formal language

A P R A C T I C A L I N T R O D U C T I O N

Adam Brooks Webber

Monmouth College

Franklin, Beedle & Associates, Inc.

8536 SW St. Helens Drive, Ste. D
Wilsonville, Oregon 97070
503/682-7668
www.fbeedle.com
Publisher Jim Leisy ([email protected])
Production Editor Tom Sumner
Editor Stephanie Welch
Order Processing Jaron Ayres

Printed in the U.S.A.

© 2008 Franklin, Beedle & Associates, Incorporated. No part of this book may be
reproduced, stored in a retrieval system, transmitted, or transcribed, in any form or by
any means—electronic, mechanical, telepathic, photocopying, recording, or otherwise—
without prior written permission of the publisher. Requests for permission should be
addressed as follows:
Rights and Permissions
Franklin, Beedle & Associates, Incorporated
8536 SW St. Helens Drive, Ste. D
Wilsonville, Oregon 97070

Library of Congress Cataloging-in Publication Data

Webber, Adam Brooks.
Formal language : a practical introduction / Adam Brooks Webber.
p. cm.
ISBN-13: 978-1-59028-197-0
1. Formal languages--Textbooks. 2. Computational complexity--Textbooks. I. Title.
QA267.3.W423 2007
005.13'1--dc22
2007027145
 c o n t e n t s
Preface viii

CHAPTER 1 Fundamentals 11
1.1 Alphabets 12
1.2 Strings 12
1.3 Languages 14
Exercises 15

CHAPTER 2 Finite Automata 7

2.1 Man, Wolf, Goat, and Cabbage 8
2.2 Not Getting Stuck 10
2.3 Deterministic Finite Automata 10
2.4 The 5-Tuple 13
2.5 The Language Accepted by a DFA 14
Exercises 15

CHAPTER 3 Closure Properties for Regular Languages 21

3.1 Closed Under Complement 22
3.2 Closed Under Intersection 23
3.3 Closed Under Union 26
3.4 DFA Proofs Using Induction 27
3.5 A Mystery DFA 30
Exercises 32

CHAPTER 4 Deterministic-Finite-Automata Applications 35

4.1 An Overview of DFA Applications 36
4.2 A DFA-based Text Filter in Java 36
4.3 Table-driven Alternatives 41
Exercises 43

CHAPTER 5 Nondeterministic Finite Automata 45

5.1 Relaxing a Requirement 46
5.2 Spontaneous Transitions 47
5.3 Nondeterminism 50
5.4 The 5-Tuple for an NFA 50
5.5 The Language Accepted by an NFA 52
Exercises 53

iii
iv CONTENTS

CHAPTER 6 NFA Applications 59

6.1 An NFA Implemented with a Backtracking Search 60
6.2 An NFA Implemented with a Bit-Mapped Parallel Search 62
6.3 The Subset Construction 65
6.4 NFAs Are Exactly as Powerful as DFAs 69
6.5 DFA or NFA? 70
6.6 Further Reading 71
Exercises 71

CHAPTER 7 Regular Expressions 75

7.1 Regular Expressions, Formally Defined 76
7.2 Examples of Regular Expressions 77
7.3 For Every Regular Expression, a Regular Language 79
7.4 Regular Expressions and Structural Induction 81
7.5 For Every Regular Language, a Regular Expression 83
Exercises 85

CHAPTER 8 Regular-ExpressionApplications 91
8.1 The egrep Tool 92
8.2 Nonregular Regexps 94
8.3 Implementing Regexps 95
8.4 Regular Expressions in Java 96
8.5 The lex Tool 97
8.6 Further Reading 100
Exercises 101

CHAPTER 9 Advanced Topics inRegular Languages 105

9.1 DFA Minimization 106
9.2 Two-way Finite Automata 108
9.3 Finite-state Transducers 109
9.4 Advanced Regular Expressions 111
9.5 Further Reading 113
Exercises 113

CHAPTER 10 Grammars 117

10.1 A Grammar Example for English 118
10.2 The 4-Tuple 120
10.3 The Language Generated by a Grammar 121
10.4 Every Regular Language Has a Grammar 122
10.5 Right-Linear Grammars 125
10.6 Every Right-Linear Grammar Generates a Regular Language 127
Exercises 128
 CONTENTS v
CHAPTER 11 Nonregular Languages 133
11.1 The Language {a b }
n n
134
11.2 The Language {xxR} 136
11.3 Pumping 137
11.4 Pumping-Lemma Proofs 139
11.5 Strategies 141
11.6 Pumping and Finite Languages 142
Exercises 143

CHAPTER 12 Context-free Languages 145

12.1 Context-free Grammars and Languages 146
12.2 Writing CFGs 147
12.3 CFG Applications: Backus-Naur Form 150
12.4 Parse Trees 151
12.5 Ambiguity 153
12.6 Extended BNF 154
12.7 Exercises 156

CHAPTER 13 Stack Machines 159

13.1 Stack Machine Basics 160
13.2 A Stack Machine for {anbn} 162
13.3 A Stack Machine for {xxR} 164
13.4 Stack Machines, Formally Defined 166
13.5 An Example: Equal Counts 168
13.6 An Example: A Regular Language 170
13.7 A Stack Machine for Every CFG 171
13.8 A CFG for Every Stack Machine 173
Exercises 177

CHAPTER 14 The Context-free Frontier 181

14.1 Pumping Parse Trees 182
14.2 The Language {anbncn} 186
14.3 Closure Properties for CFLs 187
14.4 Nonclosure Properties 189
14.5 A Pumping Lemma 191
14.6 Pumping-Lemma Proofs 193
14.7 The Language {xx} 195
Exercises 197

CHAPTER 15 Stack Machine Applications 199

15.1 Top-Down Parsing 200
15.2 Recursive-Descent Parsing 203
15.3 Bottom-Up Parsing 205
vi CONTENTS

15.4 PDAs, DPDAs, and DCFLs 207

15.5 Further Reading 209
Exercises 210

CHAPTER 16 Turing Machines 217

16.1 Turing Machine Basics 218
16.2 Simple TMs 219
16.3 A TM for {anbncn} 220
16.4 The 7-Tuple 222
16.5 The Languages Defined by a TM 222
16.6 To Halt or Not to Halt 223
16.7 A TM for {xcx | x ∈ {a, b}* } 225
16.8 Three Tapes 227
16.9 Simulating DFAs 228
16.10 Simulating Other Automata 231
16.11 Further Reading 233
Exercises 233

CHAPTER 17 Computability 237

17.1 Turing-Computable Functions 238
17.2 TM Composition 238
17.3 TM Arithmetic 240
17.4 TM Random Access 241
17.5 Functions and Languages 242
17.6 The Church-Turing Thesis 243
17.7 TM and Java Are Interconvertible 245
17.8 Infinite Models, Finite Machines 250
Exercises 251

CHAPTER 18 Uncomputability 255

18.1 Decision and Recognition Methods 256
18.2 The Language Lu 259
18.3 The Halting Problems 262
18.4 Reductions Proving a Language Is Recursive 264
18.5 Reductions Proving a Language Is Not Recursive 266
18.6 Rice’s Theorem 268
18.7 Enumerators 271
18.8 Recursively Enumerable Languages 275
18.9 Languages That Are Not RE 277
18.10 Language Classifications Revisited 278
18.11 Grammars and Computability 279
18.12 Oracles 281
18.13 Mathematical Uncomputabilities 282
Exercises 283
 CONTENTS vii
CHAPTER 19 Cost Models 289
19.1 Asymptotic Notation 290
19.2 Properties of Asymptotics 292
19.3 Common Asymptotic Functions 294
19.4 A TM Cost Model 296
19.5 A Java Cost Model 298
Exercises 307

CHAPTER 20 Deterministic Complexity Classes 311

20.1 Time Complexity Classes 312
20.2 Polynomial Time 314
20.3 Exponential Time 317
20.4 Space-Complexity Classes 320
20.5 Higher-Complexity Classes 321
20.6 Further Reading 324
Exercises 325

CHAPTER 21 Complexity Classes 331

21.1 Verification Methods 332
21.2 NP, NP-Hard, and NP-Complete 333
21.3 A Tour of Six Decision Problems 336
21.4 The Wider World of NP-Complete Problems 346
21.5 Nondeterministic Turing Machines 351
21.6 The Million-Dollar Question 354
21.7 Further Reading 355
Exercises 355

APPENDIX A From an NFA to a Regular Expression 361

A.1 The Internal Languages of an NFA 362
A.2 The d2r Function 363
A.3 Proof of Lemma 7.2 364

APPENDIX B A Time-Hierarchy Theorem 366

B.1 Another Asymptotic Notation: Little o 367
B.2 Time-Constructible Functions 367
B.3 A Cost Model for run and runWithTimeLimit 368
B.4 A Time-Hierarchy Theorem 368
B.5 Application 370

APPENDIX C Some NP-Hardness Proofs 371

C.1 Cook’s Theorem: SAT Is NP-hard 372
C.2 CNFSAT Is NP-hard 374
C.3 Vertex Cover Is NP-hard 377
C.4 Hamiltonian Circuit Is NP-hard 380
 p r e f a c e
A Word to the Instructor
This book is an undergraduate college-level introduction to formal languages, automata,
computability, and complexity. I emphasize undergraduate. There are other textbooks on this
subject, including some that I admire very much, but most assume a level of mathematical
maturity and interest that only a few of my undergraduate students actually have. Most
of these students are bright people whose primary interest in computing is practical. They
like programming and have some aptitude for it. They hope to find jobs in the computing
industry. They think (sometimes incorrectly) that they are not going to do graduate work on
the theoretical side of computer science. It is for students like these that I wrote this book.
In general, I have found that these students do not thrive on a diet of unmotivated
abstractions. This book tries to lead them gently into the beautiful abstractions at the
foundations of computer science, using concrete examples and exercises. The chapters on
finite automata include applications and implementation techniques; the chapters on stack-
based automata include a discussion of applications to compiler construction; the chapters
on computability use simple Java-like code examples. Included among the exercises are many
Java programming problems, which serve both to hold student interest and to enhance
understanding.
This book does not emphasize proof technique. There’s no doubt that the theory of
formal languages and automata is an ideal place to introduce induction and other proof
techniques for students who need to learn them. The book includes one chapter and a
number of exercises covering inductive proofs, and additional chapters devoted to pumping-
lemma proofs. However, the material is not central to the book, and the instructor can use
as much or as little of it as desired. The book includes proofs for almost all the theorems, but
usually stops short of presenting detailed inductions, and most of the exercises do not require
students to generate novel proofs. My feeling is that too much stress on proof technique
turns many students off. The concepts of formal languages and automata are among the most
beautiful and enduring in all of computer science, and have many practical applications as
well. The book tries to make these treasures accessible to all students, not just to those with a
knack for the mathematically rigorous proof.
This book deviates from the usual syllabus for its subject in one other important respect:
it does not provide a detailed treatment of PDAs and DPDAs. The reason for this is, again,
my concern for maintaining my students’ interest and enthusiasm. The general operation of
PDAs is simple enough, but there are all those special cases to worry about: the end of the
input string, the bottom of the stack, the ε-transitions, the definition of acceptance. These

viii
 PREFACE ix
are minor technicalities, to be sure, but there are enough of them that I find my students
become discouraged—and if they look to other books for guidance that only makes it worse,
because no two agree on these details. The formalism generates too much heat and not
enough light. In place of PDAs this book investigates stack machines, a simpler computational
formalism that captures the CFLs, connects neatly with CFGs, and provides a good
foundation for a discussion of parsing techniques.
With few other exceptions, the ideas, theorems, and proofs in this book are standard.
They are presented without the detailed historical attributions that may be found in
graduate-level texts. However, many of the chapters conclude with a Further Reading section,
giving references into the literature for interested students.
I designed the book to be accessible to sophomores in a computer-science major—
perhaps even to second-semester freshmen. The presentation does not rely on any college-
level mathematical prerequisites. The only computer-science prerequisite is the ability to read
simple code examples: a Java-like syntax is used, but the ability to read any C-family language
will suffice.
In some academic departments, the material on complexity—Chapters 19, 20, and 21,
and Appendices B and C—will be positioned in a separate course, together with related
material on algorithms and complexity analysis. In my experience there is more than enough
even in Chapters 1 through 18 to fill a semester-long course, so those wishing to cover
the entire subject of the book may want to consider skipping or abbreviating, particularly
Chapters 9, 11, 14, and 18.
In many CS and CE departments (such as at the University of Wisconsin—Milwaukee,
where I began the development of this book) a course in this subject is a required part of
the major. At other institutions (including Monmouth College, where I completed it) the
course has gradually fallen into disuse; it has sometimes become an elective, rarely if ever
offered. The reason for this decline, I believe, is the mismatch between the students and
the treatment of the subject. Instructors do not enjoy the drudgery of having to frog-march
undergraduates through a difficult textbook that they barely understand and quickly forget;
students enjoy the experience even less. I hope that Formal Language: A Practical Introduction
will help to rekindle the excitement of teaching and studying this subject.

Introduction
It sometimes seems that everything is a science these days: political science and beauty
science, food science and waste science, creation science and mortuary science. Just for
fun, try searching the Web for any well-formed phrase ending with the word science. Weed
science, reincarnation science, sales science—it’s hard to find one that is not in use.
Because of this oddly mixed company, the name computer science triggers some
skepticism. That’s unfortunate, because it is actually an excellent field of study for many
x PREFACE

undergraduates. Looking for a liberal-arts education? Computer science has deep connections
throughout the web of human knowledge and is highly relevant to broad, ongoing changes
in our culture. Looking for a rigorous scientific education? There are research-oriented
departments of computer science at all the world’s great universities, engaged in a dramatic
scientific adventure as our young science begins to reveal its secrets. Looking for a vocational
education? Computer science is a solid foundation for a variety of career paths in the
computer industry.
The subject of this book, formal language, is at the heart of computer science, and it
exemplifies the virtues just cited:
• Formal language is connected to many other branches of knowledge. It is where
computer science, mathematics, linguistics, and philosophy meet. Understanding
formal language helps you see this interconnectedness, so that you will never make
the mistake of thinking of computer science as a separate intellectual kingdom.
• Formal language is a rigorous branch of mathematics, with many open questions at
the frontiers. This book covers only the basics, but if you find the basics interesting
there is much more to study. Some advanced topics are identified here, with
pointers to further reading; others may be found in any graduate-level text.
• Formal language is very useful. This book stresses applications wherever they arise,
and they arise frequently. Techniques that derive from the study of formal language
are used in many different practical computer systems, especially in programming
languages and compilers.
In addition, formal language has two special virtues, not shared with most of the other
branches of computer science:
• Formal language is accessible. This book treats formal language in a way that does
not assume the reader knows any advanced mathematics.
• Finally, formal language is stable—at least in the elementary parts. Almost
everything in this book has been known for 30 years or more. That’s a very long
time in computer science. The computer you bought as a freshman may be obsolete
by the time you’re a senior, and the cutting-edge programming language you
learned may be past its prime, but the things you learn about formal language will
not lose their relevance.
No one who loves language can take much pleasure in the prospect of studying a subject
called formal language. It sounds suspiciously abstract and reductionistic. It sounds as if all
the transcendent beauty of language will be burned away, fired under a dry heat of definitions
and theorems and proofs, until nothing is left but an ash of syntax. It sounds abstract—and
it is, undeniably. Yet from this abstraction arise some of the most beautiful and enduring
ideas in all of computer science.
 PREFACE xi
This book has two major goals. The first is to help you understand and appreciate
the beautiful and enduring ideas of formal language. These ideas are the birthright of all
computer scientists, and they will profoundly change the way you think about computation.
They are not only among the most beautiful, but also among the most useful tools in
computer science. They are used to solve problems in a wide variety of practical applications,
and they are especially useful for defining programming languages and for building language
systems. The second purpose of this book is to help you develop a facility with these useful
tools. Our code examples are in Java, but they are not particularly Java-centric and should be
accessible to any programmer.
There is also a third major reason to study formal language, one that is not a primary
focus of this book: to learn the techniques of mathematical proof. When you are learning
about formal language, it can also be a good time to learn proof techniques, because the
subject is full of theorems to practice on. But this book tries to make the beautiful and useful
ideas for formal language accessible to students at all levels of mathematical interest and
ability. To that end, although the book presents and discusses many simple proofs, it does not
try to teach advanced proof techniques. Relatively few of the exercises pose challenging proof
problems. Those planning graduate-level study of theoretical computer science would be well
advised not to rely exclusively on this book for that kind of training.

Acknowledgments
Today the territory mapped in this book is part of the intellectual commons of our
discipline, but it was not always so. I am indebted to the many researchers, living and dead,
who first explored this territory. I am honored to be able to share their work with others.
I am particularly indebted to those scholars who first shared it me, for their style, grace,
and elegant precision: Scot Drysdale at Dartmouth College, and Dexter Kozen and Juris
Hartmanis at Cornell University.
Thanks to Jim Leisy, my publisher at Franklin, Beedle & Associates. His feedback and
support, both for this project and for my previous book, have been extremely valuable.
Thanks also to Stephanie Welch and Tom Sumner, my editors, for their great patience and
diligence with this project. Thanks to the anonymous reviewers who commented on early
drafts of the book. The remaining defects are, of course, entirely my own.
I have been encouraged and supported in this work by two excellent department
chairmen: K. Vairavan at the University of Wisconsin—Milwaukee, and Lyle Welch at
Monmouth College. Monmouth has been a particularly supportive environment for a
project like this, with its emphasis on the undergraduate experience. I’m proud to have been
able to teach there, and I thank the Monmouth students who endured my well-meaning
experimentation with early drafts of this book.
xii PREFACE

Finally, thanks to my family: my wife Kelly, my children Fern and Fox, my parents
Howard and Helen Webber, and my wife’s parents, Harold and Margaret Autrey. Their
loving support has been unfailing. For their sake I hope that the next book I write will be
more fun for them to read.

Web Site
The web site for this book is at http://www.webber-labs.com/fl.html. Materials there
include a full set of slides for each chapter, and all the larger code examples from the text
and exercises. There are also instructions for contacting the author to report defects, and for
accessing additional instructors-only materials.
 1
CHA P TER

Fundamentals

Algebraists use the words group, ring, and field in technical ways,
while entomologists have precise definitions for common words
like bug and fly. Although it can be slightly confusing to overload
ordinary words like this, it’s usually better than the alternative,
which is to invent new words. So most specialized fields of study
make the same choice, adding crisp, rigorous definitions for words
whose common meaning is fuzzy and intuitive.
The study of formal language is no exception. We use crisp, rigorous
definitions for basic terms such as alphabet, string, and language.

1
2 CHAPTER ONE: FUNDAMENTALS

1.1 Alphabets
Formal language begins with sets of symbols called alphabets:

An alphabet is any finite set of symbols.

A typical alphabet is the set = {a, b}. It is the set of two symbols, a and b. There is no
semantics; we do not ascribe any meaning to a or to b. They are just symbols, which could as
well be 0 and 1, or and .
Different applications of formal languages use different alphabets. If you wanted to work
with decimal numbers, you might use the alphabet {0, 1, 2, 3, 4, 5, 6, 7, 8, 9}. If you wanted
to work with the contents of computer memories, you might use the alphabet {0, 1}. If you
wanted to work with text files on a computer, you might use one of the standard machine-
text alphabets, like ASCII or Unicode.
It is a convention in the study of formal languages, which is followed in this book, to
use the first few Latin letters, like a and b, in most example alphabets. This helps you resist
the urge to ascribe meaning to the symbols. When you see a string like 1000, you naturally
think of it as representing the decimal number one thousand, while the string abbb does not
tempt you to make any irrelevant interpretations. Because the symbols in the alphabet are
uninterpreted, our results apply to all alphabets.
Another convention followed in the book is to use the symbol to stand for the
alphabet currently in use. When it is necessary to manipulate more than one alphabet at
once, subscripts are used: 1, 2, and so on.
The empty set {} is a legal alphabet, but not usually an interesting one. (Some people use
the special symbol for the empty set, but this book just uses {}.)

1.2 Strings
The symbols from an alphabet are put together into strings:

A string is a finite sequence of zero or more symbols.

For example, abbb and 010 are strings. In formal languages, unlike most programming
languages, strings are not written with quotation marks around them.
The length of a string is the number of symbols in it. To refer to the length of a string,
bracket the string with vertical lines: |abbb| = 4.
When the symbols in a string are part of a particular alphabet , we say that the string
is “a string over .” In this usage, the word “over” means “built using symbols from.” So, for
example, the set of all strings of length 2 over the alphabet {a, b} is the set of all strings of
length 2 that can be built using the symbols a and b: {aa, bb, ab, ba}.
 1.2 STRINGS 3
The special symbol ε is used to represent the string of length zero: the string of no
symbols. Note here that ε is not a symbol in any alphabet we will use; it simply stands for the
empty string, much as you would write "" in some programming languages. For example,
the set of all strings of length 2 or less over the alphabet {a, b} is {ε, a, b, aa, bb, ab, ba}.
The length of ε is zero, of course: |ε| = 0. Be careful not to confuse the empty set, {}, with
the empty string, ε, and note also that {} {ε}; the set {} contains nothing, while the set {ε}
contains one thing: the empty string.
When describing languages and proving things about them, it is sometimes necessary
to use variables that stand for strings, such as x = abbb. This is a natural concept in
programming languages; in Java one writes String x = "abbb", and it is clear from
the syntax that x is the name of a variable and abbb are the characters making up the string
to which x refers. In the notation used for describing formal languages there is not so much
syntax, so you have to rely more on context and on naming conventions. The convention
followed in the book is to use the last few Latin letters, like x, y, and z, as string variables, not
as symbols in alphabets.
For example, the following definition of concatenation uses string variables.

The concatenation of two strings x and y is the string containing all the symbols
of x in order, followed by all the symbols of y in order.

To refer to the concatenation of two strings, just write them right next to each other. For
example, if x = abc and y = def then the concatenation of x and y is xy = abcdef. For any string
x, we have xε = εx = x. So ε is the identity element for concatenation of strings, just as 0 is
the identity element for addition of natural numbers and just as 1 is for multiplication of
natural numbers.
Speaking of numbers, we’ll denote the set of natural numbers, {0, 1, …}, as N. Any
natural number n N can be used like an exponent on a string, denoting the concatenation
of that string with itself, n times. Thus for any string x,
x0 = ε (that is, zero copies of x)
x1 = x
x2 = xx
x3 = xxx
and, in general,
xn = xx . . . x
n times
When the alphabet does not contain the symbols ( and ), which is almost all the time, you
can use parentheses to group symbols together for exponentiation. For example, (ab)7 denotes
the string containing seven concatenated copies of ab: (ab)7 = ababababababab.
4 CHAPTER ONE: FUNDAMENTALS

1.3 Languages

A language is a set of strings over some fixed alphabet.

A special notation is used to refer to the set of all strings over a given alphabet.

The Kleene closure of an alphabet , written *, is the set of all strings over .

For example, {a}* is the set of all strings of zero or more as: {ε, a, aa, aaa, ...}, and {a, b}* is the set of
all strings of zero or more symbols, each of which is either a or b: {ε, a, b, aa, bb, ab, ba, aaa, ...}. This
allows a more compact way of saying that x is a string over the alphabet ; we can just write x *.
(Here the symbol is used as in standard mathematical notation for sets and means “is an element
of.”)
Except for the special case of = {}, the Kleene closure of any alphabet is an infinite
set. Alphabets are finite sets of symbols, and strings are finite sequences of symbols, but
a language is any set of strings—and most interesting languages are in fact infinite sets of
strings.
Languages are often described using set formers. A set former is written like a set, but
it uses the symbol |, read as “such that,” to add extra constraints or conditions limiting the
elements of the set. For example, {x {a, b}* | |x| 2} is a set former that specifies the set of
all strings x over the alphabet {a, b}, such that the length of x is less than or equal to 2. Thus,
{x {a, b}* | |x| 2} = {ε, a, b, aa, bb, ab, ba}
{xy | x {a, aa} and y {b, bb}} = {ab, abb, aab, aabb}
{x {a, b}* | x contains one a and two bs} = {abb, bab, bba}
{anbn | n 1} = {ab, aabb, aaabbb, aaaabbbb, ...}
That last example shows why set former notation is so useful: it allows you to describe
infinite languages without trying to list the strings in them. Unless otherwise constrained,
exponents in a set former are assumed to range over all of N. For example,
{(ab)n} = {ε, ab, abab, ababab, abababab, ...}
{anbn} = {ε, ab, aabb, aaabbb, aaaabbbb, ...}
The set former is a very expressive tool for defining languages, but it lacks precision. In
one example above we used an English-language constraint: “x contains one a and two bs.”
That’s allowed—but it makes it easy to write set formers that are vague, ambiguous, or self-
contradictory. For that matter, even if we stick to mathematical constraints like “|x| 2” and
“n 1,” we still have the problem of explaining exactly which mathematical constraints are
 EXERCISES 5
permitted and exactly what they mean. We won’t try to do that—we’ll continue to use set
formers throughout this book, but we’ll use them in an informal way, without any formal
definition of what they mean. Our quest in subsequent chapters will be to find other tools
for defining languages, formal tools that are precise and unambiguous.

Exercises
EXERCISE 1
Restate each of the following languages by listing its contents. For example, if the
language is shown as {x {a, b}* | |x| 2}, your answer should be {ε, a, b, aa, bb, ab, ba}.
a. {x {a, b, c}* | |x| 2}
b. {xy | x {a, aa} and y {aa, aaa}}
c. {}*
d. {an | n is less than 20 and divisible by 3}
e. {anbm | n < 2 and m < 3}
EXERCISE 2
List all strings of length 3 or less in each of the following languages:
a. {a}*
b. {a, b}*
c. {anbn}
d. {xy | x {a}* and y {b}*}
e. {anbm | n > m}
EXERCISE 3
Many applications of formal language theory do associate meanings with the strings in a
language. Restate each of the following languages by listing its contents:
a. {x {0, 1}* | x is a binary representation, without unnecessary leading zeros, of a
number less than 10}
b. {x {0, 1, 2, 3, 4, 5, 6, 7, 8, 9}* | x is a decimal representation, without unnecessary
leading zeros, of a prime number less than 20}
c. {x {a, b, ..., z}* | x is a two-letter word in English}
EXERCISE 4
Restate each of the following languages using set former notation.
a. the language of all strings over the alphabet {a, b} that begin with a
b. the language of all even-length strings over the alphabet {a, b, c}
c. the language of strings consisting of zero or more copies of the string ba
d. the language of strings consisting of any number of as followed by the same number
of bs, followed by the same number of cs
6 CHAPTER ONE: FUNDAMENTALS

EXERCISE 5
This exercise illustrates one of the perils of set formers: their use can lead to logical
contradictions. This example is related to Russell’s Paradox, a famous contradiction in
naive set theory discovered by Bertrand Russell in 1901, which led to an important body
of work in mathematics.
A set former is itself a string, so one can define languages that contain set formers. For
example, the set of all set formers that define the empty set would include such strings as
“{}”, “{x | |x| < 0}”, and “{x | x is not equal to x}”. (We put quotation marks around set
formers considered as strings.)
a. Give an example of a string x that is a set former that defines a language that
includes x itself.
b. Give an example of a string y that is a set former that defines a language that does
not include y itself.
c. Consider r = “{set formers x | the language defined by x does not include x}”.
(Your set former y from Part b, above, is an example of a string in the language
defined by r.) Show that assuming r is in the language defined by r leads to a
contradiction. Show that assuming r is not in the language defined by r also leads to
a contradiction.
 2
CHA P TER

Finite Automata

One way to define a language is to construct an automaton—a kind

of abstract computer that takes a string as input and produces a
yes-or-no answer. The language it defines is the set of all strings for
which it says yes.
The simplest kind of automaton is the finite automaton. The more
complicated automata we discuss in later chapters have some kind of
unbounded memory to work with; in effect, they will be able to grow
to whatever size necessary to handle the input string they are given.
But in this chapter, we begin with finite automata, and they have
no such power. A finite automaton has a finite memory that is fixed
in advance. Whether the input string is long or short, complex or
simple, the finite automaton must reach its decision using the same
fixed and finite memory.

7
8 CHAPTER TWO: FINITE AUTOMATA

2.1 Man, Wolf, Goat, and Cabbage

A man is traveling with a wolf, a goat, and a cabbage. He comes to the east bank of a river
and wants to cross. A tiny rowboat is available, but it is so small that the man can take at
most one of his possessions with him in the boat. If he leaves the goat and the cabbage alone
together, the goat will eat the cabbage. If he leaves the wolf and the goat alone together,
the wolf will eat the goat. How can the man get to the west bank without losing any of his
possessions?
In this old riddle, there are really only four moves to choose from at each step:
• Man crosses with wolf.
• Man crosses with goat.
• Man crosses with cabbage.
• Man crosses with nothing.
If we abbreviate these moves as w (wolf ), g (goat), c (cabbage), and n (nothing), then we can
represent a sequence of moves as a string over the alphabet {w, g, c, n}. A seven-move solution
to the problem is represented by the string gnwgcng: first cross with the goat, then cross back
with nothing, then cross with the wolf, then cross back with the goat (leaving the wolf on the
west side), then cross with the cabbage (leaving the goat on the east side again), then cross
back with nothing, and finally cross with the goat.
Each move is a state transition, taking the puzzle from one state to another state. This can
be shown graphically. For instance, this illustration shows that if everything is on the east side
of the river, and the g move is used, the new state has the wolf and cabbage on the east side
and the man and goat on the west side (m represents the man):
g

E: mwgc E: wc
W: W: mg

What happens if the g move is used again? Then the man rows back to the east side with the
goat, and the whole puzzle returns to the initial state. We include this transition as another
arrow:
g

E: mwgc E: wc
W: W: mg

g
 2.1 MAN, WOLF, GOAT, AND CABBAGE 9
In this way, a state transition diagram can show all legal transitions and all reachable
states, omitting those illegal states where something gets eaten:
g n

E: mwgc E: wc E: mwc
W: W: mg W: g
c
g n w
c
w
E: c E: w
W: mwg W: mgc

g g g g

E: mgc E: mgw
W: w W: c
c w
g n c
w
E: E: mg E: g
W: mwgc W: wc W: mwc

g n

An extra arrow indicates the start state, and a double circle indicates the state in which the
diagram accepts that the solution is correct.
By starting in the start state and making one transition on each symbol in the string
gnwgcng, you can check that this is indeed a minimum-length solution to the problem. The
other minimum-length solution is gncgwng. There are longer solutions too, which involve
repeated states. For example, the man can just row back and forth with the goat any number
of times before completing the solution: gggggggggncgwng. In fact, there are infinitely many
possible solutions, though only two minimum-length solutions. The language of strings
representing legal solutions to the problem can be described with reference to the diagram
above: {x {w, g, c, n}* | starting in the start state and following the transitions of x ends up in the
accepting state}.
10 CHAPTER TWO: FINITE AUTOMATA

2.2 Not Getting Stuck

One problem with the previous diagram is that it gets stuck on many nonsolution strings.
What happens, for example, if you follow the previous diagram with the string gnwn? The
last move, the final n in that string, is illegal, since it leaves the goat and the wolf alone on the
same side of the river. However, the diagram does not actually say what to do in that case; it
just leaves no alternative. The string gnwn is not in the language of legal solutions, but the
diagram shows this only by getting stuck when the final n is reached.
It is a useful property for such a diagram to be fully specified—to show exactly one
transition from every state for every symbol in the alphabet, so that it never gets stuck. A
fully specified diagram would show what to do on any string, all the way to the end. To fix
the previous diagram so that it has this property, we need to give it one additional state: a
state representing the fact that an error has been found in the solution.
In the man-wolf-goat-cabbage puzzle, errors are unrecoverable. Once eaten, the goat
cannot be restored, so no string beginning with gnwn can be a solution to the puzzle, no
matter what else follows. The error state needed for the diagram looks like this:

w,g,c,n error

The transition arrow is labeled with all the symbols in the alphabet and simply returns to the
same state. This shows that the state is a trap: once reached, it cannot be departed from. All
the transitions that were unspecified in the original diagram can now be shown as explicit
transitions to this error state. The resulting fully specified diagram is shown on the following
page.
This more elaborate diagram can handle any string over the alphabet {w, g, c, n}. By
starting in the start state and following the transitions on each character in the string, you
end up in a final state. If that final state is the accepting state, the string is a solution; if it is
any other state, the string is not a solution.

2.3 Deterministic Finite Automata

The man-wolf-goat-cabbage diagram is an example of a deterministic finite automaton (or
DFA), a kind of abstract machine for defining languages.
Here is an informal definition of a DFA; the formal definition comes a little later.

(Informal) A deterministic finite automaton or DFA is a diagram with a finite

number of states represented by circles. An arrow points to one of the states,
the unique start state; double circles mark any number of the states as accepting
 2.2 NOT GETTING STUCK 11
g n

E: mwgc E: wc E: mwc
W: W: mg W: g
c
g n w
c
w,c
g w
w,c,n
E: c E: w
c,n W: mwg W: mgc

w,n
w,g,c,n error g g g g
c,n

w,n E: mgc E: mgw

W: w W: c
w,c,n g
w,c c w
g c
n
w
E: E: mg E: g
W: mwgc W: wc W: mwc

g n

states. From every state, for every symbol in the alphabet , there is exactly one
arrow labeled with that symbol going to another state (or back to the same
state).

DFAs define languages, just as our man-wolf-goat-cabbage automaton defined the language
of solutions to the puzzle. Given any string over its alphabet, a DFA can read the string and
follow its state-to-state transitions. At the end of the string it is either in an accepting state
or not. If it is an accepting state, we say that the machine accepts the string; otherwise we
say that it rejects the string. The language defined by a DFA M is just the set of strings in *
accepted by M.
For example, here is a DFA for {xa | x {a, b}*}—that is, the language of strings over the
alphabet {a, b} that end in a:
12 CHAPTER TWO: FINITE AUTOMATA

b a

This diagram meets all the requirements for a DFA: it has a finite set of states with a single
start state, and it has exactly one transition from every state on every symbol in the alphabet
{a, b}. Unlike the man-wolf-goat-cabbage diagram, the states in this DFA are unlabeled. It
does not hurt to label the states, if they have some meaning that you want to call attention
to, or if you need to refer to the states by name for some reason. We could, for example, draw
the DFA above like this:
b a

last symbol last symbol

seen was seen was a
not a

But the labels of states (unlike the labels on the arrows) have no impact on the behavior of
the machine. They are rather like comments in a computer program.
There is one important convention you should know about before attempting the
exercises at the end of the chapter. If a DFA has more than one arrow with the same source
and destination states, like this:
a

then we usually draw it more compactly as a single arrow with a list of symbols as its label,
like this:
 2.4 THE 5-TUPLE 13

a,b

These two forms mean exactly the same thing: they both show a state transition that the
DFA can make on either a or b.

2.4 The 5-Tuple

A diagram of a DFA is reasonably easy for people to look at and understand. Mathematically,
however, a diagram of a DFA is not very useful. To study DFAs with mathematical rigor—to
prove things about them and the languages they accept—it is important to have a more
abstract definition.
A DFA has five key parts, so we describe it mathematically as a 5-tuple.

A DFA M is a 5-tuple M = (Q, , , q0, F), where

Q is the finite set of states
is the alphabet (that is, a finite set of symbols)
(Q Q) is the transition function
q0 Q is the start state
F Q is the set of accepting states

The first part of a DFA is the set Q of states; this corresponds to the set of states drawn
as circles in a DFA diagram. The different states in Q are usually referred to as qi for different
values of i; note that the definition implies that there must be at least one state in Q, namely
the start state q0.
The alphabet is the second part of a DFA. When you draw a diagram for a DFA, the
alphabet is implicit; you can tell what it must be by looking at the labels on the arrows. For
the formal definition of a DFA, however, the alphabet is explicit.
The third part of a DFA is the transition function . The definition says (Q Q).
This is the mathematical way of giving the type of the function, and it says that takes two
inputs, a state from Q and a symbol from , and produces a single output, another state in
Q. If the DFA is in state qi reading symbol a, then (qi, a) is the state to go to next. Thus the
transition function encodes all the information about the arrows in a DFA diagram.
The fourth part of the DFA is the start state q0. There must be exactly one start state.
The fifth and final part of the DFA is F, a subset of Q identifying the accepting states, those
that are double-circled in the DFA diagram. There is nothing in the definition that prevents
F from being {}—in that case there would be no accepting states, so the machine would
reject all strings and the language would be {}. There is also nothing that prevents F from
14 CHAPTER TWO: FINITE AUTOMATA

being equal to Q—in that case all states would be accepting states, so the machine would
accept all strings and the language would be *.
Consider for example this DFA for the language {xa | x {a, b}*}:
b a

q0 q1

Formally, the DFA shown is M = (Q, , , q0, F), where Q = {q0, q1}, = {a, b},
F = {q1}, and the transition function is
(q0, a) = q1
(q0, b) = q0
(q1, a) = q1
(q1, b) = q0
A DFA is a 5-tuple: it is a mathematical structure with five parts given in order.
The names given to those five parts in the definition above—Q , , and so forth—
are conventional, but the machine M could just as well have been defined as
M = ({q0, q1}, {a, b}, , q0, {q1}), without naming all the parts. The important thing
is to specify the five parts in the required order.

2.5 The Language Accepted by a DFA

A DFA’s transition function says how the DFA makes a single move: one state-to-state
transition, reading one symbol from an input string. Intuitively, we understand that the DFA
makes a sequence of these moves, one for each symbol in the string, until the end of the
string is reached. To capture this intuition about making a sequence of moves, an extended
transition function * can be defined.
Where the function takes a single symbol as its second parameter, the * function
takes a string of zero or more symbols. So for any string x * and any state qi Q, *(qi, x)
should be the state the DFA ends up in, starting from qi and making the sequence of
-transitions on the symbols in x. There is a simple recursive definition for this *:
*(q, ε) = q
*(q, xa) = ( *(q, x), a)
The first line in the definition is the base case. It just says that when the input is the empty
string, no transition is made: the final state is the same as the initial state. The second line
 EXERCISES 15
in the definition is the recursive case. It says that for a string xa, where x is any string and a
is the final symbol in the string, you first make all the transitions for x, and then one final
-transition using the final symbol a.
The definition above constructs an extended transition function * from any basic
transition function . Using these extended transition functions, it is possible to say exactly
what it means for a DFA to accept a string:

A string x * is accepted by a DFA M = (Q, , , q0, F ) if and only if

*(q0, x) F.

That is, a string x is accepted if and only if the DFA, started in its start state and taken
through the sequence of transitions on the symbols in x, ends up in one of its accepting
states. We can also define L(M), the language accepted by a DFA M:

For any DFA M = (Q, , , q0, F ), L(M) denotes the language accepted by M,
which is L(M) = {x * | *(q0, x) F }.

A language that can be recognized by a DFA is called a regular language.

A regular language is one that is L(M) for some DFA M.

A direct way to prove that a language is regular is to give a DFA that accepts it. To prove that
a language is not regular is generally much harder, and this is a topic that will be addressed
later in this book.

Exercises
EXERCISE 1
For each of the following strings, say whether it is in the language accepted by this DFA:
a b

a. b
b. ε
c. ab
d. abba
e. ababaaaab
16 CHAPTER TWO: FINITE AUTOMATA

EXERCISE 2
For each of the following strings, say whether it is in the language accepted by this DFA:
0 1

1 0

1 0

a. 0
b. 11
c. 110
d. 1001
e. 101
EXERCISE 3
Say what language is accepted by each of the following DFAs. Do not just describe the
language in English; write your answer as a set, using set former notation if necessary.
For example, if the DFA shown is
b a

then your answer might be {xa | x {a, b}*}.

a.
a,b

a,b
 EXERCISES 17
b.
a,b

a,b

c.
a,b

a,b

d.
a,b

a,b

e.
a a
b

EXERCISE 4
For each of the following languages, draw a DFA accepting that language. The alphabet
for the DFA in all cases should be {a, b}.
a. {a, b}*
b. {a}*
c. {x {a, b}* | the number of as in x is odd}
d. {ax | x {a, b}*}
e. {(ab)n}
18 CHAPTER TWO: FINITE AUTOMATA

EXERCISE 5
Draw the DFA diagram for each of the following DFAs.
a. ({q0, q1}, {a, b}, , q0, {q1}), where the transition function is
(q0, a) = q1
(q0, b) = q1
(q1, a) = q0
(q1, b) = q0
b. ({q0, q1, q2}, {0, 1}, , q0, {q0}), where the transition function is
(q0, 0) = q0
(q0, 1) = q1
(q1, 0) = q2
(q1, 1) = q0
(q2, 0) = q1
(q2, 1) = q2
c. ({q0, q1, q2}, {a, b}, , q0, {q2}), where the transition function is
(q0, a) = q1
(q0, b) = q0
(q1, a) = q1
(q1, b) = q2
(q2, a) = q2
(q2, b) = q2

EXERCISE 6
State each of the following DFAs formally, as a 5-tuple.
a.
a,b

a,b

b.
a a

b
 EXERCISES 19
c.
0,1

0,1

EXERCISE 7
For each machine M in the previous exercise, state what the language L(M) is, using set
formers.
EXERCISE 8
Evaluate each of these expressions for the following DFA:
0 1
1 0

q0 q1 q2

1 0

a. (q2, 0)
b. *(q0, 010)
c. ( *(q1, 010), 1)
d. *(q2, ε)
e. *(q2, 1101)
f. *(q2, 110111)
EXERCISE 9
Draw the diagram for a DFA for each of the following languages.
a. {x {a, b}* | x contains at least 3 as}
b. {x {a, b}* | x contains at least 3 consecutive as}
20 CHAPTER TWO: FINITE AUTOMATA

c. {x {a, b}* | the substring bab occurs somewhere in x}

d. {x {a, b}* | x has no two consecutive as}
e. {axb | x {a, b}*}
EXERCISE 10
(This exercise requires some knowledge of fundamental graph algorithms.) The following
questions ask for algorithms that operate on DFAs. You should assume that the input is a
DFA M represented as a 5-tuple.
a. Give an algorithm that determines whether L(M ) = {}.
b. Give an algorithm that finds a minimum-length string in L(M ). Assume that
L(M) {}.
c. Give an algorithm that determines whether L(M ) is finite.
 3
CHA P TER

Closure Properties for

Regular Languages
Once we have defined some languages formally, we can consider
combinations and modifications of those languages: unions,
intersections, complements, and so on. Such combinations and
modifications raise important questions. For example, is the
intersection of two regular languages also regular—capable of being
recognized directly by some DFA?

21
22 CHAPTER THREE: CLOSURE PROPERTIES FOR REGULAR LANGUAGES

3.1 Closed Under Complement

Consider the language L = {0x | x {0, 1}*}. We can show that this is a regular language by
constructing a DFA to recognize it:

q1 0,1
0

q0

1
q2 0,1

L is the language of strings over the alphabet {0, 1} that start with a 0. The complement of L,
written as L , is the language of strings over the alphabet {0, 1} that do not start with a 0. We
can show that this is a regular language too, by constructing a DFA to recognize it:

q1 0,1
0
q0

1
q2 0,1

Comparing the two DFAs, you can see that the machine for L is the same as the machine for
L, with one difference: the accepting states and the nonaccepting states have been exchanged.
Whenever the first machine accepts, the second machine rejects; whenever the first machine
rejects, the second machine accepts.
This trick for constructing a complemented DFA can be generalized to handle any
complemented language. First, here is a formal definition of the complement of a language:

If L is a language over some alphabet , the complement of L is

L = {x * | x L}.

Notice that the complement of a language is always taken with respect to the underlying
alphabet. This guarantees that the complement of a complement is the original language:
L = L.
 3.1 CLOSED UNDER COMPLEMENT 23
Now we can easily prove a useful property of the regular languages:
Theorem 3.1: If L is any regular language, L is also a regular language.
Proof : Let L be any regular language. By definition there must be some DFA
M = (Q, , , q0, F ) with L(M ) = L. Define a new DFA
M' = (Q, , , q0, Q - F ) using the same set of states, alphabet, transition
function, and start state as M, but using Q–F as its set of accepting states. Now
for any x *, M' accepts x if and only if M rejects x. So L(M' ) = L . Since M'
is a DFA that accepts L , it follows that L is regular.
This theorem shows that if you take any regular language and complement it, you
still get a regular language. In other words, you cannot leave the class of regular languages
by using the complement operation. We say that the regular languages are closed under
complement. One useful thing about such closure properties is that they give you shortcuts
for proving that a language belongs to a particular class. If you need to know whether L
is a regular language and you already know that L is, you can immediately conclude that
L is (without actually constructing a DFA for it) simply by invoking the closed-under-
complement property of the regular languages.

3.2 Closed Under Intersection

We can define the intersection of two languages in a straightforward way:

If L1 and L2 are languages, the intersection of L1 and L2 is L1 L2 = {x | x L1

and x L2}.

This definition works fine no matter what the underlying alphabets are like. If L1 and L2 are
over the same alphabet , which is the most common case, then L1 L2 is a language over
the same alphabet—it may not actually use all the symbols of , but it certainly does not use
more. Similarly, if L1 and L2 are languages over different alphabets 1 and 2, then L1 L2 is
a language over 1 2
. We will generally assume that L1 and L2 are over the same alphabet.
The regular languages are closed under intersection. As in the previous section, we can
prove this by construction: given any two DFAs, there is a mechanical way to combine them
into a third DFA that accepts the intersection of the two original languages. For example,
consider the languages L1 = {0x | x {0, 1}*} and L2 = {x0 | x {0, 1}*}: the language of
strings that start with a 0 and the language of strings that end with a 0. Here are DFAs M1
and M2 for the two languages:
24 CHAPTER THREE: CLOSURE PROPERTIES FOR REGULAR LANGUAGES

0,1 1 0

0
0
q0 q1 r0 r1

1
M1 M2

q2

0,1

We will define a DFA M3 that keeps track of where M1 is and where M2 is after each
symbol. Initially, M1 is in state q0 and M2 is in state r0; our new machine will reflect this by
starting in a state named (q0, r0 ):

0
q0 ,r0
1

Now when M1 is in state q0 and reads a 0, it goes to state q1, and when M2 is in state
r0 and reads a 0, it goes to state r1, so when M3 is in state (q0, r0 ) and reads a 0, it goes to a
new state named (q1, r1). That way it keeps track of what both M1 and M2 are doing. We can
define the transition on 1 similarly, adding this to the construction of M3:
0
q1 ,r1
0 1
q0 ,r0
0
1
q2 ,r0
1

The process of constructing M3 is not infinite, since there are only six possible states of the
form (qi, rj ). In fact, the construction is already about to repeat some states: for example,
from (q1, r1) on a 0 it will return to (q1, r1). The fully constructed DFA M3 looks like this:
 3.2 CLOSED UNDER INTERSECTION 25
0

1
q1 ,r1 q1 ,r0 1
0
0
q0,r0
0
1
q2 ,r0 q2 ,r1 0

Notice that we did not need a state for the pair (q0, r1). That is because this combination
of states is unreachable: it is not possible for M1 to be in the state q0 while M2 is in the state
r1. The constructed machine accepts in the state (q1, r1), which is exactly where both M1 and
M2 accept. Since the new machine in effect simulates both the original machines and accepts
only when both accept, we can conclude that the language accepted is L1 L2.
The construction uses an operation on sets that you may not have seen before. Given
any two sets, you can form the set of all pairs of elements—pairs in which the first is from
the first set and the second is from the second set. In mathematics this is called the Cartesian
product of two sets:

If S1 and S2 are sets, the Cartesian product of S1 and S2 is

S1 S2 = {(e1, e2) | e1 S1 and e2 S2}.

To express the general form of the construction combining two DFAs, we use the Cartesian
product of the two sets of states. For this reason, the construction used in the following proof
is called the product construction.
Theorem 3.2: If L1 and L2 are any regular languages, L1 L2 is also a regular
language.
Proof: Let L1 and L2 be any two regular languages. Since they are regular, there
must be some DFAs M1 = (Q, , 1, q0, F1) with L(M1) = L1 and
M2 = (R, , 2, r0, F2) with L(M2) = L2. Construct a new DFA
M3 = (Q R, , , (q0, r0), F1 F2), where is defined so that for all q Q,
r R, and a , we have ((q, r), a) = ( 1(q, a), 2(r, a)). Since this DFA
simulates both M1 and M2 and accepts if and only if both accept, we can
conclude that L(M3) = L1 L2. It follows that L1 L2 is a regular language.
Notice that the machine constructed for the proof includes all the states in Q R, which may
include some unreachable states. In the case of our earlier example, this is a six-state DFA
instead of the five-state DFA. The extra state is not reachable by any path from the start state,
so it has no effect on the language accepted.
26 CHAPTER THREE: CLOSURE PROPERTIES FOR REGULAR LANGUAGES

3.3 Closed Under Union

The union of two languages is defined in the usual way:

If L1 and L2 are languages, the union of L1 and L2 is

L1 L2 = {x | x L1 or x L2, or both}.

As with intersection, we will generally assume that L1 and L2 are over the same alphabet, but
the definition works fine even if they are not.
The regular languages are closed under union. There is a simple proof of this using
DeMorgan’s laws. There is also a direct proof using the subset construction, but choosing a
different set of accepting states.
Theorem 3.3: If L1 and L2 are any regular languages, L1 L2 is also a regular
language.
Proof 1: Using DeMorgan’s laws, L1 ∪ L2 = L1 ∩ L2 . This defines union in terms
of intersection and complement, for which the regular languages have already
been shown to be closed.
Proof 2: Use the product construction as in the proof of closure under
intersection, but for the accepting states use (F1 R) (Q F2). Now the
constructed DFA simulates both M1 and M2 and accepts if and only if either
or both accept. We conclude that L(M3) = L1 L2 and that L1 L2 is a regular
language.
The product construction for union is almost the same as the product construction for
intersection. The only difference is in the choice of accepting states. For intersection we make
the new DFA accept when both the original DFAs accept, while for union we make the new
DFA accept when either or both of the original DFAs accept. For example, in Section 3.2 we
produced a DFA for the intersection of the two languages {0x | x {0, 1}*} and {x0 | x
{0, 1}*}. For the union of those same two languages, we could use the same set of states and
the same transition function:
0

1
q1 ,r1 q1 ,r0 1
0
0
q0,r0
0
1
q2 ,r0 q2 ,r1 0

1
 3.4 DFA PROOFS USING INDUCTION 27
The only difference is that this DFA accepts in more states. It accepts if either or both of the
original DFAs accept.

3.4 DFA Proofs Using Induction

Rigorous proofs about DFAs often require a special technique: mathematical induction.
It is a technique that is useful in many areas of mathematics, including the foundations of
computer science. Proofs that use induction are related in an interesting way to programs
that use recursion.
Although it is not a primary aim of this book to teach proof techniques, this is too good
an opportunity to miss. Mathematical induction and DFAs are a perfect match; you can
learn a lot about DFAs by doing inductive proofs on them, and you can learn a lot about
proof technique by proving things about DFAs.
If you read the proof of Theorem 3.2 closely, you may have noticed a place where
the proof made an important step rather informally. The theorem stated that the regular
languages are closed for intersection, and the proof used the product construction. Here is
that proof again:

Proof: Let L1 and L2 be any two regular languages. Since they are regular, there
must be some DFAs M1 = (Q, , 1, q0, F1) with L(M1) = L1 and
M2 = (R, , 2, r0, F2) with L(M2) = L2. Construct a new DFA
M3 = (Q R, , , (q0, r0), F1 F2), where is defined so that for all
q Q, r R, and a , we have ((q, r), a) = ( 1(q, a), 2(r, a)). Since this
DFA simulates both M1 and M2 and accepts if and only if both accept, we can
conclude that L(M3 ) = L1 L2. It follows that L1 L2 is a regular language.
What exactly does it mean that the new DFA “simulates both M1 and M2”? Formally
speaking, it means this:
Lemma 3.1: In the product construction, for all x *, *((q0, r0), x) =
( 1*(q0, x), 2*(r0, x)).
That is, for any input string x, the new machine finds exactly the pair of final states that
the two original machines reach. Our proof of Theorem 3.2 assumed Lemma 3.1 was true
without proving it.
Does Lemma 3.1 require any proof? By construction, we know that for all q Q,
r R, and a , we have ((q, r), a) = ( 1(q, a), 2(r, a)). This is something like Lemma 3.1,
but only for a single symbol a, not for a whole string x. Whether the leap from the property
to the corresponding * property is obvious enough to be passed over without comment
is largely a matter of taste. Whenever you prove something, you have to exercise judgment
about how much detail to give. Not enough detail makes a proof unconvincing, while too
much makes it tiresome (and just as unconvincing). But even if you think Lemma 3.1 is
obvious enough not to require any proof, let’s prove it anyway.
28 CHAPTER THREE: CLOSURE PROPERTIES FOR REGULAR LANGUAGES

For any fixed length of x, it is easy to prove Lemma 3.1. For example, when |x| = 0
*((q0, r0), x)
= *((q0, r0), ε) (since |x| = 0)
= (q0, r0) (by the definition of *)
= ( 1*(q0, ε), 2*(r0, ε)) (by the definitions of 1* and 2
*)
= ( 1*(q0, x), 2*(r0, x)) (since |x| = 0)
And when |x| = 1
*((q0, r0), x)
= *((q0, r0), ya) (for some symbol a and string y)
= ( *((q0, r0), y), a) (by the definition of *)
= (( 1*(q0, y), 2*(r0, y)), a) (using Lemma 3.1 for |y| = 0)
= ( 1( 1*(q0, y), a), 2( 2*(r0, y), a)) (by the construction of )
= ( 1*(q0, ya), 2*(r0, ya)) (by the definitions of 1* and 2*)
= ( 1*(q0, x), 2*(r0, x)) (since x = ya)
Notice that we used the fact that we already proved Lemma 3.1 for strings of length 0. When
|x| = 2
*((q0, r0), x)
= *((q0, r0), ya) (for some symbol a and string y)
= ( *((q0, r0), y), a) (by the definition of *)
= (( 1*(q0, y), 2*(r0, y)), a) (using Lemma 3.1 for |y| = 1)
= ( 1( 1*(q0, y), a), 2( 2*(r0, y), a)) (by the construction of )
= ( 1*(q0, ya), 2*(r0, ya)) (by the definitions of 1* and 2*)
= ( 1*(q0, x), 2*(r0, x)) (since x = ya)
Here, we used the fact that we already proved Lemma 3.1 for strings of length 1. As you can
see, the proof for |x| = 2 is almost the same as the proof for |x| = 1. In general, you can easily
continue with proofs for |x| = 3, 4, 5, 6, and so on, each one using the fact that the lemma
was already proved for shorter strings. Unfortunately, this is a proof process that never ends.
What we need is a finite proof that Lemma 3.1 holds for all the infinitely many different
lengths of x.
To prove Lemma 3.1 for all string lengths we use an inductive proof.
Proof: By induction on |x|.
Base case: When |x| = 0, we have
*((q0, r0), x)
= *((q0, r0), ε) (since |x| = 0)
= (q0, r0) (by the definition of *)
 3.4 DFA PROOFS USING INDUCTION 29
= ( 1*(q0, ε), 2
*(r0, ε)) (by the definitions of 1
* and *)
2
= ( 1*(q0, x), 2
*(r 0
, x)) (since |x| = 0)

Inductive case: When |x| > 0, we have

*((q0, r0), x)
= *((q0, r0), ya) (for some symbol a and string y)
= ( *((q0, r0), y), a) (by the definition of *)
= (( 1*(q0, y), 2*(r0, y)), a) (by the inductive hypothesis, since |y| < |x|)
= ( 1( 1*(q0, y), a), 2( 2*(r0, y), a)) (by the construction of )
= ( 1*(q0, ya), 2*(r0, ya)) (by the definitions of 1* and 2*)
= ( 1*(q0, x), 2*(r0, x)) (since x = ya)
This proof demonstrates that Lemma 3.1 holds when |x| = 0 and then shows that whenever
it holds for some length |x|, it also holds for |x| + 1. It follows that the lemma holds for all
string lengths.
Every inductive proof has one or more base cases and one or more inductive cases using
the inductive hypothesis. In the proof of Lemma 3.1, the base case is exactly the same as the
previous proof for the special case |x| = 0. The inductive case then shows how to prove the
lemma for any |x| > 0. It is a generalized form of the previous proofs for the special cases
|x| = 1 and |x| = 2. Instead of referring to the specific length of the shorter string y for which
the lemma has already been proved, it refers to the inductive hypothesis: the assumption that
the lemma is already known to hold for any string y with |y| < |x|.
Proof with induction is a lot like programming with recursion. The proof above can be
thought of as a program for making a proof for x of any size:

void proveit(int n) {
if (n==0) {
base case: prove for empty string
}
else {
proveit(n-1);
prove for strings of length n, assuming n-1 case has been proved
}
}

Not all inductive proofs follow this pattern exactly, of course. There are as many different
ways to prove something with induction as there are to program something with recursion.
But DFA-related proofs often do follow this pattern. Many such proofs use induction on the
length of the string, with the empty string as the base case.
30 CHAPTER THREE: CLOSURE PROPERTIES FOR REGULAR LANGUAGES

3.5 A Mystery DFA

Consider this DFA:
0 1

1 0

0 1 2
1 0

What language does it accept? By experimenting with it, you can see that it rejects the
strings 1, 10, 100, 101, 111, and 1000, while it accepts 0, 11, 110, and 1001. Do you see
a pattern there? Can you give an intuitive characterization of the language, before reading
further?
First, here is a lemma that summarizes the transition function :
Lemma 3.2.1: For all states i Q and symbols c ,
(i, c) = (2i + c) mod 3.
Proof: By enumeration. Notice that we have named the states with the numbers
0, 1 and 2, which lets us do direct arithmetic on them. We have
(0, 0) = 0 = (2 0 + 0) mod 3
(0, 1) = 1 = (2 0 + 1) mod 3
(1, 0) = 2 = (2 1 + 0) mod 3
(1, 1) = 0 = (2 1 + 1) mod 3
(2, 0) = 1 = (2 2 + 0) mod 3
(2, 1) = 2 = (2 2 + 1) mod 3
Now for any string x over the alphabet {0, 1}, define val(x) as the natural number for
which x is a binary representation. (For completeness, define val(ε) = 0.) So, for example,
val(11) = 3, val(111) = 7, and val(000) = val(0) = val(ε) = 0. Using this, we can describe the
language accepted by the mystery DFA: L(M ) = {x | val(x) mod 3 = 0}. In other words, it is
the language of strings that are binary representations of numbers that are divisible by three.
Let’s prove by induction that L(M ) = {x | val(x) mod 3 = 0}. This illustrates a common
pitfall for inductive proofs and a technique for avoiding it. If you try to prove directly the
hypothesis L(M) = {x | val(x) mod 3 = 0}, you get into trouble:
Lemma 3.2.2 (weak): L(M ) = {x | val(x) mod 3 = 0}.
Proof: By induction on |x|.
Base case: When |x| = 0, we have
 3.5 A MYSTERY DFA 31
*(0, x)
= *(0, ε) (since |x| = 0)
=0 (by definition of *)
So in this case x L(M) and val(x) mod 3 = 0.
Inductive case: When |x| > 0, we have
*(0, x)
= *(0, yc) (for some symbol c and string y)
= ( *(0, y), c) (by definition of *)
= ???
Here the proof would falter, because the inductive hypothesis we’re using is not strong
enough to make progress with. It tells us that *(0, y) = 0 if and only if val(x) mod 3 = 0, but
it does not tell us what *(0, y) is when val(x) mod 3 0. Without knowing that, we can’t
make progress from here.
To make a successful proof of Lemma 3.2.2, we actually need to prove something even
stronger. We will prove that *(0, x) = val(x) mod 3. This implies the weak version of Lemma
3.2.2, because the state of 0 is the only accepting state. But it is stronger than that, because
it tells you exactly what state the DFA will end up in after reading any string in {0,1}*. Using
that trick, let’s try the proof again:
Lemma 3.2.2 (strong): *(0, x) = val(x) mod 3.
Proof: By induction on |x|.
Base case: When |x| = 0, we have
*(0, x)
= *(0, ε) (since |x| = 0)
=0 (by definition of *)
= val(x) mod 3 (since val(x) mod 3 = val(ε) mod 3 = 0)
Inductive case: When |x| > 0, we have
*(0, x)
= *(0, yc) (for some symbol c and string y)
= ( *(0, y), c) (by definition of *)
= (val(y) mod 3, c) (using the inductive hypothesis)
= (2(val(y) mod 3) + c) mod 3 (by Lemma 3.2.1)
= 2(val(y) + c) mod 3 (using modulo arithmetic)
= val(yc) mod 3 (using binary arithmetic, val(yc) = 2(val(y)) + c)
= val(x) mod 3 (since x = yc)
32 CHAPTER THREE: CLOSURE PROPERTIES FOR REGULAR LANGUAGES

We conclude by induction that *(0, x) = val(x) mod 3. Since the only

accepting state is 0, the language of strings accepted by the DFA is
L(M) = {x | val(x) mod 3 = 0}.

This technique is something you will often need to use with inductive proofs; to make
the induction go through, you will often find that you need to prove something stronger,
something more detailed, than your original hypothesis. This is a little counterintuitive at
first. It seems like it should be more difficult to prove something stronger and easier to prove
something weaker. But with induction, you get to use the thing you are trying to prove as the
inductive hypothesis, so proving something stronger can give you more to work with.

Exercises
EXERCISE 1
Draw DFAs for the following languages:
a. {0x0 | x {0, 1}*}
b. the complement of the language above
c. {xabay | x {a,b}* and y {a,b}*}
d. the complement of the language above
e. {waxayaz | w, x, y, and z are all in {a, b}*}
f. the complement of the language above
EXERCISE 2
The DFA constructed for the example in Section 3.2 accepts the language of strings over
the alphabet {0, 1} that start and end with a 0. It is not, however, the smallest DFA to do
so. Draw a DFA for the same language, using as few states as possible.
EXERCISE 3
The DFA constructed for the example in Section 3.2 has five states, but the full DFA
given by the product construction has six. Draw the full six-state DFA.
EXERCISE 4
Consider the following three DFAs and the languages they define:
b a,b

a
q r

L1 = {xay | x {a, b}* and y {a, b}*}

 EXERCISES 33
a a a,b

b b
s t u

L2 = {xbybz | x, y, and z are all in {a, b}*}

b a a,b

a b
v w x

L3 = {xaby | x {a, b}* and y {a, b}*}

Apply the product construction to create DFAs for each of the following languages.
Label each state with the pair of states to which it corresponds. You do not need to show
unreachable states.
a. L1 L2
b. L L2
c. L1 L3
d. L2 L3
EXERCISE 5
If L1 and L2 are languages, the difference of L1 and L2 is L1 - L2 = {x | x L1 and x L2}.
a. Using the product construction to combine two DFAs M1 = (Q, , 1, q0, F1) and
M2 = (R, , 2, r0, F2), what should the set of accepting states be so that the resulting
machine accepts L(M1) - L(M2)?
b. Prove that the regular sets are closed for set difference, without using the product
construction.
34 CHAPTER THREE: CLOSURE PROPERTIES FOR REGULAR LANGUAGES

EXERCISE 6
Consider this DFA M:

a,b

0 1
a,b

Prove by induction that L(M) = {x {a, b}* | |x| mod 2 = 1}.

EXERCISE 7
Prove formally that for any DFA M = (Q, , , q0, F), any strings x * and y * and
any state q Q, *(q, xy) = *( *(q, x), y). Use induction on |y|.
EXERCISE 8
Draw a DFA that accepts the language of strings that are binary representations of
numbers that are divisible by 2.
EXERCISE 9
Draw a DFA that accepts the language of strings that are binary representations of
numbers that are divisible by 4.
EXERCISE 10
Draw a DFA that accepts the language of strings that are binary representations of
numbers that are divisible by 5.
EXERCISE 11
Prove that your DFA from Exercise 10 accepts the language it is supposed to accept. Use
induction on the length of the input string.
EXERCISE 12
For any n N, n 0, define the DFA Mn = ({0, 1, …, n - 1}, {0, 1}, , 0, {0}), where
(i, c) = (2i + c) mod n. Prove that L(Mn) = {x | val(x) mod n = 0}.
 4
CHA P TER

Deterministic-
Finite-Automata
Applications
We have seen how DFAs can be used to define formal languages.
In addition to this formal use, DFAs have practical applications.
DFA-based pieces of code lie at the heart of many commonly used
computer programs.

35
36 CHAPTER FOUR: DETERMINISTIC-FINITE-AUTOMATA APPLICATIONS

4.1 An Overview of DFA Applications

Deterministic finite automata have many practical applications:
• Almost all compilers and other language-processing systems use DFA-like code to
divide an input program into tokens like identifiers, constants, and keywords and to
remove comments and white space.
• For many applications that accept typed commands, the command language is
quite complex, almost like a little programming language. Such applications use
DFA-like code to process the input command.
• Text processors often use DFA-like code to search a text file for strings that match
a given pattern. This includes most versions of Unix tools like awk, egrep, and
Procmail, along with a number of platform-independent systems such as MySQL.
• Speech-processing and other signal-processing systems often use a DFA-like
technique to transform an incoming signal.
• Controllers use DFA-like techniques to track the current state of a wide variety
of finite-state systems, from industrial processes to video games. They can be
implemented in hardware or in software.
Sometimes, the DFA-like code in such applications is generated automatically by special
tools such as lex, which we discuss more later. In this chapter we concentrate on examples of
DFA-like code constructed by hand.

4.2 A DFA-based Text Filter in Java

The last chapter introduced a DFA that accepts strings that are binary representations of
numbers that are divisible by three. Here is that DFA again:
0 1

1 0

0 1 2
1 0

Now let’s see how this DFA can be implemented in the Java programming language.
The first thing to deal with is the input alphabet. The DFA above uses the alphabet
{0, 1}, which is the alphabet of interest for this problem. But the program will work with a
typed input string, so we do not have the luxury of restricting the alphabet in this way. The
program should accept "011" (a representation for the number 3) and reject "101" (a
representation for the number 5), but it must also properly reject strings like "01i" and
"fred". The alphabet for the Java implementation must be the whole set of characters that
can occur in a Java string—that is, the whole set of values making up the Java char type.
The DFA we actually implement will have four states, like this:
 4.2 A DFA-BASED TEXT FILTER IN JAVA 37
∑

∑ - 0,1 ∑ - 0,1

0 1
∑ - 0,1
1 0

0 1 2
1 0

An object of the Mod3 class represents such a DFA. A Mod3 object has a current state,
which is encoded using the integers 0 through 3. The class definition begins like this:

/**
* A deterministic finite-state automaton that
* recognizes strings that are binary representations
* of natural numbers that are divisible
* by 3. Leading zeros are permitted, and the
* empty string is taken as a representation for 0
* (along with "0", "00", and so on).
*/
public class Mod3 {

/*
* Constants q0 through q3 represent states, and
* a private int holds the current state code.
*/
private static final int q0 = 0;
private static final int q1 = 1;
private static final int q2 = 2;
private static final int q3 = 3;

private int state;

The int variables q0, q1, q2, and q3 are private (visible only in this class), static
(shared by all objects of this class), and final (not permitted to change after initialization).
38 CHAPTER FOUR: DETERMINISTIC-FINITE-AUTOMATA APPLICATIONS

In effect, they are named constants. The int field named state will hold the current state
of each Mod3 object.
The next part of this class definition is the transition function, a method named delta.

/**
* The transition function.
* @param s state code (an int)
* @param c char to make a transition on
* @return the next state code
*/
static private int delta(int s, char c) {
switch (s) {
case q0: switch (c) {
case '0': return q0;
case '1': return q1;
default: return q3;
}
case q1: switch (c) {
case '0': return q2;
case '1': return q0;
default: return q3;
}
case q2: switch (c) {
case '0': return q1;
case '1': return q2;
default: return q3;
}
default: return q3;
}
}

The delta method is declared private (cannot be called from outside the class) and
static (is not given an object of the class to work on). It is thus a true function, having
no side effects and returning a value that depends only on its parameters. It computes the
transition function shown in the previous DFA diagram.
Next, the class defines methods that modify the state field of a Mod3 object. The first
resets it to the start state; the second applies the transitions required for a given input string.
 4.2 A DFA-BASED TEXT FILTER IN JAVA 39
/**
* Reset the current state to the start state.
*/
public void reset() {
state = q0;
}

/**
* Make one transition on each char in the given
* string.
* @param in the String to use
*/
public void process(String in) {
for (int i = 0; i < in.length(); i++) {
char c = in.charAt(i);
state = delta(state, c);
}
}

The process method makes one transition on each character in the input string. Note
that process handles the empty string correctly—by doing nothing.
The only other thing required is a method to test whether, after processing an input
string, the DFA has ended in an accepting state:

/**
* Test whether the DFA accepted the string.
* @return true if the final state was accepting
*/
public boolean accepted() {
return state==q0;
}
}

That is the end of the class definition. To test whether a string s is in the language defined by
this DFA, we would write something like
40 CHAPTER FOUR: DETERMINISTIC-FINITE-AUTOMATA APPLICATIONS

Mod3 m = new Mod3();

m.reset();
m.process(s);
if (m.accepted()) ...

To demonstrate the Mod3 class we will use it in a Java application. This is a simple filter
program. It reads lines of text from the standard input, filters out those that are not binary
representations of numbers that are divisible by three, and echoes the others to the standard
output.

import java.io.*;

/**
* A Java application to demonstrate the Mod3 class by
* using it to filter the standard input stream. Those
* lines that are accepted by Mod3 are echoed to the
* standard output.
*/
public class Mod3Filter {
public static void main(String[] args)
throws IOException {

Mod3 m = new Mod3(); // The DFA

BufferedReader in = // Standard input
new BufferedReader(new InputStreamReader(System.in));

// Read and echo lines until EOF

String s = in.readLine();
while (s!=null) {
m.reset();
m.process(s);
if (m.accepted()) System.out.println(s);
s = in.readLine();
}
}
}

To test this program, we can create a file named numbers containing the numbers zero
through ten in binary:
 4.3 A TABLE-DRIVEN ALTERNATIVES 41
0
1
10
11
100
101
110
111
1000
1001
1010

After compiling Mod3Filter (and Mod3), we can use it on the file numbers
to filter out all the numbers not divisible by 3. On a Unix system, the command
java Mod3Filter < numbers produces this output:

0
11
110
1001

4.3 Table-driven Alternatives

The Mod3 class implements the transition function in the obvious way, as a Java method
delta. The delta method branches on the current state, and then within each alternative
branches again on the current character.
An alternative implementation is to encode the transition function as a table, a two-
dimensional array delta, indexed using the current state and symbol. Instead of calling the
function delta(s,c), passing it the state s and character c as parameters, we perform
an array reference delta[s,c]. Array references are generally faster than function calls,
so this should make the DFA run faster. For example, the process method could look
something like this:

static void process(String in) {

for (int i = 0; i < in.length(); i++) {
char c = in.charAt(i);
state = delta[state, c];
}
}
42 CHAPTER FOUR: DETERMINISTIC-FINITE-AUTOMATA APPLICATIONS

Of course, the array delta must first be initialized with the appropriate transitions, so that
delta[q0,'0'] is q0, delta[q0,'1'] is q1, and so on. To avoid the possibility of
the reference delta[state,c] being out of bounds, delta will have to be initialized
with a very large array. The program uses only 4 values for state, but there are 65,536
possible values for c! That is because Java uses Unicode, a 16-bit character encoding, for
the char type. Depending on the source of the input string, we may be able to restrict this
considerably—we may know, for example, that the characters are 7-bit ASCII (with values 0
through 127). Even so, we will have to initialize the array so that delta[state,c] is q3
for every value of c other than '0' and '1'.
Instead of using a very large array, we could use a small array but handle the
exception that occurs when the array reference is out of bounds. In Java this is the
ArrayIndexOutOfBoundsException; the process method could catch this
exception and use the state q3 as the next state whenever it occurs. The definition of the
delta array and the process method would then be the following:

/*
* The transition function represented as an array.
* The next state from current state s and character c
* is at delta[s,c-'0'].
*/
static private int[][] delta =
{{q0,q1},{q2,q0},{q1,q2},{q3,q3}};

/**
* Make one transition on each char in the given
* string.
* @param in the String to use
*/
public void process(String in) {
for (int i = 0; i < in.length(); i++) {
char c = in.charAt(i);
try {
state = delta[state][c-'0'];
}
catch (ArrayIndexOutOfBoundsException ex) {
state = q3;
}
}
}
 EXERCISES 43
This is a reasonable way to solve the problem by hand. Automatically generated systems
usually use the full table with an element for every possible input character. One reason for
this is that when the full array is used, process need contain no reference to individual
states or characters. This way, any DFA can be implemented using the same process code,
just by substituting a different transition table.
Incidentally, the transition table is usually stored in a more compressed form than we
have shown. Our implementation used a full 32-bit int for each entry in the table, which
is quite wasteful. It would be relatively easy to implement this using one byte per entry,
and it would be possible to use even less, since we really need only two bits to represent
our four possible states. The degree of compression chosen is, as always, a trade-off: heavily
compressed representations take less space, but using them slows down each table access and
thus slows down each DFA transition.

Exercises
EXERCISE 1
Reimplement Mod3 using the full transition table. Assume that the characters in the
input string are in the range 0 through 127.
EXERCISE 2
Using the DFA-based approach, write a Java class Mod4Filter that reads lines of text
from the standard input, filters out those that are not binary representations of numbers
that are divisible by four, and echoes the others to the standard output.
EXERCISE 3
Using the DFA-based approach, write a Java class ManWolf that takes a string from the
command line and reports whether or not it represents a solution to the man-wolf-goat-
cabbage problem of Chapter 2. For example, it should have this behavior:

> java ManWolf gncgwng

That is a solution.
> java ManWolf ggggggggg
That is not a solution.
 5
CHA P TER

Nondeterministic
Finite Automata
A DFA has exactly one transition from every state on every symbol
in the alphabet. By relaxing this requirement we get a related but
more flexible kind of automaton: the nondeterministic finite
automaton (NFA).
NFAs are a bit harder to think about than DFAs because they do
not appear to define simple computational processes. They may seem
at first to be unnatural, like puzzles invented by professors for the
torment of students. But have patience! NFAs and other kinds of
nondeterministic automata arise naturally in many ways, as you
will see later in this book, and they too have a variety of practical
applications.

45
46 CHAPTER FIVE: NONDETERMINISTIC FINITE AUTOMATA

5.1 Relaxing a Requirement

Consider the following state machine:

a
q0 q1

a,b

This is not a DFA, since it violates the rule that there must be exactly one transition from
every state on every symbol in the alphabet. There is no transition from the q1 state on either
a or b, and there is more than one transition from the q0 state on an a. This is an example of
a nondeterministic finite automaton (NFA).
How can you tell whether such a machine accepts a given string? Consider the string aa.
Following the arrows, there are three possible sequences of moves:
1. from q0 back to q0 on the first a, and back to q0 again on the second;
2. from q0 back to q0 on the first a, then to q1 on the second; and
3. from q0 to q1 on the first a, then getting stuck with no legal move on the second.
Only the second of these is an accepting sequence of moves. But the convention is that if
there is any way for an NFA to accept a string, that string is taken to be part of the language
defined by the NFA. The string aa is in the language of the NFA above because there is a
sequence of legal moves that reaches the end of aa in an accepting state. The fact that there
are also sequences of legal moves that do not succeed is immaterial.
The language defined by the NFA above is {xa | x {a, b}*}—that is, the language of
strings over the alphabet {a, b} that end in a. The machine can accept any such string by
staying in the q0 state until it reaches the a at the end, then making the transition to q1 as its
final move.
A DFA can be thought of as a special kind of NFA—an NFA that happens to specify
exactly one transition from every state on every symbol. In this sense every DFA is an NFA,
but not every NFA is a DFA. So in some ways the name nondeterministic finite automaton
is misleading. An NFA might more properly be called a not-necessarily-deterministic finite
automaton.
The extra flexibility afforded by NFAs can make them much easier to construct. For
example, consider the language {x {0, 1}* | the next-to-last symbol in x is 1}. The smallest
DFA for the language is this one:
 5.2 SPONTANEOUS TRANSITIONS 47
0

1 1
0
0

The connection between the language and this DFA is not obvious. (With some effort you
should be able to convince yourself that the four states correspond to the four possibilities for
the last two symbols seen—00, 01, 10, or 11—and that the two accepting states are the states
corresponding to 10 and 11, where the next-to-last symbol is a 1.) An NFA for the language
can be much simpler:
0,1

1 0,1

This NFA has fewer states and far fewer transitions. Moreover, it is much easier to
understand. It can clearly accept a string if and only if the next-to-last symbol is a 1.

5.2 Spontaneous Transitions

An NFA has one additional trick up its sleeve: it is allowed to make a state transition
spontaneously, without consuming an input symbol.
In NFA diagrams the option of making a non-input-consuming transition is shown as
an arrow labeled with ε. (Recall that ε is not normally used as a symbol in an alphabet, but
stands for the empty string. That makes sense here since no symbols are consumed when an
ε-transition is made.) For example, this NFA accepts any string of all as and also accepts any
string of all bs:
48 CHAPTER FIVE: NONDETERMINISTIC FINITE AUTOMATA

q1 a

q0

q2 b

To accept the string aa, this NFA would first make an ε-transition to q1, then make a
transition back to q1 on the first a, and finally another transition back to q1 on the second a.
Notice that although q0 is not an accepting state, this NFA does accept the empty string.
Because of the ε-transitions, it has three possible sequences of moves when the input is the
empty string: it can make no moves, staying in q0; it can move to q1 and end there; or it
can move to q2 and end there. Since two of these are sequences that end in accepting states,
the empty string is in the language defined by this NFA. In general, any state that has an ε-
transition to an accepting state is, in effect, an accepting state too.
ε-transitions are especially useful for combining smaller automata into larger ones.
The example above is an NFA that accepts {an} {bn}, and it can be thought of as the
combination of two smaller NFAs, one for the language {an} and one for the language {bn}.
For another example, consider these two NFAs:

The first accepts L1 = {an | n is odd}, and the second accepts L2 = {bn | n is odd}. Suppose you
wanted an NFA for the union of the two languages, L1 L2. You might think of combining
the two smaller NFAs by unifying their start states, like this:
 5.2 SPONTANEOUS TRANSITIONS 49

a
a

b
b

But that does not work—it accepts strings like aab, which is not in either of the original
languages. Since NFAs may return to their start states, you cannot form an NFA for the
union of two languages just by combining the start states of their two NFAs. But you can
safely combine the two NFAs using ε-transitions, like this:
a

ε
a

b
ε

Similarly, suppose you wanted an NFA for the concatenation of the two languages,
{xy | x L1 and y L2}. You might think of combining the two smaller NFAs by using the
accepting state of the first as the start state of the second, like this:

a b

a b

But again that does not work—it accepts strings like abbaab, which is not in the desired
language. As before, the problem can be solved using ε-transitions:

a b
ε

a b
50 CHAPTER FIVE: NONDETERMINISTIC FINITE AUTOMATA

Using such techniques, large NFAs can be composed by combining smaller NFAs. This
property of NFAs is called compositionality, and it is one of the advantages NFAs have over
DFAs in some applications. We will see further examples of this in later chapters.

5.3 Nondeterminism
NFAs and DFAs both define languages, but only DFAs really define direct computational
procedures for testing language membership. It is easy to check whether a DFA accepts a
given input string: you just start in the start state, make one transition for each symbol in
the string, and check at the end to see whether the final state is accepting. You can carry this
computational procedure out for yourself or write a program to do it. But what about the
same test for an NFA—how can you test whether an NFA accepts a given input string? It is
no longer so easy, since there may be more than one legal sequence of steps for an input, and
you have to determine whether at least one of them ends in an accepting state. This seems
to require searching through all legal sequences of steps for the given input string, but how?
In what order? The NFA does not say. It does not fully define a computational procedure for
testing language membership.
This is the essence of the nondeterminism of NFAs:
1. For a given input there can be more than one legal sequence of steps.
2. The input is in the language if at least one of the legal sequences says so.
The first part of nondeterminism is something you may be familiar with from ordinary
programming. For example, a program using a random-number generator can make more
than one legal sequence of steps for a given input. But randomness is not the same thing
as nondeterminism. That second part of nondeterminism is the key, and it is what makes
nondeterministic automata seem rather alien to most programmers.
Because of their nondeterminism, NFAs are harder to implement than DFAs. In spite of
this higher level of abstraction, NFAs have many practical applications, as we will see in later
chapters. We will see algorithmic techniques for deciding whether an NFA accepts a given
string, and we will see that, in many applications, the compactness and compositionality of
NFAs outweigh the difficulties of nondeterminism.

5.4 The 5-Tuple for an NFA

To prove things about DFAs, we found that diagrams were not very useful. We developed a
formalism, the 5-tuple, to represent their essential structure. Now we’ll do the same thing for
NFAs.
First, we need a preliminary definition. Given any set S, you can form a new set by
taking the set of all the subsets of S, including the empty set and S itself. In mathematics this
is called the powerset of the original set:

If S is a set, the powerset of S is P(S) = {R | R S}.

 5.4 THE 5-TUPLE FOR AN NFA 51
The primary difference between a DFA and an NFA is that a DFA must have exactly one
transition from every state on every symbol, while an NFA can have any number, zero or
more. To reflect this change in the mathematical description of an NFA, we make a subtle
but important change: the output of the transition function, which is a single state for the
DFA, is a set of states for an NFA.

An NFA M is a 5-tuple M = (Q, , , q0, F), where

Q is the finite set of states
is the alphabet (that is, a finite set of symbols)
(Q ( {ε}) P(Q)) is the transition function
q0 Q is the start state
F Q is the set of accepting states

The definition says (Q ( {ε}) P(Q)); this is the mathematical way of giving
the type of the function, and it says that takes two inputs, a state from Q and a symbol
from {ε}, and produces a single output, which is some subset of the set Q. If the NFA
is in the qi state, then (qi, a) is the set of possible next states after reading the a symbol
and (qi, ε) is the set of possible next states that can be reached on an ε-transition, without
consuming an input symbol. Thus the transition function encodes all the information
about the arrows in an NFA diagram. All the other parts of the NFA—the set of states, the
alphabet, the start state, and the set of accepting states—are the same as for a DFA.
Consider for example this NFA:
a,b

q0 a q1 b q2

Formally, the NFA shown is M = (Q, , , q0, F), where Q = {q0, q1, q2}, = {a, b},
F = {q2}, and the transition function is
(q0, a) = {q0, q1}
(q0, b) = {q0}
(q0, ε) = {q2}
(q1, a) = {}
(q1, b) = {q2}
(q1, ε) = {}
(q2, a) = {}
52 CHAPTER FIVE: NONDETERMINISTIC FINITE AUTOMATA

(q2, b) = {}
(q2, ε) = {}
When the diagram has no arrow from a given state on a given symbol, as for (q1, a), the
transition function reflects this by producing {}, the empty set, as the set of possible next
states. When the diagram has a single arrow from a given state on a given symbol, as for
(q0, b), the transition function produces a singleton set such as {q0}, showing that there
is exactly one possible next state. When the diagram allows more than one transition
from a given state on a given symbol, as for (q0, a), the transition function produces a set
containing more than one possible next states.
The example above demonstrates an odd technicality about NFA alphabets. We
assumed for the example that = {a, b}, but all we really know from looking at the diagram
is that contains a and b. It could contain other symbols as well. A DFA diagram tells
you unequivocally what the alphabet must be, because it gives a transition from every state
on every symbol in the alphabet. An NFA need not give transitions on every symbol, so
the alphabet is not fully determined by the diagram. Usually this is an inconsequential
technicality; the language defined by the machine above is {a, b}*, even if we take it that
= {a, b, c}. But sometimes you do need to know the exact alphabet (as when taking the
complement of the language), and at such times an NFA diagram by itself is inadequate.

5.5 The Language Accepted by an NFA

As an NFA operates, it reads its input string and changes its state. At any point in the NFA’s
operation, its future depends on two things: the current state, and that part of the input
string that is still to be read. Taken together, these two pieces of information are called an
instantaneous description (ID) for the NFA. The following definitions are all made with
respect to some fixed NFA M = (Q, , , q0, F ):

An instantaneous description for an NFA is a pair (q, x) where

q Q is the current state
x * is the unread part of the input

The function for the NFA determines a relation on IDs; we write I J if I is an ID

and J is an ID that could follow from I after one move of the NFA.

is a relation on IDs. For any string x * and any or = ε,

(q, x) (r, x) if and only if r (q, ).

The definition permits or = ε. Thus the move can be either a move on one symbol
of input or an ε-transition. Next we define an extended relation * for sequences of zero or
more steps.
 EXERCISES 53

* is a relation on IDs, with I * J if and only if there is a sequence of zero or

more relations that starts with I and ends with J.

Notice here that * is reflexive; for any ID I, I * I by a sequence of zero moves. Using the
* relation, we can define the * function for M:

*(q, x) = {r | (q, x) * (r, ε)}

Thus, for any string x * and any state q Q, *(q, x) is the set of possible states for the
NFA to end up in, starting from q and making some sequence of legal -transitions on the
symbols in x. Using this extended transition function, it is possible to express the idea that an
NFA accepts a string if it has at least one path to an accepting state on that string:

A string x * is accepted by an NFA M = (Q, , , q0, F) if and only if

*(q0, x) contains at least one element of F.

A string x is accepted if and only if the NFA, started in its start state and taken through some
sequence of transitions on the symbols in x, has at least one accepting state as a possible final
state. We can also define L(M), the language accepted by an NFA M:

For any NFA M = (Q, , , q0, F ), L(M ) denotes the language accepted by M,
which is L(M) = {x * | *(q0, x) contains at least one element of F }.

Exercises
EXERCISE 1
For each of the following strings, say whether it is in the language accepted by this NFA:
a,b a,b

a a

a. aa
b. baabb
c. aabaab
d. ababa
e. ababaab
54 CHAPTER FIVE: NONDETERMINISTIC FINITE AUTOMATA

EXERCISE 2
For each of the following strings, say whether it is in the language accepted by this NFA:
a

b
a,b

a. bb
b. bba
c. bbabb
d. aabaab
e. ababa
f. abbaab
EXERCISE 3
Draw a DFA that accepts the same language as each of the following NFAs. Assume that
the alphabet in each case is {a, b}; even though not all the NFAs explicitly mention both
symbols, your DFAs must.
a.

a,ε

b.
a a,b

c.
a

d.
a, b
 EXERCISES 55
e.
b

f.

EXERCISE 4
Draw the diagram for an NFA for each of the following languages. Use as few states
and as few transitions as possible. Don’t just give a DFA, unless you are convinced it is
necessary.
a. {x {a, b}* | x contains at least 3 as}
b. {x {a, b}* | x starts with at least 3 consecutive as}
c. {x {a, b}* | x ends with at least 3 consecutive as}
d. {x {a, b}* | x contains at least 3 consecutive as}
e. {x {a, b}* | x has no two consecutive as}
f. {axb | x {a, b}*}
g. {x {0, 1}* | x ends in either 0001 or 1000}
h. {x {0, 1}* | x either starts with 000 or ends with 000, or both}
EXERCISE 5
Draw an NFA for each of the following languages. Hint: Try combining smaller NFAs
using ε-transitions.
a. {an | n is even} {bn | n is odd}
b. {(ab)n} {(aba)n}
c. {x {a, b}* | x has at least 3 consecutive as at the start or end (or both)}
d. {x {a, b}* | the number of as in x is odd} {bn | n is odd} {(aba)n}
e. {an | n is divisible by at least one of the numbers 2, 3, or 5}
f. {xy {a, b}* | the number of as in x is odd and the number of bs in y is even}
g. {xy | x {(ab)n} and y {(aba)m}}
EXERCISE 6
a. Draw a DFA for the set of strings over the alphabet {a} whose length is divisible by 3
or 5 (or both).
b. Draw an NFA for the same language, using at most nine states.
EXERCISE 7
Draw the diagram for each of the following NFAs.
a. ({q0, q1}, {0, 1}, , q0, {q0}), where the transition function is
56 CHAPTER FIVE: NONDETERMINISTIC FINITE AUTOMATA

(q0, 0) = {}
(q0, 1) = {q1}
(q0, ε) = {}
(q1, 0) = {q1}
(q1, 1) = {q0}
(q1, ε) = {}
b. ({q0, q1, q2}, {0, 1}, , q0, {q2}), where the transition function is
(q0, 0) = {q0, q1}
(q0, 1) = {q0}
(q0, ε) = {}
(q1, 0) = {}
(q1, 1) = {q2}
(q1, ε) = {q0}
(q2, 0) = {}
(q2, 1) = {}
(q2, ε) = {}
c. ({q0, q1, q2}, {a, b, c}, , q0, {q0}), where the transition function is
(q0, a) = {q1}
(q0, b) = {}
(q0, c) = {}
(q0, ε) = {}
(q1, a) = {}
(q1, b) = {q2}
(q1, c) = {}
(q1, ε) = {}
(q2, a) = {}
(q2, b) = {}
(q2, c) = {q0}
(q2, ε) = {}
EXERCISE 8
State each of the following NFAs formally, as a 5-tuple. The alphabet in each case should
be {a, b}.
a.
a a,b

a
 EXERCISES 57
b.
a

c.
a, b

d.

EXERCISE 9
For each machine M in the previous exercise, state what the language L(M ) is, using sets
and/or set formers.
EXERCISE 10
Evaluate each of these expressions for the following NFA:
0,1

1 q1 0,1 q2
q0

a. *(q0, 010)
b. (q2, 0)
c. *(q2, ε)
d. *(q0, 1101)
e. *(q0, 1011)
EXERCISE 11
Evaluate each of these expressions for the following NFA:
a,b

q0 a q1 b q2

ε
58 CHAPTER FIVE: NONDETERMINISTIC FINITE AUTOMATA

a. (q0, a)
b. (q0, ε)
c. *(q0, a)
d. *(q0, ab)
e. (q1, ε)
f. *(q1, ε)
EXERCISE 12
Construct NFAs with the following properties:
a. an NFA N with L(N) = {a}, that has exactly one accepting sequence of moves
b. an NFA N with L(N) = {a}, that has exactly two accepting sequence of moves
c. an NFA N with L(N) = {a}, that has infinitely many accepting sequences of moves
EXERCISE 13
Prove formally that for any NFA M = (Q, , , q0, F), any strings x * and
y *, and any state q Q,
δ*(q , xy) = δδ*(q' , y).
⊃

q' ∈ δ *(q , x)

Hint: induction is not helpful here. Use the definition of * for NFAs to express the sets
in terms of sequences of moves using the * relation.
EXERCISE 14
You’re given an integer i > 0. Give the 5-tuple for an NFA for the language
{0ix | x {0, 1}*}.
EXERCISE 15
You’re given a DFA M = (Q, , , q0, F). Show how to construct the 5-tuple for a new
NFA N with L(N) = L(M) {ε}.
EXERCISE 16
Prove Theorem 3.3 (closure of regular languages under union) again, this time using
NFAs in the construction.
EXERCISE 17
Show how to take any given NFA M = (Q, , , q0, F ) and construct another NFA
N = (Q', , ', q0', F'), so that |F'| = 1 and L(M ) = L(N). Be sure to give a convincing
proof that the two languages are equal.
EXERCISE 18
Suppose you’re given an NFA M that has a single accepting state. Show how to construct
the 5-tuple for a new NFA N with L(N ) = {xy | x L(M ) and y L(M )}. Be sure to give
a convincing proof that L(N) is the language specified.
 6
CHA P TER

NFA Applications

The problem with implementing NFAs is that, being nondetermin-

istic, they do not really define computational procedures for testing
language membership. To implement an NFA we must give a
computational procedure that can look at a string and decide whether
the NFA has at least one sequence of legal transitions on that string
leading to an accepting state. This seems to require searching through
all legal sequences for the given input string—but how?
One approach is to implement a direct backtracking search. Another
is to convert the NFA into a DFA and implement that instead. This
conversion is both useful and theoretically interesting. The fact that
it is always possible shows that in spite of their extra flexibility, NFAs
have exactly the same power as DFAs; they can define exactly the
regular languages.

59
60 CHAPTER SIX: NFA APPLICATIONS

6.1 An NFA Implemented with a Backtracking Search

Here is an NFA that accepts the language of strings over the alphabet {0, 1} that have a 1 as
the next-to-last symbol.
0,1

q0 1 q1 0,1 q2

Now let’s see how this NFA can be implemented in the Java language. As in Chapter 4, we
will write the NFA as a Java class. As in the table-based implementation in Section 4.3, we
will use a table to represent the transition function . Where (q, a) is the set of zero or more
possible next states from the state q reading the symbol a, in our Java array
delta[q][a-'0'] will be an array of zero or more possible next states. Thus the delta
array is a three-dimensional array of int values.

/**
* A nondeterministic finite-state automaton that
* recognizes strings of 0s and 1s with 1 as the
* next-to-last character.
*/
public class NFA1 {

/*
* The transition function represented as an array.
* The entry at delta[s,c-'0'] is an array of 0 or
* more ints, one for each possible move from
* the state s on the character c.
*/
private static int[][][] delta =
{{{0},{0,1}}, // delta[q0,0], delta[q0,1]
{{2},{2}}, // delta[q1,0], delta[q1,1]
{{},{}}}; // delta[q2,0], delta[q2,1]

To determine whether the NFA accepts a string, we will do the obvious thing: just
search all possible paths through the machine on that string. If we find one that ends in
an accepting state, we accept; otherwise we reject. This search is easy to implement using
recursion:
 6.1 AN NFA IMPLEMENTED WITH A BACKTRACKING SEARCH 61
/**
* Test whether there is some path for the NFA to
* reach an accepting state from the given state,
* reading the given string at the given character
* position.
* @param s the current state
* @param in the input string
* @param pos index of the next char in the string
* @return true if the NFA accepts on some path
*/
private static boolean accepts
(int s, String in, int pos) {
if (pos==in.length()) { // if no more to read
return (s==2); // accept if final state is q2
}

char c = in.charAt(pos++); // get char and advance

int[] nextStates;
try {
nextStates = delta[s][c-'0'];
}
catch (ArrayIndexOutOfBoundsException ex) {
return false; // no transition, just reject
}

// At this point, nextStates is an array of 0 or

// more next states. Try each move recursively;
// if it leads to an accepting state return true.

for (int i=0; i < nextStates.length; i++) {

if (accepts(nextStates[i], in, pos)) return true;
}

return false; // all moves fail, return false

}

The code uses some of the same tricks we saw in Chapter 4. In particular, it uses
exception handling to catch out-of-bounds array references, so that we do not need the array
to have entries for char values other than '0' and '1'. The new trick here is the recursive
62 CHAPTER SIX: NFA APPLICATIONS

search. A call to accepts(s,in,pos) tests whether an accepting state is reachable,

starting from the state s, reading the string in from the position pos. In the base case, pos
is equal to in.length, so the entire string has already been read; in that case the NFA
accepts only if s is the accepting state, state 2. In the recursive case, for each possible next
move, the code makes a recursive call to test whether an accepting state is reachable from
there.
Finally, we need a public method that clients can call to test whether a string is in the
language accepted by the NFA. We will use the name accepts for this method too, since
Java permits this kind of overloading. accepts(in) just starts up the recursive search by
making an initial call to accepts(0,in,0)—starting from the state 0 (the start state)
and the character position 0 (the first character in the string).

/**
* Test whether the NFA accepts the string.
* @param in the String to test
* @return true if the NFA accepts on some path
*/
public static boolean accepts(String in) {
return accepts(0, in, 0); // start in q0 at char 0
}
}

That is the end of the class definition. To test whether a string s is in the language
defined by this NFA, we would write something like

if (NFA1.accepts(s)) ...

This implementation is not particularly object-oriented, since the whole thing uses static
methods and fields and no objects of the NFA1 class are created. All the information
manipulated in the recursive search is carried in the parameters, so no objects were needed.

6.2 An NFA Implemented with a Bit-Mapped Parallel Search

The implementation just given uses a backtracking search. It tries one sequence of moves; if
that fails it backs up and tries another, and so on until it finds a successful sequence or runs
out of sequences to try. In effect, it tries sequences one at a time. Another approach to NFA
implementation is to search all paths simultaneously.
This is not as difficult as it sounds. We will implement this like the table-driven DFA of
Chapter 4, except that instead of keeping track of the single state the machine is in after each
character, we will keep track of the set of possible states after each character.
 6.2 AN NFA IMPLEMENTED WITH A BIT-MAPPED PARALLEL SEARCH 63
/**
* A nondeterministic finite-state automaton that
* recognizes strings with 0 as the next-to-last
* character.
*/
public class NFA2 {

/*
* The current set of states, encoded bitwise:
* state i is represented by the bit 1<<i.
*/
private int stateSet;

We use a bit-coded integer to represent a set of states. The << operator in Java shifts an
integer to the left, so 1<<0 is 1 shifted left 0 positions (thus 1<<0 is 1), 1<<1 is 1 shifted
left 1 position (thus 1<<1 is 2), 1<<2 is 1 shifted left 2 positions (thus 1<<2 is 4), and so
on. In general, an NFA implemented this way will use the number 1<<i to represent the
i state. For this NFA, there are only the three states 1<<0, 1<<1, and 1<<2. Since we are
using one bit for each state, we can form any set of states just by combining the state bits
with logical OR. For example, the set {q0, q2} can be represented as 1<<0|1<<2.
The initial state set is {q0}. This method initializes the current state set to that value:

/**
* Reset the current state set to {the start state}.
*/
public void reset() {
stateSet = 1<<0; // {q0}
}

We will use the table-driven approach for the transition function, just as we did for
DFAs. In this case, however, each entry in the table is not just a single state, but a bit-coded
set of states:

/*
* The transition function represented as an array.
* The set of next states from a given state s and
* character c is at delta[s,c-'0'].
*/
static private int[][] delta =
64 CHAPTER SIX: NFA APPLICATIONS

{{1<<0, 1<<0|1<<1}, // delta[q0,0] = {q0}

// delta[q0,1] = {q0,q1}
{1<<2, 1<<2}, // delta[q1,0] = {q2}
// delta[q1,1] = {q2}
{0, 0}}; // delta[q2,0] = {}
// delta[q2,1] = {}

The process method looks at characters in the input string and keeps track of the set
of possible states after each one. This is the only part that is significantly more complicated
than the DFA case, because for each character in the input string we have to execute a little
loop. To compute the next set of states given a symbol c, we need to take the union of the
sets (q, c) for each state q in the current set of states. The process method does that in its
inner loop; its outer loop repeats the computation once for each symbol in the input string.

/**
* Make one transition from state-set to state-set on
* each char in the given string.
* @param in the String to use
*/
public void process(String in) {
for (int i = 0; i < in.length(); i++) {
char c = in.charAt(i);
int nextSS = 0; // next state set, initially empty
for (int s = 0; s <= 2; s++) { // for each state s
if ((stateSet&(1<<s)) != 0) { // if maybe in s
try {
nextSS |= delta[s][c-'0'];
}
catch (ArrayIndexOutOfBoundsException ex) {
// in effect, nextSS |= 0
}
}
}
stateSet = nextSS; // new state set after c
}
}
 6.3 THE SUBSET CONSTRUCTION 65
All that remains is a method to test whether the NFA should accept. It accepts if the final
set of possible states includes the accepting state q2.

/**
* Test whether the NFA accepted the string.
* @return true if the final set includes
* an accepting state.
*/
public boolean accepted() {
return (stateSet&(1<<2))!=0; // true if q2 in set
}
}

That is the end of the class definition. To test whether a string s is in the language defined by
this NFA, we would write something like

NFA2 m = new NFA2();

m.reset();
m.process(s);
if (m.accepted()) ...

The interface to this is largely the same as for the DFAs we developed in Chapter 4.
Because the Java int type is a 32-bit word, our NFA2 code generalizes simply for any
NFA of up to 32 states. For NFAs of up to 64 states, the 64-bit long type could be used
easily. To implement an NFA with more than 64 states in Java would require some additional
programming. One could, for example, use an array of ⎡n/32⎤ int variables to represent
the set of n states. The process loop would be considerably more complicated, and slower,
for such large NFAs.
As the NFA2 implementation runs, the stateSet field keeps track of the set of states
that the NFA could be in. This basic idea—keeping track of the set of possible states after
each symbol is read—is the basis for a construction that lets us convert any NFA into a DFA.

6.3 The Subset Construction

For any NFA, there is a DFA that recognizes the same language. Thus, while NFAs can be
simpler to write than DFAs and can use fewer states and transitions, they are not any more
powerful than DFAs. The power of nondeterminism does not make it possible for finite-state
machines to define any additional languages. This can be proved by showing how to take any
NFA and construct a DFA that accepts the same language.
Before giving the general method of constructing a DFA from any NFA, let’s look at an
example. We begin with our NFA for {x {0, 1}* | the next-to-last symbol in x is 1}:
66 CHAPTER SIX: NFA APPLICATIONS

0,1

q0 1 q1 0,1 q2

We will construct a DFA for the same language. The new DFA will in fact simulate the NFA,
by keeping track of the states the NFA might be in after each symbol of the input string.
Being (potentially) nondeterministic, the NFA can have more than one legal sequence of
steps for an input, and so can be in more than one possible state after each symbol of the
input string. So each state of the DFA must correspond to some subset of the set of states
from the NFA. That is why this is called the subset construction.
Initially, of course, the NFA is in its start state. We will create a start state for the DFA
and label it with the set of states {q0}:

0 ...

q0

1 ...

Next, suppose the NFA is in q0 and reads a 0. What is the set of possible states after that?
It is (q0, 0) = {q0} again, so the transition on 0 in our DFA should return to the state labeled
{q0}. If the NFA is in q0 and reads a 1, its set of possible next states is (q0, 1) = {q0, q1}, so we
will add a new state labeled {q0, q1} to the DFA under construction. So far we have

0 ...

q0 1 q0,q1
0
1 ...

Next, suppose the NFA is in one of the states {q0, q1} and reads a 0. What is the set of
possible states after that? We have (q0, 0) = {q0} and (q1, 0) = {q2}, so the next set of possible
states is (q0, 0) (q1, 0) = {q0, q2}. Similarly, if the NFA is in one of the states {q0, q1} and
reads a 1, its set of possible next states is (q0, 1) (q1, 1) = {q0, q1, q2}. We add these two
new states to the DFA under construction, producing
 6.3 THE SUBSET CONSTRUCTION 67

q0 1 q0,q1
0

0 1

... 0 0 ...

q0,q2 q0,q1,q2

... 1 ...

Each step in the construction follows this same pattern. Every state of the DFA is labeled
with a set of states from the NFA. The transition on a symbol a from a state labeled R in the
DFA is determined by taking the union, over all r R, of the NFA’s (r, a). (This is exactly
what the inner loop of the process method in our NFA2 implementation computes.)
Because there are only finitely many different subsets of the set of states in the NFA; this
construction eventually stops with a finite DFA. In our example, in fact, we have already
found all the reachable subsets of states; the remaining transitions all return to sets already
added to the DFA.
The only remaining problem is to decide which of the states of the DFA should be
accepting states. Since the NFA will accept whenever its set of possible final states contains
at least one accepting state, that is what the DFA should do too. All the states labeled with
subsets containing at least one accepting state of the NFA should be accepting states of the
DFA. (This is exactly what the accepted method of our NFA2 implementation does.)
The final result is this:

q0 1 q0,q1
0
1

1
0
0

q0,q2 q0,q1,q2 1
0

If you now discard the state labels, you can see that this is the same as the DFA given in
Section 5.1 for the original language, {x {0, 1}* | the next-to-last symbol in x is 1}.
The construction just illustrated is the subset construction. The following proof expresses
it formally:
68 CHAPTER SIX: NFA APPLICATIONS

Lemma 6.1: If L = L(N) for some NFA N, then L is a regular language.

Proof: Suppose L = L(N) for some NFA N = (QN, , N
, qN, FN ). Construct a
new DFA D = (QD, , D, qD, FD ), where

Q D = P(Q N)
δD(R , a) = δδ*N (r , a), for all R ∈ Q D and a ∈ Σ

⊃
r ∈R
q D = δδ*N (q N , ε)
ε
F D = {R ∈ Q D | R ∩ F N ≠ {} }

Now by construction we have δ *D (q D , x) = q D = δ *N (q N , x) for all x *. Thus

D simulates N’s behavior on each input string x, accepting if and only if N
accepts. We conclude that L = L(N) = L(D), and it follows that L is a regular
language.

The jump from the definition of D to the conclusion that δ *D (q D , x) = q D = δ *N (q N , x)

can be made more rigorous using a routine inductive proof of the kind demonstrated in
Section 3.4. Notice that in the base case, δ *D (q D , εε) = q D = δ *N (q N , ε ). In the construction,
the start state for the new DFA is q D = δ *N (q N , ε ). Often this start state will be simply the set
q D = {q N } , as it was in our earlier example of the construction. But the start state of the DFA
will contain additional states from the NFA if the NFA has ε-transitions from its start state.
The machine constructed for the proof includes all the states in P(QN ), which may
include some unreachable states. (We saw something similar in Section 3.2 with the product
construction.) In the case of our earlier example, this will be an eight-state DFA instead of
the four-state DFA. The extra states are not reachable by any path from the start state, so
they have no effect on the language accepted.
The construction always includes a DFA state for {}, the empty set of NFA states,
because the empty set is a subset of every set. The DFA will be in this state on any string for
which the NFA has no legal sequence of moves—any string x for which δ*N (q N , x) = {} . For
example, consider this NFA, which accepts only the empty string:

q0

There is no legal sequence of moves in this NFA for any string other than the empty string.
The set of states QN = {q0}, and the powerset of that is QD = P(QN ) = {{}, {q0}}. The DFA
produced by the construction will therefore be this two-state machine:
 6.4 NFAS ARE EXACTLY AS POWERFUL AS DFAS 69

q0 0,1
0,1

No matter what the original NFA is, the construction always gives
δD({} , a) = δδ*N (r , a)= {}

⊃
r ∈ {}

so the {} state in the DFA produced by the subset construction always has transitions back
to itself for all symbols in the alphabet. In this way, the subset construction automatically
provides {} as a nonaccepting trap state. This trap state will be reachable from the start state if
and only if the original NFA can get stuck—that is, if and only if there is some input string x
for which δ *N (q N , x) = {}.
We have seen two different approaches for implementing NFAs. The subset construction
gives us a third: we can first convert the NFA into a DFA using the subset construction, then
implement that using the techniques of Chapter 4.

6.4 NFAs Are Exactly as Powerful as DFAs

The subset construction shows that NFAs are no more powerful than DFAs. It is also true,
more obviously, that NFAs are no less powerful than DFAs.

Lemma 6.2: If L is any regular language, there is some NFA N for which
L(N) = L.
Proof: Let L be any regular language. By definition there must be some DFA
D = (Q, , D, q0, F) with L(D) = L. This DFA immediately gives an equivalent
NFA N = (Q, , N, q0, F), where we define N(q, a) = { D(q, a)} for all q Q
and a , and N(q, ε) = {} for all q Q. Clearly, we have *N(q, x) = { *D(q, x)}
for all q Q and x *, and so N accepts x if and only if D
accepts x. It follows that L(N) = L(D) = L.

Although in diagrams it is clear that every DFA is an NFA, it is not quite true with
formal 5-tuples, since the functions of DFAs and NFAs have different types. However,
this difference is trivial, as the proof above illustrates; wherever a DFA has D(q, a) = r, an
equivalent NFA should have N(q, a) = {r }. The jump in the proof from N(q, a) = { D(q, a)}
to *N(q, x) = { *D(q, x)} can be made more rigorous using a routine inductive proof of the
kind demonstrated in Section 3.4.
From the previous two lemmas it follows that the set of languages that can be defined
using NFAs is exactly the set of regular languages.
70 CHAPTER SIX: NFA APPLICATIONS

Theorem 6.1: A language L is L(N ) for some NFA N if and only if L is a regular
language.
Proof: It follows immediately from Lemmas 6.1 and 6.2.

We conclude that allowing nondeterminism in finite automata can make them more
compact and easier to construct, but in the sense shown in this theorem, it neither weakens
nor strengthens them.

6.5 DFA or NFA?

We have seen a number of techniques for recognizing regular languages using DFA or NFA
ideas. No one technique is best in all circumstances.
The big advantage of DFA techniques is their speed. They run quickly and predictably,
taking time that is simply proportional to the length of the string. One disadvantage of DFA
techniques is that they are difficult to extend for nonregular constructs; NFAs are better for
such applications, as discussed below. Another potential disadvantage is that the size of the
DFA code (including the size of the transition array, if used) is proportional to the number
of states in the DFA. This is usually not a problem, but there are languages for which the
smallest DFA is exponentially larger than the smallest NFA. The subset construction gives
some intuition about how that can happen: for an NFA with n states, the subset construction
yields a DFA with 2n states and can still have exponentially many states left over, even after
the unreachable ones are eliminated. Compilers are an example of an application for which
these trade-offs favor DFAs. The first phase of most compilers is a scanner that breaks the
source file into tokens like keywords, identifiers, and constants. These token languages are
purely regular, the DFAs tend to be of modest size, and speed is critical, so DFA techniques
work well here.
NFA-based techniques using a backtracking search are widely used for other applications.
They tend to be slower, but they have their own advantages. For one thing, they do not
suffer from the problem of having exponential-size worst cases. In practice, though, this is
not as important as their major advantage, which is that they are easier to extend for non-
regular-language constructs. Many programs that use regular expressions need all the power
of regular languages and then some. This includes text-processing tools like emacs, vi, and
sed, languages like Perl and PHP, and language libraries like those for Java and the .NET
languages. In a later chapter we discuss more about regular expressions and their applications.
For now, suffice it to say that backtracking, NFA-based implementations are more easily
extended to handle the nonregular constructs required by these applications.
There are also many variations and many possible optimizations. Some systems use a
DFA implementation for purely regular tasks and an NFA only for the nonregular tasks.
 EXERCISES 71
Some systems use hybrid techniques that combine the DFA and NFA approaches, like the
bit-mapped parallel search, augmented with optimizations like the caching of frequently used
state sets and transitions.

6.6 Further Reading

Automata are implemented in an amazing variety of applications. For a look at the state
of the art in this active research area, see the proceedings of the annual Conference on
Implementation and Application of Automata. As of this writing, the most recent one was
Antipolis, Sophia. “Implementation and Application of Automata, 10th
International Conference (CIAA 2005).” Published as Lecture Notes in Computer
Science, vol. 3845. Paris: Springer-Verlag, 2005.

Exercises
EXERCISE 1
The DFA constructed for the example in Section 6.3 has four states, but the full DFA
given by the formal subset construction has eight. Draw the full eight-state DFA,
including the unreachable states.
EXERCISE 2
Convert this NFA into a DFA using the subset construction. Draw the DFA and label
each state with the corresponding subset of states from the NFA. You do not have to
show unreachable states, if any.

q0 a q1 b q2

EXERCISE 3
Convert this NFA into a DFA using the subset construction. Draw the DFA and label
each state with the corresponding subset of states from the NFA. You do not have to
show unreachable states, if any.
a,b

q0 a q1 b q2
72 CHAPTER SIX: NFA APPLICATIONS

EXERCISE 4
Convert this NFA into a DFA using the subset construction. Draw the DFA and label
each state with the corresponding subset of states from the NFA. You do not have to
show unreachable states, if any.
a,b

q0 a q1 b,ε q2

EXERCISE 5
Convert this NFA into a DFA using the subset construction. Draw the DFA and label
each state with the corresponding subset of states from the NFA. You do not have to
show unreachable states, if any.

0,1
q0 q1

0,1 1 1

0 q2

EXERCISE 6
Convert this NFA into a DFA using the subset construction. Draw the DFA and label
each state with the corresponding subset of states from the NFA. You do not have to
show unreachable states, if any.

b
q1 q4
a,ε a,ε

a b b
q0 q2 q5 q6

b a

q3
 EXERCISES 73
EXERCISE 7
Write a Java class that implements this NFA:
a,b a,b

a a

Use a table-driven, backtracking-search approach as in Section 6.1.

EXERCISE 8
Repeat the previous exercise, but use a table-driven, parallel-search approach as in
Section 6.2.
EXERCISE 9
Write a Java class that implements this NFA:
a

ε
a

b
ε

Use a table-driven, backtracking-search approach as in Section 6.1. You will need to

consider how best to modify this approach to incorporate ε-transitions.
EXERCISE 10
Repeat the previous exercise, but use a table-driven, parallel-search approach as in
Section 6.2. Again, you will need to consider how best to modify the approach to
incorporate ε-transitions.
EXERCISE 11
Rewrite the proof of Lemma 6.2, but include a rigorous inductive proof to show that
*N(q, x) = { *D(q, x)}.
EXERCISE 12
Rewrite the proof of Lemma 6.1, but include a rigorous inductive proof to show that
δ *D (q D , x) = δ*N (q N , x).
 CHA P TER

Regular Expressions
7
Most programmers and other power-users of computer systems have used tools that match text
patterns. You may have used a Web search engine with a pattern like travel cancun
OR acapulco, trying to find information about either of two travel destinations. You may
have used a search tool with a pattern like gr[ea]y, trying to find files with occurrences of
either grey or gray. Any such pattern for text works as a definition of a formal language.
Some strings match the pattern and others do not; the language defined by the pattern is the
set of strings that match. In this chapter we study a particular kind of formal text pattern
called a regular expression. Regular expressions have many applications. They also provide
an unexpected affirmation of the importance of some of the theoretical concepts from previous
chapters.
This is a common occurrence in mathematics. The first time a young student sees the mathematical
constant , it looks like just one more school artifact: one more arbitrary symbol whose definition
to memorize for the next test. Later, if he or she persists, this perception changes. In many
branches of mathematics and with many practical applications, keeps on turning up. "There
it is again!" says the student, thus joining the ranks of mathematicians for whom mathematics
seems less like an artifact invented and more like a natural phenomenon discovered.
So it is with regular languages. We have seen that DFAs and NFAs have equal definitional power.
It turns out that regular expressions also have exactly that same definitional power: they can be
used to define all the regular languages and only the regular languages. There it is again!
75
76 CHAPTER SEVEN: REGULAR EXPRESSIONS

7.1 Regular Expressions, Formally Defined

Let’s define two more basic ways of combining languages. The concatenation of two
languages is the set of strings that can be formed by concatenating an element of the first
language with an element of the second language:

If L1 and L2 are languages, the concatenation of L1 and L2 is L1L2 = {xy | x L1

and y L2}.

The Kleene closure of a language is the set of strings that can be formed by concatenating
any number of strings, each of which is an element of that language:

If L is a language, the Kleene closure of L is L* = {x1x2 ... xn | n 0, with all

xi L}.

Note that this generalizes the definition we used in Chapter 1 for the Kleene closure of an
alphabet.
There is a common mistake to guard against as you read the Kleene closure definition.
It does not say that L* is the set of strings that are concatenations of zero or more copies
of some string in L. That would be {xn | n 0, with x L}; compare that with the actual
definition above. The actual definition says that L* is the set of strings that are concatenations
of zero or more substrings, each of which is in the language L. Each of those zero or more
substrings may be a different element of L. For example, the language {ab, cd}* is the
language of strings that are concatenations of zero of more things, each of which is either ab
or cd: {ε, ab, cd, abab, abcd, cdab, cdcd, …}. Note that because the definition allows zero or
more, L* always includes ε.
Now we can define regular expressions.

A regular expression is a string r that denotes a language L(r) over some alphabet
. The six kinds of regular expressions and the languages they denote are as
follows. First, there are three kinds of atomic regular expressions:
1. Any symbol a is a regular expression with L(a) = {a}.
2. The special symbol ε is a regular expression with L(ε) = {ε}.
3. The special symbol is a regular expression with L( ) = {}.
There are also three kinds of compound regular expressions, which are built from
smaller regular expressions, here called r, r1, and r2:
4. (r1 + r2) is a regular expression with L(r1 + r2) = L(r1) L(r2)
5. (r1r2) is a regular expression with L(r1r2) = L(r1)L(r2)
6. (r)* is a regular expression with L((r)*) = (L(r))*
 7.2 EXAMPLES OF REGULAR EXPRESSIONS 77
The parentheses in compound regular expressions may be omitted, in which
case * has highest precedence and + has lowest precedence.

Regular expressions make special use of the symbols ε, , +, *, (, and ), so for simplicity
we will assume that these special symbols are not included in . Regular expressions can
look just like ordinary strings, so you will occasionally have to rely on the context to decide
whether a string such as abc is just a string or a regular expression denoting the language
{abc}.
Unfortunately, the term regular expression is heavily overloaded. It is used for the text
patterns of many different tools like awk, sed, and grep, many languages like Perl, Python,
Ruby, and PHP, and many language libraries like those for Java and the .NET languages.
Each of these applications means something slightly different by regular expression, and we
will see more about these applications later.

7.2 Examples of Regular Expressions

Using just the six basic kinds of regular expressions, you can define quite a wide variety of
languages. For starters, any simple string denotes the language containing just that string:
Regular expression: ab
Language denoted: {ab}
Notice that the formal definition permits this because a and b, by themselves, are atomic
regular expressions of type 1; they are combined into a compound regular expression of
type 5; and the parenthesis in (ab) are omitted since they are unnecessary.
Any finite language can be defined, just by listing all the strings in the language,
separated by the + operator:
Regular expression: ab + c
Language denoted: {ab, c}
Again we were able to omit all the parentheses from the fully parenthesized form ((ab) + c).
The inner pair is unnecessary because + has lower precedence than concatenation. Thus,
ab + c is equivalent to (ab) + c, not to a(b + c), which would denote a different language:
{ab, ac}.
The only way to define an infinite language using regular expressions is with the Kleene
star:
Regular expression: ba*
Language denoted: {ban} = {b, ba, baa, baaa, ...}
The Kleene star has higher precedence than concatenation. ba* is equivalent to b(a*), not to
(ba)*, which would denote the language {(ba)n}.
L((r)*) is the language of strings that are concatenations of zero or more substrings, each
of which is in the language L(r). Because this is zero or more and not one or more, we always
78 CHAPTER SEVEN: REGULAR EXPRESSIONS

have ε L((r)*). Each of the concatenated substrings can be a different element of L(r). For
example:
Regular expression: (a + b)*
Language denoted: {a, b}*
In that example the parentheses were necessary, since * has higher precedence than +. The
regular expression a + b* would denote the language {a} {bn}.
It is important to understand that (a + b)* is not the same as (a* + b*). Any confusion
about this point will cause big problems as we progress. In L((a + b)*) we concatenate zero
or more things together, each of which may be either an a or a b. That process can construct
every string over the alphabet {a, b}. In L(a* + b*) we take the union of two sets: the set of
all strings over the alphabet {a}, and the set of all strings over the alphabet {b}. That result
contains all strings of only as and all strings of only bs, but does not contain any strings that
contain both as and bs.
Occasionally the special symbol ε appears in a regular expression, when the empty string
is to be included in the language:
Regular expression: ab + ε
Language denoted: {ab, ε}
The special symbol appears very rarely. Without it there is no other way to denote the
empty set, but that is all it is good for. It is not useful in compound regular expressions, since
for any regular expression r we have L((r) ) = L( (r)) = {}, L(r + ) = L( + r) = L(r), and
L( *) = {ε}.
The subexpressions in a compound expression may be compound expressions
themselves. This way, compound expressions of arbitrary complexity may be developed:
Regular expression: (a + b)(c + d )
Language denoted: {ac, ad, bc, bd }
Regular expression: (abc)*
Language denoted: {(abc)n} = {ε, abc, abcabc, abcabcabc, ...}
Regular expression: a*b*
Language denoted: {anbm} = {ε, a, b, aa, ab, bb, aaa, aab, ...}
Regular expression: (a + b)*aa(a + b)*
Language denoted: {x {a, b}* | x contains at least two consecutive as}
Regular expression: (a + b)*a(a + b)*a(a + b)*
Language denoted: {x {a, b}* | x contains at least two as}
Regular expression: (a*b*)*
Language denoted: {a, b}*
 7.3 FOR EVERY REGULAR EXPRESSION, A REGULAR LANGUAGE 79
In this last example, the regular expression (a*b*)* turns out to denote the same language
as the simpler (a + b)*. To see why, consider that L(a*b*) contains both a and b. Those two
symbols alone are enough for the Kleene star to build all of {a, b}*. In general, whenever
L(r), then L((r)*) = *.

7.3 For Every Regular Expression, a Regular Language

We will now demonstrate that any language defined by a regular expression can be defined
by both of the other mechanisms already studied in this book: DFAs and NFAs. To prove
this, we will show how to take any regular expression and construct an equivalent NFA. We
choose NFAs for this, rather than DFAs, because of a property observed in Section 5.2: large
NFAs can be built up as combinations of smaller ones by the judicious use of ε-transitions.
To make the constructed NFAs easy to combine with each other, we will make sure they
all have exactly one accepting state, not the same as the start state. For any regular expression
r we will show how to construct an NFA N with L(N) = L(r) that can be pictured like this:

The constructed machine will have a start state, a single accepting state, and a collection
of other states and transitions that ensure that there is a path from the start state to the
accepting state if and only if the input string is in L(r).
Because all our NFAs will have this form, they can be combined with each other very
neatly. For example, if you have NFAs for L(r1) and L(r2), you can easily construct an NFA
for L(r1 + r2):

r1

ε ε

ε ε

r2

In this new machine there is a new start state with ε-transitions to the two original start states
and a new accepting state with ε-transitions from the two original accepting states (which are
no longer accepting states in the new machine). Clearly the new machine has a path from the
80 CHAPTER SEVEN: REGULAR EXPRESSIONS

start state to the accepting state if and only if the input string is in L(r1 + r2). And it has the
special form—a single accepting state, not the same as the start state—which ensures that it
can be used as a building block in even larger machines.
The previous example shows how to do the construction for one kind of regular
expression—r1 + r2. The basic idea of the following proof sketch is to show that the same
construction can be done for all six kinds of regular expressions.

Lemma 7.1: If r is any regular expression, there is some NFA N that has a single
accepting state, not the same as the start state, with L(N ) = L(r).
Proof sketch: For any of the three kinds of atomic regular expressions, an NFA
of the desired kind can be constructed as follows:

a
a ∈∑:

ε
ε:

∅:

For any of the three kinds of compound regular expressions, given appropriate
NFAs for regular subexpressions r1 and r2, an NFA of the desired kind can be
constructed as follows:
 7.4 REGULAR EXPRESSIONS AND STRUCTURAL INDUCTION 81

r1

ε ε

(r1 + r2):

ε ε

r2

r1

ε
(r1r2):

r2

ε r1 ε
(r1)*:

Thus, for any regular expression r, we can construct an equivalent NFA with a
single accepting state, not the same as the start state.

7.4 Regular Expressions and Structural Induction

The proof sketch for Lemma 7.1 leaves out a number of details. A more rigorous proof
would give the 5-tuple form for each of the six constructions illustrated and show that each
82 CHAPTER SEVEN: REGULAR EXPRESSIONS

machine actually accepts the language it is supposed to accept. More significantly, a rigorous
proof would be organized as a structural induction.
As we have observed, there are as many different ways to prove something with induction
as there are to program something with recursion. Our proofs in Chapter 3 concerning
DFAs used induction on a natural number—the length of the input string. That is the style
of induction that works most naturally for DFAs. Structural induction performs induction
on a recursively defined structure and is the style of induction that works most naturally
for regular expressions. The base cases are the atomic regular expressions, and the inductive
cases are the compound forms. The inductive hypothesis is the assumption that the proof has
been done for structurally simpler cases; for a compound regular expression r, the inductive
hypothesis is the assumption that the proof has been done for r’s subexpressions.
Using structural induction, a more formal proof of Lemma 7.1 would be organized like
this:

Proof: By induction on the structure of r.

Base cases: When r is an atomic expression, it has one of these three forms:

1. Any symbol a is a regular expression with L(a) = {a}.

2. The special symbol ε is a regular expression with L(ε) = {ε}.
3. The special symbol is a regular expression with L( ) = {}.

(For each atomic form you would give the NFA, as in the previous proof
sketch.)
Inductive cases: When r is a compound expression, it has one of these three
forms:

4. (r1 + r2) is a regular expression with L(r1 + r2) = L(r1) L(r2)

5. (r1r2) is a regular expression with L(r1r2) = {xy | x L(r1) and y L(r2)}
6. (r1)* is a regular expression with
L((r1)*) = {x1x2 ... xn | n 0, with all xi L(r1)}
By the inductive hypothesis, there is an NFA that has a single accepting state,
not the same as the start state, for r1 and r2.
(For each compound form you would give the ε-NFA, as in the previous proof
sketch, and show that it accepts the desired language.)

A proof using this style of induction demonstrates that Lemma 7.1 holds for any atomic
regular expression and then shows that whenever it holds for some expressions r1 and r2,
 7.5 FOR EVERY REGULAR LANGUAGE, A REGULAR EXPRESSION 83
it also holds for (r1 + r2), (r1r2), and (r1)*. It follows that the lemma holds for all regular
expressions.

7.5 For Every Regular Language, a Regular Expression

There is a way to take any NFA and construct an equivalent regular expression.

Lemma 7.2: If N is any NFA, there is some regular expression r with

L(N) = L(r).

The construction that proves this lemma is rather tricky, so it is relegated to Appendix A. For
now, a short example will suffice.
Consider again this NFA (which is also a DFA) from Chapter 3:
0 1

1 0

0 1 2

1 0

We proved that this machine accepts the language of strings that are binary representations of
numbers that are divisible by three. Now we will construct a regular expression for this same
language. It looks like a hard problem; the trick is to break it into easy pieces.
When you see an NFA, you normally think only of the language of strings that take it
from the start state to end in any accepting state. But let’s consider some other languages that
are relevant to this NFA. For example, what is the language of strings that take the machine
from state 2 back to state 2, any number of times, without passing through states 0 or 1? Any
string of zero or more 1s would do it, and that is an easy language to give a regular expression
for:
1*
Next, a bigger piece: what is the language of strings that take the machine from 1 back to
1, any number of times, without passing through state 0? The machine has no transition that
allows it to go from 1 directly back to 1. So each single trip from 1 back to 1 must follow
these steps:
1. Go from 1 to 2.
2. Go from 2 back to 2, any number of times, without passing through 0 or 1.
3. Go from 2 to 1.
The first step requires a 0 in the input string. The second step is a piece we already have
a regular expression for: 1*. The third step requires a 0 in the input string. So the whole
language, the language of strings that make the machine do those three steps repeated any
84 CHAPTER SEVEN: REGULAR EXPRESSIONS

number of times, is
(01*0)*
Next, a bigger piece: what is the language of strings that take the machine from 0 back to
0, any number of times? There are two ways to make a single trip from 0 back
to 0. The machine can make the direct transition from state 0 to state 0 on an input
symbol 0 or it can follow these steps:
1. Go from 0 to 1.
2. Go from 1 back to 1, any number of times, without passing through 0.
3. Go from 1 to 0.
The first step requires a 1 in the input string. The second step is a piece we already have a
regular expression for: (01*0)*. The third step requires a 1 in the input string. So the whole
language, the language of strings that make the machine go from state 0 back to state 0 any
number of times, is
(0 + 1(01*0)*1)*
This also defines the whole language of strings accepted by the NFA—the language of strings
that are binary representations of numbers that are divisible by three.
The proof of Lemma 7.2 is a construction that builds a regular expression for the
language accepted by any NFA. It works something like the example just shown, defining the
language in terms of smaller languages that correspond to restricted paths through the NFA.
Appendix A provides the full construction. But before we get lost in those details, let’s put all
these results together:

Theorem 7.1 (Kleene’s Theorem): A language is regular if and only if it is L(r)

for some regular expression r.
Proof: Follows immediately from Lemmas 7.1 and 7.2.

DFAs, NFAs, and regular expressions all have equal power for defining languages.
 EXERCISES 85

Exercises
EXERCISE 1
Give a regular expression for each of the following languages.
a. {abc}
b. {abc, xyz}
c. {a, b, c}*
d. {ax | x {a, b}*}
e. {axb | x {a, b}*}
f. {(ab)n}
g. {x {a, b}* | x contains at least three consecutive as}
h. {x {a, b}* | the substring bab occurs somewhere in x}
i. {x {a, b}* | x starts with at least three consecutive as}
j. {x {a, b}* | x ends with at least three consecutive as}
k. {x {a, b}* | x contains at least three as}
l. {x {0, 1}* | x ends in either 0001 or 1000}
m. {x {0, 1}* | x either starts with 000 or ends with 000, or both}
n. {an | n is even} {bn | n is odd}
o. {(ab)n} {(aba)n}
p. {x {a}* | the number of as in x is odd}
q. {x {a, b}* | the number of as in x is odd}
r. {x {a, b}* | the number of as in x is odd} {bn | n is odd} {(aba)n}
s. {an | n is divisible by at least one of the numbers 2, 3, or 5}
t. {x {a, b}* | x has no two consecutive as}
u. {xy {a, b}* | the number of as in x is odd and the number of bs in y is even}
EXERCISE 2
For each of these regular expressions, give two NFAs: the exact one constructed by the
proof of Lemma 7.1, and the smallest one you can think of.
a.
b. ε
c. a
d. 0+1
e. (00)*
f. ab*
EXERCISE 3
For the following DFA, give a regular expression for each of the languages indicated.
When the question refers to a machine “passing through” a given state, that means
entering and then exiting the state. Merely starting in a state or ending in it does not
count as “passing through.”
86 CHAPTER SEVEN: REGULAR EXPRESSIONS

q0 q1

b b b b

q2 q3

a. the language of strings that make the machine, if started in q0, end in q0, without passing
through q0, q1, q2, or q3
b. the language of strings that make the machine, if started in q0, end in q2, without passing
through q0, q1, q2, or q3
c. the language of strings that make the machine, if started in q2, end in q0, without passing
through q0, q1, q2, or q3
d. the language of strings that make the machine, if started in q2, end in q2, without passing
through q0 or q1
e. the language of strings that make the machine, if started in q2, end in q0, without passing
through q0 or q1
f. the language of strings that make the machine, if started in q0, end in q2, without passing
through q0 or q1
g. the language of strings that make the machine, if started in q1, end in q1, without passing
through q0 or q1
h. the language of strings that make the machine, if started in q1, end in q2, without passing
through q0 or q1
i. the language of strings that make the machine, if started in q0, end in q0, without passing
through q0 or q1
j. the language of strings that make the machine, if started in q1, end in q1, without passing
through q0
EXERCISE 4
(This exercise refers to material from Appendix A.) Let M be this NFA:
a,b

a
q0 q1
 EXERCISES 87
Give the simplest regular expression you can think of for each of the following internal
languages of M.
a. L(M, 0, 0, 2)
b. L(M, 0, 0, 1)
c. L(M, 0, 0, 0)
d. L(M, 0, 1, 2)
e. L(M, 0, 1, 1)
f. L(M, 0, 1, 0)
g. L(M, 1, 1, 2)
h. L(M, 1, 1, 1)
i. L(M, 1, 1, 0)
j. L(M, 1, 0, 2)
j. L(M, 1, 0, 1)
l. L(M, 1, 0, 0)

EXERCISE 5
(This exercise refers to material from Appendix A.) Let M be this NFA:
b

q1
a

a
q0

b
q2

Give the simplest regular expression you can think of for each of the following internal
languages of M. (Use the d2r construction if necessary.)
a. L(M,2,2,3)
b. L(M,2,2,2)
c. L(M,1,1,3)
d. L(M,1,1,2)
e. L(M,1,1,1)
f. L(M,0,0,3)
g. L(M,0,0,2)
h. L(M,0,0,1)
i. L(M,0,0,0)
88 CHAPTER SEVEN: REGULAR EXPRESSIONS

EXERCISE 6
(This exercise refers to material from Appendix A.) Let M be this DFA:
0 1

1 0

q0 q1 q2

1 0

Give the simplest regular expression you can think of for each of the following internal
languages of M. (Use the d2r construction if necessary.)
a. L(M, 2, 2, 3)
b. L(M, 2, 2, 2)
c. L(M, 1, 1, 3)
d. L(M, 1, 1, 2)
e. L(M, 1, 1, 1)
f. L(M, 0, 0, 3)
g. L(M, 0, 0, 2)
h. L(M, 0, 0, 1)
i. L(M, 0, 0, 0)
EXERCISE 7
(This exercise refers to material from Appendix A.) The following code is used to verify
the regular expression computed by d2r for the example in Section A.3:

public static void main(String args[]) {

NFA N = new NFA();
State q0 = N.getStartState();
State q1 = N.addState();
N.addX(q0,’a’,q0);
N.addX(q0,’b’,q1);
N.addX(q1,’b’,q1);
N.addX(q1,’a’,q0);
N.makeAccepting(q1);
System.out.println("final:" + N.d2r(0,1,0));
}

This code makes use of two classes, State and NFA. The NFA class represents an NFA.
When an NFA object is constructed, it initially contains only one state, the start state,
 EXERCISES 89
with no accepting states and no transitions. Additional states are added dynamically
by calling the addState method, which returns an object of the State class. Each
transition on a symbol is added dynamically by calling the addX method, giving the
source state, the symbol, and the target state. States are labeled as accepting by calling the
makeAccepting method. Finally, the d2r construction is performed by calling the
d2r method, passing it the integers i, j, and k. (States are considered to be numbered
in the order they were added to the NFA, with the start state at number 0.) The d2r
method returns a String. The code shown above creates an NFA object, initializes it as
the example NFA, and prints out the regular expression for the language it accepts.
Implement the State and NFA classes according to this specification, along with any
auxiliary classes you need. You need not implement ε-transitions in the NFA, and you
may represent ε in regular expressions using the symbol e. Test your classes with the
test code shown above, and check that your d2r returns the correct regular expression.
(Recall that, for readability, some of the parentheses in the regular expression in
Section A.3 were omitted. Your result may include more parentheses than shown there.)
EXERCISE 8
For any string x, define xR to be that same string reversed. The intuition is plain enough,
but to support proofs we’ll need a formal definition:
εR = ε
(ax)R = xRa, for any symbol a and string x
We’ll use the same notation to denote the set formed by reversing every string in a
language:
AR = {xR | x A}
Using these definitions, prove the following properties of reversal:
a. Prove that for any strings x and y, (xy)R = yRxR. Hint: Use induction on |x|.
b. Prove that for any string x, (xR)R = x. Hint: Use induction on |x| and the result from
Part a.
c. Prove that for any languages A and B, (A B)R = AR BR.
d. Prove that for any languages A and B, (AB)R = BRAR. You’ll need the result from
Part a.
e. Prove that for any language A, (A*)R = (AR)*. You’ll need the result from Part a.
f. Prove that the regular languages are closed for reversal. Hint: Using structural
induction, show that for every regular expression r, there is a regular expression
r' with L(r' ) = (L(r))R.
 8
CHA P TER

Regular-Expression
Applications
We have seen some of the implementation techniques related to
DFAs and NFAs. These important techniques are like tricks of the
programmer’s trade, normally hidden from the end user. Not so with
regular expressions; they are often visible to the end user and are part
of the user interface of a variety of useful software tools.

91
92 CHAPTER EIGHT: REGULAR-EXPRESSION APPLICATIONS

8.1 The egrep Tool

The Unix tool egrep searches a file for lines that contain a substring matching a specified
pattern and echoes all such lines to the standard output. For example, if a file named names
contains these lines:

fred
barney
wilma
betty

then this command after the % prompt searches the file for lines containing an a:

% egrep 'a' names

barney
wilma
%

In that example, egrep searched for lines containing a simple constant substring, but
the language of patterns understood by egrep can do much more. Various dialects of the
patterns understood by egrep are also used by many other tools. Unfortunately, these patterns
are often simply called regular expressions, though both in syntax and meaning they are a bit
different from the regular expressions we studied in the last chapter. To keep the two ideas
separate, this book refers to the patterns used by egrep and other tools using their common
nickname: regexps. Some of the special characters used in egrep’s regexp dialect are
* This symbol is like our Kleene star. For any regexp x, x* matches strings that are
concatenations of zero or more strings from the language specified by x.
| This symbol is like our +. For any regexps x and y, x|y matches strings that match
either x or y (or both).
() These symbols are used for grouping.
^ When this special symbol is at the start of the regexp, it allows the regexp to match
only at the start of the line.
$ When this special symbol is at the end of the regexp, it allows the regexp to match
only at the end of the line.
. This symbol matches any symbol (except the end-of-line marker).
For example, the regexp a.*y matches strings consisting of an a, followed by zero or
more other characters, followed by a y. We can search the names file for any line containing
such a string:
 8.1 THE EGREP TOOL 93
% egrep 'a.*y' names
barney
%

The regexp .(..)* matches strings consisting of an odd number of characters.

Searching the file names with this pattern produces this result:

% egrep '.(..)*' names

fred
barney
wilma
betty
%

Why did the search match all the lines, even those with an even number of characters? It did
this because egrep searches the file for all lines that contain a substring matching the specified
pattern—and any line with one or more characters contains a substring with an odd number
of characters. To match only odd-length lines, we must use the special symbols ^ and $:

% egrep '^.(..)*$' names

wilma
betty
%

One final example: Chapter 4 demonstrated a DFA-based, Java implementation of

a filter program that echoes those lines of an input file that are binary representations of
numbers that are divisible by three. In Section 7.5 a regular expression for this language was
derived: (0 + 1(01*0)*1)*. Using the corresponding regexp, egrep can do the same thing
our Java filter program did. If the file numbers contains the numbers zero through ten in
binary:

0
1
10
11
100
101
110
111
94 CHAPTER EIGHT: REGULAR-EXPRESSION APPLICATIONS

1000
1001
1010

then this egrep command selects those that are divisible by three:

% egrep '^(0|1(01*0)*1)*$' numbers

0
11
110
1001
%

8.2 Nonregular Regexps

In some ways regexps have evolved beyond the formal regular expressions from which they
developed. Many regexp dialects can define more than just the regular languages.
For example, special parentheses in some regexp dialects can do more than just group
subexpressions—they can capture the text that was matched with that subexpression so that
it can be referred to later. The grep utility (but not most versions of egrep) uses \( and \) as
capturing parentheses and uses \n as a regexp that matches the same text captured by the nth
previous capturing left parenthesis. Thus the regexp ^\(.*\)\1$ matches any line that
consists of a string followed immediately by a repetition of that same string. If the file named
test contains these lines:

abaaba
ababa
abbbabbb
abbaabb

then this grep command selects those lines that consist of repeated strings:

% grep '^\(.*\)\1$' test

abaaba
abbbabbb
%

The formal language corresponding to that example is {xx | x *}. This is not
something that can be defined with plain regular expressions. A useful intuition about regular
expressions is that, like DFAs, they can do only what you could implement on a computer
 8.3 IMPLEMENTING REGEXPS 95
using a fixed, finite amount of memory. Capturing parentheses clearly go beyond that limit,
since they must capture a string whose size is unbounded. To implement \(.*\)\1, you
would have to first store the string matched by .* somewhere, then test for a match with \1
by comparing against that stored string. Since the stored string can be arbitrarily large, this
cannot be done using a fixed, finite memory. That is, of course, just an informal argument.
We will prove formally in later chapters that {xx | x *} is not a regular language and no
mere regular expression can define it.

8.3 Implementing Regexps

There are many other tools that use regexp dialects: text-processing tools like emacs,
vi, and sed; compiler construction tools like lex; and programming languages like Perl,
Ruby, Python, PHP, Java, and the .NET languages. How do these tools implement regexp
matching? At a high level, the answer is simple: they convert the regexp into an NFA or
DFA, then simulate that automaton on each string to be tested. This roughly follows the
constructions already given in this book. Some systems convert the regexp into an NFA, and
then use that NFA to test each string for a match. As we’ve seen, an NFA is more expensive
than a DFA to simulate, but it is easier to extend for nonregular constructs like capturing
parentheses. Other systems convert the regexp into an NFA and then convert that to a DFA.
As we’ve seen, a DFA can be implemented more efficiently, but it is harder to extend beyond
the regular languages and may also be much larger than the corresponding NFA. Of course,
many optimizations and implementation tricks are possible; this book has described the
constructions without giving much consideration to efficiency.
The usual operation of a DFA makes a sequence of moves reading the entire input
string. Acceptance of the string is judged only at the end, after the whole string has been
read. Regexp matching tools generally do something a little different: they search a string
for a substring that matches the given regexp. This can still be done with a DFA, but the
implementation details are slightly different. Imagine starting a DFA at the first character
of the string. If it reaches a nonaccepting trap state or hits the end of the string without
ever reaching an accepting state, that indicates that no substring starting at the first position
matches. In that case the DFA must run again starting from the second position, then the
third, and so on, checking each position in the string for a matching substring.
If the DFA ever enters an accepting state—even if the end of the string has not been
reached—that indicates that a matching substring has been found. Since we started
at the left end of the string and worked forward, it is a leftmost match. There is still a
problem, however, since there may be more than one match starting at the same place. For
example, the string abb contains three substrings that match the regexp ab*—a, ab, and
abb—and they all start at the same place. Should the implementation report as soon as
that first accepting state is reached (after the a), or should it continue to operate and enter
accepting states again (after the first b and after the second b)? Different tools treat this
96 CHAPTER EIGHT: REGULAR-EXPRESSION APPLICATIONS

situation differently, but many tools (like lex) are required to find the longest of the leftmost
matches first, which would be the string abb. In that case, the DFA-based implementation
must continue processing the string, always remembering the last accepting state that was
entered and the position in the string that went with it. Then, as soon as the DFA enters a
nonaccepting trap state or encounters the end of the string, it can report the longest leftmost
match that was found.
Similar accommodations must be made by NFA-based automata. In particular, when
an implementation using backtracking finds a match, it cannot necessarily stop there. If the
longest match is required, it must remember the match and continue, exploring all paths
through the NFA to make sure that the longest match is found.

8.4 Regular Expressions in Java

There is a predefined package in Java, java.util.regex, that lets you work with
regular expressions. Here’s a simple example of how it works:

import java.io.*;
import java.util.regex.*;

/**
* A Java application to demonstrate the Java package
* java.util.regex. We take one command-line argument,
* which is treated as a regexp and compiled into a
* Pattern. We then use that Pattern to filter the
* standard input, echoing to standard output only
* those lines that match the Pattern.
*/
class RegexFilter {
public static void main(String[] args)
throws IOException {

Pattern p = Pattern.compile(args[0]); // the regexp

BufferedReader in = // standard input
new BufferedReader(new InputStreamReader(System.in));

// read and echo lines until EOF

String s = in.readLine();
while (s!=null) {
Matcher m = p.matcher(s);
 8.5 THE LEX TOOL 97
if (m.matches()) System.out.println(s);
s = in.readLine();
}
}
}

This application is structured like Mod3Filter from Chapter 4. But now, instead
of using a fixed DFA to test each line, it takes a parameter from the command line, treats
it as a regexp, and uses it to construct a Pattern object. A Pattern is something like
a representation of an NFA: it is a compiled version of the regexp, ready to be given an
input string to test. In the main loop, the application tests each input line by creating a
Matcher, an object that represents the NFA of the Pattern along with a particular input
string and current state. A Matcher object can do many things, such as finding individual
matches within a string and reporting their locations. But here, we just use it to test the
entire string to see whether the string matches the pattern. Using this application, we could
do our divisible-by-three filtering with this command:

% java RegexFilter '^(0|1(01*0)*1)*$' < numbers

The regex package also provides a single, static method that combines all those
steps, so you don’t have to keep track of separate Pattern and Matcher objects. To test
whether a String s is in the language defined by a regexp String r, you can just
evaluate Pattern.matches(r,s). This is easier for the programmer and makes sense
if you are going to use the regexp only once. But if (as in our example above) you need to use
the same regexp repeatedly, the first technique is more efficient; it compiles the regexp into a
Pattern that can be used repeatedly.
Scripting languages—Perl, Python, PHP, Tcl, Ruby, JavaScript, VBScript, and so on—
often have extra support for programming with regular expressions. In Perl, for example,
operations involving regular expressions don’t look like method calls; regular expressions are
used so often that the language provides a special-purpose syntax that is much more compact.
The Perl expression $x =~ m/a*b/ evaluates to true if and only if the string variable $x
matches the regexp a*b.

8.5 The lex Tool

Tools like egrep convert a regexp into an automaton, simulate the automaton, and then
discard it. That makes sense for applications that are typically run with a different and
fairly short regexp each time. But other applications need to use the same regexp every time
they run. Compilers, for example, often use a collection of regexps to define the tokens—
keywords, constants, punctuation, and so on—in the programming language to be compiled.
98 CHAPTER EIGHT: REGULAR-EXPRESSION APPLICATIONS

These regexps do not change from one compilation to the next, and it would not make sense
to perform the conversion from regexp to automaton every time the compiler is run. Instead,
the automaton can be a fixed, preconstructed part of the compiler. Tools like lex help with
this; they convert regexps into high-level-language code that can be used as a fixed part of
any application.
The lex tool converts a collection of regexps into C code for a DFA. The input file to
lex has three sections: a definition section, a rules section, and a section of user subroutines.
Lex processes this input file and produces an output file containing DFA-based C code,
which can be a stand-alone application or can be used as part of a larger program. The three
sections are separated by lines consisting of two percent signs, so a lex input file looks like
this:

definition section
%%
rules section
%%
user subroutines

The definition section can include a variety of preliminary definitions, but for the simple
examples in this book this section is empty. The user subroutines section can include C
code, which is copied verbatim to the end of the lex output file. This can be used to define
C functions used from inside the rules section, but for the simple examples in this book this
section too is empty. Thus, in our examples, the lex input file has only the rules section, like
this:

%%
rules section
%%

The rules section is a list of regexps. Each regexp is followed by some C code, which is to
be executed whenever a match is found for the regexp. For example, this lex program prints
the line Found one. for each occurrence of the string abc in the input file and ignores all
other characters:

%%
abc {fprintf(yyout, "Found one.\n");}
.|\n {}
%%
 8.5 THE LEX TOOL 99
The lex program above contains two regexps. The first is abc; whenever that is found, the C
code {fprintf(yyout, "Found one.\n");} is executed. Here, fprintf is C’s
standard input/output function for printing to a file, and yyout is the output file currently
being used by the lex-generated code. So the program generated by lex reads its input and,
whenever it sees the string abc, prints the line Found one.
Code generated by lex copies any unmatched characters to the output. Our lex program
avoids this default behavior by including a second regexp, .|\n, which matches every
character. (The regexp . matches every character except the end-of-line marker, and the
regexp \n matches the end-of-line marker.) The C code associated with the second regexp
is the empty statement, meaning that no action is taken when this match is found. The lex-
generated code finds the longest match it can, so it won’t match a single character a using the
second rule if it is the start of an abc that matches the first rule.
If the lex program above is stored in a file named abc.l, we can build a program from
it with the following Unix commands:

% flex abc.l
% gcc lex.yy.c -o abc –ll
%

The first command uses flex (the Gnu implementation of lex) to compile the lex program
into DFA-based C code, which is stored in a file named lex.yy.c. The second command
then runs the C compiler on the flex-generated code; the –o abc tells it to put the
executable program in a file named abc, and –ll tells it to link with the special library for
lex. To test the resulting program, we make a file named abctest that contains these lines:

abc
aabbcc
abcabc

Now we run abc giving it abctest as input:

% abc < abctest

Found one.
Found one.
Found one.
%

It finds the three instances of the substring abc.

For one last example, let’s again implement the divisible-by-three filter, this time using
lex. This lex program uses the same regexp we used in the previous section with egrep. Any
100 CHAPTER EIGHT: REGULAR-EXPRESSION APPLICATIONS

line matching that regexp is echoed to the output; any unmatched characters are ignored.
(The lex variable yytext gives us a way to access the substring that matched the regexp.)

%%
^(0|1(01*0)*1)*$ {fprintf(yyout, "%s\n", yytext);}
.|\n {}
%%

This lex program can be compiled into a filter program, just like the Mod3Filter we
implemented in Java:

% flex mod3.l
% gcc lex.yy.c -o mod3 –ll
%

The result is a directly executable file that filters its input file, echoing the divisible-by-
three lines to the standard output. Here, as before, the file numbers contains the binary
representations of the numbers zero through ten.

% mod3 < numbers

0
11
110
1001
%

For simple applications like those above, the code produced by lex can be compiled as a
standalone program. For compilers and other large applications, the code produced by lex is
used as one of many source files from which the full application is compiled.

8.6 Further Reading

For more information about lex, see
Levine, John R., Tony Mason, and Doug Brown. lex & yacc. Sebastopol, CA:
O’Reilly & Associates, 1992.
For more information about regexp tools in general, an excellent resource is
Friedl, Jeffrey E. F. Mastering Regular Expressions. Sebastopol, CA: O’Reilly &
Associates, 2002.
 EXERCISES 101

Exercises
EXERCISE 1
Show an egrep command that reads the standard input and echoes only those lines that
are in the language L(01*0).
EXERCISE 2
Show an egrep command that reads the standard input and echoes only those lines over
the alphabet {a, b} that have an odd number of as.
EXERCISE 3
Show an egrep command that reads the standard input and echoes only those lines that
are binary representations of numbers that are divisible by four.
EXERCISE 4
Show an egrep command that reads the standard input and echoes only those lines that
contain somewhere a decimal number in this form: one or more decimal digits, followed
by a decimal point, followed by one or more additional digits. You will need to find
more information about egrep’s regexp syntax to solve this problem correctly, and still
more to solve it elegantly.
EXERCISE 5
The lexical structure of Java programs includes the following three elements:
• Single-line comments, using //, as in these examples:
int x = 1; // NOT starting from 0 here
// s = "Hello";
• Literals of the String type, such as
"hello"
"Comments start with /*"
"He said, \"Hello.\""
• Literals of the char type, such as
'a'
'\n'
'\''
'\34'
If you are not a Java expert, you will need to refer to Java documentation for precise
definitions of these three elements. You’ll also need to learn more about egrep’s regexp
syntax to solve these problems.
a. Show an egrep command that reads the standard input and echoes only those lines
that contain single-line comments. You may assume that no String or char
literals and no block comments using /* and */ are present.
102 CHAPTER EIGHT: REGULAR-EXPRESSION APPLICATIONS

b. Show an egrep command that reads the standard input and echoes only those lines
that contain valid String literals. You may assume that no comments or char
literals are present.
c. Show an egrep command that reads the standard input and echoes only those lines
that contain valid char literals. You may assume that no comments or String
literals are present.

EXERCISE 6
The lexical structure of Java programs includes the following four elements:
• Single-line comments, as described in the previous exercise.
• Literals of the String type, as described in the previous exercise.
• Literals of the char type, as described in the previous exercise.
• Traditional comments using /* and */, as in these examples:

/* fix
this */
for (int i = 1; /* fix this */ i<100; i++)
/* hi /* // "hello" */

If you are not a Java expert, you will need to refer to Java documentation for precise
definitions of these elements. You’ll also need to learn more about the Java regexp syntax
and about the methods of the Matcher and Pattern classes. One hint: A common
beginner’s mistake when using the regex package is to forget to double up the
backslashes. Because a regexp is expressed as a Java String literal, a regexp like \s (the
character class for white space) must be written in the string as "\\s".
a. Write a Java application that reads a Java source file from standard input. It examines
the file one line at a time, identifies single-line comments by using the regex
package, and echoes only those comments (not necessarily the whole line) to
standard output. You may assume that none of the other elements described above is
present.
b. Write a Java application that reads a Java source file from standard input. It examines
the file one line at a time, identifies String literals by using the regex package,
and echoes only those literals to standard output, one on each line of output. (Note
that the input file may contain more than one String literal on a line; you must
find them all.) You may assume that none of the other elements described above
is present. Your program need not be well behaved if the input file contains errors
(such as unclosed String literals).
c. Write a Java application that reads a Java source file from standard input. It examines
the file one line at a time, identifies char literals by using the regex package,
 EXERCISES 103
and echoes only those literals to standard output, one on each line of output. (Note
that the input file may contain more than one char literal on a line; you must
find them all.) You may assume that none of the other elements described above
is present. Your program need not be well behaved if the input file contains errors
(such as unclosed char literals).
d. Write a Java application that takes the name of a Java source file as a command-line
argument. It reads the entire file into a string, identifies traditional comments by
using the regex package, and echoes only those comments to standard output. You
may assume that none of the other elements described above is present. Note that
these may be multiline comments, so it is not sufficient to process each input line
in isolation. Your program need not be well behaved if the input file contains errors
(such as unclosed traditional comments).
e. Write a Java application that takes the name of a Java source file as a command-line
argument. It reads the entire file into a string, identifies all four elements described
above by using the regex package, and echoes only them to standard output. Your
output should identify which of the four parts each element is. For example, if the
file Test.java contains this text:

char /* type is char */ p = '\''; // p is the quote

String s = "\"It's nice to be here,\" he said.";
/* Commented out:
String x = "Normal comments start with /*"; */
String w = "Single-line comments start with //";

the result should look like this:

% java AllFour Test.java

Traditional comment: /* type is char */
Char literal: '\''
Single-line comment: // p is the quote
String literal: "\"It's nice to be here,\" he said."
Traditional comment: /* Commented out:
String x = "Normal comments start with /*"; */
String literal: "Single-line comments start with //"

Your program need not be well behaved if the input file contains errors. Hint: it may
sound difficult to handle comments that contain strings, strings that contain comments,
and so on, but it isn’t. With minor modifications, you can simply combine the four
regexps from the previous four parts into one regexp. If your final regexp has four sets
104 CHAPTER EIGHT: REGULAR-EXPRESSION APPLICATIONS

of capturing parentheses, one for each of the four elements, you’ll be able to identify
which of the four was responsible for each match by using the group method of the
Matcher class.
EXERCISE 7
Write a Java application that examines a text file, looking for words that occur twice in
a row, separated by white space. (Define a “word” to be a string of one or more letters,
upper- or lowercase; define “white space” as in the Java regexp \s character class.) Your
application should take the file name as input on the command line. For each duplicated
word found, your application should print the word, the line number (counting the
first line as line 1), and the character position of the start of the first of the two copies
(counting the first character as character 1).
For example, if the file speech.txt contains this:

Four score and seven seven years ago our fathers

brought forth on on this continent, a new nation,
conceived in Liberty, and dedicated to the
the proposition that all men are created equal.

then the result should look like this:

% java Doubles speech.txt

Line 1, char 16: seven
Line 2, char 15: on
Line 3, char 40: the

In your application, read the entire file into a string, then use the regex package to find
the duplicated words. (As the example shows, you must catch pairs that are split from
one line to the next.) You’ll need to learn more about the Java regexp syntax and about
the methods of the Matcher and Pattern classes. Be careful that your application
does not find partial word pairs, like “ago our” in the example above or “equal equality.”
 9
CHA P TER

Advanced Topics in
Regular Languages
There are many more things to learn about finite automata than
are covered in this book. There are many variations with interesting
applications, and there is a large body of theory. Especially interesting,
but beyond the scope of this book, are the various algebras that arise
around finite automata. This chapter gives just a taste of some of
these advanced topics.

105
106 CHAPTER NINE: ADVANCED TOPICS IN REGULAR LANGUAGES

9.1 DFA Minimization

Given a DFA, can we find one with fewer states that accepts the same language? Can we
find the smallest DFA for a given language? Is the smallest DFA for a given language unique,
or might there be more than one? These are important questions for efficiency as well as
elegance.
Unreachable states, like some of those introduced by the subset construction, can
obviously be eliminated without affecting the language accepted. Even the reachable states of
a DFA can sometimes be combined or eliminated. For example, consider this DFA:
a

q0 a q1 a q2

b b
b

q3 q4

a,b a,b

This machine has two “trap” states, q3 and q4. When in state q3, the machine will
ultimately reject the string, no matter what the unread part contains. State q4 has exactly
the same property, so it is equivalent to q3 in an important sense and can be merged with q3
without changing the language accepted by the machine. So then we have
a

q0 a q1 a q2

b b
b

a,b
 9.1 DFA MINIMIZATION 107
The states q1 and q2 are equivalent in that same sense. When in state q2, the machine will
accept if and only if the rest of the string consists of zero or more as. State q1 has exactly the
same property, so it is equivalent to q2 and can be merged with it. We then have
a

q0 a

b
b

a,b

The result is a minimum-state DFA for the language {xay | x {b}* and y {a}*}.
That example used an important idea about the equivalence of two states. Informally, we
said that two states were equivalent when the machine’s future decision, after any remaining
input, was going to be the same from either state. Formally, let’s define a little language
L(M, q) for each state q, which is the language that would be accepted by M if q were used as
the start state:

L(M, q) = {x * | *(q, x) F}

Now we can formally define our idea of the equivalence of two states: “q is equivalent to r”
means L(M, q) = L(M, r). In our original DFA above, we had
L(M, q0) = {xay | x {b}* and y {a}*}
L(M, q1) = {x | x {a}*}
L(M, q2) = {x | x {a}*}
L(M, q3) = {}
L(M, q4) = {}
Thus q1 was equivalent to q2, and q3 was equivalent to q4.
A general procedure for minimizing DFAs is this:
1. Eliminate states that are not reachable from the start state.
2. Combine all the equivalent states, so that no two remaining states are equivalent to
each other.
108 CHAPTER NINE: ADVANCED TOPICS IN REGULAR LANGUAGES

Formally, Step 2 can be described as the construction of a new DFA whose states are
the equivalence classes of the states of the original DFA. This is sometimes called
the quotient construction. But we will not give a formal definition of the quotient
construction. Instead, we will just state without proof the important property of the
whole minimization procedure:

Theorem 9.1: Every regular language has a unique minimum-state DFA, and no
matter what DFA for the language you start with, the minimization procedure
finds it.

The minimum-state DFA is “unique” in a structural sense. Mathematicians say that the
minimized DFA is unique up to isomorphism—in this case, unique except perhaps for the
names of the states, which could of course be changed without affecting the structure of
the DFA materially. Thus our minimization procedure is both safe and effective: safe, in
that it does not change the language accepted by the DFA; effective, in that it arrives at the
structurally unique, smallest DFA for that language.
As described above, our minimization procedure is adequate for simple exercises done
by hand, but it is not obvious how to write a program to do it. Is there an algorithm that
can efficiently detect equivalent states and so perform the minimization? The answer is yes.
The basic strategy is to start with a partition of the states into two classes, accepting and
nonaccepting, and then repeatedly divide these into smaller partitions as states are discovered
to be nonequivalent. The Further Reading section below has references for this algorithm.
Minimization results are weaker for NFAs. With NFAs, you can eliminate unreachable
states and combine states using an equivalence relation similar to the one we saw for DFAs.
But the resulting NFA is not necessarily a unique minimum-state NFA for the language.

9.2 Two-way Finite Automata

DFAs and NFAs read their input once, from left to right. One way to try to make finite
automata more powerful is to allow rereading. A two-way finite automaton treats the input
like a tape; the automaton has a read-only “head” that reads a symbol at the current position
on the input tape. On each transition, the automaton moves the head either to the left
or to the right. There are two-way deterministic finite automata (2DFAs) and two-way
nondeterministic finite automata (2NFAs).
A 2DFA can be pictured like this:

x1 x2 ... xn-1 xn

2DFA
 9.3 FINITE-STATE TRANSDUCERS 109
The picture above shows an input tape containing the input string x1x2 ... xn-1xn. The input
string is framed on the tape by special symbols that serve as left and right end-markers. Like
a DFA, a 2DFA defines state transitions based on the current symbol and state. But the
transition function has an expanded role, since it now returns two things: a new state and
a direction to move the head, either L for left or R for right. For example, (q, a) = (s, L)
says that if the 2DFA is in the state q and the head is currently reading the symbol a, it then
enters the state s and moves the head one place to the left. (On every transition the head
moves one place, either left or right.) The transition function cannot move the head left past
the left end-marker or right past the right end-marker.
In a DFA the end of the computation is reached when the last input symbol is read,
but a 2DFA needs some other way to tell when the computation is finished. So instead of a
set of accepting states, a 2DFA has a single accepting state t and a single rejecting state r. Its
computation is finished when it enters either of these two states, signaling that it has reached
its decision about the input string. (It can also get into an infinite loop, so it might never
reach a decision.)
We state without proof the important results about two-way finite automata:

Theorem 9.2.1: There is a 2DFA for language L if and only if L is regular.

Theorem 9.2.2: There is a 2NFA for language L if and only if L is regular.

So adding two-way reading to finite automata does not increase their power; they can still
recognize exactly the regular languages.
The interesting thing is that with just a little more tweaking we can get automata that
are much more powerful, as we will see in later chapters. Adding the ability to write the
tape as well as read it yields a kind of machine called a linear bounded automaton, which
can define far more than just the regular languages. Adding the ability to write and to move
unboundedly far past at least one of the end-markers produces a kind of machine called a
Turing machine, which is more powerful still.

9.3 Finite-state Transducers

The machines we have seen may be thought of as computing functions with an arbitrary
string as input and a single bit as output: either yes, the string is in the language, or no it is
not. A variety of applications of finite-state machines actually want more output than this,
which leads to the idea of finite-state transducers.
Here is an example of a deterministic finite-state transducer that produces output on
each transition:
110 CHAPTER NINE: ADVANCED TOPICS IN REGULAR LANGUAGES

#,0#
#,ε q0 q1 0,0

0,ε
1,0
1,ε
#,0#
0,1

q2

1,1

The alphabet for this machine is {0, 1, #}. It works like a DFA, except that on every
transition, the machine not only reads one input symbol, but also produces a string of zero or
more output symbols. Each transition is labeled with a pair a, x. The first element in the pair,
a, is the input symbol read when the transition is made; the second element in the pair, x, is
an output string generated when the transition is made. There is no accepting state, because
the machine is not trying to recognize a language—it is trying to transform input strings into
output strings.
Let’s see what this machine does given the input 10#11#. Starting in q0, initially no
input has been read and no output produced. The following chart shows how this changes
after each move:

Input Read by M State of M Output So Far

ε q0 ε
1 q2 ε
10 q1 1
10# q0 10#
10#1 q2 10#
10#11 q2 10#1
10#11# q0 10#10#

Given the input 10#11#, the machine produced the output 10#10#. In general, if the input
to this machine is a string of binary numbers, each terminated by one or more # symbols, the
output is a string of the same binary numbers rounded down to the nearest even number, each
terminated by a single # symbol.
 9.4 ADVANCED REGULAR EXPRESSIONS 111
As this example suggests, finite-state transducers can be used as signal processors. In
particular, they are used in a variety of natural-language processing, speech recognition,
and speech-synthesis applications. Finite-state transducers come in many flavors. Some are
deterministic, some nondeterministic; some associate an output string with each transition,
others with each state.

9.4 Advanced Regular Expressions

When you write regular expressions, you quickly discover that there are many correct
answers: many different regular expressions that actually define the same language. It is
useful to be able to compare two regular expressions and decide whether they are equivalent.
We have already seen a rough outline of how this can be done. You can convert your two
expressions into NFAs, using the construction of Appendix A; then convert the NFAs into
DFAs, using the subset construction; and then minimize the DFAs, using the techniques
of this chapter. The minimum-state DFA for a language is unique, so your original regular
expressions are equivalent if and only if your two minimized DFAs are identical. This is quite
an expensive procedure, because the constructions involved can produce very large DFAs
from comparatively small regular expressions. The problem of deciding regular-expression
equivalence belongs to a class of problems known as PSPACE-complete. Without going into the
definition now, we’ll just note that PSPACE-complete problems are generally believed (but not
proved) to require exponential time. Certainly the algorithm just outlined has that property;
deciding whether two regular expressions of length n are equivalent using that approach
takes, in the worst case, time at least proportional to 2n.
We can consider extending regular expressions in various ways. As we have seen, many
practical regexp languages already do this. But there are several regular-expression extensions
that have theoretical interest as well as practical value.
One simple extension is squaring. We use the basic three atomic expressions and the
basic three compound expressions, but then add one more kind of compound expression:
(r)2, denoting the same language as (r)(r). Obviously, this extension does not add the ability
to define new languages. But it does add the ability to express some regular languages much
more compactly. For example, consider the language {0n | n mod 64 = 0}. Using basic regular
expressions, we would have to write this as
(0000000000000000000000000000000000000000000000000000000000000000)*
With squaring, we could write it more compactly as (((((0000)2)2)2)2)*.
The question of whether two regular expressions are equivalent, when squaring is
allowed, becomes more difficult to answer. It belongs to a class of problems known as
EXSPACE-complete. Without going into the definition now, we’ll just note that EXSPACE-complete
problems require exponential memory space and at least exponential time and are generally
believed (but not proved) to require more than exponential time.
112 CHAPTER NINE: ADVANCED TOPICS IN REGULAR LANGUAGES

Another possible addition to regular expressions is complement. For this extension, we

start with basic regular expressions and then allow one extra kind of compound expression:
(r)C denotes the complement of a language (with respect to some fixed alphabet), so that
L((r)C) = {x *|x L(r)}
We know that basic regular expressions can define all the regular languages, and we know
that the regular languages are closed for complement. This means that our extension
does not add the ability to define new languages. But, once again, we get the ability to
define some languages much more compactly; and, once again, that makes the problem of
deciding the equivalence of two regular expressions harder—much harder. (To be precise,
the problem now requires NONELEMENTARY TIME. Complexity classes like PSPACE, EXPSPACE, and
NONELEMENTARY TIME are discussed further in Chapter 20.)
In a regular expression, one subexpression using the Kleene star may be nested inside
another. Informally, the star height of a regular expression is the depth of this nesting:
a + b has star height 0, (a + b)* has star height 1, (a* + b*)* has star height 2, and so on.
Technically, we can define the star height h(r) for any regular expression r inductively:
h(r) = 0 for any atomic regular expression r
h(r + s) = max(h(r), h(s))
h(rs) = max(h(r), h(s))
h((r)*) = h(r) + 1
h((r)2) = h(r)
h((r)C) = h(r)
It is often possible, and desirable, to simplify expressions in a way that reduces their star
height. For example, the expression (a* + b*)*, star height 2, defines the same language as
(a + b)*, star height 1. Attempts to reduce the star height of regular expressions lead to some
interesting questions:
• Is there some fixed star height, perhaps 1 or 2, that suffices for any regular language?
• Is there an algorithm that can take a regular expression and find the minimum
possible star height for any expression for that language?
For basic regular expressions, the answers to these questions are known: no, no fixed height
suffices for all regular languages; yes, there is an algorithm for finding the minimum star
height for a regular language. However, when complement is allowed, these questions are still
open. In particular, when complement is allowed, it is not known whether there is a regular
language that requires a star height greater than 1! This generalized star-height problem is one
of the most prominent open problems surrounding regular expressions.
 EXERCISES 113

9.5 Further Reading

There is a good introduction to the theory of DFA and NFA minimization in
Kozen, Dexter. Automata and Computability. New York: Springer-Verlag, 1997.
Kozen also has a good discussion of 2DFAs and 2NFAs, with proofs of their important
properties.
The basic algorithm that minimizes a DFA with the alphabet and the state set Q in
O(| | |Q|2) time is presented in
Hopcroft, John, Rajeev Motwani, and Jeffrey Ullman. Introduction to Automata
Theory, Languages, and Computation, 2nd ed. Menlo Park, CA: Addison-Wesley,
2001.
Using the fancier partitioning algorithm due to Hopcroft, a DFA can be minimized in
O(| | |Q| log |Q|) time. See:
Hopcroft, John. “An n log n algorithm for minimizing the states in a finite
automaton.” Published in The Theory of Machines and Computations, ed. Z.
Kohavi. New York: Academic Press, 1972, 189–96.
Finite-state transducers have fallen out of most textbooks on formal languages and automata,
even as they have grown as an active area of research. See, for example,
Mohri, Mehryar. “Finite-State Transducers in Language and Speech Processing.”
Computational Linguistics 23 (1997): 269-311.
The star-height problem is simple to describe, but very difficult to solve. Unfortunately,
most of the work in this area is inaccessible to most undergraduates. If you are intested in the
problem, and if your mathematical background includes abstract algebra, combinatorics, and
computational complexity, you should begin by studying formal languages from the algebraic
perspective, as in
Holcombe, W. M. L. Algebraic Automata Theory. Cambridge:
Cambridge University Press, 2004.

Exercises
EXERCISE 1
For each of the following DFAs, list the unreachable states if any, show L(M, q) for each
q Q, and construct the minimized DFA using the procedure of Section 9.1.
114 CHAPTER NINE: ADVANCED TOPICS IN REGULAR LANGUAGES

a.
0,1

0,1
q0 q1

0 1

q2

b.
a

a
q0 q1

b b
a

q2 q3
b

a,b

c.

a
q0 q1 q2

a a a
b
b b

q3 q4 q5
b

a,b b
 EXERCISES 115
EXERCISE 2
Write a Java class that implements the finite-state transducer example from Section 9.3.
Include a main method that tests it on input from the command line.
EXERCISE 3
Using regular expressions extended with complement, find an expression with a star
height of 0 for each of these languages. Assume that = {a, b}.
a. *
b. L((a + b)*aa(a + b)*)
c. {x * | x has no two as in a row}
d. L(b*)
e. For any two extended regular expressions r1 and r2 of star height 0, the language
L(r1) L(r2).
f. For any two extended regular expressions r1 and r2 of star height 0, the language
L(r1) - L(r2).
g. L((ab)*). Hint: Almost all of these are strings that begin with ab, end with another
ab, and have no two as or bs in a row.
 10
CHA P TER

Grammars

Grammar is another of those common words for which the study

of formal language introduces a precise technical definition. For
us, a grammar is a certain kind of collection of rules for building
strings. Like DFAs, NFAs, and regular expressions, grammars are
mechanisms for defining languages rigorously.
A simple restriction on the form of these grammars yields the special
class of right-linear grammars. The languages that can be defined
by right-linear grammars are exactly the regular languages. There
it is again!

117
118 CHAPTER TEN: GRAMMARS

10.1 A Grammar Example for English

Let’s start with an example of a grammar for a familiar language: English. An article can be
the word a or the. We use the symbol A for article and express our definition this way:
A a
A the
A noun can be the word dog, cat, or rat:
N dog
N cat
N rat
A noun phrase is an article followed by a noun:
P AN
A verb can be the word loves, hates, or eats:
V loves
V hates
V eats
A sentence is a noun phrase, followed by a verb, followed by another noun phrase:
S PVP
Put them all together and you have a grammar that defines a small subset of unpunctuated
English. Call this grammar G1:
S PVP
P AN
V loves
V hates
V eats
A a
A the
N dog
N cat
N rat
How does G1 define a language? Think of each arrow as a rule that says how to modify
strings by substitution. These rules are called productions. A production of the form x y
says that wherever you see the substring x, you may substitute y. For example, if we are
 10.1 A GRAMMAR EXAMPLE FOR ENGLISH 119
given the string thedogeatsP, the production P AN permits us to replace the P with AN,
producing thedogeatsAN. By starting from S and following the productions of G1 repeatedly,
you can derive a variety of unpunctuated English sentences. For example:
S PVP ANVP theNVP thecatVP thecateatsP thecateatsAN
thecateatsaN thecateatsarat

S PVP ANVP aNVP adogVP adoglovesP adoglovesAN

adoglovestheN adoglovesthecat

S PVP ANVP theNVP thecatVP thecathatesP thecathatesAN

thecathatestheN thecathatesthedog
The language defined by G1 is the language of all the lowercase strings you can get by
applying the productions of G1, starting from S. The three examples above show that the
language of G1 includes the strings thecateatstherat, adoglovesthecat, and thecathatesthedog.
Often there is more than one place in a string where a production could be applied.
For example, if we are given the string PlovesP, there are two places where the production
P AN can be used. Replacing the leftmost P gives PlovesP ANlovesP; replacing the
rightmost one gives PlovesP PlovesAN. Because of this, there is often more than one way to
produce the same final string:
S PVP ANVP aNVP adogVP adoglovesP adoglovesAN
adoglovestheN adoglovesthecat

S PVP PVAN PVAcat PVthecat Plovesthecat ANlovesthecat

Adoglovesthecat adoglovesthecat
These two different paths both produce the same string, adoglovesthecat.
When a grammar contains more than one production with the same left-hand side, those
productions can be written in a compressed form, with the common left-hand side first,
then the arrow, then the list of alternatives for the right-hand side, separated by vertical lines.
Using this notation, our G1 looks like this:
S PVP
P AN
V loves | hates | eats
A a | the
N dog | cat | rat
This makes the grammar easier to read, but it is merely shorthand. G1 is not changed by
being written this way and still has a total of ten separate productions.
120 CHAPTER TEN: GRAMMARS

For a more abstract example, consider this grammar, which we’ll call G2:
S aS
S X
X bX
X ε
The final production for X says that an X may be replaced by the empty string, so that for
example the string abbX can become the string abb. Written in the more compact way, G2 is
S aS | X
X bX | ε
Here are some derivations using G2:
S aS aX a
S X bX b

S aS aX abX abbX abb

S aS aaS aaaS aaaX aaabX aaabbX aaabb
All derivations for G2 follow a simple pattern: first they use S aS zero or more times, then
they use S X once, then they use X bX zero or more times, and then they use X ε
once. So the string that is generated always consists of zero or more as followed by zero or
more bs. Thus, the language generated by G2 is a regular language: L(a*b*).
In all our examples, we have used productions with a single uppercase letter on the left-
hand side. Not all grammars are this simple. Productions can have any nonempty string on
the left-hand side. Exactly the same mechanism of substitution still applies; for example,
the production Sb bS says that wherever you have the substring Sb, you can substitute
bS. Productions of this kind can be very powerful, and we’ll explore them briefly in a later
chapter. But for now, all our examples and exercises will continue to use productions with a
single nonterminal symbol on the left-hand side.

10.2 The 4-Tuple

Up to this point, we have followed the convention of using S as the start symbol, with
derivations from there ending only when the resulting string is fully lowercase. This special
treatment for the Latin alphabet with its uppercase and lowercase letters is not sufficiently
general for a formal treatment of grammars. To switch to a formal treatment, we have to
include definitions of two separate sets of symbols: the terminal symbols (corresponding to
our lowercase letters) and the nonterminal symbols (corresponding to our uppercase letters).
Counting these symbol sets, a grammar has four key parts, so we describe it mathematically
as a 4-tuple.
 10.3 THE LANGUAGE GENERATED BY A GRAMMAR 121

A grammar G is a 4-tuple G = (V, , S, P), where

V is an alphabet, the nonterminal alphabet
is another alphabet, the terminal alphabet, disjoint from V
S V is the start symbol
P is a finite set of productions, each of the form x y, where
x and y are strings over V and x ε.

The first part of the grammar is an alphabet, the nonterminal alphabet V. The
nonterminal alphabet contains the symbols we have been writing as uppercase letters—the
symbols that are used in derivations but do not occur in the string that is finally derived. The
second part of a grammar is the terminal alphabet. We use the same symbol that we have
used throughout this book for the alphabet of current interest; these are the symbols that we
have been writing as lowercase letters. The sets and V are disjoint, meaning that no symbol
occurs in both. The third part of the grammar is a special nonterminal symbol S, the start
symbol. The final part of a grammar is the set of productions. Each production is of the form
x y, where x and y are both strings over V and x is not permitted to be the empty
string.
For example, consider our previous grammar for the language L(a*b*):
S aS | X
X bX | ε
Formally, the grammar is G = (V, , S, P), where V = {S, X }, = {a, b}, and the set P of
productions is
{S aS, S X, X bX, X ε}
A grammar is a 4-tuple: it is a mathematical structure with four parts given in order. The
names given to those four parts in the definition above—V, , S, and P—are conventional,
but the grammar G could just as well have been defined as G = ({S, X }, {a, b }, S, {S aS,
S X, X bX, X ε}), without naming all the parts. The important thing is to specify the
four parts in the required order.

10.3 The Language Generated by a Grammar

For DFAs, we defined a one-step transition function and then used it to define the zero-
or-more-step extended transition function *; that let us define the language of a DFA. For
NFAs, we defined a one-step relation , then extended it to a zero-or-more-step relation *;
that let us define the language of an NFA. For grammars we will do something similar: we
will first define a one-step derivation relation and then use it to define a zero-or-more-step
extended derivation relation *.
The following four definitions are all made with respect to some fixed grammar
G = (V, , S, P). Each production in the grammar specifies a possible substring substitution;
122 CHAPTER TEN: GRAMMARS

the production x y says that wherever you see the substring x, you may substitute y. When
a string w can be transformed into a string z in this way, we write w z. (This is read as
“w derives z.”) Formally:

is a relation on strings, with w z if and only if there exist strings

u, x, y, and v over V, with w = uxv, z = uyv, and (x y) P.

A sequence of -related strings, x0 x1 ... xn, is called a derivation:

An n-step derivation, for any n 0, is a sequence of strings x0 through xn such

that for all 0 i < n, xi xi+1.

Using n-step derivations, we can define the extended derivation relation *:

* is a relation on strings, with w * z if and only if there is a derivation of

zero or more steps that starts with w and ends with z.

Notice here that * is reflexive: for any string x, x * x by a zero-step derivation. Using the
* relation, we can define the language generated by G:

The language generated by the grammar G is L(G ) = {x *|S * x}

Notice here the restriction that x *. The intermediate strings in a derivation can involve
both terminal and nonterminal symbols, but only the fully terminal strings derivable from
the start symbol are in the language generated by the grammar.
All four of the preceding definitions were made with respect to some fixed grammar
G = (V, , S, P). Clearly, different grammars produce different and * relations. In those
rare cases when we need to work with two or more grammars at once, we use subscripts to
identify the relations of each: G , H , and so on.

10.4 Every Regular Language Has a Grammar

In exercises and examples we have seen many regular languages generated by grammars.
Is there a grammar for every regular language? The answer is yes, and the proof is by
construction. We will show how to take any NFA M and build a grammar G for which
L(M) = L(G).
First, an example of this construction. Consider the language L(a*bc*). This is an
NFA for the language:
 10.4 EVERY REGULAR LANGUAGE HAS A GRAMMAR 123
a c

b ε
S R T

We will construct a grammar that generates the same language. Not only will it generate the
same language, but its derivations will exactly mimic the behavior of the NFA. To make the
construction clearer, the states in the NFA are named S, R, and T, instead of the usual q0, q1,
and q2. That’s because in our construction, each state of the NFA will become a nonterminal
symbol in the grammar, and the start state of the NFA will become the start symbol in the
grammar. That settles the two alphabets and the start state for the new grammar; all that
remains is to construct the set of productions.
For each possible transition Y (X, z) in the NFA (where z is any single symbol from
or z = ε), we add a production X zY to the grammar. For our example this is

Transition of M Production in G
(S, a) = {S } S aS
(S, b) = {R } S bR
(R, c) = {R } R cR
(R, ε) = {T } R T

In addition, for each accepting state in the NFA, the grammar contains an ε-production for
that nonterminal. For our example this is
Accepting State of M Production in G
T T ε

So the complete grammar constructed for this example is

S aS
S bR
R cR
R T
T ε
To see how this G simulates M, let’s compare the behavior of M as it accepts the string
abc with the behavior of G as it generates the same string. Here are the accepting sequence of
instantaneous descriptions of M and the derivation of G, side by side:
124 CHAPTER TEN: GRAMMARS

(S, abc) (S, bc) (R, c) (R, ε) (T, ε)

S aS abR abcR abcT abc
M starts in the state S; G starts with the start symbol S. After reading the first symbol a, M
is in the state S; after the first derivation step G has produced the string aS. After reading the
first two symbols, ab, M is in the state R; after the first two derivation steps G has produced
the string abR. After the first three symbols, abc, M is in the state R; after the first three
derivation steps G has produced the string abcR. As its fourth move, M makes an ε-transition
to state T; after the first four derivation steps G has produced the string abcT. Now M has
finished reading the string abc and ends in an accepting state; G applies a final derivation step
using T ε, deriving the fully terminal string abc. Except for this final derivation step, there
is a simple relation between the NFA and the grammar: every time the NFA reads a symbol,
the grammar generates that symbol. The NFA can be in the state Y after reading a string x if
and only if the grammar can derive the string xY.
That’s how the general construction will work. Every derivation from S of a fully
terminal string has the same general form:
S z1R1 z1z2R2 z1z2z3R3 ... z1...znRn z1...zn
At every step but the last, zi (which is a single symbol or the empty string) is added to the
right-hand side of the growing terminal string, along with a rightmost nonterminal symbol.
In the last step, one of the ε-productions is used to eliminate the nonterminal symbol. This
exactly traces a possible behavior of the NFA when it reads the same string.

Theorem 10.1: Every regular language is generated by some grammar.

Proof: By construction. Every regular language is accepted by some
NFA, so let M = (Q, , , S, F ) be any NFA. (For convenience, we have named
the start state S, instead of the more common q0.) Let G be the grammar
G = (Q, , S, P), using M’s set of states as the nonterminal alphabet, M’s
alphabet as the terminal alphabet, and M’s start state S as the start symbol.
The set P of productions is defined as follows: wherever M has Y (X, ), P
contains the production X Y, and for each X F, P contains X ε.
By construction, G has X Y whenever M has (X, ) (Y, ε). We can
extend this property to all strings z *: G has a derivation X * zY if and
only if M has (X, z) * (Y, ε). (This can be proved by induction on the length
of the derivation.) By construction, the grammar has Y ε whenever the NFA
has Y F. So for all strings z *, *(S, z) contains at least one element of F if
and only if S * z. Thus L(M) = L(G).
 10.5 RIGHT-LINEAR GRAMMARS 125

10.5 Right-Linear Grammars

The construction of Theorem 10.1 generates productions of a particularly restricted form:
single-step form.

A grammar G = (V, , S, P ) is single step if and only if every production in P is

in one of these two forms:
X Y
X ε
where X V, Y V, and {ε}.

If you’re given a grammar that happens to be single step, you can easily reverse the
construction and build a corresponding NFA. For instance, here’s a single-step grammar that
generates the language L(ab*a):
S aR
R bR
R aT
T ε
We can make an NFA with three states, S, R, and T. The first three productions correspond
to three transitions in the NFA diagram, and the final production tells us to make T an
accepting state. The resulting NFA is
b

a a
S R T

So it is easy enough to convert a single-step grammar to an NFA. Even if the productions

are not of the right form, it is sometimes possible to massage them until they are. Here’s a
grammar that does not follow our restrictions on production forms—a grammar for the
language {aba}:
S abR
R a
The production S abR does not qualify because it has the string ab instead of just a single
symbol before the final nonterminal R. But this is easy to correct; we can just substitute
the two productions S aX and X bR, both of which are single step. The production
R a does not qualify either, but again, it is easy to correct; we can just substitute the two
126 CHAPTER TEN: GRAMMARS

productions R aY and Y ε, both of which are single step. The resulting single-step
grammar is
S aX
X bR
R aY
Y ε
This can then be converted into an NFA using our construction:

a b a
S X R Y

The grammar we started with this time, although not quite in the form we wanted, was
still fairly close; every production had a single nonterminal as its left-hand side and a string
containing at most one nonterminal as its right-hand side. Moreover, that nonterminal on
the right-hand side was always the rightmost symbol in the string. A grammar that follows
this restriction is called a right-linear grammar.

A grammar G = (V, , S, P ) is right linear if and only if every production in P is

in one of these two forms:
X zY
X z
where X V, Y V, and z *.

Every right-linear grammar generates a regular language and can be converted fairly easily
into an NFA that accepts that language. As in the example above, the trick is to first rewrite
the grammar to make it single step. This is always possible.

Lemma 10.1: Every right-linear grammar G is equivalent to some single-step

grammar G'.
Proof: By construction. Let G = (V, , S, P ) be any right-linear grammar. Each
production in P is of the form X z1...zn , where V or = ε. For each
such production let P' contain these n + 1 productions:

X z1K1
K1 z2K2
...
 10.6 EVERY RIGHT-LINEAR GRAMMAR GENERATES A REGULAR LANGUAGE 127
Kn-1 znKn
Kn
where each Ki is a new nonterminal symbol. Now let G' = (V', , S, P'), where
V' is the set of nonterminals used in the new productions. The new grammar G'
is single step; it remains to be proven that L(G) = L(G' ). This is straightforward.
Any derivation of a terminal string in G can be converted to a derivation of
the same terminal string in G' simply by replacing each step that used a rule
X z1...zn with the n + 1 corresponding steps in G' and vice versa.

The construction given above makes more new productions than strictly necessary. For
example, it converts a production A bC into the two productions A bK1 and
K1 C. Conversion is unnecessary in that case, of course, since the original production
was already in the required form. But it doesn’t hurt, and it makes the proof simpler to use a
single, general transformation for every production in the original grammar.

10.6 Every Right-Linear Grammar Generates a Regular Language

Using the techniques just introduced, we can take any right-linear grammar and convert it to
an NFA. This shows that the language generated by a right-linear grammar is always a regular
language.

Theorem 10.2: For every right-linear grammar G, L(G ) is regular.

Proof: By construction, using Lemma 10.1 and the reverse of the construction
of Theorem 10.1.

Incidentally, there is parallel definition for left-linear grammars:

A grammar G = (V, , S, P) is left linear if and only if all productions in P are of

one of these two forms:
X Yz or
X z,
where X V, Y V, and z *.

With a little more work, one can show that the language generated by a left-linear grammar
is also always a regular language. A grammar that is either left linear or right linear is
sometimes called a regular grammar.
A grammar does not have to be regular to generate a regular language. Consider this
grammar:
S aSaa | ε
128 CHAPTER TEN: GRAMMARS

This is neither right linear nor left linear, yet the language it generates is a simple regular
language: L((aaa)*). If a grammar is regular, you can conclude that the language it generates
is regular. But if a grammar is not regular, you can’t conclude anything about the language it
generates; the language may still be regular.
All of this leads to the next important question: are there grammars that generate
languages that are not regular? This is answered in the next chapter.

Exercises
EXERCISE 1
These questions concern grammar G1 from section 10.1:
a. Show a derivation from S for the string thedogeatsthecat.
b. List all the lowercase strings that can be derived from P.
c. How many lowercase strings are in the whole language defined by G1—the language
of all lowercase strings that can be derived from S? Show your computation.
EXERCISE 2
Modify grammar G1 from Section 10.1 to make it allow at most one of the adjectives
red, little, black, or cute, as part of any noun phrase. For example, your grammar should
generate the sentence thereddoglovestheblackcat. Make sure it also generates all the original
sentences, without adjectives.
EXERCISE 3
These questions concern grammar G2 from Section 10.1:
c. Show a derivation from S for the string bbb.
d. Show a derivation from S for the string aabb.
e. Show a derivation from S for the empty string.
EXERCISE 4
Give a regular expression for the language generated by each of these grammars.
a. S abS | ε
b. S aS | aA (Hint: This one is tricky—be careful!)
A aS | aA
c. S smellA | fishA
A y|ε
d. S aaSa | ε
EXERCISE 5
Give a grammar for each of the following languages. In each case, use S as the start
symbol.
a. L(a*)
b. L(aa*)
c. L(a*b*c*)
 EXERCISES 129
d. L((abc)*)
e. The set of all strings consisting of one or more digits, where each digit is one of the
symbols 0 through 9.
EXERCISE 6
Give a DFA that accepts the language generated by this grammar:
S A|B|C
A aB | ε
B bC | ε
C cA | ε
EXERCISE 7
State each of the following grammars formally, as a 4-tuple. (Assume, as usual, that
the terminal alphabet is the set of lowercase letters that appear in productions, the
nonterminal alphabet is the set of uppercase letters that appear in productions, and the
start symbol is S.)
a. S aS | b
b. S abS | A
A baA | ε
c. S A | B | C
A aB | ε
B bC | ε
C cA | ε
EXERCISE 8
According to the definition, what is the smallest possible grammar—with the smallest
possible set in each of the four parts? What language does it generate?
EXERCISE 9
Consider this grammar:
S aS | bbS | X
X cccX | ε
For each of the following, show all strings that can be derived from the given string in
one step. For example, for the string cS, your answer would be the three strings caS, cbbS,
and cX.
a. S
b. aXa
c. abc
d. ScS
e. aXbSc
130 CHAPTER TEN: GRAMMARS

EXERCISE 10
Consider this grammar:
S AB | ε
A aA | ε
B bB | ε
For each of the following pairs related by *, show a derivation. For instance, for the
pair S * a, one correct answer is the four-step derivation S AB aAB aB a.
a. aA * aaaA
b. aB * abbB
c. Bb * bb
d. S * ε
e. S * a
f. S * aabb
EXERCISE 11
Give a grammar for each of the following languages. You need not state it formally; just
list the productions, and use S as the start symbol.
a. The set of all strings consisting of zero or more as with a b after each one.
b. The set of all strings consisting of one or more as with a b after each one.
c. The set of all strings consisting of one or more as, with a b between each a and the
next. (There should be no b before the first or after the last a.)
d. The set of all strings consisting of zero or more as, with a b between each a and the
next. (There should be no b before the first or after the last a.)
e. The set of all strings consisting of an open bracket (the symbol [) followed by a
list of zero or more digits separated by commas, followed by a closing bracket (the
symbol ]). (A digit is one of the characters 0 through 9.)
EXERCISE 12
Using the construction of Theorem 10.1, make a right-linear grammar that generates the
language accepted by each of these NFAs.
a.
a b

a
 EXERCISES 131
b.
b b

c.

d.
a

ε
a

b
ε

e.
0,1

1 0,1
132 CHAPTER TEN: GRAMMARS

EXERCISE 13
Using the construction of Theorem 10.2, make an NFA that accepts the language
generated by each of these right-linear grammars.
a. S bS | aX
X aX | ε
b. S bS | X | ε
X aX | cY
Y ε
c. S bR | ε
R baR | ε
d. S aS | bbS | X
X cccX | d
EXERCISE 14
Given any single-step grammar G = (Q, , S, P), give a formal construction for an
NFA M with L(M) = L(G). (This is the reverse of the construction used to prove
Theorem 10.1.) You don’t have to show a proof that the languages are equal.
EXERCISE 15
Consider this grammar G:
S Sa | Sb | ε
Prove by induction on |x| that for any string x {a, b}*, S * Sx. Then use this result to
prove that L(G) = {a, b}*.
EXERCISE 16
Rewrite the proof of Theorem 10.1. This time, include a full inductive proof that G has
a derivation X * zY if and only if M has (X, z) * (Y, ε). Hint: It is easier to prove a
slightly stronger lemma, that G has an n-step derivation X * zY if and only if M has an
n-step sequence (X, z) * (Y, ε).
 11
CHA P TER

Nonregular
Languages
We have now encountered regular languages in several different
places. They are the languages that can be recognized by a DFA.
They are the languages that can be recognized by an NFA. They
are the languages that can be denoted by a regular expression. They
are the languages that can be generated by a right-linear grammar.
You might begin to wonder, are there any languages that are not
regular?
In this chapter, we will see that there are. There is a proof tool that
is often used to prove languages nonregular. It is called the pumping
lemma, and it describes an important property of all regular
languages. If you can show that a given language does not have this
property, you can conclude that it is not a regular language.

133
134 CHAPTER ELEVEN: NONREGULAR LANGUAGES

11.1 The Language {anbn}

Consider the language of strings that consist of any number of as followed by the same
number of bs. The language contains the strings ε, ab, aabb, aaabbb, and so on. In set-former
notation this language can be written simply as {anbn}.
It is simple enough to give a grammar for {anbn}:
S aSb | ε
Every derivation in this grammar of a fully terminal string has the same form:
S aSb a2Sb2 a3Sb3 ... anSbn anbn
At every step but the last, one a and one b are generated, ensuring that the number of as is
always the same as the number of bs. In the final derivation step, the ε production is used to
produce a fully terminal string anbn.
Is {anbn} a regular language? If it is, there is an NFA for it, so let’s try to develop one.
Here is an NFA for the subset {anbn | n 0}, which contains just the empty string:

Here is one for the larger subset {anbn | n 1}:

b
ε
 11.1 THE LANGUAGE {anbn} 135
Here is one for the larger subset {anbn | n 2}:

a a

b
ε

For the next larger subset {anbn | n 3}, we could use this:

a a a

b
ε

ε b

b
136 CHAPTER ELEVEN: NONREGULAR LANGUAGES

Clearly, this is not going to be a successful pattern on which to construct an NFA for the
whole language {anbn}, since for each larger value of n we are adding two more states. In
effect, we are using the states of the NFA to count how many as were seen, then to verify that
the same number of bs follows. This won’t work in general, since the F in NFA stands for
finite, but no finite number of states will be enough to count the unbounded n in {anbn}.
Of course, this failure does not prove that {anbn} is not regular, but it contains the germ
of the idea for a proof—the intuition that no fixed, finite number of states can be enough. A
formal proof follows below.

Theorem 11.1: {anbn} is not regular.

Proof: Let M = (Q, {a, b}, , q0, F ) be any DFA over the alphabet {a, b }; we
will show that L(M) {anbn}. Consider the behavior of M on an arbitrarily long
string of as. As it reads the string, M visits a sequence of states: first *(q0, ε),
then *(q0, a), then *(q0, aa), and so on. Eventually, since M has only finitely
many states, it must revisit a state; that is, there must be some i and j with i < j
for which *(q0, ai ) = *(q0, a j ). Now by appending b j to both strings, we see
that *(q0, aib j ) = *(q0, a jb j ). Thus M ends up in the same final state for both
aib j and ajb j. If that is an accepting state, M accepts both; if it is a rejecting
state, M rejects both. But this means that L(M ) {anbn}, since a jb j is in {anbn}
while aib j is not. Since we have shown that L(M ) {anbn} for any DFA M, we
conclude that {anbn} is not regular.

An interesting thing about this proof is how much one can infer about the behavior
of a completely unknown DFA M. Using only the fact that as a DFA M must have a finite
number of states, one can infer that there must be some i and j with i < j for which M
either accepts both a ib j and a jb j or rejects both. The basic insight is that with a sufficiently
long string you can force any DFA to repeat a state. This is the basis of a wide variety of
nonregularity proofs.

11.2 The Language {xxR}

Consider the language of strings that consist of any string over a given alphabet , followed
by a reversed copy of the same string. For = {a, b }, the language contains the strings ε, aa,
bb, abba, baab, aaaa, bbbb, and so on. For any string x, the notation x R refers to the reverse
of x; the language in question is just {xxR}.
It is simple enough to give a grammar for {xxR}. Here is one for = {a, b}:
S aSa | bSb | ε
Every derivation in this grammar of a fully terminal string has the same form:
 11.3 PUMPING 137
S x1Sx1R x2Sx2R x3Sx3R ... xnSxnR xnxnR
where, for all i, |xi| = i. At every step but the last, a matching pair of as or bs is generated,
ensuring that the terminal string to the left of S is always the reverse of the terminal string
to the right of S. In the final derivation step, the ε production is used to produce a fully
terminal string xnxnR.
After seeing the {anbn} example, you may already have doubts about whether {xxR} can be
a regular language. A finite state machine would have to use its current state to keep track of
the string x, then check that it is followed by a matching x R. Since there is no bound on the
length of x, this suggests that no finite number of states will suffice.

Theorem 11.2: {xx R } is not regular for any alphabet with at least two symbols.

Proof: Let M = (Q, , , q0, F) be any DFA with | | 2; we will show that
L(M) {xx }. The alphabet contains at least two symbols; call two of these
R

a and b. Consider the behavior of M on an arbitrarily long string of as. As it

reads the string, M visits a sequence of states: first *(q0, ε), then *(q0, a), then
*(q0, aa), and so on. Eventually, since M has only finitely many states, it
must revisit a state; that is, there must be some i and j with i < j for which
*(q0, a i ) = *(q0, a j ). Now by appending bba j to both strings, we see that
*(q0, a ibba j ) = *(q0, a jbba j ). Thus M ends up in the same final state for
both a ibba j and a jbba j. If that is an accepting state, M accepts both; if it is a
rejecting state, M rejects both. But this means that L(M) {xx R }, since a jbba j is
in {xx R } while a ibba j is not. Since we have shown that L(M ) {xx R } for any
DFA M, we conclude that {anbn} is not regular.

This proof was almost identical to the previous proof. We used the same insight that by
giving a sufficiently long string we could force the DFA to repeat a state—and by so doing,
we could force it to end up in the same final state for two strings that ought to be treated
differently.

11.3 Pumping
Both of the last two proofs worked by choosing a string long enough to make any given DFA
repeat a state. Both proofs just used strings of as, showing that for any given DFA there must
be some i and j with i < j for which *(q0, a i ) = *(q0, a j ).
For those proofs it was enough to find those two strings—two different strings that put
the DFA in the same state. For other proofs of nonregularity it is sometimes necessary to
carry the same argument further: to show that there are not just two but infinitely many
different strings, all of which leave the DFA in the same state. It is actually a fairly simple
extension. Let r be the state that repeats, so r = *(q0, a i ). We know that r is visited again
138 CHAPTER ELEVEN: NONREGULAR LANGUAGES

after j - i more as: *(q0, a j ) = *( q0, a i+(j-i )) = r. In fact, every time the machine reads an
additional j - i as, it will return to state r :
r = *(q0, ai)
= *(q0, ai+(j-i))
= *(q0, ai+2(j-i))
= *(q0, ai+3(j-i))
= ...
and so on. That little substring of j - i additional as can be “pumped” any number of times,
and the DFA always ends up in the same state.
All regular languages have an important property involving pumping. Any sufficiently
long string in a regular language must contain a “pumpable” substring—a substring that can
be replicated any number of times, always yielding another string in the language. Formally:

Lemma 11.1: (The Pumping Lemma for Regular Languages): For all regular
languages L there exists some k N such that for all xyz L with |y| k, there
exist uvw = y with |v| >0, such that for all i 0, xuv iwz L.
Proof: Let L be any regular language, and let M = (Q, , , q0, F ) be any DFA
for it. For k we choose k = |Q|. Consider any x, y, and z with xyz L and |y|
k. Since |y| |Q| we know there is some state r that M repeats while reading
the y part of the string xyz. We can therefore divide y into three substrings
uvw = y so that *(q0, xu) = *(q0, xuv) = r. Now v can be pumped: for all
i 0, *(q0, xuv i ) = r, and so *(q0, xuv iwz) = *(q0, xuvwz) = *(q0, xyz) F.
Therefore, for all i 0, xuviwz L.

In this proof, nothing is known about the strings x, y, and z except that xyz L and
|y| k = |Q|. Because there are at least as many symbols in y as states in Q, we know that
some state r must be visited more than once as M reads the y part of the string xyz:
In state r here And again here

x y z

That gives us a way to divide y into those three further substrings uvw = y:
In state r here And again here

x u v w z
 11.4 PUMPING-LEMMA PROOFS 139
Now v is the pumpable substring. We can replace it with any number of copies, and we know
the DFA will end up in the same state after each copy, so we know it will end in the same
(accepting) state after the whole string is read. Thus it will accept, not just that original string
xyz = xuvwz, but xuviwz for all i.
x u v v ... v w z

Take a closer look at the structure of the pumping lemma. It is a sequence of clauses that
alternate “for all” and “there exist” parts: for all regular languages L there exists some k N
such that for all xyz L with |y| k, there exist uvw = y with |v| >0, such that for all i 0,
xuviwz L. Here is the overall structure of the pumping lemma, using the symbols (“for
all”) and (“there exists”):
1. L ...
2. k ...
3. xyz ...
4. uvw ...
5. i ...
Our proof shows how the values, k and uvw, can be constructed, starting from a DFA for
the language L. But the construction used in the proof is not part of the lemma. From now
on we will forget that k might be the number of states in a DFA for L; we will close the lid
and treat the pumping lemma as a black box. The lemma merely says that suitable values k
and uvw can be found; it does not say how.
11.4 Pumping-Lemma Proofs
The pumping lemma is very useful for proving that languages are not regular. For example,
here is a pumping-lemma proof showing that {anbn} is not regular. The steps in the proof are
numbered for future reference.
1. The proof is by contradiction using the pumping lemma for regular languages.
Assume that L = {anbn} is regular, so the pumping lemma holds for L. Let k be as
given by the pumping lemma.
2. Choose x, y, and z as follows:
x = ak
y = bk
z=ε
Now xyz = akbk L and |y| k as required.
3. Let u, v, and w be as given by the pumping lemma, so that uvw = y, |v| > 0, and for
all i 0, xuviwz L.
140 CHAPTER ELEVEN: NONREGULAR LANGUAGES

4. Choose i = 2. Since v contains at least one b and nothing but bs, uv2w has more bs
than uvw. So xuv2wz has more bs than as, and xuv2wz L.
5. By contradiction, L = {anbn} is not regular.
The structure of this proof matches the structure of the pumping lemma. The alternating
and parts of the lemma make every pumping-lemma proof into a kind of game. The
parts (the natural number k and the strings u, v, and w) are merely guaranteed to exist. In
effect, the pumping lemma itself makes these moves; it is a black box that produces values for
k, u, v, and w. In Steps 1 and 3 we could only say these values are “as given by the pumping
lemma.” The parts, on the other hand, are the moves you get to make; you can choose any
values for x, y, z, and i, since the lemma is supposed to hold for all such values. In Steps 2
and 4 we chose values for x, y, z, and i so as to reach a contradiction: a string xuviwz that is
not in L, contradicting the pumping lemma.
Pumping-lemma proofs follow this pattern quite strictly. Here is another example: a
pumping-lemma proof showing that {xxR} is not regular.
1. The proof is by contradiction using the pumping lemma for regular languages.
Assume that L = {xxR } is regular, so the pumping lemma holds for L. Let k be as
given by the pumping lemma.
2. Choose x, y, and z as follows:
x = akbb
y = ak
z=ε
Now xyz = akbbak L and |y| k as required.
3. Let u, v, and w be as given by the pumping lemma, so that uvw = y, |v| > 0, and for
all i 0, xuviwz L.
4. Choose i = 2. Since v contains at least one a and nothing but as, uv2w has more
as than uvw. So xuv2wz has more as after the bs than before them, and thus
xuv2wz L.
5. By contradiction, L = {xxR } is not regular.
Notice that Steps 1, 3, and 5 are the same as before—only the language L has been changed.
In fact, Steps 1, 3, and 5 are the same in all pumping-lemma proofs. The only steps that
require creativity are Steps 2 and 4. In step 2, you choose xyz and show that you have met the
requirements xyz L and |y| k; the trick here is to choose them so that pumping in the y
part will lead to a contradiction, a string that is not in L. In Step 4 you choose i, the number
of times to pump, and show that the contradiction has been achieved: xuviwz L.
 11.5 STRATEGIES 141

11.5 Strategies
Proving that a language L is not regular using the pumping lemma comes down to those four
delicate choices: you must choose the strings xyz and the pumping count i and show that
these choices lead to a contradiction. There are usually a number of different choices that
successfully lead to a contradiction—and, of course, many others that fail.
Let A = {a n b j a n | n 0, j 1}, and let’s try to prove that A is not regular. The start is
always the same:
1. The proof is by contradiction using the pumping lemma for regular languages.
Assume that A is regular, so the pumping lemma holds for it. Let k be as given by
the pumping lemma.
2. Choose x, y, and z ...
How to proceed? The following table shows some good and bad choices for x, y, and z:
x = aaa Bad choice. The pumping lemma requires |y| k; it never applies to
y=b fixed-size examples. Since k is not known in advance, y must be some
z = aaa string that is constructed using k, such as ak.
x=ε Bad choice. The pumping lemma applies only if the string xyz A. That
y = ak is not the case here.
z = ak
x = an This is ill-formed, since the value of n is not defined. At this point the
y=b only integer variable that is defined is k.
z = an
x = ak This meets the requirements xyz A and |y| k, but it is a bad choice
y = bk+2 because it won’t lead to a contradiction. Pumping within the string y will
z = ak change the number of bs in the middle, but the resulting string can still
be in A.
x = ak This meets the requirements xyz A and |y| k, but it is a bad choice
y = bbak because it won’t lead to a contradiction. The pumping lemma can choose
z=ε any uvw = y with |v| > 0. If it chooses u = b, v = b, and w = ak, there will
be no contradiction, since for all i 0, xuviwz A.
x = akb Good choice. It meets the requirements xyz A and |y| k, and it will
y = ak lead to a contradiction because pumping anywhere in the y part will
z=ε change the number of as after the b, without changing the number
before the b.
x=ε An equally good choice.
y = ak
z = bak
142 CHAPTER ELEVEN: NONREGULAR LANGUAGES

Let’s use that last choice and continue with Step 2:

2. Choose x, y, and z as follows:
x=ε
y = ak
z = ba k
Now xyz = akbak A and |y| k as required.
3. Let u, v, and w be as given by the pumping lemma, so that uvw = y,
|v| > 0, and for all i 0, xuviwz A.
4. Choose i = ...
Now, how to choose i ? In this example it is easy, since there is only one bad choice. For
i = 1, xuviwz = xyz A and no contradiction is reached. For any other choice of i, however,
the string xuviwz will have a different number of as on each side of the b, so it will not be in
A. So we will choose i = 2 and continue:
4. Choose i = 2. Since v contains at least one a and nothing but as, uv2w has more as
than uvw. So xuv2wz has more as after the b than before it, and thus xuv2wz A.
5. By contradiction, A is not regular.
11.6 Pumping and Finite Languages
The pumping lemma applies in a trivial way to all finite languages. Here is our statement of
the pumping lemma again:
For all regular languages L there exists some k N such that for all
xyz L with |y| k, there exist uvw = y with |v| >0, such that for all i 0,
xuviwz L.

If the language L is finite then there must be some number which is greater than the length
of the longest string in L. Choosing k to be this number, it is then obviously true that “for all
xyz L with |y| k, ...”—since there are no strings in L with |y| k. The rest of the lemma
is, as mathematicians say, vacuously true.
Thus, if L is a finite language, we can’t get a contradiction out of the pumping lemma.
In fact, you may already have noticed the following:

Theorem 11.3: All finite languages are regular.

Proof: Let A be any finite language of n strings: A = {x1, ..., xn}. There is a
regular expression that denotes this language: A = L(x1 + ... + xn ). (In case
n = 0, A = L( ).) Since A is denoted by some regular expression, we conclude
that A is a regular language.

Alternative proofs using DFAs, NFAs, or right-linear grammars are only a little harder.
 EXERCISES 143

Exercises
EXERCISE 1
Prove that {anbncn} is not regular. Hint: Copy the proof of Theorem 11.1—only minor
alterations are needed.
EXERCISE 2
Prove that {anb*cn} is not regular. Hint: Copy the proof of Theorem 11.1—only minor
alterations are needed.
EXERCISE 3
Show that {xx R} is regular when | | = 1.
EXERCISE 4
Show that {xcxR | x {a, b}*} is not regular. Hint: Copy the proof of Theorem 11.2—
only minor alterations are needed.
EXERCISE 5
Show that {xx | x {a, b}*} is not regular. Hint: Copy the proof of Theorem 11.2—only
minor alterations are needed.
EXERCISE 6
Let A = {anb2n | n 0}. Using the pumping lemma for regular languages, prove that A is
not regular.
EXERCISE 7
Let B = {anbncn | n 0}. Using the pumping lemma for regular languages, prove that B is
not regular.
EXERCISE 8
Let C = {0n1m0p | n + m = p}. Using the pumping lemma for regular languages, prove that
C is not regular.
EXERCISE 9
Let D = {anbm | n m}. Using the pumping lemma for regular languages, prove that D is
not regular.
EXERCISE 10
Let E = {anbm | m n}. Using the pumping lemma for regular languages, prove that E is
not regular.
EXERCISE 11
Let F = {an | n = b2 for some integer b }. Using the pumping lemma for regular languages,
prove that F is not regular.
 12
CHA P TER

Context-free
Languages
We defined the right-linear grammars by giving a simple restriction
on the form of each production. By relaxing that restriction a bit, we
get a broader class of grammars: the context-free grammars. These
grammars generate the context-free languages, which include all
the regular languages along with many that are not regular.

145
146 CHAPTER TWELVE: CONTEXT-FREE LANGUAGES

12.1 Context-free Grammars and Languages

Many of the nonregular languages we have seen can be generated by grammars. For example,
here is a grammar for {anbn}:
S aSb | ε
This is a grammar for {xxR | x {a, b}*}:
S aSa | bSb | ε
And this is a grammar for {anb jan | n 0, j 1}:
S aSa | R
R bR | b
These grammars are not right linear nor left linear, but they do obey another, simpler
restriction: they are context free. A context-free grammar is one in which the right-hand side
of every production is a single nonterminal symbol.

A grammar G = (V, , S, P) is a context-free grammar (CFG) if and only if

every production in P is of the form
X z
where X V and z (V )*.

Why is this called context free? A production like uRz uyz specifies that R can be replaced
by y, but only in a specific context—only when there is a u to the left and a z to the right. In
a context-free grammar, all productions look like R y, so they specify a substitution for a
nonterminal that does not depend on the context of surrounding symbols in the string.

A context-free language (CFL) is one that is L(G) for some CFG G.

Because every regular language has a right-linear grammar and every right-linear grammar
is a CFG, it follows that every regular language is a CFL. But, as the examples above show,
CFGs can generate languages that are not regular, so not every CFL is regular. The CFLs
properly contain the regular languages, like this:
 12.2 WRITING CFGS 147
L(a*b*)
regular
languages

CFLs

{anbn}

12.2 Writing CFGs

Writing a CFG is a bit like writing a program; it uses some of the same mental muscles. A
program is a finite, structured, mechanical thing that specifies a potentially infinite collection
of runtime behaviors. To write a program, you have to be able to imagine how the code
you are crafting will unfold when it executes. Similarly, a grammar is a finite, structured,
mechanical thing that specifies a potentially infinite language. To write a grammar, you have
to be able to imagine how the productions you are crafting will unfold in derivations of
terminal strings. The derivation mechanism is simple enough, but it takes practice to develop
the skill of writing a CFG for a given CFL. This section gives some techniques and examples.

12.2.1 Regular Languages

If the language is regular, we already have a technique for constructing a CFG: start with an
NFA and apply the construction of Theorem 10.1. For example, consider the language
L = {x {0, 1}* | the number of 0s in x is divisible by 3}
Here is an NFA for L:
1 1 1

0 0
S T U

Now we apply the construction of Theorem 10.1. Wherever the NFA has Y (X, z), we
add the production X zY, and for each accepting state X in the NFA, we add X ε. The
result is this grammar for L:
S 1S | 0T | ε
T 1T | 0U
U 1U | 0S
148 CHAPTER TWELVE: CONTEXT-FREE LANGUAGES

It may be possible to design a smaller or otherwise more readable CFG manually,

without using the construction. Here is another grammar for L, not derived from an NFA:
S T 0T 0T0S | T
T 1T | ε
The construction of Theorem 10.1 does not always produce a pretty grammar, but it is
simple and certain—assuming you already have an NFA!

12.2.2 Balanced Pairs

CFLs often seem to involve strings that contain balanced pairs. For example, in {anbn}, each
a can be paired with a b on the other side of the string; in {xxR}, each symbol in x can be
paired with its mirror image in x R; in {anb jan | n 0, j 1}; each a on the left can be paired
with a matching a on the right. The intuition that CFLs involve strings with balanced pairs
can be made rigorous, as we will see in a later chapter. For now, we just observe that this kind
of balancing is accomplished in the grammar with a simple trick: a recursive production of
the form R xRy. Such a production generates any number of xs, each of which is matched
with a y on the other side. Look again at the grammars at the beginning of this chapter; each
uses the same trick.
For another example, consider the language {anb3n}. This is the subset of L(a*b*) where
the number of bs is three times the number of as. If you think of pairing each a with 3 bs,
you can use the trick for balanced pairs to get this grammar:
S aSbbb | ε
One more example: consider the language
L = {xy | x {a, b}*, y {c, d}*, and |x| = |y|}
If you think of pairing each symbol of x (which can be either a or b) with a symbol of y
(which can be either c or d ), you can use the trick for balanced pairs to get this grammar:
S XSY | ε
X a|b
Y c|d

12.2.3 Concatenations
A divide-and-conquer approach is often helpful for writing CFGs, just as for writing
programs. For instance, consider the language L = {a nb nc md m}. At first glance this may look
daunting, but notice that we can easily write grammars for {anbn} and {c md m}:
S1 aS1b | ε
S2 cS2d | ε
 12.2 WRITING CFGS 149
These are two separate grammars with two separate start symbols S1 and S2. Now every string
in L consists of a string generated from S1 followed by a string generated from S2. So if we
combine the two grammars and introduce a new start symbol, we get a full grammar for L:
S S1S2
S1 aS1b | ε
S2 cS2d | ε
In general, when you discover that a CFL L can be thought of as the concatenation of two
languages L1 and L2
L = L1L2 = {xy | x L1 and y L2}
you can write a CFG for L by writing separate CFGs for L1 and L2, carefully keeping the two
sets of nonterminal symbols separate and using two separate start symbols S1 and S2. A full
grammar for L is then given by combining all the productions and adding a new start symbol
S with the production S S1S2.

12.2.4 Unions
Another use of the divide-and-conquer approach is for a language that can be decomposed
into the union of two simpler languages. Consider this language:
L = {z {a, b}* | z = xxR for some x, or |z| is odd}
Taken all at once this might be difficult, but notice that the definition of L can be expressed
as a union:
L = {xxR | x {a, b}*} {z {a, b}* | |z| is odd}
We can easily give CFGs for these two parts—again, being careful to use separate sets of
nonterminal symbols and separate start symbols. A grammar for {xxR | x {a, b}*} is
S1 aS1a | bS1b | ε
And a grammar for {z {a, b}* | |z| is odd} is
S2 XXS2 | X
X a|b
So a grammar for the whole language L is
S S1 | S2
S1 aS1a | bS1b | ε
S2 XXS2 | X
X a|b
For a slightly more subtle example, consider the language L = {anbm | n m}. One way to
build a grammar for this is to think of it as a union:
150 CHAPTER TWELVE: CONTEXT-FREE LANGUAGES

L = {anbm | n < m} {anbm | n > m}

In fact, to carry the idea one step further, we can think of each of the two parts of that union
as a concatenation:
L1 = {anbn}
L2 = {bi | i > 0}
L3 = {ai | i > 0}
L = L1L2 L3L1
A string in L is a string in {anbn} followed by at least one additional b, or a string in {anbn}
preceded by at least one additional a. This expresses L as a combination of three languages,
each of which is easy to make a CFG for.
S S1S2 | S3S1
S1 aS1b | ε (S1 generates L1)
S2 bS2 | b (S2 generates L2)
S3 aS3 | a (S3 generates L3)

12.3 CFG Applications: Backus-Naur Form

Programming languages contain many “balanced pair” constructs, like the balanced
parentheses in an arithmetic expression and the balanced braces or begin/end keyword
pairs that delimit nested blocks. It is therefore not surprising to find that context-free
grammars play a central role in the definition and implementation of programming
languages.
Between 1959 and 1963, John Backus and Peter Naur developed a way to use grammars
to define the syntax of programming languages. This notation was used in their reports on
the Algol languages. It is now called Backus-Naur Form (BNF).
BNF grammars are really just CFGs. They have the same four parts: a set of terminal
symbols (called tokens), a set of nonterminal symbols, a start symbol, and a set of
productions. The productions are written with a slightly different notation:
• Instead of uppercase letters, nonterminal symbols are written as words enclosed in
angle brackets.
• Instead of the arrow symbol ( ), the symbol ::= separates the left and right sides of
each production
• Instead of ε, the empty string is written as <empty>.
Except for these cosmetic differences, a BNF grammar is a CFG. (Interestingly, although
Noam Chomsky had developed CFGs a few years earlier, Backus and Naur were unaware of
Chomsky’s work and developed their notation independently.) Here is an example of a BNF
definition for a little language of expressions:
 12.4 PARSE TREES 151
<exp> ::= <exp> - <exp> | <exp> * <exp> | <exp> = <exp> | <exp> < <exp>
| (<exp>) | a | b | c
This grammar generates expressions like a<b and (a-(b*c)). The same technique is
used to define all the constructs of a language, as in this fragment of a grammar for a C-like
language:
<stmt> ::= <exp-stmt> | <while-stmt> | <compound-stmt> |...
<exp-stmt> ::= <exp> ;
<while-stmt> ::= while ( <exp> ) <stmt>
<compound-stmt> ::= { <stmt-list> }
<stmt-list> ::= <stmt> <stmt-list> | <empty>
This grammar generates C-like statements such as

while (a<b) {
c = c * a;
a = a + a;
}

The BNF grammar for a full language may include hundreds of productions.

12.4 Parse Trees

Grammars and regular expressions generate languages; DFAs, NFAs, and other automata
recognize languages. Either way, all these formalisms define languages in a purely syntactic
way. They define sets of strings, without ascribing meaning to them.
A programming language, by contrast, is much more than just a definition of syntax.
Whether by direct interpretation or by compilation into a lower-level language, a program
specifies a computation and so has a meaning, a semantics. Programs are supposed to do
something, and that makes them very different from the formal languages studied in this
book. The BNF grammar for a programming language plays a dual role here, specifying not
only syntax but a bit of semantics as well. To see how this works, we need to look at the idea
of the parse tree.
We have treated the productions of a grammar as rules that specify how strings can be
rewritten, so that a rule R x specifies that wherever you see the nonterminal symbol R
you can replace it with x. Another way to think of the productions in a CFG is as rules for
building trees. Imagine building a tree with the start symbol at the root. Each production
R x specifies that the symbols in x may be installed as children of the nonterminal symbol
R. You add children to nonterminal nodes, always following the productions of the grammar,
and stop only when all the leaves are terminal symbols. You then read off the fringe of the
152 CHAPTER TWELVE: CONTEXT-FREE LANGUAGES

tree from left to right. These two different ways of thinking of a grammar are equivalent in
the sense that a string is produced by some derivation if and only if it is the fringe of some
parse tree.
For example, consider again this grammar for a simple language of expressions:
<exp> ::= <exp> - <exp> | <exp> * <exp> | <exp> = <exp> | <exp> < <exp>
| (<exp>) | a | b | c
The string a–b*c is in the language generated by this grammar. Our usual way to
demonstrate this has been to give a derivation, like this:
<exp> <exp> * <exp>
<exp> - <exp> * <exp>
a- <exp> * <exp>
a-b* <exp>
a-b*c
Here is a parse tree for the same string:
<exp>

<exp> * <exp>

c
<exp> - <exp>

a b

This parse tree is more than just a demonstration that the string a–b*c is in the
language—it is also a plan for evaluating the expression when the program is run. It says to
evaluate a - b, then multiply that value by c. (That’s not how most programming languages
handle a-b*c, of course—more about that below!) In general, a parse tree specifies not just
the syntax of a program, but how the different parts of the program fit together. This in turn
says something about what happens when the program runs. The parse tree (or a simplified
version called the abstract syntax tree) is one of the central data structures of almost every
 12.5 AMBIGUITY 153
compiler or other programming language system. To parse a program is to find a parse tree
for it. Every time you compile a program, the compiler must first parse it. We will see more
about algorithms for parsing in another chapter.

12.5 Ambiguity
The previous grammar for expressions is ambiguous in the sense that it permits the
construction of different parse trees for the same string. For example, consider our string
a–b*c. We built one parse tree for it, but here is another one:
<exp>

<exp> - <exp>

a
b * c

This parse tree suggests a different plan for evaluating the expression: first compute the
value b c, then subtract that value from a. This kind of ambiguity is unacceptable for
programming languages; part of the definition of the language must be a clear decision about
whether a–b*c means (a - b) c or a - (b c).
To resolve this problem, BNF grammars for programming languages are usually crafted
to be unambiguous. They not only specify the intended syntax, but do so with a unique parse
tree for each program, one that agrees with intended semantics. This is not usually difficult,
but it generally means making the grammar more complicated. For example, the following
grammar generates the same language as the previous one, but now does so unambiguously:
<exp> ::= <ltexp> = <exp> | <ltexp>
<ltexp> ::= <ltexp> < <subexp> | <subexp>
<subexp> ::= <subexp> - <mulexp> | <mulexp>
<mulexp> ::= <mulexp> * <rootexp> | <rootexp>
<rootexp> ::= (<exp>) | a | b | c
Using this grammar, a parse tree for a-b*c is
154 CHAPTER TWELVE: CONTEXT-FREE LANGUAGES

<exp>

<ltexp>

<subexp>

<subexp> - <mulexp>

<mulexp> <mulexp> * <rootexp>

<rootexp> <rootexp> c

a b

Like most programming language definitions, this gives the multiplication operator higher
precedence than the subtraction operator, so that the expression a–b*c is computed as
a - (b c). It also makes subtraction left associative, so that the expression a-b-c is
computed as (a - b) - c.
An alternative solution to the problem of ambiguity is to stick with the ambiguous,
simple grammar, but add some text describing how to choose among the possible parse
trees. In our case, we could stick with our first grammar but add a sentence saying that
each operator is at a separate level of precedence, in order =, <, -, and * from lowest
to highest, and that all except = are left associative. If the grammar is meant only for
human consumption, this approach may make sense. But often, grammars are meant to
be used directly by computer programs, like parsers and parser generators, about which
we will see more in another chapter. For such uses it is generally better for the grammar
to be unambiguous. (Incidentally, there are CFLs for which it is not possible to give an
unambiguous grammar. Such languages are called inherently ambiguous. But programming
languages rarely suffer from this problem.)

12.6 Extended BNF

Plain BNF has a few special symbols, like ::= and |, which are understood to be part of the
BNF mechanism rather than part of the language being defined. Such special symbols are
called metasymbols. A few more metasymbols can be added to BNF to help with common
patterns of language definition. For example, we might add [, ], {, }, (, and ) as metasymbols
with the following special meanings:
• [ something ] in the right-hand side of a production means that the something inside
is optional.
 12.6 EXTENDED BNF 155
• { something } in the right-hand side of a production means that the something inside
can be repeated any number of times (zero or more). This is like the Kleene star for
regular expressions.
• Parentheses are used to group things on the right-hand side, so that |, [], and {} can
be used in the same production unambiguously.
• Single quotes are used around tokens when necessary to distinguish them from
metasymbols, as in <empty-list> ::= ‘[]’.
With these new metasymbols, some common patterns can be defined more simply than
with plain BNF. For example, an if-then statement with an optional else part might be
defined like this:
<if-stmt> ::= if <expr> then <stmt> [else <stmt>]
Remember that the square brackets are not part of the language being defined; they are
metasymbols that make the else part optional. A list of zero or more statements, each
ending with a semicolon, might have a definition like this:
<stmt-list> ::= {<stmt> ;}
Again, the curly brackets are not part of the language being defined; they are metasymbols
that allow the <stmt> ; part to repeat zero or more times. A list of zero or more things, each
of which can be either a statement or a declaration and each ending with a semicolon, might
have a definition like this:
<thing-list> ::= { (<stmt> | <declaration>) ;}
The parentheses are metasymbols; they make it clear that the ; token is not part of the
choice permitted by the | metasymbol.
Plain BNF can define all these patterns easily enough, but extra metasymbols like those
above make the grammar easier to read. Much like the | metasymbol from plain BNF, the
new metasymbols allow you to express a collection of productions succinctly. Any grammar
syntax that extends BNF in this way is called an EBNF (extended BNF). Many variations
have been used, and there is no widely accepted standard. In formal language terms, an
EBNF is a bit like a CFG augmented with regular expressions in the right-hand sides of the
productions. Since we know that CFGs can already define all the regular languages, it is not
surprising that this augmentation does not really add any power. Any language construct that
can be defined with an EBNF can also be defined with plain BNF. But EBNF can define
various language constructs more conveniently.
Although EBNF grammars can be easier for people to read and write, they are not
without their drawbacks. In particular, the use of the {} metasymbols obscures the form of
the parse tree. In a previous example we used the BNF productions:
<mulexp> ::= <mulexp> * <rootexp> | <rootexp>
156 CHAPTER TWELVE: CONTEXT-FREE LANGUAGES

This could be simplified using EBNF:

<mulexp> ::= <rootexp> {* <rootexp>}
But while the BNF grammar allows only a left-associative parse tree for a sequence of
multiplications like a*b*c, the EBNF grammar is unclear about associativity. One can try
to establish conventions—asserting that, by convention, an EBNF production of the form
above implies left associativity—but many variations of such conventions have been used,
and there is no widely accepted standard.

12.7 Exercises
EXERCISE 1
Give CFGs for the following languages.
a. L(a*b*)
b. {x {0, 1}* | the number of 0s in x is divisible by 3}
c. {x {0, 1}* | |x| is divisible by 3}
d. {x {0, 1}* | x is a binary representation of a number divisible by 3}
e. {x {a, b}* | x L(a*b*)}
f. {anbm | m > 2n}
g. {anbn | n is even}
h. {(ab)nc(de)n}
i. {anbn | n is odd}
j. {aibnajbnak}
k. {anbnaib j }
l. {x {0, 1}* | the number of 0s in x is divisible by 3,
or |x| is divisible by 3, or both}
m. {anbm | m > 2n or m < n}
n. {x {a, b}* | x {anbn}}
o. {anbnc j}
p. {anb2n}
q. {anbmcmdn}
r. {anbncmdm}
s. {x {a, b}* | x = xR}
EXERCISE 2
Give a BNF grammar for each of the languages below:
a. The set of all strings consisting of zero or more as.
b. The set of all strings consisting of an uppercase letter followed by zero or more
additional characters, each of which is either an uppercase letter or one of the
characters 0 through 9.
c. The set of all strings consisting of one or more as.
 EXERCISES 157
d. The set of all strings consisting of one or more digits. (Each digit is one of the
characters 0 through 9.)
e. The set of all strings consisting of zero or more as with a semicolon after each one.
f. The set of all strings consisting of the keyword begin, followed by zero or more
statements with a semicolon after each one, followed by the keyword end. Use the
nonterminal <statement> for statements, and do not give productions for it.
g. The set of all strings consisting of one or more as with a semicolon after each one.
h. The set of all strings consisting of the keyword begin, followed by one or more
statements with a semicolon after each one, followed by the keyword end. Use the
nonterminal <statement> for statements, and do not give productions for it.
i. The set of all strings consisting of one or more as, with a comma between each a
and the next. (There should be no comma before the first or after the last a.)
j. The set of all strings consisting of an open bracket (the symbol [) followed by a list
of one or more digits separated by commas, followed by a closing bracket (the
symbol ]).
k. The set of all strings consisting of zero or more as, with a comma between each a
and the next. (There should be no comma before the first or after the last a.)
l. The set of all strings consisting of an open bracket (the symbol [) followed by a list
of zero or more digits separated by commas, followed by a closing bracket (the
symbol ]).
EXERCISE 3
Give an EBNF grammar for each of the languages of Exercise 1. Use the EBNF
extensions wherever possible to simplify the grammars. In particular, you should
eliminate recursion from the grammars wherever possible. Don’t forget to put quotation
marks around metasymbols when they are used as tokens.
EXERCISE 4
Show that each of the following BNF grammars is ambiguous. (To show that a grammar
is ambiguous, you must demonstrate that it can generate two parse trees for the same
string.)
a. This grammar for expressions:
<exp> ::= <exp> + <exp>
| <exp> * <exp>
| ( <exp> )
|a|b|c
b. This grammar:
<person> ::= <woman> | <man>
<woman> ::= wilma | betty | <empty>
<man> ::= fred | barney | <empty>
158 CHAPTER TWELVE: CONTEXT-FREE LANGUAGES

c. The following grammar for strings of balanced parentheses. (A language of any

number of different kinds of balanced parentheses is called a Dyck language. Dyck
languages play an interesting role in the theory of formal language, because they
isolate the balanced-pair concept that is the essence of the CFLs.)
<s> ::= <s> <s> | ( <s> ) | ()
d. This grammar:
<s> ::= <round> <square> | <outer>
<round> ::= ( <round> ) | ()
<square> ::= [ <square> ] | []
<outer> ::= ( <outer> ] | ( <inner> ]
<inner> ::= ) <inner> [ | )[
EXERCISE 5
For each of the grammars in Exercise 4 except the last, give an unambiguous grammar
for the same language. (The last language in that exercise is a classic example of an
inherently ambiguous language—the grammar cannot be fixed!)
EXERCISE 6
We claimed without proof that for this grammar G:
S aSb | ε
we have L(G) = {anbn}. Prove it.
EXERCISE 7
We claimed without proof that for this grammar G:
S aSa | bSb | ε
we have L(G) = {xxR | x {a, b}*}. Prove it.
 13
CHA P TER

Stack Machines
Commonly, the word stack refers to any orderly, vertical pile: a stack of
books, plates, or poker chips. All the action in a stack is at the top, where
items can be added or removed. Throughout computer science, the word
stack is used in a related technical sense: a stack is a collection of data
accessed in last-in-first-out order. Data items may be added to the top
(pushed onto the stack) or removed from the top (popped off the stack).
Stacks are ubiquitous in computer programming, and they have an
important role in formal language as well. A stack machine is a kind
of automaton that uses a stack for auxiliary data storage. The size of the
stack is unbounded—it never runs out of space—and that gives stack
machines an edge over finite automata. In effect, stack machines have
infinite memory, though they must use it in stack order.
If you travel by two paths that seem to depart in different directions, it is
a surprise to discover that they lead to the same destination. It makes that
destination feel more important—an intersection rather than a dead end.
That is the situation with context-free languages. Stack machines and
CFGs seem like two very different mechanisms for language definition—
two paths that depart in different directions. But it turns out that these
two paths lead to the same place. The set of languages that can be defined
using a stack machine is exactly the same as the set of languages that can
be defined using a CFG: the context-free languages.
159
160 CHAPTER THIRTEEN: STACK MACHINES

13.1 Stack Machine Basics

A stack machine is a very different kind of automaton for defining languages. Unlike a
DFA or NFA, it makes no state transitions; instead, it maintains a stack of symbols. We will
represent this stack as a string, treating the left end of the string as the top of the stack. For
example, the string abc represents a stack with a on top and c on the bottom; popping the a
off the top of abc leaves the stack bc, while pushing b onto abc produces the stack babc.
A stack machine is specified by a table that shows the moves it is allowed to make. For
example:
read pop push
a c abc

The entry in the first column is an input symbol (or ε—more about this below). The entry
in the second column is a stack symbol, and the entry in the third column is a string of stack
symbols. The example above says
If the current input symbol is a, and if the symbol on top of the
stack is c, you may pop the c off the stack, push the string abc on in
its place, and advance to the next input symbol.
Remember that the stack is represented as a string with the top symbol at the left end. So
when you push a string of two or more symbols onto the stack, it is the leftmost one that
becomes the new top. For example, suppose the stack machine’s next input symbol is an a,
and suppose its stack is cd. Then the move shown above can be used, leaving the string abcd
on the stack. (It first pops the c off, leaving d, then pushes the string abc on, leaving abcd.)
The new top symbol on the stack is the leftmost symbol, the a.
Every move of a stack machine pops one symbol off the stack, then pushes a string of
zero or more symbols onto the stack. To specify a move that leaves the stack unchanged, you
can explicitly push the popped symbol back on, like this:
read pop push
a c c

In this case, if the stack machine’s next input symbol is an a, and its stack is cd, then the
move shown above can be used, leaving the string cd on the stack. (It first pops the c off,
leaving d, then pushes the c back on, leaving cd.)
Every move of a stack machine pushes some string of symbols onto the stack. To specify
a move that pops but does not push, you can explicitly push the empty string, like this:
read pop push
a c ε
 13.1 STACK MACHINE BASICS 161
In this case, if the stack machine’s next input symbol is an a, and its stack is cd, then the
move shown above can be used, leaving the string d on the stack. (It first pops the c off,
leaving d, then pushes the empty string back on, still leaving d.)
The entry in the first column can be ε. This encodes a move that can be made without
reading an input symbol, just like an ε-transition in an NFA:

read pop push

ε c ab

The example above says that whenever the stack machine has a c on top of the stack, it may
pop it off and push the string ab onto the stack. This move does not advance to the next
input symbol and may even be made after all the input symbols have been read.
A stack machine starts with a stack that contains just one symbol S. On each move it can
alter its stack, but only in stack order—only by popping from the top and/or pushing onto
the top. If the stack machine decides to accept the input string, it signals this by leaving its
stack empty, popping everything off including the original bottom-of-stack symbol S.
Like an NFA, a stack machine is potentially nondeterministic: it may have more than
one legal sequence of moves on a given input string. It accepts if there is at least one legal
sequence of moves that reads the entire input string and ends with the stack empty—the initial
symbol S, and anything else that was pushed during the computation, must be popped off to
signal that the input string is accepted.
Consider this stack machine:
read pop push
1. ε S ab
2. a S ef
3. a S ε

Suppose the input string is a. The initial stack is S, so all three moves are possible. In fact,
there are three possible sequences of moves:
• If move 1 is used as the first move, no input is read and the stack becomes ab. From
there no further move is defined. This is a rejecting sequence of moves, both because the
input was not finished and because the stack is not empty.
• If move 2 is used as the first move, the a is read and the stack becomes ef; then no further
move can be made. This too is a rejecting sequence of moves, because even though all the
input was read, the stack is not empty.
• If move 3 is used as the first move, the a is read and the stack becomes empty. This is an
accepting sequence of moves: the input was all read and the stack is empty at the end.
Because there is at least one accepting sequence of moves for a, a is in the language defined
by this stack machine. (In fact, the language defined by this stack machine is just {a}.)
162 CHAPTER THIRTEEN: STACK MACHINES

13.2 A Stack Machine for {anbn}

Here’s a general strategy for using a stack machine to recognize the language {anbn}. Suppose
the stack initially contains the symbol S. Now start reading the input string from left to right:
1. For each a you read, pop off the S, push a 1, and then push the S back on.
2. In the middle of the string, pop off the S. (At this point the stack contains just a list
of zero or more 1s, one for each a that was read from the input string.)
3. For each b you read, pop a 1 off the stack.
This strategy permits only as until the S is popped and only bs after the S is popped.
Moreover, it pushes a 1 for every a and pops a 1 for every b. Thus it finishes with an empty
stack if and only if the input string is in {anbn}.
Here is our strategy, translated into a table of moves for a stack machine:
read pop push
1. a S S1
2. ε S ε
3. b 1 ε

Let’s look at the accepting sequence of moves this machine makes for the string aaabbb.
Initially, the stack machine is reading the first symbol of aaabbb and has S on its stack, as
shown here:
aaabbb S

The current input symbol and the top-of-the-stack symbol are underlined; these determine
which moves are possible. The possible moves are 1 and 2, but the accepting sequence starts
with move 1. After that move, the stack machine has advanced to the second symbol of
aaabbb and has S1 on its stack, like this:
aaabbb S

move 1: pop S, push S1, and advance to next input symbol

aaabbb S1

The full sequence is

 13.2 A STACK MACHINE FOR {anbn} 163

aaabbb S

move 1: pop S, push S1, and advance to next input symbol

aaabbb S1

move 1: pop S, push S1, and advance to next input symbol

aaabbb S11

move 1: pop S, push S1, and advance to next input symbol

aaabbb S111

move 2: pop S

aaabbb 111

move 3: pop 1 and advance to next input symbol

aaabbb 11

move 3: pop 1 and advance to next input symbol

aaabbb 1

move 3: pop 1 and advance to next input symbol

aaabbb ε

At the end, the stack machine has read all its input and emptied its stack. That is an
accepting sequence of moves. As we have already seen, there are also many nonaccepting
sequences, such as this one:
164 CHAPTER THIRTEEN: STACK MACHINES

aaabbb S

move 1: pop S, push S1, and advance to next input symbol

aaabbb S1

move 2: pop S

aaabbb 1

no legal move from here

There is no next move here, so the stack machine terminates without reaching the end of the
input string and without emptying its stack. But since there is at least one accepting sequence
of moves, the string aaabbb is in the language defined by the machine.

13.3 A Stack Machine for {xxR}

Here’s a general strategy for using a stack machine to recognize the language
{xxR | x {a, b}*}. Suppose the stack initially contains the symbol S. Now start reading the
input string from left to right:
1. For each a you read, pop off the S, push a, and then push the S back on.
2. For each b you read, pop off the S, push b, and then push the S back on.
3. In the middle of the string, pop off the S. (At this point the stack contains just a
sequence of as and bs—the reverse of the string read so far.)
4. For each a you read, pop a off the stack.
5. For each b you read, pop b off the stack.
Before the S is popped, this strategy pushes (beneath the S ) a copy of each symbol read.
Because of the stack order, the result is that (beneath the S ) the stack contains the reverse of
the string read so far. After the S is popped, our strategy matches each input symbol against
the top symbol of the stack, which is popped off. Thus it finishes with an empty stack if and
only if the second part of the input string is the reverse of the first part—that is, if and only if
the input string is in {xxR | x {a, b}*}.
Here is our strategy, translated into a table of moves for a stack machine:
 13.3 A STACK MACHINE FOR {xx R} 165

read pop push

1. a S Sa
2. b S Sb
3. ε S ε
4. a a ε
5. b b ε

Here is an accepting sequence of moves for the string abbbba:

abbbba S

move 1: pop S, push Sa, and advance to next input symbol

abbbba Sa

move 2: pop S, push Sb, and advance to next input symbol

abbbba Sba

move 2: pop S, push Sb, and advance to next input symbol

abbbba Sbba

move 3: pop S

abbbba bba

move 5: pop b and advance to next input symbol

abbbba ba

move 5: pop b and advance to next input symbol

abbbba a

move 4: pop a and advance to next input symbol

aaabbb ε
166 CHAPTER THIRTEEN: STACK MACHINES

Notice that this stack machine can use move 3 at any time (though, since it pops off the
only S, it can be used only once). However, in an accepting sequence move 3 must occur
exactly in the middle of the input string. We can think of the stack machine as making a
guess about where the middle of the string is. When it guesses that it has found the middle,
it applies move 3, then begins popping instead of pushing. This takes advantage of the
nondeterminism of the stack machine. It accepts if there is any sequence of moves that reads
the entire input string and ends with an empty stack. All those sequences that make the
wrong guess about where the middle of the string is do not accept. But if the string is in our
language, the one sequence that makes the right guess about where the middle of the string is
does accept; and that one is all it takes.

13.4 Stack Machines, Formally Defined

We have been working with informal descriptions of stack machines. Now, as always,
we need a way to describe them mathematically. This is easy to do, because our informal
description using a table of moves is already almost rigorous enough; all that remains is to
observe that the table encodes a function and the function captures the essential structure of
the machine.

A stack machine M is a 4-tuple M = ( , , S, ), where

is the stack alphabet
is the input alphabet
S is the initial stack symbol
(( {ε}) P( *)) is the transition function

The stack alphabet may or may not overlap with the input alphabet.
The transition function takes two parameters. The first parameter is an input symbol
or ε, and the second is a stack symbol. These parameters correspond to the first two
columns of the tables we used in the previous chapter: the input symbol being read (or ε
for ε-transitions) and the symbol currently on top of the stack. The output of the transition
function is a set of strings that can be pushed onto the stack. It is a set of strings, rather than
just a single string, because a stack machine is nondeterministic: at each stage there may be
any number of possible moves. For example, this machine:
read pop push
1. ε S ab
2. a S ef
3. a S ε

has this transition function:

 13.4 STACK MACHINES, FORMALLY DEFINED 167
(ε, S) = {ab}
(a, S) = {ε, ef }
As a stack machine operates, it reads its input string and changes the contents of its
stack. At any point in the stack machine’s operation, its future depends on two things: that
part of the input string that is still to be read, and the current contents of the stack. Taken
together, these two pieces of information are an instantaneous description for the stack
machine. The following definitions are all made with respect to some fixed stack machine
M = ( , , S, ):

An instantaneous description for a stack machine is a pair (x, y) where:

x * is the unread part of the input
y * is the current stack contents
As always, the left end of the string y is considered to be the top of the stack.

The function for the stack machine determines a relation on IDs; we write
I J if I is an ID and J is an ID that follows from I after one move of the stack machine.

is a relation on IDs, defined by the function for the stack machine as

follows:
1. Regular transitions: (ax, Bz) (x, yz) if and only if y (a, B)
2. ε-transitions: (x, Bz) (x, yz) if and only if y (ε, B).

Note that no move is possible when the stack machine’s stack is empty; there is never an ID I
with (x, ε) I.
Next, as usual, we define an extended relation * for sequences of zero or more steps:

* is a relation on IDs, with I * J if and only if there is a sequence of zero or

more relations that starts with I and ends with J.

Notice here that * is reflexive: for any ID I, I * I by a sequence of zero moves. Using the
* relation, we can define the language accepted by M:

The language accepted by a stack machine M is

L(M) = {x * | (x, S) * (ε, ε)}

In this definition, the stack machine starts reading its input string x with S on the stack. It
accepts the string x if it has some sequence of zero or more moves that ends with all the input
used up and the stack empty.
For example, in Section 13.3 we gave an informal description of a stack machine for the
language {xxR | x {a, b}*}:
168 CHAPTER THIRTEEN: STACK MACHINES

read pop push

1. a S Sa
2. b S Sb
3. ε S ε
4. a a ε
5. b b ε

and we showed an excepting sequence of moves for the string abbbba. Formally, this machine
is M = ({a, b, S}, {a, b}, S, ), where
(a, S) = {Sa}
(b, S) = {Sb}
(ε, S) = {ε}
(a, a) = {ε}
(b, b) = {ε}
and the accepting sequence of moves that we showed for abbbba is
(abbbba, S) (bbbba, Sa) (bbba, Sba) (bba, Sbba)
(bba, bba) (ba, ba) (a, a) (ε, ε)
Thus (abbbba, S) * (ε, ε) and so, by definition, abbbba L(M).
That one example does not, of course, establish that L(M ) = {xxR | x {a, b}*}. Proofs of
that kind tend to be quite tricky. In the case of our current stack machine, however, we can
sidestep explicit induction without too much difficulty. We first observe that only move 3
changes the number of S symbols in the stack, reducing it by 1. So move 3 must be used in
any accepting sequence, exactly once. In fact, any accepting sequence must have the form
(xy, S) * (y, Sz) 3 (y, z) * (ε, ε)
where x, y, and z are in {a, b}*, only moves 1 and 2 occur before the use of move 3, and
only moves 4 and 5 occur after it. Moves 1 and 2 push the symbol just read onto the stack;
therefore z = xR. Moves 4 and 5 pop as and bs only if they match the input; therefore z = y.
Thus any accepting sequence must have the form
(xxR, S) * (xR, SxR) (xR, xR) * (ε, ε)
and there is such an accepting sequence for any string x. We conclude that
L(M) = {xxR | x {a, b}*}.

13.5 An Example: Equal Counts

Consider this language:
A = {x {a, b}* | the number of as in x equals the number of bs in x}.
This is not the same as {anbn}, since it includes strings like abab that have the as and bs mixed
together.
 13.5 AN EXAMPLE: EQUAL COUNTS 169
For a first attempt at making a stack machine for this language, we can try the following
strategy. Whenever the machine sees an a, it pushes a 0 onto the stack. Whenever is sees a b,
it pops a 0 off the stack. So if the number of as is the same as the number of bs, the machine
ends up with just S on the stack. It pops that off as its last move. That strategy is embodied
in this stack machine, which we’ll call M1:
read pop push
1. a S 0S
2. a 0 00
3. b 0 ε
4. ε S ε

Here, moves 1 and 2 allow it to push a 0 for each a read. (Two moves are required to handle
the two possible symbols on top of the stack, S and 0.) Move 3 allows it to pop a 0 for each
b read. Move 4 allows it to pop the S off. Since move 4 is the only way to get rid of the S,
it must occur in any accepting sequence; and since it leaves the stack empty, no move can
follow it. Thus it is always the last move in any accepting sequence.
Formally, this is M1 = ({0, S}, {a, b}, S, ), where the function is
(a, S) = {0S}
(a, 0) = {00}
(b, 0) = {ε}
(ε, S) = {ε}
Here is a sequence of IDs showing that M1 accepts abab:
(abab, S) (bab, 0S) (ab, S) (b, 0S) (ε, S) (ε, ε)
This sequence shows that we have (abab, S) * (ε, ε), and so, by definition, abab L(M1).
It is clear that by using the stack as a counter, M1 checks that the number of as in the
string is the same as the number of bs. Thus everything M1 accepts is in our target language
A. Unfortunately, not everything in A is accepted by M1. Consider its behavior on the input
string abba. Starting in (abba, S), it can proceed to (bba, 0S ) and then to (ba, S ). But from
there it has no transition for reading the next input symbol, the b. Our strategy uses the
number of 0s on the stack to represent the number of as so far minus the number of bs so far.
That will fail if, as in abba, there is some prefix with more bs than as.
To handle this we need a more sophisticated strategy. We will still keep count using 0s
on the stack, whenever the number of as so far exceeds the number of bs so far. But when
the number of bs so far exceeds the number of as so far, we will keep count using 1s. That
strategy is embodied in this stack machine, which we’ll call M2:
170 CHAPTER THIRTEEN: STACK MACHINES

read pop push

1. a S 0S
2. a 0 00
3. b 0 ε
4. b S 1S
5. b 1 11
6. a 1 ε
7. ε S ε

Formally, this is M2 = ({0, 1, S }, {a, b}, S, ), where the function is

(a, S) = {0S}
(a, 0) = {00}
(b, 0) = {ε}
(b, S) = {1S}
(b, 1) = {11}
(a, 1) = {ε}
(ε, S) = {ε}
This is now flexible enough to handle strings like abba:
(abba, S) (bba, 0S) (ba, S) (a, 1S) (ε, S) (ε, ε)
In fact, the machine now accepts our language, L(M2) = A.

13.6 An Example: A Regular Language

A stack machine can easily simulate any finite automaton. Consider the language
{x {0, 1}* | the next-to-last symbol in x is 1}, which has this DFA:
0

q0 1 q1
1

1
0
0

q2 q3
0

1
 13.7 A STACK MACHINE FOR EVERY CFG 171
Our strategy for making a stack machine to simulate a DFA is simple. Until the last move,
our machine will always have exactly one symbol on its stack, which will be the current
state of the DFA. As its last move, if the DFA is in an accepting state, the stack machine will
pop that state off, ending with an empty stack. For the DFA above, our stack machine is
M = ({q0, q1, q2, q3}, {0, 1}, q0, ), where the function is
(0, q0) = {q0}
(1, q0) = {q1}
(0, q1) = {q2}
(1, q1) = {q3}
(0, q2) = {q0}
(1, q2) = {q1}
(0, q3) = {q2}
(1, q3) = {q3}
(ε, q2) = {ε}
(ε, q3) = {ε}
Notice that we are using the states of the DFA as the stack alphabet of the stack machine,
and we are using the start state q0 of the DFA as the initial stack symbol. The transition
function of the stack machine exactly parallels that of the DFA, so that the stack machine
simply simulates the actions of the DFA. For example, the DFA accepts the string 0110 by
going through a series of states (q0, q0, q1, q3, q2) ending in an accepting state (q2). The stack
machine accepts the string 0110 through the corresponding sequence of IDs:
(0110, q0) (110, q0) (10, q1) (0, q3) (ε, q2) (ε, ε)
The same kind of construction works, not just for DFAs, but for NFAs as well. You
can practice it in some of the following exercises. However, we will not give the formal
construction here, since we will now prove an even stronger result.

13.7 A Stack Machine for Every CFG

Consider the string-rewriting process defined by a CFG. From the start symbol S a sequence
of replacements is made; at each step, one nonterminal symbol in the string is replaced by a
string of symbols as specified by one production of the grammar. The set of all fully terminal
strings that can be derived in this way is the language defined by the CFG.
This string-rewriting process can be carried out by a stack machine, using the stack to
hold the string being rewritten. At each step, one nonterminal symbol in the stack string is
replaced by a string of symbols as specified by one production of the grammar. There’s just
one hitch: a stack machine can only operate on the top of the stack—the leftmost symbol
in the string. If this leftmost symbol is a nonterminal, the stack machine can replace it
with something else, tracing out a leftmost derivation. But what if the leftmost symbol is a
terminal symbol? In that case the stack machine can simply pop it off, since it is a finished
part of the string being derived.
172 CHAPTER THIRTEEN: STACK MACHINES

Using this idea, we can take any CFG and construct a stack machine for the same
language. The constructed machine starts, as always, with the start symbol S on its stack.
It carries out a derivation in the language, repeatedly using two types of moves:
1. If the top symbol on the stack is a nonterminal, replace it using any one of the
grammar’s productions for that nonterminal. This is an ε-move, not consuming any
symbols of the stack machine's input.
2. If the top symbol on the stack is a terminal that matches the next input symbol,
pop it off.
Finally, when it has completed a derivation and popped off all the resulting terminal symbols,
it is left with an empty stack, and it accepts the input string.
This is a highly nondeterministic machine. The moves of type 1 allow the machine to
trace out any derivation from S permitted by the grammar. But the moves of type 2 allow
progress only if the derived string matches the input string. The only way the stack can be
emptied is if a fully terminal string was derived. The only way the stack machine will accept
at that point is if the entire input string has been matched. In effect, the machine can try all
possible derivations; it accepts if and only if at least one of them is a derivation of the input
string.
For example, consider this CFG for {xxR | x {a, b}*}:
S aSa | bSb | ε
The constructed stack machine corresponding to this CFG is this:
read pop push
1. ε S aSa
2. ε S bSb
3. ε S ε
4. a a ε
5. b b ε

Moves 1 through 3 correspond to the three productions in the grammar. Moves 4 and 5
allow the stack machine to pop the terminals a and b off the stack if they match the input
symbols; these moves would be the same in any stack machine constructed from a CFG
whose terminal alphabet is {a, b}.
A string in the language defined by the CFG is abbbba. Here is a derivation for the
string:
S aSb abSba abbSbba abbbba
 13.8 A CFG FOR EVERY STACK MACHINE 173
Here is the corresponding accepting sequence of IDs for the stack machine. The subscripts
on each step indicate which move was performed:
(abbbba, S) 1 (abbbba, aSa) 4 (bbbba, Sa) 2 (bbbba, bSba)
5
(bbba, Sba) 2 (bbba, bSbba) 5 (bba, Sbba) 3 (bba, bba)
5
(ba, ba) 5 (a, a) 4 (ε, ε)
We can generalize this construction in the following lemma and proof sketch.

Lemma 13.1: If G = (V, , S, P) is any context-free grammar, there is some stack

machine M with L(M ) = L(G).
Proof sketch: By construction. Let M be the stack machine
M = (V , , S, ), where the function is defined as follows:
• for all v V, (ε, v) = {x | (v x) P }, and
• for all a , (a, a) = {ε}.
Now M accepts a string if and only if there is a derivation of that string in the
grammar G. Thus L(M) = L(G ).
A complete proof would have to demonstrate in detail that M accepts a string if and only
if there is a derivation of that string in the grammar G. We’ll skip that step here, but there is
a very similar proof in the next section.

13.8 A CFG for Every Stack Machine

We have just seen that it is possible to take any CFG and construct an equivalent stack
machine. The reverse is also true: there is a way to take any stack machine and construct
a CFG for the same language. In many ways this is like the previous construction; the
executions of the stack machine are exactly simulated by derivations in the CFG.
For example, consider this stack machine from Section 13.5:
read pop push
1. a S 0S
2. a 0 00
3. b 0 ε
4. b S 1S
5. b 1 11
6. a 1 ε
7. ε S ε

The language it accepts is the language of strings over the alphabet {a, b} in which the
number of as equals the number of bs. Here is a grammar that can simulate this stack
machine using productions that mimic the stack machine’s transitions:
174 CHAPTER THIRTEEN: STACK MACHINES

1. S a0S
2. 0 a00
3. 0 b
4. S b1S
5. 1 b11
6. 1 a
7. S ε
The productions are numbered here to show how they match the moves of the stack
machine. Wherever the stack machine has a move t ( , A), the grammar has a
production A t. Thus the derivations in the grammar simulate the executions of the
stack machine: each time the stack machine reads an input symbol, the grammar generates
that symbol, and the stack machine’s stack is simulated in the nonterminal part of the string
being rewritten. For example, this is an accepting sequence in the stack machine on the input
abab:
(abab, S ) 1 (bab, 0S) 3 (ab, S ) 1 (b, 0S) 3 (ε, S ) 7 (ε, ε)
(The subscripts indicate which move was used at each step.) The grammar generates the
string abab using the corresponding sequence of derivation steps:
S 1
a0S 3
abS 1
aba0S 3
ababS 7
abab
At each stage in this derivation, the string has two parts xy: x is the portion of the input
string that has been read by the stack machine, and y is the content of the stack. At the end
of the derivation the stack machine’s stack y is empty, and the string x is the fully terminal
string that was the stack machine’s input.

Lemma 13.2.1: If M = ( , , S, ) is any stack machine, there is context-free

grammar G with L(G) = L(M ).
Proof: By construction. We assume that and have no elements in common.
(This is without loss of generality since the symbols in can be renamed away
from if necessary.) Let G be the grammar G = ( , , S, P), where

P = {(A t) | A , ( {ε}), and t ( , A)}.

Now leftmost derivations in G exactly simulate executions of M, in the sense

that for any x * and y *,

S * xy if and only if (x, S ) * (ε, y)

(A detailed proof of this is given below.) Setting y = ε, this gives

 13.8 A CFG FOR EVERY STACK MACHINE 175
S * x if and only if (x, S) * (ε, ε)

and so L(G) = L(M).

In the construction, the stack symbols of the stack machine become nonterminals in
the constructed grammar, and the input symbols become terminals. That’s why we need the
assumption that the two alphabets have no symbols in common; no symbol in a grammar
can be both a terminal and a nonterminal. The assumption is “without loss of generality”
because offending stack symbols can always be renamed if necessary.
For an example demonstrating this kind of renaming, consider this stack machine for the
language {anb2n}:
read pop push
1. a S Sbb
2. ε S ε
3. b b ε

This stack machine does not satisfy our assumption, because it uses the symbol b as
both an input symbol and a stack symbol. If we applied our construction to this machine
we would end up with a grammar containing the production b b, which is not legal in a
CFG. But the problem is easy to solve by renaming: simply replace the stack symbol b with a
new symbol such as B:
read pop push
1. a S SBB
2. ε S ε
3. b B ε

Notice that we changed only the pop and push columns—only the instances of b on the stack,
not the instance of b as an input symbol. This modified stack machine now satisfies our
assumption, and we can proceed with the construction, yielding this CFG:
S aSBB | ε
B b
The proof of Lemma 13.2.1 asserts that the leftmost derivations in G exactly simulate
the executions of M. The previous examples should help to convince you that this is true, but
we can also prove the assertion in more detail using induction. We have avoided most such
proofs this far, but since this one is similar to the missing piece in our proof of Lemma 13.1,
let’s bite the bullet this time and fill in the details.
176 CHAPTER THIRTEEN: STACK MACHINES

Lemma 13.2.2: For the construction of the proof of Lemma 13.2.1, for any
x * and y *, S * xy if and only if (x, S) * (ε, y).
Proof that if S * xy then (x, S) * (ε, y): By induction on the length of the
derivation. In the base case the length of the derivation is zero. The only zero-
length derivation S * xy is with xy = S, and since x * we must have x = ε
and y = S. For these values, by definition, (x, S) * (ε, y) in zero steps.
In the inductive case, the length of the derivation is greater than zero, so it has
at least one final step. Also, there must be some leftmost derivation of the same
string, and a leftmost derivation in G always produces a string of terminals
followed by a string of nonterminals. Thus, we have

S * x'Ay' xy
for some x' *, A , and y' *. By the inductive hypothesis (which
applies since the derivation of x'Ay' is one step shorter) we know that
(x', S) * (ε, Ay'). The final step in the derivation uses one of the productions
in P, which by construction are all of the form
(A t) | A , ( {ε}), and t ( , A)

Thus the derivation can be written as

S * x'Ay' x' ty' = xy

where x' = x and ty' = y. Now since (x', S) * (ε, Ay' ) we also have
(x' , S) * ( , Ay' ), and since t ( , A) we can add a final step for
the stack machine:

(x, S ) = (x' , S) * ( , Ay' ) (ε, ty' ) = (ε, y)

This gives (x, S) * (ε, y) as required.

Proof of the other direction (if (x, S) * (ε, y) then S * xy): Similar, by
induction on the number of steps in the execution.

Putting the previous results together, we have

Theorem 13.1: A language is context free if and only if it is L(M ) for some stack
machine M.
Proof: Follows immediately from Lemmas 13.1 and 13.2.1.

So CFGs and stack machines have equivalent definitional power.

 EXERCISES 177

Exercises
EXERCISE 1
How would you change the stack machine of section 13.2 so that the language it accepts
is {anbn | n > 0}?
EXERCISE 2
Show the table of moves for a stack machine for the language {ancbn}.
EXERCISE 3
How would you change the stack machine of Section 13.3 so that the language it accepts
is {xxR | x {a, b}*, x ε}?
EXERCISE 4
Show the table of moves for a stack machine for the language {xcxR | x {a, b}*}.
EXERCISE 5
Show the table of moves for a stack machine for the language {(ab)n(ba)n}.
EXERCISE 6
Give stack machines for the following languages:
a. L(a*b*)
b. {x {0, 1}* | the number of 0s in x is divisible by 3}
c. {x {0, 1}* | |x| is divisible by 3}
d. {x {0, 1}* | x is a binary representation of a number divisible by 3}
e. {x {a, b}* | x L(a*b*)}
f. {anbm | m > 2n}
g. {anbn | n is even}
h. {(ab)nc(de)n}
i. {anbn | n is odd}
j. {aibna j bnak}
k. {anbnaib j }
l. {anbm | m n}
m. {anbnc j }
n. {anb2n}
o. {anbmc md n}
p. {anbnc md m}
q. {x {a, b}* | x = xR}
178 CHAPTER THIRTEEN: STACK MACHINES

EXERCISE 7
Using the construction of Lemma 13.1, give a stack machine for this grammar:
S S1 | S2
S1 aS1a | bS1b | ε
S2 XXS2 | X
X a|b
Show accepting sequences of IDs for your stack machine for abba and bbb.
EXERCISE 8
Using the construction of Lemma 13.1, give a stack machine for this grammar:
S T0T0T0S | T
T 1T | ε
Show accepting sequences of IDs for your stack machine for 110010 and 00011000.
EXERCISE 9
Using the construction of Lemma 13.1, give a stack machine for this grammar:
S XSY | ε
X a|b
Y c|d
Show accepting sequences of IDs for your stack machine for abcd and abbddd.
EXERCISE 10
Using the construction of Lemma 13.1, give a stack machine for this grammar:
S S1S2 | S3S1
S1 aS1b | ε
S2 bS2 | b
S3 aS3 | a
Show accepting sequences of IDs for your stack machine for aabbb and aaabb.
EXERCISE 11
Give a more elegant grammar for the language recognized by the stack machine in
Section 13.8.
EXERCISE 12
Using the construction of Lemma 13.2.1, give a CFG for this stack machine:
 EXERCISES 179

read pop push

1. a S 0S
2. a 0 00
3. b 0 ε
4. ε S ε

Show a derivation of abab in your grammar.

EXERCISE 13
Using the construction of Lemma 13.2.1, give a CFG for this stack machine:
read pop push
1. a S Sa
2. b S Sb
3. ε S ε
4. a a ε
5. b b ε

(Note that because this machine has stack and input alphabets that overlap, you will have
to rename some of the symbols in the stack alphabet to make the construction work.)
Show a derivation of abba in your grammar.
EXERCISE 14
Give an example of a stack machine for the language {anbn} that has exactly two distinct
accepting sequences of IDs on every accepted input. Then give an example of a stack
machine for the same language that has infinitely many distinct accepting sequences of
IDs on every accepted input.
EXERCISE 15
Our proof of Lemma 13.2.2 ended like this:

Proof of the other direction (if (x, S) * (ε, y) then S * xy): Similar, by
induction on the number of steps in the execution.

Fill in the missing details, proving by induction that if (x, S) * (ε, y) then S * xy.
 14
CHA P TER

The Context-free
Frontier
At this point we have two major language categories, the regular languages and the
context-free languages, and we have seen that the CFLs include the regular languages,
like this:

L(a*b*)
regular
languages

CFLs

{anbn}

Are there languages outside of the CFLs? In this chapter we see that the answer is yes,
and we see some simple examples of languages that are not CFLs.
We have already seen that there are many closure properties for regular languages. Given
any two regular languages, there are many ways to combine them—intersections, unions,
and so on—that are guaranteed to produce another regular language. The context-
free languages also have some closure properties, though not as many as the regular
languages. If regular languages are a safe and settled territory, context-free languages are
more like frontier towns. Some operations like union get you safely to another context-
free language; others like complement and intersection just leave you in the wilderness.
181
182 CHAPTER FOURTEEN: THE CONTEXT-FREE FRONTIER

14.1 Pumping Parse Trees

To show that there are some languages that are not regular, we reasoned that any sufficiently
long string will force a DFA to repeat a state. For context-free languages there is a parallel
argument, again involving repetition. We will see that we can force a parse tree to repeat a
nonterminal symbol in a certain way. Such parse trees are called pumping parse trees.

A pumping parse tree for a CFG G = (V, , S, P ) is a parse tree with two
properties:
1. There is a node for some nonterminal symbol A, which has that
same nonterminal symbol A as one of its descendants.
2. The terminal string generated from the ancestor A is longer than
the terminal string generated from the descendant A.

If you can find a pumping parse tree for a grammar, you know a set of strings that must
be in the language—not just the string the pumping parse tree yields, but a whole collection
of related strings.

Lemma 14.1.1: If a grammar G generates a pumping parse tree with a yield as

shown:

u v w x y

then L(G) includes uviwxiy for all i.

Proof: The string uvwxy is the whole terminal string derived by the pumping
parse tree. As shown, A is the nonterminal symbol that occurs as its own
descendant in the tree. The string vwx is the terminal string derived from
the ancestor A, and the string w is the terminal string derived from the
 14.1 PUMPING PARSE TREES 183
descendant A. (Satisfying the second property of pumping parse trees, v and x
are not both ε.) It is the nature of context-free grammars that any A in the tree
can be replaced with either of the two subtrees rooted at A, producing other
legal parse trees for the grammar. For example, we could replace the vwx subtree
with the w subtree, producing this parse tree for uwy:

u y

Or we could replace the w subtree with the vwx subtree, producing this parse
tree for uv2wx2y:

u v x y
A

v w x
184 CHAPTER FOURTEEN: THE CONTEXT-FREE FRONTIER

Then we could again replace the w subtree with the vwx subtree, producing this
parse tree for uv3wx3y:

u v x y
A

v x
A

v w x

Repeating this step any number of times, we can generate a parse tree for uviwxiy
for any i.

The previous lemma shows that pumping parse trees are very useful—find one, and you
can conclude that not only uvwxy L(G), but uviwxiy L(G) for all i. The next lemma
shows that they are not at all hard to find. To prove it, we will need to refer to the height of a
parse tree, which is defined as follows:

The height of a parse tree is the number of edges in the longest path from the
start symbol to any leaf.

Consider this (ambiguous) grammar, for example:

S S|S+S|S*S|a|b|c
These are parse trees of heights 1, 2, and 3, respectively:
 14.1 PUMPING PARSE TREES 185
S S S

a S + S S + S

a b S * S c

a b

We will also need the idea of a minimum-size parse tree for a given string.

A minimum-size parse tree for the string x in a grammar G is a parse tree in

G that generates x and has no more nodes than any other parse tree in G that
generates x.

The need for this definition arises because an ambiguous grammar can generate the same
string using more than one parse tree, and these parse trees might even have different
numbers of nodes. For example, the previous grammar generates the string a*b+c with the
parse tree shown above, but also with these ones:
S S

S * S S + S

a S + S S * S S

b c a b c

All three parse trees generate the same string, but the last one is not minimum size, since it
has one more node than the others. Every string in the language generated by a grammar
has at least one minimum-size parse tree; for an ambiguous grammar, some strings may have
more than one.
Now we’re ready to see why pumping parse trees are easy to find.

Lemma 14.1.2: Every CFG G = (V, , S, P) that generates an infinite language

generates a pumping parse tree.
Proof: Let G = (V, , S, P) be any CFG that generates an infinite language.
Each string in L(G) has at least one minimum-size parse tree; therefore, G
186 CHAPTER FOURTEEN: THE CONTEXT-FREE FRONTIER

generates infinitely many minimum-size parse trees. There are only finitely
many minimum-size parse trees of height |V | or less; therefore, G generates
a minimum-size parse tree of height greater than |V | (indeed, it generates
infinitely many). Such a parse tree must satisfy property 1 of pumping
parse trees, because on a path with more than |V| edges there must be some
nonterminal A that occurs at least twice. And such a parse tree must also satisfy
property 2, because it is a minimum-size parse tree; if it did not satisfy property
2, we could replace the ancestor A with the descendant A to produce a parse tree
with fewer nodes yielding the same string. Thus, G generates a pumping parse
tree.

This proof actually shows that every grammar for an infinite language generates not just one,
but infinitely many pumping parse trees. However, one is all we’ll need for the next proof.

14.2 The Language {anbncn}

We have seen that the language {anbn} is context free but not regular. Now we consider the
language {anbncn}: any number of as, followed by the same number of bs, followed by the
same number of cs.

Theorem 14.1: {anbncn} is not a CFL.

Proof: Let G = (V, , S, P) be any CFG over the alphabet {a, b, c }, and
suppose by way of contradiction that L(G) = {anbncn}. By Lemma 14.1.2, G
generates a pumping parse tree. Using Lemma 14.1.1, this means that for some
k we have akbkck = uvwxy, where v and x are not both ε and uv2wx2y is also in
L(G ). The substrings v and x must each contain only as, only bs, or only cs,
because otherwise uv2wx2y is not even in L(a*b*c*). But in that case uv2wx2y
has more than k copies of one or two symbols, but only k of the third. Thus
uv2wx2y {anbncn}. By contradiction, L(G) {anbncn}.

The tricky part of this proof is seeing that no matter how you break up a string akbkck
into substrings uvwxy, where v and x are not both ε, you must have uv2wx2y {anbncn}. If the
substrings v and/or x contain a mixture of symbols, as in this example:
a a a a a b b b b b c c c c c
u v w x y

then the resulting string uv2wx2y would have as after bs and/or bs after cs; for the example
above it would be aaaaabbaabbbbbcbbccccc, clearly not in {anbncn}. On the other hand, if the
substrings v and x contain at most one kind of symbol each, as in this example:
 14.3 CLOSURE PROPERTIES FOR CFLS 187

a a a a a b b b b b c c c c c
u v w x y

then the resulting string uv2wx2y would no longer have the same number of as, bs, and cs;
for the example above it would be aaaaaaabbbbbbccccc, clearly not in {anbncn}. Either way,
uv2wx2y {anbncn}.

14.3 Closure Properties for CFLs

Context-free languages are closed for some of the same common operations as regular
languages, including union, concatenation, Kleene star, and intersection with regular
languages. Closure under union and concatenation can be proved using constructions that
we saw, informally, back in Chapter 12.

Theorem 14.2.1: If L1 and L2 are any context-free languages, L1 L2 is also

context free.
Proof: By construction. If L1 and L2 are context free then, by definition, there
exist grammars G1 = (V1, 1, S1, P1) and G2 = (V2, 2, S2, P2) with L(G1) = L1
and L(G2) = L2. We can assume without a loss of generality that V1 and V2 are
disjoint. (Symbols shared between V1 and V2 could easily be renamed.) Now
consider the grammar G = (V, , S, P), where S is a new nonterminal symbol
and

V = V1 V2 {S},
= 1 2
, and
P = P1 P2 {(S S1), (S S2)}

Clearly L(G ) = L1 L2, so L1 L2 is a context-free language.

Theorem 14.2.2: If L1 and L2 are any context-free languages, L1L2 is also context
free.
Proof: By construction. If L1 and L2 are context free then, by definition, there
exist grammars G1 = (V1, 1, S1, P1) and G2 = (V2, 2, S2, P2) with L(G1) = L1
and L(G2) = L2. We can assume without a loss of generality that V1 and V2 are
disjoint. (Symbols shared between V1 and V2 could easily be renamed.) Now
consider the grammar G = (V, , S, P), where S is a new nonterminal symbol
and
188 CHAPTER FOURTEEN: THE CONTEXT-FREE FRONTIER

V = V1 V2 {S},
= 1 2
, and
P = P1 P2 {(S S1S2)}

Clearly L(G ) = L1L2, so L1L2 is a context-free language.

We can define the Kleene closure of a language in a way that parallels our use of the
Kleene star in regular expressions:

The Kleene closure of any language L is L* = {x1x2 ... xn | n 0, with all xi L}.

The fact that the context-free languages are closed under Kleene star can then be proved
with a CFG construction, as before:

Theorem 14.2.3: If L is any context-free language, L* is also context free.

Proof: By construction. If L is context free then, by definition, there exists
a grammar G = (V, , S, P) with L(G ) = L. Now consider the grammar
G' = (V', , S', P' ), where S' is a new nonterminal symbol and

V' = V {S'}, and

P' = P {(S' SS'), (S' ε)}

Now L(G' ) = L*, so L* is a context-free language.

One more closure property: the CFLs are closed under intersection with regular languages:

Theorem 14.2.4: If L1 is any context-free language and L2 is any regular

language, then L1 L2 is context free.
Proof sketch: This can be proved by a (somewhat hairy) construction that
combines a stack machine M1 for L1 with an NFA M2 for L2, producing a new
stack machine for L1 L2. It works a bit like the product construction: if M1’s
stack alphabet is , and M2’s state set is Q, the new stack machine can use an
element of Q as the symbol on the top of its stack to record both the M1’s
current top symbol and M2’s current state.
 14.4 NONCLOSURE PROPERTIES 189

14.4 Nonclosure Properties

In the previous section we saw that the CFLs are closed for some of the same operations
for which the regular languages are closed. Not all, however: the CFLs are not closed for
intersection or complement.

Theorem 14.3.1: The CFLs are not closed for intersection.

Proof: By counterexample. Consider these two CFGs G1 and G2:

S1 A 1 B1
A1 aA1b | ε
B1 cB1 | ε
S2 A 2 B2
A2 aA2 | ε
B2 bB2c | ε

Now L(G1) = {anbncm}, while L(G2) = {ambncn}. The intersection of the two
languages is {anbncn}, which we know is not context free. Thus, the CFLs are not
closed for intersection.

This nonclosure property does not mean that every intersection of CFLs fails to be a
CFL. In fact, we have already seen many cases in which the intersection of two CFLs is
another CFL. For example, we know that any intersection of regular languages is a regular
language—and they are all CFLs. We also know that the intersection of a CFL with a regular
language is a CFL. All the theorem says is that the intersection of two CFLs is not always a
CFL. Similarly, the complement of a CFL is sometimes, but not always, another CFL:

Theorem 14.3.2: The CFLs are not closed for complement.

Proof 1: By Theorem 14.2.1 we know that CFLs are closed for union.
Suppose by way of contradiction that they are also closed for complement. By
DeMorgan’s laws we have L1 ∩ L2 = L1 ∪ L2 . This defines intersection in terms
of union and complement, so the CFLs are closed for intersection. But this is a
contradiction of Theorem 14.3.1. By contradiction, we conclude that the CFLs
are not closed for complement.
Proof 2: By counterexample. Let L be the language {xx | x {a, b}*}, and
consider its complement, L = {x {a, b}* | x L}. Obviously, all odd-length
190 CHAPTER FOURTEEN: THE CONTEXT-FREE FRONTIER

strings are in L . Even-length strings are in L if and only if they have at least
one symbol in the first half that is not the same as the corresponding symbol
in the second half; that is, either there is an a in the first half and a b at the
corresponding position in the second half, or the reverse. In the first case the
string looks like this, for some i and j :

center

... a ... ... b ...

i symbols j symbols i symbols j symbols

At first this looks daunting: how could a grammar generate such strings? How
can we generate waxybz, where |w| = |y| = i and |x| = |z| = j? It doesn’t look
context free, until you realize that because the x and y parts can both be any
string in {a, b}*, we can swap them, instead picturing it as:

... a ... ... b ...

i symbols i symbols j symbols j symbols

Now we see that this actually is context free; it’s {way | |w| = |y|}, concatenated
with {xbz | |x| = |z|}. For the case with a b in the first half corresponding to an
a in the second half, we can just swap the two parts, getting {xbz | |x| = |z|}
concatenated with {way | |w| = |y|}.

To summarize, we can express L as the union of three sets:

L = {x {a, b}* | |x| is odd}

({way | |w | = |y|} concatenated with {xbz | |x| = |z |})
({xbz | |x | = |z|} concatenated with {way | |w | = |y|})

This CFG generates the language: O generates the odd-length part, and AB and
BA generate the two even-length parts.

S O | AB | BA
A XAX | a
 14.5 A PUMPING LEMMA 191
B XBX | b
O XXO | X
X a|b
We conclude that L is context free. But as we will show in Section 14.7, the
complement of L , L = {xx | x {a, b}*} is not context free. We conclude that the
CFLs are not closed for complement.

14.5 A Pumping Lemma

Let’s return to pumping parse trees and try to make a general-purpose pumping lemma
for CFLs. We have argued that any minimum-size tree of a height greater than |V | must
be a pumping parse tree, since on any path of more than |V | edges there must be some
nonterminal A that occurs at least twice. Of course, there may be many different places where
a nonterminal occurs as its own descendant. To make a useful pumping lemma for context-
free languages, we need to be a little more specific about how we are dividing the tree.

Lemma 14.2.1: For every grammar G = (V, , S, P), every minimum-size parse
tree of a height greater than |V | can be expressed as a pumping parse tree with
the properties shown:

height ≥ |V | + 1

height ≤ |V | + 1
A

u v w x y

Proof: Choose any path from root to leaf with more than |V | edges. Then,
working from the leaf back toward the root on this path, choose the first two
nodes that repeat a nonterminal, and let these be the nodes labeled A in the
diagram. As shown in the proof of Lemma 14.1.2, the result is a pumping parse
tree. Furthermore, the nonterminal A must have repeated within the first |V | + 1
edges, so the height of the subtree generating vwx is |V | + 1.
192 CHAPTER FOURTEEN: THE CONTEXT-FREE FRONTIER

Being able to give a bound on the height of a tree or subtree also enables us to give a
bound on the length of the string generated.

Lemma 14.2.2: For every CFG G = (V, , S, P) there exists some k greater than
the length of any string generated by any parse tree or subtree of height |V | + 1
or less.
Proof 1: For any given grammar, there are only finitely many parse trees and
subtrees of height |V | + 1 or less. Let k be one greater than the length of the
longest string so generated.
Proof 2: Let b be the length of the longest right-hand side of any production
in P. Then b is the maximum branching factor in any parse tree. So a tree or
subtree of height |V | + 1 can have at most b|V |+1 leaves. Let k = b|V |+1 + 1.

These two proofs give different values for k. That is not a problem, because what matters
here is not the value of k, but the existence of some k with the right properties. This was
also true of the value k we used in the pumping lemma for regular languages. To prove the
pumping lemma, we let k be the number of states in some DFA for the language. But once
the pumping lemma for regular languages was proved, the value of k used in the proof was
irrelevant, since the lemma itself claims only that some suitable k exists.
Putting these lemmas together with our previous results about pumping parse trees, we
get the pumping lemma for context-free languages.

Lemma 14.2.3 (The Pumping Lemma for Context-free Languages): For all
context-free languages L there exists some k N such that for all z L with
|z| k, there exist uvwxy such that
1. z = uvwxy,
2. v and x are not both ε,
3. |vwx| k, and
4. for all i, uviwxiy A.
Proof: We are given some CFL L. Let G be any CFG with L(G) = L, and let k
be as given for this grammar by Lemma 14.2.2. We are given some z L with
|z| k. Since all parse trees in G for z have a height greater than |V | + 1, that
includes a minimum-size parse tree for z. We can express this as a pumping parse
tree for z as given in Lemma 14.2.1. Since this is a parse tree for z, property
1 is satisfied; since it is a pumping parse tree, properties 2 and 4 are satisfied;
and since the subtree generating vwx is of height |V | + 1 or less, property 3 is
satisfied.
 14.6 PUMPING-LEMMA PROOFS 193
This pumping lemma shows once again how matching pairs are fundamental to the
context-free languages. Every sufficiently long string in a context-free language contains a
matching pair consisting of the two substrings v and x of the lemma. These substrings can be
pumped in tandem, always producing another string uv iwx iy in the language. Of course, one
or the other (but not both) of those substrings may be ε. In that way the pumping lemma for
context-free languages generalizes the one for regular languages, since if one of the strings is
ε, the other can be pumped by itself.

14.6 Pumping-Lemma Proofs

The pumping lemma is very useful for proving that languages are not context free. For
example, here is a pumping-lemma proof showing that {anbncn} is not context free (a fact we
already proved without using the pumping lemma). The steps in the proof are numbered for
future reference.
1. Proof is by contradiction using the pumping lemma for context-free languages.
Assume that L = {anbncn} is context free, so the pumping lemma holds for L. Let k
be as given by the pumping lemma.
2. Choose z = akbkck. Now z L and |z| k as required.
3. Let u, v, w, x, and y be as given by the pumping lemma, so that uvwxy = akbkck, v
and x are not both ε, |vwx| k, and for all i, uv iwx iy L.
4. Now consider pumping with i = 2. The substrings v and x cannot contain more
than one kind of symbol each—otherwise the string uv2wx2y would not even be in
L(a*b*c*). So the substrings v and x must fall within the string akbkck in one of these
ways:
ak bk ck
1. v x
2. v x
3. v x
4. v x
5. v x
6. v x

But in all these cases, since v and x are not both ε, pumping changes the number of
one or two of the symbols, but not all three. So uv2wx2y L.
5. This contradicts the pumping lemma. By contradiction, L = {anbncn} is not context
free.
The structure of this proof matches the structure of the pumping lemma itself. Here is
the overall structure of the pumping lemma, using the symbols (“for all”) and (“there
exists”):
194 CHAPTER FOURTEEN: THE CONTEXT-FREE FRONTIER

1. k ...
2. z ...
3. uvwxy ...
4. i 0 ...
The alternating and parts make every pumping-lemma proof into a kind of game, just
as we saw using the pumping lemma for regular languages. The parts (the natural number
k and the strings u, v, w, x, and y) are merely guaranteed to exist. In effect, the pumping
lemma itself makes these moves, choosing k, u, v, w, x, and y any way it wants. In steps
1 and 3 we could only say these values are “as given by the pumping lemma.” The parts,
on the other hand, are the moves you get to make: you can choose any values for z and i,
since the lemma holds for all such values. This pumping lemma offers fewer choices than the
pumping lemma for regular languages, and that actually makes the proofs a little harder to
do. The final step, showing that a contradiction is reached, is often more difficult, because
one has less control early in the proof.
You may have noticed that in that table in the proof—the one that shows how the strings
v and x can fall within the string akbkck—line 6 is unnecessary. It does not hurt to include it
in the proof, because all six cases lead to the same contradiction. But the substrings v and x
actually cannot be as far apart as shown in line 6, because that would make the combined
substring vwx more than k symbols long, while the pumping lemma guarantees that
|vwx| k. In many applications of the pumping lemma it is necessary to make use of this.
The following pumping-lemma proof is an example:

Theorem 14.4: L = {anbmcn | m n} is not context free.

Proof: By contradiction using the pumping lemma for context-free languages.
Assume that L = {anbmcn | m n} is context free, so the pumping lemma holds
for L. Let k be as given by the pumping lemma. Choose z = akbkck. Now
z L and |z| k as required. Let u, v, w, x, and y be as given by the pumping
lemma, so that uvwxy = akbkck, v and x are not both ε, |vwx| k , and for all
i, uv iwx iy L.
Now consider pumping with i = 2. The substrings v and x cannot contain more
than one kind of symbol each—otherwise the string uv2wx2y would not even be in
L(a*b*c*). So the substrings v and x must fall within the string akbkck in one of these
ways:
 14.7 THE LANGUAGE {xx} 195

ak bk ck
1. v x
2. v x
3. v x
4. v x
5. v x
6. v x

Because v and x are not both ε, uv2wx2y in case 1 has more as than cs. In case 2
there are either more as than cs, more bs than cs, or both. In case 3 there are
more bs than as and cs. In case 4 there are more cs than as, more bs than as, or
both. In case 5 there are more cs than as. Case 6 contradicts |vwx| k. Thus, all
cases contradict the pumping lemma. By contradiction, L = {anbmcn | m n} is
not context free.

This proof is very similar to the previous one, using the same string z and pumping
count i. This time, however, we needed to use the fact that |vwx| k to reach a contradiction
in case 6. Without that, no contradiction would be reached, since pumping v and x together
could make the number of as and the number of cs grow in step, which would still be a string
in the language L.

14.7 The Language {xx}

Consider the language of strings that consist of any string over a given alphabet , followed
by another copy of the same string: {xx | x *}. For = {a, b}, the language contains the
strings ε, aa, bb, abab, baba, aaaa, bbbb, and so on. We have seen that the related language
{xxR } is context free, though not regular for any alphabet with at least two symbols.

Theorem 14.5: {xx | x *} is not a CFL for any alphabet with at least two
symbols.
Proof: By contradiction using the pumping lemma for context-free languages.
Let be any alphabet containing at least two symbols (which we will refer to as
a and b). Assume that L = {xx | x *} is context free, so the pumping lemma
holds for L. Let k be as given by the pumping lemma. Choose z = akbkakbk. Now
z L and |z| k as required. Let u, v, w, x, and y be as given by the pumping
lemma, so that uvwxy = akbkakbk, v and x are not both ε, |vwx| k, and for all i,
uv iwx iy L.
196 CHAPTER FOURTEEN: THE CONTEXT-FREE FRONTIER

Now consider how the substrings v and x fall within the string. Because
|vwx| k, v and x cannot be widely separated, so we have these 13 cases:

ak bk ak bk
1. v x
2. v x
3. v x
4. v x
5. v x
6. v x
7. v x
8. v x
9. v x
10. v x
11. v x
12. v x
13. v x

For cases 1 through 5, we can choose i = 0. Then the string uv0wx0y is some sakbk
where |s| < 2k. In this string the last symbol of the first half is an a, while the last
symbol of the second half is a b. So uv0wx0y L.
For cases 9 through 13, we can again choose i = 0. This time the string uv0wx0y
is some akb ks where |s| < 2k. In this string the first symbol of the first half is an a,
while the first symbol of the second half is a b. So, again, uv0wx0y L.
Finally, for cases 6, 7, and 8, we can again choose i = 0. Now the string uv0wx0y
is some aksbk where |s| < 2k. But such an aksbk can’t be rr for any r, because if r
starts with k as and ends with k bs, we must have |r| 2k and so |rr| 4k, while
our |aksbk | < 4k. So, again, uv0wx0y L.
In every case we have a contradiction. By contradiction, L is not a context-free
language.

In Chapter 12 we saw the important role CFGs play in defining the syntax of
programming languages. Many languages, including Java, require variables to be declared
before they are used:
 EXERCISES 197
int fred = 0;
while (fred==0) {
...
}

The requirement that the same variable name fred occur in two places, once in a
declaration and again in a use, is not part of any grammar for Java. It is a non-context-free
construct in the same way that {xx | x *} is a non-context-free language—think of x here
as the variable name which must be the same in two places. Java compilers can enforce the
requirement easily enough, but doing so requires more computational power than can be
wired into any context-free grammar.
One final note about pumping-lemma proofs. In the proof above, as in all those shown
in this chapter, the choice of i was static. The proof above always chose i = 0, regardless of the
values of the other variables. This is not always possible. The pumping lemma permits the
choice of i to depend on the values of k, u, v, w, x, and y, just as the choice of z must depend
on k. For more challenging proofs it is often necessary to take advantage of that flexibility.
In terms of the pumping-lemma game, it is often necessary to see what moves the pumping
lemma makes before choosing the i that will lead to a contradiction.

Exercises
EXERCISE 1
If our definition of pumping parse trees included only the first, but not the second
property, the proof of Theorem 14.1 would fail to reach a contradiction. Explain exactly
how it would fail.
EXERCISE 2
Construct a grammar for an infinite language that generates a parse tree of a height
greater than |V | that is not a pumping parse tree. Show both the grammar and the parse
tree.
EXERCISE 3
Prove that {anbnc nd n} is not a CFL.
EXERCISE 4
Prove that {a nb 2nc n} is not a CFL.
EXERCISE 5
Show that {anbnc pd q} is a CFL by giving either a stack machine or a CFG for it.
EXERCISE 6
Show that {a nb mc n+m} is a CFL by giving either a stack machine or a CFG for it.
EXERCISE 7
Give a CFG for {anbn} {xxR | x {a, b}*}.
198 CHAPTER FOURTEEN: THE CONTEXT-FREE FRONTIER

EXERCISE 8
Give a stack machine for {anbn} {xxR | x {a, b}*}.
EXERCISE 9
Give a CFG for the set of all strings xy such that x {0n12n} and y is any string over the
alphabet {0, 1} that has an even number of 1s.
EXERCISE 10
Give a stack machine for the set of all strings xy such that x {0n12n} and y is any string
over the alphabet {0, 1} that has an even number of 1s.
EXERCISE 11
Give a CFG for {a ib na jb na k}*.
EXERCISE 12
Give a stack machine for {a ib na jb na k}*.
EXERCISE 13
Give a CFG for {anbn} {x {a, b}* | the number of as in x is even}.
EXERCISE 14
Give a stack machine for {anbn} {x {a, b}* | the number of as in x is even}.
EXERCISE 15
Using the pumping lemma for context-free languages, prove that {anbmcn | m n}
is not a CFL.
EXERCISE 16
Show that {anbmcn} is a CFL by giving either a stack machine or a CFG for it.
EXERCISE 17
Using the pumping lemma for context-free languages, prove that {anbmc p | n m
and m p} is not a CFL.
EXERCISE 18
Show that {anbmc p | n m} is a CFL by giving either a stack machine or a CFG for it.
EXERCISE 19
Let A = {x | x {a, b, c}* and x contains the same number of each of the three symbols}.
Using the pumping lemma for context-free languages, prove that A is not a CFL.
EXERCISE 20
Let B = {x | x {a, b, c}* and x contains the same number of at least two of the three
symbols}. Show that B is a CFL by giving either a stack machine or a CFG for it.
 15
CHA P TER

Stack Machine
Applications
The parse tree (or a simplified version, the abstract syntax tree) is
one of the central data structures of almost every compiler or other
programming language system. To parse a program is to find a parse
tree for it. Every time you compile a program, the compiler must
first parse it. Parsing algorithms are fundamentally related to stack
machines, as this chapter illustrates.

199
200 CHAPTER FIFTEEN: STACK MACHINE APPLICATIONS

15.1 Top-Down Parsing

In section 13.7 we saw a general procedure for constructing a stack machine from any
CFG. The constructed stack machine verifies that a string is in the language by finding
a derivation for that string in the given CFG. That makes it a useful point of departure
for applications that require parsing. The constructed stack machine works by repeatedly
applying productions until a fully terminal string is derived from the start symbol S. That is a
kind of top-down parsing: it builds a parse tree starting from S and working down. Of course,
our constructed stack machine doesn’t actually build anything; it just accepts or rejects by
following the steps of a leftmost derivation in the grammar. But a program implementing
such a stack machine could record the production applied for each nonterminal and thus
create the corresponding parse tree.
The construction of Section 13.7 is a good point of departure, but it is not
directly useful for parsing applications because the constructed stack machine is highly
nondeterministic. Whenever there is a nonterminal on top of the stack, the constructed
stack machine can apply any one of the productions for that nonterminal. It does not choose
deterministically which production to use for each nonterminal in the parse tree; it just relies
on nondeterminism, accepting the string if there is some choice that eventually works. For
parsing applications this could be implemented using backtracking of the kind we used when
implementing NFAs. But it would be much better for a compiler to be able to find the parse
tree in one deterministic pass over the input program, without having to search around for it.
For a certain class of grammars the nondeterminism of the top-down-parsing stack
machine can easily be avoided. For example, consider the CFL {xcxR | x {a, b}*}. It is
generated by this CFG:
S aSa | bSb | c
and the construction of Section 13.7 would give this stack machine for it:
read pop push
1. ε S aSa
2. ε S bSb
3. ε S c
4. a a ε
5. b b ε
6. c c ε

Although this stack machine is nondeterministic, closer inspection reveals that it is always
quite easy to choose the next move. For example, suppose the input string is abbcbba. There
are three possible first moves:
 15.1 TOP-DOWN PARSING 201
(abbcbba, S) 1
(abbcbba, aSa) …
(abbcbba, S) 2
(abbcbba, bSb) …
(abbcbba, S) 3
(abbcbba, c) …
But only the first of these has any future. The other two leave, on top of the stack, a terminal
symbol that is not the same as the next input symbol. From there, no further move will be
possible.
We can formulate simple rules about when to use those three moves:
• Use move 1 when the top stack symbol is S and the next input symbol is a.
• Use move 2 when the top stack symbol is S and the next input symbol is b.
• Use move 3 when the top stack symbol is S and the next input symbol is c.
These rules say when to apply those moves that derived from productions in the original
grammar. (The other stack-machine moves, the ones that read an input symbol and pop
a matching stack symbol, are already deterministic.) Our rules can be expressed as a two-
dimensional look-ahead table.
a b c $
S S aSa S bSb S c

The entry at table[A][ ] tells which production to use when the top of the stack is A and the
next input symbol is . The final column of the table, table[A][$], tells which production
to use when the top of the stack is A and all the input has been read. For our example,
table[S][$] is empty, showing that if you still have S on top of the stack but all the input has
been read, there is no way to proceed.
Using the table requires one symbol of look-ahead: when there is a nonterminal symbol
on top of the stack, we need to look ahead to see what the next input symbol will be, without
actually consuming it. With the help of this table our stack machine becomes deterministic.
Here is the idea in pseudocode:

1. void predictiveParse(table, S) {
2. initialize a stack containing just S
3. while (the stack is not empty) {
4. A = the top symbol on stack
5. c = the current symbol in input (or $ at the end)
6. if (A is a terminal symbol) {
7. if (A != c) the parse fails
8. pop A and advance input to the next symbol
9. }
10. else {
11. if table[A][c] is empty the parse fails
202 CHAPTER FIFTEEN: STACK MACHINE APPLICATIONS

12. pop A and push the right-hand side of table[A][c]

13. }
14. }
15. if input is not finished the parse fails
16. }

This treats input as a global variable: the source of the input string to be parsed.
The code scans the input from left to right, one symbol at a time, maintaining (implicitly)
the current position of the scan. Line 5 peeks at the current input symbol, but only line 8
advances to the next symbol. We assume there is a special symbol, not the same as any
terminal symbol, marking the end of the string; call this symbol $. The parameter table is
the two-dimensional table of productions, as already described.
Our predictiveParse is an example of a family of top-down techniques called
LL(1) parsers. The first L signifies a left-to-right scan of the input, the second L signifies
that the parse follows the order of a leftmost derivation, and the 1 signifies that one symbol
of input look-ahead is used (along with the top symbol of the stack) to choose the next
production. It is a simple parsing algorithm; the catch is the parse table it requires. Such
parse tables exist for some grammars but not for others—which means that LL(1) parsing is
possible for some grammars but not for others. LL(1) grammars are those for which LL(1)
parsing is possible; LL(1) languages are those that have LL(1) grammars.
LL(1) grammars can often be constructed for programming languages. Unfortunately,
they tend to be rather contorted. Consider this simple grammar and the little language of
expressions it generates:
S (S) | S + S | S * S | a | b | c
As we saw in Chapter 12, that grammar is ambiguous, and no ambiguous grammar is LL(1).
But even this reasonably simple unambiguous grammar for the language fails to be LL(1):
S S+R|R
R R*X|X
X (S) | a | b | c
In this case the grammar’s problem is that it is left recursive: it includes productions like
S S + R that replace a nonterminal with a string starting with that same nonterminal.
No left-recursive grammar is LL(1). This LL(1) grammar for the same language manages
to avoid left recursion, at some cost in clarity:
S AR
R +AR | ε
A XB
 15.2 RECURSIVE-DESCENT PARSING 203
B *XB | ε
X (S) | a | b | c
Once we have an LL(1) grammar, we still need to construct the LL(1) parse table for it. That
construction can be rather subtle. Here, for example, is the parse table that goes with the
grammar above:
a b c + * ( ) $
S S AR S AR S AR S AR
R R +AR R ε R ε
A A XB A XB A XB A XB
B B ε B *XB B ε B ε
X X a X b X c X (S)

It is instructive to consider these entries carefully. For example, the production S AR is in

table[S][a] not because it generates the terminal symbol a directly, but because using it
can eventually generate a string that starts with a. The production B ε is in table[B][)]
not because it generates the terminal symbol ), but because it can reveal a ) beneath the B on
the stack.
There is an algorithm for constructing an LL(1) parse table, given an LL(1) grammar.
We will not see that algorithm here, but the interested reader should see “Further Reading”
(Section 15.5) at the end of this chapter.

15.2 Recursive-Descent Parsing

An LL(1) parser can also be implemented using recursive functions, with one function for
parsing each nonterminal in the grammar. Consider again our simple grammar for
{xcxR | x {a, b}*}:
S aSa | bSb | c
We first implement a utility routine called match:

void match(x) {
c = the current symbol in input
if (c!=x) the parse fails
advance input to the next symbol
}

All match does is check that the expected nonterminal really does occur next in the input
and advance past it. Now we implement a function for deriving a parse tree for S:
204 CHAPTER FIFTEEN: STACK MACHINE APPLICATIONS

void parse_S() {
c = the current symbol in input (or $ at the end)
if (c=='a') { // production S aSa
match('a'); parse_S(); match('a');
}
else if (c=='b') { // production S bSb
match('b'); parse_S(); match('b');
}
else if (c=='c') { // production S c
match('c');
}
else the parse fails
}

This code is an LL(1) parser, not really that different from predictiveParse, our
previous, table-driven implementation. Both are left-to-right, top-down parsers that choose a
production to apply based on the current nonterminal and the current symbol in the input.
But now the information from the parse table is incorporated into the code, which works
recursively. This is called a recursive-descent parser.
This is part of a recursive-descent parser for our LL(1) grammar of expressions:

void parse_S() {
c = the current symbol in input (or $ at the end)
if (c=='a' || c=='b' ||
c=='c' || c=='(') { // production S AR
parse_A(); parse_R();
}
else the parse fails
}

void parse_R() {
c = the current symbol in input (or $ at the end)
if (c=='+') { // production R +AR
match('+'); parse_A(); parse_R();
 15.3 BOTTOM-UP PARSING 205
}
else if (c==')' || c=='$') { // production R ε
}
else the parse fails
}

Additional functions parse_A, parse_B, and parse_X constructed on the same lines
would complete the parser.
Compare the recursive-descent parser with the previous table-driven LL(1) parser. Both
are top-down, LL(1) techniques that work only with LL(1) grammars. The table-driven
method uses an explicit parse table, while recursive descent uses a separate function for each
nonterminal. The table-driven method uses an explicit stack—but what happened to the
stack in the recursive-descent method? It’s still there, but hidden. Where a table-driven, top-
down parser uses an explicit stack, a recursive-descent parser uses the language system’s call
stack. Each function call implicitly pushes and each return implicitly pops the call stack.

15.3 Bottom-Up Parsing

In the last two sections we saw some top-down parsing methods, which develop a parse
tree by starting at the root. It is also possible to build a parse tree from the bottom up. An
important bottom-up parsing method is shift-reduce parsing, so called because it involves two
basic kinds of moves:
1. (shift) Push the current input symbol onto the stack and advance to the next input
symbol.
2. (reduce) There is a string on top of the stack that is the right-hand side of some
production. Pop it off and replace it with the nonterminal symbol from the left-
hand side of that production.
The reduce move is exactly the reverse of what our predictive parser did. The predictive
parser replaced a nonterminal symbol on the stack with the right-hand side of some
production for that nonterminal. A bottom-up parser replaces the right-hand side of some
production on the stack with its nonterminal symbol.
For example, a shift-reduce parser for our grammar
S aSa | bSb | c
would parse the string abbcbba using a sequence of moves like this:
206 CHAPTER FIFTEEN: STACK MACHINE APPLICATIONS

Input Stack Next Move

abbcbba$ ε shift
abbcbba$ a shift
abbcbba$ ba shift
abbcbba$ bba shift
abbcbba$ cbba reduce by S c
aaacbbb$ Sbba shift
abbcbba$ bSbba reduce by S bSb
abbcbba$ Sba shift
abbcbba$ bSba reduce by S bSb
abbcbba$ Sa shift
abbcbba$ aSa reduce by S aSa
abbcbba$ S

In this table the current input symbol is underlined and, for reduce moves, the salient
substring of the stack is underlined as well. As you can see, the parser does not get around to
the top of the parse tree, the root symbol S, until the final step. It builds the parse tree from
the bottom up.
A popular kind of shift-reduce parser is the LR(1) parser. The L signifies a left-to-right
scan of the input, the R signifies that the parse follows the order of a rightmost derivation
in reverse, and the 1 signifies that one symbol of input look-ahead is used (along with the
string on top of the stack) to select the next move. These techniques are considerably trickier
than LL(1) techniques. One difficulty is that reduce moves must operate on the top-of-the-
stack string, not just the top-of-the-stack symbol. Making this efficient requires some close
attention. (One implementation trick uses stacked DFA state numbers to avoid expensive
string comparisons in the stack.)
Grammars that can be parsed this way are called LR(1) grammars, and languages that
have LR(1) grammars are called LR(1) languages. In spite of their complexity, LR(1) parsers
are quite popular, chiefly because the class of LR(1) grammars is quite broad. The LR(1)
grammars include all the LL(1) grammars, and many others as well. Programming language
constructs can usually be expressed with LR(1) grammars that are reasonably readable;
making a grammar LR(1) usually does not require as many contortions as making it LL(1).
(For example, LR(1) grammars need not avoid left-recursive productions.)
LR(1) parsers are complicated—almost always too complicated to be written without
the help of special tools. There are many tools for generating LR(1) parsers automatically.
 15.4 PDAS, DPDAS, AND DCFLS 207
A popular one is the Unix tool called yacc, which works a bit like the lex tool we saw in
Chapter 8. It converts a context-free grammar into C code for an LR(1) parser. In the yacc
input file, each production in the grammar is followed by a piece of C code, the action,
which is incorporated into the generated parser and is executed by the parser whenever a
reduce step is made using that production. For most compilers, this action involves the
construction of a parse tree or abstract syntax tree, which is used in subsequent (hand-coded)
phases of compilation.
Although they are complicated, LR(1) techniques have very good efficiency—like LL(1)
techniques, they take time that is essentially proportional to the length of the program being
parsed. Beyond LR(1) techniques there are many other parsing algorithms, including some
that have no restrictions at all and work with any CFG. The Cocke-Kasami-Younger (CKY)
algorithm, for example, parses deterministically and works with any CFG. It is quite a simple
algorithm too—much simpler than LR(1) parsing. The drawback is its lack of efficiency; the
CKY algorithm takes time proportional to the cube of the length of the string being parsed,
and that isn’t nearly good enough for compilers and other programming-language tools.

15.4 PDAs, DPDAs, and DCFLs

The DFA and NFA models are quite standardized, and most textbooks give more or less the
same definitions for those formalisms. For stack-based automata the differences are more
noticeable.
One widely studied kind of automaton is the push-down automaton (PDA). A PDA is
like a finite-state machine augmented with an unbounded stack of symbols. Each move made
by a PDA depends not only on the current state and input symbol, but also on the symbol
that appears on the top of its stack. Each move made by a PDA not only changes the current
state like a finite-state machine, but also changes the stack like a stack machine.
PDAs can be illustrated with state-transition diagrams. A state transition must show not
only the input symbol a to be read, but also the symbol Z to be popped off the stack and the
string of symbols x to be pushed back on in its place. Such a transition might be illustrated
like this:

a,Z/x
q r

This says that if the PDA is in the state q, and the current input symbol is a, and the symbol
on top of the stack is Z, the PDA can go to the state r and replace the Z on the stack with the
string of symbols x.
There is considerable variety among PDA models. Some accept a string by emptying
their stack, like a stack machine; some accept by ending in an accepting state, like an NFA;
some must do both to accept. Some start with a special symbol on the stack; some start with
208 CHAPTER FIFTEEN: STACK MACHINE APPLICATIONS

a special symbol marking the end of the input string; some do both; some do neither. All
these minor variations end up defining the same set of languages; the languages that can be
recognized by a PDA are exactly the CFLs.
The formal definition for a PDA is more complicated than the formal definition for
a stack machine, particularly in the transition function. The set of languages that can be
recognized is the same as for a stack machine. So why bother with PDAs at all? One reason is
because they are useful for certain kinds of proofs. For example, the proof that the CFLs are
closed for intersection with regular languages is easier using PDAs, since you can essentially
perform the product construction on the two state machines while preserving the stack
transitions of the PDA. Another reason is that they have a narrative value: they make a good
story. You start with DFAs. Add nondeterminism, and you have NFAs. Does that enlarge
the class of languages you can define? No! Then add a stack, and you have PDAs. Does that
enlarge the class of languages you can define? Yes! What is the new class? The CFLs!
But perhaps the most important reason for studying PDAs is that they have an
interesting deterministic variety. When we studied finite automata, we saw that the transition
function for an NFA gives zero or more possible moves from each configuration, while the
transition function for a DFA always gives exactly one move until the end of the string is
reached. One can think of DFAs as a restricted subset of NFAs. Restricting PDAs to the
deterministic case yields another formalism for defining languages: the deterministic PDA
(DPDA). For technical reasons, DPDAs are usually defined in a way that allows them to get
stuck in some configurations, so a DPDA always has at most one possible move. But like a
DFA, it never faces a choice of moves, so it defines a simple computational procedure for
testing language membership.
When we studied finite automata, we discovered that adding nondeterminism did not
add any definitional power; NFAs and DFAs can define exactly the same set of languages,
the regular languages. For PDAs, the case is different; DPDAs are strictly weaker than PDAs.
DPDAs define a separate, smaller class of languages: the deterministic context-free languages.

A deterministic context-free language (DCFL) is a language that is L(M) for some

DPDA M.

The DCFLs include all the regular languages and more:

 15.5 FURTHER READING 209

L(a*b*)
regular
languages {xx R | x ∈{a,b}* }

DCFLs
CFLs
{anbn}

For example, the language {xxR | x {a, b}*} is a CFL but not a DCFL. That makes
intuitive sense; a parser cannot know where the center of the string is, so it cannot make
a deterministic decision about when to stop pushing and start popping. By contrast, the
language {xcxR | x {a, b}*} is a DCFL.
The DCFLs have their own closure properties, different from the closure properties of
CFLs. Unlike the CFLs, the DCFLs are not closed for union: the union of two DCFLs is not
necessarily a DCFL, though of course it is always a CFL. Unlike the CFLs, the DCFLs are
closed for complement: the complement of a DCFL is always another DCFL. These different
closure properties can be used to help prove that a given CFL is not a DCFL. Such proofs are
usually quite tricky, since there is no tool like the pumping lemma specifically for DCFLs.
Part of the reason why we view the regular languages and the CFLs as important
language classes is that they keep turning up. Regular languages turn up in DFAs, NFAs,
regular expressions, and right-linear grammars, as well as many other places not mentioned
in this book. CFLs turn up in CFGs, stack machines, and PDAs. DCFLs get this kind of
validation as well. It turns out that they are the same as a class we have already met; the
DCFLs are the same as the LR(1) languages. These are exactly the languages that can be
quickly parsed using deterministic, bottom-up techniques.

15.5 Further Reading

There is a vast literature on parsing techniques. For a good place to start, see the classic
Aho, Alfred V., Ravi Sethi, and Jeffrey D. Ullman. Compilers: Principles,
Techniques, and Tools. Menlo Park, CA: Addison-Wesley, 1986.
For more information about yacc, a useful reference is
Levine, John R., Tony Mason, and Doug Brown. lex & yacc. Sebastopol, CA:
O’Reilly & Associates, 1992.
A classic treatment of PDAs and DPDAs may be found in
Hopcroft, John, Rajeev Motwani, and Jeffrey Ullman. Introduction to Automata
Theory, Languages, and Computation, 2nd ed. Menlo Park, CA: Addison-Wesley,
2001.
210 CHAPTER FIFTEEN: STACK MACHINE APPLICATIONS

Exercises
EXERCISE 1
Implement a predictive parser in Java for the language of expressions discussed in
Section 15.1, using the grammar and parse table as given. Your Java class PParse
should have a main method that takes a string from the command line and reports the
sequence of productions (used at line 12 in the predictiveParse pseudocode) used
to parse it. If the parse fails, it should say so. For example, it should have this behavior:

> java PParse a+b*c

S -> AR
A -> XB
X -> a
B ->
R -> +AR
A -> XB
X -> b
B -> *XB
X -> c
B ->
R ->
> java PParse a++
S -> AR
A -> XB
X -> a
B ->
R -> +AR
parse fails

(Unix systems usually preprocess input from the command line and give special
treatment to symbols like (, ), and *. On these systems, you should enclose the
argument to PParse in quotation marks, as in java PParse "a+b*c".)
EXERCISE 2
Complete the pseudocode for the recursive-descent parser begun in Section 15.2.
EXERCISE 3
Implement a recursive-descent parser in Java for the language of expressions discussed
in Sections 15.1 and 15.2, using the grammar given. Your Java class RDParse should
have a main method whose visible behavior is exactly like that of the main method of
PParse in Exercise 1.
 EXERCISES 211
EXERCISE 4
A stack machine is deterministic if it never has more than one possible move. To be
deterministic, a stack machine must meet two requirements:
1. For all {ε} and A , | ( , A)| 1. (No two entries in the table have the
same “read” and “pop” columns.)
2. For all A , if | (ε, A)| > 0 then for all c , | (c, A)| = 0. (If there is an ε-move
for some stack symbol A, there can be no other move for A.)
For example, consider this machine for the language L(a*b):
read pop push
1. ε S aS
2. ε S b
3. a a ε
4. b b ε

This is not deterministic, because moves 1 and 2 are both possible whenever there is an S
on top of the stack. A deterministic stack machine for the same language would be
read pop push
1. a S S
2. b S ε

For each of the following languages, give a deterministic stack machine:

a. L(01*0)
b. {xcxR | x {a, b}*}
c. {anbn | n > 0}
EXERCISE 5
Deterministic stack machines, as defined in the previous exercise, are not powerful
enough to recognize all the DCFLs, or even all the regular languages.
a. Prove that the language {a, ε} cannot be defined by a deterministic stack machine.
b. Let L be any language for which there are two strings x and y, with x L, xy L,
and y ε. Prove that L cannot be defined by a deterministic stack machine.
(Languages with no two such strings are said to have the prefix property.)
EXERCISE 6
To increase the power of the deterministic stack machines defined in Exercise 4, add an
additional kind of move:
212 CHAPTER FIFTEEN: STACK MACHINE APPLICATIONS

read pop push

1. $ A x

Here, $ in the “read” column signifies a move that can be made only when all the input
has been read. The entry above says that if all the input has been read and if there is an
A on top of the stack, you may pop the A and push x in its place. Using this additional
kind of move (along with the others), give deterministic stack machines for the following
languages. (These languages lack the prefix property, so they cannot be defined using
plain deterministic stack machines.)
a. L((0 + 1) * 0)
b. {anbn}
EXERCISE 7
Using the extended definition of a deterministic stack machine from Exercise 6, prove by
construction that every regular language is defined by some deterministic stack machine.
(Hint: Start from a DFA for the language.) Include a detailed proof that the stack
machine you construct works as required.
EXERCISE 8
(Note to instructors: Read the sample solution before assigning this one!)
In this exercise you will write a recursive-descent parser for a language of quantified
Boolean formulas. Your parser will compile each string into an object representing the
formula. Each such object will implement the following interface:

/*
* An interface for quantified boolean formulas.
*/
public interface QBF {
/*
* A formula can convert itself to a string.
*/
String toString();
}

The simplest of these formulas are variables, which are single, lowercase letters, like x
and y. These will be compiled into objects of the Variable class:

/*
* A QBF that is a reference to a variable, v.
 EXERCISES 213
*/
public class Variable implements QBF {
private char v; // the variable to which we refer
public Variable(char v) {
this.v = v;
}
public String toString() {
return "" + v;
}
}

Any QBF can be logically negated using the ~ operator, as in ~a. Logical negations will
be compiled into objects of the Complement class:

/*
* A QBF for a complement: not e.
*/
public class Complement implements QBF {
private QBF e; // the QBF we complement
public Complement(QBF e) {
this.e = e;
}
public String toString() {
return "~(" + e + ")";;
}
}

Any two QBFs can be logically ANDed using the * operator, as in a*b. These formulas
will be compiled into objects of the Conjunction class:

/*
* A QBF for a conjunction: lhs and rhs.
*/
public class Conjunction implements QBF {
private QBF lhs; // the left operand
private QBF rhs; // the right operand
public Conjunction(QBF lhs, QBF rhs) {
this.lhs = lhs;
this.rhs = rhs;
214 CHAPTER FIFTEEN: STACK MACHINE APPLICATIONS

}
public String toString() {
return "(" + lhs + "*" + rhs + ")";
}
}

Similarly, any two QBFs can be logically ORed using the + operator, as in a+b. These
formulas will be compiled into objects of the Disjunction class:

/*
* A QBF for a disjunction: lhs or rhs.
*/
public class Disjunction implements QBF {
private QBF lhs; // the left operand
private QBF rhs; // the right operand
public Disjunction(QBF lhs, QBF rhs) {
this.lhs = lhs;
this.rhs = rhs;
}
public String toString() {
return "(" + lhs + "+" + rhs + ")";
}
}

Any QBF can be placed in the scope of a universal quantifier, as in Ax(x+~x).

Syntactically, this is the symbol A, followed by a variable, followed by a QBF. The
formula is true if for all bindings of x (that is, for both x=true and x=false) the
inner formula is true. Universally quantified formulas are compiled into objects of the
Universal class:

/*
* A QBF that is universally quantified: for all v, e.
*/
public class Universal implements QBF {
private char v; // our variable
private QBF e; // the quantified formula
public Universal(char v, QBF e) {
this.v = v;
this.e = e;
 EXERCISES 215
}
public String toString() {
return "A" + v + "(" + e + ")";
}
}

Similarly, any QBF can be placed in the scope of an existential quantifier, as in

Ex(x+~x). Syntactically, this is the symbol E, followed by a variable, followed by a
QBF. The formula is true if for some binding of x (that is, for at least one of x=true
and x=false) the inner formula is true. Existentially quantified formulas are compiled
into objects of the Existential class:

/*
* A QBF that is existentially quantified:
* there exists v such that e.
*/
public class Existential implements QBF {
private char v; // our variable
private QBF e; // the quantified formula
public Existential(char v, QBF e) {
this.v = v;
this.e = e;
}
public String toString() {
return "E" + v + "(" + e + ")";
}
}

This is a BNF grammar for the language, using <QBF> as the starting nonterminal:
<QBF> ::= A <V> <QBF> | E <V> <QBF> | <BF>
<BF> ::= <A> <R>
<R> ::= + <A> <R> | <empty>
<A> ::= <X> <B>
<B> ::= * <X> <B> | <empty>
<X> ::= ~ <X> | ( <QBF> ) | <V>
<V> ::= a | b | c | … | z
The alphabet of the language is exactly as shown, with no spaces. The grammar
establishes the precedence: ~ has highest precedence, and quantifiers have lowest
216 CHAPTER FIFTEEN: STACK MACHINE APPLICATIONS

precedence. Notice that a QBF may be a parenthesized subformula, as in (a+b)*c.

Parentheses serve to override the operator precedence, but they are not compiled into
separate objects; the formulas Axx and Ax(x) should compile into exactly the same
objects.
Write a Java application QBFParser, using the classes and grammar above, that
compiles a QBF formula entered on the command line. Use a recursive-descent parser,
with one parsing method for each nonterminal in the BNF grammar. Write it so that
each of the parsing methods for a nonterminal returns an object of one of the classes
above, representing the subformula compiled from that nonterminal. Your application
should print out the final object compiled, using this as the main method:

public static void main(String[] args) {

QBF q = parse(args[0]);
System.out.println(q);
}

The QBF classes already have toString methods that will convert them back into
strings for printing. So, if your compiler works correctly, the printed output will look
much like the input, except that there will be exactly one pair of parentheses for each
operator and each quantifier. Thus, you should see

> java QBFParser 'Axx'

Ax(x)
> java QBFParser 'AxEyAzx+y*z'
Ax(Ey(Az((x+(y*z)))))
> java QBFParser '((x))'
x

Your code will compile QBF formulas, but it won’t do anything with them other than
print them out. Just throw an Error if the string does not parse. You should not have
to add anything to the QBF classes. (A later exercise, at the end of Chapter 20, will add
methods to the QBF classes, so that a formula can decide whether or not it is true.)
 16
CHA P TER

Turing Machines

We now turn to the most powerful kind of automaton we will study:

the Turing machine. Although it is only slightly more complicated
than a finite-state machine, a Turing machine can do much more. It
is, in fact, so powerful that it is the accepted champion of automata.
No more powerful model exists.
The Turing machine is the strongest of computational mechanisms,
the automaton of steel. But like all superheroes it does have one
weakness, revealed in this chapter: it can get stuck in an infinite
loop.

217
218 CHAPTER SIXTEEN: TURING MACHINES

16.1 Turing Machine Basics

A Turing machine (TM) can be pictured like this:

TM state
machine

... B B x1 x2 ... xn-1 xn B B ...

The TM’s input is presented on a tape with one symbol at each position. The tape extends
infinitely in both directions; those positions not occupied by the input contain a special
blank-cell symbol B. The TM has a head that can read and write symbols on the tape and
can move in both directions. The head starts at the first symbol in the input string.
A state machine controls the read/write head. Each move of the TM is determined by
the current state and the current input symbol. To make a move, the TM writes a symbol at
the current position, makes a state transition, and moves the head one position, either left
or right. If the TM enters an accepting state, it halts and accepts the input string. It does not
matter what the TM leaves on its tape or where it leaves the head; it does not even matter
whether the TM has read all the input symbols. As soon as the TM enters an accepting state,
it halts and accepts. This mechanism for accepting is quite different from that of DFAs and
NFAs. DFAs and NFAs can make transitions out of their accepting states, proceeding until
all the input has been read, and they often have more than one accepting state. In a TM,
transitions leaving an accepting state are never used, so there is never any real need for more
than one accepting state.
We will draw TMs using state-transition diagrams. Each TM transition moves the head
either to the right or to the left, as illustrated in these two forms:

a/b,R
q r

a/b,L
q r

The first says that if the TM is in the state q and the current tape symbol is a, the TM can
write b over the a, move the head one place to the right, and go to the state r. The second is
the same except that the head moves one place to the left.
As you can see, TMs are not much more complicated than DFAs. TMs are deterministic
and have no ε-transitions. Unlike a stack machines, TMs do not have a separate location for
 1.2 SIMPLE TMS 219
storage. They use one tape both for input and for scratch memory. But they have a far richer
variety of behaviors than DFAs or stack machines and (as we will see) can define a far broader
class of languages.

16.2 Simple TMs

TMs can easily handle all regular languages. In fact, a TM that always moves the head to the
right works much like a DFA.
For example, consider the language L(a*b*c*). This TM for the language just makes one
state transition on each symbol, much like a DFA:
b/b,R

b/b,R
a/a,R

c/c,R
B/ B,R c/c,R

B/B,R

c/c,R

B/B,R

This machine does not take advantage of a TM’s ability to move the head in both directions;
it always moves right. It also does not take advantage of a TM’s ability to write on the tape;
it always rewrites the same symbol just read. Of course, since it never moves to the left, it
makes no difference what it writes on the tape. Any symbols written are left behind and
never revisited. So it could just as well write a B over every symbol, erasing as it reads.
TMs can also easily handle all context-free languages. For example, consider the language
{a b }. Here is a series of steps for a TM that accepts this language:
n n

1. If the current symbol is B, go to Step 5. If the current symbol is an a, write a B over

it and go to Step 2.
2. Move right past any as and bs. At the first B, move left one symbol and go to Step 3.
3. If the current symbol is b, write a B over it and go to Step 4.
4. Move left past any as and bs. At the first B, move right one symbol and go to Step 1.
5. Accept.
This TM works by repeatedly erasing the first symbol (a) and the last symbol (b) from
the input string. If the input string was in {anbn}, this process eventually erases the entire
input string, leaving a blank tape. Here is an illustration of the TM, with states numbered to
match the steps above.
220 CHAPTER SIXTEEN: TURING MACHINES

a/a,R
b/b,R

B/B,R a/ B,R
5 1 2

B/ B,L
B/ B,R

b/ B,L
4 3

a/a,L
b/b,L

It is also possible to take any stack machine and convert it into an equivalent TM that
uses the infinite tape as a stack. But, as in the example above, it is often easier to find some
non-stack-oriented approach.

16.3 A TM for {anbncn}

Consider the language {anbncn}, which as we have seen is not a context-free language. Here is
a series of steps for a TM that accepts this language:
1. If the current symbol is B, go to Step 7. If the current symbol is a, write an X over it
and go to Step 2.
2. Move right past any as and Ys. At the first b, write Y over it and go to Step 3.
3. Move right past any bs and Zs. At the first c, write Z over it and go to Step 4.
4. Move left past any as, bs, Zs, and Ys. At the first X, move right one symbol and go to
Step 5.
5. If the current symbol is a, write an X and go to step 2. If the current symbol is Y go
to Step 6.
6. Move right past any Ys and Zs. At the first B, go to Step 7.
7. Accept.
The strategy followed by this TM is a simple one. Suppose for example that our input string
is aaabbbccc. In Steps 1, 2, and 3 it overwrites the first a with X, then overwrites the first b
with Y and the first c with Z. The resulting string is XaaYbbZcc. In Step 4 the TM moves
back left to the X, then one place right to the first remaining a. Now, using Steps 5, 2, and 3,
it overwrites another a, b, and c, producing the string XXaYYbZZc. Then the whole process
repeats one more time, producing the string XXXYYYZZZ. After that, Step 5 finds that the
first symbol to the right of the Xs is a Y. This indicates that all the as have been overwritten.
 16.3 A TM FOR {anbncn} 221
Step 6 then checks that there is nothing left but Ys and Zs, indicating that the number of bs
and cs was the same as the number of as. Then in step 7 the machine accepts.
If the input string is not in {anbncn} the TM gets stuck somewhere and does not reach
Step 7. For example, if the input string is acb, the TM overwrites the a with an X, but it
cannot handle the c it sees in Step 2 and gets stuck there. Try it for yourself on some other
strings like aababbccc and aaabbbbccc, and you will see that if the input string does not have
the same number of as, bs, and cs, or if it is not in L(a*b*c*) at all, the TM cannot reach
Step 7.
Here is an illustration of the TM for {anbncn} that we just outlined. The states are
numbered to match the steps in the outline.

1
B/B,R

a/X,R b/b,R
Z/Z,R

a/a,R b/Y,R
Y/Y,R 2 3

7 c/Z,L
a/X,R

X/X,R
5 4

a/a,L
Y/Y,R b/b,L
Y/Y,L
Z/Z,L

B/B,R 6

Y/Y,R
Z/Z,R
222 CHAPTER SIXTEEN: TURING MACHINES

16.4 The 7-Tuple

A TM M is a 7-tuple M = (Q, , , , B, q0, F ), where

Q is the finite set of states
is the input alphabet
is the tape alphabet, with and Q = {}
(Q Q {L, R}) is the transition function
B is the blank symbol, B ,B
q0 Q is the start state
F Q is the set of accepting states

Note that the tape alphabet includes all of the input alphabet, plus at least one additional
symbol B. The requirement that Q = {} means no state is also a symbol—not an onerous
requirement, but a necessary one for the definition of IDs that follows.
The transitions of the TM are defined by the transition function . The function
takes as parameters a state q Q (the current state) and a symbol X (the symbol at the
current position of the head). The value (q, X) is a triple (p, Y, D), where p Q is the
next state, Y is the symbol to write at the current head position, and D is one of the
directions L or R, indicating whether the head should move left or right. Note that a TM
is deterministic in the sense that it has at most one transition at any point. However, the
function need not be defined over its whole domain, so there may be some q and X with no
move (q, X).

16.5 The Languages Defined by a TM

An instantaneous description for a TM represents its entire configuration in a single string:
the contents of the tape, the state, and the position of the head. The following definitions are
all made with respect to some fixed TM M = (Q, , , , B, q0, F ):

An instantaneous description for a TM is a string xqy, where x * represents

the tape to the left of the head, q is the current state, and y * represents the
tape at and to the right of the head. In x, leading Bs from the tape are omitted,
except when there are only Bs to the left of the head, in which case x = B.
Likewise, in y, trailing Bs from the tape are omitted, except when there are only
Bs at and to the right of the head, in which case y = B. We define a function
idfix(z) that normalizes an ID string z in this way, removing (or adding) leading
and trailing Bs as necessary.

Although the tape extends infinitely in both directions, it contains only Bs in those positions
that are outside the input string and not yet written by the TM. So, because the leading Bs
on the left and the trailing Bs on the right are suppressed, an ID is always a finite string. The
 16.6 TO HALT OR NOT TO HALT 223
state q in the ID string tells what state the TM is in and also shows the position of the head.
The function for a TM determines a relation on IDs; we write I J if I is an ID
and J is an ID that follows from I after one move of the TM.

is a relation on IDs, defined by the function for the TM. For any
x *, c , q Q, a , and y *, we have
1. Left moves: if (q, a) = (p, b, L) then xcqay idfix(xpcby)
2. Right moves: if (q, a) = (p, b, R) then xcqay idfix(xcbpy)

For example, consider making a left move. Suppose the machine is in the configuration
given by the ID xcqay. So it is in the state q, reading the tape symbol a. Thus (q, a) gives
the transition. If that transition is (q, a) = (p, b, L), the machine can write a b over the a, go
to the state p, and move the head one position to the left. So after the move the ID is xpcby:
the a has been replaced by a b, and the head has moved left so that it is now reading the tape
symbol c.
Next, as usual, we define an extended relation * for sequences of zero or more steps:

* is a relation on IDs, with I * J if and only if there is a sequence of zero or

more relations that starts with I and ends with J.

Notice here that * is reflexive: for any ID I, I * I by a sequence of zero moves. Using the
* relation, we can define the language accepted by M:

The language accepted by a TM M is

L(M) = {x * | idfix(q0x) * ypz for some p F}

In this definition, the TM starts in its start state reading the first symbol of the input
string x (or reading B, if x is ε—the initial ID normalized by idfix will be either Bq0x if
x ε or Bq0B if x = ε). It accepts x if it has a sequence of zero or more moves that goes to an
accepting state, regardless of what is left on the tape, and regardless of the final position of
the head.

16.6 To Halt or Not to Halt

By the definition above, a TM accepts its input if it ever reaches an accepting state. Although
nothing in the definition of a TM specifically forbids transitions that leave accepting states,
there is no point in including such transitions, since they do not affect the language accepted.
The machine is considered to halt if it ever reaches an accepting state. It also halts on a given
input if it gets stuck—if it enters an ID I in a nonaccepting state, for which there is no ID J
with I J. There is also a third possibility: a TM may run forever on a given input. The TM
may always have a move, yet never reach an accepting state.
224 CHAPTER SIXTEEN: TURING MACHINES

Consider this TM, for example:

1/1,R

q r

B/B,L

0/0,R

Formally this is M = ({q, r, s}, {0, 1}, {0, 1, B}, , B, q, {s}), where the transition function is
given by
(q, 1) = (r, 1, R)
(q, 0) = (s, 0, R)
(r, B) = (q, B, L)
Given the input string 0, M accepts (and thus halts):
Bq0 0sB
In general, M will accept a string if and only if it begins with a 0, so L(M) is the regular
language L(0(0 + 1)*). But M also exhibits the other two possible outcomes. Given the input
string ε, M rejects by halting in a nonaccepting state—the start state, since there is no move
from there:
BqB ?
Given the input string 1, M does not accept or reject, but runs forever without reaching an
accepting state:
Bq1 1rB Bq1 1rB Bq1 ...
Of course, it is possible to make a TM for the language L(0(0 + 1)*) that halts on all inputs.
In general, though, we will see that it is not always possible for TMs to avoid infinite loops.
We might say that the possibility of running forever is the price TMs pay for their great
power.
In previous chapters we have seen occasional glimpses of nonterminating computations.
Stack machines and NFAs can in some sense run forever, since they can contain cycles of
ε-transitions. But those nonterminating computations are never necessary. We saw how to
construct an equivalent DFA for any NFA—and DFAs always terminate. Similarly, it is
possible to show that for any stack machine with a cycle of ε-transitions you can construct an
equivalent one without a cycle of ε-transitions—which therefore always terminates. Thus, the
possibility of having a computation run forever has been a minor nuisance, easily avoidable.
 16.7 A TM FOR {xcx | x ∈ {a, b}* } 225
With TMs, however, the situation is different. There are three possible outcomes when a
TM is run: it may accept, it may reject, or it may run forever. A TM M actually partitions *
into three subsets: those that make M accept (which we have named L(M )), those that make
M reject, and those that make M run forever. Instead of just defining L(M ), then, a TM
really defines three languages:

The language accepted by a TM M is

L(M) = {x * | idfix(q0x) * ypz for some p F }.

The language rejected by a TM M is

R(M) = {x * | x L(M) and there is some ID I with
idfix(q0x) * I and no J with I J}.

The language on which a TM M runs forever is

F(M) = {x * | x L(M ) and x R(M )}.

That third possibility is critical. There is a special name for TMs that halt on all inputs;
they are called total TMs.

A TM M is a total TM if and only if F(M) = {}.

In the next few chapters we will see some important results about the nonterminating
behavior of TMs. We will see that if we restrict our attention to total TMs we lose power. In
other words, we will see that these two sets of languages are not the same:

A recursively enumerable (RE) language is one that is L(M ) for some TM M.

A recursive language is one that is L(M ) for some total TM M.

These may seem like odd names—what does recursion have to do with TMs? But it turns
out that these two sets of languages were independently identified by a number of different
researchers working in different areas of mathematics. Although we are defining these sets in
terms of TMs, the standard names come from mathematical studies of computability using
the theory of recursive functions.

16.7 A TM for {xcx | x ∈ {a, b}* }

Consider the non-context-free language {xcx | x {a, b}*}. A TM for this language can work
by checking each symbol in the first x against the corresponding symbol in the second x.
To keep track of where it is in this process, our TM will mark the symbols that have been
checked. (To mark a symbol, the TM just overwrites it with a marked version of the same
symbol, for instance by overwriting a with a'.) Here is the plan:
226 CHAPTER SIXTEEN: TURING MACHINES

1. If the first symbol is a c, go to Step 6.

2. Mark the first symbol and remember what it was.
3. Move right up to the c, then right past any marked symbols to the first unmarked
a or b.
4. Mark it and check that it is the same symbol we saw in Step 2.
5. Move back left to the first unmarked symbol in the string, and go to Step 1.
6. Check that all input after the c has been marked.
Here is a TM that implements this strategy.
a/a,R a'/a',R
b/b,R b'/b',R

c/c,R
q s

a/a',L
a/a',R

a'/a',R c/c,L a'/a',L

p v u b'/b',L
b'/b',R

b/b',R a/a,L
b/b,L
b/b',L

r t
c/c,R

a/a,R a'/a',R
b/b,R b'/b',R

c/c,R
B/B,R
w x

a'/a',R
b'/b',R

This TM always halts, showing that {xcx | x {a, b}*} is a recursive language.
This example demonstrates two important higher-level techniques of TM construction.
The first is the idea of marking cells of the tape. As the machine illustrates, a cell can be
marked simply by overwriting the symbol there with a marked version of that symbol. In our
machine, the input alphabet is {a, b, c }, but the tape alphabet is {a, b, c, a', b', B}—it includes
marked versions of the symbols a and b.
The second is the technique of using the TM states to record some finite information. In
our example, we needed the machine to remember whether a or b was marked in the state p.
So we constructed the machine with two paths of states from p to u. The machine is in the
 16.8 THREE TAPES 227
state q or s if it saw a and is expecting to see a matching a in the second half of the string; it
is in the state r or t if it saw b and is expecting to see a matching b. In a high-level description
of the machine we can simply say that it “remembers” whether a or b was seen; we can always
implement any finite memory using the states of the TM.

16.8 Three Tapes

To handle more challenging problems, we will introduce a larger cousin of the basic TM.
This one works just like the basic model, except that it has three independent tapes, each
with its own independent read/write head. Recall that a one-tape TM is a 7-tuple like this:
M = (Q, , , , B, q0, F )
A three-tape TM may be defined with the same kind of 7-tuple, changing only the transition
function. For a three-tape TM, the transition function would take four inputs—the
current state and the symbol being read by the each of the three heads—and the value of
(q, X1, X2, X3) = (p, Y1, D1, Y2, D2, Y3, D3) would give the new state, plus a symbol to write
and a direction to move for each of the three heads.
The three-tape TM model is easier to program than our basic model, but no more
powerful. For any three-tape TM we can construct an equivalent basic TM that simulates it.
Imagine the three tapes stretched out side by side, each with its own head position:

... B a p p l e B ...

... B p e a r B B ...

... B l e m o n B ...

We can encode all this information on one tape, using an enlarged tape alphabet:

... (B,B,B) (a',p,l) (p,e,e ) (p,a,m) (l,r,o) (e,B',n) (B,B,B) ...

Each symbol in this new alphabet is a triple, encoding all the information about three
symbols from the original alphabet and using markings to indicate the presence or absence of
the head at that position. For example, the symbol (a', p, l ) indicates that, in the three-tape
machine, there is an a on the first tape, a p on the second tape, and an l on the third tape; the
228 CHAPTER SIXTEEN: TURING MACHINES

head of the first tape is at this location, while the heads of the other two tapes are elsewhere.
This alphabet of triples is, of course, much larger than the original alphabet. The original
alphabet is first doubled in size by the addition of a marked version of each symbol; for
example, = {a, B} becomes 2 = {a, a', B, B'}. This enlarged alphabet is then cubed in size
by being formed into 3-tuples; for example, 2 = {a, a', b, b' } becomes
3
=( 2 2 2
)
= {(a, a, a), (a, a, a'), (a, a, B), (a, a, B'), (a, a', a), (a, a', a' ), (a, a', B), (a, a', B'), ...}
and so on, for a total of 64 symbols.
Now let M3 be any three-tape TM. We can construct an ordinary TM M1 to simulate
M3. M1 will use the alphabet of triples, as shown above, and will use (B, B, B) as its blank
symbol. To simulate one move of the three-tape TM, M1 will use a two-pass strategy like this:
1. Make a left-to-right pass over the tape until all three marked symbols have been
found. Use our state to record M3’s state and the symbols being read at each of
M3’s three heads. This information determines the move M3 will make. If M3
halts, halt M1 as well—in an accepting state if M3 was in an accepting state or in a
nonaccepting state if not.
2. Make a right-to-left pass over the tape, carrying out M3’s actions at each of its
three head positions. (This means writing what M3 would write and also moving
the marks to record M3’s new head positions.) Leave the head at the leftmost cell
containing any marked symbol. Go back to step 1 for the next move.
By repeating these passes, M1 simulates M3. Of course, it makes far more moves than M3.
It uses a far larger alphabet than M3. It uses far more states than M3. But none of that matters
for questions of computability. M1 reaches exactly the same decisions as M3 for all input
strings: accepting where M3 accepts, rejecting where M3 rejects, and running forever where
M3 runs forever. Using this construction we can conclude the following:

Theorem 16.1: For any given partion of a * into three subsets L, R, and F,
there is a three-tape TM M3 with L(M3) = L, R(M3) = R, and F(M3) = F, if and
only if there is a one-tape TM M1 with L(M1) = L, R(M1) = R, and F(M1) = F.
Proof sketch: As indicated above, given any three-tape TM, we can construct a
one-tape TM with the same behaviors. In the other direction, any one-tape TM
can easily be simulated by a three-tape TM that simply ignores two of its three
tapes.

16.9 Simulating DFAs

We have given examples of TMs that define a variety of languages: regular, context free,
and noncontext free. We can now demonstrate what these examples suggest: that all regular
 16.9 SIMULATING DFAS 229
languages and all context-free languages are recursive.
Ordinarily, we would prove such claims by simple constructions. For example, we might
give a construction that takes any DFA and constructs an equivalent TM. Such constructions
are certainly possible, but we’ll tackle something more interesting: a single TM that can
simulate any DFA. This TM will take a pair of inputs: a DFA (encoded as a string) and the
string for that DFA to read. It will decide whether the given DFA accepts the given string.
We’ll specify it as a three-tape TM, knowing that (by Theorem 16.1) there is an equivalent
one-tape TM.
Our TM will use the alphabet {0, 1}, like any good computer. First, we need to represent
the alphabet of the DFA. We’ll number the symbols of the DFA’s alphabet arbitrarily as
= { 1, 2, ...}. The TM will use the string 1i to represent each symbol i. Using 0 as a
separator, we can represent any string over any alphabet this way. For example, suppose
the DFA’s alphabet is {a, b}. We’ll number a as 1 and b as 2. Then the string abba can be
represented in our TM’s alphabet as 11012012011 = 101101101.
Next, we need to represent the states of the DFA. We’ll number the states as
Q = {q1, q2, ...}, choosing the start state for q1 and numbering the other states arbitrarily.
Then we can use the string 1i to represent state qi.
Next, we need to encode the transition function of the DFA as a string. We will encode
each transition (qi, j) = qk as a string 1i01j01k and encode the entire function as a list of
such transitions, in any order, using 0 as a separator. For example, consider this DFA M:
b b

q1 q2

Numbering a as 1
and b as , we have this transition function for M:
2

(q1, 1) = q2
(q1, 2) = q1
(q2, 1) = q1
(q2, 2) = q2
that is encoded as this string:
101011 0 101101 0 110101 0 11011011
(The spaces shown are not part of the string, but emphasize the division of the string into a
list of four transitions separated by 0s.)
Finally, we need to represent the set of accepting states. We are already representing each
state qi as 1i, so we can represent the set of accepting states as a list of such strings, using 0
as the separator. Putting this all together, we can represent the entire DFA by concatenating
230 CHAPTER SIXTEEN: TURING MACHINES

the transition-function string with the accepting-states string, using 00 as a separator. For the
example DFA M, this is
101011 0 101101 0 110101 0 11011011 00 11
This gives us a way to encode a DFA as a string of 1s and 0s. It should come as no
surprise that this can be done—programmers are accustomed to the idea that everything
physical computers manipulate is somehow encoded using the binary alphabet. Our goal
now is to define a TM that takes one of these encoded DFAs along with an encoded input
string for that DFA and decides by a process of simulation whether the DFA accepts the
string.
We’ll express this as a three-tape TM (knowing that there is an equivalent, though more
complex, one-tape TM). Our TM will use the first tape to hold the DFA being simulated
and use the second tape to hold the DFA’s input string. Both tapes are encoded using the
alphabet {0, 1}, as explained in the previous section. The third tape will hold the DFA’s
current state, representing each state qi as 1i. This shows the initial configuration of the
simulator, using the example DFA M, defined above, and the input string abab.

... B 101011 0 101101 0 110101 0 11011011 00 11 B ...

... B 1 0 11 0 1 0 11 B ...

... B 1 B ...

As before, the spaces are added in the illustration to emphasize the encoding groups.
To simulate a move of the DFA, our TM will perform one state transition and erase one
encoded symbol of input. In our example above, the machine is in the state q1 (as shown by
the third tape) reading the input 1 = a (as shown by the second tape). The first transition
recorded on the first tape says what to do in that situation: go to the state 2. So after one
simulated move, the machine’s state will be this:

... B 101011 0 101101 0 110101 0 11011011 00 11 B ...

... B B B 11 0 1 0 11 B ...

... B 11 B ...

Of course, it takes many moves of the TM to accomplish this simulation of one move of
the DFA. That one-move simulation is repeated until all the input has been erased; then the
final state of the DFA is checked to see whether or not the DFA accepts the string. This is the
 16.10 SIMULATING OTHER AUTOMATA 231
strategy in more detail:
1. If the second tape is not empty, go to Step 2. Otherwise, the DFA’s input has all
been read and its computation is finished. Search the list of accepting states on the
first tape for a match with the DFA’s current state on the third tape. If a match is
found, halt in an accepting state; if it is not found, halt in a nonaccepting state.
2. Search the first tape for a delta-function move 1i01j01k, where 1i matches the
current state on the third tape and 1j matches the first input symbol on the second
tape. (Since a DFA has a move for every combination of state and symbol, this
search will always succeed.) The 1k part now gives the next state for the DFA.
3. Replace the 1i on the third tape with 1k, and erase the 1j and any subsequent
separator from the second tape, by writing Bs over them. This completes one move
of the DFA; go to Step 1.
The TM that simulates a DFA uses only a fixed, finite portion of each of its tapes. It
takes the encoded machine as input on its first tape and never uses more; it takes the encoded
input string on its second tape and never uses more; and it stores only one encoded state on
its third tape, which takes no more cells than there are states in the DFA. The existence of
this TM shows, indirectly, that all regular languages are recursive. (For any regular language,
we could construct a specialized version of the DFA-simulating TM, with a DFA for that
language wired in.)
One important detail we are skipping is the question of what to do with improper
inputs. What should the DFA simulator do with all the strings in {0, 1}* that don’t properly
encode DFAs and inputs? If our simulator needs to reject such strings, it will have to check
the encoding before beginning the simulation. For example, since the input is supposed to
be a DFA, it will have to check that there is exactly one transition from every state on every
symbol in the alphabet.

16.10 Simulating Other Automata

An interpreter is a program that takes another program as input and carries out the
instructions of that input program. This is a familiar concept from software development;
there are BASIC interpreters, Java bytecode interpreters, and Perl interpreters, written in
diverse languages. We have just seen how to construct a TM that acts as a DFA interpreter.
The same techniques can also be used to build interpreters for all our other kinds of
automata.
We could, for example, make a three-tape TM that interprets any stack machine. (Use
the first tape for an encoding of the stack machine to be simulated, the second for the input
string, and the third for the working stack.) The existence of this TM shows, indirectly,
that all context-free languages are recursive. Because we have already seen an example of a
language that is recursive but not context free, we conclude that the recursive languages are a
232 CHAPTER SIXTEEN: TURING MACHINES

proper superset of the CFLs. This picture captures our language classes so far:
L(a*b*)
regular
languages {anbncn}

CFLs
recursive
{anbn}
languages

Using the same techniques, a TM can even act as a TM-interpreter. Such a TM is called
a universal Turing machine. To construct a universal TM, we first design an encoding of TMs
as strings of 1s and 0s, whose principal content is an enumeration of the transition function.
Then we can construct a three-tape universal TM, taking an encoded TM as input on the
first tape and encoded input string on the second tape and using the third tape to record its
current state. The second tape, containing the input to the simulated TM, also serves as that
TM’s working tape, so we need to use marked symbols there to record the location of the
simulated TM’s head. Simulating a move is simply a matter of looking up the appropriate
transition on the first tape, making the necessary change on the second tape, and updating
the state on the third tape. If the simulated TM has no next move, we make our simulator
halt—in an accepting state if the simulated TM accepted or in a rejecting state if it rejected.
Of course, if the simulated TM runs forever, our simulating TM will too.
Like all the TMs described in the last few sections, our universal TM is merely sketched
here. In fact, we have not given a detailed construction of any TM since Section 16.7. We
did not give a complete, formal construction for converting three-tape TMs to standard
TMs; we just presented a sketch. We did not really construct any TMs to simulate other
automata; we just sketched how certain three-tape TMs could be constructed and then relied
on the previous construction to conclude that equivalent standard TMs exist. In effect, we
demonstrated that all these TMs could be constructed, without actually constructing them.
When you want to show that something can be done using a Turing machine, rough
outlines are often more effective than detailed constructions. That is because it is very
difficult to figure out what a TM does just by inspection—very difficult even for small TMs,
and inhumanly difficult for large ones. Merely showing a large TM and claiming that it
recognizes a certain language is not a good way to convince anyone that the language can be
recognized by a TM. Instead, it is more convincing to give less detail—to describe in outline
how a TM can be constructed. Once you’re convinced it can be done there is no point in
actually doing it!
No point, that is, except for fun. People have actually constructed many universal TMs
 EXERCISES 233
in detail. It serves virtually no purpose, but it is an interesting puzzle, especially if you try to
do it with the smallest possible number of states and the smallest possible alphabet. There is a
fine introduction to this puzzle in Marvin Minsky’s classic book, noted below.

16.11 Further Reading

Marvin Minsky illustrates a small (but not record-holding) universal Turing machine using
seven states and an alphabet of four symbols in this out-of-print classic:
Minsky, Marvin. Computation: Finite and Infinite Machines. Upper Saddle River,
NJ: Prentice Hall, Inc., 1967.
Although slightly dated, it is an elegant and very rewarding book.

Exercises
These exercises refer to ordinary, one-tape TMs. Show completed TMs in detail. (Yes, we
just asserted that high-level descriptions are more readable and that there is no point in
constructing large TMs in detail, except for fun. Conclusion: These exercises must be fun.)
EXERCISE 1
Give a TM for the language L(a*b*).
EXERCISE 2
Give a TM for the language accepted by this DFA:
0 1

1 0

1 0

EXERCISE 3
Give a TM for {x {a, b}* | x contains at least three consecutive as}.
EXERCISE 4
Give a TM for the language generated by this grammar:
S aSa | bSb | c
EXERCISE 5
The TM in Section 16.3 for {anbncn} has two states than can be merged. Rewrite it with
these states merged, so that it has one less state in total.
EXERCISE 6
Give a TM for {anbnc2n}.
234 CHAPTER SIXTEEN: TURING MACHINES

EXERCISE 7
Give a TM that accepts the same language as the TM in Section 16.6 but always halts.
EXERCISE 8
Construct a TM that accepts the language {anbncn} and runs forever on all strings that
are not accepted. Thus, for your TM M you will have L(M) = {anbncn}, R(M ) = {}, and
F(M) = {x {a, b, c}* | x {anbncn}}. Hint: Modify the machine from Section 16.3.
EXERCISE 9
Construct a TM that accepts the language {x {a, b, c}* | x {anbncn }}. Hint: Modify
the machine from Section 16.3.
EXERCISE 10
The TM of Section 16.7 was chosen to illustrate the concept of marking cells of the tape.
A simpler TM for the same language can be implemented if the symbols in the first half
of the string are not marked but simply erased (by being overwritten with a B). Using
this idea, reimplement the TM. (You should be able to eliminate a state.)
EXERCISE 11
Construct a TM for the language {0i1j0k | i > 0, j > 0, and i + j = k}.
EXERCISE 12
Construct a TM for the language {anbnaib j }.
EXERCISE 13
Construct a TM for the language L(a*b* + b*a*).
EXERCISE 14
Construct a TM that runs forever on inputs in L(a*b* + b*a*), but halts on all other
inputs in {a, b}*. (It doesn’t matter whether or not it accepts the other inputs.)
EXERCISE 15
Consider the language of encoded DFAs from Section 16.9. By removing the
requirement that there be exactly one transition from every state on every symbol in the
alphabet, we have a similar language of encoded NFAs. Give a regular expression for this
language.
EXERCISE 16
Construct a TM that accepts the language of strings that encode NFAs that have at least
one accepting state; that is, your TM must check its input for membership in the regular
language of the previous exercise and must check that the NFA encoded has at least one
accepting state.
 EXERCISES 235
EXERCISE 17
Construct a TM that takes an input of the form 1i0x, where x is an encoding of a DFA
as in Section 16.9. Your TM should halt leaving only the string 1j on the tape, where
(q1, i) = qj. Other than this string 1j, the final tape should contain nothing but Bs. The
behavior of your TM on improperly formatted inputs does not matter, and it does not
matter whether the TM accepts or rejects.
EXERCISE 18
Construct a TM that takes an input of the form 1i0x, where x is an encoding of a DFA as
in Section 16.9. Your TM should accept if and only if the DFA accepts the one-symbol
string i. The behavior of your TM on improperly formatted inputs does not matter.
 17
CHA P TER

Computability

Computability is the mysterious border of our realm. Within the

border is our neatly ordered kingdom, occupied by solvable problems.
These problems range from the trivially simple to the hideously
complicated, but all are algorithmically solvable given unbounded
time and resources. The shape of this border was first clearly perceived
in the 1930s.

237
238 CHAPTER SEVENTEEN: COMPUTABILITY

17.1 Turing-Computable Functions

Consider any total TM—any TM that always halts. But instead of focusing on the state in
which it halts (accepting or not), think of the final contents of the tape as the TM’s output.
Viewed this way, any total TM implements a function with string input and output. This
doesn’t change any of the details of TM operation that we saw in the previous chapter; it only
changes how we look at the result of the TM’s execution.

A function f: * * is Turing computable if and only if there is some total

TM M such that for all x *, if M starts with x on the tape, M halts with f (x)
on the tape.

In this definition, the final state of the TM is not considered, nor does it matter where the
TM’s head is positioned when it halts.
Let’s look at an example. In Java and the other C-family languages we write the
expression x&y to express the bitwise AND of two integer operands x and y. We’ll show that
this is Turing computable, and to make this more interesting, we’ll extend it to allow the
operands to be arbitrarily long strings of bits. Imagine first that the TM’s input is presented
with the operands already stacked up, using Σ = { 00 , 01 , 10 , 11 } as the input alphabet. To show
that this and function is Turing computable, we need to find a TM that, given any x *
as input, halts with and(x) on its tape. For example, given 00 11 10 01 11 10 as input, our TM should
halt with 010010 on its tape, since and ( 00 11 10 01 11 10 )=010010 .
But nothing could be simpler:
0
0 /0,R
1
0 /0,R
0
1 /0,R
1
1 /1,R

This and machine simply proceeds from left to right over its input, overwriting each stacked-
up pair of input bits with the corresponding output bit. When the end of the input is
reached, the machine halts, having no transition on a B. This machine implements our and
function, demonstrating that the function is Turing computable. Notice that the and TM has
no accepting states, so L(and ) = {}. But we’re not interested in the language it defines; we’re
interested in the function it implements.

17.2 TM Composition
You might argue that we cheated in that last example by assuming that the input was
presented in stacked form, using an input alphabet completely different from the output
alphabet. What if the input is presented as two, linear, binary operands? What if our
input is not 00 11 10 01 11 10 but 011011#010110, where some special symbol # separates the two
 17.2 TM COMPOSITION 239
operands? It’s simple: we can preprocess the input into the stacked form we need. We can
define a function stacker that converts linear inputs into stacked form, so that for example
stacker(011011#010110)= 00 11 10 01 11 10 . This function is Turing computable; here’s a machine to
compute it:
0 0
0 / 0 ,L
1 1
0 / 0 ,L
0 0
0/0,L 1 / 1 ,L
1 1
1/1,L 1 / 1 ,L

0 0
0/ 0 ,R
1 1
0/0,R #/#,L 0/ 0 ,R
0 0
1/1,R B/ 00 ,R 1/ 1 ,R
1 1
#/#,R / 1 ,R
0 / 00 ,R 1
0/0,R
0/B,L 1 / 10 ,R 1/1,R
#/#,R
B/B,L B/B,L
b

1/B,L B/ 01 ,R
0 / 01 ,R
#/B,L 1 / 11 ,R

#/#,L

0/0,L 0 0
/ 0 ,L
0
B/B,R 1/1,L 1 1
y z 0 / 0 ,L
0 0
1 / 1 ,L
1 1
0 0 1 / 1 ,L
0/ 0 ,L
1 1
0/ 0 ,L
0 0
1/ 1 ,L
1 1
1/ 1 ,L
0/ 00 ,L
1/ 10 ,L

The stacker machine uses the following strategy. In the start state it skips to the rightmost
symbol of the string. Then for each symbol in the second operand it makes a pass through
the main loop of the machine, starting and ending in the state b. On each such pass it erases
the rightmost remaining symbol of the second operand and overwrites the rightmost 0 or 1
of the first operand with the corresponding stacked symbol. (It treats B as equivalent to 0 in
the first operand, in case it is shorter than the second operand.)
When the second operand has been erased, the separating # is also erased and the
machine enters the state y. There it makes a final pass over the first operand. If the first
operand was longer than the second, it will still contain some unstacked 0s and/or 1s; in the
240 CHAPTER SEVENTEEN: COMPUTABILITY

state y they are replaced with stacked versions, as if the second operand had been padded
with leading 0s.
In its final state z, the TM halts, leaving the stacked version of its input on the tape. (If
the input is improperly formatted—if it does not include exactly one separator #—then all
bets are off, and the output of the stacker TM will not be a neatly stacked binary string. But
we don’t care about the output for such cases.) In the state z the stacker TM’s head is at the
leftmost symbol of the stacked string, which is exactly where our and TM wants it. So we
can simply compose the two machines, replacing the final state of the stacker TM with the
start state of the and TM. This gives us a TM that implements the composition of the two
functions: linearAnd(y) = and(stacker(y)).
Using TM composition is a bit like using subroutines in a high-level language. It is an
important technique for building computers of more elaborate functions.

17.3 TM Arithmetic
We turn now to the problem of adding binary numbers. As before, we’ll assume that
the operands are stacked binary numbers of arbitrary length. The add function will
take an input of that form and produce the simple binary sum as output. For example,
add ( 00 11 10 01 11 10 )=110001 (or, in decimal, 27 + 22 = 49).
The add function is Turing computable:
0 0
0 / 0 ,R
1 1 0
0 / 0 ,R 0 /0,L
0 0 1
1 / 1 ,R 0 /1,L
1 1 0
1 / 1 ,R 1 /1,L

B/B,L B/B,R
c0

0
0/1,L 1
1 /0,L
B/1,L

1
0 /0,L
c1 0
1 /0,L
1
1 /1,L

The add machine uses the following strategy. In the start state it skips to the rightmost
symbol of the string. Then for each pair of stacked symbols, working from right to left, it
overwrites that pair with the corresponding sum bit. The state of the TM records the carry-in
 17.4 TM RANDOM ACCESS 241
from the previous bit; it is in the state c0 if the carry-in was 0 and the state c1 if the carry-in
was 1. When all the bits have been added the TM encounters the B to the left of the input. If
the final carry-out was a 1, it writes a leftmost 1 over the B; then it halts.
As before, we can combine this machine with the stacker machine to make an adder that
takes its operands linearly: linearAdd(x) = add(stacker(x)).
To decrement a binary number (that is, to subtract one from it) is a simple operation
closely related to addition. We can decrement a binary number simply by adding a
string of 1s to it and ignoring the carry-out from the final bit position. For example,
decrement(100) = 011; this is the same as linearAdd(100#111) = 1011, except for the final
carry-out. If you start with the add machine above, but assume that the bottom bit in each
stacked pair is a 1 and ignore the final carry-out, you get this decrement machine:
0/0,R
1/1,R 0/1,L

B/B,L B/B,R
c0 z

1/0,L

B/B,R
c1 nz

0/0,L
1/1,L

The extra final state in the decrement machine may come in handy, since it indicates whether
the number that has just been decremented was already zero. If it was zero, there was no
carry-out from the leftmost bit, so the final state is z. If it was not zero, the final state is nz.
So our decrement machine can be used not only as a decrement-function computer, but also as
a decrement-and-test component in larger machines.

17.4 TM Random Access

A TM’s memory is like a tape. To get to a particular cell you have to move the TM’s head
there, one step at a time. By contrast, modern physical computers use addressable memories.
To refer to a particular memory word you just give its integer address. This seems like a
significant difference; can TMs implement this kind of addressing?
Let i and s be any two binary strings, and define ith(i#s) = the symbol at the position i
in s, treating i as a binary number and indexing s from the left starting at 0. (If i |s| then
define ith(i#s) = ε.) For example,
242 CHAPTER SEVENTEEN: COMPUTABILITY

ith(0#100) = 1 (bit 0, counting from the left, in 100)

ith(1#100) = 0 (bit 1, counting from the left, in 100)
ith(10#100) = 0 (bit 2, counting from the left, in 100)
ith(11#100) = ε (bit 3—no such bit)
The ith function is Turing computable:
0/B,R
0/0,L 1/B,R
0/0,R 1/1,L #/B,R 0/B,R
1/1,R B/B,L 0/1,L B/B,R 1/B,R
0/0',L
1/1',L
#/#,R B/B',L #/#,L B/B,R
c0 z

0'/0,L
0'/0,R
1'/1,L
1'/1,R 1/0,L B'/B,L
B'/B,R
B/B,R
nz c1
0/0,R
1/1,R 0/0,L
#/#,R 1/1,L
B/B,R

Notice that our decrement machine is embedded here, in the states labeled c0, c1, nz, and z.
The ith machine starts by marking the first symbol of the string s. Then, in its main loop, it
repeatedly decrements i and moves the mark one place to the right. When the decrementing
step finds that i was zero (entering the state z), the currently marked symbol is the one that
should be output; the ith machine then erases everything except the marked symbol and
halts.

17.5 Functions and Languages

Our main interest in this book is to define languages formally, but in this chapter we took
a detour to examine a related problem: how to use TMs to implement general functions on
strings. The TM model we used was exactly the same, and the only difference was in how we
looked at the results of a TM’s execution.
In many ways, function implementation and language definition are just two different
perspectives on the same computational problem. For every language L we can define a
corresponding function, such as
f(x) = 1 if x L, 0 if x L
 17.6 THE CHURCH-TURING THESIS 243
With such a correspondence, L is recursive if and only if f is Turing computable. The
constructions that prove this are simple. For example, given a total TM with L(M ) = L, we
could build a new TM M' that imitates M up to the point where M halts. Then, if M halts in
an accepting state, M' erases its tape, writes a 1, and halts; while if M halts in a nonaccepting
state, M' erases its tape, writes a 0, and halts. This new M' computes f.
Similarly, for every function f we can define a corresponding language, such as
{x#y | y = f(x)}
Again, it is easy to show that the language is recursive if and only if the function is Turing
computable.
The examples of this chapter illustrated that TMs can be used to
• implement Boolean logic;
• get the effect of subroutine calls by composition;
• perform binary arithmetic; and
• index memory using binary addresses.
These operations are among the most important building blocks of modern computer
systems. The fact that these primitives can be implemented using TMs is evidence
for the extraordinary power of TMs, whether for language definition or for function
implementation. In spite of their apparent simplicity, it turns out that TMs can do anything
that can be done with a high-level programming language on a modern computer.

17.6 The Church-Turing Thesis

Many familiar parts of mathematics are concerned with computational procedures: given
two numbers in N, there is a procedure for finding their greatest common denominator;
given a decimal number, there is a procedure for finding its inverse; given a second-degree
polynomial, there is a procedure for finding its roots; given a compass and a straight edge,
there is a procedure for constructing a regular pentagon. But in the early 1900s, the attention
of mathematicians was drawn to problems that seemed to resist all efforts to find effective
computational procedures: given an assertion in first-order logic, where’s the procedure for
deciding whether it is true or false? Given a polynomial equation, where’s the procedure for
finding whether it has integer solutions? (See the next chapter for the answers.)
Wrestling with such questions led to the more fundamental question: what exactly
is an effective computational procedure? During the 1920s and 1930s, a number of
mathematicians around the world were struggling with this problem. A number of very
different formalisms were developed. Emil Post developed Post systems; Kurt Gödel worked
on the theory of -recursive functions; Alonzo Church and Stephen Kleene developed the
-calculus; Moses Schönfinkel and Haskell Curry developed combinator logic; and, in 1936,
Alan Turing invented Turing machines. These different formalisms operate on different kinds
of data. For instance, Turing machines work on strings, the -calculus works on -terms, and
-recursive functions work on numbers in N.
244 CHAPTER SEVENTEEN: COMPUTABILITY

Although these mathematicians all started out in different directions, they all arrived at
the same place. With suitable conversions between the different kinds of data—for instance,
representing the natural numbers of -recursive functions as strings of digits for Turing
machines—it turns out that all those formalisms for computation are interconvertible.
Any Turing machine can be simulated by a Post system and vice versa, any Post system
can be simulated by a -term and vice versa, and so on. People say that any formalism for
computation that is interconvertible with Turing machines is Turing equivalent. (It would
also be correct to say that they are “Post-system equivalent,” “ -calculus equivalent,” and
so on, but Turing machines are the most popular of the formalisms, perhaps because they
are the most like physical computers.) All Turing-equivalent formalisms have the same
computational power as Turing machines; they also have the same lurking possibility of
infinite computation.
In 1936, Alonzo Church and Alan Turing both suggested that their equivalent
formalisms had in fact captured the elusive idea of “effective computational procedure.”
In effect, they were suggesting that computability by a total TM be taken as the definition
of computability. (Note that the definition of Turing computable requires a total Turing
machine. A computational procedure that sometimes goes on forever is naturally not
considered to be effective!) This has come to be known as Church’s Thesis, or the Church-
Turing Thesis. It is called a thesis, not a theorem, because it is not the kind of thing that is
subject to proof or disproof. But it is a definition that has come to be generally accepted,
and today it marks the border of our happy realm. One thing that makes Church’s Thesis so
compelling is the effect we have mentioned so often before: it just keeps turning up. Many
researchers walking down different paths of mathematical inquiry found the same idea of
effective computation waiting for them at the end. That was true in the 1930s, and it is even
more true today—because today we have the additional evidence of all modern programming
languages and all the physical computer systems that run them. They too are interconvertible
with Turing machines, as we will see in the next section. Java, ML, Prolog, and all the
rest just add more support for Church’s Thesis; they are Turing-equivalent formalisms for
computation. “Computable by a Java program that always halts” is the same as “Turing
computable,” and according to Church’s Thesis, that’s what computable means, period.
There are several different words in common use for this same concept of computability.
A function that is Turing computable (like addition on binary numbers) is called a
computable function. A language that is recognized by some total Turing machine (like
{anbncn}) is called a recursive language. When there is some total Turing machine that decides
whether an input has a given property (like being a prime number or the square of an
integer), we say the property is decidable. But beware: some authors prefer a different usage,
and some just invent new terms!
An important related term is algorithm. Its meaning is quite close to that of effective
computational procedure, as defined by Church’s Thesis. The usage of algorithm tends to be
 17.7 TM AND JAVA ARE INTERCONVERTIBLE 245
broader, however. There are fault-tolerant, distributed algorithms; probabilistic algorithms;
and interactive algorithms. These kinds of algorithms don’t compute functions, so they’re
not a good fit for Church’s Thesis. In casual use the word algorithm sometimes even refers to
things like recipes, which are essentially noncomputational.

17.7 TM and Java Are Interconvertible

For any Java program there is an equivalent Turing machine, and for any Turing machine
there is an equivalent Java program. This can be proved in both directions by construction,
but the constructions are not at all symmetric. High-level languages are much more
conveniently expressive than TMs, and that makes the construction invitingly simple in one
direction and forbiddingly complicated in the other direction. We’ll start by illustrating the
inviting direction: a Java implementation of a TM.
We’ll make a Java implementation of the TM from Section 16.3 for the language
{anbncn}. First, we need to encode the machine’s definition:

/**
* A Turing machine for the language a^n b^n c^n.
*/
public class TManbncn {
/*
* A constant for the tape blank.
*/
private static final char B = 0;

/*
* The transition function. A char is used to
* identify each state, and one of the chars
* 'R' or 'L' is used to identify the head directions.
* Each transition delta(q,a)=(p,b,D), where q is the
* current state, a is the symbol being read, p is
* the next state, b is the symbol to write, and D is
* the direction to move the head, is represented by
* a char array of length five: {q,a,p,b,D}.
* Individual move arrays occur in the delta array in
* no particular order, so looking up a move requires
* a linear search.
*/
private static final char[][] delta = {
{1,'a',2,'X','R'}, // delta(q1,a) = (q2,X,R)
246 CHAPTER SEVENTEEN: COMPUTABILITY

{1,B,7,B,'R'}, // etc.
{2,'a',2,'a','R'},
{2,'Y',2,'Y','R'},
{2,'b',3,'Y','R'},
{3,'b',3,'b','R'},
{3,'Z',3,'Z','R'},
{3,'c',4,'Z','L'},
{4,'a',4,'a','L'},
{4,'b',4,'b','L'},
{4,'Z',4,'Z','L'},
{4,'Y',4,'Y','L'},
{4,'X',5,'X','R'},
{5,'a',2,'X','R'},
{5,'Y',6,'Y','R'},
{6,'Y',6,'Y','R'},
{6,'Z',6,'Z','R'},
{6,B,7,B,'R'}
};

/*
* A String containing the char for each accepting
* state, in any order.
*/
private static final String accepting = "\7";

/*
* The initial state.
*/
private static final char initial = 1;

Formally, a TM is M = (Q, , , , B, q0, F). In this simulation, we encode only the last four
parts. The tape alphabet is taken to be Java’s char type, and the input alphabet is taken to
be char – {B}. The state set is not explicity represented, but the states are represented by
char values. Our machine uses the states 1 through 7.
The class definition continues:
/*
* The TM's current tape and head position. We always
* maintain 0 <= head < tape.length(), adding blanks
* to the front or rear of the tape as necessary.
 17.7 TM AND JAVA ARE INTERCONVERTIBLE 247
*/
private String tape;
private int head;

/*
* The current state.
*/
private char state;

/**
* Decide whether the TM accepts the given string. We
* run the TM until it either enters an accepting state
* or gets stuck. If the TM runs forever on the given
* string, this will of course never return.
* @param s the String to test
* @return true if it accepts, false if not
*/
public boolean accepts(String s) {
state = initial;
head = 0;
tape = s;

// establish 0 <= head < tape.length()

if (head==tape.length()) tape += B;

while (true) {
if (accepting.indexOf(state)!=-1) return true;
char[] move = lookupMove(state,tape.charAt(head));
if (move==null) return false;
executeMove(move[2],move[3],move[4]);
}
}
The accepts method simulates the TM on a given input string. Its while loop simulates
one move of the TM for each iteration. The loop has two exits. If the machine is in an
accepting state, the simulation halts and accepts; if there is no next move from the current
configuration, the simulation halts and rejects. (To test for an accepting state, it uses the
indexOf method of the String class: s.indexOf(c) will return -1 if and only if the
char c does not occur in the String s.)
248 CHAPTER SEVENTEEN: COMPUTABILITY

To complete the class definition, we need the methods executeMove and

lookupMove:

/**
* Execute one move of the TM.
* @param newstate the next state
* @param symbol the char to write
* @param dir the direction ('L' or 'R') to move
*/
void executeMove(char newstate, char symbol, char dir) {

// write on the tape

tape = tape.substring(0,head) + symbol +
tape.substring(head+1);

// move the head, maintaining the invariant

// 0 <= head < tape.length()

if (dir=='L') {
if (head==0) tape = B + tape; else head -= 1;
}
else {
head += 1;
if (head==tape.length()) tape += B;
}

// go to the next state

state = newstate;

/**
* Find the move for the given state and symbol.
* @param state current state
* @param symbol symbol at current head position
* @return the five-element char[] from the delta table
*/
char[] lookupMove(char state, char symbol) {
for(int i = 0; i<delta.length; i++) {
 17.7 TM AND JAVA ARE INTERCONVERTIBLE 249
char[] move = delta[i];
if (move[0]==state && move[1]==symbol) return move;
}
return null;
}

That concludes the implementation of the TManbncn simulator. To test the String s for
membership in {anbncn}, we just create a TManbncn object and use its accepts method,
like this:

TManbncn m = new TManbncn();

if (m.accepts(s)) ...

Of course, there are much more efficient ways for a Java program to test a string for
membership in {anbncn}. Previous chapters examined implementations of DFAs, NFAs, and
stack machines, and those techniques had important practical applications. Unlike them, our
TM implementation has no practical value. It is just an exercise that helps you understand
how TMs work, and it illustrates one direction of the Turing-equivalence construction.
In the previous chapter we showed how to construct TMs that interpret other automata
encoded as input strings. Ultimately, we saw that it is possible to construct universal TMs
that interpret other TMs encoded as input strings. Our Java implementation is not quite
like that, since the TM being interpreted is hardwired into the program, in the form of those
definitions of the variables B, delta, accepting, and initial. But it would be easy
enough to modify the code to read the values of those variables from a file. Then it would be
an implementation of a universal Turing machine.
The other direction—constructing a TM equivalent to a given Java program—is also
possible, but requires a great deal more work. A TM for Java could have two parts: a Java
compiler (converting the Java source to a machine language) and an interpreter for that
machine language. The Java compiler could be written in machine language, so we only really
need the second part, the machine-language interpreter. We have already seen that TMs can
perform the basic operations of such an interpreter: Boolean logic, binary arithmetic, indexed
addressing, and so on. Following this outline, it would be possible (though inhumanly
complicated) to build a TM that serves as a Java interpreter, just as it was possible (and rather
simple) to build a Java program that serves as a TM interpreter.
Of course, there’s nothing in all this that is unique to Java. A similar argument
about interconvertibility can be made for any high-level programming language. Our
demonstration of this has been informal; rigorously proving that TMs are equivalent in
power to a particular high-level language is a long and tedious chore. But such proofs have
250 CHAPTER SEVENTEEN: COMPUTABILITY

been done and are universally accepted. There is nothing that high-level languages can do
that TMs cannot do. All modern programming languages turn out to be Turing equivalent,
adding still more support to the Church-Turing Thesis.

17.8 Infinite Models, Finite Machines

Actually, a Turing machine can do a better job of implementing high-level languages than
an ordinary computer. A physical computer is always limited by its finite memory. You may
install as much memory as the machine can hold; you may buy a big array of disks and
implement an even larger virtual memory; but no matter what you do, there will still be
programs that demand more memory (and more time) than you have. Any implementation
of Java on a physical computer introduces memory limitations. In the abstract, high-level
languages like Java are Turing equivalent; the physical computers that implement them are
not.
Any physical computer has finite memory, and therefore has finitely many possible
states. In that sense, physical computers are more like DFAs than Turing machines—DFAs
with enormous, but still finite, state sets. Nothing we can build has unbounded memory,
nor do we have unbounded time for computation. Yet almost everyone would agree that, at
an intuitive level, TMs are much better than DFAs as models for physical computers. It’s a
curious situation. Unbounded models are patently unrealistic, so why do they feel so natural?
Part of the explanation is that the number of states that can be taken on by a physical
computer is so large that it usually appears unbounded to the human user. For example, a
word processor running in finite memory can represent only finitely many different texts.
Every possible text, profound or nonsensical, in every state of completion, corresponds to
one of the states that can be taken on by the machine on which it is being written. But
this limitation is of theoretical interest only. In practice, it does not chafe—a writer does
not think of writing as the act of selecting one of these finitely many representable texts.
Similarly, programmers do not think of programs as DFAs that drive the machine through
finitely many possible states.
In any event, all of mathematics is filled with such idealizations. Our ordinary
experience of the world includes no perfect points, lines, or circles, but that does not make
geometry less useful. It includes no infinite numbers and no infinitesimals, but that does
not make calculus less useful. Though it includes no infinite computations, the concept of
computability captured by the Turing machine is no less important. In particular, the notion
of uncomputability captured by the Turing machine is a strong one—for if a computation
is impossible for an idealized, unbounded machine, it is certainly impossible for a physical
computer as well. We will consider the question of uncomputability more deeply in the next
chapter.
 EXERCISES 251

Exercises
EXERCISE 1
The linearAnd function can be implemented with fewer states and a smaller tape alphabet,
if implemented directly and not as a composition of and with stacker. Show how.
EXERCISE 2
Let increment be the binary increment function, increment(x) = add(stacker(x#1)).
Give a TM that implements increment. Note that the definition requires it to treat
ε as a representation for 0, so that increment (ε) = 1. Hint: Actually using stacker in
your implementation would be doing it the hard way. The increment function can be
implemented with just three states, using just {0, 1, B} as the tape alphabet.
EXERCISE 3
Let roundup be the binary, round-up-to-even function, roundup(x) = x if x represents
an even number, or increment (x) if x represents an odd number. Give a TM that
implements roundup. Treat ε as a representation for zero, so that roundup(ε) = ε.
EXERCISE 4
Let rounddown be the binary, round-down-to-even function, rounddown(x) = x if
x represents an even number, or decrement (x) if x represents an odd number. Give
a TM that implements rounddown. Treat ε as a representation for zero, so that
rounddown(ε) = ε.
EXERCISE 5
For any x {a, b}*, let mid(x) be the middle symbol of x or ε if x has even length. Give a
TM that implements mid.
EXERCISE 6
Give a TM that implements subtraction on unsigned, binary numbers, assuming stacked
inputs as for the add TM of Section 17.3. Assume that the second operand is less than or
equal to the first, so that the result can be represented as an unsigned, binary number.
EXERCISE 7
Give a TM that implements multiplication by two on unary numbers, using
= {1}. Given an input string 1i, your TM should halt leaving 12i (and nothing else) on
the tape.
EXERCISE 8
Give a TM that implements division by two on unary numbers, using = {1}. Given
an input string 1i, your TM should halt leaving 1i/2 (and nothing else) on the tape. If i is
odd, round up.
252 CHAPTER SEVENTEEN: COMPUTABILITY

EXERCISE 9
For any x {0, 1}*, let val(x) be the natural number for which x is a binary representa-
tion; for completeness, let val(ε) = 0. Give a TM that implements a binary-to-unary
conversion function, btu(x) = 1val(x).
EXERCISE 10
Let utb be a unary-to-binary conversion function (an inverse of btu from the previous
exercise). Give a TM that implements utb.
EXERCISE 11
Write a Java method anbncn, so that anbncn(s) returns true if String s is
in {anbncn} and returns false if not. Make it as efficient as possible (which of course
means not implementing it as a TM simulation). Then answer in as much detail as you
can, how does the speed of your anbncn compare with that of the accepts method
in TManbncn?
EXERCISE 12
Write a Java implementation of the TM implementing the linearAdd function, as
described in Sections 17.2 and 17.3. Include a main method that reads the input
string from the command line and writes the output string (that is, the final contents of
the tape, not including leading and trailing Bs) to standard output. For example, your
program should have this behavior:
> java TMlinearAdd 1010#1100
10110
EXERCISE 13
Write a Java implementation of a universal Turing machine. Start with the code for
TManbncn, but alter it to make it read the TM to be simulated from a file. You can
design your own encoding for storing the TM in a file, but it should be in a text format
that can be read and written using an ordinary text editor. Your class should have a main
method that reads a file name and a text string from the command line, runs the TM
encoded in the given file on the given string, and reports the result. For example, if the
file anbncn.txt contains the encoding of a TM for the language {anbncn}, then your
program would have this behavior:
> java UTM anbncn.txt abca
reject
> java UTM anbncn.txt aabbcc
accept
 EXERCISES 253
EXERCISE 14
Complete the following short story:
The judge sighed as he sat down before the computer screen. He was administering
his eighth Turing test of the day, and he was tired of it. Six had been machines, their
classification easily revealed by their responses to simple conversational gambits like “So
... you married?” or “How 'bout them Yankees?” There had been only
one human this morning, and the judge had been sure of the classification in less than
a minute. He had always had a knack for this kind of testing, and the more he did, the
quicker he got.
But this morning it was getting tedious. One more test, he thought, and I can break for
lunch. He rubbed his eyes, sat up straighter in his chair, and quickly typed:
Would you say that your mind is finite?
Slowly, the response came back ...
 18
CHA P TER

Uncomputability

The Church-Turing Thesis gives a definition of computability, like a

border surrounding the algorithmically solvable problems.

L(a*b* )
regular
languages {anbncn}

CFLs
recursive
{anbn}
languages

Beyond that border is a wilderness of uncomputable problems. This

is one of the great revelations of 20th-century mathematics: the
discovery of simple problems whose algorithmic solution would be
very useful but is forever beyond us.

255
256 CHAPTER EIGHTEEN: UNCOMPUTABILITY

18.1 Decision and Recognition Methods

In this chapter, to make examples that are easier for practical programmers to read, we will
switch from using Turing machines to using a more familiar Java-like syntax. Although this
syntax looks like Java, there is a critical difference: we will assume that there are none of
the limitations of actual Java. In particular, we will assume that there are no bounds on the
length of a string or the size of an integer. So be aware that from here on, our code examples
are intended to be illustrative rather than strictly executable.
Total TMs correspond to decision methods in our Java-like notation.

A decision method takes a String parameter and returns a boolean value.

(It always returns, and it does not run forever.)

Here’s one for the language {ax | x *}:

boolean ax(String p) {
return (p.length()>0 && p.charAt(0)=='a');
}

(For those not familiar with Java, the expression p.length() is the length of the String
object referred to by p, and the expression p.charAt(i) yields the ith character of the
String p, counting from zero.) Here are decision methods for the languages {} and *:

boolean emptySet(String p) {
return false;
}

boolean sigmaStar(String p) {
return true;
}

As with Turing machines, we will refer to the language accepted by a decision method m
as L(m). So L(emptySet) = {} and L(sigmaStar) = *. Decision methods give us an
alternative, equivalent definition of a recursive language: a language is recursive if and only if
it is L(m) for some decision method m.
For methods that might run forever, we will use the broader term recognition method.
These correspond to TMs that are not necessarily total.

A recognition method takes a String parameter and either returns a

boolean value or runs forever.
 18.1 DECISION AND RECOGNITION METHODS 257
With these definitions, a decision method is a special kind of recognition method, just as a
total TM is a special kind of TM. Here is a recognition method for the language {anbncn}:

boolean anbncn1(String p) {
String as = "", bs = "", cs = "";
while (true) {
String s = as+bs+cs;
if (p.equals(s)) return true;
as += 'a'; bs += 'b'; cs += 'c';
}
}

This is a highly inefficient way of testing whether the string p is in {anbncn}, but we are not
concerned with efficiency. We are at times concerned with avoiding nontermination; the
recognition method anbncn1 loops forever if the string is not in the language. Thus it only
demonstrates that {anbncn} is RE. We know that {anbncn} is a recursive language, so there is
some decision method for it, such as this one:

boolean anbncn2(String p) {
String as = "", bs = "", cs = "";
while (true) {
String s = as+bs+cs;
if (s.length()>p.length()) return false;
else if (p.equals(s)) return true;
as += 'a'; bs += 'b'; cs += 'c';
}
}

Recognition methods give us an alternative, equivalent definition of a recursively enumerable

language: a language is RE if and only if it is L(m) for some recognition method m.
We will need the high-level-language equivalent of a universal TM. Real Java programs,
whether applets or applications, consist of one or more complete class defininitions, which
are compiled in a separate step before being run. We’ll assume something simpler: that
individual methods are run on individual input strings, interpreting directly from the source.
We’ll assume the existence of a run method with this specification:

/**
* run(rSource, in) runs a recognition method, given
* its source code. If rSource contains the source
* code for a recognition method r, and in is any
258 CHAPTER EIGHTEEN: UNCOMPUTABILITY

* string, then evaluating run(rSource, in) has the

* same effect as evaluating r(in).
*/
boolean run(String rSource, String in) {
...
}

For example, in this fragment:

String s = "boolean sigmaStar(String p) {return true;}";

run(s,"abc");
run would return true, because that’s what sigmaStar returns when run with the input
string "abc". But in this fragment:

String s =
"boolean ax(String p) { " +
" return (p.length()>0 && p.charAt(0)=='a'); " +
"} ";
run(s,"ba");

run would return false, since ax("ba") returns false. Finally, in this fragment:
String s =
"boolean anbncn1(String p) { " +
" String as = \"\", bs = \"\", cs = \"\"; " +
" while (true) { " +
" String s = as+bs+cs; " +
" if (p.equals(s)( return true; " +
" as += 'a'; bs += 'b'; cs += 'c'; " +
" } " +
"} ";
run(s,"abbc");
run would never return, because anbncn1("abbc") runs forever. Implementing run
would be a lot of work—see any textbook on compiler construction for the details—but it is
clearly possible.
The run method doesn’t quite fit our definition of a recognition method, because it
takes two input strings. We could make it fit by redefining so that it takes one delimited
input string: something like run(p+'#'+in) instead of run(p,in). That’s the kind
of trick we had to use in Chapter 17 to give a Turing machine more than one input; recall
 18.2 THE LANGUAGE LU 259
linearAdd(x#y). But to keep the code more readable, we will just relax our definitions a bit,
allowing recognition and decision methods to take more than one parameter of the String
type. With this relaxation, the run method qualifies as a recognition (but not a decision)
method.

18.2 The Language Lu

Programmers quickly learn about the perils of infinite computation. Perhaps you have
written a program with a bug like this one:

int j = 0;
for (int i = 0; i < 100; j++) {
j += f(i);
}

Oops! That postincrement expression should read i++, not j++. As it stands, the variable i
is never assigned any value other than 0, so the loop never stops. Not noticing this mistake,
you run the program. Perhaps that function f is something complicated, so you expect the
computation to take a few minutes anyway. Ten minutes go by, then twenty, then thirty.
Eventually you ask yourself the question programmers around the world have been asking
since the dawn of the computer age: is this stuck in an infinite loop, or is it just taking a long
time? There is, alas, no sure way for a person to answer such questions—and, as it turns out,
no sure way for a computer to find the answer for you.
The existence of our run method shows that a particular language is RE:
L(run) = {(p, in) | p is a recognition method and in L(p)}
The language is RE because we have a recognition method that accepts it—the run method.
A corresponding language for TMs is this:
{m#x | m encodes a TM and x is a string it accepts}
The language is RE because we have a TM that accepts it—any universal TM. In either case,
we’ll call the language Lu—remember u for universal.
Is Lu recursive, that is, is it possible to write a decision method with this specification?

/**
* shortcut(p,in) returns true if run(p,in) would
* return true, and returns false if run(p,in)
* would return false or run forever.
*/
boolean shortcut(String p, String in) {
...
}
260 CHAPTER EIGHTEEN: UNCOMPUTABILITY

The shortcut method would be just like the run method, but it would always produce
an answer—not run forever, even when run(p,in) would. For example, in this fragment:

String x =
"boolean anbncn1(String p) { " +
" String as = \"\", bs = \"\", cs = \"\"; " +
" while (true) { " +
" String s = as+bs+cs; " +
" if (p.equals(s)) return true; " +
" as += 'a'; bs += 'b'; cs += 'c'; " +
" } " +
"} ";
shortcut(x,"abbc");

shortcut would return false, even though anbncn1("abbc") runs forever.

How would you go about implementing something like shortcut? Presumably, it
would have to simulate the input program as run does, but somehow detect infinite loops.
Some infinite loops are easy to detect, of course. It is simple enough for an interpreter to
notice constructs like while(true) {}. But most infinite loops are subtler than that,
which is why they are such a plague for programmers. In the case of anbncn1, it is possible
to imagine a program clever enough to reason that the loop will not terminate once s is
longer than p; it could then insert the extra exit,

if (s.length()>p.length()) return false;

But is it possible to write shortcut so that it can detect and avoid all possible infinite
loops? It would certainly be valuable; imagine a debugging tool that could reliably alert you
to infinite computations! Unfortunately, we can prove that no such shortcut method
exists.
Such proofs are tricky; it isn’t enough to say that we tried really hard but just couldn’t
think of a way to implement shortcut. We need a proof that no such implementation is
possible. The proof is by contradiction. Assume by way of contradiction that Lu is recursive,
so some implementation of shortcut exists. Then we could use it to make this decision
method:

/**
* nonSelfAccepting(p) returns false if run(p,p)
* would return true, and returns true if run(p,p)
* would return false or run forever.
*/
 18.2 THE LANGUAGE LU 261
boolean nonSelfAccepting(String p) {
return !shortcut(p,p);
}

This decision method determines what the program would decide, given itself as input—
then returns the opposite. So the language defined by nonSelfAccepting is simply that
set of recognition methods that do not accept, given themselves as input. For example,

nonSelfAccepting(
"boolean sigmaStar(String p) {return true;}"
);

returns false. That’s because sigmaStar accepts all strings, including the one that starts
boolean sigmaStar .... In other words, sigmaStar accepts itself, and therefore
nonSelfAccepting returns false given sigmaStar as input. On the other hand,

nonSelfAccepting(
"boolean ax(String p) { " +
" return (p.length()>0 && p.charAt(0)=='a'); " +
"} "
);

tests whether the string boolean ax ... is accepted by the decision method ax.
Since the string begins with b, ax returns false. It does not accept itself, and therefore
nonSelfAccepting returns true given ax as input.
Now comes the tricky part. What happens if we call nonSelfAccepting, giving it
itself as input? We can easily arrange to do this:

nonSelfAccepting(
"boolean nonSelfAccepting(p) { " +
" return !shortcut(p,p); " +
"} "
)

What does nonSelfAccepting return, given itself as input? If it accepts itself, that
means shortcut determined it was not self-accepting—which is a contradiction. If it
rejects itself, that means shortcut determined it was self-accepting—also a contradiction.
Yet it must either accept or reject; it cannot run forever, because shortcut is a decision
method. We’re left with a contradiction in all cases. By contradiction, our original
assumption must be false: no program satisfying the specifications of shortcut exists. In
other words:
262 CHAPTER EIGHTEEN: UNCOMPUTABILITY

Theorem 18.1: Lu is not recursive.

This is our first example of a problem that lies outside the borders of computability. Lu is
not recursive; equivalently, we can say that the shortcut function is not computable and
the machine-M-accepts-string-x property is not decidable. This verifies our earlier claim that
total TMs are weaker than general TMs. No total TM can be a universal TM.

18.3 The Halting Problems

Consider the sentence “This sentence is false.” It’s a well-known paradox: if it’s true it must
be false, if it’s false it must be true. It achieves this paradox through self-reference. This is easy
to do in English: a sentence can refer to itself as “this sentence.” It’s also fairly easy to do with
computational procedures of all kinds: we can pass a method its own source as input or give a
TM the string encoding of itself as input. Like the “this sentence is false” paradox, our proof
that Lu is not recursive made use of self-reference.
For modern programmers, the idea of self-reference is not a big stretch. Every day we
work with programs that take other programs as input: compilers, interpreters, debuggers,
and the like. It is not uncommon to write such tools in the same language on which they
operate; there are C compilers written in C, ML compilers written in ML, and so on.
Such a compiler could be given itself as input without causing the modern programmer
any existential discomfort. In the 1930s, however, the idea that effective computational
procedures might have the power and the paradox that comes with self-reference was a major
insight.
Here’s another example using a similar self-reference contradiction. Consider this
language:
{(p, in) | p is a recognition method and in is a string for which it halts}
It is easy to see that this language is RE:

/**
* haltsRE(p,in) returns true if run(p,in) halts.
* It just runs forever if run(p,in) runs forever.
*/
boolean haltsRE(String p, String in) {
run(p,in);
return true;
}

A corresponding language for TMs is this:

{m#x | m encodes a TM and x is a string on which it halts}
In either case, we’ll call the language Lh—remember h for halting.
 18.3 THE HALTING PROBLEMS 263
Is Lh recursive; that is, is it possible to write a decision method with this specification?

/**
* halts(p,in) returns true if run(p,in) halts, and
* returns false if run(p,in) runs forever.
*/
boolean halts(String p, String in) {
...
}

The halts method would be just like the haltsRE method, but it would always produce
an answer—not run forever, even when run(p,in) would.
From our results about Lu, you might guess that Lh is not going to be recursive either.
Intuitively, the only way to tell what p will do when run on in is to simulate it—and if
that simulation runs forever, we won’t get an answer. But that kind of intuitive argument
is not strong enough to constitute a proof. How do we know there isn’t some other way of
determining whether p halts, a way that doesn’t involve actually running it?
The proof that this is impossible is by contradiction. Assume by way of contradiction
that Lh is recursive, so some implementation of halts exists. Then we could use it to make
this program:

/**
* narcissist(p) returns true if run(p,p) would
* run forever, and runs forever if run(p,p) would
* halt.
*/
boolean narcissist(String p) {
if (halts(p,p)) while(true) {}
else return true;
}

This method determines whether the program p will contemplate itself forever—that is, it
defines the language of recognition methods that run forever, given themselves as input.
Now comes that trick using self-reference. What happens if we call narcissist,
giving it itself as input? We can easily arrange to do this:

narcissist(
"boolean narcissist(p) { " +
" if (halts(p,p)) while(true) {} " +
264 CHAPTER EIGHTEEN: UNCOMPUTABILITY

" else return true; " +

"} "
)

Does it run forever? No, that can’t be right; if it did we would have a contradiction, since
it would be saying that it halts. So then does it halt (returning true)? No, that can’t be right
either; if it did it would be saying that it runs forever. Either way we have a contradiction.
By contradiction, we conclude that no program satisfying the specifications of halts exists.
This proves the following:

Theorem 18.2: Lh is not recursive.

At this point, we have identified our first two languages that are not recursive, and our
picture looks like this:

L(a*b*)
regular
languages {anbncn}

CFLs Lu
Lh
recursive
languages {anbn}

The non-recursive languages don’t stop there, however. It turns out that there are
uncountably many languages beyond the computability border. The question of whether
a program halts on a given input is a classic undecidable problem: a halting problem. It has
many variations: does a program halt on a given input? Does it halt on any input? Does it
halt on every input? That last variation would be the most useful. It would be nice to have
a program that could check over your code and warn you about all possible infinite loops.
Unfortunately, all these variations of the halting problem are undecidable. In fact, as we will
see below, most questions about the runtime behavior of TMs (or computer programs) turn
out to be undecidable.

18.4 Reductions Proving a Language Is Recursive

Suppose you are planning a trip from your house to a friend’s house. You formulate a plan:
1. I will drive my car to the airport.
2. I will fly to my friend’s airport.
3. My friend will pick me up.
Suppose that you know steps 1 and 3 are possible; that leaves you with step 2. You have
 18.4 REDUCTIONS PROVING A LANGUAGE IS RECURSIVE 265
reduced the original problem A (making a trip from house to house) to another problem B
(finding a flight from airport to airport). The reduction shows that if you can get a flight, you
can make the trip.
In other words, reducing problem A to problem B shows that A is no harder than B. Of
course, we have not ruled out the possibility that A might be easier than B, because there
might be other, simpler solutions. For example, you and your friend might already be in the
same city. In that case driving to the airport, getting on a plane, flying around, landing at the
same airport, and driving to your friend’s house is a correct solution but far from optimal.
The same idea is used in the design of algorithms: solve one problem by using the
existing solution to another. Given an original problem A, a reduction is a solution of this
form:
1. Convert the instance of problem A into an instance of problem B.
2. Solve that instance of problem B.
3. Convert the solution of the instance of problem B back into a solution of the
original instance of problem A.
If steps 1 and 3 are no harder than step 2, we can conclude that problem A is no harder
than problem B. (But problem A might still be easier than problem B, since there may be an
easier, completely different algorithm.)
Such reductions can be used to prove that a language is recursive. Suppose we are given
a language L1. One way to prove that L1 is recursive is to write a decision method (or a total
TM) for it. But another way is to give a reduction to some language L2 that we already know
is recursive:
1. Given a string x1 to be tested for membership in L1, convert it into another string x2
to be tested for membership in L2.
2. Decide whether x2 L2.
3. Convert that decision about x2 back into a decision about x1.
If steps 1 and 3 are computable—if those conversions can be computed effectively, without
infinite looping—and if L2 is already known to be recursive, a reduction like this proves that
L1 is recursive too.
For example, let’s use a reduction to prove that L1 = {x {a, b, d}* | x {anbnd n }} is
recursive. We’ll reduce it to the decision problem for L2 = {anbncn}.

a. boolean decideL1(String x1) {

b. String x2="";
c. for (int i = 0; i < x1.length(); i++) {
d. char ith = x1.charAt(i);
e. if (ith=='d') x2+='c';
f. else x2+=ith;
266 CHAPTER EIGHTEEN: UNCOMPUTABILITY

g. }
h. boolean b = anbncn2(x2);
i. return !b;
j. }

Lines b through g are step 1; they translate the string x1 into a new string x2, in this case by
replacing the ds with cs. Line h is step 2; it checks whether x2 L2, in this case using our
previously implemented decision method anbncn2. (In general, we don’t need to show all
the code for step 2; if we know that L2 is recursive, we know some decision method for it
exists, and that’s enough.) Line i is step 3; it converts the answer for x2 L2 into an answer
for x1 L1, in this case by logical negation. Steps 1 and 3 obviously cannot run forever, so
this reduction shows that if L2 is recursive (which we know it is) then L1 is recursive.
Here is another example of a proof using a reduction: we’ll prove that {anbn} is recursive
by reduction to the recognition problem for {anbncn}:

a. boolean anbn(String x1) {

b. String x2=x1;
c. for (int i = 0; i < x1.length()/2; i++)
d. x2+='c';
e. boolean b = anbncn2(x2);
f. return b;
g. }

This reduces the problem of deciding membership in {anbn} to the problem of deciding
membership in {anbncn}. As we have already shown that {anbncn} is recursive, we conclude that
{anbn} is recursive. Of course, we already knew that; in fact, {anbn} is not only recursive but
also context free. This illustrates the one-way nature of the inference you can draw from a
reduction. By reducing from a new problem to a problem with a known solution, we show
that the new problem is no harder than the known one. That reduction does not rule out the
possibility that the new problem might be easier.

18.5 Reductions Proving a Language Is Not Recursive

A reduction from A to B shows that A is no harder than B. Equivalently, we can say the
reduction shows that B is no easier than A. This can be used to prove that a language is not
recursive. Suppose we are given a language L1. One way to prove that is not recursive is by
contradiction, as we did in the previous sections for Lu and Lh. But another, often easier,
way is to give a reduction from some language L2 that we already know is not recursive. A
reduction in that direction—a reduction from a nonrecursive language L2 to the new language
L1—lets us conclude that L1 is not recursive, since it is no easier than L2. For example, define
 18.5 REDUCTIONS PROVING A LANGUAGE IS NOT RECURSIVE 267
Le = {p | p is a recognition method that never returns true}
In other words, Le is the set of recognition methods p for which L(p) = {} (remember e for
empty). Then we have the following:

Theorem 18.3.1: Le is not recursive.

Proof: By reduction from the halting problem. Assume by way of contradiction
that Le is recursive; then there is some decision method empty for it. We can
use it to write the following decision method:

1. boolean halts(String p, String x) {

2. String x2 =
3. "boolean f(String z) { " +
4. " run(\""+p+"\",\""+x+"\"); " +
5. " return true; " +
6. "} ";
7. boolean b = empty(x2);
8. return !b;
9. }

For any given program p and input string x, halts constructs a string x2
containing the source of a new decision method f. The f method ignores its
input string z. Instead of looking at z, it runs p on x, then returns true. Now
if p runs forever on x, f will not return, so the language of strings z it accepts
is {}. But if p halts on x, f will return true, so the language of strings z it
accepts is *. Thus x2 Le if and only if p runs forever on x. We conclude that
halts is a decision method for Lh. But this is a contradiction, since Lh is not
recursive. By contradiction, we conclude that Le is not recursive.

The reduction shows that Le is no easier to decide than Lh. But we know Lh is not
recursive; so we conclude that Le is not recursive. This is a very important application for
reductions. To show that a language is recursive is comparatively easy: you only need to give
a program that recognizes it and always halts. But to show that a language is not recursive is
generally more difficult. Often, the best choice is to give a reduction from a problem that is
already known to be nonrecursive.
Here’s another example along the same lines. Define
Lr = {p | p is a recognition method and L(p) is regular}
For example, the string

boolean sigmaStar(String p) {return true;}

268 CHAPTER EIGHTEEN: UNCOMPUTABILITY

is in Lr, because L(sigmaStar) = *, a regular language, while our previous decision

method anbn is not in Lr, because L(anbn) is not a regular language.

Theorem 18.3.2: Lr is not recursive.

Proof: By reduction from the halting problem. Assume by way of contradiction
that Lr is recursive; then there is some decision method regular for it. We
can use it to write the following decision method:

1. boolean halts(String p, String x) {

2. String x2 =
3. "boolean f(String z) { " +
4. " run(\""+p+"\",\""+x+"\"); " +
5. " return anbn(z); " +
6. "} ";
7. boolean b = regular(x2);
8. return !b;
9. }

For any given program p and input string x, halts constructs a string x2
containing the source of a new decision method f. The f method runs p on x,
then returns anbn(z). Now if p runs forever on x, f will not return, so the
language of strings z it accepts is {}—a regular language. But if p halts on x, f
will return anbn(z), so the language of strings z it accepts is {anbn}—not a
regular language. Thus x2 Lr if and only if p runs forever on x. We conclude
that halts is a decision method for Lh. But this is a contradiction, since Lh is
not recursive. By contradiction, we conclude that Lr is not recursive.

18.6 Rice’s Theorem

The previous two proofs clearly have considerable similarities. In both cases we used a
reduction from the halting problem to show that some set of programs was not recursive.
This can be generalized to a very powerful theorem, Rice’s theorem, which is usually stated
like this: All nontrivial properties of the RE languages are undecidable. Using our terminology,
we’ll state it like this:

Theorem 18.4 (Rice’s theorem): For all nontrivial properties , the language

{p | p is a recognition method and L(p) has the property }

is not recursive.
 18.5 REDUCTIONS PROVING A LANGUAGE IS NOT RECURSIVE 269
Here are some examples of languages covered by Rice’s theorem:
Le = {p | p is a recognition method and L(p) is empty}
Lr = {p | p is a recognition method and L(p) is regular}
{p | p is a recognition method and L(p) is context free}
{p | p is a recognition method and L(p) is recursive}
{p | p is a recognition method and |L(p)| = 1}
{p | p is a recognition method and |L(p)| 100}
{p | p is a recognition method and hello L(p) }
{p | p is a recognition method and L(p) = *}
Of course, Rice’s theorem also applies to the corresponding languages expressed in terms of
TMs (or any other Turing-equivalent formalism).
Rice’s theorem requires the property in question to be nontrivial. A property is trivial
if no RE languages have it or if all RE languages have it; for trivial properties the language
in question is either the empty set or the set of all recognition methods, both of which are
recursive. For example, Rice’s theorem does not apply to these two examples:
{p | p is a recognition method and L(p) is RE}
{p | p is a recognition method and L(p) *}
The first uses a property which is by definition true of all RE languages, so it is just the set of
all recognition methods. The second uses a property that is true of no RE languages, so it is
just the empty set.
The proof of Rice’s theorem is a generalization of the proofs we just saw for Le and Lr .
Proof: Let be any nontrivial property of the RE languages, and define
Α = {p | p is a recognition method and L(p) has the property }

We will show that Α is nonrecursive, by reduction from the halting problem.

Assume by way of contradiction that Α is recursive. Then there is some decision
method falpha that recognizes Α . Now either {} has the property or it does
not; we will deal with these two cases separately.
First, suppose {} has the property. Because is nontrivial, there must be at
least one RE language Y that does not have the property, and since Y is RE
there is some decision method fy that decides Y. Now construct

1. boolean halts(String p, String x) {

2. String x2 =
3. "boolean f(String z) { " +
270 CHAPTER EIGHTEEN: UNCOMPUTABILITY

4. " run(\""+p+"\",\""+x+"\"); " +

5. " return fy(z); " +
6. "} ";
7. boolean b = falpha(x2);
8. return !b;
9. }
For any given program p and input string x, halts constructs a string x2
containing the source of a new recognition method f. The f method runs p
on x, then returns fy(z). Now if p runs forever on x, f will not return, so
the language of strings it accepts is {}—a language with the property. But if
p halts on x, f will return fy(z), so the language of strings z it accepts is Y
—a language that does not have the property. Thus x2 Α if and only if p
runs forever on x. Under the assumption that falpha decides Α, we conclude
that halts decides Lh. But this is a contradiction, since Lh is not recursive. By
contradiction, we conclude that Α is not recursive.
That takes care of the case in which {} has the property. The other possibility
is that {} does not have the property. In that case, there must be at least one
RE language Y that does have the property, and since Y is RE there is some
decision method fy that decides Y. Using fy, we can construct halts as
before, but return b instead of !b on line 8. Again, by contradiction, we
conclude that Α is not recursive.
We can appeal to Rice’s theorem to make simple nonrecursiveness proofs for many
languages. One of our previous examples was
{p | p is a recognition method and |L(p)| = 1}
To show that this is not recursive, we merely observe that it is a language of the form covered
by Rice’s theorem and that the property in question (|L(p)| = 1) is nontrivial, because some
RE languages have one element and others don’t.
Rice’s theorem is a very useful tool. It applies to languages that are sets of recognition
methods (defined using TMs, high-level languages, or any other Turing-equivalent
formalism), selected for some nontrivial property of the language they accept. Rice’s theorem
covers many nonrecursive sets of programs, including the whole list of examples at the
beginning of this section. But although it can be very useful, it is not the right tool for every
nonrecursiveness proof. For example, consider these languages:
{p | p is a method that prints "hello world"}
{p | p is a method that never gets an uncaught exception}
{p | p is a method that produces no output}
 18.7 ENUMERATORS 271
These languages are not recursive, but you can’t prove them so using Rice’s theorem, because
they are not sets of recognition methods and they are not defined in terms of some property
of the language they accept. Each one needs a custom proof, like those we developed earlier
in this chapter.
Most questions about the runtime behavior of programs turn out to be undecidable,
even if you can’t apply Rice’s theorem directly. By contrast, most questions about program
syntax turn out to be decidable. For example, the question of whether a given method
contains the infinitely looping statement while(true){} is easy to decide—indeed,
the set of strings that contain while(true){} is a regular language. Likewise, questions
about the structure of TMs are generally decidable, like the question of whether a given TM
contains ten states.
To summarize: questions about program behavior are usually undecidable, and questions
about program syntax are usually decidable. But this is only a rough guideline, and there are
exceptions either way. Consider this language:
{(p, x) | p is a method that executes at least ten statements
when run with the input x}
The question of whether a method p executes at least ten statements when run with a given
input is easy to decide. You just start simulating p on x and count the number of statements
executed. If p returns before you get to 10, it isn’t in the language; if you get to 10, it is in
the language (and you can stop simulating). It’s a test that always gives you an answer within
ten statements, so there’s no trouble about infinite loops. The language is recursive, even
though it is a set of programs defined in terms of their runtime behavior.

18.7 Enumerators
We have generally treated Turing machines as language recognizers: machines that take an
input string and try to determine its membership in a language. When Alan Turing published
the 1936 paper introducing his model for effective computation, his concept was slightly
different. He envisioned his tape-memory automata as language enumerators: machines that
take no input but simply generate, on an output tape, a sequence of strings.
This is just another way of defining a language formally: L(M ) for such a machine is the
language of strings that eventually appear on the output tape:
L(M) = {x | for some i, x is the ith string in M ’s output sequence}
Like all Turing machines, enumerators may run forever. In fact, they must, if they enumerate
infinite languages; and they may even if they enumerate finite languages. In Java-like
notation we can think of these original Turing machines as enumerator objects:

An enumerator class is a class with an instance method next that takes no input
and returns a string (or runs forever).
272 CHAPTER EIGHTEEN: UNCOMPUTABILITY

An enumerator object may preserve state across calls of next, so next may (and generally
does) return a different string every time it is called. For an enumerator class C, we’ll take
L(C) to be the set of strings returned by an infinite sequence of calls to the next method of
an object of the C class.
For example, this is an enumerator class AStar with L(AStar) = {a}*:

class AStar {
int n = 0;

String next() {
String s = "";
for (int i = 0; i < n; i++) s += 'a';
n++;
return s;
}
}

For this enumerator, the jth call to next (counting from 0) will return a string of j as. In
general, an enumerator class won’t be limited to enumerating the strings in order of their
length, but this one just happens to be specified that way.
For a more interesting example, suppose we have a method isPrime that tests whether
a number is prime. Then we can implement this enumerator class:

class TwinPrimes {
int i = 1;

String next() {
while (true) {
i++;
if (isPrime(i) && isPrime(i+2))
return i + "," + (i+2);
}
}
}

This is an enumerator for a language of twin primes. A twin prime is a pair of numbers
(i, i+2) that are both prime. So a series of calls to next would return

3,5
5,7
 18.7 ENUMERATORS 273
11,13
17,19
29,31

and so on. There is a famous conjecture that there are infinitely many twin primes, but no
one has been able to prove it. So the language enumerated by TwinPrimes may or may
not be infinite. If it is not, there is some largest pair of twin primes, and a call made to next
after that largest pair has been returned will run forever.
Here’s an enumeration problem whose solution we’ll need later on: make an enumerator
class for the set of all pairs of numbers from N, {(j, k) | j 0, k 0}. (As usual, we'll represent
the natural numbers j and k as decimal strings.) This is a bit trickier. This enumerates all the
pairs (0, k):

class BadNatPairs1 {
int k = 0;

String next() {
return "(0," + k++ + ")";
}
}

But that’s not good enough. This enumerates all the pairs (j, k) where j = k, and that’s not
good enough either:

class BadNatPairs2 {
int j = 0;
int k = 0;

String next() {
return "(" + j++ + "," + k++ + ")";
}
}

Instead we need to visit the ( j, k) pairs in an order that eventually reaches every one. This is
one way:

class NatPairs {
int n = 0;
int j = 0;
274 CHAPTER EIGHTEEN: UNCOMPUTABILITY

String next() {
String s = "(" + j + "," + (n-j) + ")";
if (j<n) j++;
else {j=0; n++;}
return s;
}
}

It enumerates the ( j, k) pairs in a triangular order, like this:

j=4

j=3

j=2
etc.
j=1

j=0
k=0 k=1 k=2 k=3 k=4 k=5

so every point in the space, every pair ( j, k) in the language, is reached eventually.
For a last example, imagine defining an enumerator class SigmaStar that enumerates
*. If = {a, b}, a SigmaStar object might return these strings for a sequence of calls to
next:

""
"a"
"b"
"aa"
"ab"
"ba"
"bb"
"aaa"

and so on. The exact order in which strings are returned doesn’t matter for our purposes.
SigmaStar is not too difficult to implement, and this is left as an exercise.
Enumerators generate the strings in a language. You can also think of them as numbering
the strings in the language—since the strings are generated in a fixed order. For example, we
 18.8 RECURSIVELY ENUMERABLE LANGUAGES 275
could define a method sigmaStarIth(i) that takes a natural number i and returns the
String that is the ith one enumerated by SigmaStar:

String sigmaStarIth(int i) {
SigmaStar e = new SigmaStar();
String s = "";
for (int j = 0; j<=i; j++) s = e.next();
return s;
}

By the way, nothing in our definitions requires enumerators to generate a unique string on
every call, so sigmaStarIth does not necessarily give a one-to-one mapping from N to
*. In mathematical terms it is merely a surjection (an onto function): for every string s *
there is at least one i such that sigmaStarIth(i) = s.
We’ll make use of both sigmaStarIth and NatPairs in the next section.

18.8 Recursively Enumerable Languages

It turns out that Turing’s original enumeration machines are equivalent to our TM model, in
the following sense:

Theorem 18.5: A language is RE if and only if it is L(M) for some enumeration

machine M.

The term recursively enumerable makes more sense in this light!

We know that a language is RE if and only if it has a recognition method. So to prove
the theorem we will give two constructions. First, given an enumerator class AEnumerate
with L(AEnumerate) = A, we construct a recognition method aRecognize, showing
that A is an RE language. This is straightforward.

boolean aRecognize(String s) {
AEnumerate e = new AEnumerate();
while (true)
if (s.equals(e.next())) return true;
}

This method returns true if and only if s is eventually enumerated by AEnumerate. Thus
L(aRecognize) = L(AEnumerate).
The construction in the other direction is a bit trickier. Given a recognition method
aRecognize for some RE language A, we need to show that we can construct an
276 CHAPTER EIGHTEEN: UNCOMPUTABILITY

enumerator class AEnumerate for the same language. Here is an attempt that doesn’t quite
work. It uses the SigmaStar enumerator to generate each possible string in *. To find the
next string in A, it tries successive strings in * until it finds one that aRecognize accepts:

class BadAEnumerate {
SigmaStar e = new SigmaStar();

String next() {
while (true) {
String s = e.next();
if (aRecognize(s)) return s;
}
}
}

That would work just fine if aRecognize were total—a decision method and not just a
recognition method. But if aRecognize runs forever on one of the strings generated by
SigmaStar, next will get stuck. It will never get around to testing any of the subsequent
strings.
To solve this problem we will introduce a time-limited version of run. Define
runWithTimeLimit(p,in,j) so that it returns true if and only if p returns true
for in within j steps of the simulation. (Various definitions of a “step” are possible here:
a statement, a virtual machine instruction, and so on. It doesn’t matter which one you
use. For TMs the meaning of “step” would be more obvious: one move of the TM.) This
runWithTimeLimit can be total, because it can return false if p exceeds j steps without
reaching a decision about in.
Using runWithTimeLimit, we can make an enumerator for A. For each ( j, k) pair,
we check whether the jth string in * is accepted by aRecognize within k steps. This is
where we can use the NatPairs enumerator and the sigmaStarIth method from the
previous section. NatPairs enumerates all the ( j, k) pairs we need, and sigmaStarIth
finds indexed strings in *.
class AEnumerate {
NatPairs e = new NatPairs();

String next() {
while (true) {
int (j,k) = e.next();
String s = sigmaStarIth(j);
 18.9 LANGUAGES THAT ARE NOT RE 277
if (runWithTimeLimit(aRecognize,s,k)) return s;
}
}
}

(This uses some syntax that is not proper Java. The NatPairs method next returns a
string giving a pair of natural numbers such as "(13,7)", and it would take a few real Java
statements to break that string apart and convert the numeric strings back into integer values
j and k.)
Clearly AEnumerate only enumerates strings accepted by aRecognize. Further,
every string s accepted by aRecognize is eventually enumerated by AEnumerate,
because every such s = sigmaStarIth(j) for some j, and every such s is accepted
by aRecognize within some number of steps k, and AEnumerate tries all such
pairs (j,k). Thus, L(AEnumerate) = L(aRecognize). That completes the proof of
Theorem 18.5.

18.9 Languages That Are Not RE

We have seen examples of nonrecursive languages like Lh and Lu. Are there languages that are
not only nonrecursive, but also nonRE? Yes, and they are easy to find.

Theorem 18.6: If a language is RE but not recursive, its complement is not RE.
Proof: Let L be any language that is RE but not recursive. Assume by way of
contradiction that L is also RE. Then there exist recognition methods lrec
and lbar for both languages, so we can implement

boolean ldec(String s) {
for (int j = 1; ; j++) {
if (runWithTimeLimit(lrec,s,j)) return true;
if (runWithTimeLimit(lbar,s,j)) return false;
}
}

If s L, then runWithTimeLimit(lrec,s,j) returns true for some j,

so ldec(s) returns true. If s L, then
runWithTimeLimit(lbar,s,j)
returns true for some j, so ldec(s) returns false. Thus ldec is a decision
method for L. But this is a contradiction, because L is not recursive. By
contradiction, L is not RE.
278 CHAPTER EIGHTEEN: UNCOMPUTABILITY

From this we conclude that the RE languages are not closed for complement. Notice that the
recursive languages are closed for complement; given a decision method ldec for L, we can
construct a decision method for L very simply:

boolean lbar(String s) {return !ldec(s);}

But when a language is RE but not recursive, this kind of construction fails. If the
recognition method lrec(s) runs forever, !lrec(s) will too.
At the beginning of this chapter we saw that the languages Lh and Lu are RE but not
recursive. From Theorem 18.6 it follows that these complements are not RE:
Lu = {(p, s) | p is not a recognition method that returns true for s}
Lh = {(p, s) | p is not a recognition method that halts given s}

18.10 Language Classifications Revisited

Here’s a summary picture of our language categories, with an example language for each of
the five regions:

L(a*b*)
regular
languages {anbncn}

CFLs Lu
recursive Lu
languages
{anbn}
RE languages

When a language is recursive, there is an effective computational procedure that can

definitely categorize all strings. Given a positive example it will decide yes; given a negative
example it will decide no. A language that is recursive, a property that is decidable, a function
that is computable: these terms all refer to computations like those performed by total TMs,
computations that always halt.
By contrast, languages that are RE but not recursive are those for which only the positive
examples can be definitely categorized. A TM for such a language will definitely answer
yes, given any string in the language. But given a string that is not in the language, the TM
may run indefinitely without reaching a decision. You get a definite yes if the answer is yes,
but you don’t get a definite no if the answer is no. A property like that is sometimes called
semi-decidable. For example, the language Lh is RE but not recursive, and the corresponding
 18.11 GRAMMARS AND COMPUTABILITY 279
property (p halts on s) is said to be semi-decidable. If p does halt on s, a simple simulation
will always give you a definite yes answer, but if not, neither simulation nor any other
approach will always give you a definite no answer.
When a language is not RE, that means that there is no procedure for categorizing
strings that gives a definite yes answer on all positive examples. Consider the language Lh , for
example. One kind of positive example for this language would be a pair (p, s) where p is a
recognition method that runs forever given s. But there is no way to identify such pairs—
obviously, you can’t simulate p on s, see if it runs forever, and then say yes!

18.11 Grammars and Computability

We defined grammars using productions of the general form x y, where x and y are any
strings with x ε. But all our examples have followed a more restricted form: they have been
context-free grammars, grammars with a single nonterminal symbol on the left-hand side of
every production. Using productions that are not context free, it is possible to define more
languages.
For example, consider this grammar:
S aBSc | abc | ε
Ba aB
Bb bb
Here are some derivations for this grammar:
S ε
S abc
S aBSc aBabcc aaBbcc aabbcc
S aBSc aBaBScc aBaBabccc aaBBabccc aaBaBbccc
aaaBBbccc aaaBbbccc aaabbbccc

The language generated is {anbncn}: recursive but not context free.

In the late 1950s, Noam Chomsky proposed four classifications for grammars, using
syntactic restrictions on the forms of productions. These are
• Type 0 (unrestricted): All forms allowed.
• Type 1 (context sensitive): All productions are of the form xAz xyz, where
y ε. Such a production is called context sensitive because it allows A to be replaced
with y, but only in a particular context, with x to the left and z to the right. (The
production S ε is also allowed if S does not appear on the right side of any
production.)
• Type 2 (context free): As in our definition in Section 12.1.
• Type 3 (right linear): As in our definition in Section 10.5.
280 CHAPTER EIGHTEEN: UNCOMPUTABILITY

Grammars of these four types define a hierarchy of four classes of languages called the
Chomsky hierarchy:

Type 3 Type 2 Type 1 recursive Type 0

(regular) (CFL) (CSL) (RE)

As the illustration suggests, there is a remarkable correspondence between the Chomsky

grammar classifications and some of the major classes of formal languages. We have already
seen that a language is regular if and only if it is generated by some type-3 grammar and that
a language is context free if and only if it is generated by some type-2 grammar. The type-0,
unrestricted grammars also generate a familiar class of languages: the RE languages. Thus
unrestricted grammars are yet another Turing-equivalent formalism. In fact they are closely
related to the Post systems already mentioned.
Type-1 grammars also generate an interesting class of languages, one that we have not
mentioned before: the context-sensitive languages (CSLs). The CSLs are a superset of the CFLs
and a subset of the recursive languages. Recursive languages encountered in practice usually
turn out to be CSLs. There are languages that are recursive but not context sensitive, but
they’re hard to find.
There is another interesting way to define the CSLs. Start with the TM model, add
nondeterminism, and impose the restriction that writing anything over a B is not permitted.
In effect, the TM’s infinite-tape feature is eliminated—it is restricted to using that part
of the tape initially occupied by the input string. The resulting automaton is called a
nondeterministic linear-bounded automaton (NLBA). It turns out that a language is accepted
by some NLBA if and only if it is a CSL! So just like our other language classes, the CSLs
arise unexpectedly in several different ways.
Even the CFLs, familiar as they are, harbor some computability surprises. Rice’s theorem
shows that for all nontrivial properties , the language
{p | p is a recognition method and L(p) has the property }
is not recursive. Although there is nothing as categorical as Rice’s theorem for CFGs, there
are a number of interesting properties for which the language
 18.12 ORACLES 281
{G | G is a CFG and L(G) has the property }
is not recursive. For example, these languages are not recursive:
{G | G is a CFG and L(G) = *}
{G | G is a CFG and L(G ) is a CFL}
Similarly, the question of whether two CFGs generate the same language is undecidable, as is
the question of whether the intersection of two CFGs is context free.
Formal language is an area where computer science and linguistics overlap. Although
the Chomsky hierarchy is well known in computer science, Noam Chomsky is even better
known in linguistics, where his theories on human language acquisition have been very
influential.

18.12 Oracles
Here are two languages that Rice’s theorem shows are nonrecursive:
Le = {p | p is a recognition method and L(p) = {}}
Lf = {p | p is a recognition method and L(p) = *}
The first is the set of recognition methods that always return false or run forever; the second
is the set of recognition methods that always return true. Neither one is recursive, and neither
is RE. But there is a sense in which one is much harder to recognize than the other.
To the purely practical mind, that last statement is meaningless. Undecidable is
undecidable; everything beyond the computability border is out of reach, so why worry
about additional distinctions? But for those with active mathematical imaginations, the
unreachable territory beyond the computability border has a fascinating structure.
Earlier in this chapter, we proved that Lh is not recursive. Later, we showed that Le is
not recursive, using a reduction from Lh. In other words, we showed that if there were a way
to decide Le, we could use that to decide Lh. Our conclusion at the time was that, since we
already know there is no decision method for Lh, there couldn’t be one for Le either. But
what if there were some way of deciding membership in Le? We know that empty can’t be a
decision method in our ordinary sense, but what if it decides Le somehow anyway? Then our
construction would allow us to decide Lh as well.
Turing machines with such impossible powers are called oracle machines. They are just
like ordinary Turing machines, except that they have a one-step way of checking a string’s
membership in a particular language. The language of this oracle can be anything we choose.
Giving a TM an oracle for a nonrecursive language like Le increases the range of languages
it can decide. In particular, our previous construction shows that given an oracle for Le, the
language Lh is recursive. With a different construction you can also show the reverse: given an
oracle for Lh, the language Le is recursive.
But an oracle for Lh would not put an end to uncomputability. In particular, although
it can decide the halting problem for ordinary Turing machines, it cannot decide the halting
282 CHAPTER EIGHTEEN: UNCOMPUTABILITY

problem for Turing machines with Lh oracles! That would require a stronger oracle, whose
addition to TMs would make the halting problem even harder, requiring a still stronger
oracle, and so on. The result is an infinite hierarchy of oracles. It turns out that Le is recursive
given an oracle for Lh, but Lf is not—it requires a more powerful oracle. In that sense, Lf is
harder to decide than Le.

18.13 Mathematical Uncomputabilities

All the nonrecursive languages we have looked at have been languages of programs, like the
languages covered by Rice’s theorem. These languages are of natural interest to programmers,
but uncomputability turns up in many other domains as well. In particular, it turns up at the
foundations of mathematics.
Mathematics can be thought of as a structure of theorems, each of which can be proved
from simpler theorems by mechanically following the rules of logic. The whole edifice of
mathematics can be built up in this way, starting from the foundational axioms of arithmetic.
The foundational axioms are not proved, but are accepted as true because they are simple
and self-evident. If you think of mathematics this way, it is important to know that the
axiom foundation is firm. The axioms should be consistent, meaning that they lead to no false
theorems. They should be complete, meaning that all true theorems can be proved.
David Hilbert was a champion of this view of mathematics. He is widely regarded as one
of the greatest mathematicians of his time. He contributed to many different branches of
mathematics, but he is especially well known for a list of 23 open problems that he presented
in a lecture at an international conference in Paris in 1900. Hilbert’s list drove a great deal of
20th-century mathematical research, and a solution to any problem on the list brought fame
to the mathematician who solved it. One of the problems on Hilbert’s list (the second) was
the problem of proving that the foundational axioms of mathematics are consistent. He was
convinced that such a proof was possible. In general, he was confident that finite proof or
disproof was always possible for well-formed mathematical conjectures. At the time, most of
the world’s mathematicians would have agreed with him.
But it turned out far otherwise. In 1930, Kurt Gödel published his two famous
incompleteness theorems. First, he showed that there is a mathematical assertion that says, in
effect,
This assertion has no proof.
That kind of thing is easy enough to write in English and not hard to do with a computer
program. But Gödel’s genius was to express that self-referential idea in the extremely limited
language of formal mathematical assertions about arithmetic in N. If the assertion is false,
then you have a proof of it—a proof of something false, showing that the axioms are not
consistent. If the assertion is true, then you have an assertion that is true but has no proof,
showing that the axioms are not complete. Gödel’s results apply not just to a particular
 18.13 MATHEMATICAL UNCOMPUTABILITIES 283
axiomatic system, but to any axiomatic system capable of expressing the same basic things
about arithmetic. No such system can be both consistent and complete. Gödel’s second
incompleteness theorem then goes on to show that no consistent axiomatic system that
captures arithmetic can prove its own consistency. Gödel’s results came as quite a surprise to
those who expected a quick solution to Hilbert’s second problem.
For long years before Hilbert, mathematicians had been wishing for an effective
computational procedure to find proofs. It was a dream of Gottfried Wilhelm von Leibnitz,
who demonstrated a mechanical calculator in 1673 and hoped to bring its computational
advantages to the realm of logic. Hilbert considered this to be the central challenge of
mathematical logic. Gödel’s results showed that some true theorems must remain unprovable,
but left open the question of whether the provable ones could be identified computationally.
In 1936, both Alonzo Church and Alan Turing independently added a related result: the
provability of theorems of arithmetic is undecidable. Turing’s proof related arithmetic to
Turing machines, showing that you can write
Turing machine M halts on input x.
as a mathematical assertion about arithmetic in N.
Since then, many other mathematical problems have been found to lie on the far
side of the computability border. One example is the problem of finding solutions to
Diophantine equations. A Diophantine equation is a polynomial equation using integer
variables and constants; an example is the equation x2 + y2 = z2, whose integer solutions
are the Pythagorean triples like (x=3, y=4, z=5). The problem of finding solutions to such
equations was another of Hilbert’s (the tenth) and remained open until 1970, when Yuri
Matiyasevich showed that it could not be solved. Specifically, he showed that for every Turing
machine M there is some Diophantine equation with a variable x that has solutions exactly
where x L(M). In effect, such equations cannot really be solved; they encode computational
procedures, possibly nonterminating ones, that can only be simulated. This shows again the
close ties between computer science and the foundations of mathematics.

Exercises
EXERCISE 1
Implement Java decision methods for the following languages:
a. {abc}
b. L(a*b*)
c. {anb2n}
d. {xx | x {a, b}*}
e. {x {0, 1}* | |x| is a prime number}
f. {x {0, 1}* | x is a binary representation of an odd number}
284 CHAPTER EIGHTEEN: UNCOMPUTABILITY

EXERCISE 2
For each of these Java decision methods, give a total Turing machine for the same
language.
a. boolean ax(String p) {
return (p.length()>0 && p.charAt(0)=='a');
}
b. boolean emptySet(String p) {
return false;
}
c. boolean sigmaStar(String p) {
return true;
}
d. boolean d1(String p) {
String s = "";
int n = p.length();
for (int i = 0; i<n; i++) s = p.charAt(i) + s;
return p.equals(s);
}
EXERCISE 3
For each of these Java recognition methods, give a Java decision method that accepts the
same language.
a. boolean r1(String p) {
while (true) {}
return true;
}
b. boolean r2(String p) {
if (p.length()<10) return true;
else return r2(p);
}
c. boolean r3(String p) {
while (true) {
if (p.length()<10) p += 'a';
}
return (p.length()>=10);
}
d. boolean r4(String p) {
 EXERCISES 285
String s = "";
while (true) {
s += 'a';
if (s.equals(p)) return true;
}
return false;
}
EXERCISE 4
Given the shortcut and halts methods as described in Sections 18.2 and 18.3,
what would be the outcome of these code fragments? (Of course, we proved no such
methods exist; but assume here that we have some way of meeting their specifications.)
a. String s =
"boolean sigmaStar(String p) { " +
" return true; " +
"} ";
shortcut(s,"fourscore and seven years ago");
b. String s =
"boolean sigmaStar(String p) { " +
" return true; " +
"} ";
halts(s,"fourscore and seven years ago");
c. String s =
"boolean emptySet(String p) { " +
" return false; " +
"} ";
shortcut(s,s);
d. String s =
"boolean emptySet(String p) { " +
" return false; " +
"} ";
halts(s,s);
e. String s =
"boolean r1(String p) { " +
" while (true) {} " +
" return true; " +
"} ";
shortcut(s,"abba");
286 CHAPTER EIGHTEEN: UNCOMPUTABILITY

f. String s =
"boolean r1(String p) { " +
" while (true) {} " +
" return true; " +
"} ";
halts(s,"abba");
EXERCISE 5
Consider the nonSelfAccepting method used in Section 18.2 to prove by
contradiction that shortcut cannot be implemented. If we reimplement it like this:
boolean nonSelfAccepting(String p) {
return !run(p,p);
}
does it lead in the same way to a contradiction proving that run cannot be
implemented? Carefully explain why or why not.
EXERCISE 6
Prove that this language is not recursive:
{p | p is a recognition method that returns false for the empty string}
Hint: Follow the pattern in Section 18.5; you can’t use Rice’s theorem for this one.
EXERCISE 7
Prove that this language is not recursive:
{p | p is a method that takes no parameters and returns the string “Hello world”}
Hint: Follow the pattern in Section 18.5; you can’t use Rice’s theorem for this one.
EXERCISE 8
For each of the following languages, decide whether Rice’s theorem applies, and explain
why it does or does not.
a. {p | p is a recognition method and L(p) contains the empty string}
b. {p | p is a recognition method and L(p) *}
c. {m | m encodes a Turing machine M with abba L(M )}
d. {p | p is a recognition method that contains the statement x=1;}
e. {p | p is a recognition method that, when run on the empty string, executes the
statement x=1;}
f. {m | m encodes a Turing machine M for which L(M ) is finite}
 EXERCISES 287
EXERCISE 9
Classify the following languages as either
A. regular,
B. context free but not regular,
C. recursive but not context free, or
D. not recursive,
and explain for each why you chose that answer. You do not have to give detailed
proofs—just give one or two sentences that outline how you would prove your
classification.
a. {m#x | m encodes a TM and x is a string it accepts}
b. {anbncn}
c. {x {a, b}* | the number of as in x is the same as the number of bs in x }
d. {anb2n}
e. {p | p is a recognition method that is syntactically legal (that is, it can be parsed
using a standard grammar for Java)}
f. the set of strings that contain the Java statement while(true){}
g. {(p, x) | p is a recognition method that, when run on the input x, executes the
statement while(true){}}
h. {p | p is a recognition method and L(p) {}}
i. {m | m encodes a TM M with L(M) {}}
j. {xx | x ∈ {a, b, c }* }
k. {(m, x) | m encodes a TM M that, when run on the input x, visits a state q7}
l. L(a*b*) L((a + ba*b)*)
m. {xxR | x {0, 1}*}
EXERCISE 10
Taking the alphabet to be = {a, b }, implement a Java enumerator class SigmaStar.
Generate the strings in order of their length: first ε, then the strings of length 1, then
the strings of length 2, and so on. Among strings of the same length any fixed order is
acceptable.
EXERCISE 11
Repeat Exercise 10, but add a String parameter to the SigmaStar constructor, and
take to be the set of characters in that String. You may assume that the string is not
empty, so {}.
EXERCISE 12
Implement a Java decision method for {xx | x {a, b}*}.
EXERCISE 13
Implement a Java enumerator class for {xx | x {a, b}*}.
288 CHAPTER EIGHTEEN: UNCOMPUTABILITY

EXERCISE 14
In the proof of Theorem 18.5, how often will each string in the language recognized by
aRecognize be enumerated by AEnumerate? Explain.
EXERCISE 15
Suppose you’re given a Java decision method javaRec that decides the language {p | p
is a Java recognition method}. (The javaRec method would verify that the string p
compiles without error as a Java method, that it has a single String parameter, and
that it returns a boolean value.) Using this, write an enumerator class for {p | p is a
Java recognition method}.
EXERCISE 16
Using javaRec defined as in Exercise 15, write an enumerator class for the
nonrecursive language {p | p is a Java recognition method and ε L(p)}. Hint: This
is like the construction of AEnumerate in Section 18.8. You’ll need NatPairs,
runWithTimeLimit, and your solution to Exercise 15.
EXERCISE 17
Using javaRec defined as in Exercise 15, write an enumerator class for the
nonrecursive language {p | p is a Java recognition method and L(p) {}}. Hint: You may
assume you have an enumerator class NatTriples that enumerates natural-number
triples (i, j, k).
EXERCISE 18
Let AEnumerate be an enumerator class with the special property that it enumerates
the strings of an infinite language A in order of their length, shortest first. Give a decision
method for A, thus proving that A is recursive.
EXERCISE 19
Give a grammar for the language {anbncnd n}, and show a derivation of the string
aaabbbcccddd.
EXERCISE 20
Give a grammar for the language {xx | x {a, b}*}, and show a derivation of the string
aabaab.
EXERCISE 21
Using decision methods, show that an oracle for Lh makes Lu recursive.
EXERCISE 22
Using decision methods, show that an oracle for Lu makes Lh recursive.
EXERCISE 23
Using decision methods, show that an oracle for Le makes Lu recursive.
EXERCISE 24
Using decision methods, show that an oracle for Lu makes Le recursive. (Hint: See the
construction of AEnumerate in Section 18.8.)
 19
CHA P TER

Cost Models

Up to this point, we have focused on computability without

considering efficiency. As practical programmers, however, we are
naturally interested in understanding the resource requirements
of the code we write. How much time a program needs and how
much memory space it needs are important practical questions. A
cost model is a tool for answering such questions systematically. In
this chapter, we define cost models for TMs and for our Java-like
programs.

289
290 CHAPTER NINETEEN: COST MODELS

19.1 Asymptotic Notation

If you’ve studied algorithms and data structures, you’re probably already familiar with the
big-O notation and its close cousins, and . If not, here's a quick introduction.
The big-O notation is used to say something trenchant about a complicated function f
by relating it to a simpler function g.

Let f and g be any two functions over N. We say

f(n) is O(g(n))
if and only if there exist natural numbers c and n0 so that for every
n n0, f(n) c g(n).

Typically, the function f is the complicated function whose rate of growth we’re trying to
characterize; the function g is some simple function, like g(n) = n2, whose rate of growth is
easy to grasp. The assertion f(n) is O(g(n)) is a way of saying that f grows no faster than g.
If you multiply g(n) by some constant c, and plot the result, it will be above the plot of f(n)
almost everywhere—like this:

c g(n)

f (n)

f (n) is O(g(n))

n0

Some exceptions to f(n) c g(n) may occur among small values of n, those less than n0; the
assertion “f(n) is O(g(n))” says that some such constant n0 exists, but does not say what it is,
just as it does not say what the constant of proportionality c is. As the illustration suggests,
the big-O notation describes upper bounds. When you assert that f (n) is O(g(n)) you are
saying that f grows no faster than g; that doesn’t rule out the possibility that f actually grows
much slower than g. For example, if f(n) is O(n2), then by definition it is also true (though
less informative) to say that f(n) is O(n3), O(n4), or O(n100).
When you want to describe a lower bound on a function—to say that f grows at least as
fast as g—there is a related asymptotic notation using the symbol .
 19.1 ASYMPTOTIC NOTATION 291

Let f and g be any two functions over N. We say

f(n) is (g(n))
if and only if there exist natural numbers c and n0 so that for every
n n0, f (n) ≥ 1c g(n).

Thus, if you divide g(n) by some constant c and plot the result, it will be below the plot of
f (n) almost everywhere—like this:

f (n)

f (n) is Ω(g(n))
1 g(n)
c

n0

Again, some exceptions may occur among small values of n, those less than n0.
As the illustration suggests, the notation describes lower bounds. When you assert
that f (n) is (g(n)) you are saying that f grows at least as fast as g; that doesn’t rule out the
possibility that f actually grows much faster than g. For example, if f (n) is (n3), then by
definition it is also true (though less informative) to say that f (n) is (n2), (n), or (1).
If you inspect the definitions for O and , you'll see that it is possible for a function
f(n) to be both O(g(n)) and (g(n)) for the same function g. In fact, this is a very important
situation: when your upper and lower bounds are the same, you know that you have an
asymptotic characterization of the function f that is as informative as possible. There is a
special notation for this:

Let f and g be any two functions over N. We say

f(n) is (g(n))
if and only if f(n) is O(g(n)) and f (n) is (g(n)).

The important thing to observe here is that the constants involved, c and n0, are not the
same for both the upper bound and the lower bound. For the upper bound O(g(n)) we have
one pair of constants c and n0, and for the lower bound (g(n)) we have a different pair of
constants—call them c' and n0':
292 CHAPTER NINETEEN: COST MODELS

f (n)

c g(n)

f (n) is Θ(g(n))
1 g(n)
c'

n0 n'0

The notation is used to describe tight bounds. When you assert that f(n) is (g(n)) you
are saying that f grows at the same rate as g, neither faster nor slower.

19.2 Properties of Asymptotics

The asymptotic notations have properties that allow many functions to be simplified. That’s
exactly what makes them useful; they can describe a function that is difficult to understand
by relating it to one that is easy to understand. One important simplifying property is that
constant factors can generally be eliminated.

Theorem 19.1.1, O version: For any nonzero k N, f (n) is O(k·g(n)) if and

only if f(n) is O(g(n)).
Proof: If f(n) is O(k·g(n)) then, by definition, there exist natural numbers c and
n0 so that for every n n0, f(n) ck·g(n). Let c' = ck. Now, for every n n0,
f(n) c'·g(n). Thus, by definition, f (n) is O(g(n)). (The other direction of the
proof is trivial, because g(n) k·g(n).)

This property says, for example, that O(3n2) and O(n2) are equivalent. Thus you should
never rest with the conclusion that a function is O(3n2), because O(n2) is equivalent and
simpler.
The same theorem also applies for and , with similar proofs:

Theorem 19.1.1, version: For any nonzero k N, f (n) is (k·g(n)) if and

only if f(n) is (g(n)).
Theorem 19.1.1, version: For any nonzero k N, f(n) is (k·g(n)) if and
only if f(n) is (g(n)).
 19.2 PROPERTIES OF ASYMPTOTICS 293
Another important simplification is that in a sum of terms, all but the fastest growing
can be eliminated.

Theorem 19.1.2, O version: Let {g1, …, gm} be any finite set of functions over
N in which g1 is maximal, in the sense that every gi(n) is O(g1(n)). Then for any
function f, f(n) is O(g1(n) + g2(n) + … + gm(n)) if and only if f (n) is O(g1(n)).
Proof: If f(n) is O(g1(n) + g2(n) + … + gm(n)) then, by definition, there exist
natural numbers c0 and n0 so that for every n n0,

f (n) ≤ c0(g1(n) + ... + gm(n))

We are given that every term gi(n) is O(g1(n)), so by definition there exist
natural numbers {c1, …, cm} and {n1, …, nm} such that, for every i and every
n max(n0, …, nm), gi(n) cig1(n). Therefore, for every n max(n0, …, nm),

f (n) ≤ c0(g1(n) + ... + gm(n))

f (n) ≤ c0(c1 g1(n) + ... + cm g1(n))
f (n) = c0(c1 + ... + cm )g1(n)

By choosing c' = c0(c1 + … + cm) and n' = max(n0, …, nm), we can restate
this as follows: for every n n', f(n) c'·g1(n). Thus, by definition, f (n) is
O(g1(n)). (In the other direction, the proof is trivial, using the fact that
g1(n) g1(n) + g2(n) + … + gm(n).)

Because of this property, sums of terms inside the big O can almost always be simplified.
You just identify the fastest growing term and discard the others. For example, you would
never conclude that a function is O(n2 + n + log n + 13), because O(n2) is equivalent and
simpler. If you ever find yourself writing “+” inside the big O, think twice; it’s almost always
possible to simplify using Theorem 19.1.2.
Theorem 19.1.2 also applies to the and forms:

Theorem 19.1.2, version: Let {g1, …, gm} be any finite set of functions over
N in which g1 is maximal, in the sense that every gi(n) is O(g1(n)). Then for any
function f, f(n) is (g1(n) + g2(n) + … + gm(n)) if and only if f(n) is (g1(n)).
Theorem 19.1.2, version: Let {g1, …, gm} be any finite set of functions over
N in which g1 is maximal, in the sense that every gi(n) is O(g1(n)). Then for any
function f, f(n) is (g1(n) + g2(n) + … + gm(n)) if and only if f(n) is (g1(n)).
294 CHAPTER NINETEEN: COST MODELS

Interestingly, we still select the maximal term to keep (defined using O), even in the
and versions of this theorem. You'll see how this works out if you do Exercise 7.
Another property that is very useful for analyzing algorithms is this:

Theorem 19.1.3, O version: If f1(n) is O(g1(n)), and f2(n) is O(g2(n)), then

f1(n) f2(n) is O(g1(n) g2(n)).
Theorem 19.1.3, version: If f1(n) is (g1(n)), and f2(n) is (g2(n)), then
f1(n) f2(n) is (g1(n) g2(n)).
Theorem 19.1.3, version: If f1(n) is (g1(n)), and f2(n) is (g2(n)), then
f1(n) f2(n) is (g1(n) g2(n)).

The proofs of these are left as an exercise (Exercise 4). They come in handy for analyzing
loops; if you know that the time taken by each iteration of a loop is O(g1(n)), and you know
that the number of iterations is O(g2(n)), then you can conclude that the total amount of
time taken by the loop is O(g1(n) g2(n)).
One final property: the O and forms are symmetric:

Theorem 19.1.4: For any functions f and g over N, f (n) is O(g(n)) if and only if
g(n) is (f(n)).
Proof: Follows directly from the definitions of O and .

This symmetry is interesting but not often useful, because O and are typically used to
characterize complicated functions by relating them to simple functions. Reversing the
direction is not usually a good idea; it might be mathematically correct to characterize a
simple function by relating it to complicated function, but it's rarely helpful.

19.3 Common Asymptotic Functions

If you start with the functions typically encountered when analyzing the resource
requirements of computer programs and you apply the simplifications described above, you
end up with a kind of standard vocabulary of functions. These functions occur so often in
asymptotic analysis that they have special names. Here are some of the most common. (The
descriptions below refer to tight bounds, but the same terms are also used for upper and
lower bounds.)
Θ(1): constant
Here, g(n) = 1: a constant function that does not grow as n grows. Theorem 19.1.1 tells us
that (1) is equivalent to (k) for any constant k. But simpler is better, and the convention
is to write (1) instead of any other constant.
 19.3 COMMON ASYMPTOTIC FUNCTIONS 295
Θ(log n): logarithmic
In computer science we often use base-2 logarithms, log2 n; in calculus the natural log is more
common, loge n. But it turns out that for any two constant bases a and b, loga n and logb n
differ by only a constant factor: loga n = (loga b) logb n. Theorem 19.1.1 tells us that constant
factors can be ignored, so the base of the logarithm is not significant inside asymptotics. The
convention is to omit the base altogether and simply write (log n).
Θ(n): linear
Here, g(n) = n. As always, constant factors are ignored. You would never write (3n), because
the equivalent (n) is simpler.
Θ(n log n): loglinear
The product of n times log n arises quite often in the analysis of algorithms. It grows just
slightly faster than n, but not nearly as fast as n2.
Θ(nk): polynomial
Important special cases of polynomial include (n) (linear), (n2) (quadratic), and
(n3) (cubic). Theorem 19.1.2 allows us to drop all but the highest-order term from any
polynomial. You never write (n2 + 4n + 2), because the equivalent (n2) is simpler.
Θ(2n): exponential
It is common to refer to this simply as exponential, without mentioning the base. That’s
actually a bit sloppy, because (unlike the base of logarithms) the base of the exponential is
significant asymptotically: (3n) is not the same as (2n). We’ll see a less sloppy definition of
exponential time in the next chapter.
The common asymptotic functions above are listed in order of their rates of growth,
from the slowest growing to the fastest growing. We state without proof these useful facts
about them:

Theorem 19.2:
• For all positive real constants a and b, a is O((log n)b).
• For all positive real constants a and b, and for any logarithmic base, (log n)a is
O(nb).
• For all real constants a > 0 and c > 1, na is O(cn).
1
Note that the theorem covers cases with fractional exponents. For example, n=n and so, by
2

Theorem 19.2, log n is O( n ) .

296 CHAPTER NINETEEN: COST MODELS

19.4 A TM Cost Model

We define the time and space required by a Turing machine M on input x as follows:
time(M, x) = the number of moves made by M on the input x *
space(M, x) = the length of tape visited by M on the input x *
This cost model is obvious, and it’s hard to think of any significantly different alternatives.
One simplification is to remove specific strings from the picture and concentrate only on
the length of the input string. The worst-case cost is one way to do this:
worst-case-time(M, n) = maximum time(M, x) of all x with |x| = n
worst-case-space(M, n) = maximum space(M, x) of all x with |x| = n
Similarly, we might also define cost functions that characterize the best case for strings of
length n or functions that characterize an average over all strings of length n. Here, we will
focus on worst-case cost. It is a widely used measure in practical programming, because it
corresponds to the kind of performance guarantee we often want to make: an upper bound
on the amount of time or space that a program will require.
Example: Linear Time and Space
Consider this TM for L(a*b*c*), which was introduced in Chapter 16.
b/b,R

b/b,R
a/a,R

c/c,R
B/B,R c/c,R

B/B,R

c/c,R

B/B,R

It is easy to see how many moves this machine will make on any x L(M ), because it
always moves right and never gets stuck until after the first B at the end of the input. So
for x L(M), time(M, x) = |x| + 1 and space(M, x) = |x| + 2 (including as “visited” the final
position of the head, from which the final state has no transition). On the other hand, the
 19.4 A TM COST MODEL 297
machine stops early when x L(M), because it gets stuck before reaching the accepting state.
For example, it stops after one move on any input string beginning with ba, no matter how
long: time(M, bax) = 1. We conclude that the worst-case time and space costs are linear: both
worst-case-time(M, n) and worst-case-space(M, n) are (n).
Example: Quadratic Time, Linear Space
Consider this TM for {anbn}, which was introduced in Chapter 16.
a/a,R
b/b,R

B/B,R a/B,R
5 1 2

B/B,L
B/B,R

b/B,L
4 3

a/a,L
b/b,L

The machine contains two states with self-edges: 2 and 4. For each state, the self-edges
of that state all move the head in the same direction and all transition on a symbol that is
not B. The machine never writes a nonB over a B, so the number of nonB symbols on the
tape is never greater than the original length m. Therefore, when any state is entered, there
will be O(m) moves in the worst case before that state is exited. Aside from the self-edges,
the machine has only one cycle: the states 1, 2, 3, and 4. On each iteration of this cycle,
it overwrites one a and one b. It cannot iterate more than m/2 times before all the input
symbols have been overwritten, so the number of iterations is O(m) in the worst case. In the
worst case, then, the machine makes O(m) iterations of the cycle, each of which takes O(m)
time. Thus, worst-case-time(M, m) is O(m2). The machine visits the whole part of the tape on
which the input is presented, but not beyond that, so worst-case-space(M, m) is (m).
With a little more work it is possible to show that the time bound is tight: worst-case-
time(M, m) is (m2). It is even possible to derive an exact formula for worst-case-time,
without using asymptotics (see Exercise 8). But finding exact formulas for worst-case time
and space isn’t easy, even for this relatively simple TM. For larger TMs it becomes dauntingly
difficult, and for TMs that are described but not actually constructed it is impossible.
Fortunately, it is often possible to give asymptotic bounds on the time and space resources
298 CHAPTER NINETEEN: COST MODELS

required, based only on an informal description of the machine. For this, asymptotics are
more than just a convenience—they are indispensable.

19.5 A Java Cost Model

For higher-level languages there are no cost models as obvious as those for TMs. How
should we define time(d, x) and space(d, x) for a Java decision method d and input string x?
This is a surprisingly delicate question. We want to model the time and space costs of a
computation realistically, so that our conclusions will usefully predict experimental results on
real computers. At the same time, we want our models to be simple and robust, so that our
conclusions will not have to be discarded just because we’ve changed our experimental set-up,
for example by installing more memory, a new processor, or a new compiler. Unfortunately,
there is an irreducible conflict between these two goals, realism and robustness. The most
realistic cost model for time would predict runtime exactly, not in terms of an abstraction
like “moves” but in actual physical time, time in seconds. To do that, though, would
require an extremely complex definition for time(d, x), accounting for every detail about
the particular compiler, processor, memory, and operating system being modeled. At the
other extreme, time(d, x) = 1 would be a very robust cost model for time, but it would be
completely unrealistic, asserting that every computation takes the same amount of time!
Somewhere in between these two extremes is the cost model we’re looking for.
Our solution to this conundrum will be to use the asymptotic notation. It was developed
long before there were computer programs to analyze, but it turns out to be just the right
tool for this purpose. We’ll won’t try to give those exact, extremely complex definitions
of time(d, x) and space(d, x). Instead, we’ll state some simple assumptions about them
asymptotically, using . That way, we'll be able to reach robust conclusions about time and
space requirements, conclusions that will be correct for any implementation that meets our
asymptotic assumptions.
For the Java-like language we’ve been using, a key issue is the implementation of int
values. Starting in Chapter 17, we have assumed that int values are somehow implemented
without Java’s usual 32-bit limit. This is very important for implementing recognition
methods. For one thing, all interesting formal languages involve strings of unbounded length,
and to manipulate these we need to have int values that can index them. In addition, it is
convenient to be able to treat the int values themselves as strings being tested for language
membership—to be able to write a recognition method that takes an int input and defines
the language of even numbers, prime numbers, perfect numbers, or whatever.
So we’ll assume that int values are implemented like String values, as references
(pointers) to immutable blocks of memory that are allocated whenever a new int value
is constructed. Thus, passing an int as a parameter takes constant time, because only
the reference is passed. An int assignment, like a String assignment, takes constant
time, because only the reference is assigned. But modifying an int, as in x = x + 1,
 19.5 A JAVA COST MODEL 299
necessitates the construction of a new int and takes time proportional to the memory size
taken up by the int. (This is true even if only one bit of the int is modified; x = x | 1
would still have to make a complete copy of x.)
With that in mind, here is a basic asymptotic cost model for our Java-like language:
1. All values manipulated by a program take space proportional to the length of their
representation as Java constants. This means in particular that the space required by
an integer constant is proportional to the number of digits required to represent
it—so the space required to represent any n N is (log n).
2. The time required for any operator is (n), where n is the total size of the data read and
written. For example, the time required to compute a + 1 is taken to be (log a).1
3. The overhead of making a method call is (1) in time and space. We assume that
parameter passing, method dispatch, and returning a value take only constant time
and that the initial allocation of memory for any method takes only constant space.
4. Predefined methods have the following resource requirements:
• s.length() takes (1) time and space. It just returns (by reference) the
stored length of the string.
• s.charAt(i) takes (log i) time and (1) space. Clearly, it must take log i
time just to look at all of the index value i; we assume that it does not need
more than that to access the character at index i.
We’ll state asymptotic assumptions about other predefined methods as we need them.
This asymptotic model does not define time(d, x) and space(d, x) exactly. It only says that,
whatever their exact definitions, these cost functions satisfy the given assumptions about
asymptotic bounds. As we did with TMs, we’ll also consider the corresponding worst-case
cost functions: worst-case-time(d, n) is the maximum time(d, x) for any |x| = n, and likewise
for worst-case-space(d, n).
Example: Constant Time and Space
Let’s analyze some decision methods using this cost model. Here’s one for the language
{ax | x *}:

boolean ax(String p) {
return (p.length()!=0 && p.charAt(0)=='a');
}

Let n be the input size, in this case simply n = |p|. Using our basic cost-model assumptions,
we can break down the worst-case time cost like this:

1. This is not quite realistic for integer multiplication and division. Simple algorithms for multiplication and division
of n-bit operands require (n2) time. Advanced algorithms can improve on this, approaching (n) time. To keep
things simple here, we’ll use (n) for all operators, even including *, /, and %.
300 CHAPTER NINETEEN: COST MODELS

Time Cost
1. boolean ax(String p) { (1)
2. return ( (1)
3. p.length() (1)
4. != (1)
5. 0 (1)
6. && (1)
7. p.charAt(0) (1)
8. == (1)
9. 'a'); (1)
10. } +_________
(1)

It turns out that the time complexity does not depend on n; every operation here takes
constant time and is executed at most once. (In general, the != operator takes time
proportional to the size of the data it reads; but it is reasonable to assume that the !=
operator on line 4 does not need not read the whole number to tell whether it is zero.) To get
the total worst-case-time(ax, n), we add that column of asymptotics. Using Theorem 19.1.2,
this is simply a matter of selecting the fastest-growing asymptotic in the column. The result:
worst-case-time(ax, n) is (1).
Turning to the space analysis, we can again consider this line by line.

Space Cost
1. boolean ax(String p) { (1)
2. return ( (1)
3. p.length() (1)
4. != (1)
5. 0 (1)
6. && (1)
7. p.charAt(0) (1)
8. == (1)
9. 'a'); (1)
10. } +_________
(1)

The space taken by the string p is, of course, (n). But in our model of execution this
value is passed as a reference—it is not copied into new memory space. Similarly, the
space required for p.length() on line 3 is (log n). But in our model of execution
the length method just returns, by reference, the stored length of the string, so no new
memory space is used. We conclude that worst-case-space(ax, n) is also (1).
 19.5 A JAVA COST MODEL 301
Like ax, most methods manipulate a large (but constant) number of values that require
constant space each. These add up to (1) space, which is already given as the overhead of
calling the method in the first place. So we can simplify the analysis of space complexity by
ignoring all these constant-space needs. The analysis can focus instead on expressions that
construct new values that require nonconstant space (like String, array, or unbounded
int values) and on expressions that call methods that require nonconstant space. If there
are none, as in ax, then the space complexity is (1). In a similar way, the analysis of time
complexity can be simplified by ignoring constant-time operations repeated for any constant
number of iterations.
Example: Loglinear Time, Logarithmic Space
Here’s another example, a decision method for {anbn}:

boolean anbn(String x) {
if (x.length()%2 != 0) return false;
for (int i = 0; i < x.length()/2; i++)
if (x.charAt(i) != 'a') return false;
for (int i = x.length()/2; i < x.length(); i++)
if (x.charAt(i) != 'b') return false;
return true;
}

The previous example developed a tight bound on the worst-case complexity directly. This
time, it will be more instructive to develop upper and lower bounds separately. First, a lower
bound. Clearly, the worst case for this code is when x {anbn}, so that all the loops run to
completion and none of the early returns is taken. Assuming that x {anbn}, we can state a
lower bound for the time spent on each line:

Time Cost Iterations Total

1. boolean anbn(String x) {
2. if (x.length()%2 != 0) (log n) 1 (log n)
3. return false;
4. for (int i = 0;
5. i < x.length()/2; (log n) (n) (n log n)
6. i++) (1) (n) (n)
7. if (x.charAt(i) != 'a') (1) (n) (n)
8. return false;
9. for (int i = x.length()/2; (log n) 1 (log n)
10. i < x.length(); (log n) (n) (n log n)
11. i++) (log n) (n) (n log n)
302 CHAPTER NINETEEN: COST MODELS

12. if (x.charAt(i) != 'b') (log n) (n) (n log n)

13. return false;
14. return true;
15. } +_______
__
(n log n)

The “Total” column here shows the total time spent on each line. In this case, it is computed
for each line by simple multiplication: the minimum time for one iteration of the line
multiplied by the minimum number of iterations of that line. This technique is very useful,
but it does have its limitations. Consider, for example, lines 6 and 7. Because i can be as
low as 0 there, we can’t do better than (1) as a lower bound on the time spent for a single
iteration of those lines. (For the similar lines 11 and 12 we can claim the stronger (log n)
lower bound, because there we know that i n/2.) Of course, i doesn’t stay 0—and the
total time for lines 6 and 7 is more than just the minimum time for any iteration times
the number of iterations. With a more detailed analysis we could get a stronger total lower
bound for these lines. But in this case there is no need. Adding up the “Total” column is
simply a matter of choosing the fastest growing function in the column (as Theorem 19.1.2
showed), and that is (n log n). This turns out to be the strongest lower bound possible.
We know it is the strongest lower bound possible, because it matches the upper bound:

Time Cost Iterations Total

1. boolean anbn(String x) {
2. if (x.length()%2 != 0) O(log n) 1 O(log n)
3. return false;
4. for (int i = 0;
5. i < x.length()/2; O(log n) O(n) O(n log n)
6. i++) O(log n) O(n) O(n log n)
7. if (x.charAt(i) != 'a') O(log n) O(n) O(n log n)
8. return false;
9. for (int i = x.length()/2; O(log n) 1 O(log n)
10. i < x.length(); O(log n) O(n) O(n log n)
11. i++) O(log n) O(n) O(n log n)
12. if (x.charAt(i) != 'b') O(log n) O(n) O(n log n)
13. return false;
14. return true;
15. } +_______
__
O(n log n)
 19.5 A JAVA COST MODEL 303
The upper bounds for lines involving i take advantage of the fact that i n. The total upper
bound is O(n log n) time. This agrees with the lower bound, so we conclude that worst-case-
time(anbn, n) is (n log n).
This may come as a surprise. Clearly anbn(x) must, in the worst case, examine every
symbol of x, so we expect it to require (n) time. But where did that extra log factor come
from? In this case, our cost model makes it unavoidable; accessing the ith character of a
string takes time proportional to log i. If we introduced some way of iterating over the
characters in a String without actually indexing each one, we might eliminate that extra
log factor. (For example, we might model an implementation of strings as linked lists of
characters.) But as it stands, we’re stuck with it. It might be worse; don’t forget that for the
Turing machine in the previous section, deciding this language took (n2) time.
Turning to the space analysis, we can again examine this line by line. As already noted,
constant-space needs can be ignored, because they are subsumed by the constant-space
overhead of every method call. Also, the number of times each line is iterated is unimportant,
because the memory used by a statement on one iteration can be reused on the next. (If only
time could be reused the way space can!)

Space Cost
1. boolean anbn(String x) {
2. if (x.length()%2 != 0)
3. return false;
4. for (int i = 0;
5. i < x.length()/2; (log n)
6. i++) (log n)
7. if (x.charAt(i) != 'a')
8. return false;
9. for (int i = x.length()/2; (log n)
10. i < x.length();
11. i++) (log n)
12. if (x.charAt(i) != 'b')
13. return false;
14. return true;
15. } +_________
(log n)

The values constructed at lines 5 and 9 are the same: n/2. The largest value constructed at
line 6 is n/2, and the largest value constructed at line 11 is n. These all take (log n) space.
Thus the total space requirement, worst-case-space(anbn, n), is (log n).
304 CHAPTER NINETEEN: COST MODELS

Example: Quadratic Time, Linear Space

Consider this method:

String matchingAs(String s) {
String result = "";
for (int i = 0; i < s.length(); i++)
r += 'a';
return r;
}

A line-by-line analysis of lower bounds on the worst-case time gives

Time Cost Iterations Total

1. String matchingAs(String s) {
2. String result = "";
3. for (int i = 0;
4. i < s.length(); (log n) (n) (n log n)
5. i++) (1) (n) (n)
6. r += 'a'; (1) (n) (n)
7. return r;
8. } +_______
__
(n log n)

As in the previous example, this produces some rather weak lower bounds. In this case, the
weakness is significant. For upper bounds, we get

Time Cost Iterations Total

1. String matchingAs(String s) {
2. String result = "";
3. for (int i = 0;
4. i < s.length(); O(log n) O(n) O(n log n)
5. i++) O(log n) O(n) O(n log n)
6. r += 'a'; O(n) O(n) O(n2)
7. return r;
8. } +_______
__
O(n2)
 19.5 A JAVA COST MODEL 305
The upper bounds for lines involving i and r take advantage of the fact that i n and
|r| n. The total upper bound is O(n2) time. This does not match our lower bound of
(n log n).
It is usually desirable to refine an analysis until the upper and lower bounds meet. For
subtle algorithms this can be very difficult. Sometimes, it isn’t even clear whether to try
refining the lower bound upward, the upper bound downward, or both. But in this case there
is an obvious suspect: the weak analysis of the lower bound at line 6. The length of the string
r there is initially 0, so for a single iteration of line 6 we can’t state a stronger lower bound
than (1). But r doesn’t stay that short, and the total time for line 6 is more than just the
minimum time for any iteration times the number of iterations. Let t(k) be the time taken
for an iteration of line 6 when the length of the string contructed is k. Then the total worst-
case time at line 6, for an input of length n, is
n
g(n)= Σ t(k)
k =1

where t(k) is (k). Using the definition of , we know that for all sufficiently large n, there is
some c for which
n n n
g(n)=
k =1
Σ t(k) ≥ Σ
k =1
1 k=1
c c Σ k = 1c
k =1
n2+n
2

It follows that g(n), the total worst-case time at line 6, is (n2). That gives a total lower
bound of (n2). That matches the upper bound, so worst-case-space(matchingAs, n) is
(n2).
The worst-case space analysis is more straightforward:

Space Cost
1. String matchingAs(String s) {
2. String result = "";
3. for (int i = 0;
4. i < s.length();
5. i++) (log n)
6. r += 'a'; (n)
7. return r;
8. } +_________
(n)

The largest value constructed at line 6 has the same length as the input string s.
306 CHAPTER NINETEEN: COST MODELS

Example: Exponential Time, Linear Space

Consider this decision method for the set of prime numbers:

boolean isPrime(int x) {
if (x < 2) return false;
for (int i = 2; i < x; i++)
if (x % i == 0) return false;
return true;
}

The analysis must be carried out in terms of the input size n, which is the amount of space
required to represent the input integer x. The relation between n and x is determined by the
representation. If we’re using decimal, n is roughly log10 x; if we’re using binary, it’s roughly
log2 x, and so on. Let’s just say that for some constant base b, x is (bn). Then we have this
for an upper bound on the worst-case time:

Time Cost Iterations Total

1. boolean isPrime(int x) {
2. if (x < 2) return false; O(n) O(1) O(n)
3. for (int i = 2;
4. i < x; O(n) O(bn) O(nbn)
5. i++) O(n) O(bn) O(nbn)
6. if (x % i O(n) O(bn) O(nbn)
7. == O(1) O(bn) O(bn)
8. 0) O(1) O(bn) O(bn)
9. return false;
10. return true;
11. } +_______
__
O(nbn)

An analysis of the lower bounds, which we omit, gives a weaker result for line 5 but the same
total, (nbn). Thus, worst-case-time(isPrime, n) is (nbn): exponential time.
If we performed our analysis in terms of the value of x, instead of the length of the string
representing x, we would get a much less intimidating result: worst-case-time(isPrime, x) is
O(x log x). But in the study of formal language, we always treat complexity as a function of
the size of the input string, whether it’s a string of letters or, as in this case, a string of digits.
Turning to the space complexity, we have:
 EXERCISES 307
Space Cost
1. boolean isPrime(int x) {
2. if (x < 2) return false;
3. for (int i = 2;
4. i < x;
5. i++) (n)
6. if (x % i (n)
7. ==
8. 0)
9. return false;
10. return true;
11. } +_________
(n)

The largest value constructed at line 5 is x, and the largest remainder at line 6 is x/2 -1.
Thus the total space requirement, worst-case-space(isPrime, n), is (n).

Exercises
EXERCISE 1
True or false:
a. 2n is O(n)
b. 2n is (n)
c. 2n is (n)
d. 3n2 + 3n is O(n3)
e. 3n2 + 3n is (n3)
f. 3n2 + 3n is (n3)
g. n is O(2n)
h. n2 is (n2 + 2n + 1)
i. log n2 is (log n3)
j. n2 is O(n (log n)2)
k. 2n is O(22n)
l. 3n is (5n)
EXERCISE 2
Simplify each of the following as much as possible. For example,
O(3a2 + 2a) simplifies to O(a2). (Hint: A few are already as simple as possible.)
a. O(5)
b. O(log3 a)
308 CHAPTER NINETEEN: COST MODELS

c. O(log a + 1)
d. O(log a2)
e. O((log a)2)
f. (a + log a)
g. (log a + 1)
h. (max(a2, a))
i. (2n + n100)
j. (a2 + 2a)
k. (3n2 + 2n + log n + 1)
l. (n2n)
m. (3n + 2n)
EXERCISE 3
Using the definitions only, without using the simplification theorems proved in
Section 19.2, prove the following:
a. n2 + 5 is O(n2)
b. n3 + 2n2 is (n)
c. a2 + 2a + 12 is (a2)
EXERCISE 4
Give a detailed proof of all three cases for Theorem 19.1.3.
EXERCISE 5
Give a detailed proof of Theorem 19.1.4.
EXERCISE 6
Give a detailed proof of Theorem 19.1.1, for the and versions.
EXERCISE 7
Give a detailed proof of Theorem 19.1.2, for the and versions.
EXERCISE 8
These questions concern the TM M for the language L(M ) = {anbn}, discussed in
Section 19.4.
a. Derive an exact formula for time(M, anbn), and show that it is (n2).
b. Show that space(M, anbn) is (n).
c. Show that worst-case-space(M, m) is O(m), and conclude (using the result from
Part b) that it is (m).
d. Show that worst-case-time(M, m) is O(m2), and conclude (using the result from
Part a) that it is (m2). (Note that this m is the total length of the string; your proof
must encompass strings that are not in L(M ).)
 EXERCISES 309
EXERCISE 9
These questions concern the TM from Section 16.3 (for the language {anbncn}).
a. Show that worst-case-space(M, m) is O(m).
b. Show that worst-case-space(M, m) is (m), and conclude that it is (m).
c. Show that worst-case-time(M, m) is O(m2).
d. Show that worst-case-time(M, m) is (m2), and conclude that it is (m2).
EXERCISE 10
These questions concern the TM from Section 16.7 (for the language
{xcx | x {a, b}*}).
a. Show that worst-case-space(M, n) is O(n).
b. Show that worst-case-space(M, n) is (n), and conclude that it is (n).
c. Show that worst-case-time(M, n) is O(n2).
d. Show that worst-case-time(M, n) is (n2), and conclude that it is (n2).
EXERCISE 11
These questions concern the TM from Section 17.2 (the stacker machine). Assume that
all input strings are in the required form x#y, where x {0, 1}* and y {0, 1}*.
a. Show that space(M, x#y) is (|x#y|), and thus worst-case-space(M, n) = (n).
b. Derive a bound on worst-case-time(M, n).
c. Derive a bound on best-case-time(M, n).
EXERCISE 12
These questions concern the TM from Section 17.4 (the ith machine). Assume the input
string is in the required form.
a. Show that worst-case-time(M, n) is (n2n).
b. Show that worst-case-space(M, n) is (2n).
EXERCISE 13
Derive tight asymptotic bounds on the worst-case time and space used by this
contains method. Show enough detail to make a convincing case that your bounds
are correct.
boolean contains(String s, char c) {
for (int i = 0; i < s.length(); i++)
if (s.charAt(i)==c) return true;
return false;
}

EXERCISE 14
Derive tight asymptotic bounds on the worst-case time and space used by this decision
method for the language {xx R}. Show enough detail to make a convincing case that your
bounds are correct.
310 CHAPTER NINETEEN: COST MODELS

boolean xxR(String s) {
if (s.length() % 2 != 0) return false;
for (int i = 0; i < s.length() / 2; i++)
if (s.charAt(i) != s.charAt(s.length()-i-1))
return false;
return true;
}
EXERCISE 15
a. Let g(n) be a function with
n
g(n)= Σ t(k)
k =1

where t(k) is (log k). Show that g(n) is (n log n). (Hint: Use this asymptotic
version of Stirling’s approximation: log n! is (n log n).)
b. Using that result, derive a tight asymptotic bound on the worst-case runtime of
the following method. Show your derivation for the upper bound first, then your
derivation for the matching lower bound.
boolean containsC(String s) {
int i = s.length();
while (i > 0) {
i--;
if (s.charAt(i)=='c') return true;
}
return false;
}
EXERCISE 16
Derive an asymptotic tight bound on the worst-case space used by this fact method,
and then derive an asymptotic upper bound on the worst-case time. (Hint: Use the result
of Exercise 15, Part a.) Make sure your results are expressed in terms of the length of the
input, not in terms of its integer value. Show enough detail to make a convincing case
that your bounds are correct.
int fact(int p) {
int sofar = 1;
while (n > 0) sofar *= p--;
return sofar;
}
 20
CHA P TER

Deterministic
Complexity Classes
Have you ever thought about what you would do if you had all the
money in the world? It’s fun to imagine. It can also be instructive,
when you realize that some of the things you want, money can’t buy.
When we considered questions of computability in previous chapters,
we were asking the same kind of question: “What languages can be
decided, with unlimited time and space?” Of course, we never have
unlimited time and space, but it’s fun to think about. It can also be
instructive, when you realize that some of the languages you want
to decide can’t be decided, no matter how much time and space you
have to spend.
Now we turn to a different, more pragmatic kind of question: what
languages can be decided on a limited budget? We’ll use the worst-
case-time and worst-case-space measures developed in the previous
chapter to classify languages according to the resource budgets
required to decide them.

311
312 CHAPTER TWENTY: DETERMINISTIC COMPLEXITY CLASSES

20.1 Time Complexity Classes

Any function g(n) over N can serve as an asymptotic upper bound on runtime and so can be
used to define a class of formal languages:

TIME (g(n)) = {L | L is decided by some decision method p, where

worst-case-time(p, n) is O(g(n))}.

The resulting class of languages is called a complexity class: languages classified according to
the resources required to decide them.
The previous chapter analyzed the worst-case time requirements of individual methods.
Now, we are considering the time requirements of whole languages—of all possible
decision methods for a given language. To prove that a language L is in TIME (g(n)) can
be straightforward. We need only give an example of a decision method p, show that it
decides L, and show (using the techniques of the previous chapter) that its worst-case time is
O(g(n)). For example, in the previous chapter we analyzed a decision method anbn for the
language {anbn}. We showed that worst-case-time(anbn, n) is O(n log n). That proves that the
language can be decided in O(n log n) time:
{anbn} TIME (n log n).
Since the language can be decided within that time budget, it can certainly also be
decided within any more generous budget, so we can also conclude that {anbn} TIME (n2),
{anbn} TIME (n3), and so on. In general, if f (n) is O(g(n)), then TIME (f(n)) TIME (g(n)).
On the other hand, to prove that a language L is not in TIME (g(n)) is usually much more
difficult. It isn’t enough to say that, although you tried really hard, you just couldn’t think
of an O(g(n)) decision method. You have to prove that no such method can exist. Here’s an
example:

Theorem 20.1: {anbn} TIME (log n).

Proof: Suppose by way of contradiction that {anbn} has some decision method
p for which worst-case-time(p, n) is O(log n). This is not enough time even to
access every symbol in the input string, which would take (n) time. So it must
be true that, for some some n0, p leaves at least one symbol uninspected in every
input string x with |x| n0. In particular, there must be at least one symbol
in the string an0 bn0 that p does not examine. Because it never examines that
symbol, p must execute identically on a mutated version of an0 bn0 with the
unexamined symbol flipped (from a to b or from b to a). Thus p must accept
both an0 bn0 and the mutated version of an0 bn0 . But the mutated version of
an0 bn0 is not in {anbn}. So L(p) {anbn}, contradicting our assumption. By
contradiction, no such decision method p exists, and {anbn} TIME (log n).
 20.1 TIME COMPLEXITY CLASSES 313
We have now established the existence of two distinct complexity classes. The class
TIME (logn) is not empty; we’ve seen that it includes the language {ax | x *}. TIME (n log n)
is not the same as TIME (log n); we’ve seen that it includes the language {a b }, which is not in
n n

the smaller class. So we have

{ax | x ∈∑*}

TIME (log n)

TIME (n log n)

{anbn}

It is not surprising to observe that the more time you have for the decision, the more
languages you can decide.
A time-hierarchy theorem states this observation more rigorously. We know that if
f (n) is O(g(n)), then TIME(f(n)) TIME (g(n)). A time-hierarchy theorem establishes some
additional conditions on f and g that allow you to conclude that the inclusion is strict—that
TIME (f (n)) TIME (g(n)). It’s a very useful kind of theorem: it often allows you to conclude
that two complexity classes are distinct, without the need to find and prove special theorems
like Theorem 20.1, above.
In fact, our time-hierarchy theorem automatically establishes an infinite hierarchy
of complexity classes. For example, it can be used to show that if a < b then
TIME (na) TIME (nb), producing an infinite hierarchy of polynomial complexity classes:

TIME (1)
TIME (n)
TIME (n2)
TIME (n3)
...

This polynomial time hierarchy is just one simple consequence of our time-hierarchy
theorem. Roughly speaking, our time-hierarchy theorem says that anything more than an
extra log factor is enough to decide additional languages. So there is actually a far richer
314 CHAPTER TWENTY: DETERMINISTIC COMPLEXITY CLASSES

structure of time-complexity classes than those pictured above. For example, there is an
infinite hierarchy of distinct classes like TIME (n log2 n) and TIME (n log4 n) that all lie inside the
diagram above, between TIME (n) and TIME (n2).
Unfortunately, this rich and powerful theorem is rather difficult to state and prove in
detail. Interested readers should consult Appendix B.

20.2 Polynomial Time

We have seen the notion of Turing equivalence: many different models of computation, from
lambda calculus to modern programming languages, turn out to be interconvertible with
Turing machines. The concepts of computability are robust with respect to such conversions.
The class of recursive languages and the class of RE languages do not change when the
underlying computational model is changed. By contrast, some of the concepts of complexity
are more fragile.
For example, TIME (f(n)) defined using Java is not generally the same as TIME (f (n))
defined using TMs. It isn’t hard to see why. Some operations (like s.charAt(i)) are
cheaper under our cost model for Java, and other operations (like sequential access to all the
symbols of the input string) are cheaper for TMs. The good news is that conversions among
the standard models of computation produce only a polynomial change in runtime. For
example, a TM can simulate a digital processor in such a way that the number of TM moves
is some polynomial function of the number of processor clock cycles. So any language that
can be decided in polynomial time, measured in processor clock cycles, can also be decided
in polynomial time measured in TM moves—though not, generally, the same polynomial.
If you lump all the polynomial-time complexity classes together, you get a combined class
that is robust with respect to the underlying model of computation. Polynomial time on a
Turing machine is the same as polynomial time in Java. This robust class of polynomial-time-
decidable languages is called P :

= TIME (nk )
⊃

P
k

The robustness of P allows decision problems to be described more naturally and

informally. Usually, we omit any mention of the model of computation, with the
understanding that we are looking for an algorithm that could be implemented for any of the
standard Turing-equivalent models. Likewise, the exact encoding of the decision algorithm’s
input can often be omitted. For example, we used a lot of ink in Chapter 16 showing exactly
how to encode a DFA as a string over the alphabet {0, 1}, then sketching how to make a TM
to simulate that DFA. Sometimes that level of detail is necessary, but if we’re only interested
in knowing whether the language is in P or not it may be preferable to describe the decision
problem more informally, like this:
 20.2 POLYNOMIAL TIME 315

DFA-acceptance
Instance: a DFA M and a string x.
Question: does M accept x?

To show that DFA-acceptance can be solved in polynomial time, it suffices to sketch an

algorithm, using any reasonable model of computation:

boolean dfaAcceptance(DFA M, String x) {

State s = M.startState();
for (int i = 0; i < x.length(); i++)
s = M.delta(s,x.charAt(i));
return s.isAccepting();
}

This is too high-level to satisfy our definition of a decision method; in particular, it

takes something other than a string as input, and it suppresses the details of how a DFA
could be represented as a string. Justifiably so: any competent programmer can imagine
implementations for the classes DFA and State, implementations in which every method
runs in polynomial time. Given such an implementation, the whole decision method clearly
takes polynomial time. In fact, it would probably have been enough merely to describe the
algorithm in English: “Simulate M on x, making one move for each symbol, and return true
if and only if M ends in an accepting state.” With this algorithm in mind, we conclude that
DFA-acceptance is in P—an informal way of saying that, for any reasonable string encoding
and model of computation,
{x | x encodes a DFA M and a string x with x L(M)} P.

Informal descriptions are more readable because they suppress details. They are also more
susceptible to error, for the same reason. When you read or write these descriptions, you
should be vigilant, carefully thinking about how you might represent each problem instance
as a string. For example, consider this decision problem:

Primality
Instance: a number x N.
Question: is x prime?

How is the natural number x represented? The informal description doesn’t say, but it is
conventional to assume that natural numbers are represented compactly, using a string of
numerals with a positional notation like binary or decimal. That’s like our cost model for
Java. In fact, we’ve already seen a Java decision method for this problem:
316 CHAPTER TWENTY: DETERMINISTIC COMPLEXITY CLASSES

boolean isPrime(int x) {
if (x < 2) return false;
for (int i = 2; i < x; i++)
if (x % i == 0) return false;
return true;
}

As we showed in Section 19.5, this takes exponential time— (nbn), if integers are
represented in base b and n is the length of the string of base b numerals used to represent x.
That is not polynomial time; when b > 1, bn is not O(nk) for any fixed k. So this decision
method does not show that Primality is in P. (It turns out that it is in P, but proving it
requires a much more subtle algorithm. See the Further Reading at the end of this chapter.)
P includes some of the major classes we’ve already studied:

Theorem 20.2: {L | L is context free} P.

Proof: Polynomial-time parsing algorithms for CFGs, like the CKY algorithm
(see Section 15.3), are well known. But not all languages in P are context free;
for example, it is easy to decide {anbncn} in polynomial time.

Because P includes all context-free languages, we know it includes all regular languages.
Because it includes all regular languages, we know it includes all finite languages—a fact we’ll
make use of in the next section.
People sometimes characterize the complexity class P as the set of tractable decision
problems. That’s probably an overstatement, because almost all the algorithms used in real
computer systems have not just polynomial complexity, but very low polynomial complexity.
For most applications that work with large data-sets, useful algorithms can’t take much more
than linear time, and it’s hard to think of practical examples in which problems requiring
even, say, O(n4) time are considered tractable. (For that matter, it’s hard to think of any
commercial setting in which the constant factor hidden by the big-O notation is actually
immaterial. If you slow down a program by a factor of ten, customers don’t say, “Who cares?
It’s only a constant factor!”) P certainly includes all the tractable decision problems, but also
includes many whose decision computations are, by most everyday measures, prohibitively
expensive.
But if in practice the word tractable isn’t a perfect description for the inside of P, the word
intractable certainly seems like a good description for the outside of P. That’s another reason
for the importance of this class. Languages that are recursive, yet beyond the polynomial-
time boundary, have an interesting intermediate status. The time required to decide such
a language grows more rapidly than n4, more rapidly than n4000, more rapidly than nk for
any fixed value of k. These are languages that can be decided in theory but not in practice;
 20.3 EXPONENTIAL TIME 317
languages whose computation we can specify but cannot afford to carry out, except on the
very shortest input strings.

20.3 Exponential Time

The time-hierarchy theorem guarantees the existence of languages beyond the border
of P. For example, it can be used to prove that P TIME (2n). This class TIME (2n), like the
individual, polynomial-time complexity classes, is not robust with respect to changes in the
underlying computational model; what takes O(2n) time with one model might take O( 2n2 )
time with another. To make a robust, exponential-time complexity class, we define

k
= TIME (2n )
⊃
EXPTIME
k

This is what it means to say that a language requires exponential time to decide: the
decision algorithm takes time proportional to some constant base raised to a power that is a
polynomial function of the length of the input string.
The time-hierarchy theorem demonstrates that two classes are distinct by constructing
a language that is provable in one but not in the other. In this way, it shows that there is
at least one language in EXPTIME that is not in P. But the language it constructs is highly
artificial—it’s provably hard to decide, but it isn’t clear why you would ever want to decide
it in the first place! Luckily, we don’t have to rely on the time-hierarchy theorem to illustrate
EXPTIME. In fact, we’ve already seen one example of a language that is in EXPTIME but not in P.
It’s the language decided by our runWithTimeLimit method. Recall that we defined
runWithTimeLimit(p,in,j) so that it returns true if and only if the recognition
method p returns true for the input string in within j steps. As an informal decision
problem, this is

Run-with-time-limit
Instance: a recognition method p, a string in, and a natural number j.
Question: does p accept in within j moves?

This decision problem can be solved in exponential time, as shown in the following theorem:

Theorem 20.3.1: Run-with-time-limit is in EXPTIME.

Proof: Our decision method runWithTimeLimit is described
in more detail in Appendix B, which shows that evaluating
runWithTimeLimit(p,in,j) takes (j log j) time. This is log-
linear in j—but of course the magnitude of j is exponential in the size of the
problem instance. Restating this in terms of the size n of the problem instance,
318 CHAPTER TWENTY: DETERMINISTIC COMPLEXITY CLASSES

runWithTimeLimit takes O(nbn) time. This is still within EXPTIME,

however, so we conclude that Run-with-time-limit is in EXPTIME.

We can also prove that the decision problem cannot be solved using polynomial time:

Theorem 20.3.2: Run-with-time-limit is not in P.

Proof: Suppose by way of contradiction that Run-with-time-limit is in P. Then
there is some decision method rWTL that does what runWithTimeLimit
does, but in polynomial time. We will show that rWTL could be used to decide
any EXPTIME language in polynomial time.
Let L be any language in EXPTIME. By definition there exists some decision
method string Ldec that decides L, with some exponent k, so that Ldec’s
worst-case time on inputs of length n is O( 2nk ). This means that there exist
some c and n0 for which, for all strings x with |x| n0, Ldec decides x within
b = c2nk steps. The set of strings x L with |x| < n0 is finite, and therefore in P.
Therefore, there exists a polynomial-time decision method LdecShort for
these strings.
Using these definitions for k, c, n0, and LdecShort, the following method
LdecPoly decides L:

1 boolean LdecPoly(String x) {
2 int n = x.length();
3 if (n < n0) return LdecShort(x);
4 int b = c * power(2,power(n,k));
5 return rWTL(Ldec,x,b);
6 }

Line 3, using LdecShort, takes polynomial time. The exponentiations at

line 4 can be done in polynomial time (this is left as an exercise). The rWTL
method called at line 5 takes polynomial time. Therefore LdecPoly takes
polynomial time, and it follows that L P. Because L is an arbitrary language in
EXPTIME, we must conclude that EXPTIME P. But this contradicts the hierarchy
theorem. By contradiction, Run-with-time-limit is not in P.

This proof uses the technique of reduction, which we first encountered in Chapter 18.
In that chapter it didn’t matter how efficient the reduction was, as long as it was computable.
The proof above is different; it depends critically on the fact that the reduction itself requires
only polynomial time. Lines 2 through 4 accomplish the reduction; you can think of them as
 20.3 EXPONENTIAL TIME 319
defining a function f that converts any instance of L into an equivalent instance of Run-with-
time-limit. Such reductions are very useful for proving things about complexity classes; they
even have their own special notation.

A P B (read as “A is polynomial-time reducible to B”) if there exists some

polynomial-time computable function f such that for all strings x over the
alphabet of A, x A if and only if f(x) B.

Our proof of Theorem 20.3.2 shows that for all languages L EXPTIME, L P Run-with-
time-limit. In that sense, Run-with-time-limit is at least as hard as any EXPTIME problem. There
is a term for languages that have this property:

A is EXPTIME-hard if, for all L EXPTIME, L P

A.

As the proof of Theorem 20.3.2 shows, Run-with-time-limit is EXPTIME-hard. If any EXPTIME-

hard problem were found to be in P, we could use the polynomial-time reduction to show
that all EXPTIME languages are in P. That would contradict the hierarchy theorem, which
guarantees that not all EXPTIME problems are in P. Conclusion: If you know that a language is
EXPTIME-hard, you can conclude that it is not in P.
Combining the results of Theorems 20.3.1 and 20.3.2, we see that Run-with-time-limit
is in EXPTIME and is at least as hard as any other problem in EXPTIME. There is a term for
languages that have both of these two properties:

A is EXPTIME-complete if A is in EXPTIME and A is EXPTIME-hard.

EXPTIME-completeproblems are, in a sense, the hardest problems in EXPTIME. The proofs of

Theorems 20.3.1 and 20.3.2 together show that Run-with-time-limit is EXPTIME-complete.
Beyond EXPTIME lie other, still more expensive, complexity classes:
2EXPTIME = (b = c2 ) (doubly exponential time)
2n
k
⊃ ⊃

TIME
k

(b = c2 ) (triply exponential time)

k
2n
2
3EXPTIME = TIME
k
and so on. The time-hierarchy theorem guarantees that each of these classes is occupied by
languages that cannot be decided in less time. There’s even a union of all the classes in this
exponential-time hierarchy:
⊃

ELEMENTARY TIME = nEXPTIME

n
We’ll look at some examples of languages in these super-high-complexity classes in the
last section of this chapter.
320 CHAPTER TWENTY: DETERMINISTIC COMPLEXITY CLASSES

20.4 Space-Complexity Classes

Up to this point we have dealt with time complexity; the whole topic of complexity
classes can be revisited with a focus on space complexity instead. Any g(n) can serve as an
asymptotic upper bound on space, and so can define another class of formal languages:

SPACE (g(n))
= {L | L is decided by some decision method p,
where worst-case-space(p, n) is O(g(n))}.

Basic results for space complexity parallel those for time complexity, so we’ll just summarize
them here.
• There is a separate hierarchy theorem for space. From basic definitions we know that if
f(n) is O(g(n)), then SPACE (f (n)) SPACE (g(n)). The space hierarchy theorem establishes
some additional conditions on f and g that allow you to conclude that the inclusion is
strict—that SPACE (f(n)) SPACE (g(n)).
• The space-hierarchy theorem is strong enough to show, for example, that if a < b then
SPACE (na) SPACE (nb), producing an infinite hierarchy of polynomial-complexity classes
for space.
• Corresponding to the time-complexity class P (polynomial time) there is a space-
complexity class PSPACE (polynomial space) that is robust with respect to changes in the
underlying computational model:
SPACE(nk)
⊃

PSPACE =
k
• Corresponding to the time-complexity class EXPTIME (exponential time) there is a space-
complexity class EXPSPACE (exponential space) that provably contains problems not
solvable in PSPACE:
k
= SPACE(2n )
⊃

EXPSPACE
k

• Corresponding to the time-complexity classes EXPTIME-hard and EXPTIME-complete are

the space-complexity classes EXPSPACE-hard and EXPSPACE-complete. A language A is
EXPSPACE-hard if for all L EXPSPACE, L P A. (A polynomial-time reduction is still used.)
A language is EXPSPACE-complete if it is both EXPSPACE-hard and an element of EXPSPACE.
• Corresponding to the exponential-time hierarchy, there is an exponential-space hierarchy,
in which each class provably contains languages not decidable with less space: 2EXPSPACE,
3EXPSPACE, and so on.
The space-complexity classes are related to the time-complexity classes we’ve already
seen. Some of these relations are clear:
 20.4 SPACE-COMPLEXITY CLASSES 321
Theorem 20.4.1: For any g(n), TIME(g(n)) SPACE (g(n)).

Proof: According to our cost model, simply creating values taking (g(n)) space
must take (g(n)) time.

This holds, not just for our Java decision methods, but for any reasonable cost model and
for a variety of models of computation. For example, simply visiting g(n) cells of a Turing
machine’s tape requires making at least g(n) moves.
As a direct consequence of Theorem 20.4.1, we have

Theorem 20.4.2: P PSPACE and EXPTIME EXPSPACE.

To find inclusions in the other direction—a space-complexity class included in a time-

complexity class—we have to reach up to a higher exponential:

Theorem 20.4.3: For any g(n), there exists some constant k such that
SPACE (g(n))TIME (2k⋅g(n)).

Proof sketch: This proof is easier to do using the TM model. The basic idea is
to show that a TM that operates within an m-cell region of its tape cells cannot
have more than 2km distinct IDs, for some constant k. So a modified version
of the machine, limited to 2km steps, must produce the same decision as the
original.

Again, this is true for any reasonable cost model and for a variety of models of
computation.
As a direct consequence of Theorem 20.4.3, we have

Theorem 20.4.4: PSPACE EXPTIME and EXPSPACE 2EXPTIME.

Some of the relations among time- and space-complexity classes are fascinating puzzles—
puzzles that have withstood many years of research effort by some of the most brilliant
computer scientists on the planet. For example, most researchers believe that
P PSPACE and EXPTIME EXPSPACE, but no one has been able to prove either.

20.5 Higher-Complexity Classes

This section offers a guided tour of some of the higher-complexity classes. We state these
results without proof, providing instead some pointers into the literature for those who are
interested.
322 CHAPTER TWENTY: DETERMINISTIC COMPLEXITY CLASSES

PSPACE-complete Problems
These are the hardest problems in PSPACE. They seem to be intractable, requiring exponential
time, but no one has been able to prove this. A classic, PSPACE-complete problem is QBF: the
Quantified Boolean Formulas problem.

QBF: Quantified Boolean Formulas

Instance: a formula that uses only Boolean variables, the logical
connectives , , and , and the quantifiers and .
Question: is the formula true?

For example, the QBF instance x y (x y) is true: for all Boolean values that can be
assigned to x, there exists an assignment for y that makes x y true. That’s a positive instance,
an element of the QBF language. A negative instance is x y (x y); it’s a well-formed
instance, but it is not true that for all x there exists a y that makes x y true.
Imagine yourself facing QBF as a programming assignment: write a program that lets the
user type in a QBF formula, evaluates it, and tells whether it is true or false. There’s a simple
kind of PSPACE solution to this problem: you just evaluate the input formula using all possible
combinations of truth assignments for the variables and collect the results. (To work out the
details, try Exercise 8.) But because there are exponentially many different combinations of
truth assignments for the variables, this approach takes exponential time. It’s a simple, brute-
force solution, yet no solution with better asymptotic worst-case time complexity is known.
A more familiar PSPACE-complete program (at least to the readers of this book) is the
problem of deciding whether two regular expressions are equivalent.

Regular-expression-equivalence
Instance: two regular expressions r1 and r2, using only the three basic
operations: concatenation, +, and Kleene star.
Question: is L(r1) = L(r2)?

Regular-expression-equivalence is PSPACE-complete, as is NFA-equivalence:

NFA-equivalence
Instance: two NFAs M1 and M2.
Question: is L(M1) = L(M2)?

The same result does not hold for DFAs. Two DFAs can be converted to minimum-state
DFAs and compared for equivalence in polynomial time.
 20.4 SPACE-COMPLEXITY CLASSES 323
EXPTIME-complete Problems
We have already shown that Run-with-time-limit is EXPTIME-complete. That might seem like
a rather artificial problem, but many more natural EXPTIME-complete problems have been
found. An example is the game of checkers:

Generalized-checkers
Instance: an n-by-n checkerboard with red and black pieces.
Question: can red guarantee a win from this position?

Notice that the game is generalized to boards of unbounded size. Of course, even for n = 8
(as in American checkers) the number of problem instances—the number of possible
configurations of pieces on the 8-by-8 checkerboard—is large enough to make the game
challenging for humans. The language of winning positions is too large for human players
to memorize. But it is still a finite number, and that makes the fixed-size game uninteresting
from the perspective of formal language; it’s just a finite (though very large) language,
decidable by a simple (though very large) DFA. Generalized to n-by-n checkerboards, the
language becomes infinite, and deciding it turns out to be intractable. The exact rules of
checkers (or draughts) vary around the world, but under one reasonable definition of the
rules, Generalized-checkers has been shown to be EXPTIME-complete. Generalized versions of
chess and go have also been shown to be EXPTIME-complete.
EXPSPACE-complete Problems
In Section 9.4, several extensions to regular expressions were discussed. One was squaring,
defining L((r)2) = L(rr). The question of whether two regular expressions are equivalent,
when squaring is allowed, becomes more difficult to answer. You can of course convert all
the squaring into explicit duplication; then you’ve got two basic regular expressions that can
be compared as already discussed. But that first step, eliminating the squaring, can blow
up the expression exponentially—not a promising first step, expecially considering that the
rest of the algorithm needs space that is polynomial in that new, blown-up length! In fact,
the regular-expression-equivalence problem is known to be EXPSPACE-complete when the
expressions can use these four operations: concatenation, +, Kleene star, and squaring.
Instead of squaring, regular expressions can be extended with an intersection operator
&, so that L(r1 & r2) = L(r1) L(r2). This again makes the equivalence problem EXPSPACE-
complete. Note that neither of these extensions changes the set of languages you can define
using regular expressions; it remains the set of regular languages. The only change is that,
using the extensions, you can define some regular languages much more compactly.
324 CHAPTER TWENTY: DETERMINISTIC COMPLEXITY CLASSES

A Problem Requiring 2EXPTIME

In Section 18.13 we saw that number theory—the theory of natural numbers with
multiplication and addition—is undecidable. A similar, but simpler, theory arises if you make
just one small change: leave out multiplication. This theory is called Presburger arithmetic.

Presburger-arithmetic
Instance: a formula that uses only the constants 0 and 1, variables over
the natural numbers, the + operator, the = comparison, the logical
connectives , , and , and the quantifiers and .
Question: is the formula true?

Mojzesz Presburger explored this theory in the late 1920s. He gave a set of axioms for this
arithmetic and proved that they are consistent (not leading to any contradictions) and
complete (capable of proving all true theorems). He proved that this arithmetic is decidable:
an algorithm exists to decide it, so the language is recursive. But it is known that no
algorithm that decides Presburger-arithmetic can take less than doubly exponential time. So
this is an example of a language that is in 2EXPTIME, but not in EXPTIME.
A Problem Requiring Nonelementary Time
Beyond 2EXPTIME lies an infinite hierarchy of higher exponentials: 3EXPTIME, 4EXPTIME, and
so on, each one occupied by problems that cannot be solved in less time. The union of all
the classes in this exponential-time hierarchy is the class called ELEMENTARY TIME : the class of
languages decidable in k-fold exponential time, for any k. That’s an impressively generous
time budget—but it still isn’t enough to handle all the recursive languages. There are
languages that are recursive, but provably not in ELEMENTARY TIME.
A simple example is the regular-expression-equivalence problem, when regular
expressions are extended with a complement operator C, so that L((r)C) = {x * | x L(r)}.
We know that the regular languages are closed for complement, so this extension does not
add the ability to define new languages. But, once again, we get the ability to define some
languages much more compactly; and, once again, that makes the problem of deciding the
equivalence of two regular expressions harder, in this case, much harder. When complement is
allowed, the problem of deciding the equivalence of two regular expressions provably requires
nonelementary time.

20.6 Further Reading

The question of whether Primality is in P was an open problem for many years. Practical
algorithms for primality testing were known, and widely used in cryptography, but they were
probabilistic; they involved the use of random numbers and with very small probability took
a very long time. No deterministic polynomial-time algorithm was known until
 EXERCISES 325
Agrawal, M., N. Kayal, and N. Saxena. “PRIMES in P.” Annals of Mathematics
160, no. 2 (2004) 781–93.
On QBF, regular-expression equivalence, and NFA equivalence, see
Garey, Michael, and David Johnson. Computers and Intractability: A Guide to the
Theory of NP-Completeness. New York: W. H. Freeman and Company, 1979.
The original results on the complexity of generalized board games can be found in
Robson, J. M. “N by N Checkers is EXPTIME complete.” SIAM Journal on
Computing 13, no. 2 (1984): 252–67.
Robson, J. M. “The Complexity of Go.” Proc. International Federation of
Information Processing (1983): 13–417.
Fraenkel, S., and D. Lichtenstein. “Computing a perfect strategy for n*n chess
requires time exponential in n.” Proceedings of the 8th International Colloquium
on Automata, Languages, and Programming (1981): 278–93.
The equivalence problem for regular expressions with exponentiation is discussed in
Sipser, Michael. Introduction to the Theory of Computation. Belmont, CA: PWS
Publishing, 1997, 313–17.
For regular expressions with intersection, see
Fuerer, Martin. “The Complexity of the Inequivalence Problem for Regular
Expressions with Intersection.” ICALP (1980): 234-45.
For regular expressions with complement, see
Meyer, A. R., and L. Stockmeyer. “Nonelementary word problems in automata
and logic.” Proceedings of the AMS Symposium on Complexity of Computation.
(1973).

Exercises
EXERCISE 1
a. For fixed k, is O(nk) the same as O(nk log n)? Prove your answer.

b. Is P the same as TIME(nk log n) ? Prove your answer.

⊃

EXERCISE 2

a. For fixed k, is O(2nk ) the same as O(3nk ) ? Prove your answer.

k
TIME(3n
⊃

b. Is EXPTIME the same as ) ? Prove your answer.

k
326 CHAPTER TWENTY: DETERMINISTIC COMPLEXITY CLASSES

EXERCISE 3
Suppose that an incredibly fast computer began executing a decision method at the
time of the big bang. Assume that the big bang was 15 billion years ago, and that the
computer has executed one step per yoctosecond ever since—that is, 1024 steps per
second. What is the largest input of the size n for which it could have reached a decision
by now, assuming that the decision required
a. n steps
b. n2 steps
c. n13 steps
d. 2n steps
e. 10n steps
f. 22n nsteps
2
g. 22 steps
EXERCISE 4
(This exercise refers to concepts discussed with the time-hierarchy theorem in
Appendix B.) Give an implementation and analysis to prove that each of the following
functions over N is time constructible.
a. n3
b. 2n
c. n log2 n (Implement this as n log2n and assume n > 0.)
d. 22n / log2 n (Implement this as 22n , and assume n > 0.)
n log2n
EXERCISE 5
(Some parts require the time-hierarchy theorem from Appendix B. You may assume that
all the functions you need are time constructible.)
a. Is TIME (n3) the same as TIME (n3 (log n)3)? Prove your answer.
b. Is TIME (n3) the same as TIME (3n3)? Prove your answer.
c. Is TIME (2n) the same as TIME (2n+2)? Prove your answer.
d. Is TIME (2n) the same as TIME (n2n)? Prove your answer.
e. Is TIME (n log n) the same as TIME (n n )? Prove your answer.
EXERCISE 6
In the proof of Theorem 20.3.1, the decision method LdecPoly has this statement:

int b = c * power(2,power(n,k));

Give an implementation of the power method, and using it show that the runtime for
this statement is polynomial in n.
 EXERCISES 327
EXERCISE 7
For each of the following QBF instances, decide whether it is true or false, and prove it.
a. ∀x(x)
b. ∃x(x)
c. ∀x(x x)
d. ∃x(x x)
e. ∀x∃y((x y) (x y))
f. ∃x∀y((x y) (x y))
g. ∃x∀y∃z(x y z)
h. ∃x∀y∃z((x y) (y z))
EXERCISE 8
In this exercise, you will write a Java application DecideQBF that takes a QBF instance
from the command line and prints true or false depending on whether the input
string is in the language of true, syntactically legal, quantified Boolean formulas.
Start with the syntax, the QBF classes, and the parser from Chapter 15, Exercise 8.
Then add an eval method to the QBF interface and to each of the QBF classes, so that
each formula knows how to decide whether it is true or not. (You will need to pass the
context of current bindings in some form as a parameter to the eval method.) The
eval method should throw an Error if it encounters scoping problems—that is, if
there is a use of a free variable, as in Ax(y), or a redefinition of a bound variable, as
in Ax(Ex(x)). Of course, the parser will also throw an Error if the string cannot
be parsed. All these Errors should be caught, and your DecideQBF should decide
false for such inputs. It should decide true only for inputs that are legal, true, quantified
Boolean formulas. For example, you should have

> java DecideQBF 'Ey(y)'

true
> java DecideQBF 'ExAy((x+y)*(~x+~y))'
false
> java DecideQBF 'x'
false -- free variable x
> java DecideQBF 'hello'
false -- syntax error

EXERCISE 9
The following decision problems involving basic regular expressions can all be solved
by algorithms that begin by converting the regular expression into an NFA, as in the
construction of Lemma 7.1.
328 CHAPTER TWENTY: DETERMINISTIC COMPLEXITY CLASSES

a. Give an algorithm and analysis demonstrating that Regular-expression-equivalence is

in EXPTIME.
b. Give an algorithm and analysis demonstrating that the following decision problem is
in P.

Regular-expression-emptiness
Instance: a regular expression r using only the three basic
operations: concatenation, +, and Kleene star.
Question: is L(r) = {}?

c. Give an algorithm and analysis demonstrating that the following

decision problem is in EXPTIME.

Regular-expression-universality
Instance: a regular expression r using only the three basic
operations: concatenation, +, and Kleene star.
Question: is L(r) = *?

d. Give an algorithm and analysis demonstrating that the following

decision problem is in P.

Regular-expression-finiteness
Instance: a regular expression r using only the three basic
operations: concatenation, +, and Kleene star.
Question: is L(r) finite?

EXERCISE 10
Consider the following decision problems:

Extended-regular-expression-equivalence
Instance: extended regular expressions r1 and r2 that may use
concatenation, +, Kleene star, and complement.
Question: is L(r1) = L(r2)?

Extended-regular-expression-emptiness
Instance: an extended regular expression r that may use
concatenation, +, Kleene star, and complement.
Question: is L(r) = {}?
 EXERCISES 329
This chapter stated (without proof ) that Extended-regular-expression-equivalence requires
nonelementary time. Now, prove that Extended-regular-expression-emptiness requires
nonelementary time if and only if Extended-regular-expression-equivalence requires
nonelementary time.
EXERCISE 11
State the following assertions about natural numbers as formulas in Presburger
arithmetic. Then state whether the assertion is true.
a. The successor of any number is not equal to that number.
b. There is a number that is greater than or equal to any other number.
c. Adding zero to any number gives you that number again.
d. No number is less than three.
e. Zero is not the successor of any number.
f. Between every number and its successor there lies another number.
g. For any x 0, 3x > 2.
EXERCISE 12
Using Java decision methods, show that P is closed for
a. complement
b. union
c. intersection
d. reversal
e. concatenation
f. Kleene star (This last is more difficult than the others. Try for a solution using
dynamic programming.)
EXERCISE 13
Using Java decision methods, show that PSPACE is closed for
a. complement
b. union
c. intersection
d. reversal
e. concatenation
f. Kleene star
EXERCISE 14
Prove or disprove each of the following assertions.
a. The P relation is transitive: if A P B and B P C then A P
C.
b. The P relation is reflexive: for all languages L, L P L.
c. The P relation is symmetric: if A P B then B P A.
d. If A P B and B P then A P.
e. If A P B and B P A then both A and B are in P.
f. If A is EXPTIME-hard and A P B then B is EXPTIME-hard.
330 CHAPTER TWENTY: DETERMINISTIC COMPLEXITY CLASSES

g. If A is PSPACE-complete and A P B then B is PSPACE-complete.

h. If L P Run-with-time-limit then L EXPTIME.
i. If L P Run-with-time-limit then L is EXPTIME-complete.
j. Generalized-chess 2EXPTIME.
EXERCISE 15
Prove that any language L requires nonelementary time if and only if it requires
nonelementary space.
 21
CHA P TER

Complexity Classes

In this chapter we focus on the class NP—a class defined using a

model of computation that, as far as we know, cannot possibly be
realized as a physical computer. This might seem, at first, like an
arbitrary construction—another puzzle divised by professors for the
torment of students. It does not seem like the kind of thing that
belongs in a book whose subtitle is “A Practical Introduction.”
Where’s the practical value in learning about NP? But we ask for your
indulgence for a few pages. NP turns out to have great commercial
significance. It is also at the heart of the most important open
problem in computer science—for the solution of which there is, on
standing offer, a reward of $1 million.
If that’s not practical, what is?

331
332 CHAPTER TWENTY-ONE: COMPLEXITY CLASSES

21.1 Verification Methods

In this chapter we consider nondeterministic complexity classes and in particular the class NP.
But we start with a purely deterministic computation called a verification method. (In
Section 21.5 we’ll see how nondeterminism fits into the picture.)
In form, a verification method is like a decision method, but with two string parameters.

A verification method takes two String parameters, the instance and the
certificate, and returns a boolean value.

We have already stretched our definition of a decision method to allow multiple string
parameters, on the understanding that if we wanted to use the strict definition, we could
rewrite any method that takes two strings (x,y) to make it take a single string x+c+y,
for some separator char c not occurring in x. So, in that sense, a verification method is no
different from a decision method, and we can continue to use all our old definitions.
The important difference between verification methods and decision methods lies in
how we use them to define languages. We use a verification method to define, not the set of
instance/certificate pairs, but the set of positive instances: the set of instances for which some
certificate exists.

The language defined by a verification method v(x,p) is

L(v) = {x | for some certificate p, v(x,p)}.

The instance string x is the string being tested for language membership; it is in the language
if there exists at least one string p that the verification method accepts as a certificate for x.
For example, this is a verification method for the language {zz | z *}:

boolean xxVerify(String x, String p) {

return x.equals(p+p);
}

In this example, the certificate is a string p such that x = pp. The instance string x is in
the language {zz | z *} if and only if such a string p exists. In this case, of course, the
certificate is telling us something we could easily have figured out anyway; a decision method
for the language is not that much more difficult to write. But in other cases, the information
in the certificate appears to make writing a verification method for a language much simpler
than writing a decision method. Stretching the definition to allow int inputs, consider this
verification method:
 21.2 NP, NP-HARD, AND NP-COMPLETE 333
boolean compositeVerify(int x, int p) {
return p > 1 && p < x && x % p == 0;
}

This is a verification method for the language of composite numbers; a composite number is
a positive integer with an integer factor other than one or itself. Here, verification seems to
be much easier than decision; checking a proposed factor of x seems to be much easier than
deciding from scratch whether such a factor exists.
As these examples show, a verification method by itself is not a computational procedure
for determining language membership. It does not tell you how to find the certificate; it
only checks it after you’ve found it. Of course, you could search for it blindly. Given any
verification method aVerify for a language A, you could make a recognition method for
A by enumerating the possible certificates p and trying them all. Suppose SigmaStar is an
enumerator class that enumerates *—then we could implement a recognition method this
way:

boolean aRecognize(String x) {
SigmaStar e = new SigmaStar();
while (true) {
String p = e.next();
if (aVerify(x,p)) return true;
}
}

But this is not a decision method; if there is no string p for which aVerify(x,p) returns
true, it simply runs forever. To make this into a decision method, we need some finite bound
on the set of certificates that must be checked.

21.2 NP, NP-Hard, and NP-Complete

A finite bound is part of the definition of the language class called NP. Roughly speaking,
NP is the set of languages that have polynomial-size certificates verified by polynomial-time
verifiers.

NP is the set of all languages L for which there exists a polynomial-time

verification method v and a constant k, so that
L = {x | for some certificate p with |p| |x|k, v(x,p) }

This is a restriction on our general definition of a verification method: not only must it run
in polynomial time, but if it accepts any certificate for a string x, it must accept at least
334 CHAPTER TWENTY-ONE: COMPLEXITY CLASSES

one whose size is polynomial in x. We’ll call this an NP verification method, to distinguish it
from the unrestricted kind. The time and space requirements of a verification method are
measured in terms of the input size n = |x| + |p|. But here, because |p| itself is polynomial
in |x|, we can say that the worst-case time and space of an NP verification method are simply
polynomial in n = |x|.
Let’s examine this class NP. Where does it fit among our other complexity classes? For one
thing, it clearly contains all of P.

Theorem 21.1.1: P NP.

Proof: Let A be any language in P. Then A has a polynomial-time decision

method aDec. We can use this to construct a polynomial-time verification
method for A:

boolean aVerify(String x, String p) {

return aDec(x);
}

This method ignores the certificate string and decides whether x A using
aDec. If x A, it will return false in polynomial time, no matter what
certificate is given. If x A, it will return true in polynomial time, for any
certificate p, and in particular for the empty string. So it meets our definition
for an NP verifier using k = 0.

With a slightly more elaborate construction, we can show that NP is contained in EXPTIME.

Theorem 21.1.2: NP EXPTIME.

Proof: Let A be any language in NP. Then A has an NP verification method

aVerify. Let k be the constant from the definition, so that if x A,
there is some certificate p with |p| |x|k for which aVerify(x,p). Let
SigmaStar be an enumerator class that enumerates * in order of length,
starting with ε. Then, using our definition for k, the following method
aDecide decides A:
boolean aDecide(String x) {
int b = power(x.length(),k);
SigmaStar e = new SigmaStar();
while (true) {
String p = e.next();
 21.2 NP, NP-HARD, AND NP-COMPLETE 335
if (p.length() > b) return false;
if (aVerify(x,p)) return true;
}
}

If there is a certificate for x, there must be some certificate p with |p| |x|k,
so this method will find it and return true. If not, this method will exhaust all
the certificates with |p| |x|k and return false. The number of iterations of the
loop—the number of calls to the polynomial-time verifier—in the worst case for
an input of size n is the number of possible certificates p with |p| nk, which is
k
|Σ|n . Thus, A EXPTIME.

In this proof, the verifier for A runs in polynomial time, so the language of instance/
certificate pairs is in P. But what about the language A itself—the language of instances for
which a certificate exists? It’s still possible that A is in P, using a decision method with a better
strategy than searching though all possible certificates. But our construction above shows
only that A is in EXPTIME.
We can make hardness and completeness definitions for NP as usual:

A is NP-hard if, for all L NP, L P A.

We have already seen quite a few NP-hard problems. For example, Generalized-checkers is
EXPTIME-hard and NP EXPTIME, so Generalized-checkers is NP-hard. But Generalized-checkers
is not known to be in NP. Of particular interest are those problems that are both NP-hard and
in NP.

A is NP-complete if A is in NP and A is NP-hard.

NP-complete problems are, in a sense, the hardest problems in NP. We have not yet seen any
examples of NP-complete problems, but we’ll see several in the following pages.
Putting Theorems 21.1.1 and 21.1.2 together gives us a rough idea of where NP fits in
among our other complexity classes: P NP EXPTIME. But that’s a very wide range; problems
in P are (at least nominally) tractable, while EXPTIME, as we have seen, provably contains
intractable problems. So what about those problems in NP ? Are they all tractable, or are the
harder ones (in particular, the NP-complete problems) intractable? That’s the heart of the
matter—and the most important open question in computer science today. Most researchers
believe that P NP, so that the NP-complete problems are intractable, requiring more than
polynomial time. But no one has been able to prove it. No one has been able either to find
polynomial-time algorithms for these problems or to prove that no such algorithms exist.
336 CHAPTER TWENTY-ONE: COMPLEXITY CLASSES

21.3 A Tour of Six Decision Problems

In this section we will explore six different NP-complete problems. These decision problems
form a chain, with each reduced in polynomial time to the next.
SAT
This is the classic NP-complete problem: the first problem proved to be NP-complete, and still
one of the most widely used and studied.

Satisfiability (SAT )
Instance: a formula that uses only Boolean variables, with operators
for logical AND, logical OR, and logical NOT.
Question: is there some assignment of values to the variables that
makes the formula true?

This is related to the QBF problem we’ve already seen, but simpler; a SAT instance is like
a QBF instance with no quantifiers. (In SAT, all the variables are implicitly existentially
quantified.) Mathematicians often use the operator symbols , , and and with precedence
in that order and parentheses as needed; for example:
a b true, under the assignment a=true and b=true
a a false, under all assignments
(a b) (a b) true, under the assignment a=true and b=false
In most programming languages we use symbols that are easier to type. In Java, we might
give the same examples as
a & b
a & !a
(a | b) & !(a & b)
But the concrete syntax isn’t important here. Both the mathematical notation and the Java
notation are just context-free languages, for which simple grammars can be given, and a
string that encodes a SAT instance in one of these languages can be parsed and checked
syntactically in polynomial time. We might say that the language of syntactically correct
SAT instances is in P. But that still leaves the satisfiability question: SAT is a subset of
the syntactically correct instances, containing those that have at least one satisfying truth
assignment. This subset is not context free, and it is not believed to be in P.
It is, however, in NP. We can show this by outlining a verification method for it. Here,
the certificate is a truth assignment; a verification method need only verify that it makes the
formula true.
 21.3 A TOUR OF SIX DECISION PROBLEMS 337
Lemma 21.1.1: SAT NP.

Proof sketch: We need to show that an NP verification method vSAT(x,p)

exists for Satisfiability. The instance string is x, of course, a Satisfiability formula,
whose syntax we can check in polynomial time. The certificate p can be just
a string over the alphabet {T, F}, encoding an assignment of truth values to
each of the variables in x, in some fixed order (such as in the order of their first
appearance in x). A formula of length n cannot contain more than n distinct
variables, so the size of the certificate is polynomial (in fact, linear) in the size
of the instance. All vSAT has to do is to evaluate the formula using the truth
assignment given by the certificate. This can easily be done in polynomial time.

The NP verification method for SAT only needs to check, in polynomial time, whether the
particular truth assignment given by the certificate makes the formula true. Checking a truth
assignment is easy; finding one seems much harder. There is a simple procedure, of course: a
blind search for a certificate, as in the proof of Theorem 21.1.2. But that takes O(2n) time in
the worst case. And although many researchers have sought a polynomial-time solution, none
has been found.
If there is a polynomial-time solution, then there is a polynomial-time solution for every
problem in NP, and so P = NP. The following theorem was proved by Stephen Cook in 1971.

Lemma 21.1.2 (Cook’s Theorem): SAT is NP-hard.

Proof sketch: For any L NP, there is a polynomial-time verification method,
which we’ll take to be in the form of a Turing machine M. For any such machine
we can give a polynomial-time reduction f (x) that generates a Boolean formula,
so that f(x) SAT if and only if x L. The Boolean formula, though large and
complicated, is only polynomially larger than x. A more detailed sketch of this
important proof may be found in Appendix C.

We have now proved that SAT is in NP and is NP-hard. Putting these results together, we
conclude

Theorem 21.2.1: SAT is NP-complete.

CNFSAT
Sometimes, the simplicity of Turing machines makes them more useful than decision
methods for proofs—as in the proof of Cook’s Theorem. Similarly, simplified versions of
SAT, as long as they are still NP-complete, are often more useful for other proofs than SAT
itself.
338 CHAPTER TWENTY-ONE: COMPLEXITY CLASSES

A SAT formula is in conjunctive normal form (CNF ) when it is the logical AND
of one or more clauses, each of which is the logical OR of one or more literals,
each of which is either a positive literal (a single Boolean variable) or a negative
literal (a logical NOT operator applied to a single Boolean variable).

Restricting instances to CNF produces a different decision problem:

CNF Satisfiability (CNFSAT )

Instance: a SAT instance in conjunctive normal form.
Question: is there some assignment of values to the variables that
makes the formula true?

CNFSAT instances use all the same operators as SAT instances, but restrict the order in
which they may be applied: variables may be negated using , then those pieces (the literals)
may be combined using , and finally those pieces (the clauses) may be combined using .
For example, here are some CNFSAT instances, written with a mathematical syntax:
(a) (b) true, under the assignment a=true
and b=true
(a) ( a) false, under all assignments
(a b) ( a b) true, under the assignment a=true
and b=false
( b) ( c) (a b c) ( a b c) false, under all assignments
This concrete syntax is simpler than that concrete syntax we used for SAT. (In fact,
depending on the details, it might well be regular—see Exercise 10.) But the concrete syntax
is not the important question here. For any reasonable representation of CNFSAT instances,
the important question is the satisfiability question: is there a truth assignment for the
variables that makes the whole formula true?
It is easy to see that CNFSAT is in NP, because it can use largely the same verification
method as SAT.
Lemma 21.2.1: CNFSAT NP.

Proof: The restricted syntax is context free and so can be checked in polynomial
time. With that change, an NP verification method can be constructed the same
way as for SAT, in the proof of Lemma 21.1.1.
In fact, checking the certificates could be simpler for CNFSAT than for general SAT. The
verification method only needs to check that, using the given truth assignment, each clause
contains at least one true literal. The more interesting result is that, in spite of the restricted
syntax, CNFSAT is still NP-hard.
 21.3 A TOUR OF SIX DECISION PROBLEMS 339
Lemma 21.2.2: CNFSAT is NP-hard.
Proof: By reduction from SAT. See Appendix C.

We have now shown that CNFSAT is in NP and is NP-hard. Putting these results together,
we have

Theorem 21.2.2: CNFSAT is NP-complete.

We now have two NP-complete problems. Each can be reduced to the other in polynomial
time: SAT P CNFSAT, and CNFSAT P SAT. In that sense, CNFSAT is exactly as difficult
to decide as SAT itself. It’s simpler without being any easier.
3SAT
It turns out that we can simplify satisfiability even more, without losing NP-completeness. We
can require that every clause contain exactly the same number of literals:

A CNF formula is in k-CNF if each clause contains exactly k distinct literals.

Restricting instances to 3-CNF produces a decision problem that is structurally very simple,
yet still NP-complete:

3-CNF Satisfiability (3SAT)

Instance: a SAT instance in 3-CNF.
Question: is there some assignment of values to the variables that
makes the formula true?

This is just a restricted syntax for SAT instances, so proving membership in NP is easy:

Lemma 21.3.1: 3SAT NP.

Proof: The restricted syntax is context free and so can be checked in polynomial
time. With that change, an NP verification method can be constructed the same
way as for SAT, in the proof of Lemma 21.1.1.

Proving NP-hardness is also straightforward. We already know that CNFSAT is NP-hard, so we

only need to show that CNFSAT P 3SAT.

Lemma 21.3.2: 3SAT is NP-hard.

Proof: By reduction from CNFSAT. Each clause c in a CNFSAT instance
consists of one or more literals: c = (l1 l2 … ln ). Depending on the number
340 CHAPTER TWENTY-ONE: COMPLEXITY CLASSES

n of literals in the clause, we can replace each clause c with one or more 3SAT
clauses as follows.
If n = 1, c = (l) for some single literal l. Create two new variables x and y, not
occurring elsewhere in the formula, and replace c with these four clauses:
(l x y) (l x y) (l x y) (l x y)
A truth assignment satisfies these if and only if it makes l true, so this
transformation preserves satisfiability. (Note that we cannot simply use
(l l l ), because 3-CNF requires three distinct literals in each clause.)
If n = 2, c = (l1 l2). Create one new variable x, not occurring elsewhere in the
formula, and replace c with these two clauses:
(l1 l2 x) (l1 l2 x)
A truth assignment satisfies these if and only if it makes l1 l2 true, so this
transformation preserves satisfiability.
If n = 3, c is already a 3-CNF clause, and no transformation is necessary.
If n 4, c = (l1 l2 l3 … ln ). Create n - 3 new variables, x3 through
xn-1, not occurring elsewhere in the formula, and replace c with these n - 2
clauses:
(l1 l2 x3)
( x3 l3 x4 )
( x4 l4 x5 )
( x5 l5 x6 )
…
( xn-1 ln-1 ln )
(Except for the first and last, these clauses are all of the form ( xi li xi+1 ).)
Suppose the original clause is satisfiable. Then there is some truth assignment
that makes at least one of the li true. This can be extended to a truth assignment
that satisfies all the replacement clauses, by choosing xk = true for all k i, and
xk false for all k > i. Conversely, suppose there is some truth assignment that
satisfies all the replacement clauses. Such a truth assignment cannot make all
the li false; if it did, the first replacement clause would require x3 = true, and so
the second would require x4 = true, and so on, making the last clause false. Since
a satisfying assignment for the replacement clauses must make at least one li
true, it also satisfies the original clause. Therefore, this transformation preserves
satisfiability.
 21.3 A TOUR OF SIX DECISION PROBLEMS 341
Applying this transformation to each clause in the original formula produces
a 3CNF formula ', with CNFSAT if and only if ' 3SAT. The entire
construction takes polynomial time. (In fact, the size of ' is linear in the size
of .) Therefore, CNFSAT P 3SAT. Because CNFSAT is NP-hard, we conclude
that 3SAT is NP-hard.

Incidentally, it turns out that 3 is the smallest value of k for which kSAT is NP-hard. The
satisfiability of 1-CNF and 2-CNF formulas can be decided in linear time.
We have now shown that 3SAT is in NP and is NP-hard. Putting these results together, we
have

Theorem 21.2.3: 3SAT is NP-complete.

Vertex Cover
Now for something (apparently) completely different. The Vertex Cover problem concerns a
property of graphs and does not, at first glance, appear to have anything to do with logic.
Imagine that you are running a courier service, and you have a map of cities connected
by roads. You want to establish transfer stations in the cities, enough so that there is at least
one at one end of every road. But you may not have enough money to build one in every
city. For example, suppose this is your map:
Turington

Gödel City
Kleene Valley

Cookville

Postonia

If you can afford to build k = 5 stations, there’s no problem: just put one in every city. But
suppose you can only afford k = 2. Is it still possible to place at least one at one end of every
road? Yes. If you build one in Turington and one in Postonia, that will suffice. On the other
hand, if you can only afford k = 1, there is no solution.
This is a concrete example of the Vertex Cover problem. In the abstract, a problem
instance has two parts: a graph and a goal k N. The graph G = (V, E) consists of a set V of
vertices and a set E of edges. Each edge is just a pair of vertices. So for the example above,
this would be
342 CHAPTER TWENTY-ONE: COMPLEXITY CLASSES

G = (V, E )
v z V = {v, w , x , y , z }
x
E = {(v, w ), (w, x ), (x, y ), (v, y ), (y, z )}
w

A vertex cover of a graph is a subset of the vertex set, V' V, such that for all
(u, v) E, u V' or v V'. Of course every graph has a vertex cover V' = V,
but the decision problem asks, is there a vertex cover V' of size k or less?

Vertex Cover
Instance: a graph G = (V, E) and a number k N.
Question: does G have a vertex cover of size k or less?

It is easy to see that Vertex Cover is in NP. The certificate can just be a proposed vertex
cover, and we have only to check that it is a correct cover and has size k or less. As usual, you
would have to specify some encoding of the problem instance as a string and check that the
input string is a legal encoding. But abstracting away those details, we have

Lemma 21.4.1: Vertex Cover NP.

Proof: We can make an NP verification method for Vertex Cover. The instance
is a graph G = (V, E) and a number k N, and the certificate for an instance is
some V' V. Check the syntax of the problem instance and certificate. Then
make a pass over the edge set E. For each (u, v) E, check that at least one of
u or v occurs in V'. For any reasonable encoding of the graph, these tests can be
completed in polynomial time.

Vertex Cover is also NP-hard. The proof, by reduction from 3SAT, is rather surprising:
what could Vertex Cover have to do with Boolean satisfiability? But we can show how to
take any 3-CNF formula and convert it into an instance of Vertex Cover: a graph G and
constant k such that G has a vertex cover of size k or less, if and only if is satisfiable.

Lemma 21.4.2: Vertex Cover is NP-hard.

Proof: By reduction from 3SAT. See Appendix C.

We have now shown that Vertex Cover is in NP and is NP-hard. Putting these results
together, we have

Theorem 21.2.4: Vertex Cover is NP-complete.

 21.3 A TOUR OF SIX DECISION PROBLEMS 343
There’s an interesting relative of Vertex Cover called Edge Cover. An edge cover is a subset
of the edge set, E' E, such that every v V is touched by some edge in E'.

Edge Cover
Instance: a graph G = (V, E ) and a number k N.
Question: does G have an edge cover of size k or less?

Recall that our first example of Vertex Cover used this graph:
y

v z
x

Examining the same graph for edge covers, we see that at k = 5, we can use every edge,
E' = E, to make an edge cover; at k = 3 there is still a solution at E' = {(y, z), (v, w), (x, y)};
while at k = 2 there is no solution.
The surprising thing here is that, while Vertex Cover is NP-complete, Edge Cover is in P.
We won’t present it here, but there’s a fairly simple, polynomial-time algorithm for finding a
minimum-size edge cover for any graph. This is a good example of how unreliable intution
can be as a guide to decision problems. Edge Cover may seem like a close relative of Vertex
Cover, but it’s in a different and (apparently) much simpler complexity class. Vertex Cover
seems at first glance to be completely unrelated to Boolean-satisfiability problems; yet Vertex
Cover, SAT, CNFSAT, and 3SAT are all NP-complete, so they can all be reduced to each other
in polynomial time.
Hamiltonian Circuit
The next is a classic route-finding problem: the problem of finding a path that starts and
ends at the same point and visits every vertex in the graph exactly once. (Think of a delivery
truck that needs to start at the warehouse, visit a list of destinations, and return to the
warehouse.)

Hamiltonian Circuit
Instance: a graph G = (V, E).
Question: is there a path that starts and ends at the same vertex,
arriving at every vertex exactly once?
344 CHAPTER TWENTY-ONE: COMPLEXITY CLASSES

For example, consider these three graphs:

The first two have Hamiltonian circuits; the third does not.
It is easy to see that Hamiltonian Circuit is in NP. The certificate can just be a list of
vertices in some order. We have only to check that it is a Hamiltonian circuit for the graph.
As usual, you would have to specify some encoding of the problem instance as a string and
check that the input string is a legal encoding. But abstracting away those details, we have

Lemma 21.5.1: Hamiltonian Circuit NP.

Proof: We can make an NP verification method for Hamiltonian Circuit. The

instance is a graph G = (V, E), and the certificate for an instance is some
sequence of vertices. Check the syntax of the problem instance and certificate.
Then make a pass over the certificate, checking that it is a Hamiltonian circuit:
that it visits every vertex exactly once, moves only along edges actually present in
the graph, and ends where it began. For any reasonable encoding of the graph,
these tests can be completed in polynomial time.

Hamiltonian Circuit is also NP-hard. The proof is by reduction from Vertex Cover.
Given any instance of Vertex Cover, consisting of a graph G and a natural number k, we can
construct a new graph H such that H has a Hamiltonian circuit if and only if G has a vertex
cover of size k or less.

Lemma 21.5.2: Hamiltonian Circuit is NP-hard.

Proof: By reduction from Vertex Cover. See Appendix C.

We have now shown that Hamiltonian Circuit is in NP and is NP-hard. Putting these
results together, we have

Theorem 21.2.5: Hamiltonian Circuit is NP-complete.

Many variations of this problem are also NP-complete. For example, the problem remains NP-
complete if we allow paths that visit every vertex exactly once, but don’t end up where they
started. It remains NP-complete if we consider graphs with one-way edges (directed graphs).
On the other hand, there’s an interesting relative of Hamiltonian Circuit called Euler
Circuit. A Hamiltonian circuit must visit each vertex exactly once; an Euler circuit must use
each edge exactly once.
 21.3 A TOUR OF SIX DECISION PROBLEMS 345

Euler Circuit
Instance: a graph G = (V, E).
Question: is there a path that starts and ends at the same vertex and
follows every edge exactly once?

Alhough this sounds like a close relative of Hamiltonian Circuit, it is actually much simpler.
We won’t present it here, but there is a simple algorithm to find an Euler circuit for a graph
or prove that none exists. (A graph has an Euler circuit if and only if it is fully connected and
every vertex has an even number of incident edges.) So Euler Circuit is in P.
Traveling Salesman
Imagine you’re planning the route for a delivery truck. It leaves the distribution center in the
morning, makes a long list of deliveries, and returns to the distribution center at night. You
have only enough gas to drive k miles. Is there a route you can use to make all your deliveries
without running out of gas?
This is a Traveling Salesman Problem.

Traveling Salesman Problem (TSP )

Instance: a graph G = (V, E) with a length for each edge and a total
length target k N.
Question: is there a path that starts and ends at the same vertex,
visits every vertex exactly once, and has a total length k?

Lemma 21.6.1: TSP NP.

Proof: We can make an NP verification method for TSP. The instance is a graph
G = (V, E) and a number k; the certificate for an instance is some sequence
of vertices. Check the syntax of the problem instance and certificate. Then
make a pass over the certificate, checking that it visits every vertex once, moves
only along edges actually present in the graph, ends where it began, and has a
total length k. For any reasonable encoding of the graph, these tests can be
completed in polynomial time.

Our TSP is so close to Hamiltonian Circuit that a polynomial-time reduction is simple.

Lemma 21.6.2: TSP is NP-hard.

Proof: By reduction from Hamiltonian Circuit. Any instance G = (V, E) of
Hamiltonian Circuit can be converted to an instance of TSP in polynomial time
346 CHAPTER TWENTY-ONE: COMPLEXITY CLASSES

by simply assigning a length of 1 to each edge in E and choosing k = |V |. With

these lengths, a Hamiltonian circuit of G is a TSP path of a total length k, and a
TSP path of total length k must be k edges long and so must be a Hamiltonian
circuit. Thus Hamiltonian Circuit P TSP. Because Hamiltonian Circuit is NP-
hard, we conclude that TSP is NP-hard.

We have now shown that TSP is in NP and is NP-hard. Putting these results together, we
have
Theorem 21.2.6: TSP is NP-complete.
There are many variations of this problem. In some, the goal is a closed circuit; in
others, the goal is a path from a specified target to a specified destination. In some the graph
is undirected; in others the graph is directed. In some, the goal is to visit every vertex at
least once; in others, exactly once. In some the edge lengths are Euclidean, or at least are
guaranteed to obey the triangle inequality (length(x, z) length(x, y) + length(y, z)); in
others the lengths are unconstrained. In some, the graph is always complete, with an edge
between every pair of vertices; in others the graph contains some edges but not others. All
these variations turn out to be NP-complete.
Unfortunately, there’s no standard terminology for these problems. In different contexts,
Traveling Salesman Problem might mean any one of these variations.

21.4 The Wider World of NP-Complete Problems

In the previous section we introduced six NP-complete problems. For all of them, it was easy
to show membership in NP; the difficult part for most was showing NP-hardness. We started
with SAT and sketched the proof of Cook’s Theorem, showing that SAT is NP-hard. From
there we developed NP-hardness results for the others using a chain of polynomial-time
reductions:
SAT P CNFSAT P 3SAT P Vertex Cover P Hamiltonian Circuit P TSP
Through this chain, SAT can be reduced in polynomial time to any of the others. Through
Cook’s Theorem, any of the others can be reduced back to SAT. So they all stand or fall
together. If a polynomial-time solution is found for any one of them, then all of them have
a polynomial-time solution. If a proof is found for any one of them, that it is not in P, then
none of them have a polynomial-time solution.
This is a remarkable state of affairs. It would be remarkable even if only these six
problems were involved. In fact, the number of known NP-complete problems is now well
into the thousands, with a vast tree of polynomial-time reductions rooted at SAT. There are
NP-complete decision problems in many different domains:
 21.4 THE WIDER WORLD OF NP-COMPLETE PROBLEMS 347
• Logic problems like SAT
• Mathematical problems involving sets and equation-solving
• Graph problems: covering problems like Vertex Cover, colorability problems, and
subgraph-finding problems
• Strategy problems for games
• Network problems involving routing, capacity, and connectivity
• Database problems, including compression, allocation, and retrieval
• Compiler problems, including code generation and register allocation
• Scheduling problems: the coordination of limited resources in diverse domains such
as business, manufacturing, construction, health care, and education
• Layout problems such as disk allocation, warehouse space allocation, and cutting-
machine optimization

An amazing variety of problems in all kinds of domains, some very far removed from
computer science, have turned out to be NP-complete. All of them stand or fall together.
For many of these decision problems there is a corresponding optimization problem. For
example:

Decision Problem Optimization Problem

Does the graph G have a vertex cover of Find a minimum-size vertex
size k or less? (Vertex Cover) cover for G.
Is there a circuit of the vertices of G, of Find a shortest circuit of G.
total length k or less? (TSP)

Can G be colored using no more than k Find a coloring of G using

colors, so that no two adjacent vertices the minimum number of
get the same color? colors, so that no two adjacent
vertices get the same color.
Is there a way to fill a knapsack from a Find the most valuable way
given inventory, so that the total value to fill a knapsack from a given
of the items is at least k and the total inventory, so that the total
weight is no more than w? weight is no more than w.
Is there a way to cut a sheet of metal Find the least wasteful way to
into a set of required parts, so that less cut a sheet of metal into a set
than k square millimeters is wasted? of required parts.
348 CHAPTER TWENTY-ONE: COMPLEXITY CLASSES

When studying formal language we focus on decision problems, but in other contexts you
are much more likely to encounter the corresponding optimization problems. People often
fail to distinguish between a decision problem and its corresponding optimization problem.
This is technically sloppy: P and NP are sets of languages, so it is not really correct to speak of
an optimization problem as being NP-complete. However, this casual usage is often justified
by the close connection between the two kinds of problems.
Consider, for example, our Vertex Cover decision problem and its corresponding
optimization problem, Vertex Cover Optimization. There is a simple recursive solution for
Vertex Cover Optimization:

minCover(G)
if G has no edges, return {}
else
let (v1, v2) be any edge in G
let V1 = minCover(G - v1)
let V2 = minCover(G - v2)
let n1 = |V1|
let n2 = |V2|
if n1 < n2 return V1 {v1}
else return V2 {v2}

Here, the expression G - v refers to the graph G, with the vertex v and all its incident edges
removed. In the base case, when there are no edges in the graph, the empty set is clearly the
minimum-size vertex cover. Otherwise, we choose any edge (v1, v2) in G. At least one of
the two ends of this edge must be in any cover for G, so we try both. We remove v1 and the
edges it covers and then recursively find a minimum cover V1 for the rest of the graph; then
V1 {v1} must have the minimum size of any cover for G that contains v1. Then we do the
same for the other vertex, so that V2 {v2} has the minimum size of any cover for G that
contains v2. Because we know that at least one of v1 or v2 must be part of any vertex cover, we
can conclude that the smaller of V1 {v1} and V2 {v2} must be a minimum vertex cover for
G. (If they are the same size, it would be correct to return either one.)
This is a reasonably simple solution, but very expensive. Our minCover calls itself
recursively, twice. That makes a binary tree of recursive calls, whose height is (in the worst
case) equal to the number of edges in the graph. That will take exponential time. Now, if we
could find n1 and n2 some other way—find the size of a minimum-size vertex cover, without
actually finding the cover—then we could reorganize the code so that minCover calls itself
recursively, only once:
 21.4 THE WIDER WORLD OF NP-COMPLETE PROBLEMS 349
minCover(G)
if G has no edges, return {}
else
let (v1, v2) be any edge in G
let n1 = minCoverSize(G - v1)
let n2 = minCoverSize(G - v2)
if n1 < n2 return minCover(G - v1) {v1}
else return minCover(G - v2) {v2}

This makes just one recursive call for each edge in the graph. In fact, not counting the time
spent in minCoverSize, this is now a polynomial-time solution for Vertex Cover Optimization.
That’s the basis of the following proof:

Theorem 21.3: Vertex Cover P if and only if Vertex Cover Optimization can be
computed in polynomial time.
Proof: Suppose Vertex Cover Optimization can be computed in polynomial
time. We can immediately use this to decide any Vertex Cover instance (G, k):
in polynomial time, compute a minimum-size cover U for G, and return true if
|U| k, false if |U| > k.
In the other direction, suppose Vertex Cover P. We can compute the size of a
minimum-size vertex cover like this:

minCoverSize(G)
for each i from 0 to the number of vertices in G
if (G, i) Vertex Cover return i

This minCoverSize function makes only polynomially many tests for

membership in Vertex Cover. Therefore, with Vertex Cover P, this is a
polynomial-time computation. We can then use it, as already shown, to
complete a polynomial-time solution for Vertex Cover Optimization.

This is a common situation: either the decision problem and the corresponding optimization
problem can be done in polynomial time, or neither can. Such optimization problems stand
or fall along with the thousands of decision problems for NP-complete languages.
One interesting feature of optimization problems, not shared by decision problems, is
that they are natural targets for approximation. For example:
350 CHAPTER TWENTY-ONE: COMPLEXITY CLASSES

Optimization Problem Approximation Problem

Find a minimum-size vertex Find a reasonably small vertex cover
cover for G. for G.
Find a shortest circuit of G. Find a reasonably short circuit of G.
Find a coloring of G using the Find a coloring of G using a
minimum number of colors, so reasonably small number of colors,
that no two adjacent vertices get so that no two adjacent vertices get
the same color. the same color.

Find the most valuable way Find a reasonably valuable way to fill
to fill a knapsack from a given a knapsack from a given inventory,
inventory, so that the total so that the total weight is no more
weight is no more than w. than w.

Find the least wasteful way to Find a way to cut a sheet of metal
cut a sheet of metal into a set of into a set of required parts, with a
required parts. reasonably small amount of waste.

In practice, many applications will be satisfied by any reasonably good solution, even if it
is not provably optimal. That’s good news, because while no polynomial-time solutions
are known for the optimization problems, we can sometimes find a reasonably close
approximation in polynomial time. (Of course, what makes a solution reasonably close to
optimal depends on the application.)
A good algorithm for finding an approximate solution for one NP-related optimization
problem does not necessarily translate to others. Each problem has unique approximation
properties. For example, consider Traveling Salesman Problem. In some variations, the
edge lengths are guaranteed to obey the triangle inequality (length(x, z) length(x, y) +
length(y, z)); in others the lengths are unconstrained. From the formal-language point of
view it makes no difference; both versions are NP-complete. But if you are trying to find an
approximation algorithm, it makes a great deal of difference. There is a good algorithm for
finding a circuit that is guaranteed to be no more than 3/2 times the length of the shortest
circuit—if the triangle inequality is obeyed. If the triangle inequality is not obeyed, you
can prove that no such approximation is possible in polynomial time, unless P = NP. (See
Exercise 16.)
But we will not explore these approximation algorithms further here. Our focus here is
on formal language. In the study of formal language we are chiefly concerned with deciding
language membership. In this black-and-white world, the NP-complete problems are all
 21.5 NONDETERMINISTIC TURING MACHINES 351
treated as interchangeable; each may be reduced to the others in polynomial time. Indeed,
we tend to think of them as just different faces of the same problem—different avatars of the
one, big, abstract, NP-complete superproblem.

21.5 Nondeterministic Turing Machines

Our Turing machines are deterministic, but one can also define nondeterministic Turing
machines (NDTMs). The big change from the deterministic version is in the transition
function. In the deterministic version there is at most one move from any configuration:
(q, a) = (p, b, D). In the nondeterministic version, we define (q, a) as a set of zero or more
such triples, one for each of the possible moves when in the state q reading a. We also change
the definitions of and * accordingly. The definition of L(M ) can remain the same, using
the new * relation. (That is, L(M ) = {x * | idfix(q0x) * ypz for some p F }.)
We’ve seen this kind of nondeterminism in several other contexts. When we explored
it in the context of finite-state automata, we saw that it made no difference in definitional
power; a language is definable using a DFA if and only if it is definable using an NFA.
On the other hand, in the context of stack machines, nondeterminism made a significant
difference. Nondeterminism was built into our basic definition of stack machines, and some
of the exercises in Chapter 15 explored the serious limitations imposed by trying to make
them deterministic. So how about Turing machines: does making them nondeterministic
change the class of languages that we can define? The answer is no: NDTMs can define
exactly the same languages that TMs can define.
To prove this, we’ll introduce a deterministic way to try any finite path of computation
through an NDTM. Say that an NDTM has been ordered if, for each state q and symbol
a, the set (q, a) has been arbitrarily indexed, starting from (q, a)0. Then we can define
a method runGuided(N,s,guide) that simulates one finite path of computation
of an ordered NDTM N on an input string s. The third parameter, the guide string, is
some string in {0, 1}*; on each move, one or more bits of guide are used to select the
transition. (Specifically, on each transition (q, a), we can choose k = | (q, a)|, take the
next max( log2 k ,1) bits from guide, treat them as the binary encoding of a number i,
and use (q, a)i mod k as the next move. Note that even on a deterministic move, at least one
bit is used—so runGuided always returns, after at most |guide| moves.) The method
runGuided returns one of three string values describing the outcome of that computation:
“reject” if it rejects, “accept” if it accepts, or “unfinished” if the guide string is exhausted
before N halts on s. Clearly this can be implemented as a polynomial-time method; the
details are left as an exercise.
Using this primitive, we can see that allowing nondeterminism does not change our
results concerning computability.

Theorem 21.4.1: A language is RE if and only if it is L(N) for some NDTM N.

352 CHAPTER TWENTY-ONE: COMPLEXITY CLASSES

Proof: Every TM is (with cosmetic changes in type) an NDTM, so in one

direction this is trivial: a recursively enumerable language is by definition
L(M) for some TM M, which can be taken as an NDTM that happens to have
| (q, a)| 1 for all q and a.

In the other direction, suppose L is L(N) for some NDTM N. Let

BinaryStar be an enumerator class for {0, 1}*. We construct a
recognition method for L as follows:

boolean lrec(String s) {
BinaryStar e = new BinaryStar();
while (true) {
String guide = e.next();
String result = runGuided(N,s,guide);
if (result.equals("accept")) return true;
}
}

If there is an accepting computation of N on s, there is at least one guide

string corresponding to that computation, and this method will eventually try
that guide string and return true. Otherwise there is no guide string for any
accepting computation, and the method will run forever. Therefore,
L(lrec) = L(N) = L. Because L has a recognition method, it is recursively
enumerable.

A similar construction, with a little more attention to detail, can be used to make a decision
method for L(N ) from any total NDTM N. (See Exercise 17.) So our definition of a
recursive language is likewise unaffected by the addition of nondeterminism.
On the other hand, the addition of nondeterminism seems to have a major effect on
our complexity classes. To explore this, we’ll need to revisit our cost model. We will say that
the cost of an NDTM computation is the cost of the most expensive of any of the possible
sequences on the given input:
time(M, x) = the maximum number of moves made by M on input x *
space(M, x) = the maximum length of tape visited by M on input x *
Building from that foundation, what kinds of complexity classes do we get? Surprisingly, one
at least is already familiar: the class of languages defined by some polynomial-time NDTM is
exactly the class NP. We prove this in two parts.
 21.5 NONDETERMINISTIC TURING MACHINES 353
Theorem 21.4.2: If L NP, L = L(N) for some polynomial-time NDTM N.
Proof: Let L be an arbitrary language in NP. Then L has a verification method
vL. Because this is an NP-verification method, it runs in polynomial time, and
we have a polynomial bound t(x) on the length of any viable certificate y. Now
construct an NDTM N with two phases of computation. In the first phase, N
skips to the end of its input string x, writes a separator symbol #, and then writes
a certificate string, choosing up to t(x) symbols to write nondeterministically.
Then it returns the head to the starting position, the first symbol of x. In the
second phase, N performs the same computation as vL(x, y), verifying the
instance/certificate pair. Now L(N) = L, and N runs in polynomial time.

The NDTM constructed by this machine does all of its nondeterministic computation up
front, capturing it in the certificate. Verifying the certificate is purely deterministic. The same
idea works in reverse: given any NDTM N, we can construct a verification algorithm that
works by capturing all the nondeterminism in the certificate. In fact, we’ve already seen the
construction: it’s our runGuided method.

Theorem 21.4.3: If N is a polynomial-time NDTM, L(N ) NP.

Proof: Suppose N is an NDTM. We construct a verification method vL using

the runGuided method previously defined:

boolean vL(String x, String p) {

return runGuided(N,x,p).equals("accept");
}

Here, the certificate is a guide string, and verifying it is simply a matter of

checking that the computation it specifies is an accepting one. Since N is a
polynomial-time NDTM, there is some polynomial-length guide string p for
any x L(N ), and runGuided takes polynomial time on it. Thus, vL is an
NP-verification method for L, and so L NP.

Putting those results together, we see that NDTMs and verification methods are
interchangeable ways to define the complexity class NP. (“There it is again!”) In fact, NP was
originally defined in terms of NDTMs; NP stands for Nondeterministic Polynomial time.
NDTMs can be used to define many other nondeterministic complexity classes, both
for space and time: NPSPACE, NEXPTIME, and so on. But of all these, the most important open
questions revolve around NP.
354 CHAPTER TWENTY-ONE: COMPLEXITY CLASSES

21.6 The Million-Dollar Question

We are now coming to the end of our study of complexity classes and to the end of this
book. We have seen only a few of the most important complexity classes. Researchers have
investigated the properties of hundreds of different classes, some more important than others,
and there are many open problems concerning them. But a few problems involving the most
important classes stand out as particularly tantalizing. We know for example that
P NP PSPACE EXPTIME

And we know that P EXPTIME, so at least one of the three inclusions above must be proper.
But we don’t know which one, or ones!
The one most important open problem in this field today is the question of P versus
NP. Roughly speaking, the P class represents problems for which we can find the answer in
polynomial time; the NP class represents problems for which we can verify a given answer in
polynomial time. Our intuition tells us that finding an answer can be much more difficult
than verifying one. Finding a proof feels harder than checking someone else’s proof; solving
a jigsaw puzzle takes hours of work, while checking someone’s solution takes just a glance.
(Same thing for Sudoku puzzles—and yes, decision problems related to Sudoku have been
shown to be NP-complete.) To this intuition, we add a long experience of failure—the failure
to find a polynomial-time solution to any of the NP-hard problems. Many of these problems
are of great commercial importance, and many programmers over the last three decades have
struggled with them, but to no avail.
For most researchers, this all adds up to a feeling that these problems are truly
intractable, though we can’t yet prove it. In fact, modern computing depends on this
supposed intractability in many ways. For example, many of the cryptographic techniques
that protect financial transactions on the Internet depend on this assumption. If a
polynomial-time solution to an NP-hard problem were ever found, there would be quite a
scramble to find other ways to secure network transactions!
But just because many researchers have this same feeling is no reason to consider the
solution inevitable. The history of mathematics and computing contains a number of widely
believed hunches and conjectures that turned out to be false. Indeed, we’ve seen a major one
already in this book: in the early 1900s, most mathematicians shared David Hilbert’s feeling
that a proof of the completeness and correctness of number theory was just around the
corner, until Kurt Gödel proved them all wrong. We’ve learned not to be too complacent; a
consensus of hunches is no substitute for a proof.
All in all, it’s a dramatic situation. There is a tension that has existed since the problem
was first identified over 30 years ago, and it has grown as the list of NP-complete problems has
grown. To resolve this growing tension a solid proof is needed: ideally, either a polynomial-
time decision method for some NP-complete problem, or a proof that no such method exists.
Certain fame awaits anyone who can solve this problem. Fame, and fortune: the P versus NP
 EXERCISES 355
problem is one of seven key problems for which the Clay Mathematics Institute has offered
prizes—$1 million for each problem. The prizes were announced in 2000, in Paris, 100
years after Hilbert announced his famous list of problems in that same city. So far, the prizes
remain unclaimed.

21.7 Further Reading

An excellent discussion of NP, along with a very useful catalog of several hundred (seemingly)
intractable problems, can be found in
Garey, Michael, and David Johnson. Computers and Intractability: A Guide to the
Theory of NP-Completeness. New York: W. H. Freeman and Company, 1979.

Exercises
EXERCISE 1
For each of the following verification methods, describe the language it defines, and write
a decision method for it.
a. boolean xxVerify(String x, String p) {
return x.equals(p+p);
}
b. boolean compositeVerify(int x, int p) {
return p > 1 && p < x && x % p == 0;
}
c. boolean vc(String x, String p) {
return x.equals(p+p+p);
}
d. boolean vd(String x, String p) {
return x.equals(p);
}
e. boolean ve(int x, int p) {
return x = 2*p;
}
f. boolean vf(int x, int p) {
return x = p*p;
}
EXERCISE 2
Assuming that L is any NP-complete language, use the definitions to prove the following
statements:
a. For all A NP, A P L.
b. For all A, if A is NP-hard then L P A.
c. For all A, if A is NP-complete then A P L.
356 CHAPTER TWENTY-ONE: COMPLEXITY CLASSES

d. For all A, if A is NP-complete then L P A.

e. If L P, then all NP-complete languages are in P, and P = NP.
f. If L P, then no NP-complete language is in P, and P NP.
EXERCISE 3
Consider the following statement: For any NP-complete language L, and for all A NP,
L P A. Can you prove it true? Can you prove it false? How or why not?
EXERCISE 4
Here is an example of a concrete syntax for SAT.
S &(S,S) | |(S,S) | !(S) | V
V AV | A
A a|b|…|z
Write a Java application DecideSAT that decides SAT. Your program should parse a
Boolean formula from the command line using the syntax above. (The operators are all
prefix and fully parenthesized in this syntax, so writing a recursive descent parser from
this grammar is straightforward.) Then your program should decide whether the formula
is satisfiable. If so, it should print true, followed by a satisfying assignment. If not (or if
the formula is illegal) it should print false. For example:
> java DecideSAT '&(fred,ginger)'
True with [(fred=true)(ginger=true)]
> java DecideSAT '&(way,!(way))'
False
Feel free to implement this with a reasonable limit on the number of distinct variables—
the exponential-time search will make giant formulas useless anyway.
EXERCISE 5
Convert these formulas to CNF. (That is, make an equivalent formula using exactly the
same variables, but in conjunctive normal form.)
a. a
b. (a b)
c. (a b) (a b)
d. (a b c)
e. a (b c)
f. (a b) (c d)
g. a b (c d)
EXERCISE 6
(This exercise uses material from Appendix C, Section C.2.) Write a Java application
SAT2CNF to convert SAT formulas to CNF. Your program should parse a Boolean
formula from the command line using the syntax from Exercise 4. Then your program
 EXERCISES 357
should convert the formula to CNF, using the transformations in the proof of
Theorem C.1, and print out the resulting CNF formula. For example:
> java SAT2CNF '!(&(fred,ginger))'
|(!(fred),!(ginger))
> java SAT2CNF '|(&(a,b),c)'
&(|(a,c),|(b,c))
You should be able to make this work using only the transformations in the proof of the
theorem: the two DeMorgan’s laws, the double negative law, the distributive law, and the
commuted version of the distributive law.
EXERCISE 7
(This exercise uses material from Appendix C, Section C.2.) Apply the transformation
used in the proof of Lemma 21.2.2, converting the formulas from Exercise 5 into
formulas in CNF. (Unlike Exercise 5, this will involve the addition of new variables.)
EXERCISE 8
(This exercise uses material from Appendix C, Section C.2.) Write a Java application
SAT2CNFP to generate a CNFSAT instance from a SAT instance. Your program should
parse a Boolean formula from the command line using the syntax from Exercise 4.
Then your program should convert the formula to a CNFSAT instance, using the
transformations in the proof of Lemma 21.2.2, and print out the resulting CNF
formula. For example:

> java SAT2CNFP '!(&(fred,ginger))'

|(!(fred),!(ginger))
> java SAT2CNFP '|(&(a,b),c)'
&(&(|(#0,a),|(#0,b)),|(!(#0),c))

This is similar to Exercise 6, but not the same; you should not use the standard
distributive law, but should use instead the special transformation from the proof of
Lemma 21.2.2 and the commuted version of it, which introduce extra variables as
necessary but achieve a polynomial-time reduction. (Our sample output above names the
extra variables #i, but you can name them whatever you like.)
EXERCISE 9
Use the construction from the proof of Lemma 21.3.2 to convert these CNF formulas to
3-CNF. (This may involve the introduction of new variables.)
a. a
b. (a b c)
c. (a b)
d. (a b)
358 CHAPTER TWENTY-ONE: COMPLEXITY CLASSES

e. (a b) (c d)
f. (a b c d)
g. (a b c d e f g)
EXERCISE 10
We define a concrete syntax for CNF formulas using the operator symbols , , and .
There is exactly one pair of parentheses around each clause and no other parentheses in
the formula. Variable names are strings of one or more lowercase letters.
a. Show that this language is regular by giving a left-linear grammar for it.
b. Show that the restriction of this language to 3-CNF is regular by giving a left-linear
grammar for it.
EXERCISE 11
(This exercise uses material from Appendix C, Section C.1.) The proof of Cook’s
Theorem defines a function e = f (a, b, c, d). This function defines a symbol in an ID for
a Turing machine in terms of the previous-step symbols at and near that position:

... a b c d ...

... e ...

Give pseudocode for computing this function f using the transition function for the
TM.
EXERCISE 12
(This exercise uses material from Appendix C, Section C.3.) Convert a a to a
3SAT instance, show the Vertex Cover instance constructed for it as in the proof of
Lemma 21.4.2, and show a minimum-size vertex cover for it. (Because the formula is not
satisfiable, you should find that there is no vertex cover of size 2m + n.)
EXERCISE 13
Consider the following decision problem:

Monotone 3SAT
Instance: a formula in 3-CNF with the property that in each clause,
either all literals are positive or all are negative.
Question: is there some assignment of values to the variables that
makes the formula true?

For example, (a b d) ( a b d) is a monotone formula, a legal instance of

Monotone 3SAT. By contrast, ( a b d) is a 3SAT formula but not a monotone one.
 EXERCISES 359
Prove that Monotone 3SAT is NP-complete. (Hint: To prove NP-hardness, use a reduction
from 3SAT. Your reduction will need to convert any 3-CNF formula into a monotone
3-CNF formula ' such that is satisfiable if and only if ' is. This will require the
introduction of new variables.)
EXERCISE 14
Consider the following decision problem:

1-in-3SAT
Instance: a formula in 3-CNF.
Question: is there some assignment of values to the variables that
makes exactly one literal in each clause true?

Here, the instances are the same as for 3SAT, but the question is different; we need to
decide whether the formula has an assignment that makes exactly one literal in every
clause true. (Our regular satisfiability problems ask whether there is an assignment that
makes at least one literal in every clause true.)
Prove that 1-in-3SAT is NP-complete. (Hint: To prove NP-hardness, use a reduction from
3SAT. For each clause in the original formula, come up with a set of new clauses that
have a 1-in-3 satisfying assignment if and only if the original clause has a plain satisfying
assignment. This will require the introduction of new variables.)
EXERCISE 15
The pseudocode in the proof of Theorem 21.3 is inefficient. (We were only interested in
showing polynomial time.) Show an implementation of minCoverSize that makes only
logarithmically many calls to the decision method for Vertex Cover.
EXERCISE 16
The Traveling Salesman Optimization problem is this: given a graph G with a length for
each edge, find a path that starts and ends at the same vertex, visits every vertex exactly
once, and has a minimum total length. An approximate solution to this optimization
problem would be an algorithm for finding a reasonably short circuit. We’ll say that a
circuit is reasonably short if it is no more than k times longer than the minimum-length
path, for some constant k > 1. (Such a constant k is called a ratio bound.)
Now consider the following construction. Given a graph G = (V, E) and a constant k > 1,
construct a new graph G' with the same vertices. For each edge in G, add the same edge
to G' with length 1. For each pair of vertices not connected by an edge in G, add that
edge to G' with length k|V |. (Thus G' is a complete graph, with an edge of some length
between every pair of vertices.)
a. Prove that if G has a Hamiltonian circuit then any Traveling Salesman Optimization
solution for G' has a total length |V |.
360 CHAPTER TWENTY-ONE: COMPLEXITY CLASSES

b. Prove that if G does not have a Hamiltonian circuit then any Traveling Salesman
Optimization solution for G' has a total length > k|V |.
c. Using these results, prove that for any constant k > 1, if there is an approximate
solution for Traveling Salesman Optimization with a ratio bound k that runs in
polynomial time, then P = NP.
(Note that our construction builds a graph G' with lengths that may violate the triangle
inequality. For graphs that obey the triangle inequality, although TSP remains NP-
complete, there are approximation techniques with good ratio bounds.)
EXERCISE 17
An NDTM N is total if, for all inputs x *, all legal sequences of moves either accept
(by entering an accepting state) or reject (by entering a configuration from which there is
no possible next move).
Using the ideas from the proof of Theorem 21.4.1, show how to construct a decision
method ldec from any total NDTM N, so that L(ldec) = L(N ).
EXERCISE 18
Using your solution to Exercise 17, show that NP PSPACE. (You do not need to show
implementations for things like runGuided and BinaryStar, but you should
explain your assumptions about their space requirements.)
EXERCISE 19
Implement the runGuided method for NDTMs, as outlined in Section 21.5. Then,
using that and your ldec method from Exercise 17, write a Java implementation of
a universal Turing machine for NDTMs. Your program should read the NDTM to be
simulated from a file. You can design your own encoding for storing the NDTM in a
file, but it should be in a text format that can be read and written using an ordinary
text editor. Your class should have a main method that reads a file name and a text
string from the command line, uses your ldec method to decide whether the string is
in the language defined by the NDTM, and reports the result. For example, if the file
anbncn.txt contains the encoding of an NDTM for the language {anbncn}, then your
program would have this behavior:

> java UNDTM anbncn.txt abca

reject
> java UNDTM anbncn.txt abc
accept

If you did Exercise 13 from Chapter 17 (a Java implementation of a universal TM for

deterministic machines) you should be able to reuse much of that code.
 A
AP PEN DIX

From an NFA to a
Regular Expression
For every NFA, there is an equivalent regular expression. In Chapter 7
we showed an example of how to construct a regular expression for
an NFA, but skipped the general construction and proof. In this
appendix we tackle the proof. Warning: There are mathematical
proofs whose elegance compels both agreement and admiration.
They are so simple and concise that they can be understood in their
entirety, all at once. Unfortunately, this is not one of them!

361
362 APPENDIX A: FROM AN NFA TO A REGULAR EXPRESSION

A.1 The Internal Languages of an NFA

When you see an NFA, you normally think only of the language it accepts: the language
of strings that take it from the start state to end in any accepting state. But to construct a
regular expression from an NFA, as demonstrated in Section 7.5, we need to consider other
languages relevant to a given NFA.

Let M be any NFA with n states numbered q0 through qn-1. Define the internal
language L(M, i, j, k) = {x | (qi, x) * (qj, ε), by some sequence of IDs in which
no state other than the first and the last is numbered less than k}.

L(M, i, j, k) is the set of all strings that take the NFA from the state i to the state j, without
passing through any state numbered less than k. For example, let M be this DFA:
0 1

1 0
q0 q1 q2

1 0

• L(M, 2, 2, 3) is the set of strings x for which (q2, x) * (q2, ε), where states
other than the first and the last are all numbered 3 or higher. Since there are
no states in M numbered 3 or higher, there can be no IDs in the sequence
other than the first and last. So there are only two strings in this language:
ε (which produces the one-ID sequence (q2, ε)) and 1 (which produces the
two-ID sequence (q2, 1) (q2, ε)). Thus L(M, 2, 2, 3) = L(1 + ε).
• L(M, 2, 2, 2) is the set of strings x for which (q2, x) * (q2, ε), where states
other than the first and the last are all numbered 2 or higher. Since q2 is the
only state that can appear in the sequence, the strings in this language are
just strings of any number of 1s. Thus L(M, 2, 2, 2) = L(1*).
• L(M, 2, 1, 2) is the set of strings x for which (q2, x) * (q1, ε), where states
other than the first and the last are all numbered 2 or higher. Since q2 is the
only state other than the last that can appear in the sequence, the strings in
this language are just strings of any number of 1s with a single 0 at the end.
Thus L(M, 2, 1, 2) = L(1*0).
• L(M, 0, 0, 0) is the set of strings x for which (q0, x) * (q0, ε), where states
other than the first and the last are all numbered 0 or higher. Since all states
in M are numbered 0 or higher, this is just the language of strings that
take M from its start state to its accepting state. Thus, for this machine,
L(M, 0, 0, 0) = L(M ).
 A.2 THE D2R FUNCTION 363

A.2 The d2r Function

Next we will write a recursive function d2r(M, i, j, k) that constructs a regular expression for
any internal language L(M, i, j, k). Later, this will be used in the construction of a regular
expression for the entire language L(M ).
First, consider the cases of L(M, i, j, k) where k is greater than or equal to the number of
states in M. Such internal languages are easy to give regular expressions for because, like the
first example above, they do not permit any sequences other than the one-transition sequence
qiqj and, in the case of i = j, the zero-transition sequence qi. So let r = a1 + a2 + …, for all
ap {ε} with qj (qi, ap ). (In case there is no such ap, let r = .) Then the regular
expression for L(M, i, j, k), where k is greater than or equal to the number of states in M, is
either r (if i j) or r + ε (if i = j). This will be the base case in our recursive definition of the
function d2r. It is a base case that applies to any call with a sufficiently large value of k. As we
look at the recursive cases below, bear this in mind: recursive calls that use increasing values of
k make progress toward this base case.
Now consider the cases of L(M, i, j, k) where k is less than the number of states in M.
These are by definition strings that can take M through a sequence of one or more states
starting from qi and ending in qj , where states other than the first and last are all numbered
k or higher. These strings are permitted to take M through state qk , but are not required to,
so they can be divided into two cases: those that have sequences that actually do visit qk , and
those that have sequences that don’t. (Because M may be nondeterministic, some strings may
fall into both cases.) Those that have sequences that don’t visit qk are in L(M, i, j, k + 1), so
we will be able to generate regular expressions for them recursively, as r1 = d2r(M, i, j, k + 1).
It remains to consider those strings that have sequences of moves that actually do visit qk
at least once. These sequences start with qi, reach qk one or more times, and eventually end in
qj. Picture such a sequence with all its occurrences of qk made explicit:
qi ... qk ... qk ... ... qk ... qj
Such a sequence can be treated in three parts: first a sequence from qi to qk ; then zero or
more sequences from qk back to qk ; and finally a sequence from qk to qj. Because we have
made all the occurrences of qk explicit endpoints of these subsequences, qk does not occur as
an intermediate state in any of the subsequences. All of the subsequences use only intermediate
states numbered k + 1 or higher; so we will be able to generate regular expressions for them
recursively. For sequences from qi to qk, we can use r2 = d2r(M, i, k, k + 1); for sequences
from qk to qk, we can use r3 = d2r(M, k, k, k + 1); and for sequences from qk to qj, we can use
r4 = d2r(M, k, j, k + 1). The whole regular expression for L(M, i, j, k), where k is less than the
number of states in M, can then be written as r1 + (r2)(r3)*(r4). The resulting implementation
for d2r is
364 APPENDIX A: FROM AN NFA TO A REGULAR EXPRESSION

d2r(M, i, j, k)
if k (number of states in M ) then
let r = a1 + a2 + …, for all ap {ε} with qj (qi, ap),
or if there are no transitions from qi to qj
if i = j then return r + ε else return r
else
let r1 = d2r(M, i, j, k + 1)
let r2 = d2r(M, i, k, k + 1)
let r3 = d2r(M, k, k, k + 1)
let r4 = d2r(M, k, j, k + 1)
return r1 + (r2)(r3)*(r4)

This is a complete procedure for constructing a regular expression for any internal language
of an NFA.

A.3 Proof of Lemma 7.2

Now we can restate and prove Lemma 7.2.
Lemma 7.2: If N is any NFA, there is some regular expression r with
L(N) = L(r).
Proof: By construction. Let M = (Q, , , q0, F ) be any NFA. We have a
recursive function, d2r, that constructs a regular expression for any internal
language L(M, i, j, k). L(M ) is the language of strings that take M from q0 to
end in any accepting state, using any intermediate states. Thus

L(M )= L(M,0,j,0)
⊃

qj ∈ F

So to construct a regular expression for L(M), compute d2r(M, 0, j, 0) for each

qj F, then combine them all with + into one, big, regular expression. (In the
case of F = {}, use as the regular expression.) The resulting regular expression r
has L(r) = L(M ).

For a full example of this construction, consider this NFA:

a b

b
q0 q1

a
 A.3 PROOF OF LEMMA 7.2 365
There is only one accepting state, q1, so L(M) = L(M, 0, 1, 0). The evaluation of
d2r(M, 0, 1, 0) uses the recursive case, returning r1 + (r2)(r3 )*(r4), where

r1 = d2r(M, 0, 1, 1)
r2 = d2r(M, 0, 0, 1)
r3 = d2r(M, 0, 0, 1)
r4 = d2r(M, 0, 1, 1)

Now the evaluation of d2r(M, 0, 1, 1), for both r1 and r4, uses the recursive case,
returning r5 + (r6)(r7)*(r8), where

r5 = d2r(M, 0, 1, 2) = b
r6 = d2r(M, 0, 1, 2) = b
r7 = d2r(M, 1, 1, 2) = b + ε
r8 = d2r(M, 1, 1, 2) = b + ε

Those four are all base cases, as shown. Finally the evaluation of d2r(M, 0, 0, 1), for both r2
and r3, uses the recursive case, returning r9 + (r10)(r11)*(r12), where

r9 = d2r(M, 0, 0, 2) = a + ε
r10 = d2r(M, 0, 1, 2) = b
r11 = d2r(M, 1, 1, 2) = b + ε
r12 = d2r(M, 1, 0, 2) = a

Those are also base cases, so we can now assemble the whole expression, omitting some
unnecessary parentheses to make the result slightly more readable:
d2r(M, 0, 1, 0)
= r1 + r2r3*r4
= r5 + r6r7*r8 + r2r3*(r5 + r6r7*r8)
= r5 + r6r7*r8 + (r9 + r10r11*r12)( r9 + r10r11*r12)*(r5 + r6r7*r8)
= b + b(b + ε)*(b + ε)
+ (a + ε + b(b + ε)*a)(a + ε + b(b + ε)*a)*(b + (b + ε)*(b + ε))
The language accepted by the NFA in this example is the language of strings over the
alphabet {a, b} that end with b; a simple regular expression for that language is (a + b) * b. It
is true, but not at all obvious, that this short regular expression and the long one generated
by the formal construction are equivalent. As you can see, the regular expression given by
the construction can be very much longer than necessary. But that is not a concern here.
The point of Lemma 7.2 is that the construction can always be done; whether it can be done
concisely is another (and much harder) question. It can always be done—so we know that
there is a regular expression for every regular language.
 B
AP PEN DIX

A Time-Hierarchy
Theorem
The more time you allow for a decision, the more languages you can
decide. It’s not a surprising claim, perhaps, but it isn’t easy to prove.

366
 B.2 TIME-CONSTRUCTIBLE FUNCTIONS 367

B.1 Another Asymptotic Notation: Little o

To begin, we need to state precisely what “more time” means. To do that, we’ll introduce a
new asymptotic notation, using little o instead of big O.

Let f and g be any two functions over N. We say

f(n) is o(g(n))
if and only if for every c > 0 there exists some n0, so that for every
n n0, f (n) ≥ 1c g(n).

To say that f(n) is o(g(n)) is to say that f grows more slowly than g. That’s stricter than big O;
little o rules out the possibility that f and g grow at the same rate. Where big O is like , little
o is like <. That will serve as our definition of “more time”: g(n) allows more time than f (n) if
f(n) is o(g(n)).
In Chapter 19, Theorem 19.2 summarized the comparative rates of growth of some
common asymptotic functions. It turns out that all the results given there for big O can be
tightened to little o. We can state the following:
Theorem B.1:
• For all positive real constants a and b, a is o((log n)b).
• For all positive real constants a and b, and for any logarithmic base, (log n)a is o(nb).
• For all real constants a > 0 and c > 1, na is o(cn).

B.2 Time-Constructible Functions

Next, we’ll need to restrict the claim to complexity functions that can be implemented
reasonably efficiently.

A function f is time constructible if there exists some method

int fval(int n)
with fval(n) = f(n), where time(fval, n) is O(f (n)).

A function f is time constructible if we can compute what f(n) is within O(f (n)) time. (This
use of a time bound is unusual in that it measures time as a function of the input number,
not as a function of the size of the representation of that number.) For example, the function
n2 is time constructible, because we can implement it as

int nSquared(int n) {
return n * n;
}
368 APPENDIX B: A TIME-HIERARCHY THEOREM

This computes n2 and takes (log n) time, which is certainly O(n2). Most commonly
occurring functions, like n, n log n, nk, and 2n, are easily shown to be time constructible. (In
fact, it is challenging to come up with a function that is (n) but not time constructible.)

B.3 A Cost Model for run and runWithTimeLimit

When we introduced the methods run and runWithTimeLimit back in Chapter 18,
we were concerned only with computability, not efficiency. Now, to use these methods to
establish a time-hierarchy theorem, we must add some assumptions about them to our cost
model.
First, the run method: we do not model a separate compilation step, but treat all
execution as interpreting the source code. If m is a method and mSource is a string
containing the source code for m, it makes no difference whether we write m(in) or
run(mSource,in)—they do the same thing and are assumed to take asymptotically the
same space and time.
In Chapter 18 we said that runWithTimeLimit(p,in,n) returns the same thing
run(p,in) returns, if that completes within n steps, but returns false otherwise. We left
the concept of a runtime “step” undefined; now we stipulate that it must be some constant-
time unit of computation, according to our cost model. There is some constant c N,
c > 0, such that, if time(run, (p,in)) c n, then runWithTimeLimit(p,in,n) =
run(p,in).
Simply running a method p for n steps takes (n) time. However,
runWithTimeLimit(p,in,n) takes more than (n) time, because of the
overhead involved in enforcing the time limit. Think of it as a countdown: after each
step, runWithTimeLimit decrements the original counter n, and stops the execution
(returning false) if that counter ever reaches 0. This decrement-and-test after each step
takes O(log n) time. Thus, in the worst case, runWithTimeLimit(p,in,n) takes
(n log n) time.

B.4 A Time-Hierarchy Theorem

Now we can state a time-hierarchy theorem—a formal version of our claim that the more
time you allow for a decision, the more languages you can decide.

Theorem B.2 (Time Hierarchy): If g is time constructible and f (n) is o(g(n)),

then TIME (f(n)) TIME (g(n) log n).

Roughly speaking, this says that any more than an extra log factor of time allows you to
decide more languages. Of course, you can still decide all the languages you could have
decided without that extra log factor; clearly TIME(f(n)) TIME (g(n) log n). The tricky part is
to show that the inclusion is proper—that there exists at least one language
 B.4 A TIME-HIERARCHY THEOREM 369
A TIME (g(n) log n) that is not in TIME ( f (n)). Such a language A is the one decided by this
method:

1. boolean aDec(String p) {
2. String rejectsItself =
3. "boolean rejectsItself(String s) {" +
4. " return !run(s,s); " +
5. "} ";
6. int t = gVal(p.length());
7. return runWithTimeLimit(rejectsItself,p,t);
8. }

Here, gVal is a method that computes g(n) and does it in O(g(n)) time; we know that such
a method exists because g is time constructible.
Let A be the language decided by aDec. Because gVal is the method given by the time
constructability of g, line 6 runs in O(g(n)) time. Line 7 requires O(g(n) log g(n)) time, so
the whole method aDec requires O(g(n) log g(n)) time. This shows A TIME (g(n) log g(n)).
It remains to show that A TIME (f(n)).
Toward that end, we prove a lemma about the method aDec:

Lemma B.1: For any language in TIME ( f(n)), there is some decision method
string p for which aDec(p) = !run(p,p).
Proof: For any decision method string p, two conditions are sufficient to
establish that aDec(p) = !run(p,p): first, that p is O( f(n)), and second,
that p is sufficiently long. For if p is O( f(n)), and given that f(n) is o(g(n)), then
time(run, (rejectsItself,p)) is o(g(|p|)). This means that there exists
some n0 such that for every |p| n0,
time(run, (rejectsItself,p)) g(|p|).
Because g(|p|) = t, we can conclude that when p is sufficiently long, the
runWithTimeLimit at line 7 does not exceed its time limit. Therefore it
produces the same result as a plain run:

runWithTimeLimit(rejectsItself,p,t)
= run(rejectsItself,p)
= !run(p,p).
Now let B be any language in TIME (f(n)). By definition, there must exist an
O(f(n)) decision method for B. Its source code may not have |p| n0, but it is
easy to construct a sufficiently long equivalent method, simply by appending
370 APPENDIX B: A TIME-HIERARCHY THEOREM

spaces. Thus, B always has some decision method string p that meets our two
conditions, so that aDec(p) = !run(p,p).

Lemma B.1 shows that for every decision method for a language in TIME ( f(n)), there
is at least one string p about which aDec makes the opposite decision. It follows that the
language A, the language decided by aDec, is not the same as any language in TIME ( f(n)).
Therefore A TIME (f(n)), and this completes the proof of Theorem B.2.
The proof of this time-hierarchy theorem uses a variation of the same technique
used in Section 18.2 to prove that Lu is not recursive. Note the similarity between the
rejectsItself method used in this proof and the nonSelfAccepting method
used there.

B.5 Application
The time-hierarchy theorem is a tool for proving that one complexity class is a proper subset
of another. For example, we can use it to prove that TIME (2n) is a proper subset of TIME(22n).
To apply the theorem, we choose f (n) = 2n and g(n) = 22n / log2 n. (Note the separation here:
we choose a g that is at least a log factor below the time-complexity function we’re ultimately
interested in.) First, to invoke the time-hierarchy theorem, we need to know that g is time
constructible. Next, we need to know that f (n) is o(g(n)). By definition, this is true if for
/
every c > 0 there exists some n0, so that for every n n0, 2n ≥ 1c 22n log2 n . Simplifying, we
see that this inequality is satisfied whenever we have 2n / log2 n c, which is clearly true for
sufficiently large n, for any constant c. So we can apply the time-hierarchy theorem and
conclude that TIME( f (n)) TIME(g (n) log n). Substituting back in for f and g, we conclude
that TIME (2n) is a proper subset of TIME (22n).
This time-hierarchy theorem cannot be used to prove that time-complexity classes are
distinct when they differ by exactly one log factor. For example, though it is true that TIME (n)
is a proper subset of TIME (n log n), our time-hierarchy theorem cannot be used to prove it.
The theorem kicks in only when the complexity functions differ by more than a log factor, in
the little-o sense.
 C
AP PEN DIX

Some NP-Hardness
Proofs
This appendix presents the proofs of some of the trickier NP-hardness
results from Chapter 21.

371
372 APPENDIX C: SOME NP-HARDNESS PROOFS

C.1 Cook’s Theorem: SAT Is NP-hard

This is one of the most important results in complexity theory. The details of the proof are
too hairy even for this appendix, but the following proof sketch should give you a good idea
of how it can be done.

Lemma 21.1.2 (Cook’s Theorem): SAT is NP-hard.

Proof sketch: Let L be an arbitrary language in NP. We need to show that
L P SAT. In other words, we need to show that there is a polynomial-time
function f that maps any string x into a Boolean formula f (x), with the property
that x L if and only if f(x) SAT. That sounds like a tall order—knowing
almost nothing about a language L, we have to show how to map its decision
problem to the SAT decision problem.
The one thing we do know about L is that it is in NP. Therefore, there exists
some polynomial-time, verification algorithm for L, using polynomial-size
certificates. Now, the proof is simpler if we take this algorithm to be given as a
Turing machine M. Then we can be a little more specific about the mapping
function f we’re looking for; f (x) should be satisfiable if and only if the TM M
accepts the string x#y for some certificate string y.
We know that M is an NP-verifier, so if it has not accepted within some
polynomially bounded number of steps, it never will. Thus, for each string x we
can precompute a number t(x), depending only on the length of x, to serve as
a strict bound on the number of steps. Using this information, we can be still
more specific about the mapping function f we’re looking for; f (x) should be
satisfiable if and only if M accepts x#y for some certificate string y, within t(x)
moves.
We can model any such accepting computation as a sequence of exactly t(x) IDs,
on the understanding that once M halts, the subsequent IDs in this sequence are
unchanged. We can also make each ID in this sequence exactly 2t(x) symbols
wide, padding with blanks as necessary and placing the initial head position
in the center. (We know that no ID needs to be wider than 2t(x), because the
head can’t make more than t(x) moves to the left or right.) If you lay out such a
sequence of t(x) IDs, each of size 2t(x), you get a grid like this:
 C.1 COOK’S THEOREM: SAT IS NP-HARD 373
1 2 2t(x)
1 B ... B q0 x1 ... x2 # y1 ... y2 B ... B

2 ... ... ... ...

... ... ... ...

... ... ... ...

...

...
t(x) ... ... ... ...

In terms of this grid, we can be still more specific about the mapping function f
we’re looking for; f (x) should be satisfiable if and only if the grid can be filled so
that
1. the first line is a starting ID of M on x (for some y), as shown in the illustration
above,
2. each line follows from the line before it according to the transition function of M
(or is the same as the line before it, when M has halted), and
3. the final row is an ID in an accepting state.

Now we’re getting close to something we can express as a Boolean formula.

The one remaining snag is the alphabet; cells in the grid contain ID symbols,
which are either tape symbols or states, but the only values we can talk about
in a Boolean formula are Boolean. No problem: we simply encode each value
in Q using a fixed-length string over the binary alphabet {T, F}. Let d
be the length of the string of Ts and Fs used to encode each cell (presumably,
d= log2|Q ∪ Γ | ). Then the (i, j)th cell of the grid can be encoded using d
Boolean variables, bi,j,1 through bi,j,d . The whole encoding of the grid requires
d 2t(x)2 such variables—a number which is polynomial in |x|.

Now our mapping function f (x) can construct a big formula using these Boolean
variables. The formula is a conjunction of three parts, encoding requirements 1
through 3 above.
f (x)= starts moves accepts
374 APPENDIX C: SOME NP-HARDNESS PROOFS

Here’s where it gets technical, and we’ll skip these details. Roughly speaking:
1. starts enforces the requirement that the first line of the grid is a starting ID of M on
x. This establishes some of the values b1,j so that the start state is q0, the tape to the
left is all Bs, and the tape at and to the right of the head contains x#. It leaves the
y part, the certificate, unconstrained; ultimately, our formula will be satisfiable if
there is some y part that makes the whole system true.
2. moves enforces the requirement that each line follow from the next according to
M’s transition function. In fact, because of the way a Turing machine works, each
individual cell bi+1,j is a simple function of its near neighbors in the previous row,
bi,j-1 through bi,j+2:

... a b c d ...

... e ...

This function e = f(a, b, c, d ) is determined by the transition function of M. (When

this move might affect e, one of the inputs is a state, showing the current state
and head position. If none of the inputs is a state, the head is not near enough to
have any effect on this cell, and so f(a, b, c, d) = b.) With the inputs and outputs
encoded as Boolean values, this function has a fixed, finite number of Boolean
inputs and outputs, and so can be encoded as a Boolean formula. The conjunction
of all those Boolean formulas, one for every cell in the grid except for the first row,
gives a formula that is satisfiable if and only if the whole grid correctly follows from
the first row according to the transition function of M.
3. accepts enforces the requirement that the final row shows M in an accepting state. At
least one of the bt(x),j encodes an accepting state.
The resulting formula f(x) is polynomial in the size of the grid, which is polynomial
in the size of x. The starts part determines every cell of the first row, except the
certificate y. For any choice of certificate y, the moves part fully determines every
other cell in the grid. And the accepts part is satisfied if and only if M accepts x#y.
Therefore, the whole formula f (x) is satisfiable if and only if there is some certificate
y for which M accepts x#y. Therefore, x L if and only if f (x) SAT, and so
L P SAT. Because L is an arbitrary language in NP, this proves that SAT is NP-hard.

C.2 CNFSAT Is NP-hard

For proving that CNFSAT is NP-hard, there’s good news and bad news. The good news is
that we don’t have to repeat all the work of the previous NP-hardness proof. If we can just
 C.2 CNFSAT IS NP-HARD 375
prove that SAT P CNFSAT, that will be enough; we’ve already proved that for every L NP,
L P SAT, so by transitivity we’ll have L P CNFSAT.
The bad news is that the polynomial-time reduction from SAT to CNFSAT isn’t as
simple as it first appears. There’s a tempting approach that doesn’t quite work:

Theorem C.1: For any SAT formula, there is an equivalent formula in CNF.
Proof sketch: Any SAT formula can be transformed to CNF by using simple
laws of Boolean logic. First, all the operators can be moved inward until they
apply directly to variables, thus forming the literals for CNF. For this we can use
DeMorgan's laws and the law of double negation:
• (x y) = ( x y)
• (x y) = ( x y)
• x=x

Then, all the operators can be moved inward until they apply directly to
literals, thus forming the clauses for CNF. For this we can use the distributive
law:
• (x y) z = (x z) (y z)

Repeated application of this law (and the commuted version,

x (y z) = (x y) (x z)) completes the transformation to CNF.

This is only a proof sketch; a more complete proof would have to define the structure of
SAT instances more closely and then use a structural induction to demonstrate that the
transformations proposed actually do handle all possible SAT instances. (You can experiment
with this for yourself; see Exercise 6 in Chapter 21.)
The theorem shows that any SAT instance can be converted into an equivalent
CNFSAT instance '. Then, of course, SAT if and only if ' CNFSAT. So why doesn’t
this prove that SAT P CNFSAT? The problem is that, although it is a reduction from SAT
to CNFSAT, it is not always a polynomial-time reduction. In particular, that distributive
law is trouble. It makes duplicate terms; as written above, the right-hand side contains two
copies of the term z. For formulas that require this operation repeatedly, the result can be
an exponential blow-up in the size of the formula (and in the time it takes to construct it).
Consider a term of this form:
(a1 b1) (a2 b2) … (ak bk)
for some k. After the conversion, the CNF formula will have one clause for each way to
choose one variable from every pair. For example:
(a1 b1) (a2 b2) =
376 APPENDIX C: SOME NP-HARDNESS PROOFS

(a1 a2) (a1 b2) (b1 a2) (b1 b2)

(a1 b1) (a2 b2) (a3 b3) =
(a1 a2 a3) (a1 a2 b3) (a1 b2 a3) (a1 b2 b3)
(b1 a2 a3) (b1 a2 b3) (b1 b2 a3) (b1 b2 b3)
The length of the original SAT formula is linear in k, while the number of clauses in
the CNF formula is 2k. So, although our Theorem C.1 shows that every formula has an
equivalent CNF formula, it does not give us a polynomial-time reduction from SAT to
CNFSAT.
To show SAT P CNFSAT, we need a more subtle reduction, one that does not rely
on converting each SAT instance into an equivalent CNF formula '. Luckily, we don't
really need the two formulas to be equivalent. We need something rather more flexible:
that ' is satisfiable if and only if is. As long as that is true, the two formulas can be quite
different; in particular, they can use different variables. That’s the leverage we need to make a
polynomial-time reduction.

Lemma 21.2.2: CNFSAT is NP-hard.

Proof sketch: We will show SAT P CNFSAT. From any SAT formula we
can construct a CNF formula ' so that is satisfiable if and only if ' is.
First, all the operators can be moved inward until they apply directly to
variables, thus forming the literals for CNF, as in the proof of Theorem C.1.
These transformations preserve satisfiability (and in fact produce an equivalent
formula).

Next, all the operators can be moved inward until they apply directly to
literals, thus forming the clauses for CNF. For this we use a special variation of
the distributive law, one that introduces a new variable di:

(x y) z = (di x) (di y) ( di z)

(Each application of this law generates a new variable di, distinct from any other
variable in .) The result of applying this law is no longer an equivalent formula,
because it contains a new variable. However, suppose the left-hand side is
satisfiable. Then there is some truth assignment that makes either x y or z true.
We can extend this truth assignment to one that makes the right-hand side true,
by making di = z. Conversely, suppose the right-hand side is satisfiable; then
there is some truth assignment that makes all three clauses true. This assignment
must make either di or di false; therefore it must make either x y or z true.
So it is also a satisfying truth assignment for the left-hand side. Therefore, this
transformation preserves satisfiability.
 C.3 VERTEX COVER IS NP-HARD 377
Repeated application of this law (and the commuted version, x (y z) =
(di x) ( di y) ( di z)) completes the construction of a CNF formula
' that is satisfiable if and only if the original formula is satisfiable. The entire
construction takes polynomial time, so we conclude that SAT P CNFSAT.
Because SAT is NP-hard, we conclude that CNFSAT is NP-hard.

This is only a proof sketch; a more complete proof would have to define the structure
of SAT instances more closely and then use a structural induction to demonstrate that
the transformations proposed actually do convert all possible SAT instances and can be
implemented to do it in polynomial time. (You can experiment with this for yourself; see
Exercise 8 in Chapter 21.)

C.3 Vertex Cover Is NP-hard

Before we get to this proof, let’s look at an example of the construction we’ll use. Suppose our
formula is this:
= (x y z) ( x y z)
This formula has n = 3 variables and m = 2 clauses. For each variable, we’ll create two
vertices, one for the positive literal and one for the negative, with an edge between them, like
this:
x x y y z z

Then, for each clause, we’ll add three vertices connected in a triangle. Our example has two
clauses, so we’ll add this:

Finally, we’ll add edges from each triangle vertex to the vertex for the corresponding literal
in that clause. For instance, the first clause is (x y z), so our first triangle will get edges
from its three corners to the three vertices for x, y, and z. We end up with this graph:
x x y y z z
378 APPENDIX C: SOME NP-HARDNESS PROOFS

Now consider a minimum-size vertex cover for this graph. It must include two of the
three vertices from each triangle—there’s no other way to cover those three triangle edges.
It must also include one vertex for each variable, the positive or negative literal, in order
to cover those edges at the top of the illustration. So no cover can contain fewer than
2m + n = 7 vertices. Here’s an example at that minimum size, with the covered vertices
circled:
x x y y z z

The covered literals give us a satisfying assignment for : we covered x, y, and z, so let
x = true, y = true, and z = true. By construction, that makes at least one literal in each
clause true. In fact, any cover of the size 2m + n gives a satisfying truth assignment, and any
satisfying truth assignment gives a cover of the size 2m + n. This correspondence is the main
idea in the following proof.

Lemma 21.4.2: Vertex Cover is NP-hard.

Proof: By reduction from 3SAT. Let be any 3-CNF formula. Let m be the
number of clauses in it, and let n be the number of variables. We will construct a
graph G , as follows.

First, for each of the n variables xi, create a variable component, consisting of two
vertices, labeled xi and xi, with an edge between them:

xi xi

This gives us a vertex for each possible literal. Next, for each of the m clauses
(l1 l2 l3 ), create a clause component: three vertices connected in a triangle, and
three connecting edges (the curved lines below) from each to the corresponding
vertex already created for that literal:
 C.3 VERTEX COVER IS NP-HARD 379
l1 l2 l3

That completes the construction of the graph G . To make a complete Vertex

Cover instance we also need a constant k; choose k = 2m + n. The size of this
instance and the time to construct it are clearly polynomial in the size of .
Now we claim that is satisfiable if and only if G has a vertex cover of size k
or less. First, suppose is satisfiable. Then we can use a satisfying assignment
to construct a vertex cover V'. First, from each variable component, add either
xi or xi to V', whichever is true under the satisfying assignment. These vertices
cover all the edges in the variable components. They also cover at least one
of the connecting edges for each clause component—because, in a satisfying
assignment, there is at least one true literal in each clause. For each clause
component, select one connecting edge that is already covered (there may be
more than one), and add the other two triangle vertices to V'. This covers the
three triangle edges and the other two connecting edges. Thus all the edges in
the graph are covered. We used a vertex for each variable, plus two for each
clause: |V'| = k = 2m + n.
On the other hand, suppose G has a vertex cover V' of size k or less. This
must contain at least two vertices from each clause component—there’s no
other way to cover the three edges of the triangle. It must also contain at least
one vertex from each variable component—there’s no other way to cover the
variable component’s edge. That makes k = 2m + n vertices already, so we
must have |V'| = k. There is no room for any other vertices in V'. Because V'
contains exactly one of xi or xi for each variable, we can use it to make a truth
assignment for : make xi = true if xi V', and make xi = false if xi V'.
Because V' contains exactly two of the triangle vertices from each clause
component, one of the connecting edges from each clause component must
be covered from the other end; therefore, each clause contains at least one true
literal under this truth assignment, so it is a satisfying assignment for .
380 APPENDIX C: SOME NP-HARDNESS PROOFS

We have shown that 3SAT P Vertex Cover. Because 3SAT is NP-hard, we

conclude that Vertex Cover is NP-hard.

C.4 Hamiltonian Circuit Is NP-hard

We will prove this result by reduction from Vertex Cover. Given any instance of Vertex Cover,
consisting of a graph G and a natural number k, we will show how to construct a new graph
H such that H has a Hamiltonian circuit if and only if G has a vertex cover of size k or less.
The construction makes repeated use of a special subgraph that we’ll call a widget.
(Our constructed graph H will have one widget for every edge in G.) The widget has twelve
vertices, connected like this:

The widgets will be connected in the larger graph only by their corners. As such, there are
only two ways a widget can participate in a Hamiltonian circuit of that larger graph. The
circuit might enter at one corner, visit every vertex of the widget, and leave at the opposite
corner on the same side, like this:

We’ll call that a one-pass path. The alternative is that the circuit might make two separated
passes through the widget, visiting one side on one pass and the other side on the other, like
this:
 C.4 HAMILTONIAN CIRCUIT IS NP-HARD 381
We’ll call that a two-pass path. Our reduction will use the fact that there is no other way to
include a widget in a Hamiltonian path:

Lemma C.1: If a widget graph is connected in a larger graph only at its corners,
then a Hamiltonian circuit entering at one of the corners must exit at the
opposite corner on the same side, having visited either all the vertices in the
widget or all the vertices on that side of the widget and none on the other.
Proof: By inspection.

To convince yourself this is true, try making some paths through a widget. You’ll quickly see
why most of your choices are restricted by the need to visit every vertex exactly once.
In the following construction, there is one widget in the new graph for each edge in the
original graph. The widgets are connected together in such a way that Hamiltonian circuits
in the new graph correspond to vertex covers of the original graph. If an edge is covered from
just one end, our Hamiltonian circuit will visit its widget by a one-pass path; if it is covered
from both ends, by a two-pass path.

Lemma 21.5.2: Hamiltonian Circuit is NP-hard.

Proof: By reduction from Vertex Cover. Let G = (V, E ) and k N be any
instance of Vertex Cover. If k |V| there is always a vertex cover, so we can
reduce these to any small, positive instance of Hamiltonian Circuit. Otherwise,
k < |V|, and we construct a graph H as follows.
First, for each pair of vertices x and y connected by an edge in G, add a widget
to H, labeled at the corners as follows:

the y side
yxL yxR

xyL xyR
the x side

Next, add k vertices v1 through vk to H.

Next, for each vertex x in G, let y1…yn be any ordering of all the vertices
connected to x in G, and add edges (xy1R, xy2L), (xy2R, xy3L), and so on through
(xyn-1R, xynL). These edges connect all the x sides (of widgets that have x sides)
into one path.
382 APPENDIX C: SOME NP-HARDNESS PROOFS

Finally, for each vertex x in G and for each of the new vertices zi in H, add
the edges from zi to the start of the x-side path (that is, from zi to xy1L) and
from the end of the x-side path to zi (that is, from xynR to zi ). These edges
complete paths leading from any zi, through any x-side path, to any zk. That
completes the construction. The size of H and the time to construct it are clearly
polynomial in the size of G.
Now we claim that the constructed graph H has a Hamiltonian circuit if and
only if G has a vertex cover of size k or less. First, suppose H has a Hamiltonian
circuit. This must include every zi. By construction the only way to get from one
zi to the next is by a ui-side path for some variable ui, so the Hamiltonian circuit
must include k different ui-side paths. Let U = {u1, …, uk } be the set of variables
whose ui-side paths are used in the circuit. Every vertex of every widget must be
included in the circuit, so every widget must occur on one or two ui-side paths
for some ui U. Therefore, every edge in G must have one or two vertices in U.
So U is a vertex cover for G.
On the other hand, suppose U = {u1, …, uk } is some vertex cover for G. Then
we can make a circuit in H as follows. Start at z1; take the u1-side path to z2;
then take the u2-side path to z3; and so on. At the end of the uk-side path, return
to z1. Because U is a vertex cover for G, every edge has either one or two ends in
U; therefore every widget occurs either once or twice in this circuit. If once, use
the one-pass path through that widget; if twice, use the two-pass paths. In this
way we visit every vertex exactly once, so this is a Hamiltonian circuit.
We have shown that Vertex Cover P Hamiltonian Circuit. Because Vertex Cover
is NP-hard, we conclude that Hamiltonian Circuit is NP-hard.

Perhaps a quick example of this construction will help to elucidate the proof. Here is the
graph G for an instance of Vertex Cover; next to it is the graph H constructed by the proof,
with k = 1.
baL b side baR bcL b side bcR dcL d side dcR
a

b abL a side abR cbL a side cbR cdL c side cdR

c

z1
 C.4 HAMILTONIAN CIRCUIT IS NP-HARD 383
As you can see, G has no vertex cover of size k = 1, just as H has no Hamiltonian circuit. Any
set containing a single vertex fails to cover all the edges in G and corresponds to a path that
fails to visit all vertices in H. For example, the set {c} covers only two of the three edges in G,
just as the path from z1 through the c sides and back visits only two of the three widgets in
H.
On the other hand, if we allow k = 2, then G has a cover and H has a Hamiltonian
circuit:
z1

baL b side baR bcL b side bcR dcL d side dcR

a

b abL a side abR cbL a side cbR cdL c side cdR

c

z1

The illustration shows the cover for G (with circled vertices) and shows (with solid lines) a
corresponding Hamiltonian circuit in H: from z1 through the b sides to z2, then through the
c sides and back to z1. Notice how the widgets for the edges (a, b) and (c, d ) are visited using
the one-pass path, since those edges are covered from only one end, while the widget for
(b, c) is visited using the two-pass paths, since it is covered from both ends.
 i n d e x
A B
abstract syntax tree, 152 backtracking search, 60–62, 96, 200
algorithm Backus, John, 150
approximation, 350, 359–360 Backus-Naur form. See BNF
Cocke-Kasami-Younger (CKY), 207 base case, 29
for deciding Primality, 316, 324 BNF (Backus-Naur form), 150–151, 215–216
for DFA minimization, 108, 113 bottom-up parsing, 205–207, 209
for edge cover minimization, 343
for finding an Euler circuit, 345 C
for finding minimum star height, 112 capturing parentheses, 94
for regular expression equivalence, 111 Cartesian product, 25
parsing. See parsing certificate of language membership, 332–337
tractable, 316 CFG (context-free grammar). See also grammar
using reduction, 265 conversion from stack machine, 173–176
verification, 353, 372 conversion to stack machine, 171–173
vs. effective computational procedure, 244 definition, 146
alphabet for regular languages, 147–148
binary, 229, 373 in the Chomsky hierarchy, 279
definition, 2 undecidable questions for, 280
in a 3-tape TM, 227 with balanced pairs, 148
in Java, 36 with concatenations, 148–149
input, in a stack machine, 166 with unions, 149–150
input, of a TM, 222 CFL (context-free language)
Kleene closure of, 4 closure under concatenation, 187–188
nonterminal, for a grammar, 121 closure under intersection with regular language, 188
of a DFA, 13 closure under Kleene star, 188
of a regular expression, 76 closure under union, 187
of an NFA, 51, 52 definition, 146
renaming, 174–175 not closed for complement, 189–191
stack, in a stack machine, 166 not closed for intersection, 189
stacked, 238–240 checkers, 323
tape, of a TM, 222 chess, 323
terminal, for a grammar, 121 Chomsky hierarchy, 279–280
ambiguity Chomsky, Noam, 150, 279, 281
inherent, in context-free languages, 154, 158 Church, Alonzo, 243, 244, 283
in grammars, 153–154 Church-Turing Thesis, 244
approximation problems, 349–350 Clay Mathematics Institute, 355
asymptotic bounds closure
common, 294–295 Kleene, of a language, 188
little-o, 367 Kleene, of an alphabet, 4
lower (big- ), 290 of CFLs under concatenation, 187–188
simplification of, 292–294 of CFLs under intersection with regular languages, 188
tight (big- ), 291 of CFLs under Kleene star, 188
upper (big-O), 290 of CFLs under union, 187
384
 INDEX 385
of regular languages under complement, 23 Diophantine equations, 283
of regular languages under intersection, 23–25 disjoint sets, 121
of regular languages under union, 26–27 doubly exponential time. See 2EXPTIME
properties of DCFLs, 209 DPDA (deterministic PDA), 208
properties of P, 329 draughts, 323
properties of PSPACE, 329 Dyck language, 158
CNF (conjunctive normal form), 338
CNF Satisfiability (CNFSAT), 338 E
Cocke-Kasami-Younger (CKY) algorithm, 207 EBNF (Extended BNF), 154
complement in regular expressions, 112, 324 Edge Cover, 343
complement of a language, 22 effective computational procedure, 243
complete axiom system, 282 egrep, 92–94
complexity class, 312 ELEMENTARY TIME, 112, 319, 324
compositionality, 50 empty set, 2
computable function, 244, 278 empty string, 2
concatenation enumerator, 271–275
of languages, 76, 148 ε-transitions
of strings, 3 cycles of, 224
conjunctive normal form (CNF), 338 for NFA composition, 79
consistent axiom system, 282 in NFAs, 47
context-free grammar. See CFG in stack machines, 161, 167
context-free language. See CFL in the subset construction, 68
context-sensitive grammar, 279 Euler Circuit, 345
context-sensitive language (CSL), 280 exponential growth, 295
Cook, Stephen, 337 EXPSPACE (exponential space)
Cook’s Theorem, 337, 346, 372–374 completeness, 111, 320, 323
CSL (context-sensitive language), 280 definition, 320
cubic growth, 295 hardness, 320
Curry, Haskell, 243 EXPTIME (exponential time)
completeness, 319, 323
D definition, 317
DCFL (deterministic CFL), 208–209 hardness, 319
decidable property, 244, 278
decision method, 256 F
declaration before use, 196 finite languages, 142
derivation relations, 122 finite-state transducer, 109–111
deterministic CFL (DCFL), 208–209 flex, 99
deterministic finite automaton. See DFA
deterministic PDA (DPDA), 208 G
deterministic stack machine, 211–212 go, 323
DFA (deterministic finite automaton) Gödel, Kurt, 282, 354
applications, 36 grammar. See also CFG (context-free grammar)
encoded as a string, 229–231 conversion from NFA to, 122–124
formal definition, 13–14 derivation in, 122
informal definition, 10–11 formal definition, 121
language accepted by, 14–15 language generated by, 121–122
minimization, 106–108 left-linear form, 127
simulated by a TM, 229–231 regular, 127
difference of sets, 33 right-linear form, 126–127, 279
386 INDEX

single-step form, 125–126 N

unrestricted, 279–280 natural numbers, 3
grep, 94 Naur, Peter, 150
NDTM (nondeterministic Turing machine), 351
H NFA (nondeterministic finite automaton)
halting problems, 262–264, 267–270, 281, 283 conversion from regular expression, 79–83
Hamiltonian Circuit, 343, 380 conversion to regular expression, 83–84, 361–365
Hilbert, David, 282, 354 equivalence, 322
formal definition, 51
I informal definition, 46–50
ID (instantaneous description) internal languages of, 362
for a stack machine, 167 language accepted by, 52–53
for a TM, 222 minimization, 108
for an NFA, 52 regular languages defined by, 69–70
incompleteness theorems, 282 NLBA (nondeterministic linear-bounded automaton), 280
inductive case, 29 non-context-free grammars, 279
inductive hypothesis, 29 nondeterminism
inductive proof technique, 27–32 captured in a certificate, 353
infinite loop, 109, 224, 244, 259. See also loop, infinite in NFAs, 50
instantaneous description. See ID in stack machines, 166
intersection of languages, 23 in Turing machines, 280, 351
Java regexp support, 96–97 nondeterministic finite automaton. See NFA
Kleene closure nondeterministic linear-bounded automaton (NLBA), 280
of a language, 76, 188 nonRE languages, 277–278
of an alphabet, 4 nonterminal alphabet, 121
Kleene, Stephen, 243 nonterminal symbol, 120
Kleene’s Theorem, 84 nontrivial property of an RE language, 269
NP (nondeterministic polynomial time)
L completeness, 335, 346
language, 4 definition, 333
left recursion in grammars, 202, 206 hardness, 335
left-linear grammar, 127 versus P, 354–355
von Leibnitz, Gottfried Wilhelm, 282
length of a string, 2 O
lex, 97–100 oracles, 281–282
linear bounded automaton, 109
linear growth, 295 P
LL(1), 202–205 P(polynomial time)
logarithmic growth, 295 definition, 314
loglinear growth, 295 versus NP, 354–355
look-ahead table, 201–203 parallel search, 62–65
loop, infinite. See infinite loop parse tree
LR(1), 206–207, 209 definition, 151–153
height, 184
M minimum-size, 185
man, wolf, goat, and cabbage puzzle, 8–11 pumping, 182–186, 191
Matcher class in Java, 97 parsing
Matiyasevich, Yuri, 283 bottom up, 205–207, 209
Minsky, Marvin, 233 LL(1), 202–205
 INDEX 387
LR(1), 206–207, 209 relation to CSLs, 280
recursive descent, 203–205 recursive-descent parsing, 203–205
top down, 200–205 recursively enumerable languages. See RE languages
Pattern class in Java, 97 reduction
PDA (push-down automaton), 207–209 to prove a language not recursive, 266–268
polynomial growth, 295 to prove a language recursive, 264–266
polynomial space. See PSPACE regexp
polynomial time. See P in egrep, 92
polynomial-time reducibility, 319 in grep, 92–94
Post, Emil, 243 in Java, 96
powerset, 50 in lex, 97–100
Presburger arithmetic, 324 in Perl, 97
Presburger, Mojzesz, 324 regular expression
product construction, 25–26 conversion from NFA, 83–84, 361–365
productions in a grammar, 118, 121 conversion to NFA, 79–83
PSPACE (polynomial space) definition, 76
completeness, 111, 322 equivalence, 322, 323
definition, 320 extended with complement, 112
Pumping Lemma extended with squaring, 111
for context-free languages, 191–197 star height of, 112
for regular languages, 137–142 regular languages
pumping parse tree, 182–186, 191 closure under complement, 23
push-down automaton (PDA), 207–209 closure under intersection, 23–25
closure under reversal, 89
Q closure under union, 26–27
QBF (Quantified Boolean Formula), 322 defined by DFAs, 15
quadratic growth, 295 defined by grammars, 122, 127
quotient construction, 108 defined by NFAs, 68
defined by regular expressions, 84
R having unique minimum-state DFAs, 108
ratio bound, 359–360 including all finite languages, 142
RE (recursively enumerable) languages reverse
and enumeration machines, 275 of a language, 89
and NDTMs, 351 of a string, 89, 136
and recognition methods, 257 Rice’s Theorem, 268–271
and unrestricted (type-0) grammars, 280 right-linear grammar, 126–127, 279
definition, 225 run method, 257–259, 368
not closed for complement, 278 runGuided method, 351
Rice’s theorem and, 268 Run-with-time-limit decision problem, 317–319
trivial and nontrivial properties of, 269 runWithTimeLimit method, 276, 317–318, 368
recognition method, 256–257 Russell, Bertrand, 6
recursive languages Russell’s Paradox, 6
and computable functions, 243
and decidable properties, 244 S
defined by decision methods, 256 SAT (Satisfiability), 336–338
defined by total TMs, 225, 278 Schönfinkel, Moses, 243
including all regular languages and all CFLs, 229 semi-decidable property, 278
proved by reduction, 264–266 set former, 4
388 INDEX

single-step grammar, 125–126 informal definition, 218–221

space language accepted by, 223
complexity classes, 320 languages defined by, 225
cost function for a TM, 296 oracles, 281
cost function for an NDTM, 352 simulating any DFA, 229–231
cost function for Java, 299 three-tape, 227–228
hierarchy theorem, 320 total, 225, 256
squaring in regular expressions, 111 universal, 232–233, 257
stack machine top-down parsing, 200–205
conversion from CFG, 171–173 total TM, 225, 256
conversion from NFA, 170–171 transition function implementation, 41
conversion to CFG, 173–176 trap state, 10, 68
defining context-free languages, 176 Traveling Salesman Optimization, 359
deterministic, 211–212 Traveling Salesman Problem (TSP), 345–347, 360
formal definition, 166 triply exponential time (3EXPTIME), 319
informal definition, 160–166 trivial property of an RE language, 269
language accepted by, 167 TSP (Traveling Salesman Problem), 345–347, 360
simulated by a TM, 231 Turing equivalence, 244
star height, 112 Turing machine. See TM
start symbol, 120, 121 Turing, Alan, 243, 244, 271, 283
string, 2 Turing-computable function, 238
string exponentiation, 3 2EXPTIME (doubly exponential time), 319, 324
string variable, 3 two-way finite automaton, 108–109
structural induction, 82
subset construction, 65–69 U
union of languages, 26, 149
T universal TM, 232–233, 257
terminal alphabet, 121 unrestricted grammars, 279–280
terminal symbol, 120
3EXPTIME (triply exponential time), 319 V
3SAT (3-CNF Satisfiability), 339 verification method, 332
time Vertex Cover, 341–343, 347, 377–380
complexity classes, 312 Vertex Cover Optimization, 348
cost function for a TM, 296
cost function for an NDTM, 352 W
cost function for Java, 299 worst-case-space
hierarchy theorem, 313–314, 367–370 cost function for a TM, 296
TM (Turing machine) cost function for Java, 299
composition, 238–240 worst-case-time
computing arithmetic functions with, 240–241 cost function for a TM, 296
formal definition, 222 cost function for Java, 299
implementing random access, 241–242