·-

Turbo Pascal Numerical Methods Toolbox

Borland's No-Nonsense License Statement!

This software is protected by both United States copyright law and international treaty provisions. Therefore,
you must treat this software just like Q book, with the following single exception. Borland International authorizes
you to make archival copies of the software for the sole purpose of backing-up our software and protecting
your investment from loss.
By saying, "just like a book," Borland means, for example, that this software may be used by any number of
people and may be freely moved from one computer location to another, so long as there is no possibility of it
being used at one location while it's being used at another. Just like a book that can't be read by two different
people in two different places at the same time, neither can the software be used by two different people in two
different places at the same time. (Unless, of course, Borland's copyright has been violated.)
Borland International grants you (the licensed owner of the Turbo Pascal Numerical Toolbox) the right to
incorporate toolbox routines into your programs. You may distribute your programs that contain Numerical
Toolbox routines in executable form without restriction or fee, but you may not give away or sell any part of
the actual Numerical Methods Toolbox source code. You are not, of course, restricted from distributing your
own source code.
Sample programs are provided on the Numerical Methods Toolbox diskettes as examples of how to use the
various toolbox features. You may edit or modify these sample programs and incorporate them into the
programs that you write. Use of these sample programs is governed by the same conditions and restrictions as
outlined in the first paragraph above.
WARRANTY
With respect to the physical diskette and physical documentation enclosed herein, Borland International, Inc.
("Borland") warrants the same to be free of defects in materials and workmanship for a period.of 60 days from
the date of purchase. In the event of notification within the warranty period of defects in material or workman-
ship, Borland will replace the defective diskette or documentation. If you need to return a product. call the
Borland Customer Service Department to obtain a return authorization number. The remedy for breach of
this warranty shall be limited to replacement and shall not encompass any other damages, including but not
limited to loss of profit, and special, incidental, consequential, or other similar claims.
Borland International, Inc. specifically disclaims all other warranties, expressed or implied, including but not
limited to implied warranties of merchantability and fitness for a particular purpose with respect to defects in
the diskette and documentation, and the program license granted herein in particular, and without limiting
operation of the program license with respect to any particular application, use, or purpose. In no event shall
Borland be liable for any loss of profit or any other commercial damage, including but not limited to special,
incidental, consequential or other damages.
GOVERNING LAW
This statement shall be construed, interpreted, and governed by the laws of the state of California.

First Edition
Printed in USA
987654321
Turbo Pascal
Numerical Methods
Toolbox~

Copyright © 1986
All Rights Reserved
BORLAND INTERNATIONAL, INC.
4585 SCOTTS VALLEY DRIVE
SCOTTS VALLEY, CALIFORNIA 95066
USA
Table of Contents

Introduction ............................................................................................................... 1
Toolbox Functions ...................................................................................................... 1
About this Manual ..................................................................................................... 2
On the Distribution Disks ........................................................................................ 3
System Requirements ................................................................................................ 3
Acknowledgements .................................................................................................... 4

Chapter 1. ROUTINE BEGINNINGS ................................................................. 5

Using the Toolbox: An Example ............................................................................... 5
The Distribution Disks ............................................................................................. 7
Installation .................................................................................................................. 8
The Graphics Demos ............................................................................................... 12
Data Types and Defined Constants ....................................................................... 12
Compiler Directives ................................................................................................ 13

Chapter 2. ROOTS TO EQUATIONS IN ONE VARIABLE .......................... 15

Stopping Criteria ..................................................................................................... 17
Root of a Function Using the Bisection Method (BISECT.lNC) ....................... 18
Description .......................................................................................................... 18
User-Defined Function ...................................................................................... 18
Input Parameters ................................................................................................ 18
Output Parameters ............................................................................................. 19
Syntax of the Procedure Call ............................................................................. 19
Comments ............................................................................................................ 19
Sample Program .................................................................................................. 19
Example .......................................................................................................... 20
Root of a Function Using the Newton-Raphson Method (RAPHSON.lNC) ...... 21
Description ........................................................................................................... 21
User-Defined Functions ..................................................................................... 21
Input Parameters ................................................................................................. 21
Output Parameters ............................................................................................. 22
Syntax of the Procedure Call ............................................................................. 22
Comments ........................................................................................................... 22
Sample Program ................................................................................................. 22
Example .......................................................................................................... 23
Root of a Function Using the Secant Method (SECANT.lNC) ......................... 25
Description .......................................................................................................... 25
User-Defined Function ...................................................................................... 25
Input Parameters ................................................................................................ 25
Output Parameters ............................................................................................. 26
Syntax of the Procedure Call ............................................................................. 26
Comments ......... :................................................................................................. 26
Sample Program ................................................................................................. 26
Example .......................................................................................................... 27
Real Roots of a Real Polynomial Equation Using the Newton-Horner
Method with Deflation (NEWTDEFL.lNC) .................................................... 28
Description .......................................................................................................... 28
User-Defined Types ........................................................................................... 28
Input Parameters ................................................................................................ 28
Output Parameters ............................................................................................. 29
Syntax of the Procedure Call ............................................................................. 30
Comments ........................................................................................................... 30
Sample Program ................................................................................................. 30
Input Files ...................................................................................................... 30
Example .......................................................................................................... 30
Complex Roots of a Complex Function Using Miiller's Method
(MULLER.INC) ................................................................................................... 33
Description .......................................................................................................... 33
User-Defined Types ........................................................................................... 33
User-Defined Procedure .................................................................................... 33
Input Parameters ................................................................................................ 33
Output Parameters ............................................................................................. 34
Syntax of the Procedure Call ............................................................................. 34
Comments ........................................................................................................... 34
Sample Program ................................................................................................. 35
Example .......................................................................................................... 35

Turbo Numerical Methods Toolbox

Complex Roots of a Complex Polynomial Using Laguerre's Method and Deflation
(LAGUERRE.INC) ............................................................................................... 37
Description .......................................................................................................... 37
User-Defined Types ............................................................................................ 37
Input Parameters ................................................................................................ 37
Output Parameters ............................................................................................. 38
Syntax of the Procedure Call ............................................................................. 38
Comments ............................................................................................................ 38
Sample Program .................................................................................................. 39
Input Files ...................................................................................................... 39
Example .......................................................................................................... 39

Chapter 3. INTERPOLXfION ............................................................................ 43

Polynomial Interpolation Using Lagrange's Method (LAGRANGE.lNC) ....... .45
Description .......................................................................................................... 45
User-Defined Types ............................................................................................ 45
Input Parameters ................................................................................................ 45
Output Parameters ............................................................................................. 46
Syntax of the Procedure Call ............................................................................. 46
Sample Program .................................................................................................. 46
Input Files ...................................................................................................... 46
Example .......................................................................................................... 47
Interpolation Using Newton's Interpolary Divided-Difference Method
(DIVDIF.INC) ....................................................................................................... 49
Description .......................................................................................................... 49
User-Defined Types ............................................................................................ 49
Input Parameters ................................................................................................ 49
Output Parameters ............................................................................................. 90
Syntax of the Procedure Call ............................................................................. 50
Sample Program .................................................................................................. 50
Input Files ...................................................................................................... 50
Example .......................................................................................................... 50
Free Cubic Spline Interpolation (CUBE-FRE.lNC) .: ....................................... 52
Description .......................................................................................................... 52
User-Defined Types ............................................................................................ 52
Input Parameters ................................................................................................ 52
Output Parameters ............................................................................................. 53
Syn tax of the Procedure Call ............................................................................. 53
Sample Program .................................................................................................. 53
Input Files ...................................................................................................... 54
Example .......................................................................................................... 54
Clamped Cubic Spline Interpolation (CUBE_CLA.lNC) .................................. 57
Description .......................................................................................................... 57
User-Defined Types ............................................................................................ 57

Table of Contents III

Input Parameters ................................................................................................ 57
Output Parameters ............................................................................................. 58
Syntax of the Procedure Call ............................................................................. 58
Sample Program ................................................................................................. 59
Input Files ...................................................................................................... 59
Example .......................................................................................................... 59

Chapter 4. NUMERICAL DIFFERENTIATION ............................................ 63

First Differentiation Using Two-Point, Three-Point, or Five-Point Formulas
(DERIV.INC) ........................................................................................................ 66
Description .......................................................................................................... 66
User-Defined Types ........................................................................................... 66
Input Parameters ................................................................................................ 66
Output Parameters ............................................................................................. 67
Syntax of the Procedure Call ............................................................................. 67
Comments ........................................................................................................... 67
Sample Program ................................................................................................. 68
Input Files ...................................................................................................... 68
Example .......................................................................................................... 68
Second Differentiation Using Three-Point or Five-Point Formulas
(DERIV2.INC) ....................................................................................................... 71
Description ........................................................................................................... 71
User-Defined Types ............................................................................................ 71
Input Parameters ................................................................................................. 71
Output Parameters ............................................................................................. 72
Syntax of the Procedure Call ............................................................................. 72
Comments ........................................................................................................... 72
Sample Program ................................................................................................. 73
Input Files ...................................................................................................... 73
Example .......................................................................................................... 73
Differentiation with a Cubic Spline Interpolant (INTERDRV.lNC) ................ 76
Description .......................................................................................................... 76
User-Defined Types .: ......................................................................................... 76
Input Parameters ................................................................................................ 76
Output Parameters ............................................................................................. 77
Syntax of the Pn;)cedure Call ............................................................................. 77
Sample Program ................................................................................................. 77
Input Files ...................................................................................................... 78
Example .......................................................................................................... 78
Differentiation of a User-Defined Function (DERIVFN.lNC) ......................... 80
Description .......................................................................................................... 80
User-Defined Types ........................................................................................... 80
User-Defined Function ...................................................................................... 80
Input Parameters ................................................................................................ 80

IV Turbo Numerical Methods Toolbox

Output Parameters ............................................................................................. 81
Syntax of the Procedure Call ............................................................................. 81
Comments ............................................................................................................ 81
Sample Program .................................................................................................. 81
Input Files ...................................................................................................... 81
Example .......................................................................................................... 82
Second Differentiation of a User-Defined Function (DERIV2FN.INC) .......... 83
Description .......................................................................................................... 83
User-Defined Types ............................................................................................ 83
User-Defined Function ....................................................... ;.............................. 83
Input Parameters ................................................................................................ 83
Output Parameters ............................................................................................. 84
Syntax of the Procedure Call ............................................................................. 84
Comments ............................................................................................................ 84
Sample Program .................................................................................................. 84
Input Files ...................................................................................................... 85
Example .......................................................................................................... 85

Chapter 5. NUMERICAL INTEGRATION ...................................................... 87

Integration Using Simpson's Composite Algorithm (SIMPSON.INC) ............. 89
Description .......................................................................................................... 89
User-Defined Function ...................................................................................... 89
Input Parameters ................................................................................................ 89
Output Parameters ............................................................................................. 90
Syntax of the Procedure Call ............................................................................. 90
Sample Program .................................................................................................. 90
Example .......................................................................................................... 90
Integration Using the Trapezoid Composite Rule (TRAPZOID.INC) .............. 92
Description .......................................................................................................... 92
User-Defined Function ...................................................................................... 92
Input Parameters ................................................................................................ 92
Output Parameters ............................................................................................. 93
Syntax of the Procedure Call ............................................................................. 93
Sample Program .................................................................................................. 93
Example .......................................................................................................... 93
Integration Using Adaptive Quadrature and Simpson's Rule
(ADAPSIMP.INC) ................................................................................................. 95
Description .......................................................................................................... 95
User-Defined Function ...................................................................................... 95
Input Parameters ................................................................................................ 95
Output Parameters ............................................................................................. 96
Syntax of the Procedure Call ............................................................................. 96
Comments ............................................................................................................ 96
Sample Program .................................................................................................. 96

Table of Contents v
Example .......................................................................................................... 97
Integration Using Adaptive Quadrature and Gaussian Quadrature
(ADAPGAUS.INC) ................................................................................................ 98
Description .......................................................................................................... 98
User-Defined Function ...................................................................................... 98
Input Parameters ................................................................................................ 98
Output Parameters ............................................................................................. 99
Syntax of the Procedure Call ............................................................................. 99
Comments ........................................................................................................... 99
Sample Program ................................................................................................ 101
Example ......................................................................................................... 101
Integration Using the Romberg Algorithm (ROMBERG.INC) ........................ 102
Description ......................................................................................................... 102
User-Defined Function ..................................................................................... 102
Input Parameters ............................................................................................... 102
Output Parameters ............................................................................................ 103
Syntax of the Procedure Call ............................................................................ 103
Sample Program ................................................................................................ 103
Example ......................................................................................................... 103

Chapter 6. MATRIX ROUTINES ....................................................................... 105

Determinant of a Matrix (DET.INC) ................................................................... 107
Description ......................................................................................................... 107
User-Defined Types .......................................................................................... 107
Input Parameters ............................................................................................... 107
Output Parameters ............................................................................................ 108
Syntax of the Procedure Call ............................................................................ 108
Sample Program ................................................................................................ 108
Input File ...................................................................................................... 108
Example ......................................................................................................... 109
Inverse of a Matrix (INVERSE.INC) .................................................................. 110
Description ......................................................................................................... 110
User-Defined Types ............................................ ;............................................. 110
Input Parameters ............................................................................................... 110
Output Parameters ............................................................................................ III
Syntax of the Procedure Call ............................................................................ III
Sample Program ................................................................................................ III
Input Files ..................................................................................................... III
Example ......................................................................................................... 112
Solving a System of Linear Equations with Gaussian Elimination
(GAUSELIM.INC) ............................................................................................... 114
Description ......................................................................................................... 114
User-Defined Types .......................................................................................... 114
Input Parameters ............................................................................................... 114

vi Turbo Numerical Methods Toolbox

Output Parameters ............................................................................................ 115
Syntax of the Procedure Call ............................................................................ 115
Sample Program ................................................................................................. 115
Input File ...................................................................................................... 115
Example ......................................................................................................... 116
Solving a System of Linear Equations with Gaussian Elimination and
Partial Pivoting (PARTPIVI'.1 NC) ...................................................................... 117
Description ......................................................................................................... 117
User-Defined Types ........................................................................................... 117
Input Parameters ............................................................................................... 117
Output Parameters ............................................................................................ 118
Syntax of the Procedure Call ........................................................................ :... 118
Sample Program ................................................................................................. 118
Input File ...................................................................................................... 118
Example ......................................................................................................... 119
Solving a System of Linear Equations with Direct Factoring
(DIRFACT.INC) .................................................................................................. 120
Description ........................................................................................................ 120
User-Defined Types .......................................................................................... 120
Procedure LU-Decompose Input Parameters ............................................... 121
Procedure LU-Decompose Output Parameters ............................................ 121
Syntax of the Procedure Call ....................................................................... 121
Procedure LU_Solve Input Parameters .......................................................... 121
Procedure LU_Solve Output Parameters ...................................................... 122
Syntax of the Procedure Call ...................................................................... 122
Sample Program ................................................................................................ 122
Input File ..................................................................................................... 122
Example ........................................................................................................ 123
Solving a System of Linear Equations with the Iterative Gauss-Seidel Method
(GAUSSIDL.INC) ............................................................................................... 126
Description ........................................................................................................ 126
User-Defined Types .......................................................................................... 126
Input Parameters .............................................................................................. 127
Output Parameters ........................................................................................... 127
Syntax of the Procedure Call ........................................................................... 128
Sample Program ................................................................................................ 128
Input File ..................................................................................................... 128
Example ........................................................................................................ 129

Chapter 7. EIGENVALUES AND EIGENVECTORS .................................... 131

Real Dominant Eigenvalue and Eigenvector of a Real Matrix Using the
Power Method (POWER.INC) .......................................................................... 133
Description ........................................................................................................ 133
User-Defined Types .......................................................................................... 133

Table of Contents vii

Input Parameters ............................................................................................... 133
Output Parameters ............................................................................................ 134
Syntax of the Procedure Call ............................................................................ 134
Comments .......................................................................................................... 134
Sample Program ................................................................................................ 135
Input File ...................................................................................................... 135
Example ......................................................................................................... 135
Real Eigenvalue and Eigenvector of a Real Matrix Using the Inverse
Power Method (INVPOWER.INC) .................................................................... 137
Description ......................................................................................................... 137
User-Defined Types .......................................................................................... 137
Input Parameters ............................................................................................... 137
Output Parameters ............................................................................................ 138
Syntax of the Procedure Call ............................................................................ 138
Comments .......................................................................................................... 139
Sample Program ................................................................................................ 139
Input File ...................................................................................................... 139
Example ......................................................................................................... 140
Real Eigenvalues and Eigenvectors of a Real Matrix Using the Power
Method and Wielandt's Deflation (WIELANDT.lNC) ................................... 143
Description ......................................................................................................... 143
User-Defined Types .......................................................................................... 143
Input Parameters ............................................................................................... 143
Output Parameters ............................................................................................ 144
Syntax of the Procedure Call ............................................................................ 145
Comments .......................................................................................................... 145
Sample Program ................................................................................................ 146
Input File ...................................................................................................... 146
Example ......................................................................................................... 146
The Complete Eigensystem of a Symmetric Real Matrix Using the Cyclic
Jacobi Method (JACOBI.INC) ............................................................................ 149
Description ......................................................................................................... 149
User-Defined Types .......................................................................................... 149
Input Parameters ............................................................................................... 149
Output Parameters ............................................................................................ 150
Syntax of the Procedure Call ............................................................................ 150
Comments .......................................................................................................... 151
Sample Program ................................................................................................ 151
Input File ...................................................................................................... 151
Example ......................................................................................................... 152

Chapter 8. INITIAL VALUE AND BOUNDARY VALUE METHODS ..... 155

Solution to an Initial Value Problem for a First-Order Ordinary Differential
Equation Using the Runge-Kutta Method (RUNGE_l.INC) .......................... 159

viii Turbo Numerical Methods Toolbox

Description ........................................................................................................ 159
User-Defined Types .......................................................................................... 159
User-Defined Function .................................................................................... 160
Input Parameters .............................................................................................. 160
Output Parameters ........................................................................................... 160
Syntax of the Procedure Call ............................................................................ 161
Comments ........................................................................................................... 161
Sample Program ................................................................................................. 161
Example ........................................................................................................ 162
Solution to an Initial Value Problem for a First-Order Ordinary Differential
Equation Using the Runge-Kutta-Fehlberg Method (RKF_1.1NC) .............. 163
Description ............................................................ ,........................................... 163
User-Defined Types ................. :........................................................................ 163
User-Defined Function .................................................................................... 163
Input Parameters .............................................................................................. 164
Output Parameters ........................................................................................... 164
Syntax of the Procedure Call ........................................................................... 164
Comments .......................................................................................................... 165
Sample Program ................................................................................................ 165
Example ........................................................................................................ 166
Solution to an Initial Value Problem for a First-Order Ordinary Differential
Equation Using the Adams-Bashforth/Adams-Moulton Predictor/Corrector
Scheme (ADAMS_l.INC) ................................................................................... 168
Description ........................................................................................................ 168
User-Defined Types .......................................................................................... 169
User-Defined Function .................................................................................... 169
Input Parameters .............................................................................................. 169
Output Parameters ........................................................................................... 169
Syntax of the Procedure Call ........................................................................... 170
Comments .......................................................................................................... 170
Sample Program ............................ :................................................................... 170
Example ......................................................................................................... 171
Solution to an Initial Value Problem for a Second-Order Ordinary Differential
Equation Using the Runge-Kutta Method (RUNGE-2.1NC) ......................... 172
Description ........................................................................................................ 172
User-Defined Types .......................................................................................... 173
User-Defined Function .................................................................................... 173
Input Parameters .............................................................................................. 173
Output Parameters ............................................................................................ 174
Syntax of the Procedure Call ................................................................ :........... 174
Comments ................ ;.......................................................................................... 174
Sample Program ................................................................................................ 175
Example ........................................................................................................ 175

Table of Contents ix
Solution to an Initial Value Problem for an nth-Order Ordinary Differential
Equation Using the Runge-Kutta Method (RUNGE_N.lNC) ........................ 178
Description ......................................................................................................... 178
User-Defined Types .......................................................................................... 180
User-Defined Function ..................................................................................... 180
Input Parameters ............................................................................................... 180
Output Parameters ............................................................................................ 181
Syntax of the Procedure Call ............................................................................ 181
Comments .......................................................................................................... 181
Sample Program ................................................................................................ 182
Example ......................................................................................................... 182
Solution to an Initial Value Problem for a System of Coupled First-Order
Ordinary Differential Equations Using the Runge-Kutta Method
(RUNGE_Sl.INC) ................................................................................................ 186
Description ............................... ;......................................................................... 186
User-Defined Types .: ........................................................................................ 188
User-Defined Functions ................................................................................... 188
Input Parameters ............................................................................................... 189
Output Parameters ............................................................................................ 189
Syntax of the Procedure Call ............................................................................ 190
Comments .......................................................................................................... 190
Sample Program ................................................................................................ 191
Example ......................................................................................................... 191
Solution to an Initial Value Problem for a System of Coupled Second-Order
Ordinary Differential Equations Using the Runge-Kutta Method
(RUNGE_S2.INC) ............................................................................................... 196
Description ......................................................................................................... 196
User-Defined Types .......................................................................................... 199
User-Defined Functions ................................................................................... 199
Input Parameters .............................................................................................. 200
Output Parameters ............................................................................................ 201
Syntax of the Procedure Call ............................................................................ 201
Comments .......................................................................................................... 201
Sample Program ............................................................................................... 203
Example ........................................................................................................ 203
Solution to Boundary Value Problem for a Second-Order Ordinary Differential
Equation Using the Shooting and Runge-Kutta Methods (SHOOT2.1NC) .. 208
Description ........................................................................................................ 208
User-Defined Types ......................................................................................... 209
User-Defined Function .................................................................................... 209
Input Parameters .............................................................................................. 209
Output Parameters ............................................................................................ 210
Syntax of the Procedure Call ............................................................................ 210
Comments .......................................................................................................... 210

x Turbo Numerical Methods Toolbox

Sample Program ................................................................................................. 211
Example ......................................................................................................... 211
Solution to a Boundary Value Problem for a Second-Order Ordinary Linear
Differential Equation Using the Linear Shooting and Runge-Kutta Methods
(LINSHOT2.INC) ................................................................................................ 215
Description ......................................................................................................... 215
User-Defined Types ........................................................................................... 216
User-Defined Function ..................................................................................... 216
Input Parameters ............................................................................................... 216
Output Parameters ............................................................................................ 216
Syntax of the Procedure Call ............................................................................ 217
Comments ........................................................................................................... 217
Sample Program ................................................................................................. 218
Example ......................................................................................................... 218

Chapter 9. LEAST.SQUARES APPROXIMATION ........................................ 221

Least-Squares Approximation (LEAST.lNC) ..................................................... 222
Description ........................................................................................................ 222
User-Defined Types .......................................................................................... 223
Input Parameters .............................................................................................. 223
Output Parameters ........................................................................................... 224
Syntax of the Procedure Call ........................................................................... 224
Comments .......................................................................................................... 224
POLY.LSQ .................................................................................................... 224
FOURIER.LSQ ........................................................................................... 225
POWER.LSQ ............................................................................................... 225
EXP.LSQ ...................................................................................................... 225
LOG.LSQ ..................................................................................................... 226
USER.LSQ ................................................................................................... 226
Sample Program ................................................................................................ 227
Input Files .................................................................................................... 227
Example ........................................................................................................ 227

Chapter 10. FAST FOURIER TRANSFORM ROUTINES .......................... 233

The Application Programs .................................................................................... 235
Data Sampling ....................................................................................................... 239
User-Defined Types ............................................................................................... 240
Fast Fourier Transform Algorithms ......................................................... :........... 241
Procedure TestInput ......................................................................................... 241
Description ................................................................................................... 241
Input Parameters ......................................................................................... 241
Output Parameters ...................................................................................... 241
Syntax of the Procedure Call ...................................................................... 241
Procedure MakeSinCosTable ........................................................................... 242

Table of Contents xi
Description ................................................................................................... 242
Input Parameters ......................................................................................... 242
Output Parameters ...................................................................................... 242
Syntax of the Procedure Call ...................................................................... 242
Procedure FFT ................................................................................................. 242
Description ................................................................................................... 242
Input Parameters ......................................................................................... 243
Output Parameters ...................................................................................... 243
Syntax of the Procedure Call ...................................................................... 243
Fast Fourier Transform Applications ................................................................... 244
COMPFFT.INC ............................................................................................... 244
Description ................................................................................................... 244
Input Parameters ......................................................................................... 244
Output Parameters ...................................................................................... 244
Syntax of the Procedure Call ...................................................................... 245
Comments .................................................................................................... 245
REALFFT.INC ................................................................................................. 245
Description ................................................................................................... 245
Input Parameters ......................................................................................... 245
Output Parameters ...................................................................................... 246
Syntax of the Procedure Call ...................................................................... 246
Comments .................................................................................................... 246
COMPCNVL.INC ........................................................................................... 246
Description ................................................................................................... 246
Input Parameters ......................................................................................... 247
Output Parameters ...................................................................................... 247
Syntax of the Procedure Call ...................................................................... 247
Comments .................................................................................................... 247
REALCNVL.INC ............................................................................................. 248
Description ................................................................................................... 248
Input Parameters ......................................................................................... 248
Output Parameters ...................................................................................... 248
Syntax of the Procedure Call ...................................................................... 249
Comments .................................................................................................... 249
COMPCORR.INC ........................................................................................... 249
Description ................................................................................................... 249
Input Parameters ......................................................................................... 250
Output Parameters ...................................................................................... 250
Syntax of the Procedure Call ...................................................................... 250
Comments .................................................................................................... 250
REALCORR.INC .............................................................................................. 251
Description .................................................................................................... 251
Input Parameters .......................................................................................... 251
Output Parameters ...................................................................................... 252

xii Turbo Numerical Methods Toolbox

Syntax of the Procedure Call ...................................................................... 252
Comments ..................................................................................................... 252
Sample Program ................................................................................................ 252
Input File ..................................................................................................... 253
Example ........................................................................................................ 253

Chapter 11. GRAPHICS PROGRAMS ............................................................. 261

Function of the Least-Squares Graphics Demonstration Program .................. 262
Function of the Fourier Transform Graphics Demonstration Program ........... 264
Printing ................................................................................................................... 265
Rebuilding LSQIBM.COM ............................................................................. 267
Rebuilding FFTIBM.COM ............................................................................. 268
Rebuilding for the Hercules Card ................................................................... 269
Rebuilding for the EGA Card ......................................................................... 269
Using the Math Coprocessor ........................................................................... 270

REFERENCES ................................................................................................... 271

INDEX .................................................................................................................. 273

Table of Contents xiii

xiv Turbo Numerical Methods Toolbox
I ntroductian

The Turbo Numerical Methods Toolbox is a reference manual for both the student
of numerical analysis and the professional needing efficient routines. An elemen-
tary background in calculus and linear algebra is assumed, although many of the
algorithms use only high-school-level mathematics. A general knowledge of Turbo
Pascal® is also assumed. If you need to brush up on your knowledge of Pascal, we
suggest looking at the Turbo Pascal Reference Manual and/or the Turbo Pascal
Tutor Manual.
Before you begin using a particular routine, read through this brief introductory
chapter and then refer to the chapter that interests you.

Toolbox Functions

The Turbo Pascal Numerical Methods Toolbox provides routines for

• Finding solutions to equations
• Interpolations
• Calculus
• l'iumerical derivatives and integrals
• Matrix operations: inversions, determinants, and eigenvalues
• Differential equations
• Least-squares approximations
• Fourier transforms

About this Manual

The major areas in numerical analysis are represented in this Toolbox, with each
chapter focusing on a particular problem. Each routine begins with a general
description of the implemented algorithm or numerical method. (References to
numerical analysis texts are provided for each numerical procedure.) User-supplied
types, functions, and input and output parameters are defined, and the syntax of
the procedure call is provided. If appropriate, a "Comments" section is also pro-
vided.
Finally, every algorithm in the Toolbox is accompanied by a general-purpose
program that handles all the necessary I/O, while allowing you to try each algo-
rithm without building any code. Handily, these sample programs will often reduce
the coding your own application may require.
As an example, let's say you want to find the roots to an equation in one variable.
First, you would read the introduction to Chapter 2, "Roots to Equations in One
Variable," and choose the numerical method best suited to your particular problem.
Second, you would run the sample program for the desired numerical method to
determine the necessary input and output. Third, you would write a Turbo Pascal
function defining your equation, using the function already coded in the sample
program as a guide. Fourth, you would run the sample program with your function
substituted for the original one. Of course, if these algorithms are to be part of a
larger program, you must build all the interfaces to the other parts of the system;
but this should only be done after you gain experience with the particular numeri-
cal method.
Several books are referred to throughout the text; complete references are listed
at the back of the book in the section entitled "References."
The body of this manual is printed in normal typeface; other typefaces serve to
illustrate the following:
Alternate This type displays program examples and procedure and function
declarations.
Italics This type emphasizes certain concepts, first-mentioned terms, and
mathematical expressions.
Boldface This type marks the reserved words of Turbo Pascal in text and in
program examples.

2 Turbo Numerical Methods Toolbox

On the Distribution Disks

The routines for this Toolbox are contained on three packed disks. Their contents
and general installation instructions are covered in Chapter l.

System Requirements

All routines will run in standard Turbo Pascal version 3.0. (They will also run in
version 2.0, but you must make one change to use the sample programs; see the
section entitled "Installation" in Chapter 1.) All sample programs will run on an
IBM® PC or compatible machine using DOS 2.0 or greater.
A small portion of the Toolbox uses graphics (see Chapter 11). These programs
are for PC-DOS users only, requiring either an IBM PC Color Graphics Adapter,
an IBM Enhanced Graphics Adapter, or a Hercules Monochrome Graphics
Adapter. Recompiling these requires the Turbo Pascal Graphix Toolbox® version
1.06A or later, and Turbo Pascal 3.0. They can also be recompiled for the EGA and
several other cards using version 1.07A of the Graphix Toolbox.
We strongly recommend that anyone serious about numerical analysis invest in
hardware and software to run Turbo Pascal with support for the Intel 8087 numeri-
cal processing chip:
• Hardware: an 8087 chip plugged into the motherboard of a PC, XT® or equiva-
lent, or an 80287 chip in an AT® or equivalent.
• Software: TURBO-87.COM, the version of Turbo Pascal designed to take advan-
tage of the 8087 chip.
For machines running the Intel 8088 CPU, the increase in execution speed of
programs using real-number arithmetic is often a factor of ten or more, while for
80286 machines, the increase is only about a factor of two (Fried 1985). Perhaps
more important than speed is the increase in accuracy -16 significant figures accu-
racy for Turbo Pascal with 8087 support versus 11 significant figures for standard
Turbo. Since round-off errors are a serious concern in numerical analysis, the
increased accuracy is of great value.
All of the examples in this manual were run using Turbo-87. If you run them
without the Turbo-87, you will usually get less accurate answers.

Introduction 3
AcktWwlellgements

We refer to several products throughout the manual; following is a list of each one
and its respective company.
• Turbo Pascal, Turbo Pascal Graphix Toolbox, SideKick, SuperKey, and
Reflex: The Database Manager are registered trademarks and Turbo Pascal
Numerical Methods Toolbox is a trademark of Borland International, Inc.
• IBM, XT, and AT are registered trademarks of International Business Machines
Corp.
• MS-DOS is a registered trademark of Microsoft Corp.

4 Turbo Numerical Methods Toolbox

c H A p T E R 1
Routine Beginnings

This chapter provides you with everything you need to start using the routines in
this Toolbox. We'll discuss how to unpack the disks for use and list the files avail-
able once the disks are unpacked. We also briefly discuss data types and defined
constants used in the Toolbox, and the setting of compiler directives.
First, though, before we thrust you into the middle of numerical madness, let's
take a look at one way to use this Toolbox.

Using the Toolbox: An Example

In late 1986 and early 1987, the America's Cup 12-meter yacht championship was
held. The 12-meter yachts are just large sailboats, but the competition is so intense
that the only way to be competitive is to use dozens of people, spend millions of
dollars, design a special boat, and spend a couple of years training for the race. The
race has become so sophisticated that many of the sailboats have on-board com-
puters and other electronic equipment.
To keep stride with other challengers, one yacht's crew used personal com-
puters, and of course, Borland software. They used Turbo Pascal to design the
boat's hull. They used Reflex: The Analyst® to maintain their databases and to dis-
play plots while the boat was sailing. And when it came time to do some mathemat-
ical modeling, again they turned to Borland for its inimitable software and chose
the Toolbox.

5
Simply speaking, the problem they had was one of "precision monitoring." It
takes a crew of very highly skilled sailors to compete in America's Cup races, but
even the best skippers cannot act with sufficient precision to win. A typical race
lasts for several hours, and the winner usually wins by only a few feet.
The electronic equipment on a boat can sense with reasonable accuracy all of the
crucial variables: boat velocity, wind velocity, boat direction, boat position, and so
on. This data must then be made available to the skipper in a coherent form, and
he/she must decide at what angle to place the rudder based on that information.
The problem is too complex to rely on intuition alone.
Even just displaying the velocity is more complex than you might think at first.
When sailing on the ocean, the waves are big enough that the velocity is in constant
flux. Fortunately, the fluctuations due to the waves represents a steadily periodic
force. By using the Fourier transforms in Chapter 10 of the Toolbox, the crew was
able to identify the periodic portion of the velocity and subtract it out. The result:
the velocity as a function of time but with the wave fluctuations eliminated. The
graph of this modified velocity is much smoother, and allows the skipper to tell
much more quickly and accurately whether the boat is accelerating or decelerating.
To measure the acceleration quantitatively, the crew used the fact that the accel-
eration is the derivative of the velocity. They were able to do this easily with the
differentiation routines in Chapter 4 of the Toolbox. They were also able to directly
measure the distance travelled by using the integration routines in Chapter 5, and
the fact that distance is the integral of the speed.
Perhaps the most difficult problem in navigating a sailboat is aiming the rudder.
You can't just aim the boat in the direction that you want to go, rather you have to
pick a direction that you can sail rapidly, depending on the wind direction. An
experienced skipper can judge this pretty well, but not well enough. Every boat is
a little different, and the best way to handle one boat is not necessarily the best way
to handle another.
So, the team ran extensive trial races with the boat to gather data on how the
boat performed in various circumstances. The data was collected automatically by
electronic instruments on board, and stored digitally on floppy disks. They then
used Reflex to manage the data and to display graphs. But they lacked the tools to
relate their data to the data they would have under actual racing conditions.
In order to predict the behavior of their boat in an actual race, the team created a
model from their collected data using the least-squares routines in Chapter 9 of the
Toolbox. With the least-squares routines, you can create a multiparameter model
and then find the values of the parameters that make the model most accurately fit
the data. With a mathematical model of the boat's behavior, the team was then able
to predict how the boat would perform under circumstances similar but not identi-
cal to its practices.

6 Turbo Numerical Methods Toolbox

This, of course, is just one of many applications of this Toolbox. Now, let's go on
to the fundamentals.

. The Distribution Disks

All of the Toolbox routines are contained on three disks. (Note, however, that
MS-DOS® users will receive two disks; Disk 3 is for PC-DOS users only.) Each
disk has packed files corresponding to chapters in the manual. Use the program
UNPACK.EXE to unpack the files, described in the next section, "Installation."
The files for each chapter are self-contained and do not require any files from
any other chapter, with these exceptions:
\

• All files require Turbo Pascal (not included).

• Most files require COMMON.INC, located on Disk l.
• The files for Chapter 11 require files from Chapters 9 and 10, as well as the
Turbo Pascal Craphix Toolbox (not included).
The numerical analysis routines are in the files with the .INC suffix. The files
with the .PAS suffix are demonstration programs. To run a demonstration program,
get into Turbo Pascal and load the .PAS file of your choice. The menus are self-
explanatory. The .DAT files contain input data for specific .PAS files.
If you're a PC-DOS user with an IBM color graphics monitor or compatible, you
can run LSQIBM.COM or FFfIBM.COM from Disk 3 to see a quick graphic
demonstration of the power and usefulness of the Toolbox. These routines require
the files SAMP11A.DAT, SAMPllB.DAT, 4X6.FON, 8X8.FON, and ERROR.MSC
to be on the current directory. (These files are also on Disk 3.)
Contents of the enclosed disks:
Disk 1
README
README.COM (program to display README file)
UNPACK.EXE (installation program to unpack chapters)
COMMON.INC (used throughout the Toolbox)
COMMON2.INC (for Turbo Pascal 2.0 users)
CHAP2 (packed file with routines for Chapter 2)
CHAP3 (packed file with routines for Chapter 3)
CHAP4 (packed file with routines for Chapter 4)
CHAP5 (packed file with routines for Chapter 5)
CHAP6 (packed file with routines for Chapter 6)
CHAP7 (packed file with routines for Chapter 7)

Routine Beginnings 7
Disk 2
UNPACK.EXE (installation program to unpack chapters)
CHAP8 (packed file with routines for Chapter 8)
CHApg (packed file with routines for Chapter 9)
CHAPIO (packed file with routines for Chapter 10)
CHAP11 (packed file with routines for Chapter 11; PC-DOS users only)
Disk 3 (PC-DOS users only)
README
README.COM (program to display README file)
LSQIBM.COM (requires IBM graphics monitor)
FFTIBM.COM (requires IBM graphics monitor)
LSQHERC.COM (requires Hercules graphics card)
FFTHERC.COM (requires Hercules graphics card)
SAMP11A.DAT (data file for LSQ*.COM)
SAMP11B.DAT (data file for FFT*.COM)
14X9.FON (from the Graphix Toolbox)
4X6.FON (from the Graphix Toolbox)
8X8.FON (from the Graphix Toolbox)
ERROR.MSG (from the Graphix Toolbox)

Installation

The files CHAP2 through CHAP11 are packed files corresponding to the chapters
in this manual. In order to use these files, you must first unpack them with
UNPACK.EXE. The syntax is as follows:
UNPACK packed-file-name target-drive
For example, the files for Chapter 2 can be extracted and put on the current
directory on drive C by placing Disk 1 in drive A, changing the logged drive to A by
typing A: at the DOS prompt, and then typing
UNPACK CHAP2 c:
Wildcards are okay to use for the packed file name, and a directory can be
specified with the target drive if it ends with a backslash (\). For example, all of the
packed files on disk can be placed in a directory C:\NUMERIC by doing the
following:
UNPACK CHAP* C:\NUMERIC\
if the directory C:\NUMERIC already exists.

8 Turbo Numerical Methods Toolbox

You may wish to copy the packed files onto your hard disk, and then unpack
them as you need them.
Note: These files are not copy protected. All files are ordinary DOS files; there
are no hidden files. The unpacking program only extracts ordinary text files - it
will not create directories, modify the distribution disk, create hidden or protected
files, or do anything unexpected.
Contents of the packed files:
CHAP2 (packed file) "Roots to Equation in One Variable"
BISECT.lNC N EWTDEFL. PAS
BISECT.PAS RAPHSON.lNC
LAGUERRE.lNC RAPHSON .PAS
LAGUE RRE. PAS RAPHSON2.PAS
MULLER.lNC SECANT.lNC
MULLERPAS SECANT.PAS
NEWTDEFL.lNC

CHAP3 (packed file) "Interpolation"

CUBE_CLA.lNC SAMPLE3B.DAT
CUBE_CLAPAS SAMPLE3C.DAT
CUBE-FRE.lNC SAMPLE3D.DAT
CUBE_FRE.PAS SAMPLE3E.DAT
DIVDIF.INC SAMPLE3F.DAT
DIVDI F. PAS SAMPLE3G.DAT
LAGRANGE.lNC SAMPLE3H.DAT
LAGRANGE. PAS SAMPLE31.DAT
SAMPLE3A.DAT

CHAP4 (packed file) "Numerical Differentiation"

DERIV.lNC DERIV2FN.lNC
DERIV.PAS DERIV2FN .PAS
DERIV2.1NC INTERDRV.lNC
DERIV2.PAS INTERDRV.PAS
DERIVFN.lNC SAMPLE4A.DAT
DERIVFN.PAS SAMPLE4B.DAT

Routine Beginnings 9
CHAP5 (packed file) "Numerical Integration"
ADAPGAUS.lNC ROMBERG.PAS
ADAPGAUS.PAS SIMPSON.lNC
ADAPSIMP.lNC SIMPSON.PAS
ADAPSIMP.PAS TRAPZOID.lNC
ROMBERG.INC TRAPZOID.PAS

CHAP6 (packed file) "Matrix Routines"

DET.INC INVERSE.INC
DET.PAS INVERSE.PAS
DIRFACT.INC PARTPIVT.INC
DIRFACT.PAS PARTPIVT.PAS
GAUSELIM.INC SAMPLE6A.DAT
GAUSELIM.PAS SAMPLE6B.DAT
GAUSSIDL.INC SAMPLE6C.DAT
GAUSSIDL.PAS SAMPLE6D.DAT

CHAP7 (packed file) "Eigenvalues and Eigenvectors"

INVPOWER.INC POWERPAS
INVPOWERPAS SAMPLE7ADAT
JACOBI.INC WIELANDT.lNC
JACOBI. PAS WIELANDT.PAS
POWER.INC

CHAPS (packed file) "Initial Value and Boundary Value Methods"

ADAMS_l.INC RUNGE-2.PAS
ADAM S_I. PAS RUNGE_N.INC
LINSHOT2.1NC RUNGE-N.PAS
LINSHOT2.PAS RUNGE_Sl.INC
RKF_l.INC RUNGE_Sl.PAS
RKF_l.PAS . RUNGE_S2.INC
RUNGE_l.INC SHOOT2.INC
RUNGE_I. PAS SHOOT2.PAS
RUNGE-2.INC

10 Turbo Numerical Methods Toolbox

CHAP9 (packed file) "Least-Squares Approximations"
EXP.LSQ POLY.LSQ
FOURIERLSQ POWERLSQ
LEAST.INC SAMPLE9A.DAT
LEAST. PAS USERLSQ
LOG.LSQ

CHAPIO (packed file) "Fast Fourier Transform Routines"

COMPCNVL.INC FFTPROGS.PAS
COMPCORRINC REALCNVL.INC
COMPFFT.INC REALCORR.INC
FFT87B2.INC REALFFT.IN C
FFT87B4.INC SAMPIOA.DAT
FFTB2.INC SAMPI0B.DAT
FFTB4.INC SAMPIOC.DAT

CHAPll (packed file) "Graphics Programs" IPC-DOS only

FFTDEMO.PAS LEAST.MOD
GENERIC.LSQ LSQDEMO.PAS
GRAPHIX.EGA
GRAPHIX.HGC
IOCHECK.INC

All sample programs call the include file COMMON.INC from the disk. This
file includes procedures that are common to all sample programs. When copying
any of the sample programs to a disk, be sure to also copy the file COMMON.INC
to that disk or the sample programs will not compile.
To use the sample programs with Turbo Pascal version 2.0, rename COM-
MON2.INC (which is on Disk 1) to COMMON.INC. (You may wish to preserve a
copy of tlie original COMMON.INC file by first copying it to a file called COM-
MON3.INC). If you run the sample programs with version 2.0 and do not make
this change, the programs will compile but will handle I/O errors incorrectly.
We have made the sample programs general and easy to use. For example,
numerical input can originate from the keyboard (where improper input is trap-
ped) or from a text file; output can be sent to the printer, screen, or text file; other
refinements are also included. Since, to a beginner, the supporting code may
obscure the simplicity of calling the procedure, we have included a minimal sample
program for Newton-Raphson's method of root-finding (RAPHSON2.INC).

Routine Beginnings II
The Graphics Denws

Because graphic displays are often an essential part of numerical analysis, we have
included two demonstration programs (for PC-DOS users only) that involve dis-
play numerical results. As previously stated, graphics hardware is not necessary for
this Toolbox, but it is required for these two graphics programs. The programs are
built with subsets of the Turbo Pascal Graphix Toolbox; there are separate versions
for systems with the Hercules Monochrome Graphics card and the IBM Color
Graphics card (or good emulations of these cards).
The demonstration programs are on Disk 3. For instructions about how to run or
recompile them, see Chapter 11.

Data Types and Defined Constants

Data types that might be confused with those in the calling program have been
prefixed with the letters TN (for Turbo Numerical); for example, TNmatrix or
TNvector. You must define these variable types in your top-level program for two
reasons. First, you will probably need to use this type in your top-level program,
and the type must be defined to have the same scope as the Toolbox procedure.
Second, you will want to dimension arrays based on your particular needs. For
example, the Lagrange procedure requires the definition
type TNvector = array[O .. TNArraySize] of Real;
The identifier TNArraySize is never referred to in any of the include files. It
should be optimized by the user, although we have set a default value in each of the
sample programs. It may be replaced with an integer or byte.
TNNearlyZero is the only defined constant that must be changed when standard
Turbo Pascal is used (instead of Turbo-87); it should be changed to IE - 7. (With-
out changing this constant, you will get a syntax error when compiling with stan-
dard Turbo Pascal.) For Turbo-87, the default value of this constant is IE -15.
(There are a few exceptions to these default values; where appropriate, the "Com-
ments" section of each routine indicates the exceptions.)

12 Turbo Numerical Methods Toolbox

Compiler Directives

Aside from the usual default values of the compiler directives in standard Turbo
Pascal, we have set the compiler directive to {$R + } in all include files that use
arrays, and to {$I-} in all sample programs. The first directive checks to see that all
array-indexing operations are within the defined bounds and all assignments to
scalar and subrange variables are within range. The latter directive disables I/O
error-checking. All the sample programs have their own I/O error-checking proce-
dures (loaded in the file COMMON.lNC), so that the {$I-} directive must remain
disabled in the sample programs. The array checker {$R +} should always be
active, since the performance penalty is slight and the advantages are significant.

Routine Beginnings 13
14 Turbo Numerical Methods Toolbox
c H A p T E R 2
Roots to Equations in One Variable

The routines in this chapter are for finding the roots of a single equation in one real
variable. A typical problem is to solve
x * exp(x) - 10 = 0 .
In general, the routines find a value of x, where x is a scalar real variable,
satisfying
f(x) = 0.0
where f is a real-valued function that you program in Pascal.
All of the methods are approximate methods, meaning that they find an approxi-
mate value of x that makesf(x) close to zero. Because of round-off error, it is usually
not possible to find the exact value of x. Furthermore, they are all iterative
methods, meaning that you specify some initial guess that is some value for x,
which you think is reasonably close to the solution. The routine repeats some calcu-
lations that replace the guess x with a more accurate guess until the required level
of accuracy is achieved.
The bisection method (BISECT.INC) returns an approximation to a root of a real
continuous function of the real variable x. This method always converges (as long as
the function changes signs at a root), but may do so relatively slowly.
The Newton-Raphson method (RAPHSON.INC) also returns an approximation
to a root of a real functionf of the real variable x. When this algorithm converges, it
is usually faster than the bisection method. If more than one root of a polynomial
equation is desired, then use Newton-Horner's method (NEWfDEFL.INC).

15
The secant method (SECANT.lNC) is similar to the Newton-Raphson method,
but doesn't require knowledge of the first derivative of the function. Consequently,
it is more flexible than the Newton-Raphson method, though somewhat slower.
Newton-Horner's method (NEWfDEFL.lNC) applies Newton's method to
real polynomials. It also uses deflation techniques to attempt to approximate all the
real roots of a real polynomial. Both the Newton-Horner and Newton-Raphson
methods are faster than the bisection and secant methods, but are undefined if
If'(x)1 < = TNNearlyZero. This is less of a problem on machines with a high-
precision math coprocessor, since TNNearlyZero is smaller.
The Newton-Horner and Newton-Raphson methods both converge around mul-
tiple roots, although convergence is slow. These algorithms depend upon an initial
approximation of the root. If the initial approximation is not sufficiently close to the
root, the Newton methods may not converge. In some instances, an initial choice
may lead to successive iterations that oscillate indefinitely about a value of x usu-
ally associated with a relative minimum or relative maximum off. In either case,
the bisection method could be used to determine the root or to determine a close
approximation to the root that can be employed as an initial approximation in the
Newton-Raphson or Newton-Horner methods.
Muller's method (MULLER.lNC) returns an approximation to a root (possibly
complex) of a complex function of the complex variable x. Although Muller's
method can approximate the roots of polynomials, we recommend that you use
Newton-Horner's method, the secant method, or (in the case of complex polyno-
mials) Laguerre's method to find the roots of polynomials.
Laguerre's method (LAGUERRE.lNC) attempts to approximate all the real and
complex roots of a real or complex polynomial. Laguerre's method is very reliable
and quick, even when converging to a multiple root. This is the best general
method to use with polynomials.
A caution when solving polynomial equations: Polynomials can be ill-
conditioned, in the sense that small changes in the coefficients may lead to large
changes in the roots.

16 Turbo Numerical Methods Toolbox

Stopping Criteria

All the root-finding routines use the function TestForRoot to determine if a root has
been found.
function TestForRoot{X, OldX, V, Tol : Real) : Boolean;
(********************************************************************)
(* Here are four stopping criteria. If you wish to *)
(* change the active criteria, simply comment off the current *)
{* criteria (including the appropriate or) and remove the comment *)
{* brackets from the criteria (including the appropriate or) you *)
(* wish to be active. *)
(********************************************************************)
begin
TestForRoot := (***************************)
(ABS{V) <= TNNearlyZero) (* V=O *)
(* *)
or (* *)
(* *)
(ABS{X - OldX) < ABS{OldX*Tol)) (* relative change in X *)
(* *)
(* *)
{* or *) (* *)
{* *) (* *)
{* (ABS{X - OldX) < Tol) *) (* absolute change in X *)
{* *) (* *)
{* or *) (* *)
{* *) (* *)
(* (ABS{V) <= Tol) *) (* absolute change in V *)
(***************************)
end; { procedure TestForRoot }
The four separate tests provided by function TestForRoot may be used in any
combination. The default criteria tests the absolute value of Y and the relative
change in X. If you wish to change the active criteria, simply comment off the
current criteria (including the appropriate or) and remove the comment brackets
from the criteria (including the appropriate or) you wish to be active.
The first criterion simply checks to see if Y is zero (TNNearlyZero is defined at
the beginning of the procedure). This criterion should usually be kept active.
The second criterion examines the relative change in X between iterations. To
avoid division by zero errors, QldX has been multiplied through the inequality.
The third criterion checks the absolute change in X between iterations.
The fourth criterion determines the absolute difference between Y and the
allowable tolerance. Note: The parameter Tol(erance) means something different in
each test. Be sure you know which tests are active when you input a value for Tol.

Roots to Equations in One Variable 17

Root of a Function Using the Bisection Method
(BISECT. INC)

Description

This method (Burden and Faires 1985, 28 ff.) provides a procedure for finding a
root of a real continuous function f, specified by the user on a user-supplied real
interval [a,b]. The functionsfia) and fib) must be of opposite signs. The algorithm
successively bisects the interval and converges to the root of the function. You must
also specify the desired accuracy to which the root should be approximated.

User-Defined Function

funct;on TNTargetF(x : Real) : Real;

The procedure Bisect determines the roots of this function.

Input Parameters

LeftEnd:Real; Left end of the interval

RightEnd:Real; Right end of the interval
To 1 : Rea 1 ; Indicates accuracy of solution
Maxlter:Real; Maximum number of iterations permitted
The preceding parameters must satisfy the following conditions:
1. LeftEnd < RightEnd.
2. TNTargetF(LeftEnd) * TNTargetF(RightEnd) < 0; the endpoints must have
opposite signs.
3. Tol > O.
4. MaxIter ~ O.

18 Turbo Numerical Methods Toolbox

Output Pararneters

Answer:Real; An approximate root of TNTargetF

fAnswer:Real; The value of the function at the value Answer
Iter: Integer; Number of iterations to find answer
Error:Byte; 0: No error
1: Iter > MaxIter
2: Endpoints are of the same sign
3: LeftErul > RightErul
4: Tol :::;; 0
5: MaxIter < 0
If Error = 1 (maximum number of iterations exceeded), Answer is set to the last
x value tested and fAnswer is set to TNTargetF(Answer). If Error> 1, then the other
output parameters are not defined.

Syntax of the Procedure Call

Bisect(LeftEnd, RightEnd, Tol, MaxIter, Answer, fAnswer, Iter, Error);

The procedure Bisect determines the roots of function TNTargetF.

Comrnents

If a root occurs at a relative maximum or relative minimum, the bisection method

will be unable to locate that value of p if P does not occur as an endpoint of a
subinterval.
Convergence is determined with the Boolean function TestForRoot described at
the beginning of this chapter.

Sample Program

The sample program BISECT.PAS provides I/O functions that demonstrate the
bisection algorithm. To modify this program for your own function, simply change
the definition of function TNTargetF. Note that the file BISECT.INC is included
after the function TNTargetF is defined.

Roots to Equations in One Variable 19

Example
Problem. Determine the solution to the equation cos(x) = x.
1. Write the following code for function TNTargetF into BISECT.PAS:
{----------- HERE IS THE FUNCTION ------------}
funct;on TNTargetF(x : Real) : Real;
beg;n
TNTargetF := Cos (x) - x;
end; { function TNTargetF }

{---------------------------------------------}
2. Run BISECT.PAS:
Left endpoint: 0
Right endpoint: 100
Tolerance (> 0, default = 1.000E-08): 1E-6
Maximum number of iterations (>= 0, default = 100)? 100
Direct output to one of the following:
(S)creen
(P)rinter
(F) il e

Left endpoint: O.OOOOOOOOOOOOOOE+OOO

Right endpoint: 1.00000000000000E+002
Tolerance: 1.00000000000000E-006
Maximum number of iterations: 100
Number of iterations: 28
Calculated root: 7.39085301756859E-001
Value of the function
at the calculated root: -2.82073423951701E-007

20 Turbo Numerical Methods Toolbox

Root of a Function Using the Newton-Haphson Method
(RAPHSON.INC)

Description

This example uses Newton-Raphson's algorithm (Burden and Faires 1985,42 fr.) to
find a root of a real user-specified function when the derivative of the function and
an initial guess are given. The algorithm constructs the tangent line at each iterate
approximation of the root. The intersection of the tangent line with the x-axis
provides the next iterate value of the root. You must specify the desired tolerance to
which the root should be approximated.

User-Defined Functions

function TNTargetF(x : Real) : Real;

function TNDerivF(x : Real) : Real;

The procedure Newton Raphson determines the roots of the function

TNTargetF.
The function TNDerivF must be the first derivative of function TNTargetF.

Input Parameters

Guess:Real; User's initial approximation to the root

Tol:Real; Tolerance in answer (see "Comments")
MaxIter: Integer; Maximum number of iterations permitted
The preceding parameters must satisfy the following conditions:
1. Tal> 0
2. MaxIter > 0

Roots to Equations in One Variable 21

Output Parameters

Root: Rea 1; Approximate root.

Va 1ue: Rea 1; Value of the function at the approximate root.
Deri v: Rea 1; Value of the derivative at the approximated root.
Iter: Integer; Number of iterations needed to find the root.
Error:Byte; 0: No error.
1: Iter < MaxIter.
2: The slope is zero (see "Comments").
3: Tol :$ O.
4: MaxIter < O.
If a root is found, it is returned along with the value of the function at the root
(which, of course, should be close to zero) and the value of the derivative at the
root. If Error :$ 2, the data from the last iteration is returned.

Syntax of the Procedure Call

Newton_Raphson{Guess, Tol, MaxIter, Root, Value, Deriv, Iter, Error);

Comments

Newton's method involves division by the value of the derivative of the function.
Should the algorithm attempt to do any calculations at a point where the deriva-
tive is less than TNNearlyZero, the routine will stop and return an error message
(Error = 2).
Convergence is determined with the Boolean function TestForRoot described at
the beginning of this chapter.

Sample Program

The sample program RAPHSON .PAS provides I/O functions that demonstrate the
Newton-Raphson algorithm. Note that the file RAPHSON.INC is included after
the functions TNTargetF and TNDerivF are defined.
The program RAPHSON2.PAS also provides I/O functions that demonstrate the
Newton-Raphson method. It is an extremely bare-bones program and is provided

22 Turbo Numerical Methods Toolbox

for the newcomer to Turbo Pascal who wants to see a simple, straightforward appli-
cation of a Toolbox routine.

Example
Problem. Determine the solution to the equation cos(x) = x.
1. Code the following two functions into RAPHSON.PAS (or RAPHSON2.PAS):
{---------- HERE IS THE FUNCTION -------------}
function TNTargetF{x : Real) : Real;
begin
TNTargetF := Cos (x) - x;
end; { function TNTargetF }

{---------------------------------------------}
{-------- HERE IS THE DERIVATIVE -------------}
function TNDerivF{x : Real) : Real;
begin
TNDerivF := -Sin{x) - 1;
end; { function TNDerivF }

{---------------------------------------------}
2. Run RAPHSON.PAS:
Initial approximation to the root: 0
Tolerance (> 0, default = 1.000E-08): 1E-6
Maximum number of iterations (>= 0, default = 100): 100
Direct output to one of the following:
{S)creen
{P)rinter
{F)i1e

Initial approximation: O.OOOOOOOOOOOOOOE+OOO

Tolerance: 1.00000000000000E-006
Maximum number of iterations: 100

Number of iterations: 5
Calculated root: 7.39085133215161E-001
Value of the function
at the calculated root: O.OOOOOOOOOOOOOOE+OOO
Value of the derivative
of the function at the
calculated root: -1.67361202918321E+000

Roots to Equations in One Variable 23

Here is the RAPHSON2.PAS version of the same function:
Initial approximation to the root: 0
Tolerance: 1E-6
Maximum number of iterations: 100
Error = 0
Number of iterations: 5
Root: 7.39085133215161E-001
Value of the
function at the root: O.OOOOOOOOOOOOOOE+OOO
Derivative of the
function at the root: -1.67361202918321E+000

24 Turbo Numerical Methods Toolbox

Root of a Function Using the Secant Metlwd
(SECANT. INC)

Description

This example uses the secant method (Gerald and Wheatley 1984, 11-13) to find a
root of a user-specified real function given two initial real approximations to the
root. The secant method constructs a secant through the two points specified by
the initial approximations. The intersection of this line and the x-axis is used as the
next best approximation to the root. The approximation to the root and its prede-
cessor are used to construct the next secant line. The process continues until a root
is approximated with specified accuracy or until a specified number of iterations
have been exceeded.

User-Defined Function

function TNTargetF(x : Real) : Real;

The procedure Secant will determine the roots of this function.

Input Parameters

Guessl:Real; User's first approximation to the root

Guess2:Real; User's second approximation to the root
Tol:Real; Indicates accuracy in solution
MaxIter: Integer; Maximum number of iterations permitted
The preceding parameters must satisfy the following conditions:
1. Tal> 0
2. MaxIter;::: 0

Roots to Equations in One Variable 25

Output Parameters

Root: Rea 1; Approximate root.

Va 1ue: Rea 1; Value of the function at the approximate root.
Iter: Integer; Number of iterations needed to find the root.
Error:Byte; 0: No error.
1: Iter> Max Iter.
2: The slope is zero (see "Comments").
3: Tal :s; O.
4: MaxIter < O.
If a root is found, it is returned with the value of the function at the root (which,
of course, should be nearly zero). If Error :s; 2, then the data from the last iteration
is returned.

Syntax of the Procedure Call

Secant(Guessl, Guess2, Tol, MaxIter, Root, Value, Iter, Error);

The procedure Secant determines the roots of the function TNTargetF.

Comments

The secant algorithm constructs a line through two points and finds the intersec-
tion of that line with the x-axis. If the line has a slope whose absolute values are
less than TNNearlyZero (that is, the two points have the same y-value), then it has
no intersection with the x-axis (or infinitely many if it lies on the x-axis) and the
algorithm will no longer continue. If this happens, Error 2 is returned. Error 2 will
also be returned if the absolute difference of the two initial approximations (Guessl
and Guess2) is less than TNNearlyZero.
Convergence is determined with the Boolean function TestForRoat described at
the beginning of this chapter.

Sample Program

The sample program SECANT.PAS provides I/O functions that demonstrate the
secant algorithm. Note that the file SECANT.lNC is included after the function
TNTargetF is defined.

26 Turbo Numerical Methods Toolbox

Example
Problem. Determine the solution to the equation cos(x) = x.
1. Write the following code for procedure TNTargetF into SECANT.PAS:
{----------- HERE IS THE FUNCTION ------------}
function TNTargetF{x : Real) : Real;
begin
TNTargetF := Cos (x) - x;
end; { function TNTargetF }

{---------------------------------------------}
2. Run SECANT.PAS:
First initial approximation to the root: 0
Second initial approximation to the root:
Tolerance (> 0, default = 1.000E-08): 1E-8
Maximum number of iterations (>= 0, default 100): 100
Direct output to one of the following:
{S)creen
(P)rinter
(F) il e

First initial approximation: O.OOOOOOOOOOOOOOE+OOO

Second initial approximation: 1.00000000000000E+000
Tolerance: 1.00000000000000E-008
Maximum number of iterations: 100
Number of iterations: 6
Calculated root: 7.39085133215161E-001
Value of the function
at the calculated root: O.OOOOOOOOOOOOOOE+OOO

Roots to Equations in One Variable 27

Real Roots of a Real Polynomial Equation Using the
Newton-Horner Metlwd with Deflation (NEWIDEFL.INC)

Description

This example uses Newton-Horner's algorithm and deflation (see RAPHSON.INC

in this chapter for a description of Newton's method). Newton-Horner is the New-
ton-Raphson method applied to polynomials (Burden and Faires 1985, 42 £I). Defla-
tion is used to find several roots of a user-specified real polynomial given an initial
guess specified by the user. This procedure approximates a real root and then
removes the corresponding linear factor from the given polynomial. The newly
obtained (deflated) polynomial is then analyzed for a real root. This process con-
tinues until a quadratic remains, the remaining roots are complex, or the algorithm
is unable to approximate the remaining real roots. Should the polynomial contain
two complex roots, they may be determined using the quadratic formula. You must
specify (at most) the tolerance to which the roots should be approximated.

User-Defined Types

TNvector = array[O .. TNArraySize] of Real;

TNlntVector = array[O .. TNArraySize] of Integer;

Input Parameters

Ini tDegree: Integer; Degree of user-defined polynomial

Ini tPoly:TNvector; Coefficients of user-defined polynomial
Guess:Real; User's initial approximation
Tol:Real; Indicates accuracy in solution
MaxIter:Integer; Maximum number of iterations permitted

28 Turbo Numerical Methods Toolbox

The preceding parameters must satisfy the following conditions:
1. InitDegree > 0
2. Tol > 0
3. Maxlter ~ 0
4. InitDegree =:; TNArraySize

TNArraySize fixes an upper bound on the number of elements in each vector. It

is used in the type definition of TNvector. TNArraySize is not a variable name and
is never referenced by the procedure; hence there is no test for condition 4. If
condition 4 is violated, the program will crash with an Index Out of Range error
(assuming the directive {$R + } is active).

Output Parameters

Degree: Integer: Degree of the deflated polynomial (> 2 if some of the roots are
not approximated).
NumRoots: Integer: Number of roots found.
Poly:TNvector; Coefficients of the deflated polynomial.
Root:TNvector: Real part of all roots found.
Imag:TNvector: Imaginary part of all roots found (nonzero for 2 at most).
Value:TNvector: Value of the polynomial at each approximate root.
Deriv:TNvector: Value of the derivative at each found root.
Iter:TNIntVector: Number of iterations required to find each root.
Error:Byte: 0: No error.
1: Maximum number of iterations exceeded.
2: The slope is zero (see "Comments").
3: Degree =:; O.
4: Tol =:; O.
5: Maxlter < O.
If a root is found, it is returned with the value of the polynomial at that root
(which should be close to zero) and with the value of the derivative at that root. If
the last two roots are complex (only two can be complex, since they are evaluated
by the quadratic formula), then the value and derivative at those points are arbi-
trarily set to zero. If all the roots have not been found, then the unsolved deflated
polynomial is also returned.

Roots to Equations in One Variable 29

Syntax of the Procedure Call

Newt-Horn_Defl(InitDegree, InitPoly, Guess, Tol, MaxIter, Degree,

NumRoots, Poly, Root, Imag, Value, Deriv, Iter, Error);

Comments

Newton's method involves division by the derivative of the function. Should the
algorithm attempt to do any calculations at a point where the absolute values of the
derivative are less than TNNearlyZero, the routine stops and returns an error mes-
sage (Error = 2).
Convergence is determined with the Boolean function TestForRoot described at
the beginning of this chapter.

Sample Program

The sample program NEWTDEFL.PAS provides I/O functions that demonstrate

the Newton-deflation algorithm.

Input Files
It is possible to input the coefficients from a text file. The format for the text file is
as follows:
1. The degree of the polynomial
2. The coefficients in descending order, beginning with the leading coefficient
and decreasing to the constant term

Spaces or carriage returns can be used to separate the data. It does not matter
whether the file ends with a carriage return; for example, the polynomial
F(x) =x 3
- 2x
could be entered in a text file as
310 -20

Example
Problem. Determine the roots to the 7th degree polynomial:
Xfi + x5 - 49x·' + 69x3 + 120x2 + 98x - 240

30 Turbo Numerical Methods Toolbox

Run NEWTDEFL.PAS:
(K)eyboard or (F)ile input of data? K
Degree of the pol ynomi a1 « = 30)? 6
Input the coefficients of the polynomial
where Poly[n] is the coefficient of x~n
Poly[6] 1
Poly[5] 1
Poly[4] -49
Poly[3] 69
Poly[2] 120
Poly[l] 98
Po 1y [0] -240
Initial approximation to the root: 0
Tolerance (> 0, default = 1.000E-08): 1E-8
Maximum number of iterations (>= 0, default 100): 100
Direct output to one of the following:
(S)creen
(P)rinter
(F) i1 e

Initial Polynomial:
Poly[6]: 1.00000000000000E+000
Poly[5]: 1.00000000000000E+000
Poly[4]: -4.90000000000000E+001
Poly[3]: 6.90000000000000E+001
Poly[2]: 1.20000000000000E+002
Poly[l]: 9.80000000000000E+001
Poly[O]: -2.40000000000000E+002
Initial approximation: O.OOOOOOOOOOOOOOE+OOO
Tolerance: 1.00000000000000E-008
Maximum number of iterations: 100
Number of calculated roots: 6
Root 1
Number of iterations: 7
Calculated root: 3.00000000000000E+000
Value of the function
at the calculated root: 4.83169060316868E-013
Value of the derivative
of the function at
the calculated root: -7.47999999999999E+002

Roots to Equations in One Variable 31

Root 2
Number of iterations: 7
Calculated root: 1.00000000000000E+000
Value of the function
at the calculated root: O.OOOOOOOOOOOOOOE+OOO
Value of the derivative
of the function at
the calculated root: 3.60000000000000E+002
Root 3
Number of iterations: 32
Calculated root: -8.00000000000000E+000
Value of the function
at the calculated root: O.OOOOOOOOOOOOOOE+OOO
Value of the derivative
of the function at
the calculated root: -6.43500000000000E+004
Root 4
Number of iterations: 25
Calculated root: 5.00000000000000E+000
Value of the function
at the calculated root: O.OOOOOOOOOOOOOOE+OOO
Value of the derivative
of the function at
the calculated root: 3.84800000000000E+003
Root 5
Number of iterations: 0
Calculated root: -1.00000000000000E+000 + -1.00000000000000E+000i
Value of the function
at the calculated root: O.OOOOOOOOOOOOOOE+OOO
Value of the derivative
of the function at
the calculated root: O.OOOOOOOOOOOOOOE+OOO
Root 6
Number of iterations: 0
Calculated root: -1.00000000000000E+000 + 1.00000000000000E+000i
Value of the function
at the calculated root: O.OOOOOOOOOOOOOOE+OOO
Value of the derivative
of the function at
the calculated root: O.OOOOOOOOOOOOOOE+OOO

32 Turbo Numerical Methods Toolbox

Complex Roots of a Complex Function Using Muller's
Metlwd (MULLERINC)

Description

This example uses Muller's method (Burden and Faires 1985, 71-75) to find a
possibly complex root of a user-defined complex function. The algorithm finds a
root of a parabola defined by three distinct points of the given function. This
approximation to the root and its two predecessors are used to construct the next
parabola. This is repeated until the convergence criteria is satisfied. Muller's
method has the advantage of nearly always converging; however, it is slow because
it uses complex arithmetic. You must create a complex function, input an initial
guess (which need not be very accurate), the tolerance in the answer, and the
maximum number of iterations.

User- Defined Types

TNcomplex = record
Re, Im:Real;
end;

User- Defined Procedure

procedure TNTargetF(x:TNcomplex; var y:TNcomplex);

The Muller procedure approximates a complex root of this function.

Input Parameters

Guess:TNcomplex; An initial guess

Tol:Real; Indicates accuracy in solution
MaxIter: Integer; Maximum number of iterations

Roots to Equations in One Variable 33

The preceding parameters must satisfy the following conditions:
1. Tol > 0
2. MaxIter;::: 0

Output Parameters

Answer:TNcomplex; An approximate root of the function

yAnswer: TNcomp 1ex; Value of the function at the approximate root
Iter:Integer; Number of iterations required to find the root
Error:Byte; 0: No error
1: Iter> MaxIter
2: Parabola could not be formed (see "Comments")
3: Tol :5 0
4: MaxIter < 0
If Error :5 2, then the information from the last iteration is output.

Syntax of the Procedure Call

Muller(Guess, Tol, MaxIter, Answer, yAnswer, Iter, Error);

The procedure Muller approximates a complex root of function TNTargetF.

Comments

Muller's method involves constructing a parabola from three points. If they all lie
on a line whose slope in absolute value is less than TNNearlyZero, then a parabola
that intersects the x-axis cannot be constructed. Such a construction will halt the
algorithm and return Error = 2. Fortunately, this does not commonly occur.
Convergence is determined with the Boolean function TestForRoot described at
the beginning of this chapter. Complex arithmetic is used.

34 Turbo Numerical Methods Toolbox

Sample Program

The sample program MULLERPAS provides I/O functions that demonstrate

Miiller' s method.
The user-defined function is contained in the procedure TNTargetF. It is neces-
sary to separately define the real and complex parts of the function. To define the
complex function F(x), you must code the following definitions:
y.Re : = Re[F(x.Re + ix.Im)];
y.Im : = Im[F(x.Re + ix.Im)];
where i is the square root of - 1.
For example, the complex function F(x) : = exp(x) would be coded like this:
y.Re : = exp(x.Re) * cos(x.Im);
y.Im : = exp(x.Re) * sin(X.Im);
Note that the procedure TNTargetF is defined before MULLER.lNC is
included.

Example
Problem. Find a solution to the complex equation cos(x) = x.
1. First, code the following procedure TNTargetF into MULLERPAS:
(*------------- HERE IS THE FUNCTION ------------------*)
procedure TNTargetF(x : TNcomplex; var y : TNcomplex);
beg;n { this is the complex function y = Cos (x) - x }
y.Re := Cos(x.Re)*(Exp(-x.lm) + Exp(x.lm))/2 - x.Re;
y.lm := Sin(x.Re)*(Exp(-x.lm) - Exp(x.lm))/2 - x.lm;
end; { procedure TNTargetF }

(*-----------------------------------------------------*)
2. Run MULLERPAS:
Initial approximation to the root:
Re(Approximation)= -4
Im(Approximation)= 4
Tolerance (> 0, default = 1.000E-08): 1E-6
Maximum number of iterations (>= 0, default = 100): 100

Roots to Equations in One Variable 35

Direct output to one of the following:
(S}creen
(P}rinter
(F) il e

Initial approximation: -4.00000000000000EtOOO t 4.00000000000000EtOOOi

Tolerance: 1.00000000000000E-006
Maximum number of iterations: 100

Number of iterations: 18
Calculated root: -9.10998745393294EtOOO t 2.95017086170180EtOOOi
Value of the function
at the calculated root: -1.42534872793476E-011 t 3.75033337718378E-011i

36 Turbo Numerical Methods Toolbox

Complex Roots of a Complex Polynomial Using Laguerre's
Method and Deflation (LAGUERRE.INC)

Description

This example uses Laguerre's method (Ralston and Rabinowitz 1978, 380-383) and
linear deflation to find the possibly complex roots of a complex (or real) polynomial.
You must input the coefficients of the polynomial, an initial guess, the tolerance
with which to find the answer, and the maximum number of iterations.

User-Defined Types

TNcomplex = record
Re, Im:Real;
end;
TNIntVector = array[O •• TNArraySize] of Integer;
TNCompVector = array[O •• TNArraySize] of TNcomplex;

Input Parameters

Degree: Integer; Degree of the user's polynomial

Poly:TNvector; Coefficients of the user's polynomial
InitGuess:TNcompl ex; Initial guess of the root
Tol :Real; Indicates accuracy in solution
MaxIter: Integer; Maximum number of iterations
The preceding parameters must satisfy the following conditions:
1. degree> 0
2. Tol> 0
3. MaxIter ~ 0

4. degree ~ TNArraySize

Roots to Equations in One Variable 37

TNArraySize fixes an upper bound on the number of elements in each vector. It
is used in the type definition of TNvector. TNArraySize is not a variable name and
is never referenced by the procedure; hence there is no test for condition 4. If
condition 4 is violated, the program will crash with an Index Out of Range error
(assuming the directive {$R + } is enabled).

Output Parameters

Degree:Integer; Degree of the deflated polynomial

Poly:Integer; Coefficients of deflated polynomial
NumRoots:Integer; Number of approximate roots
Roots :TNCompVector; Approximate roots
yRoots :TNCompVector; Value of the polynomial at the approximate root
Iter:TNlntVector; Number of iterations required to find each root
Error:Byte; 0: No error
1: Iter ~ Max Iter
2: Degree:::;; 0
3: Tal:::;; 0
4: MaxIter < 0

Syntax of the Procedure Call

Laguerre(Degree, Poly, InitGuess, Tol, MaxIter, NumRoots,

Roots, yRoots, Iter, Error);

Comments

For some polynomials, certain starting values (Guess) will not yield convergence. If
the routine does not converge to a solution, try a different starting value. Note that
convergence is slower around multiple roots than around single roots.
Convergence is determined with the Boolean function TestForRaat described at
the beginning of this chapter.

38 Turbo Numerical Methods Toolbox

Sample Program

The sample program LAGUERRE.PAS provides I/O routines that demonstrate

Laguerre's method.

Input Files
It is possible to input the coefficients from a text file. The format for the text file is
as follows:
1. The degree of the polynomial
2. The real and imaginary parts of the coefficients in descending order, begin-
ning with the leading coefficient and descending to the constant term

Spaces or carriage returns can be used to separate the data. It does not matter
whether the file ends with a carriage return; for example, the polynomial
F(x) = x-l - (2 + 2i)x3 + 4ix2 + (2 - 2i)x -1
where i represents the square root of - 1, could be entered in a text file like this:
410 -2 -2042 -2 -10

Example
Problem. Find all the roots to the complex polynomial
F(x) = l - (2 + 2i)x3 + 4ix2 + (2 - 2i)x - 1
where i is the square root of - 1.
Run LAGUERRE.PAS:
(K)eyboard or (F)ile input of data? K
Degree of the polynomial «= 30)? 4
Input the coefficients of the polynomial
where Poly[n] is the coefficients of x~n
Re(Poly[4]) = 1
Im(Poly[4]) = 0
Re(Poly[3]) = -2
Im(Poly[3]) = -2
Re(Poly[2]) 0
Im(Poly[2]) = 4

Roots to Equations in One Variable 39

Re{Poly[l]) = 2
Im{Poly[l]) = -2
Re{Poly[O]) = -1
Re{Poly[O]) = 0

Initial approximation:
Re{Approximation) = 0
Im{Approximation) = 0
Tolerance (> 0, default = 1.000E-08): 1E-6
Maximum number of iterations (>= 0, default = 100): 100
Direct output to one of the following:
{S)creen
{P)rinter
(F) il e

Initial Polynomial:
Poly[4]: 1.00000000000000E+000 + O.OOOOOOOOOOOOOOE+OOOi
Poly[3]: -2.00000000000000E+000 + -2.00000000000000E+000i
Poly[2]: O.OOOOOOOOOOOOOOE+OOO + 4.00000000000000E+000i
Poly[l]: 2.00000000000000E+000 + -2.00000000000000E+000i
Poly[O]: -1.00000000000000E+000 + O.OOOOOOOOOOOOOOE+OOOi

Initial approximation: O.OOOOOOOOOOOOOOE+OOO + O.OOOOOOOOOOOOOOE+OOOi

Tolerance: 1.00000000000000E-006
Maximum number of iterations: 100
Root 1
Number of iterations: 14
Calculated root: 9.99999949924438E-001 + 1.30979567319146E-008i
Value of the function at
the calculated root: -2.44249065417534E-015 + -4.67073566002583E-015i
Root 2
Number of iterations: 2
Calculated root: 1.00000001879739E+000 + -3.01914119849697E-009i
Value of the function at
the calculated root: -3.33066907387547E-016 + -1.03591657956252E-015i
Root 3
Number of iterations: 6
Calculated root: 2.54057206426756E-007 + 9.99999844722996E-001i
Value of the function at
the calculated root: -1.57873714101697E-013 + -8.07179393811825E-014i

40 Turbo Numerical Methods Toolbox

Root 4
Number of iterations: 2
Calculated root: -5.97412717946260E-008 + 1.00000003900209E+OOOi
Value of the function at
the calculated root: -9.32587340685131E-015 + -3.90669101062423E-015i
The exact roots of this polynomial are
x = 1
x= 1
x = i
x = i

Roots to Equations in One Variable 41

42 Turbo Numerical Methods Toolbox
c H A p T E R 3
Interpolation

Interpolation is useful when some values of a function are known but others are
required. For example, suppose values are known for a functionJ(x) at x = 2.3,2.4,
2.5, 2.6, 2.7, 2.8, and the value ofJ(x) is desired at x = 2.415. The routines in this
chapter provide the means to model to given values of J(x) with an appropriate
function, so that the function can be evaluated at other arbitrary points.
The goal of interpolation is to approximate the value of the function at a speci-
fied value of x, given N values of the function at N distinct points. This approxima-
tion will be a polynomial determined from the input data. The value of the
polynomial at x will be returned as the approximation to the value ofJ(x).
The Lagrange method (LAGRANGE.INC) accepts points in any order. The x-
values need not be equally spaced. An interpolating polynomial is explicitly calcu-
lated. Although an interpolating polynomial can be useful for computing deriva-
tives (and more), the Lagrange method is a lengthy process. Furthermore, high-
degree polynomials may cause significant round-off error in some interpolations.
Newton's general divided-difference algorithm (DIVDIF.INC) does not require
input to have equally spaced x-values, nor is it necessary that the x-values be
in either ascending or descending order. For large amounts of data, the divided-
difference routine (DIVDIF.INC) is more accurate than Lagrangian interpola-
tion (LAGRANGE.INC).
If there are many input points, the Lagrange (LAGRANGE.INC) and the
divided-difference (DIVDIF.INC) methods may result in high-degree polynomials
whose oscillatory nature can produce an inaccurate approximation. This is espe-

43
cially true if the interpolation occurs at a value near the midpoint between adjacent
input x-values. In such cases, the cubic spline methods (CUBEJ'RE.INC and
CUBE_CLA.INC) are preferable.
The cubic spline methods require that the x-values be entered in ascending
order. The clamped cubic spline method (CUBE_CLA.lNC) may yield more accu-
rate results than the free cubic spline method (CUBEJ'RE.lNC), but requires
knowledge of the first derivative of the function at the endpoints of the input data.
When this information is not available; the free cubic spline routine should be
used.
The values at which interpolation is to occur should lie in the closed interval
bounded by the extreme values of the input x-values. The preceding methods will
not give accurate approximations to values outside this interval (extrapolation).

44 Turbo Numerical Methods Toolbox

Polynomial Interpolation Using Lagrange's Metlwd
(LAGRANGE. INC)

Description

This example provides an interpolation algorithm (Burden and Faires 1985, 84 If;
Horowitz and Sahni 1984,429-430). Given a set of N data points (x,y), the routine
uses Lagrange polynomials to construct a polynomial to fit the data points. The
degree of the polynomial is at most N - l.
Note: The nature of high-degree polynomials may cause significant error if the
algorithm is used with large amounts of data (about N > 25). In such cases, DIV
DIF.INC, CUBEJ'RE.INC, or CUBE_CLA.INC should be used. You must sup-
ply the data points and the x-values at which interpolation will take place.

User- Defined Types

TNvector = array[O .. TNArraySize] of Real;

TNmatrix = array[O .. TNArraySize] of TNvector;

Input Parameters

The parameters for Lagrange:

NumPoi nts: Integer; Number of data points
XData:TNvector; The x-coordinates of the data points
YData:TNvector; The y-coordinates of the data points
Numlnter: Integer; Number of interpolations
Xlnter:TNvector The x-coordinates at which interpolation is to take place
The preceding parameters must satisfy the following conditions:
1. The x-coordinates of the data points (Xlnter) must be unique.
2. NumPoints, Numlnter ~ TNArraySize.
3. NumPoints > O.

Interpolation 45
TNArraySize fixes an upper bound on the number of elements in each vector. It
is used in the type definition of TNvector. TNArraySize is not a variable name and
is never referenced by the procedure; hence there is no test for condition 2. If
condition 2 is violated, the program will crash with an Index Out of Range error
(assuming the directive {$R + } is active).

Output Parameters

YInter:TNvector; The interpolated values at Xlnter

Poly:TNvector; The coefficients of the interpolating polynomial
Error:Byte; 0: No error
1: X-values of the data points not unique
2: NumPoints < 1

Syntax of the Procedure Call

Lagrange(NumPoints, XData, YData, NumInter, XInter, YInter, Poly, Error);

Sample Program

The sample program LAGRANGE.PAS provides I/O functions that demonstrate

the Lagrange interpolating algorithm.

Input Files
Data may be entered from a text file. The x and y coordinates should be separated
by a space and followed by a carriage return. For example, data values of sqr(x)
could be entered in a text file as:
11
24
39
416
525

46 Turbo Numerical Methods Toolbox

Example
Problem. Construct and use an interpolating polynomial for the cosine function
between x = 1 degree and x = 20 degrees.
Run LAGRANGE.PAS:
(K)eyboard or (F)ile entry of the data points? F

File name? SAMPLE3A.DAT

(K)eyboard or (F)ile entry of the interpolated points? F

File name? SAMPLE3B.DAT

Direct output to one of the following:

(S)creen
(P)rinter
(F) il e

The data:
1.0000000 9.99847695156391E-001
2.0000000 9.99390827019096E-001
3.0000000 9.98629534754574E-001
4.0000000 9.97564050259824E-001
5.0000000 9.96194698091746E-001
6.0000000 9.94521895368273E-001
7.0000000 9.92546151641322E-001
8.0000000 9.90268068741570E-001
9.0000000 9.87688340595138E-001
10.0000000 9.84807753012208E-001
11.0000000 9.81627183447664E-001
12.0000000 9.78147600733806E-001
13.0000000 9.74370064785235E-001
14.0000000 9.70295726275996E-001
15.0000000 9.65925826289068E-001
16.0000000 9.61261695938319E-001
17.0000000 9.56304755963035E-001
18.0000000 9.51056516295154E-001
19.0000000 9.45518575599317E-001
20.0000000 9.39692620785908E-001
The polynomial:
Poly[19]= -1.72986376643586E-028
Poly[18]= 3.57504241395844E-026
Poly[17]= -3.43926153199199E-024
Poly[16]= 2.04507188280402E-022
Poly[15]= -8.41710490427928E-021
Poly[14]= 2.54454663251946E-019
Poly[13]= -5.85115478257286E-018
Poly[12]= 1.04567701944119E-016
Poly[ll]= -1.47131277721604E-015
Poly[10]= 1.64108936708876E-014
Poly[9]= -1.45382580196402E-013
Poly[8]= 1.02034999744286E-012
Poly[7]= -5.63354870368428E-012
Poly[6]= 2.41324946244329E-011
Poly[5]= -7.90989844502363E-011

Interpolation 47
Poly[4] = 4.05827439744465E-009
Poly[3]= -3.31023555675263E-010
Poly[2]= -1.52308331145220E-004
Poly[l]= -2.53687934078865E-010
Poly[O]= 1.00000000007368E+000

X Interpolated Y value Actual values

1.500 9.99657324975249E-001 9.99657324975557E-00 1
2.500 9.99048221581889E-001 9.99048221581858E-001
3.500 9.98134798421861E-001 9.98134798421867E-001
4.500 9.96917333733130E-001 9.96917333733128E-001
5.500 9.95396198367178E-001 9.95396198367179E-001
6.500 9.93571855676587E-001 9.93571855676587E-001
7.500 9.91444861373810E-001 9.9144486137381OE-001
8.500 9.89015863361917E-001 9.89015863361917E-001
9.500 9.86285601537231E-001 9.8628560 1537231E-001
10.500 9.83254907563954E-001 9.83254907563955E-001
11.500 9.79924704620830E-001 9.79924704620830E-001
12.500 9.76296007119933E-001 9.76296007119933E-001
13.500 9.72369920397676E-001 9.72369920397677E-001
14.500 9.68147640378107E-001 9.68147640378108E-001
15.500 9.63630453208623E-001 9.63630453208623E-00 1
16.500 9.58819734868193E-001 9.58819734868193E-001
17.500 9.53716950748226E-001 9.53716950748227E-001
18.500 9.48323655206200E-001 9.48323655206199E-001
19.500 9.42641491092201E-001 9.42641491092178E-001
20.500 9.36672189247006E-001 9.36672189248398E-00 1

The data is taken from a function of which the derivative could be computed
exactly. Though the actual values in the right-hand column are not displayed on
screen, they are shown here to indicate the accuracy of the routine.

48 Turbo Numerical Methods Toolbox

Interpolation Using Newton's Interpolary
Divided-Difference Metlwd (DWDIEINC)

Description

This example provides an interpolation algorithm. Given a set of data points (x,y),
the routine uses Newton's interpolary divided-difference equation to interpolate
between the points (Burden and Faires 1985,100-102). The data points must have
unique x-values, but these values need not be evenly spaced nor set in any particu-
lar order. You must supply the data points and the x-values at which interpolation is
to take place.

User-Defined Types

TNvector = array[O •. TNArraySize] of Real;

TNmatrix = array[O •• TNArraySize] of TNvector;

Input Parameters

NumPoi nts: Integer; Number of data points

XData:TNvector; The x-coordinates of the data points
YData:TNvector; The y-coordinates of the data points
NumInter: Integer; Number of interpolations
XInter:TNvector The x-coordinates at which interpolation is to take place
The preceding parameters must satisfy the following conditions:
1. The x-coordinates of the data points (Xlnter) must be unique.
2. NumPoints, Numlnter ~ TNArraySize.
3. NumPoints > o.
TNArraySize fixes an upper bound on the number of elements in each vector. It
is used in the type definition of TNvector. TNArray Size is not a variable name and
is never referenced by the procedure; hence there is no test for condition 2. If
condition 2 is violated, the program will crash with an Index Out of Range error
(assuming the directive {$R + } is active).

Interpolation 49
Output Parameters

Ylnter:TNvector; The interpolated values at Xlnter

Error:Byte; 0: No error
1: X-values of the data points not unique
2: NumPoints < 1

Syntax of the Procedure Call

Divided-Difference(NumPoints, XData, YData, Numlnter, Xlnter, Ylnter, Error);

Sample Program

The sample program DIVDIF.PAS provides I/O functions that demonstrate New-
ton's interpolary divided-difference algorithm.

Input Files
Data may be entered from a text file. The x and y coordinates should be separated
by a space and followed by a carriage return. For example, data values of sqr(x}
could be entered in a text file as
11
24
39
416
525

Example
Problem. Interpolate the cosine function between x = Ix and x = 20x.
Run DIVDIF.PAS:
(K)eyboard or (F)ile entry of the data points? F
File name? SAMPLE3C.DAT
(K)eyboard or (F)ile entry of the interpolated points? K

50 Turbo Numerical Methods Toolbox

Number of points (0-50}?15

Point 1: 1.5
Point 2: 2.5
Point 3: 3.5
Point 4: 4.5
Point 5: 5.5
Point 6: 6.5
Point 7: 7.5
Point 8: 8.5
Point 9: 9.5
Point 10: 10.5
Point 11: 11.5
Point 12: 12.5
Point 13: 13.5
Point 14: 14.5
Poi nt 15: 15.5

Direct output to one of the following:

(S}creen
(P}rinter
(F) il e

X y
12.000 0.9781476
8.000 0.9902681
1.000 0.9998477
10.000 0.9848078
5.000 0.9961947
15.000 0.9659258
4.000 0.9975641
3.000 0.9986295
7.000 0.9925462
14.000 0.9702957

X Interpolated Y value Actual Values

1.500 9.99656668284607E-001 9.99657324975557E-001
2.500 9.99047982204853E-001 9.99048221581858E-001
3.500 9.98134846782587E-001 9.98134798421867E-001
4.500 9.96917355869352E-001 9.96917333733128E-001
5.500 9.95396200633579E-001 9.95396198367179E-001
6.500 9.93571893532269E-001 9.93571855676587E-001
7.500 9.91444906399794E-001 9.9144486137381OE-001
8.500 9.89015879894104E-001 9.890 15863361917E-001
9.500 9.86285623948171E-001 9.86285601537231E-OOI
10.500 9.83254980952454E-001 9.83254907563955E-OOl
11. 500 9.79924765142406E-001 9.79924 704620830E-00 1
12.500 9.76295923083642E-001 9.76296007119933E-001
13.500 9.72369781236267E-001 9. 72369920397677E-OO 1
14.500 9.68147757339141E-001 9.68147640378108E-001
15.500 9.63629212784399E-001 9.63630453208623E-OO 1

The data is taken from a function of which the derivative could be computed
exactly. Though the values in the right-hand column are not displayed on screen,
they are shown here to indicate the accuracy of the routine.

Interpolation 51
Free Cubic Spline Interpolation (CUBEJRE.INC)

Description

This example constructs a smooth curve through a given set of data points. The
curve is a cubic spline interpolant with the following properties:
1. It passes through every data point.
2. It is continuous.
3. Its first derivative is continuous.
4. Its second derivative is continuous.

The second derivative is assumed to be zero at both endpoints (thus the cubic
spline is "free") of the interval determined by the data (Burden and Faires 1985,
117 fI). Cubics that join adjacent data points are of the following form:
S[i](x) = CoefO[i] + Coefl[i](x - x[i]) + Coef2[i](x - x[i]?
+ Coef3[i](x- X[i])3
where i ranges between 1 and the number of data points minus 1, the x[i]' s are the
x-coordinates of the input data, and x[i] :5 x < x[i + 1]. The interpolated values of
f(x) are found by evaluating the ith cubic polynomial at x, where
x[i] :5 x :5 x[i + 1].

User- Defined Types

TNvector = array[O •• TNArraySize] of Real;

Input Parameters

NumPoi nts: Integer; Number of data points

XData:TNvector; The x-coordinates of the data points
YData: TNvector; The y-coordinates of the data points
Numlnter: Integer; Number of interpolations
Xlnter:TNvector; X-coordinates of points at which to interpolate

52 Turbo Numerical Methods Toolbox

The preceding parameters must satisfy the following conditions:
1. X data points must be unique.
2. X data points must be in ascending order.
3. NumPoints, Numlnter ~ TNArraySize.
4. NumPoints > 1.

TNArraySize fixes an upper bound on the number of elements in each vector. It

is used in the type definition of TNvector. TNArraySize is not a variable name and
is never referenced by the procedure; hence there is no test for condition 3. If
condition 3 is violated, the program will crash with an Index Out of Range error
(assuming the directive {$R + } is active).

Output Parameters

CoefO:TNvector; Coefficient of the constant term

Coefl: TNvector; Coefficient of the linear term
Coef2:TNvector; Coefficient of the squared term
Coef3: TNvector; Coefficient of the cubed term
YInter:TNvector; Interpolated values at Xlnter
Error:Byte; 0: No error
1: X-values of the data points not unique
2: X-values of the data points not in ascending order
3: NumPoints < 2

Syntax of the Procedure Call

CubicSplineFree(NumPoints, XData, YData, Numlnter, Xlnter,

CoefO, Coefl, Coef2, Coef3, Ylnter, Error);

Sample Program

The sample program CUBE-FRE.PAS provides I/O functions that demonstrate

the free cubic spline algorithm.

Interpolation 53
Input Files
Data may be entered from a text file. The x and y coordinates should be separated
by a space and followed by a carriage return. For example, data values of sqr(x)
could be entered in a text file as
11
24
39
416
525

Example
Problem. Construct an interpolating spline for the following figure:

2 3 4 5 6

Because a cusp occurs at x = 3.55, we will construct two splines, one for each
side of the cusp.
Run CUBE_FRE.PAS:
(K)eyboard or (F)ile entry of the data points? F
File name? SAMPLE3D.DAT
(K)eyboard or (F)ile entry of the interpolated points? F
File name? SAMPLE3E.DAT

54 Turbo Numerical Methods Toolbox

Direct output to one of the following:
(S)creen
(P)rinter
(F) il e

Data: x y
1: 0.0000000000 2.8000000000
2: 0.1000000000 2.7000000000
3: 0.2000000000 2.6000000000
4: 0.6000000000 2.2000000000
5: 1.0000000000 1.8000000000
6: 1.4000000000 1.6000000000
7: 1.8000000000 1.4000000000
8: 2.0000000000 1.4200000000
9: 2.2000000000 1.4000000000
10: 2.6000000000 1.5000000000
11: 3.0000000000 1.8000000000
12: 3.4000000000 2.4000000000
13: 3.4500000000 2.6000000000
14: 3.5000000000 2.8000000000
15: 3.5500000000 2.9000000000

Splines: CoefO Coefl Coef2 Coef3

1: 2.800000 -0.9988332302 0.0000000000 -0.1166769808
2: 2.700000 -1.0023335396 -0.0350030942 0.5833849040
3: 2.600000 -0.9918326113 0.1400123770 -0.4010771215
4: 2.200000 -1.0723397281 -0.3412801689 1.3053237227
5: 1.800000 -0.7188084763 1.2251082984 -1.6952177695
6: 1.600000 -0.5524263669 -0.8091530249 2.3505473551
7: 1.400000 -0.0714860563 2.0115038012 -5.7703675978
8: 1.420000 0.0406713524 -1.4507167575 3.7367999767
9: 1.400000 -0.0911993534 0.7913632286 0.1540878869
10: 1.500000 0.6158534153 0.9762686929 -1. 6022555777
11: 1. 800000 0.6277856923 -0.9464380003 7.8174344240
12: 2.400000 3.6230038155 8.4344833084 -17 .8911923822
13: 2.600000 4.3322682035 5.7508044511 -247.9233704257
14: 2.800000 3.0479233704 -31. 4377011128 209.5846740851

Interpolated Points: X y
1: 0.3000000000 2.5018157855
2: 0.5000000000 2.3042222482
3: 1.2000000000 1. 6916808945
4: 1.6000000000 1.4759529845
5: 2.1000000000 1.4132967676
6: 2.3000000000 1.3989477848
7: 2.5000000000 1.4480232575
8: 2.7000000000 1.5697457729
9: 2.9000000000 1.7293593063
10: 3.2000000000 1.9502390938
11: 3.3000000000 2.1142270171

Interpolation 55
Second half of the figure:
(K)eyboard or (F)ile entry of the data points? F

File name? SAMPLE3F.DAT

(K)eyboard or (F)ile entry of the interpolated points? F

File name? SAMPLE3G.DAT

Direct output to one of the following:

(S)creen
(P)rinter
(F) il e

Data: X y
1: 3.5500000000 2.9000000000
2: 3.6000000000 2.8000000000
3: 3.6500000000 2.6500000000
4: 3.8000000000 2.5000000000
5: 4.0000000000 2.3500000000
6: 4.3000000000 2.2000000000
7: 4.8000000000 1.9500000000
8: 5.3000000000 1.6000000000
9: 5.6000000000 1.3000000000
10: 5.8000000000 1.2000000000
11: 6.0000000000 0.0000000000
Splines: CoefO Coefl Coef2 Coef3
1: 2.9000000000 -1. 6719664279 0.0000000000 -131.2134288293
2: 2.8000000000 -2.6560671441 -19.6820143244 256.0671441466
3: 2.6500000000 -2.7037649955 18.7280572976 -49.1308266290
4: 2.5000000000 -0.4016786037 -3.3808146854 8.1960385189
5: 2.3500000000 -0.7704798556 1.5368084259 -2.1173630243
6: 2.2000000000 -0.4200828166 -0.3688182960 0.4179678583
7: 1.9500000000 -0.4754252188 0.2581334916 -1.4145661079
8: 1.6000000000 -1.2782163082 -1.8637156703 9.3036778805
9: 1.3000000000 0.1155473174 6.5095944222 -47.9366550462
10: 1.2000000000 -3.0330135193 -22.2523986055 37.0873310092

Interpolated Points: X y
1: 3.7000000000 2.5554905401
2: 3.9000000000 2.4342200313
3: 4.1000000000 2.2862027357
4: 4.2000000000 2.2404374617
5: 4.5000000000 2.1045744477
6: 4.6000000000 2.0520666406
7: 5.0000000000 1.8539237670
8: 5.2000000000 1.7105990402
9: 5.5000000000 1.3442375346
10: 5.7000000000 1.3287140209
11: 5.9000000000 0.7112619930

56 Turbo Numerical Methods Toolbox

Clamped Cubic Spline Interpolation (CUBE_CLAINC)

Description

Tl1is example constructs a smooth curve through a given set of data points. The
curve is a cubic spline interpolant with the following properties:
1. It passes through every data point.
2. It is continuous.
3. Its first derivative is continuous.
4. Its second derivative is continuous.

The first derivative at the endpoints of the interval determined by the input data
is defined by the user (Burden and Faires 1985, l22 ff.). (This is what makes the
cubic spline "clamped.") The cubics that join adjacent data points are of the follow-
ing form:
S[i](x) = CoefO[i] + Coefl[i](x - x[i]) + Coef2[i](x - X[i])2
+ Coef3[i](x - x[i]?
where i ranges between 1 and the number of data points minus 1, the x[i]' s are the
x-coordinates of the input data, and x[i] S x < x[i + 1]. The interpolated values of
f(x) are found by evaluating the ith cubic polynomial at x, where x[i] S x S
x[i + 1].

User-Defined Types

TNvector = array[O .. TNArraySize] of Real;

Input Parameters

NumPoi nts : Integer; Number of data points

XData:TNvector; The x-coordinates of the data points
YData:TNvector; The y-coordinates of the data points
DerivLE:Real; Derivative of the function at the left endpoint
DerivRE:Real; Derivative of the function at the right endpoint

Interpolation 57
NumInter: Integer; Number of interpolations
XInter:TNvector; X-coordinates of points at which to interpolate
The preceding parameters must satisfy the following conditions:
1. X data points must be unique.
2. X data points must be in ascending order.
3. NumPoints, Numlnter S TNArraySize.
4. NumPoints > 1.

TNArraySize fixes an upper bound on the number of elements in each vector. It

is used in the type definition of TNvector. TNArraySize is not a variable name and
is never referenced by the procedure; hence there is no test for condition 3. If con-
dition 3 is violated, the program will crash with an Index Out of Range error
(assuming the directive {$R + } is active).

Output Parameters

CoefO:TNvector; Coefficient of the constant term

Coef1:TNvector; Coefficient of the linear term
Coef2: TNvector; Coefficient of the squared term
Coef3: TNvector; Coefficient of the cubed term
YInter:TNvector; Interpolated values at Xlnter
Error: Integer; 0: No error
1: X-values of the data points not unique
2: X-values of the data points not in ascending order
3: NumPoints < 2

Syntax of the Procedure Call

CubicSplineClamped(NumPoints, XData, YData, DerivLE, DerivRE, NumInter,

XInter, CoefO, Coefl, Coef2, Coef3, YInter, Error);

58 Turbo Numerical Methods Toolbox

Sample Program

The sample program CUBE_CLA.PAS provides I/O functions that demonstrate

the clamped cubic spline interpolation algorithm.

Input Files
Data may be entered from a text file. The x- and y -coordinates should be separated
by a space and followed by a carriage return. The last two values in the file must be
the derivatives of the function at the endpoints. For example, data values of sqr(x)
could be entered in a text file as
11
24
39
416
525
210
Note that the last two values are the derivatives of sqr(x) at the endpoints x =1
and x = 5.

Example
Problem. Construct an interpolating spline for the following figure:

2 3 4 5 6

Interpolation 59
Because a cusp occurs at x = 3.55, we will construct two splines, one for each
side of the cusp.
Run CUBE_CLA.PAS:
(K}eyboard or (F}ile entry of the data points? F

File name? SAMPLE3H.DAT

(K}eyboard or (F}ile entry of the interpolated points? F

File name? SAMPLE3E.DAT

Direct output to one of the following:

(S}creen
(P}rinter
(F) i1 e

Data: x y
1: 0.0000000000 2.8000000000
2: 0.1000000000 2.7000000000
3: 0.2000000000 2.6000000000
4: 0.6000000000 2.2000000000
5: 1.0000000000 1.8000000000
6: 1.4000000000 1.6000000000
7: 1.8000000000 1.4000000000
8: 2.0000000000 1.4200000000
9: 2.2000000000 1.4000000000
10: 2.6000000000 1.5000000000
11: 3.0000000000 1.8000000000
12: 3.4000000000 2.4000000000
13: 3.4500000000 2.6000000000
14: 3.5000000000 2.8000000000
15: 3.5500000000 2.9000000000

Derivative at X= O.OOOOOOOOOOOOOOE+OOO -1.33333333333333E+000

Derivative at X= 3. 55000000000000E+000 3.00000000000000E+000

Sp 1i nes: CoefO Coefl Coef2 Coef3

1: 2.8000000000 -1.3333333333 5.7579845570 -24.2465122365
2: 2.7000000000 -0.9091317890 -1.5159691140 6.0728700429
3: 2.6000000000 -1.0301395105 0.3058918989 -0.5763578064
4: 2.2000000000 -1.0620777385 -0.3857374687 1.3523295373
5: 1.8000000000 -0.7215495356 1.2370579761 -1.7079603429
6: 1.6000000000 -0.5517241193 -0.8124944355 2.3545118344
7: 1.4000000000 -0.0715539872 2.0129197658 -5.7757491499
8: 1.4200000000 0.0405240212 -1.4525297241 3.7495480911
9: 1.4000000000 -0.0905420975 0.7971991306 0.1353902832
10: 1.5000000000 0.6122045428 0.9596674704 -1.5379470688
11: 1.8000000000 0.6417239262 -0.8858690121 7.5788979919
12: 2.4000000000 3.5708997526 8.2088085781 7.4639274157
13: 2.6000000000 4.4477600660 9.3283976905 -365.6719802043
14: 2.8000000000 2.6380599835 -45.5223993401 655.2239934014

60 Turbo Numerical Methods Toolbox

Interpolated Points: X y
1: 0.3000000000 2.4994686101
2: 0.5000000000 2.3029267570
3: 1.2000000000 1.6915087292
4: 1.6000000000 1.4759914934
5: 2.1000000000 1.4132766530
6: 2.3000000000 1.3990531718
7: 2.5000000000 1.4482408301
8: 2.7000000000 1.5692791819
9: 2.9000000000 1.7285068643
10: 3.2000000000 1.9535412087
11: 3.3000000000 2.1174192125
Second half of figure:
(K)eyboard or (F)ile entry of the data points? F
File name? SAMPLE3I.DAT

(K)eyboard or (F)ile entry of the interpolated points? F

File name? SAMPLE3G.DAT

Direct output to one of the following:

(S)creen
(P)rinter
(F)ile

Data: X y
1: 3.5500000000 2.9000000000
2: 3.6000000000 2.8000000000
3: 3.6500000000 2.6500000000
4: 3.8000000000 2.5000000000
5: 4.0000000000 2.3500000000
6: 4.3000000000 2.2000000000
7: 4.8000000000 1.9500000000
8: 5.3000000000 1.6000000000
9: 5.6000000000 1.3000000000
10: 5.8000000000 1.2000000000
11: 6.0000000000 0.0000000000

Derivative at X= 3.55000000000000EtOOO -4.00000000000000EtOOO

Derivative at X= 6.00000000000000EtOOO: -1.70000000000000EtOOl
Sp 1i nes: CoefO Coefl Coef2 Coef3
1: 2.9000000000 -4.0000000000 80.2233303937 -804.4666078741
2: 2.8000000000 -2.0111665197 -40.4466607874 413.3998236224
3: 2.6500000000 -2.9553339213 21.5633127559 -56.8516885392
4: 2.5000000000 -0.3238290709 -4.0199470867 9.4454622054
5: 2.3500000000 -0.7983524409 1.6473302365 -2.1760736673
6: 2.2000000000 -0.3974941891 -0.3111360640 0.2122488846
7: 1.9500000000 -0.5494435897 0.0072372629 -0.6167001671
8: 1.6000000000 -1.0047314521 -0.9178129877 3.1119483153
9: 1.3000000000 -0.7151931996 1.8829404961 -4.0348724916
10: 1.2000000000 -0.4462017001 -0.5379829989 -136.1550425028

Interpolation 61
Interpolated Points: X y
1: 3.7000000000 2.5490351248
2: 3.9000000000 2.4368630843
3: 4.1000000000 2.2844619846
4: 4.2000000000 2.2388141319
5: 4.5000000000 2.1097537107
6: 4.6000000000 2.0584802174
7: 5.0000000000 1.8354671712
8: 5.2000000000 1.6919117155
9: 5.5000000000 1.3872367766
10: 5.7000000000 1.2432752125
11: 5.9000000000 1.0138449575

62 Turbo Numerical Methods Toolbox

c H A p T E R 4
Numerical Differentiation

Differentiation is a process used in calculus to quantify the rate of change of a given

function. The derivative of a real-valued function of a real variable is another real-
valued function of a real variable. For example, suppose you are driving down the
freeway in your car and f(t) gives the distance traveled at time t. Typical values
might be

x f(x)
1.0 45.0
1.1 49.2
1.2 54.5
1.3 59.8
1.4 65.1
1.5 70.4
The units are in hours and miles, and the data refers to a trip that started at
noon.f(l.O) = 45.0, so the distance traveled by one o'clock is 45.0 miles, andf(l.5)
= 70.4, so by half past one you will be 70.4 miles from where you were at noon.
The derivative of this distance function gives the velocity function. The car's
velocity at one o'clock is the value of the derivative at x = l.0. From the previous
data, it is impossible to compute the derivative exactly, but it is possible to approxi-
mate the derivative. The car traveled 49.2 - 45.0 = 4.2 miles in the six minutes
after one o'clock (1.1 - 1.0 = 0.1 hours = 6 minutes). Thus, the average velocity
of the car during thos~ six minutes is 4.2 / 0.1 = 42 miles per hour. This gives an
approximation to the velocity at one o'clock.

63
Each method described in this chapter approximates derivatives of a real func-
tion of one real variable.
The routines DERIV.lNC, DERIV2.1NC, and INTERDRV.lNC compute
derivatives of a function that is represented by tabular data. Consequently, their
accuracy depends heavily upon the precision and spacing of the data points.
The routines DERIVFN.lNC and DERIV2FN.lNC compute derivatives of a
user-defined function. Consequently, the accuracy of the values calculated with
these routines is limited by the precision of the computer.
Differentiation consists of subtracting two very close numbers and dividing by a
very small number; hence, it is extremely sensitive to round-off error. The accuracy
of the first derivative is approximately the square root of the precision with which
real numbers are represented; the accuracy of the second derivative is approxi-
mately equal to the fourth root. Thus, the precision of the first derivative will be
about IE - 8 when run with the 8087 math coprocessor, or about IE - 4 when run
without the 8087 math coprocessor. The precision of the second derivative will be
about IE - 4 with the 8087, or IE - 2 without it.
The first derivative of a function that is represented by a table of values can be
approximated in DERIV.lNC via a two-point formula, a three-point formula, or a
five-point formula. The accuracy of the formula increases with the number of
points used in the formula. In order to use the five-point formula, however, the
domain values of the data points (that is, the x-coordinates) must be equally spaced.
This is not required for the two-point and three-point formulas. Derivatives can
only be approximated at data points.
The second derivative of a function that is represented by a table of values can
be approximated in DERIV2.1NC via a three-point formula or a five-point for-
mula. The domain values of the data points must be equally spaced (regardless of
whether the three-point formula or five-point formula is used). Second derivatives
can only be approximated at data points.
The routine INTERDRV.lNC approximates a function by constructing a free
cubic spline to a set of data points. Cubic splines avoid the undesirable oscillatory
behavior of other interpolating polynomials. The derivative of the cubic spline at a
given domain value, which may be different from the input data values, will then
approximate the corresponding derivative of the function.
The first derivative of a user-supplied function is approximated in DERIV
FN.lNC via a three-point formula. The approximation is refined with Richardson
extrapolation. The derivative can be approximated at any point within the domain
of the function.

64 Turbo Numerical Methods Toolbox

The second derivative of a user-supplied function is approximated in
DERIV2FN.INC via a three-point formula. The approximation is refined with
Richardson extrapolation. The second derivative can be approximated at any point
within the domain of the function.

Numerical Differentiation 65
First Differentiation Using Two-Point, Three-Point, or
Five-Point Formulas (DERIV.INC)

Description

This example contains several algorithms for approximating the derivative of a

functionj(x), given several data points (x,j(x)). The user must specify whether a
two-point, three-point, or five-point formula should be used. Two points are used
in the two-point formula, three in the three-point formula, and five in the five-
point formula. The user must supply the data points (x,j(x)) and the x-values of the
data points at which to approximate the derivative. Note: Derivatives can only be
approximated at x-values corresponding to input data points.

User-Defined Types

TNvector = array[l •• TNArraySize] of Real;

Input Parameters

NumPoi nts : Integer; Number of data points

XData : TNvector; X-coordinates of data points
YData : TNvector; Y-coordinates of data points
Poi nt : Byte; Two-point, three-point, or five-point differentiation
NumDeriv : Integer; Number of points at which the derivative is to be approxi-
mated
XDeriv : TNvector; X-coordinates of data points at which the derivative is to be
approximated
The preceding parameters must satisfy the following conditions:
1. XData points must be unique.
2. XData points must be entered in ascending order.
3. At least two points are needed for two-point differentiation, three for
three-point differentiation, and five for five-point differentiation.
4. Point must equal two, three, or five.

66 Turbo Numerical Methods Toolbox

5. XData points must be equally spaced for five-point differentiation.
6. XDeriv points must be a subset of the XData points.
7. NumPoints, NumDeriv ::::;; TNArraySize.

TNArraySize represents the number of elements in each vector. It is used in the

type definition of TNvector. TNArraySize is not a variable name and is never refer-
enced by the procedure; hence there is no test for condition 7. If condition 7 is
violated, the program will crash with an Index Out of Range error (assuming the
directive {$R + } is active).

Output Parameters

YDeriv : TNvector; Approximation to the first derivative at the points in XDeriv

Error: Byte; 0: No errors
1: WARNINGI Not all the derivatives were computed (see
"Comments")
2: X-values not unique
3: X-values not in ascending order
4: Not enough data
5: Point not equal to 2, 3, or 5
6: X-values not equally spaced for the five-point formula

Syntax of the Procedure Call

First_Derivative(NumPoints, XData, YData, Point, NumDeriv, XDeriv, YDeriv, Error);

Comments

If an x-value at which the derivative is to be approximated is not among the data

points, the value - 9.999999999E35 is arbitrarily assigned to the derivative at that
point and Error = 1 is returned. When using five-point differentiation with only
five points, there is not enough information to approximate the derivative at the
first, second, fourth, or fifth points. Likewise, if only six points are input, there is
insufficient information for approximating the derivative at the second and fifth
data points. Should an attempt be made to approximate the derivative at any of
these points, the value of 9.999999999E35 is arbitrarily assigned the derivative at
that point and Error = 1 is returned.

Numerical Differentiation 67
Since numerical differentiation is prone to round-off errors, TNNearlyZero is
different in this routine. The values of TNNearlyZero are TNNearlyZero = IE - 13
if using the 8087 math coprocessor and TNNearlyZero = IE - 6 if not using the
8087.

Sample Program

The sample program DERIV.PAS provides I/O functions that demonstrate differ-
entiation with two-point, three-point, and five-point formulas.

Input Files
Data points may be entered from a text file. The x- and y- coordinates should be
separated by a space and followed by a carriage return. For example, data values of
sqr(x) could be entered in a text file as
11
24
39
416
525
Derivative points may also be entered from a text file. Every derivative point
must be followed by a carriage return. For example, to determine the derivatives of
the preceding points, create the following file of derivative points:
1
2
3
4
5

Example
Problem. Approximate the first derivative off(x) = sqr(x) * cos (x) at several points
between one and two radians. The output from three runs is given. Actual values of
the derivatives to eight significant figures are also given.

68 Turbo Numerical Methods Toolbox

Run DERrV.PAS:
(K)eyboard or (F)ile entry of the data points? F
File name? SAMPLE4A.DAT
(K)eyboard or (F)ile entry of the derivative pOints? K

Number of X values (0-100)? 5

Poi nt 1: 1.1
Point 2: 1.3
Point 3: 1.5
Point 4: 2.0
Point 5: 2.2

2-, 3-, or 5-point differentiation (default = 5)? 2

Direct output to one of the following:
(S)creen
(P)rinter
(F) il e

Input Data: X y
1.0000000 5.40302305868140E-001
1.1000000 5.48851306924949E-001
1.2000000 5.21795166446410E-001
1.3000000 4.52073020375553E-001
1.4000000 3.33135600084472E-001
1.5000000 1.59158703752332E-001
1.6000000 -7.47507770912994E-002
1.7000000 -3.72360588514066E-001
1.8000000 -7.36134786805602E-001
1.9000000 -1.16707533637725E+000
2.0000000 -1.66458734618857E+000
Output using two-point differentiation:

<* --------------------------- *>

<* WARNING *>
<* --------------------------- *>
X Derivative at X
1.100 8.54900105680900E-002
1.300 -6.97221460708569E-001
1.500 -1.73976896332140E+000
2.000 -4.97512009811320E+000
2.200 No derivative calculated

Numerical Differentiation 69
Output using three-point differentiation:

<* --------------------------- *>

<* WARNING *>
<* --------------------------- *>
X Derivative at X
1.100 -9.25356971086502E-002
1.300 -9.43297831809691E-001
1.500 -2.03943188587886E+000
2.000 -5.30797739931155E+000
2.200 No derivative calculated

Output using five-point differentiation:

<* --------------------------- *>

<* WARNING *>
<* --------------------------- *>
Actual Values
X Derivative at X X Derivative at X
1.100 -8.08749392678299E-002 1.100 -0.0804494
1.300 -9.32986606435738E-001 1.300 -0.9329163
1.500 -2.03221450709712E+000 1.500 -2.0321521
2.000 -5.30200229054730E+000 2.000 -5.3017771
2.200 No derivative calculated 2.200 -6.5025275

The data is taken from a function of which a derivative could be computed

exactly. Though the values in the right-hand columns (under "Actual Values") are
not displayed on screen, they are shown here to indicate the accuracy of the rou-
tine.
The warning signal indicates that some derivatives were not calculated.
The derivative is not approximated for x = 2.2 in any of the examples because
x = 2.2 is not among the data points.

70 Turbo Numerical Methods Toolbox

Second Differentiation Using Three-Point or Five-Point
Formulas (DER1V2.INC)

Description

This example contains two algorithms that approximate the second derivative of a
functionf(x) when several data points (x,f(x)) are specified. You decide whether to
use a three-point or five-point formula (Gerald and Wheatley 1984, 236-237);
three points are used in the three-point formula, and five in the five-point formula.
You must supply the data points (x,j(x)) and the x-values of the data points at which
the second derivative is to be approximated. The second derivative may only be
approximated at x-values that were input as data points.

User-Defined Types

TNvector = array[l •• TNArraySize] of Real;

Input Parameters

NumPoints : Integer; Number of data points

XData : TNvector; X-coordinates of the data points
YData : TNvector; Y-coordinates of the data points
Point: Byte; Three-point or five-point differentiation
NumDeriv : Integer; Number of points at which the derivative is to be approxi-
mated
XDeriv : TNvector; X-coordinates of points at which the derivative is to be approx-
imated
The preceding parameters must satisfy the following conditions:
1. XData points must be unique.
2. XData points must be entered in ascending order.
3. At least three points for three-point differentiation and five points for five-
point differentiation.
4. Point must equal 3 or 5.

Numerical Differentiation 71
5. XData points must be equally spaced.
6. XDeriv points must be a subset of the XData points.
7. NumPoints, NumDeriv :5 TNArraySize.

TNArraySize represents the number of elements in each vector. It is used in the

type definition of TNvector. TNArraySize is not a variable name and is never refer-
enced by the procedure; hence there is no test for condition 7. If condition 7 is
violated, the program will crash with an Index Out of Range error (assuming the
directive {$R + } is active).

Output Parameters

YDeri v : TNvector; Approximation to the second derivative at the XDeriv points

Error : Byte; 0: No errors
1: WARNING! At least one derivative was not approximated
(see "Comments")
2: X-values not unique
3: X-values not in increasing order
4: Not enough data
5: Point not equal to 3 or 5
6: X-value points not equally spaced

Syntax of the Procedure Call

Second-Derivative(NumPoints, XData, YData, Point, NumDeriv, XDeriv, YDeriv, Error);

Comments

If an x-value at which the second derivative is approximated is not among the data
points, the value - 9.9999999E35 is arbitrarily assigned to the derivative at that
point and Error = 1 is returned. When using five-point second differentiation with
only five data points, there is insufficient information for approximating the second
derivative at the second and fourth data points. Should an attempt be made to
approximate the second derivative at these points, the value 9.9999999E35 is arbi-
trarily assigned to the second derivative at that point and Error = 1 is returned.

72 Turbo Numerical Methods Toolbox

Since numerical differentiation is extremely prone to round-off error, TNNear-
lyZero is different in this routine. The values of TNNearlyZero are TNNearlyZero
= IE - 13 if using the 8087 math coprocessor and TNNearlyZero = IE - 6 if not
using the 8087.

Sample Program

The sample progam DERIV2.PAS provides I/O functions that demonstrate

second-order differentiation with three-point and five-point formulas.

Input Files
Data points may be entered from a text file. The x- and y-coordinates should be
separated by a space and followed by a carriage return. For example, data values of
sqr(x) could be entered in a text file as
11
24
39
416
525
Derivative points may also be entered from a text file. Every derivative point
must be followed by a carriage return. For example, to determine the second deriv-
atives of the preceding points, create the following file of derivative points:
1
2
3
4
5

Example
Problem. Approximate the second derivative of f(x) = sqr(x) * cos(x) at several
points between x = 1 and x = 2 radians. The output from two runs is given. Actual
values of the second derivatives to eight significant figures are also given.

Numerical Differentiation 73
Run DERIV2.PAS:
(K)eyboard or (F)ile entry of the data points? F

File name? SAMPLE4A.DAT

(K)eyboard or (F)ile entry of the derivative points? K

Number of X values (0-100)?5

Point 1: 1.1
Point 2: 1.3
Point 3: 1.5
Point 4: 2.0
Point 5: 2.2

3- or 5-point second differentiation (default = 5)? 3

Direct output to one of the following:
(S)creen
(P)rinter
(F) i 1e

Input Data: X y
1.0000000 5.40302305868140E-001
1.1000000 5.48851306924949E-001
1.2000000 5.21795166446410E-001
1.3000000 4.52073020375553E-001
1.4000000 3.33135600084472E-001
1.5000000 1.59158703752332E-001
1.6000000 -7.47507770912994E-002
1.7000000 -3.72360588514066E-001
1.8000000 -7.36134786805602E-001
1.9000000 -1. 16707533637725E+000
2.0000000 -1.66458734618857E+000
Output using three-point second differentiation:

<* --------------------------- *>

<* WARNING *>
<* --------------------------- *>
X 2nd Derivative at X
1.100 -3.56051415353479E+000
1.300 -4.92152742202240E+000
1.500 -5.99325845114913E+000
2.000 -6.65714602396721E+000
2.200 No 2nd derivative calculated

74 Turbo Numerical Methods Toolbox

Output using five-point second differentiation:

<* --------------------------- *>

<* WARNING *>
<* --------------------------- *>
Actual Values
X 2nd Derivative at X X 2nd Derivative at X
1.100 -3.61167369644119E+000 1.100 -3.5629714
1.300 -4.92756964541465E+000 1.300 -4.9275779
1.500 -6.00263647117236E+000 1.500 -6.0026542
2.000 -6.59765691992321E+000 2.000 -6.4420857
2.200 No 2nd derivative 2.200 -5.4434251
calculated
The data is taken from a function of which the derivative could be computed
exactly. Though the values in the right-hand columns (under "Actual Values") are
not displayed on screen, they are shown here to indicate the accuracy of the rou-
tine.
The warning signal indicates that some second derivatives were not calculated.
The second derivative is not approximated at x = 2.2 for either run because
x = 2.2 is not among the input x-value points.

Numerical Differentiation 75
Differentiation with a Cubic Spline Interpolant
(INTERDRV.INC)

Description

This example contains an algorithm for approximating the first and second deriva-
tives of a function given several data points (x,f(x)). The algorithm assumes that a
free cubic spline interpolant (Burden and Faires 1985, 117-122) is an adequate ap-
proximation to the functionf(x), so that the slope of the interpolant at any value x.
is an adequate approximation tof'(x). See Chapter 3 (CUBE_FRE.lNC) for mor~
information on free cubic splines. The user must supply the data points (x,f(x)) and
the x-values at which to approximate the derivatives. Derivatives may be approxi-
mated at any x-value contained in the closed interval determined by the data
points. This routine will likely give significant errors if interpolation (Gerald and
Wheatley 1984, 227-231) is attempted outside the range of x-values (extrapo-
lation).

User-Defined Types

TNvector = array[l .• TNArraySize] of Real;

Input Parameters

NumPoints : Integer; Number of data points

XData : TNvector; X-coordinates of data points
YData : TNvector; Y-coordinates of data points
NumDeriv : Integer; Number of points at which the derivative is to be approxi-
mated
XDeri v : TNvector; X-coordinates of points at which the derivative is to be approx-
imated

76 Turbo Numerical Methods Toolbox

The preceding parameters must satisfy the following conditions:
1. XData points must be unique.
2. XData points must be in ascending order.
3. NumPoints ~ 2.
4. NumPoints, NumDeriv S TNArraySize.

TNArraySize represents the number of elements in each vector. It is used in the

type definition of TNvector. TNArray Size is not a variable name and is never refer-
enced by the procedure; hence there is no test for condition 4. If condition 4 is
violated, the program will crash with an Index Out of Range error (assuming the
directive {$R + } is active).

Output Parameters

Ylnter : TNvector; Interpolated y-values at the XDeriv points

YDeri v : TNvector; Approximation to the first derivative at the x-values in XDeriv
YDeriv2 : TNvector; Approximation to the second derivative at the x-values in
XDeriv
Error: Byte; 0: No errors
1: X-values not unique
2: X-values not in ascending order
3: NumPoints < 2

Syntax of the Procedure Call

Interpolate-Derivative(NumPoints, XData, YData, NumDeriv,

XDeriv, Ylnter, YDeriv, YDeriv2, Error);

Sample Program

The sample program INTERDRV.PAS provides I/O functions that demonstrate

differentiation with a cubic spline interpolant.

Numerical Differentiation 77
Input Files
Data points may be entered from a text file. The x- and y-coordinates should be
separated by a space and followed by a carriage return. For example, data values of
sqr(x) could be entered in a text file as
11
24
39
416
525
Derivative points may also be entered from a text file. Every derivative point
must be followed by a carriage return. For example, to determine the derivatives of
the preceding points, create the following file of derivative points:
1
2
3
4
5

Example
Problem. Determine the first and second derivative ofJ(x) = sqr(x) * cos (x) at
several points between one and two radians. Actual values of the derivatives to
eight significant figures are given here.
Run INTERDRY.PAS:
(K)eyboard or (F)ile entry of data points? F
File name? SAMPLE4B.DAT
(K)eyboard or (F)ile entry of derivative points? K
Number of derivative points (0-100)?5
Point 1: 1.1
Poi nt 2: 1. 3
Point 3: 1.55
Point 4: 1.95
Point 5: 2.20

Direct output to one of the following:

(S)creen
(P)rinter
(F) il e

78 Turbo Numerical Methods Toolbox

Input Data: X y
1.000 0.5403023
1.100 0.54885l3
1.200 0.5217952
1.300 0.4520730
1.400 0.3331356
1.500 0.1591587
1.600 -0.0747508
1.700 -0.3723606
1.800 -0.7361348
1.900 -1.1670753
2.000 -1.6645873

Using free cubic spline interpolation:

X Value at X 1st Deriv at X 2nd Deriv at X
1.100 5.48851300000000E-001 -5.86015666816468E-002 -4.32274700
1.300 4.52073000000000E-001 -9.31377366861404E-001 -4.98862501
1.550 4.99429267146238E-002 -2.33770918101853E+000 -6. 19118137
1.950 -1.41057141673716E+000 -5.01018588841893E+000 -4.20790661
2.200 -2.57545316779455E+000 -3.43222090956677E+000 16.83162644

The data is taken from a function of which the derivative could be computed
exactly. The actual values are shown here:

X Value at X 1st Deriv at X 2nd Deriv at X

1.1 0.5488513 -0.0804494 -3.5629715
1.3 0.4520730 -0.9329164 -4.9275779
1.55 0.0499596 -2.3375165 -6.2070293
1.95 -1.4076126 -4.9760746 -6.5786348
2.20 -2.8483454 -6.5025275 -5.4434252
Note the poor results obtained at values outside the range of input data
(x = 2.2). Also note the large error in the second derivatives near the endpoints of
the interval determined by the data.

Numerical Differentiation 79
Differentiation of a User-Defined Function
(DERIVFN.INC)

Description

Given a user-defined function fix), this example will approximate the first deriva-
tive of the function at a set of x values. The formula
f' (x) = [f(x + ~X) - f(x - ~X)]/2*~

gives a first approximation to the derivative. Richardson extrapolation is then used

to refine the approximation (Burden and Faires 1985,137-152).

User- Defined Types

TNvector = array[l •• TNArraySize] of Real;

User-Defined Function

function TNTargetF(X Real) Real;

Input Parameters

NumDeriv : Integer; Number of points at which the derivative is to be approximated

XDeriv : TNvector; X-coordinates of points at which the derivative is to be approxi-
mated
To 1erance : Real; Indicates accuracy of solution
The preceding parameters must satisfy the following conditions:
l. NumDeriv::::; TNArraySize
2. Tolerance > TNNearlyZero

TNArraySize represents the number of elements in each vector. It is used in the

type definition of TNvector. TNArraySize is rwt a variable name and is never refer-
enced by the procedure; hence there is no test for condition 1. If condition 1 is

80 Turbo Numerical Methods Toolbox

violated, the program will crash with an Index Out of Range error (assuming the
directive {$R + } is active).

Output Parameters

YDer;v : TNvector; Approximation to the first derivative at the x-values in XDeriv

Error: Byte; 0: No errors
1: Tolerance < TNNearlyZero

Syntax of the Procedure Call

F;rstDer;vat;ve(NumDer;v, XDer;v, YDer;v, Tolerance, Error};

The procedure FirstDerivative approximates the first derivative of function
TNTargetF.

Comments

Since numerical differentiation is extremely prone to round-off errors, TNNear-

lyZero is different in this routine. Its values are TNNearlyZero = IE -10 if using
the 8087 math coprocessor and TNNearlyZero = IE - 5 if not using the 8087.

Sample Program

The sample program DERIVFN.PAS provides I/O functions that find the first
derivative of a function at a set of points.

Input Files
Derivative points may be entered from a text file. Every derivative point must be
followed by a carriage return. For example, to determine the derivatives at x-values
1 through 5, create the following file of derivative points:
1
2
3
4
5

Numerical Differentiation 81
Example
Problem. Determine the first derivative off(x) = sqr(x) * cos(x) at several points
between 1 and 2.2. Actual values of the derivatives to eight significant figures are
given here.
First, write the function into the DERIVFN.PAS program:
(* ----- here is the function to differentiate -------------------- *)
funct;on TNTargetF(X : Real) : Real;
beg;n
TNTargetF := Sqr(X)*Cos(X);
end; { function TNTargetF }

(* ---------------------------------------------------------------- *)
Run DERIVFN.PAS:
(K)eyboard or (F)ile entry of derivative points? K
Number of points (0-100)? 5
Poi nt 1: 1.1
Point 2: 1.3
Point 3: 1.55
Point 4: 1.95
Point 5: 2.2

Tolerance (> 0, default = 1.000E-02)? 1E-4

Direct output to one of the following:
(S)creen
(P)rinter
(F) il e

Tolerance = 1.00000000000000E-004
Actual Values
X Derivative at X X Value at X 1st Deriv at X
1.100 -8.04494385380506E-002 1.1 0.5488513 -0.0804494
1.300 -9.32916380187812E-001 1.3 0.4520730 -0.9329164
1.550 -2.33751652942968E+000 1.55 0.0499596 -2.3375165
1.950 -4.97607456093019E+000 1.95 -1.4076126 -4.9760746
2.200 -6.50252751001358E+000 2.20 -2.8483454 -6.5025275

The data is taken from a function of which the derivative could be calculated
exactly. Though the values in the three right-hand columns (under "Actual Values")
are not displayed on screen, they are shown here to indicate the accuracy of the
routine.

82 Turbo Numerical Methods Toolbox

Second Differentiation of a User-Defined Function
(DERIV2FN.INC)

Description

Given a user-defined function fix), this example will approximate the second deriv-
ative of the function at a set of x values. The three-point formula
f" (x) = [f(x + /lX) - 2f(x) + f(x - AX)]/AX2
gives a first approximation to the second derivative. Richardson extrapolation is
then used to refine the approximation (Burden and Faires 1985, 142-152).

User-Defined Types

TNvector = array[l .. TNArraySize] of Real;

User- Defined Function

function TNTargetF(X Real) Real;

Input Parameters

NumDeri v : Integer; Number of points at which the derivative is to be approximated

XDeriv : TNvector; X-coordinates of points at which the derivative is to be approxi-
mated
To 1erance : Real; Indicates accuracy in solution

Numerical Differentiation 83
The preceding parameters must satisfy the following conditions:
1. NumDeriv $; TNArraySize
2. Tolerance ~ TNNearlyZero

TNArraySize represents the number of elements in each vector. It is used in the

type definition of TNvector. TNArraySize is not a variable name and is never refer-
enced by the procedure; hence there is no test for condition 1. If condition 1 is
violated, the program will crash with an Index Out of Range error (assuming the
directive {$R + } is active}.

Output Parameters

YDeri v : TNvector; Approximation to the second derivative at the x-values in XDeriv

Error: Byte; 0: No errors
1: Tolerance < TNNearlyZero

Syntax of the Procedure Call

SecondDerivative(NumDeriv, XDeriv, YDeriv, Tolerance, Error);

SecondDerivative approximates the derivative of function TNTargetF.

Comments

Since numerical differentiation is extremely prone to round-off errors, TNNear-

lyZero is different in this routine. Its values are TNNearlyZero = IE - 4 if using
the 8087 math coprocessor and TNNearlyZero = IE - 2 if not using the 8087.

Sample Program

The sample program DERIV2FN.PAS provides I/O functions that find the second
derivative of a function at a set of points.

84 Turbo Numerical Methods Toolbox

Input Files
Derivative points may be entered from a text file. Every derivative I,Joint must be
followed by a carriage return. For example, to determine the second derivatives at
x-values 1 through 5, create the following file of derivative points:
1
2
3
4
5

Example
Problem. Determine the second derivative of fix) = sqr(x) * cos (x) at several
points between 1 and 2.2. Actual values of the derivatives to eight significant fig-
ures are given here.
First, write the function into the DERIV2FN .PAS program:
(* ----- here is the function to differentiate -------------------- *)

function TNTargetF(X : Real) : Real;

begin
TNTargetF := Sqr(X)*Cos(X);
end; { function TNTargetF }

(* ---------------------------------------------------------------- *)
Run DERIV2FN.PAS:
(K)eyboard or (F)ile entry of derivative pOints? K
Number of points (0-100)? 5
Point 1: 1.1
Point 2: 1.3
Point 3: 1.55
Point 4: 1.95
Point 5: 2.2
Tolerance (> 0, default = 1.000E-02)? 1E-4
Direct output to one of the following:
(S)creen
(P)rinter
(F)ile

Tolerance = 1.00000000000000E-004

Numerical Differentiation 85
Actual Values
X 2nd Derivative at X X Value at X 2nd Deriv at X
1.100 -3.56297143915941E+000 1.1 0.5488513 -3.5629715
1.300 -4.92757787674466E+000 1.3 0.4520730 -4.9275779
1.550 -6.20702925534123E+000 1.55 0.0499596 -6.2070293
1.950 -6.57863484485542E+000 1.95 -1.4076126 -6.5786348
2.200 -5.44342524529641E+000 2.20 -2.8483454 -5.4434252

The data is taken from a function of which the derivative could be calculated
exactly. Though the values in the three right-hand columns (under "Actual Values")
are not displayed on screen, they are shown here to indicate the accuracy of the
routine.

86 Turbo Numerical Methods Toolbox

c H A p T E R 5
Numerical Integration

Integration is another concept used in calculus. It is just the opposite of differentia-

tion, for which routines are provided in Chapter 4. Differentiation tells you the
changes in a function, where integration tells you how to add those changes to get
the original function.
Integration is most easily understood in terms of areas under curves. Given a
function f(x) and real numbers a and b with a < b, the area under the curve
y = f(x) and above the x-axis between x = a and x = b is given by the integral of
f(x) from a to b.
As with derivatives, the laws of calculus are required to compute integrals
exactly. The routines in this chapter provide very accurate approximations.
Several methods are described here that approximate the value of a deRnite
integral of a real function of one real variable. Both limits of integration must be
Rnite.
The trapezoid method (TRAPZOID.lNC) and Simpson's method (SIMPSON.
INC) return an approximation of the integral when a number of equal length
subintervals are speciRed. For a given number of subintervals, Simpson's method
is preferre~ over the trapezoid method whenever the function being integrated is
sufficiently smooth.
It is sometimes possible to approximate the deRnite integral to within a user-
speciRed accuracy with fewer function evaluations using adaptive schemes. Adap-
tive schemes determine the length of each subinterval by the local behavior of the

87
integrand. Simpson's method (ADAPSIMP.INC) and the Gaussian quadrature
method (ADAPGAUS.INC) are used with adaptive schemes. The Gaussian quad-
rature method permits, in some instances, the integrand to possess a singularity at
an endpoint of integration, since the function is evaluated at points that are not the
endpoints of the interval of integration.
The Romberg method (ROMBERG.INC) uses the trapezoid method and Rich-
ardson extrapolation to approximate the integral. It returns an approximation
within a user-specified accuracy. Except for extremely oscillatory functions or
functions that possess an endpoint singularity, this method is fastest and most
accurate. If the function oscillates substantially or possesses an endpoint singular-
ity, the adaptive Gaussian quadrature routine is preferred.

88 Turbo Numerical Methods Toolbox

Integration Using Simpson's Composite Algorithm
(SIMPSON. INC)

Description

This example uses Simpson's composite algorithm (Burden and Faires 1985, 156-
167) to approximate the definite integral of a functionJ\x) over an interval [a, b].
The interval is divided into N subintervals of equal length. The curve in each
subinterval is approximated by a second-degree Lagrange polynomial. The integral
of the resulting polynomial is then calculated. The sum of the integrals of the N
Lagrange polynomials approximates the integral of the function! over the interval
[a, bJ. You must supply the function, the limits of integration, and the number of
subintervals.

User- Defined Function

function TNTargetF(x : Real) : Real;

The procedure Simpson approximates the integral of this function.

Input Parameters

LowerL i mi t : Rea 1 ; Lower limit of integration

UpperLi mit : Real; Upper limit of integration
NumIntervals : Integer; Number of subintervals over which to apply Simpson's rule

The preceding parameters must satisfy the following condition:

NumI ntervals > 0

Numerical Integration 89
Output Parameters

Integral: Real; Approximation to the integral of the function

Error: Byte; 0: No errors
1: Numlntervals :5 0

Syntax of the Procedure Call

Simpson(LowerLimit, UpperLimit, NumIntervals, Integral, Error);

Simpson approximates the integral of TNTargetF.

Sample Program

The sample program SIMPSON.PAS provides I/O functions that demonstrate

Simpson's composite algorithm.

Example
Problem. Approximate the integral exp(3x) + sqr(x)/3 from 0 to 5 using Simpson's
composite algorithm.
1. Code function TNTargetF:
funct;on TNTargetF(x : Real) : Real;
(**************************************************************************)
(**** THIS IS THE FUNCTION TO INTEGRATE ****)
(**************************************************************************)
beg;n
TNTargetF .- Exp(3*X) + Sqr(X)/3;
end; { function TNTargetF }

90 Turbo Numerical Methods Toolbox

2. Run SIMPSON.PAS:
Lower limit of integration? 0
Upper limit of integration? 5
Number of intervals (> O)? 100
Direct output to one of the following:
(S)creen
(P)rinter
(F) il e

Lower limit: O.OOOOOOOOOOOOOOE+OOO

Upper limit: 5.00000000000000E+000
Number of intervals: 100
Integral: 1.08968620446199E+006
To eight significant figures, the correct answer is 1,089,686.2.

Numerical Integration 91
Integration Using the Trapezoid Composite Rule
(TRAPZOID.INC)

Description

This example uses the trapezoid composite rule (Burden and Faires 1985,154-167)
to approximate the definite integral of a function fix) over an interval [a, b]. The
interval is divided into N subintervals of equal length. In each subinterval the
function is approximated by a straight line. The sum of the integrals of the result-
ing trapezoids approximates the integral of the function f over the interval [a, b].
You must supply the function, the limits of integration, and the number of subinter-
vals.

User-Defined Function

funct;on TNTargetF(x : Real) : Real;

The procedure Trapezoid approximates the integral of this function.

Input Parameters

Lower Li mi t : Real; Lower limit of integration

UpperLimit : Real; Upper limit of integration
NumInterva 1s : Integer; Number of subintervals over which to apply the trapezoid
rule
The preceding parameters must satisfy the following condition:
Numlntervals > 0

92 Turbo Numerical Methods Toolbox

Output Parameters

Integral: Real; Approximation to the integral of the function

Error: Byte; 0: No errors
1: NumI ntervals S 0

Syntax of the Procedure Call

Trapezoid(LowerLimit, UpperLimit, NumIntervals, Integral, Error);

Trapezoid approximates the integral of TNTargetF.

Sample Program

The sample program T~APZOID.PAS provides I/O functions that demonstrate the
trapezoid composite rule.

Example
Problem. Approximate the integral exp(3x) + sqr(x)/3 from 0 to 5 using the trape-
zoid composite rule.
1. Code function TNTargetF:
function TNTargetF(x : Real) : Real;
(*****************************************************************************)
(**** THIS IS THE FUNCTION TO INTEGRATE ****)
(*****************************************************************************)
begin
TNTargetF .- Exp(3*X) + Sqr(X)/3;
end; { function TNTargetF }

Numerical Integration 93
2. Run TRAPZOID.PAS:
Lower limit of integration? 0
Upper limit of integration? 5
Number of intervals (> O)? 100
Direct output to one of the following:
(S)creen
(P)rinter
(F) il e

Lower limit: O.OOOOOOOOOOOOOOE+OOO

Upper limit: 5.00000000000000E+000
Number of intervals: 100
Integral: 1.09172838320798E+006
To eight significant figures, the correct answer is 1,091,728.3.

94 Turbo Numerical Methods Toolbox

Integration Using Adaptive Quadrature and Simpson's Rule
(ADAPSIMP.INC)

Description

This example contains an algorithm for approximating the definite integral of a

functionf(x) over an interval [a,b] within a specified tolerance. By increasing the
number of subintervals in regions of large functional variation (adaptive quadra-
ture), the desired degree of accuracy can be reached (Burden and Faires 1985,
153-167). The integral within each subinterval is calculated with Simpson's rule.
The adaptive quadrature approximates the integral over a subinterval twice: once
over the whole subinterval, and again as the sum of the integral over each half of
the subinterval. The algorithm halts when the fractional difference between these
two approximations is less than the tolerance. You must supply the function, the
limits of integration, and the tolerance with which to approximate the integral.

User- Defined Function

function TNTargetF(x : Real) : Real;

The procedure Adaptive-.Simpson approximates the integral of this function.

Input Parameters

Lower Limit: Real; Lower limit of integration

UpperLi mi t : Real; Upper limit of integration
Tolerance: Real; Indicates accuracy in solution
MaxInterval s : Integer; Maximum number of subintervals

The preceding parameters must satisfy the following conditions:

1. Tolerance > 0
2. Maxlntervals > 0

Numerical Integration 95
Output Parameters

Integral: Real; Approximation to the integral of the function

NumInterval s : Integer; Number of subintervals used
Error : Byte; 0: No errors
1: Tolerance S 0
2: Maxlntervals s,O
3: Numlntervals ~ Maxlntervals

Syntax of the Procedure Call

Adaptive-Simpson(LowerLimit, UpperLimit, Tolerance, MaxIntervals,

Integral, NumIntervals, Error);
Adaptive_Simpson approximates the integral of TNTargetF.

Comments

Adaptive quadrature is a recursive routine. In order to avoid recursive procedure

calls (which slow down the execution), a stack is created on the heap to simulate
recursion. Should you attempt to evaluate the integral to a very high degree of
accuracy with a large number of subintervals, you may get run-time error $FF,
Heap/Stack collision. If this happens, remove any RAM-resident software (for
example, SideKick®, SuperKey®, or a print buffer). If the problem remains, the
adaptive Simpson routine cannot be used to approximate the integral to the
desired accuracy.
Adaptive quadrature uses the New/Dispose procedures to manipulate the heap
and should not be used in any program that uses Mark/Release to manipulate the
heap.

Sample Program

The sample program ADAPSIMP.PAS provides I/O functions that demonstrate the
adaptive quadrature method with Simpson's rule.

96 Turbo Numerical Methods Toolbox

Example
Problem. Approximate the integral exp(3x)
quadrature and Simpson's rule.
+ sqr(x)/3 from °
to 5 using adaptive

1. Code function TNTargetF:

function TNTargetF(x : Real) : Real;
(****************************************************************************)
(** THIS IS THE FUNCTION TO INTEGRATE ***)
(****************************************************************************)
begin
TNTargetF := Exp(3*X) + Sqr(X)/3;
end; { function TNTargetF }
2. Run ADAPSIMP.PAS:
Lower limit of integration? 0
Upper limit of integration? 5
Tolerance (> 0, default = 1.000E-08): 1E-8
Maximum number of subintervals (> 0, default 1000): 1000
Direct output to one of the following:
(S)creen
(P)rinter
(F) il e

Lower 1imit: O.OOOOOOOOOOOOOOE+OOO

Upper 1imi t: 5.00000000000000E+000
Tolerance: 1.00000000000000E-008
Maximum number of subintervals: 1000
Number of subintervals used: 511
Integral: 1.08968601332498E+006
To eight significant figures, the correct answer is 1,089,686.0.

Numerical Integration 97
Integration Using Adaptive Quadrature and Gaussian
Quadrature (ADAPGAUS.INC)

Description

This example contains an algorithm for approximating the integral of a functionf(x)

over an interval [a,b] within a specified tolerance. By increasing the number of
subintervals in regions of large functional variation (adaptive quadrature), the
desired degree of accuracy can be reached. The integral within each subinterval is
approximated by applying Gaussian quadrature (Burden and Faires 1985,184-188)
with a 16th degree Legendre polynomial. Adaptive quadrature (Burden and Faires
1985, 172-176) approximates the integral over a subinterval twice: once over the
whole subinterval, and again as the sum of the integral over each half of the subin-
terval. The algorithm halts when the fractional difference between these two
approximations is less than the tolerance. You must supply the function, the limits
of integration, and the tolerance with which to approximate the integral.

User- Defined Function

funct;on TNTargetF(x : Real) : Real;

The procedure Adaptive_Gauss_Quadrature approximates the integral of this

function.

Input Parameters

LowerLimit : Real; Lower ~imit of integration

UpperLimit : Real; Upper limit of integration
To 1erance.: Real; Indicates accuracy in solution
MaxInterval s : Integer; Maximum number of subintervals

The preceding parameters must satisfy the following conditions:

1. Tolerance > 0
2. Maxlntervals > 0

98 Turbo Numerical Methods Toolbox

Output Parameters

Integral: Real; Approximation to the integral of the' function

NumIntervals : Integer; Number of subintervals used
Error: Byte; 0: No errors
1: Tolerance :5 0
2: Maxlntervals :5 0
3: Numlntervals ~ Maxlntervals

Syntax of the Procedure Call

Adaptive_Gauss_Quadrature(LowerLimit, UpperLimit, Tolerance, MaxIntervals,

Integral, NumIntervals, Error);
Adaptive_Gauss_Quadrature approximates the integral of TNTargetF.

Comments

Adaptive quadrature is a recursive routine. In order to avoid recursive procedure

calls (which slow down execution), a stack is created on the heap to simulate recur-
sion. Should you attempt to evaluate the integral to a very high degree of accuracy
with a large number of subintervals, you may get run-time error $FF, Heap/Stack
collision. If this happens, remove any RAM-resident software (for example, Side-
Kick, SuperKey, or a print buffer). If the problem remains, the adaptive Gaussian
quadrature routine cannot be used to approximate the integral to the desired accu-
racy.
Adaptive quadrature uses the New/Dispose procedures to manipulate the heap
and should not be used in any program that uses Mark/Release to manipulate the
heap.
Gaussian quadrature uses orthogonal polynomials (in this case, Legendre poly-
nomials) to approximate an integral. Generally, a higher degree polynomial will
yield a more accurate result, but will take more time to compute. The 16th degree
Legendre polynomial used in ADAPGAUS.lNC is very efficient. The values of its
zeros and weight factors follow (Abramowitz and Stegun 1972).

Numerical Integration 99
The following condition is satisfied by the numbers that follow it:
Integral from -1 to 1 ofJ(x) dx
equals
Sum from i = 1 to NumLegendreTerms of
Legendre[i].Weight * J(Legendre[i].Root)
for an arbitrary functionJ(x).
Legendre[I] ........................ ............... Root: 0.0950125098376370440185
Weight: 0.189450610455068496285
Legendre[2] ...................... ................. Root: 0.281603550778258913230
Weight: 0.182603415044923588867
Legendre[3] ... .................................... Root: 0.458016777657227386342
Weight: 0.169156519395002538189
Legendre[4] ................. ...................... Root: 0.617876244402643748447
Weight: 0.149595988816576732081
Legendre[5] ....................................... Root: 0.755404408355003Q33895
Weight: 0.124628971255533872052
Legendre[6] ....................................... Root: 0.865631202387831743880
Weight: 0.095158511682492784810
Legendre[7] ....................................... Root: 0.944575023073232576078
Weight: 0.062253523938647892863
Legendre[8] ... .................................... Root: 0.989400934991649932596
Weight: 0.027152459411754094852
Legendre[9] ....................................... Root: - 0.0950125098376370440185
Weight: 0.189450610455068496285
Legendre[10] ..................................... Root: - 0.281603550778258913230
Weight: 0.182603415044923588867
Legendre[ll] ..................................... Root: - 0.458016777657227386342
Weight: 0.169156519395002538189
Legendre[12] ..................................... Root: - 0.617876244402643748447
Weight: 0.149595988816576732081
Legendre[13] ..................................... Root: - 0.755404408355003033895
Weight: 0.124628971255533872052
Legendre[14] ..... ; ............................... Root: - 0.865631202387831743880
Weight: 0.095158511682492784810
Legendre[15] ..................................... Root: - 0.944575023073232576078
Weight: 0.062253523938647892863
Legendre[16] ..................................... Root: - 0.989400934991649932596
Weight: 0.027152459411754094852

100 Turbo Numerical Methods Toolbox

Sample Program

The sample program ADAPGAUS.PAS provides I/O functions that demonstrate

the adaptive quadrature method with Gaussian quadrature.

Example
Problem. Approximate the integral exp(3x) + sqr(x)/3 from
quadrature with Gaussian quadrature algorithm.
°to 5 using adaptive
1. Code function TNTargetF:
funct;on TNTargetF(x : Real) : Real;
(*****************************************************************************)
(** THIS IS THE FUNCTION TO INTEGRATE ***)
(*****************************************************************************)
beg;n
TNTargetF := Exp(3*X) + Sqr(X)/3;
end; { function TNTargetF }
2. Run ADAPGAUS.PAS:
Lower limit of integration? 0
Upper limit of integration? 5
Tolerance (> 0, default = 1.000E-08): 1E-8
Maximum number of subintervals (> 0, default 1000): 1000
Direct output to one of the following:
(S)creen
(P)rinter
(F) il e

Lower limit: O.OOOOOOOOOOOOOOE+OOO

Upper 1imit: 5.00000000000000E+000
Tolerance: 1.00000000000000E-008
Maximum number of subintervals: 1000
Number of subintervals used: 1

Integral: 1.08968601304609E+006
To eight significant figures, the correct answer is 1,089,686.0.

Numerical Integration 101

Integration Using the Romberg Algorithm
(ROMBERG. INC)

Description

This example contains an algorithm (Burden and Faires 1985,177-182) for approxi-
mating the integral of a functionf(x) over an interval [a, b] within a specified toler-
ance. The trapezoid rule is used to generate a preliminary approximation, and
Richardson extrapolation (Burden and Faires 1985,148-152) is subsequently used
to improve the approximation. Extrapolation continues until the fractional differ-
ence between successive approximations of the integral is less than the tolerance.
You must supply the function, the limits of integration, and the tolerance with
which to approximate the integral.

User- Defined Function

funct;on TNTargetF(x : Real) : Real;

The procedure Romberg approximates the integral of this function.

Input Parameters

LowerLi mi t : Real; Lower limit of integration

UpperLimit : Real; Upper limit of integration
To 1erance : Real; Indicates accuracy in solution
Max Iter : Integer; Maximum number of iterations allowed

The preceding parameters must satisfy the following conditions:

1. Tolerance > 0
2. MaxIter > 0

102 Turbo Numerical Methods Toolbox

Output Paralneters

Integral: Real; Approximation to the integral of the function

Iter: Integer; Number of iterations
Error: Byte; 0: No errors
1: Tolerance :5 0
2: MaxIter :5 0
3: Iter;::: MaxIter

Syntax of the Procedure Call

Romberg(LowerLimit, UpperLimit, Tolerance, MaxIter, Integral, Iter, Error);

Romberg approximates the integral of TNTargetF.

Sample Program

The sample program ROMBERG.PAS provides I/O functions that demonstrate the
Romberg algorithm.

Example
Problem. Approximate the integral exp(3x} + sqr(x}/3 from 0 to 5 using the Rom-
berg algorithm.
1. Code function TNTargetF:
function TNT~rgetF(x : Real) : Real;
(*************************************************************************)
(**** THIS IS THE FUNCTION TO INTEGRATE ****)
(*************************************************************************)
begin
TNTargetF := Exp(3*X) + Sqr(X)/3;
end; { function TNTargetF }

Numerical Integration 103

2. Run ROMBERG.PAS:
Lower limit of integration? 0
Upper limit of integration? 5
Tolerance (> 0, default =' 1.000E-08): 1E-8
Maximum number of iterations: (> 0, default = 100) 100
Direct output to one of the following:
{S)creen
{P)rinter
(F) il e

Lower 1imit: O.OOOOOOOOOOOOOOE+OOO

Upper 1i mit: 5.00000000000000E+000
Tolerance: 1.00000000000000E-008
Maximum number of iterations: 100
Number of iterations: 7
Integral: 1.08968601696675E+006
To eight significant figures, the correct answer is 1,089,686.0.

104 Turbo Numerical Methods Toolbox

c H A p T. E R 6
Matrix Routines

This chapter provides routines for dealing with systems of linear equations. An
example of a system of linear equations is as follows:
2X+Y+Z=7
X-Y+Z=2
X+Y-Z=O
Matrix algebra is a collection of notations that constitutes a technique for han-
dling such systems. With matrix algebra, the preceding system would be written
Ax = b
where

[~
1
A= -1
1

In Pascal, x and b are represented as one-dimensional arrays, and A is repre-

sented as a two-dimensional array. In matrix notation, the solution is given by
x = A-I b
where A - I is the inverse to A.
The determinant is an indicator of whether the matrix can be inverted. For
example, the equations
3X - 3Y =4
-2X + 2Y = 5

105
cannot be solved. Even for different values of the right-hand side, the equations
can only be solved in certain exceptional cases. (If you change 4 and 5 to 3 and - 2,
then there are infinitely many solutions; but there are none if you change 4 and 5 to
3 and - 3.0001.)
Following is a description of several routines that operate on matrices and sys-
tems of linear equations.
The determinant of a square matrix is found via DET.INC.
The inverse of a nonsingular matrix is found via INVERSE.INC.
The direct techniques implemented to solve a system of N linear equations in N
unknowns are Gaussian elimination (GAUSELIM.INC), Gaussian elimination
with partial pivoting (PARTPIVT.INC), and direct factorization (DIRFACT.INC) ..
The Gauss-Seidel method (GAUSSIDL.INC), an iterative technique that con-
verges to the solution, is seldom used for solving small systems, since the time
required for sufficient accuracy exceeds that required for the preceding direct
techniques.
In general, Gaussian elimination with partial pivoting is the fastest, most accu-
rate algorithm (see Chapter 9, LEAST.INC, for an application of PARTPIVT.INC).
The following special cases may warrant the use of one of the other routines:
• If you are considering systems where round-off is minimal (that is, small systems
whose coefficients are all of nearly the same magnitude), Gaussian elimination
without pivoting may be used. It is somewhat faster than its pivoting counterpart
(PARTPIVT.INC).
• When considering sparse coefficient matrices, the Gaussian elimination routine
with partial pivoting is the most efficient and accurate routine. If the matrix is
small and the nonzero coefficients do not differ wildly from each other, regular
Gaussian elimination (GAUSELIM.INC) can usually be used safely.
• For large, dense matrices, the iterative technique (GAUSSIDL.INC) is the most
efficient; it creates less round-off error than the direct methods. However, the
Gauss-Seidel algorithm has' its own weaknesses (see the section, "Solving a Sys-
tem of Linear Equations with the Iterative Gauss-Seidel Method," for more
details).
• When it is necessary to solve several systems with the same coefficient matrix
but a different vector of· constant terms, the direct factorization method
(DIRFACT.INC) is the most efficient. This is because it does not require reduc-
tion of the coefficient matrix for each vector of constants. (See Chapter 7 for an
application of DIRFACT.INC.)

106 Turbo Numerical Methods Toolbox

Determinant of a Matrix (DET.INe)

Description

The determinant of an N X N matrix can be computed by the following algorithm

(Gerald and Wheatley 1984, 110-111):
1. Use elementary row operations to make the matrix upper triangular (that
is, all the elements below the main diagonal are zero).
2. Find the product of the main diagonal elements - this will be the determi-
nant.

User- Defined Types

TNvector = array[l .. TNArraySize] of Real;

TNmatrix = array[l .• TNArraySize] of TNvector;

Input Parameters

Dimen : Integer; Dimension of the data matrix

Data: TNmatrix; The square matrix
The preceding parameters must satisfy the following conditions:
1. Dimen > 0
2. Dimen S; TNArraySize

TNArraySize sets an upper bound on the number of elements in each vector. It

is used in the type definition of TNvector and TNmatrix. TNArraySize is not a
variable name and is never referenced by the procedure; hence there is no test for
condition 2. If condition 2 is violated, the program will crash with an Index Out of
Range error.

Matrix Routines 107

Output Parameters

Det : Rea 1; Determinant of the data matrix

Error: Byte; 0: No errors
1: Dimen < 1

Syntax of the Procedure Call

Determinant(Dimen, Data, Det, Error);

Sample Program

The sample program DET.PAS provides I/O functions that demonstrate how to
find the determinant of a matrix.

Input File
Data may be input from a text file. All entries in the text file should be separated
by a space or carriage return, and it does not matter if the text file ends with a
carriage return. The format of the text file should be like this:
1. The dimension of the matrix
2. The elements of the matrix in row order; that is,
[1, 1], [1, 2] ... [1, N], [2, 1] ... [2, N] ... [N, N],
where N is the dimension of the matrix

For example, a text file containing the matrix

could look like this:

2
2 3
-4 0

108 Turbo Numerical Methods Toolbox

Example
Problem. Find the determinant of the following matrix:

-i ! ~ =~:~ 1
[o 2 2
0
1
3
-3.0
-4.0
Run DET.PAS:
(K)eyboard or (F)ile input of data? F

File name? SAMPLE6A.DAT

Direct output to one of the following:

(S)creen
(P)rinter
(F) ile

The matrix:
1.00000000 2.00000000 0.00000000 -1.00000000
-1.00000000 4.00000000 3.00000000 -0.50000000
2.00000000 2.00000000 1.00000000 -3.00000000
0.00000000 0.00000000 3.00000000 -4.00000000

Determinant = -2.10000000000000E+001

Matrix Routines 109

Inverse of a Matrix (INVERSE.INC)

Description

The inverse of an N X N matrix A is an N X N matrix A- \ such that A -lA equals

the identity matrix (Burden and Faires 1985, 306-316). Gauss-Jordan elimination
(Gerald and Wheatley 1984, 96-98) is used to transform the original matrix into
the identity matrix. The same elementary row operations that reduce A to the
identity matrix transform the identity matrix into the inverse of the original matrix
A. If one or more of the main diagonal elements of the transformed original matrix
(that is, after Gauss-Jordan elimination) is zero, then the original matrix A is singu-
lar and its inverse does not exist.

User- Defined Types

TNvector = array[l .• TNArraySize] of Real;

TNmatrix = array[l .. TNArraySize] of TNvector;

Input Parameters

Dimen : Integer; Dimension of the data matrix

Data: TNmatrix; The elements of the square matrix
The preceding parameters must satisfy the following conditions:
1. Dimen > 0
2. Dimen ~ TNArraySize

TNArraySize fixes an upper bound on the number of elements in each vector. It

is used in the type definition of TNvector and TNmatrix. TNArraySize is not a
variable name and is never referenced by the procedure; hence there is no test for
condition 2. If condition 2 is violated, the program will crash with an Index Out of
Range error.

110 Turbo Numerical Methods Toolbox

Output Paranreters

INV : TNmatrix; The inverse of the data matrix

Error: Byte; 0: No errors
1: Dimen < 1
2: No inverse exists

Syntax of the Procedure Call

Inverse(Dimen, Data, INV, Error);

Sample Program

The sample program INVERSE.PAS provides I/O functions that demonstrate how
to find the inverse of a matrix.

Input Files
Data may be input from a text file. All entries in the text file should be separated
by a space or carriage return, and it does not matter if the text file ends with a
carriage return. The format of the text file should be as follows:
1. The dimension of the matrix
2. The elements of the matrix in row order; that is,
[1, 1], [1, 2] ... [1, N], [2, 1] ... [2, N] ... [N, N],
where N is the dimension of the matrix

For example, a text file containing the matrix

could look like this:

2
2 3
-4 0

Matrix Routines III

Example
Problem. Invert the following matrix:
1 2
-1 4 3
0
-1.0]
-0.5
[ 221 -3.0
003 -4.0
Run INVERSKPAS:
(K)eyboard or (F)ile input of data? F
File name? SAMPLE6A.DAT
Direct output to one of the following:
(S)creen
(P)rinter
(F)ile

The matrix:
1.000000000 2.000000000 0.000000000 -1.000000000
-1.000000000 4.000000000 3.000000000 -0.500000000
2.000000000 2.000000000 1.000000000 -3.000000000
0.000000000 0.000000000 3.000000000 -4.000000000

Inverse:
-1.952380952 0.190476190 1.571428571 -0.714285714
0.761904762 0.047619048 -0.357142857 0.071428571
-1.904761905 0.380952381 1.142857143 -0.428571429
-1.428571429 0.285714286 0.857142857 -0.571428571

To continue this example, reinvert the matrix just obtained:

(K)eyboard or (F)ile input of data? F
File name? SAMPLE6B.DAT
Direct output to one of the following:
(S)creen
(P)rinter
(F) il e

The matrix:
-1.952380952 0.190476190 1.571428571 -0.714285714
0.761904762 0.047619048 -0.357142857 0.071428571
-1.904761905 0.380952381 1.142857143 -0.428571429
-1.428571429 0.285714286 0.857142857 -0.571428571

Inverse:
1.000000000 2.000000000 0.000000000 -1.000000000
-1.000000000 4.000000000 3.000000000 -0.500000000
2.000000000 2.000000000 1.000000000 -3.000000000
-0.000000000 -0.000000000 3.000000000 -4.000000000

112 Turbo Numerical Methods Toolbox

The coefficients of the original matrix are returned to fourteen significant figures
(only ten are displayed). The coefficients will be less precise if this example is run
on a machine without an 8087 math coprocessor.

Matrix Routines 113

Solving a System of Linear Equations with Gaussian
Elimination (GAUSELIM.INC)

Description

The solution to a system of N linear equations, AX = B, in N unknowns may be

found by simultaneously performing Gaussian elimination (Burden and Faires
1985,291-304) on the matrix containing the coefficients of the equations (the coef-
ficient matrix A) and the vector containing the constant terms of the equations (the
constant vector B). First, elementary row operations are used to make A upper
triangular (that is, all the elements below the main diagonal are zero). Backward
substitution (whereby X[N] is calculated and used to calculate X[N -1], which is
then used to calculate X[N - 2], and so on) is then used to compute the solution
vector X. If one or more of the elements on the main diagonal of the upper triangu-
lar matrix is zero, then no unique solution to the system exists.

User- Defined Types

TNvector = array[l •• TNArraySize] of Real;

TNmatrix = array[l .. TNArraySize] of TNvector;

Input Parameters

Dimen : Integer; Dimension of the coefficients matrix

Coefficients: TNmatrix; The square matrix containing the coefficients of the equa-
tions
Constants: TNvector; The constant terms of each equation
The preceding parameters must satisfy the following conditions:
1. Dimen > 0
2. Dimen:5 TNArraySize

TNArraySize sets an upper bound on the number of elements in each vector. It

is used in the type definition of TNvector and TNmatrix. TNArraySize is not a
variable name and is never referenced by the procedure; hence there is no test for

114 Turbo Numerical Methods Toolbox

condition 2. If condition 2 is violated, the program will crash with an Index Out of
Range error.

Output Parameters

So 1ut ion : TNvector; Solution to the set of equations.

Error: Byte; 0: No errors.
1: Dimen < 1.
2: Coefficients matrix is singular; no unique solution exists.

Syntax of the Procedure Call

Gaussian_Elimination(Dimen, Coefficients, Constants, Solution, Error);

Sample Program

The sample program GAUSELIM.PAS provides I/O functions that demonstrate

how to solve a system of linear equations with Gaussian elmination.

Input File
Data may be input from a text file. All entries in the text file should be separated
by a space or carriage return, and it does not matter if the text file ends with a
carriage return. The format of the text file should be as follows:
1. The dimension of the coefficient matrix
2. The elements of the matrix in row order; that is,
[1, 1], [1, 2], ... , [1, N], [2, 1], "', [2, N], ... , [N, N],
where N is the dimension of the matrix
3. The elements of the constant vector, in the order [l], ... ,[N]

Matrix Routines 115

For example, to solve the system
2x + 3y = 10
-4x = 10
a text file could be created to look like this:
2
2 3
-4 0
10
10

Example
Problem. Solve the following linear system:
w + 2x + Oy - z = 10.0
- w + 4x + 3y - 0.5z = 21.5
2w + 2x + y - 3z = 26.0
3y - 4z = 37.0
Run GAUSELIM.PAS:
(K)eyboard or (F)ile input of data? F
File name? 5AMPLE6A.DAT
Direct output to one of the following:
(5) creen
(P)rinter
(F) il e

The coefficients:
1.000000000 2.000000000 0.000000000 -1.000000000
-1.000000000 4.000000000 3.000000000 -0.500000000
2.000000000 2.000000000 1.000000000 -3.000000000
0.000000000 0.000000000 3.000000000 -4.000000000

The constants:
1.00000000000000E+001
2. 15000000000000E+001
2.60000000000000E+001
3.70000000000000E+001

The solution:
-1.00000000000000E+000
2.00000000000000E+000
3.00000000000000E+000
-7.00000000000000E+000

116 Turbo Numerical Methods Toolbox

Solving a System of Linear Equations with Gaussian
Elimination and Partial Pivoting (PARTPIVT.INC)

Description

The solution to a system of N linear equations, AX = B, in N unknowns may be

found by simultaneously performing Gaussian elimination (Burden and Faires
1985, 291-304) on the matrix containing the coefficients of the equations (the coef-
ficient matrix A) and the vector containing the constant terms of the equations (the
constant vector B). However, excessive round-off errors can occur when elements
on the main diagonal are small compared to the elements below them in the same
column. To avoid this, partial pivoting (maximal column pivoting) is performed
(Burden and Faires 1985, 324-327); that is, row interchanges are performed so that
each main diagonal element is greater than or equal to the elements below it in the
same column. (See Chapter 9 for an application of PARTPIVf.lNC.)

User-Defined Types

TNvector = array[l .. TNArraySize] of Real;

TNmatrix = array[l .• TNArraySize] of TNvector;

Input Parameters

Dimen : Integer; Dimension of the coefficients matrix

Coefficients: TNmatrixi The square matrix containing the coefficients of the equa-
tions
Constants: TNvector; The constant terms of each equation
The preceding parameters must satisfy the following conditions:
1. Dimen > 0
2. Dimen::; TNArraySize

TNArraySize sets an upper bound on the number of elements in each vector. It

is used in the type definition of TNvector and TNmatrix. TNArraySize is not a
variable name and is never referenced by the procedure; hence there is no test for

Matrix Routines 117

condition 2. If condition 2 is violated, the program will crash with an Index Out of
Range error.

Output Parameters

So 1ut ion : TNvector; Solution to the set of equations.

Error : Byte; 0: No errors.
1: Dimen < l.
2: Coefficients matrix is singular; no unique solution exists.

Syntax of the Procedure Call

Partial_Pivoting(Dimen, Coefficients, Constants, Solution, Error);

Sample Program

The sample program PARTPIVT.PAS provides I/O functions that demonstrate how
to solve a system of linear equation with Gaussian elimination and partial pivoting.

Input File
Data may be input from a text file. All entries in the text file should be separated
by a space or carriage return, and it does not matter if the text file ends with a
carriage return. The format of the text file should be as follows:
1. The dimension of the matrix
2. The elements of the matrix in row order; that is,
[1, 1], [1, 2], ... , [1, N], [2, 1], ... , [2, N], ... , [N, N],
where N is the dimension of the matrix
3. The elements of the constant vector, in the order [l], ... ,[N]

118 Turbo Numerical Methods Toolbox

For example, to solve the system
2x + 3y = 10
-4x = 10
a text file could be created to look like this:
2
2 3
-4 0
10
10

Example
Problem. Solve the following linear system:
w + 2x + Oy - z = 10
-w + 4x + 3y - 0.5z = 21.5
2w + 2x + Y - 3z = 26
3y - 4z = 37
Run 'PARTPIVf.PAS:
{K)eyboard or {F)ile input of data? F
File name? SAMPLE6A.DAT
Direct output to one of the following:
{S)creen
{P)rinter
(F) i 1e

The coefficients:
1.000000000 2.000000000 0.000000000 -1.000000000
-1.000000000 4.000000000 3.000000000 -0.500000000
2.000000000 2.000000000 1.000000000 -3.000000000
0.000000000 0.000000000 3.000000000 -4.000000000

The constants:
1.00000000000000Et001
2. 15000000000000Et001
2.60000000000000Et001
3.70000000000000Et001

The solution:
-1.00000000000000EtOOO
2.00000000000000EtOOO
3.00000000000000EtOOO
-7.00000000000000EtOOO

Matrix Routines 119

Solving a System of Linear Equations with Direct Factoring
(DIRFACT.INC)

Description

The solution to a system of N linear equations, AX = B, in N unknowns can be

computed by factoring the matrix containing the coefficients of the N equations
(the coefficient matrix A) into an upper triangular matrix U (that is, all the elements
below the main diagonal are zero) and a lower triangular matrix L (that is, all the
elements above the main diagonal are zero) such that A = LU. Partial pivoting is
used to reduce round-off error. A record of the pivoting permutations are recorded
in a permutation matrix P, so that the equation is actually A = PLU. Forward
substitution (analogous to backward substitution; see "Solving a System of Linear
Equations with Gaussian Elimination") is used to solve the equations LZ = B
(actually LZ = PB, where P is the pivoting permutation matrix) and UX = Z
(where X is the solution to the N linear equations, and Z is an intermediate solu-
tion). If the coefficient matrix cannot be factored into nonsingular triangular
matrices, then no unique solution exists.
This module includes two procedures to perform this algorithm. Procedure
LUJJecompose performs the LV decomposition of a matrix, and procedure
LU-.Solve performs forward and backward substitution to solve the linear equa-
tions. Both procedures are in the include file DIRFACT.INC.
The most efficient way to calculate the solutions to several systems with the
same coefficient matrix but different constant vectors is to first decompose the
coefficient matrix A into Land U (Burden and Faires 1985, 342-349). Then per-
form backward substitution on this decomposed matrix and each of the constant
vectors B. Thus, the coefficient matrix is decomposed only once.

User-Defined Types

TNvector = array[l .. TNArraySize] of Real;

TNmatrix = array[l •• TNArraySize] of TNvector;

120 Turbo Numerical Methods Toolbox

Procedure LU-Decompose Input Parameters

Dimen : Integer; Dimension of the coefficients matrix

Coefficients: TNmatrix; Square matrix containing the coefficients of the equations

The preceding parameters must satisfy the following conditions:

l. Dimen > 0
2. Dimen ~ TNArraySize

TNArraySize fixes an upper bound on the number of elements in each vector. It

is used in the type definition of TNvector and TNmatrix. TNArraySize is not a
variable name and is never referenced by the procedure; hence there is no test for
condition 2. If condition 2 is violated, the program will crash with an Index Out of
Range error.

Procedure LU-Decompose Output Parameters

Decomp : TNmatrix; The LV decomposition of the coefficients matrix.

Permute: TNmatrix; A permutation matrix that records the effects of pivoting.
Error: Byte; 0: No errors.
1: Dimen < l.
2: The coefficients matrix is singular.

Syntax of the Procedure Call

LU_Decompose(Dimen, Coefficients, Decomp, Permute, Error);

Procedure LU-Solve Input Parameters

Dimen : Integer; Dimension of the coefficients matrix

Decomp : TNmatrix; The LV decomposition of the coefficients matrix
Constants: TNmatrix; The constant terms of each equation
Permute: TNmatri x; A permutation matrix that records the effects of pivoting

Matrix Routines 121

The preceding parameters must satisfY the following conditions:
1. Dimen > 0
2. Dimen::; TNArraySize

TNArraySize fixes an upper bound on the number of elements in each vector. It

is used in the type definition of TNvector and TNmatrix. TNArraySize is not a
variable name and is never referenced by the procedure; hence there is no test for
condition 2. If condition 2 is violated, the program will crash with an Index Out of
Range error.

Procedure LU...solve Output Parameters

Solution: TNvector; Solution to each system of equations

Error: Byte; 0: No errors
1: Dimen < 1

Syntax of the Procedure Call

LU_Solve(Dimen, Decomp, Constants, Permute, Solution, Error);

Sample Program

The sample program DIRFACT.PAS provides I/O functions that demonstrate how
to solve a system of linear equations with the method of direct factoring.

Input File
Data may be input from a text file. All entries in the text file should be separated
by a space or carriage return, and it does not matter if the text file ends with a
carriage return. The format of the text file should be as follows:
l. The dimension of the matrix
2. The elements of the matrix in row order; that is,
[1, 1], [1, 2], ... , [1, N], [2, 1], ... , [2, N], ... , [N, N],
where N is the dimension of the matrix
3. The elements of the first constant vector, in the order [1], ... ,[N], with each
element followed by a carriage return

122 Turbo Numerical Methods Toolbox

4. The elements of any additional constant vectors, in the order [l], ... ,[N],
with each element followed by a carriage return

For example, to solve the systems

2x + 3y = 10 2x + 3y = 1
-4x = 10 -4x = 2
a text file could be created to look like this:
2
2 3
-4 0
10
10
1
2

Example
Problem. Given the following set of coefficients:
2w + x + 5y - 8z
7w + 6x + 2y + 2z
- 1w - 3x - lOy + 4z
2w + 2w + 2y + z
compute solutions for each of the five constant vectors:
-15 14 -13
[ 17
0
-1~
50
-5 -12
1 84
-51
30
-15
5]
17 1 37 10
Run DIRFACT.PAS:
(K)eyboard or (F)ile input of data? F
File name? SAMPLE6C.DAT
Direct output to one of the following:
(S)creen.
(P)rinter
(F) il e

Matrix Routines 123

The coefficients:
2.000000000 1.000000000 5.000000000 -8.000000000
7.000000000 6.000000000 2.000000000 2.000000000
-1.000000000 -3.000000000 -10.000000000 4.000000000
2.000000000 2.000000000 2.000000000 1.000000000

The constants:
O.OOOOOOOOOOOOOOE+OOO
1.70000000000000E+001
-1.00000000000000E+001
7.00000000000000E+000

The solution:
9.99999999999999E-001
1.00000000000000E+000
9.99999999999999E-001
9.99999999999999E-001

The constants:
-1.50000000000000E+001
5.00000000000000E+001
-5.00000000000000E+000
1.70000000000000E+001

The solution:
2.00000000000000E+000
4.99999999999999E+000
1.85358974546113E-015
3.00000000000000E+000

The constants:
1.40000000000000E+001
1.00000000000000E+000
-1.20000000000000E+001
1.00000000000000E+000

The solution:
1.00000000000000E+000
-1.00000000000000E+000
1.00000000000000E+000
-1.00000000000000E+000

The constants:
-1.30000000000000E+001
8.40000000000000E+001
-5.10000000000000E+001
3.70000000000000E+001

The solution:
3.99999999999999E+000
5.00000000000001E+000
6.00000000000000E+000
7.00000000000000E+000

124 Turbo Numerical Methods Toolbox

The constants:
5.00000000000000E+OOO
3.00000000000000E+OOl
-1.50000000000000E+OOl
1.OOOOOOOOOOOOOOE+OOl

The solution:
-1.01506105108586E-015
5.00000000000000E+OOO
O.OOOOOOOOOOOOOOE+OOO
O.OOOOOOOOOOOOOOE+OOO

Matrix Routines 125

Solving a System of Linear Equations with the Iterative
Gauss-Seidel Metlwd (GAUSSIDL.INC)

Description

The solution to a system of N linear equations, AX = B, in N unknowns can be

approximated by the Gauss-Seidel iterative technique (Burden and Faires 1985,
424-432). The equation AX = B is transformed into X = TX + C. Given an initial
approximation Xo' the sequence Xm = TXm _ 1 + C is generated with the following
formula:
i-I N

I
j=I
A[i,j] X",[j] - I
j=i+l
(A[i,j] Xm_Jj]) + B[i]
X,Ji] =
A[i,i]

The algorithm halts when the fractional difference for each element of the vector
X between two iterations is less than a specified tolerance.
If A is diagonally dominant (that is, each of the diagonal terms is greater than or
equal to the sum of the off-diagonal terms in the same row), then the sequence will
converge to the solution X. If the matrix A is not diagonally dominant, then the
sequence may converge to the solution, but more likely it will not. You must supply
the tolerance with which to approximate a solution.

User- Defined Types

TNvector = array[l .. TNArraySize] of Real;

TNmatrix = array[l .. TNArraySize] of TNvector;

126 Turbo Numerical Methods Toolbox

Input Parameters

Dimen : Integer; Dimension of the coefficients matrix

Coeffi ci ents : TNmatri X; The square matrix containing the coefficients of the equa-
tions
Constants: TNvector; The constant terms of the equation
To1 : Real; Indicates accuracy in solution
MaxIter : Real; Maximum number of iterations
The preceding parameters must satisfy the following conditions:
J. Dimen > O.
2. Dimen =::; TNArraySize.
3. Tal> O.
4. MaxIter ~ O.
5. The coefficients matrix may not contain a zero on the main diagonal.

TNArraySize sets an upper bound on the number of elements in each vector. It

is used in the type definition of TNvector and TNmatrix. TNArraySize is not a
variable name and is never referenced by the procedure; hence there is no test for
condition 2. If condition 2 is violated, the program will crash with an Index Out of
Range error.

Output Parameters

Solution: TNvector; Solution to the set of equations.

Iter: Real; The number of iterations required to find the solution.
Error: Byte; 0: No errors.
1: Iter> MaxIter and matrix is not diagonally dominant.
2: Iter > MaxIter and matrix is diagonally dominant.
3: Dimen < 1.
4: Tal =::; O.
5: MaxIter < O.
6: Zero on the diagonal of the coefficients matrix.
7: Sequence is diverging.

Matrix Routines 127

If the coefficients matrix is diagonally dominant, then the Gauss-Seidel method
will converge to a solution. If the coefficients matrix is not diagonally dominant,
then the Gauss-Seidel may converge to a solution, but more likely it will not. Error
7 can only occur when the coefficients matrix is not diagonally dominant. If Error 1
is returned, it is likely that convergence is not possible; if Error 2 is returned,
convergence is possible but will take more than MaxIter iterations.
If the diagonal of the coefficients matrix contains a zero (Error 6), then the
Gauss-Seidel method may not be used to solve the system of equations.
If the system of equations is under-determined, the Gauss-Seidel method will
still converge to a (nonunique) solution. The Gauss-Seidel method cannot distin-
guish between unique and nonunique solutions. If you suspect that your system of
equations is under-determined, use one of the direct methods (for example,
GAUSELIM.INC) to attempt a solution; Gaussian elimination will give an error if
it is under-determined. Alternatively, you could use DET.INC to find the determi-
nant; if the determinant is zero, then the system is under-determined.

Syntax of the Procedure Call

Gauss-Seidel(Dimen, Coefficients, Constants, Tol, MaxIter, Solution, Iter, Error);

Sample Program

The sample program GAUSSIDL.PAS provides I/O functions that demonstrate

how to solve a system of linear equations with the iterative Gauss-Seidel method.

Input File
Data may be input from a text file. All entries in the text file should be separated
by a space or carriage return, and it does not matter if the text file ends with a
carriage return. The format of the text file should be as follows:
l. The dimension of the matrix
2. The elements of the matrix in row order; that is,
[1, 1], [1, 2], "', [1, N], [2, 1], ... , [2, N], ... , [N, N],
where N is the dimension of the matrix
3. The elements of the first constant vector, in the order [l], ... ,[N]

128 Turbo Numerical Methods Toolbox

For example, to solve the systems
20x + 3y = 10
-4y = 10
a text file could be created to look like this:
2
20 3
o -4
10
10

Example
Problem. Solve the following linear system to within a tolerance of IE - 12:
10v + w + 2x - 3y + 2z = 29
4v + SOw + x + z = 35
- 2v + 5w - 30x + Y + z = 25
6v + 4w + lOy + 3z = 46
- 3v - 2w - x + 6y + 25z = - 106
Run GAUSSIDL.PAS:
(K)eyboard or (F)ile input of data? F

File name? SAMPLE6D.DAT

Tolerance (> 0, default = 1.000E-08): 1E-12

Maximum number of iterations (> 0, default = 100): 100

Direct output to one of the following:

(S)creen
(P)rinter
(F)ile

The coefficients:
10.000000000 1.000000000 2.000000000 -3.000000000 2.000000000
4.000000000 50.000000000 1.000000000 0.000000000 1.000000000
-2.000000000 5.000000000 -30.000000000 1.000000000 1.000000000
6.000000000 4.000000000 0.000000000 10.000000000 3.000000000
-3.000000000 -2.000000000 -1.000000000 6.000000000 25.000000000

Matrix Routines 129

The constants:
-2.90000000000000E+001
3.50000000000000E+001
-2.50000000000000E+001
-4.60000000000000E+001
-1.06000000000000E+002
Tolerance: 1.00000000000000E-012
Maximum number of iterations: 100

Number of iterations: 15
The result:
-2.99999999999997E+000
9.99999999999999E-001
9.99999999999998E-001
-1.99999999999999E+000
-4.00000000000000E+000

130 Turbo Numerical Methods Toolbox

c H A p T E R 7
Eigenvalues and Eigenvectors

The routines in this chapter can find the eigenvalues and eigenvectors. A scalar c is
an eigenvalue (or characteristic value) of a square matrix A if there is a nonzero
vector v satisfying
Av = cv
The vector v is called the eigenvector corresponding to c.
The eigenvalues and eigenvectors of a matrix provide a lot of information about
the matrix. If a matrix is written in terms of a basis of eigenvectors, then it is
diagonal, meaning. that its only nonzero terms are on the main diagonal.
Each procedure in this chapter attempts to approximate at least' one real eigen-
value (and associated eigenvector) of a real square matrix. The eigenvector is nor-
malized so that the element with the largest magnitude is 1.
The power method (POWER.INC) approximates the eigenvalue that is largest
in magnitude (dominant eigenvalue). The iterative process will converge slowly or
not at all if the dominant eigenvalue is not simple or if it has nearly the same
magnitude as the next most-dominant eigenvalue.
The inverse power method (INVPOWER.INC) approximates the eigenvalue
nearest to a user-supplied real value. This process usually converges more rapidly
than the power method, and may be used to refine the approximate value of the
eigenvalue determined by the latter method (POWER.INC).

131
The Wielandt method (WIELANDT.lNC) attempts to approximate a user-
specified number of eigenvalues of a given matrix. The power method (POWER.
INC) is first used to approximate the dominant eigenvalue of the matrix. Deflation
is employed to form a deflated, square matrix (that is, a square matrix whose
dimension is one less than the original matrix). The eigenvalues of the deflated
matrix are identical to those of the original matrix except for the determined domi-
nant eigenvalue. Eigenvectors of the remaining eigenvalues from the original
matrix are also contained in the deflated matrix. The dominant eigenvalue of the
new deflated" matrix is then determined using the power method. Wielandt's
method is susceptible to round-off error, thus it may be desirable to use its results
as input to the inverse power method (INVPOWER.lNC).
The cyclic Jacobi method QACOBI.INC) approximates all the eigenvalues of a
symmetric matrix. The iterative process uses orthogonal plane rotations to reduce
the given matrix into a diagonal form. Although Jacobi's method is only applicable
to symmetric matrices, it is much more efficient and accurate than Wielandt's
method.

132 Turbo Numerical Methods Toolbox

Real Dominant Eigenvalue and Eigenvector of a Real
Matrix Using the Power Method (POWERINC)

Description

The power method (Burden and Faires 1985, 452-456) approximates the dominant
real eigenvalue of a matrix and its associated eigenvector. The dominant eigen-
value is the eigenvalue of the largest absolute magnitude. Given a square matrix A
and a real nonzero vector v, a vector w is constructed by the matrix operation
Av = w. The vector w is normalized by dividing by its element of the largest
absolute magnitude q. If the absolute difference between each of the correspond-
ing elements in wand v is less than a specified tolerance, then the procedure halts.
Otherwise, v is set equal to w, and the operation repeats until a solution is found.
The magnitude q is the dominant eigenvalue, and w will be the associated eigen-
vector of the matrix A.
You must supply the matrix A, an initial approximation to the eigenvector v, and
the tolerance.

User-Defined Types

TNvector = array[l •• TNArraySize] of Real;

TNmatrix = array[l .. TNArraySize] of TNvector;

Input Parameters

Dimen : Integer; Dimension of the matrix Mat

Mat: TNmatrix; The matrix
GuessVector : TNvector; Initial approximation to the eigenvector
MaxIter : Integer; Maximum number of iterations
Tolerance: Real; Indicates accuracy in solution

Eigenvalues and Eigenvectors 133

The preceding parameters must satisfy the following conditions:
1. Dimen > 1
2. Dimen S TNArraySize
3. Tolerance> 0
4. MaxIter > 0

TNArraySize fixes an upper bound on the number of elements in each vector. It

is used in the type definition of TNvector and TNmatrix. TNArraySize is not a
variable name and is never referenced by the procedure; hence there is no test for
condition 2. If condition 2 is violated, the program will crash with an Index Out of
Range error (assuming the directive {$R + } is active).

Output Parameters

Eigenval ue : Real; Approximation to the dominant eigenvalue of the matrix

Eigenvector: TNvector; Approximate eigenvector associated with the dominant
eigenvalue
Iter: Integer; Number of iterations required to find the solution
Error: Byte; 0: No errors
1: Dimen S 1
2: Tolerance S 0
3: MaxIter S 0
4: Iter;:::: MaxIter

Syntax of the Procedure Call

Power(Dimen, Mat, GuessVector, MaxIter, Tolerance,

Eigenvalue, Eigenvector, Iter, Error);

Comments

The power method will not converge if the initial approximation (Guess) to the
eigenvector is orthogonal to the dominant eigenvector. If the initial approximation
is orthogonal, then the power method will converge to a different eigenvector with-
out warning. If you suspect this has happened, run the routine with several differ-
ent initial approximations.

134 Turbo Numerical Methods Toolbox

The power method may not converge to repeated eigenvalues with linearly
dependent eigenvectors. Repeated eigenvalues with linearly independent eigen-
vectors do not pose a problem.
The eigenvectors are normalized such that the element of largest absolute mag-
nitude in each vector is equal to one.

Sample Program

The sample program POWER.PAS provides I/O functions that demonstrate the
power method of approximating eigenvalues.

Input File
Data may be input from a text file. Entries in the text file should be separated by
spaces or carriage returns, and it does not matter if the text file ends with a carriage
return. The format of the text file should be as follows:
1. Dimension of the matrix
2. Elements of the matrix, in the order
[1, 1], [1, 2], ... , [1, N], ... , [N, 1], ... , [N, N],
where N is the dimension of the matrix

For example, to find the dominant eigenvalue of the matrix

you could first create the following text file:

4
1
2
3
4

Example
Problem. Find the dominant eigenvalue of the matrix:

[ ~ 1~ ~]
024
using the initial guess (1, 2, 3).

Eigenvalues and Eigenvectors 135

Run POWER.PAS:
(K)eyboard or (F)ile entry of data? K
Dimension of the matrix (1-60)? 3
Matrix[l, 1]: 2
Matrix[l, 2]: 10
Matrix[l, 3]: 0
Matrix[2, 1]: 0
Matrix[2, 2]: 1
Matrix[2, 3]: 0
Matrix[3, 1]: 0
Matrix[3, 2]: 2
Matrix[3, 3]: 4
Now input an initial guess for the eigenvector:
Vector[l]: 1
Vector[2]: 2
Vector[3]: 3
Tolerance (> 0, default = 1.000E-06): 1E-8
Maximum number of iterations (> 0, default 100): 100
Direct output to one of the following:
(S)creen
(P)rinter
(F)ile
The matrix:
2.00000000000000E+00 1.00000000000000E+01 O.OOOOOOOOOOOOOOE+OO
O.OOOOOOOOOOOOOOE+OO 1.00000000000000E+00 O.OOOOOOOOOOOOOOE+OO
O.OOOOOOOOOOOOOOE+OO 2.00000000000000E+00 4.00000000000000E+00
Tolerance: 1.00000000000000E-008
Maximum number of iterations: 100
Number of iterations: 12
The approximate eigenvector:
-2.30295155112597E-014
3. 15544362088405E-030
1.00000000000000E+000
The associated eigenvalue: 4.00000000000000E+000
The exact solution is
Eigenvalue = 4
Eigenvector = (0, 0, 1)

136 Turbo Numerical Methods Toolbox

Real Eigenvalue and Eigenvector of a Real Matrix Using
the Inverse Power Method (INVPOWERINC)

Description

Where the power method converges to the dominant real eigenvalue of a matrix
(see POWER.INC), the inverse power method (Burden and Faires 1985, 459-462)
converges to the real eigenvalue nearest to a user-supplied real value. Given a
square matrix A, an initial approximation p to the eigenvalue, and an initial approx-
imation v to the eigenvector, the linear system (A - pI)w = v (where I is the
identity matrix) is solved via LV decomposition (see Chapter 6, "Solving a System
of Linear Equations with Direct Factoring"). The vector w is normalized by divid-
ing through by the element q with the largest absolute magnitude. If the absolute
difference between each of the corresponding elements in v and w is less than a
specified tolerance, then the procedure halts. Otherwise, v is set equal to w, and
the previous matrix equation is solved again. The process repeats until a solution is
reached. The eigenvalue of A closest to p will be (l/q + p), and w will be the
associated eigenvector.
You must supply the matrix A, the initial approximations p and v, and the toler-
ance.

User-Defined Types

TNvector = array[l •• TNArraySize] of Real;

TNmatrix = array[l •• TNArraySize] of TNvector;

Input Parameters

Dimen : Integer; Dimension of the matrix Mat

Mat: TNmatri x; The matrix
GuessVector : TNvector; Initial approximation (Guess) of the eigenvector
ClosestVal : Real; The approximate eigenvalue
Max Iter : Integer; Maximum number of iterations
To 1erance : Real; Indicates accuracy of solution

Eigenvalues and Eigenvectors 137

The preceding parameters must satisfy the following conditions:
1. Dimen> 1
2. Dimen S TNArray Size
3. Tolerance > 0
4. MaxIter > 0

TNArray Size sets an upper bound on the number of elements in each vector. It
is used in the type definition of TNvector and TNmatrix. TNArraySize is not a
variable name and is never referenced by the procedure; hence there is no test for
condition 2. If condition 2 is violated, the program will crash with an Index Out of
Range error (assuming the directive {$R +} is active).

Output Parameters

Eigenval ue : Real; Approximation to the eigenvalue closest to ClosestVal

Eigenvector: TNvector; Approximation to the eigenvector associated with Eigen-
value
Iter: Integer; Number of iterations required to find the solution
Error : Byte; 0: No errors
1: Dimen S 1
2: Tolerance S 0
3: MaxIter S 0
4: Iter ~ MaxIter
5: Eigenvalue/Eigenvector not calculated (see "Comments")

Syntax of the Procedure Call

InversePower(Dimen, Mat, GuessVector, ClosestVal, MaxIter,

Tolerance, Eigenvalue, Eigenvector, Iter, Error);

138 Turbo Numerical Methods Toolbox

Comments

The inverse power method approximates the solution of a system of linear equa-
tions. If the matrix (Mat - Eigenvalue * I) is singular, where I is the identity rnatrix,
the method will not converge to a solution and Error 5 will be returned. If this
occurs, run the routine again with a slightly different initial approximation,
ClosestVal.
The power method may not converge to repeated eigenvalues with linearly
dependent eigenvectors. Repeated eigenvalues with linearly independent eigen-
vectors do not pose a problem.
The inverse power method is sensitive to the initial approximation (ClosestVal).
If ClosestVal is not close to an eigenvalue or lies midway between two eigenvalues,
the algorithm will converge very slowly, if at all.
The eigenvectors are normalized such that the element of the largest absolute
magnitude in each vector is equal to one.

5mnple Program

The sample program INVPOWER.PAS provides I/O functions that demonstrate

the inverse power method of approximating eigenvalues.

Input File
Data may be input from a text file. Entries in the text file should be separated by
spaces or carriage returns, and it does not matter if the text file ends with a carriage
return. The format of the text file should be as follows:
1. Dimension of the matrix
2. Elements of the matrix, in the order
[1, 1], [1, 2], ... , [1, N], "', [N, 1], ... , [N, N],
where N is the dimension of the matrix
3. Elements of the initial guess, in the order [1], [2], ... , [N],
where N is the dimension of the matrix

Eigenvalues and Eigenvectors 139

For example, to find an eigenvalue of the matrix

with an initial guess of (11, 10), you could first create the following text file:
4
1
2
3
4
11
10

Example
Problem. Suppose you know that two of the eigenvalues of the matrix

2 10 0]
010
[
024
are approximately 1.999 and 0.7. Use the inverse power method with an initial
guess of (1, 2, 3) to refine these approximations.
Run INVPOWER.PAS with 1.999 as the approximate eigenvalue:
(K)eyboard or (F)i1e entry of data? K
Dimension of the matrix (1-30)? 3
Matrix[1. 1]: 2
Matrix[1. 2]: 10
Matrix[1. 3]: 0
Matrix[2. 1]: 0
Matrix[2. 2]: 1
Matrix[2. 3]: 0
Matrix[3. 1]: 0
Matrix[3. 2]: 2
Matrix[3. 3]: 4
Now input an initial guess for the eigenvector:
Vector[1]: 1
Vector[2]: 2
Vector[3]: 3
Approximate eigenvalue (default = 5.2857): 1.999
To1 erance (> O. default = 1.0.00E-06): 1E-8
Maximum number of iterations (> O. default = 200): 200

140 Turbo Numerical Methods Toolbox

Direct output to one of the following:
{S}creen
{P}rinter
{F)ile

The matrix:
2.00000000000000E+000 1.00000000000000E+00I O.OOOOOOOOOOOOOOE+OOO
O.OOOOOOOOOOOOOOE+OOO 1.00000000000000E+000 O.OOOOOOOOOOOOOOE+OOO
O.OOOOOOOOOOOOOOE+OOO 2.00000000000000E+000 4.00000000000000E+000
Approximate eigenvalue: 1.99900000000000E+000
Tolerance: 1.00000000000000E-008
Maximum number of iterations: 200

Number of iterations: 4
The approximate eigenvector:
1.00000000000000E+000
9. 12736381850482E-014
-5. 13983108970145E-014
The associated eigenvalue: 2.00000000000091E+000
Run INVPOWERPAS with 0.7 as the approximate eigenvalue:
{K}eyboard or {F}ile entry of data? K
Dimension of the matrix {1-30}? 3
Matrix[l, 1]: 2
Matrix[l, 2]: 10
Matrix[l, 3]: 0
Matrix[2, 1]: 0
Matrix[2, 2]: 1
Matrix[2, 3]: 0
Matrix[3, 1]: 0
Matrix[3, 2]: 2
Matrix[3, 3]: 4
Now input an initial guess for the eigenvector:
Vector[l]: 1
Vector[2]: 2
Vector[3]: 3
Approximate eigenvalue {default = 5.2857}: 0.7
Tolerance {> 0, default = 1.000E-06}: 1E-8
Maximum number of iterations {> 0, default = 200}: 200
Direct output to one of the following:
{S}creen
{P}rinter
{F)ile

Eigenvalues and Eigenvectors 141

The matrix:
2.00000000000000E+000 1.00000000000000E+001 O.OOOOOOOOOOOOOOE+OOO
O.OOOOOOOOOOOOOOE+OOO 1.00000000000000E+000 O.OOOOOOOOOOOOOOE+OOO
O.OOOOOOOOOOOOOOE+OOO 2.00000000000000E+000 4.00000000000000E+000
Approximate eigenvalue: 7.00000000000000E-001
Tolerance: 1.00000000000000E-008
Maximum number of iterations: 200

Number of iterations: 12
The approximate eigenvector:
1.00000000000000E+OOO
-1.00000002395103E-001
6.66666682633328E-002

The associated eigenvalue: 9.99999976048973E-001

The exact solutions are
Eigenvalue = 2; Eigenvector = (1, 0, 0)
Eigenvalue = 1; Eigenvector = (1, - 0.1, 2/30)

142 Turbo Numerical Methods Toolbox

Real Eigenvalues and Eigenvectors of a Real Matrix Using
the Power Method and Wielandt's Deflation
(WIELANDT.INC)

Description

Wielandt's deflation is a technique that approximates each real eigenvalue and

related eigenvector of a matrix (Burden and Faires 1985,452-456). Once the domi-
nant real eigenvalue/vector of a matrix has been approximated with the power
method (see "Real Dominant Eigenvalue and Eigenvector of a Real Matrix Using
the Power Method"), the next most dominant real eigenvalue/vector is approxi-
mated by removing the dominant solution. This deflates the matrix. The deflated
matrix will have the same eigenvalues as the original matrix (except for the
removed ones). The eigenvectors of the deflated matrix will be related to the
eigenvectors of the original matrix. (They will not be identical because the dimen-
sion of the deflated matrix is less than the dimension of the original matrix.) The
power method then approximates the dominant eigenvalue of the deflated matrix.
This process is repeated until the appropriate number (user-supplied) of eigen-
values/vectors have been approximated.
You must supply the matrix, the number of eigenvalues/vectors to approximate,
and the tolerance with which to approximate the eigenvalues/vectors.

User-Defined Types

TNvector = array[l •. TNArraySize] of Real;

TNmatrix = array[l •. TNArraySize] of TNvector;
TNIntVector = array[l •. TNArraySize] of Integer;

Input Parameters

Dimen : Integer; Dimension of the matrix Mat

Mat: TNmatrix; The matrix
Guess: TNvector; Initial approximation (Guess) of an eigenvector

Eigenvalues and Eigenvectors 143

MaxEigens : Integer; Number of eigenvalues/vectors to find (at most, Dimen), (see
"Comments")
Max Iter : Integer; Maximum number of iterations
Tolerance: Real; Indicates accuracy in solution
The preceding parameters must satisfy the following conditions:
l. Dimen> 1
2. Dimen ::;; TNArray Size
3. Tolerance > 0
4. MaxIter> 0
5. MaxEigens > 0
6. MaxEigens ::;; Dimen

TNArraySize sets an upper bound on the number of elements in each vector. It

is used in the type definition of TNvector and TNmatrix. TNArraySize is not a
variable name and is never referenced by the procedure; hence there is no test for
condition 2. If condition 2 is violated, the program will crash with an Index Out of
Range error (assuming the directive {$R + } is active).

Output Parameters

NumEigens : Integer; The number of eigenvectors returned (will be ::;;

MaxEigens).
Ei genva 1ues : TNvector; The first NumEigens eigenvalues of the matrix.
Eigenvectors: TNmatrix; The eigenvectors associated with the eigenvalues.
Iter: TNIntVector; Number of iterations required to find each eigenvalue/
vector.
Error: Byte; 0: No errors.
1: Dimen ::;; l.
2: Tolerance ::;; O.
3: MaxIter ::;; O.
4: MaxEigens ~ 0, MaxEigens > Dimen.
5: Iter ~ MaxIter.
6: Warning! Not a fatal error!
The last two eigenvalues aren't real.

144 Turbo Numerical Methods Toolbox

Syntax of the Procedure Call

Wielandt(Dimen, Mat, Guess, MaxEigens, MaxIter, Tolerance,

NumEigens, Eigenvalues, Eigenvectors, Iter, Error);

Comments

It is often unnecessary to determine the complete eigensystem of a matrix. The

parameter MaxEigens prevents the routine from approximating more eigenvalues/
vectors than needed. For example, if the four most dominant eigenvalues of a
20 X 20 matrix are desired, set MaxEigens equal to 4. The algorithm will halt
when it has approximated the four most dominant eigenvalues, thus saving a con-
siderable amount of time. Note, however, that the dimension of the vector
eigenvalues and the matrix eigenvectors must still be TNArraySize (that is, the
same as the dimension of the matrix).
The power method may not converge to repeated eigenvalues with linearly
dependent eigenvectors. Repeated eigenvalues with linearly independent eigen-
vectors do not pose a problem.
The eigenvectors are normalized such that the element of the largest absolute
magnitude in each vector is equal to one.
This routine stores much information on the heap. If you try to compute all the
eigenvalues of a large matrix (say, all 20 of a 20 X 20 matrix), you may get run-time
error $FF, Heap/Stack collision. If this happens, the dimension of TNvector and
TNmatrix should be reduced as much as possible. If this is not possible, then
remove any RAM-resident software (for example, SideKick, SuperKey, or a print
buffer).
It is difficult to determine why the power method doesn't converge to a particu-
lar eigenvector; usually the eigenvalue is complex, or eigenvectors of repeated
eigenvalues are linearly dependent. However, when Wielandt's deRation has
deRated the matrix to a 2 X 2, it is easy to determine if the eigenvalues of the
2 X 2 are real or complex. If the last two eigenvalues are real, then they (and their
associated eigenvectors) are returned; if the last two eigenvalues are complex,
Error 6 is returned. (Error 6 is only a warning; it is not a fatal error.) It is returned
to give you some information about the undetermined eigenvectors.
This procedure uses the New/Dispose procedures to manipulate the heap and
should not be used in any program that uses Mark/Release to manipulate the heap.

Eigenvalues and Eigenvectors 145

Sample Program

The sample program WIELANDT.PAS provides I/O functions that demonstrate

Wielandt's method of approximating eigensystems.

Input File
Data may be input from a text me. Entries in the text me should be separated by
spaces or carriage returns, and it does not matter if the text me ends with a carriage
return. The format of the text me should be as follows:
1. Dimension of the matrix
2. Elements of the matrix, in the order
[1, 1], [1, 2], ... , [1, N], ... , [N, 1], "', [N, N],
where N is the dimension of the matrix

For example, to find the dominant eigenvalue of the matrix

you could first create the following text me:

4
1
2
3
4

Example
Problem. Find all real eigenvalues and eigenvectors of the matrix

2 100]
[
o 10
o 24
using an initial guess of (1, 2, 3).

146 Turbo Numerical Methods Toolbox

Run Wielandt.PAS:
(K)eyboard or (F)ile entry of data? K
Dimension of the matrix (1-10)? 3
Matrix[l, 1]: 2
Matrix[l, 2]: 10
Matrix[l, 3]: 0
Matrix[2, 1]: 0
Matrix[2, 2]: 1
Matrix[2, 3]: 0
Matrix[3, 1]: 0
Matrix[3, 2]: 2
Matrix[3, 3]: 4
Now input an initial guess for the eigenvector:
Vector[l]: 1
Vector[2]: 2
Vector[3]: 3
Tolerance (> 0, default = 1.000E-06): 1E-6
Maximum number of eigenvalues/eigenvectors to find «= 3, default 3): 3
Maximum number of iterations (> 0, default = 200): 200
Direct output to one of the following:
(S)creen
(P)rinter
(F) i1 e

The matrix:
2.00000000000000E+000 1.00000000000000E+001 O.OOOOOOOOOOOOOOE+OOO
O.OOOOOOOOOOOOOOE+OOO 1.00000000000000E+000 O.OOOOOOOOOOOOOOE+OOO
O.OOOOOOOOOOOOOOE+OOO 2.00000000000000E+000 4.00000000000000E+000

Tolerance: 1.00000000000000E-006
Maximum number of eigenvalues/eigenvectors to ,find: 3
Maximum number of iterations: 200

Number of iterations: 10
The approximate eigenvector:
-8.32731765653921E-007
4.60590249431080E-015
1.00000000000000E+000
The associated eigenvalue: 4.00000000000004E+000

Number of iterations: 0
The approximate eigenvector:
1.00000000000000E+000
O.OOOOOOOOOOOOOOE+OOO
O.OOOOOOOOOOOOOOE+OOO

Eigenvalues and Eigenvectors 147

The associated eigenvalue: 2.00000000000000E+000

Number of iterations: 0
The approximate eigenvector:
1.00000000000000E+000
-9.99999888969117E-002
6.66666592646070E-002
The associated eigenvalue: 9.99999999999991E-001
The exact solution is
Eigenvalue = 4; Eigenvector = (0, 0, 1)
Eigenvalue = 2; Eigenvector = (1, 0, 0)
Eigenvalue = 1; Eigenvector = (1, - 0.1, 2/30)

148 Turbo Numerical Methods Toolbox

The Complete Eigensystem of a Symmetric Real Matrix
Using the Cyclic Jacobi Method UACOBI.INC)

Description

The eigensystem of a symmetric matrix can be computed much more simply and
efficiently than the eigensystem of an asymmetric matrix. The cyclic Jacobi method
(Atkinson and Harley 1983, 154-160) is an iterative technique for approximating
the complete eigensystem of a symmetric matrix to within a given tolerance. It
consists of multiplying the matrix A by a series of rotation matrices RI . The rotation
matrices are chosen so that the elements of the upper triangular part of A (exclud-
ing the diagonal) are systematically annihilated; that is, Rl is chosen so that A[l, 2]
becomes zero, R2 is chosen so that A[l, 3] becomes zero, and so on. Since the matrix
is symmetric, this will also annihilate the lower triangular part of A. Because each
rotation wili probably change the value of elements annihilated in previous rota-
tions, the method is iterative. Eventually, the matrix will be diagonalized. The
eigenvalues will be the elements of the main diagonal of the diagonal matrix; the
eigenvectors will be the corresponding rows of the matrix created by the product of
the rotation matrices RI •

User-Defined Types

TNvector = array [1. . TNArraySize] of Real;

TNmatrix = array[l .• TNArraySize] of TNvector;

Input Parameters

Dimen : Integer; Dimension of the matrix Mat

Mat: TNmatrix; The symmetric matrix
MaxIter : Integer; Maximum number of iterations
Tolerance: Real; Accuracy in solution

Eigenvalues and Eigenvectors 149

The preceding parameters must satisfy the following conditions:
1. Dimen> 1.
2. Dimen :5 TNArraySize.
3. Tolerance > O.
4. MaxIter > O.
5. Mat must be symmetric.

TNArraySize sets an upper bound on the number of elements in each vector. It

is used in the type definition of TNvector and TNmatrix. TNArraySize is not a
variable name and is never referenced by the procedure; hence there is no test for
condition 2. If condition 2 is violated, the program will crash with an Index Out of
Range error (assuming the directive {$R + } is active).

Output Parameters

Ei genva 1ues : TNvector; Approximation to the eigenvalues of the matrix

Eigenvectors: TNmatrix; Approximation to the eigenvectors associated with the
eigenvalues
Iter: Integer; Number of iterations required to find eigenvalues/vectors
Error: Byte; 0: No errors
1: Dimen :5 1
2: Tolerance :5 0
3: MaxIter :5 0
4: Mat not symmetric
5: Iter 2: MaxIter

Syntax of the Procedure Call

Jacobi(Dimen, Mat, MaxIter, Tolerance, Eigenvalues, Eigenvectors, Iter, Error);

150 Turbo Numerical Methods Toolbox

Comments

For symmetric matrices, the Jacobi method is preferred to Wielandt's deRation.

Unlike the power (POWER.lNC) and inverse power (INVPOWER.lNC)
methods, the efficiency of the Jacobi method is not affected by repeated eigen-
values with linearly dependent eigenvectors.
The eigenvectors are normalized such that the element of largest absolute mag-
nitude in each vector is equal to one.

Sample Program

The sample program JACOBI.PAS provides I/O functions that demonstrate

Jacobi's method of approximating the eigensystem of symmetric matrices.

Input File
Data may be input from a text file. Entries in the text file should be separated by
spaces or carriage returns, and it does not matter if the text file ends with a carriage
return. The format of the text file should be as follows:
1. Dimension of the matrix
2. Elements of the matrix, in the order
[1, 1], [1, 2], "', [1, N], ... , [N, 1], ... , [N, N],
where N is the dimension of the matrix

For example, to find the dominant eigenvalue of the matrix

you could first create the following text file:

4
1
2
2
1

Eigenvalues and Eigenvectors 151

Example
Problem. Find the complete eigensystem of the symmetric matrix

[ ~ i ={ =~ 1
-3 -1
-1 -3
1
2
2
1
Run JACOBI.PAS:
{K)eyboard or {F)ile entry of data? F

File name? SAMPLE7A.DAT

Tolerance (> 0, default = 1.000E-06): 1E-8

Maximum number of iterations (> 0, default = 200): 200

Direct output to one of the following:

{S)creen
(P) ri nter
(F) il e

The matrix:
1.000000000 2.000000000 -3.000000000 -1.000000000
2.000000000 1.000000000 -1.000000000 -3.000000000
-3.000000000 -1.000000000 1.000000000 2.000000000
-1.000000000 -3.000000000 2.000000000 1.000000000
Tolerance: 1.00000000000000E-008
Maximum number of iterations: 200

Number of iterations: 4
The approximate eigenvector:
1.00000000000000E+000
1.00000000000000E+000
-1.00000000000000E+000
-1.00000000000000E+000

The associated eigenvalue: 7.00000000000000E+000

The approximate eigenvector:
-9.99999999977159E-001
9.99999999977775E-001
1.00000000000000E+000
-9.99999999999384E-001
The associated eigenvalue: 1.00000000000000E+000
The approximate eigenvector:
1.00000000000000E+000
-9.99999556935431E-001
9.99999999977774E-001
-9.99999556913205E-001

152 Turbo Numerical Methods Toolbox

The associated eigenvalue: -2.99999999999990E+000

The approximate eigenvector:

9.99999556935431E-001
9.99999999999384E-001
9.99999556934815E-001
1.00000000000000E+000
The associated eigenvalue: -1.00000000000010E+000
The exact solution is
Eigenvalue = 7; Eigenvector = (1, 1, -1, -1)
Eigenvalue = 1; Eigenvector = (-1, 1, 1, -1)
Eigenvalue = - 3; Eigenvector = (1, - 1, 1, - 1)
Eigenvalue = -1; Eigenvector = (1, 1, 1, 1)

Eigenvalues and Eigenvectors 153

154 Turbo Numerical Methods Toolbox
c H A p T E R 8
Initial Value and Boundary Value Metlwds

A differential equation is like an ordinary equation except that the unknown is a

function, and derivatives of the function appear in the equation. For example,
f"(x) + f(x) =0
is a differential equation.f"(x) is the second derivative off(x). The solutions are the
functions of the form
f(x) = a * cos(x) + b * sin(x)
The function is uniquely determined by suitable initial conditions, such as
f(O) = 3
f'(O) = 4

in which case the solution is

f(x) = 3 * cos(x) + 4 * sin(x)
The routines in this chapter solve differential equations that are ordinanJ and
linear. A differential equation is ordinary if there is only an independent variable
(that is, the unknown function is a function of only one variable), and thus the
derivatives are ordinary derivatives and not partial derivatives. A differential equa-
tion is linear if the unknown function and its derivatives appear linearly in the
equation.
This chapter describes routines that specifically solve: (1) initial value problems
for nth-order ordinary differential equations, (2) initial value problems for systems
of coupled first-order and second-order ordinary differential equations, and (3)

155
boundary value problems for second-order ordinary differential equations.
Note that these routines work only with ordinary differentiai equations, not par-
tial differential equations. All of the routines in this chapter can solve problems
involving nonlinear equations (except the linear-shooting routine LINSHOT
2.1NC).
Two one-step techniques that solve initial value problems for first-order ordinary
differential equations are implemented. The first technique employs the
fourth-order Runge-Kutta method (RUNGE_I.lNC), also known as the classical
Runge-Kutta method. The second employs the Runge-Kutta-Fehlberg method
(RKF_l.INC).
Each one-step technique approximates the value of the dependent variable at a
mesh point, which is a value of the independent variable, by using only the infor-
mation obtained from the preceding mesh point. The Runge-Ku~ta method
employs equally spaced mesh points. On the other hand, the Runge-Kutta-
Fehlberg method varies the spacing of the mesh points in order to control the local
truncation error. This produces a corresponding bound on the global error.
The Adams-BashforthIAdams-Moulton predictor/corrector method (ADAMS_I.
INC) is a multistep method that uses information obtained at several preceding
mesh points to approximate the value of the dependent variable at the current
mesh point. The procedure employs the Adams-Bashforth four-step method to
obtain a predictor. It is subsequently used as input for the Adams-Moulton three-
step method to obtain a corrector. The corrector is the approximate value of the
solution. Mesh points are equally spaced, and the starting values for the process are
determined by the one step, fourth-order Runge-Kutta method.
The Runge-Kutta methods are the most reliable and should be used when you
are uncertain of the behavior of the differential equation (for example, if the solu-
tion to the differential equation is not very smooth). If you want the output to be
evenly spaced (in x) or do not require a high degree of accuracy, use the classical
Runge-Kutta method. Otherwise, the Runge-Kutta-Fehlberg method is the best
general purpose routine to use, since it provides control over the accuracy of the
solution.
The Adams-Bashforth/Adams-Moulton method achieves the same accuracy (for
equally spaced mesh points) as the fourth-order Runge-Kutta formula, but if is
significantly faster. Consequently, the Adams-Bashforth/Adams-Moulton method is
the most desirable method if you are reasonably certain that the differential equa-
tion is well-behaved.
Initial value problems for first-order ordinary differential equations are guaran-
teed to have a unique solution on the interval a, b if the function
x' = f(t, x)

156 Turbo Numerical Methods Toolbox

is continuous over the interval a, h, and if the function satisfies the Lipshitz condi-
tion. The Lipshitz condition states that there exists a positive number L such that
If(t, x2) - f(t, Xl) I ~ Llx2 - XII
for all a ~ t ~ h, - 00 < X < 00.
Initial value problems for second-order ordinary differential equations can be
solved via a fourth-order Runge-Kutta method (RUNGE--2.INC). This procedure
reduces a given differential equation to a system of two, first-order ordinary differ-
ential equations. The solution to this system is approximated at equally spaced
mesh points with the fourth-order Runge-Kutta method.
Initial value problems for second-order ordinary differential equations are guar-
anteed to have a unique solution on the interval a, h if the function
x" = f(t, x, x')
is continuous over the interval a, h and if the function satisfies the Lipshitz condi-
tion. For a second-order differential equation, the Lipshitz condition states that
there exists a positive number L such that
If(t, x2' X' 2) - f(t, Xl' X' 2) I ~ L( IX2 - XII + Ix'2 - x'II)
for all a ~ t ~ h, - 00 < X < 00, - 00 < X < 00. I

The Runge-Kutta method can be generalized for any order ordinary differential
equation. The file RUNGE_N.INC contains an algorithm that can solve an initial
value problem for an nth-order differential equation with the fourth-order Runge-
Kutta formulas. The Lipshitz condition can be generalized for any order ordinary
differential equation. (For details, consult the reference book listed in the section,
"Solution to an Initial Value Problem for a First-Order Ordinary Differential Equa-
tion Using the Runge-Kutta Method.")
Although RUNGE-N.INC can be used to solve initial value problems for first-
order and second-order ordinary differential equations, we recommend that
RUNGE_l.INC and RUNGE--2.INC be used instead. The notation used by these
routines is somewhat simpler than the general case. There is no significant differ-
ence in computation time between the general program (RUNGE_N.INC) and the
specific programs (RUNGE_l.INC and RUNGE--2.INC).
Systems of coupled equations may also be solved with Runge-Kutta techniques.
A system of up to ten first-order ordinary differential equations can be solved with'
the file RUNGE_Sl.INC. A system of up to ten second-order ordinary differen-
tial equations can be solved with the file RUNGE_S2.INC. The algorithms in
both these files are based on the classical Runge-Kutta method with uniform spac-
ing between mesh points; hence, they do not allow for accuracy control (as
in the Runge-Kutta-Fehlberg method). (The Lipshitz condition for systems of
equations is given in the reference in the sections about RUNGE_Sl.INC and
RUNGE_S2.INC.)

Initial Value and Boundary Value Methods 157

Boundary value problems for second-order ordinary differential equations
(where the value of the dependent variable is specified at the two endpoints of
interval) can be solved using slwoting techniques. Shooting techniques converge
onto the slope of the function at one boundary. This reduces the boundary value
problem to a series of initial value problems. The series concludes when the initial
value problem satisfies the boundary condition at the other boundary.
If the second-order differential equation is linear (that is, linear in the dependent
variable(s), not necessarily linear in the independent variable), the linear-shooting
method (LINSHOT2.INC) may be used. A linear combination of solutions to two
initial value problems yields the solution to the boundary value problem.
If the second-order differential equation is nonlinear, the routine SHOOT2.1NC
must be used. The secant method generates a sequence of solutions with different
values of the first derivative until the appropriate boundary condition, subject to a
desired accuracy, is satisfied. Although SHOOT2.1NC may be used to solve linear
boundary value problems, LINSHOT2.1NC is more efficient for the linear case .
. Boundary value problems for second-order differential equations are guaranteed
to have a unique solution on the interval a, b if the function
y" = !(x, y, y')
and the two partial derivatives a!lay, apay' are continuous on the interval [a, b].
Furthermore, apay must be positive and apay' must be bounded for all x, y, y'
a :5 x :5 b, - 00 < y < 00, - 00 < y" < 00 •
The convergence to the appropriate initial value of the first derivative is not
assured for nonlinear boundary value problems. A good guess of the derivative
boundary condition is often required and may involve considerable trial and error.
Interpolation techniques (see Chapter 3) may be used to approximate the solu-
tion of values of the independent variable that are not mesh points.

158 Turbo Numerical Methods Toolbox

Solution to an Initial Value Problemfor a First-order
Ordinary Differential Equation Using the Runge-Kutta
Method (RUNGEJ..INC)

Description

This example uses the Runge-Kutta method (Burden and Faires 1985,220-227) to
approximate the solution to a first-order ordinary differential equation with a speci-
fied initial condition.
Given a function of the form
dx/dt = TNTargetF(t, x)
which satisfies the conditions given at the beginning of this chapter, and an initial
condition
x[LowerLimit] = Xlnitial
and spacing
h = (UpperLimit - LowerLimit)/Numlntervals
the fourth-order Runge-Kutta method approximates x in the interval [LowerLimit,
UpperLimit].
The fourth-order Runge-Kutta formulas consist of the following:
F1 h * TNTargetF(t, x[t])
=
F2 h * TNTargetF(t + h/2, x[t] + Fl/2)
=
F3 h * TNTargetF(t + h/2, x[t] + F2/2)
=
F4 h * TNTargetF(t + h, x[t] + F3)
=
x[t + 1] = x[t] + (F1 + 2 * F2 + 2 * F3 + F4)/6
where t ranges from LowerLimit to UpperLimit in steps of h. These formulas give a
4
truncation error of order h •
You must supply LowerLimit, UpperLimit, Xlnitial, Numlntervals, and TNTar-
getF.

User- Defined Types

TNvector = array[l .. TNArraySize] of Real;

Initial Value and Boundary Value Methods 159

User-Defined Function

TNTargetF{t, X : Real) : Real;

dx/dt = TNTargetF(t, x)
The function TNTargetF(t, x) is a user-defined function that calculates the deriv-
ative dx/dt.

Input Pararneters

LowerLimit : Real; Lower limit of interval

UpperL imi t : Real; Upper limit of interval
XInitial : Real; Value of X at LowerLimit
NumReturn : Integer; Number of (t, x) pairs returned from the procedure
NumInterval s : Integer; Number of subintervals used in calculations

The preceding parameters must satisfy the following conditions:

1. NumReturn > 0
2. Numlnteroals ~ NumReturn
3. LowerLimit ¢ UpperLimit

Output Pararneters

TVal ues : TNvector; Values of t between the limits

XVal ues : TNvector; Values of X approximated at the values in TValues
Error : Byte; 0: No errors
1: NumReturn < 1
2: Numlntervals < NumReturn
3: LowerLimit = UpperLimit

160 Turbo Numerical Methods Toolbox

Syntax of the Procedure Call

InitialCondlstOrder(LowerLimit, UpperLimit, Xlnitial, NumReturn,

Numlritervals, TValues, XValues, Error);
The procedure InitialConditionlstOrder integrates the first-order differential
equation.

Comments

This procedure will compute Numlntervals values in its calculations; however, you
will rarely need to use all the values. The vectors TValues and XValues will contain
only NumReturn values at roughly equally spaced t-values between the lower and
upper limits. (They will be equally spaced only when Numlntervals is a multiple of
NumReturn.) Thus, you can ensure a highly accurate solution (by making Numln-
tervals large) without generating an excessive amount of output (by making Num-
Return small).
The Runge-Kutta method uses the New/Dispose procedures to manipulate the
heap and should not be used in any program that uses Mark/Release to manipulate
the heap.
Warning: A stiff differential equation occurs when there are at least two very
different scales of the independent variable on which the dependent variable{s) is
changing; for example, y = x + e - lOOx. The Runge-Kutta method may generate a
numerical solution that bears no resemblance to the exact solution of the differen-
tial equation. This unstable numerical solution usually grows exponentially and
may be oscillatory. However, if the exact solution of the differential equation grows
as the independent variable increases, the instability may be difficult to detect. If a
suspected instability has been encountered, reduce the interval size (Numlnter-
vals).

Sample Program

The sample program RUNGE_l.PAS provides I/O functions that demonstrate the
Runge-Kutta method of solving initial value problems. Note that the file
RUNGE_l.INe is included after the function TNTargetF is defined.

Initial Value and Boundary Value Methods 161

Example

Problem. Solve the following initial value problem with the Runge-Kutta method:
X' = xlt + t - 1 I:5t:52
x(l) = 1
1. Code the equation into the program RUNGE_l.PAS:
funct;on TNTargetF(t, X : Real) : Real;
(************************************************************************)
(**** THIS IS THE FIRST-ORDER DIFFERENTIAL EQUATION ****)
(************************************************************************)

beg;n
TNTargetF := x/t + t - 1
end; { function TNTargetF }
2. Run RUNGE_I.PAS:
Lower limit of interval? 1

Upper limit of interval? 2

X value at t = 1.00000000E+00:

Number of values to return (1-500)? 10

Number of intervals (>= 10, default = 10)? 100

Direct output to one of the following:

(S)creen
(P)rinter
(F) il e

Lower limit: 1.00000000000000E+000

Upper limit: 2.00000000000000E+000
Value of X at 1.0000: 1.00000000000000E+000
Number of intervals: 100

t X
1.00000000 1.00000000000000E+000
1.10000000 1.10515880220649E+000
1.20000000 1.22121413182916E+000
1.30000000 1.34892645616477E+000
1.40000000 1.48893886869362E+000
1.50000000 1.64180233779216E+000
1.60000000 1.80799419315265E+000
1.70000000 1.98793197313186E+000
1.80000000 2. 18198400310574E+000
1.90000000 2.39047761619428E+000
2.00000000 2.61370563879444E+000
The exact solution is
x=l - t * In(t)
X(2) = 2.6137056

162 Turbo Numerical Methods Toolbox

Solution to an Initial Value Problemfor a First-Order
Ordinary Differential Equation
Using the Runge-Kutta-Fehlberg Method (RKFJ..INC)

Description

This example uses the Runge-Kutta-Fehlberg method (Burden and Faires 1985,
230-235) to approximate a solution within a specified tolerance to a first-order
ordinary differential equation with a specified initial condition.
Where the Runge-Kutta method (see RUNGE_l.INe) uses a constant spacing
h, the Runge-Kutta-Fehlberg method varies the spacing so that the solution can be
approximated with accuracy.
Given a function of the form
dx/dt = TNTargetF(t, x)
which satisfies the conditions given at the beginning of this chapter, and an initial
condition
x[LowerLimit] = Xlnitial
both the fourth-order and fifth-order Runge-Kutta formulas are used to approxi-
mate x in the interval [LowerLimit, UpperLimit]. The number of subintervals is
continually increased until the fractional difference between the results of the
5
fourth-order and fifth-order formulas (which give a truncation error of h4 and h ,
respectively) in each subinterval is less than the specified tolerance.
You must supply LowerLimit, UpperLimit, Tolerance, and TNTargetF.

User- Defined Types

TNvector = array[l .. TNArraySize] of Real;

User- Defined Function

TNTargetF(t, X : Real) : Real;

dx/dt = TNTargetF(t, x)

Initial Value and Boundary Value Methods 163

Input Parameters

LowerLimit : Real; Lower limit of interval

UpperLimit : Real; Upper limit of interval
XInitial : Real; Value of X at LowerLimit
To 1erance : Real; Maximum tolerable fractional difference between iterate
values
NumReturn : Integer; Number of (t, x) values to be returned

The preceding parameters must satisfy the following conditions:

1. Tolerance > 0
2. NumReturn > 0
3. LowerLimit ¢ UpperLimit

Output Parameters

TVal ues : TNvector; Values of t at which X was approximated

XVal ues : TNvector; Values of X at the values in TValues
Error : Byte; 0: No errors
1: Tolerance :5 0
2: NumReturn :5 0
3: LowerLimit = UpperLimit
4: Tolerance not reached

Syntax of the Procedure Call

RungeKuttaFehlberg{LowerLimit, UpperLimit, XInitial, Tolerance,

NumReturn, TValues, XValues, Error);

The procedure RungeKuttaFehlberg integrates the first-order differential equa-

tion TNTargetF.

164 Turbo Numerical Methods Toolbox

Comments

This procedure will compute more values in its calculations than it will return in
the vectors TValues and XValues. The vectors TValues and XValues will contain
only NumReturn values at subintervals between the lower and upper limits. More
values will be returned in regions of large functional variation than in regions· of
small functional variation. Thus, you can ensure a highly accurate solution (by
making the Tolerance small) without generating an excessive amount of output (by
making NumReturn small).
The Runge-Kutta-Fehlberg method improves the accuracy in the solution by
reducing the spacing between successive values of t. However, if the Tolerance is
too small, the spacing required to reach Tolerance may be beyond the machine's
limit of precision. Consequently, the routine will not converge to a solution that
meets the required Tolerance and Error 5 will be returned.
The Runge-Kutta-Fehlberg method uses the New/Dispose procedures to manip-
ulate the heap and should not be used in any program that uses Mark/Release to
manipulate the heap.
Warning: A stiff differential equation occurs when there are at least two very
different scales of the independent variable on which the dependent variable(s) is
changing; for example, y = x + e- 1OOx • The Runge-Kutta-Fehlberg method may
generate a numerical solution that bears no resemblance to the exact solution of the
differential equation. This unstable numerical solution usually grows exponentially
and may be oscillatory. However, if the exact solution of the differential equation
grows as the independent variable increases, the instability may be difficult to
detect. If a suspected instability has been encountered, reduce the interval size
(NumI ntervals).

Sample Program

The sample program RKF_l.PAS provides I/O functions that demonstrate the
Runge-Kutta-Fehlberg method of solving initial value problems. Note that the file
RKF_l.INC is included after the function TNTargetF is defined.

Initial Value and Boundary Value Methods 165

Example
Problem. Use the Runge-Kutta-Fehlberg method to solve the following initial
value problem with a tolerance of lE-6:
X' = x/t +t - 1 1 $; t $; 2
x(l} = 1
1. Code the differential equation into the program RKF_l.PAS:
function TNTargetF(t, X : Real) : Real;
(************************************************************************)
(**** THIS IS THE FIRST-ORDER DIFFERENTIAL EQUATION ****)
(************************************************************************)

begin
TNTargetF := x/t + t - 1;
end; { function TNTargetF }
2. Run RKF_l.PAS:
Lower limit of interval?

Upper limit of interval? 2

X value at t = 1.00000000E+00:

Number of values to return (1-500)? 10

Tolerance (> 0, default = 1.000E-06)? 1E-6

Direct output to one of the following:

(S)creen
(P)rinter
(F) il e

Lower limit: 1.00000000000000E+000

Upper limit: 2.00000000000000E+000
Value of X at 1.0000: 1.00000000000000E+000
Tolerance: 1.00000000000000E-006

t X
1.00000000 1.00000000000000E+000
1.10000000 1.10515881708653E+000
1.20000000 1.22121416069278E+000
1.30000000 1.34892649817459E+000
1.40000000 1.48893892310351E+000
1.50000000 1.64180240395245E+000
1.60000000 1.80799427050390E+000
1.70000000 1.98793206119471E+000
1.80000000 2.18198410146987E+000
1.90000000 2.39047772450816E+000
2.00000000 2.61370575675625E+000

166 Turbo Numerical Methods Toolbox

Now solve the same problem with a smaller tolerance, l.OOOE-OB:
Lower limit of interval? 1

Upper limit of interval? 2

X value at t = 1.00000000E+00:

Number of values to return (1-500)? 10

Tolerance (> 0, default = 1.000E-06)? 1E-8

Direct output to one of the following:

{S)creen
(P)rinter
(F) il e

Lower limit: 1.00000000000000E+000

Upper limit: 2.00000000000000E+000
Value of X at 1.0000: 1.00000000000000E+000
Tolerance: 1.00000000000000E-008

t X
1.00000000 1.00000000000000E+000
1.12208941 1. 12982837391732E+000
1.20585321 1.22836146826667E+000
1.29271260 1.33921121906568E+000
1.38286653 1.46405185209736E+000
1.47648998 1.60468229863568E+000
1.57374241 1.76304147973215E+000
1.67477301 1.94122165006705E+000
1.77972398 2. 14148082447423E+000
1.88873280 2.36625482837546E+000
2.00193373 2.61816928222327E+000
The exact solution is
x= t 2 - t In(t)
X(2) = 2.6137056
X(2.00193373) = 2.61B1693
In the first run, a solution could be approximated within tolerance with a spacing
of 0.1. In the second run, the algorithm had to vary the spacing in order to approxi-
mate a solution within the tolerance.

Initial Value and Boundary Value Methods 167

Solution to an Initial Value Problemfor a First-Order
Ordinary Differential Equation
Using the Adams-Bashforth/Adams-Moulton
Predictor/Corrector Scheme (ADAMSJ..INC)

Description

This example approximates the solution to a first-order ordinary differential equa-

tion with a specified initial condition using the four-step Adams-Bashforth/Adams-
Moulton formulas (Burden and Faires 1985, 238-247). Runge-Kutta methods are
one-step methods, because each calculation uses information from only one pre-
vious point. The Adams' formulas use information from four previous points, thus
the four-step method.
Given a function of the form
dx/dt = TNTargetF(t, x)
which satisfies the conditions given at the beginning of this chapter, and an initial
condition
x[LowerLimit] = Xlnitial
and spacing
h = (UpperLimit - LowerLimit)/Numlntervals
the fourth-order Runge-Kutta formula (see RUNGE_l.INC) is used to find approx-
imations at the first three points in the interval [LowerLimit, UpperLimit]. Then
the following explicit Adams-Bashforth formula:
xJi + 1] = x[i] + h/24 * { 55 * TNTargetF(t[i], x[i])
- 59 * TNTargetF(t[i - 1], x[i -1])
+ 37 * TNTargetF(t[i - 2], x[i - 2])
- 9 * TNTargetF(t[i - 3], x[i - 3]) }
and the following implicit Adams-Moulton formula:
x[i + 1] = x[i] + h/24 * { 9 * TNTargetF(t[i + 1], xJi + 1])
+ 19 * TNTargetF(t[i], x[i])
- 5 * TNTargetF(t[i - 1], x[i -1])
+ TNTargetF(t[i - 2], x[i - 2]) }
approximate (predict) and refine (correct) all other points in the interval.

168 Turbo Numerical Methods Toolbox

You must supply UpperLimit, LowerLimit, Xlnitial, Numlntervals, and TNTar-
getF.

User-Defined Types

TNvector = array[l •• TNArraySize] of Real;

User- Defined Function

TNTargetF(t, X : Real) : Real;

dxldt = TNTargetF(t, x)

Input Parameters

LowerL imi t : Real; Lower limit of interval

UpperLimit : Real; Upper limit of interval
XInitial : Real; Value of X at LowerLimit
NumReturn : Integer; Number of (t, x) values to be returned from the procedure
NumInterva 1s : Integer; Number of subintervals to be used in calculations

The preceding parameters must satisfy the following conditions:

1. NumReturn > 0
2. Numlntervals ~ NumReturn
3. LowerLimit ~ UpperLimit

Output Parameters

TVal ues : TNvector; Values of t between the limits

XVal ues : TNvector; Values of X determined at the values in TValues
Error: Byte; 0: No errors
1: NumReturn < 1
2: Numlntervals < NumReturn
3: LowerLimit = UpperLimit

Initial Value and Boundary Value Methods 169

Syntax of the Procedure Call

Adams(LowerLimit, UpperLimit, XInitial, NumReturn,

NumIntervals,TValues, XValues, Error);
The procedure Adams integrates the first-order differential equation TNTargetF.

Comments

This procedure will compute Numlntervals values in its calculations; however, you
will rarely need to use the values. The vectors TValues and XValues will contain
only NumReturn values at roughly equally spaced t-values between the lower and
upper limits. (They will be equally spaced only when Numlntervals is a multiple of
NumReturn.) Thus, you can ensure a highly accurate solution (by making Numln-
tervals large) without generating an excessive amount of output (by making Num-
Return small).
The Adams-Bashforth/Adams-Moulton method uses the New/Dispose proce-
dures to manipulate the heap and should not be used in any program that uses
Mark/Release to manipulate the heap.
Warning: A stiff differential equation occurs when there are at least two very
different scales of the independent variable on which the dependent variable(s) is
changing; for example, y = x + e- • The Adams-Bashforth/Adams-Moulton
1OOx

method may generate a numerical solution that bears no resemblance to the exact
solution of the differential equation. This unstable numerical solution usually
grows exponentially and may be oscillatory. However, if the exact solution of the
differential equation grows as the independent variable increases, the instability
may be difficult to detect. If a suspected instability has been encountered, reduce
the interval size (Numlntervals).

Sample Program

The sample program ADAMS_l.PAS provides I/O functions that demonstrate the
Adams-Bashforth/Adams-Moulton predictor/corrector method of solving initial
value problems. Note that the file ADAMS_l.lNC is included after the function
TNTargetF is defined.

170 Turbo Numerical Methods Toolbox

Example
Problem. Solve the following initial value problem with the Adams-Bashforth/
Adams-Moulton method:
X' = x/t + t - 1 l:5t:52
x(l) = 1
1. Code the differential equation into the program ADAMS_l.PAS:
function TNTargetF{t, X : Real} : Real;
{**************************************************************************}
{**** THIS IS THE FIRST-ORDER DIFFERENTIAL EQUATION ****}
{**************************************************************************}

begin
TNTargetF := x/t + t - 1;
end; { function TNTargetF }
2. Run ADAMS_l.PAS:
Lower limit of interval? 1

Upper limit of interval? 2

X value at t = 1.00000000E+00:

Number of values to return {1-500}? 10

Number of intervals {>= 10, default = 10}? 100

Direct output to one of the following:

{S}creen
(P}rinter
(F) ile

Lower limit: 1.00000000000000E+000

Upper limit: 2.00000000000000E+000
Value of X at 1~0000: 1.00000000000000E+000
Number of intervals: 100

t X
1.00000000 1.00000000000000E+000
1.10000000 1.10515880229293E+000
1.20000000 1.22121413201736E+000
1.30000000 1.34892645643801E+000
1.40000000 1.48893886904034E+000
1.50000000 1.64180233820416E+000
1.60000000 1.80799419362396E+000
1.70000000 1.98793197365806E+000
1.80000000 2. 18198400368348E+000
1.90000000 2.39047761682098E+000
2.00000000 2.61370563946811E+000
The exact solution is
X = t t In(t)
2
-

x(2) = 2.6137056
Initial Value and Boundary Value Methods 171
Solution to an Initial Value Problem/or a Second-Order
Ordinary Differential Equation
Using the Runge-Kutta Metlwd (RUNGE....2.INC)

Description

This example approximates the solution to a second-order ordinary differential

equation with specified initial conditions using the two variable Runge-Kutta for-
mulas (Burden and Faires 1985, 261-269).
Given a function of the form
d2x/dt2 = TNTargetF(t, x, x')
where x' indicates dx/dt (which satisfies the Lipshitz condition given at the begin-
ning of this chapter), the initial conditions
x[LowerLimit] = InitialValue
x'[LowerLimit] = InitialDeriv
and spacing
h = (UpperLimit - LowerLimit)/Numlntervals
rewrite the second-order differential equation as two, first-order differential equa-
tions:
x' = y
y' = TNTargetF(t, x, y)
Then the fourth-order, two-variable Runge-Kutta method can be used to approx-
imate simultaneously x and y (x and x').
The fourth-order Runge-Kutta formulas for these equations consist of the follow-
ing:
FIx = h * y[t]
Fly = h * TNTargetF(t, x[t], y[t])
F2x = h * (y[t] + F1y/2)
F2y = h * TNTargetF(t + h/2, x[t] + F1x/2, y[t] + F1y/2)
F3x = h * (y[t] + F2y/2)
F3y = h * TNTargetF(t + h/2, x[t] + F2x/2, y[t] + F2y/2)
F4x = h * (y[t] + F3y)
F4y = h * TNTargetF(t + h, x[t] + F3x, y[t] + F3y)

172 Turbo Numerical Methods Toolbox

x[t+ 1] = x[t] + (FIx + 2 * F2x + 2 * F3x + F4x)/6
y[t + 1] = y[t] + (Fly + 2 * F2y + 2 * F3y + F4y)/6
where t ranges from LowerLimit to UpperLimit in steps of h. These formulas give a
truncation error of order h4 •
You must supply LowerLimit, UpperLimit, Xlnitial, Numlntervals, and TNTar-
getF.

User- Defined Types

TNvector = array[l .. TNArraySize] of Real;

User-Defined Function

TNTargetF(t, x, XPrime : Real) : Real;

2
dx /dt 2
= TNTargetF(t, x, dx/dt)

Input Parameters

LowerLimit : Real; Lower limit of interval

UpperLimit : Real; Upper limit of interval
InitialValue: Real; Value of X at LowerLimit
InitialDeriv : Real; Derivative of X at LowerLimit
NumReturn : Integer; Number of (t, x) values returned from the procedure
NumInterval s : Integer; Number of subintervals used in the calculations

The preceding parameters must satisfy the following conditions:

1. NumReturn > 0
2. Numlntervals;::= NumReturn
3. LowerLimit ~ UpperLimit

Initial Value and Boundary Value Methods 173

Output Parameters

TValues : TNvector; Values of t between the limits

XV a1ues : TNvector; Values of X determined at the values in TValues
XDeri vVa 1ues : TN vector ; Values of the first derivative of X determined at the values
in TValues
Error: Byte; 0: No errors
1: NumReturn < 1
2: Numlntervals < NumReturn
3: LowerLimit = UpperLimit

Syntax of the Procedure Call

InitialCond2ndOrder(LowerLimit, UpperLimit, InitialValue, InitialDeriv,

NumReturn, NumIntervals, TValues, XValues,
XDerivValues, Error);
The procedure InitialCond2ndOrder integrates the second-order differential
equation TNTargetF.

Comments

This procedure will compute Numlntervals values in its calculations; however, you
will rarely need to use all these values. The vectors TValues, XValues, and XDeriv-
Values will contain only NumReturn values at roughly equally spaced t-values
between the lower and upper limits. (They will be equally spaced only when
Numlntervals is a multiple of NumReturn.) Thus, you can ensure a highly accurate
solution (by making Numlntervals large) without generating an excessive amount
of output (by making NumReturn small).
Warning: A differential equation occurs when there are at least two very differ-
ent scales of the independent variable on which the dependent variable(s) is chang-
ing; for example, y = x + e -IOOx. The Runge-Kutta method may generate a
numerical solution that bears no resemblance to the exact solution of the differen-
tial equation. This unstable numerical solution usually grows exponentially and
may be oscillatory. However, if the exact solution of the differential equation grows
as the independent variable increases, the instability may be difficult to detect. If a
suspected instability has been encountered, reduce the interval size (Numlnter-
vals).

174 Turbo Numerical Methods Toolbox

Sample Program

The sample program RUNGE-2.PAS provides I/O functions that demonstrate the
Runge-Kutta method of solving initial value problems for second-order ordinary
differential equations. Note that the file RUNGE-2.INC is included after the func-
tion TNTargetF is defined.

Example
Problem. A weight with mass m lies on a frictionless table and is connected to a
spring with spring constant k:

'4-Wall

F( w)
k
m

. . It ~
FnctlOn ess surlace

If the weight is subject to a driving force F sin(oo t) (00 represents the frequency
of the driving force and t is time), the equation of motion of the mass is as follows:
m d2x/dt2 + k x = F sin(oo t)
Given
m = 2kg
F=9N
k = 32 N/m
00 = 5 cycles/sec
x(O) = 0 m
dx(O)/dt = -2.5 m/sec
find the position and velocity of the block from t = 0 second to t = 2 seconds.
1. Rewrite the preceding second-order differential equation:
d 2x/dt2 = F/m sin(oo t) - kim x

Initial Value and Boundary Value Methods 175

2. Code this second-order differential equation into the program RUNGE--2.PAS:
function TNTargetF(t : Real;
X : Real;
XPrime : Real) : Real;
(*************************************************************************)
(**** THIS IS THE SECOND-ORDER DIFFERENTIAL EQUATION ****)
(*************************************************************************)

begin
TNTargetF := 9/2 * Sin (5 * t) - 32/2 * x;
end; { function TNTargetF }
3. Run RUNGE--2.PAS:
Lower limit of interval? 0
Upper limit of interval? 2

Enter X value at t = O.OOOOOOOOE+OO: 0

Enter derivative of X at t = O.OOOOOOOOE+OO: -2.5

Number of values to return (1-500)? 10

Number of intervals (>= 10, default = 10)? 100

Direct output to one of the following:

(S)creen
(P)rinter
(F) il e

Lower 1imit: O.OOOOOOOOOOOOOOE+OOO

Upper 1i mit: 2.00000000000000E+000
Value of X at 0.0000: O.OOOOOOOOOOOOOOE+OOO
Value of X' at 0.0000: -2.50000000000000E+000
Number of intervals: 100

t Value of X Derivative of X
0.00000000 O.OOOOOOOOOOOOOOE+OOO -2.50000000000000E+000
0.20000000 -4.20735284275848E-001 -1.35075642830665E+000
0.40000000 -4.54648724216734E-001 1.04036531118478E+000
0.60000000 -7.05605786993375E-002 2.47497991717220E+000
0.80000000 3.78400378699554E-001 1.63411037473655E+000
1.00000000 4.79461767300631E-001 -7.09151289407567E-001
1.20000000 1.39708469016311E-001 -2.40042152228323E+000
1.40000000 -3.28491796183335E-001 -1.88475529635974E+000
1.60000000 -4.94677974769030E-001 3.63745224811839E-001
1.80000000 -2.06059519715175E-001 2.27781864414105E+000
2.00000000 2. 72008842396951E-001 2.09767516082021E+000

176 Turbo Numerical Methods Toolbox

The exact solution is
F sin{oo t)
x=-------

F., cos{oo t)
dx/dt = - - - - - - -
2 2
m (oo - (0 ) 0

where 00
0
is the natural frequency of the system
2
000 = kim
The period of oscillation is given by
t = 2 'Tr/oo = 1.257 sec
The data is taken from a function of which the derivative could be computed
exactly. Following are the actual values:
t Values of X Derivative of X
0.0 O.OOOOOOOOOOOOE + 000 - 2.500000000000E + 000
0.2 - 4.207354924039E - 001 ...:. 1.350755764670E + 000
0.4 - 4.546487134128E - 001 1.040367091367E + 000
0.6 -7.056000402993E - 002 2.474981241501E + 000
0.8 3.784012476539E - 001 1.634109052159E + 000
1.0 4.794621373315E - 001 -7.091554636580E - 001
1.2 1.397077490994E - 001 - 2.400425716625E + 000
1.4 - 3.284932993593E - 001 - 1.884755635858E + 000
1.6 -4.946791233116E - 001 3.637500845215E - 001
1.8 - 2.060592426208E - 001 2.277825654711E + 000
2.0 2.720105554446E - 001 2.097678822691E + 000

Initial Value and Boundary Value Methods 177

Solution to an Initial Value Problemfor an nth-Order
Ordinary Differential Equation Using the Runge-Kutta
Metlwd (RUNGE...N.INC)

Description

This example integrates an nth-order ordinary differential equation with specified

initial conditions using the generalized Runge-Kutta formulas· (Burden and Faires
1985,261-269).
Given a function of the form
dnx/dt" = TNTargetF(t, x, x(l), ... , x(n-l»)
where xlJ> indicates dJx/dt J, which satisfies the general Lipshitz condition (the Lip-
shitz condition for first-order and second-order ordinary differential equations is
given at the beginning of this chapter, and initial condition

x[LowerLimit] = a 1
x(l)[LowerLimit] = a2

x(n-l)[LowerLimit] = an
and spacing
h = (UpperLimit - LowerLimit)/Numlnteroals
rewrite the nth-order differential equation as n first-order differential equations:
X(l) = Y
I

X(2) = y(I)I = Y2

X(3) = y(I)2 = Y3

x(n-I) = Y
(1)
n-2 = Yn - I

X(n) = y(I)n_I = TNTargetF(t, x, YI' Y

2
, ... , Yn-I)
Then the fourth-order general Runge-Kutta method can be used to approximate
simultaneously the y's (x and its derivatives).

178 Turbo Numerical Methods Toolbox

The general Runge-Kutta formulas for these equations consist of the following:

FIx = h * yJt]
FIY l = h * Yz[t]

FIy z
lI
_ = h * y _Jt] lI

FIYn _ l = h * TNTargetF(t, x[t], yJt], ... , yn_Jt])

F2x = h * (yJt] + Fly/2)
F2Y l = h * (Yz[t] + Fly/2)

F2y z = h * (yn_Jt] + FIy n_/2)

lI
_

F2Y n_l = h * TNTargetF(t + h/2, x[t] + Flx/2, yl[t] + Fly/2, ... , yn_Jt]
+ FIy n_/2)
F3x = h * (yJt] + F2y/2)
F3Yl = h * (Yz[t] + F2y/2)

F3YIl_2 = h * (y _Jt] + F2y n_/2)

lI

F3Y l
II
_ = h * TNTargetF(t + h/2, x[t] + F2x/2, yJt] + F2y/2, ... , yn_Jt]
+ F2Yn_/2)
F4x = h * (yJt] + F3y)
F4Yl = h * (Yz[t] + F3yz)

F4YIl_2 = h * (ylI_Jt] + F3YII_)

F4Y l
II
_ = h * TNTargetF(t + h, x[t] + F3x, yJt] + F3y l, ... , yn_Jt]
+ F3YIl_)
x[t + 1] = x[t] + (FIx + 2 * F2x + 2 * F3x + F4x)/6
yJt+ 1] = yJt] + (FlYl + 2 * F2Yl + 2 * F3Yl + F4y)/6
Y2[t+ 1] = Yz[t] + (FIY 2 + 2 * F2y z + 2 * F3y z + F4Y z)/6

YIl_2[t + 1] = y 2[t] + (FIY 2 + 2

lI
_
II
_ * F2y n_z + 2 * F3Y n_2 + F4YIl_2)/6
ylI)t+ 1] = ylI_Jt] + (FIY"_l + 2 * F2Y II _l + 2 * F3Y l +
II
_ F4y n_l)/6

where t ranges from LowerLimit to UpperLimit in steps of h. These formulas give a

truncation error of order h-t.

Initial Value and Boundary Value Methods 179

You must supply the order, limits, initial values, and TNTargetF. The order may
be arbitrarily large.

User-Defined Types

TNvector = array[O •• TNRowSize] of Real;

TNmatrix = array[O •• TNColumnSize] of TNvector;

TNRowSize is an upper bound for the number of values returned for a particular
variable (NumReturn). TNColumnSize is an upper bound for the order of the differ-
ential equation (Order).

User-Defined Function

TNTargetF(V : TNvector) : Real;

The elements of V are defined as

V[O] corresporids to t
V[l] corresponds to x
V[2] corresponds to first derivative of x
V[3] corresponds to second derivative of x

This is the differential equation:

dnx/dt = TNTargetF(t, x, X(l), ... x(n-l») where n is the order of the equation.

The procedure Initial Condition integrates this nth-order differential equation.

Input Parameters

Order: Integer; Order of the differential equation

LowerLimit : Real; Lower limit of interval
UpperLimit : Real; Upper limit of interval
InitialValues : TNvector; Values of X and its derivatives at LowerLimit
NumReturn : Integer; Number of (t, x, X(l), ... , X (n») values returned from the pro-
cedure
NumInterval s : Integer; Number of subintervals used in the calculations

180 Turbo Numerical Methods Toolbox

The preceding parameters must satisfy the following conditions:
l. NumReturn > 0
2. Numlntervals 2: NumReturn
3. Order> 0
4. LowerLimit ~ UpperLimit

Output Parameters

SolutionValues : TNmatrix: Values of t, x and the derivatives of x between the limits

Error: Byte: 0: No errors
1: NumReturn < 1
2: Numlntervals < NumReturn
3: Order < 1
4: LowerLimit = UpperLimit

Syntax of the Procedure Call

InitialCondition(Order, LowerLimit, UpperLimit, InitialValues,

NumReturn, Numlntervals, SolutionValues, Error):

Comments

The first row of SolutionValues will be the values of t between the limits, the
second row of SolutionValues will be the values of x between the limits, the third
row of SolutionValues will be the values of X(l) between the limits, and so on.
This procedure will compute Numlntervals values in its calculations; however,
you will rarely need to use all those values. The rows of SolutionValues will contain
only NumReturn values at roughly equally spaced t-values between the lower and
upper limits. (They will be equally spaced only when Numlntervals is a multiple of
NumReturn.) Thus, you can ensure a highly accurate solution (by making Numln-
tervals large) without generating an excessive amount of output (by making Num-
Return small).
There are no bounds on the order of the differential equation.

Initial Value and Boundary Value Methods 181

This routine stores much information on the heap. If you try to solve a high-
order differential equation very precisely (that is, both Order and Numlntervals are
large), you may get run-time error $FF, Heap/Stack collision. If this happens, the
dimension of TNvector and TNmatrix should be reduced as much as possible. If
this is not possible, remove any RAM-resident software (for example, SideKick,
SuperKey, or a print buffer).
The Runge-Kutta method uses the New/Dispose procedures to manipulate the
heap and should not be used in any program that uses Mark/Release to manipulate
the heap.
Warning: A stiff differential equation occurs when there are at least two very
different scales of the independent variable on which the dependent variable(s) is
changing; for example, y = x + e -100%. The Runge-Kutta method may generate a
numerical solution that bears no resemblance to the exact solution of the differen-
tial equation. This unstable numerical solution usually grows exponentially and
may be oscillatory. However, if the exact solution of the differential equation grows
as the independent variable increases, the instability may be difficult to detect. If a
suspected instability has been encountered, reduce the interval size (Numlnter-
vals).

Sample Program

The sample program RUNGE_N.PAS provides I/O functions that demonstrate the
Runge-Kutta method of solving initial value problems for high-order ordinary dif-
ferential equations. Note that the file RUNGE-N.lNC is included after the func-
tion TNTargetF is defined.

Example
Problem. Find the solution to the following fourth-order ordinary differential equa-
tion from t = 0 to t = 1:
4
d x(t)/dt = - 4 x(t) d x(t)/dt
3 3

x(O) = 1
dx(O)/dt = - 1
d x(O)/dt = 2
2 2

3 3
d x(O)/dt = - 6

182 Turbo Numerical Methods Toolbox

1. Code the equation into the program RUNGE_N.PAS:
function TNTargetF(V : TNvector) : Real;
(***********************************************)
(* THIS IS THE DIFFERENTIAL EQUATION *)
(***********************************************)
(* *)
(* dn x (1) (n-1) *)
(* = TNTargetF(t, x, x , ..• x *)
(* *)
n
(* dt *)
(* *)
(* where n is the order of the equation. *)
(* *)
(* The elements of V are defined: *)
(* V[O] corresponds to t *)
(* V[l] corresponds to X *)
(* V[2] corresponds to 1st derivative of X *)
(* V[3] corresponds to 2nd derivative of X *)
(* *)
(* *)
(* *)
(***********************************************)

begin
TNTargetF := -4 * V[l] * V[4];
end; { function TNTargetF }
2. Run RUNGE-N.PAS:
Order of the equation (1-10)? 4
Lower limit of interval? 0
Upper limit of interval?
Enter X value at t O.OOOOOOOE+OOO: 1
Derivative 1 of X at t O.OOOOOOOE+OOO: -1
Derivative 2 of X at t O.OOOOOOOE+OOO: 2
Derivative 3 of X at t O.OOOOOOOE+OOO: -6
Number of values to return (1-100)? 10
Number of intervals (>= 10, default = 10)? 100
Direct output to one of the following:
(S)creen
(P)rinter
(F) il e

Lower limit: O.OOOOOOOOOOOOOOE+OOO

Upper limit: 1.00000000000000E+000
Number of intervals: 100

Initial Value and Boundary Value Methods 183

Initial conditions at lower limit:
X[l]= 1.00000000000000E+000
X[2]= -1.00000000000000E+000
X[3]= 2.00000000000000E+000
X[4]= -6.00000000000000E+000
t Value X[l]
0.00000000 1.00000000000000E+000
0.10000000 9.09090909737517E-001
0.20000000 8.33333334189337E-001
0.30000000 7.69230770157394E-001
0.40000000 7.14285715280102E-001
0.50000000 6.66666667788519E-001
0.60000000 6.25000001337168E-001
0.70000000 5.88235295769619E-001
0.80000000 5.55555557625526E-001
0.90000000 5.26315792064849E-001
1.00000000 5.00000003213983E-001

t Val ue X[2]
0.00000000 -1.00000000000000E+000
0.10000000 -8.26446283273189E-001
0.20000000 -6.94444446826215E-001
0.30000000 -5.91715977923112E-001
0.40000000 -5.10204082090465E-001
0.50000000 -4.44444443661452E-001
0.60000000 -3.90624997971428E-001
0.70000000 -3.46020758007957E-001
0.80000000 -3.08641970911504E-001
0.90000000 -2.77008304743045E-001
1.00000000 -2.49999993429933E-001
t Value X[3]
0.00000000 2.00000000000000E+000
0.10000000 1.50262961438149E+000
0.20000000 1. 15740742373768E+000
0.30000000 9.10332288053840E-001
0.40000000 7.28862989793594E-001
0.50000000 5.92592607536865E-001
0.60000000 4.88281263842229E-001
0.70000000 4.07083261374878E-001
0.80000000 3.42935540127152E-001
0.90000000 2.91587706310718E-001
1.00000000 2.50000010753535E-001

t Value X[4]
0.00000000 -6.00000000000000E+000
0.10000000 -4.09808076056272E+000
0.20000000 -2.89351855059016E+000
0.30000000 -2.10076680857258E+000
0.40000000 -1.56184925333600E+000
0.50000000 -1. 18518520443061E+000
0.60000000 -9.15527359078898E-001
0.70000000 -7. 18382215400418E-001
0.80000000 -5.71559223064178E-001
0.90000000 -4.60401631119694E-001
1.00000000 -3.75000005740567E-001

184 Turbo Numerical Methods Toolbox

X[I] are the values of x(t}.
X[2] are the values of dx(t}jdt.
2
X[3] are the values of d 2x(t}jdt •
3
X[4] are the values of d\(t}jdt •
The exact solution is
x(t} = (t+ Ifl
dx(t}jdt = - (t + I}-2
d 2x(t}jdt2 = 2(t + I}-3
d3x(t}jdt3 = - 6(t + If4

x(I) = 0.5
dx(I}jdt = - 0.25
d2x(I}jdt2 = 0.25
d3x(I}jdt3 = - 0.375

Initial Value and Boundary Value Methods 185

Solution to an Initial Value Problem/or a System 0/
Coupled First-Order Ordinary Differential Equations Using
the Runge-Kutta Metlwd (RUNGE-S1.INC)

Description

This example integrates a system of coupled first-order ordinary differential equa-

tions with specified initial conditions using the generalized Runge-Kutta formulas
(Burden and Faires 1985, 261-269).
Given m first-order ordinary differential equations in the form
dx/dt = TNTargetF1(t, Xl' X 2
' ... , xJ
dx/dt = TNTargetF2(t, Xl' X 2
' ... , xJ

dx m Idt = TNTargetFm(t, Xl' X 2

' ... , Xm )

which satisfies the Lipshitz condition (the Lipshitz condition for first-order and
second-order ordinary differential equations is given at the beginning of this chap-
ter; consult the previous book reference for details of the Lipshitz condition for
systems), and initial conditions
xJLowerLimit] = a l
xJLowerLimit] = a 2

x.JLowerLimit] = am
and spacing
h = (UpperLimit - LowerLimit)/Numlnteroals
the fourth-order general Runge-Kutta method can be used to approximate simulta-
neously the xj's.

186 Turbo Numerical Methods Toolbox

The general Runge-Kutta formulas for these equations are as follows:

FIxl = h * TNTargetFI(t, xl[t], x2[t], ... , x.Jt])

FIx2 = h * TNTargetF2(t, xJt], x2[t], ... , x.Jt])

FIx", = h * TNTargetFm(t, xJt], x2[t], "', x.Jt])

F2xl = h * TNTargetFI(t + h/2, xl[t] + FIx/2, x2[t] + FIx/2, ... , x.Jt]
+ FIx,/2)
F2x2 = h * TNTargetF2(t + h/2, xJt] + FIx/2, x2[t] + FIx/2, ... , x.Jt]
+ FIx,/2)

F2x", = h * TNTargetFm(t + h/2, xJt] + FIx/2, x2[t] + FIx/2, ... , x.Jt]

+ FIx,/2)
F3x l = h * TNTargetFI(t + h/2, xJt] + F2x/2, x2[t] + F2x/2, ... , x.Jt]
+ F2x,/2)
F3x2 = h * TNTargetF2(t + h/2, xJt] + F2x/2, x2[t] + F2x/2, ... , x.Jt]
+ F2x,/2)

F3x", = h * TNTargetFm(t + h/2, xl[t] + F2x/2, x2[t] + F2x/2, ... , x.Jt]

+ F2x rn /2)
F4xl = h * TNTargetFI(t + h, xJt] + F3x l, x2[t] + F3x 2, ... , x.Jt] + F3x,J
F4x2 = h * TNTargetF2(t + h, xJt] + F3x l, x2 [t] + F3x 2 , ... , x.Jt] + F3x",)

F4x", = h * TNTargetFm(t + h, xl[t] + F3x l, x2 [t] + F3x2, "', x.Jt] + F3x,,)

xJt + 1] = xl[t] + (FIx l + 2*F2x l + 2*F3x l + F4x)/6
xJt + 1] = x2 [t] + (FIx 2 + 2*F2x2 + 2*F3x2 + F4x 2 )/6

xm [t+ 1] = xm [t] + (FIx m + 2*F2x m + 2*F3xm + F4x m )/6

Initial Value and Boundary Value Methods 187

where t ranges from LowerLimit to UpperLimit in steps of h. These formulas give a
truncation error of order h4 •
You must supply the number of differential equations, the limits, initial values,
and TNTargetF's.
This procedure can solve a system of up to ten differential equations (see "Com-
ments" for information about how to increase this limit).

User-Defined Types

TNvector .; array[O .. TNRowSize] of·Real;

TNmatrix = array[O •• TNColumnSize] of TNvector;
TNRowSize is an upper bound for the number of values returned for a particular
variable (NumReturn). TNColumnSize is an upper bound for the number of differ-
ential equations (NumEquations).

User-Defined Functions

function TNTargetFl(V TNvector) Real;

function TNTargetF2(V TNvector) Real;
function TNTargetF3(V TNvector) Reali
function TNTargetF4(V TNvector) Real;
function TNTargetF5(V TNvector) Real;
function TNTargetF6(V TNvector) Real;
function TNTargetF7(V TNvector) Real;
function TNTargetF8(V TNvector) Real;
function TNTargetF9(V TNvector) Real;
function TNTargetF10(V : TNvector) : Real;
These are the differential equations:
dx/dt = TNTargetFj(t, Xl' X 2 ' ... , x IO )
where j ranges from 1 to 10.

188 Turbo Numerical Methods Toolbox

The elements of the vector V are defined as follows:
V[O] = t
V[l] = Xl
V[2] = X2

V[10] = XlO

The procedure defined in RUNGE_Sl.INC solves this system of coupled differ-

ential equations (a maximum of ten equations). All ten functions must be defined,
even if your system contains less than ten equations.

Input Parameters

NumEquations : Integer; Number of first-order differential equations

LowerL imi t : Real; Lower limit of interval
UpperLimit : Real; Upper limit of interval
InitialValues: TNvector; Values of Xl' X2 ' ... , Xm at LowerLimit
NumReturn : Integer; Number of (t, Xl' X2 ' ... , xJ values returned from the
procedure
NumInterval s : Integer; Number of subintervals used in the calculations
The preceding parameters must satisfy the following conditions:
l. NumReturn > 0
2. Numlntervals ~ NumReturn
3. NumEquations > 0
4. LowerLimit ~ UpperLimit

Output Parameters

SolutionValues : TNmatrix; Values oft, Xl' X2 ' ... xm between the limits
Error: Byte; 0: No errors
1: NumReturn < 1
2: Numlnteroals < NumReturn
3: NumEquations < 1
4: LowerLimit = UpperLimit

Initial Value and Boundary Value Methods 189

Syntax of the Procedure Call

InitialConditionSystem(NumEquations, LowerLimit, UpperLimit,

InitialValues, NumReturn, Numlntervals,
SolutionValues, Error):

Comments

The first row of SolutionValues will be the values of t between the limits, the
second row of SolutionValues will be the values of Xl between the limits, the third
row of SolutionValues will be the values of x2 between the limits, and so on.
All ten user-defined functions are called from the procedure. If your system has
less than ten equations, you must still define all ten functions or the program will
not compile. The superfluous functions should be defined as follows (TNTargetFIO
is used as an example):
funct;on TNTargetF10(V : TNvector} : Real;

beg;n
end: { function TNTargetF10 }
If you need to solve a system with more than ten equations, then edit the include
file RUNGE_Sl.lNC. The following line should be added to the end of procedure
Step:
F[ll] := Spacing * TNTargetFll(CurrentValues};
More statements (for F[12], and so on) may be added as necessary. All new
functions (for example, TNTargetFll) must be defined in your top-level program.
Note: Before making any changes to the include file, make sure you have a backup
copy.
This procedure will compute Numlntervals values in its calculations; however,
you will rarely need to use these values. The rows of SolutionValues will contain
only NumReturn values at roughly equally spaced t-values between the lower and
upper limits. (They will be equally spaced only when Numlntervals is a multiple of
NumReturn.) Thus, you can ensure a highly accurate solution (by making Numln-
tervals large) without generating an excessive amount of output (by making Num-
Return small).
This routine stores much information on the heap. If you try to accurately solve a
large system (that is, if both NumEquations and Numlntervals are large), you may
get run-time error $FF, Heap/Stack collision. If this happens, the dimension of
TNvector and TNmatrix should be reduced as much as possible. If this is not possi-
ble, then remove any RAM-resident software (for example, SideKick, SuperKey, or
a print buffer).

190 Turbo Numerical Methods Toolbox

The Runge-Kutta method uses the New/Dispose procedures to manipulate the
heap and should not be used in any program that uses Mark/Release to manipulate
the heap.
·Warning: A stiff differential equation occurs when there are at least two very
different scales of the independent variable on which the dependent variable(s) is
changing; for example, y = x + e- 1OOx • The Runge-Kutta method may generate a
numerical solution that bears no resemblance to the exact solution of the differen-
tial equation. This unstable numerical solution usually grows exponentially and
may be oscillatory. However, if the exact solution of the differential equation grows
as the independent variable increases, the instability may be difficult to detect. If a
suspected instability has been encountered, reduce the interval size (Numlnter-
vals).

Sample Program

The sample program RUNGE_S1.PAS provides I/O functions that demonstrate the
Runge-Kutta method of solving initial value problems for systems of first-order
ordinary differential equations. Note that the file RUNGE_S1.INC is included
after the TNTargetF functions are defined.

Example
Problem. A weight with mass m lays on a frictionless table and is connected to a
spring with spring constant k:

~Wall

F( w)
k
m

Frictionlts surface

Initial Value and Boundary Value Methods 191

If the mass is subject to a driving force F sin(w t) (w represents the frequency of
the driving force and t is time), the equation of motion of the mass is as follows:
2
m d xldt
2
+kx = F sin(w t)
Given
m = 2kg
F=9N
k = 32 N/m
w = 5 cycles/sec
x(O) = 0 m
dx(O)/dt = - 2.5 m/sec
find the position and velocity of the block from t = 0 second to t = 2 seconds.
1. Write the second-order ordinary differential equations as a system of two cou-
pled first-order ordinary differential equations:
dx/dt =x 2
dx/dt = (F/m) sin(w t) - (kim) Xl

2. Code these equations into the program RUNGE_Sl.PAS:

funct;on TNTargetF1(V : TNvector) : Real;
(**************************************************)
(* THIS IS THE FIRST DIFFERENTIAL EQUATION *)
(**************************************************)
(* *)
(* dx [1] *)
(* ----- = TNTargetF1(t, x[1], x[2], ..• x[m]) *)
(* dt *)
(* *)
(* The vector V is defined: *)
(* V[O] = t *)
(* V[1] = X[1] *)
(* V[2] = X[2] *)
(* *)
(* *)
(* *)
(* V[m] = X[m] *)
(* *)
(* where m is the number of coupled equations. *)
(**************************************************)

beg;n
TNTargetFl := V[2];
end; { function TNTargetF1 }
funct;on TNTargetF2{V TNvector): Real;

192 Turbo Numerical Methods Toolbox

(**************************************************)
(* THIS IS THE SECOND DIFFERENTIAL EQUATION *)
(**************************************************)
(* *)
(* dx[2] *)
(* TNTargetF2(t, x[l], x[2], ... x[m]) *)
(* dt *)
(* *)
(* The vector V is defined: *)
(* V[O] t *)
(* V[l] X[l] *)
(* V[2] X[2] *)
(* *)
(* *)
(* *)
(* V[m] X[m] *)
(* *)
(* where m is the number of coupled equations. *)
(**************************************************)

begin
TNTargetF2 := 9/2 * Sin(5 * V[O]) - 32/2 * V[l];
end; { function TNTargetF2 }
function TNTargetF3(V : TNvector) : Real;
(**************************************************)
(* THIS IS THE THIRD DIFFERENTIAL EQUATION *)
(**************************************************)
(* *)
(* dx[3] *)
(* TNTargetF3(t, x[l], x[2], ..• x[m]) *)
(* dt *)
(* *)
(* The vector V is defined: *)
(* V[0] t *)
(* V[l] X[l] *)
(* V[2] X[2] *)
(* *)
(* *)
(* *)
(* V[m] X[m] *)
(* *)
(* where m is the number of coupled equations. *)
(**************************************************)

begin
end; { function TNTargetF3 }
Functions TNTarget4 to TNTargetlO should be defined like the function
TNTargetF3.

Initial Value and Boundary Value Methods 193

3. Run RUNGE_S1.PAS:
Number of first order equations: (1-10)? 2

Lower limit of interval? 0

Upper limit of interval? 2

Enter X[l] value at t = O.OOOOOOOOE+OO: 0

Enter X[2] value at t = O.OOOOOOOOE+OO: -2.5
Number of values to return (1-100)? 10

Number of intervals (> = 10, default=10)? 100

Direct output to one of the following:

(S)creen
(P)rinter
(F) il e
Lower limit: O.OOOOOOOOOOOOOOE+OOO
Upper limit: 2.00000000000000E+000
Number of intervals: 100
Initial conditions at lower limit:
X[l]= O.OOOOOOOOOOOOOOE+OOO
X[2]= -2.50000000000000E+000
t Value X[l]
0.00000000 O.OOOOOOOOOOOOOOE+OOO
0.20000000 -4.20735284275848E-001
0.40000000 -4.54648724216734E-001
0.60000000 -7.05605786993375E-002
0.80000000 3.78400378699554E-001
1.00000000 4.79461767300631E-001
1.20000000 1.39708469016311E-001
1.40000000 -3.28491796183335E-001
1.60000000 -4.94677974769030E-001
1.80000000 -2.06059519715175E-001
2.00000000 2.72008842396951E-001
t Value X[2]
0.00000000 -2.50000000000000E+000
0.20000000 -1.35075642830665E+000
0.40000000 1.04036531118478E+000
0.60000000 2.47497991717220E+000
0.80000000 1.63411037473655E+000
1.00000000 -7.09151289407567E-001
1.20000000 -2.40042152228323E+000
1.40000000 -1.88475529635974E+000
1.60000000 3.63745224811839E-001
1.80000000 2.27781864414105E+000
2.00000000 2.09767516082021E+000
X[l] are the values of x(t), the position. X[2] are the values of dx(t)/dt, the veloc-
ity.

194 Turbo Numerical Methods Toolbox

The exact solution is
F sin{w t)
x=-------

F", cos{w t)
dxldt = -------
2 2
m (wo - ( )

where Ulo is the natural frequency of the system:

2
wo = kim
The period of oscillation is given by
T = 2 -rrlw = 1.257 sec
The data is taken from a function of which the derivative could be computed
exactly. The actual values are as follows:
t Values of X Derivative of X
0.0 O.OOOOOOOOOOOOE + 000 - 2.500000000000E + 000
0.2 - 4.207354924039E - 001 - 1.350755764670E + 000
0.4 - 4.546487134128E - 001 1.040367091367E + 000
0.6 - 7.056000402993E - 002 2.474981241501E + 000
0.8 3.784012476539E -001 1.634109052159E + 000
1.0 4.794621373315E -001 -7.091554636580E - 001
1.2 1.397077490994E - 001 - 2.400425716625E + 000
1.4 - 3.284932993593E - 001 - 1.884755635858E + 000
1.6 - 4.946791233116E - 001 3.6375008452l5E - 001
1.8 - 2.060592426208E - 001 2.277825654711E + 000
2.0 2.720105554446E - 001 2.097678822691E + 000

Initial Value and Boundary Value Methods 195

Solution to an Initial Value Problemfor a System of
Coupled Second-Order Ordinary Differential Equations
Using the Runge-Kutta Metlwd (RUNGE-S2.INC)

Description

This example integrates a system of coupled second-order ordinary differential

equations with specified initial conditions using the generalized Runge-Kutta for-
mulas (Burden and Faires 1985,261-269),
Given m coupled second-order ordinary differential equations of the form
2 2
d x/dt = TNTargetF1(t, Xl' X'I' X2' X' 2' .. ,' Xm ' X' J
2 2
d x/dt = TNTargetF2(t, Xl' X'I' X 2' X' 2' .. " Xrn ' X' J

2 2
d x",Idt = TNTargetFm(t, Xl' X'I' X2' X' 2' .. ,' X rn ' X' J
where x') indicates dx/dt, which satisfies the Lipshitz condition (the Lipshitz con-
dition for first-order and second-order ordinary differential equations is given at
the beginning of this chapter; consult the previous book reference for details of the
Lipshitz condition for systems), and initial condition
xJLowerLimit] =a l
x' JLowerLimit] = hI
X' 2[LowerLimit] = h2 .

xJLowerLimit] = am x'JLowerLimit] = hm
and spacing
h = (UpperLimit - LowerLimit)/Numlntervals

196 Turbo Numerical Methods Toolbox

rewrite each of the second-order differential equations as two, first-order differen-
tial equations:
dx/dt =Y I

dy/dt = TNTargetFI(t, Xl' YI' X2' Y2, "', Xm ' yJ

dx/dt = Y2
dx/dt = TNTargetF2(t, Xl' YI, X2' Y2, .. ,' Xm ' yJ

dx)dt = Ym
dx)dt = TNTargetFm(t, Xl' YI, X2' Y2' "', Xm' yJ
Then the fourth-order general Runge-Kutta method can be used to approximate
the xj's and the yj's simultaneously,
The general Runge-Kutta formulas for these equations are as follows:
Flxl = h * Yl
FIY I = h * TNTargetFI(t, xl[t], yJt], x2[t], y2[t], "', xJt], yJt])
Flx2 = h * Y2
FIY2 = h * TNTargetF2(t, xJt], yJt], x2[t], y2[t], "', xJt], yJt])

Flxm = h * Ym
FIY m = h * TNTargetFm(t, xJt], yJt], x2[t], y2[t], .. ,' xJt], yJt])

F2x1 = h * (YI + Fly/2)

F2YI = h * NTargetFI(t + h/2, xJt] + Flx/2, yl[t] + Fly/2, x2[t]
+ Flx/2, Y2[t] + Fly/2, "', xJt] + Flx)2, yJt] + Fly)2)
F2x2 = h * (Y2 + Fly/2)
F2Y2 = h * NTargetF2(t + h/2, xJt] + Flx/2, yJt] + Fly/2, x2[t]
+ Flx/2, Y2[t] + Fly/2, .. ,' xJt] + Flx)2, yJt] + Fly)2)

F2x m = h * (Yrn + Fly)2)

F2Yrn = h * TNTargetFm(t + h/2, xl[t] + Flx/2, yl[t] + Fly/2, x2[t]
+ Flx/2, Y2[t] + Fly/2, "', x.Jt] + Flx,/2, y.Jt] + Fly,/2)

Initial Value and Boundary Value Methods 197

F3x l ::::: h * (Yl + F2y/2)
F3Yl ::::: h * TNTargetFI(t + h/2, xl[t] + F2x/2, Yl[t] + F2y/2, X2[t]
+ F2x/2, Y2[t] + F2y/2, ... , xJt] + F2x)2, yJt] + F2y)2)
F3x 2 ::::: h * (Y2 + F2y/2)
F3Y2 ::::: h * NTargetF2(t + h/2, xJt] + F2x/2, Yl[t] + F2y/2, X2[t]
+ F2x/2, Y2[t] + F2y/2, ... , x)t] + F2x)2, yJt] + F2y)2)

F3x m ::::: h * (Ym + F2y)2)

F3y", ::::: h * TNTargetFm(t + h/2, xl[t] + F2x/2, Yl[t] + F2y/2, X2[t]
+ F2x/2, Y2[t] + F2y/2, ... , x)t] + F2x)2, y)t] + F2y)2)

F4xl ::::: h * (Yl + F3Yl)

F4Yl ::::: h * TNTargetFI(t + h, xl[t] + F3x l, yJt] + F3Yl' X2[t] + F3x 2, Y2[t]
+ F3Y2' ... , X,Jt] + F3x"" yJt] + F3y,J
F4x2 ::::: h * (Y2 + F3Y2)
F4Y2 ::::: h * TNTargetF2(t + h, xl[t] + F3x l, yJt] + F3Yl' X2[t] + F3x 2, Y2[t]
+ F3Y2' ... , X.Jt] + F3x"" y"Jt] + F3y,J

F4x", ::::: h * (y", + F3y)

F4y", ::::: h * TNTargetFm(t + h, xJt] + F3x l, yJt] + F3Yl' X2[t] + F3x 2, Y2[t]
+ F3Y2' ... , X)t] + F3x"" y,Jt] + F3y)

xJt + 1] ::::: xJt] + (Flx l + 2 * F2xl + 2 * F3x l + F4x l)/6

yJt+ 1] ::::: yJt] + (FlYl + 2 * F2Yl + 2 * F3Yl + F4y)/6
x2[t + 1] ::::: x2[t] + (Flx 2 + 2 * F2x2 + 2 * F3x2 + F4x2)/6
Y2[t + 1] ::::: Y2[t] + (FlY 2 + 2 * F2Y2 + 2 * F3Y2 + F4Y2)/6

x [t + 1] ::::: xm [t] + (FIx m + 2

til
* F2x + 2 * F3x + F4x )/6
III til m

yJt + 1] ::::: yJt] + (Fly", + 2 * F2y", + 2 * F3y", + F4y,)/6

where t ranges from LowerLimit to UpperLimit in steps of h. These formulas give a
truncation error of order h4.

198 Turbo Numerical Methods Toolbox

You must supply the number of equations, limits, initial values, and TNTargetF's.
This procedure can solve a system of up to ten, second-order ordinary differen-
tial equations (see "Comments" for information about how to increase this limit).

User-Defined Types

TNData = record
x : Real;
xDeriv : Real;
end; { TNData record}
TNvector = array[O •• TNRowSize] of TNData;
TNmatrix = array[O •• TNColumnSize] of TNvector;
TNRowSize is an upper bound for the number of values returned for a particular
variable (NumReturn). TNColumnSize is an upper bound for the number of second-
order differential equations (NumEquations).

User- Defined Functions

function TNTargetFl(V TNvector) Real;

function TNTargetF2(V TNvector) Real;
function TNTargetF3(V TNvector) Real;
function TNTargetF4(V TNvector) Real;
function TNTargetF5(V TNvector) Real;
function TNTargetF6(V TNvector) Real;
function TNTargetF7(V TNvector) Real;
function TNTargetF8(V TNvector) Real;
function TNTargetF9(V TNvector) Real;
function TNTargetFIO(V : TNvector) : Real;

Initial Value and Boundary Value Methods 199

Here are the differential equations:
2
d x/dl = TNTargetFj(t, Xl' X'I' X2 ' X' 2' ... , X IO ' X'1O)

where j ranges from 1 to 10.

The elements of the vector V are defined as follows:

V[O].X =t
V[l].x = x[l]
V[l].xDeriv = x'[l]
V[2].x = x[2]
V[2].xDeriv = x ' [2]

V[10].x = x[10]
V[10].xDeriv = x ' [10]
The procedure defined in RUNGE_S2.INC solves this system of coupled differ-
ential equations (a maximum of ten equations). All ten functions must be defined,
even if your system contains less than ten equations.

Input Parameters

NumEquations : Integer; Number of second-order differential equations

LowerLimit : Real; Lower limit of interval
UpperLimit : Real; Upper limit of interval
InitialValues : TNvector; Values of Xj'S and X'j'S at LowerLimit
NumReturn : Integer; Number of (t, Xl' X'I' X2 ' X' 2' "', Xm ' X' J values returned
from the procedure
NumInterval s : Integer; Number of subintervals used in the calculations

200 Turbo Numerical Methods Toolbox

The preceding parameters must satisfy the following conditions:
1. NumReturn > 0
2. Numlntervals;::: NumReturn
3. NumEquations > 0
4. LowerLimit ~ UpperLimit

Output Parameters

SolutionValues : TNmatrix; Values oft, Xi' and x'} between the limits
Error: Byte; 0: No errors
1: NumReturn < 1
2: Numlntervals < NumReturn
3: NumEquations < 1
4: LowerLimit = UpperLimit

Syntax of the Procedure Call

InitialConditionSystem2(NumEquations, LowerLimit, UpperLimit,

InitialValues, NumReturn, Numlntervals,
SolutionValues, Error);

Comments

The first row of SolutionValues will be the values of t between the . limits, the
second row of Solution Values will be the values of Xl and XiI between the limits, the
third row of Solution Values will be the values of x2 and x' 2 between the limits, and
so on.
All ten user-defined functions are called from the procedure. If your system has
less than ten equations, you must still define all ten functions or the program will
not compile. The superfluous functions should be defined as follows (TNTargetF10
is used as an example):
function TNTargetFIO(V : TNvector) : Real;
begin
end; { function TNTargetFIO }

Initial Value and Boundary Value Methods 201

If you need to solve a system with more than ten equations, then edit the include
file RUNGE-S2.1NC. The following lines should be added to the end of procedure
Step:
F[ll].xDeriv := Spacing * CurrentValues[ll].xDeriv;
F[ll].x := Spacing * TNTargetFll(CurrentValues);
More statements (for F[12], and so on) may be added as necessary. All new
functions (for example, TNTargetFll) must be defined in your top-level program.
Note: Before making any changes to the include file, make sure you have a backup
copy.
The procedure will compute Numlntervals values in its calculations; however,
you will rarely need to use these values. The rows of SolutionValues will contain
only NumReturn values at roughly equally spaced t-values between the lower and
upper limits. (They will be equally spaced only when Numlntervals is a multiple of
NumReturn.) Thus, you can ensure a highly accurate solution (by making Numln-
tervals large) without generating an excessive amount of output (by making Num-
Return small).
This routine stores much information on the heap. If you try to accurately solve a
large system (that is, both NumEquations and Numlntervals are large), you may get
run-time error $FF, Heap/Stack collision. If this happens, the dimension of TNvec-
tor and TNmatrix should be reduced as much as possible. If this is not possible,
then remove any RAM-resident software (for example, SideKick, SuperKey, or'a
print buffer).
The Runge-Kutta method uses the New/Dispose procedures to manipulate the
heap and should not be used in any program that uses Mark/Release to manipulate
the heap.
Warning: A' stiff differential equation occurs when there are at least two very
different scales of the independent variable on which the dependent variable(s) is
changing; for example, y = x + e- 1oo The Runge-Kutta method may generate a
%.

numerical solution that bears no resemblance to the exact solution of the differen-
tial equation. This unstable numerical solution usually grows exponentially and
may be oscillatory. However, if the exact solution of the differential equation grows
as the independent variable increases, the instability may be difficult to detect. If a
suspected instability has been encountered, reduce the interval size (Numlnter-
vals).

202 Turbo Numerical Methods Toolbox

Sample Program

The sample program RUNGE_S2.PAS provides I/O functions that demonstrate

the Runge-Kutta method of solving initial value problems for systems of first-order
ordinary differential equations. Note that the file RUNGE_S2.INC is included
after the TNTargetF functions are defined.

Example
Problem.Two weights of mass m each hang from a pendulum of length l and are
connected by a spring with spring constant k:

Ceiling

k
®-~
y UUUUUUUUUUUU X

The equations of motion of these two masses are as follows:

m d 2x/dt2 = -m g x/I - k(x - y)
d y/de = -m g y/l + k(x - y)
2
m

where g is the acceleration due to gravity, t is time, and x and yare the displace-
ments of the two weights from their rest positions. Given
m = 2kg
k = 32 N/m
2
g = 9.8 m/sec
l = 0.6l25 m
x(O) = 1 m
y(O) = -1 m
dx(O)/dt = 0 m/sec
dy(O)/dt = 0 m/sec

Initial Value and Boundary Value Methods 203

find the positions and velocities of the two weights from t o second to t 2
seconds.
1. Rewrite the equations of motion as shown here:
d 2xldt2 = - g xll - klm(x - y)
d2Yldt2 = - g yll + klm(x - y)
2. Code these equations into the program RUNGE_S2.PAS:
function TNTargetFl(V : TNvector) : Real;
(*******************************************************)
(* THIS IS THE FIRST DIFFERENTIAL EQUATION *)
(*******************************************************)
(* *)
(* *)
(* d2 x[l] *)
(* TNTargetFl(t, x[l], x'[l], x[2], x'[2], *)
(* .•• , x [m], x' [m] *)
(* *)
(* dt 2 *)
(* *)
(* The elements of the vector V are defined: *)
(* V[O].x = t *)
(* V[l].x = X[l] *)
(* V[l].xDeriv X'[l] *)
(* V[2].x X[2] *)
(* V[2] .xDeriv = X' [2] *)
(* *)
(* *)
(* *)
(* V[m].x X[m] *)
(* V[m].xDeriv X'[m] *)
(* *)
(* where m is the number of coupled equations. *)
(*******************************************************)

var
t Real;
begin
t := v[O] .x;
TNTargetFl .- -9.8 * V[1].x/O.6125 - 32/2 * (V[l].x - V[2].x);
end; { function TNTargetFl }

204 Turbo Numerical Methods Toolbox

function TNTargetF2(V : TNvector) : Real;
(*******************************************************)
(* THIS IS THE SECOND DIFFERENTIAL EQUATION *)
(*******************************************************)
(* *)
(* *)
2
(* d x[2] *)
(* TNTargetF2(t, x[l], Xl [1], x[2], Xl [2], *)
(* ... , x [m], x I [m] *)
2
(* dt *)
(* *)
(* *)
(* The elements of the vector V are defined: *)
(* V[O].x t *)
(* V[l].x X[l] *)
(* V[l].xDeriv X' [l] *)
(* V[2].x X[2] *)
(* V[2].xDeriv X' [2] *)
(* *)
(* *)
(* *)
(* V[m] .x X[m] *)
(* V[m].xDeriv XI[m] *)
(* *)
(* where m is the number of coupled equations. *)
(*******************************************************)

var
t Real;
begin
t := v[O] .x;
TNTargetF2 := -9.8 * V[2].x/O.6125 + 32/2 * (V[l].x - V[2].x);
end; { function TNTargetF2 }

Initial Value and Boundary Value Methods 205

function TNTargetF3(V : TNvector) : Real;
(*******************************************************)
(* THIS IS THE THIRD DIFFERENTIAL EQUATION *)
(*******************************************************)
(* *)
(* *)
2
(* d x[3] *)
(* TNTargetF3(t, x[l], x'[l], x[2], x'[2], *)
(* ••• , x [m], x' [m] *)
(* dt 2 *)
(* *)
(* *)
(* The elements of the vector V are defined: *)
(* V[O].x t *)
(* V[l].x X[l] *)
(* V[l].xDeriv X'[l] *)
(* V[2].x X[2] *)
(* V[2].xDeriv X'[2] *)
(* *)
(* *)
(* *)
(* V[m].x X[m] *)
(* V[m].xDeriv X'[m] *)
(* *)
(* where m is the number of coupled equations. ~)
(*******************************************************)

var
t Real;
begin
end; { function TNTargetF3 }
Functions TNTargetF 4 to TNTargetF10 should be defined like function TNTargetF3.
3. Run RUNGE_S2.PAS:
Number of second order equations: (1-20)? 2
Lower limit of interval? 0
Upper limit of interval?
Enter X[l] value at t O.OOOOOOOOE+OO: 0.01
Enter X'[l] value at t O.OOOOOOOOE+OO: 0.00
Enter X[2] value at t = O.OOOOOOOOE+OO: -0.01
Enter X'[2] value at t = O.OOOOOOOOE+OO: 0.00
Number of values to return (1-100)? 10
Number of intervals (>= 10, default=10)? 100

206 Turbo Numerical Methods Toolbox

Direct output to one of the following:
(S)creen
(P)rinter
(F)ile

Lower limit: O.OOOOOOOOOOOOOOE+OOO

Upper limit: 1.00000000000000E+000
Number of intervals: 100
Initial conditions at lower limit:
X[l]= 1.00000000000000E-002
X' [l]= O.OOOOOOOOOOOOOOE+OOO
X[2]= -1.00000000000000E-002
X' [2]= O.OOOOOOOOOOOOOOE+OOO
t Value X[l] Deriv X[I]
0.00000000 1.00000000000000E-002 O.OOOOOOOOOOOOOOE+OOO
0.10000000 7.69447788485895E-003 -4.42511063153028E-002
0.20000000 1.84099813762452E-003 -6.80978317847279E-002
0.30000000 -4.86137387553900E-003 -6.05443464988731E-002
0.40000000 -9.32214486443693E-003 -2.50735962983904E-002
0.50000000 -9.48443369885917E-003 2.19586991271007E-002
0.60000000 -5.27340834792187E-003 5.88657408762406E-002
0.70000000 1.36920877108260E-003 6.86295294795966E-002
0.80000000 7.38047758874091E-003 4.67479393932010E-002
0.90000000 9.98857556718864E-003 3.31066873866277E-003
1.00000000 7.99089728515028E-003 -4.16531651968366E-002

t Value X[2] Deri v X[2]

0.00000000 -1.00000000000000E-002 O.OOOOOOOOOOOOOOE+OOO
0.10000000 -7.69447788485895E-003 4.42511063153028E-002
0.20000000 -1.84099813762452E-003 6.80978317847279E-002
0.30000000 4.86137387553900E-003 6.05443464988731E-002
0.40000000 9.32214486443693E-003 2.50735962983904E-002
0.50000000 9.48443369885917E-003 -2.19586991271007E-002
0.60000000 5.27340834792187E-003 -5.88657408762406E-002
0.70000000 -1.36920877108260E-003 -6.86295294795966E-002
0.80000000 -7.38047758874091E-003 -4.67479393932010E-002
0.90000000 -9.98857556718864E-003 -3.31066873866277E-003
1.00000000 -7.99089728515028E-003 4.16531651968366E-002
The weights move in opposite directions; the system is in one of its normal
modes. The natural frequency 00 is given by the following:
0

00,,2 = gil + 2k/m

00" = 6.928 cycles/sec
Thus the period of oscillation, t, is
t = 27r/ooo
t = 0.9069 sec

Initial Value and Boundary Value Methods 207

Solution to Boundary Value Problemfor a Second-Order
Ordinary Differential Equation Using the Slwoting and
Runge-Kutta Metlwds (SHOOT2.INC)

Description

This example uses the shooting method to approximate the solution to a second-
order ordinary differential equation with specified boundary conditions (Burden
and Faires 1985,526-531).
Given a second-order differential equation (Burden and Faires 1985, 261-269) of
the form
2 2
d y/dx = TNTargetF(x, y, y')
where y' represents dy/dx, which satisfies the conditions given at the beginning of
this chapter, boundary conditions
y[LowerLimit] = Lowerlnitial
y[UpperLimit] = Upperlnitial
and spacing
h = (UpperLimit - LowerLimit)/Numlntervals
and an initial approximation (guess) to the slope at LowerLimit
y'[LowerLimit] = InitialSlope
the shooting method first solves the second-order initial value problem (using the
method described in RUNGE-2.1NC). Based on a comparison of the solution at
UpperLimit with the boundary condition Upperlnitial, a new approximation to the
slope at LowerLimit is made. In this way, a new "shot" at the solution is made by
observing the result of the previous "shot." Subsequent iterations use information
from two previous shots and the secant method (see Chapter 2, "Roots of a Func-
tion Using the Secant Method") to approximate the slope at LowerLimit. This pro-
cess is repeated until the fractional difference between successive approximations
to the boundary condition at UpperLimit is less than a specified tolerance.
You must supply the LowerLimit, UpperLimit, LowerInitial, Upperlnitial,
InitialSlope, Numlntervals, Tolerance, and TNTargetF.

208 Turbo Numerical Methods Toolbox

User-Defined Types

TNvector = array[l •• TNArraySize] of Real;

User-Defined Function

TNTargetF(x, y, yPrime : Real) : Real;

cfy/dx2 = TNTargetF(x, y, dy/dx)

The procedure S/woting integrates this second-order differential equation.

Input Parameters

LowerLimit : Real; Lower limit of interval

UpperLimit : Real; Upper limit of interval
LowerInitial : Real; Value of y at LowerLimit
UpperInitial : Real; Value of y at UpperLimit
InitialSlope: Real; Approximation to the slope at LowerLimit
NumReturn : Integer; Number of (x, y, y') values returned from the procedure
Tolerance: Real; Indicates accuracy in solution
Max Iter : Integer; Maximum number of ilerations
NumInterval s : Integer; Number of subintervals used in calculations

The preceding parameters must satisfy the following conditions:

1. NumReturn > 0
2. Numlntervals ~ NumReturn
3. LowerLimit;c UpperLimit
4. Tolerance > 0
5. Maxlter > 0

Initial Value and Boundary Value Methods 209

Output Parameters

Iter: Integer: Number of iterations required to reach a solution

XVal ues : TNvector: Values of x between the limits
YValues : TNvector: Values of y determined at values in XValues
YDerivValues : TNvector: Values of the first derivative of y determined at values in
XValues
Error: Byte: 0: No errors
1: NumReturn < 1
2: Numlntervals < NumReturn
3: LowerLimit = UpperLimit
4: Tolerance S 0
5: MaxIter S 0
6: Iter> MaxIter
7: Convergence not possible

Syntax of the Procedure Call

Shooting(LowerLimit, UpperLimit, LowerInitial, UpperInitial, InitialSlope,

NumReturn, Tolerance, MaxIter, NumIntervals, Iter, XValues,
YValues, YDerivValues, Error):

Comments

The parameter Tolerance can be misleading. The shooting method converges to the
initial slope, which satisfies the upper boundary condition. Convergence is
achieved when the fractional difference between Upperlnitial and the upper
boundary approximation is less than Tolerance. This does not mean that every
value between the boundaries has been approximated with the same degree of
accuracy. To improve the accuracy of the entire approximation, increase the num-
ber of intervals. The example demonstrates the different effects of Tolerance and
Numlntervals.
The shooting algorithm will compute Numlntervals values in its calculations.
However, you will rarely need to use all those values. The vectors XValues,
YValues, and YDerivValues will contain only NumReturn values at roughly equally
spaced t-values between the lower and upper limits. (They will be equally spaced
only when Numlntervals is a multiple of NumReturn.) Thus, you can ensure a
highly accurate solution (by making Numlntervals large) without generating an
excessive amount of output (by making NumReturn small).

210 Turbo Numerical Methods Toolbox

Boundary value problems are notoriously difficult to solve. The shooting method
is extremely sensitive to the approximation of the initial slope. If the shooting
method does not converge onto a solution (Error 7), run the program with a differ-
ent value of the initial slope InitialSlope. You may also alleviate some stability
problems by solving the equation backwards (from UpperLimit to LowerLimit).
Considerable trial and error may be involved before a solution is found.
The Runge-Kutta method uses the New/Dispose procedures to m~nipulate the
heap and should not be used in any program that uses Mark/Release for heap
management.
Warning: A stiff differential equation occurs when there are at least two very
different scales of the independent variable on which the dependent variable(s) is
changing; for example, y = x + e • The shooting method may generate a numeri-
1OOx

cal solution that bears no resemblance to the exact solution of the differential equa-
tion. This unstable numerical solution usually grows exponentially and may be
oscillatory. However, if the exact solution of the differential equation grows as the
independent variable increases, the instability may be difficult to detect. If a sus-
pected instability has been encountered, reduce the interval size (Numlntervals).

Sample Program

The sample program SHOOT2.PAS provides I/O functions that demonstrate the
shooting method of solving boundary value problems. Note that the file SHOOT
2.1NC is included after the function TNTargetF is defined.

Example
Problem. Use the nonlinear shooting method to solve the following boundary value
problem:
Y 1/ = 192 sqr(y/y') 0:5x:51

y(l) = 1
y(2) = 16

Initial Value and Boundary Value Methods 211

1. Code the differential equation into the program:
function TNTargetF(x : Real;
y : Real;
yPrime : Real) : Real;
(*******************************************************************)
(* THIS IS THE SECOND-ORDER NONLINEAR DIFFERENTIAL EQUATION *)
(*******************************************************************)

begin
TNTargetF := 192 * Sqr(y/yPrime);
end; {function TNTargetF}
2. Run SHOOT2.PAS:
Lower limit of interval? 0
Upper limit of interval? 1
Enter Y value at X = O.OOOOOOOOE+OO: 1
Enter Y value at X = 1.00000000E+00: 16
Enter a guess for the slope at X = O.OOOOOOOOE+OO (default=1.50E+01): 15
Number of points returned (1-500)? 10
Number of intervals (>= 10, default=10)? 10
Tolerance (> 0, default = 1.000E-06)? 1E-12
Maximum number of iterations (> 0, default 100)? 100
Direct output to one of the following:
(S)creen
(P) ri nter
(F) il e

Lower limit: O.OOOOOOOOOOOOOOE+OOO

Upper limit: 1.00000000000000E+000
Value of Y at 0.0000: 1.00000000000000E+000
Value of Y at 1.0000: 1.60000000000000E+001
Initial slope at 0.0000: 1.50000000000000E+001
NumIntervals: 10
Tolerance: 1.000E-012
Maximum number of iterations: 100

212 Turbo Numerical Methods Toolbox

Number of iterations: 6

X Y Value Derivative of Y
O.OOOOOOOOOOOOOOE+OOO 1.00000000000000E+000 4.00053795390884E+000
1.00000000000000E-001 1.46417721408153E+000 5.32386904044879E+000
2.00000000000000E-001 2.07370562259973E+000 6.91162114244397E+000
3.00000000000000E-001 2.85621262766442E+000 8.78752756627335E+000
4.00000000000000E-001 3.84170902091389E+000 1.09754927855527E+001
5.00000000000000E-001 5.06259931530967E+000 1.34994802016423E+001
6.00000000000000E-001 6.55368547624580E+000 1.63834750611955E+001
7.00000000000000E-001 8.35216836918581E+000 1.96514712240017E+001
8.00000000000000E-001 1.04976483580762E+001 2.33274661179548E+001
9.00000000000000E-001 1.30321255669365E+001 2.74354587043771E+001
1.00000000000000E+000 1.60000000000094E+001 3. 19994486182108E+001
Now solve the same problem using a smaller spacing, but with a greater tolerance:
Lower limit of interval? 0
Upper limit of interval? 1

Enter Y value at X = O.OOOOOOOOE+OO: 1

Enter Y value at X = 1.00000000E+00: 16
Enter a guess for the slope at X = O.OOOOOOOOE+OO (default = 1.50E+01): 15

Number of points returned (1-500)? 10

Number of intervals (>= 10, default = 10)? 100

Tolerance (> 0, default = 1.000E-06)? 1E-6

Maximum number of iterations (> 0, default = 100)? 100

Direct output to one of the following:
(S)creen
(P)rinter
(F) il e

Lower 1i mit: O.OOOOOOOOOOOOOOE+OOO

Upper 1i mit: 1.00000000000000E+000
Value of Y at 0.0000: 1.00000000000000E+000
Value of Y at 1.0000: 1.60000000000000E+001
Initial slope at 0.0000: 1.50000000000000E+001
NumIntervals: 100
Tolerance: 1.000E-006
Maximum number of iterations: 100

Initial Value and Boundary Value Methods 213

Number of iterations: 5
X Y Value Derivative of Y
O.OOOOOOOOOOOOOOE+OOO 1.OOOOOOOOOOOOOOE+OOO 4.00000062625638E+OOO
1.OOOOOOOOOOOOOOE-OOl 1.46410005120828E+OOO 5.32400035609946E+OOO
2.00000000000000E-OOI 2.07360008576235E+OOO 6.91200027103432E+OOO
3.00000000000000E-OOI 2.85610011557157E+OOO 8.78800025750412E+OOO
4.00000000000000E-OOI 3.84160014547825E+OOO 1.09760002747783E+OOl
5.00000000000000E-OOI 5.06250017769403E+OOO 1.35000003070170E+OOl
6.00000000000000E-OOI 6.55360021337284E+OOO 1. 63840003476283E+OOI
7.00000000000000E-OOI 8.35210025321451E+OOO 1.96520003937080E+OOl
8.00000000000000E-OOI 1.04976002977125E+OOl 2.33280004439140E+OOI
9.00000000000001E-OOI 1.30321003472617E+OOl 2.74360004976014E+OOI
1.OOOOOOOOOOOOOOE+OOO 1.60000004022081E+OOl 3.20000005544507E+OOI
The exact solution is
y = (x + It
X Y Value Derivative ofY
0.0 1.0000000000 4.000000000
0.1 1.4641000000 5.324000000
0.2 2.0736000000 6.912000000
0.3 2.8561000000 8.788000000
0.4 3.8416000000 1.097600000
0.5 5.0625000000 1.350000000
0.6 6.5536000000 1.638400000
0.7 8.3521000000 1.965200000
0.8 1.0497600000 2.332800000
0.9 1.3032100000 2.743600000
1.0 1.6000000000 3.200000000
Although the tolerance is smaller (that is, more exacting) in the first case, the
accuracy of the approximation is greater in the second case. The spacing in the first
case is too large to permit a more accurate approximation.

214 Turbo Numerical Methods Toolbox

Solution to a Boundary Value Problem/or a Second-Order
Ordinary Linear Differential Equation Using the Linear
Shooting and Runge-Kutta Methods (LINSHOT2.INC)

Description

This example uses the linear shooting method to approximate the solution to a
second-order linear ordinary differential equation with specified boundary condi-
tions (Burden and Faires 1985, 519-524).
Given a second-order differential equation (Burden and Faires 1985, 261-264) of
the form
d2y/dx2 = TNTargetF(x, y, y')
which is linear in both y and y', where y' represents dy/dx, and which satisfies the
conditions given at the beginning of this chapter, boundary conditions
y[LowerLimit] = Lowerlnitial
y[UpperLimit] = Upperlnitial
and spacing
h = (UpperLimit - LowerLimit)/Numlntervals
the shooting method solves the two initial value problems (see RUNGE-2.INC):
y'[LowerLimit] =0 y[LowerLimit] = Lowerlnitial
y'[LowerLimit] =1 y[LowerLimit] = Lowerlnitial
(These values are particular to this implementation; any other nonidentical set of
initial conditions will suffice.) Since neither of these initial values of y' is likely to
be correct, the solutions generated are not likely to satisfy the boundary condition
at Upperlnitial. However, because of the linearity of the equation, an appropriate
linear combination of these two solutions will be a solution to the boundary value
problem. The linear shooting method requires that only two initial value problems
be solved, where the ordinary shooting method (SHOOT2.INC) requires that an
unknown number of initial value problems be solved before the method converges
to a solution.
You must supply the LowerLimit, UpperLimit, Lowerlnitial, Upperlnitial,
Numlntervals, and TNTargetF.

Initial Value and Boundary Value Methods 215

User-Defined Types

TNvector = array[l •• TNArraySize] of Real;

User-Defined Function

TNTargetF(x, y, yPrime : Real) : Real;

2 2
d y/dx = TNTargetF(x, y, dy/dx}
The procedure LinearSlwoting integrates this second-order differential equation.

Input Parameters

LowerLimit : Real; Lower limit of interval

UpperLimit : Real; Upper li.:nit of interval
LowerInitial : Real; Value of y at LowerLimit
UpperInitial : Real; Value of y at UpperLimit
NumInterval s : Integer; Number of subintervals used in calculations
NumReturn : Integer; Number of (x, y, y') triples returned from the procedure
The preceding parameters must satisfy the following conditions:
1. NumReturn > 0
2. Numlntervals;::: NumReturn
3. LowerLimit ~ UpperLimit

Output Parameters

XValues : TNvector; Values of x between the limits

YVa 1ues : TN vector ; Values of y determined at values in XValues
YDerivVal ues : TNvector; Values of the first derivative of y determined at values in
XValues
Error: Byte; 0: No errors
1: NumReturn < 1
2: Numlntervals < NumReturn
3: LowerLimit = UpperLimit
4: Equation not linear

216 Turbo Numerical Methods Toolbox

Syntax of the Procedure Call

LinearShooting(LowerLimit, UpperLimit, LowerInitial, UpperInitial,

NumReturn, Numlntervals, XValues, YValues,
YDerivValues, Error);

Comments

If TNTargetF is a nonlinear function, the linear shooting algorithm will usually

compute a solution (albeit an incorrect one) without returning an error message.
Error 4 (nonlinear equation) will be returned in only a few cases where the two
initial value problems happen to yield solutions with the same y-value at
x = UpperLimit.
The procedure will compute Numlntervals values in its calculations; however,
you will rarely need to use these values. The vectors XValues, YValues, and
YDerivValues will contain only NumReturn values at roughly evenly spaced inter-
vals between the lower and upper limits. (They will be exactly evenly spaced only
when Numlntervals is a multiple of NumReturn.) Thus, you can ensure a highly
accurate solution (by making Numlntervals large) without generating an excessive
amount of output (by making NumReturn small).
The Linear Shooting/Runge Kutta method uses the New/Dispose procedures to
manipulate the heap and should not be used in any program that uses Mark/
Release to manipulate the heap.
Warning: A stiff differential equation occurs when there are at least two very
different scales of the independent variable on which the dependent variable(s) is
changing; for example, y = x + e - lOOx. The Runge-Kutta method may generate a
numerical solution that bears no resemblance to the exact solution of the differen-
tial equation. This unstable numerical solution usually grows exponentially and
may be oscillatory. However, if the exact solution of the differential equation grows
as the independent variable increases, the instability may be difficult to detect. If a
suspected instability has been encountered, reduce the interval size (Numlnter-
vals).

Initial Value and Boundary Value Methods 217

Sample Program

The sample program LINSHOT2.PAS provides I/O functions that demonstrate the
linear shooting method of solving boundary value problems. Note that the file
LINSHOT2.1NC is included after the function TNTargetF is defined.

Example
Problem.Use the linear shooting method to solve the following boundary value
problem:
y" = y' /x - y/sqr(x) +1 1 =x S 10
y(1) 1
y(10) 76.974149

1. Code the differential equation into the program LINSHOT2.PAS:

funct;on TNTargetF(x : Real;
y : Real;
yPrime : Real) : Real;
(***************************************************************)
(* THIS IS THE SECOND-ORDER DIFFERENTIAL EQUATION *)
(***************************************************************)

beg;n
TNTargetF := yPrime/x - y/Sqr(x) + 1;
end; { function TNTargetF }
2. Run LINSHOT2.PAS:
Lower limit of interval? 1
Upper limit of interval? 10
Enter Y value at X = 1.00000000E+00: 1
Enter Y value at X = 1.00000000E+01: 76.974149
Number of points returned (1-500)? 9
Number of intervals (>= 9, default = 9)? 9
Direct output to one of the following:
(S)creen
(P)rinter
(F)ile
Lower limit: 1.00000000000000E+00
Upper limit: 1.00000000000000E+01
Value of Y at 1.0000: 1.00000000000000E+00
Value of Y at 10.0000: 7.69741490000000E+01
NumIntervals: 9.00000000000000E+00

218 Turbo Numerical Methods Toolbox

X Y Value Derivative of Y
1.00000000000000E+000 1.00000000000000E+000 1.00042467674563E+000
2.00000000000000E+000 2.61170356138588E+000 2.30627678512124E+000
3.00000000000000E+000 5.70207271413620E+000 3.90115296191831E+000
4.00000000000000E+000 1.04528257144925E+001 5.61367861126495E+000
5.00000000000000E+000 1.69509897305375E+001 7.39067355864438E+000
6.00000000000000E+000 2.52478687612139E+001 9.20845513089500E+000
7.00000000000000E+000 3.53773649984557E+001 1.10543869346579E+001
8.00000000000000E+000 4.73635728977226E+001 1.29209245920937E+001
9.00000000000000E+000 6. 12245068576119E+001 1.48032011472994E+001
1.00000000000000E+001 7.69741490000000E+001 1.66978931711222E+001
Now solve the same problem with a spacing of only 0.1:
Lower limit of interval? 1

Upper limit of interval? 10

Enter Y value at X = 1.00000000E+00: 1

Enter Y value at X = 1.00000000E+01: 76.974149

Number of pOints returned (1-500)? 9

Number of intervals (>= 9, default = 9)? 90

Direct output to one of the following:

(S)creen
(P)rinter
(F) 11 e

Lower limit: 1.00000000000000E+00

Upper limit: 1.00000000000000E+01
Value of Y at 1.0000: 1.00000000000000E+00
Value of Y at 10.0000: 7.69741490000000E+01
NumIntervals: 9.00000000000000E+01

X Y Value Derivative of Y
1.00000000000000E+000 1.00000000000000E+000 1.00000001942594E+000
2.00000000000000E+000 2.61370547174514E+000 2.30685275847028E+000
3.00000000000000E+000 5.70416298088411E+000 3.90138768358927E+000
4.00000000000000E+000 1.04548224122436E+001 5.61370562650429E+000
5.00000000000000E+000 1.69528103026793E+001 7.39056208402440E+000
6.00000000000000E+000 2.52494430584438E+001 9.20824053324639E+000
6.99999999999999E+000 3.53786288412165E+001 1.10540898579729E+001
7.99999999999999E+000 4.73644675641047E+001 1.29205584690303E+001
8.99999999999999E+000 6.12249787166508E+001 1.48027754364805E+001
9.99999999999998E+000 7.69741490000000E+001 1.66974149235206E+001
The exact solution is
y = x * x - x * In(x)
y(l) = 1 y'(l) = 1
y(10) = 7.6974149 y'(10) = 16.6974149
The second approximation is more accurate.

Initial Value and Boundary Value Methods 219

220 Turbo Numerical Methods Toolbox
c H A p T E R 9
Least-Squares Approximation

Given a set of data points, this chapter provides routines to model the data with a
function of a given type. The most common application of this concept is linear
regression.
With linear regression, there is some control variable, say X, and some observed
variable, say Y. X and Yare known or suspected to have some linear relationship,
say
Y=a*X+b
but the parameters a and b are unknown. Usually there is some experimental error
or some other nonlinear influence on Y, so that there are no values of a and b for
which the preceding equation holds exactly. The method of regression provides a
formula for a and b in terms of the values of X and Y such that the error is mini-
mized. The error is the sum of squares of the errors (a * X + b - Y) on each data
point. Except in certain unusual cases, there is exactly one value for a and one
value for b that makes this sum of squares the least possible. This is called the
least-squares solution.
This concept of least squares also applies' when more variables are present-
then it is often called multiple regression. Using this method, you can find the best
model for a given set of data that is linear in a given set of other data sets or
functions. Models that are nonlinear variables can also be treated as long as the
unknown parameters appear linearly.

221
Least-Squares Approximation (LEAST.INC)

Description

This model provides a method for finding a least-squares approximation (Cheney

and Kincaid 1985, 362--387) to a set of data points (x, y). The approximation must
be a linear combination of a set of basis vectors. The functional form of the approxi-
mation (polynomial, logarithmic, and so on) is therefore determined by the user, as
long as it is represented linearly. (How to represent logarithmic, and other func-
tions linearly is discussed later.)
Given a set of m data points (x, y), an m X n matrix (m ;::: n), A, is constructed,
where n is the number of basis vectors in the approximation. The elements of the
matrix are
A[i,j] = V/X j
)

where ~(X) is the jth basis vector evaluated at the data value Xli]. A vector Y is
constructed that contains the y-values of the data points. The coefficients of the
basis vectors that form the least-squares approximation will be the n vector C, such
that the Euclidean norm of(AC - Y) (represented by II AC - Y liz) is a minimum.
This requirement is converted to the requirement that

be a minimum. Here B is an n X n matrix, Z is an n vector, and R is an (m - n)

vector. The equations BC = Z are the normal equations. The previous expression
will be minimized when C solves the equation Be = Z. Gaussian elimination with
partial pivoting (see Chapter 6, "Solving a System of Linear Equations with Gauss-
ian Elimination and Partial Pivoting") is used to solve the normal equations.
The goodness offit is indicated by the standard deviation:
S.D. = ((Y[i] - F(X[i]))z/(m - n))I/Z
where F(X[i]) is the least-squares solution at the point Xli], (Y[i] - F(X[i])) is the
residual, and (m - n) is the degree of freedom of the fit.

222 Turbo Numerical Methods Toolbox

User-Defined Types

TNColumnVector = array[l .. TNColumnSize] of Real;

TNRowVector = array[l .. TNRowSize] of Real;
(TNColumnSize will usually be much larger than TNRowSize.)
TNmatrix = array[l .. TNColumnSize] of TNRowVector;
TNSquareMatrix = array[l .. TNRowSize] of TNRowVector;
TNString40 = str1ng[40];

Input Parameters

NumPoi nts : Integer; Number of data points

XOata : TNCol umnVector; X coordinates of the data points
YOata : TNRowVector; Y coordinates of the data points
NumTerms : Integer; Number of terms in the least-squares approximation
The preceding parameters must satisfy the following conditions:
1. NumPoints > 1.
2. NumTerms:5 NumPoints.
3. NumPoints:5 TNColumnSize.
4. NumTerms:5 TNRowSize.
5. The XData points cannot all be identical.

TNColumnSize and TNRowSize set an upper bound on the number of elements

in a vector. Neither of these identifiers are variable names and are never referenced
by the procedure. If conditions 3 or 4 are violated, the program will crash with an
Index Out of Range error (assuming the directive {$R +} is active).

Least-Squares Approximation 223

Output Parameters

So 1ut ion: TNRowVector; Coefficients of the basis vectors that form the least-
squares approximation
YFit : TNColumnVector; Values of the least-squares fit evaluated at the XData
values
Res i dua 1s : TN Co 1umnVector; Difference between YData and YFit values
StandardDevi ation : Real; Square root of the variance-indicates the goodness of fit
Error: Byte; 0: No error
1: NumPoints < 2
2: NumTerms < 1
3: NumTerms > NumPoints
4: Least-squares solution does not exist (see "Com-
ments")

Syntax of the Procedure Call

LeastSquares(NumPoints, XData, YData, NumTerms, Solution,

YFit, Residuals, StandardDeviation, Error);

Comments

The least-squares routine is kept in two modules (include files). One is called
LEAST.lNC and must always be included in your top-level program. The choice of
the second module will depend upon the functional form (basis vectors) to which
you fit the data. Following are the five basis modules included in this package:

POLY.LSQ
This module uses Chebyshev polynomials to fit a polynomial to the data points.
NumTerms must be one greater than the degree of the polynomial (for example, to
fit a fourth-degree polynomial, input NumTerms = 5). To get a straight-line least-
squares fit, use this module and fit a curve with only two coefficients. The elements
of the Solution vector will be as follows:
Solution[j] = aJ 1 s j S NumTerms
where a. is the coefficient of Xl-I.
}

224 Turbo Numerical Methods Toolbox

FOURIER.LSQ
This module will fit a finite Fourier series to the data points. The number of terms
in the approximation will be NumTerms. The elements of the Solution vector will
be as follows:
Solution[j] = Fj _ 1 1 :5 j :5 NumTerms
where Fj _ 1 is the (j - l)th term in the Fourier series. Following are the first few
terms in the Fourier series:
F[O] = 1
F[l] = cos{x)
F[2] = sin{x)
F[3] = cos{2x)
F[4] = sin{2x)
F[5] = cos{3x)
F[6] = sin{3x)

POWER.LSQ
This module will fit the function
y = al
where a and b are real numbers to the data points. A linear equation is obtained by
taking the log of both sides, like so:
In{y) = In{a) + b * In{x)
and expanding on basis vectors 1 and In{x). The x-values of the data points must all
be positive, and the y-values of the data points must all have the same sign. The
number of coefficients in the approximation will be two regardless of the value of
NumTerms (unless NumTerms > NumPoints, in which case Error 3 will occur).
The elements of the Solution vector will be as follows:
Solution[l] =a
Solution[2] =b

EXP.LSQ
This module will fit the function
y = aebx
where a and b are real numbers to the data points. A linear equation is obtained by
taking the log of both sides, like so:
In{y) = In{a) + bx

Least-Squares Approximation 225

and expanding on basis vectors 1 and. x. The y-values of the data points must all
have the same sign. The number of coefficients in the approximation will be two
regardless of the value of NumTerms (unless NumTerms > NumPoints, in which
case Error 3 will occur). The elements of the Solution vector will be as follows:
Solution[l] =a
Solution[2] =b

WG.LSQ
This module will fit the function
y = a In(bx)
where a and b are real numbers to the data points. A linear equation is obtained by
rewriting the equation:
y = a In(b) + a In(x)
and expanding on basis vectors 1 and In(x). The x-values of the data points must all
have the same sign. The number of coefficients in the approximation will be two
regardless of the value of NumTerms (unless NumTerms > NumPoints, in which
case Error 3 will occur). The elements of the Solution vector will be as follows:
Solution[l] =a
Solution[2] =b

USERLSQ
This module is included if you need a least-squares approximation on a set of basis
vectors different from the ones listed earlier. This module allows you to create your
own set of basis vectors. The source code contains detailed instructions of how to
flesh out the skeleton contained in USER.LSQ.
A least-squares solution may not exist for some input data and choice of basis
vectors (Error 4). The reasons for this will depend on the module you are using. For
example, it is impossible to fit an exponential function (module EXP.LSQ) to data
with y-values of differing signs; Error 4 will occur if you try. The same data could
be fit with a polynomial and no error would result. Error 4 will also occur if all the
x-values of the data are identical.

226 Turbo Numerical Methods Toolbox

Sample Program

The demonstration program LEAST.PAS contains I/O routines that allow you to
run the least-squares approximation routine. Note that there are two include com-
mands:
{$I POLY.LSQ} (* load the basis vectors *)
{$I LEAST. INC} (* load procedure LeastSquares *)
The LEAST.lNC file must always be included after the basis module. To change
the basis vectors of the approximation, simply load a different basis module with
the first INCLUDE command.

Input Files
Data may be entered from a text file. The x- and y-coordinates should be separated
by a space and followed by a carriage return. For example, data values of sqr(x)
could be entered in a text file as
11
24
39
416
525

Example
Problem. Given the following data (contained in the file SAMPLE9A.DAT), fit a
fourth-degree polynomial and a logarithmic function to the data:
O.oooooooOOOOOOOE+OO 1.33830225764886E-03
0.10000000000000E+00 4.43184841193803E-02
0.20000000000000E+00 5.39909665131879E-01
0.30000000000000E+00 2.41970724519143E+00
0.40000000000000E+00 3.98942280401433E+00
0.02000000000000E+00 2.91946925791461E-03
0.04000000000000E+00 6.11901930113775E-03
0.06000000000000E+00 1.23221916847303E-02
0.08000000000000E+00 2. 38408820146486E-02
0.12000000000000E+00 7.91545158298001E-02
0.14000000000000E+00 1.35829692336855E-01
0.16000000000000E+00 2.23945302948430E-01
0.18000000000000E+00 3.54745928462313E-01
0.22000000000000E+00 7.89501583008939E-01
0.24000000000000E+00 1.10920834679455E+00
0.26000000000000E+00 1.49727465635745E+00
0.28000000000000E+00 1.94186054983213E+00
0.32000000000000E+00 2.89691552761483E+00
0.34000000000000E+00 3.33224602891800E+00
0.36000000000000E+00 3.68270140303323E+00
0.38000000000000E+00 3.91042693975456E+00

Least-Squares Approximation 227

(The function is the left-hand side of a Gaussian distribution curve with mean =
0.5 and standard deviation = 0.1.) Note that the points do not have to be in any
particular order.
First fit the polynomial; include the proper include files in the LEAST. PAS pro-
gram.
{$I POLY.LSQ} (* load the basis vectors *)
{$I LEAST. INC} (* load procedure LeastSquares *)
Run LEAST.PAS:
(K)eyboard or (F)ile entry of data? F
File name? SAMPLE9A.DAT

Number of terms in the least squares fit «= 21)? 5

Direct output to one of the following:
(S)creen
(P)rinter
(F) il e

The Data Points:

X Y
0.200 0.0443185
0.300 0.5399097
0.400 2.4197072
0.500 3.9894228
0.100 0.0013383
0.120 0.0029195
0.140 0.0061190
0.160 0.0123222
0.180 0.0238409
0.220 0.0791545
0.240 0.1358297
0.260 0.2239453
0.280 0.3547459
0.320 0.7895016
0.340 1.1092083
0.360 1.4972747
0.380 1.9418605
0.420 2.8969155
0.440 3.3322460
0.460 3.6827014
0.480 3.9104269

*----------------------------------------*
Polynomial Least Squares Fit
*----------------------------------------*

228 Turbo Numerical Methods Toolbox

Coefficients in least squares approximation:
Coefficient 0: -3.1905595419E+00
Coefficient 1: 6.4048009604E+Ol
Coefficient 2: -4.3900537685E+02
Coefficient 3: 1.2058567475E+03
Coefficient 4: -1.0523352671E+03
X Least Squares Fit Residual
0.2000 2.1944857683E-02 -2.2373626436E-02
0.3000 5.4757594258E-Ol 7.6662774545E-03
0.4000 2.4228330082E+00 3.1257630399E-03
0.5000 . 4.0432402964E+00 5.3817492408E-02
0.1000 -7.5189129229E-02 -7.6527431486E-02
0.1200 3.9032402623E-02 3.6112933365E-02
0.1400 7.6262215339E-02 7.0143196038E-02
0.1600 6.8115144530E-02 5.5792952845E-02
0.1800 4.2165058391E-02 1.8324176377E-02
0.2200 2.6946475745E-02 -5.2208040084E-02
0.2400 7.2620878494E-02 -6.3208813842E-02
0.2600 1.7037806441E-Ol -5.3567238538E-02
0.2800 3.2758706456E-Ol -2.7158863898E-02
0.3200 8.2963179468E-Ol 4.0130211675E-02
0.3400 1. 1690007497E +00 5.9792402863E-02
0.3600 1.5568879689E+00 5.9613312500E-02
0.3800 1.9804576462E+00 3.8597096373E-02
0.4200 2.8630963140E+00 -3.3819213604E-02
0.4400 3.2762888552E+00 -5.5957173721E-02
0.4600 3.6334109560E+00 -4.9290447012E-02
0.4800 3.9014219733E+00 -9.0049664504E-03
Standard Deviation: 5.381534E-02
The fourth-degree polynomial that best fits this data is as follows:
y = -1052.34 X4 + 1205.86l - 439.005 x2 + 64.0480 x - 3.19056
Note that a fourth-degree polynomial requires five terms in the fit.
Now fit the logarithmic function; include the proper include files in the LEAST.
PAS program.
{$I LOG.LSQ} (* load the basis vectors *)
{$I LEAST.INC} (* load procedure LeastSquares *)

Run LEAST. PAS:

{K)eyboard of {F)ile entry of data? F
File name? SAMPLE9A.DAT
Number of terms in the least squares fit (<= 21)? 2

Least-Squares Approximation 229

Direct output to one of the following:
(S)creen
(P)rinter
(F) il e

The Data Points:

X y
0.200 0.0443185
0.300 0.5399097
0.400 2.4197072
0.500 3.9894228
0.100 0.0013383
0.120 0.0029195
0.140 0.0061190
0.160 0.0123222
0.180 0.0238409
0.220 0.0791545
0.240 0.1358297
0.260 0.2239453
0.280 0.3547459
0.320 0.7895016
0.340 1.1092083
0.360 1.4972747
0.380 1. 9418605
0.420 2.8969155
0.440 3.3322460
0.460 3.6827014
0.480 3.9104269

*----------------------------------------*
Logarithmic Least Squares Fit
*----------------------------------------*

230 Turbo Numerical Methods Toolbox

Coefficients in least squares approximation:
Coefficient 0: 2.5984092387E+00
Coefficient 1: 6.0253489685E+00

X Least Squares Fit Residual

0.2000 4.8470072529E-01 4.4038224117E-01
0.3000 1.5382650082E+00 9.9835534304E-01
0.4000 2.2857807630E+00 -1.3392648216E-01
0.5000 2.8655990283E+00 -1.1238237757E+00
0.1000 -1.3163793125E+00 -1.3177176147E+00
0.1200 -8.4263329487E-01 -8.4555276412E-01
0.1400 -4.4208674425E-01 -4.4820576355E-01
0.1600 -9.5117540004E-02 -1.0743973169E-01
0.1800 2.1093098801E-01 1.8709010600E-01
0.2200 7.3235557703E-01 6.5320106120E-01
0.2400 9.5844674288E-01 8.2261705054E-01
0.2600 1. 1664304540E+00 9.4248515104E-01
0.2800 1.3589932935E+00 1.0042473650E+00
0.3200 1.7059624977E+00 9.1646091473E-01
0.3400 1.8634900752E+00 7.5428172837E-01
0.3600 2.0120110257E+00 5.1473636940E-01
0.3800 2.1524997930E+00 2.1063924317E-01
0.4200 2.4125575764E+00 -4.8435795124E-01
0.4400 2.5334356148E+00 -7.9881041414E-01
0.4600 2.6489394853E+00 -1.0337619177E+00
0.4800 2.7595267806E+00 -1.1509001591E+00
Standard Deviation : 8.320742E-01
The logarithmic function that bests fits this data is as follows:
y = 2.59841 * In(6.02535x)
The standard deviation of the polynomial fit is much smaller than that of the
logarithmic fit; a fourth-degree polynomial fits this data much better than a loga-
rithmic function.

Least-Squares Approximation 231

232 Turbo Numerical Methods Toolbox
c H A p T E R 10
Fast Fourier Transform Routines

Fourier transforms are used to analyze periodic phenomena such as waves. A con-
tinuous functionfthat has period 2'TT (= 2 * 3.14159265...); that is, satisfies
f(x + 2'TT) = f(x)
for all x, can be decomposed into sines and cosines:
f(x) = a[O] + a[l] * cos(x) + b[l] * sin(x) + a[2] * cos(2x)
+ b[2] * sin(2x) + ...
This is an infinite series where the coefficients get closer and closer to zero. The
routines in this chapter can be used to calculate the coefficients.
The Fast Fourier Transform (FFT) is a particular algorithm for computing Fou-
rier transforms efficiently.
This chapter includes two kinds of units. One group consists of four variations of
the FFT method of calculating discrete Fourier transforms (FFTB2.INC, FFT
B4.INC, FFT87B2.INC, FFT87B4.INC), each optimized for certain conditions.
All are variations of the original Cooley-Tukey method. The second group consists
of six applications (COMPFFT.INC, REALFFT.INC, COMPCNVL.lNC, REAL
CNVL.INC, COMPCORR.INC, REALCORR.INC) that can each be used with
any of the FFT methods. You can select the FFT method most appropriate to the
circumstances and combine it with the appropriate application or integrate it into
another program (Brigham 1974; Nussbaumer 1982).
In each FFT unit the procedure calls have exactly the same form (although there
are different restrictions on the data) so that anyone FFT unit can be combined

233
with any of the application units without rewriting code. Each of these algorithms
will compute either a forward or an inverse transform.
Each unit contains two procedures needed to prepare for the FFf calculation:
procedure Testlnput and procedure MakeSinCosTable. TestInput examines the
input data to ensure that it satisfies certain conditions (for example, that there is
more than 1 data point). MakeSinCosTable precalculates a table of the nth roots of
unity for look up in the FFf calculation.
FFfB2.1NC contains a procedure that implements the Cooley-Tukey powers-of
two (radix 2 or base 2) Fast Fourier Transforms. It is optimized to reduce the
number of real multiplications by taking advantage of the symmetries of certain
roots of unity and by using a complex multiplication that requires three real multi-
plications and three real additions. This algorithm is appropriate when the time for
real multiplications is significantly greater than the time for real additions; for
example, when running on an 8088 machine with no numeric coprocessor.
FFf87B2.1NC implements the same algorithm as FFTB2.1NC. The difference
is that the complex multiplications are done with four real multiplications and two
real additions. By using this standard form of complex multiplication, storage over-
head and assignment statements are reduced. This algorithm is appropriate when
the time for a real multiplication is close to the time for a real addition; for example,
when running an 8088 machine with an 8087 numeric coprocessor, or an 80286
machine with an 80287 numeric coprocessor.
Standard Turbo Pascal uses 6-byte reals; Turbo-87 Pascal (which utilizes the
8087 coprocessor) uses 8-byte reals. Consequently, given the same amount of
memory, more data points can be manipulated in Turbo Pascal than in Turbo-87
Pascal. Both FFfB2.INC and FFf87B2.1NC require the number of data points to
be a power of two up to a maximum of 4,096 points when in Turbo-87, or 8,192
points when in standard Turbo Pascal.
FFfB4.1NC and FFf87B4.1NC implement powers-offour (radix 4 or base 4)
Fast Fourier Transforms. The powers-of-four method is the same as the Cooley-
Tukey algorithm except at each stage of reduction a given transform is converted
into four transforms each with one fourth the data points of its predecessor (Nus-
sbaumer 1982). When this algorithm is optimized, there are about 25 percent
fewer multiplications and slightly fewer additions than in a radix-2 algorithm. The
algorithm has the disadvantage of only being applicable to data sets where the
number of points is a power of four up to a maximum of 4,096 points whether in
Turbo-87 or standard Turbo. A reduction in execution time of about 20 percent is
accomplished when FFfB4.1NC or FFf87B4.1NC is used over its B2 counterpart.

234 Turbo Numerical Methods Toolbox

FFTB4.INC performs complex multiplication with three real multiplications
and three real additions and thus is most appropriate when multiplications take
much more time than additions.
FFT87B4.INC performs complex multiplication with four real multiplications
and two additions and thus is most appropriate with a numeric coprocessor.
Table 10-1 shows the recommended use of the four FFT algorithms for optimal
performance.

Table 10-1 Four Fast Fourier TransJorms

Number of Points Without Coprocessor With Coprocessor
FFfB2.1NC FFf&7B2.1NC
FFfB4.1NC FFf87B4.1NC

Although each of the algorithms is most efficient under different sets of

circumstances, all four FFT algorithms will work whether you have a math
coprocessor or not. The sample program defaults to the Turbo-87 radix-2 algorithm
(FFT87B2.IN C).

The Application Programs

There are six application programs that use the basic FFT routines
contained in the previously mentioned include files (COMPFFT.INC,
REALFFT.lNC, COMPCNVL.lNC, REALCNVL.INC, COMPCORR.INC, and
REALCORR.INC).
Fast Fourier Transforms are particularly useful for analyzing periodic signals.
Such a signal is represented by a function J satisfying
J(t + T) = J(t)
where t is time and T is the period. Under mild hypotheses,J can be expanded into
a Fourier series such as the following:
oc

J(t) = N- 112 I
11= - oc
F(n) exp (2'Ti' i n tiT)

where i is the square root of - 1. The term exp (2'Ti' i n tiT) is a sinusoid of period
Tin and frequency niT, and its coefficient F(n) gives the strength of that frequency
component in the original signal.

Fast Fourier Transform Routines 235

To analyze a signal on a digital computer, the signal must be discretized. Let x(n)
be computed by discretizing the function f at N equidistant points in one period.
Thus, let
x(n) = f(nTIN) n = 0,1, ... N - 1
Once we restrict attention to N points; it only makes sense to represent the
signal in terms of N of the functions
exp (2'TT i n tiT)
since the rest are redundant. For example:
exp (2'TT i ( -1) tiT) = exp (2'TT i (N -1) tiT)
for t = nTIN, n = 0,1, ... N - 1. The Fourier series for the signal is then a finite
sum, and has the form
N-I
x(n) = N- 1/2 I X(k) exp (2'TT i k nlN)
k'" 0

(The factor of N- I12 is a matter of convention. Some books do not include it in this
formula, resulting in a factor of liN in the formula for X that follows.)
The formula for the coefficients X(k) is as follows:

I
N-I
1/2
X(k) = N- x(n) exp (-i 2'TT n kiN)
n'" 0

This formula for X makes sense for any integer k. X is then periodic, satisfying
X(k + N) = X(k)
for all k. In formulas and programs, it is more convenient to let k run from t9 °
N - 1, but for analyzing signals it makes more sense to think of k as running from
(- N/2) to (N/2 - 1). This is because values of k near zero represent the low
frequency information, and values of k near or greater than N12 represent frequen-
cies that are so high that the discretization is too coarse to realize them accurately
anyway. Therefore, if k is between N12 and N, X(k) should be thought of as the
coefficient of
exp (2'TT i (k - N) tiT)
rather than
exp (2'TT i k tiT)
In other words, negative frequencies are represented on the right half of the
transform.

236 Turbo Numerical Methods Toolbox

COMPFFT.lNC simply takes the complex Fast Fourier Transform of a set of
complex data points. The complex Fourier transform is defined as
N-I
Xf = N- I12 L xn exp (- 21T if n/N) f = O.. N - 1
,,= 0
where i is the square root of - 1. The inverse Fourier transform (which may also be
calculated with COMPFFT.lNC) is defined as
N-I
xn = N- 1I2
L~ f= 0
exp ( - 21T if n/N) n = O..N - 1

where the bar stands for complex conjugation.

REALFFT.lNC provides a procedure that is optimized for a discrete Fourier
transform with all real data. It proceeds by mapping the N real data points onto
N /2 complex points, applying one of the FFT routines, then reconstructing the N
points of the desired transform. This reduces the computation time by about 25
percent compared to applying the complex FFT routine to the N real data points.
REALFFT.lNC can be used with any of the given FFT methods, but note that if a
radix-4 method is used (FFTB4.1NC or FFT87B4.1NC), N/2 must be a power of
four; so N must be of the form 2 * 4k.
COMPCNVL.lNC provides a procedure for calculating convolutions of two
complex vectors (Brigham 1974; Nussbaumer 1982). The discrete. ~onvolution of
two complex functions x and h is defined by

m = 0,1, ... N - 1
,,= 0
where subscripts are taken modulo N (circular convolution). The basic theorem that
allows us to calculate these effectively using FFTs is shown in the following:
y In
= X HIn In
m = 0,1, ... N - 1
where capital letters indicate the transforms of the functions represented by lower-
case letters. Thus the procedure 'for convolution works like this:
1. Transform both given data sets using FFTs.
2. Multiply the resulting transforms.point by point.
3. Find the inverse transform of this product using FFTs.

Fast Fourier Transform Routines 237

REALCNVL.INC provides a procedure for calculating convolutions of two real
vectors (Brigham 1984; Nussbaumer 1982). This procedure is exactly the same
as the previous procedure (COMPCNVL.INC) for complex convolution except that
only one forward Fourier transform need be performed. The procedure is as
follows:
1. Given two real vectors XReal and HReal, combine them into a complex vec-
tor XReal plus iHReal, where i is the square root of - 1.
2. Transform this complex vector.
3. Extract the transforms of the two real functions from the transform of the
complex function (using the symmetry Xf = X_ f , where the bar stands for
complex conjugation).
4. Multiply the resulting transforms point by point.
5. Find the inverse transform of this product using FFTs. REALCNVL.INC
is about 25 percent faster than its complex counterpart for the same set of
real data.

COMPCORR.INC provides a procedure for calculating the crosscorrelation of

two discrete complex functions or the autocorrelation of one discrete complex func-
tion (Brigham 1974). If x and h are the given discrete functions, then their correla-
tion is defined as
N - 1 .

Cm =L
" xn hn+m m = 0,1, ... N - 1
n = 0

where subscripts are taken modulo N (circular convolution). This can be computed
using FFTs with a method analogous to that used in COMPCNVL.INC:
c m
= Xm H N-m m = 0,1, ... N - 1
Commonly x and h are real functions; in which case the preceding formula
reduces to C = X Ii , where the bar stands for complex conjugation. Thus the
procedure fo; correGtio~ works like this:
1. Transform both given data sets using FFTs.

2. Multiply each element of the transform of the first data set by the appropri-
ate element of the transform of the second data.
3. Find the inverse transform of this product using FFTs.

REALCORR.INC provides a procedure for calculating the crosscorrelation of

two discrete real functions or the autocorrelation of one discrete real function
(Brigham 1974). This procedure is exactly the same as the previous procedure for

238 Turbo Numerical Methods Toolbox

complex correlation except that only one forward Fourier transform need be per-
formed. The procedure is as follows:
1. Given two real vectors XReal and HReal, combine them into a complex vec-
tor XReal + iHReal, where i is the square root of - 1.
2. Transform this complex vector.
3. Extract the transforms of the two real vectors from the transform of the
complex vector (using the symmetry Xf = X_ f , where the bar stands for
complex conjugation). .
4. Multiply each element of the transform of the first data set by the appropri-
ate element of the transform of the second data.
5. Find the inverse transform of this product using FFTs.

Anyone of the FFT include files can be used with any of the applications.

Data Sampling

While sampling theory is beyond the scope of this Toolbox, we would like to men-
tion several common problems associated with data sampling (Brigham 1974; Press
et al. 1986, Ch.12). The most common frustration is aliasing. A Fourier transform
only represents frequencies up to a certain limit (called the Nyquist limit, or
Nyquist frequency), which is given by half the sampling rate. (For example, if a
signal is sampled sixty times a second, the Nyquist frequency will be 30 Hz.) 'A
sample containing frequencies greater than this limit will not be properly trans-
formed. The high frequencies will falsely contribute to the transform. This contri-
bution will be indistinguishable from a contribution of a frequency below the
Nyquist frequency.
There are several ways to combat aliasing. Increasing the sampling rate will
increase the Nyquist frequency and thus reduce aliasing effects. It is also possible
to pass the signal through a low pass filter, thus eliminating the high frequencies
before sampling. If the Fourier transform of a signal does not converge to zero at
the Nyquist frequency, the transform has very likely been aliased.
The Fourier transform assumes that the sample represents a periodic function
and that the sample is an integer multiple of one period. If the latter condition is
not true, spurious frequencies will show up in the transform. For example, if a sine
wave is sampled from 0 to 270 degrees (instead of the full period), a sharp bound-
ary is created because the sine of 0 does not equal the sine of 270. High frequen-
cies will be introduced into the transform to account for that sharp boundary.

Fast Fourier Transform Routines 239

The assumption of periodicity can cause problems when convolving or correlat-
. ing two signals that are not periodic. The convolution of each point in a signal
affects the points surrounding it (the nature and extent of the affect depends on the
particular convolving function). The assumption of periodicity means that the con-
volution at one end of the signal will affect the other end of the signal. This "end
effect" can be eliminated by padding the data (on either end) with a sufficient
number of zeros.

User-Defined Types

TNvector = array[O .. TNArraySize] of Real;

TNvectorPtr = ~TNvector;

These user-defined types are different from others in this Toolbox, because they
involve pointers. Pointers are used to transcend the limitations imposed by the 64K
data segment size of Turbo Pascal. One array of 8,000 elements uses the entire data
segment (in Turbo-87). However, it is possible to store these arrays on the heap,
and to point to them with pointers that only require 4 bytes. The size of the heap
(and hence the maximum size and number of TNvectors) is determined by the
amount of memory in the machine.

240 Turbo Numerical Methods Toolbox

Fast Fourier Transform Algorithms

The following documentation generally applies to all four FFT algorithms

(FFTB2.INC, FFTB4.INC, FFT87B2.INC, FFT87B4.INC). When a difference
between the radix-2 and radix-4 algorithms needs to be described, the radix-4
information will be placed in brackets following the radix-2 information (for exam-
ple, the number of points must be a power of two [four]). The following describes
the three procedures contained in each of the include files.

Procedure TestInput

Description
This example determines the number of data points in terms of a power of two
[four]. If the number of data points is not a power of two [four], then an error is
returned.

Input Parameters
NumPoi nts : Integer; Number of data points

The preceding parameter must satisfy the following conditions:

1. NumPoints ~ 2.
2. NumPoints must be a power of two [four].

Output Parameters
NumberOfBits : Byte; Number of data points as a power of two [four]
Error: Byte; 0: No errors
1: NumPoints < 2
2: NumPoints not a power of two [four]

Syntax of the Procedure Call

TestInput(NumPoints, NumberOfBits, Error);

Fast Fourier Transform Routines 241

Procedure MakeSinCosTable

Description
This example creates a look-up table ofNumPoints/2 [3/4 NumPoints] roots of unity.
The roots of unity are defined as follows:
Rootn = exp ( - i 2'TT n/NumPoints n = O..NumPoints/2 [3/4 NumPoints]
where i is the square root of -1. These values are stored in two tables: SinTable,
containing the imaginary parts of the roots of unity, and CosTable, containing the
real parts of the roots of unity. It is faster to look up these values in a table than to
calculate them in the FFf procedure.

Input Pararneters
NumPoints : Integer; Number of data points

The preceding parameter must satisfy the following conditions:

1. NumPoints ~ 2.
2. NumPoints must be a power of two [four].

Output Pararneters
Si nTabl e : TNvectorPtr; Table of sine values
CosTabl e : TNvectorPtr; Table of cosine values

Syntax of the Procedure Call

MakeSinCosTable(NumPoints, SinTable, CosTable);

Procedure FFT

Description
This example implements the particular variation of the Cooley-Tukey algorithm.
The Fast Fourier Transform of the data XReal, Xlmag is made in place and is thus
returned in the vectors XReal, Xlmag. The inverse transform of the data can also
be calculated with this procedure.

242 Turbo Numerical Methods Toolbox

It is essential that procedures TestInput and MakeSinCosTable be called before
procedure FFT is called. TestInput will flag any errors in the data (for example,
number of points that are not a power of two [four]), and MakeSinCosTable
generates a table of sine and cosine values referenced by FFT. TestInput and
MakeSinCosTable need only be called once, even if several calls to FFT are made
within the same program (for example, when computing the discrete convolution),
as long as the number of data points is unchanged. If the number of data points
changes between two calls of FFT, TestInput and MakeSinCosTable must be called
again. (Interested readers are urged to consult the references given in the begin-
ning of the chapter for details about the Cooley-Tukey algorithm.)

Input Parameters
NumberOfBi ts : Byte; Number of data points as a power of two [four]
NumPoi nts : Integer; Number of data points
Inverse: Boolean; FALSE equals forward transform; TRUE equals inverse
transform
XRea 1 ': TNvectorPt r; Pointer to real values of the data points
XImag : TNvectorPtr; Pointer to imaginary values of the data points
SinTable: TNvectorPtr; Table of sine values
CosTable : TNvectorptr: Table of cosine values

The preceding parameters must satisfy the following conditions:

1. NumPoints;::: 2.
2. NumPoints must be a power of two [four].

Output Parameters
XReal : TNvectorPtr; Pointer to real values of the discrete Fourier transform of the
input data
XImag : TNvectorPtr; Pointer to imaginary values of the discrete Fourier transform
of the input data

Syntax of the Procedure Call

FFT(NumberOfBits, NumPoints, Inverse, XReal, XImag, SinTable, CosTable);

Fast Fourier Transform Routines 243

Fast Fourier Transform Applications

Each of the six application programs calls the three procedures contained within
FFT algorithm files.

COMPFFT.INC

Description
This example is the most basic application, performing a complex Fast Fourier
Transform. It simply calls Testlnput, MakeSinCosTable, and FFT sequentially; thus
accomplishing an in-place transformation of the complex data XReal, Xlmag.

Input Parameters
NumPoints : Integer; Number of data points
Inverse: Boolean; FALSE equals forward transform; TRUE equals inverse trans-
form
XReal : TNvectorPtr; Pointer to real values of the data points
XImag : TNvectorPtr; Pointer to imaginary values of the data points

The preceding parameters must satisfy the following conditions:

1. NumPoints ~ 2.
2. NumPoints must be a power of two [four].

Output Parameters
XReal : TNvectorptr; Pointer to real values of the discrete Fourier transform of the
input data
XImag : TNvectorPtr; Pointer to imaginary values of the discrete Fourier transform
of the input data
Error: Byte; 0: No errors
1: NumPoints < 2
2: NumPoints not a power of two [four]

244 Turbo Numerical Methods Toolbox

Syntax of the Procedure Call
Comp1exFFT(NumPoints, Inverse, XRea1, XImag, Error);

Comments
The complex Fast Fourier method uses the New/Dispose procedures to manipulate
the heap and should not be used in any program that uses Mark/Release to manip-
ulate the heap.

REALFFT.INC

Description
This example performs a complex Fast Fourier Transform of real data. The Num-
Points real data points are first mapped onto NumPoints/2 complex data points. A
complex Fast Fourier Transform of these complex points is performed by calling
TestInput, Make Sin Cos Table , and FFT. The NumPoints/2 transform is then mapped
onto NumPoints complex points. The real part of the transformation will be even,
and the imaginary part of the transformation will be odd. If you are implementing
this application with a radix-4 algorithm, be sure that the number of real data
points (NumPoints) is twice the power of four.

Input Parameters
NumPoints : Integer; Number of data points
Inverse: Boo1 ean; FALSE equals forward transform; TRUE equals inverse trans-
form
XReal : TNvectorPtr; Pointer to real values of the data points

The preceding parameters must satisfy the following conditions:

1. NumPoints ~ 4.
2. NumPoints must be a power of two (twice a power of four for a radix-4
algorithm).

At least four data points are required, because this algorithm transforms the real
vector to a complex vector half the size. If only two real data points were entered,
the routine would have to take the transform of a single complex point.

Fast Fourier Transform Routines 245

Output Parameters
XRea 1 : TNvectorPt r; Pointer to real values of the Fourier transform of the input data
XImag : TNvectorPtr; Pointer to imaginary values of the Fourier transform of the
input data
Error: Byte; 0: No errors
1: NumPoints < 4
2: NumPoints not a power of two [twice a power of four]

Syntax of the Procedure Call

RealFFT{NumPoints, Inverse, XReal, XImag, Error);

Comments
This method uses the New/Dispose procedures to manipulate the heap and should
not be used in any program that uses Mark/Release to manipulate the heap.

COMPCNVL.INC

Description
The calculation of the convolution of two complex vectors is facilitated with a Fast
Fourier Transform routine. The discrete convolution of two functions x and h is
defined by
N-I
Ym = IX
1\ = 0
n hm _ n m = 0,1, ... N - 1

where subscripts are taken modulo N (circular convolution). The basic theorem
that allows us to calculate these effectively using FFTs is as follows:
m = 0,1, ... N - 1

246 Turbo Numerical Methods Toolbox

where capital letters indicate the transforms of the functions represented by lower-
case letters. Thus the procedure for convolution works like this:
1. Transform both given data sets using FFTs.
2. Multiply the resulting transforms point by point.
3. Find the inverse transform of this product using FFTs.

Input Parameters
NumPoints : Integer; Number of data points
XReal : TNvectorPtr; Pointer to real values of the first set of data points
XImag : TNvectorptr; Pointer to imaginary values of the first set of data points
HReal : TNvectorptr; Pointer to real values of the second set of data points
HImag : TNvectorptr; Pointer to imaginary values of the second set of data points

The preceding parameters must satisfy the following conditions:

1. NumPoints ~ 2.
2. NumPoints must be a power of two [four].

Output Parameters
XReal : TNvectorptr; Pointer to real values of the convolution of XReal, Xlmag and
HReal, Hlmag
XImag : TNvectorptr; Pointer to imaginary values of the convolution of XReal, Xlmag
and HReal, Hlmag
Error: Byte; 0: No errors
1: NumPoints < 2
2: NumPoints not a power of two [four]

Syntax of the Procedure Call

ComplexConvolution(NumPoints, XReal, XImag, HReal, HImag, Error);

Comments
This method uses the New/Dispose procedures to manipulate the heap and should
not be used in any program that uses Mark/Release to manipulate the heap.

Fast Fourier Transform Routines 247

REALCNVL.INC

Description
The calculation of the convolution of two real vectors is facilitated with a Fast
Fourier Transform routine. This procedure is exactly the same as the previous
procedure for complex convolution (COMPCNVL.lNC) except that only one
Fourier transform need be performed. The procedure is as follows:
1. Given two real vectors XReal and HReal, combine them into a complex vec-
tor XReal + iHReal, where i is the square root of -1.
2. Transform this complex vector.
3. Extract the transforms of the two real functions from the transform of the
complex function (using the symmetry Xf = X_ f , where the bar stands for
complex conjugation).
4. Multiply the resulting transforms point by point.
5. Find the inverse transform of this product using FFTs. REALCNVL.lNC is
about 25 percent faster than its complex counterpart for the same set of real
data.

Input Parameters
NumPoints : Integer; Number of data points
XReal : TNvectorPtr; Pointer to real values of the first set of data points
HReal : TNvectorptr; Pointer to real values of the second set of data points

The preceding parameters must satisfy the following conditions:

1. NumPoints ~ 2.
2. NumPoints must be a power of two [four]'

Output Parameters
XReal : TNvectorPtr; Pointer to real values of the convolution of XReal and HReal
XImag : TNvectorPtr; Pointer to imaginary values of the convolution of XReal and
HReal
Error: Byte; 0: No errors
1: NumPoints < 2
2: NumPoints not a power of two [four]

248 Turbo Numerical Methods Toolbox

Syntax of the Procedure Call
RealConvolution(NumPoints, XReal, Xlmag, HReal, Error);

Comments
This method uses the New/Dispose procedures to manipulate the heap and should
not be used in any program that uses Mark/Release to manipulate the heap.

COMPCORR.INC

Description
The calculation of the correlation of two complex vectors is facilitated with a Fast
Fourier Transform routine. The discrete correlation of two complex functions x and
h is defined by
N-l
Ym = "'xh m+n
L n
m = 0,1, ... N - 1
11 = 0

where subscripts are taken modulo N (circular correlation). The basic theorem that
allows us to calculate these effectively using FFTs is as follows:
ym =X m
H N-m m = 0,1, ... N - 1
where capital letters indicate the transforms of the functions represented by lower-
case letters and - indicates the complex conjugate. (Commonly x and h are real
functions, in which case the preceding formula reduces to the more familiar
expression Grn = Xm H,1l' where the bar stands for complex conjugation. (See REAL
CORR.INC for a real version of correlation.) Thus the procedure for correlation
works like this:
1. Transform both given data sets using FFTs.
2. Multiply each element of the transform of the first data set by the appropri-
ate element of the transform of the second data.
3. Find the inverse transform of this product using FFTs.

If the functions x and h are different, the correlation is called crosscorrelation; if

the functions x and h are the same, the correlation is called autocorrelation.

Fast Fourier Transform Routines 249

Input Parameters
NumPo; nts : Integer; Number of data points
Auto : Boolean; FALSE equals crosscorrelation; TRUE equals autocorrelation
XReal : TNvectorPtr; Pointer to real values of the first set of data points
XImag : TNvectorptr; Pointer to imaginary values of the first set of data points
HReal : TNvectorPtr; Pointer to real values of the second set of data points (for cross-
correlation)
HImag : TNvectorPtr; Pointer to imaginary values of the second set of data points (for
crosscorrelation)
The preceding parameters must satisfy the following conditions:
1. NumPoints;;::: 2.

2. NumPoints must be a power of two [four].

Output Parameters
XReal : TNvectorptr; Pointer to real values of the correlation of XReal, Xlmag and
HReal, Hlmag (or the autocorrelation of XReal, Xlmag if
Auto = TRUE)
Xlmag : TNvectorPtr; Pointer to imaginary values of the correlation of XReal, Xlmag
and HReal, Hlmag (or the autocorrelation of XReal, Xlmag if
Auto = TRUE)
Error: Byte; 0: No errors
1: NumPoints < 2
2: NumPoints not a power of two [four]

Syntax of the Procedure Call

ComplexCorrelation(NumPoints. Auto. XReal. Xlmag. HReal. HImag. Error):

Comments
If you are performing an autocorrelation of the vector XReal, Xlmag, then set
Auto = TRUE. In this case, the vector HReal, Hlmag will not contain any informa-
tion but must still be passed into the procedure. Autocorrelations are faster to
compute, since only one forward transformation must be made, as opposed to two
for crosscorrelation.

250 Turbo Numerical Methods Toolbox

This method uses the New/Dispose procedures to manipulate the heap and
should not be used in any program that uses Mark/Release to manipulate the heap.

REALCORR.INC

Description
The calculation of the convolution of two real vectors is facilitated with a Fast
Fourier Transform routine. This procedure is exactly the same as the previous
procedure for complex correlation (COMPCORR.INC) except that only one for-
ward Fourier transform need be performed. The procedure is as follows:
1. Given two real vectors XReal and HReal, combine them into a complex
vector XReal + iHReal, where i is the square root of· - 1.
2. Transform this complex vector.
3. Extract the transforms of the two real vectors from the transform of the
complex vector (using the symmetry Xf = X_ f , where the bar stands for
complex conjugation).
4. Multiply each element of the transform of the first data set by the appro-
priate element of the transform of the second data.
5. Find the inverse transform of this product using FFTs.

Input Parameters
NumPoi nts : Integer; Number of data points
Auto: Bool ean; FALSE equals crosscorrelation; TRUE equals autocorrelation
XRea 1 : TNvectorPtr; Pointer to real values of the first set of data points
HRea 1 : TNvectorPtr; Pointer to real values of the second set of data points (for cross-
correlation)
The preceding parameters must satisfy the following conditions:
1. NumPoints 2! 2.
2. NumPoints must be a power of two [four].

Fast Fourier Transform Routines 251

Output Parameters
XReal : TNvectorptr; Pointer to real values of the correlation of XReal and HReal (or
the autocorrelation of XReal if Auto = TRUE)
Xlmag : TNvectorPtr; Pointer to imaginary values of the correlation of XReal and
HReal (or the autocorrelation of XReal if Auto = TRUE)
Error : Byte; 0: No errors
1: NumPoints < 2
2: NumPoints not a power of two [four]

Syntax of the Procedure Call

RealCorrelation(NumPoints, Auto, XRe31, Xlmag, HReal, Error);

Comments
If you are performing an autocorrelation of the vector XReal, then set Auto equal
to TRUE. In this case, the vector HReal will not contain any information but must
still be passed into the procedure. Autocorrelations are faster to compute, since
only one forward transformation must be made, as opposed to two for crosscorrela-
tion. This method uses the New/Dispose procedures to manipulate the heap and
should not be used in any program that uses Mark/Release to manipulate the heap.

Sample Program

The sample program FFTPROGS.PAS provides I/O functions that demonstrate

any of the application programs. The FFT algorithm routines are included with the
following statements:
(*{$I FFTB2.INC} (* radix 2, 8088 version *)
{*{$I FFTB4. INC} (* radix 4, 8088 version *)
{$I FFT87B2.INC} (* radix 2, 8087 version *)
(*{$I FFT87B4.INC} (* radix 4, 8087 version*)
As you can see, three of the include statements must be commented-off so that
only one file is included. To change which file is included, simply comment-off the
one that is currently active (in this example that would be FFT87B2.INC) and
remove the comment symbol (* from the include file of your choice.

252 Turbo Numerical Methods Toolbox

Input File
Data may be entered from a text file. The real and imaginary parts of a complex
number should be separated by a space and followed by a carriage return. Real
numbers should each be followed by a carriage return.
The application files COMPFFT.lNC, COMPCNVL.lNC, and COMP
CORR.lNC expect data to be in complex form. A data file containing a four-point
complex square wave could look like this:
00
11
11
00
The application files REALFFT.lNC, REALCNVL.lNC, and REAL
CORR.INC expect data to be in real form. A data file containing a four-point real
square wave could look like this:
o
1
1
o

Example
Problem. Perform a Fourier transform and an autocorrelation of a 32-point square
wave. Also, convolve and crosscorrelate this square wave with a saw-tooth wave
(assume you are working in Turbo-87).
1. First, make sure that the FFT file FFT87B2.1NC is included in
FFTPROGS.PAS:
(*{$I FFTB2. INC} (* radix 2, 8088 version *)
(*{$I FFTB4.INC} (* radix 4, 8088 version *)
{$I FFT87B2.INC} (* radix 2, 8087 version *)
(*{$I FFT87B4.INC} (* radix 4, 8087 version *)
The input data file SAMPI0A.DAT is as follows (note that this is in real
format):
o
o
o
o
o
o
o
o

Fast Fourier Transform Routines 253

o
o
o
1
1
1
1
1
1
1
1
1
1
1
o
o
o
o
o
o
o
o
o
o
2. Run FFTPROGS.PAS:
1. Real Fast Fourier Transform
2. Real Convolution
3. Real Autocorrelation
4. Real Crosscorrelation
5. Complex Fast Fourier Transform
6. Complex Convolution
7. Complex Autocorrelation
8. Complex Crosscorrelation
Select a number (1-8): 1
********* Real Fast Fourier Transform *********

(F)orward or (I)nverse transform? F

Enter data from (K)eyboard or (F)ile? F
File name? SAMPI0A.DAT
Direct output to one of the following:
(S)creen
(P)rinter
(F) il e

254 Turbo Numerical Methods Toolbox

Results of Real Fourier Transform:
1.94454364826301E+OOO O.OOOOOOOOOOOOOOE+OOO
-1.59057003804788E+OOO -3.14018491736755E-OI6
7.53417436515731E-OOI 5.88784672006416E-OI7
5.96901852132470E-002 -2.94392336003208E-OI6
-4.26776695296637E-OOI 3.92523114670944E-OI7
2.89883706652938E-OOI -2.94392336003208E-OI6
6.20757203331860E-002 7.85046229341887E-OI7
-2.66655959906343E-OOI -5.39719282672548E-OI7
1.76776695296637E-OOI O.OOOOOOOOOOOOOOE+OOO
6.63840517512576E-002 2.06074635202245E-OI6
-2.08522329739913E-OOI 1.76635401601925E-OI6
1.27160952826887E-OOI -1. 17756934401283E-016
7.32233047033632E-002 -1.37383090134830E-OI6
-1.83841625879619E-OOI -6.86915450674151E-OI7
1.00135954077543E-OOI 2.74766180269661E-OI6
8.37351650164209E-002 1.37383090134830E-OI6
-1.76776695296637E-OOI O.OOOOOOOOOOOOOOE+OOO
8.37351650164209E-002 -1.37383090134830E-OI6
1.00135954077543E-OOI -2.74766180269661E-OI6
-1.83841625879619E-OOI 6.86915450674151E-OI7
7.32233047033632E-002 1.37383090134830E-OI6
1.27160952826887E-OOI 1. 17756934401283E-016
-2.08522329739913E-OOI -1.76635401601925E-016
6.63840517512576E-002 -2.06074635202245E-OI6
1.76776695296637E-OOI O.OOOOOOOOOOOOOOE+OOO
-2.66655959906343E-OOI 5.39719282672548E-OI7
6.20757203331860E-002 -7.85046229341887E-OI7
2.89883706652938E-OOI 2.94392336003208E-OI6
-4.26776695296637E-OOI -3.92523114670944E-OI7
5.96901852132470E-002 2.94392336003208E-OI6
7.53417436515731E-OOI -5.88784672006416E-OI7
-1.59057003804788E+OOO 3.14018491736755E-OI6
Note that the transform of the even real-square wave is an even real function. If
you take the inverse transform of this data, you should get back the original square
wave.
3. Run FFfPROGS.PAS:
1. Real Fast Fourier Transform
2. Real Convolution
3. Real Autocorrelation
4. Real Crosscorrelation
5. Complex Fast Fourier Transform
6. Complex Convolution
7. Complex Autocorrelation
8. Complex Crosscorrelation
Select a number (1-8): 5
********* Complex Fast Fourier Transform *********

(F)orward or (I)nverse transform? I

Enter data from (K)eyboard or (F)ile? F

File name? SAMPI0B.DAT

Fast Fourier Transform Routines 255

Direct output to one of the following:
(S)creen
(P)rinter
(F) il e

Results of Complex Fourier Transform:

1.88411095042053E-015 O.OOOOOOOOOOOOOOE+OOO
1.72710170455215E-015 O.OOOOOOOOOOOOOOE+OOO
2.04112019628891E-015 -3.92523114670944E-017
2.04112019628891E-015 -1.19825492681322E-016
1.17756934401283E-015 O.OOOOOOOOOOOOOOE+OOO
1.09906472107864E-015 -3. 14264671746203E-016
1.02325930356393E-015 1.53635757788210E-016
1.02243303999922E-015 -3.53529506016654E-017
1.07086707894778E-015 O.OOOOOOOOOOOOOOE+OOO
3.60019188480208E-016 2.56022496768081E-016
-3.89233270203019E-017 1.57501605887274E-016
1.00000000000000E+000 4.33843984418077E-016
1.00000000000000E+000 O.OOOOOOOOOOOOOOE+OOO
1.00000000000000E+000 4.71273917614581E-016
1.00000000000000E+000 2.94392336003208E-016
1.00000000000000E+000 5.69158516272868E-016
9.99999999999999E-001 O.OOOOOOOOOOOOOOE+OOO
1.00000000000000E+000 3.92523114670944E-017
1.00000000000000E+000 3.92523114670944E-017
9.99999999999999E-001 -4. 13208697471334E-017
1.00000000000000E+000 O.OOOOOOOOOOOOOOE+OOO
1.00000000000000E+000 1. 56763065858929E-016
-8. 12038283536641E-017 -1.53635757788210E-016
2.33640926947802E-016 -2.78665541135090E-016
9.70253117341126E-016 O.OOOOOOOOOOOOOOE+OOO
1.21007327020357E-015 -2.95274808235175E-016
4.31446441691246E-016 -1.57501605887274E-016
1.80560632748634E-015 -2.72697621989622E-016
9.42055475210265E-016 O.OOOOOOOOOOOOOOE+OOO
2.04112019628891E-015 -3.13772311727307E-016
1.80560632748634E-015 -2.94392336003208E-016
1.96261557335472E-015 -2.55140024536113E-016
You get back the original square wave, accurate to 15 significant figures.
The autocorrelation of a square wave is simply a triangle. Let's take the
autocorrelation of the square wave.
4. Run FFTPROGS.PAS:
1. Real Fast Fourier Transform
2. Real Convolution
3. Real Autocorrelation
4. Real Crosscorrelation
5. Complex Fast Fourier Transform
6. Complex Convolution
7. Complex Autocorrelation
8. Complex Crosscorrelation
Select a number (1-8): 3
********* Real Autocorrelation *********

256 Turbo Numerical Methods Toolbox

Enter data from (K)eyboard or (F)ile? F
File name? SAMP10A.DAT
Direct output to one of the following:
(S)creen
(P)rinter
(F) il e

Results of Autocorrelation:
1.94454364826301E+OOO -6.89287897017933E-016
1.76776695296637E+OOO -5.49532360539321E-016
1.59099025766973E+OOO -6.55792559089139E-016
1.41421356237309E+OOO -4. 17055809337878E-016
1.23743686707646E+OOO -4.44089209850063E-016
1.06066017177982E+OOO -2.74766180269661E-016
8.83883476483184E-001 -3.84393694788862E-016
7.07106781186547E-001 -1. 17756934401283E-016
5.30330085889910E-001 5.37799391805346E-017
3.53553390593274E-001 -1.96261557335472E-016
1.76776695296637E-001 -1.88132137453390E-016
O.OOOOOOOOOOOOOOE+OOO -9.32242397343491E-017
-3.92523114670944E-016 4.44089209850063E-016
2.35513868802566E-016 O.OOOOOOOOOOOOOOE+OOO
4.71027737605132E-016 -1.06260198549818E-016
4.71027737605132E-016 -1.66822323735151E-016
1.57009245868377E-016 5.81728018656864E-016
-3.14018491736755E-016 3.92523114670944E-016
O.OOOOOOOOOOOOOOE+OOO 6.00281407857881E-016
-1.57009245868377E-016 2.89485797069821E-016
-3.92523114670944E-016 4.44089209850063E-016
-3. 14018491736755E-016 1. 17756934401283E-016
1.76776695296636E-001 4.39904846020120E-016
3.53553390593273E-001 -1.96261557335472E-017
5.30330085889910E-001 5.37799391805346E-017
7.07106781186547E-001 3.53270803203849E-016
8.83883476483184E-001 2.43643288684648E-016
1.06066017177982E+OOO 2.20794252002406E-016
1.23743686707646E+OOO -4. 44089209850063E-016
1.41421356237309E+OOO 1.57009245868377E-016
1.59099025766973E+OOO 5.07490473185598E-017
1.76776695296637E+OOO 3.04205413869981E-016
Keeping in mind that this is a periodic function (see "Data Sampling"), you
can see that this is a triangle wave.
Let's now convolve the square wave with a saw-tooth wave. The input file for
the saw-tooth wave (SAMPIOC.DAT) is as follows:
o
o
o
o
o
o

Fast Fourier Transform Routines 257

o
o
o
o
o
1
2
3
4
5
6
7
8
9
10
1
1
o
o
o
o
o
o
o
o
o
o
5. Run FFTPROGS.PAS:
1. Real Fast Fourier Transform
2. Real Convolution
3. Real Autocorrelation
4. Real Crosscorrelation
5. Complex Fast Fourier Transform
6. Complex Convolution
7. Complex Autocorrelation
B. Complex Crosscorrelation
Select a number (I-B): 2
********* Real Convolution *********

Enter data from (K)eyboard or (F)ile? F

The first function:
File name? SAMPI0A.DAT

258 Turbo Numerical Methods Toolbox

The second function:
File name? SAMP10C.DAT
Direct output to one of the following:
(S)creen
(P)rinter
(F) i1 e

Results of Real Convolution~

1. 16672618895780E+001 O.OOOOOOOOOOOOOOE+OOO
1. 14904851942814E+001 1.09906472107864E-015
1. 11369318036881E+001 6.28036983473510E-016
1.06066017177982E+001 4.71027737605132E-016
9.89949493661168E+000 O.OOOOOOOOOOOOOOE+OOO
9.01561146012848E+000 7.85046229341887E-016
7.95495128834866E+000 -6.28036983473510E-016
6.71751442127220E+000 1.57009245868377E-016
5.30330085889910E+000 O.OOOOOOOOOOOOOOE+OOO
3.71231060122936E+000 -1.96261557335472E-015
1.94454364826299E+000 -2.11962481922310E-015
-1.85270910124685E-014 -2.98317567149917E-015
-2.04112019628891E-014 O.OOOOOOOOOOOOOOE+OOO
-1.13046657025232E-014 -3.29719416323593E-015
-1.44448506198907E-014 -3.45420340910430E-015
-9.42055475210265E-015 -3.45420340910430E-015
~1.50728876033642E-014 O.OOOOOOOOOOOOOOE+OOO
-1.38168136364172E-014 -4.71027737605132E-016
-1.38168136364172E-014 -6.28036983473510E-016
-1.13046657025232E-014 -1.57009245868377E-016
-1. 13046657025232E-014 O.OOOOOOOOOOOOOOE+OOO
-5.02429586778808E-015 -1.09906472107864E-015
1.76776695296632E-001 6.28036983473510E-016
5.30330085889907E-001 -2.04112019628891E-015
1.06066017177982E+000 O.OOOOOOOOOOOOOOE+OOO
1.76776695296637E+000 1.33457858988121E-015
2.65165042944956E+000 2.11962481922310E-015
3.71231060122938E+000 2.66915717976242E-015
4.94974746830584E+000 O.OOOOOOOOOOOOOOE+OOO
6.36396103067893E+000 3.61121265497268E-015
7.95495128834867E+000 3.45420340910430E-015
9.72271824131504E+000 5.33831435952483E-015
Now let's crosscorrelate the square wave with the saw-tooth wave.
6. Run FFTPROGS.PAS:
1. Real Fast Fourier Transform
2. Real Convolution
3. Real Autocorrelation
4. Real Crosscorrelation
5. Complex Fast Fourier Transform
6. Complex Convolution
7. Complex Autocorrelation
8. Complex Crosscorrelation
Select a number (1-8): 4
********* Real Crosscorrelation *********

Fast Fourier Transform Routines 259

Enter data from {K)eyboard or {F)ile? F

The first function:

File name? SAMPI0A.DAT
The second function:

File name? SAMPI0C.DAT

Direct output to one of the following:
{S)creen
{P)rinter
(F) il e

Results of Real Crosscorrelation:

1.16672618895780E+001 O.OOOOOOOOOOOOOOE+OOO
9.72271824131504E+000 -6.67289294940604E-016
7.95495128834866E+000 -7.85046229341887E-016
6.36396103067893E+000 -9.02803163743171E-016
4.94974746830583E+000 O.OOOOOOOOOOOOOOE+OOO
3.71231060122937E+000 -1.49158783574959E-015
2.65165042944954E+000 -2.04112019628891E-015
1.76776695296636E+000 -1.72710170455215E-015
1.06066017177981E+000 O.OOOOOOOOOOOOOOE+OOO
5.30330085889897E-001 -2.62990486829532E-015
1. 76776695296621E-001 -3.29719416323593E-015
-1.38168136364172E-014 -2.86541873709789E-015
-1.63289615703113E-014 O.OOOOOOOOOOOOOOE+OOO
-1.06766287190497E-014 -3.06168029443336E-015
-1.25607396694702E-014 -3. 14018491736755E-015
-1.00485917355762E-014 -4.39625888431457E-015
-1. 13046657025232E-014 O.OOOOOOOOOOOOOOE+OOO
-9.42055475210265E-015 -3. 53270803203849E-016
-5.02429586778808E-015 7.85046229341887E-016
-5.02429586778808E-015 -2.74766180269661E-016
-5.33831435952483E-015 O.OOOOOOOOOOOOOOE+OOO
-3.76822190084106E-015 -5.49532360539321E-016
1.94454364826300E+000 2.04112019628891E-015
3.71231060122937E+000 4.71027737605132E-016
5.30330085889911E+000 O.OOOOOOOOOOOOOOE+OOO
6.71751442127221E+000 3. 65046496643978E-015
7.95495128834867E+000 3.29719416323593E-015
9.01561146012849E+000 4.04298808111072E-015
9.89949493661168E+000 O.OOOOOOOOOOOOOOE+OOO
1.06066017177982E+001 5.10280049072227E-015
1.11369318036881E+001 3. 14018491736755E-015
1. 14904851942814E+001 5.65233285126159E-015

260 Turbo Numerical Methods Toolbox

c H A p T E R n
Graphics Programs

The programs in this chapter can only be run by those users with PC-DOS.
There are some programs that graphically demonstrate the usefulness of the
least-squares routines in Chapter 9 and the Fourier transforms in Chapter 10. A
graphics monitor is required. Each program reads a data set from an input file, and
uses this Toolbox to display the results. You will see curves being fitted to data
using the least-squares routines and also see a signal being transformed into its
Fourier spectrum.
The programs LSQIBM.COM and FFTIBM.COM graphically illustrate the
power and utility of the Turbo Pascal Numerical Methods Toolbox. To run them,
you'll need an IBM Color Graphics Adapter (CGA) or a suitable clone. (The pro-
grams LSQHERC.COM AND FFTHERC.COM can be used to graphically illus-
trate the Toolbox on a Hercules Monochrome Graphics Adapter or compatible.)
And to print, you'll need an Epson or IBM compatible dot-matrix printer. As
explained in this chapter, these programs can be recompiled to run on other hard-
ware, including the IBM Enhanced Graphics Adapter (EGA) in its high resolution
mode. The programs can also be recompiled to take advantage of the 8087 (or
80287) math coprocessor.

261
Function of the Least-Squares Graphics Demonstration
Program

The program LSQIBM.COM demonstrates the least-squares capabilities of the

Toolbox. To run it, you must have
• 4x6.FON (Graphix fonts)
• 8x8.FON (Graphix fonts)
• ERROR.MSG (Graphix error messages)
• LSQIBM.COM
• SAMPllA.DAT (Input data file)
on the current directory. (The first three files are identical to those in the Graphix
Toolbox.)
By default the input is a file called SAMPllA.DAT that has X and Y values (in
ASCII form) separated by carriage returns. Running LSQIBM.COM will provide
five different least-squares fits to the input data. The different fits are based on the
function forms: logarithm, exponential, polynomial, power law, and finite Fourier
series. The fits are displayed graphically on the screen and can be printed on an
Epson or compatible printer.
The first plot shows the input data from SAMPllA.DAT along with three curves.
The three curves are the graphs of the power function
Y = aX b

the exponential function

Y = a exp (bX)
and the logarithm function
Y = a In(bX)
The header to the plot tells which curve corresponds to which function. The next
plot shows the same input data plotted with a five-term Fourier series:
Y= a + b * cos(x) + c * sin(x) + d * cos(2X) + e * sin(2X)
and a five-term polynomial (that is, a polynomial of degree four). The coefficients
are found using the routines from Chapter 9, and they give the least-square error
among all functions of that form. (In some cases, the problem is transformed into a
linear problem, and the error is actually the least for the transformed problem but
possibly not exactly the least for the original problem.) Again, the header to the plot
tells which curve corresponds to which function.

262 Turbo Numerical Methods Toolbox

Finally, a bar chart shows the error for each function. The data is not at all
periodic, so the Fourier series model is the worst. The five-degree polynomial gives
the best fit, but it is not much better than the fit obtained by using power, exponen-
tial, or logarithm functions.
Pressing the space bar allows you to cycle through the different screens. Press-
ing I]l exits the program. PressingIJDsends a hard copy to the printer (see the
section entitled "Printing").
You can use your own data to run the program by running LSQDEMO with two
file names, such as
LSQIBM LSQIN.DAT L~QOUT.DAT

The input data from LSQIN.DAT along with the least-squares fits and coeffi-
cients are output to a text file called LSQOUT.DAT.
A default output data file can easily be arranged by changing a constant
WriteToFile in LSQDEMO.PAS and recompiling it. (See the section, "Rebuilding
LSQIBM.COM; to recompile it, and the comment in LSQDEMO.PAS next to the
constant WriteToFile.)

Graphics Programs 263

Function of the Fourier Transform Graphics Demonstration
Program

The program FFfIBM.COM demonstrates the Fourier capabilities of the Toolbox.

To run it, you must have
• 4x6.FON (Graphix fonts)
• 8x8.FON (Graphix fonts)
• ERROR.MSG (Graphix error messages)
• FFTIBM.COM
• SAMPllB.DAT (Input data file)
on the current directory. (The first three files are identical to those in the Graphix
Toolbox.)
By default the input is a file called SAMPllB.DAT that has 1,024 real values (in
ASCII form) separated by carriage returns. These represent sample points from a
two-second signal sampled at a rate of 512 points per second. The program will
display four FFT transforms at the following sampling rates: 8 per second (16
points), 32 per second (64 points), 128 per second (256 points), and 512 per second
(1,024 points). For the last two samplings, the default data yields the same trans-
forms, demonstrating that a sample rate higher than twice the highest frequency
adds no new information (the Nyquist limit). The transforms are shown on a scale
of - 64 to + 63 cycles per second.
In addition to the transforms, the program displays the inverse transform over
the original data, illustrating the degree to which information is lost at different
sampling rates. The header tells which curve is the original data and which is the
inverse transform.
Pressing the space bar cycles through the four plots. To the right of each plot is
two smaller plots that show the coefficients for the real and imaginary parts of the
Fourier transform. Some information about the sample and the transform appears
after each plot.
The graphs that appear on the screen can be printed on an Epson or compatible
printer.
You can use your own data to run the program by running FFfDEMO with two
file names on the command line, such as
FFTIBM FFTIN.DAT FFTOUT.DAT
The 1,024 complex points of the 1,024-point FFTwill be written into the output
file FFfOUT.DAT, with one complex number per line. The real and imaginary
parts are separated by a space.

264 Turbo Numerical Methods Toolbox

A default output data file can easily be arranged by changing the constant
WriteToFile in FFTDEMO.PAS and recompiling it. (See the section, "Rebuilding
FFfIBM.COM; to recompile it, and the comment in FFfDEMO.PAS next to
Write ToFile.)

Printing

Both LSQIBM.COM and FFfIBM.COM provide the capability to print the

graphs that appear on the screen. If you run these programs on the default data
sets, the printouts will look just like the ones displayed here (see Figure 11-1,11-2,
and 11-3). This is particularly useful when you would like a permanent, visual
record of your program results.

Figure 11-1 LSQIBM.COM

Least-square fits of power, exponential, and logarithmic functions to data in SAMPllA.DAT. Note: This was run
without an 8087.

Graphics Programs 265

POWER EXP LOG S-,OUR. S-POLY.

'.11

Figure 11-2 LSQIBM.COM

Sum of squares of residuals for five fits. Left side displays bar graph comparing various fits. Right side depicts
same data fitted by 5th degree polynomial and a partial Fourier series. Note: This was run without an 8087.

Ti.e interval = 2 seconds. Sople rate = 32 per second.

TransFor. tI.e = 8.7 seconds. IllYerse tI.e = 8.5 seconds.

Press (SPACE) to continue. (Q) to quit

Figure 11-3 FFTIBM.COM

Graph depicts data ofSAMPllB.DAT. Upper right-hand side displays real transform coefficients. Mid right-hand
side displays imaginary coefficients. Dotted lines represent inverse transform of the Fourier transform, superim-
posed over original data. Inverse transform is not identical to original data because of coarseness of sample rate.
Note: This was run without an 8087.

266 Turbo Numerical Methods Toolbox

To get a printout, merely pressl]lfor Hardcopy at the menu prompt. An Epson
or IBM compatible dot-matrix printer is required. (You cannot get a graphics print-
out on a daisy-wheel printer.)
If your printer is not Epson-compatible and the Hardcopy command is not func-
tioning properly, you can rebuild the program with a different printer mode. (See
the instructions that follow for rebuilding the program.) Both LSQDEMO.PAS and
FFfDEMO.PAS have a constant called PrintMode near the beginning of the file.
Setting PrintMode to the value 1 will allow printing on the largest number of
printers, but a mode value of 4 will give the best looking output from a color display
to an Epson printer. The default value is 6, since that gives good results for the
Color Graphics Adapter, Hercules, and Enhanced Graphics Adapter.
You can also use the DOS program GRAPHICS.COM, which enables the
I PrtSc I key to print the screen in graphics mode.

Rebuilding LSQIBM.COM

The sources to LSQIBM.COM are provided. To recompile, you need Turbo Pascal
(version 3.0) and Turbo Graphix Toolbox (version 1.06A or later), as well as this
Toolbox.
To rebuild LSQIBM.COM, the following files are needed:
From CHAP9:
EXP.LSQ (for the exponential model)
FOURIER.LSQ (Fourier series)
LOG.LSQ (logarithmic)
POLY.LSQ (polynomial)
POWER.LSQ (power law)
From CHAPll:
GENERIC.LSQ
IOCHECK.lNC (a modification of COMMON.lNC)
LEAST.MOD (a modification of LEAST.lNC from Chapter 9)
LSQDEMO.PAS
From Turbo Pascal:
(These files are not included in this Toolbox.)
TURBO.COM or TURBO-87.COM (the Turbo Pascal compiler)
TURBO.MSG (compiler error messages)

Graphics Programs 267

From the Turbo Pascal Graphix Toolbox (version 1.06A or later):
(These files are not included in this Toolbox.)
AXIS.HGH
FINDWRLD.HGH
GRAPHIX.SYS (a copy of GRAPHIX.IBM)
HATCH.HGH
HISTOGRM.HGH
KERNEL.SYS
POLYGON.HGH
TYPEDEF.SYS
WINDOWS.SYS
Once you have all of these files on the current directory, enter Turbo Pascal and
compile LSQDEMO.PAS to disk. The resulting LSQDEMO.COM should be
renamed LSQIBM.COM to distinguish it from the Hercules version.

Rebuilding FFTIBM.COM

The sources to FFTIBM.COM are provided. To recompile, you need Turbo Pascal
(version 3.0) and Turbo Graphix Toolbox (version 1.06A or later), as well as this
Toolbox.
To rebuild FFTIBM.COM, the following files are needed:
From CHAPIO:
COMPFFT.lNC
FFTB2.1NC
REALFFT.lNC
From CHAPll:
4X6.FON (originally from the Graphix Toolbox)
8X8.FON (originally from the Graphix Toolbox)
ERROR.MSG (originally from the Graphix Toolbox)
FFTDEMO.PAS
IOCHECK.lNC (a modification of COMMON.INC)
From Turbo Pascal:
(These files are not included in this Toolbox.)
TURBO.COM or TURBO-87.COM (the Turbo Pascal compiler)
TURBO.MSG (compiler error messages)

268 Turbo Numerical Methods Toolbox

From the Turbo Pascal Graphix Toolbox (version 1.06A or later):
(These files are not included in this Toolbox.)
AXIS.HGH
FINDWRLD.HGH
GRAPHIX.SYS (a copy of GRAPHIX.lBM)
KERNEL.SYS
POLYGON.HGH
TYPEDEF.SYS (with MaxPlotGlb changed to the value 1,024)
WINDOWS.SYS
The system constant MaxPlotGlb in the file TYPEDEF.SYS must be changed to
have the value 1,024 instead of 100. Without this change, FFTDEMO will termi-
nate with an error message.
Once you have all of these files on the current directory, enter Turbo Pascal and
compile FFTDEMO.PAS to disk. The resulting FFTDEMO.COM should be
renamed FFTIBM.COM to distinguish it from the Hercules version.

Rebuilding for the Hercules Card

To recompile for the Hercules graphics card, simply copy GRAPHIX.HGC to

GRAPHIX.SYS, as explained in the Turbo Pascal Graphix Toolbox. The font file
14X9.FON is also required from the Graphix Toolbox, and the version of the
Graphix Toolbox must be 1.07A or later. (You can use version 1.06A of the Graphix
Toolbox if you're using the copy of GRAPHIX.HGC accompanying the other
CHAPll files.)
The resulting files LSQDEMO.COM and FFTDEMO.COM should be
renamed to LSQHERC.COM and FFTHERC.COM to distinguish them from the
IBM CGA versions.

Rebuilding for the EGA Card

To recompile for the IBM Enhanced Graphics Adapter (EGA), you must copy
GRAPHIX.EGA to GRAPHIX.SYS. A copy of GRAPHIX.EGA is included with
this Toolbox since many users purchased copies of the Turbo Graphix Toolbox
before support for the EGA was added.
LSQIBM.COM and FFTIBM.COM will run on machines with an EGA card.
However, to take advantage of the higher resolution offerred by this card, you must
copy GRAPHIX.EGA and recompile.

Graphics Programs 269

Release 1.07A of the Graphics Toolbox also supports other graphics hardware.
LSQDEMO and FFTDEMO can be recompiled for any of these systems.

Using the Math Coprocessor

To recompile to take advantage of the math coprocessor, you must use TURBO-
87.COM instead of TURBO.COM. Some increased performance in the FFr demo
program will be obtained ifFFfB2.1NC is replaced by FFT87B2.INC from Chap-
ter 10 of the Toolbox.

270 Turbo Numerical Methods Toolbox

References

Abramowitz, Milton, and Irene A. Stegun, eds. Handbook of Mathematical

Functions with Formulas, Graphs, and Mathematical Tables. Washington; D.C.:
National Bureau of Standards Applied Mathematics Series, 55, 1972.
Atkinson, L.V., and P.J. Harley. An Introduction to Numerical Methods with
Pascal. Reading: Addison-Wesley Publishing Co., 1983. This is an excellent text for
learning numerical methods, with an emphasis on the implementation of various
numerical algorithms.
Brigham, E. Oran. The Fast Fourier Transform. Englewood Cliffs: Prentice-Hall,
Inc., 1974. A very complete, easy-to-read text on the use and implementation of the
fast Fourier transform algorithm .
.The next three texts are excellent for learning numerical analysis, emphasizing
the mathematical theory underlying the algorithms in this toolbox.
Burden, Richard L., and J. Douglas Faires. Numerical Analysis, 3rd ed. Boston:
Prindle, Weber & Schmidt, 1985.
Cheney, Ward, and David Kincaid. Numerical Mathematics and Computing, 2nd
ed. Monterey: Brooks/Cole Publishing Co., 1985.
Dahlquist, Germund, and Ake Bjorck. Numerical Methods, trans. Ned Anderson.
Englewood Cliffs: Prentice-Hall, Inc., 1974.
Fried, Stephen S. "Evaluating 8087 Performance on the IBM PC." Byte. Vol. 9,
Number 9 (Special Issue), 1985, pp.l97-208.

271
Gerald, Curtis F., and Patrick O. Wheatley. Applied Numerical Analysis, 3rd ed.
Reading: Addison-Wesley Publishing Co., 1984. This is another excellent source for
learning numerical analysis.
Press, William H., Brian P. Flannery, Saul A. Teukolsky, and William T. Vetter-
ling. Numerical Recipes: The Art of Scientific Computing. New York: Cambridge
University Press, 1986. This book is user-oriented, discusses many of the subtle
problems encountered when implementing numerical methods, and has program
listings in Turbo Pascal.
Ralston, Anthony, and Philip Rabinowitz. A First Course in Numerical AnalysiS.
New York: McGraw-Hill Book Co., 1978. A well-written mathematics text that is a
step more sophisticated than the preceding ones.

272 Turbo Numerical Methods Toolbox

Index

273
$FF error, 145, 190,202 Cubic spline methods
clamped, 44, 57-62
A free, 44,52-56, 64,76-79
ADAMS_l.INC, 156, 168-170 Cyclic Jacobi method, 132, 149-153
ADAMS_l.PAS, 170-171
Adams-Bashforth/Adams-Moulton D
predictor-corrector method, 156, Data sampling, 239-240
168-171 Data types, 12
Adams-Bashforth formula, 168 Defined constants, 12
Adams-Moulton formula, 168 Deflation, 16
ADAPGAUS.INC, 88, 98-100 and Laguerre, 37-41
ADAPGAUS.PAS, 101 and Newton-Horner, 28-32
ADAPSIMP.INC, 88, 95-96 of a matrix, 132, 143-148
ADAPSIMP.PAS, 96-97 DERIV2FN.lNC, 64-65, 83-84
Adaptive schemes, 87 DERIV2FN.PAS, 84-86
quadrature, 95-101 DERIV2.1NC, 64, 71-73
Aliasing, 239 DERIV2.PAS, 73-75
Derivative, 63
B approximation of, 64-86
Backward substitution, 114, 120 DERIVFN .INC, 64, 80-81
Basis vectors, 222 DERIVFN .PAS, 81-82
BISECT.INC, 18-19 DERIV.INC, 64, 66-68
Bisection method, 15 DERIV.PAS, 68-70
root of a function using, 18-20 Determinant of a matrix, 105-109
BISECT.PAS, 19-20 DET.INC, 106-108, 128
Boundary value problems, 155-156, 158 DET.PAS, 108-109
using Linear Shooting/Runge-Kutta Diagonally dominant, 126
methods, 215-219 Diagonal matrix, 131
using Shooting/Runge-Kutta methods, Differential equations
208-214 first-order, 159-171
coupled,186-195
C linear, 155
nth order, 157, 178-185
Chebyshev polynomials, 224 ordinary, 155
Color Graphics Adapter, 3, 12 second-order, 158, 172-177, 192,
for graphics demos, 261, 267 208-219
COMMON.INC,7, 11, 13 coupled, 196-207
COMPCNVL.INC, 233, 235, 237-238 stiff, 161
application, 246-247 . systems of, 157
COMPCORR.INC, 233, 235, 238 Direct factorization of matrices, 106,
application, 249-250 120-125
COMPFFT.INC, 233, 235, 237 DIRFACT.lNC. 106, 120-122
application, 244-245 DIRFACT.PAS, 122-125
Compiler directives, 13 Distribution disks, 7-11
Convergence, rate of, 15-16 DIVDIF.lNC, 49-50
Cooley-Tukey method, 233-234 DIVDIF.PAS, 50-51
Coprocessor, 3, 270
CUBE_CLA.lNC, 57-58 E
CUBE_CLAPAS, 59-62
CUBEJRE.lNC, 52-53 Eigensystem, 149
CUBEJRE.PAS, 53-56 Eigenvalue, 131-132
Eigenvector, 131

274 Turbo Numerical Methods Toolbox

Enhanced Graphics Adapter, 3, 261, 267 Graphics
rebuilding for, 269-270 demo programs, 12, 262-267
EXP.LSQ, 225-226, 267 printing, 265-267 ,
system requirements, 3,261
F GRAPHICS.COM, 267
Fast Fourier Transform, 233 GRAPHICS.EGA, 269
algorithms, 241-243
applications, 244-260 H
sample program, 252-260 Heap/Stack collision, 145, 190, 202
FFT87B2.1NC, 233-235 Hercules Monochrome Graphics Adapter,
algorithms, 241-243 3,12,261
with the math coprocessor, 270 rebuilding for, 269
FFT87B4.1NC, 233-235,
algorithms, 241-243 I
and REALFFT.lNC, 237 Initial value problems, 155-157
FFTB2.1NC, 233"'-234 Adams-Bashforth/Adams-Moul ton,
algorithms, 241-243 168-171
rebuilding for FFTIBM.COM, 268 Runge-Kutta order five, 163-167
with the math coprocessor, 270 Runge-Kutta order four, 159-167,
FFTB4.1NC, 233-235 172-207
algorithms, 241-243 Installation, 8-11
and REALFFT.lNC, 237 Integration, 87-104
FFTDEMO.COM, 269 INTERDRV.INC, 64, 76-77
FFTDEMO.PAS, 264-269 INTERDRV.PAS, 77-79
FFTHERC.COM,261 Interpolation, 43, 158
FFTIBM.COM, 261 cubic splines, 52-62
graphics demo, 264-266 Lagrange polynomials, 45-48
rebuilding, 268-269 Newton's divided-difference method,
FFT procedure, 242-243 49-51
FFTPROGS.PAS, 252-260 INVERSE.lNC, 106, 110-111
Forward substitution, 120 Inverse of a matrix, 105-106, 110-113
Forward transform, 234 INVERSKPAS, 111-113
FOURIERLSQ, 225, 267 Inverse power method, 131-132, 137-142
Fourier series, 225 Inverse transform, 234
in graphics programs, 262 INVPOWER.INC, 131-132, 137-139
Fourier transform, 3, 233-260 INVPOWERPAS, 139-142
in graphics demo, 264-265 IOCHECK.lNC, 267
Iterative methods, 15
G cyclic Jacobi, 149-153
GAUSELIM.lNC, 106, 114-115 Gauss-Seidel, 106
GAUSELIM.PAS, 115-116
Gaussian elimination, 106 J
with partial pivoting, 106, 117-119 JACOBI.INC, 132, 149-151
Gaussian quadrature, 88
JACOBI.PAS, 151-153
using Legendre polynomials, 98-101
GAUSSIDL.INC, 106, 126-127 L
GAUSSIDL.PAS, 128-130
Gauss-Jordan elimination, 110 LAGRANGE.INC, 45-46
Gauss-Seidel iterative method, 106, Lagrange method, 43, 45-48
126-130 LAGRANGE.PAS, 46-48
GENERIC.LSQ, 267 LAGUERRE.INC, 37-38
Goodness of fit, 222 LAGUERRKPAS, 39-41

Index 275
Laguerre's method, 16 Mesh points, 156
finding roots of complex polynomial, MULLERINC, 33-35
37-41 MULLERPAS, 35-36
LEAST.IN C, 106, 222-226 Muller's method, 16
LEAST. MOD, 267 finding roots of complex function,
LEAST.PAS, 227-231 33-36
Least-squares approximation, 222-231
Least-squares solution N
graphics demo, 262-263 New/Dispose, 96, 99, 145
linear regression, 221 in Adams-Bashforth/Adams-Moulton,
multiple regression, 221 170
Linear equations, 105-106 in complex fast Fourier, 245-252
differential, 155 in Linear Shooting/Runge-Kutta, 217
with direct factoring, l20-125 in Runge-Kutta, 161, 182, 191,202,211
with Gaussian elimination, 114-119 in Runge-Kutta-Fehlberg, 165
LINSHOT2.1NC, 156, 158, 215-218 NEWTDEFL.INC, 28-30
LINSHOT2.PAS, 218-219 NEWTDEFL.PAS, 30-32
Lipshitz condition, 157 Newton-Horner method, 15-16
LOG.LSQ, 226 with deflation, 28-32
LSQDEMO.COM, 268 Newton-Raphson method, 15-16
LSQDEMO.PAS, 263 root of a function using, 21-24
rebuilding for EGA, 269 Newton's general divided-difference
rebuilding for printing, 265-267 algorithm, 43, 49-51
LSQHERC.COM, 261 Nonlinear shooting method, 158,
rebuilding for Hercules, 269 208-214
LSQIBM.COM, 261-263 Numerical differentiation, 63-65
graphics demo, 265-267 five-point formulas, 64, 66-75
rebuilding, 267 three-point formulas, 64, 66-75
rebuilding for EGA, 269 two-point formulas, 64, 66-70
LUJ)ecompose, l20-l21 Nyquist frequency, 239, 264
LU_Solve, l20-l22
p
M
Partial pivoting, 106, 117
MakeSinCosTable, 242-243 and direct factoring, l20
Mark/Release, 96, 99, 145 PARTPIVT.INC, 106, 117-118
in Adams-Bashforth/Adams-Moulton, PARTPIVT. PAS , 118-119
170 POLY.LSQ, 224, 267
in complex fast Fourier, 245-252 Polynomials
in Linear Shooting/Runge-Kutta, 217 Lagrange, 89
in Runge-Kutta, 161, 182, 191, 202, 211 Legendre, 98-100
in Runge-Kutta-Fehlberg, 165 methods to approximate roots of, 16,
Matrix 25-41
algebra, 105 POWERINC, 131-135
diagonal, 131 POWERLSQ, 225, 267
direct factorization, 106, l20 Power method, 131-136
identity, 139 and Wielandt's deflation, 143-148
nonsingular, 106, l20-l25 POWERPAS, 135-136
orthogonal, 134 Powers-of-four, 234
permutation, l20 Powers-of-two, 234
rotation, 149
square, 106, 131-133, 137
symmetric, 132, 149-153

276 Turbo Numerical Methods Toolbox

R SideKick, 96, 99, 145, 190, 202
RAPHSON2.INC, 11 SIMPSON .INC, 87, 89-90
RAPHSON.INC, 9, 21-22 SIMPSON.PAS, 90-91
RAPHSON .PAS, 22-24 Simpson's method, 87-97
REALCNVL.INC, 233, 235, 238 Splines (see Cubic spline methods)
SuperKe~96, 99,145, 190, 202
application, 248-249
REALCORR.INC, 233, 235,238 System requirements, 3
application, 251-252
REALFFf.INC, 233, 235, 237 T
application, 245-246 TestForRoot, 17
Richardson extrapolation TestInput, 241, 243
and numerical integration, 64, 80-82 TNArraySize,12
and Romberg method, 88, 102-104 TN complex, 33-34
numerical integration, 64, 80-82 TNCompVector, 37-38
RKF_l.INC, 156, 163-165 TNIntVector, 28-29
RKF_l.PAS, 165-167 TNNeariyZero, 12, 16-17
Romberg algorithm, 102-104 TNTargetF, 18
ROMBERG.INC, 88, 102-103 TN vector, 12
Romberg method, 88 Trapezoid composite rule, 92-94
using trapezoidal rule, 102-104 Trapezoid method, 87-88
ROMBERG.PAS, 103-104 TRAPZOID.lNC, 87, 92-93
Roots of an equation, 15-42 TRAPZOID.PAS, 93-94
Rotation matrix, 149-153 Turbo-87, 3-4, 12
Runge-Kutta-Fehlberg, 156, 163-167 TURBO-87.COM, 267-268, 270
RUNGE_l.INC, 156-157, 159-161 TURBO.COM, 267-268, 270
RUNGE_l.PAS, 161-162 TURBO.MSG, 267-268
RUNGE-2.INC, 157, 172-175 Turbo Pascal, 1-3, 11
RUNGE-2.PAS, 175-177 rebuilding with, 267-268
Runge-Kutta formulas, 179, 187-188, Turbo Pascal Graphix Toolbox, 3, 12
197-199 in graphics demos, 267-268
Runge-Kutta methods, 156 rebuilding, 267-269
fifth-order, 163
fourth-order, 156-157, 159-168, U
172-186 UNPACK.EXE,7-8
RUNGE-N.INC, 157, 178-182 USER.LSQ, 226
RUNGE-N .PAS, 182-185
RUNGE-Sl.INC, 157, 186-191 w
RUNGE-Sl.PAS, 191-195
RUNGE_S2.INC, 157, 196-203 WIELANDT.INC, 132, 143-145
RUNGE-S2.PAS, 203-207 Wielandt method, 132
with deflation, 143-148
S
SECANT.INC, 25-26
Secant method, 16
in nonlinear equations, 158, 208-214
root of a function using, 25-27
SECANT.PAS, 26-27
SHOOT2.INC, 158, 208-211
SHOOT2.PAS, 211-214
Shooting method, 158
linear, 215-219
nonlinear, 208-214

Index 277
278 Turbo Numerical Methods Toolbox
Borland
Software

=i 3:-0
¥J'
BORLAND
INTERNATIONAL 4585 Scotts Valley Drive, Scotts Valley, CA 95066

Available at better dealers nat1onwide.

To order by credit card, call (800) 255-8008; CA (BOO) 742-1133;
CANADA (800) 237-1136.
"11,rll.fll 11E IEBI1"
IJ.J IR.IIJ : IIRSAIIIIER
®

Whether you're running WordStar,® Lotus,® dBASE,®

or any other program, SideKick puts all these desktop
accessories at your fingertips-Instantly!
A full-screen WordStar-like Editor to jot A Monthly Calendar from 1901 through
down notes and edit files up to 25 pages 2099.
long. Appointment Calendar to remind you
A Phone Directory for names, addresses, of important meetings and appointments.
and telephone numbers. Finding a name or a A full-featured Calculator ideal for
number is a snap. business use. It also performs decimal
An Autodialer for all your phone calls. It will to hexadecimal to binary conversions.
look up and dial telephone numbers for you. An ASCII Table for easy reference.
(A modem is required to use this function.)

All the SideKick windows stacked up over Lotus 1-2-3.- Here's SideKick running over Lotus 1-2-3. In the
From bottom to top: SideKick's "Menu Window," ASCII SideKick Notepad you'll notice data that's been imported
Table, Notepad, Calculator, Appointment Calendar, Monthly directly from the Lotus screen. In the upper right you can
Calendar, and Phone Dialer. see the Calculator.

The Critics' Choice

"In a simple, beautiful implementation of WordStar's "SideKick is by far the best we've seen. It is also
block copy commands, SideKick can transport all the least expensive."
or any part of the display screen (even an area -Ron Mansfield, ENTREPRENEUR
overlaid by the notepad display) to the notepad."
"If you use a PC, get SideKick. You'll soon become
-Charles Petzold, PC MAGAZINE
dependent on it." -Jerry Pournelle, BYTE
"SideKick deserves a place in every PC."
-Gary Ray, PC WEEK
Suggested Retail Price: $84.95 (not copy protected)
Minimum system configuration: IBM PC, XT, AT, PCjr and true compatibles. PC-DOS (MS-DOS) 2.0 or greater. 128K RAM. One disk
drive. A Hayes-compatible modem, IBM PCjr internal modem, or AT&T Modem 4000 is required for the autodialer function.

SideKick is a registered trademark of BorlCJld International. Inc. dBASE is a registered trademark of

Ashton-Tate. IBM. XT. AT. and PCjr are registered trademarks of International Business Machines Corp.
AT&T is a registered trademark of AmericCJl Telephone & Telegraph Company. Lotus and 1-2-3 are
registered trademarks of Lotus Development Corp WordStar is a registered trademark of MicroPro
International Corp. Hayes is a trademark of Hayes Microcomputer Products, Inc.
Copyright 1987 Borland International BOR0060C
The Organizer For The Computer Age!
Traveling SideKick is BinderWare,@ both a binder you take with you when you travel
and a software program-which includes a Report Generator-that generates and
prints out all the information you'll need to take with you.
Information like your phone list, your client list, you simply punch three holes, fold and clip
your address book, your calendar, and your the form into your Traveling SideKick binder, and
appointments. The appointment or calendar files you're on your way. Because Traveling SideKick is
you'r~ already using in your SideKick8 can auto- CAD (Computer-Age Designed), you don't fool
matically be used by your Traveling SideKick. You around with low-tech tools like scissors, tape, or
don't waste time and effort reentering information staples. And because Traveling SideKick is
that's already there. electronic, it works this year, next year, and all the
One keystroke prints out a form like your address "next years" after that. Old-fashioned daytime
book. No need to change printer paper; organizers are history in 365 days.

What's inside Traveling SideKick

What the software program and its
Report Generator do for you before
you go-and when you get back
Belore you go:
• Prints out your calendar,
appointments, addresses, phone
directory, and whatever other
information you need from your
data files
When you return:
• Lets you quickly and easily enter all
the new names you obtained while
you were away into your
SideKick data files
It can also:
• Sort your address book by contact,
zip code or company name
• Print mailing labels
• Print information selectively
• Search files for existing addresses
or calendar engagements
Suggested Retail Price: $69.95 (not copy protected)
Minimum system configuration: IBM PC, XT, AT, Portable, PCjr, 3270 and true compatibles. PC·DOS (MS·DOS) 2.0 or later.
256K RAM mimimum.

SideKick. BinderWare and Traveling SideKick are registered trademarks of Borland

BORLAND
INTERNATIONAL
International. Inc. IBM. AT. XT. and PCjr are registered trademarks 01 International Business
Machines Corp. MS-DOS is a registered trademark of Microsoft Corp. Copyright 1987
Borland Inlernational BOR 0083A
RAM-resident
Increased productivity for IBM8PCs Dr compatibles
SuperKey's simple macros are electronic shortcuts to success.
By letting you reduce a lengthy paragraph into a single keystroke
of your choice, SuperKey eliminates repetition.
SuperKey turns 1,000 keystrokes into 1!
SuperKey can record lengthy keystroke sequences and play them back at the touch of a single key.
Instantly. Like magic.
In fact, with Super Key's simple macros, you can turn "Dear Customer: Thank you for your inquiry.
We are pleased to let you know that shipment will be made within 24 hours. Sincerely," into the
one keystroke of your choice!
SuperKey keeps your confidential files-confidential!
Without encryption, your files are open secrets. Anyone can walk up to your PC and read your
confidential files (tax returns, business plans, customer lists, personal letters, etc.).
With SuperKey you can encrypt any file, even while running another program. As long as you keep
the password secret, only you can decode your file correctly. SuperKey also implements the U.S.
government Data Encryption Standard (DES).

~ RAM resident-accepts new macro files ~ Keyboard buffer increases 16 character

even while running other programs keyboard "type-ahead" buffer to 128
~ Pull-down menus characters
~ Superfast file encryption ~ Real-time delay causes macro playback
~ Choice of two encryption schemes to pause for specified interval
~ On-line context-sensitive help ~ Transparent display macros allow
creation of menus on top of application
~ One-finger mode reduces key programs
commands to single keystroke
~ Data entry and format control using
~ Screen OFF/ON blanks out and restores "fixed" or "variable" fields
screen to protect against "burn in"
~ Command stack recalls last 256
~ Partial or complete reorganization of
characters entered
keyboard

Suggested Retail Price: $99.95 (not copy protected)

Minimum system configuration: IBM PC. Xl. AT. PCjr. and true compatibles: PC-DOS (MS-DOS)
2.0 or greater. 128K RAM. One disk drive.

SuperKey is a registered trademark 01 Borland Internationat. Inc. IBM. XT, AT. and PCjr are
registered trademarks 01 International Business Machines Corp. MS-DOS is a registered
trademark 01 Microsoft Corp. BOR 0062C
If you use an IBM® PC, you need

T U R B 0
Lightning®
Turbo Lightning teams up II you ever write a You can teach Turbo
with the Random House word, think a word, or Lightning new words
Concise Word List to say a word, you need You can teach your new Turbo
check your spelling as Turbo Lightning Lightning your name, business
you type! associates' names, street
Turbo Lightning, using the names, addresses, correct
BO,OOO-word Random House capitalizations, and any
Dictionary, checks your spelling specialized words you use
as you type. If you misspell a frequently. Teach Turbo
word, it alerts you with a Lightning once, and it
"beep." At the touch of a key, knows forever.
Turbo Lightning opens a
window on top of your Turbo Lightning is the
application program and engine that powers
suggests the correct spelling.
Just press one key and the
"
. Borland's Turbo Lightning
Libraryf'J
misspelled word is instantly Turbo Lightning brings
replaced with the correct word. electronic power to the
The Turbo Lightning Proofreader
Random House Concise Word
Turbo Lightning works List and Random House
hand-in-hand with the Thesaurus. They're at your
Random House Thesaurus fingertips-even while you're
to give you instant access running other programs. Turbo
to synonyms Lightning will also "drive"
soon-to-be-released
Turbo Lightning lets you
encyclopedias, extended
choose just the right word from
thesauruses, specialized
a list of alternates, so you
dictionaries, and many other
don't say the same thing the
popular reference works. You
same way every time. Once
get a head start with this
Turbo Lightning opens the
first volume in the Turbo
Thesaurus window, you see a Lightning Library.
list of alternate words; select
the word you want, press
ENTER and your new word will The Turbo Lightning Thesaurus
instantly replace the original
word. Pure magic!

Suggested Retail Price: $99.95 (not copy protected)

Minimum system configuration: IBM PC, XT, AT, PCjr, and true compatibles with 2 lIoppy disk drives. PC·DOS (MS·DOS) 2.0 or greater.
256K RAM. Hard disk recommended.

BORLAND
INTERNATIONAL
Turbo Lightning and Turbo Lightning Library are registered trademarks of Bortand tnternational, Inc.
IBM. XT, AT, and PCjr are registered trademarks of International Business Machines Corp, Random
House is a registered trademark of Random House, Inc, Copyright 1987 Borland International
BOR 0070B
Your Development Toolbox and Technical Reference Manual for Thrbo lightning@)

l I G H T N I N G

Lightning Word Wizard includes complete, commented Turbo

Pascal® source code and all the technical information you'll
need to understand and work with Turbo Lightning's "engine."
More than 20 fully documented Turbo Pascal procedures
reveal powerful Turbo Lightning engine calls. Harness the full power
of the complete and authoritative Random House® Concise
Word List and Random House Thesaurus.

Turbo Lightning's "Reference The ultimate collection of word

Manual" games and crossword solvers!
Developers can use the versatile on-line The excitement, challenge, competition,
examples to harness Turbo Lightning's and education of four games and three
power to do rapid word searches. Lightning solver utilities-puzzles, scrambles, spell-
Word Wizard is the forerunner of the data- searches, synonym-seekings, hidden words,
base access systems that will incorporate " crossword solutions, and more. You and
and engineer the Turbo Lightning Library® your friends (up to four people total) can
of electronic reference works. set the difficulty level and contest the high-
speed smarts of Lightning Word Wizard!

Turbo Lightning-Critics' Choice

"Lightning's good enough to make programmers and users cheer, executives of other
software companies weep." Jim Seymour, PC Week

"The real future of Lightning clearly lies not with the spelling checker and thesaurus currently
included, but with other uses of its powerful look-up engine." Ted Silveira, Profiles
"This newest product from Borland has it all." Don Roy, Computing Now!

Minimum system configuration: IBM PC. XT. AT. PCjr. Portable. and true compatibles. 256K RAM minimum. PC·DOS (MS·DOS) 2.0
or greater. Turbo Lightning software required. Optional-Turbo Pascal 3.0 or greater to edit and compile Turbo Pascal source code.

Suggested Retail Price: $69.95

(not copy protected)

Turbo Pascal, Turbo Lightning and Turbo Lightning Library are registered trademarks and Lightning Word Wizard is a trademark of Borland International, Inc. Random
House is a registered trademark of Random House, Inc. IBM, XT, AT, and PCjr are registered trademarks of International Business Machines Corp. MS-DOS is a
registered trademark of Microsoft Corp. CqJyright 1987 Borland International BOR00878
'EEI Ell
I TIE DATABASE
~r'~II: .ANABEI
The high-performance database manager
®

that's so advanced it's easy to use!

Lets you organize, analyze and report information faster than ever before! If you manage mailing lists,
customer files, or even your company's budgets-Reflex is the database manager for you!
Reflex is the acclaimed, high-performance database manager you've been waiting for. Reflex extends
database management with business graphics. Because a picture is often worth a 1000 words, Reflex
lets you extract critical information buried in mountains of data. With Reflex, when you look, you see.
The REPORT VIEW allows you to generate everything from mailing labels to sophisticated reports.
You can use database files created with Reflex or transferred from Lotus 1-2-3,8 dBASE,8 PFS: File,8
and other applications.

Reflex: The Critics' Choice

"... if you use a PC, you should know about Reflex ... may be the best bargain in software today."
Jerry Pournelle, BYTE
"Everyone agrees that Reflex is the best-looking database they've ever seen."
Adam B. Green, Info World
liThe next generation of software has officially arrived." Peter Norton, PC Week

Reflex: don't use your PC without it!

Join hundreds of thousands of enthusiastic Reflex users and experience the power and ease of use of
Borland's award-winning Reflex.

Suggested Retail Price: $149.95 (not copy protected)

Minimum system configuration: IBM PC, Xl, AT, and true compatibles. 384K RAM minimum. IBM Color Graphics Adapter, Hercules
Monochrome Graphics CArd, or equivalent. PC·DOS (MS· DOS) 2.0 or greater. Hard disk and mouse optional. Lotus 1·2·3, dBASE,
or PFS: File optional.
Reflex is a trademark of Bortand/Analytica tnc. Lotus 1-2-3 is a registered trademark of Lotus
Development Corporation. dBASE is a registered trademark of Ashton-Tate. PFS: Fite is a
registered trademark of Software Publishing Corporation. IBM. XT. AT, and IBM Color Graphics
Adapter are registered trademarks of International BUSiness Machines Corporation. Hercules
Graphics Card is a trademark of Hercules Computer Technology. MS-DOS is a registered
trademark of Microsoft Corp. Copyright 1987 Borland International BOA 0066C
REILEX: TIE WIIISII"·
Includes 22 "instant templates" covering a broad range of
business applications (listed below). Also shows you how to
customize databases, graphs, crosstabs, and reports. It's an invaluable
analytical tool and an important addition to another one of
our best sellers, Reflex: The Database Manager.
Fast-start tutorial examples:
Learn Reflex@) as you work with practical business applications. The Reflex Workshop Disk supplies
databases and reports large enough to illustrate the power and variety of Reflex features. Instructions in each
Reflex Workshop chapter take you through a step-by-step analysis of sample data. You then follow simple
steps to adapt the files to your own needs.
22 practical business applications:
Workshop's 22 "instant templates" give you a wide range of analytical tools:
Administration • Tracking Manufacturing Quality Assurance
• Scheduling Appointments • Analyzing Product Costs
• Planning Conference Facilities Accounting and Financial Planning
• Managing a Project • Tracking Petty Cash
• Creating a Mailing System • Entering Purchase Orders
• Managing Employment Applications • Organizing Outgoing Purchase Orders
Sales and Marketing • Analyzing Accounts Receivable
• Researching Store Check Inventory • Maintaining Letters of Credit
• Tracking Sales Leads • Reporting Business Expenses
• Summarizing Sales Trends • Managing Debits and Credits
• Analyzing Trends • Examining Leased Inventory Trends
• Tracking Fixed Assets
Production and Operations • Planning Commercial Real Estate Investment
• Summarizing Repair Turnaround

Whether you're a newcomer learning Reflex basics or an experienced "power user" looking for tips, Reflex:
The Workshop will help you quickly become an expert database analyst.
Minimum system configuration: IBM PC, AT, and XT, and true compatibles. PC-DOS (MS-DOS) 2.0 or greater. 384K RAM minimum. Requires Rellex:
The Database Manager, and IBM Color Graphics Adapter, Hercules Monochrome Graphics Card or equivalent.

Suggested Retail Price: $69.95

(not copy protected)

Reflex is a registered trademark and Reflex: The Workshop is a trademark of Borland/Analytica, Inc. IBM, AT, and Xl are registered trademarks 01 International Business
Machines Corp. Hercules is a trademark 01 Hercules Computer Technology. MS-DOS is a registered trademark 01 Microsoft Corp. Copyright 1987 Borland International
BOA 0088B
Version 3.0 with 8087 support and BCD reals

Free MicroCaic Spreadsheet With Commented Source Code!

FEATURES: THE CRITICS' CHOICE:
One-Step Compile: No hunting &fishing "Language deal of the century ... Turbo Pascal:
expeditions! Turbo finds the errors, takes you it introduces a new programming environment
to them, lets you correct them, and instantly and runs like magic."
recompiles. You're off and running in -Jell Duntemann, PC Magazine
record time.
"Most Pascal compilers barely fit on a disk, but
Built-in Interactive Editor: WordStar~like Turbo Pascal packs an editor, compiler, linker,
easy editing lets you debug quickly. and run-time library into just 39K bytes of
Automatic Overlays: Fits big programs into random access memory."
small amounts of memory. -Dave Garland, Popular Computing

MicroCalc: A sample spreadsheet on your disk "What I think the computer industry is headed
with ready-to-compile source code. for: well-documented, standard, plenty of
good features, and a reasonable price."
IBM~ PC Version: Supports Turtle Graphics, -Jerry Pournelle, BYTE
color, sound, full tree directories, window
routines, input/output redirection, and
much more.
LOOK AT TURBO NOW!
@' More than 500,000 users worldwide. @' Turbo Pascal named "Most
Significant Product of the Year" by
@' Turbo Pascal is the de facto industry PC WEEK.
standard.
@' Turbo Pascal 3.0-the fastest Pascal
@' Turbo Pascal wins PC MAGAZINE'S development environment on the
award for technical excellence. planet, period.

Suggested Retail Price: $99.95; CP/M~-80 version without 8081 and BCD: $69.95

Features lor 16-bit Systems: 8087 math co-processor support for intensive calculations.
Binary Coded Decimals (BCD): eliminates round-off error! Amust for any serious business application.

Minimum system configuration: 128K RAM minimum. Includes 8087 &BCD features for 16-bit MS-DOS 2.0 or later and
CP/M-86 1.1 or later. CP/M-80 version 2.2 or later 48K RAM minimum (8087 and BCD features not available). 8087
version requires 8087 or 80287 co-processor.

Turbo Pascal is a registered trademark of Bor1and International. Inc. CP/M is a registered trademark
of Digital Research Inc. IBM is a registered tradermrk of International Business Machines Corp.
MS-DOS is a registered tradermrk of Micr~ft Corp. VlbrdStar is a registered trademark of
MicroPro International. Copyright 1987 Borland International BOA 00619
VERSION 2.0

Learn Pascal From The Folks Who Created

The Turbo Pascal® Family
Borland International proudly presents Turbo Tutor, the perfect complement
to your Turbo Pascal compiler. Turbo Tutor is really for everyone-
even if you've never programmed before.

And if you're already proficient, Turbo Tutor can sharpen up the fine paints.
The manual and program disk focus on the whole spectrum of Turbo
Pascal programming techniques.

• For the Novice: It gives you a concise history of Pascal, tells you how to write a
simple program, and defines the basic programming terms you need to know.
• Programmer's Guide: The heart of Turbo Pascal. The manual covers the fine points
of every aspect of Turbo Pascal programming: program structure, data types, control
structures, procedures and functions, scalar types, arrays, strings, pointers, sets, files,
and records.
• Advanced Concepts: If you're an expert, you'll love the sections detailing such topics as
linked lists, trees, and graphs. You'll also find sample program examples for PC-DOS and
MS-DOS.@)
10,000 lines of commented source code, demonstrations of 20 Turbo Pascal features, multiple-
choice quizzes, an interactive on-line tutor, and more!

Turbo Tutor may be the only reference work about Pascal and programming you'll ever need!

Suggested Retail Price: $39.95 (not copy protected)

Minimum system configuration: Turbo Pascal 3.0. PC-DOS (MS-DOS) 2.0 or later. 192K RAM minimum (CP/M-80
version 2.2 or later: 64K RAM minimum).

Turbo Pascal and Turbo Tutor are regislered tademarks of Borland International Inc. CP1M is a
registered trademark of Digital Research Inc. MS-DOS is a registered trademark of Microsoft Corp.
Copyright 1987 Borland International BOR 0064C
Is The Perfect Complement To Turbo Pascal®
It contains a complete library of Pascal procedures that
allows you to sort and search your data and build powerful database
applications. It's another set of tools from Borland that will give
even the beginning programmer the expert's edge.

THE TOOLS YOU NEED!

TURBO ACCESS Using B+ trees: The best way to organize and search your data. Makes it
possible to access records in a file using key words instead of numbers. Now available with
complete source code on disk, ready to be included in your programs.
TURBO SORT: The fastest way to sort data using the QUICKSORT algorithm-the method
preferred by knowledgeable professionals. Includes source code.
GINST (General Installation Program): Gets your programs up and running on other
terminals. This feature alone will save hours of work and research. Adds tremendous value
to all your programs.
GET STARTED RIGHT AWAY.-FREE DATABASE!
Included on every Toolbox diskette is the source code to a working database which
demonstrates the power and simplicity of our Turbo Access search system. Modify it to suit
your individual needs or just compile it and run.
THE CRITICS' CHOICE!
liThe tools include a B+ tree search and a sorting system. I've seen stuff like this, but not as
well thought out, sell for hundreds of dollars." -Jerry Pournelle, BYTE MAGAZINE
liThe Turbo Database Toolbox is solid enough and useful enough to come recommended."
-Jeff Duntemann, PC TECH JOURNAL

Suggested Retail Price: $69.95 (not copy protected)

Minimum system conliguration: 128K RAM and one disk drive (CP/M·80: 48K). 16·bit systems: Turbo Pascal 2.0 or greater lor
MS·DOS or PC·DOS 2.0 or greater. Turbo Pascal 2.1 or greater lor CP/M·86 1.0 or greater. 8·bit systems: Turbo Pascal 2.0 or
greater lor CP/M·80 2.2 or greater.

BORLAND
INTERNATIONAL
Turbo Pascal and Turbo Database Toolbox are registered trademarks of Borland International
Inc. CP1M is a registered trademark of Dignal Research. Inc. MS-DOS is a registered
trademark of Microsoft Corp. Copyright 1987 Borland International BOR 00630
7VRBO MS'CAl.
GRAPHlx200tsox®
A Library of Graphics Routines for Use with Turbo Pasca/®
High-resolution graphics for your IBM" PC, AT," XT," PCjr", true PC compatibles, and the Heath
Zenith Z-100:" Comes complete with graphics window management.
Even if you're new to Turbo Pascal programming, the Turbo Pascal Graphix Toolbox will get you started
right away. It's a collection of tools that will get you right into the fascinating world of high-resolution
business graphics, including graphics window management. You get immediate, satisfying results. And
we keep Royalty out of American business because you don't pay any-even if you distribute your own
compiled programs that include all or part of the Turbo Pascal Graphix Toolbox procedures.

What you get includes:

• Complete commented source code on disk. • Full graphics window management.
• Tools for drawing simple graphics. • Two different font styles for graphic labeling.
• Tools for drawing complex graphics, including • Choice of line-drawing styles.
curves with optional smoothing. • Routines that will let you Quickly plot functions
• Routines that let you store and restore graphic and model experimental data.
images to and from disk. • And much, much more ...
• Tools allowing you to send screen images to
Epsone-compatible printers.

"While most people only talk about low-cost personal computer software, Borland has been doing
something about it. And Borland provides good technical support as part of the price."
John Markov & Paul Freiberger, syndicated columnists.
II you ever plan to create Turbo Pascal programs that make use of business graphics or scientific
graphics, you need the Turbo Pascal Graphix Toolbox.
Suggested Retail Price: $69.95 (not copy protected)
Minimum system configuration: IBM PC, Xl, AT, PCjr, true compatibles and the Heath Zenith Z-100. Turbo Pascal 3.0 or later. 192K
RAM minimum. Two disk drives and an IBM Color Graphics Adapter (CGA), IBM Enhanced Graphics Adapter (EGA), Hercules Graphics
Card or compatible.

BORLAND
INTERNATIONAL
Turbo Pascal and Turbo Graphix Toolbox are registered trademarks 01 Borland International.
Inc. IBM. XT. AT. and PCjr are registered trademarks 01 International Business Machines
Corporation. Hercules Graphics Card is a trademark 01 Hercules Computer Technology. Heath
Zenith Z-1 00 is a trademark 01 Zenith Data Systems Epson is a registered trademark 01
Epson Corp. Copyright 1987 Borland Intemational BOR 0068C
2VRBO PASCAl.
.EDJ2OR 2totsox®
It's All You Need To Build Your Own Text Editor
Or Word Processor
Build your own lightning-fast editor and incor- Create your own word processor. We provide all
porate it into your Turbo Pascale programs. the editing routines. You plug in the features you want.
Turbo Editor Toolbox gives you easy-to-install You could build aWordStar e -like editor with pull-down
modules. Now you can integrate a fast and powerful menus like Microsoft'se Word, and make it work as fast
editor into your own programs. You get the source as WordPerfect.e
code, the manual, and the know-how.
To demonstrate the tremendous power of Turbo Editor Toolbox, we give you the source code for
two sample editors:
Simple Editor A complete editor ready to include in your programs. With windows, block commands, and
memory-mapped screen routines.
MicroStar A full-blown text editor with a complete pull-down menu user interface, plus a lot more.
Modify MicroStar's pull-down menu system and include it in your Turbo Pascal programs.

The Turbo Editor Toolbox gives you all the

standard features you would expect to find
in any word processor:
• Wordwrap
, UN-delete last line
• Auto-indent
• Find and Find/Replace with options
• Set left and right margin
• Block mark, move, and copy
• Tab, insert and overstrike modes,
centering, etc. MicroStar's pull-down menus.

And Turbo Editor Toolbox has features that word processors selling for several hundred dollars can't begin to match.
Just to name a few:
[3" RAM-based editor. You can edit very large [3" Multiple windows. See and edit up to eight
files and yet editing is lightning fast. documents-or up to eight parts of the same
[3" Memory-mapped screen routines. In- document-all at the same time.
stant paging, scrolling, and text display. [3" Multitasking. Automatically save your
[3" Keyboard installation. Change control text. Plug in a digital clock, an appointment
keys from WordStar-like commands to any that alarm-see how it's done with MicroStar's
you prefer. "background" printing.
Best of all, source code is included for everything in the Editor Toolbox.
Suggested Retail Price: $69.95 (not copy protected)
Minimum system configuration: IBM PC, XT, AT, 3270, PCjr, and true compatibles. PC-DOS (MS-DOS) 2.0 or greater. 192K RAM.
You must be using Turbo Pascal 3.0 for IBM and compatibles.

Turbo Pascal and Turbo Editor Toolbox are registered trademarks of Borland International. Inc.
WordStar is a registered trademark of MicroPro International Corp. Word and MS-DOS are
registered trademarks of Microsoft Corp. WordPerfect is a trademark of Satellite Software
International. IBM, XT, AT, and PCjr are registered trademarks of International Business Machines
Corp. BOR 0067B
®

Secrets And Strategies 01 The Masters Are

Revealed For The First Time
Explore the world of state-of-the-art computer games with Turbo GameWorks. Using
easy-to-understand examples, Turbo GameWorks teaches you techniques to quickly create
your own computer games using Turbo Pascal.8 Or, for instant excitement, play the three
great computer games we've included on disk-compiled and ready to run.
TURBO CHESS
Test your chess-playing skills against your computer challenger. With Turbo GameWorks, you're on your way to
becoming a master chess player. Explore the complete Turbo Pascal source code and discover the secrets of
Turbo Chess.
"What impressed me the most was the fact that with this program you can become a computer chess analyst.
You can add new variations to the program at any time and make the program play stronger and stronger chess.
There's no limit to the fun and enjoyment of playing Turbo GameWorks Chess, and most important of all, with this
chess program there's no limit to how it can help you improve your game."
-George Koltanowski, Dean of American Chess, former President of
the United Chess Federation, and syndicated chess columnist.

TURBO BRIDGE
Now play the world's most popular card game-bridge. Play one-on-one with your computer or against up to
three other opponents. With Turbo Pascal source code, you can even program your own bidding or scoring
conventions.
"There has never been a bridge program written which plays at the expert level, and the ambitious user will
enjoy tackling that challenge, with the format already structured in the program. And for the inexperienced player,
the bridge program provides an easy-to-follow format that allows the user to start right out playing. The user can
'play bridge' against real competition without having to gather three other people."
-Kit Woolsey, writer of several articles and books on bridge,
and twice champion of the Blue Ribbon Pairs.

TURBO GO-MOKU
Prepare for battle when you challenge your computer to a game of Go-Moku-the exciting strategy game also
known as Pente." In this battle of wits, you and the computer take turns placing X's and D's on a grid of 19X19
squares until five pieces are lined up in a row. Vary the game if you like, using the source code available on your
disk.
Suggested Retail Price: $69.95 (not copy protected)
Minimum system configuration: IBM PC, XT, AT, Portable, 3270, PClr, and true compatibles. PC·DOS (MS·DOS) 2.0 or later. 192K
RAM minimum. To edit and compile the Turbo Pascal source code, you must be using Turbo Pascal 3.0 for IBM PCs and
compatibles.

~ " BORLAND Turbo Pascal and Turbo GameWorks are registered trademarks of Borland International. Inc.
Pente is a registered trademark of Parker Brothers. IBM. XT. AT. and PCjr are registered
'# I N T ERN A T ION A L trademarks of International Business Machiles Corporation. MS-DOS is a registered trademark
of Microsoft Corporation. Copyright 1987 Borland International BOR0065C
TURBO

the natural language of ArtiflCiallntelrlgenCe

Turbo Prolog brings fifth-generation supercomputer
power to your IBM®PC!
Turbo Prolog takes Turbo Prolog provides
programming into a new, a fully integrated pro-
natural, and logical gramming environment
environment like Borland's Turbo
With Turbo Prolog, Pascal,® the de facto
because of its natural, worldwide standard.
logical approach, both You get the
people new to programming complete Turbo
and professional programmers Prolog programming
can build powerful applica- system
tions such as expert systems, You get the 2DD-page
customized knowledge manual you're holding,
bases, natural language software that includes
interfaces, and smart L.~ ~<.<".~";;
••• ~_~
_____ the lightning-fast Turbo
_ _ _.............,,,,-,

information management systems. Prolog six-pass

Turbo Prolog is a declarative language which compiler and interactive editor, and the
uses deductive reasoning to solve free GeoBase natural query language
programming problems. database, which includes commented
source code on disk, ready to compile.
(GeoBase is a complete database designed
Turbo Prolog's development system and developed around U.S. geography.
ncludes: You can modify it or use it "as is.")
] A complete Prolog compiler that is a variation
of the Clocks in and Mellish Edinburgh
standard Prolog.
] A full-screen interactive editor.
Minimum system configuration: IBM PC, XT, AT, Portable, 3270, PCjr
] Support for both graphic and text windows. and true compatibles. PC-DOS (MS-DOS) 2.0 or later. 384K RAM
] All the tools that let you build your own minimum.
expert systems and AI applications with
unprecedented ease. Suggested Retail Price: $99.95
(not copy protected)
-.=..,
BORLAND
INTERNATIONAL
Turbo Prolog is a trademark and Turbo Pascal is a registered trademark of Borland International. Inc.
IBM. AT. XT. and PCjr are registered trademarks of International Business Machines Corp. MS-DOS is a
registered trademark of Microsoft Corp. Copyright 1987 Borland International BOR 0016D
-~
""II ""L"STM
11111111
Enhances Turbo Prolog with more than 80 tools
and over 8,000 lines 01 source code
Turbo Prolog, the natural language of Artificial Intelligence, is the
most popular AI package in the world with more than 100,000 users.
Our new Turbo Prolog Toolbox extends its possibilities.

The Turbo Prolog Toolbox enhances Turbo Prolog-our 5th-generation computer programming
language that brings supercomputer power to your IBM PC and compatibles-with its more than 80
tools and over 8,000 lines of source code that can be incorporated into your programs, quite easily.

Turbo Prolog Toolbox features include:

& Business graphics generation: boxes, circles, ellipses, bar charts, pie charts, scaled graphics
& Complete communications package: supports XModem protocol
& File transfers from Reflex,~ dBASE III,~ Lotus 1-2-3,~ Symphony~
& A unique parser generator: construct your own compiler or query language
& Sophisticated user -interface design tools
& 40 example programs
& Easy-to-use screen editor: design your screen layout and 1/0
& Calculated fields definition
& Over 8,000 lines of source code you can incorporate into your own programs

Suggested Retail Price: $99.95 (not copy protected)

Minimum system configuration: iBM PC, XT, AT or true compatibles. PC-DOS (MS-DOS) 2.0 or later. Requires Turbo Prolog 1.10
or higher. Dual-floppy disk drive or hard disk. 512K.

Turbo Prolog Toolbox and Turbo Prolog are trademarks of Borland International, Inc. Reflex
is a registered trademark of Borland/Analytica, Inc. dBASE III is a registered trademark of
Ashton-Tate. Lotus 1-2-3 and Symphony are registered trademarks of Lotus Development
Corp. IBM, XT. and AT are registered trademarks of International Business Machines Corp.
MS-DOS is a registered trademark of Microsoft Corp. BOR 0240
1UIIII IABIC®
The high-speed BASIC you've been waiting for!
You probably know us for our Turbo Pascale and Turbo Pr%g.'" Well, we've done
it again! We've created Turbo Basic, because BASIC doesn't have to be slow.

If BASIC taught you how to walk, Turbo Basic will teach you how to run!
With Turbo Basic, your only speed is "Full Speed Ahead"! Turbo Basic is a complete development
environment with a lightning fast compiler, an interactive editor and a trace debugging system. And
because Turbo Basic is also compatible with BASICA, chances are that you already know how to use
Turbo Basic. .
Turbo Basic ends the basic confusion
There's now one standard: Turbo Basic. And because Turbo Basic is a Borland product, the price is
right, the quality is there, and the power is at your fingertips. Turbo Basic is part of the fast-growing
Borland family of programming languages we call the "Turbo Family." And hundreds of thousands of
users are already using Borland's languages. So, welcome to a whole new generation of smart PC
users!
Free spreadsheet included with source code!
Yes, we've included MicroCalc, our sample spreadsheet, complete with source code. So you can get
started right away with a "real program." You can compile and run it "as is," or modify it.

A technical look at Turbo Basic

B Full recursion supported executable program, with separate windows
B Standard IEEE floating-point format for editing, messages, tracing, and execution
B Floating-point support, with full 8087 B Compile and run-time errors place you in
coprocessor integration. Software emulation source code where error occurred
if no 8087 present B Access to local, static and global variables
B Program size limited only by available B New long integer (32-bit) data type
memory (no 64K limitation) B Full 80-bit precision
B EGA and CGA support B Pull-down menus
B Full integration of the compiler, editor, and B Full window management

Suggested Retail Price: $99.95 (not copy protected)

Minimum system configuration: IBM PC. AT. XT or true compatibles. 256K. One floppy drive. PC-DOS (MS-DOS) 2.0 or lalel.

Turbo Basic and Turbo Pascal are registered trademarks and Turbo Prolog is a trademark of
Borland International. Inc. IBM. AT. and XT are regislered trademarks of International Business
Machines Corp. MS-DOS is a registered trademark of Microsolt Corp.
Copyright 1987 Borland International BOR 0265A
Includes tree
MicroCalc spreadsheet
with source code

A complete interactive development environment

With Turbo C, you can expect what only Borland delivers:
Quality, Speed, Power and Price. And with its compilation
speed of more than 7000 lines a minute, Turbo C makes
everything else look like an exercise in slow motion.
Turbo C: The C compiler for both amateurs and professionals
If you're just beginning and you've "kinda wanted to learn C," now's your chance to do it the easy way.
Turbo C's got everything to get you going. If you're already programming in C, switching to Turbo C will
considerably increase your productivity and help make your programs both smaller and faster.
Turbo C: a complete interactive development environment
Like Turbo Pascale and Turbo Prolog,'" Turbo C comes with an interactive editor that will show
you syntax errors right in your source code. Developing, debugging, and running a Turbo C
program is a snap!
Technical Specifications
[Y' Compiler: One-pass compiler generating native in- [iY' Development Environment: A powerful "Make" is
line code, linkable object modules and assembler. included so that managing Turbo C program
The object module format is compatible with the development is easy. Borland's fast "Turbo
PC-DOS linker. Supports small, medium, compact, Linker" is also included. Also includes pull-down
large, and huge memory model libraries. Can mix menus and windows. Can run from the environ-
models with near and far pointers. Includes ment or generate an executable file.
floating point emulator (utilizes 8087/80287 if [iY' Links with relocatable object modules created
installed). using Borland's Turbo Prolog into a
[Y' Interactive Editor: The system includes a powerful, single program.
interactive full-screen text editor. If the compiler [iY' ANSI C compatible.
detects an error, the editor automatically positions ~ Start-up routine ~ource ~ode included ..
the cursor appropriately in the source code. ~ Both. co~mand line and Integrated enVIronment
versions Included.

"Sieve" benchmark (25 Iterations)

Turbo C Microsoft PJ C Lattice C
Compile time 3.89 16.37 13.90
Compl7e and link time 9.94 29.06 27.79
Execution time 5.77 9.51 13.79
Object code size 274 297 301
Price $99.95 $450.00 $500.00
Benchmark run on a 6 Mhz IBM AT using Turbo eversion 1.0 and the Turbo Linker version 1.0; Microsoft eversion 4.0 and the
MS overlay linker version 3.51; Lattice eversion 3.1 and the MS object linker version 3.05.

Suggested Retail Price: $99.95* (not copy protected) ·Introductory offer good through July 1. 1987.

Minimum system configuration: IBM PC, XT, AT and true compatibles. PC-DOS (MS-DOS) 2.0 or later. One floppy drive. 320K.

Turbo C and Turbo Pascal are registered traoomarks and Turbo Prolog is a trademark 01 Borland
International. Inc. Microsoft C and MS-DOS ere registered trademarks 01 Microsoft Corp. lattice C
is a registered trademark 01 lattice. Inc. IBM, Xl, and AT are registered trademarks 01 International
Business Machines Corp. BOA 0243
EIIIEIA: "E BllfJlE'u
The solution to your most complex
equations-in seconds!
If you're a scientist, engineer, financial analyst, student, teacher, or any other professional working with
equations, Eureka: The Solver can do your Algebra, Trigonometry and Calculus problems in a snap.
Eureka also handles maximization and minimization problems, plots functions, generates reports, and
saves an incredible amount of time. Even if you're not a computer specialist, Eureka can help you
solve your real-world mathematical problems fast, without having to learn numerical approximation
techniques. Using Borland's famous pull-down menu design and context-sensitive help screens, Eureka
is easy to learn and easy to use-as simple as a hand-held calculator.
X + exp(X) = 10 solved instantly instead 01 eventually!
Imagine you have to "solve for X," where X + exp(X) = 10, and you don't have Eureka: The Solver.
What you do have is a problem, because it's going to take a lot of time guessing at "X." With Eureka,
there's no guessing, no dancing in the dark-you get the right answer, right now. (Ps: X = 2.0705799,
and Eureka solved that one in .4 of a second!)
How to use Eureka: The Solver
It's easy. You can then tell Eureka to
1. Enter your equation into the • Evaluate your solution
full-screen editor • Plot a graph
2. Select the "Solve" command • Generate a report, then send the output
3. Look at the answer to your printer, disk file or screen
4. You're done , • Or all of the above

Some 01 Eureka's key leatures Eureka: The Solver includes

You can key in: ~ A full-screen editor
~ A formula or formulas ~ Pull-down menus
~ A series of equations-and solve for ~ Context-sensitive Help
all variables ~ On-screen calculator
~ Constraints (like X has to be ~ Automatic 8087 math co-processor
< or = 2) chip support
~ A function to plot ~ Powerful financial functions
~ Unit conversions ~ Built-in and user-defined math and
~ Maximization and minimization problems financial functions
~ Interest Rate/Present Value calculations ~ Ability to generate reports complete with
~ Variables we call "What happens?," like plots and lists
"What happens if I change this variable to ~ Polynomial finder
21 and that variable to 27?" ~ Inequality solutions

Minimum system configuration: IBM PC, AT, XT, Portable, Suggested Retail Price: $99.95*
3270 and true compatibles. PC-DOS (MS-DOS) 2.0 and (not copy protected)
later. 384K.

Eureka: The Solver is a trademark of Borla1d International, Inc. IBM, AT, and XT are registered
trademarks of International Business Machnes Corp. MS-DOS is a registered trademark of
Microsoft Corp. Copyright 1987 Borland International BOR 0221A
'Introductory price expires July 1, 1987
"II'EII'PI®
IJJ Ift'lIJ : IIRSAIIIIER
THE BEBKTllfI
Release 2.0
Macintosh™
The most complete and comprehensive collection of
desk accessories available for your Macintosh!
Thousands of users already know that SideKick is the best collection of desk accessories available
for the Macintosh. With our new Release 2.0, the best just got better.

We've just added two powerful high-performance tools to SideKick-Outlook": The Outliner
and MacPlan'": The Spreadsheet. They work in perfect harmony with each other and while you
run other programs!
Outlook: The Outliner
• II's the desk accessory with more power than a stand-alone outliner
• A great desktop publishing tool, Outlook lets you incorporate both text and graphics
into your outlines
• Works hand-in-hand with MacPlan
• Allows you to work on several outlines at the same time
MacPlan: The Spreadsheet
• Integrates spreadsheets and graphs
• Does both formulas and straight numbers
• Graph types include bar charts, stacked bar charts, pie charts and line graphs
• Includes 12 example templates free!
• Pastes graphics and data right into Outlook creating professional memos and reports, complete
with headers and footers.

SideKick: The Desktop Organizer,

Release 2.0 now includes
~ Outlook: The Outliner
~ MacPlan: The Spreadsheet • 1367'1: ~~s ~
o
~ Mini word processor 1'!i.94'S SaluB
II 2'361~ Tot.IR.YfO~H

~ Calendar III
•
'"
em ExplIf'Is.c

~
[] 0319iLJ,bor
Phone Log []I
~
<4.E.6S1 H.-.... iII's
6.,,1a Ovtrtlt"
~ Analog clock 0" 18'£ Tot.,ExpfOSK

~ Alarm system IJ
•
'"
1845'1: ".tPro(11

~ Calculator
~ Report generator MacPlan does both spreadsheets and business
~ Telecommunications (new version now graphs. Paste them into your Outlook files and
supports XModem file transfer protocol) generate professional reports.

Suggested Retail Price: $99.95 (not copy protected)

Minimum system configurations: Macintosh 512K or Macintosh Plus with one disk drive. One BOOK or two 400K drives are recommended.
With one 400K drive, a limited number of desk accessories will be installable per disk.

SideKick is a registered trademark and Outlook and MacPlan are trademarks of Borland
International, Inc. Macintosh is a trademark of Mcintosh Laboratory, Inc. licensed to Apple
Computer, Inc. Copyright 1987 Borland International BOR 00690
'I £IIEI;• IAIIASEI
~.]
111 IIA1AIA"
The easy-to-use relational database that thinks like a spreadsheet.
Rellex lor the Mac lets you crunch numbers by entering lormulas and link
databases by drawing on-screen lines.
5 free ready-to-use templates are included on the examples disk:

• A sample 1040 tax application with Sched- • A client billing application set up for a law
ule A, Schedule 8, and Schedule 0, each office, but easily customized by any
contained in a separate report document. professional who bills time.
• A portfolio analysis application with linked • A parts explosion application that breaks
databases of stock purchases, sales, and down an object into its component parts
dividend payments. for cost analysis.
• A checkbook application.

Reflex for the Mac accomplishes all of these tasks without programming-using
spreadsheet-like formulas. Some other Reflex for the Mac features are:
• Visual database design. • Data types which include variable length text, number,
• "What you see is what you get" report and form layout integer, automatically incremented sequence number,
with pictures. date, time, and logical.
• Automatic restructuring of database files when data • Up to 255 fields per record.
types are changed, or fields are added and deleted. • Up to 16 files simultaneously open.
• Display formats which include General, Decimal, • Up to 16 Mac fonts and styles are selectable for
Scientific, Dollars, Percent. individual fields and labels.
• file Edit form.t OllCrlb. D,"'''' S•• n .. Milt Window
OltlB . . IOUflr"llw

<om""'''I.''''
Compon.ntOfl.llg IO_ltnembl
~:;: ___
.1
MII_ .. p,rt
ISP.~t I.to.
"Peru.S

After opening the "Overview" window, you The link lines you draw establish both You can have multiple windows open
draw link lines between databases directly visual and electronic relationships between simultaneously to view all members 01 a
onto your Macintosh screen. your databases. linked set-which are interactive and truly
relational.
Critic's Choice
"... a powerful relational database ... uses a visual approach to information management." InfoWorld
"... gives you a lot of freedom in report design; you can even import graphics." A+ Magazine
". '.' bridges the gap between the pretty programs and the power programs." Stewart Alsop, PC Letter

Suggested Retail Price: $99.95*

(not copy protected)
Minimum syslem configuration: Macirtosh 512K or Macintosh Plus with one disk drive. Second external drive recommended.
Reflex is a registered trademark of BorlandiAnalytica. Inc. MaCintosh is a trademark of MCintosh Laboratory. Inc. and is used with express permission of as owner.
Copyright 1987 Borland Internationat BOR0149A
'Introductory price expires July 1. 1987
The ultimate Pascal development environment
Borland's new Turbo Pascal for the Mac is so incredibly fast that it can
compile 1,420 lines of source code in the 7.1 seconds it took you to read this!
And reading the rest of this takes about 5 minutes, which is plenty of time for Turbo Pascal for the Mac
to compile at least 60,000 more lines of source code!
Turbo Pascal for the Mac does both Windows and "Units"
The separate compilation of routines offered by Turbo Pascal for the Mac creates modules called "Units,"
which can be linked to any Turbo Pascal program. This "modular pathway" gives you "pieces" which can
then be integrated into larger programs. You get a more efficient use of memory and a reduction in the
time it takes to develop large programs.
Turbo Pascal for the Mac is so compatible with Lisae that they should be living together
Routines from Macintosh Programmer's Workshop Pascal and Inside Macintosh can be compiled and run
with only the subtlest changes. Turbo Pascal for the Mac is also compatible with the Hierarchical File
System of the Macintosh.

The 27-second Guide to Turbo Pascal for the Mac

• Compilation speed of more than 12,000 lines Workshop Pascal (with minimal changes)
per minute • Compatibility with Hierarchical File System of
• "Unit" structure lets you create programs in your Mac
modular form • Ability to define default volume and folder names
• Multiple editing windows-up to 8 at once used in compiler directives
• Compilation options include compiling to disk or • Search and change features in the editor speed up
memory, or compile and run and simplify alteration of routines
• No need to switch between programs to compile • Ability to use all available Macintosh memory
or run a program without limit
• Streamlined development and debugging • "Units" included to call all the routines provided by
• Compatibility with Macintosh Programmer's Macintosh Toolbox

Suggested Retail Price: $99.95* (not copy protected)

·Introductory price expires July 1. 1987

Minimum system configuration: Macintosh 512K or Macintosh Plus with one disk drive.

Turbo Pascal and SideKick are registered trademarks of Borland International, Inc. and Reflex is a
registered trademark of Borlandl Analytica, Inc. Macintosh is a trademark 01 Mcintosh Laboratories, Inc. licensed
to Apple Computer with its express permission. Lisa is a registered trademark of Apple Computer, Inc. Inside
Macintosh is a copyright of Apple Computer, Inc.
Copyright 1987 Borland International BOR 0167A
-
Borland
Software
OllD.Bll fODAY I
_________ ~ I

j
j
==:115,~O
-: L
BORLAND INTERNATIONAL
IIIIi
J 4585 Scotts Valley Drive Scotts Valley, California 95066 I'

J In Ii
I To Orde~ ,.- ..., California
By Credit call
r I'
I'

I Card, ' ... , ' (800) 'i

I Call
(800)
742-1133 'I'
I 255- 8008 In Canada call
(800) 237-1136 i
i
I

--------11
I

I
BOA 0234 I
I
NOTES
NOTES
NOTES
ew from Borland's Scientific & Engineering Division, Turbo
, Pascal Numerical Methods Toolbox implements the latest
high-level mathematical methods to solve common scientific
and engineering problems. Fast.
So every time you need to calcu- • Least squares approximations
late an integral, work with Fourier • Fourier transforms
Transforms or incorporate any of the
classical numerical analysis tools into 5 free ways to look at
your programs, you don't have to "Least Squares Fit"!
reinvent the wheel. Because the As well as a free demo "Fast
Numerical Methods Toolbox is a Fourier Transforms," you also get
complete collection of Turbo Pascal "Least Squares Fit" in 5 different
routines and programs that gives you forms-which gives you 5
applied state-of-the-art math tools. different methods of fitting curves
It also includes two graphics demo to a collection of data pOints.
programs, Least Squares Fit and Fast You instantly get the picture 1 The
Fourier Transforms, to give you the
5 different forms are:
picture along with the numbers.
The Numerical Methods Toolbox is 1. Power 4. 5-term Fourier
a must for you if you're involved with 2. Exponential 5. 5-term
any type of scientific or engineering 3. Logarithm Polynomial
computing. Because it comes with They're all ready to compile
complete source code, you have total and run "as is." To modify or
control of your application. add graphics to your own
programs, you simply add Turbo
What Numerical Methods
Toolbox will do for you now: Graphix Toolbox" to your soft-
• Find solutions to equations ware library. Our Numerical
• Interpolations I
Methods Toolbox is designed to
• Calculus: numerical derivatives work hand-in-hand with our
and integrals Turbo Graphix Toolbox to make
• Matrix operations: inversions, professional graphics in your
determinants and eigenvalues own programs an instant part of
• Differential equations the picture!
Minimum s,stem conllgul.don: IBM PC. XT, AT and true compatibles. PC-DOS (MS-OOSI 2.0 or later 256K. Turbo
Pascal 2.0 or later. The graphics modules reQuire a graphics monilor with an IBM CGA. IBM EGA, or Hercules compatible
adapter card, and reQuire the Turbo Graph~ Toolbox. MS-DOS generic version will not support Turbo Graphix Toolbox
routines. An 8087 or 80287 numeric co-processor is not reQuired, bul recommended lor optimal perlormance.
Turbo Pascal Numerical Methods Toolbox ~ a trademark and Turbo Pascal and Turbo Graphix Toolbox are registered
trademarks 01 Borland International, Inc. IBM, XT, and AT are registered trademarks of International Business Machines
Corp. MS-DOS is a registered trademark 01 Microso" Corp. Hercules is a trademark 01 Hercules Compuler TechnOlogy.
Apple is a registered trademark 01 Apple Computer, Inc. Macintosh is a trademark 01 Mcintosh Laboralory. Inc. licensed
10 Apple Computer. Copyright 1986 Borland Inlernational BOA 0224