0% found this document useful (0 votes)

270 views

C++ and Object-Oriented

This document provides a preface and introduction to a book on C++ and object-oriented numeric computing for scientists and engineers. The book aims to provide an easy introduction to C++ with an emphasis on applying the language to numeric computation. It covers basic C++ features, object-oriented programming, generic programming, and C++ standard libraries. It also introduces several common numeric algorithms and shows how to implement them efficiently in C++.

Uploaded by

VgokulGokul

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

270 views

C++ and Object-Oriented

Uploaded by

VgokulGokul

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 334

This is page i

Printer: Opaque this

C++ and Object Oriented Numeric

Computing for Scientists and Engineers
Daoqi Yang

To Xiujuan, Joy, and Forrest

Preface

This book is intended to be an easy, concise, but rather complete, introduction to the ISO/ANSI C++ programming language with special emphasis on object-oriented numeric computation for students and professionals
in science and engineering. The description of the language is platformindependent. Thus it applies to dierent operating systems such as UNIX,
Linux, MacOS, Windows, and DOS, as long as a standard C++ compiler
is equipped. The prerequisite of this book is elementary knowledge of calculus and linear algebra. However, this prerequisite is hardly necessary if
this book is going to be used as a textbook for teaching C++ and all the
sections on numeric methods are skipped. Programming experience in another language such as FORTRAN, C, Ada, Pascal, Maple, or Matlab will
certainly help, but is not presumed.
All C++ features are introduced in an easy way through concepts such
as functions, complex numbers, vectors, matrices, and integrals, which are
familiar to every student and professional in science and engineering. In
the nal chapter, advanced features that are not found in FORTRAN, C,
Ada, or Matlab, are illustrated in the context of iterative algorithms for
linear systems such as the preconditioned conjugate gradient (CG) method
and generalized minimum residual (GMRES) method. Knowledge of CG,
GMRES, and preconditioning techniques is not presumed and they are explained in detail at the algorithmic level. Matrices can be stored in full (all
entries are stored), band (entries outside a band along the main diagonal
are zero and not stored to save memory), and sparse (only nonzero entires
are stored to save memory) formats and exactly one denition for CG or
GMRES is needed that is good for all three matrix storage formats. This is

This is page vii

Printer: Opaque this

viii

in contrast to a procedural language such as FORTRAN, C, Ada, or Matlab, where three denitions of CG and GMRES may have to be given for the
three matrix formats. This is one of the salient features of object-oriented
programming called inheritance: The CG and GMRES methods are dened for a base class and can be inherited by classes for full, band, and
sparse matrices. If one later decides to add, for example, a symmetric (full,
band, or sparse) storage format (only half of the entries need be stored for a
symmetric matrix to save memory), the code of CG and GMRES methods
can be reused for it without any change or recompilation.
Another notable feature is generic programming through templates, which
enables one to dene a function that may take arguments of dierent data
types at dierent invocations. For example, the same denition of CG and
GMRES can deal with matrices with entries in single, double, and extended
double (long double) precisions and complex numbers with dierent types
for real and imaginary parts. Again, using C or FORTRAN would require
dierent versions of the same function to handle dierent data types when
the implementation of the function is going to be changed later for robustness or eciency, then every version has to be changed, which is tedious and
error-prone. The operator overloading feature of C++ enables one to add
and multiply matrices and vectors using + and in the same way as adding
and multiplying numbers. The meaning of the operators is user-dened and
thus it provides more exibility than Matlab. For example, the operators
+ and can be used to add and multiply full, band, and sparse matrices.
These and other features of C++ such as information hiding, encapsulation, polymorphism, error handling, and standard libraries are explained in
detail in the text. With increasingly complicated numeric algorithms, many
sophisticated data structures can be relatively easily implemented in C++,
rather than in FORTRAN or C this is a very important feature of C++.
The C++ compiler also checks for more type errors than FORTRAN, C,
Ada, Pascal, and many other languages do, which makes C++ safer to use.
However, there are trade-os. For example, templates impose compiletime overhead and inheritance (with dynamic type binding) may impose
run-time overhead. These features could slow down a program at compileand run-times. A good C++ compiler can minimize these overheads and a
programmer can still enjoy programming in C++ without suering noticeable compile- and run-time slowdowns but with spending possibly much
less time in code development. This is evidenced by the popularity of C++
in industry for database, graphics, and telecommunication applications,
where people also care dearly about the speed of a language. The fact is
that C++ is getting faster as people are spending more time optimizing the
compiler. Some performance comparisons on nite element analysis have
shown that C++ is comparable to FORTRAN 77 and C in terms of speed.
On the other hand, C++ has the potential of outperforming FORTRAN 77
and C for CPU-intensive numeric computations, due to its built-in arithmetic compound operators, template mechanisms that can, for example,

avoid the overhead of passing functions as arguments to other functions,

and high performance libraries (like valarray), which can be optimized on
dierent computer architectures. This potential has become a reality in
many test examples.
This book consists of three parts. The rst part (Chapters 1 to 4) is
an introduction to basic features of C++, which have equivalents in FORTRAN 90 or C. When a person knows only this part, he can do whatever a
FORTRAN 90 or C programmer can do. A brief introduction is also made
to Makefile, debuggers, making a library, and how to time and prole
a program. The second part (Chapters 5 to 9) is on object-oriented and
generic programming features of C++, which can not be found in FORTRAN 90 or C. Many techniques for writing ecient C++ programs are
also included see x6.5, x7.2, x7.6, x7.2.4, x7.7, and x8.6. It is this part that
causes many people to think that C++ is \too complicated." To make it
easy to understand for a typical science and engineering student or professional, I make use of basic concepts and operations for complex numbers,
vectors, matrices, and integrals. This should be readily acceptable by a
person with elementary knowledge of calculus and matrix algebra. The
third part (Chapters 10 and 11) provides a description of C++ standard
libraries on containers (such as linked list, set, vector, map, stack, and
queue) and algorithms (such as sorting a sequence of elements and searching an element from a sequence according to dierent comparison criteria), and an introduction to a user-dened numeric linear algebra library
(downloadable from my Web page) that contains functions for preconditioned conjugate gradient and generalized minimum residual methods for
real and complex matrices stored in full, band, and sparse formats. Gauss
elimination with and without pivoting is also included for full and band
matrices. This enhances the reader's understanding of Parts 1 and 2. Furthermore, it can save the reader a great deal of time if she has to write
her own basic numeric linear algebra library, which is used on a daily basis by many scientists and engineers. Great care has been taken so that
features that are more easily understood and that are more often used
in numeric computing are presented rst. Some examples in the book are
taken from or can be used in numeric computing libraries, while others are
made up only to easily illustrate features of the C++ language. A Web page
(http://www.math.wayne.eud/~yang/book.htm) is devoted to the book on
information such as errata and downloading programs in the book.
On the numeric side, discussions and programs are included for the following numeric methods in various chapters: polynomial evaluation (x3.12),
numeric integration techniques (x3.13, x5.1, and x7.7), vector and matrix
arithmetic (x6.3 and x7.1), iterative algorithms for solving nonlinear equations (x4.7), polynomial interpolation (x7.8), iterative and direct algorithms
for solving systems of linear equations in real and complex domains (x6.6,
x11.3, and x11.4), Euler and Runge-Kutta methods for solving ordinary
dierential equations (x5.9), and a nite dierence method for solving par-

tial dierential equations (x11.5) with real and complex coecients. The
coverage of numeric algorithms is far from being complete. It is intended
to provide a basic understanding of numeric methods and to see how C++
can be applied to program these methods eciently and elegantly. Most
of them are covered at the end of a chapter. People with an interest in
learning how to program numeric methods may read them carefully, while
readers who just want to learn C++ may wish to skip them or read them
brie y.
C++ is not a perfect language, but it seems to have all the features and
standard libraries of which a programmer can dream. My experience is that
it is much easier to program than FORTRAN and C, because FORTRAN
and C give too many run-time errors and have too few features and standard
libraries. Many such run-time errors are hard to debug but can easily be
caught by a C++ compiler.
After all, C++ is just a set of notation to most people and a person does
not have to know all the features of C++ in order to write good and useful
programs. Enjoy!

How to Use This Book:

This book can be used as a textbook or for self-study in a variety of ways.

1. The primary intent of this book is to teach C++ and numeric computing at the same time, for students and professionals in science
and engineering. C++ is rst introduced and then applied to code
numeric algorithms.
2. It can also be used for people who just want to learn basic and advanced features of C++. In this case, sections on numeric computing
can be omitted, and knowledge of calculus and linear algebra is not
quite necessary. However, in some sections, the reader should know
what a vector and a matrix are, and basic operations on vectors and
matrices, such as vector addition, scalar-vector multiplication, and
matrix-vector multiplication.
3. This book can be used as a reference for people who are learning
numeric methods, since C++ programs of many numeric methods
are included. Students who learn numeric methods for the rst time
often have diculty programming the methods. The C++ code in the
book should help them get started and have a deeper understanding
of the numeric methods.
4. For experienced C++ programmers, this book may be used as a reference. It covers essentially all features of the ISO/ANSI C++ programming language and libraries. It also contains techniques for writing
ecient C++ programs. For example, standard operator overloading for vector operations, as introduced in most C++ books, may be

a few times slower than using C or FORTRAN style programming.

Techniques are given in the book so that the C++ code is no slower
than its corresponding C or FORTRAN style code. Examples are also
given to replace certain virtual functions by static polymorphism to
improve run-time eciency. Other advanced techniques include expression templates and template metaprograms.

Acknowledgments:

This book has been used as the textbook for a one-semester undergraduate
course on C++ and numeric computing at Wayne State University. The
author would like to thank his students for valuable suggestions and discussions. Thanks also go to the anonymous referees and editors whose careful
review and suggestions have improved the quality of the book. Readers'
comments and suggestions are welcome and can be sent to the author via
email (dyang@na-net.ornl.gov).

Daoqi Yang
Wayne State University
Detroit, Michigan
June 28, 2000

xii

This is page xiii

Printer: Opaque this

Contents

Preface
1 Basic Types

1.1 A Sample Program . . . . . .

1.2 Types and Declarations . . .
1.3 Basic Types . . . . . . . . . .
1.3.1 Integer Types . . . . .
1.3.2 Character Types . . .
1.3.3 Floating Point Types .
1.3.4 The Boolean Type . .
1.3.5 The Void Type . . . .
1.4 Numeric Limits . . . . . . . .
1.5 Identiers and Keywords . . .
1.5.1 Identiers . . . . . . .
1.5.2 Keywords . . . . . . .
1.6 Exercises . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.

2.1 Scopes and Storage Classications . . .

2.1.1 Local and Global Variables . . .
2.1.2 External and Register Variables .
2.2 Expressions . . . . . . . . . . . . . . . .
2.2.1 Arithmetic Expressions . . . . .
2.2.2 Relational Expressions . . . . . .

.
.
.
.
.
.

2 Expressions and Statements

.
.
.
.
.
.
.
.
.
.
.
.
.

vii
1

1
6
8
8
10
11
14
14
15
19
19
20
21

25
26
27
28
28
29

xiv

Contents

2.2.3 Logical Expressions . . . . . .

2.2.4 Bitwise Expressions . . . . . .
2.2.5 Comma Expressions . . . . . .
2.3 Statements . . . . . . . . . . . . . . .
2.3.1 Declarations and Initializations
2.3.2 Assignments . . . . . . . . . .
2.3.3 Compound Assignments . . . .
2.3.4 Increments and Decrements . .
2.3.5 Compound Statements . . . . .
2.3.6 Conditional Statements . . . .
2.3.7 Iteration Statements . . . . . .
2.4 Fibonacci Numbers . . . . . . . . . . .
2.5 Exercises . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.

Constants and Macros . . . . . . . . . . . . .

Enumerations . . . . . . . . . . . . . . . . . .
Arrays . . . . . . . . . . . . . . . . . . . . . .
Structures . . . . . . . . . . . . . . . . . . . .
Unions and Bit Fields . . . . . . . . . . . . .
Pointers . . . . . . . . . . . . . . . . . . . . .
3.6.1 Pointer Arithmetic . . . . . . . . . . .
3.6.2 Multiple Pointers . . . . . . . . . . . .
3.6.3 Osetting Pointers . . . . . . . . . . .
3.6.4 Constant Pointers . . . . . . . . . . .
3.6.5 Void and Null Pointers . . . . . . . . .
3.6.6 Pointers to Structures . . . . . . . . .
3.6.7 Pointers to Char . . . . . . . . . . . .
3.6.8 Pointers and Arrays . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . .
Functions . . . . . . . . . . . . . . . . . . . .
3.8.1 Function Declarations and Denitions
3.8.2 Function Overloading . . . . . . . . .
3.8.3 Argument Passing . . . . . . . . . . .
3.8.4 Return Values . . . . . . . . . . . . .
3.8.5 Recursive Functions . . . . . . . . . .
3.8.6 Inline Functions . . . . . . . . . . . .
3.8.7 Default Arguments . . . . . . . . . . .
3.8.8 Function Types . . . . . . . . . . . . .
3.8.9 Static Local Variables . . . . . . . . .
3.8.10 The Function main . . . . . . . . . .
Program Execution . . . . . . . . . . . . . . .
Operator Summary and Precedence . . . . . .
Standard Library on Mathematical Functions
Polynomial Evaluation . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

3 Derived Types
3.1
3.2
3.3
3.4
3.5
3.6

3.7
3.8

3.9
3.10
3.11
3.12

.
.
.
.
.
.
.
.
.
.
.
.
.

30
31
33
33
34
35
35
36
36
37
41
44
47

51
52
54
57
57
60
64
64
67
68
70
70
72
72
72
73
73
74
75
79
81
81
82
83
84
85
86
88
92
94

Contents

3.13 Trapezoidal and Simpson's Rules . . . . . . . . . . . . . . . 98

3.14 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

4 Namespaces and Files

4.1 Namespaces . . . . . . . . . . . . . . . . . .
4.1.1 Using Declarations and Directives .
4.1.2 Multiple Interfaces . . . . . . . . . .
4.1.3 Namespace Alias . . . . . . . . . . .
4.1.4 Unnamed Namespaces . . . . . . . .
4.1.5 Name Lookup . . . . . . . . . . . . .
4.2 Include Files . . . . . . . . . . . . . . . . .
4.2.1 Include Files for Standard Libraries
4.2.2 User's Own Include Files . . . . . .
4.2.3 Conditional Include Directives . . .
4.2.4 File Inclusion . . . . . . . . . . . . .
4.3 Source Files and Linkages . . . . . . . . . .
4.3.1 Separate Compilation . . . . . . . .
4.3.2 External and Internal Linkages . . .
4.3.3 Linkage to Other Languages . . . . .
4.4 Some Useful Tools . . . . . . . . . . . . . .
4.4.1 How to Time a Program . . . . . . .
4.4.2 Compilation Options and Debuggers
4.4.3 Creating a Library . . . . . . . . . .
4.4.4 Makele . . . . . . . . . . . . . . . .
4.5 Standard Library on Strings . . . . . . . . .
4.5.1 Declarations and Initializations . . .
4.5.2 Operations . . . . . . . . . . . . . .
4.5.3 C-Style Strings . . . . . . . . . . . .
4.5.4 Input and Output . . . . . . . . . .
4.5.5 C Library on Strings . . . . . . . . .
4.6 Standard Library on Streams . . . . . . . .
4.6.1 Formatted Integer Output . . . . . .
4.6.2 Formatted Floating Point Output .
4.6.3 Output Width . . . . . . . . . . . .
4.6.4 Input and Output Files . . . . . . .
4.6.5 Input and Output of Characters . .
4.6.6 String Streams . . . . . . . . . . . .
4.7 Iterative Methods for Nonlinear Equations .
4.7.1 Bisection Method . . . . . . . . . . .
4.7.2 Newton's Method . . . . . . . . . . .
4.8 Exercises . . . . . . . . . . . . . . . . . . .

5 Classes

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

113

113
117
119
120
121
121
122
122
124
126
128
129
129
129
133
134
134
136
138
139
142
142
143
144
144
145
146
146
148
149
150
153
157
158
159
165
167

173

5.1 Class Declarations and Denitions . . . . . . . . . . . . . . 173

5.2 Copy Constructors and Copy Assignments . . . . . . . . . . 180

xvi

Contents

5.3
5.4
5.5
5.6
5.7
5.8
5.9
5.10

Friends . . . . . . . . . . . . . . . . . . . . . . . . . . .
Static Members . . . . . . . . . . . . . . . . . . . . . .
Constant and Mutable Members . . . . . . . . . . . .
Class Objects as Members . . . . . . . . . . . . . . . .
Array of Classes . . . . . . . . . . . . . . . . . . . . .
Pointers to Members . . . . . . . . . . . . . . . . . . .
Numeric Methods for Ordinary Dierential Equations
Exercises . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

7.1 Class Templates . . . . . . . . . . . . . . . . . . . . . . .

7.1.1 Member and Friend Denitions . . . . . . . . . . .
7.1.2 Template Instantiation . . . . . . . . . . . . . . . .
7.1.3 Template Parameters . . . . . . . . . . . . . . . .
7.1.4 Type Equivalence . . . . . . . . . . . . . . . . . . .
7.1.5 User-Dened Specializations . . . . . . . . . . . . .
7.1.6 Order of Specializations . . . . . . . . . . . . . . .
7.2 Function Templates . . . . . . . . . . . . . . . . . . . . .
7.2.1 Function Template Parameters . . . . . . . . . . .
7.2.2 Function Template Overloading . . . . . . . . . . .
7.2.3 Specializations . . . . . . . . . . . . . . . . . . . .
7.2.4 Class Templates as Function Template Parameters
7.2.5 Member Function Templates . . . . . . . . . . . .
7.2.6 Friend Function Templates . . . . . . . . . . . . .
7.3 Template Source Code Organization . . . . . . . . . . . .
7.4 Standard Library on Complex Numbers . . . . . . . . . .
7.5 Standard Library on valarrays . . . . . . . . . . . . . . . .
7.5.1 The Type valarray . . . . . . . . . . . . . . . . . .
7.5.2 Slice Arrays . . . . . . . . . . . . . . . . . . . . . .
7.5.3 Generalized Slice Arrays . . . . . . . . . . . . . . .
7.5.4 Mask Arrays and Indirect Arrays . . . . . . . . . .
7.6 Standard Library on Numeric Algorithms . . . . . . . . .
7.6.1 Accumulate . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

6 Operator Overloading

6.1 Complex Numbers . . . . . . . . . . . . . . . . . .

6.1.1 Initialization . . . . . . . . . . . . . . . . .
6.1.2 Default Copy Construction and Assignment
6.1.3 Conversions and Mixed-Mode Operations .
6.2 Operator Functions . . . . . . . . . . . . . . . . . .
6.3 Vectors and Matrices . . . . . . . . . . . . . . . . .
6.4 Explicit and Implicit Conversions . . . . . . . . . .
6.5 Eciency and Operator Overloading . . . . . . . .
6.6 Conjugate Gradient Algorithm . . . . . . . . . . .
6.7 Exercises . . . . . . . . . . . . . . . . . . . . . . .

7 Templates

.
.
.
.
.
.
.
.
.
.

184
185
187
189
191
193
194
199

203

204
206
206
207
207
211
220
223
226
229

231

231
233
234
235
235
235
239
240
241
242
243
245
246
247
248
250
251
252
253
256
257
258
259

Contents

7.6.2 Inner Products . . . . . . . . . . . . . . . . . . . .

7.6.3 Partial Sums . . . . . . . . . . . . . . . . . . . . .
7.6.4 Adjacent Dierences . . . . . . . . . . . . . . . . .
7.7 Ecient Techniques for Numeric Integration . . . . . . . .
7.7.1 Function Object Approach . . . . . . . . . . . . .
7.7.2 Function Pointer as Template Parameter . . . . .
7.7.3 Using Dot Products and Expression Templates . .
7.7.4 Using Dot Products and Template Metaprograms .
7.8 Polynomial Interpolation . . . . . . . . . . . . . . . . . . .
7.8.1 Lagrange Form . . . . . . . . . . . . . . . . . . . .
7.8.2 Newton Form . . . . . . . . . . . . . . . . . . . . .
7.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . .

8 Class Inheritance

8.1 Derived Classes . . . . . . . . . . . . . . . . . . . . . .

8.1.1 Member Functions . . . . . . . . . . . . . . . .
8.1.2 Constructors and Destructors . . . . . . . . . .
8.1.3 Copying . . . . . . . . . . . . . . . . . . . . . .
8.1.4 Class Hierarchy . . . . . . . . . . . . . . . . . .
8.1.5 Virtual Functions . . . . . . . . . . . . . . . . .
8.1.6 Virtual Destructors . . . . . . . . . . . . . . . .
8.2 Abstract Classes . . . . . . . . . . . . . . . . . . . . .
8.3 Access Control . . . . . . . . . . . . . . . . . . . . . .
8.3.1 Access to Members . . . . . . . . . . . . . . . .
8.3.2 Access to Base Classes . . . . . . . . . . . . . .
8.4 Multiple Inheritance . . . . . . . . . . . . . . . . . . .
8.4.1 Ambiguity Resolution . . . . . . . . . . . . . .
8.4.2 Replicated Base Classes . . . . . . . . . . . . .
8.4.3 Virtual Base Classes . . . . . . . . . . . . . . .
8.4.4 Access Control in Multiple Inheritance . . . . .
8.5 Run-Time Type Information . . . . . . . . . . . . . .
8.5.1 The dynamic cast Mechanism . . . . . . . . .
8.5.2 The typeid Mechanism . . . . . . . . . . . . . .
8.5.3 Run-Time Overhead . . . . . . . . . . . . . . .
8.6 Replacing Virtual Functions by Static Polymorphism .
8.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . .

9 Exception Handling

9.1 Throw and Catch . . . . . . . . . .

9.2 Deriving Exceptions . . . . . . . .
9.3 Catching Exceptions . . . . . . . .
9.3.1 Re-throw . . . . . . . . . .
9.3.2 Catch All Exceptions . . . .
9.3.3 Order of Handlers . . . . .
9.4 Specifying Exceptions in Functions

.
.
.
.
.
.
.

xvii

.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.

262
263
263
264
264
265
266
270
273
273
275
279

283

283
286
286
287
287
288
289
291
296
296
297
300
301
302
303
304
305
306
308
309
309
316

319

319
323
325
326
326
327
327

xviii

Contents

9.5 Standard Exceptions . . . . . . . . . . . . . . . . . . . . . . 329

9.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331

10 Standard Libraries on Containers and Algorithms

10.1 Standard Containers . . . . . . . . . . . . . . . . . .

10.1.1 Vector . . . . . . . . . . . . . . . . . . . . . .
10.1.2 List . . . . . . . . . . . . . . . . . . . . . . .
10.1.3 Map and Set . . . . . . . . . . . . . . . . . .
10.1.4 Stack and Queue . . . . . . . . . . . . . . . .
10.2 Standard Algorithms . . . . . . . . . . . . . . . . . .
10.2.1 Sorting, Copying, and Replacing Algorithms .
10.2.2 Searching and Traversing Algorithms . . . . .
10.2.3 Set, Permutation, and Heap Algorithms . . .
10.3 Standard Function Objects and Adaptors . . . . . .
10.3.1 Arithmetic Function Objects . . . . . . . . .
10.3.2 Relational Function Objects . . . . . . . . . .
10.3.3 Logical Function Objects . . . . . . . . . . .
10.3.4 Standard Adaptors . . . . . . . . . . . . . . .
10.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . .

11 Linear System Solvers

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

11.1 Matrix Storage Formats . . . . . . . . . . . . . . . . . . . .

11.1.1 Full Matrices . . . . . . . . . . . . . . . . . . . . . .
11.1.2 Band Matrices . . . . . . . . . . . . . . . . . . . . .
11.1.3 Sparse Matrices . . . . . . . . . . . . . . . . . . . . .
11.2 A Class Hierarchy for Matrices . . . . . . . . . . . . . . . .
11.3 Iterative Algorithms . . . . . . . . . . . . . . . . . . . . . .
11.3.1 Conjugate Gradient Method . . . . . . . . . . . . . .
11.3.2 Generalized Minimum Residual Method . . . . . . .
11.3.3 Preconditioning Techniques . . . . . . . . . . . . . .
11.4 Gauss Elimination . . . . . . . . . . . . . . . . . . . . . . .
11.4.1 LU Decomposition . . . . . . . . . . . . . . . . . . .
11.4.2 Gauss Elimination Without Pivoting . . . . . . . . .
11.4.3 Gauss Elimination with Pivoting . . . . . . . . . . .
11.5 Finite Dierence Method for Partial Dierential Equations
11.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . .

References
Index

333

333
334
341
344
347
348
348
355
360
365
365
366
366
367
368

371

372
372
372
374
375
385
385
390
398
401
401
406
408
414
424

427
430

This is page 1
Printer: Opaque this

Basic Types

This chapter starts with a sample C++ program and then presents basic
data types for integral and oating point types, and two special types
called bool and void. Towards the end of the chapter, numeric limits are
introduced such as the largest integer and smallest double precision number
in a particular C++ implementation on a particular computer. Finally,
identiers and keywords are discussed.

1.1 A Sample Program

A complete program in C++ must have a function called main, which
contains statements between braces f and g. Each statement can extend
to more than one line, but must end with a semicolon. Comments must
be enclosed by = and =, or preceded by == extending to the rest of the
line. The rst can be used for comments that stand on many lines while
the second for a whole line or an end part of a line. One of them can be
used to comment out the other. Comments and blank lines are ignored by
the compiler and contribute only to the readability of a program.
Below is a simple program that calculates the sum of all integers between
two given integers. The user is prompted to enter these two integers from
the standard input stream (usually the terminal screen) represented by cin.
The result is output to the standard output stream (usually the terminal
screen) represented by cout. The streams cin and cout are dened in a C++
standard library called <iostream>. The program reads:

1. Basic Types

/* A sample program to illustrate some basic features of C++.

It adds all integers between two given integers and
outputs the sum to the screen.
*/
#include <iostream>
using namespace std

// include standard library for I/O

main() {
int n, m
// declare n and m to be integers
cout << "Enter two integers: \n"
// output to screen
cin >> n >> m
// input will be assigned to n, m
if (n
int
n =
m =
}

> m) {
temp = n
m
temp

double sum = 0.0

//
//
//
//

if n is bigger than m, swap them

declare temp and initialize it
assign value of m to n
assign value of temp to m

// sum has double precision

// a loop, i changes from n to m with increment 1 each time

for (int i = n i <= m i++) { // <=: less than or equal to
sum += i
// sum += i: sum = sum + i
}
cout << "Sum of integers from " << n << " to " << m
<< " is: "<< sum << '\n' // output sum to screen
}

The rst three lines in the program are comments enclosed by = and
are usually used for multiline comments. Other comments in the
program are short ones and preceded by two slashes == they can start from
the beginning or middle of a line extending to the rest of it.
Input and output are not part of C++ and their declarations are provided
in a header le called iostream that must be included at the beginning of
the program using

=, which

#include <iostream>

with the sign # standing in the rst column. The statement

using namespace std

lets the program gain access to declarations in the namespace std. All
standard libraries are dened in a namespace called std. The mechanism
namespace enables one to group related declarations and denitions together and supports modular programming see Chapter 4 for more details.
Before standard C++, these two statements could be replaced by including

1.1 A Sample Program

<iostream:h>. The compiler will not recognize the identiers cin and cout
if the standard library <iostream> is not included. Similarly, the math
library < math:h > must be included when mathematical functions like
sin (sine), cos (cosine), atan (arctangent), exp (exponential), sqrt (square
root), and log (logarithm) are used. The header les for standard libraries
reside somewhere in the C++ system, and a programmer normally does not
have to care where. When they are included in a program, the system will
nd them automatically. See x4.2 for more information on how to include
a le.
The symbol << is called the output operator and >> the input operator.
The statement
cout << "Enter two integers: \n"

outputs the string "Enter two integers:" followed by a new line to the
screen, telling the user to type in, on the keyboard, two integers separated
by a whitespace. Then the statement
cin >> n >> m

causes the computer to pause while characters are entered from the keyboard and stored in memory at locations identied by n and m. It is equivalent to the following two statements.
cin >> n
cin >> m

// first input from screen is stored in n

// second input from screen is stored in m

Notice that a variable must be declared before it can be used. The integer
variables n and m are declared by the statement:
int n, m

// declare n and m to be integers

The character 'nn' represents a newline. A character string like "Sum of

integers from " must appear between double quotation marks, while a single
character appear between single quotation marks such as 'A', '5' and 'nn'.
Note that 'nn' is a single character. See x1.3.2 for more details on characters,
x4.5 on strings, and x4.6 on input and output streams.
The words int, double, if, and for are reserved words in C++ and can
not be used for the name of a variable. The reserved word int stands for
integers while double stands for double precision oating point numbers.
The variable sum is declared to have double precision by
double sum = 0.0

// sum is initialized to 0

This statement not only declares sum to have type double, but also assigns
it an initial value 0. A variable may not have to be initialized when it is
declared.
In the if conditional statement
if (n > m) {

// if n is bigger than m, swap them

1. Basic Types
int temp = n
n = m
m = temp

// declare temp and initialize it

// assign value of m to n
// assign value of temp to m

the statements inside the braces will be executed if the condition n > m
(n is strictly larger than m) is true and will be skipped otherwise. This if
statement instructs the computer to swap (interchange) the values of n and
m if n is larger than m. Thus, after this if statement, n stores the smaller
of the two input integers and m stores the larger. Notice that a temporary
variable temp is used to swap n and m. The braces in an if statement can
be omitted when there is only one statement inside them. For example, the
following two statements are equivalent.
if (n > m) m = n + 100
if (n > m) {
m = n + 100
}

// one statement
// an equivalent statement

The second statement above can also be written in a more compact way:
if (n > m) { m = n + 100 }

The for loop in the sample program

for (int i = n i <= m i++) {
sum += i
}

rst declares variable i of type int and assigns the value of n to i. If the value
of i is less than or equal to the value of m (i.e. i <= m), the statement inside
the braces sum += i will be executed Then statement i ++ (equivalent to
i = i + 1 here) is executed and causes the value of i to be incremented by
1. The statement sum += i (equivalent to sum = sum + i, meaning sum
is incremented by the value of i) is executed until the condition i <= m
becomes false. Thus this for loop adds all integers from n to m and stores
the result in the variable sum. The braces in a for loop can also be omitted
if there is only one statement inside them. For example, the for loop above
can also be written in a more compact way:
for (int i = n i <= m i++) sum += i

Except for possible eciency dierences, this for loop can be equivalently
written as
for (int i = n i <= m i = i + 1) sum = sum + i

The compound operators in i ++ and sum + = i can be more easily

optimized by a compiler and more ecient than i = i + 1 and sum =
sum + i. For example, an intermediate value of i +1 is usually obtained and

1.1 A Sample Program

stored and then assigned to i in i = i +1 while in i ++ such an intermediate

process could be omitted and the value of i is just incremented by 1. Notice
that += is treated as one operator and there is no whitespace between +
and =. See x2.3.7 for details on the for statement and x2.3.3 for compound
operators.
Finally the value of sum is output to the screen by the statement:
cout << "Sum of integers from " << n << " to " << m
<< " is: " << sum << '\n'
// output sum to screen

First the character string "Sum of integers from " is output, then the value
of n, and then string " to ", value of m, string " is: ", value of sum, and
nally a newline character 'nn'.
Suppose the above program is stored in a le called sample:cc. To compile
it, at the UNIX or Linux command line, type
c++ -o add sample.cc

Then the machine-executable object code is written in a le called add. If

you just type
c++ sample.cc

the object code will be automatically written in a le called a:out in UNIX
(or Linux) and sample:exe in DOS. On other operating systems, compiling
or running a program may just be a matter of clicking some buttons. Now
you may type the name of the object code, namely, add or a:out at the
UNIX command line, and input 1000 and 1. You shall see on the screen:
Enter two integers:
1000 1
Sum of integers from 1 to 1000 is: 500500

In this run, the input number 1000 is rst stored in n and 1 in m. Since
n is larger than m, the if statement interchanges the values of n and m so
that n = 1 and m = 1000. The for loop adds all integers from 1 to 1000.
Here is another run of the program:
Enter two integers:
1 1000
Sum of integers from 1 to 1000 is: 500500

In the second run, the input number 1 is stored in n and 1000 in m. Since
n is smaller than m, the if statement is skipped.
A user may not have to input values from the terminal screen. Alternatively, this sample program can also be written as
#include <iostream>
using namespace std

// include input/output library

1. Basic Types

main() {
int n = 1
int m = 1000

// declare integer n with value 1

// declare integer m with value 1000

double sum = 0
for (int i = n i <= m i++) sum += i
cout << "The sum is: " << sum << '\n'
}

or in a more compact form:

#include <iostream>
using namespace std
main() {
double sum = 0
for (int i = 1 i <= 1000 i++) sum += i
cout << "The sum is: " << sum << '\n'
}

The disadvantage of the alternative forms is that, if one wishes to add

integers from 2 to 5050, for example, then the program needs to be modied
and recompiled.
By convention, C++ les are usually postxed with :cc :c, :C , :cpp, and
the like, and C++ compilers are c ++, g ++, gcc, CC , and so on.
Notice that unlike in C or FORTRAN, declarations (like double sum)
may not have to appear at the beginning of a program and can be mixed
with other statements. The rule of thumb is that a variable should not be
declared until it is used. This should increase the readability of a program
and avoid possible misuse of variables.

1.2 Types and Declarations

Consider the mathematical formula:
z = y + f(x)

For this to make sense in C++, the identiers x f y, and z must be

suitably declared. All identiers must be declared before they can be used.
Declarations introduce entities represented by identiers and their types
(e.g. int or double) into a C++ program. The types determine what kind of
operations can be performed on them and how much storage they occupy
in computer memory. For example, the declarations
int x
float y = 3.14
double z

// x is declared to be of integer type

// y is a floating point number
// z is a floating point number

1.2 Types and Declarations

double f(int)

// f is a function taking an integer as

// its argument and returning a double

will make the above formula meaningful, where + is interpreted as adding

two numbers, = assigns the value on its right-hand side to the variable on
its left-hand side, and f (x) is interpreted as a function call returning the
function value corresponding to argument x. The rst declaration above
introduces variable x to be of type int. That is, x can only store integers.
The second declaration introduces variable y to be of type float (single
precision oating point number) and assigns an initial value 3:14 to it. A
simple function denition is:
double f(int i) {
return (i*i + (i-1)*(i-1) + (i-2)*(i-2) - 5)/3.14
}

The function f () takes integer i as input and returns a double precision

number as output. It calculates the value of the mathematical expression
(i2 + (i ; 1)2 + (i ; 2)2 ; 5)=3:14 for a given integer i. In C++, the symbols
+ ; , and = mean addition, subtraction, multiplication, and division,
respectively. All these statements can be organized into a complete C++
program:
#include <iostream>
using namespace std
double f(int i) {
// function definition
return (i*i + (i-1)*(i-1) + (i-2)*(i-2) - 5)/3.14
}
main() {
int x = 4
float y = 3.14
double z = y + f(x)
cout << "The value of z is: " << z << '\n'
}

Note that the denition of the function f () can not be put inside the
function main(). See x3.8 for more details on functions.
Basic types are discussed in the next section. Additional types including
functions and user-dened types are discussed in subsequent chapters.

1. Basic Types

1.3 Basic Types

1.3.1 Integer Types

The integer types are short int, int, and long int. An int is of natural
size for integer arithmetic on a given machine. Typically it is stored in
one machine word, which is 4 bytes (32 bits) on most workstations and
mainframes, and 2 bytes (16 bits) on many personal computers. A long int
normally occupies more bytes (thus stores integers in a wider range) and
short int fewer bytes (for integers in a smaller range).
Integers are stored in binary bit by bit in computer memory with the
leading bit representing the sign of the integer: 0 for nonnegative integers
and 1 for negative integers. For example, the nonnegative int 15 may be
represented bit by bit on a machine with 4 bytes for int as
sign bit
#

0 0000000 00000000 00000000 00001111

Note that whitespaces were inserted to easily see that it occupies 4 bytes,
and the leading (leftmost) bit 0 signies a nonnegative integer. Negative
integers are normally represented in a slightly dierent way, called two's
complement see x2.2.4. Thus a computer with 32 bits for int can only store
integers
;231 ;231 + 1 : : : ;2 ;1 0 1 2 : : : 231 ; 2 231 ; 1
while a computer with 16 bits for int can only store integers
;215 ;215 + 1 : : : ;2 ;1 0 1 2 : : : 215 ; 2 215 ; 1:
Half of them are negative and the other half are nonnegative. Note that
231 = 2147483648 and 215 = 32768. Numbers out of the given range of integers on a machine will cause integer overow. When an integer over ows,
the computation typically continues but produces incorrect results (see Exercises 1.6.9 and 2.5.14 for two examples). Thus a programmer should make
sure integer values are within the proper range. Use long int if necessary
or other techniques (see Exercise 3.14.21 where digits of a large integer
are stored in an array of integers) for very large integers. For example, a
long int may occupy 6 bytes and be able to store integers:
;247 ;247 + 1 : : : ;2 ;1 0 1 2 : : : 247 ; 2 247 ; 1:
Although the number of 8-bit bytes of storage for short int, int, and
long int is machine-dependent, it is always given by the C++ operator
sizeof . In general
sizeof (short int) sizeof (int) sizeof (long int):

1.3 Basic Types

On one machine, sizeof (int) = 4 and sizeof (long int) = 6, while on

another, sizeof (int) = sizeof (long int) = 4. However, it is guaranteed
that a short int has at least 16 bits and a long int has at least 32 bits.
A new type called unsigned int for nonnegative integers does not store
the sign bit and thus stores larger integers than the plain int, which is also
called signed int. If on a machine signed int holds integers from ;32768 to
32767, then unsigned int, occupying the same number of bits but without
storing the sign bit, will hold nonnegative integers from 0 to 65535. An
integral type (short int, int, and long int) can be signed or unsigned. The
keyword long is a synonym for long int, short for short int, signed for
signed int, and unsigned for unsigned int. An int is always signed that
is, int and signed int always mean the same thing.
To know their number of bytes on your machine, compile and run the
program:
#include <iostream>
using namespace std
main() {
cout << "number of bytes in short = "
<< sizeof(short) << '\n'
cout << "number of bytes in long int = "
<< sizeof(long) << '\n'
cout << "number of bytes in int = "
<< sizeof(int) << '\n'
cout << "number of bytes in unsigned int = "
<< sizeof(unsigned int) << '\n'
}

From now on, statements such as

#include <iostream>
using namespace std

may not be explicitly included in example programs to save space and

concentrate on more important features.
Standard conversions are performed when dierent types appear in arithmetic operations. Truncations occur when there is not enough space for
converting one type into another. For example,
int i = 321
short ii = 321
long iii = i
iii = long(i)
iii = i + ii

//
//
//
//
//
//

stored in sizeof(int) bytes

stored in sizeof(short) bytes
implicit conversion from int to long
explicit conversion from int to long
implicit conversion,
it is same as iii =long(i) +long(ii)

iii = 123456789

// a big integer

1. Basic Types

ii = short(iii)
cout << ii

// conversion from long to short

// on one machine, output of ii is -13035

Note that explicit type conversion requires the use of the name of the
type. For example, long(i) converts explicitly an int i to long. Explicit
type conversion is also called cast. Notice that when the value of long
integer iii = 123456789 is assigned to short integer ii, the space occupied
by ii (sizeof(short) bytes) may not be enough to hold the value of iii.
Truncation may occur and lead to errors. For example, on one machine
with sizeof (short) = 2, this results in ii = ;13035. The negative sign is
caused by the leading (leftmost) bit in the binary representation of ii (after
over ow), which happens to be 1.
The sux U is used to explicitly write unsigned integer constants and L
for long integer constants. For example, 5 is a signed int, 5U is an unsigned
int and 5L is a long int. They may occupy dierent numbers of bytes in
memory.
unsigned int i1 = 5U
long int i2 = 5L
int i = 5

// 5U means unsigned integer 5

// 5L means long integer 5
// 5 means signed integer 5

By default, a number is a decimal number (base 10). A number preceded

by 0 (zero) is octal (base 8) and a number preceded by 0x is hexadecimal
(base 16). The letters a b c d e, and f , or their upper case equivalents,
represent hexadecimal numbers 10 11 12 13 14, and 15, respectively. For
example, 63 is a decimal number, 077 is an octal number, and 0x3f is a
hexadecimal number. Octal and hexadecimal numbers can be conveniently
used to express bit patterns see x6.4 for an example. See x4.6 on how to
print them out.

1.3.2 Character Types

A char variable is an integral type and is of natural size to represent a

character (in English and other similar languages) appearing on a computer keyboard (it occupies one byte almost universally and is assumed
always to be one byte in this book). A char can only store an integer that
ts in one byte (thus one of 28 = 256 values) it corresponds to a character in a character set, including (American) ASCII. For example, the
integral value 98 corresponds to the character constant 'b' in the ASCII
character set. Character constants should appear between single quotation
marks like 'A', '5', 'd', 'nn' (newline character), 'nt' (horizontal tab), 'nv'
(vertical tab), 'n0' (null character), 'nn' (backslash character), 'n"' (double
quotation character), and 'n"(single quotation character). Notice the special characters above that use a backslash n as an escape character. See
Exercise 1.6.5 for a few more. For example,
char cc = 'A'

// assign character 'A' to cc.

1.3 Basic Types

cc = '\n'
int i = cc
short ii = short(cc)

//
In ASCII, cc = 65.
// assign a new value to cc.
//
In ASCII, cc = 10
// i = 10, implicit type conversion
// ii = 10, explicit type conversion

A char, occupying 8 bits, can range from 0 to 255 or from ;128 to 127,
depending on the implementation. Fortunately, C++ guarantees that a
signed char ranges at least from ;128 up to 127 and an unsigned char at
least from 0 up to 255. The types char, signed char, and unsigned char are
three distinct types and the use of char could cause portability problems
due to its implementation dependency. For example,
char c = 255
int i = c

// c has all 8 bits 1

// or: int i = int(c). Now i = ?

What is the value of i now? The answer is undened. On an SGI Challenge

machine, a char is unsigned so that i = 255. On a SUN SPARC or an IBM
PC, a char is signed so that i = ;1. A signed integer with all bits equal to 1
(including the sign bit) represents the integer ;1 in two's complement see
x2.2.4 for more details. Thus a char should be used primarily for characters,
instead of small integers.
A char is output as a character rather than as a numeric value. For
example, the program segment
char c = 'A'
int i = 'A'
cout << c << 'B' << i << "CD" << '\n'

// i = 65 in ASCII

outputs AB65CD to the screen (assuming the ASCII character set is used).
Inherited from the C programming language, a constant string always
ends with a null character. For example, the string "CD" consists of three
characters: C, D, and the null character (n0), and "nn" (notice the double
quotation marks) is a string of two characters: the newline character (nn)
and the null character (n0). In contrast, 'nn' (notice the single quotation
marks) is a single character. This can be checked by the sizeof operator:
int i = sizeof("CD")
int j = sizeof("\n")
int k = sizeof('\n')

1.3.3 Floating Point Types

// i = 3
// j = 2
// k = 1

The oating point types are float double, and long double, which correspond to single precision, double precision, and an extended double precision, respectively. The number of 8-bit bytes of storage for each of them is
given by the operator sizeof . In general
sizeof (float) sizeof (double) sizeof (long double):

1. Basic Types

On many machines, they occupy 4, 8, and 12 bytes, respectively, and a

float stores about 6 decimal places of precision, a double stores about 15

decimal places, and a long double about 18. A long double with 16 bytes

can have 33 decimal places of precision. The possible values of a oating

point type are described by precision and range. The precision describes the
number of signicant decimal digits that a oating point number carries,
while the range describes the largest and smallest positive oating values
that can be taken by a variable of that type. For example, on machines
with 4 bytes for floats, 8 bytes for doubles, and 12 bytes for long doubles,
the range of float is roughly from 10;38 to 1038, the range of double is
roughly from 10;308 to 10308 , and the range of long double from 10;4932
to 104932. This means that, on such machines, a float number f is roughly
represented in the form

f = 0:d1 d2 d6 10n
where ;38 n 38 and d1 6= 0 (here 0:d1 d2 d6 is called the fractional
part of f ) a double number d is roughly represented in the form

d = 0:d1 d2 d15 10n

where ;308 n 308 and d1 6= 0 and a long double number g in the
form
g = 0:d1 d2 d18 10n
where ;4932 n 4932 and d1 6= 0. On such machines, an overow
happens when a float number f = q 10m (0:1 q < 1) with m > 38
and an underow happens for m < ;38. Similar denitions of over ow
and under ow can be made for double and extended double precisions.
The value of a variable is often set to zero when underow occurs. The
preceding discussion is true only roughly and can be made more precise in
binary representation, which is normally used in the computer see x1.4.
The numbers of bytes given as examples above are used to store oating
point types. The IEEE (Institute for Electric and Electronic Engineers)
standards currently require that at least 80 bits be used for doing internal
computation (128 bits are used on some machines for long double calculations). When the results of a computation are stored or output, accuracy
is often lost due to roundo errors. For example, the statements
float fpi = atan(1)*4
double dpi = atan(1)*4
long double ldpi = atan(1)*4

// include <math.h>

give dierent precisions of the value . Note the function call atan(1) (for
computing the arctangent of 1) gives the value (=4) of the mathematical
function arctan(x) with x = 1. Trigonometric and other functions are in
the library <math:h> see x3.11. To see their dierence when being output

1.3 Basic Types

to the screen, we can use the precision function to specify the accuracy
and use the width function to specify the number of digits in the output
(see x4.6 for more details). For example, the statements
cout.precision(30)
cout.width(30)
cout << fpi << '\n'
cout.width(30)
cout << dpi << '\n'
cout.width(30)
cout << ldpi << '\n'

// include <iostream>
// output occupies 30 characters

output the following to the screen (on my computer).

3.1415927410125732421875
3.14159265358979311599796346854
3.14159265358979323851280895941

Compared to the exact value of : 3:14159265358979323846264338:::, the

float value fpi has 7-digit accuracy, the double value dpi has 16-digit,
while the extended double value ldpi has 19-digit accuracy. They are all
calculated using the same formula atan(1) 4. The internal calculation is
done in the same way and accuracy is lost when the result is stored in
fpi dpi, and ldpi.
The IEEE standard introduces two forms of innity (Inf or Innity in
computer output): +1 and ;1, for very large and small oating point
numbers (when it makes sense to do so). For example, x=0 and y + y give 1
for a positive oating point number x within range and the biggest oating
point number y. So do x + 1, x 1, and 1=x. Here 1 is understood to
be +1. Similar explanations hold for ;1. The standard also introduces
the entity NaN (Not-a-Number) to make debugging easier. Indeterminate
operations such as 0:0=0:0 and 1 ; 1 result in NaN . So does x + NaN .
Run the program in Exercise 1.6.12 and you will see a situation where
Innity and NaN are among the output, which can be useful debugging
information.
Variables of dierent types can be mixed in arithmetic operations and
implicit and explicit conversions can be performed. For example,
double d = 3.14
int n = 2
int m = d + n
int k = int(d) + n
double c = d + n

// m = 5, implicit type conversion

// k = 5, explicit type conversion
// c = 5.14, implicit type conversion

Note that in the addition m = d + n, integer n is rst implicitly promoted

to oating point number 2:0 and the result d + n = 5:14 is then implicitly
truncated into integer 5 for m, causing loss of accuracy.

1. Basic Types

By default, a oating point constant is of double precision. For example,

the number 0:134, 1:34E;1, 0:0134E 1, or 0:0134e1 is taken to be a double,
which occupies sizeof (double) bytes. Its oat representation is 0:134F ,
or 0:134f , suxed with F or f , which occupies sizeof (float) bytes. The
number after the letter E or e means the exponent of 10. For example,
1:34e;12 means 1:34 10;12 and 1:34e12 means 1:34 1012. This is the
so-called scientic notation of real numbers.
The reason for providing several integer types, unsigned types, and oating point types is that there are signicant dierences in memory requirements, memory access times, and computation speeds for dierent types
and that a programmer can take advantage of dierent hardware characteristics. The type int is most often used for integers and double for oating
point numbers. Despite the fact that float usually requires less memory
storage than double, arithmetics in float may not always be signicantly
or noticeably faster than in double for a particular C or C++ compiler,
especially on modern computers see x4.4.1 for an example and some explanations.

1.3.4 The Boolean Type

A Boolean, represented by the keyword bool, can have one of the two values:
true and false, and is used to express the results of logical expressions
(x2.2.3). For example,
bool flag = true

// declare flag to be of bool

// ... some other code that might change the value of flag
double d = 3.14
if (flag == false) d = 2.718

The operator == tests for equality of two quantities. The last statement
above means that, if flag is equal to false, assign 2:718 to variable d.
By denition, true has the value 1 and false has the value 0 when
converted to an integer. Conversely, nonzero integers can be implicitly converted to true and 0 to false. A bool variable occupies at least as much
space as a char. For example,
bool b = 7
int i = true
int m = b + i

// bool(7) is converted to true, so b = true

// int(true) is converted to 1, so i = 1
// m = 1 + 1 = 2

1.3.5 The Void Type

A type that has no type at all is denoted by void. It is syntactically a

fundamental type, but can be used only as part of a more complicated
type. It is used either to specify that a function does not return a value or

1.4 Numeric Limits

as the base type for pointers to objects of unknown type. These points are
explained later. See, for example, x3.8 and Exercise 3.14.24.

1.4 Numeric Limits

Machine-dependent aspects of a C++ implementation can be found in
the standard library <limits>. For example, the function numeric limits
<double> :: max() gives the largest double and function numeric limits
<int> :: min() gives the smallest int that can be represented on a given
computer. The library <limits> denes numeric limits <T > as a template class (see Chapter 7) that has a type parameter T , where T can be
float, double, long double, int, short int, long int, char, and unsigned
integers. When T is, for example, float, then numeric limits <float>
gives implementation-dependent numbers for float. The following program
prints out information on float.
#include <iostream>
#include <limits>
using namespace std
main () {
cout <<
<<
cout <<
<<
cout <<
<<
cout <<
<<
cout <<
<<
cout <<
<<
cout <<
<<
cout <<
<<
cout <<
<<
cout <<
<<
cout <<
<<
cout <<

"largest float = "

numeric_limits<float>::max() << '\n'
"smallest float = "
numeric_limits<float>::min() << '\n'
"min exponent in binary = "
numeric_limits<float>::min_exponent << '\n'
"min exponent in decimal = "
numeric_limits<float>::min_exponent10 << '\n'
"max exponent in binary = "
numeric_limits<float>::max_exponent << '\n'
"max exponent in decimal = "
numeric_limits<float>::max_exponent10 << '\n'
"# of binary digits in mantissa: "
numeric_limits<float>::digits << '\n'
"# of decimal digits in mantissa: "
numeric_limits<float>::digits10 << '\n'
"base of exponent in float: "
numeric_limits<float>::radix << '\n'
"infinity in float: "
numeric_limits<float>::infinity() << '\n'
"float epsilon = "
numeric_limits<float>::epsilon() << '\n'
"float rounding error = "

1. Basic Types
<< numeric_limits<float>::round_error() << '\n'
cout << "float rounding style = "
<< numeric_limits<float>::round_style << '\n'

Similarly, implementation-dependent information on double and long double may be obtained as

main() {
double smallestDouble = numeric_limits<double>::min()
double doubleEps = numeric_limits<double>::epsilon()
long double largestLongDouble =
numeric_limits<long double>::max()
long double longDoubleEpsilon =
numeric_limits<long double>::epsilon()
}

and information on char may be obtained as

main() {
cout <<
<<
cout <<
<<
cout <<
<<
cout <<
<<
cout <<
<<
}

"number of digits in char: "

numeric_limits<char>::digits << '\n'
"char is signed or not: "
numeric_limits<char>::is_signed << '\n'
"smallest char: "
numeric_limits<char>::min() << '\n'
"biggest char: "
numeric_limits<char>::max() << '\n'
"is char an integral type: "
numeric_limits<char>::is_integer << '\n'

Some explanations are given now to the terms epsilon and mantissa. A
machine epsilon is the smallest positive oating point number such that
1 + epsilon 6= 1. That is, any positive number smaller than it will be
treated as zero when added to 1 in the computer. It is also called the unit
roundo error. When a oating point number x is represented in binary
as x = q 2m , where q = 0:q1 q2 qn with q1 = 1 and qi = 0 or 1 for
i = 2 3, : : : n, and m is an integer, then q is called the mantissa of x, m
is called the exponent of x, and n is the number of bits in the mantissa.
Due to nite precision in computer representation, not all numbers, even
within the range of the computer, can be represented exactly. A number
that can be represented exactly on a computer is called a machine number.
When a computer can not represent a number exactly, rounding, chopping,
over ow, or under ow may occur.
On a hypothetical computer called Marc-32 (see !CK99, KC96] a typical
computer should be very similar to it if not exactly the same), one machine

1.4 Numeric Limits

word (32 bits) is used to represent a single precision oating point number:

x = 1:b1b2 b23 2m :
The leftmost bit in the machine word is the sign bit (0 for nonnegative and
1 for negative), the next 8 bits are for the exponent m, and the rightmost
23 bits for the fractional part (b1 , b2 , : : :, and b23 ) of the mantissa. Notice
that, to save one bit of memory, the leading bit 1 in the fractional part is
not stored. It is usually called the hidden bit. The exponent m takes values
in the closed interval !;127 128]. However, m = ;127 is reserved for 0,
and m = 128 for 1 (if b1 = b2 = = b23 = 1) and NaN (otherwise).
Thus the exponent of a nonzero machine number must be in the range
;126 m 127. On such a machine, the single precision machine epsilon
is then 2;23
1:2 10;7 , the smallest positive machine number is 2;126

1:2 10;38 , and the largest machine number is (2 ; 2;23 ) 2127

3:4 1038 (corresponding to bi = 1 for i = 1 2, : : : 23, and m = 127).

In double precision, there are 52 bits allocated for the fractional part of
the mantissa and 11 bits for the exponent, which results in the machine
epsilon to be 2;52
2:22 10;16, the smallest positive machine number
2;1022
2:25 10;308, and the largest machine number (2 ; 2;52) 21023

1:798 10308 .
It can be shown (see !CK99, KC96]) that for any nonzero real number x
and its oating point machine representation x# (assuming x is within the
range of the computer and thus no over ow or under ow occurs), there
holds

x ; x
# epsilon

in the case of chopping, and

x ; x# 1 epsilon
x 2

in the case of rounding to the nearest machine number. That is, the relative
roundo error in representing a real number in the range of a computer with
a particular precision is no larger than epsilon in that precision.
The machine epsilon and number of bits in binary (or equivalent digits
in decimal) for representing the mantissa of a oating point number on a
particular computer are given in the template class numeric limits< T>
for a oating point type T .
In addition, C library <limits:h> has macros (see x3.1 for some rules
and examples for dening macros) such as those listed in Table 1.1, and
C library <float:h> has macros such as those listed in Table 1.2. These
macros are better avoided. However, the C++ library <limits> may not
be available on early (nonstandard) compilers. In this case, numeric limits
can be obtained as in the following program.

1. Basic Types

INT MAX
INT MIN
LONG MAX
LONG MIN
ULONG MAX
UINT MAX
SHRT MAX
USHRT MAX
SCHAR MIN
UCHAR MAX
CHAR MAX
CHAR MIN
WORD BIT
CHAR BIT

largest int
smallest int
largest long int
smallest long int
largest unsigned long int
largest unsigned int
largest short int
largest unsigned short int
smallest signed char
largest unsigned char
largest char
smallest char
number of bits in one word
number of bits in char
TABLE 1.1. Limits of integral numbers from C library <limits.h>.

DBL MAX
DBL MIN
DBL EPSILON
DBL MANT DIG
DBL DIG
DBL MAX 10 EXP
LDBL MAX
LDBL MIN
LDBL EPSILON
LDBL MANT DIG
LDBL DIG
LDBL MAX 10 EXP
FLT MAX
FLT MIN
FLT EPSILON
FLT MANT DIG
FLT DIG
FLT MAX 10 EXP

largest double
smallest double
double epsilon
number of binary bits in mantissa
number of decimal digits in mantissa
largest exponent
largest long double
smallest long double
long double epsilon
number of binary bits in mantissa
number of decimal digits in mantissa
largest exponent
largest float
smallest float
float epsilon
number of binary bits in mantissa
number of decimal digits in mantissa
largest exponent
TABLE 1.2. Limits of oating point numbers from C library <oat.h>.

1.5 Identi ers and Keywords

#include <limits.h>
#include <float.h>
main() {
int i = INT_MIN
long j = LONG_MAX
double x = DBL_MAX
long double y = LDBL_MAX
float z = FLT_MAX
double epsdbl = DBL_EPSILON
float epsflt = FLT_EPSILON
long double epsldb = LDBL_EPSILON
}

//
//
//
//
//
//
//
//

smallest int
largest long int
biggest double
biggest long double
biggest float
double epsilon
float epsilon
long double epsilon

For headers ending with the :h sux such as float:h and math:h, the
include directive
#include <float.h>

automatically gives access to declarations in float:h, and the statement

using namespace std

is not necessary.
Note that the C++ library <limits> is dierent from the C library
<limits:h>, which C++ inherited from C. All C libraries can be called from
a C++ program provided that their headers are appropriately included. See
x4.2 for more details on all C and C++ standard header les.

1.5 Identiers and Keywords

1.5.1 Identiers

In our rst program, we have used variables such as sum and m to hold
values of dierent data types. The name of a variable must be a valid
identier. An identier in C++ is a sequence of letters, digits, and the
underscore character . A letter or underscore must be the rst character
of an identier. Upper and lower case letters are treated as being distinct.
Some identiers are:
double sum = 0
double product
bool flag

// identifier sum suggests a summation

// identifier product suggests multiply
// flag certain condition (true or false)

int Count
int count
// different from Count
int this_unusually_long_identifier // legal, but too long

1. Basic Types

and
bitand
case
compl
default
dynamic cast
export
for
inline
namespace
operator
protected
return
static
template
try
union
void
xor

int _aAbBcCdD

and eq
bitor
catch
const
delete
else
extern
friend
int
new
or
public
short
static cast
this
typedef
unsigned
volatile
xor eq

asm
bool
char
const cast
do
enum
false
goto
long
not
or eq
register
signed
struct
throw
typeid
using
wchar t

auto
break
class
continue
double
explicit
oat
if
mutable
not eq
private
reinterpret cast
sizeof
switch
true
typename
virtual
while

TABLE 1.3. All C++ keywords.

// legal, but no meaning

It is good programming practice to choose identiers that have mnemonic

signicance so that they contribute to the readability and documentation
of a program. Confusing identiers should be avoided. For example, l1,
ll, lo, and lO are dierent and valid identiers but are hard to read, and
identiers Count and count can be easily misunderstood.

1.5.2 Keywords

Keywords such as int, double, and for are explicitly reserved identiers that
have a strict meaning in C++. They can not be redened or used in other
contexts. For example, a programmer can not declare a variable with the
name double. A keyword is also called a reserved word. A complete list of
C++ keywords is in Table 1.3.
Note that cin and cout, for example, are not keywords. They are part of
the input and output (I/O) library <iostream>.

1.6 Exercises

1.6 Exercises
1.6.1. Modify the program in x1.1 to compute the sum of the squares of
all integers between two given integers. That is, nd the sum n2 +
(n + 1)2 + + m2 for two given integers n and m with n < m. In
the sample program in x1.1, the variable sum should be declared as
double or long double in order to handle large values of n and m. In
this exercise, try to compute sum in two ways: as a long double and
as an int, and compare the results. On a computer with 4 bytes for
storing int, the second way calculates the sum 12 +22 +32 + +50002
as ;1270505460. Why could a negative number as the output be
possible?
1.6.2. Modify the program in x1.1 to multiply all integers between two given
small positive (e.g., less than or equal to 12) integers. When one of
them is 1 and the other is a positive integer n, the program should
nd the factorial n! = 1 2 3 (n ; 1) n. The program to nd
n! may be written as
#include <iostream>
using namespace std
main() {
int n
cout << "Enter a positive integer: \n"
cin >> n
int fac = 1
for (int i = 2 i <= n i++) fac *= i
cout << n << "! is: " << fac << '\n'

// fac = fac*i

Except for possible eciency dierence, the statement fac = i is

equivalent to fac = fac i.
This simple program can only compute n! correctly for n = 1 up
to n = 12 on a computer with sizeof (int) = 4. When n = 13,
such a computer may calculate 13! as 1932053504 while the correct
value is 13! = 6227020800. It may also compute 20! as ;2102132736
(why negative?) and 40! as 0 (why just 0?). A computer can produce
garbage very easily. Test what your computer gives you for 13! 20!,
and 40!. A user should check the correctness of computer outputs by
all means. Sometimes outputs such as the erroneous result for 13!
above can be very hard to check. Outline a procedure to determine
the correctness of your computer output of 13!, assuming you do not
know the correct value.

1. Basic Types

In Exercise 3.14.21, a technique is introduced that can compute n!

correctly for much larger n, for example, n = 3000 or larger. Notice
that the number 3000! has 9131 digits that would over ow as an
integer on any current computer.
1.6.3. If one wishes to compute the following summation
sin(1:1) + sin(1:3) + sin(1:5) + + sin(9:9)
the program in x1.1 can be modied to do so:
#include <iostream>
#include <math.h>

// input/output library
// math library for sin

main() {
double sum = 0

// sum initialized to 0

for (double d = 1.1 d <= 9.9 d += 0.2) sum += sin(d)

cout << "The sum is: " << sum << '\n'
}

This for loop declares d to be a variable of double precision with

initial value 1:1, and executes the statement sum += sin(d) for d
changing from 1:1 up to 9:9 with increment 0:2 each time. Compile
and run this program, and modify it to compute

e1:1 + e1:2 + e1:3 + + e15:5

where e is the base of the natural logarithm. See x3.11 for a complete
list of mathematical functions in the library <math:h>.
1.6.4. Write a program that outputs the largest integer, smallest integer,
and the number of bytes for storing an integer on your computer.
Also do this with short int, int, and long int.
1.6.5. Write a program that outputs the integer values corresponding to
characters A, 9, a, f, $, nn (new line), nt (horizontal tab), n0 (null
character), nn (backslash), nr (carriage return), n" (double quote),
nb (backspace), nf (formfeed), n' (single quote), nv (vertical tab), n?
(question mark), and na (alert) on your computer.
1.6.6. Write a program that outputs exactly the sentences:
He said: \I can output double quotation marks."
She asked: \Do you know how to output the newline character nn ?"
Notice that the double quotation mark, newline, and question mark
are special characters in C++. See Exercise 1.6.5

1.6 Exercises

1.6.7. A backslash symbol n at the end of a line is supposed to continue it

to the next line. Test the following program to see its eect.
#include <iostream>
using namespace std
main() {
cout << "I am outputting a string that stands on \
three lines to test the effect of a continuation line \
using a backslash\n"
}

1.6.8. Write a program that outputs the largest and smallest numbers,
epsilon, the number of decimal digits used to store the mantissa,
and the largest exponent for double precision oating point numbers
on your computer. Repeat this for float and long double as well.
1.6.9. Compile and run the program
#include <iostream>
using namespace std
main() {
long g = 12345678912345
short h = g
// beware of integer overflow
int i = g - h
cout << "long int g = " << g << '\n'
cout << "short int h = " << h << '\n'
cout << "their difference g - h = " << g - h << '\n'
}

on your computer. Does your compiler warn you about integer overow? It may not always do so. Does your computer give g ; h = 0?
Beware of over ow and truncation in converting a larger type to a
smaller one.
1.6.10. Calculate the value of on your computer following the steps in x1.3.3
in single, double, and extended double precisions. How many digits of
accuracy does your computer give in each of the three oating point
types?
1.6.11. What is the value of i in the following statement?
int i = 3.8 + 3.8

Is it 6, 7, or 8? Type conversion may lead to loss of accuracy. Does

your compiler warn you about this? It may not always do so.

1. Basic Types

1.6.12. What do you think the following program will output?

#include <iostream>
#include <float.h>
int main() {
double x = DBL_MAX
double epsilon = DBL_EPSILON
double zero = 0.0
double y = 100.2
double
double
double
double
double
double
double
cout
cout
cout
cout
cout
cout
cout
cout
cout

z0
z1
z2
z3
z4
z5
z6

<<
<<
<<
<<
<<
<<
<<
<<
<<

=
=
=
=
=
=
=

// biggest double
// double epsilon

x + x
x * 2
epsilon/9
y/zero
zero/zero
z3 - z3
x + y

"outputting results:\n"
z0 << '\n'
z1 << '\n'
z2 << '\n'
z3 << '\n'
z4 << '\n'
z5 << '\n'
z6 << '\n'
1 + z2 << '\n'

Run the program on your computer to check the results. You may
see Innity, NaN, and other unusual outputs.

This is page 25
Printer: Opaque this

Expressions and Statements

The code segment

int n
int m = 10
n = m + 5

contains three statements. The rst one is a declaration (it declares n to be

an integer i.e., it informs the compiler that n is a variable of type int and
the compiler will allocate appropriate memory space for it), the second one
is a declaration together with an initialization (it declares m to be an integer
and initializes it to 10), and the third statement is an assignment (it assigns
the value of m plus 5 to n) that contains an expression m +5 with operands
m and 5 and operator +: A declaration is a statement that introduces a
name associated with a type into the program. All statements must end
with a semicolon. In this chapter, various expressions and statements are
discussed, after starting with scopes and storage classications. The chapter
nishes by providing a sample program on calculating Fibonacci numbers.

2.1 Scopes and Storage Classications

A scope is part of a program text consisting of statements enclosed by a pair
of braces, except for the global scope that is outside all braces. Variables
dened in the global scope are called global variables, and those dened in
local scopes are called local variables. This section deals with scopes and
dierent kinds of variables.

2. Expressions and Statements

2.1.1 Local and Global Variables

Unlike C and FORTRAN, a declaration is a statement that can occur anywhere another statement can, not necessarily at the beginning of a program.
If it occurs within a block enclosed by a pair of braces, it then declares a
local variable to the block. This block introduces a scope: Variables local to
a scope are inaccessible outside the block dening the scope. For example,
int n
{
int m = 10
}
n = m
int m = 20

// this pair of braces introduces a scope

// m is accessible only within this scope
// error, m is out of scope
// OK, it declares a new variable m

The local variable m = 10 is not accessible outside the scope in which it is

dened. Thus the assignment n = m is illegal and the compiler treats m
as not having been declared. However, it is legal to dene another variable
m = 20 after that scope. Variables can have the same name if they are in
dierent scopes.
A variable declared before a block can be accessed after and within the
block, and within all inner blocks, provided it is not hidden by another
variable declared within such a block and having the same name. A variable
declared outside all blocks is said to be a global variable. When a local
variable and a global variable have the same name, the scope resolution
operator :: can be used to access the global variable. For example,
int x = 10
main() {
int x = 25
int y = ::x
{
int z = x
int x = 38
int t = ::x
t = x
}
int z = x
}

// this x is global (outside all blocks)

// local variable x, it hides global x
// y = 10, ::x refers to global variable x
//
//
//
//

z = 25 x is taken from previous scope

it hides local variable x = 25 above
t = 10, ::x refers to global variable x
t = 38, x is treated as local variable

// z = 25, z has same scope as y

There is no way to access a hidden local variable. For example, the local
variable x = 25 is hidden and inaccessible from the scope where integer t
is dened. Note that the local variables x = 25 and y have the same scope
as z , the local variable introduced in the last statement of the program.
An automatic variable is created at the time of its declaration and destroyed at the end of the block in which it is declared, as the variables y
t and z above. When the block is entered, the system sets aside memory

2.1 Scopes and Storage Classi cations

for automatic variables. When the block is exited, the system no longer
reserves the memory for these variables and thus their values are lost. A
local variable is automatic by default. It can also be declared explicitly
by the keyword auto: For example, the declaration int x = 38 above is
equivalent to: auto int x = 38 :

2.1.2 External and Register Variables

A variable dened in one le can be accessed in another le by declaring it

to be external using the keyword extern:
extern double x

// x can be accessed across files

This declaration tells the system that variable x is dened externally, perhaps in another le (see x4.3.2 for more details). However, it does not
allocate space for the external variable x and simply means that x is accessible in all les that have the statement extern double x : It must be
declared as usual in one and only one le:
double x

// declare it as usual in one of the files

It is this statement that creates an object and allocates space for x: (See
x4.4.2 on how to compile a program consisting of more than one le.)
External variables never disappear they exist throughout the execution life
of the program. To improve the modularity and readability of a program,
the use of external variables should be kept to the minimum. In particular,
a function (see x3.8) should not change an external variable within its body
rather than through its parameter list or return value. An external variable
can sometimes be localized:
{
// some code
{
extern double x
double y = x
}

// x is restricted to this scope only

The declaration of a register variable indicates that it is stored in high

speed registers. Since resource limitations and semantic constraints sometimes make this impossible, this storage classication defaults to automatic
whenever the compiler can not allocate an appropriate physical register.
The use of register variables is an attempt to improve execution speed
and the programmer may declare a few most frequently used variables as
register: For example,
for (register int i = 0 i < 10000 i++) {
// ... do something

2. Expressions and Statements

However, modern compilers can normally do a better job than a programmer as to which variable should be assigned to a register.
Thus every variable has two attributes: type and storage classication.
Three storage classications have been talked about in this section. They
are automatic, external, and register, which are represented by keywords
auto, extern, and register. How these variables are stored distinguishes them
from each other. The use of C-style static storage classication, dening
a variable local to only one le, is discouraged in C++, since its use is
subtle and the keyword static is overused in C++. However, C++ provides
namespaces for dening variables local to a le see x4.1.4.

2.2 Expressions
If we dene int x = 25 int y = 10 then x ; y is called an expression
having operands x and y and operator ; the value of which is 15: In this
case, ; is called a binary operator since it takes two operands. It can also
be used as a unary operator as in x = ;5 which takes only one operand.
Dierent expressions are introduced in this section.

2.2.1 Arithmetic Expressions

Arithmetic expressions contain arithmetic operations for addition, subtraction, multiplication, and division, whose corresponding operators are
+ ; and = respectively. + and ; are also unary operators. These operators are dened for all integer and oating point types and there are
appropriate type conversions between dierent types. The reminder of one
integer divided by another is given by the modulo operator %: Note that
the division of one integer by another has an integer value in C++, where
the reminder (zero or not) is ignored. For example,
double m = 13/4
double n = 3/4
double x = double(3)/4
double y = 3.0/4
double z = 3/4.0
int t = 13%4

//
//
//
//
//
//

m
n
x
y
z
t

=
=
=
=
=
=

3, reminder is ignored
0, it just gives quotient
0.75, 3 is converted to 3.0
0.75
0.75
1, reminder of 13/4 is 1

Left to right associativity holds for arithmetic expressions. Multiplication

and division have higher precedence than addition and subtraction, and
parentheses override this precedence as in mathematical expressions. For
example,
m = 1 + 2 + 3*4

// m = 3 + 12 = 15

2.2 Expressions
m = 1 + (2 + 3)*4
m = exp1 + (exp2 - exp3)

// m = 1 + 5*4 = 21

In general, the following two statements are equivalent,

m = exp1 + exp2 + exp3
m = (exp1 + exp2) + exp3

// left to right associativity

although it is not guaranteed that exp1 or exp2 is evaluated before the

expression exp3 is evaluated. That is, associativity and precedence only
group subexpressions together, but do not imply which subexpression is
evaluated rst in time.

2.2.2 Relational Expressions

Relational operators are > (bigger than), < (smaller than), >= (bigger
than or equal to), <= (smaller than or equal to), == (equal), and != (not

equal). The notation ! is the negation operator. The result of evaluating

these operators is of type bool it is either true or false: For example,
int x = 5
bool b = (x < 6)
bool c = (x == 0)
if (b != c) x = 17
if (b) x = 19
if (!b) x = 221

//
//
//
//
//

b = true, since x is less than 6

c = false, since x is not 0
if b is not equal to c, let x = 17
if b is true, then do x = 19
if b is false, then do x = 221

In general, a zero quantity is false and a nonzero quantity is true: Care

must be taken when testing for equality of two oating point numbers
owing to nite precision in computer representation and roundo errors
see Exercise 2.5.10. Instead of doing so directly, one usually checks if the
absolute value of their dierence is less than a small number. Notice that
m = n is an assignment while m == n is a relational expression.
Mathematical and C++ expressions may sometimes have surprisingly
dierent interpretations. For example, the mathematical expression
3<m<9
is true for m = 7 and is false for m = 20: However, consider the C++ code:
int m = 7
bool b = 3 < m < 9

// b = true

m = 20
bool c = 3 < m < 9

// c = true

Due to the left to right associativity for relational operators, the C++
expression 3 < m < 9 is equivalent to (3 < m) < 9: Thus the C++

2. Expressions and Statements

expression 3 < m < 9 always evaluates to true because 3 < m is evaluated

to either 1 (true) or 0 (false).
Another example is that the mathematical expression

x <x+y
is always true for any positive real number y and any real number x while
the C++ expression
x < x + y

evaluates to false for some large value of x (e.g., x = 5e+31 in 12-byte long
double precision) and small value of y (e.g., y = 0.1) since, in this case, x
and x + y are treated as being equal (due to nite precision and roundo
error) on the computer. Similarly, the expression x + y > x ; y is always
true for any positive numbers x and y in mathematics, but not in C++.

2.2.3 Logical Expressions

Logical operators are && (and) and jj (or). A logical expression has value
of type bool: If exp1 and exp2 are two expressions, then exp1 && exp2 has
value true if both exp1 and exp2 are true and has value false otherwise
and exp1 jj exp2 has value true if at least one of exp1 and exp2 is true and
has value false otherwise. This can be more easily illustrated by using the
so-called truth table.
true
true
false
false

&&
&&
&&
&&

true
false
true
false

= true
true
= false true
= false false
= false false

jj
jj
jj
jj

true
false
true
false

= true
= true
= true
= false

Some examples of logical expressions are:

bool
bool
bool
bool
bool
bool

b
d
e
f
g
h

=
=
=
=
=
=

true, c = false
b && c
b || c
(e == false) && c
(d == true) || c
(d = true) || c

//
//
//
//
//

d
e
f
g
h

=
=
=
=
=

false
true
false
false
true

Note that in the last statement above, d is assigned to be true in d = true

and thus (d = true) jj c = true jj false = true: Again, d == true is a
relational expression while d = true is an assignment.
The logical expression
i >= 0 && i < n

2.2 Expressions

tests if i is bigger than or equal to 0 and less than n: If i is less than 0,

then the value of the logical expression is false no matter what the value
of i < n is, and the subexpression i < n is skipped without evaluation.
This is the so-called short-circuit evaluation and should lead to a faster
evaluation of such logical expressions. It also applies to the logical or operator jj: It is guaranteed that evaluation is from left to right and the process
stops as soon as the outcome true or false is known. That is, in this example, the subexpression on the left i >= 0 is evaluated rst and then the
subexpression i < n may or may not be evaluated depending on the value
of the left subexpression. This can be more easily seen from the following
examples.
int i,
bool k
bool m
bool n

j
=
=
=

= 2
((i = 0) && (j = 3))
((i = 4) || (j = 5))
((i = 0) || (j = 6))

// i = 0, k = false, j = 2
// i = 4, m = true, j = 2
// i = 0, j = 6, n = true

In the second statement above, i is assigned to be 0 and the left subexpression (i = 0) is converted to false. Thus k = false no matter what value
the right subexpression (j = 3) has, and the right subexpression (j = 3)
is skipped without evaluation. In the third statement, i is assigned to be
4 and (i = 4) is converted to true and the right subexpression (j = 5) is
skipped.
Notice that this kind of left to right evaluation is not true for arithmetic
expressions. For example, in m = exp1 + exp2 it is not guaranteed that
exp1 is evaluated before exp2:
Referring to an example in x2.2.2, the correct way of translating the
mathematical expression 3 < m < 9 into C++ is
(3 < m) && (m < 9)

Since operator < has higher precedence than operator && the two pairs
of parentheses above may be omitted.
It is dened that the operator && has higher precedence than jj: For
example, the following two statements are equivalent,
bool m = i || j && k
bool m = i || (j && k)

although x = i jj j jj k is equivalent to x = (i jj j ) jj k (left to right

associativity). Parentheses can override precedence and make code more
readable.
A summary on precedence for all C++ operators is given in x3.10.

2.2.4 Bitwise Expressions

Bitwise operators are & (and), j (or), ^ (exclusive or), ~ (complement),

<< (left shift), and >> (right shift), which can be applied to objects of

2. Expressions and Statements

integral type (bool char short int long and their unsigned counterparts).
In these expressions, operands are rst implicitly represented in binary
(string of binary bits) and then operation is performed on them bit by bit.
Bitwise operations can be machine-dependent and may not be portable,
since, for example, dierent machines may represent integers dierently
and use dierent bytes for storing them.
Binary bitwise operations & j and^are performed by comparing their
operands bit by bit. If two bits are both 1, then & and j give 1 and ^ gives
0: If two bits are both 0, then & ^ and j all give 0: If one bit is 0 and
the other is 1, then j and^give 1 and & gives 0: This can be more easily
illustrated by the following table.
1&1 =1 1&0 =0 0&1 =0 0&0 =0
1 j 1 =1 1 j 0 =1 0 j 1 =1 0k 0 =0
1 ^1 =0 1 ^0 =1 0 ^1 =1 0 ^0 =0
For example, 13 and 7 are in binary 0 01101 and 0 00111 respectively.
Then
int a = 13 & 7

// & is performed bit by bit

denes a to have a binary representation 0 00101 which is obtained

by applying the bitwise & operator to corresponding bits of the binary
representations of 13 and 7. This gives a decimal number 22 + 20 = 5.
Similarly, the statements
int b = 13 ^ 7
int c = 13 | 7

// b = 10
// c = 15

dene b to have a binary representation 0 01010 which is 23 + 2 = 10

and c a binary representation 0 01111 which is 23 + 22 + 21 + 20 = 15:
If n and s are integers, then n << s shifts the pattern of bits of n s
places to the left, putting s zeros in the rightmost places. On the other
hand, n >> s shifts the pattern s places to the right, putting s zeros in
the leftmost places for unsigned integers. (For signed integers, s zeros or
s ones may be shifted in depending on the machine.) For example, for an
unsigned int n = 16 consider the statements
int m = n << 2
int k = n >> 2

// m = 64
// k = 4

Since 16 = 24 has a binary representation 0 010000, left shifting its bits

by 2 positions leads to a binary pattern 0 01000000 which is 26 = 64
and right shifting its bits by 2 positions leads to 0 0100 which is 22 = 4:
Note that << and >> are also input and output operators and they can be
easily distinguished from the context.
The bitwise complement operator~is unary and changes every 0 into 1
and every 1 into 0 in the binary representation of its operand. For example,

2.3 Statements

the integer 0 has all bits equal to 0 in its binary representation and~0 has
all bits equal to 1:
The value of ~n is called one's complement of n. Two's complement of
an integer n is the value of adding 1 to its one's complement. On a two's
complement machine, a nonnegative integer n is represented by its bit representation and a negative integer ;n is represented by the two's complement of n: For example, 7 is represented by the binary string 0000 0111
and ;7 is represented by 1111 1001 which is obtained from adding 1 to
1111 1000 the complement of 7.
It can be checked that the two's complement of a negative integer ;n
has value n: For example, ;7 is represented by 1111 1001: Its two's complement is 0000 0111 which is 7: In two's complement, the integer 0 has
all bits o and ;1 has all bits on. Notice that the leftmost bit represents
the sign of the integer: 1 for negative and 0 for nonnegative. On a two's
complement machine, the hardware that does addition can also do subtraction. For example, the operation a ; b is the same as a + (;b) where ;b is
obtained by taking b's two's complement.
In Exercise 2.5.15, a program is given to print out the bit representation
of any integer on a computer.

2.2.5 Comma Expressions

The expression

i = 3, i + 2

is called a comma expression with a comma used to separate subexpressions. Its value is the value of the last subexpression, and in this example
its value is i + 2 = 5: A comma expression is guaranteed to be evaluated
from left to right. Thus, the subexpression i = 3 is evaluated rst and then
the subexpression i + 2: Consequently, the statement
j = (i = 3, i + 2)

gives j = 5:

2.3 Statements
A program consists of a collection of statements. Each statement must
end with a semicolon. The null statement is simply a semicolon, which is
used when the syntax requires a statement but no action is desired. In this
section, we deal with various statements: declarations, assignments, compound statements, conditional statements, iteration statements (loops),
jump statements, and selection statements. Later we encounter return statements and function calls (see Chapter 3).

2. Expressions and Statements

2.3.1 Declarations and Initializations

The declaration
int n

declares the variable n to be an integer and reserves for it sizeof (int) bytes
of addressable memory, the content of which is in general undened. The
system may initialize n with a default value in some situations, but it is
safe to always assume the value of n is undened. To give it a meaningful
value, an assignment such as
n = 10

can be used. A variable can be declared and initialized in one statement

such as
int n = 10

This is called an initialization: All variables must be declared before they

are used. One declaration statement can declare several variables of the
same type, separated by commas, such as
double x, y, z

// declare x, y, z in one declaration

It is equivalent to declaring them separately.

The particular block of memory reserved for a variable is sometimes
called an object. For example, when we declare an integer variable n:
int n = 10

a certain contiguous region of memory (sizeof (int) bytes) is allocated for

it. This particular contiguous region of memory is called an object. The
name of this object is n. There are two numbers associated with the object
n: its value (the value stored in the block of memory), written as n, and its
location or address of the object in memory, written as &n: The symbol &
is called the address-of operator. It gives the address of an object. We can
change the value of object n through the assignment: n = 55. It simply
stores the new value 55 at the same memory location. This can be checked
by the following program.
main() {
int n = 10
cout << "The value of variable n is: " << n << '\n'
cout << "The location in memory of variable n is: " << &n
n = 55
cout << "The value of variable n is: " << n << '\n'
cout << "The location in memory of variable n is: " << &n
}

2.3 Statements

Locations of objects in memory are implemented using large integers. In

pointers are introduced as a data type to represent and manipulate
locations of objects.

x3.6,

2.3.2 Assignments

If a b c have type double then

a = b + c

is an example of an assignment, which assigns the sum of values of objects

b and c to the object named a at the address &a: Here b + c can be called
the rvalue and a the lvalue: The term lvalue comes from \left-hand side
value," and rvalue \right-hand side value." An lvalue is dened to be an
expression that refers to an object. An rvalue is obtained by reading the
content of an object and an lvalue is associated with writing: Therefore,
a statement like a + b = c is meaningless and illegal, since a + b can not be
used as an lvalue:
Also note that an initialization and an assignment are two distinct operations. An initialization allocates space in memory for a variable and
initializes a value for the variable at its location and thus creates an object.
An assignment does not create an object but simply assigns a value to an
object already created before.

2.3.3 Compound Assignments

The assignment operator may be combined with any of the binary arithmetic or bitwise operators to form a compound assignment operator. If @
is one of these operators, then x @= y means x = x @ y except for possible
eciency dierences and side eects (see Exercises 2.5.6 and 3.14.12). For
example, x == 5 means x = x = 5 (division), and x &= 5 means x = x & 5
(bitwise and). Notice that there is no whitespace between the two operators
and they form a single operator. The complete set of compound assignment
operators are += ;= = == %= &= j= ^= <<= >>=:
Compound assignment (also increment and decrement see x2.3.4) operators have the potential to be implemented eciently. For example, an
intermediate value of x ; 5 is usually obtained and stored and then assigned
to x in x = x ; 5 while in x ;= 5 such an intermediate process can be
omitted and the value of x is just decremented by 5. They may also save
some typing and be more readable as in the following two statements.
my_very_long_variable = my_very_long_variable - 5
my_very_long_variable -= 5

// messy
// better

2. Expressions and Statements

2.3.4 Increments and Decrements

When increasing or decreasing a variable by unity, more compact notation

can be used. If n is a variable, n ++ and ++ n both increase the value
of n by 1: In the postx form n ++ the original value of n is used in the
expression in which it occurs before its value is incremented. In the prex
form ++n the value of n is rst incremented by 1 and then the new value
of n is used. A similar explanation holds for the decrement operator ;; :
For example,
int i = 4, j = 4
i++
++j
int m = i++
int n = ++j
m = --i
n = j--

//
//
//
//
//
//
//

i and j are both initialized to 4

i = i+ 1. So i= 5
j = j+ 1. So j= 5
first m=i, then i=i+1. So m=5, i=6
first j=j+1, then n=j. So n=6, j=6
first i=i-1, then m=i. So m=5, i=5
first n=j, then j=j-1. So n=6, j=5

In the statement m = ;; i the value of i is decremented by 1 rst and

the new value of i is assigned to m, while in the statement n = j ;; the
value of j is rst assigned to n and then the value of j is decremented by
1:
Compound operators may sometimes cause side eects. For example, the
statement
j = (i--) + (++i)

// unpredictable result

is a legal statement but its behavior is unpredictable and thus should not
be used, since it is nondeterministic to evaluate the left operand i ;; or the
right operand ++ i rst. Suppose i = 5: Then evaluating the left operand
i ;; before the right operand ++ i would give j = 10: Reversing the order
of evaluation would give j = 12: See Exercises 2.5.6 and 3.14.12. Compound
operators should be used in a straightforward way. Otherwise they can
easily lead to less readable and error-prone code.

2.3.5 Compound Statements

A compound statement is a block of statements enclosed by a pair of braces.

For example,
{
int n = 10
n += 5
}

is a compound statement, and it creates a new scope. Notice that no semicolon is needed after the right brace symbol. If there were a semicolon,
it would be a null statement following the compound statement. Since

2.3 Statements

a compound statement is syntactically equivalent to a single statement,

compound statements can be nested. That is, a compound statement can
appear inside another compound statement. For example,
{
int n = 10
n |= 5
{
double m = 20
m *= 3.14
}

// n = n | 5 (bitwise or)
// a nested compound statement

2.3.6 Conditional Statements

if-else Statements
The simplest conditional statement is the if-statement, whose syntax is:
if (condition) statement

where condition is a logical expression having value true or false, and

statement can be a single or compound statement. The statement is executed when the value of condition evaluates to true and is skipped otherwise. The condition must be enclosed in parentheses. For example,
if (n > 0) x *= 10
if (n == 0) {
x *= 5
}
if (n) y *= 7

// when n > 0, do x *= 10
// when n is 0, do x *= 5

// when n != 0, do y *= 7

Notice that an implicit conversion is made in the third if-statement y = 7

is executed when n has a nonzero value (which is converted to true) and
skipped when n has a zero value (which is converted to false).
The if-statement should be used with caution when the conditional expression requires the comparison for variables not of integral type. For
example, if x and y are doubles, then the condition if (x == y) may be
replaced by if (fabs(x ; y) <= Small) where Small is a user-dened very
small number and fabs is a math function that gives the absolute value of
a oating point number. See Exercise 2.5.10.
The dierence between if (x == y) and if (x = y) should also be noticed.
Such small points can be hard to detect when they are buried in a program.
The if-statement can have an optional else part. Its syntax is

if (condition) statement1
else statement2

2. Expressions and Statements

where condition is a logical expression and statement1 and statement2 are

single or compound statements. statement1 is executed and statement2 is
skipped when the value of condition evaluates to true and statement2 is
executed and statement1 is skipped otherwise. For example,
if (x == 0) cout << "denominator is zero\n"
else y /= x

// warning msg
// do division

When x is not 0, a division by x is performed. Otherwise output a warning

message.
The if-else statement can be nested. For example, in the statement
if (condition1) statement1
else if (condition2) statement2
else if (condition3) statement3
else statement4

statement1 is executed if and only if condition1 is true. If it is true,

the rest is skipped if it is false statement1 is skipped, and the value
of condition2 is checked. If condition2 is true, then statement2 is executed
and statement3 and statement4 are skipped if it is false, statement2
is skipped and the value of condition3 is checked. If condition3 is true,
then statement3 is executed and statement4 is skipped if it is false, then
statement3 is skipped and statement4 is executed. The compound if-else
statement above can be equivalently written as
if (condition1) statement1
else {
if (condition2) statement2
else {
if (condition3) statement3
else statement4
}
}

An if or if-else statement denes a new scope and declaration of a new

variable is possible, even in the condition part. For example,
double x
if (int r = rand()) {
x /= r
} else {
x += r
}
double y = r

//
//
//
//
//

#include <stdlib.h>
r is declared in condition part
if r is nonzero, do a division
if r is zero, do an addition
r is also within scope here

// illegal, r is out of scope

The function rand(), declared in the standard header <stdlib:h> generates a pseudorandom number between 0 and RAND MAX: The macro
RAND MAX is also dened in <stdlib:h>. To generate a pseudorandom

2.3 Statements

number between 0 and n, using int((double(rand()) =RAND MAX ) n)

gives a more random result than simply rand()%n.
The ternary conditional operator ? : is dened as follows. The expression z ?x : y has value x when z is true and has value y when z is false. As an
example, the following ternary conditional statement and if-else statement
are equivalent.
i = a > b ? a : b

// ternary conditional statement

if (a > b) i = a
else i = b

// if-else statement

goto Statements

It is sometimes necessary to make an explicit jump from the point of current

execution to another point that has been labeled with an identier using
the goto statement:
if (condition) goto label
label: statement

The labeled statement (a colon is needed after the label) can occur before
or after the goto jump. A goto statement is useful when a jump is needed
out of deeply nested blocks. For example,
repeat: i = 9
// statement labeled by "repeat"
{
// ...
{
// ...
if (i == 0) goto repeat
if (i == 8) goto done
}
}
done: i = 1
// colon needed after label "done"

However, a succession of goto statements can lead to loss of control in a

program and should be avoided. Even a single goto statement is rarely
necessary.

switch Statements
A switch statement is often used for multiple choice situations. It tests the
value of its condition, supplied in parentheses after the keyword switch:
Then this value will be used to switch control to a case label matching the
value. If the value does not match any case label, control will go to the
default case. The break statement is often used to exit a switch statement.
For example,

2. Expressions and Statements

int i, j, k
// do something to compute the value of i
switch(i) {
case 0:
j = 6
k = 36
break
case 5:
j = 7
k = 7
break
default:
j = 8
}

// i is tested against the cases

// control comes here if i == 0
// statement of case 0
// branch out of switch statement
// control comes here if i == 5
// statement of case 5
// branch out of switch statement
// use default if i != 0 && i !=5
// default statement

In this switch statement, the value of i is tested. When i is 0, the statement corresponding to case 0 is executed. When i is 5, the statement corresponding to case 5 is executed. When i is neither 0 nor 5, the default
case is executed. Notice the appearance of the break statement after each
case. Without it, every following statement in the switch statement would
be executed, until a break statement is encountered. There are occasions
when this is desired. A default choice is optional but is sometimes useful to
catch errors. For example,
char x
double y
// do something to compute the value of x
switch(x) {
case 'A':
//
case 'a':
//
y = 3.14
break
//
case 'B':
case 'b':
//
y = 2.17
break
//
default:
//
cout << "Impossible value
}

char enclosed in single quotes

when x is A or a, let y = 3.14
branch out of switch statement
when x is B or b, let y = 2.17
branch out of switch statement
if x not A, a, B, b, report error
of x happened\n"

In general, the switch expression may be of any type for which integral
conversion is provided and the case expression must be of the same type.
A switch statement can always be replaced by an if-else statement, but a
switch statement can sometimes make the code more readable.

2.3 Statements

2.3.7 Iteration Statements

for Loops
The for loop is very versatile and convenient to use. Its general form is
for (initialize condition expression) statement

where initialize represents statements initializing the control variable or

variables, condition provides loop control to determine if statement is
to be executed, and expression indicates how the variables initialized in
initialize are to be modied and is evaluated (each time) after statement
is executed. The loop proceeds until condition evaluates to false. If the
rst evaluation of condition is false, neither expression nor statement is
evaluated at all. For example,
for (int i = 3 i < 50 i *= 2) {
cout << i << '\n'
}

It will print out integers 3, 6, 12, 24, 48. The initial value of i is 3. Since
condition i < 50 is true, 3 is printed. Then i = 2 gives the second value
of i to be 6: Since i < 50 is still true, 6 is printed. This process repeats
until i = 2 gives i = 96 at which time the condition i < 50 is no longer
true.
The following loop will not print out anything since the initial condition
is false.
for (int i = 3 i > 5 i *= 2) {
cout << i << '\n'
}

When i = 3 the condition i > 5 is false. Thus the loop exits without
executing the statement inside the braces. The expression i = 2 is also
skipped.
A for loop denes a new scope. That is, the declaration int i = 3 in the
previous for loop is out of scope outside the loop:
for (int i = 3 i < 50 i *= 2) {
cout << i << '\n'
}
i = 20

// i visible only in loop

// error, i out of scope

To avoid accidental misuse of a variable, it is usually a good idea to

introduce the variable into the smallest scope possible. In particular, it
is usually best to delay the declaration of a local variable until one can
give it an initial value. Both if-else statements and for statements provide
such a mechanism to avoid accidental misuse of variables and to increase
readability.

2. Expressions and Statements

The break statement can be used to stop a loop, and the continue statement stops the current iteration and jumps to the next iteration. For example,
for (int i = 3 i < 50 i *= 2) {
if (i == 6) continue
// jump to next iteration when i==6
cout << i << '\n'
if (i == 24) break
// break the loop when i is 24
}

It will print out integers 3, 12, 24. When i is 6, the value of i is not printed
out since the continue statement causes control to go to the next iteration
with i = 6 2 = 12: When i is 24 the loop is exited due to the break
statement. The break statement is often used to break an otherwise innite
loop. For example,
int n = 0
for ( ) {
cout << ++n << '\n'
if ( n == 100) break
}

// an infinite loop
// break infinite loop when n==100

It will print out 1 2 : : : 100: The loop breaks when n is 100: This is an
otherwise innite loop since it has no condition to check.
Comma expressions sometimes appear inside a for loop and the statement
of a for loop can be an empty statement. For example,
double sum
int i
for (sum = 0, i = 0 i < 100 sum += i, i++)
cout << "sum = " << sum << '\n'
// sum = 4950

It initializes sum = 0 and i = 0: The comma expression sum += i i ++

is executed until i < 100 is not true. This for loop can also be written as
for (sum = 0, i = 0 i < 100 sum += i++)

The iteration variable in a for loop does not have to be integral. For
example,
for (double x = 0

x < 300.14 x += 2.17) { sum += x }

A for loop can be nested just like any other compound statement. For
example,
double a = 0
for (int i = 0 i < 100 i++) {
for (int j = 200 j <= 500 j++)
}

a += i - j

2.3 Statements
long double b = 5
for (int i = 1 i < 5 i++)
for (int j = 27 j >= - 3 j--)

b *= i + j*j

They compute, respectively,

99 X
500
X

i=0 j =200

where

(i ; j ) and 5

stands for summation and

4 Y
27
Y

i=1 j =;3

(i + j 2 )

for product.

while Loops
The while loop statement has the form
while (expression) statement
The statement is executed until expression evaluates to false. If the initial
value of expression is false, the statement is never executed. For example,
int x = 0
while (x <= 100) {
cout << x << '\n'
x++
}

The statements inside the braces are executed until x <= 100 is no longer
true. It will print out 0 1 2 : : : 100: This loop is equivalent to the following
for loop.
int x
for (x = 0 x <= 100 x++) cout << x << '\n'

The following while loop will not print out anything since the initial condition is false.
int x = 10
while (x > 100) {
cout << x++ << '\n'
}

do-while Loops
The do-while statement resembles the while loop, but has the form
do statement while (expression)
The execution of statement continues until expression evaluates to false.
It is always executed at least once. For example,

2. Expressions and Statements

int x = 10
do {
cout << x++ << "\n"
} while (x > 100)

It will stop after the rst iteration and thus print out only 10:
The continue statement can also be used in a while or do-while loop to
cause the current iteration to stop and the next iteration to begin immediately, and the break statement to exit the loop. Note that continue can
only be used in a loop while break can also be used in switch statements.
For example, the code segment
int x = 10
do {
cout << x++ << '\n'
if (x == 20) break
} while (true)

// print x, then increment x

// exit loop when x is 20

should print out all integers from 10 up to 19 and the code segment
int x = 10
do {
if (x++ == 15) continue
cout << x << '\n'
if (x == 20) break
} while (true)

// test equality, then increase x

should print out all integers from 11 up to 20 except for 16: However, the
code
int x = 10
while (true) {
if (x == 15) continue
cout << x++ << '\n'
if (x == 20) break
}

rst prints out integers 10 11 12 13 14 and then goes into an innite loop
(since x will stay equal to 15 forever).

2.4 Fibonacci Numbers

In this section, an example on Fibonacci numbers is given to illustrate the
use of some C++ language features.
The sequence of Fibonacci numbers is dened recursively by

f0 = 0

f1 = 1

fn+1 = fn + fn;1 for n = 1 2 3 : : ::

2.4 Fibonacci Numbers

For instance, f2 = f1 + f0 = 1 + 0 = 1 and f3 = f2 + f1 = 1 + 1 = 2:

Fibonacci numbers have many interesting applications and properties. One
of the properties is that the sequence of Fibonacci quotients

qn = fn =fn;1

n = 2 3 4 : : :
p
converges to the golden mean, which is (1 + 5)=2
1:618:

The following program prints out the rst 40 (starting with n = 2) Fibonacci numbers and their quotients, using long int for Fibonacci numbers
and long double for Fibonacci quotients.
Making use of the for loop, the program can be written as
#include <iostream>
using namespace std
main() {
long fp = 1
long fc = 1

// previous Fibonacci number

// current Fibonacci number

for (int n = 2 n <= 40 n++) {

// main loop
cout << n << "
"
// output value of n
cout << fc<< "
"
// output Fibonacci number
cout << (long double)fc/ fp << '\n'
// Fib quotient
long tmp = fc
fc += fp
fp = tmp

// temporary storage
// update Fib number
// store previous Fib number

}
}

Inside the for loop, the variable fc is used to store the updated Fibonacci
number and fp the previous one. Adding fp to fc gives the next Fibonacci
number that is also stored in fc: A temporary storage tmp is used to store
fc before the updating and assign the stored value to fp after updating,
getting ready for the next iteration. Notice that the integer fc is converted
into long double before the division fc=fp is performed, since otherwise
the result would be just an integer.
This program works ne except the output may look messy. To have a
formatted output, it may be modied to control the space in terms of the
number of characters each output value occupies:
main() {
long fp = 1
long fc = 1
cout.width(2)
cout << "n"

// previous Fibonacci number

// current Fibonacci number
// output in a space of 2 chars

2. Expressions and Statements

cout.width(27)
// output in a space of 27 chars
cout << "Fibonacci number"
cout.width(30)
cout << "Fibonacci quotient" << "\n\n"
cout.precision(20)
// precision for Fib quotients
for (int n = 2 n <= 40 n++) {
cout.width(2)
cout << n
cout.width(27)
cout << fc
// output Fibonacci number
cout.width(30)
cout << (long double)fc/ fp << '\n'
// Fib quotient
long tmp = fc
fc += fp
fp = tmp

// temporary storage
// update Fib number
// store previous Fib number

}
}

Some of the formatted output of the program is:

Fibonacci number

Fibonacci quotient

2
3
4
5
6
7
8
9
10

1
2
3
5
8
13
21
34
55

1
2
1.5
1.6666666666666666666
1.6
1.625
1.6153846153846153846
1.6190476190476190476
1.6176470588235294118

6765

1.6180339631667065295

832040

1.6180339887482036213

39
40

63245986
102334155

1.6180339887498951409
1.6180339887498947364

The results show that the Fibonacci quotients converge very quickly to
the golden mean and the Fibonacci numbers grow extremely fast. Over ow
will occur well before n reaches 100 see Exercise 2.5.14. A technique is
introduced in Exercise 3.14.21 to store a large integer into an array of

2.5 Exercises

integers digit by digit, which enables one to compute Fibonacci numbers

for n up to thousands or larger.

2.5 Exercises
2.5.1. Add printing statements to the program at the end of x2.1.1 so that
the value of each variable in dierent scopes is output to the screen
right after it is dened. Check to see if the output values are the ones
that you have expected.
2.5.2. The function rand(), declared in the standard header <stdlib:h>,
generates a pseudorandom number between 0 and RAND MAX:
The macro RAND MAX is also dened in <stdlib:h>. Write and
run a program that outputs the value of RAND MAX on your system
and generates 100 pseudorandom numbers between 5 and 50, using a
for loop.
2.5.3. Rewrite the program in Exercise 2.5.2 using a while loop. Compare
the 100 numbers generated by this program and the program in Exercise 2.5.2.
2.5.4. Use a nested for loop to write a program to compute the sum:
100 X
300
X

i=0 j =5

cos(i2 + j ):

Check x3.11 for the cosine and square root functions in the math
library <math:h>.
2.5.5. Write a program that computes the number of (decimal) digits of a
given integer n. For example, when n = 145 the number of digits of
n is 3:
2.5.6. Write a program that contains the statements
j = i-- + ++i
k = j + j++
k *= k++

// unpredictable result
// unpredictable result
// unpredictable result

and outputs the values of j and k with i initialized to 5: Try to run the
program on dierent machines if possible. The result of such statements is machine-dependent and can cause dangerous side eects.
Experienced programmers can recognize such statements and do not
use them.
2.5.7. The two statements

2. Expressions and Statements

i = 5
j = (i = 7) + (k = i + 3)

are legal, but give dierent results on dierent machines. The value
of j is either 17 or 15 depending on whether the left subexpression
(i = 7) or the right subexpression (k = i + 3) is evaluated rst. The
value of k is either 10 or 8: Test what values j and k have under
your compiler. Again, such legal but machine-dependent statements
should be avoided.
2.5.8. Can division by zero ever occur in the following code segment?
int x = rand(), y = rand()
if ((x != 5) && (y/(x-5) < 30))
z = 2.17

// x, y are random int

// assume z declared

Then how about this (just switch the order of the subexpressions):
int x = rand(), y = rand()
if ((y/(x-5) < 30) && (x != 5))
z = 2.17

// x, y are random int

// assume z declared

Note that the order of evaluation in logical expressions can make a

dierence.
2.5.9. Are the following two statements equivalent?
for (sum = 0, i = 0 i < 100 sum += i, i++)
for (sum = 0, i = 0 i < 100 i++, sum += i)

Note that it is guaranteed that evaluation is done from left to right

in a comma expression. There are only three such expressions. What
are they?
2.5.10. Run the following program:
main() {
double sum = 0
for (double x = 0.0 x != 5.5 x += 0.1) sum += x
}

It is intended to nd the sum of 0:1 + 0:2 + + 5:4: Does it go

into an innite loop on your computer? It does on most computers
since x will never be equal to 5:5 due to computers' nite precision.
In fact, numbers such as 1=3 and 0:1 can not be represented exactly
on computers. To x this for loop, replace it by

2.5 Exercises

for (double x = 0.0 x <= 5.5 x += 0.1) sum += x

Test this loop on your computer. Equality expressions involving oating point arithmetic must be used with great care. Replacing them
by relational expressions often leads to more robust code.
2.5.11. Write a program to test the dierence between the statement
while (i++ < n) { cout << i << '\n' }

and the statement

while (++i < n) { cout << i << '\n' }

where i and n are given integers, say i = 0 and n = 10: Is there any
dierence between the following two statements?
for (char i = 'a' i <= 'z' i++) cout << i << '\n'

and
for (int i = 'a' i <= 'z' ++i) cout << i << '\n'

2.5.12. Here is an easy way to know the largest unsigned int on a computer
with two's complement:
unsigned int largeui = - 1

Test to see if largeui stores the largest unsigned int on your computer. Explain the reason. Does your compiler warn you about assigning a negative int to an unsigned int? If it does, how about:
int i = - 1
unsigned int largeui = i

2.5.13. Rewrite the rst Fibonacci number program in x2.4 using a while loop
and the second using a do-while loop so that they produce exactly
the same output as the original programs.
2.5.14. Can you modify the program in x2.4 so that it will try to print out the
rst 100 Fibonacci numbers? Your output will probably have negative
numbers for large n: This is caused by integer overow since Fibonacci
numbers grow very quickly. Computation usually continues but produces incorrect results when overow occurs. It is the programmer's
responsibility to make sure that all variables (suspected to take on
large or small values) stay within range and stop the program or take
other actions when overow happens. Rewrite the program so that it
will print out a warning message before overow occurs.

2. Expressions and Statements

2.5.15. Write a program that outputs the bit representation of an integer. On

a computer with sizeof (int) = 4 (i.e., 32 bits are used to represent
an integer) the program may be written as
main() {
int n
int mask = 1 << 31
// mask = 10000000 ... 00000000
for ( ) {
// enter an infinite loop
cout << "\nEnter an integer: \n"
// input an int
cin >> n
// input is assigned to n
cout << "bit representation of " << n << " is: "
for (int i = 1 i <= 32 i++) { // loop over 32 bits
char cc = (n & mask) ? '1': '0'
// take bit i from left to right
cout << cc
// print bit i from left to right
n <<= 1
// n = n << 1 left shift n by 1 bit
if (i%8 == 0 && i != 32) cout << ' '
// print one space after each byte
}
}
}

Note that n & mask gives the leftmost bit of n followed by 31 '0's
in bit representation. When this bit is 1, character '1' is assigned to
cc and otherwise cc = '0'. The statement n <<= 1 shifts the bit
pattern of n to the left by 1 bit. When i = 1 the leftmost bit of n
is printed out and when i = 2 the second leftmost bit is printed out,
and so on. Run the program on your computer and print out the bit
representations of 55555 and ;55555: Modify the program to handle
integers with sizeof (int) = m, where m may be dierent on dierent
machines.

This is page 51
Printer: Opaque this

Derived Types

From basic types, other types can be derived. This chapter discusses derived
types such as enumeration types for representing a specic set of values,
pointers for manipulating addresses or locations of variables, arrays for
storing a number of elements having the same type, reference types for creating aliases, structures for storing elements of dierent types, and unions
for storing elements of dierent types when only one of them is present
at a time. Functions|a special type|that may take some values as input
and produce some other values as output are also discussed. Another derived type called class is discussed in Chapter 5. All C++ operators and
their precedences, and all mathematical functions in the standard library
<math:h> are summarized towards the end of the chapter. The chapter
ends with one section on polynomial evaluations and another on numeric
integration techniques.

3.1 Constants and Macros

A value that can not be changed throughout a program is called a constant.
It is dened using the keyword const: It is illegal to redene a constant
and for this reason, a constant should be initialized when it is declared.
For example,
const char tab = '\t'
// tab character
const double pi = 3.1415926535897932385 // pi
const double e = 2.7182818284590452354 // natural log base

3. Derived Types

pi = 3.1
const int size

// illegal, a const can not be changed

// illegal, a const must be initialized

const int dimension = 3

const int numelements = 100*dimension

// const expression

Note that const changes a type in that it restricts the way in which an
object can be used. Constant expressions are allowed in the denition of
a constant such as numelements above, since the compiler can evaluate
constant expressions.
Constants are typed and thus preferable to traditional C macros, which
are not type-checked and are error-prone, such as
#define PI 3.1415926535897932385

A macro is dened by using dene preceded by # which must stand in

the rst column. A preprocessor will substitute every occurrence of PI by
3.1415926535897932385 in the program when it is compiled. An example
of a macro with an argument is:
#define SQ(x) ((x) * (x))

Then every occurrence of SQ(x) will be substituted by ((x) (x)) in the

program. For example, SQ(a + b) will be substituted by ((a + b) (a +
b)). Notice the necessity of the parentheses in the denition of a macro.
Otherwise SQ(a + b) would be just (a + b a + b) if the two inner pairs of
parentheses were omitted in the macro denition. By convention, capital
letters are usually used for macro names. Although macros are still legal
in C++, they are best avoided.
The const specier can be used together with the keyword volatile: A
volatile specier is a hint to the compiler that an object may be modied
in ways not detectable by the compiler and aggressive optimizations must
be avoided. For example, a real-time clock may be declared as:
extern const volatile int clock

This says that clock may be changed by an agent external to the program
since it is volatile and may not be modied inside the program since it is
const: Thus two successive reads of the variable clock in the program may
give two dierent values. An extern const may not have to be initialized.

3.2 Enumerations
The enumeration type enum is for holding a set of integer values specied
by the user. For example,
enum {blue,yellow,pink =20,black,red =pink +5, green =20}

3.2 Enumerations

is equivalent to
const int blue = 0, yellow = 1, pink = 20, black = 21,
red = 25, green = 20

By default, the rst member (enumerator) in an enum takes value 0 and

each succeeding enumerator has the next integer value, unless other integer
values are explicitly set. The constant pink would take value 2 if it were not
explicitly dened to be 20 in the denition. The member black has value
21 since the preceding member pink has value 20: Note that the members
may not have to take on dierent values.
An enumeration can be named and thus create a distinct type. By default enumerations can be implicitly converted to integers, but there is no
implicit conversion from integers to enumerations to avoid unpredictable
results. For example,
enum bctype {Dirichlet,
bctype x, y
//
x = Dirichlet
//
y = 0
//
y = 40
//
x = bctype(2)
int i = Neumann

Neumann, Robin}
// a named enum
declare x, y to be of type bctype
assigning a value to variable x
illegal, an int can not be assigned
to a variable of type bctype

// OK, explicit conversion, x = Robin

// OK, enum is restricted int, i = 1

if (x != y) i = Robin

// they can be compared

The rst statement above denes a new type called bctype: A variable of
this type may take on values of Dirichlet, Neumann and Robin:
An enum is a user-dened type and often used in conjunction with the
switch statement:
bctype x
// ...

// declare x to be of type bctype

compute or get the value of x

switch(x) {
// do differently based on value of x
case Dirichlet:
// ... deal with Dirichlet boundary condition
break
case Neumann: case Robin:
// ... deal with Neumann and Robin conditions at same time
break
Default:
cout << "undefined boundary condition\n"
}

3. Derived Types

Although a programmer rarely needs to know the range of an enumeration type, the range is normally larger than the set of values of the enumerators. The range is dened to be the set of all integers in the closed interval
!0 2n ; 1] if the smallest enumerator is nonnegative and in the closed interval !;2n 2n ; 1] if the smallest enumerator is negative, where n is the
nearest larger binary exponent of the largest (in magnitude) enumerator.
This is the smallest bit-eld that can hold all the enumerator values. For
example,
enum flag { red = 1, blue = 0 }
// range: 0, 1
enum onenine { one = 1, nine = 9 } // range: 0, 1,..., 15
enum myminmax{ min = -90, max = 1000}
// range: - 2048,..., 2047
onenine m0 = onenine(5)
myminmax m1 = myminmax(- 1000)
myminmax m2 = myminmax(5000)

// Ok, 5 is within range

// OK, -1000 within range
// error, out of range

Since 9 = 23 + 1 the nearest larger binary exponent is n = 4: Thus the

range of onenine is the set of all integers from 0 up to 24 ; 1 = 15:
Constants and enumerations are designed for a better type-checking
mechanism to avoid possible type errors and accidental modication of
constant values.

3.3 Arrays
For a type T , T !n] is the type \one-dimensional array of n elements of type
T" where n is a positive integer. The elements are indexed from 0 to n ; 1

and are stored contiguously one after another in memory. For example,
float vec%3]
int stg%30]
vec%0] = 1.0
vec%1] = 2.0

// array of 3 floats: vec%0],vec%1],vec%2]

// array of 30 ints: stg%0],..., stg%29]
// accessing element 0 of vec
// accessing element 1 of vec

for (int i = 0 i < 30 i++) stg%i] = i*i + 7

int j = stg%29]
// accessing the last element of stg

The rst two statements declare vec and stg to be one-dimensional arrays
with 3 and 30 elements of type float and int, respectively. A for loop is
often used to access all elements of a 1D array. A one-dimensional array
can be used to store elements of a (mathematical) vector.
Two-dimensional arrays having n rows and m columns (looking like a
matrix) can be declared as T !n]!m], for elements of type T: The row index
changes from 0 to n ; 1 and the column index from 0 to m ; 1: For example,

3.3 Arrays
double mt%2]%5]
mt%0]%0] = 5
mt%1]%4] = 5
double a = mt%0]%0]

//
//
//
//

2D array of 2 rows and

access entry at row 0,
access entry at row 1,
access entry %0]%0] of

5 columns
column 0
column 4
mt

for (int i = 0 i < 2 i++) {

for (int j = 0 j < 5 j++) mt%i]%j] = i + j
}

The rst statement above declares mt to be a two-dimensional array with 2

rows and 5 columns, whose element at row i and column j can be accessed
by using mt!i]!j ]: A nested for loop is often used to access all elements of a
2D array. A two-dimensional array can be thought of as a one-dimensional
array of one-dimensional arrays. For example, mt can be viewed as having
two \elements": mt!0] and mt!1] where mt!0] is a one-dimensional array of
ve elements representing the initial row (row 0) of mt and mt!1] is also a
one-dimensional array of ve elements but representing the second row (row
1). The symmetric notation mt!i]!j ] seems to be superior to FORTRAN's
notation mt(i j ) since mt!i] can be used as a one-dimensional array and
element j of mt!i] is mt!i]!j ]: A two-dimensional array is a natural data
structure for storing elements of a matrix.
Three and more dimensional arrays can be declared similarly. Let s1 s2
: : : sk be positive integers then the declaration of a k-dimensional array
a!s1 ]!s2 ] !sk ] will allocate spaces for s1 s2 sk elements, which are
stored contiguously in memory row by row (unlike in FORTRAN, where
they are stored column by column).
Arrays can be initialized at the time of declaration. For example,
int v%] = {1, 2, 4, 5}
// initialization of 1D array
int a%3] = {2, 4, 5}
// initialization of 1D array
int u% ]%3] = { {1, 2, 3}, {4, 5, 8} }
// initialization of 2D array
char filename%30] = "outpt" // a character array

In the declaration of v above, the compiler will count the number of elements in the initialization and allocate the correct amount of space for v
so that the number of elements can be omitted. In an initialization of a
multi-dimensional array such as u! ]!3], only the rst dimension (the number of rows for u) can be omitted. However, assignments can not be done
this way:
int vv%4]
vv = {1, 2, 4, 5}

// a declaration of a 1D array
// error, use: vv%0] =1, vv%1] =2, ...

Since multidimensional arrays are stored row by row, they can be initialized like one-dimensional arrays. For example, the following two initializations are equivalent.

3. Derived Types

int ia%2]%3] = {1, 2, 0, 4, 0, 0}

int ia%2]%3] = { {1, 2, 0}, {4, 0, 0} }

Entries that are not explicitly initialized are set to 0: Thus they are also
equivalent to:
int ia%2]%3] = { {1, 2}, {4} }

For this reason, the initialization

int ib%2]%3] = {5}

initializes the rst entry ib!0]!0] = 5 and all other entries to 0:

Here is an example of an initialization of a three-dimensional array:
double ax%4]%2]%3] = {
{ {11, 22, 0}, {55, 0,
{ {-1, -2, 0}, {-5, -6,
{ {23, -7, 8}, {0, 13,
{ {-3, 19, 0}, {9, -5,
}

0}
0}
8}
3}

},
},
},
}

//
//
//
//

ax%0]
ax%1]
ax%2]
ax%3]

In particular, ax!0]!0]!0] = 11 ax!2]!0]!2] = 8 ax!3]!1]!1] = ;5 and ax!0]

is a two-dimensional array with entries ff11 22 0g f55 0 0gg. Its entries
may be accessed by a triple loop:
for (int i = 0 i < 4 i++)
for (int j = 0 j < 2 j++)
for (int k = 0 k < 3 k++) cout << ax%i]%j]%k] << '\n'

The dimensions of an array must be known at the time it is declared

so that the compiler can allocate space for it. Otherwise it is illegal. For
example,
const int m = 2000
const int k = 2*m + 100
double b%k]

// compiler evaluates k = 4100

// OK, k is known to compiler

// ... compute the value of n at run-time

double a%n]
// error, compiler does not know value of n

The number of elements in a sequence often can not be known at compiletime, but rather is calculated at run-time. If we declare a large array, it
wastes space. If we declare a small one, it may not be big enough to hold
all the elements. For this reason arrays should not be used unless their dimensions are exactly or roughly known at compile-time. Instead, pointers
are often used and space can be allocated at run-time see x3.6. The standard libraries <valarray> and <vector> can also be conveniently used to
handle arrays whose sizes may not be known at compile-time see x7.5 and
x10.1.1.

3.4 Structures

3.4 Structures
Unlike an array that takes values of the same type for all elements, a struct
can contain values of dierent types. It can be declared as
struct point2d {
char nm
float x
float y
}

//
//
//
//

a structure of 2D points
name of the point
x-coordinate of point
y-coordinate of point

This denes a new type called point2d: Note the semicolon after the right
brace. This is one of the very few places where a semicolon is needed following a right brace. The structure point2d has three elds or members, one
char and two floats: Its size in memory is sizeof (point2d) bytes, which
may not be equal to sizeof (char) + 2 sizeof (float): Its members are
accessed by the : (dot) operator. For example,
point2d pt
pt.nm = 'f'
pt.x = 3.14
pt.y = - 3.14

//
//
//
//

declare pt of type point2d

assign char `f' to its field nm
assign 3.14 to its field x
assign -3.14 to its field y

double a = pt.x
char c = pt.nm

// accessing member x of pt
// accessing member nm of pt

A variable of a struct represents a single object and can be initialized by

and assigned to another variable (consequently, all members are copied).
For example,
point2d pt2 = pt
pt = pt2

// initialize pt2 by pt, memberwise copy

// assign pt2 to pt, memberwise copy

It can also be initialized in a way similar to arrays:

point2d pt3 = {'F', 2.17, - 3.14}

// OK, initialization

point2d pt4
pt4 = {'A', 2.7, - 3.4}

// illegal, assignment

An example of struct with an unnamed enum as a member can be found

in x7.7.4.

3.5 Unions and Bit Fields

Unions are to save space, but follow the same syntax as structures. Only
one member of a union can exist at a time and all members take up only
as much space as its largest member. It can be declared as

3. Derived Types

union val {
int i
double d
char c
}

// i, d, c can not be used at same time

// d is largest member in storage

The union val has three members: i, d and c, only one of which can exist
at a time. Thus sizeof (double) bytes of memory are enough for storing an
object of val: The exact size of val is sizeof (val): Members of a union are
also accessed by the : (dot) operator. It can be used as the following.
int n
cin >> n

// n is taken at run-time

val x
// x is a variable of type val
if (n == 1) x.i = 5
else if (n == 2) x.d = 3.14
else x.c = 'A'
double v = sin(x.d)

// error, x.d may not exist all times

When x is declared, sizeof (double) bytes of memory are created for it.
This piece of memory can be used to store an int (member x:i), a double
(member x:d), or a char (member x:c). The last statement above is an
error (a compiler can not catch such an error and a programmer should be
responsible for this) since x:d may not be dened at all times. To see this,
look at the program:
main() {
union val {
int i
double d
char c
}
val x
x.i = 5

// only x.i is defined now.

// Other members should not exist
cout << "x.i = " << x.i << ", x.d = "
<< x.d << ", x.c = " << x.c << '\n'
x.d = 6.28
// only x.d is defined now.
cout << "x.i = " << x.i << ", x.d = "
<< x.d << ", x.c = " << x.c << '\n'
x.c = 'G'

// only x.c is defined now.

3.5 Unions and Bit Fields

cout << "x.i = " << x.i << ", x.d = "
<< x.d << ", x.c = " << x.c << '\n'
}

The output of this program on my computer is:

x.i = 5, x.d = 3.10143, x.c =
x.i = 1374389535, x.d = 6.28, x.c =
x.i = 1374389575, x.d = 6.28, x.c = G

Note that only the member that is dened can be correctly printed out.
Values of other members should not be used. When x:i is dened to be
5 its value can be printed out (and used) correctly. However, this same
piece of memory is interpreted to be a double and a char when x:d and
x:c are printed out, respectively. The outputs for x:d and x:c are of course
erroneous in the rst printing statement above.
Suppose that triangle and rectangle are two structures and a figure
can be either a triangle or a rectangle but not both then a structure for
figure can be declared as
struct figure2d {
char name
bool type
union {
triangle tria
rectangle rect
}
}

// 1 for triangle, 0 for rectangle

// an unnamed union

If fig is a variable of type figure2d, its members can be accessed as

fig:name fig:type fig:tria, or fig:rect. Since a gure can not be a rectangle and a triangle at the same time, using a union can save memory
space by not storing triangle and rectangle at the same time. The member fig:type is used to indicate if a triangle or rectangle is being stored in
an object fig (e.g., fig:rect is dened when fig:type is 0). Such a member
is often called a type tag.
Note that a union without a name is used inside figure2d: If a named
union is used, for example,
struct figure2d {
char name
bool type
union torr {
triangle tria
rectangle rect
}
}

// 1 for triangle, 0 for rectangle

// a named union

3. Derived Types

then the members tria and rect will have to be accessed as fig:torr:tria
and fig:torr:rect for a figure2d object fig:
The size of a C++ object is a multiple of the size of a char which is one
byte. That is, a char is the smallest object that can be represented on a
machine. There are occasions when it is necessary to put a few small objects
into one byte to squeeze more space. A member of a struct or union can be
declared to occupy a specied number of bits. Such a member is called a bit
eld and the number of bits it occupies is called its bit width. A bit eld must
be of an integral type (typically unsigned int), and it is not possible to take
its address. The bit width is at most the number of bits in a machine word
and is specied by a nonnegative constant integral expression following a
colon. Unnamed bit elds are also allowed for padding and alignment. An
example is
struct card {
unsigned int pips : 4
unsigned int suit : 2
unsigned int kq
: 2
// ...
}

//
//
//
//

pips occupies 4 bits

suit occupies 2 bits
kq occupies 2 bits
other fields can be added

A variable of card has a 4-bit eld pips that can be used to store small
nonnegative integers 0 to 15 a 2-bit eld suit for values 0 to 3 and another
2-bit eld kq: Thus the values of pips suit and kq can be compactly stored
in 8 bits. It can be used as
card c
c.pips = 13
c.suit = 3
c.kq
= 0

// c is of type card

The layout of the bits in left-to-right or right-to-left order is machinedependent. A compiler may not assign bit elds across word boundaries.
For example, sizeof (card) = 4 on some machines with 4-byte words. (It
still saves memory in this case since storing three unsigned int separately
would require 12 bytes.) Programs manipulating bits may not be portable.
Unless saving memory is really necessary for an application, bit elds (also
unions) should be avoided.

3.6 Pointers
For a type T , T is the pointer to T: A variable of type T can hold the
address or location in memory of an object of type T: For this reason, the
number of bytes of any pointer is sizeof (int) = sizeof (double) and so
on. For example, the declaration

3.6 Pointers
int* p

// p is a pointer to int

declares the variable p to be a pointer to int. It can be used to store the

address in memory of integer variables.
If v is an object, &v gives the address of v (the address-of operator &
is introduced in x2.3.1). If p is a pointer variable, p gives the value of the
object pointed to by p: We also informally say that p is the value pointed
to by p: The operator is called the dereferencing or indirection operator.
The following example illustrates how the address-of operator & and the
dereferencing operator can be used:
int i = 5
int* pi = &i

// i is int, value of object i is 5

// pi is a pointer to int
// and assign address of i to pi

int j = *pi

// value of object pointed to by pi

// is assigned to j, so j=5

double* d = &j

// illegal

The second statement above declares pi to be a variable of type: pointer to

int, and initializes pi with the address of object i: Another way of saying
that pointer pi holds the address of object i is to say that pointer pi points
to object i: The third statement assigns pi the value of the object pointed
to by pi, to j: The fourth statement is illegal since the address of a variable
of one type can not be assigned to a pointer to a dierent type.
For a pointer variable p, the value p of the object that it points to can
change so can the pointer p itself. For example,
double d1 = 2.7, d2 = 3.1
double* p = &d1
// p points to d1, now *p = 2.7
double a = *p
// a = 2.7
p = &d2
double b = *p

// p now points to d2, *p = 3.1

// b = 3.1

*p = 5.5
double c = *p
double d = d2

// value p points to is now 5.5

// c = 5.5
// d = 5.5, since *p = 5.5

Since p is assigned to hold the address of d2 in the statement p = &d2 then

p can also be used to change the value of object d2 as in the statement
p = 5:5: When p points to d2 p refers to the value of object d2 and
assignment p = 5:5 causes d2 to equal 5:5.
A sequence of objects can be created by the operator new and the address
of the initial object can be assigned to a pointer. Then this sequence can
be used as an array of elements. For example,

3. Derived Types

int n = 100
// n can also be computed at run-time
double* a
// declare a to be a pointer to double
a = new double %n] // allocate space for n double objects
// a points to the initial object

The last two statements can also be combined into a more ecient and
compact declaration with an initialization:
double* a = new double %n]

// allocate space of n objects

In allocating space for new objects, the keyword new is followed by a type
name, which is followed by a positive integer in brackets representing the
number of objects to be created. The positive integer together with the
brackets can be omitted when it is 1: This statement obtains a piece of
memory from the system adequate to store n objects of type double and
assigns the address of the rst object to the pointer a: These objects can
be accessed using the array subscripting operator ! ] with index starting
from 0 ending at n ; 1: For example,
for (int i = 0 i < n i++) // a%0] refers to initial object
a%i] = i*i + 8
// a is used just like an array
double sum = 0
// find sum of all objects a%i]
for (int i = 0 i < n i++) sum += a%i]

The pictorial representation of the pointer a and the space allocated for it
can be drawn as

a
#

a!0]

a!1]

a!2]

a!n-1]

You can print out the values of a and a!0] to check that they are indeed the
same. This agrees with the fact that a points to the initial object a!0]: After
their use, these objects can be destroyed by using the operator delete :
delete%] a

// free space pointed to by a

The system will automatically nd the number of objects pointed to by a

(actually a only points to the initial object) and free them. Then the space
previously occupied by these objects can be reused by the system to create
other objects. Since the operator new creates objects at run-time, this is
called dynamic memory allocation. The number of objects to be created by
new can be either known at compile-time or computed at run-time, which
is preferred over the built-in arrays (x3.3) in many situations. In contrast,
creation of objects at compile-time is called static memory allocation. Thus
there are two advantages of dynamic memory allocation: objects no longer

3.6 Pointers

in use can be deleted from memory to make room to create other objects,
and the number of objects to be created can be computed at run-time.
Automatic variables represent objects that exist only in their scopes. In
contrast, an object created by operator new exists independently of the
scope in which it is created. Such objects are said to be on the dynamic
memory (the heap or the free store). They exist until being destroyed by
operator delete or to the end of the program.
Space can also be allocated for a pointer to create just one object in a
similar way:
char* pt
pt = new char
pt%0] = 'c'
*pt = 'd'
char x = pt%0]
char y = *pt
delete pt

//
//
//
//
//
//
//

pt is a pointer to character
allocate space of 1 char for pt
place 'c' at location pointed to by pt
place 'd' at location pointed to by pt
assign pt%0] to object x, x = 'd'
assign *pt to object y, y = 'd'
free space pointed to by pt

Note that when more than one object is freed, delete! ] should be used
otherwise use delete without brackets.
An object can also be initialized at the time of creation using new with
the initialized value in parentheses. For example,
double* y = new double (3.14) // *y = 3.14
int i = 5
int* j = new int (i) // *j = 5, but j does not point to i

Declarations of forms T a and T

int* ip
int *ip

are equivalent, as in

// these two declarations are equivalent

However, the following two declarations are equivalent,

int* i, j
int *i, j

// i is a pointer to int but j is an int

each of which declares i to be a pointer to int and j to be an int: This kind

of inconsistency is inherited from C, so that many C functions can be called
from C++. To avoid confusion, it is best to declare a pointer variable in a
single declaration statement.
An array of pointers and a pointer to an array can also be dened:
int* ap%10]
int (*vp)%10]

// ap is an array of 10 pointers to int

// vp is a pointer to an array of 10 int

Notice that parentheses are needed for the second statement above, which
declares vp to be a pointer to an array of 10 integers. The rst statement
declares ap to be an array of 10 pointers, each of which points to an int.

3. Derived Types

3.6.1 Pointer Arithmetic

Some arithmetic operations can be performed on pointers. If pt is a pointer,

then pt represents the object to which pt points. If pt is a pointer and space
has been allocated to store more than one object, then pt points to object
0 (the initial object) and pt + k points to object k (kth object counted
starting from 0). Notice pt + k is a special arithmetic operation that simply
refers to object k in the sequence. The value of object k is (pt + k) or pt!k]:
For example,
char* cc
cc = new char %3]

// cc is a pointer to char
// allocate 3 char for cc
// cc points to initial char

cc%0] = 'a'
cc%1] = 'b'

// or *cc = 'a' cc points to object 0

// or *(cc + 1) = 'b'
// (cc + 1) points to object 1

cc%2] = 'c'

// or *(cc + 2) = 'c'
// (cc + 2) points to object 2

char* p = &cc%0]
char dd = *p
dd = *(cc + 2)

// assign address of cc%0] to p

// dd = 'a', since *p = cc%0]
// (cc+2) points to 'c', now dd = 'c'

delete%] cc

// free space allocated by new

The object pointed to by p is cc!0] and the value of cc!0] is 'a'. Thus the
value of the object pointed to by p that is, p is 'a', which is initialized to
dd: Similarly, cc + 2 points to object 2 in the sequence and thus (cc + 2) is
the value of object 2, which is also referred to by cc!2]: The term element
instead of object is often used when referring to an individual in a sequence,
such as an array.

3.6.2 Multiple Pointers

Two-dimensional arrays and matrices can be achieved through double pointers (a pointer to a pointer is called a double pointer). For example,
int** mx
// double pointer:a pointer to a pointer
mx = new int* %n] // new space to hold n pointers to int
// mx points to initial element mx%0]
for (int i = 0 i < n i++) mx%i] = new int %m]
// create m objects for each of the n pointers
// mx%i] points to initial element mx%i]%0]

3.6 Pointers

The rst statement above declares mx to be a pointer to a pointer, called

a double pointer. The second statement allocates n objects of type int
and assigns the address of the initial element to mx. It happens that these
n objects are pointers to int. Now mx has value &mx!0]: This can be
pictorially depicted in the case of n = 3 as
mx
#

mx!0]
mx!1]
mx!2]
The third statement above (the for loop) allocates m objects of type int for
each of the n pointers mx!i] i = 0 1 : : : n ; 1: A graphical representation
of the double pointer mx in the case of n = 3 and m = 5 can now be drawn
as
mx
#

mx!0] ! mx!0]!0] mx!0]!1] mx!0]!2] mx!0]!3] mx!0]!4]

mx!1] ! mx!1]!0] mx!1]!1] mx!1]!2] mx!1]!3] mx!1]!4]
mx!2] ! mx!2]!0] mx!2]!1] mx!2]!2] mx!2]!3] mx!2]!4]
The process above may be more easily explained using the keyword typedef.
A declaration prexed by typedef introduces a new name for a type rather
than declaring a new variable. For example, the statement
typedef int* intptr

// intptr is synonym to int*

introduces intptr as a synonym to pointer to int: A typedef does not introduce a new type so that intptr means the same as int : Then the double
pointer mx can be equivalently declared as
intptr* mx = new intptr %n]
for (int i = 0 i < n i++) mx%i] = new int %m]

It says that mx is a pointer to intptr and n objects: mx!i], i = 0 1

: : : n ; 1 of type intptr are created by using the new operator. Then m
objects of type int are created for each mx!i]:
Now mx can be used in the same way as a 2D array:
for (int i = 0 i < n i++)
// mx%i] is a pointer to int
for (int j = 0 j < m j++) // used just like a 2D array
mx%i]%j] = i*i + 9
// access element mx%i]%j]

After use, the space can be freed as

3. Derived Types

for (int i = 0 i < n i++)

delete%] mx%i]
delete%] mx

// free space after use

// first free these n pointers
// then free double pointer mx

In the above, n and m can be computed at run-time and space for mx

can be allocated dynamically. This is in contrast to an array that can
not be declared unless its dimension is known at compile-time. Note that
the orders of allocating and deleting space for the double pointer mx are
opposite. The same rule applies to other multidimensional pointers. The
eect of the double pointer mx can be viewed as a matrix: mx is a pointer
that points to the initial object mx!0] of an array of n pointers: mx!i] i =
0 1 : : : n ; 1 (the rows of the matrix). Each of the n pointers subsequently
points to the initial object mx!i]!0] of an array of m integers: mx!i]!j ] j =
0 1 : : : m ; 1 (each row has m elements). The element at row i and column
j can be accessed using mx!i]!j ] the same way as in a 2D array. It can also
be alternatively accessed using the dereferencing operator as: (mx!i] + j )
((mx + i))!j ] ((mx + i) + j ) or (&mx!0]!0] + m i + j ):
Using pointers an n by n lower triangular or symmetric matrix can be
dened very conveniently (to save memory, zero or symmetric elements
above the main diagonal are not stored ):
double** tm = new double* %n]
for (int i = 0 i < n i++) tm%i] = new double %i+1]
// allocate (i+1) elements for row i
for (int i = 0 i < n i++)
// access its elements
for (int j = 0 j <= i j++) tm%i]%j] = 2.1/(i+j+1)
for (int i = 0 i < n i++) delete%] tm%i]
delete%] tm
// after using it, delete space

In the above, tm is created to store an n by n lower triangular matrix.

Since the lower triangular part of a matrix contains i + 1 elements in row i
for i = 0 1 : : : n ; 1 only i + 1 doubles are allocated for tm!i]: A picture
for illustrating this idea can be drawn as (when n = 3)
tm
#

tm!0] ! tm!0]!0]
tm!1] ! tm!1]!0] tm!1]!1]
tm!2] ! tm!2]!0] tm!2]!1] tm!2]!2]
Upper triangular matrices can also be declared using the techniques discussed here and in x3.6.3 see Exercise 3.14.12. Note that arrays can only
represent rectangular matrices. Using rectangular matrices to store triangular matrices or symmetric matrices would waste space.

3.6 Pointers

ar ; = k

ar!0]

ar!1]

ar!n-1]

ar += k
FIGURE 3.1. Pointer osetting. Originally, ar points to initial object ar0]:
Pointer arithmetic ar ; = k osets ar to the left by k positions. Then,
ar += k changes ar back to point to the initial object. That is, after osetting, ark] = (ar + k) is the initial object.

See Exercise 3.14.14 for triple and quadruple pointers, which can be
used to dynamically allocate space for three- and four-dimensional arrays.
Higher-dimensional arrays may be created similarly, but are less often used.
Double, triple, and quadruple pointers, and the like, may be called multiple
pointers.

3.6.3 Osetting Pointers

It is sometimes very useful to use an array of n elements whose index

changes from k to n ; 1 + k where k is a positive or negative integer.
The case of k = 0 corresponds to a regular array, where the index changes
from 0 to n ; 1: The idea is to oset the pointer, k elements away from its
original place. For example, let us rst assume that k is positive:
double* ar =
ar -= k

new double %n]

// ar points to 1st element
// offset ar, k elements to the left

The rst statement allocates space for a sequence of n elements and ar

points to the initial element in the sequence. The second statement is simply
a pointer arithmetic operation, which causes ar to be decremented by k:
In other words, pointer ar is oset by k positions to the left. Now ar
may not point to any valid object at all and ar!0] should never be used.
However, if ar is incremented by k, it should come back and refer to the
initial element in the sequence. That is, now (ar + k) = ar!k] represents
the initial element in the original sequence, and subscripting starting from
k has been achieved. See Figure 3.1 for a pictorial description.
The n objects can now be referred to as
for (int i = k i < n + k i++) {

3. Derived Types
ar%i] = i*i + 1

// now ar%k] is initial element

// ar%n+k-1] is the last element

Note that the objects are not moved at all in this process. Only the pointer

ar is oset (subtracted by k) and changed back (added by k) to point

to the same object. When k is negative, the pointer is actually oset jkj
elements to the right and the code above remains exactly the same. When
the memory is reclaimed, the pointer needs to be changed back rst:
ar += k
delete%] ar

// first change pointer ar back

// then delete it as usual

For example, to store the number of students in classes 1999 and 2000
in an array of 2 elements, the code can be written as
int* ns = new int %2]
ns -= 1999
ns%1999] = 36000
ns%2000] = 37000
ns += 1999
delete%] ns

//
//
//
//
//
//

ns points to initial element

offset ns 1999 elements to left
ns%1999] now is initial element
37000 students in class 2000
change ns back before deleting
delete the space

If an array of 108 elements is needed, but the index must change from
to 7 it can be done as

;100

double* ade = new double

ade += 100
for (int i = - 100 i <=
ade%i] = i + 9.8
ade -= 100
delete%] ade

%108]
// offset
7 i++)
// access
// change
// delete

ade 100 elements to right

its elements
ade back before deleting
the space

This technique is extremely useful and convenient when dealing with

band matrices, where the nonzero elements of a matrix form a band along
the main diagonal. See Chapter 11.

3.6.4 Constant Pointers

A constant pointer is a pointer that can not be redened to point to

another object that is, the pointer itself is a constant. It can be declared
and used as
int m = 1, n = 5
int* const q = &m
q = &n

// q is a const pointer, points to m

// error, constant q can not change

*q = n
int k = m

// ok, value that q points to is now n

// k = 5

3.6 Pointers

Although q is a constant pointer that can only point to object m the value
of the object that q points to can be changed to the value of n, which is 5:
Thus k is initialized to 5:
A related concept is a pointer that points to a constant object that is,
if p is such a pointer, then the value of the object pointed to by p can not
be changed. It only says that p can not be changed explicitly by using it
as lvalue: However, the pointer p itself can be changed to hold the address
of another object. It can be declared and used as
const int* p = &m
*p = n
p = &n

// p points to constant object

// error, *p can not change explicitly
// ok, pointer itself can change

There is some subtlety involved here. Look at the example:

int m = 1, n = 5
const int* p = &m
int i = *p
m = 3
int j = *p
p = &n
int k = *p

//
//
//
//
//
//

p points to m, so *p becomes 1
*p = m = 1, so i = 1
m = 3, so *p becomes 3
*p = 3, so j = 3
ok, p itself can change, *p = 5
*p = n = 5, so k = 5

Since p points to m at rst, the assignment m = 3 changes p to 3: Then

the assignment p = &n changes p to the value of n, which is 5: In other
words, p has been changed implicitly.
To avoid the subtlety above, a const pointer that points to a const object
can be declared:
int m = 1, n = 5
const int* const r = &m
// r is a const pointer that points to a const value
int i = *r
r = &n
*r = n

// i = 1, since *r = m = 1
// error, r is const pointer
// error, r points to const value

m = 3
int j = *r

// this is the only way to change *r

// j = 3

Since r is a const pointer that points to a const value m, it can not be

redened to point to other objects, and r can not be assigned to other
values. The only way to change r now is through changing m, the value
of the object to which r points.

3. Derived Types

3.6.5 Void and Null Pointers

The void pointer (void ) points to an object of unknown type. A pointer of

any type can be assigned to a variable of type void and two variables of
type void can be compared for equality and inequality. It can be declared
and used as
void* pv
int* pi
pv = pi

// pv is a void pointer
// pi is int pointer
// implicit conversion from int* to void*

Its primary use is to dene functions that take arguments of arbitrary types
or return an untyped object. A typical example is the C quicksort function
qsort(), declared in <stdlib:h>, that can sort an array of elements of any
type. See Exercise 3.14.24 or the book !Str97] for more details. However,
templates (see Chapter 7) can be more easily and eciently used to achieve
the same goal.
The null pointer points to no object at all. It is a typed pointer (unlike
the void pointer) and has the integer value 0: If at the time of declaration, a
pointer can not be assigned to a meaningful initial value, it can be initialized
to be the null pointer. For example,
double* dp = 0

// dp initialized to 0, dp is null pointer

This statement is an initialization, which declares dp to be a pointer to

double and initializes dp with the value 0: Since no object is located at
address 0, dp does not point to any object. Later, dp can be assigned to
hold the address of some object of type double:
double d = 55
dp = &d
*dp = 0

// *dp = 55.0
// it causes d = 0.0, dp still points to d

Zero (0) is an integer. Due to standard conversions, 0 can be used as a

constant of any integral, oating point, or pointer type. The type of zero
is determined by context. No object is allocated with the address 0: Thus
0 can be used to indicate that a pointer does not refer to any object. The
negation of a null pointer is true:

3.6.6 Pointers to Structures

A pointer can point to a structure. In this case, its members are accessed
using the ;> operator. Space for structure objects can be allocated by the
operator new and freed by delete: For example,
point2d ab = {'F', 3, -5}
point2d* p = &ab
char name = p->nm
double xpos = p->x

//
//
//
//

ab is of type point2d
let p point to ab
assign nm field of p to name
assign x field of p to xpos

3.6 Pointers
double ypos = p->y

// assign y field of p to ypos

p->x = 15.0
p->y = 26.0
p->nm = 'h'

// ab.x = 15
// ab.y = 26
// ab.nm = 'h'

point2d* q = new point2d

q->x = 5
delete q

// allocate space for q

// assign value to its member
// deallocate space

where the structure point2d is dened in x3.4.

A structure can contain a member that points to the structure itself. For
example,
struct stack {
char* element
stack* next
}

// elements of stack are chars

// points to next element in stack

This is allowed because next is only a pointer and occupies a xed amount
of space (sizeof (int) bytes). Such a self-referential structure is called a
dynamic data structure elements can be allocated and attached to it (or
removed from it and deallocated) using the pointer next: C++ provides a
standard library on stacks see x10.1.4.
To dene two or more structure types that refer to each other, a forward
declaration is needed:
struct BB

// forward declaration, defined later

struct AA {
char c
BB* b
}

// AA has a member pointing to BB

struct BB {
AA* a
int i
}

// definition of structure BB

// b is a pointer to BB

Without the forward declaration of BB , the use of BB in the declaration

of AA would cause a syntax error. The name of a structure or other userdened type can be used before the type is dened as long as the use does
not require the size of the type or the names of its members to be known.
This technique can be used to create many complicated data structures.

3. Derived Types

3.6.7 Pointers to Char

By a convention in C, a string constant is terminated by the null character

'n0', with value 0: Thus the size of "hello" is 6 and its type is const char!6]
with the last element equal to 'n0'. Due to its compatibility to C, C++
allows one to assign a string constant to char directly. For example,
char* str = "hella" // assign string constant to char*
str%4] = 'o'
// error, string constant can not change
char str2%] = "hella" // array of 6 char, sizeof(str2) = 6
str2%4] = 'o'
// OK, now str2 = "hello"
char* str3 = new char %5]
str3%4] = 'o'
// OK
delete%] str3

The C++ standard library <string> can be conveniently used for dealing
with strings see x4.5. C-style strings are better avoided if possible.

3.6.8 Pointers and Arrays

As in C, pointers and arrays are closely related. The name of an array can
be used as a const pointer to its initial element. For example,
int v%5] = {6,9,4,5,7}
int* q = &v%0]
// point to initial element. *q = 6
int* p = v
// point to initial element. *p = 6
int* r = &v%5]
int* s = v + 5

// point to last-plus-one element

// but *r is undefined
// pointer arithmetic, *s is undefined

It is legal to declare a pointer to the last-plus-one element of an array. Since

it does not point to any element of the array, the value that such a pointer
points to is undened. This is useful in writing many algorithms for a
variety of data structures such as vector (x10.1.1) and list (x10.1.2), where
begin() points to the initial element and end() points to the last-plus-one
element see the picture on Page 338.

3.7 References
A reference is an alternative name or alias for an object. For a type T , T &
denotes a reference to T: A reference to an object b of type T is dened as
T& r = b

// r is declared to be a reference to object b

3.8 Functions

Any operation carried out on r is then an operation on b. For example,

int i = 5
int& r = i
int x = r
r = 2

// r is a reference to int and refers to i

// x = 5 since r is an alias of i
// i = 2 since r is an alias of i

Due to the fact that r refers to i, the assignment r = 2 is equivalent to

i = 2:
Since a reference is an alternative name for an object, a reference must
be initialized and can not be changed to refer to another object.
The main use of references is in specifying arguments and return values
for functions and for overloaded operators see x3.8 and Chapter 6.

3.8 Functions
3.8.1 Function Declarations and Denitions

A function is a procedure that may take some input, perform some instructions, and return some output. The input and output must be specied to
be of certain types. For example, one of the sqrt() functions in <math:h>
takes a double as its input, calculates its square root, and returns it as the
output. It is declared as
double sqrt(double)

// take a double, return a double

When a function does not return an output, void is used to denote it.
When the output type is not explicitly stated, it is assumed to be int by
default. Not specifying the input of a function means that it takes no input.
Here are a few more examples of function declarations:
int square(int)
// take an int, return an int
double sum(double, double)
// take two double, return a double
int f()
// take no input and return an int
void g(double, double) // take two double, return nothing
h(int)
// take an int and return an int

In the declarations above, the input (called parameter or argument) and

output (called return value) just have types but no names. The names
may be included for readability, but will be ignored by the compiler. For
example,
int square(int i)
// return i*i
double pow(double base, double exp)
// return base to the power of exp

3. Derived Types

All these declarations are called prototypes: It is only the types that matter
in a prototype.
Every function that is called in a program must be dened somewhere
and only once. The instructions that a function performs must be put inside
a pair of braces. The return statement is used for a function to produce an
output. A function denition also serves as its declaration. For example,
the square() function can be dened as
int square(int x) {
return x*x
}

// definition of a function

A function contains a block of statements and may be called outside

the block. After execution of a function, program control reverts to the
statement immediately following the function call. For example,
int i = 5
int j = square(i)
int k = j

// call function square(i) with i = 5

// control comes here after call, k=25

However, when a function is called before its denition, the function declaration must appear before the statement that calls the function. Such a
declaration is called a forward declaration. For example,
main() {
int square(int)
// forward declaration
int x = square(6) // call square() before its definition
}
int square(int x) {
return x*x
}

// definition of function square()

In the denition of the function square(), x is called a parameter (or

formal argument) to the function. It does not take on any value until
square() is called, when it is replaced by the argument (or actual argument)
from the calling environment, as in j = square(i): In this call, the parameter x is replaced by the argument i: The type of i is also checked against
the type of x and some conversion may be performed when their types do
not match.
The denition of a function can not appear inside another function. For
example, the denition of square() above can not be put inside the main()
function it can be before or after main() or in another le.

3.8.2 Function Overloading

We can dene another square() function that takes a double and returns
a double:

3.8 Functions
double square(double x) {
return x*x
}

// definition of a function

So can we dene yet another taking a long double and returning a long
double. These versions of the square() function can coexist in a program.
The compiler can decide which version to use by comparing the type of the
argument with the type of the parameter. Using the same name for similar
operations on dierent types is called overloading.
Return type is not considered in overloading resolution. For example,
double a =
double b =

square(5)
square(5.0)

// call int square(int)

// call double square(double)

Neither are functions declared in dierent scopes. An example is:

int cubic(int)
double fg(int i) {
double cubic(double)
return cubic(i)
}

// no overloading occurs here

// call double cubic(double)

In call cubic(i), the function cubic(double) declared in the same scope, is

used.

3.8.3 Argument Passing

There are two argument passing mechanisms in C++: pass by value and
pass by reference. In pass by value, the argument is rst evaluated at the
time of the function call, and its value becomes the value of the parameter
during the execution of the function. The default is pass by value, as in the
following example.
int pass_val(int x) {
x = x*x
return x + 5
}

// default is pass by value

int i = 5
int j = pass_val(i)
// now i = 5, j = 30

In the function call j = pass val(i) the value of the argument i which is
5 is passed to the parameter x: The function pass val() starts with the
value x = 5 executes its block of statements, and returns value 30 to j:
Thus after the function call, i = 5 and j = 30:
In pass by reference, an argument must be a variable with allocated
location and the argument is passed to the function as a reference, so that

3. Derived Types

the parameter becomes an alias to the argument and any change made to
the parameter occurs to the argument as well. Pass by reference is signied
by the reference operator &: For example, consider an example of pass by
reference:
int pass_ref(int& x) {
x = x*x
return x + 5
}

// pass by reference

int i = 5
int j = pass_ref(i)
// now i = 25, j = 30

In the function call j = pass ref (i) the argument i is passed as a reference
to the parameter x: That is, parameter x and argument i refer to the same
object and x becomes an alias to i: Thus any change made to x inside the
function also occurs to i: The function pass ref () starts with the value
x = 5 executes its block of statements, and returns value 30 to j: However,
inside the function, x is updated to 25: This update occurs to i as well.
Thus after the function call, i = 25 and j = 30:
Pass by value for pointers can achieve the eect of pass by reference,
since two pointers point to the same object if they have the same value.
For example, compare the following three versions of the swap() function,
which is intended to swap two numbers.
void swap(int& p, int& q) {
int tmp = p
p = q
q = tmp
}
int i = 2, j = 3
swap(i, j)
// now i = 3, j = 2

// pass by reference
// a temporary variable
// swap values of p and q

void swap(int* p, int* q) {

int tmp = *p
*p = *q
*q = tmp
}
i = 2, j = 3
swap(&i, &j)
// now i = 3, j = 2

// pass by value for pointers

void swap2(int p, int q) {

int tmp = p

// pass by value

// addresses of i, j are passed

3.8 Functions

p = q
q = tmp
}
i = 2, j = 3
swap2(i, j)
// now i = 2, j = 3

In the rst function call swap() the references of i and j are passed to
the function, and p and q become aliases to i and j respectively. Thus any
change made to p and q inside the function occur to i and j as well. In the
second swap() the values of &i and &j that is, the addresses of i and j are
passed to the parameters p and q. Thus p points to object i and q points to
object j and swapping p and q is equivalent to swapping i and j: In the
third version: swap2() the values of i and j are passed to the parameters p
and q respectively, and they can not be changed by the function call. This
version is useless since it does not swap the values of i and j at all.
Now look at another function that passes by value for its rst argument
and by reference for its second:
void f(int val, int& ref) {
val++
ref++
}
int i = 1, j = 1
f(i, j)
// now j = 2, i = 1.

Pass by value does not change the value of the arguments and thus is safe.
On the other hand, pass by reference or pass by value for pointers usually
implies that the values of the arguments are going to be changed through
the function call, unless they are explicitly told not to, using the keyword
const: For example,
int g(int val, const int& ref) {
// ref is not supposed to be changed
val++
return ref + val
}

Because of the const specier, the compiler will give a warning if ref is to
be changed inside the function. For example, it is illegal to write
void w(const int& ref) {
ref = 5
// WRONG, ref is not writable
}

Since the const specier implies not to change the parameter ref, it can be
substituted by a constant argument in a call:

3. Derived Types

g(5, 100)

// OK

f(5, 100)
int i = 100
f(5, i)

// error, 2nd arg is not a variable

// OK, 2nd arg must be a variable

The name of an array is a const pointer that points to its initial element.
Thus array arguments are treated as pointers and have the same eect of
pass by reference. For example,
double h(double* const d, int n) { // d is a const pointer
double sum = 0
for (int i = 0 i < n i++) sum += d%i]
d%n - 1] = 1000
// but d%i] can change
return sum
}
double a%] = { 1, 2, 8, 20, -10, 30 }
double sum = h(a, 6)
// sum of elements in a
double d5 = a%5]
// d5 = 1000

Constant pointers that point to constant objects or simply constant

pointers can be used as safeguards to prevent accidental modication of
certain variables that are not supposed to be modied. For example,
double fg(double* const d, int n) { // d is const pointer
d = new double %n]
// error
// ...
}
double gh(const double* const d, int n) {
// d is const pointer that points to constant object
double sum = 0
for (int i = 0 i < n i++) sum += d%i]
d%n - 1] = 1000
// error, d points to constants
return sum
// d%i] can not be changed
}

In the denition of fg(), the parameter d can not be made to point to

another object since d is a const pointer. Using such a parameter indicates
that space should not be allocated inside the function denition. Instead,
the caller to this function should allocate space for it before calling. Similarly, the value of parameter d in function gh() should not be changed in
the denition.

3.8 Functions

3.8.4 Return Values

A function that is not declared void must return a value, except for the
function main() see x3.8.10. For example,
int square(int& i) {
i *= i
// error, a value must be returned.
}

Each time a function is called, new copies of its arguments and automatic
variables (like tmp in the swap() functions) are created. Since automatic
variables will be destroyed upon the return of the function call, it is an error
to return a pointer or reference to an automatic variable. The compiler
normally gives a warning message so that such errors can be easily avoided.
For example,
int* fa() {
// need to return address of int
int local = 5
return &local // error
}
// can not return address of local variable
int& fr() {
// need to return reference to int
int local = 5
return local // error
}
// can not return a reference to local var

However, it is not an error to return the value of a local pointer variable

or the reference to a global variable. For example,
int* f() {
int* local = new int
*local = 5
return local // return value of a local variable
}
int* i = f()
int j = *i
// j = 5
delete i
// free space at address i allocated in f()
int v%30]
int& g(int i) {
return v%i] // it makes g(i) refer to v%i] (v is global)
}
g(3) = 7
// now v%3] = 7, since g(3) refers to v%3]

Observe that objects created by operator new exist on the dynamic memory
(heap, or free store). They still exist after the function in which they are
created is returned. These objects can be explicitly destroyed by calling
delete: The function f () returns the address of the object created by new

3. Derived Types

to the variable i. Thus i gives the value of the object and the operator
delete frees the object at address i: The function g() returns a reference to
a global variable v and thus g(3) refers to v!3]:
For a void function, it is an error to return a value. However, an empty
return statement can be used to terminate a function call, or a call to a
void function can be made in a return statement. For example,
void h(int)
void hh(int& i) {
int k = 0
for (int j = 0 j < 100 j++) {
// do something to modify k and i
if (k == 0) return
// fcn terminates if k is zero
if (k == 2) return h(i)
}
// call void function in return statement
}

As a more meaningful example of a function that returns a pointer, an

n by m matrix multiplication with an m-column vector can be dened and

used:

double* mxvrmy(const double** const mx,

const double* const vr, int n, int m) {
double* tmv = new double %n] // tmv will store product
for (int i = 0 i < n i++) { // find the product
tmv%i] = 0
for (int j = 0 j < m j++) tmv%i] += mx%i]%j]*vr%j]
}
return tmv
// return the address of the product
}
main() {
int n = 100, m = 200
double** a = new double* %n] // create space for matrix
for (int i = 0 i < n i++) a%i] = new double %m]
double* b = new double %m]
// create space for vector
for (int i = 0 i < n i++)
// assign values to matrix
for (int j = 0 j < m j++) a%i]%j] = i*i + j
for (int j = 0 j < m j++) b%j] = 3*j + 5

double* c = mxvrmy(a, b, n, m) // matrix-vector multiply

double sum = 0
// find sum of elements
for (int i = 0 i < n i++) sum += c%i]

3.8 Functions
for (int
delete%]
delete%]
delete%]

i = 0 i < n i++)
a
b
c

delete%] a%i]
// free matrix a
// free vector b
// free vector c

The function mxvrmy() returns (the base address of) the product of the
matrix mx with vector vr: Inside the main() program, some values are
assigned to matrix a and vector b and variable c is initialized to point to
the initial element of the product vector, which is allocated and computed
inside the function mxvrmy(): The summation of all elements in the product vector c is computed and stored in sum: Notice that the space for the
product vector is allocated using new inside the function mxvrmy(), but
is freed using delete in main():

3.8.5 Recursive Functions

A function that calls itself in its denition is said to be a recursivefunction.

For example, a function that calculates the factorial n! of a nonnegative
integer n can be dened recursively:
long factorial(int n) {
if (n == 0) return 1L
return n*factorial(n-1)
}

// stopping statement

Here factorial(n) is recursively dened as n factorial(n ; 1) and the

recursion stops when it reaches factorial(0): Thus a call like factorial(2)
can be interpreted as

factorial(2) = 2 factorial(1) = 2 1 factorial(0) = 2 1 1 = 2:

Note that without the stopping statement: if (n == 0) return 1L the
program would cause an innite recursion.
Recursive functions may impose a lot of run-time overhead compared to
nonrecursive functions. For example, a recursive function call factorial(2)
requires actually three function calls (factorial(2), factorial(1), and factorial(0)). The overhead of function calls can signicantly slow down a
program when there are a large number of them.

3.8.6 Inline Functions

When a function is called, its arguments are copied and the program
branches out to a new location (on the run-time stack see x3.9) and comes
back after the instructions contained in the function are executed. To avoid

3. Derived Types

such function calling overhead, inline functions can be dened using the
keyword inline for small functions. Statements in an inline function are
substituted in the program calling it during compilation. For example, an
inline function can be dened as
inline int twice (int i) { return 2*i }

If somewhere in the program there is a statement like y = twice(x) then the

content of twice(x) will be substituted. That is, the compiler will generate
the statement y = 2 x and a run-time function call will be avoided.
This can also be achieved as normally done in C by a macro:
#define twice(i) (2*(i))

Then every occurrence of twice(x) will be substituted by (2 (x)) in the

compiling process. Compared to C-style macros, inline functions are typechecked and safer to use.
However, the inline specier is only a hint to the compiler that it should
attempt to generate code substitution. The possibility of mutually recursive
inline functions and other situations make it impossible to guarantee that
every call of an inline function is actually inlined. When it can not be
inlined, the usual function call mechanism will be used and the keyword
inline will be ignored.

3.8.7 Default Arguments

A function can have default arguments for its trailing arguments. A default
value for an argument can be specied in a function denition. For example,
void print(int v, int b = 10) { // b is defaulted to be 10
// define the function to print out integer v in base b
}
print(15,16)
print(15,2)
print(15,10)
print(15)

//
//
//
//

print
print
print
print

15
15
15
15

in
in
in
in

hexadecimal (base 16)

binary (base 2)
decimal (base 10)
decimal (base 10)

The last call is equivalent to print(15 10): That is, when a value is not
explicitly given for the second argument, 10 is assumed by default.
The default values for arguments can also be specied in a function
declaration. For example,
void print(int = 100, int = 10)
print(15,16)
print(15,2)
print(15)

// 2 default arguments

// print 15 in hexadecimal (base 16)

// print 15 in binary (base 2)
// print 15 in decimal (base 10)

3.8 Functions
print()

// print 100 in decimal (base 10)

void print(int v, int b) {

// define the function to print out integer v in base b
}

Note that default arguments are possible only for trailing arguments.
The compiler will be confused at
void print(int b = 10, int)

// error

3.8.8 Function Types

It is sometimes useful to dene a type synonym using the keyword typedef.

A declaration prexed by typedef introduces a new name for a type rather
than declaring a new variable. For example,
typedef unsigned int Uint
Uint i = 5

// Uint is unsigned int

// i is an unsigned int

typedef double (*coef)(double)

// coef means a pointer to function
coef f
// f has type of coef

A typedef does not introduce a new type. For example, Uint and unsigned
int are the same type, and coef is the same type as a pointer to a function
that takes a double and returns a double: The last statement above declares
f to be a variable of type coef. That is, f can take on values that are pointers
to a function that takes a double and returns a double:
A pointer to a function can be passed as an argument to another function
like any other pointer. For example,
double integral(coef f, double a, double b){
return f((a+b)*0.5)*(b-a) // an inaccurate evaluation
}
// of a definite integral
double square(double d) { return d*d }
coef sfp = square
// initialize sfp with `square()'
double v = integral(sfp, 2, 2.01)
// call integral()
sfp = sqrt
// assign sqrt in <math.h> to sfp
v = integral(sfp, 2, 2.01)
// call integral()

The function call integral(square 2 2:01) now gives an approximate value

of the denite integral of the function square() from 2 to 2.01. The name
of a function is treated as a pointer to the function.
Here is a more complicated example from a C header le <signal:h>:

3. Derived Types

typedef void (*SIG_TYP) (int)

// from <signal.h>
typedef void (*SIG_ARG_TYP) (int) // pointer-to-fcn type
SIG_TYP signal(int, SIG_ARG_TYP) // return pointer to fcn

This denes SIG TY P and SIG ARG TY P to be the type of pointer to

functions that take an int as argument and do not return a value, and
signal is a function returning a value of type SIG TY P and taking an int
and a SIG ARG TY P as arguments.

3.8.9 Static Local Variables

Variables dened within the body of a function are automatic variables

by default. They lose their value after the function call is completed. This
property can be overridden by use of the keyword static: A static variable,
although still local, persists across function invocations and maintains its
value after the function call is completed. Consider a function that calculates a value only when it is rst called and just returns the value without
repeating the calculation in subsequent calls:
long double cal() {
static bool first = true
static long double v
if (first == true) {
v = 5.0
// or perform some calculation to get v
first = false
}
return v
}
long double x = cal()
// 1st call, the 'if' block is executed in cal()
long double y = cal()
// 2nd call, the 'if' block is skipped in cal()

The value of the variable first is initialized to true when the function
cal() is rst called. Then it is assigned to have a new value false inside the
function. This new value will be retained and the initialization statement
static bool first = true

will be ignored in subsequent calls. Since rst = false, the if statement

in cal() is then skipped. The same explanation holds for the static local
variable v. For this reason, assignment v = 5:0 is executed only in the rst
call.
The value of a static variable can also change with each function call.
For example,
int gcd(int m, int n) {

// greatest common divisor

3.8 Functions

static int callnum = 1

cout << "gcd() has been called " << callnum++
<< " times so far.\n"
if (n == 0) return m
return gcd(n,m%n)
}

When gcd() is rst called, callnum = 1 and callnum is then given a new
value 2: This value will be maintained when the second time gcd() is called
and the initialization int callnum = 1 will be ignored. In the second call,
callnum is incremented by 1 to a new value 3: With the static variable
callnum one can know how many times a function is called in a program.
For example, in calculating the greatest common divisor gcd(215 335) the
recursive function need be called eight times. However, look at the following
two calls of the function.
int i = gcd(215,335)
int j = gcd(215,335)

// after this call, callnum = 8

// after this call, callnum = 16

At the beginning of the second call j = gcd(215 335) the value of callnum
is 9 (from the rst call i = gcd(215 335)): The value of callnum will be
incremented from 9 to 16 during the second call, since the static variable
callnum persists throughout the program.
A static variable provides a function with \memory" without introducing
a global variable that might be accidentally accessed and corrupted by other
functions.

3.8.10 The Function main

The minimal C++ program is
int main() { }

It takes no arguments and does nothing. Every C++ program must have
a function called main(): The program starts by executing this function.
The int value returned to the system indicates the program has completed
successfully or otherwise. If no value is explicitly returned, the system will
receive a zero value for successful completion and nonzero value for failure.
The function main() may take two arguments specifying the number of
arguments, usually called argc, and the array of arguments, usually called
argv for the program. The arguments are character strings having type
char !argc +1]: The name of the program is passed as argv!0] and the list
of arguments is zero-terminated that is, argv!argc] = 0.
For example, consider the simple program in x1.1. Instead of inputting
two integers after the program is run, we want to input them at the
UNIX/Linux command line as

3. Derived Types

add 1 1000

Here add is the name of the executable object code for the program. Then
add is stored in argv!0] and 1 and 1000 will be stored in argv!1] and argv!2],
respectively. The number of arguments is argc = 3: Note that argv!3] is 0
indicating the end of arguments. Now we need to modify the program to
support this input format:
#include <iostream>
#include <stdlib.h>
// library for function atoi()
int main(int argc, char* argv%]) {
if (argc != 3)
cout << "number of arguments is incorrect.\n"
int n = atoi(argv%1])
int m = atoi(argv%2])

// first input int assigned to n

// 2nd input int assigned to m

if (n
int
n =
m =
}

//
//
//
//

> m) {
temp = n
m
temp

if n > m, swap them

declare temp, initialize it
assign value of m to n
assign value of temp to m

double sum = 0
// sum initialized to 0
for (int i = n i <= m i++) sum += i
cout << "Sum of integers from " << n << " to " << m
<< " is: " << sum << '\n'
// output sum to screen
}

Note that the function atoi() in the standard header <stdlib:h> converts
a character string into an integer. A related function atof () converts a
character string into a oat.

3.9 Program Execution

When a program is run, the execution is typically arranged as shown in
Table 3.1. It consists of several portions.
The text portion contains the actual machine instructions that are
executable by the hardware. These machine instructions are generated by the compiler from the source code of the program. When
a program is executed by the operating system, the text portion is
typically read into memory from its disk le.
The data portion contains the program's data, which can be further
divided into initialized data (initialized by the program) and unini-

3.9 Program Execution

..
.
.
text initialized uninitialized heap ..
..
.
data
data

..
.
..
. stack
..
.

TABLE 3.1. A typical arrangement of a program execution.

tialized data. Uninitialized data contain items that are not initialized
by the program but are set to a default value (typically zero) before
the program starts execution. A program le consists of the program
text and initialized data. The advantage of providing an uninitialized data area is that the system does not have to allocate space in
the program le on the disk for uninitialized global variables. It also
reduces the time to read a program le from disk into memory.
The heap is used to allocate more data space dynamically for the
program while it is being executed.
The stack is used dynamically to contain the stack frames that include the return linkage for each function call and also the data elements required by the function.
A gap between the heap and the stack indicates that many operating
systems leave some room between these two portions so that they can grow
dynamically during a program execution. When there is no more room for
these two portions to grow, a run-time error will occur.

As an example, consider the following short program:

const int param = 10000
double myvalue
double f() {
double* a = new double %param]
for (int i = 0 i < param i++) a%i] = rand() + i
double sum = 0
for (int i = 0 i < param i++) sum += a%i]
return sum
}
int main() {
myvalue = f()
double s = sqrt(myvalue)
cout << "result is: " << s + 555 << '\n'
}

3. Derived Types

The variable param is stored in the initialized data area and myvalue
in uninitialized data. The constants 10000 555 and \result is: " (string
literal) can be stored as (read-only) initialized data. The variables a i and
sum are automatic variables and will be stored on the stack. When the
function f () is called in the statement myvalue = f () the system creates
the objects a i and sum on the stack and 10000 objects pointed to by a on
the heap. After the computation in the function call is nished, the objects
a i and sum are destroyed and the computed value of sum is returned
to the variable myvalue: However, the 10000 objects created on the heap
are not explicitly deleted and thus still exist throughout the rest of the
program execution. The automatic variable s is also stored on the stack
while the function main() is being executed. The machine instructions for
the functions main() f (), sqrt() and rand() are in the text segment.

3.10 Operator Summary and Precedence

This section presents a summary of operators and their precedence, some
of which are explained later in this book. Each operator is followed by
its name and an example illustrating its use. Tables 3.2 and 3.3 contain a
summary of all C++ operators and their precedence, where a class name
is the name of a class, structure, or union, and a member or Mem is the
name of one of its members. An object is an expression yielding an object,
a pointer an expression yielding a pointer, an expr or Er is an expression,
an lvalue is an expression denoting a nonconstant object, and a type or
Tp can be a basic type or a derived type (with , ( ), etc.) with possible
restrictions.
In Tables 3.2 and 3.3, each box holds operators with the same precedence.
Operators presented in a higher box have higher precedence than operators
in lower boxes. For example, the operator has higher precedence than ;
and a ; b c means a ; (b c) instead of (a ; b) c: Similarly, i <= 0 jj 100 < i
means (i <= 0) jj (100 < i) instead of the legal but nonsense expression
i <= (0 jj 100 < i): Note that there are cases when the operator precedence
does not result in obvious interpretation, such as
if (i & mask == 0 )

{ /* do something */ }

Because of the higher precedence of == over the `bitwise and' operator &
the expression is interpreted as i & (mask == 0): In this case, parentheses
should be used to yield the desired result:
if ( (i & mask) == 0 )

{ /* do something */ }

3.10 Operator Summary and Precedence

Operator Summary and Precedence

Name
Example
scope resolution
class name :: Mem
scope resolution
namespace name :: Mem
global operator
:: name
member selection
object.member
;>
member selection
pointer;>member
!]
subscripting
pointer!expr], array!expr]
()
function call
expr(expr list)
()
value construction
type(expr list)
++
postincrement
lvalue++
;;
postdecrement
lvalue;;
()
type identication
typeid(type)
()
type identication
typeid(expr)
dynamic cast
run-time cast
dynamic cast<Tp> (Er)
static cast
compile-time cast
static cast<Tp> (Er)
reinterpret cast unchecked cast
reinterpret cast<Tp>(Er)
const cast
const cast
const cast<Tp>(Er)
sizeof
size of object
sizeof (expr)
sizeof
size of type
sizeof (type)
++
preincrement
++lvalue
;;
predecrement
;;lvalue
~
bitwise complement
~expr
!
not (negation)
!expr
;
unary minus
; expr
+
unary plus
+ expr
&
address of
&lvalue

dereference
*expr
new
allocate one object
new type
new
allocate objects
new type !expr]
new
allocate, initialize
new type (expr)
new
create (place an object) new (expr) type
new
create (place objects)
new (expr) type !expr]
new
create (place, initialize) new (expr) type (expr)
delete
deallocate one object
delete pointer
delete
deallocate objects
delete! ] pointer
()
cast (type conversion) (type) expr
.
member selection
object.pointer to Mem
;>
member selection
pointer;>pointer to Mem

multiply
expr expr
=
divide
expr = expr
%
modulo (reminder)
expr % expr
Operator
::
::
::
.

TABLE 3.2. Operator summary and precedence.

3. Derived Types

Operator Summary and Precedence (continued)

Operator Name
Example
+
add (plus)
expr + expr
;
subtract (minus)
expr ; expr
<<
shift left
expr << expr
>>
shift right
expr >> expr
<
less than
expr < expr
<=
less than or equal to
expr <= expr
>
greater than
expr > expr
>=
greater than or equal to expr >= expr
==
equal
expr == expr
!=
not equal
expr != expr
&
bitwise AND
expr & expr
^
bitwise exclusive OR
expr^expr
j
bitwise inclusive OR
expr j expr
&&
logic AND
expr && expr
jj
logic inclusive OR
expr jj expr
=
simple assignment
lvalue = expr
=
multiply assignment
lvalue = expr
==
divide assignment
lvalue == expr
%=
modulo assignment
lvalue %= expr
+=
add assignment
lvalue += expr
;=
subtract assignment
lvalue ;= expr
<<=
shift left assignment
lvalue <<= expr
>>=
shift right assignment
lvalue >>= expr
&=
AND assignment
lvalue &= expr
^=
exclusive OR assignment lvalue^= expr
j=
inclusive OR assignment lvalue j = expr
?:
conditional expression
expr ? expr: expr
throw
throw exception
throw expr
,
comma (sequencing)
expr, expr
TABLE 3.3. Continuation of operator summary and precedence.

3.10 Operator Summary and Precedence

Another example is:

if (0 > i || i > 100 && abs(i) < 500) { /* ... */ }

which is equivalent to
if (0 > i || (i > 100 && abs(i) < 500))

{ /* ... */ }

where the function abs(i) gives the absolute value of an integer i (declared
in the standard header <stdlib:h>): Note that it is dierent from
if ((0 > i || i > 100) && abs(i) < 500)

{ /* ... */ }

This can be easily checked by taking i = ;1000 in the two statements.

Also, the mathematical expression 0 x < 100 can be translated into the
C++ expression
0 <= x && x < 100

or equivalently (0 <= x) && (x < 100) but never

0 <= x < 100

When in doubt, parentheses should be used it may also increase readability

of the code.
All operators are left associative, except for unary operators and assignment operators, which are right associative. For example, a ; b ; c means
(a ; b) ; c and ajjbjjc means (ajjb)jjc but a = b = c means a = (b = c)
and p++ means (p++) rather than (p)++ (i.e., q = p++ means
q = p p++).
A few grammar rules can not be expressed in terms of operator precedence and associativity. For example, the expression a = b < c ? d = e :
f = g means a = ((b < c)?(d = e) : (f = g)):
The order of evaluation of subexpressions in an expression is undened
in general. For example, in the statement
int i = f(x) + g(y)

// undefined order of evaluation

it is undened whether f (x) or g(y) is called rst and the result of

int i = 5
v%i] = i++

// undefined results

is unpredictable (it may be v!5] = 5 or v!6] = 5 depending on whether

the left-hand operand or right-hand operand is evaluated rst). A program
should not assume any order of evaluation for such expressions.
However, the three operators , (comma), && (logical and), and jj (logical
or) guarantee that their left-hand operand is evaluated before the righthand operand. See x2.2.5 and x2.2.3.
The following keywords

3. Derived Types

and
and eq bitand bitor compl not
or eq xor xor eq
not eq or
are provided for some non-ASCII character sets, where characters such
as & and^are not available. They are equivalent to the following tokens,
respectively.
&&
!=

&=
jj

&
j=

~
^=

However, they can also be used even with ASCII character sets for clarity.
For example, not eq may be more readable than != to some people.

3.11 Standard Library on Mathematical Functions

The standard library <math:h> provides the following mathematical functions.
double abs(double)
// absolute value, not in C same as C's fabs()
double fabs(double)
// absolute value, inherited from C
double ceil(double d)
// smallest integral value not less than d
double floor(double d)
// largest integer not greater than d
double sqrt(double d)
// square root of d, d must be nonnegative
double pow(double d, double e)
// d to the power of e
// error if d ==0 and e<= 0
// and error if d < 0 and e is not an int
double pow(double d, int i)
// d to the power of i, not in C
double cos(double)
double sin(double)
double tan(double)

// cosine
// sine
// tangent

double
double
double
double

//
//
//
//

acos(double)
asin(double)
atan(double)
atan2(double x, double y)

arc cosine
arc sine
arc tangent
atan(x/y)

3.11 Standard Library on Mathematical Functions

double cosh(double)
double sinh(double)
double tanh(double)

// hyperbolic cosine
// hyperbolic sine
// hyperbolic tangent

double acosh(double)
double asinh(double)
double atanh(double)

// inverse hyperbolic cosine

// inverse hyperbolic sine
// inverse hyperbolic tangent

double exp(double)
double log(double d)

//
//
//
double log2(double d) //

natural exponential, base e

natural (base e) logarithm
d must be > 0
base 10 logarithm, d must be > 0

double modf(double d, double* p)

// return fractional part of d, and store
// integral part of d in *p
double fmod(double d, double m)
// floating point remainder, same sign as d
double ldexp(double d, int i)
// d*exp(2,i)
double frexp(double d, int* p)
// return mantissa of d, put exponent in *p

In addition, <math:h> supplies these functions for float and long double
arguments. For complex arguments with real and imaginary parts in float
double and long double precisions, see x7.4. The angle returned by the
inverse trigonometric functions asin() acos(), atan() and atan2() are in
radians. The range of asin() and atan() is !;=2 =2]: The range of acos()
is !0 ]: The range of atan2() is !; ] whose principal use is to convert Cartesian coordinates into polar coordinates. For asin() and acos(), a
domain error occurs if the argument is not in !;1 1]:
The function frexp(double d int p) splits a value d into mantissa x and
exponent p such that d = x 2p and returns x where the magnitude of x is
in the interval !0:5 1) or x is zero. The function fmod(double d double m)
returns a oating modulus d(mod m): If m is nonzero, the value d ; i m
is returned, where i is an integer such that the result is zero, or has the
same sign as d and magnitude less than the magnitude of m: If m is zero,
the return value is system-dependent but typically zero.
Errors are reported by setting the variable errno in the standard library
<errno:h> to EDOM for a domain error and to ERANGE for a range
error, where EDOM and ERANGE are macros dened in <errno:h>.
For example,
int main() {
errno = 0
// clear old error state, include <errno.h>
double x = - 50
// x can also be computed at run-time

3. Derived Types
double i = sqrt(x)
if (errno == EDOM)
cerr << "sqrt() not defined for negative numbers\n"
double p = pow(numeric_limits<double>::max(), 2)
if (errno == ERANGE)
cerr << "result of pow(double,int) overflows\n"

where cerr represents the standard error stream (usually the terminal
screen), dened in <iostream>. Thus, by checking the value of the variable
errno, the user can know if a domain or range error has occurred in calling
standard functions in <math:h>.
For historical reasons, a few mathematical functions are in <stdlib:h>
rather than <math:h>:
int abs(int)
long abs(long)
long labs(long)

// absolute value
// absolute value, not in C
// absolute value

struct div_t { implementation_defined quot, rem}

struct ldiv_t { implementation_defined quot, rem}
div_t div(int n, int d)
// divide n by d, return (quotient, remainder)
ldiv_t div(long int n, long int d)
// divide n by d, return (quotient, remainder), not in C
ldiv_t ldiv(long int n, long int d)
// divide n by d, return (quotient, remainder)
double atof(const char* p)
int atoi(const char* p)
long atol(const char* p)

// convert p to double
// convert p to int
// convert p to long int

In the denition of the structures div t and ldiv t the type of quot and rem
is implementation-dependent. The functions atof , atoi, and atol convert
strings representing numeric values into numeric values, with the leading
whitespace ignored. If the string does not represent a number, zero is returned. For example, atoi("seven") = 0 atoi("77") = 77, while atoi('s')
gives the numeric value representing the lowercase letter s.

3.12 Polynomial Evaluation

In many scientic applications, the value of an nth degree polynomial

pn (x) = an xn + an;1 xn;1 + + a1 x + a0

3.12 Polynomial Evaluation

needs to be evaluated for given values of x where the coecients an an;1
: : : a1 a0 are known. A C++ function can be written to do this job having
prototype
double eval(double* a, int n, double x)

where a is a pointer that points to the coecients a0 a1 : : : an;1 an : An

inner for loop may be used to do the multiplication ak xk and an outer for
loop for adding all the terms. Thus the function eval() can be straightforwardly dened as
double eval(double* a, int n, double x) {
double sum = 0
for (int k = n k >= 0 k--) {
double xpowerk= 1
for (int i = 1 i <= k i++) {
// or call pow(x, k)
xpowerk *= x
// compute x to power of k
}
sum += a%k]*xpowerk
// add each term to sum
}
return sum
}

Notice that evaluating the rst term an xn requires n multiplications and

the second term an;1 xn;1 requires n ; 1 multiplications, and so on. Thus
the total number of multiplications to evaluate pn (x) using the function
call eval(x) is n + (n ; 1) + + 2 + 1 = n(n + 1)=2: This method is very
inecient since it duplicates the work in computing the power of x for each
term of pn (x):
An ecient method called Horner's algorithm (or nested multiplication)
is described below. The idea is to rewrite pn (x) in a nested multiplication
form:
pn (x) = ( ((an x + an;1 )x + an;2 )x + + a1 )x + a0 :
An example of writing a polynomial of degree 4 from its regular form into
nested multiplication form is:
4x4 + 3x3 + 2x2 + x + 10 = (((4x + 3)x + 2)x + 1)x + 10:
Introduce the notation:
un = an
un;1 = un x + an;1
un;2 = un;1 x + an;2
..
.
u1 = u2 x + a1
u0 = u1 x + a0

3. Derived Types

Thus pn (x) = u0 : At rst glance, un un;1 , : : : and u0 would be stored in

an array. However, it is not necessary since after computing un;1 the value
un is no longer needed, and so on. This procedure can be written into an
algorithm.
Algorithm 3.12.1 horner(a n x)
Initialize u = an :
for (i = n ; 1 n ; 2 : : : 0) f
u ux + ai
g

return u:
It takes only n multiplications. Now Horner's algorithm can be coded into
C++ as

double horner(double* a, int n, double x) {

double u = a%n]
for (int i = n - 1 i >= 0 i--) u = u*x + a%i]
return u
}

This program is not only shorter, but also more ecient than the straightforward approach eval(). Ecient algorithms require fewer operations and
may sometimes result in smaller accumulative roundo errors.
Now these two functions can be compared by evaluating an 8th degree
polynomial:
p8 (x) = x8 ; 8x7 + 28x6 ; 56x5 + 70x4 ; 56x3 + 28x2 ; 8x + 1
for x taking on 11 equally spaced values covering interval !0:99999 1:0001]:
Observe that p8 (x) = (x ; 1)8 : This expression will be compared to the
values returned by eval() and horner():
#include <iostream>
#include <math.h>
int main() {
double a%9] = {1, - 8, 28, - 56, 70, - 56, 28, - 8, 1}
for (double x = 0.99999 x <= 1.0001 x += 1.0e-5) {
cout.width(18)
cout << eval(a,9,x)
// straightforward evaluation
cout.width(18)
cout << horner(a,9,x) // Horner's algorithm
cout.width(18)
cout << pow(x-1,8) << '\n'
// direct evaluation
}
}

3.12 Polynomial Evaluation

The output on a computer with 8 bytes for double is as follows.

5.32907e-15
0
3.55271e-15
0
0
-2.66454e-15
4.44089e-15
-6.21725e-15
-5.32907e-15
8.88178e-15
2.66454e-15

2.14564e-16
0
1.32321e-15
-1.59367e-15
-1.76064e-15
-6.9931e-17
-3.5203e-15
-3.23331e-15
-1.87686e-15
-2.31586e-16
5.89264e-16

1e-40
0
1e-40
2.56e-38
6.561e-37
6.5536e-36
3.90625e-35
1.67962e-34
5.7648e-34
1.67772e-33
4.30467e-33

The correct value for p8 (x) = (x ; 1)8 should be nonnegative and very
small since x is close to 1: The third column is the best since it is a direct
calculation of (x ; 1)8 and will be used as exact values in the comparison.
Comparing the rst column (produced by eval()) with the second column
(produced by horner()), the results by Horner's algorithm are more accurate (closer to the third column) for 8 out of 11 values. Both eval() and
horner() produce negative values due to the subtractions of nearly equal
quantities in their evaluations.
Let us now enlarge the errors by 1010 times in the evaluations of eval()
and horner() so that their accuracies can be more clearly seen. Modify the
main program as
int main() {
double a%9] = {1, - 8, 28, - 56, 70, - 56, 28, - 8, 1}
for (double x = 0.99999 x <= 1.0001 x += 1.0e-5) {
cout.width(20)
cout << 1.0e10*(pow(x-1,8) - eval(a,9,x))
cout.width(20)
cout << 1.0e10*(pow(x-1,8) - horner(a,9,x)) << '\n'
}
}

The output becomes:

-5.32907e-05
0
-3.55271e-05
2.56e-28
6.561e-27
2.66454e-05
-4.44089e-05
6.21725e-05
5.32907e-05

-2.14564e-06
0
-1.32321e-05
1.59367e-05
1.76064e-05
6.9931e-07
3.5203e-05
3.23331e-05
1.87686e-05

3. Derived Types
-8.88178e-05
-2.66454e-05

2.31586e-06
-5.89264e-06

For 9 out of 11 values, the second column is no larger than the rst.
In large-scale applications, it is not enough to compute correctly, but
also eciently and elegantly. The elegance of a program should include extendibility, maintainability, and readability, while correctness should also
include accuracy, robustness, and portability. Both horner() and eval() are
correct (horner() is slightly more accurate in certain cases), but the eciency of horner() is far better for polynomials of high degrees. In nance,
evaluation of high-degree polynomials is often encountered see Exercises
3.14.18 and 4.8.12.

3.13 Trapezoidal and Simpson's Rules

In this section, a complete
example is presented to evaluate denite inR
tegrals of the form ab f (x)dx using the Trapezoidal Rule and Simpson's
Rule, based on partitioning the interval !a b] into n subintervals, where n
is a positive integer. We dene two functions trapezoidal() and simpson()
that take the integral bounds a and b the integrand f , and the number
of subintervals n as arguments, and return a value for the approximate
integral.
We want to write a main program as
#include <math.h>
#include <iostream>
typedef double (*pfn)(double) // define pfn for integrand
double trapezoidal(double a, double b, pfn f, int n)
double simpson(double a, double b, pfn f, int n)
double square(double d) { return d*d }
int main(){
double result = trapezoidal(0, 5, square, 100)
cout << "Integral using trapezoidal with n = 100 is: "
<< result << '\n'
result = simpson(0, 5, square, 100)
cout << "Integral using simpson with n = 100 is: "
<< result << '\n'
result = trapezoidal(0, 5, sqrt, 100)
cout << "Integral using trapezoidal with n = 100 is: "

3.13 Trapezoidal and Simpson's Rules

y = f (x)

6
f (xi )

f (xi;1 )

xi;1

FIGURE 3.2. The integral xxii;1 f (x)dx, representing the area of the region under
curve y = f (x) and above subinterval xi;1 xi ] is approximated by the area of
a trapezoid (with one edge drawn in dotted line), which is f (xi;1 ) + f (xi )]h=2:
<< result << '\n'
result = simpson(0, 5, sqrt, 100)
cout << "Integral using simpson with n = 100 is: "
<< result << '\n'
}

Here sqrt() is the square-root function declaredR in <math:h>

. This main
R5p
5 2
program intends to nd approximate values for 0 x dx and 0 xdx using
the Trapezoidal and Simpson's Rules. Since these integrals can be evaluated
exactly, they can be used as examples for testing the correctness of the
program. An easy modication
of the programR enables
one to nd more
p
R
complicated integrals such as 01 sin(x2 )dx and 01 1 ; sin3 xdx:
Now we need to dene the functions for the Trapezoidal and Simpson's
Rules. The idea of the Trapezoidal Rule is to rst partition the given interval !a b] into n (assuming equally spaced for simplicity) subintervals
!x0 x1 ] !x1 x2 ] : : : !xn;1 xn ] with xi ; xi;1 = hR where
h is the size
of each subinterval. Then approximate the integral xx ;1 f (x)dx on each
subinterval !xi;1 xi ] by the area of a trapezoid passing through four points
(xi;1 0), (xi;1 f (xi;1 )) (xi f (xi )) and (xi 0) (assuming f (xi;1 ) and
f (xi ) are positive, but the result also applies to general functions). The
area of such a trapezoid is (f (xi;1 ) + f (xi )) h=2. See Figure 3.2.
i

Thus,
n Z
b f (x)dx = X xi f (x)dx
a
i=1 xi;1

100

3. Derived Types

i=1

;1
f (xi;1 ) + f (xi ) h = f (x0 ) + nX
f (xn ) h
f
(
x
)
+
i
2
2
2
"

i=1

where the subinterval size h = (b ; a)=n and partitioning points are xi =

a + i h i = 0 1 : : : n with x0 = a and xn = b: This function can be
dened as

double trapezoidal(double a, double b, pfn f, int n) {

double h = (b - a)/n
// size of each subinterval
double sum = f(a)*0.5
for (int i = 1 i < n i++) sum += f(a + i*h)
sum += f(b)*0.5
return sum*h
}

Simpson's RRule is similar to the Trapezoidal Rule but approximates

the integral xx ;1 f (x)dx on each subinterval !xi;1 xi ] by the area under a quadratic curve (represented by a polynomial of degree two) passing through three points (xi;1 f (xi;1 )) (#xi f (#xi )) and (xi f (xi )) where
x#i = (xi;1 + xi )=2 is the middle point of the subinterval !xi;1 xi ]. It can
be shown that the area of such a trapezoid with one edge as a quadratic
curve is (f (xi;1 ) + 4f (#xi ) + f (xi ))h=6: Thus,
i

n Z
b f (x)dx = X xi f (x)dx
a
i=1 xi;1
n f (x ) + 4f (#x ) + f (x )
X
i;1
i
i h

6
i=1
n
n
X
X
= 31 f (xi;1 )2+ f (xi ) h + 32 f (#xi )h:
i=1
i=1

Notice that the rst summation in the formula is the same as the Trapezoidal Rule. Now it can be dened as
double simpson(double a, double b, pfn f, int n) {
double h = (b - a)/n
double sum = f(a)*0.5
for (int i = 1 i < n i++) sum += f(a + i*h)
sum += f(b)*0.5
double summid = 0.0
for (int i = 1 i <= n i++) summid += f(a + (i-0.5)*h)
return (sum + 2*summid)*h/3.0
}

3.14 Exercises

101

In x5.1, an abstract mechanism called class is used to present this example in a dierent style. Notice that this C-style denition of trapezoidal()
and simpson() passes a function pointer f for the integrand as an argument, which imposes function calling overhead for each call f () inside
trapezoidal() and simpson(): In x7.7 some ecient techniques are applied
to overcome the function calling overhead in passing a function pointer to
another function call, where integration functions are written with a type
parameter so that they can also handle arithmetics in dierent precisions
such as oat, double, and long double. For more details on mathematical
aspects of the Trapezoidal and Simpson's Rules, see !Ste99, KC96, VPR00].

3.14 Exercises
3.14.1. Write a program that denes a const integer n generates n random
numbers and stores them in an array, and computes and outputs
the average of these random numbers. Also compute and output the
maximum and minimum values of the random numbers generated.
3.14.2. The standard library <assert:h> contains a macro assert() that can
be used to ensure that the value of an expression is as expected. Here
is an example of how assert() can be used to ensure a variable n is
positive, but less than or equal to 100:
#include <assert.h>
void f(int n) {
assert(n > 0 && n <= 100)
}

If an assertion fails, the system will print out a message and abort the
program. When the macro NDEBUG is dened at the point where
<assert:h> is included, for example,
#define NDEBUG
#include <assert.h>

then all assertions are ignored. Thus assertions can be freely used during program development and later discarded (for run-time eciency)
by dening NDEBUG: Write a program to generate Fibonacci numbers (see x2.4) ensuring that they are positive using an assertion.
3.14.3. Write a function having prototype
int sumsq(int n, int m)

102

3.14.4.
3.14.5.

3.14.6.

3.14.7.

3. Derived Types

that returns the sum of squares n2 +(n +1)2 + +(m ; 1)2 + m2 for
n < m: The function should also be able to handle the case n > m:
Write a function that computes and returns the number of (decimal)
digits of a given integer n. For example, it should return 3 if n = 145:
Write a function
p outputs (e.g., the rst 1000 terms of) the sep that
quence dn = n ; n ; 1 for n = 1 2 3 : : :: Write another function
that computes dn in a mathematically equivalent but numerically
more robust way:
dn = p 1p
for n = 1 2 3 : : ::
n+ n;1
Compare the dierence in the outputs of the two functions. The second one should give more accurate results since it avoids subtracting
two nearly equal numbers and thus reduces cancellation errors.
An algorithm to nd a square root of a positive real number b is:

x0 = 1 xn+1 = 12 xn + xb n = 0 1 2 : : ::
n
Write a function that implements the algorithm.
A fast (quadratic)
p
convergence should be observed for xn ! b: Compare the accuracy
of this function with the built-in function sqrt() in <math:h>.
Write a function that takes three doubles a b and c as input and
computes the roots of the quadratic equation
ax2 + bx + c = 0:
When a = 0 and b 6= 0 the equation has only one root x = ;c=b:
When a 6= 0 its roots are given by
p

x1 = (;b + b2 ; 4ac)=(2a) x2 = (;b ; b2 ; 4ac)=(2a):

If the discriminant b2 ; 4ac is negative, the two roots are complex
numbers.
Output complex numbers in the form of r + mi, where
p
i = ;1 and r and m are real and imaginary parts. Write a main()

program to test your function.

3.14.8. The statements
a%i++] *= n
a%i++] = a%i++]*n

are not equivalent. The operand a!i ++] has a side eect of incrementing i: Except for possible implementation eciency, the rst
statement is equivalent to

3.14 Exercises

103

a%i] = a%i]*n
i = i + 1

while the second statement is equivalent to either

a%i+1] = a%i]*n
i = i + 2

or
a%i] = a%i+1]*n
i = i + 2

depending on whether the right-hand side of the assignment is evaluated before the left-hand side. The order of their evaluations is not
dened. Test what result your machine gives for the second statement
above. Such a statement is not portable and should be avoided.
3.14.9. Allocate space for a sequence, denoted by s of 2n + 1 doubles so that
the index changes from ;n to n where n is a given positive integer.
Assign s!i] = i for i = ;n ;(n ; 1) : : : ;1 0 1 : : : n ; 1 n and
print out the values of s!i] ;n i n for n = 10: Finally, delete
the space.
3.14.10. Explain the dierence between a two-dimensional array and a double
pointer. Consider further an example of a two-dimensional array:
double tda%5]%7]

The base address of the array tda is &tda!0]!0] or tda!0] but not tda:
The array name tda by itself is equivalent to &tda!0]: What can you
say for a double pointer and a one-dimensional array?
3.14.11. Write a function that takes an n by m matrix and a vector of m
components as input and computes the matrix-vector product such
that it has prototype:
void multiply(const double** const mx,
const double* const vr, int n,
int m, double* const pt)

Here the double pointer mx represents the matrix, single pointer vr

the input vector, and pt the matrix-vector product. Write a main()
program to test the function. Space should have been allocated for
mx vr and pt before calling the function.

104

3. Derived Types

3.14.12. Declare an n by n upper triangular matrix utm of floats using a

double pointer. To save memory, store only the upper part of the
matrix. The entry at row i and column j should be accessed by using
utm!i]!j ] for i = 0 1 : : : n ; 1 and j = i i + 1 : : : n ; 1: Then
assign utm!i]!j ] = 2i + j for n = 4 and print out all the entries.
Finally, deallocate space for utm: To save memory, space should not
be allocated for the lower triangular part of the matrix.
3.14.13. Write a function that takes an n by n upper triangular matrix and
a vector of n components as input and computes the matrix-vector
product. It may have either of the prototypes
double* multiplyutm(const double** const utm,
const double* const vr, int n)
void multiplyutm(const double** const utm,
const double* const vr,
int n, double* const pt)

Implement the matrix as in Exercise 3.14.12 and the vector as a

pointer. Write a main() program to test the function.
3.14.14. A three-dimensional array with dimensions p q r using a triple
pointer can be allocated and deallocated as
int p = 4, q = 2, r = 3

// or computed at run-time

double* p3d = new double %p] // allocate space

for (int i = 0 i < p i++) {
p3d%i] = new double* %q]
for (int j = 0 j < q j++) p3d%i]%j] = new double %r]
}
for (int i = 0 i < p i++)
// access its entries
for (int j = 0 j < q j++)
for (int k = 0 k < r k++) p3d%i]%j]%k] = i + j + k
for (int i = 0 i < p i++) { // delete space after use
for (int j = 0 j < q j++) delete%] p3d%i]%j]
delete%] p3d%i]
}
delete%] p3d

Note that the order of memory deallocation should be the opposite of

allocation. Allocate space for a four-dimensional array of dimensions
p q r s using a quadruple pointer:

3.14 Exercises

105

double**** p4d

Assign it some values such as p4d!i]!j ]!k]!m] = i ; j + k ; m: Finally,

deallocate the space.
3.14.15. Write a recursive function having the prototype
long fibonacci(int n)

that takes an integer n and calculates the Fibonacci number fn see

x2.4 for the denition of Fibonacci numbers. Also output the number
of recursions needed to obtain fn :
3.14.16. Write a function to print out a decimal integer v in base b having the
prototype:
void print(int v, int b = 10)

// b defaulted to be 10

with 2 b 16: Your program should print out an error message

when a user attempts to input b outside this range. Hint: an integer v
can be expressed in base b as an an;1 a1 a0 if v = a0 + a1 b + a2 b2 +
+ an bn with 0 ai < b i = 0 1 : : : n and an 6= 0:
3.14.17. Write the polynomial
p8 (x) = x8 ; 8x7 + 28x6 ; 56x5 + 70x4 ; 56x3 + 28x2 ; 8x + 1
directly into a nested multiplication form using the idea of Horner's
algorithm. Then print out values of p8 (x) for x taking on 11 equally
spaced values covering the interval !0:99999 1:0001]: You should get
the same values as the function horner() dened in x3.12.
3.14.18. A man has just started a xed-rate mortgage of total amount of S
dollars with annual percentage rate r and length of m months. But he
decides to pay o the mortgage early and pay P1 P2 , : : : and Pn;1
dollars for the rst, second, : : : and (n ; 1)th month, respectively,
and pay o the mortgage at the end of the nth month with n m:
The payo amount Pn can be computed from the formula:

nX
;1

;i

;n
r
r
Pi 1 + 12 + Pn 1 + 12

i=1
assuming that each Pi is no smaller than his regular monthly payment

so that no penalty is imposed on him. Write a function that computes

the payo amount Pn , given S , r, P1 , P2 : : : and Pn;1 : If the man
has a 15-year mortgage of 200000 dollars at rate 0:0625 but pays Pi =
2000 + 10 i dollars in month i for i = 1 2 : : : 59 and wants to pay
o the mortgage in month 60 what is his payo amount P60 ? Hint:
Horner's algorithm should be used in calculating the summation.

106

3. Derived Types

3.14.19. Apply the idea of Horner's algorithm to derive ecient algorithms

for evaluating

Bn =

bj = b0 + b0 b1 + b0 b1 b2 + + b0 b1 bn
i=0 j =0
n Y
i 1
X
1 + 1 + 1 + +
1
Dn =
=
d0 d1 dn :
i=0 j =0 dj d0 d0 d1 d0 d1 d2

The algorithms should require only n multiplications for Bn and n +1

divisions for Dn , in addition to n additions for each. Then write
functions having prototypes
double summtp(const double* const b, int n)
double sumdiv(const double* const d, int n)

for them, where b is a pointer that points to the numbers b0 b1 : : :

bn;1 bn and d points to d0 d1 : : : dn;1 dn : Use both functions to
compute the partial sums:

Em =

m 1

n=2 n!

for m = 100 500 and 1000: Hint: your answer should be closer to
0.71828182845904523536 as m gets larger.
3.14.20. The base of the natural algorithm e is
e = 2:7182818284590452353602874713526624977572 : : :
to 41 signicant digits. It is known that
n

x = 1 + 1 ! e as n ! 1:
n

Write a function that generates the sequence fxn g and prints out the
error e ; xn : A very slow (worse than linear) convergence should be
observed for xn ! e: Indeed it can be shown that je ; xn+1 j=je ;
xn j ! 1 as n ! 1: A computationally eective way to compute e to
arbitrary precision is to use the innite series:
1
X
e = n1! = 2 + 2!1 + 3!1 + 4!1 + :
n=0
To reduce roundo errors, smaller terms should be added rst in
calculating partial sums. It can be shown that

m 1

X
e :

<
e ;

n
!
(
m
+
1)!
n=0

3.14 Exercises

107

Write another function to generate the partial sums for approximating e and compare the convergence. Horner's algorithm may be used
to eciently compute this kind of partial sum.
3.14.21. When n is large, the factorial n! can easily cause overow. Write
a function to calculate n! by storing a large integer in an array of
integers, whose one element stores one digit of the large integer. To
illustrate the idea, consider the program:
#include <iostream>
// for input and output
#include <stdlib.h>
// for function exit()
const int mxdgs = 9500 // max # of digits in factorial
int main() {
int n
cout << "Enter a positive integer: \n"
cin >> n
// digits of n! will be stored in array dgs backwards,
// ie, highest digit will be stored in dgs%0], next
// highest in dgs%1], and so on.
int dgs%mxdgs]
dgs%0] = 1
for (int i = 1 i < mxdgs i++) dgs%i] = 0
for (int k = 2 k <= n k++) { // n! =2*3*4...(n-1)*n
int cary = 0
for (int j = 0 j < mxdgs j++) {
// multiply k with every digit of (k-1)! to get k!
dgs%j] = k*dgs%j] + cary
cary = dgs%j]/10
dgs%j] -= 10*cary
}
}
// counting the number of digits in n!
int count = mxdgs - 1
while (!dgs%count]) count--
// compute the number of digits in integer n
int numdgofn = 0
int m = n
while (m) {
m /= 10
numdgofn++
}

108

3. Derived Types
// check to see if size of dgs is enough
if (mxdgs - count < numdgofn + 1) {
cout << "factorial overflow, adjust parameter.\n"
exit(1)
// exit program
}
// printing n!: lowest digit dgs%count] first,
// highest digit dgs%0] last.
cout << "There are " << count + 1 << " digits in "
<< n << "!. They are:\n"
for (int i = count i >= 0 i--) cout << dgs%i]
cout << '\n'
}

The standard library function exit() declared in <stdlib:h>, is called

(and returns value 1) to terminate the program when the dgs array
is not large enough to store n!: In this program, the factorial (k ; 1)!
is stored digit by digit in an array of integers. To get k! multiply
every element of the array (every digit of (k ; 1)!) by k: The result is
still stored in the array. Looping k from 2 through n gives n!: Notice
that the digits of k! are stored in the array backwards. For example,
the three digits 120 of 5! are stored backwards as 021000 000 with
dgs!0] = 0 dgs!1] = 2 dgs!2] = 1 dgs!3] = 0 : : : dgs!mxdgs;1] = 0:
With this technique, 3000!, which has 9131 digits, can be calculated
correctly on a computer with sizeof (int) = 4: To compute the factorial of a larger (than 3000) integer, the size of the array can be
adjusted.
Write a factorial function having prototype:
facs factorial(int n, int maxdgs)

where facs is a structure

struct {
int* digits
int numdgs
}

// store the digits of factorial

// number of digits of factorial

This function returns a struct that contains the number of digits of

n! and the digits of n!. It should print out an error message and exit
when n! has more than maxdgs digits.
3.14.22. ApplyR the 2Trapezoidal and Simpson's Rules to evaluate the integral 02 e;x dx where e is the base of natural logarithm. Output
the approximate values corresponding to the number of subintervals

3.14 Exercises

109

n = 100 300 and 500: Compare your answers with 0:8820813907

(obtained from a handbook).
R

3.14.23. Compute the integral 12 ((sin x)=x)dx with ve digit accuracy after
the decimal point.
3.14.24. The standard library <stdlib:h> contains a sorting function qsort()
based on the quick sort algorithm. It can sort an array of any elements
in an order dened by the user through a pointer to a function. It
has prototype:
typedef int (*CMF)(const void* p, const void* q)
// a type for comparing p and q, it defines an order
void qsort(void* b, size_t n, size_t sz, CMF cmp)
// sort "n" elements of array "b" in an order
// defined by comparison function "cmp".
// Each element of "b" is of size "sz" (in bytes).

Here size t is an unsigned integral type (possibly unsigned long)

dened in <stddef:h> and used in the standard libraries.
If an array of point2d (see x3.4) is going to be sorted in increasing
order according to its x-coordinate, a comparison function can be
dened as
int cmp2dx(const void* r, const void* s) {
double d = static_cast<const point2d*>(r)->x static_cast<const point2d*>(s)->x
if (d > 0) return 1
else if (d < 0) return - 1
else return 0
}

The operator static cast converts between related types such as from
one pointer type to another or from an enumeration to an integral type. Since r has type const void, it needs to be converted
to const point2d before accessing its member x.
Now arrays of point2d can be generated and sorted by qsort() in the
order dened by cmp2dx() :
int main() {
const int n = 10
point2d a%n]
for (int i = 0 i < n i++) {
a%i].x = i*(5 - i) a%i].y = 5/(i+2.3) - 1
}

110

3. Derived Types
qsort(a, n, sizeof(point2d), cmp2dx)
for (int i = 0 i < n i++)
cout << a%i].x << " " << a%i].y << '\n'
}

For people who are curious about how such a sorting function is
dened, here is a denition of a related (based on the shell sort algorithm see !Knu98] and !Str97]) function:
void shsort(void* b, size_t n, size_t sz, CMF cmp) {
for (int gap = n/2 gap > 0 gap /= 2)
for (int i = gap i < n i++)
for (int j = i - gap j >= 0 j -= gap) {
char* base = static_cast<char*>(b)
char* pj = base + j*sz
char* pjpg = base + (j + gap)*sz
if (cmp(pj, pjpg) > 0) { // swap b%j], b%j+gap]
for (int k = 0 k < sz k++) {
char temp = pj%k]
pj%k] = pjpg%k]
pjpg%k] = temp
}
}
}
}

Write a program that sorts an array of point2d in decreasing order

according to its y-coordinate by using the standard library function
qsort() or the user-dened function shsort() above. Note that C++
provides a more ecient mechanism for dealing with objects of a
generic type see Chapters 7 and 10.
3.14.25. Generate an array of random numbers and mimic Exercise 3.14.24 to
sort this array in decreasing order in terms of absolute value.
3.14.26. Consider the declaration:
int* (*(*a)%10])(double)

What is the type of a? It can be analyzed by substitution as follows.

Let b = (a)!10]: Then the declaration can be written as
int* (b)(double)

3.14 Exercises

111

Then b is a function that takes a double and returns a pointer to an

int: Since b = (a)!10] this relation can be written into a declaration:
b

*(*a)%10]

where b is a type just claried above. It is equivalent to

(*a)%10]

Thus a is a pointer to an array of 10 elements, each of which is a

pointer to type b: Finally it can be concluded that a is a pointer
to an array of 10 elements, each of which is a pointer to a function
that takes a double and returns a pointer to an int: This substitution
method can be rigorously justied in C++ by using the keyword
typedef (why?).
Apply this idea to explain the type of x in the declaration:
double (*(*x())%10])(int)

using substitutions y = (x())!10] and z = x(): Hint: x is a function

that takes no arguments and returns a pointer to type z where z is
an array of 10 elements of type y and y is a pointer to a function
that takes an int and returns a double: Such confusing declarations
should be avoided in practice but are presented here for a better
understanding of pointers, arrays, and functions.

112

3. Derived Types

This is page 113

Printer: Opaque this

Namespaces and Files

C++ supports modular programming through a mechanism called namespace. A namespace is a logical unit that contains related declarations and
denitions. The idea of modular programming is to divide a large program
into small and logically related parts for easy management and information hiding. Dividing a large program into dierent parts and storing them
in dierent les can also help to achieve modular programming (this is
more so in C and FORTRAN 77). A few useful tools (some in UNIX and
Linux) are also presented for managing les, creating a library, proling a
program, debugging a program, and timing a program. Towards the end
of the chapter, two standard libraries on character strings and input and
output streams are given. Finally, iterative algorithms for solving nonlinear
equations are described.

4.1 Namespaces
The keyword namespace is used to group related declarations and denitions in C++. For example, we can group vector-related declarations in
a namespace called Vec and matrix-related declarations in a namespace
called Mat:
namespace Vec {
const int maxsize = 100000
double onenorm(const double* const, int)
double twonorm(const double* const, int)

// 1-norm
// 2-norm

114

4. Namespaces and Files

double maxnorm(const double* const, int)

// max norm

}
namespace Mat {
double onenorm(const double** const, int)
double twonorm(const double** const, int)
double maxnorm(const double** const, int)
double frobnorm(const double** const, int)
}

//
//
//
//

1-norm
2-norm
max norm
Frobenius

For a vector v = !v0 v1 : : : vn;1 ] its one, two, and maximum (also
denoted by 1) norms are dened as
nX
;1

kv k1

= jv0 j + jv1 j + + jvn;1 j =

kv k2

n;
= jv0 j2 + jv1 j2 + + jvn;1 j = =
jvi j2

kv k1

i=0

jvi j
v
u
1
uX
t

2 1 2

i=0

= maxfjv0 j jv1 j : : : jvn;1 jg = 0max

jv j
i<n i

respectively. For example, for the vector u4 = !1 ;2 0 10]

ku4 k1

= j1j + j ; 2j + j0j + j10j = 13

p
;
ku4 k2 = j1j2 + j ; 2j2 + j0j2 + j10j2 )1=2 = 105
ku4 k1 = maxfj1j j ; 2j j0j j10jg = 10:
1
For an n by n matrix A = (aij )nij;=0
, its one, 1, and Frobenius norms
are dened as

kAk1

= max

kAk1
kAkF

= max

jai0 j

i=0
n;1

X
j =0

nX
;1

jai1 j : : :

i=0
nX
;1

ja0j j

j =0

nX
;1

i=0
nX
;1

ja1j j : : :

jain;1 j

j =0

= ja00 j2 + ja01 j2 + + jan;1n;1 j

B3 = 4

= 0max
j<n

nX
;1

= 0max
i<n

v
u
1
1
uX X
t
j

2 1 2

1 ;2 0
;4 3 5 5
2 0 6

jaij j

i=0
nX
;1

jan;1j j

respectively. For example, for the matrix

n; n;

i=0 j =0

j =0

jaij j

aij j2

4.1 Namespaces

115

its one, max, and Frobenius norms are:

kB3 k1

= maxfj1j + j ; 4j + j2j j ; 2j + j3j + j0j j0j + j5j + j6jg

= maxf7 5 11g = 11
kB3 k1 = maxfj1j + j ; 2j + j0j j ; 4j + j3j + j5j j2j + j0j + j6jg
= maxf3 12 8g = 12
p
kB3 kF = j1j2 + j; 2j2 + j0j2 + j; 4j2 + j3j2 + j5j2 + j2j2 + j0j2 + j6j2
p
= 95:
That is, the 1-norm of a matrix is the maximum of the absolute column
sums, 1-norm is the maximum of the absolute row sums, and the Frobenius
norm is the square root of the square sum of all entries. The 2-norm of
a matrix is more involved and not mentioned any further. These norms
provide certain measurements of the magnitude of a vector or matrix.
A namespace is a scope. Members of a namespace can be accessed by
using the :: operator. In particular, the functions of namespace Vec can be
dened this way.
double Vec::onenorm(const double* const v, int size) {
if (size > maxsize) cout << "vector size too large.\n"
double norm = fabs(v%0])
// fabs() in <math.h>
for (int i = 1 i < size i++) norm += fabs(v%i])
return norm
}
double Vec::maxnorm(const double* const v, int size) {
if (size > maxsize) cout << "vector size too large.\n"
double norm = fabs(v%0])
for (int i=1 i<size i++) norm = max(norm, fabs(v%i]))
return norm
// max() in <algorithm>
}
double Vec::twonorm(const double* const v, int size) {
if (size > maxsize) cout << "vector size too large.\n"
double norm = v%0]*v%0]
for (int i = 1 i < size i++) norm += v%i]*v%i]
return sqrt(norm)
}

The function fabs() declared in <math:h>, gives the absolute value of its
argument and max() declared in <algorithm>, returns the maximum of its
two arguments. The name maxsize can be used directly without using the
qualication operator :: inside the denitions of members of namespace Vec.
This is because maxsize is a member of Vec. Note that in the third function
denition above, the variable norm could cause over ow while sqrt(norm)

116

4. Namespaces and Files

is still within the range of the machine. A robust 2-norm formula is given
in Exercise 4.8.5 that will not cause over ow if kvk2 is within range.
Functions for the matrix namespace Mat can be dened similarly. However, when members from another namespace are needed, the qualication
operator :: should be used. For example, use Vec::onenorm to refer to the
function onenorm() in the namespace Vec. In this way, name clashes can
be avoided. That is, members of dierent namespaces can have the same
names. For example,
double Mat::maxnorm(const double** const a, int size) {
double norm = Vec::onenorm(a%0],size)
// 1-norm of row 0
for (int i = 1 i < size i++)
norm = max(norm, Vec::onenorm(a%i],size)) //a%i]: row i
return norm
}
double Mat::onenorm(const double** const a, int size) {
double norm = 0
for (int j = 0 j < size j++) {
double temp = 0
// store column abs sum
for (int i = 0 i < size i++) temp += fabs(a%i]%j])
norm = max(norm, temp)
}
return norm
}
double Mat::frobnorm(const double** const a, int size) {
double norm = 0
for (int i = 0 i < size i++)
for (int j = 0 j < size j++) norm += a%i]%j]*a%i]%j]
return sqrt(norm)
}

See Exercise 4.8.6 for a robust denition of the Frobenius matrix norm.
However, for matrices whose entries are not large enough to cause over ow
when summing the squares of all entries, this straightforward evaluation of
the Frobenius norm should be more ecient.
Now the namespaces Vec and Mat can be applied to evaluate norms of
matrices and vectors:
// declarations and definitions of Vec and Mat may be put here
int main() {
int n = 1000
double** a = new double* %n]
for (int i = 0 i < n i++) a%i] = new double %n]

4.1 Namespaces

117

for (int i = 0 i < n i++)

for (int j = 0 j < n j++) a%i]%j] = 1.0/(i + j + 1)
cout << Mat::maxnorm(a, n) << '\n'
cout << Vec::maxnorm(a%2], n) << '\n'

// matrix max norm

// vector max norm

Notice that a!2] refers to row 2 of matrix a which can be treated as a

1
vector. A matrix Hn = (hij )nij;=0
with entries hij = 1=(i + j + 1) is called
a Hilbert matrix with dimension n:
The declarations in a namespace are called the interface part and the
denitions of the members are called the implementation part. A normal
user need only see the interface and gure out how to use it. The implementation part is usually placed somewhere else. For large programs, it has
the advantage that the user will not be confused by the implementation
details and can just concentrate on the interface. The larger a program is,
the more useful namespaces are to express logical separation of its parts.
Since a namespace is a scope, the usual scope rules hold for it. If a name
is previously declared in a namespace, it can be used directly inside the
namespace and its members. For example, the member maxsize of Vec can
be used directly in the denition of Vec::onenorm(). A name from another
namespace can be used when qualied by the name of its namespace. For
example, the name of namespace Vec must be used to qualify its function
onenorm() in order for it to be recognized in namespace Mat:

4.1.1 Using Declarations and Directives

When namespace Mat or other programs have to use members of namespace Vec intensively, repeated qualication using Vec:: can be tedious and
distracting, especially if a namespace has a long name. This qualication
can be eliminated by applying the keyword using:
double Mat::maxnorm(const double** const a, int size) {
using Vec::onenorm
// use onenorm from namespace Vec
double norm = onenorm(a%0],size) // Vec::onenorm is used
for (int i = 1 i < size i++)
norm = max(norm, onenorm(a%i],size))
return norm
}

The using declaration brings Vec::onenorm into the current scope and then
it can be used directly without qualifying by the name of its namespace.
Note that a function name overloading also happened here. Since Vec's
function onenorm() takes dierent arguments from Mat :: onenorm() the
compiler can gure out that Vec::onenorm() is called. If both Vec and Mat

118

4. Namespaces and Files

had exactly the same members, the using-declaration would cause a name
clash.
If a member of Mat depends on several members of Vec, then the usingdeclaration has to be used for each member. This can be replaced by the
using-directive:
namespace Mat {
double onenorm(const double** const, int)
double twonorm(const double** const, int)
double maxnorm(const double** const, int)
double frobnorm(const double** const, int)
using namespace Vec

//
//
//
//

1-norm
2-norm
max norm
Frobenius

// make all names from Vec available

The using directive brings all members of namespace Vec into the scope of
namespace Mat: Then all names of Vec can be used directly in namespace
Mat and the denitions of its members.
By the using-directive for namespace Vec, the code for members of Mat
can then be simplied as if everything of Vec had the same scope. For
example, Mat :: maxnorm() can be dened as:
double Mat::maxnorm(const double** const a, int size) {
double norm = onenorm(a%0],size) // Vec::onenorm is used
for (int i = 1 i < size i++)
norm = max(norm, onenorm(a%i],size))
return norm
}

If two namespaces have two functions that have the same prototype and
both namespaces are used in a third namespace, it is possible to choose one
function over the other and to avoid name clash. For example,
namespace A {
int f(int)
}
namespace B {
int f(int)
}
namespace C {
using namespace A
using namespace B
using A::f
int g(int i) {

// use A's member f() instead of B::f()

4.1 Namespaces
return i + f(i)

119

// A::f(i) is called

}
int h(int i) {
return i + B::f(i)
}

// B::f(i) is called

In the denition of C :: g(), the function A :: f () is used because of the using

declaration of A :: f: Here the functions g() and h() are dened inside the
namespace C . Generally speaking, this is not preferred since it makes the
interface of namespace C obscure, especially when C has a lot of members.
Namespaces enable codes written by dierent people to be combined
easily. For example, two programmers might have used the same name for
their variables with dierent meanings. Codes written by dierent people
can be grouped into dierent namespaces and name clashes can be avoided.
The using-declaration and using-directive can be used to easily access
names in dierent namespaces when names do not clash.

4.1.2 Multiple Interfaces

If the namespace Mat needs to be expanded to include new function, data,

or type members, we can combine old and new members to form a new
version:
namespace Mat {
double onenorm(const double** const, int)
double twonorm(const double** const, int)
double maxnorm(const double** const, int)
double frobnorm(const double** const, int)

//
//
//
//

1-norm
2-norm
max norm
Frobenius

double* eigenvalue(const double** const, int)

// find its eigenvalues
double** inverse(const double** const, int)
// find inverse matrix
void gausselim(const double** const, double* const, int)
// Gauss elimination
}

These two versions of the namespace Mat can coexist in a program so that
each can be used where it is most appropriate. For example, for a user
who just needs to evaluate matrix norms, the rst version is sucient to
bring the functions of Mat into her scope (including what she does not
need may be confusing). Minimal interfaces lead to systems that are easier
to understand, have better information hiding, are easier to modify, and
compile faster.

120

4. Namespaces and Files

The second version is equivalent to:

namespace Mat {
double* eigenvalue(const double** const, int)
// find its eigenvalues
double** inverse(const double** const, int)
// find inverse matrix
void gausselim(const double** const, double* const, int)
// Gauss elimination
}

provided the rst version of Mat has been declared. That is, a namespace
can have multiple interfaces, but the compiler will combine all of them
together for the namespace. Another example is:
namespace A {
struct pt { double x double y }
int f(pt)
// A has two members pt and f()
}
namespace A {
void g()
}

// now A has three members pt, f(), and g()

The second version of namespace A is equivalent to:

namespace A {
struct pt { double x double y }
int f(pt)
void g()
}

4.1.3 Namespace Alias

Aliases for long namespace names can be created. For example,

namespace Numerical_Integration {
double lowerbound
double (*integrand)(double)
// .....
}

// name is very long

// lower integral bound
// integrand function

Numerical_Integration::lowerbound = 1.0

The long name of this namespace takes time to type and is not quite
readable. A short name NI can be provided for the namespace:
namespace NI = Numerical_Integration

4.1 Namespaces

121

NI::lowerbound = 1.0
namespace my_module {
using NI::lowerbound
// .....
}

The short name NI is very convenient to use.

4.1.4 Unnamed Namespaces

Sometimes we do not want the name of a namespace to be known outside

a local context. An unnamed namespace, which is a namespace without a
name, can be used. For example,
namespace {
int a
void f(int& i) { i++ }
int g() { return 5 }
}

Its members are global names in the le (more precisely, translation unit
see x4.3) where it is dened. It is equivalent to
namespace anyname {
int a
void f(int& i) { i++ }
int g() { return 5 }
}
using namespace anyname

where anyname is some name unique to the scope in which the namespace
is dened. Unnamed namespaces in dierent translation units (les see
x4.3) are distinct. There is no way to refer to a member of an unnamed
namespace from another translation unit. Unnamed namespaces can be
used to avoid using one of the meanings of static in C programs (see x4.3.2).

4.1.5 Name Lookup

If a function can not be found in the scope where it is used, looking up of

the namespaces of its arguments will follow. For example,
namespace Matrix {
struct mat {
double** elem
int size
}
double maxnorm(const mat& a) {/* it may be defined here */}

122

4. Namespaces and Files

double onenorm(const mat& a)

}
double f(Matrix::mat a) {
double m = maxnorm(a)
double n = onenorm(a)
return m+n
}

// Matrix::maxnorm() is called
// Matrix::onenorm() is called

In the call maxnorm(a) the function maxnorm() is not dened in the

scope. Then the compiler looks for it in its argument a's namespace Matrix:
It can also be explicitly qualied by the name of the namespace as Matrix ::
maxnorm(a): This lookup rule not only saves a lot of typing compared with
explicit qualication, but also makes the code look more concise.

4.2 Include Files

This section deals with various issues on how to include standard library
header les and user-dened header les.

4.2.1 Include Files for Standard Libraries

The way in which a C program includes its standard libraries' header les
is still valid in C++. For example,
#include <math.h>
// math library, inherited from C
#include <stdio.h>
// I/O library, from C
int main() {
printf("Hello World!\n")
// prints string to screen
printf("square root of pi is: %f", sqrt(atan(1)*4))
}

where the output function printf () has the prototype (it is an example of
a function that takes an arbitrary number of arguments, indicated by an
ellipsis):
int printf(const char* ...)

// from <stdio.h>

It prints out strings and values of variables to the screen according to some
specied format. Here %f means to print out the value of sqrt(atan(1) 4)
as a float and % is a format control character.
All C++ standard libraries, including those inherited from C, are in a
namespace called std: However, there is no namespace in C. The declarations of standard C I/O facilities from the C header le <stdio:h> are
wrapped up in the namespace std like this:

4.2 Include Files

123

// in header file stdio.h

namespace std {
int printf(const char* ...)
// ..... declarations of C's other stuff in stdio.h
}
using namespace std

The using-directive above makes every declaration in stdio:h globally available in every le that includes the header stdio:h:
Similarly, the declarations in C's <math:h> le are put in the namespace
std, and made globally available. New C++ header les such as <cstdio>
and <cmath> (note: without the .h sux) are dened for users who do
not want the names globally available. For example, the C++ header le
cstdio looks like
// in header file cstdio
namespace std {
int printf(const char* ...)
// ..... declarations of C's other stuff in cstdio
}

Without the using-directive as opposed to <stdio:h>, all the declarations

in <cstdio> are restricted to the namespace std:
In general, for each C standard library <X:h>, there is a corresponding
standard C++ header <cX>, inherited from C. If we use C++'s headers,
our C code above should go this way:
#include <cmath>
// C++ math library
#include <cstdio>
// C++ I/O (input/output) library
int main() {
std::printf("Hello World!\n") // print string to screen
std::printf("square root of pi is: %f",
std::sqrt(std::atan(1)*4))
}

That is, explicit namespace qualication such as std :: printf () and std ::
sqrt() must be used here. If we do not want the namespace qualier std ::
we can either use C-style include les with a sux :h or do this:
#include <cmath>
// C++ math library
#include <cstdio>
// C++ I/O (input/output) library
int main() {
using namespace std // make declarations in std available
printf("Hello World!\n")
// print string to screen
printf("square root of pi is: %f", sqrt(atan(1)*4))
}

124

4. Namespaces and Files

<assert:h>
<ctype:h>
<errno:h>
<float:h>
<limits:h>
<locale:h>
<math:h>
<setjmp:h>
<signal:h>
<stdarg:h>
<stddef:h>
<stdio:h>
<stdlib:h>
<string:h>
<time:h>

diagnostics, denes assert() macro

character handling
errors and error handling
oating point limits
integer limits
localization
mathematics
nonlocal jumps
signal handling
variable arguments
common denitions
standard input and output
general utilities
string handling
date and time

TABLE 4.1. Standard C header les.

Note that C++ has an alternative input and output library with headers
<iostream> and <fstream>: Details of the standard libraries <iostream>
and <fstream> are talked about in x4.6.
Table 4.1 lists standard C header les, while Tables 4.2 and 4.3 contain
standard C++ header les.

4.2.2 User's Own Include Files

By convention, standard libraries are included using angle brackets, and

user's own header les are included using double quotation marks. For
example, if a user's own header le my:h is like this:
// in file my.h
namespace MatVec {
// .....
}
struct point2d {
double x, y
}

// declarations of matrix vector library

// note semicolon is needed after struct

then this header le should be included in another le using double quotes:
// in file my.cc
#include "my.h"

// user's own header file

4.2 Include Files

general algorithms
set of Booleans
diagnostics, denes assert() macro
C-style character handling
C-style error handling
C-style oating point limits
C-style integer limits
C-style localization
mathematical functions (real numbers)
complex numbers and functions
nonlocal jumps
C-style signal handling
variable arguments
common denitions
C-style standard input and output
general utilities
C-style string handling
C-style date and time
C-style wide-character string functions
wide-character classication
double-ended queue
exception handling
streams to and from les
function objects
input and output stream manipulators
standard iostream bases
forward declarations of I/O facilities
standard iostream objects and operations
input stream
iterators and iterator support
numeric limits, dierent from <climits>
doubly linked list
represent cultural dierences
associative array
allocators for containers
dynamic memory management
numeric algorithms
output stream
queue
set
streams to and from strings

TABLE 4.2. Standard C++ header les.

125

126

4. Namespaces and Files

standard exceptions
stack
stream buers
string, dierent from <cstring>
run-time type identication
operators and pairs
numeric vectors and operations
one-dimensional array

TABLE 4.3. Continuation of Standard C++ header les.

#include <cmath>
#include <iostream>

// C++ math library

// C++ I/O (input and output) library

int main() {
point2d A
{
using namespace std // cmath, iostream available here
A.x = sqrt(5.0)
A.y = log(5.0)
cout << "A.x = " << A.x << " " << "A.y = " << A.y
}
}

Spaces are signicant within " " or < > of an include directive. For
example,
#include " my.h"
#include <iostream >

will not nd the header les "my:h" and <iostream>.

A simple way to partition a large program into several les is to put the
denitions in a number of :cc les and to declare the types needed for them
to communicate in a :h le that each :cc le includes. As a rule of thumb,
a header le may contain items listed in Table 4.4, with an example in the
second column. Conversely, a header le should not contain items listed in
Table 4.5 (also with an example in the second column). Some of the items
in these two tables are discussed later in this book.

4.2.3 Conditional Include Directives

The directive

4.2 Include Files

named namespaces
type denitions
data declarations
constant denitions
enumerations
function declarations
inline function denitions
name declarations
include directives
macro denitions
template declarations
template denitions
condition compile directives
comments

127

namespace MatVec f /* ... ... */ g

struct point2d f oat x, y g
extern double b
const double e = 2.718
enum color f red, yellow, green g
extern int strlen(const char*)
inline char get(char*p)freturn *p++g
struct Matrix
#include <cmath>
#dene VERSION 12
template <class T> class A
template <class T> class B f/*...*/g
#ifdef MatVec H
/* This is a matrix-vector library */

TABLE 4.4. Items that a header le may contain.

ordinary function defns

data denitions
aggregate denitions
unnamed namespaces
exported template defns

char get(char* p) f return *p++ g

int aa double bb = 4.5
int table! ] = f 1, 2, 3g
namespace f int size = 100 /* ... */ g
export template <class T> f(T t)f/*...*/g

TABLE 4.5. Items that a header le should not contain.

#ifdef identifier

checks if identifier has been and remains dened. It can be used to include
or skip some statements from it up to
#endif

depending on whether identifier has been and remains dened. For example, C and C++ can share the following header le.
#ifdef __cplusplus
namespace std {
#endif

// this line will be skipped

// if __cplusplus is not defined

// declarations of standard library functions for C and C++

int printf(const char* ...)
#ifdef __cplusplus
}
using namespace std
#endif

// skipped if __cplusplus not defined

128

4. Namespaces and Files

The identier plusplus is dened when compiling (actually preprocessing)

a C++ program and thus the function declarations such as printf () are put
in the namespace std: When compiling a C program, however, plusplus
is not dened and there is no concept of namespace the lines related to
namespace will not be compiled.
Similarly, the conditional compilation directive
#ifndef identifier

checks if identifier is still undened.

A directive of the form
#if expression

checks whether the (constant) expression evaluates to true: Similar directives are #else and #elif (else if). They may be used as
#define DEBUG 1
// debugging information can be printed out.
#if DEBUG
cout << "debug: value of a is: " << a << '\n'
#endif

After the debugging process, DEBUG can be dened to be 0 so that unnecessary information will not be printed out. In fact, this printing statement
above will then be ignored by the preprocessor when the le is compiled.
This use of directives actually serves as comments that can be easily turned
on and o.

4.2.4 File Inclusion

If a le includes a header le, the preprocessor simply copies the header
le into the le at the point indicated before compilation begins. There
are situations where several header les are necessary and one includes
another. Then there is a danger that some header les are included in a le
more than once. For example, if head1:h and head2:h both include head:h
and head1:h also includes head2:h, to avoid chaining, head:h may take the
form:
#ifndef _HEAD_H
#define _HEAD_H
void my_function(double)
// declare/define other things that need be
// declared/defined in head.h
#endif

4.3 Source Files and Linkages

129

If the le head:h has already been included by the preprocessor, the macro
HEAD H will have been dened and the body of the le will not be
included more than once. For safety, every header le may be designed like
this.
If somewhere the macro HEAD H need be undened, use the directive
#undef as
#undef _HEAD_H

4.3 Source Files and Linkages

4.3.1 Separate Compilation

A le is a traditional unit of storage in a le system and a traditional unit

of compilation. A large program usually contains many les. When a le or
something it depends on is changed, all of it has to be recompiled. Breaking
a large program into suitable-size les can save tremendous compilation
time in code development, since les (and those they have included) that
have not been changed need not be recompiled.
When a source le is compiled, it is rst preprocessed by a C++ preprocessor that is, macros are substituted and #include directives bring in
headers. The result from preprocessing a le is called a translation unit.
Then the compiler works on this unit according to C++ language rules.
Thus strictly speaking, there is a distinction between a source le and a
translation unit, although most of the time these two names are used interchangeably.
To ensure successful separate compilation of the source les, the declarations in these les must be consistent. Note that an object can be declared
many times (although it can only be dened once). After all source les
have been compiled, their object codes will be linked by a linker also called
a loader: The linker can catch inconsistencies between separately compiled
les. See x4.4.2 on how to compile and link dierent les on UNIX or Linux.

4.3.2 External and Internal Linkages

With separate compilation, some names may be declared or dened in

one le and used in another. A name that can be used in translation units
dierent from the one in which it is dened is said to have external linkage:
For example, consider the following two les file1:cc and file2:cc:
// file1.cc
double x = 3.141
double f(double y) { return y + 2.178 }
int main() { /* ... ... */ }

130

4. Namespaces and Files

// file2.cc
extern double x
// another file may define x
extern double f(double)
// another file may define f
void g(double z) { x = f(z) }

The variable x and function f () used in file2:cc are the ones dened in
file1:cc. Thus x and f () are said to have external linkage. External variables have external linkage. Note that x and f are both global and external
according to denitions in x2.1. A name that can be referred to only in
the translation unit in which it is dened is said to have internal linkage:
Thus the function g() in file2:cc has internal linkage and can not be used in
file1:cc. The keyword extern in front of a function such as double f (double)
on the second line in file2:cc can be omitted.
An object must be dened only once in a program, although it can be
declared many times as long as the declarations agree exactly with each
other. For example, consider the following two les.
// file3.cc
int i = 5
float a = 3.14
extern char c

// this declaration is also a definition

// this defines a to be 3.14
// a declaration only

int main() { /* ... ... */ }

// file4.cc
int i = 10
extern double a
extern char c
char c = 'c'

// i is defined again, illegal

// definition of c

There are two errors in these two les: the variable i is dened twice and
a is declared twice with dierent types. Note c is declared three times
but dened only once, which is not an error. These two errors can not
be caught by the compiler since each le is compiled separately. They are
linkage errors that should be detected by the linker.
At this moment, it may be a good idea to look at a few more examples
of declarations and denitions:
double d
int f(int)
extern int i
struct s

//
//
//
//

definition
declaration
declaration
declaration

namespace mv { int m = 100 }

// definition
inline int g(int n) { return n } // definition
const int j = 5
// definition

4.3 Source Files and Linkages

extern const int k = 5
typedef unsigned int Uint
struct p2d { double x, y }

131

// definition
// definition
// definition

The rst statement is a denition since d is initialized by default. The names

f , i, and s are declarations and must be dened somewhere. However, the
names mv g j k Uint and p2d are denitions since they introduce and
dene some entities into the program. Notice that Uint is just dened to
be a synonym to unsigned int (it does not introduce a new type) and k is a
constant that can be accessed in another le in which it is declared (using
the statement extern const int k).
By default, const typedef, and macros have internal linkage in C++. For
example, it is legal (but confusing) to have the following two les.
// file5.cc
#define DEBUG 1
const double d = 3.14
typedef unsigned int Uint
extern const int i = 100

//
//
//
//

internal
internal
internal
external

// file6.cc
#define DEBUG 0
const double d = 2.17
typedef unsigned long int Uint
extern const int i

//
//
//
//

internal linkage
internal linkage
internal linkage
declaration

linkage
linkage
linkage
linkage

However, an inline function must be dened in every le in which it is used

with exactly the same denition. For example,
// file7.cc
inline int f(int i) { return i + i*i }
// file8.cc
inline int f(int i) { return i + i*i }

The compiler needs to know its denition in order to make code substitution. An inline function can not be made extern to simplify the implementation of C++ compilers.
These specications may be very confusing and error-prone. The best
way to handle them is to put them in a common header le and include it
in dierent source les. An example is:
// file9.h
a header file
#include <iostream>
// include a standard library
#include <cmath>
// include a standard library
#define DEBUG 1
const double d = 2.17

// user's stuff

132

4. Namespaces and Files

extern long double pi

typedef unsigned int Uint
struct s { int i, j }
enum color { red, green, blue}
namespace matvec { /* ... */ }
inline int f(int i) { return i + i*i }
extern double g(long double, int i)
// file10.cc
#include "file9.h"

for definitions of user's functions

long double pi = std::atan(1)*4

double g(long double d, int i) {
// ...
return d + i
}

// definition of g

// file11.cc
a source file the main program
#include "file9.h"
int main() {
double a = g(pi, 5) + f(10) + d
// ...
std::cout << "a = " << a << '\n'
}

Here all common declarations and denitions are put in header le file9:h
and source les file10:cc and file11:cc both include this header. A struct,
enum, named namespace, inline function, and so on, must be dened exactly once in a program. Strictly speaking, this rule is violated when their
denitions are put in a header le and included in dierent source les.
For convenience, the language allows this since the denitions appear in
dierent translation units and are exactly the same. See x4.2.2 for what
should be and what should not be put in header les.
It is worthwhile to note that there are dierences in C and C++ in terms
of linkage. For example, consider these two les:
// file12.cc
float y
float f() { return 5 }
int main() { g() }
// file13.cc

4.3 Source Files and Linkages

133

float y
void g() { y = f() }

There are three errors as C++ programs: y is dened twice (the declarations
of y are also denitions since the compiler initializes y to a default value)
in the two les, and f () in file13:cc and g() in file12:cc are not declared.
However, they are correct C programs.

4.3.3 Linkage to Other Languages

Sometimes a programmer has to deal with codes written in C++ and other
languages such as C and FORTRAN. It can be dicult in general, since
dierent languages may dier in the layout of function arguments put on
the run-time stack, in the use of machine registers, in the layout of builtin types such as integers and arrays, and so on. For example, the layout
of two-dimensional arrays is row by row in C++ and C, but column by
column in FORTRAN, and arguments to functions are passed exclusively
by reference in FORTRAN.
A user can specify a linkage convention in an extern declaration. For
example,
extern "C" char* strcpy(char*, const char*)
// C linkage convention
extern "C" {
int f(int)
int g(int)
double sqrt(double)
}

// C linkage convention
// group several declarations
// in one block

Note that the letter C in double quotes following extern means to link
these function calls according to the C linkage convention. Here C does
not refer to a language but to a linkage convention. It can also be used to
link FORTRAN and assembler codes that conform to C implementation
conventions.
In general, C code can be mixed with C++ code statement by statement,
provided the C code does not contain C++ keywords as user-dened names.
This enables many C programs to be called directly from and mixed with
C++ programs. Vaguely speaking, C is a subset of C++. However, dealing
with a mixture of C++ and FORTRAN programs is platform-dependent.
For example, a FORTRAN function name sum() could be compiled to a
symbol name sum, sum , sum , or SUM, depending on the compiler. For
examples on how to mix C/C++ and FORTRAN programs, see !LF93,
Arn98].

134

4. Namespaces and Files

4.4 Some Useful Tools

This section discusses a few useful tools (some in UNIX and Linux) for
timing, debugging, and proling a program, and for managing les and
creating libraries.

4.4.1 How to Time a Program

C++ inherits from C the time and date library with header in <ctime>
and <time:h>.
They have a function clock() that provides access to the underlying machine clock and returns an approximation to the number of CPU \clock
ticks" used by the program up to the current time. The returned value
can be divided by CLOCKS PER SEC to convert it into seconds. The
macro CLOCKS PER SEC is dened in <ctime> or <time:h>, but the
rate at which the clock time runs is machine-dependent. If the CPU clock
is not available, the value ;1 is returned. The function time() returns the
calendar time, normally expressed as the number of seconds elapsed since 1
January 1970. Other units and starting dates are possible. The prototypes
of these two functions and related declarations are given in <ctime> and
<time:h> and look like the following.
#define CLOCKS_PER_SEC
typedef long clock_t
typedef long time_t
clock_t clock(void)
time_t time(time_t* tp)

100

// machine dependent

// return number of CPU clock ticks

// return calendar time

double difftime(time_t t1, time_t t0)

// return t1 - t0
char* ctime(const time_t *tp)
// convert calendar time to a string

When the variable tp in function time(tp) is not the null pointer, the returned value also gets stored in the object pointed to by tp: Note that the
calendar time (wall time) and CPU clock time may not be the same in a
time-shared operating system such as UNIX and Linux.
For example, the following program computes the wall times and CPU
times for a number of multiplications in data types of double and float.
#include <iostream>
#include <time.h>

// for output
// for measuring time

int main() {
int n = 100000000
double d, dpi = 3.1415926535897932385

4.4 Some Useful Tools

135

float f, fpi = 3.1415926535897932385

time_t tm0 = time(0)
// wall time at this point
clock_t ck0 = clock()
// clock ticks at this point
for (int i = 0 i< n i++) d = (double(i) + dpi)*dpi
time_t tm1 = time(0)
// wall time at this point
clock_t ck1 = clock()
// clock ticks at this point
cout <<"wall time = " <<difftime(tm1,tm0) <<" seconds.\n"
cout <<"CPU time = " <<double(ck1 -ck0)/CLOCKS_PER_SEC
<< " seconds.\n"
for (int i = 0 i< n i++) f = (float(i) + fpi)*fpi
time_t tm2 = time(0)
// wall time at this point
clock_t ck2 = clock()
// clock ticks at this point
cout <<"wall time = " << difftime(tm2,tm1) <<" seconds.\n"
cout <<"CPU time = " << double(ck2 -ck1)/CLOCKS_PER_SEC
<< " seconds.\n"
cout << "The current time is: " << ctime(&tm2) << "\n"
}

A run on a Pentium II PC running Linux (RedHat 5.1) with a GNU

C++ compiler (gcc version egcs-2.91.66) produces the results:
wall time = 8 seconds.
CPU time = 7.23 seconds.
wall time = 4 seconds.
CPU time = 3.61 seconds.
The current time is: Sat Jun

3 15:56:45 2000

A run on an SGI workstation running UNIX (IRIX64) with a GNU C++

compiler (gcc version 2.8.1) produces the results:
wall time = 8 seconds.
CPU time = 7.69 seconds.
wall time = 10 seconds.
CPU time = 8.73 seconds.
The current time is: Fri Nov 12 14:16:39 1999

The computing speed generally depends on the underlying machine and

compiler. It is surprising to observe that arithmetic in single precision may
not be faster than in double precision on some machines. Some old C and
C++ compilers promote float to double so that arithmetic in single precision can actually be slower than in double precision.

136

4. Namespaces and Files

4.4.2 Compilation Options and Debuggers

When a program consists of dierent les, for example, main:cc, pgm0:cc,

pgm1:cc, and pgm2:cc, the UNIX/Linux command with the ;c option compiles them:
c++ -c main.cc pgm0.cc pgm1.cc pgm2.cc

If there are no errors, corresponding object les ending with :o will be

created. Then the ;o option links these object les to form a machineexecutable object le:
c++ -o pgm main.o pgm0.o pgm1.o pgm2.o

If ;o pgm is left out above, the executable object le will be created in
the default le a:out instead of pgm: These two commands can also be
combined into one:
c++ -o pgm main.cc pgm0.cc pgm1.cc pgm2.cc

Below are some useful options for the C++ compiler.

-c
-o name
-O
-g
-p
-I dir
-D name=def
-S
-M

compile only, generate corresponding :o les

put executable object le in name
attempt code optimization
generate code suitable for the debugger
generate code suitable for the proler
look for #include les in directory dir
place at the top of each source le the line
#dene name def
generate assembler code in corresponding :s les
create a makele

Your compiler may not support all of these options and may provide others.
Also your compiler may not be named c ++:
A debugger allows the programmer to know the line number in a program where a run-time error has occurred, and to see the values of variables
and expressions at each step. This can be very helpful in nding out why
a program is not running successfully. The debuggers gdb and dbx are generally available on UNIX and Linux systems. The code must be compiled
with the ;g option. To use gdb on the object code pgm for example, at
the command line type
gdb

Then type run pgm (followed by the argument list if pgm has any) and
it will run the program. Errors, if existing, will be displayed. Then typing
where gives the exact line number and the function in which the error has
occurred. Below are some useful commands.

4.4 Some Useful Tools

run pgm !arglist]

print exp
c
next
step
help name
bt
quit

137

run program pgm with argument list arglist

display the value of the expression exp
continue running the program after stopping
execute next program line after stopping
execute next line (step into the line)
show information about GDB command name
backtrace, display the program stack
exit the debugger

The UNIX/Linux manpage may give more information on how to use it.
See !LF93] for debuggers dbx and xdb, and other tools.
In UNIX, the ;p option used in compiling a program produces code that
counts the number of times a function is called. The proler prof then
generates an execution prole, which can be very useful when working to
improve execution time eciency. Look at the example:
#include
#include
#include
#include

//
//
//
//

for
for
for
for

output
measuring time
pseudorandom number
sort()

int main() {
int n = 100000
double* dp = new double %n]
srand(time(0))
// seed random number generator
for (int i = 0 i< n i++) dp%i] = rand()%1000
sort(dp, dp+n)

// sort array in increasing order

The function srand() seeds the random number generator and causes
the sequence generated by repeated calls to rand() to start in a dierent
place. The function sort() sorts a sequence in increasing order by default
and is discussed further in x10.2.1.
Suppose the program is called sortrand:cc. Compile it with the ;p option:
c++ -o sortrand -p sortrand.cc

The executable object code is written in le sortrand. Next run the program
by issuing the command
sortrand

Then the command

138

4. Namespaces and Files

prof sortrand

causes the following to be printed on the screen.

%time
57.5
21.5
16.0
2.8
2.2
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0

cumsecs #call
1.04
1
1.43
1.72 100000
1.77 100000
1.81 100000
1.81
1
1.81
1
1.81
1
1.81
1
1.81
1
1.81
1
1.81
3
1.81
2
1.81
1
1.81
1
1.81
1
1.81
1

ms/call
1040.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00

name
_main
mcount
.rem
.mul
_rand
.udiv
.umul
_cfree
_exit
_free
_gettimeofday
_malloc
_on_exit
_profil
_sbrk
_srand
_time

This execution prole gives the number of times that a function is called,
average number of milliseconds per call, and cumulative times in seconds.
For example, the function rand() was invoked 100000 times, took 0:00
milliseconds per call, and took total 1:81 seconds, while the function main()
was called once, took 1040:00 milliseconds per call, and took total 1:04
seconds.

4.4.3 Creating a Library

The UNIX/Linux operating system provides a utility called the archiver to

create and manage libraries. The name of the command is ar: By convention, UNIX library les end in :a (:lib in DOS). For example, the standard
C++ library may be in the le libstdc++:a in the directory =usr=lib: Issue
the UNIX/Linux command
ar t /usr/lib/libstdc++.a

The key t will display the titles (names) of the les in the library.
A programmer may create his or her own library. Suppose there are
les matvec:cc mesh:cc and fem:cc that contain matrix-vector operation,
mesh generation, and nite element analysis codes. To create a library, rst
compile these source les and obtain the corresponding :o les. Then type
the two commands:
ar ruv libmvmf.a matvec.o mesh.o fem.o

4.4 Some Useful Tools

139

ranlib libmvmf.a

The keys ruv in the rst command stand for replace, update, and verbose,
respectively. The library libmvmf:a is then created if it does not already
exist. If it exists, then the :o les replace those of the same name already
existing in the library. If any of these :o les are not already in the library,
they are added to it. The ranlib command simply randomizes the library
that is useful for the loader (linker).
If there is a program consisting of two les: main:cc and iofile:cc, and
these two les call functions in the library libmvmf:a then the library can
be made available to the compiler by issuing the following command.
c++ -o pgm main.cc iofile.cc libmvmf.a

The functions that are invoked in main:cc and iofile:cc but not dened
there will be searched for rst in libmvmf:a and then in C++ standard
libraries. Only those functions of libmvmf:a that are used in main:cc and
iofile:cc will be loaded into the nal executable le pgm: Typing all the
commands can be tedious when there are a lot of les and libraries. A
UNIX utility named make can achieve this by specifying all the commands
in a le called Makefile see x4.4.4.

4.4.4 Makele

The make utility (available always in UNIX and Linux and often in DOS
as well) provides easy management of compiling source les and linking
their object les, and convenient access to libraries and associated header
les. It is ecient to keep a moderate or large size program in several les
and compile them separately. If one of the source les needs to be changed,
only that le has to be recompiled. The use of the make utility can greatly
facilitate the maintenance of the source les in a program written in C++,
C, FORTRAN, Java, and others.
The make command reads a le with a default name makefile or Makele. The le contains dependency lines (e.g., how an object le depends
on a source le and some header les) and action lines (e.g., how a source
le is to be compiled with what options). As an example, let us write
a Makele for a program that contains a header le matvec:h and two
source les main:cc and matvec:cc all in the current directory. The le
main:cc contains the function main() that calls some functions dened in
matvec:cc: The header matvec:h is included in both main:cc and matvec:cc:
The Makefile for such a simple program can be just as
mat: main.o matvec.o
# mat depends on main.o,matvec.o
c++ -o mat main.o matvec.o
# link object code mat
main.o: main.cc matvec.h
cc -c -O main.cc

# main.o rely on main.cc,matvec.h

# compile main.cc and optimize

140

4. Namespaces and Files

matvec.o: matvec.cc matvec.h # what matvec.o depends on ?

cc -c -O matvec.cc
# compile with option -O

A comment in Makefile begins with symbol # and extends to the rest of

the line (just like == in a C++ program). The rst line is a dependency
line that indicates that mat depends on object les main:o and matvec:o.
The second line is an action line that indicates how the object les are
to be linked and the resulting executable object code is written in a le
named mat: A dependency line must start in column 1 while an action line
must begin with a tab character. A dependency line and the action lines
following it are called a rule. So there are three rules mat main:o, and
matvec:o in this Makefile: Note that the name of a rule must be followed
by a colon. Then the UNIX/Linux command
make mat

makes the rule mat in Makefile, which depends on two other rules main:o
and matvec:o. These two rules will be made rst according to their action
lines. That is, the source les main:cc and matvec:cc are to be compiled
with the ;O option. Then the rule mat will cause the object code main:o
and matve:o to be linked and the nal executable object code to be written
in le mat:
If the source le main:cc is to be changed, but matvec:h and matvec:cc
remain unchanged, then the command make mat will only cause the le
main:cc to be recompiled and leave the third rule matvec:o alone. However,
if the header le matvec:h is to be changed, then main:cc and matvec:cc
will be recompiled since main:o and matvec:o both depend on matvec:h
by the second and third rules.
Similarly, the command make main:o will cause the rule main:o to be
made in this case only main:cc will be compiled. By default, the command
make without an argument simply makes the rst rule in the Makefile.
For this Makefile, the commands make and make mat are equivalent.
Certain rules and macros are built into make: For example, each :o le
depends on the corresponding :c le and the second rule in the above
Makefile is equivalent to
main.o: main.cc matvec.h
cc -c -O $*.cc

# main.o rely on main.cc, matvec.h

# compile and attempt to optimize

where $ :cc expands to main:cc when main:o is being made. It stands for
all :cc les on the dependency line. The system also echoes a message on
the screen about the action. To suppress the message, the symbol @ can
be used:
main.o: main.cc matvec.h
@cc -c -O $*.cc

# main.o rely on main.cc, matvec.h

# compile main.cc with option -O

4.4 Some Useful Tools

141

A user can also dene macros in a makefile: The following is a makefile

that I use for compiling my code in les main:cc fem:cc and fem:h that
call functions in libraries matveclib:a and meshlib:a and include headers
matvec:h and mesh:h in two dierent directories.
#
#
#
#

Makefile for source files fem.cc and main.cc that

include fem.h. Sources main.cc and fem.cc call functions
in my libraries matveclib.a and meshlib.a
fem.h includes two headers matvec.h and mesh.h

# a macro BASE for my home directory.

# Value of the macro can be accessed by $BASE
BASE = /home/yang
# macro for name of my compiler
CC = c++
# macro for compilation options
CFLAGS = -O
# directories for include files matvec.h and mesh.h
# which are included in fem.h
INCLS = -I$(BASE)/c++/matvec -I$(BASE)/c++/mesh
# paths for my libraries matveclib.a and meshlib.a
# '\' means a continued line
LIBS = $(BASE)/c++/matvec/matveclib.a \
$(BASE)/c++/mesh/meshlib.a
# name of executable file
EFILE = $(BASE)/c++/fem/xx
HDRS = fem.h
OBJS = main.o fem.o
fem: $(OBJS)
@echo "linking ..."
$(CC) $(CFLAGS) -o $(EFILE) $(OBJS) $(LIBS)
$(OBJS): $(HDRS)
$(CC) $(CFLAGS) $(INCLS) -c $*.cc
clean:
rm -f $(OBJS) xx

142

4. Namespaces and Files

# main.o depends on main.cc and fem.o depends on fem.cc

main.o: main.cc
fem.o: fem.cc

The les main:cc and fem:cc include the header le fem:h, all of which are
in the same directory =home=yang=c ++=fem as the Makefile: Furthermore, fem:h includes two header les matvec:h and mesh:h which reside,
respectively, in two dierent directories:
/home/yang/c++/matvec
/home/yang/c++/mesh

The macro INCLS means to search for header les in these directories.
When main:cc and fem:cc are compiled, these header les are searched
for by the system. When main:o and fem:o are linked to generate the
executable object le xx the libraries matveclib:a and meshlib:a are loaded
through the macro LIBS . Note that paths for C++ standard libraries
such as < cmath > and < complex > and their header les need not be
indicated since the system should nd them automatically. Only paths for
user-dened libraries and header les have to be indicated.
At UNIX or Linux command, type
make

It will compile main:cc and fem:cc and link main:o and fem:o into the
excutable object code xx: The last two lines in this Makefile indicate that
main:o depends also on main:cc besides fem:h and fem:o also on fem:cc:
Thus any change to main:cc will cause only main:o to be regenerated. The
command make clean will remove the les dened in macro OBJS and
the le xx:
UNIX also provides a utility called SCCS (Source Code Control System)
for managing dierent versions of source code (or any other text les) see
!LF93].

4.5 Standard Library on Strings

The C++ standard library < string > provides a type called string to
manipulate character strings. Only a few examples are presented here to
illustrate its use. Details can be found in !Str97, LL98].

4.5.1 Declarations and Initializations

A string can be initialized by a sequence of characters, by another string,

or by a C-style string. However, it can not be initialized by a character or
an integer. For example,

4.5 Standard Library on Strings

#include <string>
using namespace std
int main() {
string s0
string s00 = ""
string s1 = "Hello"
string s2 = "world"
string s3 = s2
string s33(5,'A')

143

// empty string
// empty string
// initialization

// 5 copies of 'A', s33 = "AAAAA"

string s333 = 'A'

string s3333 = 8

// error, no conversion from char

// error, no conversion from int

s3 = 'B'

// OK, assign a char to string

Notice that assigning a char to a string is allowed, although it is illegal to

initialize a string by a char:

4.5.2 Operations

Certain arithmetic and operations are dened for string: For example,
addition of strings means concatenation. The operator += can be used to
append to the end of a string. Strings can be compared to each other and to
string literals. The function call s:empty() returns true if string s is empty
and false otherwise, and s:max size() gives the maximum size that a string
can have. The function calls s:size() and s:length() both give the size of
a string s: The size of a string must be strictly less than string :: npos
which is the largest possible value for string size (it is equal to 4294967295
on one machine). Elements in a string s can be accessed by the subscripting
operator ! ] (index changes from 0 up to s:size() ; 1). For example,
if (s2 == s1 ) {
// test for equality of s1, s2
s2 += '\n'
// append '\n' to end of s2
} else if (s2 == "world") {
s3 = s1 + ", " + s2 + "!\n" // s3 = "Hello, world!\n"
}
s3%12] = '?'
char h = s3%12]

// assign '?' to s3%12]

// h = '?'

int i = s3.length()
int j = s3.size()

// size of s3, i = 14
// size of s3, j = 14

The function call s:substr(n m) gives a substring of s consisting of

m characters starting from s!n] s:replace(n m t) replaces m characters

144

4. Namespaces and Files

starting from s!n] by string t s:insert(n t) inserts a string t at s!n] and

s:find(t) returns the position of substring t in string s and returns npos if t
is not a substring of s: Note that if npos is used as a character position, an
exception range error will be thrown (see Chapter 9). They can be used
as
// now s3 = "Hello, world?\n"
string s4 = s3.substr(7,5) // s4 ="world"
s3.replace(0,5,"Hi")
// s3 = "Hi, world?"
s3.insert(0,"Hi")
// insert "Hi" at s3%0]
unsigned int pos = s3.find("wor")
// find position of substring "wor"
unsigned int pos2 = s3.find("HW")
// pos2 = 4294967295 on one machine
s3.erase(0,2)
// erase two characters from s3

A related function, rfind() searches for a substring from the end of a string.
The function s:erase(n m) erases m characters from s starting from s!n].
The usual comparison operators == != > < >= and <= are provided
based on lexicographical ordering. In addition, many standard algorithms
can be applied to strings see x10.2.

4.5.3 C-Style Strings

Conversion from string to C-style strings (char) is provided. For example,

const char* p = s3.c_str()
const char* q = s3.data()
char c = q%0]

// convert to C-style string

// convert to an array of chars
// c = first character in q

char* s6 = "Hello"
string s7 = s6

// from C-style string to string

The function s:data() converts a string s to an array of characters while

s:c str() produces a C-style string (char or const char).
Note that a C++ string is not an array of characters. That is, for a
string s &s!0] is dierent from s:

4.5.4 Input and Output

Strings can be read from and written to the screen or a le using the input
and output operators. For example,
string s10
cout << "Enter your city and state names\n"
cin >> s10
cout << "Hi, " << s10 << "!\n"

4.5 Standard Library on Strings

145

If you type in your city and state as

Detroit Michigan

the output on the screen is Hi Detroit! By default, a whitespace character

terminates the input. Thus the second word Michigan above is not read
to string s10: To read the whole line, the function getline() can be used:
cout << "Enter your name\n"
getline(cin, s10)
cout << "Hi, " << s10 << "\n"

// read a whole line from cin

The function
getline(istream istm, string strg, char eol)

reads a line from an input stream istm terminated by character eol to the
string strg. If no eol argument is provided, a newline character 'nn' is used
as the delimiter.

4.5.5 C Library on Strings

The header <cstring > or <string:h> contains functions operating on

C-style strings. For example,
char* strcpy(char*p, const char*q)
// copy from q into p,
char* strcat(char*p, const char*q)
// append from q to p,
char* strncpy(char*p, const char*q,
// copy n chars from q
char* strncat(char*p, const char*q,
// append n chars from

including terminator
including terminator
int n)
into p
int n)
q into p

size_t strlen(const char* p)

// length of p
int strcmp(const char*p, const char*q)
// compare p, q
int strncmp(const char*p, const char*q, int n)
// compare first n characters in p and q

The pointer argument above is assumed to be nonzero, and the array of

char it points to is assumed to be terminated by 0. The strn-functions
pad with 0 if there are not n characters to copy. The string comparisons
return 0 if the strings are equal, a negative number if the rst string is
lexicographically before the second, and a positive number otherwise. The
type size t is an unsigned integral type dened in the standard headers
<stddef:h> and <stdlib:h>:
The header < cctype > or < ctype:h > contains functions dealing with
ASCII characters. For example,

146
int
int
int
int
int
int

4. Namespaces and Files

isalpha(int)
isupper(int)
islower(int)
isdigit(int)
isspace(int)
isalnum(int)

int toupper(int)
int tolower(int)

//
//
//
//
//
//

Is 'a'..'z' 'A'..'Z' ?
Is 'A'..'Z' ?
Is 'a'..'z' ?
Is digit '0'..'9' ?
Is space ' ' '\t' '\v' ?
isalpha() | isdigit()

// to uppercase
// to lowercase

These functions take int as argument since a character is treated as a small

integer in C and C++. Equivalent functions for wide characters can be
found in <cwtype> and <wtype:h>. The type wchar t for wide characters
is dened in <stddef:h> and <cstddef>.
The header <cstdio> or <stdio:h> contains a function called sprintf ()
which can concatenate strings and numeric numbers together:
char s11%50]
int ii = 5
int jj = 80
sprintf(s11, "%s%d%s%d%s", "ex", ii, "h", jj, "\n")
cout << s11
// ex5h80 and a newline are printed out

In the specication "%s%d%s%d%s" above, s means a string and d means

an integer (a digit). This function is useful when we want to generate a
string for an output le name, which represents, for instance, the example
number and grid size such as ex5h80 see x4.6.4 and x4.6.6.

4.6 Standard Library on Streams

The C++ standard libraries <iostream> and <fstream> provide functions for reading from and writing to the screen or les, which are illustrated through examples in this section. More details can be found in
!Str97, LL98].

4.6.1 Formatted Integer Output

The function setf () can set the format in which an integer is output. For
example,
int main() {
int i = 123456789
cout << i << " " << i << '\n'
// decimal
cout.setf(ios_base::oct, ios_base::basefield) // octal
cout << i << " " << i << '\n'
// print i in octal

4.6 Standard Library on Streams

147

cout.setf(ios_base::hex, ios_base::basefield) // hex

cout << i << " " << i << '\n'
// print i in hex
cout.setf(ios_base::dec, ios_base::basefield) // decimal
cout << i << " " << i << '\n'
}

The identiers ios base, ios base :: dec (decimal), ios base :: oct (octal, base
8), ios base :: hex (hexadecimal, base 16), and ios base :: basefield (which
base in the output?) are dened in <iostream>. Since the name ios instead
of ios base was used in prestandard C++, many standard C++ compilers
may continue to support ios: This program will output the value of i in
decimal, octal, hexadecimal, and decimal formats:
123456789 123456789
726746425 726746425
75bcd15 75bcd15
123456789 123456789

Setting showbase will enable one to see which base is used. By default, a
number is decimal. An octal number is preceded by 0, while a hexadecimal
number is preceded by 0x: For example, the program segment
int i = 123456789
cout.setf(ios_base::showbase)
// show base of output
cout << i << " " << i << '\n'
// decimal, default
cout.setf(ios_base::oct, ios_base::basefield)
cout << i << " " << i << '\n'
// octal, base 8
cout.setf(ios_base::hex, ios_base::basefield)
cout << i << " " << i << '\n'
// hex, base =16
cout.setf(0, ios_base::basefield) // reset to default
cout << i << " " << i << '\n'

will output to the screen:

123456789 123456789
0726746425 0726746425
0x75bcd15 0x75bcd15
123456789 123456789

Setting ios base :: basefield to 0 resets the base to default, that is, decimal. This technique can also print out an octal or hexadecimal integer
in decimal. For example, let i = 012345 or i = 0x12345 above. Setting
showbase shows the base of an integer. Similarly, setting showpoint prints
trailing zeros and setting showpos prints explicitly + for positive integers.
Once set, these format ags remain valid until they are unset by calling
the function unsetf () as in cout:unsetf (ios base :: showbase): To set or
unset several ags in one statement, the bitwise or operator can be used as
in cout:unsetf (ios base :: showbase j ios base :: showpos):

148

4. Namespaces and Files

4.6.2 Formatted Floating Point Output

Floating point output is controlled by a format and a precision:

The general format presents a number that best preserves its value
in the space available the precision species the maximum number
of digits.
The scientific format presents a number with one digit before the
decimal point and an exponent the precision species the maximum
number of digits after the decimal point.
The fixed format presents a number with an integral part and a
fractional part the precision species the maximum number of digits
after the decimal point.
The general format is the default and the default precision is six digits. The
function setf () can set the format and precision() can set the precision.
For example,

double d = 12345.6789
cout << d << " " << d << '\n'
// default
cout.setf(ios_base::scientific, ios_base::floatfield)
cout << d << " " << d << '\n'
// scientific
cout.setf(ios_base::fixed, ios_base::floatfield)
cout << d << " " << d << '\n'
// fixed
cout.setf(0, ios_base::floatfield) // reset
cout << d << " " << d << '\n'
// to default

outputs
12345.7 12345.7
1.234568e+04 1.234568e+04
12345.678900 12345.678900
12345.7 12345.7

The identiers ios base :: scientific (for scientic format), ios base :: fixed
(for xed format), and ios base :: floatfield are dened in <iostream>.
Setting ios base :: floatfield to 0 resets the format back to the default.
Note that the default precision of six digits is applied to the whole number
of the general format, while only to the fractional part of scientic and xed
formats. To increase the accuracy of an output, the precision() function
can be used. For example,
double d = 12345.67890123456789
cout.setf(ios_base::scientific, ios_base::floatfield)
cout.setf(ios_base::uppercase) // E in scientific format
cout.precision(15)
// precision = 15
cout << d << " " << 1000*d << '\n'

4.6 Standard Library on Streams

cout.precision(10)
cout << d << " " << 1000*d << '\n'

149

// precision = 10

The output is:

1.234567890123457E+04 1.234567890123457E+07
1.2345678901E+04 1.2345678901E+07

Since the output is in the scientic format, the precision means the number
of digits after the decimal point. Notice that E instead of e appears in the
exponential part of the scientic format by setting ios base :: uppercase:

4.6.3 Output Width

The function width() can control the number of characters that the next
output number or string occupies. If the number is not long enough, whitespace by default or some other \padding" character specied by the function
fill() will be used. For example,
double d = 12345.678987654321
cout.width(25)
// output width 25 chars
cout.precision(15)
cout << d << '\n'
cout.width(25)
// output width 25 chars
cout.precision(8)
cout.fill('#')
// use '#' for padding extra space
cout << d << '\n'

It will produce the following output, which is aligned to the right by default:
12345.6789876543
################12345.679

Adjustment of the alignment can be made by setting adjustfield to left

(left alignment), internal (internal alignment, put ll characters between
sign and value), or right (right alignment). For example,
double d = - 12345.678987654321
cout.precision(15)
cout.setf(ios_base::left, ios_base::adjustfield)
cout.width(25)
cout << d << '\n'

The alignment is adusted to the left in this output.

The width() function sets the minimum number of characters. If the number has more digits, more characters will be output in order not to lose accuracy. The eld width can be reset to its default by calling cout:width(0):

150

4. Namespaces and Files

4.6.4 Input and Output Files

The C++ header <fstream> provides functions for inputting from and
outputting to les. In particular, they have types ifstream for input le
stream, ofstream for output le stream, and fstream for both input and
output. The open mode of a le can be
in : to open for reading.
out : to open for writing.
app : to append.
binary : to open in binary mode, rather than in text mode.
ate : to open and seek to end of le.
trunc : to truncate a le to 0-length.
For example, the declarations
ifstream infile("data", ios_base::in)
// open file for input
ofstream outfile(result, ios_base::out)
// open file for output

declare the variable inle to be an input le stream that opens the le data
for reading, and outle to be an output le stream that opens a le whose
name is stored in string result for writing.
The following is a complete program that reads an integer and a double
from a le called data, generates a string for an output le name, and writes
to the le.
#include <stdio.h>
#include <fstream>

// for sprintf()
// for input/output files

int main() {
ifstream infile("data", ios_base::in) // open file to read
int i
double d
infile >> i
// read from "data"
infile >> d
char result%20]
// output file name
sprintf(result, "%s%d%s%d%s", "ex",i,"-", int(1/d), "h")
ofstream outfile(result,ios_base::out) // open output file
outfile << "example number: " << i << '\n'
outfile << "grid size: " << d << '\n'

4.6 Standard Library on Streams

infile.close()
outfile.close()

151

// close input file stream

// close output file stream

A le needs to be closed by calling function close() when it is no longer

used. There is a maximum number of les that can be opened at a time. If
the input le data has only two lines:
5
0.01

then the program rst reads the rst number 5 to variable i and then reads
the second number 0:01 to variable d. The function sprintf () causes the
variable result to have value ex5-99h. The output of the program is written
in a le called \ex5-99h":
example number: 5
grid size: 0.01

The bitwise or operator j can be used to choose a le to have more than
one open mode. For example, to open a le for both writing and appending,
or to open a le for both writing and reading, we can do:
ofstream wafile("data2", ios_base::out | ios_base::app)
// open file for write and append
fstream rwfile("data3", ios_base::in | ios_base::out)
// open file for read and write

To generate formatted output to les, the functions setf (), precision(),

and width() can be applied to le streams in the same way as described
earlier for cout: For example,
ofstream ouf("res", ios_base::out) // open output file
ouf.setf(ios_base::scientific, ios_base::floatfield)
ouf.precision(10)
ouf.width(30)
ouf << 12345.67890 << '\n'

Finally, here is an example on how to read a le with comments. Suppose

there is a le called data4 that contains a comment line as the rst line
and some comments preceded by a slash sign = on every following line:
4
0.01

data4: input file for a test example

/ example number
/ grid size

The input le with comments is more readable since it indicates that 4
represents an example number and 0:01 a grid size. The program, which
must ignore the rst line and everything after the slash sign on other lines,
may be written as

152

4. Namespaces and Files

#include <string>
#include <stdlib.h>
#include <fstream>
int main(int argc, char* argv%]) {
if (argc != 3)
cout << "input and output files must be specified\n"
ifstream infile(argv%1], ios_base::in)
if (!infile)
cout << "can not open file: " << argv%1] << " for read\n"
string s
int i
double d
infile.ignore(80,'\n')

// ignore the first line

getline(infile,s,'/')
i = atoi(s.c_str())
infile.ignore(80,'\n')

// get a line terminated by '/'

// convert to an int
// ignore rest of line, comments

getline(infile,s,'/')
d = atof(s.c_str())

// get next line terminated by '/'

// convert to a float

ofstream outfile(argv%2], ios_base::out)

if (!outfile)
cout << "can not open file: "<< argv%2]<< " for write\n"
outfile.setf(ios_base::scientific, ios_base::floatfield)
outfile << "example number: " << i << '\n'
outfile << "grid size:
" << d << '\n'
infile.close()
outfile.close()
}

The function sm:ignore(streamsize n int ct) skips from an input stream

sm at most n characters terminated by the character ct, where streamsize
is an integral type for the size of a stream, possibly int. The default terminating character for ignore() is end-of-le, represented by EOF which
is dened in <iostream>. The default number of characters ignored is
one. Thus a call sm.ignore() without any arguments means to throw the
next character away. Compile this program and suppose the object code is
written in a:out. Then the UNIX/Linux command
a.out data4 result

4.6 Standard Library on Streams

153

will read from le data4 and write the output into a le called result.

4.6.5 Input and Output of Characters

There are many operations dened for reading and writing characters. For
example, the function call
ism.read(Ch* p, streamsize n)

reads at most n characters (including whitespace and newline characters)

from input stream ism (of type istream) and stores them in p!0] p!1] : : :
where Ch can be char, unsigned char, or signed char, and streamsize is
an integral type for the size of a stream, typically int. It returns a reference
of type basic istream to the input stream object to which it is applied.
The function ignore() is similar to read() except that it does not store the
characters read but simply ignores them. The function call ism:gcount()
returns the number of bytes read from ism in the last call read(). Similarly,
the function call
osm.write(const Ch* p, streamsize n)

writes at most n characters (including whitespace and newline characters)

stored in p!0] p!1] : : : into output stream osm. It returns a reference of
type basic ostream to the output stream object to which it is applied. These
two functions have similar return types to the input and output operators
of Cmpx in x6.1. The types basic istream and basic ostream are dened in
<iostream>. The two functions can be used to write a program that reads
from an input stream and changes every character into its corresponding
upper case:
#include <iostream>
#include <fstream>
#include <ctype.h>
main() {
ifstream infile("readfile", ios_base::in)
char c
int n = 0, m = 0
while ( infile.read(&c, 1) ) { // read 1 char at a time
switch(c) {
case ',' : n++ break
case '\n': m++ break
}
c = toupper(c)
cout.write(&c, 1)
// write c on screen
}
cout << "There are " << n << " commas and "

154

4. Namespaces and Files

<< m << " lines in the input file.\n"

It also counts the number of commas and newlines in the input stream.
Suppose a le named readle contains the following.
The function read(p, n) reads at most n characters and
stores them in a character array p%0], p%1], up to
p%n -1]. The returned object will be converted to false
when end of file is encountered.

Then the program outputs the following on the screen:

THE FUNCTION READ(P, N) READS AT MOST N CHARACTERS AND
STORES THEM IN A CHARACTER ARRAY P%0], P%1], UP TO
P%N -1]. THE RETURNED OBJECT WILL BE CONVERTED TO FALSE
WHEN END OF FILE IS ENCOUNTERED.
There are 3 commas and 4 lines in the input file.

The functions read() and write() may read or write many characters at a
time. When only one character is read or written, the functions get(Ch&)
and put(Ch) are better suited. Using get() and put() the program above
can be rewritten as
main() {
ifstream infile("readfile", ios_base::in)
char c
int n = 0, m = 0
while ( infile.get(c) ) {
// read 1 char at a time
switch(c) {
case ',' : n++ break
case '\n': m++ break
}
c = toupper(c)
cout.put(c)
// write c on screen
}
cout << "There are " << n << " commas and "
<< m << " lines in the input file.\n"
}

The function get() is overloaded. Calling get() without any arguments returns the character just read.
Another version of get() is:
ism.get(Ch* p, streamsize n, Ch delimiter = '\n')

which reads at most n ; 1 characters from input stream ism and places
them into p!0] p!1] : : : p!n ; 2]: The third argument delimiter is a character stating that the reading of characters should be terminated when it

4.6 Standard Library on Streams

155

is encountered. The default delimiter character is 'nn'. The terminating

character itself is not read into array p: Thus the user may have to remove
the next character before performing a second read. The pointer p must
point to an array of at least n characters, since get() places 0 at the end of
the characters (C-style strings). The function call
ism.getline(Ch* p, streamsize n, Ch delimiter = '\n')

is very similar to get(), except that it discards the delimiter rather than
leaving it as the next character to be read in the input stream. The call to
gcount() returns the number of characters actually extracted by the last
call of get() or getline(). Here is an example using get() and gcount() to
read a le and count the number of characters on each line and the total
number of lines in the le:
void f(istream& ism) {
int n = 0
char line%1024]
while ( ism.get(line, 1024, '\n') ) {
n++
cout << "there are " << ism.gcount() << " characters "
<< "on line " << n << '\n'
ism.ignore()
// ignore terminate character: '\n'
}
cout << "There are " << n << " lines in input file.\n"
}
main() {
ifstream infe("readfile", ios::in)
f(infe)
}

Without ignoring the newline character on each line (through the function
call ism:ignore() see x4.6.4), this would cause an innite loop. The output
of this program on the input le readle above is:
there
there
there
there
There

are
are
are
are
are

56 characters on
54 characters on
57 characters on
34 characters on
4 lines in input

line 1
line 2
line 3
line 4
file.

Notice the newline character on each line is not read by get(): The function
f () above can also be alternatively written using getline():
void f(istream& ism) {
int n = 0
char line%1024]

156

4. Namespaces and Files

while ( ism.getline(line,
n++
cout << "there are " <<
<< "on line " << n
}
cout << "There are " << n

1024, '\n') ) {
ism.gcount() << " characters "
<< '\n'
<< " lines in input file.\n"

Thus the function getline() is normally preferred over get() when one line
needs to be read at a time. Notice this function is dierent from getline()
discussed in x4.6.4. The output of this program on readle is:
there
there
there
there
There

are
are
are
are
are

57 characters on
55 characters on
58 characters on
35 characters on
4 lines in input

line 1
line 2
line 3
line 4
file.

The newline character on each line of the le readle is also read by
getline():
There are three other istream operators that puts back, ungets, or peeks
a character in an input stream:
ism.putback(Ch c)
ism.unget()
ism.peek()

//
//
//
//

put character c back into stream ism

put most recently read character back
return next character (or EOF)
but do not extract it from ism

The following program illustrates how they might be used.

void f(istream& ism) {
char c, d
while (ism.get(c)) {
switch(c) {
case '/':
// look for comments starting with //
if (ism.peek() == '/')
// if there are two slashes,
ism.ignore(1024, '\n') // ignore rest of the line
break
case '<':
// look for operator <<
d = ism.get()
if (d != '<') ism.putback(d)
// put d back
break
case '&':
// look for operator &&
if (ism.get() != '&') ism.unget()
break
}
}
}

4.6 Standard Library on Streams

157

4.6.6 String Streams

The standard library <sstream> contains two types istringstream (for input string streams) and ostringstream (for output string streams), which
can be used to read from a string and write to a string using input and
output streams.
To illustrate how ostringstream can be used, consider a le named data
that contains only two lines:
5
0.01

A program can be written that reads the two numbers from this le, generates a string such as "ex5-99h" from the two numbers just read, opens a
le with this string as its name, and nally writes some information to this
le:
#include <string>
#include <fstream>
#include <sstream>
using namespace std
main() {
ifstream infe("data", ios_base::in)
int i
double d
infe >> i
infe >> d
// read from input file
ostringstream tmp
// declare variable tmp
tmp << "ex" << i << "-" << int(1/d) << "h"
string ofname = tmp.str()
// convert it into string
ofstream otfe(ofname.c_str(), ios_base::out)
otfe << "example number : " << i << '\n'
otfe << "grid size: " << d << '\n'
infe.close()
otfe.close()
}

It declares tmp to be a variable of type ostringstream and writes some

numeric values and characters into tmp using the standard output operator.
The function call tmp:str() converts an ostringstream object tmp into
a C++ string. Then c str() is called to convert a C++ string into a
C string (char), which is used as the output le name. Notice that an
ostringstream object can be initialized and the statements

158

4. Namespaces and Files

ostringstream tmp
// declare variable tmp
tmp << "ex" << i << "-" << int(1/d) << "h"

are equivalent to:

ostringstream tmp("ex")
// initialize variable tmp
tmp << i << "-" << int(1/d) << "h"

Compiling and running this program produces a le named ex5-99h:

example number: 5
grid size: 0.01

This is exactly what was achieved in x4.6.4 using the C function sprintf ():
An istringstream is an input stream reading from a string. It can be
used to convert a string (with whitespace characters) into dierent parts,
some of which may have numeric values. For example,
main() {
string s = "ex 5 - 9.9 h"
istringstream ism(s)
// initialize ism from string s
int i
double d
string u, v, w
ism >> u >> i >> v >> d >> w

// i = 5, d = 9.9

This achieves more than what C functions atoi(), atof (), and atol() do.
The libraries <string> and <sstream> are very convenient to use when
manipulating strings. Compared to C-style strings (char), the user does
not need to do memory allocation and deallocation. However, they may
cause performance problems when used extensively, since they are built on
the top of C-style strings.

4.7 Iterative Methods for Nonlinear Equations

In this section, the bisection and Newton's methods are presented to iteratively nd a solution to a nonlinear equation f (x) = 0 in a given interval
!a b]: If f is a continuous function on !a b] and f (a) and f (b) have dierent signs (one is positive and the other is negative), then the Intermediate
Value Theorem guarantees that there is at least one root r in (a b) such
that f (r) = 0: The goal of an iterative method is to nd a sequence of
numbers frn g such that rn converges to a root r when n ! 1: On a computer, an iterative process has to be stopped after nite many iterations.
The stopping criterion can be jrm ; rm;1 j < or jf (rm )j < for some
given small numbers and . When either criterion is satised, rm may

4.7 Iterative Methods for Nonlinear Equations

159

be taken as an approximation to an exact root r: Due to the approximate

nature of an iterative method and computer roundo errors, rm may never
satisfy f (rm ) = 0 exactly, no matter how many iterations are used. The
value f (rm ) is called a residual.

4.7.1 Bisection Method

The bisection method assumes that f is continuous on !a b] and f (a)f (b) <
0: It rst computes the middle point c = (a + b)=2 and then tests if
f (a)f (c) < 0: If it is true, then f must have a root in the smaller interval !a c]: If it is false, then f must have a root in !c b]: In either case,
rename the smaller interval as !a b] which contains a root but whose size is
reduced by half. Repeat this process until jb ; aj or jf (c)j is small, and then
the middle point c is taken as an approximate root of f . This algorithm
can be written easily in a recursive program as
double bisctn0(double a, double b, double (*f)(double),
double delta, double epsn) {
double c = (a + b)*0.5
// middle point
if (fabs(b-a)*0.5 < delta || fabs(f(c)) < epsn) return c
(f(a)*f(c) < 0) ? b = c : a = c
return bisctn0(a, b, f, delta, epsn)
}

Note that the function can also be alternatively declared as

typedef double (*pfn)(double)
// define a function type
double bisctn0(double a, double b, pfn f,
double delta, double epsn)

With the help of the function pointer pfn, this equivalent declaration may
be more readable to C++ beginners. Also the ternary operator ? : is used
to make the code look concise. It may also be equivalently written using
an if -statement:
if (f(a)*f(c) < 0)
b = c
else
a = c

Observe the close resemblance of the code to the algorithm. It recursively

calls the function itself to shrink the interval size by half in each recursion
and stops when the interval size or residual is small.
Now this program can be tested to nd solutions to:
(a) f (x) = x;1 ; 2x !a b] = !0 1]
(b) f (x) = 2;x + ex + 2 cos x ; 6 !a b] = !1 3]

160

4. Namespaces and Files

(c) f (x) = (x3 + 4x2 + 3x + 5)=(2x3 ; 9x2 + 18x ; 2) !a b] = !0 4]:
Common declarations and denitions can be put in a header le called
ch4it.h:
// file ch4it.h
#include <iostream>
#include <math.h>
#include <stdlib.h>

// for input and output

// for math functions
// for exit()

const double delta = 1.0e-8

const double epsn = 1.0e-9
double bisctn0(double, double, double (*)(double),
double, double)
double fa(double x)
double fb(double x)
double fc(double x)

The source code containing function denitions can be put in another

le called ch4it.cc:
// file ch4it.cc
#include "ch4it.h"

// include header file

double bisctn0(double a, double b, double (*f)(double),

double delta, double epsn) {
// ... put its definition here
}
double fa(double x) {
// test function (a)
if (x) return 1.0/x - pow(2,x)
else {
cout << "division by zero occurred in function fa()."
exit(1)
}
}
double fb(double x) {
// test function (b)
return pow(2,-x) + exp(x) +2*cos(x) - 6
}
double fc(double x) {
// test function (c)
double denorm = ((2*x-9)*x +18)*x -2 // nested multiply
if (denorm) return (((x+4)*x+ 3)*x +5)/denorm
else {
cout << "division by zero occurred in function fc()."

4.7 Iterative Methods for Nonlinear Equations

161

exit(1)
}
}

The functions fa() and fc() may involve division by zero. When it occurs,
the program is terminated by calling the function exit() declared in header
<stdlib:h>.
A main program can be written to nd the roots of the functions:
// file mainch4it.cc
#include "ch4it.h"
int main() {
// find a root of fa()
double root = bisctn0(1.0e-2, 1, fa, delta, epsn)
cout << "Approximate root of fa() by bisctn0() is: "
<< root << '\n'
cout << "Fcn value at approx root (residual) is: "
<< fa(root) << '\n'
// find a root of fb()
root = bisctn0(1, 3, fb, delta, epsn)
cout << "\nApproximate root of fb() by bisctn0() is: "
<< root << '\n'
cout << "Fcn value at approx root (residual) is: "
<< fb(root) << '\n'
// find a root of fc()
root = bisctn0(0, 4, fc, delta, epsn)
cout << "\nApproximate root of fc() by bisctn0() is: "
<< root << '\n'
cout << "Fcn value at approximate root (residual) is:"
<< fc(root) << '\n'
}

Compiling and running the program gives the following output.

Approximate root of fa() by bisctn0() is: 0.641186
Fcn value at approx root (residual) is: -1.48037e-09
Approximate root of fb() by bisctn0() is: 1.82938
Fcn value at approx root (residual) is: 9.4115e-09
Approximate root of fc() by bisctn0() is: 0.117877
Fcn value at approximate root (residual) is:2.24202e+09

Observe that the program nds approximate roots for functions fa() and
fb() with very small residuals. However, the residual for fc() is so large!

162

4. Namespaces and Files

A closer look reveals that fc() does not have a root in the given interval,
although it changes sign at the endpoints (why?). It is a good idea to
check the correctness of computer output whenever possible. In the case of
nding a solution to an equation, a small residual normally implies a good
approximate answer.
The function bisctn0() resembles the algorithm closely and has worked
ne on three test problems. However, it is not very ecient and robust in
the following sense. First, it invokes three function evaluations: f (a) once
and f (c) twice, in each recursion. Second, the product f (a) f (c) may
cause over ow or under ow while f (a) and f (c) are within range. Third,
the middle point c can be updated in a more robust way as c = a +(b ; a)=2:
There are examples in which the middle point computed by c = (a + b)=2
moves outside the interval !a b] on a computer with limited nite precision.
It is numerically more robust to compute a quantity by adding a small
correction to a previous approximation. An improved version can be dened
as
double bisctn1(double a, double
double u, double
double e = (b - a)*0.5
double c = a + e
double w = f(c)

b, double (*f)(double),
delta, double epsn) {
// shrink interval size
// middle point
// fcn value at middle point

if (fabs(e) < delta || fabs(w) < epsn) return c

((u>0 && w<0) || (u<0 && w>0)) ? (b=c):(a=c, u=w)
return bisctn1(a, b, f, u, delta, epsn)
}

This version bisctn1() requires only one function evaluation f (c) per
recursion (u = f (a) is passed as an argument to the next recursion) and
avoids the unnecessary product f (a) f (c): However, these two functions
bisctn0() and bisctn1() have a common disadvantage. That is, both can
lead to an innite recursion (why?). A safeguard is to put a limit on the
number of recursions. From x3.8.9, a static local variable may be used to
count the number of recursions and lead to the next version:
double bisctn2(double a, double b, double (*f)(double),
double u, double delta, double epsn,
int maxit) {
static int itern = 1
double e = (b - a)*0.5
// shrink interval size
double c = a + e
// middle point
double w = f(c)
// fcn value at middle point
if (fabs(e)<delta || fabs(w)<epsn || itern++ > maxit)
return c

4.7 Iterative Methods for Nonlinear Equations

163

((u>0 && w<0) || (u<0 && w>0)) ? (b=c) : (a=c, u=w)

return bisctn2(a, b, f, u, delta, epsn, maxit)
}

This new version bisctn2() achieves the goal that it stops when the number of recursions is larger than a prescribed number maxit, but has a side
eect. Since static local variable itern persists in all calls to bisctn2() it
may cause the function to terminate prematurely in subsequent calls. For
example,
double root = bisctn2(1.0e-2, 1, fa, fa(1.0e-2), 1.0e-8,
1.0e-9, 40)
root = bisctn2(1, 3, fb, fb(1), 1.0e-8, 1.0e-9, 40)
root = bisctn2(0, 4, fc, fc(0), 1.0e-8, 1.0e-9, 40)

The rst call on fa takes 27 recursions to stop. The static local variable
itern = 27 at the beginning of the second call on fb which is terminated
prematurely after 14 recursions when itern = 41. The third call on fc is
terminated after only the rst recursion.
To overcome this diculty, one more argument can be passed (by reference) to the function to count the number of iterations. It can be written
as
double bisectionr(double a, double b, double (*f)(double),
double u, double delta, double epsn,
int maxit, int& itern) {
/*********************************************************
bisection algm: recursive version, returns an approximate
root of a fcn in a given interval
a, b: endpoints of interval in which a root lies
f:
function is defined in interval %a,b] or %b,a].
u:
= f(a)
delta: root tolerance, return when updated interval is
narrower than it
epsn: residual tolerance, return when residual is
smaller than it
maxit: maximum number of iterations allowed
itern: iteration count, it must be initialized to 1.
********************************************************/
double e = (b - a)*0.5
double c = a + e
double w = f(c)

// shrink interval size

// middle point
// fcn value at middle point

if (fabs(e)<delta || fabs(w)<epsn || itern++ > maxit)

return c
((u>0 && w<0) || (u<0 && w>0)) ? (b=c) : (a=c, u=w)

164

4. Namespaces and Files

return bisectionr(a,b,f,u,delta,epsn,maxit,itern)
}

The reference variable itern must be initialized to 1 before each call and
stores the number of iterations after a call is nished. For example,
int itrn = 1
double root = bisectionr(1.0e-2, 1, fa, fa(1.0e-2),
1e-8, 1e-9, 500, itrn)
cout << "Number of iterations used = " << itrn << '\n'
itrn = 1
// itrn must be initialized to 1
root = bisectionr(1, 3, fb, fb(1), 1e-8, 1e-9, 500, itrn)
cout << "Number of iterations used = " << itrn << '\n'

Run this program to know that the rst call takes 27 iterations and the
second call takes 28 iterations to stop. They produce the same approximate
roots as bisctn0():
The procedure above is generally called an incremental approach to software development. It starts with a simple version representing the main
idea of an algorithm. From there, more ecient and robust versions can be
built.
A nonrecursive version of the bisection algorithm can now be easily coded
as
double bisection(double a, double b, double (*f)(double),
double delta, double epsn, int maxit) {
/**********************************************************
bisection algm: nonrecursive version, returns approximate
root of a fcn in a given interval
a, b: endpoints of interval in which a root lies
f:
function is defined in interval %a,b] or %b,a].
delta: root tolerance, return when updated interval is
narrower than it
epsn: residual tolerance, return when residual is
smaller than it
maxit: maximum number of iterations allowed
*********************************************************/
double u = f(a)
double e = b - a
double c

// fcn value at left pt

// interval length
// store middle point

for (int k = 1 k <= maxit k++) {

// main iteration
e *= 0.5
// shrink interval by half
c = a + e
// update middle pt
double w = f(c)
// fcn value at middle pt

4.7 Iterative Methods for Nonlinear Equations

165

if (fabs(e) < delta || fabs(w) < epsn) return c

((u>0 && w<0) || (u<0 && w>0)) ? (b=c) : (a=c, u=w)
}
return c
}
// end bisection()

All the disadvantages in early recursive versions bisctn0() bisctn1() and

bisctn2() are avoided in this nonrecursive version. Both bisectionr() and
bisection() provide safeguards against entering innite loops. This is a good
consideration since roundo errors and numeric instability can lead to unpredictable behaviors.

4.7.2 Newton's Method

Let f (x) be a function of independent variable x f 0 (x) its rst-order derivative, and an initial approximation xp to a root r of f: Newton's algorithm
tries to nd a better approximation xn to this root r and repeat this process until convergence. The idea is to rst compute the root of its linear
approximation at xp and then let xn be this root. The linear approximation
of f (x) at xp is:

Lf (xp ) = f (xp ) + f 0 (xp )(x ; xp ):

This is exactly the rst two terms of its Taylor's expansion at xp : Geometrically, the curve f (x) is approximated near the point (xp f (xp )) by its
tangent line at this point. To nd the root of this linear approximation, set
Lf (xp ) = 0 that is,

f (xp ) + f 0 (xp )(x ; xp ) = 0

and solve for x to get x = xp ; f 0 (xp );1 f (xp ): Then let this x be the next
approximate root xn : That is,

xn = xp ; f 0 (xp );1 f (xp ):

This is the iterative formula of Newton's method. If jxn ; xp j < or

where and are two given small positive numbers, then stop
and take xn to be the approximate root of f (x) near r: Otherwise let xp =
xn and repeat this process. In order to prevent entering an innite loop,
this process should also be stopped when the total number of iterations has
exceeded a prescribed number. It can be coded into a C++ program as
jf (xn )j <

typedef double (*pfn)(double)

// define a function type

double newton(double xp, pfn f, pfn fd, double delta,

double epsn, int mxt) {
/**********************************************************

166

4. Namespaces and Files

newton's algorithm: finds an approximate root of f(x) = 0

xp:
an initial guess of a root
f:
the function whose root is to be found
fd:
the derivative function of f
mxt:
maximum number of iterations allowed
delta: program stops when distance of two iterates is < it
epsn: program stops when residual is less than it.
**********************************************************/
double v = f(xp)
// fcn value at initial guess
double xnew
// store new iterate
for (int k = 1 k <= mxt k++) {
// main iteration loop
double derv = fd(xp)
// derivative at xp
if (!derv) {
cout << "Division by 0 occurred in newton().\n"
exit(1)
// stop if divisor == 0
}
xnew = xp - v/derv
// compute new iterate
v = f(xnew)
// fcn value at new iterate
if (fabs(xnew - xp)<delta || fabs(v)<epsn) return xnew
xp = xnew
}
// end main iteration loop
return xnew
}
// end newton()

This function can be tested on a problem of nding a root of f (x) = x3 ;

5x2 +3x +7 near 5: Compiling and running the program (with appropriate
declarations and header les),
double f(double x) {
return ((x - 5)*x + 3)*x + 7
}

// nested multiply

double fder(double x) {
return (3*x - 10)*x + 3
}
int main() {
double root = newton(5, f, fder, 1e-8, 1e-9, 500)
cout << "Approx root near 5 by newton method is: "
<< root << '\n'
cout << "Fcn value at approximate root (residual) is: "
<< f(root) << '\n'
}

4.8 Exercises

167

gives the following output

Approx root near 5 by newton method is: 3.65544
Fcn value at approximate root (residual) is: 3.71647e-14

Thus Newton's method nds an approximate root with very small residual to this polynomial function. In general, the initial approximation xp
(also called initial guess) may not have to be very close to a root r: But a
good initial guess can speed up the convergence and even aect the convergence or divergence of the method. That is, Newton's method may converge
with one initial guess and diverge with another. Once a good initial guess
is obtained, Newton's method converges more rapidly than the bisection
method.
If f (r) = 0 and f 0 (r) 6= 0 then r is called a simple root of f: Newton's
method is designed for nding a simple root due to the following observation. If f 0 (r)
0 and xp is an approximation to r then the iterative
formula of Newton's method would involve a very small number or zero in
the denominator.
Let the second derivative f 00 be continuous and r a simple root of f:
Then there exists a neighborhood of r and a constant C such that if xp is
in the neighborhood, the next iterate xn is closer to r and
jr ; xn j C jr ; xp j2 :

See !KC96] for a proof. Such a convergence is called quadratic:

The root-nding functions can be put in a namespace. The common
declarations are put in a header le and object codes in a library. A user can
just include the header le and call the library for root-nding applications.
Notice that the bisection and Newton's methods are programmed in traditional C or FORTRAN style in this section. They require passing a function pointer as argument, which is hard to be inlined or optimized, and
may impose function calling overhead for each function evaluation f (x)
inside them. In x7.7 some ecient techniques are applied to overcome the
function calling overhead in passing a function pointer to another function
call.

4.8 Exercises
4.8.1. Implement matrix and vector namespaces as outlined in x4.1. Add
functions for vector addition, vector-scalar multiplication, matrix addition, and matrix-vector multiplication. Compute one, two, and max;1 and one, maximum, and Frobenius
imum norms of vector v = (vi )ni=0
n
;
1
norms of matrix m = (aij )ij=0 and the matrix-vector product mv
where vi = sin(i) and aij = cos(i2 + j ) for dierent matrix and vector sizes n = 100 500 and 1000: Manage the les by using a Makele
if you are on a UNIX/Linux system.

168

4. Namespaces and Files

4.8.2. Implement a matrix-vector library including vector and matrix norms,

and additions and multiplications in single, double, and long double
precisions. Put common declarations and type denitions in a header
le, implementations in a source le, and the main program in another source le. Test your program on the matrix m and vector v
dened in Exercise 4.8.1 with dierent sizes and precisions.
Note that three versions of the same functions in single, double, and
long double precisions seem to be tedious and hard to maintain. When
the implementation is to be changed, for example, the straightforward
evaluation of 2-norms is changed to a more robust evaluation (see
Exercise 4.8.5), then every version of the function has to be changed.
This has been done in a routine way in languages such as FORTRAN
77 and 90, and C. Using templates one version suces see Chapter
7.
4.8.3. Modify the program in x4.4.1 and compare the computing speeds in
double, single, and extended double precisions on your machine.
4.8.4. Compute the golden mean (x2.4) in float double and long double.
Write the results in a le with formatted output (e.g., scientic format, 30-character output width, 20 digits of precision, aligned to the
left).
4.8.5. To derive a robust vector 2-norm formula for a vector v = !v0 v1 : : :
vn;1 ] that will not cause over ow if kvk2 is within range, let r0 = 0
and ri2+1 = ri2 + vi2 for i = 0 1 2 : : : n ; 1: That is,
r12 = v02 r22 = v02 + v12 : : : rn2 = v02 + v12 + + vn2 ;1 :
p
Then kvk2 = rn2 = rn : From the recursive relation ri2+1 = ri2 + vi2
to avoid possible
over ow in summing the squares ri2 + vi2 comp
putepri+1 = ri 1 + (vi =ri )2 when ri is large and compute ri+1 =
jvi j 1 + (ri =vi )2 when jvi j is large. Thus an alternative but more
robust denition for Vec::twonorm() is:
double Vec::twonorm(const double* const v, int size) {
if (size > maxsize) cout << "vector size too large.\n"
double norm = fabs(v%0])
for (int i = 1 i < size i++) {
double avi = fabs(v%i])
if (norm < 100 && avi < 100)
norm = sqrt(norm*norm + avi*avi)
else if (norm>avi) norm *= sqrt(1 + pow(avi/norm,2))
else norm = avi*sqrt(1 + pow(norm/avi,2))
}
return norm
}

4.8 Exercises

169

This robust denition involves more operations and should be slower

than the straightforward evaluation as dened in x4.1. Write a program that compares the time eciency of these two denitions, on a
;1 where vi = cos(i) and n = 900000: Adjust the
vector v = (vi )ni=0
value of n if your computer takes too long or too short. A robust
algorithm is achieved sometimes at the expense of eciency.
4.8.6. Using the idea of Exercise 4.8.5, the function Mat :: frobnorm() (see
x4.1) can be dened in a more robust way:
double Mat::frobnorm(const double** const a, int size) {
double norm =Vec::twonorm(a%0],size) //2-norm of row 0
for (int i = 1 i < size i++) {
double avi = Vec::twonorm(a%i], size)
if (norm < 100 && avi < 100)
norm = sqrt(norm*norm + avi*avi)
else if (norm > avi) norm *= sqrt(1+pow(avi/norm,2))
else norm = avi*sqrt(1 + pow(norm/avi,2))
}
return norm
}

It can also be dened without using the function call Vec::twonorm(),

which leads to a possibly more ecient version:
double Mat::frobnorm(const double** const a, int size) {
double norm = 0
for (int i = 0 i < size i++) {
for (int j = 0 j < size j++) {
double avi = fabs(a%i]%j])
if (norm < 100 && avi < 100)
norm = sqrt(norm*norm + avi*avi)
else if (norm > avi)
norm *= sqrt(1 + pow(avi/norm,2))
else norm = avi*sqrt(1 + pow(norm/avi,2))
}
}
return norm
}

Write a program that compares the time eciency of the three versions of the matrix Frobenius norm function, on an n n matrix
evaluated repeatedly m times. You may take n = 1000 and m = 100:
Adjust the values of n and m to t your computer.

170

4. Namespaces and Files

4.8.7. Time the program computing the rst 40 Fibonacci numbers in x2.4
and the recursive program in Exercise 3.14.15. To clearly see the time
eciency (or ineciency) of recursive and nonrecursive function calls,
invoke each function inside a loop for m times, where m is a large
integer (e.g., m = 20000). Output the results directly into a le. Write
your own Makele for compiling and managing the programs if you
are using UNIX or Linux.
4.8.8. Write a function that takes a decimal int as argument and prints out
(to the screen and a le) the integer in hexadecimal, octal, and binary
formats.
4.8.9. Call all the functions (dened in x4.7.1) for the bisection method
inside a main() program to nd a root of
(a) f (x) = x;1 ; tan x !a b] = !0 2]
(b) f (x) = x ; e;x !a b] = !1 2]:
Comment on the advantages and disadvantages of these methods.
4.8.10. Apply Newton's method to nd a root of the function f (x) = 1 ; x ;
e;x near x = 1 and a root of g(x) = tan x ; x near x = 4:5 (test
what you get if starting from x = 5 instead).
4.8.11. Amounts A1 A2 : : : An are placed in an investment account (e.g., a
bank savings account) at the beginning of month 1 2 : : : n respectively. Let Bn be the balance in the account just after amount An
has been invested at the end of month n: The average annual return
rate r (assuming monthly interest compounding) of this investment
account is a root to the equation:

Bn =

i=1

n;i
r
:
Ai 1 + 12

Write a function that computes the average annual return rate r,

given Bn , A1 , A2 : : : and An : If a man invests 100 dollars each
month for the rst year, 200 300 400 and 500 dollars each month
for the second, third, fourth, and fth years, respectively, and has
balance 28000 dollars right after the 60th investment, what is his
average annual return rate (assuming monthly compounding)? Hint:
Newton's or the bisection method may be used to nd the root and
Horner's algorithm should be used in calculating the summation.
4.8.12. Amounts A1 A2 : : : An are placed in an investment account (e.g., a
mutual fund account) at the beginning of month 1 2 : : : n respectively. Let Bn be the balance in the account just after amount An has
been invested at the end of month n: The average annual return rate

4.8 Exercises

171

r (assuming continuous interest compounding it is very close to daily

compounding) of this investment account is a root to the equation:

Bn =

i=1

Ai er(n;i)=12 :

Write a function that computes the average annual return rate r,

given Bn , A1 , A2 : : : and An :
4.8.13. The library function system() in <stdlib:h> provides access to operating system commands. A string can be passed to it as the argument
and this string is treated as an operating system command. For example, the program
#include <stdlib.h>
int main() {
system("date")
}

executes the command date that causes the current date to be output
to the screen. Write a program on UNIX/Linux that lists all les with
a sux :cc using the following statements.
char command%30]
sprintf(command, "ls %s", "*.cc") // include <stdio.h>
system(command)
// include <stdlib.h>

Can you write a program that moves every :cc le into a corresponding
:cpp le?
A related function is asm() whose argument is a string literal representing a piece of assembly code that will be inserted into the generated code at the place where it is specied.

172

4. Namespaces and Files

This is page 173

Printer: Opaque this

Classes

The C++ construct class provides an encapsulation mechanism so that a

user can encapsulate data and functions manipulating the data together
to dene new types, which can be used as conveniently as built-in types.
This can also increase the readability and maintainability of a program,
and separate implementation details from its user interface.
Starting with this chapter, advanced features of C++ on object-oriented
programming and generic programming are introduced. However, in this
chapter, only basic concepts about class are discussed. This chapter ends
with a section on numeric methods for ordinary di
erential equations.

5.1 Class Declarations and Denitions

A class is a user-dened type that may contain data members and function
members that manipulate the data. There can also be functions, called
friends, which are not members of the class but have access to members
of the class. A class may have a private part that may be accessed only by
members and friends of the class, and a public part that can be accessed
freely in the program. A member of a class is private by default and can
also be explicitly declared private. The words class private, public and
friend are reserved words. Thus a class provides certain access restriction
for some of its members and achieves encapsulation and information hiding.
Objects of a class are created and initialized by its member functions
specically declared for this purpose such functions are called constructors.

174

5. Classes

A constructor is recognized by having the same name as the class itself.

A member function can be specically declared to \clean up" objects of a
class when they are destroyed such a function is called a destructor.
As an example, consider dening pt2d to be a class with x and y coordinates of a two-dimensional point as data members and a few functions
manipulating the data:
class pt2d {
private:
double x
double y
public:
pt2d(double r, double s) {
x = r y = s
}

//
//
//
//
//
//

class for 2D points

private members
x coordinate
y coordinate
public members
constructor

void move(double, double)

void move(pt2d)
void draw() const
friend double norm(pt2d)
}

//
//
//
//

move to new position

move to new position
a const member
a friend

Notice a semicolon is needed after a class declaration. A class declaration is actually a class denition since it denes a new type. The members x and y are private and can only be accessed by member functions
pt2d::pt2d(), pt2d::move(), pt2d::draw(), and friend norm(), where the function pt2d::pt2d() is dened inside the class declaration and others are dened below. The member pt2d :: pt2d() whose name is identical to the name
of the class, is called a constructor and all objects of class pt2d must be
created by the constructor. For example, the statements
pt2d p(5.4, 6)
pt2d q(5, 0)

// call constructor
// call constructor

create an object called p of type pt2d with x-coordinate 5:4 and y-coordinate
6, and an object q with coordinates x = 5 and y = 0: A destructor is not
necessary for this class since the constructor does not allocate space for it
using the operator new. Many examples in later sections need destructors
see, for example, x5.2. Classes are di
erent from other types in that an
object of a class must be constructed through a constructor. In this sense,
objects of a class are di
erent from other objects.
The class scope operator :: must be used for denitions of members outside a class declaration. A member function such as pt2d::draw(), declared
using the const sux, can only read but can not modify members of the
class. The keyword const here is part of the type and must not be omitted
in the denition when dened outside the class declaration. In contrast,
nonconstant member functions imply that they modify members of the

5.1 Class Declarations and Denitions

175

class and should not be invoked on constant class objects. Denitions of

other function members and the friend can be:
inline void pt2d::move(double r, double s) {
x = r y = s
// move point to (r, s)
}
inline void pt2d::move(pt2d p) {
x = p.x y = p.y
}

// move point to p

inline void pt2d::draw() const {

// draw coordinates
cout << '(' << x << ',' << y << ')'
// const member
}
inline double norm(pt2d p) {
return sqrt(p.x*p.x + p.y*p.y)
}

// friend definition
// distance to origin

Notice the di
erence between the ways that members are accessed by a
function member such as move() and a friend such as norm(). In move()
data members x and y can be used directly (they refer to members of an
object for which the member function is called), while the class qualier
must be used in norm() such as p:x and p:y: Dening move() to be a
member function and norm() a friend here is only to show the di
erence
on how to dene member functions and friends. The function norm() may
also be alternatively dened as a member and move() as friends. A friend
does not belong to a class but has access to members of the class. More
details on friends are given in x5.3. Member functions dened inside a class
declaration are inline by default. They can also be alternatively made
inline outside the class declaration like pt2d :: move() above. Again, inline
is only a suggestion to the compiler for code substitution to avoid function
calling overhead. Short member functions and friends can be made inline
for possible run-time eciency. Long denitions should be put outside a
class declaration to make the interface look clean.
The following main program denes a few objects of class pt2d and calls
member functions and friend to show how this class can be used.
int main() {
pt2d a(0, 0)
pt2d b(5, 0)
pt2d c(3, -40)
a.move(23, -3.5)
b.move(c)
c.draw()

// create object a = (0,0)

// create object b = (5, 0)
// create object c = (3, -40)
// move a to (23, -3.5)
// move b to (3, -40)
// draw point c

176

5. Classes

double dist = norm(c)

// distance from c to origin

In the above, member functions move() and draw() are called using the
member selection operator : following a class object, while a friend norm()
is called in the same way as an ordinary function.
The following statements in f () are illegal.
void f(pt2d a) {
double ax = a.x
pt2d d
}

// can not access private member

// error, initializer missing

That is, private members of a class may only be accessed by its member functions and friends as member x can be used in member function
move() and friend norm() but not in other functions such as main() and
f () above to prevent them from accidental modication or misuse. Since
the constructor of pt2d requires two arguments, an object can not be constructed (for d in f () above) without specifying them. The protection of
private members in a class leads to a safer user-dened type that is also
easier to maintain. If the representation of a class is to be changed, only
its member functions and friends need be changed. User code depends only
on the public interface and need not be changed.
It is sometimes convenient to provide several constructors the compiler
can choose the correct one to use according to the type of its arguments
just as in function overloading. For example, the class pt2d can be dened
to have three constructors:
class pt2d {
public:
// in addition to other members
pt2d() { x = y = 0 }
// default constructor
pt2d(double p) { x = p y = 0 }
pt2d(double p, double q) { x = p y = q }
}

Now objects of pt2d can be more conveniently constructed as

pt2d a
pt2d b(5)
pt2d c(3, -40)

// a = (0,0), default constructor

// b = (5, 0)
// c = (3, -40)

A constructor that takes no arguments is called a default constructor. Thus

the object a above is constructed by the default constructor of class pt2d:
These three constructors can also be combined into one:
class pt2d {
public:
// in addition to other members
pt2d(double p = 0, double q = 0) { x = p y = q }
}

5.1 Class Declarations and Denitions

177

This is nothing but providing default arguments to a function.

An object of a class can be used to initialize another object and assigned
to another object. For example,
pt2d b(5)
pt2d c(3, -40)
pt2d d = b
c = d

//
//
//
//

b = (5, 0)
c = (3, -40)
initialize d to b
assign d to c

In the above, d is constructed by a memberwise copy from object b which is

called copy construction or copy initialization. The assignment c = d causes
the content of d to be assigned to c by a memberwise copy such an assignment is called copy assignment. Copy construction and copy assignment
are di
erent operations in that a copy construction creates a new object
while a copy assignment potentially modies an object already created.
By denition, a struct is a class with all members public. For example,
struct pt3d {
double x, y, z
}

// all members are public

The members x y and z are public by default and can be accessed by

any function. Constructors, destructor, and other member functions may
also be dened for it. When a class does not provide a constructor, the
compiler tries to generate a default constructor for it to use. For example,
in the statement
pt3d p

the compiler creates an object p of type pt3d and initializes the members x,
y, and z to some default value. But there are classes for which the compiler
can not generate a default constructor see x5.6.
The keyword this refers to a constant pointer that points to the object for
which a member function is invoked. For a nonconstant member function
of class X , the pointer this is of type X const, and for a constant member
function, this has type const X const. For example, the member functions
move() and draw() can also be implemented as
inline void pt2d::move(double r, double s) {
this->x = r this->y = s
// move point to (r, s)
}
inline void pt2d::draw() const {
// draw coordinates
cout << '(' << this->x << ',' << this->y << ')'
}

Using this when referring to members is usually unnecessary. Its main use
is for dening member functions that manipulate pointers directly. Useful
examples are given in x5.2 and the next chapter.

178

5. Classes

As another example, we now dene a class for denite integrals that

contains two data members for the lower and upper integral bounds, another data member as a pointer to a function for the integrand, and a few
function members and a friend for manipulating and evaluating the definite integral such as the Trapezoidal Rule and Simpson's Rule. Numeric
integration was discussed in x3.13 in the procedural programming style as
in C and FORTRAN 90. Now class can be used to achieve encapsulation
of data and functions and information hiding. It can be declared as
typedef double (*pfn)(double)

// define a function pointer

class integral {
// members by default are private
double lower
// lower integral bound
double upper
// upper integral bound
pfn integrand
// integrand function
public:
// public members
integral(double a, double b, pfn f){
// a constructor
lower = a upper = b integrand = f
}
double lowbd() const { return lower } // const fcn member
double upbd() const { return upper }
// const fcn member
void changebd(double, double)
// nonconst member
double trapezoidal(int) const
// const fcn member
friend double simpson(integral, int)
// a friend
}
// note semicolon

The members lower upper, and integrand are private and can be accessed
only by its function members such as integral :: lowbd() and friends such as
simpson(): Thus public functions integral :: lowbd() and integral :: upbd()
are provided to return the lower and upper integral bounds. When a class
has a constructor, all objects of the class must be constructed by calling a
constructor. For example, the declaration
integral di(0, 5.5, sqrt)

declares di to be of type integral and creates an object with initialization

lower = 0:0 upper = 5:5 and integrand = sqrt where sqrt() is the squareroot function in <math:h>: A destructor is not needed for the class since
the user does not explicitly allocate space for it using the operator new:
Denitions of other function members and the friend can be:
inline void integral::changebd(double a, double b) {
lower = a
// change integral bounds to a, b
upper = b
}
double integral::trapezoidal(int n) const {

// const member

5.1 Class Declarations and Denitions

179

double h = (upper - lower)/n

// size of subinterval
double sum = integrand(lower)*0.5
for (int i = 1 i < n i++) sum += integrand(lower + i*h)
sum += integrand(upper)*0.5
return sum*h
}
double simpson(integral ig, int n) {
double h = (ig.upper - ig.lower)/n
double sum = ig.integrand(ig.lower)*0.5
for (int i = 1 i < n i++)
sum += ig.integrand(ig.lower + i*h)
sum += ig.integrand(ig.upper)*0.5

// a friend

double summid = 0.0

for (int i = 1 i <= n i++)
summid += ig.integrand(ig.lower + (i-0.5)*h)
return (sum + 2*summid)*h/3.0
}

The
following main program can be used to evaluate integrals
R7
and 3 sin(x)dx.

sin(x)dx

int main(){
integral di(0, 5, sin)
// sin from <math.h>
double result = di.trapezoidal(100) // call a fcn member
cout << "Integral from " <<di.lowbd() <<" to "<< di.upbd()
<< " is approximately = " << result << '\n'
di.changebd(3, 7)
// change bounds
result = di.trapezoidal(100)
cout << "Integral from " <<di.lowbd() <<" to "<< di.upbd()
<< " is approximately = " << result << '\n'
result = simpson(di,200)

// invoke a friend

More complicated denite integrals can be approximated similarly. Ecient

integration techniques are discussed in x7.7.
Protecting private members in a class leads to a safer user-dened type
that may also be easier to maintain. If the representation of the integral
class is changed to:
class integral {
double bound2]

// lower, upper bounds

180

5. Classes

typedef double (*pfn)(double)

pfn integrand
// .... other members
}

// a member that is a type

// integrand function

then only the member functions and friends have to be changed and the
user code as in the main() function above does not have to be changed.
This slight modication of the representation for integral has hardly any
advantage, but there are many situations in which a change of class representation can improve eciency or portability. For example, a matrix class
can be changed to be represented by a valarray (see x7.5) to make use of
the high performance computing library.

5.2 Copy Constructors and Copy Assignments

By default, a memberwise copy is provided for copy initialization and copy
assignment for a class, for which a copy constructor and assignment operator are not explicitly dened such as classes pt2d and integral in x5.1. This
kind of compiler-generated copy constructor and copy assignment should
be used when a class does not have a destructor that deallocates space using
delete since they may be more ecient than user-dened copy operations.
For example,
pt2d d = b
c = d

// initialize d to b
// assign d to c

where every member of b is copied to d in the copy initialization in the rst

statement, and every member of d is copied to c in the assignment in the
second statement. Such default memberwise initialization and assignment
were already mentioned in x3.4 for structures.
If the construction of an object of a class has allocated space using the
new operator, then a destructor is needed to clean up the object when it
goes out of scope, using the operator delete: The name of the destructor
for a class X must be eX (): For example,
class triangle {
pt2d* vertices
// pt2d defined in previous section
public:
triangle(pt2d, pt2d, pt2d)
// constructor
~triangle() { delete] vertices } // destructor
double area() const
// a function member
}
triangle::triangle(pt2d v0, pt2d v1, pt2d v2) {
vertices = new pt2d 3] // 3 vertices in a triangle
vertices0] = v0

5.2 Copy Constructors and Copy Assignments

181

vertices1] = v1
vertices2] = v2
}
void f() {
pt2d x(1.0,2.0)
pt2d y(3.0)
pt2d z
pt2d z2(7.0)
triangle t1(x,y,z)
triangle t2(x,y,z2)
}

When f () is called, the triangle constructor is called for t1 and t2 to allocate

space for their vertex points. At the return of f () the space for the vertices
are deallocated by the destructor automatically. A destructor need not
be called explicitly by the user. Instead, the system calls the destructor
implicitly when an object goes out of scope. Note that t1 is constructed
before t2 and t2 is destroyed before t1: That is, objects are destroyed in
the reverse order of construction.
Since triangle has a destructor that deallocates space using delete the
default memberwise copy constructor and assignment operator would not
work correctly in this case:
void g(pt2d x, pt2d y, pt2d z) {
triangle t1(x,y,z)
triangle t2 = t1
// copy initialization, trouble
triangle t3(x,y,z)
t3 = t1
// copy assignment, trouble
}

In this code segment, the copy initialization made the member vertices of
t2 equal to that of t1 (since they are pointers, being equal implies that
they point to the same object), and the copy assignment made the member
vertices of t3 equal to that of t1: Consequently, the member vertices of t1
t2 and t3 point to the same object. At the time t1, t2 and t3 are destroyed
when g() is returned, the same pointer is deleted three times (its behavior is
undened) and results in an error. In general, for a class with a destructor
that deletes space allocated by its constructor, the user must dene a copy
constructor and a copy assignment. For a class T , they have the form
T::T(const T &)
T& T::operator=(const T&)

// copy constructor
// copy assignment

where operator is a keyword that allows the assignment operator = to be

overloaded. In the case of the triangle class, they can be dened as
triangle::triangle(const triangle & t) { // copy constructor

182

5. Classes

vertices = new pt2d 3]

for (int i = 0 i < 3 i++) verticesi] = t.verticesi]
}
triangle& triangle::operator=(const triangle& t) {
if (this != &t)
// beware of self-assignment like t = t
for (int i = 0 i < 3 i++) verticesi] = t.verticesi]
return *this
}

These two member functions must also be declared in the class triangle:
Since this is a constant pointer that points to the object for which a member function is invoked, this is then the object. The user-dened copy
assignment returns a reference to the object. Notice that new space is allocated in the copy initialization and no new objects are created in the copy
assignment, which clearly reects the fact that initialization and assignment are two distinct operations. With the user-provided copy constructor
and assignment above, the function g() above should work correctly now.
In particular, at the time t1 t2 and t3 are destroyed, the three pointers,
for which spaces are allocated when they are constructed, are deleted. The
redenition of the assignment operator is called operator overloading for =.
Operator overloading is discussed in more detail in the next chapter.
As another example, a class called triple (consisting of three numbers)
can be declared as
class triple {
float* data
public:
triple(float a, float b, float c)
~triple() { delete] data }
triple(const triple& t)
triple& operator=(const triple& t)

// a triple of numbers

//
//
//
//

constructor
destructor
copy constructor
copy assignment

friend triple add(const triple&, const triple&) // friend

void print() const {
// print out members
cout << data0] << " " << data1] << " "
<< data2] << '\n'
}
}

Member denitions inside a class declaration are inline by default. Other

members and friend (also made inline since they are short) can be dened
as follows.
inline triple::triple(float a, float b, float c) {
data = new float 3]
data0] = a data1] = b data2] = c

5.2 Copy Constructors and Copy Assignments

183

}
inline triple::triple(const triple& t) { // copy constructor
data = new float 3]
for (int i = 0 i < 3 i++) datai] = t.datai]
}
inline triple& triple::operator=(const triple& t) {
if (this != &t)
// beware of self assignment
for (int i = 0 i < 3 i++) datai] = t.datai]
return *this
}
inline triple add(const triple& t, const triple& s) {
return triple(t.data0] + s.data0], t.data1] + s.data1],
t.data2] + s.data2])
}

Now the class triple may be used as

int main() {
triple aaa(5,6,7)
triple bbb(10,20,30)
triple ccc = aaa
triple ddd = add(aaa,bbb)
ccc = add(aaa,ddd)
ccc = ccc
ccc.print()

// create an object
// copy construction
// copy construction
// copy assignment
// self-copying is allowed
// print out members of ccc

If a class duple (consisting of n numbers) is to be dened, the copy

assignment operation may need to take care of assignments of duples with
di
erent sizes. To be more precise, consider
class duple {
//
int size
//
double* data
//
public:
duple(int n, double* a) {
//
data = new double size = n]
for (int i = 0 i < size i++)
}
~duple() { delete] data }
duple(const duple&)
duple& operator=(const duple&)

a duple of 'size' numbers

how many numbers?
store the numbers
constructor
datai] = ai]
// destructor
// copy constructor
// copy assignment

184

5. Classes

friend duple add(const duple& t, const duple& s)

void print() const
}
duple& duple::operator=(const
if (this != &t) {
if (size != t.size) {
delete] data
data = new double size
}
for (int i = 0 i < size
}
return *this
}

duple& t) {

// copy assignment

// if sizes are different

// delete old data
= t.size] //space for new data
i++) datai] = t.datai]

When a duple t of one size is to be assigned to a duple s of another size,

the space for s is rst deleted and then allocated with the same number of
elements in t: This will make the following code work correctly:
double xx] = { 5, 6, 7}
double* zz = new double 10]
for (int i = 0 i < 10 i++) zzi] = i + 1
duple d1(3, xx)
duple d2(8, zz)
d1 = d2

// d1 contains 3 numbers
// d2 contains 8 numbers
// d1 is resized to have size 8

In the assignment d1 = d2, the object d1 is resized to have the same

number of elements as d2 according to the denition of the copy assignment
operator =. Exercise 5.10.5 asks for the denitions of other members in

duple:

5.3 Friends
A friend can belong to many classes and have access to private members of
objects of these classes, while a function, data, or other member can only
belong to one class. A friend declaration can be put in either the private
or public part of a class it does not matter where. For example, a matrixvector multiply function can be declared to be a friend of classes Mat and
Vec:
class Mat
// forward declaration
class Vec {
// .... in addition to other members
friend Vec multiply(const Mat&, const Vec&)
}

5.4 Static Members

185

class Mat {
// .... in addition to other members
friend Vec multiply(const Mat&, const Vec&)
}

Matrix and vector classes are discused in more detail in Chapters 6 and 7.
A friend is not a class member and thus does not have a this pointer. A
member function of one class can be a friend of another class. A class can
be a friend of another in this case all functions of the class become friends.
For example,
class X {
void f()
int g(int)
}
class Y {
// .... in addition to other members
friend void X::f()
// f() of X becomes friend of Y
}
class Z {
// .... in addition to other members
friend class X
// all functions of X become
}
// friends of class Z

5.4 Static Members

Suppose we want to dene a class for vectors with default (unit, zero, or
other) vectors of certain sizes. One way to do this is to use global variables
to store the default vectors. However, too many global variables will make
code unmanageable except to its original programmer. Global variables
should be kept to the minimum. C++ provides static members for this
purpose without introducing global variables or functions. A member that
is part of a class but is not part of the objects of the class, is called a
static member: There is exactly one copy of a static member per class.
Note that there is one copy per object for ordinary nonstatic members.
Similarly, a function that needs access to members of a class but need not
be invoked for a particular object, is called a static member function. For
example,
class Vec {
private:
int size
double* entry
static Vec defaultVec
public:

// holding default vector

186

5. Classes

Vec(int sz = 0, double* et = 0) // constructor

Vec(const Vec&)
// copy constructor
Vec& operator=(const Vec&)
// copy assignment
static void setDefault(int, double*)
// function for setting default vector
}
Vec::Vec(int sz, double* et) { // if sz = 0 use default size
size = sz ? sz: defaultVec.size
entry = new double size]
if (et == 0) {
// if et is null pointer, use default
for (int i = 0 i < size i++)
entryi] = defaultVec.entryi]
} else {
for (int i = 0 i < size i++) entryi] = eti]
}
}

In the constructor, the default vector size and entries are used when
it is not supplied with any arguments. The copy constructor and assignment can be dened accordingly. Static data and function members such
as Vec::defaultVec and Vec::setDefault() must be dened somewhere in the
program. Since static members belong to the class, are shared by all objects
of the class, and do not belong to a particular object, they are qualied by
the name of the class using the scope resolution operator. For example,
double ad] = {5, 6, 7, 8, 9}
Vec Vec::defaultVec(5, ad)
// defaultVec initialized by ad of size 5
void Vec::setDefault(int sz, double* p) {
Vec::defaultVec = Vec(sz, p)
// reset default vector
}
// note class qualification Vec::defaultVec

Note that a static data member such as Vec::defaultVec is dened in the

same way as a variable is declared and initialized, very much like the way
a function is dened. Although it is private, it may not have to be dened
through a member function. A nonstatic data member can not be dened
this way. Still, private static members can not be accessed publicly. The
default vector can be changed when appropriate. For example,
void f() {
Vec a
Vec b(3)

// default vec of size 5, a ={5,6,7,8,9}

// default vector of size 3, b = {5,6,7}

double* ey = new double 8]

for (int i = 0 i < 8 i++ ) eyi] = 1

5.5 Constant and Mutable Members

Vec::setDefault(8, ey)
Vec a2
Vec b2(3)

187

// reset defaultVec to unit

// new default vec of size 8, a2={1,...,1}

// new default vec of size 3, b2={1,1,1}

double* enb = new double 9]

for (int i = 0 i < 9 i++ ) enbi] = 8*i*i + 100
Vec c(9,enb)
// c is a vector constructed from enb.
}

Since a and a2 are not supplied with arguments, default vectors (with default sizes and default entries) are constructed. Note the default vector is
reset by calling the static member Vec::setDefault() before the declaration
of a2: Thus a and a2 are constructed di
erently. Similarly, b and b2 are constructed di
erently although their declarations look exactly the same. In
contrast, c is constructed according to the arguments supplied to the constructor. From this example, it can be seen that it would be a waste of space
if each object such as b and c had a copy of static member Vec::defaultVec.
That is, although class Vec has member defaultVec, an object of Vec does
not have its own copy of this member. Instead, all objects of Vec share this
member. In other words, the member defaultVec belongs only to class Vec.
In contrast, every object of Vec has a copy of nonstatic members such as
size and entry:

5.5 Constant and Mutable Members

Constant member functions are not supposed to modify objects of a class.
However, if a data member is declared to be mutable then it can be changed
by a const member function. For example, the denite integral class in x5.1
can be redened so that it has a data member for storing an approximate
value of the integral by the Trapezoidal Rule:
class integral {
private:
double lower
double upper
double (*integrand)(double)
mutable double value
// store value from trapezoidal
mutable bool value_valid // value valid or not
public:
integral(double a, double b, double (*f)(double))
void changebd(double, double)
double trapezoidal(int = 100) const
// const member
}

188

5. Classes

From the user's point of view, the function trapezoidal() does not change
the state of an object of integral. This is why it is declared to be a const
function. The reason for declaring a mutable data member value for storing
an approximate value of the integral is that trapezoidal() might be an
expensive process and repeated invocation of it can just return the denite
integral value without repetitively calculating it. Now trapezoidal() and
other members of integral can be dened as
inline integral::integral(double a, double b,
double (*f)(double)) {
lower = a upper = b integrand = f value_valid = false
}
inline void integral::changebd(double a, double b) {
lower = a upper = b value_valid = false
}
double integral::trapezoidal(int n) const { // const member
if (value_valid == false || n != 100) {
double h = (upper - lower)/n
// compute integral
double sum = integrand(lower)*0.5
for (int i = 1 i < n i++) sum += integrand(lower+i*h)
sum += integrand(upper)*0.5
value = sum*h
value_valid = true

// update mutable member

}
return value
}

This newly dened class integral may be used as

integral di(7.7, 9.9, log)
double d = di.trapezoidal()
double e = di.trapezoidal()

// do the calculation
// just returns value

Observe that the mutable member value is modied by a const member

function trapezoidal(): In the rst call to trapezoidal() the approximate
value of the integral is computed and assigned to variable d, while in the
second call this value is just returned to variable e without recomputing
it. The computation is not repeated in the second call since value valid is
set to true in the rst call. This technique is especially useful for adaptive
integration methods in which the number of subintervals n is not specied
as an argument to trapezoidal(), but rather is computed by an adaptive
algorithm. Since such adaptive algorithms are often expensive, avoiding
repetitive execution of them can potentially save a lot of time.
A mutable member of a const object can also be modied. For example,

5.6 Class Objects as Members

integral di2(1.1, 2, log)
const integral di3(1.1, 5, exp)
double dd = di2.trapezoidal()
double ee = di3.trapezoidal()

//
//
//
//

di2.changebd(5, 9)
di3.changebd(5, 9)

// OK
// error, const object di3

189

log() in <math.h>
a const object
const member modifies di2
OK, although di3 is const

The mutable member value of the const object di3 is modied by the
const member function trapezoidal(): Regular (non-mutable) members of
a const object can not be modied. Thus di3:changebd() is an illegal call
since changebd() modies regular members and di3 is a const object.
An example of a class with a const data member is given at the end of
x5.6. A class with a static const data member is presented in x9.1.

5.6 Class Objects as Members

Objects of one class can be members of another class. The initialization of
such members can be done di
erently from other members. For example,
dene a class for line segments (in two dimensions) with a direction that
contains members of another class:
class line {
pt2d oneend
// one
pt2d otherend
// the
bool direction
// = 1
public:
// = 0
line(pt2d, pt2d, bool)
line(pt2d, bool)
}

end of the line segment

other end of the line segment
if from oneend to otherend
otherwise
// constructor
// constructor

line::line(pt2d a,pt2d b,bool dir): otherend(b), oneend(a) {

direction = dir
}
line::line(pt2d b, bool dir): oneend(), otherend(b) {
direction = dir
}

The rst constructor takes two points and a bool as arguments. Initialization for class object members (endpoints of the line segment) is done in
a member initializer list in the denition of the constructor. The member
initializers are preceded by a colon and di
erent member initializers are separated by commas. Initialization for ordinary members (such as direction
of line) is done inside the braces, which can also be done in the initializer

190

5. Classes

list. The rst constructor can create an object of class line with oneend
equal to class object a otherend equal to b, and direction equal to dir:
Constructors of the members are called before the body of the containing
class' own constructor is executed. The members' constructors are called
in the order in which they are declared in the class rather than in the order
in which they appear in the initializer list. The members' destructors are
called in the reverse order of construction. For example, the construction
for member oneend as an object of class pt2d is done before otherend when
constructing an object of class line.
In the denition of the second constructor above, the endpoint oneend of
line is taken to be the default point (the origin by the default constructor
of class pt2d see x5.1). Thus no argument is specied for it. In this case,
it can be equivalently written as one of the following.
line::line(pt2d b, bool dir): otherend(b) {
direction = dir
}
line::line(pt2d b, bool dir): otherend(b), direction(dir) { }

There is an alternative way of dening constructors for some classes. For

example, the rst constructor can also be dened as
line::line(pt2d a, pt2d b, bool dir): otherend(b) {
oneend = a
direction = dir
}

Here otherend is initialized to b, but oneend is rst initialized to the origin

(default construction for pt2d) and then a copy of a is assigned to it. Thus
there is an eciency advantage to using the initializer syntax, which avoids
default construction and initializes oneend to a directly. It can even be less
eciently dened as
line::line(pt2d a, pt2d b, bool dir)
oneend = a
otherend = b
direction = dir
}

// less efficient

These three versions are equivalent except for eciency.

However, member initializers are essential for member objects of classes
without default constructors, for const members, and for reference members. For example,
class AA {
const int i
line n

5.7 Array of Classes

191

line& rn
AA(int j,pt2d p,bool d,line& c): i(j), n(p,d), rn(c) { }
}

Since i is a const member, n has a type of class line that does not have a
default constructor, and rn is a reference member, they must be initialized
in the initializer list, instead of inside the braces. Since constants and references must be initialized, a class with a const or reference member can
not be default-constructed:
struct BB {
const int i
const float f
double& d
}
BB b
// error, no default constructor is provided
struct CC {
int i
float f
double d
}
CC c
// OK, use compiler-generated default constructor

Objects of classes without constructors or destructors can also be members of a union. A named union is dened as a struct, where every member
has the same address. A union can have member functions but not static
members. In general, a compiler does not know which member of a union is
used. Thus a union may not have members with constructors or destructors.

5.7 Array of Classes

If a class does not have a constructor or has a default constructor, then
arrays of that class can be dened. For example,
pt2d ap100]

denes an array of 100 two-dimensional points. Each of the points has been
initialized to the origin according to the default constructor of class pt2d
(see page 176).
However, if a class has constructors that do not provide default values,
an array of objects of such a class can not be declared. For example, if a
vector class is dened as
class Vec {
double* en
int size

192

5. Classes

public:
Vec(int, double* = 0)
void f(int)
}
Vec::Vec(int s, double*
en = new double size
if (d) {
for (int i = 0 i <
} else {
for (int i = 0 i <
}
}

// constructor
// a test function

d) {
= s]
// if d is not the null pointer
size i++) eni] = di]
// otherwise, use default
size i++) eni] = 0

then an n by n square matrix can not be dened directly as an array of n

vectors, each of which represents a row of n elements:
int n = 100
Vec matrixn]

// error, Vec has no default constructor

where each matrixi] i = 0 1 : : : n ; 1 is intended to be a Vec, but their

sizes can not be specied.
Instead, an array of n pointers (a double pointer), each of which points
to an object of class Vec, can be declared. For example,
void Vec::f(int n) {
Vec** tm = new Vec* n]
for (int i = 0 i < n i++) {
tmi] = new Vec(n)
}

//
//
//
//

allocate space for matrix

tmi] points to an
object of zero-vector of
size n, represent row i

for (int i = 0 i < n i++)

for (int j = 0 j < n j++) // assign values to entries
tmi]->enj] = 1.0/(i + j + 1)
for (int i = 0 i < n i++ ) delete tmi]
delete] tm
// free space after use
}

Thus tm is an array of n pointers, each of which points to an object of

Vec that contains n elements representing a row of a matrix. Since tmi]
is a pointer, its member should be accessed by using the operator ;>
(in a member function or friend, since en is a private member of Vec).
For example, tmi];>enj ] represents the entry at row i and column j
of the matrix. This kind of awkward notation can be replaced by using
tmi]j ] through operator overloading see Chapter 6. Because tmi] is a
single pointer that points to only one object (although this object contains

5.8 Pointers to Members

193

n elements in a row), a simple delete, instead of delete ], should be used

when deallocating space for it. This technique can be very useful and may
also be used to dene a triangular matrix as an array of pointers see x11.3
for an application.

5.8 Pointers to Members

It is sometimes useful to deal with pointers to member functions of a class.
A pointer to a member can be thought of as a position of the member in
an object of the class, like an index in an array. It is di
erent from pointers
to ordinary functions. Here is a simple example:
class X {
double g(double a) { return a*a + 5.0 }
double h(double a) { return a - 13 }
public:
void test(X*, X)
}
typedef double (X::*pf)(double)
void X::test(X* p, X q) {
double g5 = p->g(5)
double h5 = p->h(5)
double g10 = q.g(10)
double h10 = q.h(10)
pf m1 = &X::g
pf m2 = &X::h
double g6 = (p->*m1)(6)
double h6 = (p->*m2)(6)
double g12 = (q.*m1)(12)
double h12 = (q.*m2)(12)

// define member pointer

//
//
//
//
//

p is
call
call
call
call

pointer, q has type X

member directly
member directly
member directly
member directly

//
//
//
//
//
//

m1 is a pointer to member
m2 is a pointer to member
call thru pointer to member
call thru pointer to member
call thru pointer to member
call thru pointer to member

Here pf is dened to be a type of a pointer to a member of class X: A

pointer to a member of a class can be obtained by applying the address-of
operator & to a qualied class member name, as in &X :: g and &X :: h: The
operators ; > and : can be used to bind pointers to members to class
objects. If m is a pointer to a member, then p;>m binds m to the object
pointed to by p, and b: m binds m to the object b: This technique can be
used to evaluate integrals with di
erent integrand functions. Pointers to
members can be used in object-oriented nite element analysis code where
di
erent coecient functions and basis functions have to be evaluated.

194

5. Classes

A static member of a class is not associated with a particular object. Thus

a pointer to a static member is just a pointer to an ordinary function. For
example,
struct Y {
static void f(int)
}
void (*p)(int) = &Y::f
// OK, p is a pointer to an ordinary function
void (Y::*pm)(int) = &Y::f
// illegal, ordinary pointer assigned to pm

Here p is declared to be a pointer to an ordinary function that takes an int

as argument and returns nothing, and pm is a pointer to a member function
of class Y that takes an int and returns nothing. The rst initialization is
legal since the address of static member Y :: f can be assigned to an ordinary
function pointer p, while the second is an error since the address of a static
member is assigned to a class member pointer pm.

5.9 Numeric Methods for Ordinary Dierential

Equations
Many physical problems are modeled by di
erential equations. In this section, some basic numeric methods for solving ordinary di
erential equations
are presented.
A rst-order ordinary di
erential equation with a given initial value is
usually written in the form

dx = f (t x) t < t T

0
dt
x(t0 ) = x0

(5.1)
(5.2)

where x(t) is the unknown function depending on t (often representing

time) and f (t x) is a given (source) function depending on t and x. The
independent variable t changes from t0 to T and the initial value of x at
t = t0 is given as x0 . An example is:

dx = (1 ; et )x 0 < t 2
dt
1 + et
x(0) = 3

(5.3)
(5.4)

The exact solution x(t) can not be found in general. Instead, approximations to the solution x(t) are sought at certain points t0 < t1 < : : : <
tN = T for a positive integer N: Let the approximate solution at t = tk be

5.9 Numeric Methods for Ordinary Di erential Equations

195

denoted by xk k = 0 1 : : : N with x0 being the given exact initial value.

The objective of a numeric method is to nd values x1 x2 : : : xN which
approximate the exact solution values x(t1 ) x(t2 ), : : : and x(tN ) respectively. One simple such method is to approximate the derivative dx=dt at
t = tk by a nite di
erence (xk+1 ; xk )=hk where hk = tk+1 ; tk is called
the grid size. This is because a derivative is dened to be the limit of
di
erence quotients:

dx = lim x(tk+1 ) ; x(tk ) :

dt t=t

hk 0

The smaller the grid size hk is, the more accuracy may be expected of
this approximation in the absence of roundo
errors and numeric instability.
That is, from (5.1),

xk+1 ; xk
hk

dx = f (t x(t )) f (t x )

k
k
k k
dt t=t
k

which leads to the so-called explicit Euler's method:

xk+1 = xk + hk f (tk xk ):

(5.5)

Since x0 is given, x1 can be obtained directly from the formula with k = 0:

Then x2 can be solved by letting k = 1 and using x1 : This procedure can
go on until xN is obtained. For simplicity, equally spaced points are often
used that is, tk = t0 + kh k = 0 1 : : : N with grid size h = (T ; t0 )=N:
Euler's method (5.5) may be programmed as
class ode {
double tini
// initial time
double ison
// initial solution
double tend
// end time
double (*sfn)(double t, double x)
// source function
public:
ode(double t0, double x0, double T,
double (*f)(double, double)) {
// constructor
tini = t0 ison = x0 tend = T sfn = f
}
double* euler(int n) const
// Euler with n subintervals
}
double* ode::euler(int n) const {
// explicit Euler
double* x = new double n + 1]
// approximate solution
double h = (tend - tini)/n
// grid size
x0] = ison
// initial solution
for (int k = 0 k < n k++)
xk+1] = xk] + h*sfn(tini + k*h, xk])

196

5. Classes

return x
}

Here data (tini for initial time, ison for initial solution, tend for ending
simulation time, and sfn for the source function) about the initial value
problem and method euler() are encapsulated into a class called ode: The
private data members can only be accessed by function member euler() and
others dened later. Copy construction and assignment and a destructor
need not be dened for such a class. The user needs to supply the number
of subintervals N to the function euler() that returns a pointer to the
approximate solution values x0 x1 : : : xN :
For the example (5.3)-(5.4), the exact solution happens to be found as
x(t) = 12et=(1 + et )2 . It can be used to check the accuracy of the approximate solution by Euler's method:
double f(double t, double x) {
// source function
return x*(1 - exp(t))/(1 + exp(t))
}
double exact(double t) {
// exact solution
return 12*exp(t)/pow(1 + exp(t), 2)
}
int main() {
ode exmp(0, 3, 2, f)
double* soln = exmp.euler(100)

// x(0) = 3, T = 2
// 100 subintervals

double norm = 0
double h = 2.0/100
// grid size
for (int k = 1 k <= 100 k++)
// compute error
norm = max(norm, fabs(exact(k*h) - solnk]))
cout << "max norm of error by euler's method = "
<< norm << '\n'
}

The maximum norm of the error between the exact solution and the approximate solution is 0:00918355: With 200 subintervals (h = 1=100), the
error is 0:0045747: When h = 1=200, the error is 0:00228309: It can be observed that the error decreases linearly with the grid size h: That is, when
the grid size h is halved, the error is reduced by a factor of two. Mathematically, it can be shown that the error is proportional to h. Such a method
is said to have rst-order. For some problems, the proportionality constant
is reasonably small and the explicit Euler's method gives good accuracy.
However, for some other problems (see Exercise 5.10.17), the proportionality constant is very large, which causes the method to be very inaccurate.
In other words, the explicit Euler's method is not very stable.

5.9 Numeric Methods for Ordinary Di erential Equations

197

A modication to the explicit Euler's method is as follows.

xk+1 = xk + hk f (tk+1 xk+1 ):
(5.6)
This is called the implicit Euler's method. The only di
erence is that the
evaluation of the source function f (t x) is now at t = tk+1 : This scheme
requires some nonlinear methods such as Newton's algorithm (see x4.7)
to solve for xk+1 from (5.6). Its advantage is that it provides an accurate
solution for a wider range of problems than the explicit Euler's method.
See Exercise 5.10.17. In other words, the implicit Euler's method is more
stable than its explicit counterpart, but requires more computational work.
Based on Euler's method, a predictor-corrector scheme can be constructed
as

xk = xk + hk f (tk xk )
xk+1 = xk + hk f (tk+1 xk )

(5.7)
(5.8)

in which a prediction xk is computed in (5.7) based on the explicit Euler's

method and a correction xk+1 is obtained in (5.8) based on the implicit
Euler's method. Its advantage is to avoid solving a nonlinear equation as
opposed to the implicit Euler's method, and to provide more stability than
the explicit Euler's method.
Runge{Kutta methods are explicit since they do not require solving nonlinear equations. A second-order Runge{Kutta method is:

xk+1 = xk + (F1 + F2 )=2

(5.9)

where

F1 = hk f (tk xk )
F2 = hk f (tk+1 xk + F1 ):
A fourth-order Runge{Kutta method is:
xk+1 = xk + (F1 + 2F2 + 2F3 + F4 )=6
where

F1
F2
F3
F4

(5.10)

= hk f (tk xk )
= hk f (tk + hk =2 xk + F1 =2)
= hk f (tk + hk =2 xk + F2 =2)
= hk f (tk+1 xk + F3 ):

It can be shown that the error between the exact solution and numeric
solution from (5.9) is proportional to h2 (second-order accuracy) while the
error for (5.10) is h4 (fourth-order accuracy) with the proportionality constant depending on high-order derivatives of the exact solution. In contrast,

198

5. Classes

Euler's methods (5.5), (5.6), and (5.7)-(5.8) have accuracy proportional to

h (rst-order accuracy). When the solution to a problem is well behaved
(e.g., its high-order derivatives are not very large), high-order methods such
as fourth-order Runge{Kutta methods are preferred. For problems with a
transit area where the solution changes very abruptly such as Exercise
5.10.17 with = ;500 the implicit Euler's method may be preferred.
Now several more functions can be added to class ode for these methods.
For example,
class ode {
public:
// ... in addition to other members
double* eulerpc(int n) const // predictor--correct euler
double* rk2(int n) const
// second-order Runge--Kutta
}
double* ode::eulerpc(int n) const {
double* x = new double n + 1]
double h = (tend - tini)/n
x0] = ison
for (int k = 0 k < n k++) {
xk+1] =xk] +h*sfn(tini + k*h, xk])
xk+1] =xk] +h*sfn(tini +(k+1)*h, xk+1])
}
return x
}

// predictor
// corrector

double* ode::rk2(int n) const {

double* x = new double n + 1]
double h = (tend - tini)/n
x0] = ison
for (int k = 0 k < n k++) {
double tp = tini + k*h
double f = h*sfn(tp, xk])
xk+1] = xk] + 0.5*(f + h*sfn(tp + h, xk] + f))
}
return x
}

See AP98, CK99, KC96] for more details on numeric methods for ordinary di
erential equations. Notice that a function pointer is passed to
the constructor of class ode, which may impose a lot of function calling
overhead on function evaluations of sfn() inside member functions euler()
eulerpc() and rk2(): In x7.7, ecient techniques for avoiding such function
calling overhead are discussed.

5.10 Exercises

199

5.10 Exercises
5.10.1. Extend the class pt2d dened in x5.1, to three-dimensional points.
The new class should have several constructors including a default
constructor, member functions move() draw() and a friend for calculating the distance from a three-dimensional point to the origin.
5.10.2. Use the class for denite integrals dened in x5.1 to evaluate the
integrals in Exercises 3.14.22 and 3.14.23.
5.10.3. Add printing statements in the constructors and destructor of the
class triangle declared in x5.2. For example,
triangle::triangle(pt2d v0, pt2d v1, pt2d v2) {
cout << "allocating space in triangle\n"
vertices = new pt2d 3]
// ...
}
triangle::~triangle() {
cout << "deleting space in triangle\n"
delete] vertices
}

Write a main program calling the function g() dened in x5.2, and
compare the outputs without and with user-dened copy constructor
and assignment. You should see space being allocated the same number of times as being deleted with user-dened copy constructor and
assignment. What do you think will happen without them?
5.10.4. Dene the member function triangle :: area() declared in x5.2, to
compute the area of a triangle. Note the area of a triangle in a twodimensional space with vertices A, B and C is 12 j(B:x ; A:x)(C:y ;
A:y) ; (B:y ; A:y)(C:x ; A:x)j: Then use the triangle class to nd
the area of the triangle passing through three two-dimensional points
a(5 5) b(7 8) and c(20 ;4):
5.10.5. Dene all the function and friend members in class duple as discussed
in x5.2. When dening add() for the addition of two duples, an error
message can be rst printed out and the program then exited if the
two duples have di
erent sizes. Test a few examples on all the functions, especially copy construction and copy assignment, and print
out the data to see if they work as expected.
5.10.6. Dene a vector class that contains a pointer for the entries, an integer
for the size of the vector, and one, two, and maximum norm functions.
Test your denitions on a few simple vectors.

200

5. Classes

5.10.7. Dene a matrix class that contains a double pointer for the entries,
two integers for the dimension (numbers of rows and columns) of the
matrix, and one, maximum, and the Frobenius norm functions. Test
your denitions on a few simple matrices.
5.10.8. Dene a friend function for both the matrix class and the vector class
in Exercises 5.10.6 and 5.10.7 for matrix-vector multiplication. Test
your denitions on a few simple matrices and vectors.
5.10.9. Dene a class for a triangle given three vertices, each of which is represented by an object of another class for three-dimensional points.
Using member initializer lists, dene two constructors for it, one with
three points as arguments for the three vertices and another with two
points (the third vertex is defaulted to be the origin). Also dene a
member function that moves a triangle to a new location and another
member function that prints out the coordinates of the three vertices
of a triangle. Create a few triangle objects, move them to new locations, and print out the coordinates of the triangles before and after
moving them.
5.10.10. Implement the root-nding problem as discussed in x4.7 using a class
with the bisection method as a member function and Newton's method
as a friend. Then apply it to Exercises 4.8.9 and 4.8.10.
5.10.11. A numeric quadrature has the form:
b

f (x)dx

nX1
i=0

wi f (xi )

where xi are called quadrature points and wi are called weights. In a

Gauss quadrature, xi and wi are chosen such that the quadrature is
exact if f (x) is a polynomial of degree p for p as big as possible. Such
xi are called Gauss points and such p is called the degree of exactness.
It can be shown that the degree of exactness of a Gauss quadrature
is 2n ; 1 where n is the number of Gauss points. An example of a
Gauss quadrature is:
r
r
Z 1
3
8
5
5
f (x)dx 9 f (0) + 9 f (; 5 ) + 9 f ( 35 ):
;1
Implement this Gauss quadrature as a class with Gauss points and
weights as static members. Check that this quadrature is exact when
the integrand f (x) is a polynomial of degree less than or equal to 5:
5.10.12. In Gauss{Lobatto quadratures, the endpoints of the interval are also
quadrature points. An example of a Gauss{Lobatto quadrature is:
r
r
Z 1
5
1
5
1
f (t)dt 6 f (;1) + 6 f (; 5 ) + 6 f ( 15 ) + 16 f (1):
;1

5.10 Exercises

201

Write a class for numerically integrating an integral ab f (x)dx with

n (equally spaced) subintervals xi;1 xi ] (i = 1 : : : n), on each of
which the above Gauss{Lobatto quadrature is applied after a linear
transformation from xi;1 xi ] onto ;1 1]: Hint: x(t) = (d ; c)t=2 +
(d + c)=2 transforms ;1 1] onto c d]:
R

5.10.13. Test the class integral dened in x5.5 to see if repeated invocations
to function trapezoidal() with the default number of subintervals really do not cause recomputation of the denite integral. The member
value of integral is dened to be mutable so that it can be modied
by const function trapezoidal(). Alternatively, value may also be dened to be a regular (instead of mutable) member and trapezoidal()
a nonconstant member function. For what kind of classes does the
alternative way seem better?
5.10.14. Dene an n by n lower triangular matrix as a double pointer (or
say an array of n pointers) to Vec see x5.7 for the denition of Vec.
Each of the n pointers can be allocated space for a Vec of certain
number of elements representing a row of the matrix. Assign value
1=(i + j +1:0) to the entry at row i and column j for i = 0 1 : : : n ; 1
and j = 0 1 : : : i: Finally, deallocate space for the matrix.
5.10.15. Dene a type that is a pointer to a member of class integral as discussed in x5.1. Use pointers to members to call function trapezoidal()
and evaluate a few denite integrals this way.
5.10.16. Implement the implicit Euler's method and the fourth-order Runge{
Kutta method for the class ode discussed in x5.9. Test your implementation on the example (5.3)-(5.4).
5.10.17. Compare the accuracy of explicit, implicit, and predictor-corrector
Euler's methods, and the second- and fourth-order Runge{Kutta methods for the initial value problem: dx=dt = x 0 < t < 2 x(0) = 1
with di
erent values of = ;10 ;50 ;100 ;500 and di
erent grid
sizes h = 1=50 1=100 1=200 1=400 respectively. Its exact solution is
x(t) = et : The accuracy should decrease when gets smaller (negatively).
5.10.18. Suppose that Ms. Li deposits $10000 in an investment account that
pays an annual interest rate of 5% compounded continuously. She
then withdraws $1000 from the account each year in a continuous
way starting in year 10. Then A(t) the amount of money in the
account in year t, satises the di
erential equation

dA =
dt

0:05A
for t < 10
0:05A ; 1000 for t 10:

202

5. Classes

How long will the money last? That is, at what time t, will the account
balance A(t) = 0? (Answer: t = 44.79 years.)
To maintain a certain life standard, she decides to adjust the withdrawal amount to ination and withdraws $1000e0:03(t;10) in year
t for t 10 (assuming an annual ination rate of 3% compounded
continuously) then the di
erential equation becomes:

dA =
dt

0:05A
for t < 10
0:05A ; 1000e0:03(t;10) for t 10:

How long will the money last in this case? (Answer: 30.00 years.)

This is page 203

Printer: Opaque this

Operator Overloading

Suppose that v1 v2 and v3 are three vectors and m is a matrix. Mathematically we can write

v3 = v1 + v2
v2 = v1 + m v3
provided the dimensions of the matrix and vectors are compatible. In C++,
we can dene classes for vectors and matrices and redene the meanings of
+ and = such that vector addition and matrix-vector multiplication can
be written exactly in the same way as the mathematical expressions above.
This kind of extension of operators such as + and = from built-in types
to user-dened types is called operator overloading. Operator overloading
enables the C++ code of many mathematical methods to resemble their algorithms, which can make programming in C++ easier and C++ programs
more readable. In this chapter, various issues on operator overloading are
discussed. Examples are complex numbers, vectors, and matrices, which
are building blocks of many scientic programs. Note that standard C++
libraries include complex numbers (x7.4) and vectors (x7.5 and x10.1.1). A
simpler and easier-to-understand version is presented here to illustrate how
operators are overloaded. A deferred-evaluation technique is also presented
to improve the eciency of overloaded composite operators. In the nal
section, operator overloading is applied to solve systems of linear equations
using the conjugate gradient method.

204

6. Operator Overloading

6.1 Complex Numbers

The name of an operator function is the keyword operator followed by
the operator itself for example, operator+ operator, and operator =. In
this section, a class called Cmpx for complex numbers is dened and some
basic operator functions are provided, some of which are class members and
friends and others are just ordinary functions. The purpose is only to show
the idea of operator overloading and a user should use the standard library
<complex> (see x7.4) for computations involving complex numbers. It is
declared as
class Cmpx {
// class for complex numbers
private:
double re
// real part of complex number
double im
// imaginary part
public:
Cmpx(double x=0, double y=0) {re=x im=y}
// constructor
Cmpx& operator+=(Cmpx)
// operator +=, z1 += z2
Cmpx& operator-=(Cmpx)
// operator -=, z1 -= z2
friend Cmpx operator*(Cmpx, Cmpx) // binary *, z = z1 * z2
friend ostream& operator<<(ostream&, Cmpx) // cout << z
friend istream& operator>>(istream&, Cmpx&) // cin >> z
}
Cmpx
Cmpx
Cmpx
Cmpx

operator+(Cmpx)
operator-(Cmpx)
operator+(Cmpx, Cmpx)
operator-(Cmpx, Cmpx)

//
//
//
//

unary +, eg z1
unary -, eg z1
binary +, eg z
binary -, eg z

=
=
=
=

+ z2
- z2
z1 + z2
z1 - z2

Note that the operators += ;=, +, ;, , <<, and >> are overloaded,
among which + and ; are overloaded as both unary and binary operators.
The operators + and ; are declared as ordinary functions since they do
not need direct access to private members of class Cmpx: The types ostream and istream are dened in <iostream>. The members, friends, and
ordinary functions can be dened as (the order in which they are dened
does not matter)
inline Cmpx operator+(Cmpx z) {
return z
}

// unary +, eg z1 = + z2

inline Cmpx operator-(Cmpx z) {

return 0 - z
}

// unary -, eg z1 = - z2
// make use of binary // 0 converted to Cmpx(0)

inline Cmpx& Cmpx::operator+=(Cmpx z) { // add-assign, y += z

6.1 Complex Numbers

re += z.re im += z.im
return *this

205

// increment real and imaginary

}
inline Cmpx& Cmpx::operator-=(Cmpx z) { // eg y -= z
re -= z.re im -= z.im
// decrement real and imaginary
return *this
}
inline Cmpx operator+(Cmpx a, Cmpx b) { // binary +, z=z1+z2
return a += b
// make use of +=: a += b return a
}
inline Cmpx operator-(Cmpx a, Cmpx b) { // binary -, z=z1-z2
return a -= b
// equivalently: a -= b return a
}
inline Cmpx operator*(Cmpx a, Cmpx b) { // eg z = z1 * z2
return Cmpx(a.re*b.re - a.im*b.im, a.re*b.im + a.im*b.re)
}
ostream& operator<<(ostream& s, Cmpx z) {
s << "(" << z.re << ", " << z.im << ")"
return s
}

// output
// if z=Cmpx(2,5)
// output (2,5)

istream& operator>>(istream& s, Cmpx& z) { // pass reference

s >> z.re >> z.im
// input two numbers, written to z
return s
// if input 3 6, then z = Cmpx(3,6)
}

With the denition for Cmpx and the overloaded operators, we can now
write:
int main() {
Cmpx a(1,1)
//
Cmpx b = a
//
Cmpx c = a + b
//
c -= b
//
cout << c << '\n'
//
c = - b
//
cout << - c + a*b << '\n'
}

construction of an object
copy initialization
addition and copy initialization
subtract and assign
output the complex number c
unary operator and assignment

Note that operators + and ; do not directly manipulate the representation of an object but rely on operators += and ;= : This can be very useful

206

6. Operator Overloading

when dealing with user-dened types with a lot of data such as matrices
and vectors. The usual precedence rules still hold for overloaded operators.
For example, ;c + a b means ;c + (a b) instead of (;c + a) b: A few
more remarks are given below.

6.1.1 Initialization

The constructor with default arguments

class Cmpx {
public:
// ... in addition to other members
Cmpx(double x = 0, double y = 0) { re =x im =y }
}

is equivalent to the following three constructors

class Cmpx {
public:
// ... in addition to other members
Cmpx() { re = 0 im = 0 }
// default constructor
Cmpx(double x) { re = x im = 0 }
Cmpx(double x, double y) { re = x im = y }
}

Thus the statements

Cmpx a
Cmpx b(3)
Cmpx f = 3

construct a to be the complex number with real and imaginary parts zero
and construct b and f to be the complex number with real part 3 and
imaginary part 0. In the third statement above, 3 is rst constructed into
Cmpx(3) = Cmpx(3 0) and then copied to f by the compiler-generated
copy constructor.

6.1.2 Default Copy Construction and Assignment

The default compiler-generated copy constructor and assignment work just

ne for this simple class and are preferred over user-dened copy constructor and assignment. A default copy constructor and assignment simply copy
all members. The user could also dene a copy constructor and assignment:
class Cmpx {
public:
// ... in addition to other members
Cmpx(const Cmpx& c) { re = c.re im = c.im }
Cmpx& operator=(const Cmpx& c) {
re = c.re im = c.im

6.2 Operator Functions

207

return *this
}
}

However, this is unnecessary, error-prone, and usually less ecient than

the compiler-generated default copy constructor and assignment.

6.1.3 Conversions and Mixed-Mode Operations

If a and b are of Cmpx, then in a = 2+ b the operand 2 is rst converted to

Cmpx(2) by the constructor and then added to b. To avoid such conversion
and increase eciency, mixed-mode operations may be dened:
class Cmpx {
public:
// ... in addition to other members
Cmpx& operator+=(double a) {
re += a
// only real part is changed
return *this
// imaginary not affected here
}
}
inline Cmpx operator+(Cmpx a, double b) {
return a += b
// use Cmpx::operator+=(double)
}
inline Cmpx operator+(double a, Cmpx b) {
return b += a
// use Cmpx::operator+=(double)
}

Now adding a double to a Cmpx does not touch the imaginary part of the
complex number and thus is simpler and more ecient than adding two

Cmpx:

6.2 Operator Functions

The following operators may be overloaded by a user.
+

j
;=

<<
>=
;>

>>
&&
]

*
!

=
= = %=
>>= <<=
jj
++
()
new

^=
==

&=
!=

;;

new ]

delete

&
+=
j=
<=
,
delete ]

The following operators can not be overloaded by a user.

208

6. Operator Overloading

(scope resolution operator)

: (member selection operator)

: (member selection through pointer to function)
A binary operator can be dened by either a nonstatic member function
taking one argument or a nonmember function (a friend or an ordinary
function) taking two arguments. If both are dened, overloading resolution determines which one to use or if there is an ambiguity. For any binary operator @, aa@bb can be interpreted as either aa:operator@bb or
operator@(aa bb): The binary operators + and ; are dened as nonmember functions for Cmpx in x6.1. They could also be alternatively dened
as member functions:
class Cmpx {
public:
// ... in addition to other members
Cmpx operator+(Cmpx)
// binary plus as a member
Cmpx operator-(Cmpx)
// binary minus as a member
}
inline Cmpx Cmpx::operator+(Cmpx bb) {
return bb += *this
// no temporary object needed
}
inline Cmpx Cmpx::operator-(Cmpx bb) {
Cmpx temp = *this
// a temporary object is used
return temp -= bb
// temp -= bb return temp
}

When they are dened as class members, aa + bb means aa:operator+(bb)

and aa + bb is only a shorthand notation for the explicit function call
aa:operator+(bb): Thus if aa is of type Cmpx, aa +1 means aa + Cmpx(1)
which in turn is aa:operator+(Cmpx(1)): But 1+ aa will not be legal, since
it means 1:operator +(aa): Here 1 is not an object of class Cmpx and the
conversion Cmpx(1) is not tried to make a Cmpx out of 1 since otherwise
it would be possible to change the meaning of operations on built-in types
such as 1+1: Member functions can be invoked for objects of their class and
no user-dened conversions are applied. Similarly, the statement 1 += aa
is illegal. As a general rule, an operator function intended to accept a
basic type as its rst operand can not be a member function. Thus binary
operators are most commonly dened as friends or ordinary functions. On
the other hand, operators that require lvalue operands such as = = and
++, are most naturally dened as members.
Note that a temporary object temp is rst created by copy construction
and then returned (after being decremented by bb) in the denition of
the binary minus operator when dened as a member function, while no
temporary object is created in the operator ;=: Thus a ;= b is not only

6.2 Operator Functions

209

a shorthand notation, but also more ecient than a = a ; b: Attempting

to avoid such a temporary object in dening the member operator; may
easily result in an error:
Cmpx Cmpx::operator-(Cmpx bb) { // WRONG !
return *this -= bb
// this changes the object
}
Cmpx a(9, 5.7)
Cmpx b(3)
Cmpx c = a - b

// a also equals Cmpx(6, 5.7)

This denition not only assigns a ; b to c but also changes the object a to
a ; b in a statement c = a ; b since a ; b simply means a:operator ; (b)
and this ;= bb in the denition of ; as a member function above also
changes the underlying object. A temporary object may need be created
even for member operator + when passing by reference is used for large
objects such as vectors and matrices see Exercise 6.7.2 and x6.3.
A unary operator, whether prex or postx, can be dened by either
a nonstatic member function taking no arguments or a nonmember function taking one argument. For any prex unary operator @, @aa can be
interpreted as either aa:operator@() or operator@(aa): For any postx
unary operator @, aa@ can be interpreted as either aa:operator@(int) or
operator@(aa int) where the int argument is used only to indicate that
the operator is postx, distinguishing it from prex. The unary operators
+ and ; are dened as nonmember functions for Cmpx in x6.1. They can
also be dened alternatively as member functions:
class Cmpx {
public:
Cmpx operator+()
Cmpx operator-()
}

// ... in addition to other members

inline Cmpx Cmpx::operator+() {

return *this
}

// unary + as member, y=+z

inline Cmpx Cmpx::operator-() {

return Cmpx(- re, - im)
}

// unary - as member, y=-z

Prex and postx operators ++ are now dened for Cmpx and their use
are illustrated:
class Cmpx {
public:

// ... in addition to other members

210

6. Operator Overloading

Cmpx operator++()
Cmpx operator++(int)
}

// prefix
// postfix

inline Cmpx Cmpx::operator++() { // prefix as in: ++z

re++ im++
// increment both real and imaginary
return *this
}
inline Cmpx Cmpx::operator++(int i) {
Cmpx temp = *this
re++ im++
return temp
}
Cmpx aa = 5.0
Cmpx bb = ++aa
Cmpx cc = aa++

// postfix as in: z++

// bb = Cmpx(6,1), aa = Cmpx(6,1)
// cc = Cmpx(6,1), aa = Cmpx(7,2)

A user can also dene ++ to increment only the real part of a complex number, which reects the exibility of operator overloading on user-dened
types. If a is an integer, ++a means a += 1 which in turn means a = a +1:
For user-dened operators, this is not true unless they are dened this way.
For example, a compiler will not generate a denition of Z :: operator+=()
from the denitions of Z :: operator +() and Z :: operator =():
The operators = ] () and ;> must be dened as nonstatic member
functions to ensure that their rst operand is an lvalue. The subscripting
operator ] is used in x6.3, while the function call operator () is used in
x6.3, x7.6, x10.2.1, and x10.2.2. The dereferencing operator ;> can be used
to create \smart pointers" and is discussed in Str97, LL98].
The operators = (assignment), & (address-of), and , (sequencing) have
predened meanings when applied to class objects. They can be made inaccessible to general users by making them private:
class X {
private:
void operator=(const X&)
void operator&()
void operator,(const X&)
}
void f(X a, X b) {
a = b
&a
a, b
}

// error, operator = is private

// error, operator & is private
// error, operator , is private

6.3 Vectors and Matrices

211

This is sometimes necessary to reduce possible errors or to improve eciency (see x7.5 for an example).
When an operator is overloaded for many operations with the same base
type (e.g., + can be overloaded to add two complex numbers, one complex
number and one double, etc.) or di
erent types (e.g., + can add complex
numbers, vectors, and matrices), overloading resolution (see x3.8.2, x6.4,
and x7.2.2) then comes in to determine which one to use or if there are
ambiguities. When the denition of an operator can not be found in the
scope in which it is used, then the namespaces of its arguments will be
looked up (see x4.1.5).

6.3 Vectors and Matrices

In this section, a matrix class called Mtx a vector class called V tr, and
their associated operations such as matrix-vector multiply are dened. The
declaration of V tr is:
class Vtr {
int lenth
double* ets
public:
Vtr(int, double*)
Vtr(int = 0, double d = 0)
Vtr(const Vtr&)
~Vtr(){ delete] ets }

// number of entries
// entries of the vector
//
//
//
//

constructor
all entries equal d
copy constructor
destructor

int size() const { return lenth } // return length

Vtr& operator=(const Vtr&)
// overload =
Vtr& operator+=(const Vtr &)
// v += v2
Vtr& operator-=(const Vtr &)
// v -= v2
double maxnorm() const
// maximum norm
double twonorm() const
// 2-norm
double& operator](int i) const { return etsi] }
// subscript, eg v3] = 1.2
friend Vtr operator+(const Vtr&) // unary +, v = + v2
friend Vtr operator-(const Vtr&) // unary -, v = - v2
friend Vtr operator+(const Vtr&, const Vtr&)
// binary +, v = v1 + v2
friend Vtr operator-(const Vtr&, const Vtr&)
// binary -, v = v1 - v2
friend Vtr operator*(double, const Vtr&)
// vec-scalar multiply
friend Vtr operator*(const Vtr&, double)

212

6. Operator Overloading

// vec-scalar multiply
friend Vtr operator/(const Vtr&, double)
// vec-scalar divide
friend Vtr operator*(const Vtr&, const Vtr&)
// vector multiply
friend double dot(const Vtr&, const Vtr&)
// dot (inner) product
friend ostream& operator<<(ostream&, const Vtr&)
// output operator
}

For two vectors v = v0 v1 : : : P

vn;1 ] and w = w0 w1 : : : wn;1 ] their
;1 vi wi (here wi means the complex
dot product is a scalar v w = ni=0
conjugate of wi ) and their vector product is a vector v w = v0 w0 v1 w1
: : : vn;1 wn;1 ]:
The default copy constructor and assignment will not work correctly in
this case, since space need be allocated using new in the construction of
an object of class Vtr and deallocated in its destructor. The operator ] is
overloaded and returns a reference here so that the elements of a vector
can be referred to more conveniently. In particular, if v is of type Vtr, then
its ith element can be referred to as vi], instead of v:etsi] and assignment
to vi] is also possible:
double a = vi]
vi] = 10

// better notation than v.etsi]

// operator] returns a reference

Here are the denitions of the members and friends of class Vtr:
inline void error(char* v) {
cout << v << ". program exited\n"
exit(1)
}

// auxiliary fcn
// include <iostream>
// include <stdlib.h>

Vtr::Vtr(int n, double* abd) {

// constructor
ets = new double lenth = n]
for (int i = 0 i < lenth i++) etsi]= *(abd +i)
}
Vtr::Vtr(int n, double a) {
// constructor
ets = new double lenth = n]
for (int i = 0 i < lenth i++) etsi] = a
}
Vtr::Vtr(const Vtr & v) {
// copy constructor
ets = new double lenth = v.lenth]
for (int i = 0 i < lenth i++) etsi] = vi]
}

6.3 Vectors and Matrices

213

Vtr& Vtr::operator=(const Vtr& v) {

// overload =
if (this != &v) {
// beware of self-assignment
if (lenth != v.lenth ) error("bad vector sizes")
for (int i = 0 i < lenth i++) etsi] = vi]
}
return *this
}
Vtr& Vtr::operator+=(const Vtr& v) { // add-assign, u += v
if (lenth != v.lenth ) error("bad vector sizes")
for (int i = 0 i < lenth i++) etsi] += vi]
return *this
}
Vtr& Vtr::operator-=(const Vtr& v) { // subtract-assign
if (lenth != v.lenth ) error("bad vector sizes")
for (int i = 0 i < lenth i++) etsi] -= vi]
return *this
}
inline Vtr operator+(const Vtr& v) {
return v
}

// usage: u = + v
// unary +

inline Vtr operator-(const Vtr& v) {

return Vtr(v.lenth) - v
}

// usage: u = - v
// unary -

Vtr operator+(const Vtr& v1, const Vtr& v2) {

// addition
if (v1.lenth != v2.lenth ) error("bad vector sizes")
Vtr sum = v1 // would cause error without copy constructor
sum += v2
// eg, v1 would be changed in v =v1 +v2
return sum
// since sum.ets would equal v1.ets
}
Vtr operator-(const Vtr& v1, const Vtr& v2) {
// subtract
if (v1.lenth != v2.lenth ) error("bad vector sizes")
Vtr sum = v1
// create a temporary object
sum -= v2
return sum
// members of sum returned
}
// sum is then destroyed
Vtr operator*(double scalar, const Vtr& v) {
Vtr tm(v.lenth)
// scalar-vector multiply

214

6. Operator Overloading

for (int i = 0 i < v.lenth i++) tmi] = scalar*vi]

return tm
}
inline Vtr operator*(const Vtr& v, double scalar) {
return scalar*v
// scalar-vector multiply
}
Vtr operator*(const Vtr& v1, const Vtr& v2) { // multiply
if (v1.lenth != v2.lenth ) error("bad vector sizes")
int n = v1.lenth
Vtr tm(n)
for (int i = 0 i < n i++) tmi] = v1i]*v2i]
return tm
}
Vtr operator/(const Vtr & v, double scalar) {
if (!scalar)
// vector scalar divide
error("division by zero in vector-scalar division")
return (1.0/scalar)*v
}
double Vtr::twonorm() const {
// two (euclidean) norm
double norm = ets0]*ets0]
for (int i = 1 i < lenth i++) norm += etsi]*etsi]
return sqrt(norm)
}
double Vtr::maxnorm() const {
// maximum norm
double norm = fabs(ets0])
for (int i = 1 i < lenth i++)
norm = max(norm, fabs(etsi])) // max() in <algorithm>
return norm
}
double dot(const Vtr &v1, const Vtr& v2) { // dot product
if (v1.lenth != v2.lenth ) error("bad vector sizes")
double tm = v10]*v20]
for (int i = 1 i < v1.lenth i++) tm += v1i]*v2i]
return tm
}
ostream& operator<<(ostream& s, const Vtr& v ) { // output
for (int i =0 i < v.lenth i++ ) {
s << vi] << " "

6.3 Vectors and Matrices

if (i%10 == 9) s << "\n"
}
return s

215

// print 10 elmts on a line

Note that without the user-dened copy constructor, erroneous results

would occur in statements such as v = v1 + v2 which would cause the
summation v1 + v2 to be assigned to both v1 and v: The reason is that in
the denition of the binary operator + above, the variable sum is copyconstructed from v1 and thus sum:ets would equal v1:ets with a compilergenerated copy constructor. Thus any update to sum would occur to entries
of v1 as well. Compared to class Cmpx in x6.1, temporary objects are created for V tr in the denitions of member functions such as the binary
operator +. This may make operation v1 += v2 noticeably more ecient
than a seemingly equivalent one: v1 = v1+ v2: The code is also written such
that the program will exit if two vectors of di
erent sizes are accidentally
added or subtracted.
Now vector operations in C++ can be written in a similar way to conventional mathematical notation. For example,
void f() {
int m = 500
double* v = new double m]
for (int i = 0 i < m i++ ) vi] = i*i + 10
Vtr v1(m, v)
// create v1 from v
Vtr v2(m)
// zero vector of size m
for (int i = 0 i < m i++ ) v2i] = 5*i - 384
Vtr v3(m, 5.8)
// every entry equals 5.8
Vtr v4 = - v1 + 3.3*v3
v4 += v2
cout << v4

// resemble mathematics

Vtr v5 = - v1*v4
double a = dot(v1, v5)

// vector multiply
// dot product

// output vector

Below is a class Mtx for matrices, in which some operations such as

matrix addition and matrix-vector multiplication are dened. To compare
the di
erence in dening member functions and friends, matrix addition
is implemented as a member function and matrix subtraction as a friend.
This implementation detail should not have much e
ect on how the Mtx
class is going to be used.
class Mtx {
private:
int nrows

// number of rows

216

6. Operator Overloading

int ncols
double** ets
public:
Mtx(int n, int m, double**)
Mtx(int n, int m, double d = 0)
Mtx(const Mtx &)
~Mtx()

// number of columns
// entries of matrix
//
//
//
//

constructor (n by m)
all entries equal d
copy constructor
destructor

Mtx& operator=(const Mtx&)

// overload =
Mtx& operator+=(const Mtx&)
// overload +=
Mtx& operator-=(const Mtx&)
// overload -=
Vtr operator*(const Vtr&) const
// matrix vec multiply
double* operator](int i) const { return etsi] }
// subscript, entry at row i and column j is i]j]
double& operator()(int i, int j) { return etsi]j] }
// subscript, entry at row i and column j is (i,j)
// Implement plus as a member fcn. minus could also be
// so implemented. But I choose to implement minus as
// a friend just to compare the difference
Mtx& operator+() const
// unary +, m1 = + m2
Mtx operator+(const Mtx&) const
// binary +, m = m1+m2
friend Mtx operator-(const Mtx&) // unary -, m1 = - m2
friend Mtx operator-(const Mtx&,const Mtx&) // binary }

6.3 Vectors and Matrices

217

Mtx::Mtx(const Mtx & mat) {

// copy constructor
ets = new double* nrows = mat.nrows]
ncols = mat.ncols
for (int i = 0 i< nrows i++) {
etsi] = new double ncols]
for (int j = 0 j < ncols j++) etsi]j] = mati]j]
}
}
inline Mtx::~Mtx(){
// destructor
for (int i = 0 i< nrows i++) delete] etsi]
delete] ets
}
Mtx& Mtx::operator=(const Mtx& mat) {
// copy assignment
if (this != &mat) {
if (nrows != mat.nrows || ncols != mat.ncols)
error("bad matrix sizes")
for (int i = 0 i < nrows i++)
for (int j = 0 j < ncols j++)
etsi]j] = mati]j]
}
return *this
}
Mtx& Mtx::operator+=(const Mtx& mat) { // add-assign
if (nrows != mat.nrows || ncols != mat.ncols)
error("bad matrix sizes")
for (int i = 0 i < nrows i++)
for (int j = 0 j < ncols j++)
etsi]j] += mati]j]
return *this
}
Mtx& Mtx::operator-=(const Mtx& mat) { // subtract-assign
if (nrows != mat.nrows || ncols != mat.ncols)
error("bad matrix sizes")
for (int i = 0 i < nrows i++)
for (int j = 0 j < ncols j++)
etsi]j] -= mati]j]
return *this
}
inline Mtx& Mtx::operator+() const {

// usage: m1 = + m2

218

6. Operator Overloading

return *this
}
inline Mtx operator-(const Mtx & mat) {
return Mtx(mat.nrows, mat.ncols) - mat
}

// m1 = - m2

Mtx Mtx::operator+(const Mtx & mat) const { // m = m1 + m2

if (nrows != mat.nrows || ncols != mat.ncols)
error("bad matrix sizes")
Mtx sum = *this
// user-defined copy constructor
sum += mat
// is important here
return sum
// otherwise m1 would be changed
}
Mtx operator-(const Mtx& m1, const Mtx& m2) { // m=m1-m2
if(m1.nrows != m2.nrows || m1.ncols != m2.ncols)
error("bad matrix sizes")
Mtx sum = m1
sum -= m2
return sum
}
Vtr Mtx::operator*(const Vtr& v) const { // u = m*v
if (ncols != v.size())
error("matrix and vector sizes do not match")
Vtr tm(nrows)
for (int i = 0 i < nrows i++)
for (int j = 0 j < ncols j++)
tmi] += etsi]j]*vj]
return tm
}

Then the following code can be conveniently written.

int main() {
int k = 300
double** mt = new
for (int i = 0 i
for (int i = 0 i
for (int j = 0
Mtx
Mtx
Mtx
for

double* k]
< k i++) mti] = new double k]
< k i++)
j < k j++) mti]j] = 2*i*j + i + 10

m1(k, k, mt)
// construct m1 from mt
m2(k, k, 5)
// construct m2, all entries = 5
m3(k, k)
// construct m3, all entries = 0
(int i = 0 i < k i++)
// update entries of m3

6.3 Vectors and Matrices

219

for (int j = 0 j < k j++) m3i]j] = 1/(2*i + j + 5.7)

m3 += - m1 + m2
m1 -= m3

// resemble mathematics
// very readable

Vtr vv(k)
for (int i = 0 i < k i++ ) vvi] = 5*i + 3
vv = m3*vv
// resemble mathematics
}

Since the function call operator () is overloaded, a matrix entry can also
be accessed using FORTRAN notation such as m3(5 7) = 10:
Note that pass by reference has been used when possible in the denition
of member functions and friends of classes Mtx and V tr: This is more
ecient when dealing with large objects than pass by value. In the latter
case operands are copied according to the copy constructor. Pass by value
is used in Cmpx, where only two doubles of a complex number are copied.
Since matrices and vectors often have many entries, copying their members
would impose signicant overhead. For example, if the binary plus operator
is dened through pass by value:
Mtx Mtx::operator+(Mtx mat) {
// usage: m1 + m2
Mtx sum = *this
// user-defined copy constructor
sum += mat
// is important here
return sum
// otherwise m1 would be changed
}

then in the statement m = m1 + m2 (or equivalently m = m1:operator +

(m2)) members of the argument m2 will be copied to the parameter mat
by calling the copy constructor. The return value sum is also copied to
the calling environment (as a temporary object, which is then assigned to
variable m by the copy assignment operator this temporary object can
often be optimized away). Using pass by reference just creates an alias,
instead of copying all the entries of a matrix. It would be more ecient if
the return value could be passed by reference. However, pass by reference
for a local variable (such as sum in the denition of the binary + above)
can not be used (see x3.8.4).
Pointers can not be used because it is not possible to redene the meaning
of an operator acting on a pointer or any other built-in type.
The vector class V tr (especially its template form in Chapter 7) may be
used to replace one-dimensional built-in arrays (see x3.3) since it provides
dynamic memory allocation and does not require the user to allocate and
deallocate space explicitly. For example,
double g(int n) {
Vtr a(n)
// allocate dynamic memory
for (int i = 0 i < n i++)

220

6. Operator Overloading

ai] = std::sin(i*i)
return a.twonorm()

// can be used like an array

// space freed automatically

Besides, it provides basic vector operations and the user can easily extend
it to meet specic needs. However, it may not be as ecient as the standard library <vector> (see x10.1.1) and <valarray> (see x7.5), which also
provide many more operations than Vtr.
Since C++ standard libraries do not include matrices, the matrix class
Mtx (especially its template form in Chapter 7) may be used in place of
two-dimensional built-in arrays (see x3.3). For example,
double h(int n, int m) {
Mtx mx(n, m)
for (int i = 0 i < n i++)
for (int j = 0 j < m j++)
mxi]j] = 1.2/(i + j + 1)
return mx.maxnorm()
}

// allocate dynamic memory

// for an n by m array
// can be used like array
// space freed implicitly

From a user's point of view, no double pointers and memory management

need be manipulated directly.

6.4 Explicit and Implicit Conversions

A constructor with one argument also denes a conversion. For example,
class Cmpx {
public:
// ... in addition to other members
Cmpx(double a)
// constructor from double to complex
Cmpx(const Cmpx &) // copy constructor
}
void f(Cmpx)
Cmpx z = 2.2
f(5.5)

// initialize z to Cmpx(2.2)
// call fcn f with argument Cmpx(5.5)

The constructor rst implicitly converts 2:2 into Cmpx(2:2) and then initializes z with the complex number Cmpx(2:2): This is called implicit conversion. For some types, this conversion may not be desirable. For example,
conversion from int to an enumeration type may cause truncation. Implicit
conversion can be suppressed by declaring a constructor to be explicit.
With an explicit constructor, conversion can not be done implicitly. For
example,
class Cmpx {
public:

// ... in addition to other members

6.4 Explicit and Implicit Conversions

explicit Cmpx(double a)
Cmpx(const Cmpx &)
}
void f(Cmpx)
Cmpx y = 2.2
Cmpx z = Cmpx(2)
f(5.5)
f(Cmpx(5.5))
f(z)

//
//
//
//
//

221

// explicit constructor
// copy constructor

error, implicit conversion disallowed

OK, explicit conversion
error, can not convert 5.5 to Cmpx(5.5)
OK
OK

Conversion operators can also be dened to convert one type to another.

If X is a class and T is a type, then the member function X :: operator T ()
denes a conversion from X to T: For example, one can dene a class for
4-bit nonnegative integers, called tiny, that can be assigned to and from
integers:
class tiny {
char v
void assign (int i) {
// for range checking
if (i & ~0xf) { cout << "range error\n" exit (1) }
v = i
}
public:
tiny(int i) { assign(i) }
// constructor
tiny& operator=(int i) { assign(i) return *this }
operator int() const { return v } // conversion operator
}

Note that 0xf is a hexadecimal number and (f )16 = (1111)2 = (15)10

where (n)b represents number n in base b: The bitwise expression i &
e 0xf yields a value having a bit representation with the four low-order
bits zero and all other bits the same as the high-order bits of i: The hex
number ~0xf is called a mask for high-order bits beyond the four loworder bits, and 0xf a mask for the four low-order bits. Range checking is
made when a tiny is initialized by an int and when an int is assigned to a
tiny: No range checking is needed when copying a tiny so that the default
copy constructor and assignment work just ne for this simple class. Now
implicit conversion from tiny to int is handled by the conversion operator
tiny :: operator int(): Note that the type being returned is part of the
name of the conversion operator and can not be repeated in the function:
class tiny {
public:
// ... in addition to other members
operator int() const { return v }
// correct
}

222

6. Operator Overloading

class tiny {
public:
// ... in addition to other members
int operator int() const { return v } // wrong !!!
}
// no return type needed

Now variables of tiny and int can be converted to each other freely:
tiny t = 12
int i = t
cout << int(t)
tiny tt = 16

//
//
//
//

int is assigned to tiny

tiny is converted to int
convert t to int and print it out
range error occurs, program exits.

Note that ambiguity can arise with user-dened conversions and userdened operators. For example,
tiny operator+(tiny t, tiny s) {
return tiny(int(t) + int(s))
}

// overload + for tiny

void f(tiny t, int i) {

t + i // error, ambiguous: int(t) + i or t + tiny(i) ?
}

The user should replace t + i by int(t) + i or t + tiny(i) according to what

is intended.
In some cases, a value of the desired type can be constructed by repeated
use of constructors or conversion operators. This must be done by explicit
conversion. Only one level of user-dened implicit conversion is legal. Userdened conversions are considered only if they are necessary to resolve a
call. For example,
struct X { X(int) X(double) }
struct Y { Y(int) Y(double) }
struct Z { Z(X) }
X f(X)
Y f(Y)
Z g(Z)
X h(X)
X h(double)
f(5)
f(X(5))
f(Y(3.14))
f(3.14)
g(5.6)
h(5)

//
//
//
//
//
//

// X has two constructors

// Y has two constructors
// Z has one constructor

error, ambiguous: f(X(5)) or f(Y(5)) ?

OK
OK
error
error, g(Z(X(5.6))) is not tried.
OK, h(double(5)) is used, h(X(5)) not used

In the call g(5:6) two user-dened conversions g(Z (X (5:6))) would be

needed and thus are not tried. Only one level of user-dened implicit con-

6.5 Eciency and Operator Overloading

223

version is legal. In the call h(5), the standard conversion h(double(5)) for
the argument from int to double is preferred over the user-dened conversion h(X (5)):

6.5 Eciency and Operator Overloading

When dealing intensively with large objects such as matrices and vectors,
eciency may be a predominating factor to be considered in many scientic
computing applications. It was mentioned in x6.3 that creating temporary
objects or copying large objects could slow down a program.
In addition, multiple loops in composite operations can also a
ect efciency. For example, in the so-called vector saxpy (mnemonic for scalar
a x plus y) operation z = a x + y for vectors x y z and scalar a, one
loop is needed for vector-scalar multiplication a x another loop for vector
addition a x + y and yet another for copying the result to vector z when
using the vector class V tr dened in x6.3. Also temporary objects may
need be created for holding a x and a x + y in this process (besides the
construction and destruction of temporary objects for automatic variables
in the denitions of operators and + for V tr). That is, a few loops are
needed for a simple saxpy operation. A much more ecient and traditional
way for dening the saxpy operation is to write only one loop without any
temporary objects, as is normally done in FORTRAN 77 and C:
inline void saxpy(double a,const Vtr& x,const Vtr& y,Vtr& z){
for (int i = 0 i < z.size() i++) zi] = a*xi] + yi]
}

But this would sacrice readability and defeat the purpose of operator
overloading. That is, instead of writing
z = a*x + y

one would have to write a less expressive statement

saxpy(a, x, y, z)

in order to improve eciency.

How can one achieve the same eciency as in the traditional way while
still using operator overloading?
Below is a technique to achieve this. The idea is to defer the evaluations
of intermediate suboperations in a composite operation until the end of the
operation. First, dene a class called Sax for scalar-vector multiplication,
which does not do the evaluation but simply constructs an object holding
the references to the scalar and vector to be multiplied:
struct Sax {
const double& a

// reference is used to avoid copying

224

6. Operator Overloading

const Vtr& x
Sax(const double& d, const Vtr& v) : a(d), x(v) { }
}
inline Sax operator*(const double& d, const Vtr& v) {
return Sax(d,v)
// overload operator *
}

Then dene a class called Saxpy for vector saxpy operations (Again, to
avoid creating temporary objects, evaluations of suboperations are deferred).
struct Saxpy {
const double& a
// reference is used to avoid copying
const Vtr& x
const Vtr& y
Saxpy(const Sax& s,const Vtr& u) : a(s.a),x(s.x),y(u) { }
}
inline Saxpy operator+(const Sax& s, const Vtr& u) { //a*x+y
return Saxpy(s, u)
// overload +
}
inline Saxpy operator+(const Vtr& u, const Sax& s) { //x+a*y
return Saxpy(s, u)
// overload +
}

Finally, dene a copy constructor and a copy assignment from Saxpy to

V tr in which the evaluation of the saxpy operation is done:
class Vtr {
int lenth
double* ets
public:
Vtr(int, double*)
Vtr(int, double = 0)
Vtr(const Vtr&)
Vtr(const Saxpy&)
~Vtr(){ delete] ets }

// number of entries
// entries of vector
//
//
//
//
//

constructor
constructor
copy constructor
constructor from Saxpy
destructor

Vtr& operator=(const Vtr&)

// overload assignment
Vtr& operator=(const Saxpy&)
// overload assignment
int size() const { return lenth }
double& operator](int i) const { return etsi] }
}
Vtr::Vtr(const Saxpy& sp) {

6.5 Eciency and Operator Overloading

225

ets = new double lenth = sp.x.size()]

for (int i = 0 i < lenth i++)
etsi] = sp.a*sp.xi] + sp.yi]
}
Vtr& Vtr::operator=(const Saxpy& sp) {
for (int i = 0 i < lenth i++)
etsi] = sp.a*sp.xi] + sp.yi]
return *this
}

With the new denition of class V tr the usual scalar-vector multiplication
operator (see x6.3) should not be dened, since otherwise ambiguity would
arise. Notice that V tr is used when dening Saxpy and Saxpy is used
when dening V tr. A forward declaration is actually needed to resolve this
chaining. With a forward declaration for V tr, the name of class V tr can be
used in Sax and Saxpy before V tr is dened, as long as its use does not
require the size of V tr or its members to be known.
Now a saxpy operation can be evaluated without creating any temporary
V tr objects and with only one loop:
int main() {
int k = 50000
double* p = new double k]
for (int i = 0 i < k i++ ) pi] = i
Vtr v(k, p)
// create a Vtr object using p
Vtr w(k, 5)
// create another Vtr object
Vtr u = 5*v + w
v = u + 3.14*w

// efficient saxpy operation

In fact, a saxpy operation z = a x + y is now equivalent to

z:operator=(Saxpy(Sax(a x) y)):

Due to inlining for operators and + above, this should be as ecient as
the traditional saxpy() function. This deferred evaluation technique can be
generalized into what are called expression templates see x7.7.3.
Run-time comparisons can be made for the three versions of the saxpy
operation based on the straightforward operator overloading as dened in
x6.3, the deferred-evaluation operator overloading as dened in this section,
and the traditional procedural (C or FORTRAN) style dened as function
saxpy() early in this section. For vectors of size 50000, saxpy operations are
done repeatedly 1000 times. The three versions take 152 57 and 72 seconds,
respectively, on a Pentium PC (100 MHz) running Linux with a GNU C++
compiler. The same operations take 104 25 and 25 seconds, respectively,

226

6. Operator Overloading

on a SUN Ultra Workstation (167 MHz) running Solaris with a GNU C++
compiler. They take 67 14 and 19 seconds, respectively, on a Pentium
II PC (400 MHz) running Linux with a GNU C++ compiler. On an SGI
machine running UNIX, they take 78, 18, and 26 seconds, respectively. On
these four di
erent machines, the straightforward operator overloading is
three to ve times as slow as the deferred-evaluation operator overloading.
The latter is even faster than the traditional procedural function call on
three of the four machines, which is hard to explain.

6.6 Conjugate Gradient Algorithm

In this section, the conjugate gradient algorithm is presented and coded in
C++ for solving systems of linear equations. The details of this algorithm
are given in Chapter 11.
Let h i denote the usual dot product in Rn and A = (aij ) a real n by
n symmetric and positive denite matrix that is,

aij = aji
for all i j = 0 1 : : : n ; 1
(6.1)
n
hA i > 0
for all 2 R and 6= 0:
(6.2)
Given an initial guess x0 , the conjugate gradient algorithm iteratively
constructs a sequence fxk g, which converges to the solution of the matrix
equation Ax = b: It can be stated as follows.
Algorithm 6.6.1 Taking an initial guess x0 2 Rn and setting r0 = b ;
Ax0 p0 = r0 construct the sequence xk , for k = 0 1 2 : : :
hrk rk i
k = hAp

(6.3)
k pk i
xk+1 = xk + k pk
(6.4)
rk+1 = rk ; k Apk
(6.5)
rk+1 i

k = hrkh+1
(6.6)
rk rk i
pk+1 = rk+1 +
k pk :
(6.7)
Equation (6.4) shows how a more accurate solution xk+1 is obtained from
an old iterate xk : The vectors pk are called search directions and the scalars
k search lengths. It can be shown easily that rk+1 = rk ; k Apk = b ;
Axk+1 : Thus rk+1 represents the residual corresponding to the approximate
solution xk+1 :
This algorithm can be implemented with only one matrix-vector multiplication and two vector dot products, two scalar-vector multiplications,
and one vector saxpy operation per iteration. The iterative process can
be stopped when a norm of the residual rk = b ; Axk is smaller than a

6.6 Conjugate Gradient Algorithm

227

prescribed small number or the total number of iterations has exceeded

a given number. The closer is the initial guess x0 to the exact solution, the
faster is the convergence. However, the initial guess x0 is often taken to the
zero vector when not much information is known about the exact solution.
It can be shown that, for Hermitian positive denite coecient matrix A
the iterates xk converge to the exact solution within n iterations no matter
what the initial guess x0 is, in the absence of roundo
errors (i.e., in exact
arithmetics).
Making use of the vector and matrix classes dened in previous sections,
it can be coded as
class Mtx {
public:
// ... in addition to other members
int CG(Vcr& x, const Vcr& b, double& eps, int& iter)
/* Conjugate gradient method for Ax = b.
it returns 0 for successful return and 1 for breakdowns
A: symmetric positive definite coefficient matrix
x: on entry: initial guess on return: approximate soln
b: right side vector
eps: on entry:
stopping criterion, epsilon
on return: absolute residual in two-norm
for approximate solution
iter: on entry: max number of iterations allowed
on return: actual number of iterations taken.
*/
}
int Mtx::CG(Vcr& x, const Vcr& b, double& eps, int& iter) {
if (nrows != b.size())
cout << "matrix and vector sizes do not match\n"
const int maxiter = iter
Vcr r = b - (*this)*x
// initial residual
Vcr p = r
// p: search direction
double zr = dot(r,r)
// inner prod of r and r
const double stp = eps*b.twonorm() // stopping criterion
if (r.twonorm() == 0) {
eps = 0.0
iter = 0
return 0
}

// if initial guess is true soln,

// return. Otherwise division by
// zero would occur.

for (iter = 0 iter < maxiter iter++) { // main loop

Vcr mp = (*this)*p
// matrix-vector multiply

228

6. Operator Overloading
double alpha = zr/dot(mp,p)
x += alpha*p
r -= alpha*mp
if (r.twonorm() <= stp) break
double zrold = zr
zr = dot(r,r)
p = r + (zr/zrold)*p

//
//
//
//

divisor=0 only if r=0

update iterative soln
update residual
stop if converged

// dot product
// zrold=0 only if r=0

}
eps = r.twonorm()
if (iter == maxiter) return 1
else return 0
}

// end CG()

With operator overloading, the main loop in the code resembles the algorithm very closely. Now the conjugate gradient algorithm can be tested
to solve a linear system with a 300 by 300 Hilbert coecient matrix (which
is symmetric and positive denite) and a known solution vector:
int main() {
int n = 300
Mtx a(n, n)
// n by n Hilbert matrix
for (int i = 0 i < n i++)
for (int j = 0 j < n j++) ai]j] = 1/(i + j + 1.0)
Vcr t(n)
// exact solution vector of size n
Vcr x(n)
// initial guess and solution vector
for (int i = 0 i < n i++)
// true solution
ti] = 1/(i + 3.14)
int iter = 300
double eps = 1.0e-9
int ret
if (ret
cout <<
cout <<
cout <<
<<

= a.CG(x, a*t, eps, iter)

// call CG algorithm
== 0) cout << "CG returned successfully\n"
iter << " iterations are used in CG method.\n"
"Residual = " << eps << ".\n"
"Two-norm of exact error vector = "
(x - t).twonorm() << '\n'

The output of this program is:

CG returned successfully
11 iterations are used in CG method.
Residual = 1.31454e-09.
Two-norm of exact error vector = 4.44765e-06

6.7 Exercises

229

The maximum number of iterations allowed can be taken to be the dimension of the matrix. In Chapter 11, this algorithm is used to solve matrix
equations with real and complex matrices in di
erent precisions and di
erent matrix storage formats.

6.7 Exercises
6.7.1. Test and run the class for complex numbers dened in x6.1. Add more
operations such as division, complex conjugate, complex equality operator == and inequality operator !=, prex and postx operators
;;, modulus, argument, and nth roots.
6.7.2. Modify the binary + and ; operator functions for Cmpx as dened
in x6.1 so that they have prototype:
Cmpx operator+(const Cmpx&, const Cmpx&)
Cmpx operator-(const Cmpx&, const Cmpx&)

6.7.3.

6.7.4.

6.7.5.

6.7.6.

The only di
erence is that pass by reference is now used in argument
passing. Dene these two functions and make sure your denitions
do not cause unwanted side e
ects.
Test and run the class for vectors dened in x6.3. Then comment out
the user-dened copy constructor and copy assignment and try to
add two vectors v = v1+ v2: What would happen to vector v1 in this
vector addition? Could any other vector operations also be a
ected
by depending on the compiler-generated copy constructor and copy
assignment?
Redene the class for vectors dened in x6.3 so that it provides range
checking for vector indexes. That is, if v is a vector, then accessing
vi] with i < 0 or i v:size() will give a run-time error and exit the
program.
Dene a new class called Cvec for vectors of complex numbers using
the simple denition in x6.1 and Exercise 6.7.1 for complex numbers.
Provide basic operations (similar to those in x6.3 for real vectors) for
complex vectors such as vector addition, vector multiplication, vectorscalar multiplication, and dot product. Note that the dot product
of two complex vectors includes complex conjugation of the second
vector.
Test and run the class for matrices dened in x6.3. Add more operations such as matrix-matrix multiplication, matrix-scalar multiplication, and one, maximum, and Frobenius matrix norms.

230

6. Operator Overloading

6.7.7. Implement the Gauss quadrature in Exercise 5.10.11 as the dot product of two vectors one vector represents the weights and the other
represents the function values at quadrature points. Evaluating denite integrals as dot products enables one to deal easily with integrals
with the integrand being the product or sum of several functions (see
x7.7).
6.7.8. Following the idea of Exercise 6.7.7, implement the Gauss{Lobatto
quadrature problem in Exercise 5.10.12 as evaluating dot products
on subintervals of a given interval a b].
6.7.9. Mimic the class tiny dened in x6.4 to dene a class for 6-bit nonnegative integers that can mix freely with integers in arithmetic operations. Provide range checking when assigning an int to a variable
of such a class and the conversion operator from it to int:
6.7.10. Test and run the deferred-evaluation operator overloading code for
vector saxpy operations as dened in x6.5. Compare the run-time on
your machine for the straightforward operator overloading as dened
in x6.3, the deferred-evaluation operator overloading as dened in
x6.5, and the traditional procedural-style function call saxpy () as
dened in x6.5.
6.7.11. Extend the vector class dened in x6.5 using deferred-evaluation operator overloading, so that it will include ecient vector addition and
scalar-vector multiplication (without any temporary vector objects
and with a minimum number of loops).
6.7.12. Dene a matrix class based on deferred-evaluation operator overloading as discussed in x6.5 to support ecient matrix-vector gaxpy (general A x plus y) operation z = A x + y where x y z are vectors
and A a matrix. The implementation should not use any temporary
vector objects and should apply only a minimum number of loops.
6.7.13. Apply the conjugate gradient algorithm to solve a system of linear
equations with a symmetric and positive denite coecient matrix,
for example, with a Hilbert matrix of dimension 500 and a given
right-hand side vector. Compute the two-norm of the residual of the
numeric solution to check its accuracy.

This is page 231

Printer: Opaque this

Templates

C++ provides a mechanism, called template, that allows a type to be

a parameter in the denition of a class or function. For example, a class
template allows a uniform and general denition for vectors whose elements
can be int, float, double, long double, and so on. A function template
permits its arguments and return value to have type parameters. This is
di
erent from function overloading that requires several similar denitions
to be given.
A template is often called a parameterized type that directly supports
generic programming. What a template provides is referred to as compiletime polymorphism, since the compiler will generate di
erent forms of a
template according to the types of the actual arguments used. As function
overloading, class and function templates impose no run-time overhead.
Besides talking about class and function templates, this chapter also
presents standard libraries on complex numbers, valarrays, and numeric
algorithms. Surprisingly, templates can be used to improve run-time eciency. Several ecient template techniques are given. The chapter ends
with a section on polynomial interpolation.

7.1 Class Templates

The keyword template is used when dening a template class or function.
For example, a simple template class for vectors with a type parameter T
can be declared as

232

7. Templates

template<class T> class Vcr {

int lenth
T* vr
public:
Vcr(int, const T* const)
Vcr(int = 0, T = 0)
Vcr(const Vcr&)
~Vcr(){ delete] vr}

// number of entries
// entries of vector
//
//
//
//

constructor
constructor
copy constructor
destructor

int size() const { return lenth }

// number of entries
Vcr& operator=(const Vcr&)
// assignment
T& operator](int i) { return vri] } // subscripting
Vcr& operator+=(const Vcr&)
// v += v2
T maxnorm() const
// maximum norm
template<class S>
// dot product
friend S dot(const Vcr<S>&, const Vcr<S>&)
}

The template<class T > prex species that a template is being declared

with the template type parameter T . The scope of T extends to the end
of the declaration. Here T means a type name that need not be a class.
If one prefers, a template can be equivalently declared using the keyword
typename:
template<typename T> class Vcr

// a template declaration

Notice the friend dot() is itself a template with template parameter S .

Once members and friend of this template class are dened, the compiler
can automatically generate di
erent classes. They can be used as
Vcr<double> dv(10)
Vcr<int> iv
Vcr<float*> fv(15)

// vector of 10 doubles,
// empty vector of int
// a vector of 15 (float*)s

double* b = new double 10]

for (int i = 0 i < 10 i++)
Vcr<double> dv2(10, b)
dv2 += dv
double c = dot(dv,dv2)

bi] = i
// a vector of 10 doubles
// add and assign
// dot product of two vectors

Note that the type parameter is replaced by specic type names bracketed by < > and is put after the template class name. The compiler automatically generates denitions for classes Vcr< double >, Vcr< int >,
and Vcr<float > from the class template Vcr<T >. In particular, the
compiler-generated class Vcr<double> is a class for vectors whose entries
are doubles and works exactly like the class Vtr dened in x6.3. The use

7.1 Class Templates

233

of a template class does not impose any run-time overhead compared to

an equivalent \hand-written" class such as Vtr, although certain compilation time is needed to generate specic classes from a class template. Such
compilation time should be in general less than compiling di
erent versions
of a class for di
erent element types when not using a template. Besides,
templates can make user code shorter and have space eciency.
It may sometimes be a good idea to debug a particular class such as
Vtr before turning it into a class template such as Vcr<T >. In this way,
many design problems and code errors can be handled in the context of
a concrete example, which is generally easier than dealing with a general
template class. However, for experienced programmers, dening a template
directly may save a lot of code development time.

7.1.1 Member and Friend Denitions

Member functions of a template class are themselves templates. In the case

of Vcr< T >, its member functions are templates parameterized by the
parameter of the template class. (In general, a member function or friend
can be parameterized by other parameters see x7.2.5 and x7.2.6) They
can be dened either inside the class declaration like the destructor and
subscript operator above or outside as templates. For example,
template<class T> Vcr<T>::Vcr(int n, const T* const abd) {
vr = new T lenth = n]
for (int i = 0 i < lenth i++) vri]= *(abd +i)
}
template<class T> Vcr<T>& Vcr<T>::operator=(const Vcr& v) {
if (this != &v) {
if (lenth != v.lenth) cout << "bad assignment of vector"
for (int i = 0 i < lenth i++) vri] = vi]
}
return *this
}
template<class T> T Vcr<T>::maxnorm() const {
T nm = abs(vr0])
for (int i = 1 i < lenth i++) nm = max(nm, abs(vri]))
return nm
}
template<class T>
// dot product
T dot(const Vcr<T>& v1, const Vcr<T>& v2) {
if (v1.lenth != v2.lenth) cout << "bad vector sizes\n"
T tm = v10]*v20]

234

7. Templates

for (int i = 1 i < v1.lenth i++) tm += v1i]*v2i]

return tm
}

Other member functions can be dened similarly. Although dot<S > is

declared using parameter S , it can be dened using any parameter such as
T.
Within the scope of Vcr<T>, qualication with <T> to refer to Vcr<T>
is redundant and can be omitted, although it is not an error to keep the redundance. So the constructor and assignment operator can be alternatively
dened as
template<class T> Vcr<T>::Vcr<T>(int n, const T* const abd) {
vr = new T lenth = n]
for (int i = 0 i < lenth i++) vri]= *(abd +i)
}
template<class T> Vcr<T>& Vcr<T>::operator=(const Vcr<T>& v){
if (this != &vec) {
if (lenth != v.lenth) cout << "bad assignment of vector"
for (int i = 0 i < lenth i++) vri] = vi]
}
return *this
}

However, for friends and ordinary function templates, the qualication with
<T> is not redundant and must not be omitted.

7.1.2 Template Instantiation

The process of generating a class declaration from a template class and a

specic template argument is often called template instantiation. A version of a template for a particular template argument is called a template
specialization. For example, for the program,
int main() {
Vcr<float> fv(20)
Vcr<double> dv(500)
Vcr<double> dv2(500, 3.14)
dv += dv2
}

the compiler will generate declarations for Vcr<float> and Vcr<double>,

for their constructors and destructors, and for the operator Vcr<double>::
operator+=(const Vcr<double> &). Other member and nonmember functions are not used and should not be generated. Specializations are generated when they are used. The user can also force an explicit instantiation:

7.1 Class Templates

235

template class Vcr<long double>

// a class
template short& Vcr<short>::operator](int) // a member fcn

When a class template is explicitly instantiated, every member and friend

is also instantiated.

7.1.3 Template Parameters

A template can take several parameters and a template parameter can be

used in the denition of subsequent template parameters. For example,
template<class C, class T, C val, int i> class Sometype {
T vi]
int sz
C d = val
public:
Sometype(): sz(i) {cout << "member d is:" << d << '\n'}
}
Sometype<int, double, 3, 25> di32

Here val of type C and i of type int are also parameters of the template
class. Such parameters can be a constant expression, the address of an
object or function with external linkage, or a nonoverloaded pointer to a
member. It is an error if their values can not be determined at compile
time.

7.1.4 Type Equivalence

Given a template class, di

erent types (specializations) can be generated
by supplying di
erent template arguments. For example,
Vcr<long double> ldv
Vcr<unsigned int> uiv(33)
Vcr<int> iv(33)
typedef unsigned int Uint
Vcr<Uint> uintv(33)

The classes Vcr<int> and Vcr<unsigned int> are di

erent types, while
Vcr<Uint> and Vcr<unsigned int> are the same, since typedef simply
creates a type synonym but does not introduce a new type.

7.1.5 User-Dened Specializations

A close look at the template class Vcr<T> and the friend dot<T> () reveals
that it does not work correctly when the type parameter T is a complex

236

7. Templates

number, since the maximum-norm function maxnorm() should return a

real number and the dot-product function dot<T> () should include complex conjugation. This could be xed by including conjugation even for
dot products of real numbers and letting maxnorm() always return a long
double. But this would sacrice eciency. A user-dened specialization is
a way to solve this problem. Now I dene a version (a specialization) of the
template class Vcr<T> when its elements are complex numbers the C++
standard library <complex> (dened as a class template itself see x7.4) is
used in this denition.
// ******* a specialization of Vcr<T> for
template<class T> class Vcr< complex<T> >{
int lenth
complex<T>* vr
public:
Vcr(int, const complex<T>* const)
//
Vcr(int = 0, complex<T> = 0)
//
Vcr(const Vcr&)
//
~Vcr() { delete] vr }
//

complex numbers
// use <complex>

constructor
constructor
copy constructor
destructor

int size() const { return lenth }

complex<T>& operator](int i) const { return vri] }
Vcr& operator+=(const Vcr&)
// add-assign
Vcr& operator=(const Vcr&)
// assignment
T maxnorm() const
// maximum norm
template<class S> friend complex<S>
// dot product
dot(const Vcr<complex<S> >&, const Vcr<complex<S> >&)
}

Note that this specialization is itself a template class such a specialization is called a partial specialization. The template<class T> prex species that this template class has a type parameter T . The <complex<T>>
after the template class name means that this specialization is to be used
when the template argument in Vcr<T> is a complex number. Its members
and nonmembers must be redened to accommodate special situations of
vectors of complex numbers. They may be dened as
template<class T>
Vcr< complex<T> >::Vcr(int n, const complex<T>* const abd) {
vr = new complex<T> lenth = n]
for (int i = 0 i < lenth i++) vri] = *(abd +i)
}
template<class T>
Vcr< complex<T> >::Vcr(const Vcr & vec) { // copy constructor

7.1 Class Templates

237

vr = new complex<T> lenth = vec.lenth]

for (int i = 0 i < lenth i++) vri] = veci]
}
template<class T>
T Vcr< complex<T> >::maxnorm() const {
T nm = abs(vr0])
for (int i = 1 i < lenth i++)
nm = max(nm, abs(vri]))
return nm
}

// maximum norm

// use <algorithm>

template<class T>
// dot product
complex<T> dot(const Vcr<complex<T> >& v1,
const Vcr<complex<T> >& v2) {
if (v1.lenth != v2.lenth) cout << "bad vector sizes\n"
complex<T> tm = v10]*conj(v20])
// conjugate of v2
for (int i =1 i <v1.lenth i++) tm += v1i]*conj(v2i])
return tm
}

The prototypes and denitions for maxnorm() and dot() are very di
erent
now from the previous ones.
Now we can write the following function main() to test it.
int main(){
int n = 300
complex<double>* aa = new complex<double> n]
for (int j = 0 j < n j++) aaj] = complex<double>(5, j)
Vcr< complex<double> > v1(n, aa)
// vector v1
Vcr< complex<double> > v2(n)
// vector v2
for (int j = 0 j < n j++) v2j] =complex<double>(2, 3+j)
cout << "norm = " << v1.maxnorm() << '\n' // max norm
cout << "dot = " << dot(v1,v2)
<< '\n' // dot product
}

Here is another example of partial specialization:

template <int height, int width> class screen {
// a class for screen with a height and width
// height and width are two template parameters
}
template <int height> class screen<height, 80> {
// a partialization for screens with 80-character width

238

7. Templates

// which may be implemented more efficiently

// Now only height is parameter, width = 80.
}

A complete specialization is one that does not take any type parameters.
For example, a complete specialization for Vcr<T> is:
template<> class Vcr< complex<double> >{
int lenth
//
complex<double>* vr
//
public:
Vcr(int, complex<double>*)
Vcr(int = 0, complex<double> = 0)
Vcr(const Vcr&)
//
~Vcr() {delete] vr }
//

number of entries
entries of vector

copy constructor
destructor

Vcr& operator=(const Vcr&)

int size() { return lenth }
complex<double>& operator](int i) const { return vri] }
Vcr& operator+=(const Vcr&)
// add-assign
double maxnorm() const
// maximum norm
template<> friend
// dot product
complex<double> dot(const Vcr&, const Vcr&)
}

The template<> prex species that this class is a specialization for Vcr<
T> that does not take any type parameters. It is a complete specialization.
The same explanation holds for template function dot(). All members and
friends must be dened for a specialization. However, only two denitions
are presented here (the rest are omitted to save space):
Vcr< complex<double> >&
Vcr< complex<double> >::operator+=(const Vcr& vec) {
if (lenth != vec.lenth ) cout << "bad vector sizes\n"
for (int i = 0 i < lenth i++) vri] += veci]
return *this
}
template<>
complex<double> dot(const Vcr<complex<double> >& v1,
const Vcr<complex<double> >& v2) {
if (v1.lenth != v2.lenth ) cout << "bad vector sizes\n"
complex<double> tm = v10]*conj(v20])
// conjugation
for (int i = 1 i < v1.lenth i++) tm += v1i]*conj(v2i])
return tm
}

7.1 Class Templates

239

All specializations of a template must be declared in the same namespace as the template itself. If a specialization is explicitly declared such as
Vcr< complex<T > >, it (including all of its members) must be dened
somewhere. In other words, the compiler will not generate denitions for
it.
Similarly, a partial specialization for vectors of pointers can be dened
as
// ************* a specialization of Vcr<T> for pointers
template<class T> class Vcr<T*>{
int lenth
// number of entries
T** vr
// entries of vector
public:
Vcr(int, T**)
Vcr(const Vcr&)
// copy constructor
T* & operator](int i) const { return vri] }
Vcr& operator=(const Vcr&)
// ... other members and definitions
}

Specialization is a way of specifying alternative and often ecient implementations for specic type parameters. However, the user interface is not
changed in other words, with or without the specializations, the user still
uses a Vcr of complex numbers and a Vcr of pointers the same way. See Exercise 7.9.4 for an alternative approach to specializing vector maxnorm().

7.1.6 Order of Specializations

One specialization is more specialized than another if one argument list

that matches its specialization pattern also matches the other but not vice
versa. For example,
template<class T> class Vcr
// general
template<class T> class Vcr< complex<T> >
// specialized for complex numbers
template<> class Vcr< complex<float> >
// specialized for single precision complex numbers

In the above, the second class is more specialized than the rst one, and
the third is more specialized than the second. The most specialized version
will be preferred over others in declarations of objects, pointers, and in
overload resolution. For example, in the declaration,
Vcr< complex<double> > cdv(100)

the second class is preferred over the rst and general class.

240

7. Templates

7.2 Function Templates

A function template denes a function that takes type parameters. In x7.1,
the basic idea of a function template is illustrated through a dot product
example dot<T > () (which is a function template that happens to be a
friend of a template class) in the context of vectors. This section presents
more details on function templates.
Assume we want to sort a general array of elements in increasing order
a sort function can be dened as a function template:
template<class T> void sort(T*, int)

// function template

void f(int* vi, double* vd, string* vs, int n) {

sort(vi, n)
// sort(int*, int), T = int
sort(vd, n)
// sort(double*, int), T = double
sort(vs, n)
// sort(string*, int), T = string
}

When a function template is called, the type of the function arguments

determines which version (specialization) of the template is used. In other
words, the template arguments are deduced from the function arguments.
For example, in the call sort(vi n), the template argument is deduced as
T = int from the type of vi.
Based on the shell sort algorithm Knu98, Str97], the sort function template can be dened as
template<class T> void sort(T* v, int n) {
for (int gap = n/2 gap > 0 gap /= 2)
for (int i = gap i < n i++)
for (int j = i - gap j >= 0 j -= gap)
if (vj + gap] < vj]) {
T temp = vj]
// swap vj] and vj+gap]
vj] = vj + gap]
vj + gap] = temp
}
}

Note that the operator < is used for comparison. However, not every type
has such an operator. Improvements are given in x7.2.3 and x7.2.4. Another
way is to overload the operator <. For example, an array of point2d can
be sorted according to its x-coordinate as
struct point2d {
double x, y
}
inline bool operator<(const point2d& p, const point2d& q) {

7.2 Function Templates

return p.x < q.x

241

// compare x coordinates

}
int main() {
const int n = 10000
point2d an]
for (int i = 0 i < n i++) {
ai].x = i*(5 - i) ai].y = 5/(i+2.3) - 1
}
sort(a, n)
// sort array a of n elements
}

Compare this denition with C-style polymorphic functions that rely

on a pointer to function for comparison (see Exercise 3.14.24 or the book
Str97]). This template denition sort<T >() is shorter and most likely
faster since function calls through pointers are hard to be inlined. Running
times for this function template and the shell sort function shsort() dened
in Exercise 3.14.24 for the array of 10000 point2d on a Pentium PC are 70
and 109 seconds, respectively. They are 27 and 86 seconds on a SUN Ultra2
workstation, 10 and 19 seconds on a Pentium II PC, and 18 and 30 seconds
on an SGI workstation, respectively. On these four machines, this function
template is much faster.

7.2.1 Function Template Parameters

A compiler can deduce the types of the parameters of a function template, provided the function argument list identies them uniquely. However, when a template parameter can not be deduced from the template
function arguments, it must be specied explicitly within a pair of angle
brackets:
template<class T> T* f(T)
int i = 5
double d = 3.4
f(i)
f(5)
f(d)

// template parameter T = int

// template parameter T = int
// template parameter T = double

template<class T, class S> T g(S)

g<int, double>(d)
g<double>(i)
g(i+d)
g<double, double>(i+d)

//
//
//
//

T = int, S = double
T = double, S = int
error: can not deduce T and S
T = double, S = double

242

7. Templates

As with default function arguments, only trailing template parameters can

be omitted. The call g<double>(i) is equivalent to g<double int>(i) since
the compiler can deduce S = int by the type of i. However, the call g(i + d)
is ambiguous since there is no way to deduce template parameters T and
S from it.
Note that class template parameters are not allowed to be deduced and
must be specied explicitly since ambiguities arise for classes with multiple
constructors.

7.2.2 Function Template Overloading

As functions can be overloaded, function templates can also be overloaded.

For example,
template<class T> T abs(T)
// version 1, template
template<class T> T abs(complex<T>) // version 2, template
double abs(double)
// version 3, ordinary fcn
abs(2.0)
abs(2)
complex<double> z(2,3)
abs(z)

// abs(double), version 3
// abs<int>(int), version 1
// abs<double>(complex<double>)

In the case of more than one overloaded function template or specialization,

the most specialized one will be chosen and used according to the arguments
of the function call. The detailed overloading resolution rules are as follows.
1. Find the set of function template specializations that match the
function call. For the function call abs(z ) above, the specializations
abs < double > (complex < double >) (version 2 with T = double)
and abs<complex<double> > (complex<double>)(version 1 with
T = complex<double>) are candidates.
2. If more than one template function can be called, consider the most
specialized ones. For the call abs(z ), the specialization abs<double>
(complex<double>) (version 2) is preferred over the specialization
abs<complex<double> > (complex<double>) (version 1), because
any call that matches abs<complex<T> > (version 2) also matches
abs<T> (version 1).
3. Do overload resolution for this set of template function specializations
and ordinary functions. If a template function argument has been
determined by template argument deduction (x7.2.1), that argument
can not have standard type conversions, promotions, or user-dened
conversions. For the call abs(2), abs<int> (int) is an exact match
(without type conversions, etc.) and preferred over abs(double).

7.2 Function Templates

243

4. If an ordinary function and a template specialization are equally good

matches, the function is preferred. For example, the function call
abs(double) is preferred over abs<double> (double) (version 1 with
T = double) for the call abs(2:0).
5. If more than one equally good match is found, the call is ambiguous
and an error. If no match is found, the call is also an error.
The following is another example.
template<class C> C absmax(C a, C b) {
return (abs(a) > abs(b)) ? a: b
}
absmax(-1, 2)
absmax(2.7, 5.6)
absmax(5, 6.9)

//
//
//
//

absmax<int>(1, 2)
absmax<double>(2.7,5.6)
ambiguous, absmax<double>(5,6.9)
or absmax<int>(5,6.9)?

The call absmax(5 6:9) is ambiguous and illegal, since no standard conversion is made for its arguments. This ambiguity can be resolved either by
explicit qualication:
absmax<double>(5, 6.9)

or by dening an ordinary function, to which standard type conversion

applies:
inline double absmax(double a, double b) {
return absmax<double>(a,b)
}

If a function template also takes basic types as its arguments, for example,
template<class T, class S, class R> R f(T& t, S s, double d)

then standard conversions apply to the basic type argument d, although no

standard conversion can be implied for the arguments t and s.

7.2.3 Specializations

The sort program given early in this section does not sort arrays of char
correctly since it will compare the addresses of the rst char for each element of the array. Specialization can be used to dene a comparison function that compares di
erent types of elements correctly. We now use the
standard library <vector>, dened as a class template (see x10.1.1), for a
dynamic array of elements. For example,

244

7. Templates

template<class T> bool less(T a, T b) {

return a < b
// a general template
}
template<> bool less<const char*>(const char* a,
const char* b) {
return strcmp(a,b)
// specialization
}
// use string-compare in <cstring>
template<class T> void sort(vector<T>& v) {
const unsigned long n = v.size()
for (int gap = n/2 gap > 0 gap /= 2)
for (int i = gap i < n i++)
for (int j = i - gap j >= 0 j -= gap)
if (less(vj+gap], vj])) swap(vj], vj+gap])
}

Here the function template swap(), dened in the library <algorithm>,

is used to swap the values of two variables. As for class templates, the
template<> prex implies a specialization that does not have a type parameter. The < const char > after the template function name means
that this specialization is to be used when the template argument T =
const char. Now the function template sort<T >() should sort numbers
and strings correctly. Since the template argument can be deduced from
the function argument list, this specialization can also be written as
template<> bool less(const char* a, const char* b) {
return strcmp(a,b)
// specialization, using <cstring>
}

If complex numbers are going to be compared according to their absolute

values, another specialization can be dened as
template<class T> bool less(complex<T> a, complex<T> b) {
return abs(a) < abs(b)
// specialization, use <complex>
}

With di
erent specializations of template less<T > (), a vector of strings,
a vector of complex numbers, or a vector of point2d can be sorted in increasing order or any other order dened by the specializations.
Another example of function template specialization is the useful vector
dot product, applied to real and complex arrays:
template<class T> T dot(T* a, T* b, int n) {
T init = 0
for (int i = 0 i < n i++) init += (*a++)*(*b++)
return init
}

7.2 Function Templates

245

template<class T> // specialization for complex vectors

complex<T> dot(complex<T>* a, complex<T>* b, int n) {
complex<T> init = 0
for (int i = 0 i < n i++) init += (*a++)*conj(*b++)
return init
}

7.2.4 Class Templates as Function Template Parameters

When comparing vectors of char or other types according to di

erent criteria such as case-sensitive and case-insensitive comparisons, it is convenient
to pass the criterion as a class template to the compare-function template.
For example, dene a function template compare<T > ():
template<class T, class S>
int compare(const vector<T>& v1, const vector<T>& v2) {
for (int i = 0 i < v1.size() && i < v2.size() i++)
if (!S::eq(v1i], v2i])) return S::st(v1i], v2i])
return v2.size() - v1.size()
}

Here the template parameter S can be a class whose member functions

S :: eq() and S :: st() are dened to represent the comparison or sorting
criteria. For example,
template<class T> class Ncomp {
// normal compare
public:
static int eq(T a, T b) { return a == b } // equal
static int st(T a, T b) { return a < b } // smaller than
}
class Nocase {
// compare by ignoring case
public:
static int eq(char a, char b) {
// equal
return toupper(a) == toupper(b)
// #include <ctype.h>
}
static int st(char a, char b) {
// smaller than
return toupper(a) < toupper(b)
// case-insensitive
}
}

The class Nocase changes characters into their corresponding uppercase

equivalents (using std :: toupper()) before comparing. Each class generated
from a class template gets a copy of each static member of the class template. Now vectors can be compared according to rules specied by the
template argument:

246

7. Templates

void f(vector<char> v1, vector<char> v2) {

compare<char, Ncomp<char> >(v1, v2) // case-sensitive
compare<char, Nocase>(v1, v2)
// case-insensitive
}

Passing comparison criteria as template parameters has the advantage that

it allows several operations to be passed as a single argument and the
comparison operators eq() and st() are trivial to be inlined. This should
increase run-time eciency compared to C- or FORTRAN-style passing
function pointers see x7.2 and x7.6 for run-time comparisons.
Default template arguments can be used as well:
template<class T, class S = Ncomp<T> > // default parameter
int compare(const vector<T>& v1, const vector<T>& v2) {
// define the function template as before
}

Then the normal comparison criterion need not be specied each time it is
used:
void g(vector<char> v1, vector<char> v2) {
compare(v1, v2)
// use Ncomp<char>
compare<char, Nocase>(v1, v2)
// use Nocase
}

Classes used as template parameters for expressing criteria or policies

are often called \traits". See Exercise 7.9.4 for another example of traits.

7.2.5 Member Function Templates

A class or class template can have function members that are parameterized
by general parameters. For example,
template<class S> class Cmpx {
S re, im
public:
template<class T> Cmpx(const Cmpx<T>& c)
: re(c.re), im(c.im) { }
}

The copy constructor of this class template is a template function that

denes a type conversion from Cmpx<T> to Cmpx<S>. It can be used as
void f(Cmpx<float> cf) {
Cmpx<double> cd = cf
}

// float to double conversion

7.2 Function Templates

247

7.2.6 Friend Function Templates

A friend function of a class or a class template may be an ordinary function,

a function template, or an instantiation of a function template. In x7.1, the
function template dot<S> () is declared to be a friend of class template
Vcr<T>:
template<class T> class Vcr {
// ... in addition to other members or friends
template<class S>
// dot product
friend S dot(const Vcr<S>&, const Vcr<S>&)
}

This means that all instantiations of the function template dot<S> ()

are friends to any instantiation of class template Vcr<T>. In particular, dot<double> () is a friend of class instantiations Vcr<double> and
Vcr<float>.
It is also possible to rst dene or declare a function template such as
dot<T> () and then declare an instantiation of it to be a friend of a class
or class template such as class Vcr<T>. For example,
template<class T> class Vcr // template forward declaration
template<class T>
// template forward declaration
friend T dot(const Vcr<T>&, const Vcr<T>&)
template<class T> class Vcr {
// class template definition
// ... in addition to other members or friends
friend T dot<T>(const Vcr&, const Vcr&)
}

This means that only one corresponding instantiation of dot<T>() is a

friend to each instantiation of Vcr<T>. In particular, dot<double>() is a
friend of class Vcr<double>, but not Vcr<float>. Notice the instantiation
syntax in the friend declaration dot<T>(). When it is omitted:
template<class T> class Vcr {
// class template definition
// ... in addition to other members or friends
friend T dot(const Vcr&, const Vcr&)
}

this would be interpreted as declaring an ordinary function dot() to be a

friend to all instantiations of class template Vcr<T>. It could be dened
as
float dot(const Vcr<float> &v1, const Vcr<float>& v2) {

248

7. Templates

if (v1.lenth != v2.lenth ) error("bad vector sizes")

float tm = v10]*v20]
for (int i = 1 i < v1.lenth i++) tm += v1i]*v2i]
return tm
}
double dot(const Vcr<double>& v1, const Vcr<double>& v2) {
// ...
}

Similarly, a class, a class template, or an instantiation of a class template

can be a friend of another class or class template.

7.3 Template Source Code Organization

Specializations of class and function templates are generated only when
needed (x7.1.2). Thus the denition of templates must be known to the
source code that uses the templates. A trivial strategy is to put everything
in one le so that the compiler can easily generate code as needed. But
this is applicable only to small programs. There are two other ways of
organizing the source code.
Consider the simple example for a function template that prints some
user-specied (error) messages and terminates the program. Assume the
denition of the template is put in a le called error:cc.
// file error.cc
#include <cstdlib>
#include <iostream>
template<class T> void error(const T & t) {
std::cout << t << "\nProgram exited.\n"
std::exit(1)
}

The rst way is to include this template denition le in all les that use
the template. For example,
// file f1.cc
#include "error.cc"
int* a = new (nothrow) int 1000000]
// return 0 when out of memory
if (!a) error("no more space.")
// file f2.cc
#include "error.cc"
int* b = new (nothrow) int 1000000]

7.3 Template Source Code Organization

249

if (!b) error(*b)

That is, the denition of the template and all declarations that it depends
on are included in di
erent compilation units. The operator new throws
an exception bad alloc when there is no more memory available. However,
using (nothrow) causes it to return 0 instead see x9.5. The compiler will
generate appropriate specializations according to how the template is used.
This strategy treats templates the same way as inline functions.
The second way is to include only declarations of templates and explicitly export its denitions to di
erent compilation units. Split the original
error:cc le into two:
// file error2.h: only declarations of the template
template<class T> void error(const T & )
// file error2.cc: definition of template and export it
#include <cstdlib>
#include <iostream>
#include "error2.h"
export template<class T> void error(const T & t) {
std::cout << t << "\nProgram exited.\n"
std::exit(1)
}

Now the declaration of the template can be included in di

erent les:
// file f3.cc
#include "error2.h"
int* a = new (nothrow) int 1000000]
if (!a) error("no more space.")
// file f4.cc
#include "error2.h"
int* b = new (nothrow) int 1000000]
if (!b) error(*b)

This strategy treats templates the same way as noninline (ordinary) functions. The template denition in le error2:cc is compiled separately, and
it is up to the compiler to nd the denition of the template and generate
its specializations as needed. Note that the keyword export means accessible from another compilation unit. This can be done by adding export to
the denition or declaration of the template. Otherwise the denition must
be in scope wherever the template is used.

250

7. Templates

7.4 Standard Library on Complex Numbers

Complex numbers are provided as a template class in the C++ standard
library <complex>. It is dened as
template<class T> class std::complex {
T re, im
public:
complex(const T& r = T(), const T& i = T()) // constructor
template<class S> complex(const complex<S>&)// constructor
complex<T>& operator=(const T&)
template<class S> complex<T>& operator=(const complex<S>&)
// also operators +=, -=, *=, /=
T real() const { return re }
T imag() const { return im }
}

Notice the two template member functions above, which enable construction and assignment from other types of complex numbers, for example, from complex<float> to complex<double>. However, specializations will be dened to restrict implicit conversions, for example, from
complex<double> to complex<float>, in order to avoid truncation.
Other functions acting on complex numbers can be dened as
template<T> complex<T> operator+(const complex<T>&,
const complex<T>&)
template<T> complex<T> operator+(const T&,const complex<T>&)
template<T> complex<T> operator+(const complex<T>&,const T&)
// also binary operators: -, *, /, ==, !=
template<T> complex<T> operator+(const complex<T>&)
// also unary operator: template<T> T real(const complex<T>&)
// real part
template<T> T imag(const complex<T>&)
// imaginary part
template<T> complex<T> conj(const complex<T>&) // conjugate
template<T> T abs(const complex<T>&)
// absolute value
template<T> T arg(const complex<T>&)
// argument
template<T> complex<T> polar(const T& abs, const T& arg)
// polar form (abs,arg)
template<T> T norm(const complex<T>&)
// square of abs()
template<T> complex<T> sin(const complex<T>&)

// sine

7.5 Standard Library on valarrays

251

// also: cos, tan, sqrt, sinh, cosh, tanh, exp,

//
log, log10, asin, acos, atan
template<T>
template<T>
template<T>
template<T>

complex<T>
complex<T>
complex<T>
complex<T>

pow(const
pow(const
pow(const
pow(const
const

complex<T>&, int)
complex<T>&, const T&)
T&, const complex<T>&)
complex<T>&,
complex<T>&)

The function polar() constructs a complex number given its absolute value
and argument. Note the misleading function name of norm(), which gives
the square of the absolute value of a complex number. The input and output
operators << and >> are also provided. Specializations for complex<float>,
complex<double>, and complex<long double> are dened to prevent certain conversions and to possibly optimize the implementations. For example,
class complex<float> {
float re, im
public:
complex(float r = 0.0, float i = 0.0)
complex(const complex<float>&)
explicit complex(const complex<double>&)
explicit complex(const complex<long double>&)
// ... other functions
}

The explicit constructors declared above prevent implicit conversions

from complex<double> and complex<long double> to complex<float>.
For example,
complex<float> cf(3,4)
complex<double> cd(5,6)
cd = cf
cf = cd
cf = complex<float>(cd)

// OK, implicit conversion

// illegal, no implicit conversion
// OK, explicit conversion

7.5 Standard Library on valarrays

C++ provides a standard library called <valarray> for optimized vector
arithmetics in high performance computing. Since the design priority of this
library is speed of computing, other features such as exibility, generality,
and type checking are sacriced. For example, it does not check to see if
the sizes of two vectors match before adding or multiplying them, and does
not provide operations for various norms of a vector and dot products.
Matrices are not directly supported, but a user can manipulate matrices

252

7. Templates

and submatrices through the (one-dimensional) valarray mechanism. If

your main concern is easy-to-use and easy-to-modify in a general setting,
you may want to write your own library (see Chapter 11) to t your need.

7.5.1 The Type valarray

Some members of the type valarray are dened as

template<class T> class std::valarray
public:
valarray()
//
explicit valarray(size_t n)
//
valarray(const T* p, size_t n) //
valarray(const T& v, size_t n) //
valarray(const valarray&)
//
~valarray()
//

{
valarray with size = 0
valarray with size = n
size =n, elements *p
size =n, all entries =v
copy constructor
destructor

valarray& operator=(const valarray&)

valarray& operator=(const T& v)
T operator](size_t) const
T& operator](size_t)

//
//
//
//

copy assignment
all elements = v
subscript, rvalue
subscripting

// scalar-vector operations: +=, -=, *=, /=, ^=, &=,

//
|=, <<=, >>=
valarray& operator+=(const T& v) // add v to every element
// unary operator: -, +, ~, !
valarray& operator-()

// unary minus

size_t size() const

// number of elements
void resize(size_t n, const T& v= T()) // all entries = v
T sum() const
// sum of all elements
valarray
valarray
valarray
valarray
}

cshift(int i) const
// cyclic shift of entries
shift(int i) const
// logical shift
apply(T f(T)) const
// apply f to each entry
apply(T f(const T&)) const
// apply f()

Here size t is an unsigned integral type (possibly unsigned long) dened

in <cstddef> and used in the standard libraries. In functions shift(i) and
cshift(i), left shift of elements by i positions is performed when i > 0, and
right shift when i < 0.
Here is an example for illustrating some usages of valarray.
double* pd = new double 50000]

7.5 Standard Library on valarrays

253

for (int i = 0 i <50000 i++ ) pdi] = i*i

valarray<double> v0(pd,50000)
// vector of 50000 doubles
int permu] =
valarray<int>
valarray<int>
valarray<int>
valarray<int>
valarray<int>

{1,2,3,4,5}
v(permu,5)
v2 = v.shift(2)
v3 = v.cshift(2)
v4 = v.shift(-2)
v4 = v.cshift(-2)

//
//
//
//
//
//

permutation
vector of 5 int
v2 = {3,4,5,0,0}
v2 = {3,4,5,1,2}, cyclic
right shift,v2={0,0,1,2,3}
right shift,v2={4,5,1,2,3}

Other operators and functions are provided as nonmember functions:

// +, -, *, /, ^, &, |, <<, >>, &&, ||, ==, !=, <, >, <=, >=,
// sin, cos, tan, asin, acos, atan, atan2, sinh, cosh, tanh,
// pow, exp, log, log10, sqrt, abs
valarray<T> operator+(const valarray<T>&,const valarray<T>&)
valarray<T> operator+(const T&, const valarray<T>&)
valarray<T> operator+(const valarray<T>&, const T&)
valarray<T> abs(const valarray<T>&)
T min(const valarray<T>&)
T max(const valarray<T>&)

// smallest value
// largest value

Note the function abs() needs a specialization in order to apply to a valarray

of complex numbers.
The output operator is not provided for valarray since di
erent users
do it di
erently, especially in di
erent applications. Neither is the input
operator provided.

7.5.2 Slice Arrays

Slices are introduced in <valarray> for manipulating matrices through

one-dimensional arrays. A slice describes every nth element of some part
of a valarray, for a positive integer n. It is dened as
class std::slice {
// ... representation, starting index, size, stride
public:
slice()
slice(size_t start, size_t size, size_t stride)
size_t start() const
// index of first element
size_t size() const
// number of elements in slice
size_t stride() const // entry i is at start()+i*stride()
}

254

7. Templates

A stride is the distance (in number of elements) between two adjacent

elements in the slice. For example, the 4 by 3 matrix A can be laid out as
a vector B of size 12 :
2

a00
6 a
A = 64 a10
20
a30
0

a01
a11
a21
a31
1

a02
a12 77
a22 5
a32
2

10 11

B = a00 a01 a02 a10 a11 a12 a20 a21 a22 a30 a31 a32 :
Then slice(0 3 1) of vector B describes row 0 of matrix A (consisting of 3
elements starting from 0 with stride 1 that is, B 0] = a00 , B 1] = a01 , and
B 2] = a02 ) and slice(3 3 1) describes row 1. Similarly, the slice(0 4 3)
of vector B describes column 0 of matrix A and slice(2 4 3) column 2.
That is, a slice(s z d) denes a mapping from three nonnegative integers
fs z dg into an index set I = fi : i = s + j d j = 0 1 : : : z ; 1g that
describes some elements vi] i 2 I , of a valarray v. In particular, a matrix
can be represented through a valarray and its rows and columns can be
described by slices. The idea is that a row or column of a matrix can be
accessed by using one index j for j = 0 1 : : : z ; 1, where z is the number
of elements in the row or column. Note that the row by row layout has
been used for representing a matrix through a valarray. FORTRAN-style
column by column layout can be used similarly. However, it may be better
to get used to the C++ style (row by row layout, array index starting from
0 instead of 1, and the symmetric notation mi]j ] instead of m(i j )) if one
wants to program in C++ extensively.
A type called slice array is dened in <valarray> to refer to elements
of a valarrary described by slices:
template <class T> class std::slice_array {
private:
slice s
valarray<T>* p
// implementation-dependent
slice_array()
// prevent construction
slice_array(const slice_array&) // prevent construction
slice_array& operator=(const slice_array&)
public:
void operator=(const valarray<T>&)
void operator=(const T& t)
// assign t to each entry
void operator*=(const T&)
// scalar-vector multiply

7.5 Standard Library on valarrays

255

// also: +=, -=, /=, %=, ^=, &=, |=, <<=, >>=
~slice_array()
}

Since construction and copy operations are dened as private, a user can
not directly create a slice array. Instead, subscripting a valarray is dened
to be the way to create a slice array corresponding to a given slice:
template <class T> class std::valarray {
public:
// in addition to other members
valarray(const slice_array<T>&)
valarray operator](slice) const
slice_array<T> operator](slice)
valarray& operator=(const slice_array<T>&)
}

Once a slice array is dened, all operations on it go to the valarray:

void create(valarray<double>& vd) {
slice s(1, vd.size()/2, 2)
// slice for elements vd1], vd3]...
slice_array<double>& sad = vds]
// subscripting a valarray by a slice
sad0] = 3.1415926
// vd1] = 3.1415926
sad = 2.2
// assign 2.2 to vd1], vd3], ...
float a = sad1]
// a = vd3]
float b = sad2]
// b = vd5]
valarray<double> ods = vds]
// ods has only odd-index entries of vd
}

Note that it is illegal to copy slice arrays. The reason for this is to allow
alias-free optimizations and prevent side e
ects. For example,
slice_array<double> create2(valarray<double>& vd) {
slice_array<double> sad0 // error, no default constructor
slice s(1, vd.size()/2, 2)
// slice for odd index entries of vd
slice_array<double> sad1 = vds]
// illegal, no copy construction
sad1 = sad0
return sad0
}

// illegal, no copy assignment

// illegal, can not make a copy

256

7. Templates

7.5.3 Generalized Slice Arrays

A slice can be used to describe rows and columns of a matrix and multidimensional arrays represented through a valarray. In some applications,
submatrices need be manipulated. Generalized slices gslice and arrays
gslice array are introduced for this purpose.
An n-slice or generalized slice gslice contains n di
erent dimensions (n
slices) and is dened as
class std::gslice {
// ... representation, starting index, n sizes, n strides
public:
gslice()
gslice(size_t start, const valarray<size_t> nsize,
const valarray<size_t> nstride)
size_t start() const
// index of first entry
valarray<size_t> size() const
// number of elements in all dimensions
valarray<size_t> stride() const
// strides for all dimensions
}

That is, an n-slice gslice(s z d), where z = fzk g d = fdk g 0 P

k < n,
;1 jk
denes a mapping from (s z d) into an index set I = fi : i = s + nk=0
dk where jk = 0 1 : : : zk ; 1g that describes some elements vi] i 2 I ,
of a valarray v. For example, consider a 3 by 2 submatrix A32 (inside the
box in the picture below) in the lower-right corner of a 4 by 3 matrix A
represented row by row through a valarray B :
2

a00
6 a
A = 64 a10
20
a30
0

a01
a11
a21
a31
1

a02
a12
a22
a32
2

7
7
5

10 11

B = a00 a01 a02 a10 a11 a12 a20 a21 a22 a30 a31 a32 :
Since every row of the submatrix A32 can be represented by a slice of size
2 and stride 1 and every column by a slice of size 3 and stride 3, a 2-slice
can be dened to describe the submatrix:
size_t z] = {2,3}
// slice(s,z0],d0]) for a row
size_t d] = {1,3}
// slice(s,z1],d1]) for a column
{\it valarray}<size_t> vz(z,2) // size vector

7.5 Standard Library on valarrays

257

{\it valarray}<size_t> vd(d,2) // stride vector

gslice gs2(4,vz,vd)
// starting index is 4

Thus gs2 denes a mapping from f(4 vz vd)g into the index set I = fi : i =
4+ j0 +3j1 j0 = 0 1 j1 = 0 1 2g for the submatrix A32 . In particular, the
subset fi : i = 4+j0 +3j1 j0 = 0 1 j1 = 0g = f4 5g describes row 0 of A32
(i.e, B 4] = a11 B 5] = a12 ), and fi : i = 4+ j0 +3j1 j0 = 0 j1 = 0 1 2g =
f4 7 10g describes column 0 of A32 (i.e., B 4] = a11 B 7] = a21 B 10] =
a31 ). The idea is that the submatrix A32 can now be conveniently accessed
using two indexes j0 and j1 for j0 = 0 1 : : : z0 ; 1 and j1 = 0 1 : : : z1 ; 1,
where z0 is the number of elements in each row and z1 the number of
elements in each column of the submatrix.
The template class gslice array is dened in the same way and o
ers
the same set of members. This implies that a gslice array can not be constructed directly and can not be copied by the user. Instead, subscripting
a valarray is dened to be the way to create a gslice array corresponding
to a given n-slice:
template <class T> class std::valarray {
public:
// in addition to other members
valarray(const gslice_array<T>&)
valarray operator](const gslice&) const
gslice_array<T> operator](const gslice&)
valarray& operator=(const gslice_array<T>&)
}

7.5.4 Mask Arrays and Indirect Arrays

A mask array is dened from a valarray v through a mask valarray<bool>,

with true indicating an element of the valarray v is selected. Subscripting
a valarray through a mask creates a mask array:
template <class T> class std::valarray {
public:
// in addition to other members
valarray(const mask_array<T>&)
valarray operator](const valarray<bool>&) const
mask_array<T> operator](const valarray<bool>&)
valarray& operator=(const mask_array<T>&)
}

Similarly, an indirect array is dened from a valarray v through an index

array (a permutation) valarray<size t>. Subscripting a valarray through
an index array valarray<size t> creates an indirect array:
template <class T> class std::valarray {

258

7. Templates

public:
// in addition to other members
valarray(const indirect_array<T>&)
valarray operator](const valarray<size_t>&) const
indirect_array<T> operator](const valarray<size_t>&)
valarray& operator=(const indirect_array<T>&)
}

Here is an example on how to use a mask array and indirect array:

void g(valarray<double>& vd) {
bool ba4] = {false, true, false, true}
valarray<bool> mask(ba,4)
// a mask
valarray<double> vd2 = sin(vdmask]) // vd20]=sin(vd1])
// vd21]=sin(vd3])
size_t ia4] = {3, 1, 0, 2}
valarray<size_t> index(ia,4)
// an index array
valarray<double> vd3 = sin(vdindex]) //vd30]=sin(vd3])
// vd31] =sin(vd1]), vd32]=sin(vd0]), vd33]=sin(vd2])
}

An irregular subset of an array can be selected by a mask array and a

reordering of an array can be achieved by an indirect array. An index array
must be a permutation that is, it can not contain the same index more
than once. For example, the behavior of the indirect array is undened if
the above index array is dened as
size_t ia4] = {3, 3, 0, 2}

The same set of members for mask array and indirect array are dened
as for slice array. In particular, mask array and indirect array can not
be directly constructed or copied by a user.

7.6 Standard Library on Numeric Algorithms

The standard library <numeric> contains a few numeric algorithms for
performing operations on sequences such as the built-in array (x3.3), valarray (x7.5), and vector (x10.1.1). Other standard algorithms such as sorting
a sequence of elements and nding the maximum value of elements in a
sequence are discussed in x10.2.
The numeric algorithms are function templates and can be summarized
as
accumulate()
accumulate results on a sequence
inner product()
accumulate results on two sequences
partial sum()
generate a sequence of \partial sums"
adjacent dierence() generate sequence of \adjacent di
erences"

7.6 Standard Library on Numeric Algorithms

259

7.6.1 Accumulate

The function template accumulate() calculates the sum of the elements in

a sequence given a starting and an ending pointer (or iterator see x10.1.2),
and an initial sum. A general binary operation, instead of summation, can
also be specied as the fourth argument. It is dened as
template<class In, class T>
T accumulate(In first, In last, T init) {
while (first != last) init = init + *first++
return init
// summation
}
template<class In, class T, class BinOp>
T accumulate(In first, In last, T init, BinOp op) {
while (first != last) init = op(init, *first++)
return init
// general operation
}

An explanation of the meaning of (p++) is given in x3.10. We can use

accumulate() to, for example, calculate the sum or product of elements or
their absolute values of an array. It can also be used to calculate the one,
two, and maximum norms (x4.1) of a vector. A general binary operation
can be passed as a pointer to a function. It can be used as
double SumAbs(double a, double b) { return a + fabs(b) }
int main() {
double ai5]
for (int i = 0 i < 5 i++) aii] = -1
double d = 0
d = accumulate(ai, ai + 5,
//
d = accumulate(ai, ai + 5,
//

d)
sum by default, d = -5
0.0, SumAbs)
sum of absolute values, d = 5

The rst and second arguments of accumulate() must point to the rst and
the last-past-one elements of a sequence, respectively. With the operation
SumAbs(), it actually computes the one-norm of a sequence. It can also
be used to compute, for example, the sum of x-coordinates of an array of
point2d:
struct point2d { double x, y }
double Sumx(double a, point2d p) { return a + p.x }

260

7. Templates

double f(point2d* p, int n) { // sum of x-coordinates

return accumulate(p, p + n, 0.0, Sumx)
}

Note that the second version of the function template accumulate() takes
a binary operation as its fourth argument. The fourth argument can be
either a binary function such as SumAbs() above, or a class object with
the function call operator () overloaded. For example,
template<class T> struct Multiply {
// overload operator()
T operator()(const T& a, const T& b) const { return a*b }
}
int main() {
double ai5]
for (int i = 0 i < 5 i++) aii] = -1
double d =
accumulate(ai, ai + 5, 1.0, Multiply<double>())
}

//d=-1

Notice an object rather than a type is needed for the fourth argument of

accumulate(). Thus Multiply<double> () is used instead of Multiply<

double>. An object of a class with the function call operator () overloaded
is called a function-like object, or a function object. Note that from the
denition of accumulate(), the operator () is passed to its fourth argument
op in such a way that op can be either a function or a function object. More
examples on function objects are presented in x7.7 and x10.2.
Passing a function object can be more easily inlined than a call through
a pointer to a function. In the following example, we measure the times
they take in performing the same arithmetic operations.
template<class T> T SumFcn(T a, T b) { return a + b }
template<class T> struct SumFO {
T operator()(const T& a, const T& b) const {return a + b}
}
int main() {
int n = 10000
double* ai = new double n]
for (int i = 0 i < n i++) aii] = i - 500
double d = 0
time_t tm1 = time(0)

// #include <time.h>

for (int i = 0 i < 10000 i++)

d = accumulate(ai, ai + n, 0.0)

// repeat 10000 times

// use default sum

7.6 Standard Library on Numeric Algorithms

261

time_t tm2 = time(0)

// current time
cout << "time in default summation = "
<< tm2 -tm1 << " seconds" << '\n'
for (int i = 0 i < 10000 i++)
// pass pointer to fcn
d = accumulate(ai, ai + n, 0.0, SumFcn)
time_t tm3 = time(0)
// measure time
cout << "time in passing function = "
<< tm3 -tm2 << " seconds" << '\n'
for (int i = 0 i < 10000 i++)
// pass fcn object
d = accumulate(ai, ai + n, 0.0, SumFO<double>())
time_t tm4 = time(0)
// measure time
cout << "time in passing fcn object = "
<< tm4 -tm3 << " seconds" << '\n'
}

On a Pentium PC running Linux, the running times are 24, 47, and 33
seconds, respectively, while on an SGI PowerChallenge they are 8, 16, and
14 seconds, and on a SUN Ultra2 workstation 10, 32, and 16 seconds. Thus
it can be more ecient to use a function object or a template argument
than passing a function pointer to another function. This implies that the
FORTRAN- and C-style passing functions as arguments to other functions
should be avoided if possible when a large number of such function calls
are needed. However, both of them can be much less ecient than plain
function calls. This suggests that a user may wish to dene his or her own
version of accumulate() and other functions when portability and readability are not as important as eciency. For example,
template<class In, class T>
// user-defined version
T accumulate_multiply(In first, In last, T init) {
while (first != last) init *= *first++
// multiply
return init
}
template<class In, class T>
// user-defined version
T accumulate_abs(In first, In last, T init) {
while (first != last) init += abs(*first++)
return init
// sum of absolute values
}

These user-dened function templates do not require a function object or

pointer to a function passed as an argument and should be more ecient
than the standard accumulate() with a function object or function pointer
as its fourth argument.

262

7. Templates

7.6.2 Inner Products

The function inner product() calculates the inner product or dot product
of two sequences of oating point numbers. It is dened as
template<class In, class In2, class T>
T inner_product(In first, In last, In2 first2, T init) {
while (first != last) init = init + *first++ * *first2++
return init
}
template<class In,class In2,class T,class BinOp,class BinOp2>
T inner_product(In first, In last, In2 first2, T init,
BinOp op, BinOp2, op2) {
while (first != last)
init = op(init, op2(*first++, *first2++))
return init
}

Note that only the beginning of the second sequence is passed as an argument and its size is implicitly assumed to be at least as long as the
rst sequence. Range checking is not provided here. The rst version of
inner product() does not apply to vectors of complex numbers since complex conjugation is needed for the second sequence. For inner products of
complex vectors, the user may wish to dene her own template instead of
using the second version for eciency reasons. For example,
void f(valarray<double>& v1,double* v2,vector<double>& v3){
double d = inner_product(&v10], &v1v1.size()], v2, 0.0)
double e = inner_product(v3.begin(),v3.end(),&v10], 0.0)
}
template<class In, class In2, class T>
// user-defined
T inner_product_cmpx(In first, In last, In2 first2, T init){
while (first != last)
init = init + *first++ * conj(*first2++)
return init
}
void g(valarray<complex<double> >& v1, complex<double>* v2){
complex<double> d = 0
d = inner_product_cmpx(&v10], &v1v1.size()], v2, d)
}

For a vector vr (x10.1.1), the functions vr:begin() and vr:end() point to the
rst and last-past-one elements of vr, respectively. However, a valarray va
does not have functions va:begin() and va:end(). Fortunately, it forms a

7.6 Standard Library on Numeric Algorithms

263

sequence in the sense that va0] va1] : : : vava:size() ; 1] are contiguous

in memory, and thus standard algorithms can be applied.

7.6.3 Partial Sums

Given a sequence a0 a1 : : : an;1 , where n is a positive integer, the function

template partial sum() produces a sequence of partial sums, dened as
a0 a0 + a1 a0 + a1 + a2 : : : a0 + a1 + + an;1 . A general binary operation
instead of summation can also be specied as an extra argument in the
same way as in accumulate(). Its prototype is:
template<class In, class Out>
// summation by default
Out partial_sum(In first, In last, Out res)
template<class In, class Out, class BinOp>
// general op
Out partial_sum(In first, In last, Out res, BinOp op)

It can be used as
void f(valarray<double>& v1, vector<double> v2) {
partial_sum(&v10], &v1v1.size()], &v10])
partial_sum(v2.begin(), v2.end(), v2.begin())
}

If v1 has elements 1 2 5 14, then the rst call above changes it into the
partial sum sequence 1 3 8 22.

7.6.4 Adjacent Dierences

Given a sequence a0 a1 : : : an;1 , where n is a positive integer, the function
adjacent dierence() produces a sequence of adjacent di
erences, dened
as a0 a1 ; a0 a2 ; a0 ; a1 : : : an;1 ; a0 ; a1 ; ; an;2 . A general binary

operation instead of subtraction can also be specied as an extra argument

in the same way as in partial sum() and accumulate(). Its prototype is:
template<class In, class Out>
// subtract by default
Out adjacent_difference(In first, In last, Out res)
template<class In, class Out, class BinOp> // general op
Out adjacent_difference(In first,In last,Out res,BinOp op)

It can be used as
void f(valarray<double>& v1, vector<double> v2) {
adjacent_difference(&v10], &v1v1.size()], &v10])
adjacent_difference(v2.begin(), v2.end(), v2.begin())
}

264

7. Templates

If v1 has elements 1 3 8 22, then the rst call above changes it into the
adjacent di
erence sequence 1 2 5 14. It is easy to see that the algorithms
adjacent dierence() and partial sum() are inverse operations.

7.7 Ecient Techniques for Numeric Integration

This section presents several ecient techniques for numeric integration.
The techniques include using function objects instead of passing function
pointers as arguments to function calls, using function pointers as template
parameters, expression templates, and template metaprograms. These techniques are very general and also apply to many other situations. Two numeric integration methods, the Trapezoidal Rule and Simpson's Rule, are
described in earlier sections. In x3.13 they are coded using traditional procedural programming style while in x5.1 a class is used for encapsulation.
The programs in x3.13 and x5.1 work ne except for run-time eciency,
when they are called frequently. In this section, they are improved for runtime performance and generality by using templates.

7.7.1 Function Object Approach

As pointed out in x7.2 and x7.6, using templates and function objects can
improve run-time eciency for programs that traditionally require passing
pointer-to-functions as arguments. Using a template andRa function object,
the function trapezoidal(), for numerically evaluating ab f (x)dx, can be
written and used as
template<class Fo>
double trapezoidal(double a, double b, Fo f, int n) {
double h = (b - a)/n
// size of each subinterval
double sum = f(a)*0.5
for (int i = 1 i < n i++) sum += f(a + i*h)
sum += f(b)*0.5
return sum*h
}
class integrand {
// define a class for my integrand
public:
double operator()(double x) { return exp(-x*x) }
}
int main() {
cout << trapezoidal(0, 1, integrand(), 100) << '\n'
}

7.7 Ecient Techniques for Numeric Integration

265

Recall that an object of a class with the function call operator () overloaded is called a function object. Thus the object integrand(), instead
of the class name integrand, is passed to the function call trapezoidal().
The compiler will then instantiate trapezoidal(), substitute integrand for
the template parameter Fo, which results in integrand::operator() being inlined into the denition of trapezoidal(). Note that trapezoidal() also accepts
a pointer-to-function as its third argument, but this would be hard to be
inlined and may not improve the run-time.
This technique can be put in a more general form that also takes a
template parameter for the precision of integration:
template<class T, class Fo>
T trapezoidal(T a, T b, Fo f, int n) {
T h = (b - a)/n
// size of each subinterval
T sum = f(a)*0.5
for (int i = 1 i < n i++) sum += f(a + i*h)
sum += f(b)*0.5
return sum*h
}
template<class T> class integrand2 {
public:
// define a class for my integrand
T operator()(T x) { return exp(-x*x) }
}
int main() {
cout << trapezoidal(0.0, 1.0, integrand2<double>(), 100)
}

With this form, integrals such as 01 e;x2 dx can be integrated by the Trapezoidal Rule in double, oat, long double, or other precisions.
R

7.7.2 Function Pointer as Template Parameter

The technique here to improve run-time eciency still uses a function

pointer, but in a very di
erent way, in which the function pointer is used
as a template parameter. Thus it can be inlined to avoid function call
overheads. It can be illustrated by implementing the Trapezoidal Rule:
template<double F(double)>
// F is a template parameter
double trapezoidal(double a, double b, int n) {
double h = (b - a)/n
// size of each subinterval
double sum = F(a)*0.5
for (int i = 1 i < n i++) sum += F(a + i*h)
sum += F(b)*0.5
return sum*h

266

7. Templates

}
double myintegrand(double x) {
return exp(-x*x)
}
int main() {
cout << trapezoidal<myintegrand>(0, 1, 100) << '\n'
}

With this approach, the integrand is still implemented as a function. But

it is not passed as an argument to trapezoidal(). Instead, it is taken as a
template parameter, which will be inlined. It can be generalized to
template<class T, T F(T)>
T trapezoidal(T a, T b, int n) {
T h = (b - a)/n
// size of each subinterval
T sum = F(a)*0.5
for (int i = 1 i < n i++) sum += F(a + i*h)
sum += F(b)*0.5
return sum*h
}
template<class T> T myintegrand2(T x) {
return exp(-x*x)
}
int main() {
cout << trapezoidal<double, myintegrand2>(0,1,100) << '\n'
}

7.7.3 Using Dot Products and Expression Templates

For integrals such as ab f (x) + g(x)]dx and ab f (x)g(x)h(x)dx, the integrand is a product (or summation) of di
erent functions such as f , g, and h,
and the C++ programs presented in previous sections can not be directly
applied, since code such as
R

double result0 = trapezoidal(0, 1, f + g, 100)

double result2 = trapezoidal(0, 1, f*g*h, 100)

is illegal. One way is to write another function that returns the product or
sum of some functions:
double F(double x, double f(double), double g(double)) {
return f(x) + g(x)
}

7.7 Ecient Techniques for Numeric Integration

267

double G(double x, double f(double), double g(double),

double h(double)) {
return f(x)*g(x)*h(x)
}

But the original integration function

template<class Fo>
double trapezoidal(double a, double b, Fo f, int n)

can no longer be applied to F or G (due to the complicated prototypes of

F and G), and thus several versions of it are needed. For a typical nite
element analysis code, the integration of the product (or sum) of many
functions (such as basis functions and their derivatives, coecient functions
and their derivatives) need be evaluated, and this approach would lead
to many overloaded integration functions that are hard to manage and
maintain.
In this subsection, dot products are used to evaluate numeric integration.
As an example, consider a typical Gauss quadrature:
Z

f (x)dx

nX1
i=0

wi f (xi )

(7.1)

where x0 x1 , : : :, xn;1 ] are given Gauss points and w0 w1 , : : :, wn;1 ] are
the corresponding weights see Exercise 5.10.10. If vector wgts contains the
weights wi and vector fxi contains the function values at Gauss points, then
the dot product function can be called to give the approximate integral:
double dot(const Vtr& u, const Vtr& v) {
double dotprod = 0
int n = u.size()
for (int i = 0 i < n i++) dotprod += ui]*vi]
return dotprod
}
Vtr fxi(n)
for (int i = 0 i < n i++) fxii] = f(xi])
double d = dot(fxi, wgts) // dot() evaluates integrals

Here Vtr is a vector class (e.g., the standard library vector<double> or

the class Vcr dened in Chapter 6). When an integrand is the product of
many functions such as f g h, a vector-multiply function can be dened:
Vtr operator*(const Vtr& u, const Vtr& v) {
int n = u.size()
Vtr w(n)
for (int i = 0 i < n i++) wi] = ui]*vi]

268

7. Templates

return w
}
Vtr fghxi = fxi*gxi*hxi
double d0 = dot(fghxi, wgts)
double d2 = dot(fxi, wgts)

// evaluate integral of fgh

// evaluate integral of f

where fxi, gxi, and hxi are the vectors consisting of the function values of
f , g, and h, respectively, evaluated at the Gauss points x0 x1 : : : xn;1 ]

of a given Gauss quadrature. This is a traditional way to evaluate integrals

with integrands as products (or sums or a combination of them) of many
di
erent functions. It works ne and is very convenient to use, but performs
very poorly due to the extra work in computing the loop for fxigxi, storing
the result into a temporary vector tmp, computing another loop to obtain
tmp hxi, and storing it in another temporary vector see x6.5 for similar
situations.
Expression templates can be applied to improve run-time eciency of
such vector multiplications by deferring intermediate operations and avoiding temporary objects. It is similar to the technique discussed in x6.5, but
is more general. First dene a class to hold left operand, operator, and
right operand in an operation such as X Y :
template<class LeftOpd, class Op, class RightOpd>
struct LOR {
LeftOpd fod
// store left operand
RightOpd rod
// store right operand
LOR(LeftOpd p, RightOpd r): fod(p), rod(r) { }
double operator](int i) {
return Op::apply(fodi], rodi])
}
}

Here the template parameter Op refers to a class whose member apply()

denes a specic operation on elements of the left and right operands.
It defers intermediate evaluations and thus avoids temporary objects and
loops. Evaluation of the operation such as X Y takes place only when the
operator ] is called on an object of type LOR.
Then dene a vector class, in which the operator = is overloaded that
forces the evaluation of an expression:
class Vtr {
int lenth
double* vr
public:
Vtr(int n, double* d) { lenth = n vr = d }

7.7 Ecient Techniques for Numeric Integration

269

// assign an expression to a vector and

// do the evaluation
template<class LeftOpd, class Op, class RightOpd>
void operator=(LOR<LeftOpd, Op, RightOpd> exprn) {
for (int i = 0 i < lenth i++) vri] = exprni]
}
double operator](int i) const { return vri] }
}

Notice the function call exprn ] forces the operation Op :: apply() to evaluate elements of the left and right operands.
Next dene a class on how left and right operands are going to be evaluated:
// define multiply operation on elements of vectors
struct Multiply {
static double apply(double a, double b) {
return a*b
}
}

Finally overload the operator for vector multiplication:

template<class LeftOpd>
LOR<LeftOpd, Multiply, Vtr> operator*(LeftOpd a, Vtr b) {
return LOR<LeftOpd, Multiply, Vtr>(a,b)
}

The multiplication is actually deferred. Instead a small object of type LOR

is returned. For example, X Y gives an object of LOR (without actually
multiplying X and Y ), where Y is of type Vtr, but X can be of a more
general type such as Vtr and LOR.
They can now be used as
int main() {
double a] = { 1, 2, 3, 4, 5}
double b] = { 6, 7, 8, 9, 10}
double c] = { 11, 12, 13, 14, 15}
double d5]
Vtr X(5, a), Y(5, b), Z(5, c), W(5, d)
W = X * Y * Z
}

270

7. Templates

When the compiler sees the statement W = X Y Z it constructs the

object LOR<X Multiply Y> from X Y and constructs the object

LOR< LOR<X Multiply Y> Multiply Z>

from LOR<X Multiply Y> Z for the expression X Y Z . When it is
assigned to object W , the operation Vtr::operator= is matched as
W.operator=(LOR< LOR<X, Multiply, Y>, Multiply, Z> exprn) {
for (int i = 0 i < lenth i++) vri] = exprni]
}

where exprni] is then expanded by inlining the LOR :: operator ]:

Multiply::apply(LOR<Vtr,Multiply,Vtr>(X,Y)i], Zi])

which is further expanded into

Multiply::apply(Xi], Yi]) * Zi]
= Xi] * Yi] * Zi].

The nal result of W = X Y Z is:

for (int i = 0 i < lenth i++) Wi] = Xi] * Yi] * Zi]

Thus there are no temporary vector objects or extra loops.

Similarly, vector multiplication V = W X Y Z will be nally expanded
by the compiler into
for (int i = 0 i < lenth i++)
Vi] = Wi] * Xi] * Yi] * Zi]

That is, the expression template technique eciently computes the product
of an arbitrary number of vectors with a single loop and without temporary
vector objects.
For references on expression templates, see Vel95, Fur97]. Some software libraries using expression templates to achieve run-time eciency are
PETE Pet], POOMA Poo], and Blitz++ Bli]. Besides, the Web address
http://www.oonumerics.org contains a lot of useful information on objectoriented numeric computing.

7.7.4 Using Dot Products and Template Metaprograms

In x7.7.3, numeric integration is evaluated through dot products, where

vector multiplication can be eciently computed by using expression templates. A typical numeric integration formula (see Exercise 5.10.11 and
x7.7.3) has very few terms in the summation, which implies that dot products of small vectors are usually evaluated. A dot product function is normally written using a loop:

7.7 Ecient Techniques for Numeric Integration

271

double dot(const Vtr& u, const Vtr& v) {

double dotprod = 0
int n = u.size()
for (int i = 0 i < n i++) dotprod += ui]*vi]
return dotprod
}

where the size of the vectors may not have to be known at compile-time.
For vectors of a small size (say 4), an alternative denition:
inline double dot4(const Vtr& u, const Vtr& v) {
return u0]*v0] + u1]*v0] + u2]*v2] + u3]*v3]
}

can boost performance, since it removes loop overhead and low-level parallelism can be easily done on the summation operation. Besides, it is easier
to make this small function to be inlined (reducing function-call overhead)
and its data to be registerized. All these factors can make dot4() much
more ecient than the more general dot() for vectors of size 4.
To improve performance, template metaprograms can be used so that
the dot products of small vectors are expanded by the compiler in the
form of dot4(). To illustrate the idea of template metaprograms, consider
the Fibonacci numbers dened by a recursive relation: f0 = 0 f1 = 1,
fn = fn;1 + fn;2 , for n = 2 3 : : :. They can be recursively computed by
the compiler:
template<int N> struct fib{
enum { value = fib<N-1>::value + fib<N-2>::value}
}
template<> struct fib<1>{
enum { value = 1 }
}
template<> struct fib<0>{
enum { value = 0 }
}
const int f3 = fib<3>::value
const int f9 = fib<9>::value

The compiler can generate that f 3 = 2 and f 9 = 34. The recursive template behaves like a recursive function call (but at compile-time) and the
specializations fib<1> and fib<0> stop the recursion.
This recursive template technique can be generalized to compute dot
products. First dene a class for small vectors:
template<int N, class T>

272

7. Templates

class smallVtr {
T vrN]
// array of size N and type T
public:
T& operator](int i) { return vri] }
}

Then dene a dot product function that calls a template metaprogram:

template<int N, class T>
inline T dot(smallVtr<N,T>& u, smallVtr<N,T>& v) {
return metaDot<N>::f(u,v)
}

where the template metaprogram metaDot can be dened as

template<int M> struct metaDot {
template<int N, class T>
static T f(smallVtr<N,T>& u, smallVtr<N,T>& v) {
return uM-1]*vM-1] + metaDot<M - 1>::f(u,v)
}
}
template<> struct metaDot<1> {
template<int N, class T>
static T f(smallVtr<N,T>& u, smallVtr<N,T>& v) {
return u0]*v0]
}
}

The specialization metaDot<1> stops the recursion.

Now dot can be called on small vectors:
int main() {
smallVtr<4, float> u, v
for (int i = 0 i < 4 i++) {
ui] = i + 1
vi] = (i + 2)/3.0
}
double d = dot(u,v)
cout << " dot = " << d << '\n'
}

where the inlined function call dot(u v) is expanded during compilation as

dot(u,v)
= metaDot<4>::f(u,v)
= u3]*v3] + metaDot<3>::f(u,v)
= u3]*v3] + u2]*v2] + metaDot<2>::f(u,v)
= u3]*v3] + u2]*v2] + u1]*v1] + metaDot<1>::f(u,v)

7.8 Polynomial Interpolation

273

= u3]v3] + u2]v2] + u1]v1] + u0]v0]

It is this compile-time expanded code that gives better performance than

a normal function call with a loop inside.
This template metaprogram technique could result in code that is several
times faster than a normal function call when computing dot products of
small vectors.
In large-scale applications, it is not enough to compute correctly, but
also eciently and elegantly. The elegance of a program should also include extendibility, maintainability, and readability, not just the look of
the program. The code for dot products of small vectors using the template metaprogram technique is also elegant in the sense that it can be
easily extended to other situations, for example, dot products of complex
vectors, and can be maintained easily. Writing it in FORTRAN or C styles
would possibly result in tens of versions (dot products of vectors of sizes 2,
3, 4, 5, and so on, and for real and complex small vectors in di
erent precisions). The Trapezoidal Rule itself is also written in ecient and elegant
ways in this section and many other functions such as Simpson's Rule can
be as well see Exercise 7.9.11.

7.8 Polynomial Interpolation

Given a set of n +1 values fyi gni=0 of a function f (x) at n +1 distinct points
fxi gni=0 , where n is a positive integer, how can one approximate f (x) for x 6=
xi i = 0 1 : : : n? Such a problem arises in many applications in which the
n +1 pairs fxi yi g may be obtained from direct measurement, experimental
observations, or an expensive computation, but function values f (x) at
other points are needed. This section deals with polynomial interpolation,
in which a polynomial pn (x) of degree n is constructed to approximate f (x)
that satises the conditions f (xi ) = pn (xi ) i = 0 1 : : : n.

7.8.1 Lagrange Form

The approximating polynomial pn (x) can be written in the Lagrange form:

pn (x) = y0 L(0n)(x) + y1 L(1n)(x) + + yn L(nn) (x) =

n
X
i=0

yi L(in) (x)

(n)

where each Li (x) is a polynomial of degree n and depends only on the

interpolation nodes fxi gni=0 , but not on the given function values fyi gni=0 .
They are dened as

L(in)(x) =

n
Y

=0
j =i
j

x ; xj
xi ; xj

i = 0 1 : : : n ; 1:

274

7. Templates

That is, each L(in) (x) is the product of all factors (x ; xj )=(xi ; xj ) for
j 6= i. They have the following property.

L(in) (xj ) = 0 for j 6= i and L(in)(xi ) = 1:

Thus pn (x) agrees with f (x) at the given interpolation nodes fxi gni=0 . Function value f (x) for x other than the interpolation nodes can be approximated by pn (x).
When n = 1, the interpolation polynomial p1 (x) is a line passing through
(x0 y0 ) and (x1 y1 ), and has the form:
x ; x1
x ; x0
(1)
p1 (x) = L(1)
0 (x)y0 + L1 (x)y1 =
x0 ; x1 y0 + x1 ; x0 y1 :
This is the so-called linear interpolation, which is used frequently in many
scientic works. When n = 2, p2 (x) is a quadratic polynomial passing
through (x0 y0 ), (x1 y1), and (x2 y2 ), and has the form:

p2 (x) =

2
X

i=0

L(2)
i (x)yi

x0 )(x ; x2 ) y + (x ; x0 )(x ; x1 ) y :
; x1 )(x ; x2 )
y0 + (x(x ;
= (x(x ;
x
)(
x
;
x
)
;
x0 )(x1 ; x2 ) 1 (x2 ; x0 )(x2 ; x1 ) 2
0
1
0
2
1
Since some applications require more accuracy than others, a template
function can be written so that a user can conveniently choose single, double, long double, or other precisions:
template<class T>
T lagrange(const vector<T>& vx, const vector<T>& vy, T x) {
int n = vx.size() - 1
T y = 0
for (int i = 0 i <= n i++) {
T temp = 1
for (int j = 0 j <= n j++)
if (j != i) temp *= (x - vxj])/(vxi] - vxj])
y += temp*vyi]
}
return y
}

The vectors vx and vy store, respectively, the given x and y coordinates of

the interpolation points. The function template computes the approximate
function value at x. In particular, the inner
for loop calculates L(in)(x) and
Pn
the outer for loop gives the summation i=0 yi L(in) (x).
As an example, consider the interpolation problem that gives xi = 1 +
i=4:0 and yi = ex for i = 0 1 2 3. The function template can then be used
to nd an approximate function value at x = 1:4:
i

7.8 Polynomial Interpolation

275

int main() {
const int n = 4
vector<float> px(n)
vector<float> py(n)
for (int i = 0 i < n i++) {
pxi] = 1 + i/4.0 pyi] = exp( pxi] )
}
float x = 1.4
float approximation = lagrange(px, py, x)
}

Compared to the exact function value e1:4, the error of this approximation
is 0:0003497 (on my machine).

7.8.2 Newton Form

The Newton form is more ecient and seeks pn (x) as

pn (x) = c0 + c1 (x ; x0 )+ c2(x ; x0 )(x ; x1 )+ + cn (x ; x0 ) (x ; xn;1 )

where c0 , c1 , : : :, cn are constants to be determined. For example, the linear
and quadratic interpolation polynomials in Newton's form are

p1 (x) = c0 + c1 (x ; x0 )
p2 (x) = c0 + c1 (x ; x0 ) + c2 (x ; x0 )(x ; x1 ):
Before presenting an algorithm on how to nd ci i = 0 1 : : : n, we talk
about how Newton's form can be evaluated eciently, assuming all the ci
are known. Write pn (x) into a nested multiplication form:
pn (x) = c0 + d0 (c1 + d1 (c2 + + dn;3 (cn;2 + dn;2 (cn;1 + dn;1 (cn ))) ))
where d0 = x ; x0 d1 = x ; x1 : : : dn;1 = x ; xn;1 , and introduce the
notation:

un = cn
un;1 = cn;1 + dn;1 un
un;2 = cn;2 + dn;2 un;1
..
.
u1 = c1 + d1 u2
u0 = c0 + d0 u1

Then pn (x) = u0 . This procedure can be written into an algorithm:

u
cn

276

7. Templates

for (i = n ; 1 n ; 2 : : : 0) f
u
ci + di u

== that is: u
ci + (x ; xi ) u

return u
This is the so-called Horner's algorithm (see x3.12), which is very ecient
in evaluating polynomials.
Below is a procedure on how to nd the coecients ci . The requirement
pn (xi ) = f (xi ) leads to

f (x0 ) = pn (x0 ) = c0
f (x1 ) = pn (x1 ) = c0 + c1 (x1 ; x0 )
f (x2 ) = pn (x2 ) = c0 + c1 (x2 ; x0 ) + c2 (x2 ; x0 )(x2 ; x1 )
..
.

Thus c0 = f (x0 ) and c1 = (f (x1 ) ; f (x0 ))=(x1 ; x0 ). Dene the divided

di
erences of order 0 as

f xi ] = f (xi )

i = 0 1 : : : n

and of order 1 as

; f xi ]
f xi xj ] = f xxj ] ;
x
j

i j = 0 1 : : : n and i 6= j

and of order j as

f xi xi+1 : : : xi+j ]

= f xi+1 xi+2 : : : xix+j ] ;;fxxi xi+1 : : : xi+j;1 ] i j = 0 1 : : : n:
i+j
i
Then c0 = f x0 ], c1 = f x0 x1 ], and
c2 = f (x2 )(x; f;(xx0 ))(;x c1;(xx2 ); x0 )
2
0
2
1
f (x2 ) ; f (x1 ) + f (x1 ) ; f (x0 ) ; c1 (x2 ; x0 )
x2 ; x1
= x2 ; x1
x2 ; x0
f x0 x1 ]
= f x1 xx2 ] ;
2 ; x0
= f x0 x1 x2 ]:
Similarly, it can be shown that ci = f x0 x1 : : : xi ] for i = 0 1 : : : n.
That is, all the coecients ci can be expressed as divided di
erences and

7.8 Polynomial Interpolation

x0
x1
x2
x3

277

f x0 ] f x0 x1 ] f x0 x1 x2 ] f x0 x1 x2 x3 ]

f x1 ] f x1 x2 ] f x1 x2 x3 ]
f x2 ] f x2 x3 ]
f x3 ]

TABLE 7.1. A table of divided di erences in the case of n = 3.

the interpolation polynomial pn (x) is:

pn (x) = f x0 ] + f x0 x1 ](x ; x0 ) + f x0 x1 x2 ](x ; x0 )(x ; x1 ) +
+f x0 x1 : : : xn ](x ; x0 )(x ; x1 ) (x ; xn;1 ):
If a table of function values (xi f (xi )) is given, then a table of divided
di
erences can be constructed recursively column by column as indicated in
Table 7.1 for the case of n = 3. In particular, the rst two columns list the
given x and y coordinates of the interpolation points, divided di
erences
of order 1 are computed and put in the third column, divided di
erences
of order 2 are computed and put in the fourth column, and so on. When
the table is completed, the divided di
erences in the top row will be used
to compute pn (x).
For example, given a table of interpolation nodes and corresponding function values:

x
3 1 5 6
f (x) 1 ;3 2 4
the table of divided di
erences can be computed according to Table 7.1 as
3 1 2
;3=8 7/40
1 ;3 5/4 3/20
5 2 2
6 4
Then the interpolation polynomial can be written as
7 (x ; 3)(x ; 1)(x ; 5):
p3 (x) = 1 + 2(x ; 3) ; 38 (x ; 3)(x ; 1) + 40
When n is large, calculating the divided di
erence table can not be done
by hand. To derive an ecient algorithm for doing this, introduce the notation cij = f xi xi+1 : : : xi+j ]. Then Table 7.1 of divided di
erences
becomes (in the case of n = 4):

278

7. Templates

x0
x1
x2
x3
x4

c00
c10
c20
c30
c40

c01 c02 c03 c04

c11 c12 c13
c21 c22
c31

This actually leads to the following algorithm:

for (k = 0 1 : : : n) f
ck0
f (xk )
g

for (j = 1 2 : : : n) f
for (i = 0 1 : : : n ; j ) f
cij
(ci+1j;1 ; cij;1 )=(xi+j ; xi )
g

Then pn (x) can be obtained as

pn (x) = c00 + c01 (x ; x0 ) + c02 (x ; x0 )(x ; x1 ) +
+c0n(x ; x0 )(x ; x1 ) (x ; xn;1 ):
A closer look at the algorithm reveals that only c00 c01 : : : c0n are
needed in constructing pn (x), and other cij for i 6= 0 are just intermediate
results. A space-ecient version of the algorithm requires only an array
b = b0 b1 : : : bn] of size n + 1. First store the given function value f (xi )
into bi , for i = 0 1 : : : n. Note that b0 = c00 . Then compute divided
di
erences ci1 for i = 0 1 : : : n ; 1 and store them in b1 b2 : : : bn. Now
b0 = c00 and b1 = c01 are the rst two desired coecients for pn (x).
Next compute divided di
erences ci2 for i = 0 1 : : :n ; 2 and store them
in b2 b3 : : : bn. Now b0 = c00 , b1 = c01 , and b2 = c02 . Continue this
process until c0n is computed and stored in bn . Now the revised and more
space-ecient algorithm for calculating the coecients of pn (x) reads:
for (k = 0 1 : : : n) f
bk
f (xk )
g

for (j = 1 2 : : : n) f
for (i = n n ; 1 : : : j ) f
bi
(bi ; bi;1 )=(xi ; xi;j )
g

7.9 Exercises

279

After this algorithm computes all bk , the interpolation polynomial pn(x)

can be nally constructed in Newton's form:

pn (x) = b0 + b1 (x ; x0 ) + b2(x ; x0 )(x ; x1 ) +

+bn(x ; x0 )(x ; x1 ) (x ; xn;1 ):
Combining this algorithm and the one for nested multiplication leads
to the following code for computing Newton's form of the interpolation
polynomial pn(x).
template<class T>
T newton(const vector<T>& vx, const vector<T>& vy, T x) {
vector<T> b = vy
int n = vx.size() - 1
// find coefficients in Newton's form
for (int j = 1 j <= n j++)
for (int i = n i >= j i--)
bi] = (bi] - bi-1])/(vxi] - vxi-j])
// evaluate interpolation polynomial at x
T u = bn]
for (int i = n - 1 i >= 0 i--) u = bi] + (x - vxi])*u
return u
}

Applying this program to the example presented at the end of x7.8.1 produces about the same result as lagrange().
The advantage of Newton's form is that it is very ecient and the coefcients ci i = 0 1 : : : n, can be used when later there are more interpolation points available. For example, if later one more point (xn+1 f (xn+1 ))
is given, then pn+1 (x) = pn (x) + cn+1 (x ; x0 ) (x ; xn ) gives cn+1 =
(f (xn+1 ) ; pn (xn+1 ))=((xn+1 ; x0 ) (xn+1 ; xn )).

7.9 Exercises
7.9.1. Test the class template for vectors Vcr<T> presented in x7.1 and
add more operations such as vector multiplication and addition, and
2-norm (see x6.3).
7.9.2. Test the class template for vectors Vcr<T> together with its specializations Vcr<complex<T> > and Vcr<complex<double> > presented in x7.1. Also add more operations such as vector multiplication
and addition, and 2-norm (see x6.3).

280

7. Templates

7.9.3. Turn the matrix class Mtx dened in x6.3 into a class template.
7.9.4. Traits can be used as an alternative to certain template specializations
or function overloading. For example, the maxnorm() function for
arrays of real or complex numbers can be dened as
template<class T> struct RTrait {
typedef T RType // for return type of maxnorm()
}
template<class T> struct RTrait< complex<T> > {
typedef T RType // return type for complex arrays
}
// a specialization of RTrait<T>
template<class T>
// definition of maxnorm
typename RTrait<T>::RType maxnorm(T u], int n) {
typename RTrait<T>::RType nm = 0
for (int i = 0 i < n i++) nm = max(nm, abs(ui]))
return nm
}

7.9.5.
7.9.6.

7.9.7.
7.9.8.

Here the trait class RTrait denes the return type for the function
maxnorm() but for complex vectors, the return value of a norm function must be a real number. The keyword typename must be used to
instruct the compiler that RTrait<T>:: RType is a type name. One
denition of the function handles real and complex vectors. Rewrite
the vector class Vcr<T > in x7.1 using this approach for the member function maxnorm(). This approach would be much better if
maxnorm() were a large function. Can it be applied to vector dot
products?
Apply the sort function template sort<T>() dened in x7.2 to sort
a vector of complex numbers in decreasing order according to their
absolute values, by overloading the operator < for complex numbers.
Apply the sort function template sort<T>() dened in x7.2.3 to sort
a vector of complex numbers in decreasing order according to their
absolute values, by using a specialization for the comparison template
less<T>().
Write a specialization for the template less<T>() so that the sort
function template sort<T>() in x7.2.3 will sort a vector of point2d
in decreasing order according to the y-coordinate.
Apply function template compare<T>(), dened in x7.2.4, to compare two vectors of char with case-sensitive and case-insensitive comparisons, respectively.

7.9 Exercises

281

7.9.9. Write a function object to be used as an argument of the function

template accumulate(), presented in x7.6, for calculating the 2-norm
of a vector of complex numbers. Write a function template, similar to
accumulate(), that calculates directly the 2-norm of a vector of complex numbers. Compare the eciency of the two function templates.
7.9.10. Write the deferred-evaluation operator overloading for the vector
saxpy operation as presented in x6.5 into a template so that it can
handle vectors of single, double, and long double precisions.
7.9.11. Write Simpson's Rule for numeric integration into the more ecient
and elegant forms as described in x7.7.
7.9.12. Write a template metaprogram to compute the factorial of small integers (e.g, less than 12 see Exercise 1.6.2) during compilation.
7.9.13. Use expression templates to write a program to add an arbitrary
number of vectors without introducing temporary vector objects or
extra loops.
7.9.14. Write the conjugate gradient algorithm (x6.6) into a member function
of a class template so that it can be used to solve real linear systems
Ax = b for symmetric and positive denite matrix A and vectors b
and x in single, double, and long double precisions.
7.9.15. Write the conjugate gradient algorithm (x6.6) into a member function
of a class template with specializations for vector dot products so that
it can be used to solve real and complex linear systems Ax = b for
Hermitian and positive denite matrix A and vectors b and x in single,
double, and long double precisions for real and imaginary parts.
7.9.16. Given a set of values of function f (x):

x
5 7
6
6.6
f (x) 1 ;23 ;50 ;4
apply Lagrange and Newton forms of the interpolation polynomial to
approximate f (x) for x = 5:5 and x = 6:2.

282

7. Templates

This is page 283

Printer: Opaque this

Class Inheritance

A new class can be derived from an existing class. The new class, called
derived class, then inherits all members of the existing class, called base
class, and may provide additional data, functions, or other members. This
relationship is called inheritance. The derived class can then behave as a
subtype of the base class and an object of the derived class can be assigned
to a variable of the base class through pointers and references. This is the
essential part of what is commonly called object-oriented programming.
This chapter deals with various issues in class inheritance. The disadvantage
of object-oriented programming is that it may sometimes greatly a
ect runtime performance. The last section of the chapter presents techniques on
what can be done when performance is a big concern.

8.1 Derived Classes

Suppose we want to dene classes for points in one, two, and three dimensions. A point in 1D can be dened as
class Pt {
private:
double x
public:
Pt(double a = 0) { x = a }
void draw() const { cout << x }
}

// class for 1D point

// x coordinate
// constructor
// draw the point

284

8. Class Inheritance

The function draw() is dened to simply print out the coordinate. Since
a 2D point has both x and y coordinates, it can be dened to inherit
the x coordinate from the class Pt and provide a new member for the y
coordinate:
class Pt2d: public Pt {
// class for 2D point
private:
// inherits x from Pt
double y
// y coordinate
public:
Pt2d(double a = 0, double b = 0): Pt(a), y(b) { }
void draw() const {
// draw x coordinate
Pt::draw()
// by Pt's draw()
cout << " " << y
// draw y coordinate
}
}

By putting :public Pt after the name of the class Pt2d when it is declared,
the new class Pt2d is dened to be a derived class from the base class Pt:
The derived class Pt2d is also called a subclass and the base class Pt a
superclass. Pt2d has members of class Pt in addition to its own members.
The derived class Pt2d is often said to inherit properties from its base
class Pt: This relationship is called inheritance: It can be represented
graphically by an arrow from the derived class to the base:

Pt2d
Now a Pt2d is also a Pt (a two-dimensional point can be regarded as
a one-dimensional point by just looking at the x coordinate), and Pt2d
(pointer to Pt2d) can be used as Pt : However, a Pt is not necessarily a
Pt2d and a Pt can not be used as a Pt2d : The keyword public in the
denition of Pt2d means that the derived class Pt2d has a public base Pt,
which in turn means that a Pt2d can be assigned to a variable of type
Pt by any function without an explicit type conversion. (Inheritance using
private and protected bases is discussed in x8.3.) The opposite conversion,
from a pointer to base class to a pointer to derived class, must be explicit.
For example,
void f(Pt p1, Pt2d p2) {
Pt* q1 = &p2
Pt2d* q2 = &p1
}

// OK. Every Pt2d is a Pt

// error, not every Pt is a Pt2d

8.1 Derived Classes

285

void g(Pt p1, Pt2d p2) {

// statements below are legal
Pt* q1 = &p2
// Every Pt2d is a Pt
Pt2d* q2 = static_cast<Pt2d*>(q1)
// explicit type conversion
q2->draw()

// x, y of p2 are printed

Pt2d* r2 = static_cast<Pt2d*>(&p1)
// explicit type conversion
r2->draw()

// OK, but y may be garbage

}
int main() {
Pt a(5)
Pt2d b(4, 9)
g(a, b)

// test what is printed out

// create a 1D point
// create a 2D point
// call g() on a and b

During the second draw in calling g(), the 1D point a does not have a
y-coordinate and thus some garbage should be printed out. The rst draw
in g() on the 2D point b should print out both coordinates of b correctly.
An object of a derived class can be treated as an object of its base
class when manipulated through pointers and references. The opposite must
be through explicit type conversion using static cast (at compile-time) or
dynamic cast (at run-time see x8.5), although the result of such conversions can not be guaranteed in general. For example, the second cast in the
function g() above tries to convert a pointer to a base class to a pointer to
a derived class and may result in an object with an undened y coordinate.
The operator static cast converts between related types such as from one
pointer type to another or from an enumeration to an integral type. There
is another cast, called reinterpret cast, which converts between unrelated
types such as an integer to a pointer. These casts may not be portable, can
be even dangerous, but are sometimes necessary. They should be avoided
when possible.
Note that the function draw() may draw garbage for the y coordinate
in r2; >draw() above. A closer look reveals that r2 is cast from a onedimensional point. It would be nice if the system could detect the type of an
object and invoke the draw function Pt :: draw() for one-dimensional points
and invoke Pt2d :: draw() for two-dimensional points. In x8.1.5, virtual
functions are introduced so that for a 1D point, the function Pt :: draw() will
be invoked and for a 2D point, Pt2d :: draw() will be invoked automatically.

286

8. Class Inheritance

8.1.1 Member Functions

Member functions of a derived class can not access the private part of a
base class, although a derived class contains all members of the base class.
For example, the function Pt2d :: draw() can not be dened as
void Pt2d::draw() const { // incorrect
cout << x << " " << y // x is not accessible
}

A member of a derived class that is inherited from the private part of a

base class may be called an invisible member of the derived class. Thus an
invisible member of a class can not be accessed by members or friends even
of the same class. This is for information hiding, since otherwise a private
member of a class could be accessed freely just by deriving a new class from
it.
However, the hidden member x of the derived class Pt2d is accessible
through the public member function Pt :: draw() :
void Pt2d::draw() const { // draw the point
Pt::draw()
// draw x coordinate by Pt's draw()
cout << " " << y
// draw y coordinate
}

Note the class qualier Pt :: must be used to refer to a member function of

class Pt:

8.1.2 Constructors and Destructors

In dening the constructor of a derived class, the constructor of the base

class must be called in the initializer list (like class object members see
x5.6) if the base class has a constructor. For example,
Pt2d::Pt2d(double a = 0, double b = 0): Pt(a), y(b) { }

It can also be dened alternatively as

Pt2d::Pt2d(double a = 0, double b = 0): Pt(a) { y = b }

But it can not be dened as

Pt2d::Pt2d(double a = 0, double b = 0) { Pt(a) y = b }

The third denition is wrong since the construction of base class Pt is not
put in the initializer list.
A derived class constructor can specify initializers for its own members
and immediate bases only. The construction of an object of a derived class
starts from the base class, then the members, and then the derived class
itself. Its destruction is in the opposite order: rst the derived class itself,

8.1 Derived Classes

287

then its members, and then the base class. Members and bases are constructed in order of declarations in the derived class and destructed in the
reverse order.
In the denition of the destructor of a derived class, only spaces allocated
in the derived class need be explicitly deallocated spaces allocated in the
base class are freed implicitly by the destructor of the base class.

8.1.3 Copying

Copying of class objects is dened by the copy constructor and assignment.

For example,
class Pt {
public:
// ... in addition to other members
Pt(const Pt& p) { x = p.x }
// copy constructor
Pt& operator=(const Pt& p) {
// copy assignment
if (this != &p) x = p.x
return *this
}
}
Pt2d p2
Pt p1 = p2
p1 = p2

// construct p1 from Pt part of p2

// assign Pt part of p2 to p1

Since the copy functions of Pt do not know anything about Pt2d only the
Pt part (x-coordinate) of a Pt2d object is copied and other parts are lost.
This is commonly called slicing: It can be avoided by passing pointers and
references see x8.1.5.

8.1.4 Class Hierarchy

A derived class can be a base class of another derived class. For example,
class Pt3d: public Pt2d {
// point in 3D
private:
// inherits x,y from Pt2d
double z
// z coordinate
public:
Pt3d(double a = 0, double b = 0, double c = 0)
: Pt2d(a, b), z(c) { }
void draw() const {
// draw the point
Pt2d::draw()
// draw x,y by Pt2d's draw()
cout << " " << z
// draw z coordinate
}
}

288

8. Class Inheritance

A set of related classes derived through inheritance is called a class hierarchy.

A derived class can inherit from two or more base classes. This is called
multiple inheritance see x8.4.

8.1.5 Virtual Functions

Virtual functions can be declared in a base class and may be redened

(also called overridden) in each derived class when necessary. They have
the same name and same set of argument types in both base class and
derived class, but perform di
erent operations. When they are called, the
system can guarantee the correct function be invoked according to the type
of the object at run-time. For example, the class Pt can be redened to
contain a virtual function:
class Pt {
private:
double x
public:
Pt(double a = 0) { x = a }
virtual void draw() const { cout << x }
}

// virtual fcn

The member Pt :: draw() is declared to be a virtual function. The declarations of Pt2d and Pt3d remain unchanged. The keyword virtual indicates
that draw() can act as an interface to the draw() function dened in this
class and the draw() functions dened in its derived classes Pt2d and Pt3d.
When draw() is called, the compiler will generate code that chooses the
right draw() for a given Pt object (Pt2d and Pt3d objects are also Pt
objects). For example, if a Pt object is actually a Pt3d object, the function
Pt3d :: draw() will be called automatically by the system.
A virtual function can be used even if no class is derived from the class,
and a derived class need not redene it if the base class version of the
virtual function works ne. Now functions can be dened to print out a set
of points (some may be 1D or 2D while others may be 3D):
void h(const vector<Pt*>& v) {
for (int i = 0 i < v.size() i++) {
vi]->draw()
cout << '\n'
}
}
int main() {
Pt a(5)
Pt2d b(4, 9)

// include <vector>

// test what is printed out

8.1 Derived Classes

289

Pt3d c(7, 7, 7)

vector<Pt*> v(3)
v0] = &a
v1] = &b
v2] = &c
h(v)
}

For library <vector> see x10.1.1. This will work even if the function h() was
written and compiled before the derived classes Pt2d and Pt3d were even
conceived of, and the points (1D, 2D, or 3D) in the argument of h() can
be generated dynamically at run-time by another function. Since draw() is
declared to be virtual in the base class, the system guarantees that in the
call vi];>draw() the function Pt :: draw() is invoked if vi] is a pointer to
Pt, Pt2d :: draw() is invoked if vi] is a pointer to Pt2d, and Pt3d :: draw() is
invoked if vi] is a pointer to Pt3d: This is a key aspect of class inheritance.
When used properly, it is a cornerstone of object-oriented design.
Getting the right correspondence between the function (e.g., among different versions of draw()) and the type (e.g., among Pt, Pt2d, or Pt3d)
of an object is called run-time polymorphism: A type with virtual functions is called a polymorphic type: Run-time polymorphism is also called
dynamic binding: To get polymorphic behavior, member functions called
must be virtual and objects must be manipulated through pointers or references. When manipulating objects directly (rather than through pointers
or references), their exact types must be known at compile-time, for which
run-time polymorphism is not needed and templates may be used. See x8.6
for examples that can be implemented using either templates or virtual
functions. In contrast, what templates provide is often called compile-time
polymorphism, or static polymorphism.
When run-time polymorphism is not needed, the scope resolution operator :: should be used, as in Pt :: draw() and Pt2d :: draw(): When a virtual
function is inline, function calls that do not need run-time polymorphism
can be specied by using the scope resolution operator :: so that inline substitution may be made. This is reected in the denitions of Pt2d :: draw()
and Pt3d :: draw():
Pointers to class members were discussed in x5.8. A function invoked
through a pointer to a member function can be a virtual function, and
polymorphic behavior can be achieved through pointers to virtual member
functions.

8.1.6 Virtual Destructors

A virtual destructor is a destructor that is also a virtual function. Proper

cleanup of data can be ensured by dening a virtual destructor in a base

290

8. Class Inheritance

class and overriding it in derived classes. For example, consider the following
inheritance without a virtual destructor.
class B {
double* pd
public:
B() {
pd = new double 20]
cout << "20 doubles allocated\n"
}
~B() {
delete] pd
cout << "20 doubles deleted\n"
}
}
class D: public B {
int* pi
public:
D(): B() {
pi = new int 1000]
cout << "1000 ints allocated\n"
}
~D() {
delete] pi
cout << "1000 ints deleted\n"
}
}

// constructor of B

// destructor of B

// derive D from B

// constructor of D

// destructor of D

Then the code

int main() {
B* p = new D
// `new' constructs a D object
delete p
// `delete' frees a B object since p is a pointer to B
}

will produce the output

20 doubles allocated
1000 ints allocated
20 doubles deleted

Proper cleanup is not achieved here. The reason is that the new operator
constructed an object of type D (allocated 20 doubles and 1000 ints when
D's constructor was called), but the delete operator cleaned up an object of
type B pointed to by p (freed 20 doubles when B's destructor was implicitly

8.2 Abstract Classes

291

called). In other words, run-time polymorphism did not apply to the destructor of class D. If it were applied, the system would have detected the
type of the object pointed to by p and called the destructor of the correct
type (class D).
This problem can be easily solved by declaring the destructor of the base
class B to be virtual :
class B {
double* pd
public:
B() {
pd = new double 20]
cout << "20 doubles allocated\n"
}
virtual ~B() {
// a virtual destructor
delete] pd
cout << "20 doubles deleted\n"
}
}

The denition of the derived class D is unchanged. Now the main program
above produces the following desired output.
20 doubles allocated
1000 ints allocated
1000 ints deleted
20 doubles deleted

From this, the orders of construction and destruction of objects of derived

and base classes can also be clearly seen see x8.1.2.
In general, a class with a virtual function should have a virtual destructor,
because run-time polymorphism is expected for such a class. However, other
base classes such as class B above may also need a virtual destructor for
proper cleanups.

8.2 Abstract Classes

Some virtual functions in a base class can only be declared but can not be
dened, since the base class may not have enough information to do so and
such virtual functions are only meant to provide a common interface for
the derived classes. A virtual function is called a pure virtual function if
it is declared but its denition is not provided. The initializer \= 0" makes
a virtual function a pure one. For example,
class Shape {
public:

292

8. Class Inheritance

virtual
virtual
virtual
virtual
}

void draw() = 0
void rotate(int i) = 0
bool is_closed() = 0
~Shape() { }

//
//
//
//

pure virtual function

rotate i degrees
pure virtual function
empty virtual destructor

The member functions rotate() draw(), and is closed() are pure virtual
functions recognized by the initializer \= 0". Note the destructor~Shape()
is not a pure virtual function since it is dened, although it does nothing.
This is an example of a function that does nothing but is useful and necessary. The class Shape is meant to be a base class from which other classes
such as Triangle and Circle can be derived. Its member functions rotate()
draw() and is closed() can not be dened since there is not enough information to do so.
A class with one or more pure virtual functions is called an abstract class:
No objects of an abstract class can be created. For example, it is illegal to
dene:
Shape s

// object of an abstract class can not be defined

An abstract class can only be used as an interface and as a base for derived
classes. For example, a special shape: Circle can be derived from it as
class Circle: public Shape {
// derived class for circles
private:
Pt2d center
// center of circle
double radius
// radius of circle
public:
Circle(Pt2d, double)
// constructor
void rotate(int) { }
// override Shape::rotate()
void draw()
// override Shape::draw()
bool is_closed() { return true }
// a circle is closed
}

The member functions of Circle can be dened since there is enough information to do so. These functions override the corresponding functions
of Shape: The function Circle :: rotate() does nothing since a circle does
not change when being rotated, and Circle :: is closed() returns true since
a circle is a closed curve. Circle is not an abstract class since it does not
contain pure virtual functions, and consequently objects of Circle can be
created.
A pure virtual function that is not dened in a derived class remains a
pure virtual function. Such a derived class is also an abstract class. For
example,
class Polygon: public Shape {
// abstract class
public:
bool is_closed() { return true }

8.2 Abstract Classes

293

// override Shape::is_closed()
// but draw() & rotate() still remain undefined.
}
class Triangle: public Polygon { // not abstract any more
private:
Pt2d* vertices
public:
Triangle(Pt2d*)
~Triangle() { delete] vertices }
void rotate(int)
// override Shape::rotate()
void draw()
// override Shape::draw()
}

Class Polygon remains an abstract class since it has pure virtual functions draw() and rotate(), which are inherited from its base class Shape
but have not been dened. Thus objects of Polygon can not be created.
However, Triangle is no longer an abstract class since it does not contain
any pure virtual functions (denitions of its members rotate() and draw()
are omitted here) and objects of Triangle can be created.
Note that a virtual destructor is dened for the base class Shape. This
will ensure proper cleanup for the derived class Triangle which frees dynamic memory space in its destructor. In general, a class with a virtual
function should have a virtual destructor, although other base classes may
also need one see the example in x8.1.6.
Below is a design problem for iterative numeric methods on solving linear
systems of algebraic equations Ax = b where A is a square matrix, b is the
right-hand side vector, and x is the unknown vector. In x6.6 the conjugate
gradient (CG) method is implemented for a Hermitian and positive denite
matrix A and in x11.3 the generalized minimum residual (GMRES) method
is introduced for any nonsingular matrix A: These two iterative methods
require only matrix-vector multiplication and some vector operations in
order to solve the linear system Ax = b: Details of CG and GMRES are
not needed here. Instead, a class hierarchy is designed to apply CG and
GMRES to full, band, and sparse matrices A: In some applications, most
entries of the matrix A are not zero and all entries of A are stored. Such
a matrix storage format is called a full matrix. In some other applications,
the entries of A are zero outside a band along the main diagonal of A: Only
entries within the band (zero or nonzero) may be stored to save memory
and such a format is called a band matrix. Yet in other applications most
entries of A are zero and only nonzero entries are stored to further save
memory such a format is called a sparse matrix. Three di
erent classes
need be dened for these three matrix storage formats, of which the full
matrix format is given in x6.3 where every entry of a matrix is stored.

294

8. Class Inheritance

Details of these matrix storage formats are not needed here, although they
can be found in x11.1.
The objective of such a design is that the CG and GMRES methods
are dened only once, but are good for all three matrix storage formats,
instead of providing one denition for each storage format. This should
also be extendible possibly by other users later to other matrix formats
such as symmetric sparse, band, and full matrices (only half of the entries
need be stored for symmetric matrices to save memory). Since CG and
GMRES depend on a matrix-vector product, which must be done di
erently
for each matrix storage format, the correct matrix-vector multiplication
function has to be called for a particular matrix format. To provide only
one denition for CG and GMRES, they can be dened for a base class
and inherited for derived classes representing di
erent matrix formats. To
ensure the correct binding of matrix-vector multiplication to each matrix
storage format, a virtual function can be utilized. To be more specic,
dene a base class
class AbsMatrix {
public:
virtual ~AbsMatrix() { }

// base class
// a virtual destructor

virtual Vtr operator*(const Vtr &) const = 0

// matrix vector multiply
int CG()
int GMRES()
}
int AbsMatrix::CG() { /* ... definition omitted here */ }
int AbsMatrix::GMRES() { /* ... definition omitted here */ }

where Vtr is a vector class as dened in Chapter 6, which can also be

thought of as the standard vector<double> here. This matrix class serves
as an interface for all derived classes. Since there is not enough information
to dene the matrix-vector multiplication operator , it is declared as a
pure virtual function, which will be overridden in the derived classes. For
such a class, it is natural to dene a virtual destructor to ensure that
objects of derived classes are properly cleaned up, although this virtual
destructor does nothing. The class AbsMatrix is actually an abstract class
so that no objects of AbsMatrix can be created. With the matrix-vector
multiply operator (although it is not dened yet), the functions CG()
and GMRES () can be fully dened, which can be inherited and used in
derived classes once the matrix-vector multiply operator is dened. For
simplicity, the arguments of CG() and GMRES () and their denitions
are not given here (nor are they needed since only the design problem is
considered in this section).
Derived classes for full, band, and sparse matrices can now be dened:

8.2 Abstract Classes

295

class FullMatrix: public AbsMatrix {

public:
// ... in addition to other members
Vtr operator*(const Vtr&) const
// multiply a full matrix with a vector
}
class SparseMatrix: public AbsMatrix {
public:
// ... in addition to other members
Vtr operator*(const Vtr&) const
// multiply a sparse matrix with a vector
}
class BandMatrix: public AbsMatrix {
public:
// ... in addition to other members
Vtr operator*(const Vtr&) const
// multiply a band matrix with a vector
}

The matrix-vector multiply operator can be dened (but omitted here) for
each of the derived classes and thus CG() and GMRES () can be invoked
for objects of FullMatrix, BandMatrix, and SparseMatrix: For example,
void f(FullMatrix&
fm.CG()
bm.CG()
sm.CG()
}

fm, BandMatrix&
// call CG() on
// call CG() on
// call CG() on

bm, SparseMatrix& sm) {

FullMatrix fm
BandMatrix bm
SparseMatrix sm

void g(vector<AbsMatrix*>& vm) {

for (int i = 0 i < vm.size() i++) {
vmi]->GMRES()
// the object pointed to by vmi] could be FullMatrix,
// BandMatrix, or SparseMatrix. Its exact type may not
// be known at compile-time.
}
}

During such an invocation, the matrix-vector multiply operator inside

CG() and GMRES () is guaranteed to be called correctly according to
the type of the objects, since the operator is declared to be a virtual
function in the base class. Note that in the call vmi];>GMRES () vmi]
may be a pointer to FullMatrix, BandMatrix, or SparseMatrix. This
technique can be combined with templates so that one denition of CG()
and GMRES () can be applied to di
erent matrix storage formats with
di
erent precisions for real and complex matrices. This is the main subject
of Chapter 11.

296

8. Class Inheritance

8.3 Access Control

8.3.1 Access to Members

With derived classes, the keyword protected is introduced. A member (it

can be a function, type, constant, etc., as well as a data member) of a class
can be declared to be private, protected, or public.
A private member can be used only by members and friends of the
class in which it is declared,
A protected member can be used by members and friends of the class
in which it is declared. Furthermore, it becomes a private, protected,
or public member of classes derived from this class as a member of
a derived class, it can be used by members and friends of the derived
class and maybe by other members in the class hierarchy depending
on the form of inheritance. See x8.3.2 for more details.
A public member can be used freely in the program. It also becomes a
private, protected, or public member of classes derived from this class,
depending on the form of inheritance. See x8.3.2 for more details.
For example,
class B {
private:
int i
protected:
float f
public:
double d
void g1(B& b) { f = b.f }
}
void g()
B bb
bb.i =
bb.f =
bb.d =
}

// a private member
// a protected member
// a public member
// f and b.f are accessible

{
5
3.14
2.71

//
//
//
//

by default constructor
error, can not access bb.i
error, can not access bb.f
bb.d is accessible freely

The speciers protected and private mean the same when inheritance is
not involved.
Protected members of a class are designed for use by derived classes and
are not intended for general use. They provide another layer of information
hiding, similar to private members. The protected part of a class usually
provides operations for use in the derived classes, and normally does not
contain data members since it can be more easily accessed or abused than

8.3 Access Control

297

the private part. For this reason the data member x in the class Pt in x8.1
is better kept private than protected. Despite this, the derived classes Pt2d
and Pt3d can still print out the value of x.

8.3.2 Access to Base Classes

A base class can be declared private protected or public in dening a

derived class. For example,
class
class
class
class

B { /* ... */ }
X: public B { /* ... */ }
Y: protected B { /* ... */ }
Z: private B { /* ... */ }

//
//
//
//

a
B
B
B

class
is a public base
is a protected base
is a private base

Public derivation makes the derived class a subtype of its base class and
is the most common form of derivation. Protected and private derivations
are used to represent implementation details. No matter what form the
derivation is, a derived class contains all the members of a base class (some
of which may have been overridden), although members inherited from the
private part of the base class are not directly accessible even by members
and friends of the derived class (they are called invisible members of the
derived class in x8.1.1). The form of derivation controls access to the derived
class's members inherited from the base class and controls conversion of
pointers and references from the derived class to the base class. They are
dened as follows. Suppose that class D is derived from class B:
If B is a private base, its public and protected members become
private members of D: Only members and friends of D can convert
a D to a B :
If B is a protected base, its public and protected members become
protected members of D: Only members and friends of D and members and friends of classes derived from D can convert a D to a

B:
If B is a public base, its public members become public members of
D and its protected members become protected members of D: Any
function can convert a D to a B : In this case, D is called a subtype
of B .

The following example illustrates the use of protected members and public derivation.
class B {
private:
int i
protected:
float f

298

8. Class Inheritance

public:
double d
void g1(B b) { f = b.f }
}
class X: public B {
protected:
short s
public:
void g2(X b) { f = b.f }
}

// f and b.f are accessible

// by a member of B
// public derivation

// f and b.f are accessible

// by members of X

Since X has a public base B it has three public members: d and g1()
(inherited from B ), and g2() (its own), two protected members: f (inherited
from B ) and s (its own), but no private members. X also inherits a member
i from base B but this member is invisible in X: The protected member
f of class B and member f of the object b of type B are accessible in the
member function B :: g1(B b): Similarly, the protected (inherited from base
class) member f of class X and member f of the object b of type X are
accessible in the member function X :: g2(X b):
However, it is illegal to dene a member function g3() of X this way:
class X: public B {
protected:
short s
public:
void g2(X b) { f = b.f }
void g3(B b) { f = b.f }
}

// public derivation

// f and b.f are accessible

// error: b.f not accessible

In the denition of X :: g3(B b), the protected member f of object b of type

B is not accessible. Member b:f of object b of type B is only accessible by
members and friends of class B: Note the only di
erence between X :: g2(X )
and X :: g3(B ) is the type of their arguments.
A derived class inherits members (e.g., B :: f above) that are protected
and public in its base class. These inherited members (e.g., X :: f ) of the

derived class are accessible by members and friends of the derived class.
However, the derived class can not access protected members (e.g., b:f in
the denition of X :: g3(B )) of an object of the base class, just as it can not
access protected members of objects of any other class.
No matter what type of inheritance it is, a private member of a base class
can not be used in derived classes to achieve a strong sense of information
hiding, since otherwise a private member of any class could be accessed
and possibly abused easily by dening a derived class from it.

8.3 Access Control

299

Access to an inherited member may be adjusted to that of the base class

by declaring it public or protected in the derived class using the scope
resolution operator :: as in the following example.
class B {
private:
int i
protected:
float f
public:
double d
}
class Y: protected B {
protected:
short s
public:
B::d
}

// protected derivation

// B::d is adjusted to be public

Now the member d of Y inherited from the protected base B is public in

Y instead of protected by default. Note that the private member i of B
can not be adjusted in Y since it is not visible in Y:
If a derived class adds a member with the same name as a member of a
base class, the new member in the derived class hides the inherited member
from the base class. The inherited member may be accessible using the
scope resolution operator preceded by the base class name. For example,
class B {
public:
double d
}
class Z: private B {
private:
double d
public:
void f(Z& z) {
z.d = 3.1415926
z.B::d = 2.718
}
}

// private derivation
// this d hides B::d

// Z::d instead of B::d

// z's member B::d

The derivation specier for a base class can be left out. In this case, the
base defaults to a private base for a class and a public base for a struct:
class

XX: B { /* ... */ }

// B is a private base

300

8. Class Inheritance

struct YY: B { /* ... */ }

// B is a public base

8.4 Multiple Inheritance

A class can be directly derived from two or more base classes. This is
called multiple inheritance: In contrast, derivation from only one direct
base class is called single inheritance: This section talks about multiple
inheritance. Consider a class Circle in Triangle (circle inscribed in a triangle) derived from classes Circle and Triangle :
class Circle {
// ...
public:
virtual void draw()
virtual void add_border()
double area()
void dilate(double d)
}
class Triangle {
// ...
public:
virtual void draw()
double area()
void refine()
}

//
//
//
//

draw the circle

add border to circle
find its area
enlarge circle d times

// draw the triangle

// find its area
// refine into 4 triangles

class Circle_in_Triangle: public Circle, public Triangle {

// ...
// two bases
public:
void draw() {
// override Circle::draw()
//
and Triangle::draw()
Circle::draw()
// call draw() of Circle
Triangle::draw()
// call draw() of Triangle
}
void add_border()
// override Circle::add_border()
}

The derived class Circle in Triangle inherits properties from both Circle
and Triangle: It can be used as
void f(Circle_in_Triangle& ct) {
ct.dilate(5.0)
// call Circle::dilate()
ct.refine()
// call Triangle::refine()
ct.draw()
// call Circle_in_Triangle::draw()

8.4 Multiple Inheritance

301

}
double curvature(Circle*)
vector<double> angles(Triangle*)

// curvature of a curve
// angles of a triangle

void g(Circle_in_Triangle* pct) {

double c = curvature(pct)
vector<double> a = angles(pct)
}

Since public derivation is used for both base classes, any function can convert a Circle in Triangle to Circle or Triangle:

8.4.1 Ambiguity Resolution

When two base classes have members with the same name, they can be
resolved by using the scope resolution operator. For example, both Circle
and Triangle have a function named area(): They must be referred to with
the class names:
void h(Circle_in_Triangle* pct) {
double ac = pct->Circle::area()
double at = pct->Triangle::area()
double aa = pct->area()
pct->draw()
}

//
//
//
//

OK
OK
ambiguous, error
OK

Overload resolution is not applied across di

erent types. In particular,
ambiguities between functions from di
erent base classes are not resolved
based on argument types. A using-declaration can bring di
erent functions
from base classes to a derived class and then overload resolution can be
applied. For example,
class A {
public:
int g(int)
float g(float)
}
class B {
public:
char g(char)
long g(long)
}
class C: public A, public B {
public:

302

8. Class Inheritance

using A::g
using B::g
char g(char)
C g(C)
}
void h(C& c) {
c.g(c)
c.g(1)
c.g(1L)
c.g('E')
c.g(2.0)
}

// bring g() from A

// bring g() from B
// it hides B::g(char)

//
//
//
//
//

C::g(C) is called
A::g(int) is called
B::g(long) is called
C::g(char) is called
A::g(float) is called

8.4.2 Replicated Base Classes

With the possibility of derivation from two bases, a class can be a base
twice. For example, if both Circle and Triangle are derived from a class
Shape then the base class Shape will be inherited twice in the class
Circle in Triangle:
class Shape {
protected:
int color
public:
// ...
virtual void draw() = 0
}
class Circle: public Shape {
// ...
public:
void draw()
}
class Triangle: public Shape {
// ...
public:
void draw()
}

// color of shape

// a pure virtual function

// inherit color for Circle

// draw circle

// color for Triangle

// draw the triangle

class Circle_in_Triangle: public Circle, public Triangle {

// ...
// two bases
public:
void draw()
// override Circle::draw()

8.4 Multiple Inheritance

}

303

and Triangle::draw()

This causes no problems if the colors of a Circle and a Triangle for an

object of Circle in Triangle can be di
erent. Indeed, two copies of Shape
are needed to store the colors. This class hierarchy can be represented as

Shape

Circle

Triangle

@
I
;
@@;;

Circle in Triangle
To refer to members of a replicated base class, the scope resolution operator must be used. For example,
void Circle_in_Triangle::draw() {
int cc = Circle::color
// or Circle::Shape::color
int ct = Triangle::color
// or Triangle::Shape::color
// ...
Circle::draw()
Triangle::draw()
}

A virtual function of a replicated base class can be overridden by a

single function in a derived class. For example, Circle in Triangle :: draw()
overrides Shape :: draw() from the two copies of Shape:

8.4.3 Virtual Base Classes

Often a base class need not be replicated. That is, only one copy of a
replicated class need be inherited for a derived class object. This can be
done by specifying the base to be virtual: Every virtual base of a derived
class is represented by the same (shared) object. For example, if the circle
and triangle in Circle in Triangle must have the same color, then only
one copy of Shape is needed for storing the color information. The base
class Shape then should be declared to be virtual :
class Shape {
int color
public:
// ...
virtual void draw() = 0
}

// color of shape

// pure virtual function

304

8. Class Inheritance

class Circle: public virtual Shape { // virtual base Shape

// ...
// inherit color
public:
//
for Circle
void draw()
// draw circle
}
class Triangle: public virtual Shape { // virtual base Shape
// ...
// inherit color
public:
//
for Triangle
void draw()
// draw triangle
}
class Circle_in_Triangle: public Circle, public Triangle {
// ...
// two bases
public:
void draw()
// override Circle::draw()
}
//
and Triangle::draw()

This class hierarchy can be represented as

Shape

;; @I@
;
@
Circle
Triangle
@I

@ ;;
@;
Circle in Triangle
Compare this diagram with the inheritance graph in x8.4.2 to see the
di
erence between ordinary inheritance and virtual inheritance. In an inheritance graph, every base class that is specied to be virtual will be
represented by a single object of that class. The language ensures that a
constructor of a virtual base class is called exactly once.

8.4.4 Access Control in Multiple Inheritance

A name or a base class is called accessible if it can be reached through

any path in a multiple inheritance graph, although often it can be reached
through several paths. For example,
class B {
public:
double d

8.5 Run-Time Type Information

305

static double sdm

}
class D1: public virtual B { /* ... */ }
class D2: public virtual B { /* ... */ }
class D12: protected D1, public D2 { /* ... */ }
D12* pd12 = new D12
B* pb = pd12
double m = pd12->d

// accessible through public base D2

The member pt2d; >d is accessible publicly since D12 has a public base
D2 and d is a public member of D2. It would be inaccessible if both D1
and D2 were protected bases. However, ambiguities might arise when a
single entity was accessible through more than one path. Again, the scope
resolution operator can be used to resolve such ambiguities. For example,
class X1: public B { /* ... */ }
class X2: public B { /* ... */ }
class X12: protected X1, private X2 {
void f()
}
void X12::f() {
X12* p = new X12
// ... assign some value to *p
double i = p->d
// illegal, ambiguous
// X12::X1::B::d or X12::X2::B::d?
double j = p->X1::B::d
double k = p->X2::B::d

// OK
// OK

double n = p->sdm

// OK
// only one B::sdm in an X12 object

There are two copies of member d in an object of X 12, but there is only
one static member sdm in X 12:

8.5 Run-Time Type Information

Due to run-time polymorphism and assignment of pointers or references
from one type to another, the type of an object may be lost. Recovering
the run-time type information (or RTTI for short) of an object requires the

306

8. Class Inheritance

system to examine the object to reveal its type. This section introduces two
mechanisms that can be used to recover the run-time type of an object.

8.5.1 The dynamic cast Mechanism

The dynamic cast mechanism can be used to handle cases in which the
correctness of a conversion can not be determined by the compiler. Suppose
p is a pointer. Then
dynamic_cast<T*>(p)

examines the object pointed to by p at run-time. If this object is of class T

or has a unique base class of type T then it returns a pointer of type T
to the object otherwise it returns 0. If the value of p is 0 (null pointer),
then dynamic cast<T > (p) returns 0: For example,
class A { /* ... */ }
class B { /* ... */ }
class C: public A, protected B { /* ... */ }
void f(C* p) {
A* q1 = p
A* q2 = dynamic_cast<C*>(p)

// OK
// OK

B* p1 =p
// error, B is protected base
B* p2 =dynamic_cast<C*>(p) // OK, but 0 is assigned to p2
}
void g(B* pb) {
// assume A, B have virtual functions
if (A* pa = dynamic_cast<A*>(pb)) {
// if pb points to an object of type A, do something
} else {
// if pb does not point to an A object, do something else
}
}

Since B is a protected base of C , function f () can not directly convert a

pointer to C into a pointer to B . Although using dynamic cast makes it
legal, the null pointer is returned. In g(), converting B into A requires
that A and B have virtual functions so that a variable of A or B may be
of another type, for example, C .
Casting from a derived class to a base class is called an upcast because
of the convention of drawing an inheritance graph growing from the base
class downwards. Similarly, a cast from a base class to a derived class is
called a downcast: A cast from a base to a sibling class (like the cast from
B to A above) is called a crosscast: A dynamic cast requires a pointer or
reference to a polymorphic type to do a downcast or crosscast. The result of

8.5 Run-Time Type Information

307

the return value of a dynamic cast of a pointer should always be explicitly

tested, as in function g() above. If the result is 0, it means a failure in the
conversion indicated.
For a reference r, then
dynamic_cast<T&>(r)

tests to see if the object referred to by r is of type T . When it is not,

it throws a bad cast exception. To test a successful or failed cast using
references, a suitable handler should be provided see Chapter 9 for details.
For example,
void h(B* pb, B& rb) { // assume A, B have virtual functions
A* pa = dynamic_cast<A*>(pb)
if (pa) {
// when $pa$ is not null pointer
// if pb points to an A object, do something
} else {
// when $pa$ is null pointer
// if pb does not point to an A object, do something else
}
A& ra = dynamic_cast<A&>(rb)

// rb refers to an A object?

}
int main() {
try{
A* pa = new A
B* pb = new B
h(pa, *pa)
h(pb, *pb)
} catch (bad_cast) {
// exception handler
// dynamic_cast<A&> failed, do something.
}
}

A dynamic cast can cast from a polymorphic virtual base class to a

derived class or a sibling class, but a static cast can not since it does not
examine the object from which it casts. However, dynamic cast can not cast
from a void (but static cast can) since nothing can be assumed about the
memory pointed to by a void : Both dynamic cast and static cast can not
cast away const which requires a const cast: For example,
void cast(const A* p, A* q) {
q = const_cast<A*>(p)
// OK
q = static_cast<A*>(p)
// error, can not cast away const
q = dynamic_cast<A*>(p)
// error, can not cast away const
}

308

8. Class Inheritance

8.5.2 The typeid Mechanism

The dynamic cast operator is enough for most problems that need to know
the type of an object at run-time. However, situations occasionally arise
when the exact type of an object needs to be known. The typeid operator
yields an object representing the type of its operand it returns a reference
to a standard library class called type info dened in header <typeinfo>.
The operator typeid() can take a type name or an expression as its operand
and returns a reference to a type info object representing the type name or
the type of the object denoted by the expression. If the value of a pointer
or a reference operand is 0, then typeid() throws a bad typeid exception.
For example,
void f(Circle* p, Triangle& r, Circle_in_Triangle& ct) {
typeid(r)
// type of object referred to by r
typeid(*p)
// type of object pointed to by p
if (typeid(ct) == typeid(Triangle) { /* ... */ }
if (typeid(r) != typeid(*p) { /* ... */ }
cout << typeid(r).name()

// print out type name

The function type info::name() returns a character string representing

the type name. Equality and inequality of objects of type info can be compared. The implementation-independent part of the class type info, presented in the standard header <typeinfo>, has the form:
class type_info {
private:
// prevent copy initialization and assignment
type_info(const type_info&)
type_info& operator=(const type_info&)
public:
virtual ~typeid()
//polymorphic type
bool operator==(const type_info&) const // can be compared
bool operator!=(const type_info&) const // can be compared
bool before(const type_info&) const
// can be sorted
const char* name() const
// name of type
}

There can be more than one type info object for each type in a system.
Thus comparisons for equality and inequality should be applied to type info
objects instead of their pointers or references.

8.6 Replacing Virtual Functions by Static Polymorphism

309

8.5.3 Run-Time Overhead

Using run-time type information (dynamic cast or typeid) involves overhead since the system examines the type of an object. It should and can
be avoided in many cases in which virtual functions suce. Use virtual
functions rather than dynamic cast or typeid if possible since virtual functions normally introduce less run-time overhead. Even a virtual function
can cause a lot of performance penalty when it is small (i.e., it does not
perform a lot of operations) and called frequently (e.g., inside a large loop).
When a virtual function contains a lot of code (i.e., performs a lot of computation) or is not called frequently, the overhead of virtual function dispatch
will be insignicant compared to the overall computation.
In large-scale computations, static (compile-time) checking may be preferred where applicable, since it is safer and imposes less overhead than
run-time polymorphism. On one hand, run-time polymorphism is the cornerstone of object-oriented programming. On the other hand, too much
run-time type checking can also be a disadvantage due to the run-time
overhead it imposes, especially in performance-sensitive computations. In
x8.6, some techniques are described to replace certain virtual functions by
using templates.

8.6 Replacing Virtual Functions by Static

Polymorphism
Virtual functions provide a nice design technique in software engineering.
However, virtual functions may slow down a program when they are called
frequently, especially when they do not contain a lot of instructions. In this
section, two techniques are presented to replace certain virtual functions
by static polymorphism, in order to improve run-time eciency.
The idea of the rst technique BN94] is to dene a base template class
and dene functions operating on this base class. When dening a derived
class from the base template class, the template parameter of this base class
is taken to be the type of the derived class. Then the functions dened for
the base class can be used for derived classes as well.
Consider the example in which a function CG() is dened for a base
matrix class but may be called on derived classes for full, band, and sparse
matrices. A virtual function is used in x8.2 to dene matrix-vector multiplication, upon which the function CG() depends. Then this virtual function
is overridden in derived classes for full, band, and sparse matrices, which
enables one version of CG() to be called for di
erent derived classes. Here
the same goal can be achieved without using a virtual function. Rather, a
matrix-vector multiplication is dened for a base template class that refers
to the template parameter, which will be a derived class. The following
code illustrates this idea.

310

8. Class Inheritance

class Vtr {

/* a vector class */

}

template<class T> class Matrix {

// base class
const T& ReferToDerived() const {
return static_cast<const T&>(*this)
}
public:
// refer operation to derived class
Vtr operator*(const Vtr& v) const {
return ReferToDerived()*v
}
// define common functionality in base class that can
// be called on derived classes
void CG(const Vtr& v) { // define CG here
Vtr w = (*this)*v
// call matrix vector multiply
cout << "calling CG\n"
}
}
class FullMatrix: public Matrix<FullMatrix> {
//
encapsulate storage information for full matrix
public:
Vtr operator*(const Vtr& v) const {
// define (full) matrix vector multiply here
cout << "calling full matrix vector multiply\n"
}
}
class BandMatrix: public Matrix<BandMatrix> {
//
encapsulate storage information for band matrix
public:
Vtr operator*(const Vtr& v) const {
// define (band) matrix vector multiply here
cout << "calling band matrix vector multiply\n"
}
}
class SparseMatrix: public Matrix<SparseMatrix> {
//
encapsulate storage information for sparse matrix
public:
Vtr operator*(const Vtr& v) const {
// define (sparse) matrix vector multiply here
cout << "calling sparse matrix vector multiply\n"

8.6 Replacing Virtual Functions by Static Polymorphism

311

}
}
void f(const Vtr& v) {
FullMatrix A
A.CG(v)
// Calling CG() on full matrix
BandMatrix B
B.CG(v)

// Calling CG() on band matrix

SparseMatrix S
S.CG(v)

// Calling CG() on sparse matrix

Although the code above compiles and runs, it has been greatly simplied and full denitions of CG() and matrix-vector multiply functions are
not given. The function CG() is dened for the base class Matrix which
depends on the matrix-vector multiply operator : The operator is syntactically dened for the base Matrix, but actually refers to the operator
belonging to the template parameter T: When T is instantiated by, for
example, FullMatrix in the denition of the class FullMatrix this operator now is the FullMatrix-vector multiply operator, which can be fully
dened according to its matrix structure.
In this approach, the type of an object is known at compile-time and thus
there is no need for virtual function dispatch. One version of the function
CG() can be called for di
erent classes, which is achieved without using
virtual functions.
A simple but more concrete and complete example using this technique
is now given. First dene a function sum() for a base template matrix that
adds all entries of a matrix. Then dene a derived class for full matrices
that stores all entries of a matrix and another derived class for symmetric
matrices that stores only the lower triangular part of a matrix to save
memory. The goal is to give only one version of sum() that can be called
for full and symmetric matrices, without using virtual functions. This can
be achieved as
template<class T> class Mtx {
// base matrix
private:
// refer to derived class
T& ReferToDerived() {
return static_cast<T&>(*this)
}
// entry() uses features of derived class
double& entry(int i, int j) {
return ReferToDerived()(i,j)

312

8. Class Inheritance

}
protected:
int dimn
// dimension of matrix
public:
// define common functionality in base class that can
// be called on derived classes
double sum() {
// sum all entries
double d = 0
for (int i = 0 i < dimn i++)
for (int j = 0 j < dimn j++) d += entry(i,j)
return d
}
}
class FullMtx: public Mtx<FullMtx> {
double** mx
public:
FullMtx(int n) {
dimn = n
mx = new double* dimn]
for (int i=0 i<dimn i++) mxi] = new double dimn]
for (int i=0 i<dimn i++)
for (int j=0 j<dimn j++)
mxi]j] = 0
// initialization
}
double& operator()(int i, int j) { return mxi]j] }
}
class SymmetricMtx: public Mtx<SymmetricMtx> {
// store only lower triangular part to save memory
double** mx
public:
SymmetricMtx(int n) {
dimn = n
mx = new double* dimn]
for (int i=0 i<dimn i++) mxi] = new double i+1]
for (int i=0 i<dimn i++)
for (int j = 0 j <= i j++)
mxi]j] = 0
// initialization
}
double& operator()(int i, int j) {
if (i >= j ) return mxi]j]
else return mxj]i]
// due to symmetry
}
}

8.6 Replacing Virtual Functions by Static Polymorphism

313

void g() {
FullMtx A(2)
A(0,0) = 5 A(0,1) = 3 A(1,0) = 3 A(1,1) = 6
cout << "sum of full matrix A = " << A.sum() << '\n'
SymmetricMtx S(2)
// just assign lower triangular part
S(0,0) = 5 S(1,0) = 3 S(1,1) = 6
cout << "sum of symmetric matrix S = " << S.sum() << '\n'
}

Here the member function sum() depends on entry(i j ) that refers to an

entry at row i and column j of a matrix through a function call operator
(): But entry() is not dened for FullMtx and SymmetricMtx and the
function call operator () is not dened for the base class Mtx:
The second technique in this section does not use inheritance at all, but
only uses templates. It is much simpler in many situations. Taking the
matrix solver CG() as an example, this technique denes CG() as a global
template function whose template parameter can be any type that denes
a multiply operator with a vector:
class Vtr {

/* a vector class */

}

template<class M>
// define CG as a global function
void CG(const M& m, const Vtr& v) {
// define CG here for matrix m and vector v
Vtr w = m*v
// call matrix vector multiply
cout << "calling CG\n"
}
class FullMatrix {
//
encapsulate storage information for full matrix
public:
Vtr operator*(const Vtr& v) const {
// define (full) matrix vector multiply here
cout << "calling full matrix vector multiply\n"
}
}
class BandMatrix {
//
encapsulate storage information for band matrix
public:
Vtr operator*(const Vtr& v) const {
// define (band) matrix vector multiply here
cout << "calling band matrix vector multiply\n"
}

314

8. Class Inheritance

}
class SparseMatrix {
//
encapsulate storage information for sparse matrix
public:
Vtr operator*(const Vtr& v) const {
// define (sparse) matrix vector multiply here
cout << "calling sparse matrix vector multiply\n"
}
}
void ff(const Vtr& v) {
FullMatrix A
CG(A, v)
// Calling CG() on full matrix
BandMatrix B
CG(B, v)

// Calling CG() on band matrix

SparseMatrix S
CG(S, v)

// Calling CG() on sparse matrix

If this technique is applied to the second example above, where the

entries of a matrix are sum()ed, each of the classes for FullMtx and
SymmetricMtx will need to dene a function entry():
template<class M> double sum(const M& m) {
double d = 0
for (int i = 0 i < m.dimn i++)
for (int j = 0 j < m.dimn j++) d += m.entry(i,j)
return d
}
class FullMtx {
double** mx
public:
FullMtx(int n) {
dimn = n
mx = new double* dimn]
for (int i=0 i<dimn i++) mxi] = new double dimn]
for (int i=0 i<dimn i++)
for (int j=0 j<dimn j++)
mxi]j] = 0
// initialization
}
int dimn
double& operator()(int i, int j) { return mxi]j] }

8.6 Replacing Virtual Functions by Static Polymorphism

315

double entry(int i, int j) const {

return const_cast<FullMtx&>(*this)(i,j)
}
}
class SymmetricMtx {
// store only lower triangular part to save memory
double** mx
public:
SymmetricMtx(int n) {
dimn = n
mx = new double* dimn]
for (int i=0 i<dimn i++) mxi] = new double i+1]
for (int i=0 i<dimn i++)
for (int j = 0 j <= i j++)
mxi]j] = 0
// initialization
}
int dimn
double& operator()(int i, int j) {
if (i >= j ) return mxi]j]
else return mxj]i]
// due to symmetry
}
double entry(int i, int j) const {
return const_cast<SymmetricMtx&>(*this)(i,j)
}
}
void gg() {
FullMtx A(2)
A(0,0) = 5 A(0,1) = 3 A(1,0) = 3 A(1,1) = 6
cout << "sum of full matrix A = " << sum(A) << '\n'
SymmetricMtx S(2)
// just assign lower triangular part
S(0,0) = 5 S(1,0) = 3 S(1,1) = 6
cout << "sum of symmetric matrix S = " << sum(S) << '\n'
}

When function entry() is large and accesses local information of FullMtx

and SymmetricMtx, this may cause a lot of code duplication. In this case,
the rst technique might be better. This example is only for illustrating
the idea if just for summing entries of a matrix, the function entry() may
not be needed.
Using virtual functions as shown in x8.2, the derived classes FullMatrix
BandMatrix and SparseMatrix may not have to be known when the
base class AbsMatrix is compiled. That is, if one wants to add one more

316

8. Class Inheritance

class SymmetricMatrix to this class hierarchy, the code for the base class
AbsMatrix need not be recompiled. However, using the techniques here
without virtual functions, the complete type of all classes must be known at
compile-time. Thus, in this aspect, there may be some design disadvantage
to static polymorphism. Besides, not all virtual functions can be replaced
by templates.

8.7 Exercises
8.7.1. Implement the class hierarchy for points in one, two, and three dimensions, as outlined in x8.1. Add more functions such as moving a point,
nding the equation of the line passing through two points, and evaluating the distance between two points. Create a vector of (pointers
to) points and print them out using run-time polymorphism.
8.7.2. Implement the class hierarchy for points in one, two, and three dimensions, as outlined in x8.1, using pointers for their data members.
For example, declare a pointer to double for the x-coordinate in Pt
and for the y-coordinate in Pt2d: Consequently, user-dened copy
constructor and copy assignment need be dened for them, and a
virtual destructor is needed for the base class Pt:
8.7.3. When should a class have a virtual destructor?
8.7.4. Rewrite the base class B and derived class D in x8.1.6 so that B will
allocate space for d doubles in its constructor and D for i integers,
where d and i must be passed as arguments to the constructors.
8.7.5. Summarize all C++ language features that support polymorphism
(in compile- and run-times), and discuss their advantages and disadvantages.
8.7.6. Dene an abstract class called Shape containing a pure virtual function distance() that takes a two-dimensional point as argument and
is meant to nd the distance between the point and Shape. Derive
three classes for Triangle, Circle and Rectangle respectively, and
override the virtual function distance() in each of the derived classes.
The distance between a point and a gure is dened to be the minimum distance between the given point and any point in the gure.
Create a vector of pointers to Shape that actually point to objects of
Triangle, Circle or Rectangle and create a vector of points. Call
distance() to nd the distance between each point to each of the
Shape objects.
8.7.7. Test the class hierarchy for full, band, and sparse matrices as outlined
in x8.2 in which one denition of CG() and GMRES () is given for

8.7 Exercises

317

three matrix storage formats. Instead of writing code implementing

each function, you may just let, for example, CG() call a matrixvector multiply operator and print out \Solving linear system using
CG," band-matrix vector multiply operator print out \multiplying
band matrix with vector," and sparse-matrix vector multiply operator
print out \multiplying sparse matrix with vector." This is enough to
know if the right functions are called. Then create a vector of pointers to AbsMatrix which actually point to objects of FullMatrix,
BandMatrix, or SparseMatrix and call CG() and GMRES () on
each of the matrix objects. Check the printouts to see if dynamic
binding is working correctly.
8.7.8. Dene an abstract class called Shape containing a pure virtual function inside() that takes a point as argument and is meant to check
if Shape contains the point in a two-dimensional Euclidean space.
Derive three classes for Triangle, Circle and Rectangle respectively, and override the virtual function inside() in each of the derived
classes. Create a vector of pointers to Shape that actually point to objects of Triangle, Circle or Rectangle and create a vector of points.
Call function inside() to see which point is inside which gure.
Here is a hint on how to check if a point p(x y) is contained inside
a triangle with vertices p0 (x0 y0) p1 (x1 y1 ) and p2 (x2 y2): Let q in
boldface be a vector from the origin to point q for any point q, and
pipj be a vector from point pi to point pj : Then p can be represented
as p = p0 + p0 p1 +
p0 p2 : Applying vector cross-product with p0 p1
to its both sides yields p p0 p1 = (p0 + p0 p1 +
p0 p2 ) p0 p1 =
p0 p0p1 +
p0p2 p0p1: That is, (p ; p0) p0p1 =
p0p2 p0p1:
Using the denition of vector cross-products gives
x0 )(y1 ; y0 ) ; (x1 ; x0 )(y ; y0 ) :

= (x(x ;
;
x0 )(y1 ; y0 ) ; (x1 ; x0 )(y2 ; y0 )
2
Now check to see if points p and p2 are on the same side of p0p1 by
checking if
> 0: Similarly, check to see if p and p1 are on the same
side of p0 p2 and if p and p0 are on the same side of p1 p2: These
three conditions are true if and only if p is inside the triangle.
8.7.9. Generalize Exercise 8.7.8 to three-dimensional Euclidean spaces with
three derived classes for tetrahedron, ball and cube, respectively.
The idea of checking to see if a point is inside a triangle given in
Exercise 8.7.8 applies to tetrahedra. Such techniques are often used
in computer-aided geometric design.
8.7.10. Dene an abstract class Shape, derive three classes Triangle, Circle
and Rectangle from it, and dene a function
bool intersect(Shape* s, Shape* t)

318

8. Class Inheritance

that determines if the two shapes s and t overlap. To achieve this, suitable virtual functions may need be declared for the base class Shape
and overridden in derived classes Triangle, Circle and Rectangle:
Hint: this is a so-called double dispatch problem. An answer to this
exercise can be found in Van98].
8.7.11. Run the codes in x8.6 to see if they work as expected. Rewrite the
codes into more general templates with a template parameter for the
type of the entries of the matrix, so that they can handle real and
complex matrices in single, double, and other precisions.

M. Bertero, P. Boccacci, Christine de Mol - Introduction To Inverse Problems in Imaging-CRC Press (2021)
No ratings yet
M. Bertero, P. Boccacci, Christine de Mol - Introduction To Inverse Problems in Imaging-CRC Press (2021)
358 pages
Probability With Applications in Engineering, Science, and Technology, 2nd (Instructor's Solution Manual) - Matthew A. Carlton
100% (1)
Probability With Applications in Engineering, Science, and Technology, 2nd (Instructor's Solution Manual) - Matthew A. Carlton
400 pages
Message Passing Interface (MPI)
No ratings yet
Message Passing Interface (MPI)
14 pages
Linear Algebra Answers
No ratings yet
Linear Algebra Answers
364 pages
Examination of The Factors Leading To Early Bilingualism
No ratings yet
Examination of The Factors Leading To Early Bilingualism
24 pages
Essential Checking of Finite Element Analysis
No ratings yet
Essential Checking of Finite Element Analysis
13 pages
An CLP PDF
No ratings yet
An CLP PDF
545 pages
Introduction To Matlab
No ratings yet
Introduction To Matlab
45 pages
Functional Programming Languages
No ratings yet
Functional Programming Languages
12 pages
C Palindrome Program, C Program For Palindrome - Programming Simplified
No ratings yet
C Palindrome Program, C Program For Palindrome - Programming Simplified
4 pages
Matlab Matlab Toolbox Deep Learning Toolbox Neural Network Toolbox Libraries Functions How To Use
No ratings yet
Matlab Matlab Toolbox Deep Learning Toolbox Neural Network Toolbox Libraries Functions How To Use
5 pages
EX03. 2D Static Analysis of Rectangular Plate With Hole
No ratings yet
EX03. 2D Static Analysis of Rectangular Plate With Hole
2 pages
Numerical Examples
No ratings yet
Numerical Examples
44 pages
Numerical Methods
0% (1)
Numerical Methods
2 pages
Project Name: P-Adic Integers
No ratings yet
Project Name: P-Adic Integers
51 pages
Course Code: Course Title TPC Version No. Course Pre-Requisites/ Co-Requisites Anti-Requisites (If Any)
No ratings yet
Course Code: Course Title TPC Version No. Course Pre-Requisites/ Co-Requisites Anti-Requisites (If Any)
4 pages
Programming Contest Guide
100% (1)
Programming Contest Guide
240 pages
Dies Tel
No ratings yet
Dies Tel
28 pages
Binary Indexed Tree
No ratings yet
Binary Indexed Tree
9 pages
Nurlita Sari - Jurnal Akhir Perc.8 Fiskom
No ratings yet
Nurlita Sari - Jurnal Akhir Perc.8 Fiskom
10 pages
Modern C++ Design
0% (1)
Modern C++ Design
6 pages
Computational-tools-and-software-MATLAB-Python
No ratings yet
Computational-tools-and-software-MATLAB-Python
5 pages
Ogata - Exercicios Resolvidos
No ratings yet
Ogata - Exercicios Resolvidos
30 pages
The Least-Squares Finite Element Method Theory and Applications in Computational Fluid Dynamics and Electromagnetics
No ratings yet
The Least-Squares Finite Element Method Theory and Applications in Computational Fluid Dynamics and Electromagnetics
424 pages
Skript QFT2
No ratings yet
Skript QFT2
107 pages
Coding Theory: A Bird S Eye View: Continued Block Codes: Basics
No ratings yet
Coding Theory: A Bird S Eye View: Continued Block Codes: Basics
32 pages
Parallel
No ratings yet
Parallel
13 pages
Cplusplus
No ratings yet
Cplusplus
169 pages
Guide To Graph Algorithms Sequential Parallel and Distributed Compress
No ratings yet
Guide To Graph Algorithms Sequential Parallel and Distributed Compress
475 pages
Scilab Manual
No ratings yet
Scilab Manual
700 pages
Review of Conic Sections PDF
No ratings yet
Review of Conic Sections PDF
9 pages
Solid Modeling: Evolution of Geometric Modeling
No ratings yet
Solid Modeling: Evolution of Geometric Modeling
18 pages
On Rice S Theorem: 1 Properties
No ratings yet
On Rice S Theorem: 1 Properties
7 pages
Onishchik, A. L. Lectures On Real Semisimple Lie Algebras and Ther Rerpesentations
No ratings yet
Onishchik, A. L. Lectures On Real Semisimple Lie Algebras and Ther Rerpesentations
96 pages
KMP String Matching Algorithm
No ratings yet
KMP String Matching Algorithm
8 pages
Computational Data Science: Advanced Programme in
No ratings yet
Computational Data Science: Advanced Programme in
15 pages
Cannon Strassen DNS Algorithm
No ratings yet
Cannon Strassen DNS Algorithm
10 pages
COMSOL Release Notes
No ratings yet
COMSOL Release Notes
142 pages
AA V1 I1 PCG Lanczos Eigensolver
No ratings yet
AA V1 I1 PCG Lanczos Eigensolver
2 pages
MSC Syllabus
No ratings yet
MSC Syllabus
27 pages
NumericalMethodsforEngineers PDF
No ratings yet
NumericalMethodsforEngineers PDF
160 pages
Tiger Language Specification
No ratings yet
Tiger Language Specification
5 pages
Pointers and Dynamic Objects: COMP171 Fall 2006
No ratings yet
Pointers and Dynamic Objects: COMP171 Fall 2006
55 pages
MATH 22 Calculus For Business
No ratings yet
MATH 22 Calculus For Business
2 pages
Distance Based Models
No ratings yet
Distance Based Models
58 pages
Download Full Discrete Mathematics Special Indian Edition Semyour Lipschutz PDF All Chapters
100% (8)
Download Full Discrete Mathematics Special Indian Edition Semyour Lipschutz PDF All Chapters
77 pages
Homology (Mathematics)
No ratings yet
Homology (Mathematics)
5 pages
Ogata - Exercicios Resolvidos
100% (1)
Ogata - Exercicios Resolvidos
28 pages
DSP Lab Manual Final PDF
No ratings yet
DSP Lab Manual Final PDF
102 pages
5 Time & Space Complexity 1.1!03!08 2024
No ratings yet
5 Time & Space Complexity 1.1!03!08 2024
31 pages
Rings of Continuous Functions: July 2007
No ratings yet
Rings of Continuous Functions: July 2007
26 pages
Foulds Optimization
No ratings yet
Foulds Optimization
511 pages
Ugc Net Exam Daa PDF
No ratings yet
Ugc Net Exam Daa PDF
94 pages
Realanal PDF
No ratings yet
Realanal PDF
282 pages
Let Us C Y Kanitkar 01 - First Few Pages
No ratings yet
Let Us C Y Kanitkar 01 - First Few Pages
7 pages
Substitutional Analysis
From Everand
Substitutional Analysis
Daniel Edwin Rutherford
No ratings yet
Group Theory in Solid State Physics and Photonics: Problem Solving with Mathematica
From Everand
Group Theory in Solid State Physics and Photonics: Problem Solving with Mathematica
Wolfram Hergert
No ratings yet
C Plus Plus and Object Oriented Numeric Computing For Scientists and Engineers Daoqi Yang PDF
No ratings yet
C Plus Plus and Object Oriented Numeric Computing For Scientists and Engineers Daoqi Yang PDF
334 pages
C++ Fem Object Oriented
No ratings yet
C++ Fem Object Oriented
17 pages
C
No ratings yet
C
100 pages
C++ Functional Programming for Starters: A Practical Guide with Examples
From Everand
C++ Functional Programming for Starters: A Practical Guide with Examples
William E. Clark
No ratings yet
Ch-1 Som Workbook 2017 Answer Key by S K Mondal
No ratings yet
Ch-1 Som Workbook 2017 Answer Key by S K Mondal
1 page
New Doc 2017-10-17
No ratings yet
New Doc 2017-10-17
1 page
SOM 2018 Workbook Answer Key by S K Mondal PDF
No ratings yet
SOM 2018 Workbook Answer Key by S K Mondal PDF
1 page
ISRO Scientist/ Engineer SC' Syllabus, Exam Pattern 2018
No ratings yet
ISRO Scientist/ Engineer SC' Syllabus, Exam Pattern 2018
3 pages
Syllab
No ratings yet
Syllab
5 pages
Applied Mathematical Modelling: Amale Mahi, El Abbas Adda Bedia, Abdelouahed Tounsi
No ratings yet
Applied Mathematical Modelling: Amale Mahi, El Abbas Adda Bedia, Abdelouahed Tounsi
20 pages
Verbal Ability
100% (1)
Verbal Ability
17 pages
AFT Calc1
100% (1)
AFT Calc1
18 pages
Mathematics - Problem Sheet Level 2
No ratings yet
Mathematics - Problem Sheet Level 2
4 pages
Thermodynamical Analysis Tool in MATLAB
No ratings yet
Thermodynamical Analysis Tool in MATLAB
116 pages
Rockets and Mission Analysis: Orbital Mechanics Hohmann Transfer Orbit Summary
No ratings yet
Rockets and Mission Analysis: Orbital Mechanics Hohmann Transfer Orbit Summary
24 pages
Method of Characteristics and Velocity of Sound Reacting: For Gases
No ratings yet
Method of Characteristics and Velocity of Sound Reacting: For Gases
2 pages
N Revision Test Set II 2007
No ratings yet
N Revision Test Set II 2007
7 pages
Quad
No ratings yet
Quad
55 pages
66b21c6d06eac .Exam Answers
No ratings yet
66b21c6d06eac .Exam Answers
13 pages
Simple Present Tense
No ratings yet
Simple Present Tense
59 pages
Sample Question Paper ENGLISH
0% (1)
Sample Question Paper ENGLISH
7 pages
21 American English Phrases That Are
No ratings yet
21 American English Phrases That Are
5 pages
REVISION
No ratings yet
REVISION
3 pages
HTML (Hyper Text Markup Language)
100% (1)
HTML (Hyper Text Markup Language)
53 pages
NLP Lecture 6
No ratings yet
NLP Lecture 6
57 pages
Quarter 1 Teacher's Guide Week 6 I. Objectives A. Expressive Objectives
No ratings yet
Quarter 1 Teacher's Guide Week 6 I. Objectives A. Expressive Objectives
17 pages
Together 1 Worksheets Unit2
No ratings yet
Together 1 Worksheets Unit2
11 pages
PDF of Antropologia Culturale 3Rd Edition Conrad P Kottak Full Chapter Ebook
100% (13)
PDF of Antropologia Culturale 3Rd Edition Conrad P Kottak Full Chapter Ebook
69 pages
Part 1 Listening Task 1 The Origins of The Roman Alphabet
No ratings yet
Part 1 Listening Task 1 The Origins of The Roman Alphabet
25 pages
Phrases and Clauses
No ratings yet
Phrases and Clauses
26 pages
Fragments Workshop
No ratings yet
Fragments Workshop
5 pages
Vector Ecuaciones Ejercicios Curso Matematicas: Anuncios Google
No ratings yet
Vector Ecuaciones Ejercicios Curso Matematicas: Anuncios Google
6 pages
On The Historicity of The Exodus What E
No ratings yet
On The Historicity of The Exodus What E
22 pages
The Mushroom Books - Guide To Identification and Study The Common Mushrooms
100% (7)
The Mushroom Books - Guide To Identification and Study The Common Mushrooms
276 pages
Department of Education: Division Diagnostic Test Grade 1 Mathematics
100% (1)
Department of Education: Division Diagnostic Test Grade 1 Mathematics
5 pages
Score Ukh Batch 4
No ratings yet
Score Ukh Batch 4
40 pages
よつばと! (２目)
No ratings yet
よつばと! (２目)
4 pages
Household Survey Questionnaire
No ratings yet
Household Survey Questionnaire
5 pages
3rd QTR Module 3 in English 2 Synonyms and Antonyms GrADE 2
No ratings yet
3rd QTR Module 3 in English 2 Synonyms and Antonyms GrADE 2
25 pages
English Language Profeciency - Noun
No ratings yet
English Language Profeciency - Noun
4 pages
A Reformed Druid Anthology-07-Miscellany
No ratings yet
A Reformed Druid Anthology-07-Miscellany
30 pages
Activity Sheet Q2W5
No ratings yet
Activity Sheet Q2W5
12 pages
Lý thuyết tiếng tổng hợp - EN57.048
No ratings yet
Lý thuyết tiếng tổng hợp - EN57.048
87 pages
Useful Vocabulary - Unit 1 - CPE
No ratings yet
Useful Vocabulary - Unit 1 - CPE
9 pages
Treasure Hunt Competition 2017
No ratings yet
Treasure Hunt Competition 2017
5 pages
Weber County Sheriff's Office (Utah) - NEW ICE 287 (G) MOA (10/15/09)
No ratings yet
Weber County Sheriff's Office (Utah) - NEW ICE 287 (G) MOA (10/15/09)
20 pages
Mock Interview Session
No ratings yet
Mock Interview Session
102 pages