithm is that we can prune the recursion
tree as soon as the sum of elements related to the partial solution is equal
to x (i.e., when (12.1) is satisfied); or when it is greater than x, where it
will not be possible to satisfy (12.1) by expanding the partial solution,
since the elements of S are positive.
Firstly, the method includes a base case (in lines 3 and 4) in order to
print the subset associated with the partial solution if it satisfies (12.1).
This allows us to prune the tree at an internal node, of depth i, of the
recursion tree. In the base case T can only contain the first i elements of
S. In other words, only the first i components of the partial solution are
meaningful. If the n−i last entries can take arbitrary values, it would
be necessary to pass the value of i to the printing method, and specify
the “true” size of the partial solution. Instead, the algorithm calls the
method print_subset_binary in Listing 12.1, since these last values
will always be zero, indicating that the elements are not in the partial
solution (T ). This is achieved by initializing the partial solution with n
zeros (in print_subset_sum_wrapper), denoting the empty set; and by
updating the partial solution appropriately so that its n−i last items
Listing 12.7 Backtracking code for solving the subset sum problem.
1 def print_subset_sum(i, sol , psum, elements , x):
2 # Base case
3 if psum == x:
4 print_subset_binary(sol , elements )
5 elif i < len(elements ):
6 # Generate candidates
8
9 # Check if recursion tree can be pruned
10 if psum + k * elements [i] <= x:
11
12 # Expand partial solution
13 sol[i] = k
14
15 # Update sum related to partial solution
16 psum = psum + k * elements [i]
17
18 # Try to expand partial solution
19 print_subset_sum(i + 1, sol , psum, elements , x)
20
21 # not necessary:
22 #psum = psum - k*elements [i]
23
24 # Make sure a 0 indicates the absence of an element
25 sol[i] = 0
26
27
28 def print_subset_sum_wrapper(elements , x):
29 sol = [0] * (len(elements ))
30 print_subset_sum(0, sol , 0, elements , x)
are always zeros when reaching a base case (this is implemented through
the assignment in line 25).
If the partial solution does not satisfy (12.1) the method can simply
terminate if the partial solution has n elements, without carrying out any
operation. Thus, with the condition in line 5 the algorithm only continues
expanding partial solutions if i< n. In that case it uses a loop in order
to process the two possibilities of ignoring (k=0) or including (k=1) the
i-th element of S in the partial solution. Since k is a number we can
use it within the expression psum + k*elements[i] to indicate the new
sum associated with the partial solution. Notice that when k=0 it does
not change, while the method adds the i-th element of S when k=1.
+2 ?
0 2
+6 ?
0 6 2
8
+3 ?
0 3 6 2 5 {2, 6}
9
+5 ?
0 5 3 8 6 11 2 7 5 10
{3, 5}
Recursion tree of the procedure that solves the subset sum

Figure 12.9
problem for S = {2, 6, 3, 5} and x = 8.
Therefore, the condition in line 10 makes sure that the new sum is less
than or equal to x before including the element, and proceeding to make
additional recursive calls. If the condition is True the method expands
the partial solution with the corresponding candidate (line 13), updates
psum (line 16), and carries out the recursive call (line 19). Afterwards,
it is not necessary to restore the value of psum (in line 22) since in the
first iteration of the loop k=0, which implies that the value of psum is
not modified. Lastly, the algorithm sets the value of the partial solution
to zero at position i when terminating, which is necessary in order to
print the partial solutions correctly in the base case (its n−i last items
must always be zero).
Finally, Figure 12.9 shows the recursion tree of the method
print_subset_sum, for S = {2, 6, 3, 5} and x = 8. The partial solutions
are constructed at each level as in the methods in Section 12.2.1. When
descending through a left branch the algorithm does not include the i-th
element of S in T , but it incorporates it when advancing down a right
branch. In this case the numbers next to the nodes indicate the sum of
the items included in the partial solutions (i.e., psum) when invoking the
method. Observe that the tree is pruned as soon as it finds a solution that
satisfies (12.1), or if it detects that the sum of the elements in T is greater
than x. Lastly, a call to print_subset_sum_wrapper([2,6,3,5],8)
produces the correct result:
{3,5}
{2,6}
Finally, observe that without the assignment in line 25, the partial so-
lution would hold a 1 in its third position when reaching the base case
associated with the subset {2, 6}, and the method would print {2,6,3}.
12.5 PATH THROUGH A MAZE

This section describes a backtracking algorithm for finding a path
through a maze, such as the one illustrated in Figure 12.10(a). A maze
is defined as a rectangular array of cells that can either be empty or
constitute a wall. We will code it by using a list of lists (of characters),
denoted as M, where the first list will contain the cells of the top row
of the maze (from left to right), the second list will represent the row
below the first one, and so on. In order to define a maze a user can write
each row as a line of characters, separated by a space bar character, as
shown in (b). Specifically, an ‘E’ indicates an empty cell, which can be
traversed by a path, and a ‘W’ indicates a wall. Since it is time con-
suming to enter large mazes manually, it is convenient to store the array
of characters in a file that can be loaded automatically when running
the program (this is also beneficial for debugging). The method will also
receive parameters that indicate the cells where a path enters and exits
the maze. In the figure these cells are the top-left and bottom-right ones,
respectively. Thus, the initial cell in the example is M[0][0].
The backtracking algorithm carries out an exhaustive search for
paths that start at the initial cell and arrive at the final cell. Having
reached some particular cell in an intermediate step, the algorithm needs
to analyze every possible path when advancing to a neighboring cell go-
ing upwards, downwards, rightwards, or leftwards. Since the final cell
is in the bottom-right corner of the maze the algorithm can begin by
searching paths that start going downwards, then towards the right,
subsequently upwards, and finally towards the left. For this particular
order of search the method takes the steps shown in Figure 12.10(c).
Observe that when arriving at an empty cell the method first tries to
expand the path downwards. If it cannot advance it will continue trying
to move towards the right, then upwards, and finally towards the left.
If it cannot proceed in any of the four directions it must backtrack to a
previous cell, and continue to search for other alternative paths. In the
example the initial cell is at position (0, 0), corresponding to row 0 and
column 0. The method will mark that cell as being part of the current
path, and update M[0][0] to ‘P’. The first move advances to position
(1, 0), and also marks the corresponding cell with a ‘P’. The method
Enter
E W E E E E E E W E W W¶
E E E W W W E W W E E E¶
W W E E W E E E W W W E¶
W E E W W E W E E E E E¶
E E W W E E E E W E W W¶
E W W E E W W W W E E E¶
E E E E W W E E E E W E¶
E W E W W W W W W W W E¶
E W E E E E W E E E W W¶
E W W E W E E E W E E E¶
W W E E W W E W W E W E¶
E E E W W E E E W E W E¶
Exit
(a) (b)
Free cell Wall Search progress
Explored cell Cell in the final path
Enter Search order: ↓, →, ↑, ← Enter Search order: ↓, →, ↑, ←
Exit Exit
(c) (d)
Problem of finding a path through a maze, and solution

Figure 12.10
through backtracking when searching in a particular order.
Enter Search order: →, ↑, ←, ↓ Enter Search order: ←, ↑, →, ↓
Exit Exit
Figure 12.11 Different paths through a maze, depending on the search

order.
then tries to advance downwards, but hits a wall located at (2, 0). Since
a path with that cell would not be valid, the algorithm continues by
trying to advance towards the right, which it can do since the cell at
(1, 1) is empty. Eventually the path reaches the cell at (6,0), where the
method will also begin by trying to advance downwards. The path ends
up advancing to the cell at (9, 0), where it hits a dead end. In that case,
not only are there walls below and to the right, but going upwards is
not allowed because the cell on top is already part of the path. Further-
more, going towards the left would mean exiting the limits of the maze.
Since the algorithm cannot advance in any direction, it backtracks to the
cell at (8, 0), where it has not yet tried to advanced towards the right,
upwards, or leftwards. However, these possibilities are also not allowed.
This also occurs when returning to the cell at (7, 0). After backtracking
to the cell at (6, 0), it explores new paths that go towards the right, after
having exhaustively explored all of the possible paths going downwards.
This process is repeated until the algorithm finds the exit cell. The so-
lution is shown in (d), where the explored cells appear with a shaded
background.
Backtracking algorithms can stop searching for solutions when they
find the first one, or they can continue computing every solution. We
will see an algorithm that halts as soon as it finds a path through the
maze. In this regard, the order in which it examines the possible paths
(c, r)
Problem
↓ ←
→ ↑
(c, r + 1) (r, c + 1) (c, r − 1) (c − 1, r)
Subproblem Subproblem Subproblem Subproblem
Figure 12.12 Decomposition of the problem of finding a path through a

maze.
can lead to different solutions, as illustrated in Figure 12.11. Observe

that the explored cells are different in each case.
The solution to this problem is clearly not a permutation of cells.
This can be seen, for instance, since the length of a path (i.e., a solution)
is not fixed. Furthermore, although a path consists of a subset of cells,
these appear ordered in a solution. Therefore, it is not appropriate to
model a solution as a subset. Instead, we will use a matrix of characters
to code a solution, where the cells in a particular path will be marked
with a ‘P’ character. Thus, a partial solution will be the input maze
that also contains a path from the initial cell to some other free cell.
Having arrived at this last cell, the method has to consider including four
candidate cells in the maze. Therefore, the problem can be decomposed
recursively as shown in Figure 12.12, where r and c denote a particular
row and column, respectively. Every node of the recursion tree will have
four possible children, which correspond to the four directions in which
it is possible to advance. However, the algorithm will have to check that
the new cell is free before expanding the path through a recursive call.
Listing 12.8 shows a possible implementation of a backtracking al-
gorithm that solves the problem. Firstly, the parameters of the wrapper
method are the initial maze, and four integers that indicate the coordi-
nates of the entering (enter_row and enter_col) and exiting (exit_row
and exit_col) cells. It also declares the list incr, which specifies the
search directions. Each tuple (x, y) in the list indicates how to increment
the column c and row r of a cell in order to advance to a neighboring
one. In other words, the new cell would be (c + x, r + y). The order of
Listing 12.8 Backtracking code for finding a path through a maze.

1 def find_path_maze(M, row , col , incr, exit_row , exit_col ):
2 # Base case: check if path found
3 if row == exit_row and col == exit_col :
4 return True # Solution found
5 else:
6 sol_found = False
8 k = 0
9 while not sol_found and k < 4:
10
11 # New candidate cell
12 new_col = col + incr[k][0]
13 new_row = row + incr[k][1]
14
15 # Test candidate validity
16 if (new_row >= 0 and new_row < len(M)
17 and new_col >= 0 and new_col < len(M[0])
18 and M[new_row ][new_col ] == 'E'):
19
20 # Add to path (partial solution )
21 M[new_row ][new_col ] = 'P'
22
23 # Try to expand path starting at new cell
24 sol_found = find_path_maze(
25 M, new_row , new_col , incr,
26 exit_row , exit_col )
27
28 # Mark as empty if new cell not in solution
29 if not sol_found:
30 M[new_row ][new_col ] = 'E'
31
32 k = k + 1
33
34 return sol_found
35
36
37 def find_path_maze_wrapper(M, enter_row , enter_col ,
38 exit_row , exit_col ):
39 # search directions
40 incr = [(0, 1), (1, 0), (0, -1), (-1, 0)]
41
42 M[enter_row][enter_col] = 'P'
43 return find_path_maze(M, enter_row , enter_col , incr,
44 exit_row , exit_col )
search for this algorithm is therefore: down, right, up, and left. Observe
that adding a unit to a row implies going down (since the first row is
located at the top of the maze), while adding one to a column signifies
moving towards the right. Lastly, the wrapper method indicates that the
initial cell will form part of the solution path, and returns the (Boolean)
result of calling the recursive backtracking method. The parameter M of
find_path_maze is both the initial maze and the partial solution. The
first two parameters of the recursive function indicate the coordinates of
the last cell of the partial path M, and the algorithm will try to expand
it advancing to one of its neighboring cells. Finally, the method receives
incr, and the coordinates of the exiting cell.
The recursive function find_path_maze returns True if it finds a
path through the maze, where the solution would be stored precisely
in M. It declares the variable sol_found, initialized as False, which is
set to True at the base case if the algorithm finds a complete solution.
Otherwise, it uses a while loop to generate the four candidate cells,
but one that terminates as soon as a solution is found. The variables
new_col and new_row constitute the new candidate (lines 12 and 13).
Subsequently, the algorithm examines whether it is valid. In particular,
it has to be within the limits of the maze (lines 16 and 17), and it has
to be empty (line 18). If the cell does not violate the constraints of the
problem the method incorporates it to the path (line 21), and calls the
recursive method in lines 24–26, where the output is stored in sol_found.
If a solution is not found the cell will not belong to the path. Therefore,
it sets its value back to ‘E’, indicating that it is empty (line 30). This
is necessary since the path in the final maze is determined through the
cells marked as ‘P’. Without this condition all of the explored cells would
contain a ‘P’ character. Lastly, the method can simply return the value
of sol_found after exiting the loop (line 34).
Finally, Listing 12.9 shows additional code that can be used in
order to execute the program and draw the maze. The method
read_maze_from_file reads a text file defining a maze and returns the
corresponding list of lists. The basic iterative procedure draw_maze uses
the Matplotlib package to depict a maze. Finally, the last lines of the
code read a maze from a file, and draw it only if there exists a path
through it, where the initial and final cells are the top-left and bottom-
right ones, respectively.
Auxiliary code related to the backtracking methods for find-

Listing 12.9
ing a path through a maze.
1 import matplotlib .pyplot as plt
2 from matplotlib .patches import Rectangle
3
4 def read_maze_from_file (filename ):
5 file = open(filename , 'r')
6 M = []
7 for line in file.readlines ():
8 M.append ([x[0] for x in line.split(' ')])
9 file.close ()
10 return M
11
12 gray = (0.75, 0.75, 0.75)
13 black = (0, 0, 0)
14 red = (0.75, 0, 0)
15 green = (0, 0.75, 0)
16
17 def draw_maze (M, enter_row , enter_col , exit_row , exit_col ):
18 nrows = len(M)
19 ncols = len(M[0])
20 fig = plt.figure ()
21 fig.patch.set_facecolor ('white')
22 ax = plt.gca()
23
24 if enter_row is not None and enter_col is not None:
25 ax.add_patch (Rectangle ((enter_col , nrows - enter_row ),
26 1, -1, linewidth =0, facecolor =green,
27 fill=True))
28 if exit_row is not None and exit_col is not None:
29 ax.add_patch (Rectangle ((exit_col , nrows - exit_row ),
30 1, -1, linewidth =0, facecolor =red ,
31 fill=True))
32
33 for row in range(0, nrows):
34 for col in range(0, ncols):
35 if M[row][col] == 'W':
36 ax.add_patch (Rectangle ((col , nrows - row), 1, -1,
37 linewidth =0, facecolor =gray))
38 elif M[row][col] == 'P':
39 circ = plt.Circle ((col + 0.5, nrows - row - 0.5),
40 radius =0.15, color=black, fill=True)
41 ax.add_patch (circ)
42
43 ax.add_patch (Rectangle ((0, 0), ncols, nrows, edgecolor =black ,
44 fill=False))
45 plt.axis('equal')
46 plt.axis('off')
47 plt.show()
48
49 M = read_maze_from_file ('maze_01 .txt') # some file
50 # Enter at top-left , exit at bottom-right
51 if find_path_maze_wrapper (M, 0, 0, len(M) - 1, len(M[0]) - 1):
52 draw_maze (M, 0, 0, len(M) - 1, len(M[0]) - 1)
6 1 4 5 9 6 3 1 7 4 2 5 8
8 3 5 6 1 7 8 3 2 5 6 4 9
2 1 2 5 4 6 8 9 7 3 1
8 4 7 6 8 2 1 4 3 7 5 9 6
6 3 4 9 6 8 5 2 3 1 7
7 9 1 4 7 3 5 9 6 1 8 2 4
5 2 5 8 9 7 1 3 4 6 2
7 2 6 9 3 1 7 2 4 6 9 8 5
4 5 8 7 6 4 2 5 9 8 1 7 3
Sudoku puzzle Solution
Figure 12.13 An instance of the sudoku puzzle and its solution.
12.6 THE SUDOKU PUZZLE

The sudoku puzzle is a constraint satisfaction problem where the goal is
to fill a 9 × 9 grid with the digits from 1 to 9 (they can be any nine dif-
ferent symbols, since their numerical values are irrelevant). A particular
digit can appear only once in each row, each column, and each of nine
3 × 3 non-overlapping subgrids, also called boxes, which cover the 9 × 9
grid. Figure 12.13 shows an instance of the problem, which consists of
a partially filled grid with fixed initial digits. The problem is well-posed
when the puzzle has a unique solution (the initial grid must contain at
least 17 digits).
We will use a list of lists of digits to implement the grid, where a
zero represents an empty cell. The partial solutions will be partially filled
grids, and the algorithm must enumerate the cells somehow in order to
expand the partial solutions with new digits. The algorithm described
below considers expanding them in row-major order, where it will begin
inserting valid digits (i.e., candidates) in the top row from left to right,
then in the second row, and so on. Thus, the algorithm will have found
a solution if it is able to place a valid digit in the bottom-right cell.
In a recursive case the method generates nine possible candidates to
include in an empty cell. Thus, a node of the recursion tree could have
up to nine potential children, as shown in Figure 12.14(a). However, the
problem requires handling a scenario that has not appeared in previous
problems. When expanding the partial solution some of the digits in the
grid will be the initial fixed ones. When the algorithm processes a cell
Fixed initial digit

1 2 8 9
⋯
(a) (b)
Figure 12.14 Recursive cases for the sudoku solver.
with one of these digits it must skip it (it cannot expand the partial
solution for that cell), and carry out a recursive call that processes the
following cell. This requires incorporating a second recursive case that
is illustrated in (b).
Listings 12.10 and 12.11 show a backtracking procedure that solves
the problem, together with several auxiliary methods. The recursive
method solve_sudoku_all_solutions assumes that a sudoku might
not be well-posed, and could potentially have several solutions (or even
0). Thus, it prints all of the valid solutions to a given sudoku grid. For
example, if the top row of the sudoku in Figure 12.13 is replaced by an
empty row, there are 10 different ways to fill the sudoku grid with digits.
The inputs to the recursive procedure are the coordinates (row and
col) of a cell where it will try to include digits in order to expand
the partial solution represented by the list of lists of digits S. Since the
procedure expands the partial solution in row-major order, starting at
cell (0, 0), it will have obtained a valid solution when row is equal to 9. In
that base case it can simply print the sudoku grid (line 4). Otherwise, the
method checks whether the current cell is not empty (i.e., if it contains
one of the initial fixed digits). If the result is True it skips the cell, and
continues by invoking the recursive method (line 14) with the coordinates
of the next cell in the row-major order (computed in line 11).
In the second recursive case it uses a loop to generate the nine pos-
sible candidates to include in the empty cell. Afterwards, in line 20 it
analyzes whether it is feasible to incorporate candidate k in the partial
solution S, at cell (row,col). If it is, the method includes the candidate in
the partial solution (line 23), and continues carrying out a recursive call
with the next cell (line 29). When the loop finishes, the method needs
to undo the changes made to the cell, leaving it empty (line 32). This is
necessary for future explorations of solutions. Note that the cell has to
Listing 12.10 Code for solving a sudoku puzzle.

1 def solve_sudoku_all_sols(row , col , S):
2 # Check if sudoku is complete
3 if row == 9:
4 print_sudoku(S) # print the completed sudoku
5 print ()
6 else:
7 # Check if digit S[row][col] is an initial fixed symbol
8 if S[row][col] != 0:
9
10 # Advance to a new cell in row -major order
11 (new_row , new_col ) = advance (row , col)
12
13 # Try to expand the partial solution
14 solve_sudoku_all_sols(new_row , new_col , S)
15 else:
16 # Generate candidate digits
17 for k in range (1, 10):
18
19 # Check if digit k is a valid candidate
20 if is_valid_candidate(row , col , k, S):
21
22 # Include digit in cell (row ,col)
23 S[row][col] = k
24
25 # Advance to a new cell in row -major order
26 (new_row , new_col ) = advance (row , col)
27
28 # Try to expand the partial solution
29 solve_sudoku_all_sols(new_row , new_col , S)
30
31 # Empty cell (i,j)
32 S[row][col] = 0
33
34
35 # Compute the next cell in row -major order
36 def advance (row , col):
37 if col == 8:
38 return (row + 1, 0)
39 else:
40 return (row , col + 1)
be free the next time the algorithm processes it after backtracking to an

earlier cell.
Listing 12.11 Auxiliary code for solving a sudoku puzzle.

1 import math
2
3 # Check if the digit at cell (row ,col) is valid
4 def is_valid_candidate(row , col , digit , S):
5 # Check conflict in column
7 if k != col and digit == S[row][k]:
8 return False
9
10 # Check conflict in row
12 if k != row and digit == S[k][col]:
13 return False
14
15 # Check conflict in box
16 box_row = math.floor (row / 3)
17 box_col = math.floor (col / 3)
19 for m in range(0, 3):
20 if (row != 3 * box_row + k
21 and col != 3 * box_col + m):
22 if digit == S[3 * box_row + k][3 * box_col + m]:
23 return False
24
25 return True
26
27 # Read a sudoku grid from a text file
28 def read_sudoku(filename ):
29 file = open(filename , 'r')
30 S = [[None] * 9] * 9
31 i = 0
32 for line in file.readlines():
33 S[i] = [int(x) for x in line.split(' ')]
34 i = i + 1
35 file.close ()
36 return S
37
38 # Print a sudoku grid on the console
39 def print_sudoku(S):
40 for s in S:
41 print (*s)
42
43 S = read_sudoku('sudoku01_input.txt') # Some file
44 solve_sudoku_all_sols(0, 0, S)
The code also includes: (a) a function that returns a tuple with the
coordinates of the next cell in row-major order; (b) a function that de-
termines whether the digit placed at some row and column violates the
constraints of the problem, where the variables box_row and box_col
are the top-left cells of the 3 × 3 boxes; (c) a function for reading a su-
doku grid from a text file, where each row contains the nine initial digits
of a sudoku row, separated by space bar characters; (d) a method for
printing the grid; and (e) code for reading and solving a sudoku.
12.7 0-1 KNAPSACK PROBLEM

Backtracking algorithms are also used to solve optimization problems. In
the 0-1 knapsack problem there are n objects with values vi , and weights
wi , for i = 0, . . . , n−1. The goal of the problem is to introduce a subset of
these objects in a knapsack with total weight capacity C, such that the
sum of their values is maximized. Formally, the optimization problem
can be written as:
maximize n−1
∑ xi vi
x i=0
n−1
subject to ∑ xi wi ≤ C,
i=0
xi ∈ {0, 1}, i = 0, . . . , n − 1.
The vector or list x = [x1 , . . . , xn ] is the variable of the problem, whose

components can be either 0 or 1. In particular, the i-th object is intro-
duced in the knapsack when xi = 1. Therefore, x plays the same role as
the binary list in the subset generation or subset sum problem. The sum
n−1
∑i=0 xi vi is called the objective function, and simply adds the values
n−1
of the objects that are in the knapsack. The constraint ∑i=0 x i wi ≤ C
indicates that the sum of their weights cannot exceed the capacity C.
The following subsections describe two approaches that perform an
exhaustive search for the optimal solution. The first is a standard
backtracking algorithm that prunes the recursion tree when the partial
solution violates the constraints of the problem (i.e., when the sum of the
weights exceeds the capacity of the knapsack). The second approach uses
an algorithm design paradigm known as branch and bound, which is
closely related to backtracking. The method enhances the search by also
pruning the recursion tree as soon as the algorithm detects that it will
not be able to improve the best solution found in previous steps (by
expanding a particular partial solution).
12.7.1 Standard backtracking algorithm

Since the solution is a subset of the n objects we can develop an al-
gorithm whose recursion tree has the same binary structure as the one
related to the algorithms for the subset generation or subset sum prob-
lem. Figure 12.15 illustrates the recursion tree for weights w = [3, 6, 9, 5],
values v = [7, 2, 10, 4], and knapsack capacity C = 15, which will be in-
put parameters to the recursive method. The nodes represent method
calls with specific partial solutions as parameters, where the numbers
depicted inside the nodes specify the remaining capacity in the knap-
sack, and the sum of the values of the objects included in it, as shown in
(a). For a given index i in the partial solution, the algorithm can discard
introducing the associated object in the knapsack (left child), or it can
include it (right child). If it discards the object both of the numbers in
the child node do not vary. However, if an element is introduced in the
knapsack its remaining capacity decreases by the weight of the object
wi , while the total value of the objects in the knapsack increases by vi .
The algorithm solves an optimization problem. Thus, it needs to keep
track of the best solution found in previous steps, and must update it
if it encounters a better one. The full recursion tree is shown in (b),
where the shaded nodes indicate the method calls that update this best
solution. Assuming that the initial optimal value is a negative number,
the method would first update the optimal solution when reaching the
first leaf. In that case the partial solution encodes the empty set, where
the sum of values is obviously zero. Continuing the tree traversal, the
method updates the best solution after introducing the last object in
the knapsack (w3 = 5, and v3 = 4). Thus, the remaining capacity at
the leaf node is 10, and the total value of the partial solution is 4. The
process continues, and updates the best solution three more times (the
corresponding best values are 10, 14, and 17). The last update naturally
determines the optimal solution to the problem. In this case it consists
of the subset formed by the first and third objects, where the optimal
sum of values is 17. Lastly, it is also important to note how the method
prunes the binary recursion tree. In particular, the remaining capacity
associated with a valid partial solution (and node) cannot be negative.
Listing 12.12 shows a possible implementation of the backtracking
algorithm. Firstly, similarly to other methods, its input parameters in-
x = remaining capacity
y = partial value
0: do not include object i 1: include object i
x x − wi
y y + vi
(a)
w = [3, 6, 9, 5]
15
v = [7, 2, 10, 4]
0
C = 15 1
15 12
0 7
0
15 9 12 6
0 2 7 9
15 6 9 0 12 3 6 -3
0 10 2 12 7 17 9 19
15 10 6 1 9 4 0 -5 12 7 3 -2 6 1
0 4 10 14 2 6 12 16 7 11 17 21 9 13
Optimal solution: [1, 0, 1, 0]

(b)
Recursion tree of a backtracking algorithm for the 0-1 knap-

Figure 12.15
sack problem.
Listing 12.12 Backtracking code for solving the 0-1 knapsack problem.
1 def knapsack_0_1(i, w_left , current_v , sol ,
2 opt_sol , opt_v , w, v, C):
3 # Check base case
4 if i == len(sol):
5 # Check if better solution has been found
6 if current_v > opt_v:
7 # Update optimal value and solution
8 opt_v = current_v
9 for k in range (0, len(sol)):
10 opt_sol [k] = sol[k]
11 else:
14
16 if k * w[i] <= w_left :
17
19 sol[i] = k
20
21 # Update remaining capacity and partial value
22 new_w_left = w_left - k * w[i]
23 new_current_v = current_v + k * v[i]
24
26 opt_v = knapsack_0_1(i + 1, new_w_left ,
27 new_current_v , sol ,
28 opt_sol , opt_v , w, v, C)
29
30 # return value of optimal solution found so far
31 return opt_v
32
33
34 def knapsack_0_1_wrapper(w, v, C):
35 sol = [0] * (len(w))
36 opt_sol = [0] * (len(w))
37 total_v = knapsack_0_1(0, C, 0, sol , opt_sol , -1, w, v, C)
38 print_knapsack_solution(opt_sol , w, v, C, total_v )
clude the partial solution sol and the index i related to the object that
may be introduced in the knapsack. The method also receives the lists
of weights w, values v, and the capacity C. With these parameters it is
possible to compute the remaining capacity and the accumulated sum of
values of the objects in the knapsack, in i+1 steps. However, it is more
efficient to include two extra parameters that contain these quantities.

In particular, w_left stores the remaining capacity, while current_v
holds the sum of the values of the objects. The advantage is that they
can be updated in a single step.
Furthermore, the method needs to store the best solution found as it
is executed. The parameter opt_sol stores this information, which will
be a copy of the partial solution when reaching a leaf of the recursion
tree. Lastly, opt_v contains the optimum sum of values associated with
opt_sol. This parameter is also redundant in the sense that it is possible
to compute its value given opt_sol. However, it is more efficient to pass
this information as a parameter.
As usual, the recursive function begins by analyzing the base case. If
the method has completed a partial solution (line 4) it checks whether
the current value of the partial solution is greater than the best value
found at any previous step (line 6). If the result is True it updates opt_v,
and overwrites the values of opt_sol with the ones in sol. The method
will later return opt_v in line 31. The values of the list opt_sol will be
available after executing the method since it is a mutable object.
In the recursive case the method generates the candidates (zero or
one) to include in the binary partial solution list, in line 13. It then uses
the condition k*w[i]<=w_left to check if it can prune the recursion tree.
On the one hand, if k = 0 the object is not introduced in the knapsack
and the condition is always True since w_left is nonnegative. However,
if k = 1 the algorithm can only include the object in the knapsack if
it fits. In other words, wi must be less than or equal to the remaining
capacity of the knapsack. The method would continue by updating the
partial solution, the remaining knapsack capacity, and the sum of the
values of the objects in the knapsack. The function carries out a recursive
call with these new parameters, storing the result in opt_v, which ends
up returning in line 31.
Regarding the wrapper function, i is initially 0, w_left stores the
total capacity C, while current_v is 0 since the knapsack is empty.
The optimal value opt_v can be either 0 or some negative number. In
Figure 12.15 we have assumed that it is negative. This simply implies
that the method will update the optimal solution even when it consists
of an empty set (at the leftmost leaf).
Finally, Listing 12.13 shows a simple iterative code that can be used
to print a solution to the problem. Its optimal value can be computed
from the partial solution and the list of values. However, the method
includes such value as a parameter since the function knapsack_0_1
Listing 12.13 Auxiliary code related to the 0-1 knapsack problem.

1 def print_knapsack_solution(sol , w, v, C, opt_value):
2 n = len(sol)
3 k = 0
4 while k < n and sol[k] == 0:
5 k = k + 1
6
7 total_weight = 0
8 if k < n:
9 print ('(', w[k], ',', v[k], ')', sep='', end='')
10 total_weight = total_weight + w[k]
11
12 for i in range(k + 1, n):
13 if sol[i] == 1:
14 total_weight = total_weight + w[i]
15 print (' + ', sep='', end='')
16 print ('(', w[i], ',', v[i], ')', sep='', end='')
17
18 print(' => ', '(', total_weight ,
19 ',', opt_value , ')', sep='')
20
21
22 w = [3, 6, 9, 5] # List of object weights
23 v = [7, 2, 10, 4] # List of object values
24 C = 15 # Weight capacity of the knapsack
25 knapsack_0_1_wrapper(w, v, C)
returns it, and is therefore readily available in knapsack_0_1_wrapper.

The last lines simply define an instance of the problem, and compute its
solution.
12.7.2 Branch and bound algorithm

Branch and bound can be viewed as a variant of backtracking that per-
forms a more efficient exhaustive search of solutions to discrete and com-
binatorial optimization problems. The idea consists of using bounds in
order to prune the recursion tree, not only when a partial solution does
not satisfy the constraints of the problem, but also when it is certain
that expanding the partial solution will not lead to a better solution
than the best one found in previous steps.
The key to the branch and bound algorithm that we will examine for
the 0-1 knapsack problem relies on using an extra parameter to store the
maximum sum of values that could be obtained by expanding a partial
solution. If at some method call this value is smaller than the best value
found by the algorithm in previous steps, it will not continue to expand
the partial solution, since it will not be able to obtain a better solution.
This allows us to prune the recursion tree at more nodes, which can lead
to a considerably more efficient search.
Figure 12.16 illustrates the recursion tree for the same weights, val-
ues, and knapsack capacity as in Figure 12.15. In this case, the numbers
inside the nodes indicate the partial value, and a bound on the maximum
possible value that we can obtain by expanding the associated partial
solution, as shown in (a). Initially, the partial value is 0, and the bound
is the sum of all of the values of the objects (i.e., 23 = 7 + 2 + 10 + 4), as
illustrated in (b), since at first we have to contemplate the case where
every object fits in the knapsack.
Each internal node of depth i can have two children. Descending
through the left branch implies not introducing object i in the knapsack.
Therefore, the partial value does not change, but the bound is decreased
by vi , since the object will not contribute its value to the total sum of
a complete solution. Instead, the object is introduced in the knapsack
when descending through the right branch. This implies adding vi to the
partial value, but the bound remains unaltered.
Similarly to the example in Figure 12.15, the method updates the
best solution found so far in the shaded leaves. In addition, the nodes
drawn with a dotted contour indicate method calls that are not carried
out because the sum of the weights of the objects exceeds the knapsack’s
capacity. Furthermore, the figure shows a new type of node with a lighter
dashed contour. The algorithm also discards the method calls associated
with these nodes since the value of the bound is smaller than the best
sum of values encountered previously. For instance, the best value after
reaching the fourth leaf is 14. Afterwards, consider reaching the node
with partial value 2 and bound 16. Discarding the third object implies
reducing the bound by 10. This means that the maximum sum of values
that it is possible to obtain by expanding the partial solution is 16 − 10 =
6. Since 6 < 14 (depicted underneath the node), the algorithm avoids
calling the method, pruning the recursion tree. Observe that it has the
same structure as the recursion tree in Figure 12.15, but it contains
less nodes since it prunes the tree on more occasions. In practice, this
enhancement can have a dramatic effect regarding efficiency.
Listing 12.14 shows a branch and bound code for solving the problem
that is very similar to the one in Listing 12.12. The main difference be-
tween them is the new (fourth) parameter max_v that stores the bound.
x = partial value
y = bound on maximum possible value
−vi (do not include object i) +vi (include object i)
x x + vi
y − vi y
(a)
w = [3, 6, 9, 5]
0
v = [7, 2, 10, 4]
23
C = 15 −7 +7
0 7
16 23
−2 +2 −2 +2
0 2 7 9
14 16 21 23
−10 +10 −10 +10 −10 +10 −10 +10
0 10 2 12 7 17 9 19
4 14 6 16 11 21 13 23
6 < 14 11 < 14 13 < 17
−4 +4 −4 +4 −4 +4 −4 +4
0 4 10 14 12 16 17 21
0 4 10 14 12 16 17 21
12 < 14
Optimal solution: [1, 0, 1, 0]
(b)
Recursion tree of a branch and bound algorithm for the 0-1

Figure 12.16
knapsack problem.
Listing 12.14 Branch and bound code for solving the 0-1 knapsack prob-
lem.
1 def knapsack_0_1_bnb(i, w_left , current_v , max_v , sol ,
2 opt_sol , opt_v , w, v, C):
3 # Check base case
4 if i == len(sol):
5 # Check if better solution has been found
6 if current_v > opt_v:
7 # Update optimal value and solution
8 opt_v = current_v
9 for k in range (0, len(sol)):
10 opt_sol [k] = sol[k]
11 else:
14
16 # according to (capacity ) constraint
17 if k * w[i] <= w_left :
18
19 # Update maximum possible value
20 new_max_v = max_v - (1 - k) * v[i]
21
23 # according to optimal value
24 if new_max_v > opt_v:
25
27 sol[i] = k
28
29 # Update remaining capacity
30 # and partial value
31 new_w_left = w_left - k * w[i]
32 new_current_v = current_v + k * v[i]
33
35 opt_v = knapsack_0_1_bnb(i + 1, new_w_left ,
36 new_current_v ,
37 new_max_v , sol ,
38 opt_sol , opt_v ,
39 w, v, C)
40
41 # return value of optimal solution found so far
42 return opt_v
Auxiliary code for the branch and bound algorithm related

Listing 12.15
to the 0-1 knapsack problem.
1 def knapsack_0_1_branch_and_bound_wrapper (w, v, C):
2 sol = [0] * (len(w))
3 opt_sol = [0] * (len(w))
4 total_v = knapsack_0_1_bnb(0, C, 0, sum(v), sol ,
5 opt_sol , -1, w, v, C)
6 print_knapsack_solution(opt_sol , w, v, C, total_v )
7
8
9 w = [3, 6, 9, 5] # List of object weights
10 v = [7, 2, 10, 4] # List of object values
11 C = 15 # Weight capacity of the knapsack
12 knapsack_0_1_branch_and_bound_wrapper (w, v, C)
Note that it is initialized to the sum of all of the values of the objects
in the wrapper method. In the recursive case, after making sure that a
partial solution with a new candidate is valid (in line 17), the method
computes the new value of the bound new_max_v in line 20. Observe
that when k = 0 the new bound is decreased by vi , while when k = 1 it
remains unaltered. Subsequently, the method uses another if statement
to check if it can prune the tree according to the new bound and the
optimal value of a solution computed in earlier calls (line 24). The rest
of the code is analogous to the backtracking algorithm. Finally, List-
ing 12.15 contains an associated wrapper method, defines an instance of
the problem (through w, v, and C), and solves it.
12.8 EXERCISES
Exercise 12.1 — There are numerous ways to build algorithms that
generate subsets of elements. The methods described in Section 12.2.1
were based on binary recursion trees. The goal of this exercise is to
implement alternative procedures that also print all of the subsets of n
items provided in a list. However, they must generate a subset at each
node of the recursion tree illustrated in Figure 12.17. Observe that there
are exactly 2n nodes, and the labels 0, 1, and 2 represent indices of the
elements of an initial input list [a, b, c]. Furthermore, instead of using
binary lists to indicate the presence of items in a subset, the partial
solutions will contain precisely these indices. For example, the list [0, 2]
{}
0 2
1
{a} {b}
1 {c}
2 2
{a, b}
{a, c} {b, c}
2
{a, b, c}
Figure 12.17Alternative recursion tree of an algorithm that generates all

of the subsets of three items.
r
r
r
r
Figure 12.18 One solution to the four-rooks puzzle.
will represent {a, c}. Therefore, partial solutions will also correspond to
complete solutions.
Exercise 12.2 — Implement a backtracking algorithm that prints all

of the solutions to the n-rooks puzzle. It is analogous to the n-queens
puzzle, but uses rooks instead of queens. Since rooks can only move
vertically and horizontally, two (or more) rooks cannot appear in the
same row or column. However, several can appear on a same diagonal.
Figure 12.18 shows a solution to the puzzle.
Exercise 12.3 — Implement a backtracking function that counts the

number of valid solutions to a sudoku puzzle that may not be well-posed
(i.e., it may have 0, 1, or more solutions).
8 1 6 15
3 5 7 15
4 9 2 15
15 15 15 15 15
Figure 12.19 A 3 × 3 magic square.
N N n
N N
n nn
N N n
N N n
(a) (b)
Figure 12.20 Chess knight moves.
Exercise 12.4 — A magic square is an n×n grid of the first n2 positive

integers such that the sum of the elements in each row, column, and
diagonal is equal. The particular sum is called the “magic constant,”
and is n(n2 + 1)/2 for a general n. Figure 12.19 shows an example of a
3×3 magic square, where the magic constant is 15. Design a backtracking
algorithm that prints every possible 3 × 3 magic square. Note: it is not
necessary to implement a method that finds magic squares for a general
n. What problem could we run into when computing magic squares of
larger values of n?
Exercise 12.5 — A knight can jump from a square to another one in L-

shaped patterns, as shown in Figure 12.20(a), where (b) shows a sequence
of four moves. The “knight’s tour” problem consists of determining a
sequence of moves of a knight on an n × n chessboard, when staring from
Figure 12.21 An instance and solution to the traveling salesman problem.
a specific square, such that it visits every square only once. Implement
a backtracking algorithm that provides one knight’s tour, and test it for
n = 5 and n = 6. Besides n, the method will receive the coordinates of
an initial square on the chessboard. Finally, it is not necessary to search
for a “closed” tour that ends at a square that is one move away from the
initial square.
Exercise 12.6 — The “traveling salesman problem” is a classical dis-

crete optimization problem. Given n cities, the goal is to find the shortest
path that visits each one exactly once and returns to the starting city.
Figure 12.21 shows the shortest path for 10 cities. Implement a back-
tracking algorithm that solves the problem. Assume that the locations
of the cities are specified in a text file, where each line contains the x
and y coordinates of the cities on a two-dimensional map. The file can
be read with the loadtxt method from the NumPy package. The start-
ing city will be the one specified in the first line of the file. In addition,
consider Euclidean distances between cities (the salesman travels from
one city to another by following a straight line). These can be computed,
for example, with the method pdist in the SciPy package. In particular,
the recursive backtracking method should receive a matrix of distances
between the cities (their locations are not necessary when using this
distance matrix). Finally, test the code for n ≤ 10.
Exercise 12.7 — The following problem is known as “tug of war.”

Given a nonempty set of n numbers, where n is even, the goal consists
of dividing it into two subsets of n/2 elements in order to minimize the
absolute difference between the sum of the numbers in each subset. For
set 3 5 9 14 20 24 set 3 5 9 14 20 24
solution 0 1 1 0 0 1 solution 1 2 5
(a) (b)
Figure 12.22 Two ways to represent a solution for the tug of war problem.
example, given the set {3, 5, 9, 14, 20, 24}, the optimal way to partition
it leads to the two subsets: {5, 9, 24}, and {3, 14, 20}. The sums of their
elements are 38 and 37, and the absolute difference between these sums
is 1.
Implement a backtracking algorithm based on generating subsets
that solves this problem. Assuming that the input set is coded as a list,
the solution would be a binary list of length n with n/2 zeros, and n/2
ones. The positions of the zeros (or ones) would indicate the locations of
the elements of a particular subset. In the example, one solution could
be the list [0, 1, 1, 0, 0, 1] (the list [1, 0, 0, 1, 1, 0] would be equivalent),
representing the subset {5, 9, 24}.
In addition, implement a more efficient strategy where the solution
is a list s of length n/2 whose elements appear in increasing order, and
correspond to the locations of the elements of a particular subset. For
example, the subset {5, 9, 24} would be represented by the list [1, 2, 5].
Figure 12.22 shows these two ways to represent a solution.
Exercise 12.8 — In the problem described in Section 12.4 the goal

consisted of printing every subset T of a set of n positive integers S,
for which the sum of the elements of T was equal to some integer value
x. Design an alternative recursive function that instead computes the
(valid) subset T with the least cardinality (i.e, number of elements). In
particular, it will store the optimal subset in a list, and will return the
optimal cardinality. For example, for S = {1, 2, 3, 5, 6, 7, 9} and x = 13, the
method would store the subset {6, 7}, and return the value 2. In addition,
note that in this optimization problem the optimal subset may not be
unique, but the optimal cardinality is. For instance, for S = {2, 6, 3, 5}
and x = 8, both {2, 6} and {3, 5} would be optimal subsets. Code the
partial (and optimal) solution as a list of binary digits, similarly to the
method in Listing 12.7. Also, prune the tree considering the validity of a
solution and the best cardinality found. Finally, code a wrapper method
that computes the optimal subset, and prints it if it exists (it may not
be possible to find a subset T of S whose elements add up to x).
Further reading
MONOGRAPHS IN RECURSION
This book has provided a broad coverage of recursion, containing essen-
tial topics for designing recursive algorithms. However, the reader can
find additional aspects, examples, and implementation details in other
references. In particular, the following book:
• Jeffrey Soden Rohl. Recursion Via Pascal. Cambridge Computer Sci-
ence Texts. Cambridge University Press, 1st edition, August 1984
contains numerous examples in Pascal of recursive algorithms on data
structures (linked lists, trees, or graphs) implemented through point-
ers. In contrast, the current book avoids pointers since they are not
used explicitly in Python. The suggested reference includes backtrack-
ing algorithms for generating additional combinatorial entities such as
combinations, compositions, and partitions. Lastly, it contains a chapter
on recursion elimination, which offers low-level explanations on how to
transform recursive programs into equivalent iterative versions.
Another book that focuses entirely on recursion is:
• Eric S. Roberts. Thinking Recursively. Wiley, 1st edition, January 1986,
which contains examples in Pascal, and a more recent edition:
• Eric S. Roberts. Thinking Recursively with Java. Wiley, 1st edition,
February 2006,
where the code is in Java. The book contains a chapter on recursive data
types, and another on the implementation of recursion from a low-level
point of view.
Java programmers can also benefit from:
• Irena Pevac. Practicing Recursion in Java. CreateSpace Independent
Publishing Platform, 1st edition, April 2016,
which contains examples related to linked lists, linked trees, and graph-
ical problems.
403
404 Further reading
DESIGN AND ANALYSIS OF ALGORITHMS

The current book contains problems that can be solved by applying al-
gorithm design techniques such as divide and conquer, or backtracking.
There are numerous excellent texts that analyze more advanced prob-
lems, describe other design techniques like greedy algorithms or dynamic
programming, or rely on advanced data structures in order to construct
efficient algorithms. The following list of books is only a small subset of
the broad literature on algorithm design (and analysis) techniques:
• Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Clif-
ford Stein. Introduction to Algorithms. The MIT Press, 3rd edition,
2009.
• Robert Sedgewick and Kevin Wayne. Algorithms. Addison-Wesley Pro-
fessional, 4th edition, 2011.
• Anany V. Levitin. Introduction to the Design and Analysis of Algo-
rithms. Addison-Wesley Longman Publishing Co., Inc., Boston, MA,
USA, 3rd edition, 2012.
• Michael T. Goodrich, Roberto Tamassia, and Michael H. Goldwasser.
Data Structures and Algorithms in Python. Wiley Publishing, 1st edi-
tion, 2013.
More ambitious readers can explore texts aimed at training for pro-
gramming competitions (e.g., the ACM International Collegiate Pro-
gramming Contest, or the International Olympiad in Informatics). Well-
known references include:
• Steven S. Skiena and Miguel Revilla. Programming Challenges: The
Programming Contest Training Manual. Springer-Verlag New York,
Inc., Secaucus, NJ, USA, 2003.
• Steven S. Skiena. The Algorithm Design Manual. Springer Publishing
Company, Incorporated, 2nd edition, 2008.
Other textbooks focus exclusively on the analysis of algorithms, such
as:
• Jeffrey J. McConnell. Analysis of Algorithms: An Active Learning Ap-
proach. Jones and Bartlett Publishers, Inc., USA, 1st edition, 2001
Among these,
• Robert Sedgewick and Philippe Flajolet. An Introduction to the Anal-
ysis of Algorithms. Addison-Wesley Professional, 2nd edition, 2013
is considerably more advanced, and is used by the author in a massive
open online course offered by coursera.org. Similarly, this and other on-
line learning platforms offer excellent courses on algorithms.
Further reading 405
FUNCTIONAL PROGRAMMING
Recursion is omnipresent in functional programming. Thus, program-
mers should have mastered the contents of this book in order to be
competent in this programming paradigm. Popular references include:
• Harold Abelson and Gerald J. Sussman. Structure and Interpretation
of Computer Programs. MIT Press, Cambridge, MA, USA, 2nd edition,
1996.
• Richard Bird. Introduction to Functional Programming Using Haskell.
Prentice Hall Europe, April 1998.
• Martin Odersky, Lex Spoon, and Bill Venners. Programming in Scala:
A Comprehensive Step-by-Step Guide. Artima Incorporation, USA, 1st
edition, 2008.
The book by Odersky et al. is used in a highly recommended course also
offered at coursera.org.
PYTHON
It is assumed that the reader of this book has some programming expe-
rience. The following popular texts can be useful to readers interested
in learning more Python features, looking up implementation details,
or developing alternative and/or more efficient recursive variants of the
examples covered throughout the book:
• Mark Pilgrim. Dive Into Python 3. Apress, Berkely, CA, USA, 2009.
• Mark Summerfield. Programming in Python 3: A Complete Introduc-
tion to the Python Language. Addison-Wesley Professional, 2nd edition,
2009.
• Mark Lutz. Learning Python. O’Reilly, 5th edition, 2013.
RESEARCH IN TEACHING AND LEARNING RECURSION

The current book incorporates several ideas that stem from research in
teaching and learning recursion. In particular, it focuses on the declar-
ative abstract level of problem decomposition and induction, instead of
focusing on a computational model (e.g., a trace mental model) or iter-
ative/imperative thinking. This is suggested, for example, in:
• D. Ginat and E. Shifroni. Teaching recursion in a procedural environ-
ment – how much should we emphasize the computing model? SIGCSE
Bull., 31(1):127–131, 1999.
• R. Sooriamurthi. Problems in comprehending recursion and suggested
solutions. SIGCSE Bull., 33(3):25–28, 2001.
406 Further reading
The diagrams and methodology (the template is similar to the one in

Sooriamurthi’s paper) used in Chapter 2 have been introduced precisely
in order to focus on declarative thinking. In addition, the structure of the
book also reflects this goal. Elements such as tracing, recursion trees, the
program stack, or the relationship between iteration and tail recursion
are not introduced until Chapters 10 and 11.
The book also includes examples from articles where I have addressed
teaching mutual recursion, and tail (and nested) recursion through func-
tion generalization:
• M. Rubio-Sánchez, J. Urquiza-Fuentes, and C. Pareja-Flores. A gentle
introduction to mutual recursion. SIGCSE Bull., 40(3):235–239, 2008.
• Manuel Rubio-Sánchez. Tail recursive programming by applying gen-
eralization. In Proceedings of the Fifteenth Annual Conference on In-
novation and Technology in Computer Science Education, ITiCSE ’10,
pages 98–102. ACM, 2010.
Finally, there is rich literature on research in teaching and learning
recursion. The following recent surveys contain hundreds of references
related to the topic:
• Christian Rinderknecht. A survey on teaching and learning recursive
programming. Informatics in Education, 13(1):87–119, 2014.
• Renée McCauley, Scott Grissom, Sue Fitzgerald, and Laurie Murphy.
Teaching and learning recursive programming: a review of the research
literature. Computer Science Education, 25(1):37–66, 2015.
Index
Accumulator variable, 337, 347, Strassen’s, 190, 204

374 Alphabet, 273
Ackermann function, 342, 351 Arithmetic series, 62
Activation Asymptotic notation, 73, 76
record, 306 big-O, 73
tree, 300, 327, 334, 351 big-Omega, 75
Active big-Theta, 75
frame, 306
subroutine, 306 Backtracking, 353
Addition, see sum template, 353
Algorithm, 8 Backward substitution, see
bisection, 157, 336 expansion method
breath-first search, 313 Base, 115
brute-force, 353 Base case, 4, 34, 295
depth-first search, 213, 313, misconception, 24, 35, 295
316 missing, 36, 52
Euclid’s, 164, 300 redundant, 35, 132
Hoare’s partitioning, 148, 150, Base conversion, 117, 340, 348
179 Basketball, 251
Horner’s, 124 Binomial coefficient, 58, 131, 226,
in-place, 179 242, 245, 331
Karatsuba’s, 183, 203 Bits, 115, 131, 183, 251
Newton’s method, 169 Branch and bound, 388, 393
out-of-place, 179
Calculator, 273
searching
Call stack, 306
binary, 142, 160, 168
Candidate, 354, 361, 380, 384
linear, 139
Catalan number, 248, 251
quickselect, 148, 155
Ceiling, 66, 80
sorting, 172, 173
Characteristic polynomial, 90
counting sort, 167, 174
Combination, 240
insertion sort, 132
Combinatorics, 235
merge sort, 174
Compiler, 273
quicksort, 148, 177
Complete solution, 354
selection sort, 121
407
408 Index
Computational complexity, 70 misconceptions, 83

Computational problem, 7 Expression, 274, 279, 282
Constraint satisfaction problem, parenthesized, 279, 287
353
Control Factor, 279, 286
flow, 292 Factorial, 4, 35, 238, 337
stack, 306 Fibonacci
Counting problem, 235 function, 4, 11, 17, 27–29,
244, 260, 265, 268, 288,
Data structure, 6 300, 318, 351
FIFO, 305 number, 4, 250, 258, 288, 323
LIFO, 305 sequence, 4
list, 6, 17 File system, 313
queue, 305, 313 Floor, 66, 80
stack, 305, 313, 316 Formal
tree, 6 grammar, 273, 279
Debug, 26, 52, 295, 327–330, 337, language, 273
362 Fractal, 220, 224, 230, 233
Declarative programming, 25, 26 Koch, 330
Dependency graph, 322, 331 Function generalization, 344
Dequeue, 306 Functional programming, 25
Descendant, 7, 29
Difference equation, 89 General method for difference
Digital root, 343 equations, 89
Digits, 5, 47, 118, 131, 132, 134, Generality, 36
167, 203 Geometric series, 63
Discrete state space, 354 Grammar, see formal grammar
Divide and conquer, 28, 51, 84, Greatest common divisor, 164
133, 171 Guido van Rossum, 337
Domino tiles, 251
Heap, 312
Dot product, 68
Hilbert curve, 233
Droste effect, 3
Dynamic programming, 217, 322 Imperative programming, 25, 341
Induction, 20, 41
Efficiency, see computational
hypothesis, 21
complexity
proof, 20
Enqueue, 306
Inorder, 216
Execution time, 70
Iteration, 25, 333, 337
Exhaustive search, 353, 355
method, see expansion
Expansion method, 80
method
Index 409
Knapsack, 388 Optimization problem, 217, 353,

Knight’s tour, 399 388, 393, 400, 401
Koch Order of growth, 71, 76
curve, 220 Overlapping subproblems, 219,
snowflake, 220 317
Limit, 59 Palindrome, 120, 217, 230, 320,

List, 132, 167, 168, 203, 350 332
head, 30 Parity, 52, 129, 254
tail, 30 Partial solution, 354
Logarithm, 58, 76 Partially filled array, 18
Partitioning scheme, 148, 155
Magic square, 399 Pascal’s triangle, 125, 131
Majority element, 180 Passive flow, 295
Manhattan paths, 244 Permutation, 236, 364, 369
Master theorem, 84, 176, 304 with repetition, 245
Matplotlib, 193, 223, 382 Pitfall, 36, 37, 52
Matrix, 67, 129 Polygon, 245
block, 40 Polynomial, 167, 203
product, 69, 186, 190, 204 Pop, 305
rotation, 69, 223 Postorder, 216, 299
symmetric, 69 Power, 58, 106, 129, 240, 350
transpose, 203 Precondition, 16
Matryoshka dolls, 3 Preorder, 216, 299, 357
Maze, 377 Problem, 37
McCarthy 91 function, 342 decomposition, 5, 7, 37, 171
Memoization, 219, 317, 331, 332 instance, 5, 8
Mental model, 41, 291 size, 5, 32, 110, 171
Methodology, see recursive design Procedure, 47
template Product, 65, 130, 183, 203
Moor traversal, 205 Production rule, 273, 279
Multiplayer game, 255, 288 Program stack, 26, 306, 308–310
Multiplication, see product Programming paradigm, 25
principle, 239 Push, 305
Pyramid, 43, 63, 248, 252
N-queens puzzle, 354, 368
N-rooks puzzle, 398 Rabbit population, 256, 258
NumPy, 129, 208, 223 Radix, see base
Recurrence relation, 57, 76
Optimal solution, 388, 392
homogeneous, 90
410 Index
nonhomogeneous, 94 trace, 309

Recursion, see recursion Staircase climbing, 242
depth, 36, 312 String, 119, 120, 136, 276
infinite, 14, 36, 52, 53, 312 Sublist, 13, 230
tree, 297, 326, 356, 359, 364 Subproblem, 5, 8, 37
height, 303, 310 graph, 322
pruned, 358, 374, 393 Subsequence, 13, 217, 230, 332
Recursive Subset, 359, 372, 388, 397
call, 14 Substring, 217, 320
case, 4, 37, 41 Sudoku, 384, 398
descent parser, 274, 282 Sum, 4, 11, 12, 14, 19, 42, 60, 113,
design 130, 350
diagram, 37, 41 partial, 62
template, 31 Swamp traversal, 205, 226
leap of faith, 22
type, 27 Term, 279, 281, 284, 285
binary, 28 Test, 52
final, 133 Token, 273, 274, 279
indirect, 28, 253 Towers of Hanoi, 209
linear, 27, 105, 133, 300 cyclic, 268, 288
multiple, 28, 171, 205 variant, 229, 268, 288
mutual, 28, 253 Traceback module, 309
nested, 29, 342 Transpose, 68, 203
tail, 27, 300, 333, 337 Traveling salesman problem, 400
Resistor, 127 Tree, 313
Return address, 306 binary, 251, 304, 359, 389
Romanesco broccoli, 1 binary search, 143, 215, 326
Root finding, 157, 169 method, 304
Ruler, 227 traversal, 213
Triangulation, 245
Scalability, 336 Trigonometry, 66
Short-circuit evaluation, 135, 137 Tromino tiling, 192
Sierpiński Tug of war, 400
carpet, 224
Variation with repetition, 239, 250
triangle, 3, 230
Vector, 68, 221
Skyline, 196
Vowel, 131
Spin-out® brainteaser, 289
Stack Water treatment plant, 263
frame, 306, 308, 310, 335 Woodcutter problem, 158, 350
overflow, 26, 312, 336 Worst case, 75

Introduction Recursive Programming

Uploaded by

Introduction Recursive Programming

Uploaded by

Introduction to

© 2018 by Taylor & Francis Group, LLC

No claim to original U.S. Government works

Printed on acid-free paper

International Standard Book Number-13: 978-1-4987-3528-5 (Paperback)

Chapter 1 Basic Concepts of Recursive Programming 1

1.1 RECOGNIZING RECURSION 1

Chapter 2 Methodology for Recursive Thinking 31

2.1 TEMPLATE FOR DESIGNING RECURSIVE ALGO-

2.2 SIZE OF THE PROBLEM 32

Chapter 3 Runtime Analysis of Recursive Algorithms 57

3.1 MATHEMATICAL PRELIMINARIES 57

Chapter 4 Linear Recursion I: Basic Algorithms 105

4.1 ARITHMETIC OPERATIONS 106

4.1.1 Power function 106

Chapter 5 Linear Recursion II: Tail Recursion 133

5.1 BOOLEAN FUNCTIONS 134

5.7 THE WOODCUTTER PROBLEM 158

Chapter 6 Multiple Recursion I: Divide and Conquer 171

6.1 IS A LIST SORTED IN ASCENDING ORDER? 172

Chapter 7 Multiple Recursion II: Puzzles, Fractals, and

7.1 SWAMP TRAVERSAL 205

Chapter 8 Counting Problems 235

8.1 PERMUTATIONS 236

Chapter 9 Mutual Recursion 253

9.1 PARITY OF A NUMBER 254

Chapter 10 Program Execution 291

10.1 CONTROL FLOW BETWEEN SUBROUTINES 292

10.3.1 Stack frames 306

Chapter 11 Tail Recursion Revisited and Nested Recursion 333

11.1 TAIL RECURSION VS. ITERATION 333

Chapter 12 Multiple Recursion III: Backtracking 353

12.1 INTRODUCTION 354

12.2.1 Subsets 359

FURTHER READING 403

Recursion is one of the most fundamental concepts in computer science

ily learn iteration, understanding and controlling how programs work. In

The book proposes numerous exercises at the end of the chapters,

divide and conquer, or backtracking). Additionally, since the book pro-

BOOK CONTENT AND ORGANIZATION

putational problems in large discrete state spaces. The strategy is usu-

POSSIBLE COURSE ROAD MAPS

1.1 Examples of recursive entities. 2

2.1 General template for designing recursive algorithms. 32

2.5 General diagram for thinking about recursive cases

3.1 Graphical mnemonic for determining the quadratic

3.6 Orders of growth typically used in computational

4.1 Conversion of 142 into its representation (1032) in

5.1 Decomposition related to the binary search algorithm. 141

5.6 Example of Hoare’s partition method. 151

6.1 Merge sort algorithm. 174

7.1 Swamp traversal problem. 206

7.8 Concrete example of the decomposition of the in-

8.1 Possible permutations (lists) of the ﬁrst four positive

8.9 Manhattan paths problem. 244

9.1 Calls of mutually recursive methods. 254

9.9 Decompositions of the three problems for the mutu-

10.1 Sequence of method calls and returns for the code

10.12 Recursion tree associated with the recurrence

11.1 State of the program stack when running the base

12.1 One solution to the four-queens puzzle. 354

12.2 Partial solutions within complete solutions that are

12.21 An instance and solution to the traveling salesman

3.1 Concrete values of common functions used in com-

11.1 Program state related to the iterative factorial func-

1.1 Python code for adding the ﬁrst n natural numbers. 16

2.8 Incomplete Python code for adding the ﬁrst n posi-

5.3 Linear-recursive function that determines if two

6.8 Karatsuba’s fast algorithm for multiplying two non-

9.7 Mutually recursive functions for tokenizing a math-

10.14 Code for generating a Koch fractal based on the

To iterate is human, to recurse divine.

ECURSION is a broad concept that is used in diverse disciplines

1.1 RECOGNIZING RECURSION

Tree branches Branching rivers

Romanesco broccoli Spiral Droste eﬀect