Métodos Numéricos en Ing. Química para Programar

Contents
1 On models 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Classification of chemical process models . . . . . . . . . 4
1.3 Lumped parameter, steady state models . . . . . . . . . . 5
1.3.1 Example of a stagewise separation process . . . . 5
1.3.2 Process flow sheet simulation . . . . . . . . . . . . 9
1.3.3 Example of a multicomponent flash . . . . . . . . . 11
1.3.4 Example of a phenomenalogical model . . . . . . . 13
1.3.5 Example of reactors in series . . . . . . . . . . . . . 14
1.4 Lumped parameter, dynamic models . . . . . . . . . . . . 15
1.4.1 Example of cooling a molten metal . . . . . . . . . 15
1.4.2 Ozone decomposition . . . . . . . . . . . . . . . . . 16
1.5 Distributed parameter, steady state models . . . . . . . . 17
1.5.1 Heat transfer through a tappered fin . . . . . . . . 17
1.6 Distributed parameter, dynamic models . . . . . . . . . . 20
1.6.1 Heat transfer through a tappered fin . . . . . . . . 20
2 Single nonlinear algebraic equation 22

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2 Bisection method . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 Regula-falsi method . . . . . . . . . . . . . . . . . . . . . . 29
2.4 Secant method . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.5 Newton’s method . . . . . . . . . . . . . . . . . . . . . . . . 31
2.6 Muller’s method . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.7 Fixed point iteration . . . . . . . . . . . . . . . . . . . . . . 34
2.8 Error analysis and convergence acceleration . . . . . . . . 37
2.8.1 Convergence of Newton scheme . . . . . . . . . . . 40
2.9 Deflation technique . . . . . . . . . . . . . . . . . . . . . . . 41
2.10 Parameter continuation . . . . . . . . . . . . . . . . . . . . 43
2.10.1 Euler-Newton continuation . . . . . . . . . . . . . . 43
2.10.2 Homotopy continuation . . . . . . . . . . . . . . . . 43
i
CONTENTS ii
2.11 Software tools . . . . . . . . . . . . . . . . . . . . . . . . . . 43

2.11.1 MATLAB . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.11.2 Mathematika . . . . . . . . . . . . . . . . . . . . . . . 44
2.12 Exercise problems . . . . . . . . . . . . . . . . . . . . . . . 45
2.12.1 Multicomponent, isothermal flash model . . . . . . 45
2.12.2 Compressible flow in a pipe . . . . . . . . . . . . . 46
2.12.3 A model for separation processes . . . . . . . . . . 47
2.12.4 Peng-Robinson Equation of State . . . . . . . . . . . 49
2.12.5 Transient heat conduction in semi-infinite slab . . 50
2.12.6 Turbulent flow in a parallel pipeline system . . . . 51
3 Review of Linear Algebra 53

3.1 Matrix notation . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.1.1 Basic operations . . . . . . . . . . . . . . . . . . . . . 54
3.2 Matrices with special structure . . . . . . . . . . . . . . . . 57
3.3 Determinant . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.3.1 Laplace expansion of the determinant . . . . . . . 59
3.4 Direct methods . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.4.1 Cramers rule . . . . . . . . . . . . . . . . . . . . . . . 60
3.4.2 Matrix inverse . . . . . . . . . . . . . . . . . . . . . . 61
3.4.3 Gaussian elimination . . . . . . . . . . . . . . . . . . 64
3.4.4 Thomas algorithm . . . . . . . . . . . . . . . . . . . 70
3.4.5 Gaussian elimination - Symbolic representaion . . 71
3.4.6 LU decomposition . . . . . . . . . . . . . . . . . . . 75
3.5 Iterative methods . . . . . . . . . . . . . . . . . . . . . . . . 78
3.5.1 Jacobi iteration . . . . . . . . . . . . . . . . . . . . . 80
3.5.2 Gauss-Seidel iteration . . . . . . . . . . . . . . . . . 81
3.5.3 Successive over-relaxation (SOR) scheme . . . . . . 82
3.5.4 Iterative refinement of direct solutions . . . . . . . 84
3.6 Gram-Schmidt orthogonalization procedure . . . . . . . . 85
3.7 The eigenvalue problem . . . . . . . . . . . . . . . . . . . . 86
3.7.1 Wei-Prater analysis of a reaction system . . . . . . 86
3.8 Singular value decomposition . . . . . . . . . . . . . . . . 86
3.9 Genaralized inverse . . . . . . . . . . . . . . . . . . . . . . . 86
3.10 Power iteration . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.11 Software tools . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.11.1 Lapack, Eispack library . . . . . . . . . . . . . . . . . 86
3.11.2 MATLAB . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.11.3 Mathematica . . . . . . . . . . . . . . . . . . . . . . . 86
3.12.1 Laminar flow through a pipeline network . . . . . 86
CONTENTS iii
4 Nonlinear equations 89
4.1 Newton’s method . . . . . . . . . . . . . . . . . . . . . . . . 90
4.2 Euler-Newton continuation . . . . . . . . . . . . . . . . . . 96
4.3 Arc-length continuation . . . . . . . . . . . . . . . . . . . . 96
4.4 Quasi-Newton methods . . . . . . . . . . . . . . . . . . . . 96
4.4.1 Levenberg-Marquadt method . . . . . . . . . . . . . 96
4.4.2 Steepest descent method . . . . . . . . . . . . . . . 96
4.4.3 Broyden’s method . . . . . . . . . . . . . . . . . . . 96
4.5.1 Turbulent flow through a pipeline network . . . . 96
5 Functional approximations 98
5.1 Approximate representation of functions . . . . . . . . . 99
5.1.1 Series expansion . . . . . . . . . . . . . . . . . . . . 99
5.1.2 Polynomial approximation . . . . . . . . . . . . . . 99
5.2 Approximate representation of data . . . . . . . . . . . . . 104
5.2.1 Least squares approximation . . . . . . . . . . . . . 105
5.3 Difference operators . . . . . . . . . . . . . . . . . . . . . . 107
5.3.1 Operator algebra . . . . . . . . . . . . . . . . . . . . 108
5.3.2 Newton forward difference approximation . . . . . 109
5.3.3 Newton backward difference approximation . . . . 112
5.4 Inverse interpolation . . . . . . . . . . . . . . . . . . . . . . 116
5.5 Lagrange polynomials . . . . . . . . . . . . . . . . . . . . . 118
5.6 Numerical differentiation . . . . . . . . . . . . . . . . . . . 122
5.6.1 Approximations for first order derivatives . . . . . 122
5.6.2 Approximations for second order derivatives . . . 124
5.6.3 Taylor series approach . . . . . . . . . . . . . . . . . 126
5.7 Numerical integration . . . . . . . . . . . . . . . . . . . . . 126
5.7.1 Romberg Extrapolation . . . . . . . . . . . . . . . . 129
5.7.2 Gaussian quadratures . . . . . . . . . . . . . . . . . 132
5.7.3 Multiple integrals . . . . . . . . . . . . . . . . . . . . 132
5.8 Orthogonal functions . . . . . . . . . . . . . . . . . . . . . 132
5.9 Piecewise continuous functions - splines . . . . . . . . . . 132
6 Ordinary differential equations - Initial value problems 133

6.1 Model equations and initial conditions . . . . . . . . . . . 133
6.1.1 Higher order differential equations . . . . . . . . . 134
6.2 Taylor series expansion . . . . . . . . . . . . . . . . . . . . 135
6.2.1 Alternate derivation using interpolation polynomials 135
6.2.2 Stability limits . . . . . . . . . . . . . . . . . . . . . . 137
6.2.3 Stiff differential equations . . . . . . . . . . . . . . 139
6.3 Multistep methods . . . . . . . . . . . . . . . . . . . . . . . 141
CONTENTS iv
6.3.1 Explicit schemes . . . . . . . . . . . . . . . . . . . . 141

6.3.2 Implicit schemes . . . . . . . . . . . . . . . . . . . . 142
6.3.3 Automatic stepsize control . . . . . . . . . . . . . . 143
6.4 Runge-Kutta Methods . . . . . . . . . . . . . . . . . . . . . 145
6.4.1 Explicit schemes . . . . . . . . . . . . . . . . . . . . 146
6.4.2 Euler formula revisited . . . . . . . . . . . . . . . . 147
6.4.3 A two-stage (v = 2) Runge-Kutta scheme . . . . . . 147
6.4.4 Semi-implicit & implicit schemes . . . . . . . . . . . 150
6.4.5 Semi-Implicit forms of Rosenbrock . . . . . . . . . 151
6.5 Dynamical systems theory . . . . . . . . . . . . . . . . . . 152
7 Ordinary differential equations - Boundary value problems 153

7.1 Model equations and boundary conditions . . . . . . . . . 153
7.2 Finite difference method . . . . . . . . . . . . . . . . . . . . 156
7.2.1 Linear problem with constant coefficients . . . . . 156
7.2.2 Linear problem with variable coefficients . . . . . . 157
7.2.3 Nonlinear problem . . . . . . . . . . . . . . . . . . . 159
7.3 Quasilinearization of nonlinear equations . . . . . . . . . 161
7.4 Control volume method . . . . . . . . . . . . . . . . . . . . 163
7.5 Shooting method . . . . . . . . . . . . . . . . . . . . . . . . 163
7.6 Collocation methods . . . . . . . . . . . . . . . . . . . . . . 163
7.7 Method of weighted residuals . . . . . . . . . . . . . . . . 163
A ChE Network A.1

A.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . A.1
A.2 Userid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.5
A.3 Overview of the network . . . . . . . . . . . . . . . . . . . . A.5
A.4 The client-server model . . . . . . . . . . . . . . . . . . . . A.8
A.4.1 File servers . . . . . . . . . . . . . . . . . . . . . . . . A.8
A.4.2 License servers . . . . . . . . . . . . . . . . . . . . . A.9
A.4.3 X servers . . . . . . . . . . . . . . . . . . . . . . . . . A.11
A.4.4 News server . . . . . . . . . . . . . . . . . . . . . . . A.13
A.4.5 Gopher service . . . . . . . . . . . . . . . . . . . . . A.15
A.4.6 World wide web service . . . . . . . . . . . . . . . . A.16
A.5 Communication . . . . . . . . . . . . . . . . . . . . . . . . . A.16
A.5.1 Software tools . . . . . . . . . . . . . . . . . . . . . . A.16
A.5.2 Connection to the AIX machines from OS/2 machines A.17
A.5.3 Connection to the AIX machines from a home com-
puter . . . . . . . . . . . . . . . . . . . . . . . . . . . A.19
A.5.4 Connection to the AIX machine from a DOS machine
using tn3270 . . . . . . . . . . . . . . . . . . . . . . . A.20
A.5.5 File transfer with Kermit . . . . . . . . . . . . . . . A.21
CONTENTS v
A.5.6 File transfer with ftp . . . . . . . . . . . . . . . . . . A.22

A.5.7 File transfer from AIX to OS/2 machines . . . . . . A.23
A.6 Operating systems . . . . . . . . . . . . . . . . . . . . . . . A.23
A.6.1 How to signon and logout . . . . . . . . . . . . . . . A.23
A.6.2 Customizing your environment - the .profile file . A.24
A.6.3 File management . . . . . . . . . . . . . . . . . . . . A.24
A.6.4 How to get online help . . . . . . . . . . . . . . . . . A.25
A.7 Editors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.26
A.7.1 Emacs - the ultimate in editors . . . . . . . . . . . . A.26
A.8 Fortran compilers . . . . . . . . . . . . . . . . . . . . . . . . A.30
A.9 Debuggers . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.31
A.10Application programs . . . . . . . . . . . . . . . . . . . . . A.32
A.10.1 ASPEN PLUS . . . . . . . . . . . . . . . . . . . . . . . A.32
A.10.2 Xmgr . . . . . . . . . . . . . . . . . . . . . . . . . . . A.32
A.10.3 TEX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.33
A.10.4 pine - the mail reader . . . . . . . . . . . . . . . . . A.33
A.10.5 tin - the news reader . . . . . . . . . . . . . . . . . . A.34
A.11Distributed Queueing System . . . . . . . . . . . . . . . . . A.34
A.11.1 Example of a CPU intensive MATLAB job . . . . . . A.35
A.11.2 Example of a FLOW3D job . . . . . . . . . . . . . . . A.36
A.12Printing reports, graphs etc. . . . . . . . . . . . . . . . . . A.36
A.12.1 Using the network printer from AIX machines . . A.38
A.12.2 Using the network printer from OS/2 machines . . A.39
A.13Anonymous ftp service . . . . . . . . . . . . . . . . . . . . A.39
B MATLAB primer B.1

B.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . B.1
B.2 Starting a MATLAB session . . . . . . . . . . . . . . . . . . B.2
B.2.1 Direct access on AIX machines . . . . . . . . . . . . B.2
B.2.2 Remote access from OS/2 machines or home com-
puter . . . . . . . . . . . . . . . . . . . . . . . . . . . B.3
B.3 MATLAB basics . . . . . . . . . . . . . . . . . . . . . . . . . B.4
B.3.1 Using built in HELP, DEMO features . . . . . . . . . B.5
B.3.2 Data entry, line editing features of MATLAB . . . . B.8
B.3.3 Linear algebra related functions in MATLAB . . . . B.13
B.3.4 Root finding . . . . . . . . . . . . . . . . . . . . . . . B.15
B.3.5 Curve fitting . . . . . . . . . . . . . . . . . . . . . . . B.16
B.3.6 Numerical integration, ordinary differential equations B.16
B.3.7 Basic graphics capabilities . . . . . . . . . . . . . . B.17
B.3.8 Control System Toolbox . . . . . . . . . . . . . . . . B.19
B.3.9 Producing printed output of a MATLAB session . . B.21
B.3.10 What are m-files? . . . . . . . . . . . . . . . . . . . . B.23
CONTENTS vi
B.3.11 Programming features . . . . . . . . . . . . . . . . . B.25

B.3.12 Debugging tools . . . . . . . . . . . . . . . . . . . . . B.26
C Unix primer C.1

C.1 Introduction to the shell and the desktop . . . . . . . . . C.1
C.1.1 The ".profile" file . . . . . . . . . . . . . . . . . . . . C.2
C.2 Managing files . . . . . . . . . . . . . . . . . . . . . . . . . . C.4
C.2.1 Making sense of the directory listing - the "ls" command
C.4
C.2.2 Changing permission on files - the "chmod" command C.6
C.2.3 Moving files . . . . . . . . . . . . . . . . . . . . . . . C.6
C.2.4 Copying files . . . . . . . . . . . . . . . . . . . . . . C.7
C.2.5 Changing ownership of files - the "chown" command C.8
C.2.6 Compressing files - the "compress" command . . C.8
C.2.7 Removing files - the "rm" command . . . . . . . . C.8
C.3 Managing processes . . . . . . . . . . . . . . . . . . . . . . C.9
C.3.1 Examining jobs or processes - the "ps" command C.9
C.3.2 Suspending jobs or processes . . . . . . . . . . . . C.10
C.3.3 Terminating jobs or processes - the "kill" command C.10
C.3.4 Initiating background jobs - the "nohup" command C.10
C.3.5 Script files & scheduling jobs - the "at" command C.11
C.4 List of other useful AIX command . . . . . . . . . . . . . . C.12
Bibliography R.1
List of Tables
2.1 Example of fixed point iteration . . . . . . . . . . . . . . . 37

2.2 Feed composition & equilibrium ratio of a natural gas mix-
ture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.1 Loss of precision and need for pivoting . . . . . . . . . . . 71
5.1 Convergence of P2n+1 (x) to er f (x) at selected values of x 100

5.2 Saturation temperature vs. pressure from steam tables . 105
5.3 Inverse interpolation . . . . . . . . . . . . . . . . . . . . . . 118
5.4 Summary of difference approximations for derivatives . 124
A.1 List of machines in the chemical and materials engineering

network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.7
A.2 List of software on the chemical and materials engineering
network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.12
A.3 List of ftp commands . . . . . . . . . . . . . . . . . . . . . A.22
A.4 List of frequently used emacs functions and their key bind-
ings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.29
B.1 List of MATLAB Ver 4.0 help topics . . . . . . . . . . . . . B.6

B.2 General purpose MATLAB Ver 4.0 commands . . . . . . . B.7
B.3 Linear algebra related functions in MATLAB . . . . . . . . B.14
B.4 Graphics related function in MATLAB . . . . . . . . . . . . B.18
B.5 List of functions from control system tool box . . . . . . B.22
B.6 Program control related help topics . . . . . . . . . . . . . B.27
C.1 List of other useful AIX command . . . . . . . . . . . . . . C.13
vii
List of Figures
1.1 Scheme of things - a personal view . . . . . . . . . . . . . 2

1.2 Three stage separation process . . . . . . . . . . . . . . . . 5
1.3 Linear and nonlinear equilibrium model . . . . . . . . . . 8
1.4 Example of material balance equations in a process flow
sheet. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.5 Multicomponent, isothermal flash process . . . . . . . . . 12
1.6 Reactors in series . . . . . . . . . . . . . . . . . . . . . . . . 14
1.7 Heat transfer from a molten metal . . . . . . . . . . . . . 16
1.8 Heat transfer through a fin . . . . . . . . . . . . . . . . . . 18
2.1 Graphs of some simple functions . . . . . . . . . . . . . . 23

2.2 Graphical representation of some simple root finding algo-
rithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3 MATLAB implementation of the bisection algorithm . . . 27
2.4 MATLAB implementation of the secant algorithm . . . . 30
2.5 Graphical representation of Muller’s scheme . . . . . . . 33
2.6 Graphical representation of fixed point scheme . . . . . . 36
2.7 Graphical illustration of mean value theorem . . . . . . . 38
2.8 Graphical illustration deflation technique . . . . . . . . . 42
2.9 Multicomponent flash function . . . . . . . . . . . . . . . . 45
2.10 Turbulent flow in a parallel pipe . . . . . . . . . . . . . . . 51
3.1 MATLAB implementation of inverse of an upper triangular

matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.2 Naive Gaussian elimination shceme . . . . . . . . . . . . . 67
3.3 MATLAB implementation of naive Gaussian elimination . 68
3.4 Thomas algorithm . . . . . . . . . . . . . . . . . . . . . . . 72
3.5 MATLAB implementation of Thomas algorithm . . . . . . 73
3.6 MATLAB implementation of LU decomposition algorithm 77
3.7 MATLAB implementation of Gauss-Seidel algorithm . . . 83
3.8 Laminar flow in a pipe network . . . . . . . . . . . . . . . 86
viii
LIST OF FIGURES 1
4.1 MATLAB implementation of Newton scheme . . . . . . . 93

4.2 CSTR in series example - function & Jacobian evaluation 95
5.1 Error distribution of (x, n) := |er f (x) − P2n+1 (x)| for

different levels of truncation . . . . . . . . . . . . . . . . . 100
5.2 MATLAB implementation illustrating steps of functional ap-
proximation . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.3 Structure of Newton forward difference table for m equally
spaced data . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.4 Example of a Newton forward difference table . . . . . . 112
5.5 Structure of Newton backward difference table for 5 equally
spaced data . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.6 Example of a Newton backward difference table . . . . . 115
5.7 Example of a inverse interpolation . . . . . . . . . . . . . . 116
5.8 Structure of divided difference table for 4 unequally spaced
data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.9 MATLAB implementation of Lagrange interpolation polyno-
mial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.10 Illustration of Romberg extrapolation . . . . . . . . . . . . 131
5.11 MATLAB implementation of quadrature evaluation . . . . 132
6.1 Spring and dash pot model . . . . . . . . . . . . . . . . . . 139

6.2 Stepsize control strategies for multistep methods . . . . 144
6.3 MATLAB implementation of ozone decomposition model 145
6.4 Results of ozone decomposition model shows a stiff system
behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
7.1 One dimensional finite difference grid of equally spaced

data points . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
7.2 One dimensional control volume discretization . . . . . . 163
A.1 Anatomy of a Motif window . . . . . . . . . . . . . . . . . . A.4

A.2 Logical dependencies between various machines in the net-
work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.6
A.3 Network File System server and clients . . . . . . . . . . . A.9
A.4 Network Information Service . . . . . . . . . . . . . . . . . A.10
A.5 X-Windows client-server model . . . . . . . . . . . . . . . . A.14
A.6 Anatomy of an emacs window . . . . . . . . . . . . . . . . A.27
C.1 Output of the "ls" command . . . . . . . . . . . . . . . . . C.5

To see a World in a Grain of Sand,
And a Heaven in a Wild Flower,
Hold Infinity in the palm of your hand,
And Eternity in an hour
— WILLIAM BLAKE
Chapter 1
The physical, mathematical and

computational models
1.1 Introduction
In the Trekie language, the prime directive of a chemical engineer should

be explore and understand physical and chemical processes involved in
converting a raw material into a useful product. Use this knowledge in
designing, constructing and operating chemical process plants. This def-
inition is as arbitrary as anything else that one might propose. In fact if
one substitutes the word chemical in the above definition with any other
(such a mechanical or electrical) it would remain equally valid. This is
because the basic principles and the scientific methodology we use to
uncover such principles remain the same in any field of engineering or
science. A broader, although highly personal, view of the our attempt
to understand the physical world, describe it in the language of math-
ematics, and finally investigate its consequence by means of analytical,
graphics or numerical methods is shown in Figure 1.1.
A mathematical model is at best an approximation to the physical
world. Such models are constructed based on certain conservation prin-
ciples and/or empirical observations. Those curious about the nature of
the physical laws should read a delightful little book by Feynman (1967)
on the character of physical laws. As a matter of convenience mathemat-
ical models can be classified as linear or non-linear, steady-state or dy-
1
1.1. INTRODUCTION 2
For Chemical engineers

this world consists of process
plants involving transport,
reaction & equilibrium
Physical world processes
Conceptual modelling Empirical observations

based on based on
thought experiments (physcical experiments)
Introduce concepts like force, energy etc.
(need a good eye)
(need a good mind)
Construct Conservation laws Construct empirical

(e.g. mass, energy) relationships like y=f(x)
or between observed (y) and
Phenomenological models manipulated (x) variables
(e.g. equation of state) (curve fitting)
Classification I
Linear models
In light of ignorance we tend tend
to assume linearity
Mathematical model
(dating back to flat earth model!) (an approximation to the physical world)
Non-linear models Classification II
In chemical process models
nonlinearity arises primarily in the Steady state models
description of equilibrium models Lumped - algebraic equations
(phase behavior), reaction rate Distributed - ordinary differential equations(BVP)
models or fluid flow models. - partial differential equations (elliptic)
Before the computer era (<1950)
graphical methods were often used Dynamic models
to solve such problems. Lumped - ordinary differential equations(IVP)
They have been replaced largely Distributed - partial differential equations (parabolic)
by computer simulation tools.
Numerical model
Input (an approximation to the mathematical model) Output
examples
ASPEN, HYSIM,
PROCESS, FLOW3D,
HCOMP, SPEEDUP,
DREM
Figure 1.1: Scheme of things - a personal view

1.1. INTRODUCTION 3
namic, lumped or distributed. Examples to illustrate each type of model

are provided later in this chapter. In general, non-linearity is found to
occur quite naturally and frequently in nature; it is also very difficult to
analyse non-linear models without the aid of a computer.
A numerical model (or a computer simulation tool) is an approxima-
tion to the mathematical model. Although the importance of mathemat-
ical modelling in chemical engineering was recognized since the early
1920’s, it was the text by Bird et al. (1960) on Transport Phenomena
that has proved to be the major inspiration in exploring the link between
the physical world and the mathematical one for transport processes in-
volving momentum, heat and mass transfer. Since then a number of
outstanding texts have appeared that explore this relationship for reac-
tion and equilibrium processes as well. While such studies form the core
of chemical engineering curriculum, the importance of sharpening our
mathematical abilities, and the need to incorporate these as part of the
curriculum was recognized and aided by the appearance of early text
books by Amundson (1966) and Jenson & Jeffreys (1963). These dealt
specifically with mathematical applications of chemical engineering. The
texts by Lapidus (1962) and Rosenbrock (1966) served a similar purpose
in introducing digital computational methods in the analysis of chemical
processes.
We are now at a new threshold; computers have become quite per-
vasive. Significant advances have been made in our ability to analyse
non-linear systems. The advances in both the hardware and software
technology have been revolutionary. As a result of these advances, com-
puter aided design and analysis has become a standard tool as evidenced
by the success of several commercial packages such as ASPEN PLUS, PRO-
CESS, HYSIM (steady state process simulators), FLOW3D, FLUENT (fluid
dynamics simulators), HCOMP, DREM (multiphase and transient pipeline
flow simulators) etc. In addition to such simulators that are specific for
certain classes of problems, general purpose mathematical tools such as
MATLAB (for matrix linear algebra functions), Mathematika and MAPLE
(for symbolic computation) provide easy access to a vast array of mathe-
matical functions and the ability to process them both symbolically and
numerically. Such packaged tools tend to accomplish the following: (i)
codify the most advanced algorithms, (ii) assemble a vast database (in
the case of physical properties) and knowledge base in the case of math-
ematical functions (in MAPLE and Mathematika) and (iii) make these ac-
cessible to the end user through an intuitive interface. While this puts a
lot of power at the hands of the end user, in order to use them wisely and
interpret the results correctly, the users are expected to have a sound
knowledge of the relationship between the physical world and the mathe-
1.2. CLASSIFICATION OF CHEMICAL PROCESS MODELS 4
matical model and that between the mathematical model and the numer-
ical approximation. One is well served to remember the cliche garbage
in, garbage out!
In this course we examine the link between the mathematical and
the numerical model. In so doing modern computational tools are used
quite liberally. Concepts of a networked computer environment are dis-
cussed in Appendix A. A basic introduction to MATLAB can be found
in Appendix B. MATLAB is used throughout in illustrating various algo-
rithms.
1.2 Classification of chemical process models
In modelling chemical processes, one is interested in tracking material

and energy of process streams from the raw material stage to the finished
product state. The state of a stream is characterized by the concentra-
tion of the various species that it carries and its temperature, pressure
and flow rates. Applying the laws of conservation of mass, energy and
momentum allows us to track changes in the state of the system. Typi-
cally we subject the raw material streams to either physical treatment to
add or remove chemical species exploiting such property differences as
density, solubility, volatility, diffusivity etc. (transport and equilibrium
processes) or, chemical treatment to alter the chemical structure (reac-
tion processes).
If the state variables are assumed to be independent of time and spa-
tial position, then we often have a lumped parameter, steady state model
resulting in a set of coupled algebraic equations. If they are assumed to
have no spatial variation, but are time dependent, then we have lumped
parameter, dynamic models which result in ordinary differential equa-
tions of the initial value type. If there is no time dependence, but there is
a spatial variation and that too restricted to one dimension (for reasons
of symmetry or scale), then we have ordinary differential equations of
the boundary value type. If both spatial and time dependence are im-
portant, then we end up with partial differential equations, which are
further classified into parabolic, elliptic and hyperbolic equations. The
classification outlined in this paragraph are illustrated with specific ex-
amples in the next sections. The examples are drawn from transport,
equilibrium and reaction processes. The objective is to sensitize you to
the model building process in the hope that you would begin to appre-
ciate the relationship between the physical world and the mathematical
model that represents it.
1.3. LUMPED PARAMETER, STEADY STATE MODELS 5
L, x0 x1 x2 L, x3
1 2 3
V, y1 y2 y3 V, y4
Figure 1.2: Three stage separation process
1.3 Lumped parameter, steady state models
1.3.1 Example of a stagewise separation process

Consider a stagewise separation process shown in Figure 1.2. Suppose
we wish to process a gas stream at a rate of V kmole/s containing a
pollutant at a mole fraction of y4 . We wish to remove the pollutant
by scrubbing it with a solvent in a counter-current 3-stage separation
device. The liquid rate is, say, L kmole/s and it contains a pollutant con-
centration of x0 (which may be zero for a pure solvent stream). Only the
pollutant transfers from the gas phase to the liquid phase and we make
use of the solubility differences between the inert carrier gas and the
pollutant. So far we have made an attempt to describe a physical world.
Is the description adequate to formulate a mathematical model? How
do we know that such a model should result in a steady state, lumped
parameter model? The answer is no, we don’t! We need to further de-
fine and refine the problem statement. For a process engineer this is the
most important step - viz. understand the objective of the task and the
nature of the process (the physical world) to formulate a mathematical
model. Let us continue with the description of the problem.
The state variables in this problem are (L, V , x0 , x1 , x2 , x3 , y1 , y2 , y3 , y4 ).
By focusing only the steady state operation, we remove the dependence
of state variables on time. Such a model cannot clearly answer any ques-
tions concerning start up or shutdown of this process. Next, we assume
that in each stage the gas and liquid are mixed thoroughly so that there
is no spatial variation of concentration within the equipment. This is the
so-called lumped parameter approximation.
We further assume that the streams leaving a stage are in thermody-
namic equilibrium. This implies that for a given inlet streams, no matter
what we do inside the process equipment, the exit concentrations cannot
be changed as they have reached an invariant state. To state it another
way, there is a unique relationship, y = f (x), between the exit concen-
trations of each stage and this relationship could be determined in a

laboratory and entered into a database. Often this relation is expressed
as, y = Kx where K is called the equilibrium ratio; at extremely low con-
centration range K may be assumed to be a constant (results in a linear
model), while at higher concentrations the equilibrium ratio may itself be
a function of concentration, K(x) (results in a non-linear model). While
experience and experimentation suggest that such relationships do ex-
ist, study of equilibrium thermodynamics takes this one step further in
attempting to construct predictive models for the function, y = f (x) by
examining the equilibrium process at a molecular level. Let us continue
with the assumption that the equilibrium model is given by
yn = Kxn n = 1, 2, 3 (1.1)
where we have introduced the subscript n to indicate the equilibrium

relationship is valid for each stage of the separation process. This yields
us three equations, but recall that the state of this 3-stage separation
train is described by 10 variables: (L, V , x0 , x1 , x2 , x3 , y1 , y2 , y3 , y4 ). At
this stage we ask ourselves if there are other laws or principles that this
system should obey. Conservation laws such as mass, momentum and
energy conservation should come to mind. In the present case our ob-
jective has been narrowly focused on tracking the concentration of the
pollutant in each of the three stages. In particular we have not concerned
ourselves with flow and heat effects. Let us speculate briefly what these
effects might be! Heat transfer effects might include heat of absorption,
while flow effects will include imperfect mixing in a stage. The later
in fact has serious consequence in negating two of our earlier assump-
tions: viz. lumped parameter system implying concentration is spatially
uniform in each stage and the exit streams are in thermodynamic equi-
librium. Nevertheless, we still proceed with the assumption of perfect
mixing; a model description that takes into accounts such realities often
becomes intractable. Neglecting heat and flow effects, we have only mass
conservation principle. Applying this for the pollutant species around
each of the three stages, we obtain,
Stage 1: V (y2 − y1 ) = L(x1 − x0 )

Stage 2: V (y3 − y2 ) = L(x2 − x1 ) (1.2)
Stage 3: V (y4 − y3 ) = L(x3 − x2 )
Note that in each of these equations, the left hand side represents the
amount of pollutant that has been removed from the gas phase and the
right hand side represents the same amount of material absorbed into
the liquid phase. Now we have a total of six equations, but still ten
variables. Hence we conclude that we have four degrees of freedom.
This implies that we can choose four of the variables and the remaining
six variables must be determined by satisfying the six equations.
One can also write an overall balance, taking all three stages as one
group:
Overall: V (y4 − y1 ) = L(x3 − x0 ) (1.3)
This, however, is not an independent equation since summing equations

(1.2) produces equation (1.3). This will be used later in introducing con-
cepts of linear independence and rank of matrices.
Specifications: a 3-stage linear system
Let us assume that we pick the four variables associated with the inlet
streams to be specified, viz. (L, V , x0 , y4 ). Defining S = L/KV (a known
value) and eliminating variables (y1 , y2 , y3 ) from equations (1.2) we get
the following system of three equations
    
(1 + S) −1 0 x1 Sx0
    
 −S (1 + S) −1   x2  =  0  (1.4)
0 −S (1 + S) x3 y4 /K
in the unknowns (x1 , x2 , x3 ). This can be represented in compact matrix

form as
Tx = b (1.5)
where T represents the tridiagonal matrix

   
d1 c1 0 (1 + S) −1 0
   
T =  a1 d2 c2  =  −S (1 + S) −1 
0 a2 d3 0 −S (1 + S)
and      
x1 b1 Sx0
     
x =  x2  b =  b2  =  0 
x3 b3 y4 /K
First variation: n-stage separation sequence
Once we have expressed the mathematical model in a symbolic, matrix

form as in equation (1.5), we can generalize the model to any number of
y (a) linear model (b) nonlinear model
y = f(x)
x
K
y=
1
x
x)
1
K(
K=constant
y=
1
x x1 x
Figure 1.3: Linear and nonlinear equilibrium model
stages. Suppose there are n stages, then we merely have,

     
d1 c1 0 ··· 0 x1 b1
 a   x   
 1 d2 c2 ··· 0   2   b2 
 .   ..   .. 
T = ..  x=  b= 
  .   . 
     
 0 0 an−2 dn−1 cn−1   xn−1   bn−1 
0 ··· 0 an−1 dn xn bn
where ai = −S, di = (1 + S), ci = −1. Efficient algorithms for solving
such system will be developed in Chapter 3.
Second variation: nonlinear equilibrium model
Instead of assuming the equilibrium ratios, K in equation (1.1) to be

constant as in figure 1.3a, if they are found to be dependent on concen-
trations x, then we have a nonlinear system of equations. One can then
interpret the K(x) values to be the slopes of the chord as shown in fig-
ure 1.3b, which is no longer a constant, but depends on x. This implies
that, S(x) = L/K(x)V and hence T becomes a function of x. Thus the
elements in T cannot be determined without knowing the solution x.
An intutive approach to resolving this dilemma in solving such systems,
T (x)x = b might be to make an initial guess for x old and use this guess
to evaluate, K(x old ), S(x old ) and hence T (x old ) and obtain a new solu-
tion for x new by solving the linearized system, T (x old )x new = b. One
can repeat this procedure until the difference between the new and the
old values of x becomes vanishingly small. Although there are numer-
ous variations on this scheme, a large class of non-linear problems are
solved within the conceptual frame work (i) estimating an initial guess
(ii) devising an algorithm to improve the estimate and (iii) checking for
convergence of the result.
Third variation: alternate specification
In all of the previous cases we considered the inlet streams to be spec-

ified viz. (L, V , x0 , y4 ). This would be typical for performance analysis
problems where the output of an existing process is desired, given its
inlet conditions. A design engineer, who gets into this game at an ear-
lier stage, might face an alternate problem. For example, environmental
regulations might dictate that the exit concentration of the pollutant y1
be below a certain acceptable level. Thus the four degrees of freedom
might be used up in specifying (V , x0 , y4 , y1 ). Assuming once again a
linear equilibrium model (K constant), the system of equations (1.2) in
the unknown set (L, x2 , x3 ) can be written as:
f1 (L, x2 ; V , y1 , x0 ) := V (Kx2 − y1 ) − L(y1 /K − x0 ) = 0

f2 (L, x2 , x3 ; V , y1 , x0 ) := KV (x3 − x2 ) − L(x2 − y1 /K) = 0 (1.6)
f3 (L, x2 , x3 ; V , y4 ) := V (y4 − Kx3 ) − L(x3 − x2 ) = 0
In spite of assuming a linear equilibrium model, the above set of equa-

tions are non-linear! Why? Although the mathematical model has re-
mained the same for various specifications, we have nice tridiagonal ma-
trix structure for some specifications while no such structure exists for
others.
1.3.2 Process flow sheet simulation

Consider the flow sheet shown in figure 1.4. It is an extremely simple
unit consisting of a reactor and a separator. We are given the mole frac-
tions of components in the recycle stream and the exit stream from the
reactor. We are asked to determine the molar rates of CO and H2 in the
inlet stream, the recyle rate R and the product rate, P . In analysing this
problem we setup a series of material balance equations. Focusing on
the reactor loop (loop 1) shown in dashed line in figure 1.4, we can write
the following three component material balance equations:
C-balance: x + R(0.302 + 0.004) = (0.274 + 0.095) × 275
or
x + 0.306R = 101.475
mole fr of H2 = 0.694
mole fr of CO= 0.302
R (recycle rate) mole fr of CH3OH= 0.004
Reactor Condenser
x: CO P (rate): CH3OH
y: H2
loop 1
loop 2 Find {x, y, R, P}

F = 275 mol/min
mole fr of H2 = 0.631
mole fr of CO= 0.274
mole fr of CH3OH= 0.095
Figure 1.4: Example of material balance equations in a process flow

sheet.
H2 -balance: y + R(0.694 + 0.008) = (0.631 + 2 × 0.095) × 275
or
y + 0.702R = 225.775
O-balance: x + R(0.302 + 0.004) = (0.274 + 0.095) × 275
or
x + 0.306R = 101.475
Note that the O balance equation is redundant and in the language of

linear algebra, these three equations do not form a linearly independent
set of equations. So we proceed to construct additional equations by
examining material balance around the full flow sheet (loop 2). These
give rise to:
C-balance: x=P
H2 -balance: y = 2P
These five equations can be arranged in a matrix form as,

   
1 0 0.306 0   101.475
  x  
 0 1 0.702 0    225.775 
  y   
 1 0 0.306 0  = 101.475  (1.7)
  R   
 1 0 0 −1   0 
P
0 1 0 −2 0
Recognizing the redundancy between the first and third equations and
also combining equations four and five to eliminate P , we can write the
above set in an alternate form as
    
1 0 0.306 x 101.475
    
 0 1 0.702   y  =  225.775  (1.8)
−2 1 0 R 0
1.3.3 Example of a multicomponent flash
Next, we examine a model for a multicomponent, isothermal flash pro-

cess. This also results in a lumped, steady state model description. It
is also an example of how a potentially large system system of algebraic
equations can be reduced to a single equation in one unknown through
clever manipulations. Thus root finding algorithms could be used effi-
ciently to solve this system. A sketch of the process is shown in figure
1.5. A feed stream of known flow rate, F , composition (mole fractions),
{zi |i = 1 · · · N}, temperature, TF and pressure, PF is flashed into a drum
maintained at a temperature and pressure of (T , P ), respectively. Under
right conditions, the feed will split into a vapor phase and a liquid phase.
The objective is to predict the flow rate and compositions of the vapor,
(V , yi ) and the liquid (L, xi ) phases. Each exit stream contains (N + 1)
unknowns. The assumptions are that the process is operating under
steady conditions, perfect mixing takes place inside the drum (lumped
approximation) and the exit streams are in thermodynamic equilibrium.
The model equations are as follows:
Thermodynamic equilibrium - (empirical model)
yi = Ki (T , P )xi i = 1···N (1.9)
Component material balance - (mass conservation)

Observe the
F zi = V yi + Lxi i = 1···N (1.10) nonlinear terms in
this equation: viz.
product of
unknowns V & yi
and L & xi
V
yi i=1 .. N
40
Non-physical
30 solution
20
10
F f(ψ)
0
Zi i=1 .. N
TF 10
PF Real solution
20
30
3 2 1 0 1 2 3
L ψ
xi i=1 .. N
Figure 1.5: Multicomponent, isothermal flash process
Overall material balance
F =V +L (1.11)
Mole fraction constraints

N
X N
X
yi = xi = 1 (1.12)
i i
A simple count indicates that we have written down (2N + 3) equations

for the (2N + 2) unknowns; but it is easy to verify that summing equa-
tions (1.10) over all components and using the mole fraction constraint,
results in equation (1.11). Thus, equation (1.11) is not an independent
one. Although these equations could be solved as a system of nonlinear
algebraic equations, a much more efficient scheme is to eliminate all,
except one variable and reduce the system into a single equation. First
eliminate yi from equation (1.10) using (1.9) to obtain
F zi = (Ki V + L)xi or xi = F zi /(Ki V + L)
Rearrange equation (1.12) as,

N
X N
X
(xi − yi ) = 0 or (1 − Ki )xi = 0
i i
Combine the last two equations as,

N
X (1 − Ki )F zi
=0
i
Ki V + L
Eliminate L from above equation using (1.11) and define ψ = V /F to get

the final form of the flash equation as.
N
X (1 − Ki )zi
=0 (1.13)
i
(Ki − 1)ψ + 1
This is a single equation in one unknown, viz. ψ. In general the number

of roots that a nonlinear equation posses cannot be known a priori. A
possible sketch of the function is shown in figure 1.5b. Since ψ is defined
as the fraction of feed that appears as vapor, (V /F ), the physical world
dictates that it must lie between (0, 1) and it is sufficient if the search
for the root is limited to this range. The flash equation (1.13)may posses
other roots outside the range of interest (0, 1). Such roots are valid
mathematical solutions of the problem, they are not physically relevant.
1.3.4 Example of a phenomenalogical model

In the previous two examples, models were built based on conservation
laws. Models based on empirical observations are also quit common.
The Pressure-Volume-Temperature (PVT) behavior of gases, for exam-
ple, could be modeled by the ideal gas law viz. P V = nRT . A more
refined model, called the Peng-Robinson equation of state is used widely
in chemical engineering literature. It is given by the following equations:
RT a(T )
P= − (1.14)
(V − b) V (V + b) + b(V − b)
where
R 2 Tc2
a(T ) = 0.45724 α(Tr , ω)
Pc
RTc
b = 0.0778
Pc
p
α1/2 = 1 + m(1 − Tr )
m = 0.37464 + 1.54226ω − 0.26992ω2
Here Tc , Pc are the critical temperature and pressure of the component
and ω is the accentric factor. These are properties of a component.
a0 a1 ai-1 ai aN-1 aN
F
1 i N
V V V
Figure 1.6: Reactors in series
We define Tr = T /Tc as the reduced temperature, and Z = P V /RT as

the compressibility factor. Equation (1.14) can be rearranged as a cubic
equation in Z as follows:
Z 3 − (1 − B)Z 2 + (A − 3B 2 − 2B)Z − (AB − B 2 − B 3 ) = 0 (1.15)
where A = aP /R 2 T 2 and B = bP /RT . For the class of problems where

the pressure and temperature (P , T ) are given and the material is iden-
tified ( i.e., Tc , Pc , ω are known), the coefficients (A, B) in equation (1.15)
can be calculated and hence the cubic equation can be solved to find the
roots, Z. This allows the determination of the volume (or density) from
Z = P V /RT .
1.3.5 Example of reactors in series

An example from chemical reaction engineering process that gives rise
to a system of nonlinear equations is that of a continuously stirred tank
reactor in series. A sketch is shown in figure 1.6 Consider an isother-
mal, irreversible second order reaction. The composition in each reactor
is assumed to be spatially homogeneous due to thorough mixing. The
reaction rate expression is given by,
r = kV a2i
where ai is the exit concentration of component A from the i−th reactor,

k is the reaction rate constant and V is the volume of the reactor. A
material balance under steady state conditions on the i − th reactor
results in,
kV a2i = F (ai−1 − ai ) (1.16)
Letting β = kV /F , we have the following n-simultaneous nonlinear equa-
tions.
1.4. LUMPED PARAMETER, DYNAMIC MODELS 15
fi := βa2i + ai − ai−1 = 0 i = 1···N (1.17)

While we have constructed N equations, there are (N + 2) variables in
total. They are [a0 · · · aN β]. Hence we have two degrees of freedom. In
analysing an existing reactor train, for example, one might regard (β, a0 )
to be known and solve for the remaining N variables including the exit
concentration aN (and hence the conversion). In a design situation one
might wish to achieve a specific conversion and hence regards (a0 , aN )
as knowns and solve for remaining N variables including β (and hence
the volume V ).
1.4 Lumped parameter, dynamic models
Lumped parameter, dynamic models arise typically when the spatial vari-
ation of the state variables can be ignored for some reason, but time vari-
ation cannot be ignored. Let us consider an example from heat transfer.
1.4.1 Example of cooling a molten metal

A sample of molten metal at an inital temperature of Ti is placed in a
crucible (at an initial temperature of T∞ ) and allowed to cool by convec-
tion. A sketch is shown in figure 1.7. Let T1 (t) be the temperature of
the molten metal at any time t and T2 (t) be the temperature of the cru-
cible. The argument used to justify neglecting spatial variation is that
the thermal conductivity of the two materials are sufficiently large to
keep the temperature of each material uniform within its boundaries.
The conservation law statement is:
{rate of accumulation} = {rate in} - {rate out}+ {rate of generation}
Applying this first to the molten metal,
d
(m1 Cp1 T1 ) = −h1 A1 (T1 − T2 ) (1.18)
dt | {z }
heat loss from 1 to 2
Energy balance on the crucible results in,
d
(m2 Cp2 T2 ) = h1 A1 (T1 − T2 ) − h2 A2 (T2 − T∞ ) (1.19)
dt | {z } | {z }
heat gain by 2 from 1 heat loss from 2 to ∞
These two equations can be presented using matrix notation as follows:

dθ
= Aθ + b
dt
1.4. LUMPED PARAMETER, DYNAMIC MODELS 16
Insulation
Molten metal
T1
T2
T∞
Figure 1.7: Heat transfer from a molten metal
where
" #  h A h1 A1
  
T1 − m11Cp1 m1 Cp1 0
θ= A= h A
1
h A +h A
 b= h2 A2 T∞ 
T2 + m21Cp1 − 1 m12 Cp2 2 m2 Cp2
2 2
The initial condition is

" #
Ti
θ(t = 0) =
T∞
This problem depends on several parameters which are assumed to be

known. A1 is the heat trasfer area at the metal-crucible interface, A2
is the area at the crucible-air interface. (h1 , h2 ) are the corresponding
heat transfer coefficients, (m1 , m2 ) are the corresponding mass of the
materials, (Cp1 , Cp2 ) are the specific heats of the two materials. Since all
of these are assumed to be known constants, the problem is linear.
1.4.2 Ozone decomposition
A number of mechanisms (some involving 40 steps and 40 equations)

have been proposed to model the decomposition of ozone in the atmo-
sphere. Let us consider a simple two-step model.
O3 + O2 z O + 2O2
O3 + O → 2O2
1.5. DISTRIBUTED PARAMETER, STEADY STATE MODELS 17
In the early stages of research, we were mainly concerned with identi-

fying the mechanisms of ozone depletion in the atmosphere. For lack
of better data, the compositions were assumed to be spatially homoge-
neous in the atmosphere, although we know now that there can be spatial
variations. For the present purpose we will assume the compositions to
be spatially uniform. Let y1 be the composition of O3 and y2 be that of
O. The model equations are,
dy1
= f1 (y1 , y2 ) = −y1 − y1 y2 + κy2 (1.20)
dt
dy2
= f2 (y1 , y2 ) = y1 − y1 y2 − κy2 (1.21)
dt
The initial compositions are y(t = 0) = [1.0, 0.0]. The parameters are
= 1/98 and κ = 3.0. This is a system of two non-linear ordinary dif-
ferential equations. It is an interesting problem in the limit of → 0. In
the reaction analysis literature, the consequence of this limit is known
as the quasi-steady-state-approximation. The physical interpretation is
that the second reaction is much faster than the first one so that it can
be assumed to have reached the equilibirum state at every instant of
time. The second equation becomes an albegraic one. In the applied
mathematics literatture it is called the singular perturbation problem.
From the computational point of view this limit gives rise to a phenom-
ena called stiff systems. We will explore these features further in later
chapters.
1.5 Distributed parameter, steady state models
1.5.1 Heat transfer through a tappered fin

Let us examine an example from transport processes. Consider the use
of a fin to enhance the rate of heat transfer. Basically, a fin provides a
large heat transfer surface in a compact design. In the design and perfor-
mance analysis of fins one might be interested in a variety of questions
such as what is the efficiency of the fin? (as a corollary what is a useful
definition of fin efficiency?), How many fins are required to dissipate a
certain heat load?, What is the optimal shape of the fin that maximizes
the heat dissipation for minimum weight of fin material? How long does
it take for the fin to reach a steady state? etc. You will learn to develop
answers to these questions in a heat transfer course. Our interest at this
stage is to develop a feel for the model building process. A sketch of a
fin is shown in figure 1.8. Let us first examine the steady state behavior
(a) Planar fin (b) Cylindrical fin
W
L
hP δx
T(x) T∞
T0
δx
hP δx
t0
y z
qA|x qA|x+δx
x
Figure 1.8: Heat transfer through a fin
of a planar fin shown in figure 1.8a. The base of the fin is maintained
at a uniform temperature of T0 and the ambient temperature is T∞ . The
state variable that we are interested in predicting is the temperature of
the fin, T . In general in might be a function of all three spatial positions
i.e., T (x, y, z). (Note time, t is eliminated by assuming steady state). If
we know something about the length scales of the fin and the material
property of the fin, we make further assumptions that will reduce the
complexity of the problem. Let us also assume that the fin is made of a What types of
homogeneous material i.e., its thermal conductivity, k is independent of materials might
position. If the length, L, of the fin is much larger than the thickness, t0 , violate this
assumption?
then we might argue that the temperature variation in the y direction
will be smaller than that in the x direction. Thus we can assume T to
be uniform in the y direction. Next, we examine what happens in the
z direction? This argument is somewhat subtle as it is based on sym-
metries in the system. The basic premise here is that symmetric causes
produce symmetric effects. An excellent and easily accessible exposi-
tion on this topic can be found in Golubiksky and Stewart (1993). First
we assume that the ends of the fin in the z direction are at infinity (or
W >> L) so that the end effects can be neglected. Since the temperature
gradient within the fin is caused by the driving force T0 and T∞ which
are independent of z direction, we can expect the fin to respond in a
similar way - viz. T to be independent of z. Note that the end effect,

however small it may be, is always present in a planar fin. By making W
large compared to L we reduce the error caused by the two-dimensional
effect near the ends. On the other hand the azimuthal symmetry in the
circular fin (figure 1.8b) make the problem truly one dimensional with
temperature varying only in the radial direction. Now that we have a bet-
ter feel for what kinds of arguments or assumptions make this problem
one-dimensional, let us proceed with constructing the model equation.
Since the temperature variation is present only in the x direction, we
take an elemental control volume of thickness δx and identify the input
and output sources of energy into this control volume. See figure 1.8.
Energy enters by conduction mechanism at a rate of (qA)|x through the
left boundary at x and leaves at a rate of (qA)|x+δx through the right
boundary at x + δx. Heat is also lost by convection through the upper
and lower boundaries, which is represented by hP δx(T − T∞ ). Here q
is the heat flux (J/s · m2 ) by conduction. This is given by another phe-
dT
nomenological model called the Fourier law: q = −k dx . k in the Fourier
law defines a material property called thermal conductivity (J/m·s ·o C).
A is the cross-sectional area which is allowed to be a function of x (ta-
pered fin), h is the heat transfer coefficient (J/s · m2 ·o C). P = 2W is
the perimeter (m). The conservation law statement is:
{rate of accumulation} = {rate in} - {rate out}+ {rate of generation}
In symbolic terms, it is given by,
0 = (qA)|x − (qA)|x+δx − hP δx(T − T∞ )
Dividing by δx and taking the limit of δx → 0, we obtain,
d(qA)
0=− − hP (T − T∞ )
dx
Using the Fourier law to replace q in terms of T ,

d dT
kA − hP (T − T∞ ) = 0 (1.22)
dx dx
Equation (1.22) is a second order ordinary differential equation. The

physical description dictates that two conditions be specified at the two
ends of the fin, viz.
T (x = 0) = T0
T (x = L) = T∞ (1.23)

dT = 0
dx x=L
1.6. DISTRIBUTED PARAMETER, DYNAMIC MODELS 20
This problem can be solved to obtain T (x) provided the geometrical pa-
rameters {A(x), P }, the material property, k and the heat transfer envi-
ronment {h, T∞ , To } are known. The problem is nonlinear if the thermal
conductivity is a fucntion of temperature, k(T ). In order to determine
the effectiveness of the fin, one is interested in the total rate of heat
transfer, Q through the fin. This is can be computed in one of two ways
as given by,
ZL
dT

Q= hP [T (x) − T∞ ] dx = −k · A(x = 0)
o dx x=0
1.6 Distributed parameter, dynamic models
1.6.1 Heat transfer through a tappered fin

As an example of a distributed, dynamic model let us re-examine the fin
problem, but during the early transient phase. Let the fin be initially at
the uniform ambient temperature of T∞ . At time t = 0 suppose the base
of the fin at x = 0 is brought to a temperature of To . One migth ask
questions like, how long will it take for the fin to reach a steady state?
what will be the temperature at the tip of the fin at a given time? etc.
Now we have the temperature as a function of both position and time,
i.e., T (x, t). We can represent the rate of accumulation within a control
d(ρAδxCp T )
volume symbolically as, dt where the term in paranthesis is the
energy (J) at any time within the control volume, Cp is the specific heat
of the fin material (J/kg ·o C) and ρ is the density of the material. Thus
the transient energy balance becomes,
∂(ρAδxCp T )
= (qA)|x − (qA)|x+δx − hP δx(T − T∞ )
∂t
Dividing by δx and taking the limit of δx → 0, we obtain,
∂(ρACp T ) ∂(qA)
=− − hP (T − T∞ )
∂t ∂x
Finally using the Fourier’s law to replace q in terms of T ,

∂(ρACp T ) ∂ ∂T
= kA − hP (T − T∞ ) (1.24)
∂t ∂x ∂x
Note that we have switched to partial derivatives as T is a function of

both (x, t) and equation (1.24) is a partial differential equation. In ad-
dition to the boundary conditions specified in equation (1.23), we need
1.6. DISTRIBUTED PARAMETER, DYNAMIC MODELS 21
an initial condition at t = 0 to complete the problem specification. The

conditions are
IC : T (x, t = 0) = T∞ ,
BC1 : T (x = 0, t) = T0 , (1.25)
BC2 : T (x = L, t) = T∞
Detection is, or ought to be, an exact science, and
should be treated in the same cold and unemotional
manner. You have attempted to tinge it with ro-
manticism, which produces much the same effect
as if you worked a love-story or an elopement into
the fifth proposition of Euclid.
— SIR ARTHUR CONAN DOYLE
Chapter 2
Single nonlinear algebraic

equation
2.1 Introduction
In this chapter we consider the problem of finding the root of a single

nonlinear algebraic equation of the form,
f (x) = 0 (2.1)
where f (x) is a continuous function of x and the equation is satisfied

only at selected values of x = r , called the roots. The equation can be a
simple polynomial as we saw with the Peng-Robinson equation of state in
section 1.3.4 or a more complicated function as in the multicomponent
flash example discussed in section 1.3.3. If the equation depends on
other parameters, as is often the case, we will represent them as
f (x; p) = 0 (2.2)
where p represents a set of known parameter values. Clearly, we can

graph the function f vs. x for a range of values of x. The objective
of such an exercise is to graphically locate the values of x = r where
the function crosses the x axis. While such a graphical approach has
an intuitive appeal, it is difficult to generalize such methods to higher
22
2.1. INTRODUCTION 23
(a) f(x) := sin(x), f(r)=0 at r=nπ, n=0,1,.. (b) f(x):= x3 + 4 x2 + x -6, f(r)=0 at r=-3,-2,1
1
10
0.5 7.5
5
0 2.5
0
-0.5
-2.5
-5
-1
-7.5 -5 -2.5 0 2.5 5 7.5 -3 -2 -1 0 1 2 3
(c) f(x):= x3 + 3 x2 - 4, f(r)=0 at r=-2,-2,1 (d) f(x):= x3 - x2 + x - 1, f(r)=0 at r=1
6
10
4 5
2 0
0 -5
-2 -10
-4 -15
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
Figure 2.1: Graphs of some simple functions
dimensional systems. Hence we seek to construct computational algo-

rithms that can be generalized and refined successively to handle a large
class of nonlinear problems.
Before we embark on such a task, some of the potential problems are
illustrated with specific examples. In the general nonlinear case, there
is no way to know a priori, how many values of r can be found that will
satisfy the equation, particularly if the entire range of x in (−∞, ∞) is
considered. For example, the simple equation
f (x) := sin(x) = 0
has infinitely many solutions given by r = nπ where n is any integer.

The graph of the function shown in figure 2.1a illustrates this clearly.
Often, the physical description of the problem that gave rise to the math-
ematical function, will also provide information on the range of values
of x that are of interest. For example, in the multicomponent flash equa-
tion, the problem was so formulated that the dependent variable had the
physical interpretation of fraction of feed in vapor; hence this fraction
must be between (0, 1). Although the mathematical equation may have
many other roots, outside of this range, they would lack any physical
meaning and hence would not be of interest.
Algebraic theory tells us that the total number of roots of a polyno-
mial is equal to the degree of the polynomial, but not all of them may be
real roots. Furthermore, if the coefficients of the polynomial are all real,
2.2. BISECTION METHOD 24
then any complex root must occur in pairs. Consider the three cubic
equations given below:
f (x) := x 3 + 4 x 2 + x − 6 = 0 r = −3, −2, 1
f (x) := x 3 + 3 x 2 − 4 = 0, r = −2, −2, 1

f (x) := x 3 − x 2 + x − 1 = 0 r =1
These are graphed in figures 2.1b,c,d respectively. In the first case there
are three distinct roots. The function has a non-zero slope at each value
of the root and such roots are called simple roots. In the second case
we have a degeneracy or a non-simple root at r = −2. This problem
manifests itself in a graphical representation with a zero slope of the
function at the multiple root, r = −2. If the coefficients of the polyno-
mial were slightly different, the curve could have moved slightly upward
giving rise to two distinct roots or downwards yielding no roots in this
region. Algebraically, we can see that the root is a multiple one with a
multiplicity of 2 by factoring the function into (x + 2)(x + 2)(x − 1). In
the third case there is only a single real root.
We begin by constructing some simple algorithms that have an in-
tuitive appeal. They are easy to represent graphically and symbolically
so that one can appreciate the connection between the two represen-
tations. Subsequently we can refine the computational algorithms to
meet the challenges posed by more difficult problems, while keeping the
graphical representation as a visual aid.
There are essentially three key steps in any root finding algorithm.
They are:
step 1: Guess one or more initial values for x.
step 2: Iterate using a scheme to improve the initial guess.
step 3: Check convergence - i.e., has the improvement scheme of

step 2 produced a result of desired accuracy?
The crux of the algorithm is often in the second step and the objective
in devising various clever schemes is to get from the initial guess to the
final result as quickly as possible.
2.2 Bisection method
The bisection algorithm is quite intuitive. A graphical illustration of this

algorithm is shown in figure 2.2a. In step 1, we make two guesses x1 and
(a) Bisection (b) Regula-Falsi

f2 f2
f3
Discard x1
x1 r x1 x3
x3 x2 r x2
Discard x2
f3 (keep the two values
f1 (keep the two values f1 that bracket the root, r)
that bracket the root, r)
( x2 − x1 )
x3 = x1 − f 1
x3=(x1+x2)/2 (f 2 − f 1 )
(c) Secant method f2 (d) Newton method

f3 f2
Discard x1
x1 r x1 slope = f’1
r
x3 x2 x2
(keep the two most f1
f1
recent values)
(x x )
x3 = x1 − f 1 2 − 1
(f 2 − f 1 )
{ δ
x2 = x1 −
f1
f ’1
Figure 2.2: Graphical representation of some simple root finding algo-

rithms
x2 and calculate the functions values f1 = f (x1 ) and f2 = f (x2 ). If the

function values have opposite signs, it implies that it passes through
zero somewhere between x1 and x2 and hence we can proceed to the
second step of producing a better estimate of the root, x3 . If there is
no sign change, it might imply that there is no root between (x1 , x2 ). So What else might it
we have to make a set of alternate guesses. The scheme for producing a imply?
better estimate is also an extremely simple one of using the average of
the two initial guesses, viz.
x1 + x2
x3 = (2.3)
2
In figure 2.2a, (x1 , x3 ) bracket the root r ; hence we discard x2 , or better

still, store the value of x3 in the variable x2 so that we are poised to
repeat step 2. If the situation were such as the one shown in figure 2.2b,
then we would discard x1 or better still, store the value of x3 in the

variable x1 and repeat step 2. In either case, (x1 , x2 ) will have better
guesses than the original values.
The final step is to check if we are close enough to the desired root
r so that we can terminate the repeated application of step 2. One test
might be to check if the absolute difference between two successive val-
ues of x is smaller than a specified tolerance, i.e.,
|xi+1 − xi | ≤
Another test might be to check if the absolute value of the function fi

at the end of every iteration is below a certain tolerance, i.e.,
|f (xi )| ≤
In addition, we might wish to place a limit on the number of times step

2 is repeated. A MATLAB function, constructed in the form of a m-file,
is shown in figure 2.3.
Example - Multicomponent flash equation
Let us consider the multicomponent, isothermal flash equation devel-

oped in section §1.3.3. The equation is
N
X (1 − Ki )zi
=0 (2.4)
i
(Ki − 1)ψ + 1
where (Ki , zi , N) are treated as known parameters and ψ is the unknown

variable. Implementation of this function in MATLAB is shown below as
an m-file. See Appendix B for
an introduction to
MATLAB.
function r=bisect(Fun,x,tol,trace)
%BISECT find the root of "Fun" using bisection scheme
% Fun - the name of the external function
% x - vector of length 2, (initial guesses)
% tol - error criterion
% trace - print intermediate results
%
% Usage bisect(’flash’,[0,1])
% flash is the name of the external function.
% [0,1] is the initial guess
%Check inputs
if nargin < 4, trace=0; end
if nargin < 3, tol=eps; trace=0; end
if (length(x) ˜= 2)
error(’Please provide two initial guesses’)
end
f = feval(Fun,x); %Fun is assumed to accept a vector
if (prod(sign(f))) > 0, %Check if roots are bracketed

error(’No sign change - no roots’)
end;
for i = 1:100 %Set max limit on iterations

x3 = (x(1) + x(2))/2; %Update the guess
f3 = feval(Fun,x3); %Cal. f(x3)
%Check if x2 or x1 should be discarded

if sign(f(1)*f3) < 0, x(2)=x3; else x(1)=x3; end;
if abs(f3) < tol, r=x3; return; end %Check convergence

if trace, fprintf(1,’%3i %12.5f %12.5f\n’, i,x3,f3); end
end
error(’Exceeded maximum number of iterations’)
Figure 2.3: MATLAB implementation of the bisection algorithm

function f=flash(psi)
% K is a vector of any length of equil ratios.
% z is the feed composition (same length as K)
% K, z are defined as global in main
% psi is the vapor fraction.
global K z
if ( length(K) ˜= length(z) )
error(’Number of K values & compositions do not match’)
end
n=length(psi);
for i = 1:n
f(i)=sum( ((K-1).*z) ./ (1+(K-1)*psi(i)) );
end
Observe that this function, while being concise, is fairly general to han-
dle any number of components N and a vector of guesses psi of any
length and return a vector of function values, one corresponding to each
element in the guessed variable psi. Assuming that you have such a
function defined in a file named flash.m, you are encouraged to work
through the following exercise using MATLAB.
» global K z
» z=[.25 .25 .25 .25] %define a 4-component system
» K=[2 1.5 0.5 0.1] %define equilibrium values
» bisect(’flash’,[0,.1]) %find the root using bisect ans=0.0434
» x=0:0.05:1; %create a vector of equally spaced data
» y=flash(x); %evaluate the function
» plot(x,y) %plot the function
In this section, we have developed and implemented the bisection al-

gorithm as a function in MATLAB and used it to solve an example prob-
lem from multicomponent flash. The function bisect can be used to
solve any other root finding problem as long as you define the problem
you want to solve as another MATLAB function along the lines of the
example function flash.m.
While we managed to solve the problem, we did not concern our-
selves with questions such as, (i) how many iterations did it take to con- Is it always
verge? and, (ii) can we improve the iteration scheme in step 2 to reduce possible to find
the number of iterations? By design, the bisection scheme will always such a guess?
Consider the
pathological case
of r = −2 in
figure 2.1c!
2.3. REGULA-FALSI METHOD 29
converge provided an acceptable set of initial guesses have been chosen.

This method, however, converges rather slowly and we attempt to devise
algorithms that improve the rate of convergence.
2.3 Regula-falsi method
Instead of using the average of the two initial guesses as we did with
the bisection scheme, we can attempt to approximate the function f (x)
by straight line (a linear approximation) since we know two points on
the function f (x). This is illustrated graphically in figure 2.2b with the
dashed line approximating the function. We can then determine the root,
x3 of this linear function, f˜(x̃). The equation for the dashed straight
line in figure 2.2b is
x̃ − x1 x2 − x 1
f˜(x̃) := =
f˜ − f1 f2 − f 1
Now we can determine the value of x̃ = x3 where the linear function

f˜ = 0 in the above equation. This results in,
x2 − x 1
x3 = x1 − f1 (2.5)
f2 − f 1
which is used as the iterative equation in step 2. When we evaluate the

original function at x3 , f (x3 ) clearly will not be zero (unless the scheme
has converged) as shown in figure 2.2b; but x3 will be closer to r than
either x1 or x2 . We can then retain x3 and one of x1 or x2 in such a
manner that the root r remains bracketed. This is achieved by following
the same logic as in the bisection algorithm to discard the x value that
does not bracket the root.
The MATLAB function bisect can be easily adapted to implement
Regula-Falsi method by merely replacing equation (2.3) with equation
(2.5) for step 2.
2.4 Secant method
The secant method retains the linear approximation procedure of the

regula-falsi method, but differs by retaining the two most recent values
of x viz. x2 and x3 and always discarding the oldest value x1 . This
simple change from the Regula-Falsi scheme produces a dramatic dif-
ference in the rate of convergence. A MATLAB implementation of the
2.4. SECANT METHOD 30
function r=secant(Fun,x,tol,trace)
%SECANT find the root of a function "Fun" using secant scheme
% Fun - the name of the external function
% x - vector of length 2, (initial guesses)
%
% Usage secant(’flash’,[0,1])
% Here flash is the name of the external function.
% [0,1] is the initial guess
%Check inputs
if (length(x) ˜= 2)
error(’Please provide two initial guesses’)
end
f = feval(Fun,x); %Fun is assumed to accept a vector
for i = 1:100 %Set max limit on iterations

x3 = x(1) - f(1)*(x(2)-x(1))/(f(2)-f(1)) ; %Update(step 2)
f3 = feval(Fun,x3); %Cal. f(x3)
%Keep the last two values

x(1) = x(2);f(1) = f(2); x(2) = x3; f(2) = f3;
if abs(f3) < tol, r=x3; return; end %Check for convergence

if trace, fprintf(1,’%3i %12.5f %12.5f\n’, i,x3,f3); end
end
error(’Exceeded maximum number of iterations’)
Figure 2.4: MATLAB implementation of the secant algorithm

2.5. NEWTON’S METHOD 31
secant method is shown in figure 2.4. The convergence rate of the se-
cant method was analyzed by Jeeves (1958).
2.5 Newton’s method
The Newton method is by far the most powerful and widely used algo-
rithm for finding the roots of nonlinear equations. A graphical repre-
sentation of the algorithm is shown in figure 2.2d. This algorithm also
relies on constructing a linear approximation of the function; But this is
achieved by taking the tangent to the function at a given point. Hence
this scheme requires only one initial guess, x1 . The linear function f˜(x̃)
shown by dashed line in figure 2.2d is,
f˜ − f1
f˜(x̃) :=
x̃ − x1
The root of this linear equation, is at (x̃ = x2 , f˜ = 0). Under these

conditions one can solve the above equation for x2 as,
f1
x2 = x1 − (2.6)
f10
which forms the iterative process (step 2). Note that this algorithm re-
quires that the derivative of the function be evaluated at every iteration,
which can be a computationally expensive operation.
While we have relied on the geometrical interpretation so far in con-
structing the algorithms, we can also derive Newton’s scheme from a
Taylor series expansion of a function. This is an instructive exercise,
for it will enable us to generalize the Newton’s scheme to higher dimen-
sional ( i.e., more than two equations) systems as well as provide some
information on the rate of convergence of the iterative scheme.
The Taylor series representation of a function around a reference
point, xi is,
X∞
f (k) (xi ) k
f (xi + δ) = δ (2.7)
k=0
k!
where f (k) (xi ) is the k−th derivative of the function at xi and δ is a small
displacement from xi . While the infinite series expansion is an exact
representation of the function, it requires all the higher order derivative
of the function at the reference point. We can construct various levels of
2.6. MULLER’S METHOD 32
approximate representations of the function, f˜ by truncating the series

at finite terms. For example a three term expansion (k = 0, 1, 2) is
2δ
f˜(xi + δ) = f (xi ) + f 0 (xi )δ + f 00 (xi ) + O(δ3 )
2!
where the symbol O(δ3 ) stands as a reminder of the higher order terms
(three and above in this case) that have been neglected. The error in-
troduced by such omission of higher order terms is called truncation
error. In fact to derive the Newton scheme, we neglect the quadratic
term O(δ2 ) also. In figure 2.2d, taking the reference point to be xi = x1 ,
the displacement to be δ = xi+1 − xi , and recognising that f˜(xi + δ) = 0
we can rewrite the truncted two-term series as,
0 = f (xi ) + f 0 (xi )(xi+1 − xi ) + O(δ2 )
which can be rearranged as
f (xi )
xi+1 = xi −
f 0 (xi )
This is the same as equation (2.6).
2.6 Muller’s method
This scheme can be thought of as an attempt to generalize secant method

and is important at least from the point of illustrating such generaliza-
tions. Instead of making two guesses and constructing an approximate
linear function as we did with the secant method, we can choose three
inital guesses and construct a quadratic approximation to the original
function and find the roots of the quadratic. A graphical representation
of this is shown in figure 2.5. The three initial guesses are (x0 , x1 , x2 )
and the corresponding function values are represented by (f1 , f2 , f3 )
respectively. We construct a second degree polynomial as,
p2 (v) = av 2 + bv + c
where v = (x−x0 ). Note that the polynomial is represented as a function

of a new independent variable v, which is merely a translation of the
original independent variable x by x0 . An alternate view is to regard
v as the distances measured from reference point x0 , so that v = 0
at this new origin. (a, b, c) are the coefficients of the quadratic that
must be determined in such a way that p2 (v) passes through the three
2.6. MULLER’S METHOD 33
(a) Muller’s scheme

f1
f0 f(x)
p2(v)
x2
r x0 x1
h1 = x1 - x0
f2 h2 = x0 - x2
Figure 2.5: Graphical representation of Muller’s scheme
data points (x0 , f0 ), (x1 , f1 ) and (x2 , f2 ). Defining h1 = (x1 − x0 ) and

h2 = (x0 − x2 ) and requiring that the polynomial pass through the three
points, we get,
p2 (0) = a(0)2 + b(0) + c = f0 ⇒ c = f0

p2 (h1 ) = ah21 + bh1 + c = f1
p2 (−h2 ) = ah22 − bh2 + c = f2
The reason for coordinate shift should be clear by now. This enables c
to be found directly. Solving the remaining two equations we obtain a
and b as follows:
γf1 − f0 (1 + γ) + f2
a=
γh21 (1 + γ)
f1 − f0 − ah21
b=
h1
where γ = h2 /h1 . So far we have only constructed an approximate rep-
resentation of the original function, f (x) ≈ p2 (v). The next step is to
find the roots of this approximate function, p2 (v) = 0. These are given
by, √
−b ± b2 − 4ac
v = r̃ − x0 =
a2
2.7. FIXED POINT ITERATION 34
This can be rearranged as,
2c
r̃ = x0 − √ (2.8)
b± b2 − 4ac
Since p2 (v) is a quadratic, there are clearly two roots r̃ in equation (2.8).
In order the take the root closest to x0 we choose the largest denominator
in equation (2.8). In summary, the sequential procedure for impliment-
ing Muller’s scheme is as follows:
• Guess (x0 , x1 , x2 )
• Compute (f0 = f (x0 ), f1 = f (x1 ), f2 = f (x2 )).
• Compute h1 = (x1 − x0 ), h2 = (x0 − x2 ), γ = h2 /h1
• Compute c = f (x0 ).
• Compute a = (γf1 − f0 (1 + γ) + f2 )/(γh21 (1 + γ))
• Compute b = (f1 − f0 − ah21 )/(h1 ).
• Compute the roots from equation (2.8).
• From (x0 , x1 , x2 ) discard the point farthest from r̃ and substitue

the new root in its place and repeat.
Note that Muller’s method converges almost quadratically (as does New-
ton’s scheme), but requires only one additional function evaluation at
every iteration which is comparable to the computational load of the se-
cant method. In particular derivative evaluation is not required, which is
a major advantage as compared to Newton’s method. Also, this scheme
can converge to complex roots even while starting with real initial guesses
as long as provision is made for handling complex arithmetic in the com-
puter program. MATLAB handles complex arithmetic quite naturally.
2.7 Fixed point iteration
Another approach to construct an update scheme (for step 2) is to rear-

range the given equation f (x) = 0 into a form,
x = g(x)
Then, starting with a guess xi , we can evaluate g(xi ) from the right hand
side of the above equation and the result itself is regarded as a better
estimate of the root, i.e.,
xi+1 = g(xi ) i = 0, 1 · · · (2.9)
Given, f (x) = 0 it is not difficult rewrite it in the form x = g(x); nor is

this process unique. For example, we can always let g(x) = x + f (x).
Such an iterative scheme need not always converge. Let us examine the
possible behavior of the iterates with a specific example. In particular we
will illustrate that different choices of g(x) lead to different behavior.
Consider the function
f (x) = x 2 − x − 6 = 0 (2.10)
which has roots at r = −2 and r = 3. In the first case let us rewrite it as

p
x = x+6
A geometrical interpretation is that we are finding

√ the intersection of two
curves, y = x (the left hand side) and y = x + 6 (the right hand side).
See figure 2.6 for a graphical illustration.
p Starting with an initial guess,
say x0 = 4, we compute x1 = g(x0 ) = x0 + 6. This is tantamount to
stepping between the y = x and y = g(x) curves as shown in figure
2.6a. It is clear that the sequence will converge monotonically to the
root r = 3. The table 2.1 shows the first ten iterates, starting with an
inital guess of x0 = 5.
Observe that the slope of the function at the root is g 0 (r = 3) <
1. We will show shortly that the condition for convergence is indeed
|g 0 (r )| < 1. As an alternate formulation consider rewritting equation
(2.10) as x = g(x) = 6/(x − 1). Now, g(x) has a singularity at x = 1. A
graphical illustration is shown in figure 2.6b. Using this new g(x), but
starting at the same initial guess x0 = 4 the sequence diverges initially in
an oscillatory fashion around the root r = 3, but eventually is attracted
to the other root at r = −2, also in an oscillatory fashion. Observe
that the slopes at the two roots are: g 0 (3) = −3/2 and g 0 (−2) = −2/3.
Both are negative and hence the oscillatory behavior. The one with abso-
lute magnitude greater than unity diverges and the other with absolute
magnitude less than unity converges. Finally consider the formulation
x = g(x) = (x 2 − 6). The behavior for this case is shown in figure 2.6c.
For reasons indicated above, the sequence will not converge to either
root! The following shows the MATLAB implementation for generating
the iterative sequence for the first case. Enter this into a file called g.m.
(a) g ( x ) =
x + 6 6
(b) g ( x ) = ( x − 1 )
x
y=
x
y=
|g’(r=3) = -3/2 | > 1 r
x0
y=g(x) r
)
g’(r=3) = 1/6 < 1 =g(x
r x1 y
|g’(r=-2) = -2/3| < 1
x0
(c) g ( x ) = ( x2 − 6 )
y=g(x)
x
y=
g’(r=3) = 6 > 1
r
x0
|g’(r=-2) = -4| > 1
Figure 2.6: Graphical representation of fixed point scheme

2.8. ERROR ANALYSIS AND CONVERGENCE ACCELERATION 37
Iteration Iterate
Number
x1 5.0000000
x2 3.3166248
x3 3.0523147
x4 3.0087065
x5 3.0014507
x6 3.0002418
x7 3.0000403
x8 3.0000067
x9 3.0000011
x10 3.0000002
p
Table 2.1: The first ten iterates of xi+1 = xi + 6 starting with x0 = 5
function x=g(x)
for i=1:10
fprintf(1,’%2i %12.5e\n’,i,x); %print the iterates
x=sqrt(x+6); %also try x=6/(x-1) and x=(xˆ2-6) here
end
Invoke this function from within MATLAB with various initial guesses,
e.g., try initial guess of 5 by entering,
» g(5)
2.8 Error analysis and convergence acceleration
A simple error analysis can be developed for the fixed point iterative
scheme which will provide not only a criterion for convergence, but also
clues for accelerating convergence with very little additional computa-
tional effort. We are clearly moving away from the realm of intuition to
the realm of analysis! Consider the fixed point iteration xi+1 = g(xi ).
After convergence to the root r we will have r = g(r ). Subtracting the
two equations we get,
(xi+1 − r ) = g(xi ) − g(r )

xi
g(xi) - g(r)
g’(ξ)
xi - r
ξ
r
r < ξ <xi
Figure 2.7: Graphical illustration of mean value theorem
Multiplying and diving the right hand side by (xi − r ) ,
g(xi ) − g(r )
(xi+1 − r ) = (xi − r )
(xi − r )
Now the difference ei = (xi −r ) can be interpreted as the error at iterative

step i and hence the above equation can be written as,
ei+1 = g 0 (ξ) ei (2.11)
where we have used the mean value theorem to replace slope of the
chord by the tangent to the curve at some suitable value of x = ξ, i.e.,
(g(xi ) − g(r ))
= g 0 (ξ)
(xi − r )
A geometrical interpretation of the mean value theorem is shown in fig-

ure 2.7. From equation (2.11), it is clear the error will decrease with every
iteration if the slope |g 0 (ξ)| < 1; otherwise the error will be amplified
at every iteration. Since the error in the current step is proportional to
that of the previous step, we conclude that the rate of convergence of
the fixed point iteration is linear. The development has been reasonably
rigorous so far. We now take a more pragmatic step and assume that
g 0 (ξ) = K is a constant in the neighbourhodd of the root r . Then we
have the sequence, Will K be the same
constant at every
e2 = Ke1 , e3 = Ke2 = K 2 e1 , e4 = Ke3 = K 3 e1 ··· iteration?
and hence we can write a general error propagation solution as,
en = K n−1 e1 or xn − r = K n−1 e1 (2.12)
It should be clear now that en → 0 as n → ∞ only if |K| < 1. . We refer Well, if we know
to equation (2.12) as the error propagation solution since it provides a the error in the
solution of estimating the error at any step n, provided the error at the first step, none of
this analysis would
first step e1 and K are known.
be necessary!
We can develop a convergence acceleration scheme using the error so-
r = x1 − e1 would
lution (2.12) to estimate the three unknowns (r , K, e1 ) in the second form do it!
of equation (2.12). Once we have generated three iterates, (xn , xn+1 , xn+2 ),
we can use equation (2.12) to write down,
xn = r + K n−1 e1
xn+1 = r + K n e1 (2.13)
n+1
xn+2 = r +K e1
Now we have three equations in three unkowns which can be solved to

estimate r (and K, e1 as well). If K were to remain a true constant with
every iteration, r would be the correct root; since K is not a constant
in general, r is only an estimate of the root, hopefully a better estimate
than any of (xn , xn+1 , xn+2 ). Now let us proceed to construct a solution
for r from the above three equations. We will define a first order forward
difference operator ∆ as,
∆xn = xn+1 − xn
Think of the symbol ∆ as defining a new rule of operation just like a

d
derivative operator dx defines a rule. When ∆ operates on xn it is com-
puted using the rule shown on the right hand side. Now, if we apply the
operator ∆ to xn+1 we should have,
∆xn+1 = xn+2 − xn+1 .
If we apply the ∆ operator twice (which is equivalent to defining higher

order derivatives), we should get,
∆(∆xn ) = ∆2 xn = (∆xn+1 ) − (∆xn ) = xn+2 − 2xn+1 + xn .
You can verify that using equation (2.13) in the above definitions, we get,
(∆xn )2
= K n−1 e1 = xn − r
∆2 x n
and hence r is given by,
2 2
(∆xn )2 xn+1 − 2xn xn+1 + xn
r = xn − = xn −
∆ 2 xn xn+2 − 2xn+1 + xn
Thus the three iterates (xn , xn+1 , xn+2 ) can be plugged into the right
hand side of the above equation to get a better estimate of r .
Example of convergence acceleration
Let us apply this convergence acceleration procedure to the first three

iterates of table 2.1.
x1 = 5.0000000
x2 = 3.3166248
x3 = 3.0523147
∆x1 = (x2 − x1 ) = −1.6833752
∆x2 = (x3 − x2 ) = −0.26431010
∆2 x1 = (∆x2 − ∆x1 ) = 1.41906510
(∆x1 )2 (−1.6833752)2
r = x1 − = 5.000000 − = 3.0030852
∆ 2 x1 1.41906510
Compare this with the fourth iterate produce in the original sequence
x4 = 3.0087065.
2.8.1 Convergence of Newton scheme
The Newton scheme given by equation (2.6) can be thought of as a fixed

point iteration scheme where g(x) has been specified in a special manner
as,
f (xn )
xn+1 = xn − 0 = g(xn )
f (xn )
Hence,
(f 0 )2 − f f 00 f f 00

g 0 (x) = 1 − = 0 2 < 1
(f 0 )2 (f )
Since f (r ) = 0 by definition, g 0 (r ) = 0 (barring any pathological situa-

tion such as f 0 (r ) = 0) and the inequality should hold near the root r .
Thus the Newton method is guaranteed to converge as long as we have
a good initial guess. Having progressed this far, we can take the next
2.9. DEFLATION TECHNIQUE 41
step and ask the question about the rate of convergence of the Newton
method. A Taylor series expansion of g(x) around r is,
g 00 (r )
g(xn ) = g(r ) + g 0 (r )(xn − r ) + (xn − r )2 + · · ·
2
Recognizing that en+1 = xn+1 − r = g(xn ) − r , en = (xn − r ) and
g 0 (r ) = 0, the truncated Taylor series expansion can be rearranged as,
g 00 (r ) 2
en+1 = en
2
which shows that the error at any step goes down as the square of the
previous step - i.e., quadratically! This manifests itself in the form of
doubling accuracy at every iteration.
2.9 Deflation technique
Having found a root, r , of f (x) = 0, if we are interested in finding addi-

tional roots of f (x) = 0, we can start with a different initial guess and
hope that the new initial guess lies within the region of attraction of a
root different from r . Choosing a different initial guess does not guar-
antee that the iteration scheme will not be attracted to the root already
discovered. In order to ensure that we stay away from the known root, r ,
we can choose to deflate the original function by constructing a modified
function,
g(x) = f (x)/(x − r )
which does not have r as a root. For a single equation the concepts are
best illustrated with a graphical example. Consider the illustration in
figure 2.8a where the original function,
f (x) := (x − 2) sin(2x)e−0.8x
can be seen to have several roots including one at r = 2. A sketch of the

deflated function, g(x) = f (x)/(x − 2) is shown in figure 2.8b. Since
r = 2 turns out to be a simple root of f (x) = 0, the deflated function
g(x) = 0 does not contain the already discovered root at x = 2. Hence
starting with a different initial guess and applying an iterative method
like the secant or Newton scheme on the function g(x) will result in
convergence to another root. This process can obviously be repeated by
deflating successively found roots. For example if we know two roots r1
and r2 then a new function can be constructed as
f (x)
h(x) = .
(x − r1 )(x − r2 )
2.9. DEFLATION TECHNIQUE 42
0.1 0.2
(a) (c) f ( x ) = ( x − 2 ) 2 sin ( 2x )e− 0.8x
0.05
0.1
0
0
-0.05 deflated root
deflated root
-0.1 f ( x ) = ( x − 2 ) sin ( 2x )e− 0.8x -0.1
0 2 4 6 8 10 -0.2
0 2 4 6 8 10
0.1 0.2
deflated root
deflated root
(b) (d)
0.05
0.1
0
0
-0.05
-0.1
-0.1
g( x ) = f ( x ) ( x − 2 ) g( x ) = f ( x ) ( x − 2 )
0 2 4 6 8 10 -0.2
0 2 4 6 8 10
Figure 2.8: Graphical illustration deflation technique
The successive application of this approach is of course susceptible to

propagation of round off errors. For example if the roots r1 , r2 are known Can you think of a
to only a few significant digits, then the definition of the deflated func- way to alleviate
tion h(x) will inherit these errors and hence the roots of h(x) = 0 will this problem?
not be as accurate as those of the original equation f (x) = 0.
Another problem of the deflation technique pertains to non-simple

roots. A sketch of the function,
f (x) := (x − 2)2 sin(2x)e−0.8x
is shown in figure 2.8c. It is immediately clear that r = 2 is a double

root - i.e., occurs with a multiplicity of two. Hence the deflated function What
g(x) = f (x)/(x − 2) still has r = 2 as a simple root as seen in figure computational
2.8d. problem might this
pose?
2.10. PARAMETER CONTINUATION 43
2.10 Parameter continuation
2.10.1 Euler-Newton continuation

2.10.2 Homotopy continuation
2.11 Software tools
2.11.1 MATLAB
The MATLAB function for determining roots of a polynomial is called
roots. You can invoke it by entering,
» roots(c)
where c is a vector containing the coefficients of the polynomial in the

form,
pn (x) = c1 x n + c2 x n−1 + · · · + cn x + cn+1 .
Let us consider the factored form of the polynomial p3 (x) = (x + 2)(x +
i)(x − i) so that we know the roots are at (−2, ±i). To check whether
MATLAB can find the roots of this polynomial we need to construct the
coefficients of the expanded polynomial. This can be done with the con-
volve fucntion conv(f1,f2) as follows.
» f1 = [1 2] %Here we define coeff of (x+2) as [1 2]

» f2 = [1 i] %Here we define coeff of (x+i) as [1 i]
» f3 = [1 -i] %Here we define coeff of (x-i) as [1 -i]
» c=conv(conv(f1,f2),f3) % c contains coeff of polynomial
» r=roots(c) %returns roots of polynomial defined by c
Note that the function roots finds all of the roots of a polynomial,
including complex ones.
The MATLAB function for finding a real root of any real, single non-
linear algebraic equation (not necessarily a polynomial) is called fzero.
You can invoke it by entering,
» fzero(’fn’,x)
where fn is the name of a m-file that defines the function, x is the initial
guess for the root. This fzero is not based on a very robust algorithm.
If the function you want to solve has singularities, or multiple roots, the
scheme fails to converge, often without producing any appropriate er-
ror or warning messages. Hence use with caution. After it produces an
2.11. SOFTWARE TOOLS 44
answer check that it is the correct result by evaluating the function at

the root. As an example try the multicomponent flash problem consid-
ered previously. You are encouraged to try the following steps during a
MATLAB session.
Note that the
desired result is
»global K z; % define K,z to be global root=0.0949203.
»K=[2 1.5 0.5 0.2]; % define K values But starting with
»z=[.25 .25 .25 .25]; % define z values different initial
»root=fzero(’flash’,.5) % solve x0=0.5, ⇒ ans=0.0949 guesses, MATLAB
»flash(root) % check solution produces different
results! Why? Try
»root=fzero(’flash’,-.85)% solve x0=-0.85, ⇒ ans=-1.0000
plotting the
»flash(root) % check solution function over the
»root=fzero(’flash’,1.85)% solve x0=1.85, ⇒ ans=1.6698 range ψ ∈ [−5, 5]
»flash(root) % check solution in MATLAB and see
»root=fzero(’flash’,1.90)% solve x0=1.90, ⇒ ans=2.0 if you can
»flash(root) % check solution understand
MATLAB behavior!
(Clue: sign change)
2.11.2 Mathematika
Mathematika is another powerful software package for mathematical
analysis including symbolic processing. It also has an inteactive envi-
ronment; i.e., commands, functions are executed as soon as you enter
them. If you invoke Mathematica using a Graphical User Interface (GUI)
then, the ploting functions will display the graphs. Otherwise you are
limited to the use of computational features of Mathematica. A complete
reference to Mathematica can be found in the book by Wolfram (1988).
Example of solving the flash equation with Mathematica
The Mathematica function for finding the roots of an equation is called

Solve. In MATLAB we constructed a m-file to define a function. In
Mathematica this is done in one line within the workspace. Anything
delimited by (* *) is treated as comments and ignored by Mathematica.
You may wish to work through the following exercise.
Observe that first
letter of all
(* Define the multicomponent flash function f[psi] Mathematica
for a 4-component system. functions is in
K[[i]], z[[i]] are called lists in Mathematica. uppercase.
Treat them as arrays.
2.12. EXERCISE PROBLEMS 45
10 0.05
0
5
-0.05
f(ψ) 0 -0.1
-0.15
-5
-0.2
-10
-4 -2 0 2 4 0 0.1 0.2 0.3 0.4 0.5
ψ ψ
Figure 2.9: Multicomponent flash function
Sum[f(i), {i,N}] is a function that sums f(i)

over the index range {1,N} *)
f[psi_]:=Sum[(K[[i]]-1)*z[[i]]/((K[[i]]-1)*psi+1), {i,4}]
K={2., 1.5, .5, .2} (* Define K values *)

z={.25, .25, .25, .25} (* Define z values *)
Solve[f[x] == 0, x ] (* Solve f(x)=0 *)
Plot[f[x],{x,-5,5}, Frame -> True] (* Plot f[x] over 5 to 5 *)
Plot[f[x],{x,0,.5}, Frame -> True] (* Plot f[x] over 0 to 0.5 *)
Note that no initial guess is needed. Mathematica finds all the three
roots. They are:
{{x -> -1.57727}, {x -> 0.0949203}, {x -> 1.66984}}
Ofcourse, only the second root is of interest. Others do not have any
physical relevance. Note also that while ploting functions, Mathematica
samples the function at sufficient number of data points ( i.e., x-values)
to provide a smooth function. Graphs of the flash equation produced by
Mathematica are shown in figure 2.9.
2.12 Exercise problems
2.12.1 Multicomponent, isothermal flash model

A sketch of a multicomponent flash process is shown in figure 1.5. The
following equation, which was derived in Chapter 1, models the multi-
component flash process. This is a single non-linear algebraic equation
Component i zi Ki
Carbon dioxide 1 0.0046 1.650
Methane 2 0.8345 3.090
Ethane 3 0.0381 0.720
Propane 4 0.0163 0.390
Isobutane 5 0.0050 0.210
n-Butane 6 0.0074 0.175
Pentanes 7 0.0287 0.093
Hexanes 8 0.0220 0.065
Heptanes+ 9 0.0434 0.036
Table 2.2: Feed composition & equilibrium ratio of a natural gas mixture
in the unknown ψ, which represents the fraction of feed that goes into
the vapor phase.
N
X (1 − Ki )zi
=0 (2.14)
i=1
(Ki − 1)ψ + 1
This equation has several roots, but not all of them have any physical For a bonus point
meaning. Only the root for ψ ∈ [0, 1] is of interest. you may choose to
The test data in Table 2.2, (taken from Katz etal) relate to the flashing find the number of
roots for the test
of a natural gas stream at 1600 psia and 120o F . Determine the fraction
data above!
ψ using the secant algorithm given in figure 2.4 and another root finding
function that is provided in MATLAB named fzero.
2.12.2 Compressible flow in a pipe

In a fluid mechanics course you might come across the Weymouth equa-
tion, which is used for relating the pressure drop vs. flow rate in a
pipeline carrying compressible gases. It is given by,
" #0.5
To (P12 − P22 )
Qo = 433.54 d2.667 η (2.15)
Po Lσ T
where
Qo is the gas flow rate = 2000000 SCFD

To is the standard temperature = 520o R
Po is the standard pressure = 14.7 psia

P1 is the upstream pressure, (?), psia
P2 is the downstream pressure, (21.7), psia
L is the length of pipe = 0.1894 miles
σ is the specific gravity of gas (air=1) = 0.7
T is the actual gas temperature = 530o R
d is the diameter of the pipe, (?) inches
η is the efficiency = 0.7 (a fudge factor!)
1. If the diameter of the pipe is 4.026 inches, determine the upstream

pressure using the secant (initial guess of [5, 45]) and fzero (initial
guess of 25) functions. Compare the flops which stands for the
floating point operations.
2. Suppose the maximum pressure the pipeline can withstand is only

24.7 psia. Other conditions remaining the same as in previous part,
determine the diameter of the pipe that should be used using the
secant (initial guess of [4, 8]) and fzero (initial guess of 6) functions.
Compare the flops.
2.12.3 A model for separation processes

Consider a stagewise separation process shown in Figure 1.2. A model
for this process was developed in Chapter 1. The variables of interest
are (L, V , x0 , x1 , x2 , · · · , xn , y1 , y2 , y3 , · · · , yn+1 ). Under the assump-
tion of linear equilibrium model, yi = Kxi it is possible to successively
eliminate all of the variables and obtain the following single, analytical
expression relating the input, (x0 , yn+1 ), the output, xn , the separation
factor S = L/KV and the number of stages n.
h i
[x0 − xn ] (1/S)n+1 − (1/S)
= (2.16)
x0 − yn+1 /K (1/S)n+1 − 1
The equation is called the Kremser-Brown-Souders equation. We have a

single equation relating six variables, viz. (x0 , xn , yn+1 , K, S, n). Given
any five of these variables, we can solve for the 6th one. Your task is to
formulate this problem as a root finding problem of the type
f (x; p) = 0
where the unknown variable is associated with x and p is the parameter

set consisting of the remaining five known values. Write an m-file to
represent the function in the form
h i
(x0 − xn ) (1/S)n+1 − (1/S)
f (x; p) = − =0 (2.17)
(x0 − yn+1 /K) (1/S)n+1 − 1
1. In a typical design problem you might be given the flow rates, (say
L = 10, V = 10), the inlet compositions (say, x0 = 0.8, yn+1 = 0)
and a specified recovery (xn = 0.1615). Your task is to determine
the number of stages (n) required to meet the specifications. Take
the equilibrium ratio to be K = 0.8. Here the unknown variable x
is associated with n and the others form the parameter set. Solve
for n using secant and bisection methods using initial guesses of
[10, 30]. Report the number of iterations required for convergence
to the MATLAB built-in convergence tolerance of eps = 10−16 . You
can use the secant.m and bisect.m algorithms outlined in figures
2.4,2.3. You must construct a m-file to define the function rep-
resented by equation (2.17) in terms of the unknown. Here is a
sample function for the first case.
function f=KBS1(x)
% Kremser-Brown-Souders equation
% number of stages is unknown i.e. solve for x=n
K=0.8; L=10; V=10;

x0 = 0.8; ynp1= 0; xn=0.1615; S=L/K/V; %Known values
m=length(x);
for i=1:m
n=x(i);
f(i) = (x0-xn)/(x0-ynp1/K) - ( (1/S)ˆ(n+1) - (1/S)) ...
/ ( (1/S)ˆ(n+1) - 1);
end
Make sure that you understand what the above function does! In
the next two parts you will have to modify this function to solve
for a different unknown! Create a file named KBS1.m and enter the
above function. Then to solve the problem from within MATLAB
enter
» secant(’KBS1’,[10,30],eps,1)
You may wish to plot the function to graphically locate the root
using
» x=10:1:30;
» y=KBS1(x);
» plot(x,y)
or, you can do the same in just one line using
» plot([10:1:30],KBS1(10:1:30))
2. In a performance analysis problem, you will be analyzing an exist-

ing process with a known number of stages (say, n = 10). Suppose
x0 = 0.8, yn+1 = 0, L = 10, xn = 0.01488. Find the amount of gas
V that can be processed. Use an initial guess of [5, 20] and [5, 30]
with both secant and bisect algorithms. Record and comment on
your observations.
3. In another variation of the performance analysis problem, the amount

of gas to be processed (V = 10) may be given. You will have to de-
termine the exit composition xn . Take n = 20, x0 = 0.8, yn+1 =
0, L = 10. Try initial guesses of [0, .2] and [0, 1] on both bisection
and secant algorithms. Record and comment on your observations.
2.12.4 Peng-Robinson Equation of State
The phase behavior of fluids can be predicted with the help of equations
of state. The one developed by Peng & Robinson is particularly well
tuned, accurate and hence is widely used. The equation is given below.
RT a(T )
P= − (2.18)
(V − b) V (V + b) + b(V − b)
where
R 2 Tc2 RTc √ p
a(T ) = 0.45724 α(Tr , ω), b = 0.0778 , α = 1 + m(1 − Tr )
Pc Pc
m = 0.37464 + 1.54226ω − 0.26992ω2 , Tr = T /Tc , and Z = P V /RT .
Whenever (P , T ) are given it is convenient to write equation (2.18) as a

cubic equation in Z
Z 3 − (1 − B)Z 2 + (A − 3B 2 − 2B)Z − (AB − B 2 − B 3 ) = 0 (2.19)
where A = aP /(R 2 T 2 ), B = bP /(RT ).

Use equation (2.19) to compute the density of CO2 in gmole/lit at
P = 20.684MP a and T = 299.82o K. The critical properties required
for CO2 are Tc = 304.2o K, Pc = 7.3862MP a and ω = 0.225, R =
8314P a m3 /kmol o K.
a) Use the function roots(c) in MATLAB to find all the roots of the cubic
equation (2.19) in terms of Z. In MATLAB, how does the function
roots differ from the function fzero?
b) Use the secant method to find the real roots of the above equation.
c) After finding the compressibility Z from each of the above meth-

ods, convert it into molar density and compare with the experimen-
tal value of 20.814gmole/lit
d) Consider the case where you are given (P , V ) and you are asked to
find T . Develop and implement the Newton iteration to solve for
this case. Use the above equation to compute the temperature of
CO2 in o K at P = 20.684 × 106 P a and V = .04783lit/gmole.
Compare the number of iterations required to obtain a solution to
a tolerance of ||f || < 10−15 using an initial guess of T = 250 by
Newton method with that required by the secant method with an
initial guess of [200,310].
e) Suppose that you are given (T , V ) and you are asked to find P ,
which form of equation will you choose? Eqn. (2.18) or Eqn.(2.19)?
What method of solution would you recommend?
2.12.5 Transient heat conduction in semi-infinite slab

Many engineering problems can be cast in the form of determining the
roots of a nonlinear algebraic equation. One such example arises in de-
termining the time required to cool a solid body at a given point to a
predetermined temperature level.
Consider a semi-infinite solid, initially at a temperature of Ti = 200o C
and one side of it is suddenly exposed to an ambient temperature of
Ta = 70o C. The heat transfer coefficient between the solid and sur-
roundings is h = 525W /(m2o C). The thermal conductivity of the solid
is k = 215W /mo C and the thermal diffusivity of the solid is α = 8.4 ×
L1=1000 m D1=0.1m ε1=0.000046m

QT=0.045 m3/s
A B
L2=1100 m D2=0.05m ε2=0.0
Figure 2.10: Turbulent flow in a parallel pipe
10−5 m2 /s. Determine the time required to cool the solid at a distance
of x = 4cm measured from the exposed surface, to T = 120o C. The
temperature profile as a function of time and distance is given by the
following expression.
h i √
θ = 1 − er f (ξ) − e(hx/k+τ) [1 − er f (ξ + τ)]
(T −Ti ) √x
where the dimensionless temperature, θ = (Ta −Ti ) , and ξ = 2 αt
and
h2 αt
τ= k2
, t is the time, x is the distance and er f is the error function.
2.12.6 Turbulent flow in a parallel pipeline system

Consider the flow of an incompressible (ρ = 1000kg/m3 ), Newtonian
fluid (µ = 0.001P a · s) in a parallel pipe system shown in figure 2.10.
The lengths, diameters, roughness for the pipes as well as the total flow
rate are as shown in figure 2.10. Your task is to determine the individual
flow rates in each of the pipe segments 1 and 2. The equation to be
satisfied is obtained based on the fact that the pressure drop between
points A and B is the same. The equation is
L1 v12 L2 v22
f1 (v1 ) = f2 (v2 ) (2.20)
D1 2 D2 2
where v1 , v2 are the velocities in the two pipes and f1 , f2 are the friction
factors given by the Churchill equation.
" #1/12
8 12 1
fi (vi ) = 8 ∗ +
Rei (A + B)1.5
where
16 16
1 37530 D i vi ρ
A = 2.457 ln 0.9
B= and Rei =
(1/Rei ) + 0.27(i /Di ) Rei µ
Finally the mass balance equation provides another constraint as,

π
(D12 v1 + D22 v2 ) = QT
4
This problem can be formulated as two equation in two unknowns (v1 , v2 ),
but your task is to pose this as a single equation in one unknown, v1 , by
rearranging equation 2.20 as,
L1 v12 L2 v22
F (v1 ) = f1 (v1 ) − f2 (v2 ) =0
D1 2 D2 2
i.e., for a given guess of v1 , write a m-file that will calculate F (v1 ). Then
carryout the following calculations.
• Solve the problem using secant algorithm with initial guess of [4.5,5.5].
[Ans:v1 = 4.8703]
• Suppose the total flow rate is increased to 0.09 m3 /s, what will be
the new velocities in the pipes.
• Consider the case where L1 = 1000, L2 = 900 and D1 = 0.1, D2 =

0.09, other values being the same. Is there is flow rate QT for
which the velocities in both pipes will be the same? If so what is
it? [Ans:QT = 0.0017]
• Plot v1 vs QT and v2 vs QT for the case above over a suitable range

of QT .
• Discuss the pros and cons of implementing the Newton method for
this problem.
Mathematics, rightly viewed, possesses not only
truth, but supreme beauty – a beauty cold and aus-
tere, like that of sculpture.
— BERTRAND RUSSELL
Chapter 3
System of linear algebraic

equation
Topics from linear algebra form the core of numerical analysis. Almost
every conceivable problem, be it curve fitting, optimization, simulation
of flow sheets or simulation of distributed parameter systems requiring
solution of differential equations, require at some stage the solution of a
system (often a large system!) of algebraic equations. MATLAB (acronym
for MATrix LABoratory) was in fact conceived as a collection of tools to
aid in the interactive learning and analysis of linear systems and was
derived from a well known core of linear algebra routines written in
FORTRAN called LINPACK.
In this chapter we provide a quick review of concepts from linear

algebra. We make frequent reference to MATLAB implimentation of var-
ious concepts throughout this chapter. The reader is encouraged to try
out these interactively during a MATLAB session. For a more complete
treatment of topics in linear algebra see Hager (1985) and Barnett (1990).
The text by Amundson (1966) is also an excellent source with specific
examples drawn from Chemical Engineering. For a more rigorous, ax-
iomatic introduction within the frame work of linear opeartor theory
see Ramakrishna and Amundson (1985).
53
3.1. MATRIX NOTATION 54
3.1 Matrix notation
We have already used the matrix notation to write a system of linear

algebraic equations in a compact form in sections §1.3.1 and §1.3.2.
While a matrix, as an object, is represented in bold face, its constituent
elements are represented in index notation or as subscripted arrays in
programming languages. For example the following are equivalent.
A = [aij ], i = 1, · · · , m; j = 1, · · · , n
where A is an m × n matrix. aij represents an element of the matrix A

in row i and column j position. A vector can be thought of as an object
with a single row or column. A row vector is represented by,
x = [x1 x2 · · · xn ]
while a column vector can be represented by,

 
y1
 y 
 2 
y=  .. 

 . 
ym
These elements can be real or complex.
Having defined objects like vectors and matrices, we can extend the
notions of basic arithmetic opeartions between scalar numbers to higher
dimensional objects like vectors and matrices. The reasons for doing so
are many. It not only allows us to express a large system of equations
in a compact symbolic form, but a study of the properties of such ob-
jects allows us to develop and codify very efficient ways of solving and
analysing large linear systems. Packages like MATLAB and Mathematica
present to us a vast array of such codified algorithms. As an engineer
you should develop a conceptual understanding of the underlying prin-
ciples and the skills to use such packages. But the most important task
is to indentify each element of a vector or a matrix, which is tied closely
to the physical description of the problem.
3.1.1 Basic operations

The arithmetic opeartions are defined both in symbolic form and using
index notation. The later actually provides the algorithm for implement-
ing the rules of operation using any programing language. The addition
operation between two matrices is defined as,
addition: A=B+C ⇒ aij = bij + cij

This implies an element-by-element addition of the matrices B and C.

Clearly all the matrices involved must have the same dimension. Note
that the addition operation is commutative as with its scalar counter
part. i.e.,
A+B=B+A
Matrix addition is also associative, i.e., independent of the order in which
it is carried out, e.g.,
A + B + C = (A + B) + C = A + (B + C)
The scalar multiplication of a matrix involves multiplying each element

of the matrix by the scalar, i.e.,
scalar multiplication: kA = B ⇒ k aij = bij
Subtraction operation can be handled by combining addition and sccalar

multiplication rules as follwos:
subtraction: C = A + (−1)B = A − B ⇒ cij = aij − bij
The product between two matrices B (of dimension n × m) and C (of

dimension m × r ) is defined as, MATLAB syntax for
the product
m
X operator between
multiplication: A=BC ⇒ aij = bik ckj matrices is
k=1
A=B*C
and the resultant matrix has the dimension n × r . The operation in-
dicated in the index notation is carried out for each value of the free
indices i = 1 · · · n and j = 1 · · · r . The product is defined only if the
dimensions of B, C are compatible - i.e., number of columns in B should
equal the number of rows in C. This implies that while the product B
C may be defined, the product C B may not even be defined! Even when
they are dimensionally compatible, in general
BC 6= BC
i.e., matrix multiplication is not commutative.

Multiplying a scalar number by unity leaves it unchanged. Extension
of this notion to matrices resutls in the definition of identity matrix, MATLAB function
  for producing an
1 0 ··· 0 identity matrix of
 0 1 ··· 0  (
  1 i=j size N is
I= ..

 ⇒ δij = I=eye(N)
 . 0 1 0  0 i 6= j
0 0 ··· 1
Multiplying any matrix A with an identity matrix I of appropriate dimen-

sion leaves the original matrix unchanged, i.e.,
AI = A
This allows us to generalize the notion of division with scalar numbers

to matrices. Division operation can be thought of as the inverse of the MATLAB function
multiplciation operation. For example, given a number, say 2, we can for finding the
define its inverse, x in such a way that the product of the two numbers inverse of a matrix
A is
produce unity. i.e., 2 × x = 1 or x = 2−1 . In a similar way, given a matrix
B=inv(A)
A, can we define the inverse matrix B such that
AB = I or B = A−1
The task of developing an algorithm for finding the inverse of a matrix

will be addressed late in this chapter.
For a square matrix, powers of a matrix A can be defined as,
A2 = AA A3 = AAA = A2 A = AA2
Note that Ap Aq = Ap+q for positve integers p and q. Having extended MATLAB operator
the definition of powers, we can extend the definition of exponential for producing the
from scalars to square matrices as follows. For a scalar α it is, n-th power of a
matrix A is,
α2 X∞
αk Aˆn
α
e =1+α+ + ··· = while the syntax
2 k=0
k! for producing
element-by-
For a matrix A the exponential matrix can be defined as, element power
is,
X∞
A2 Ak A.ˆn.
eA = I + A + + ··· = Make sure that you
2 k=0
k!
understand the
difference between
One operation that does not have a direct counter part in the scalar world
these two
is the transpose of a matrix. It is defined the result of exchanging the operations!
rows and columns of a matrix, i.e.,
MATLAB function
B = A0 ⇒ bij = aji exp(A)
evaluates the
It is easy to verify that exponential
(A + B)0 = A0 + B0 element-by-
element
Something that is not so easy to verify, nevertheless true, is while
expm(A)
(AB)0 = B0 A0 evaluates the true
matrix exponential.
3.2. MATRICES WITH SPECIAL STRUCTURE 57
3.2 Matrices with special structure
A diagonal matirx D has non-zero elements only along the diagonal.

 
d11 0 ··· 0
 0 d 0 
 22 · · · 
D=  .. ..


 . 0 . 0 
0 0 · · · dnn
A lower triangular matrix L has non-zero elements on or below the di-

agonal,  
l11 0 ··· 0
 l 
 21 l22 · · · 0 

L =  .. 
.. 
 . . 0 
ln1 ln2 · · · lnn
A upper triangular matrix U has non-zero elements on or above the
diagonal,  
u11 u12 · · · u1n
 0 u22 · · · u2n 
 
U = .. 
.. 
 0 0 . . 
0 0 · · · unn
A tridiagonal matrix T has non-zero elements on the diagonal and one
off diagonal row on each side of the diagonal
 
t11 t12 0 ··· 0
 t 
 21 t22 t23 0 0 
 .. .. .. 
 . . . 
T = 0 0 
 . 
 . 
 . 0 tn−1,n−2 tn−1,n−1 tn−1,n 
0 ··· 0 tn,n−1 tn,n
A sparse matrix is a generic term to indicate those matrices without any

specific strucutre such as above, but with a small number (typically 10
to 15 %) of non-zero elements.
3.3 Determinant
A determinant of a square matrix is defined in such a way that a scalar

value is associated with the matrix that does not change with certain row
or column operations on the matrix - i.e., it is one of the scalar invariants
of the matrix. In the context of solving a system of linear equations the
3.3. DETERMINANT 58
determinant is also useful is knowing whether the system of equations is

solvable uniquely. It is formed by summing all possible products formed
by choosing one and only one element from each row and column of the
matrix. The precise definition, taken from Amundson (1966), is
X
det(A) = |A| = (−1)h (a1l1 a2l2 · · · anln ) (3.1)
Each term in the summation consists of a product of n elements selected

such that only one element appears from each row and column. The
summation involves a total of n! terms accounted for as follows: for the
first element l1 in the product there are n choices, followed by (n − 1)
choices for the second element l2 , (n − 2) choices for the third element
l3 etc. resulting in a total of n! choices for a particular product. Note
that in this way of counting, the set of second subscripts {l1 , l2 , · · · ln }
will contain all of the numbers in the range 1 to n, but they will not be in
their natural order {1, 2, · · · n}. hence, h is the number of permutations
required to arrange {l1 , l2 , · · · ln } in their natural order. MATLAB function
This definition is neither intutive nor computationaly efficient. But it for computing the
is instructive in understanding the following properties of determinants. determinant of a
square matrix is
1. The determinant of a diagonal matrix D, is simply the product of det(A)
all the diagonal elements, i.e.,
n
Y
det(D) = dkk
k=1
This is the only product term that is non-zero in equation (3.1).

2. A little thought should convince you that it is the same for lower
or upper triangular matrices as well, viz.
n
Y
det(L) = lkk
k=1
3. It should also be clear that if all the elements of any row or column
are zero, then the determinant is zero.
4. If every element of any row or column of a matrix is multiplied
by a scalar, it is equivalent to multiplying the determinant of the
original matrix by the same scalar, i.e.,

ka11 ka12 · · · ka1n a11 a12 · · · ka1n

a a22 · · · a2n
21 a21 a22 · · · ka2n
. .. = = k det(A)
. .. .. .. ..
. . . . . .

an1 an2 · · · ann an1 an2 · · · kann
3.3. DETERMINANT 59
5. Replacing any row (or column) of a matrix with a linear combination

of that row (or column) and another row (or column) leaves the
determinant unchanged.
6. A consequence of rules 3 and 5 is that if two rows (or columns) of

a matrix are indentical the determinant is zero.
7. If any two rows (or columns) are interchanged, it results in a sign

change of the determinant.
3.3.1 Laplace expansion of the determinant

A definition of determinant that you might have seen in an earlier linear
algebra course is
( Pn
aik Aik for any i
det(A) = |A| = Pk=1
n (3.2)
k=1 akj Akj for any j
where Aik , called the cofactor, is given by,
Aik = (−1)i+k Mik
and Mik , called the minor, is the determinant of (n−1)×(n−1) submatrix

of A obtained by deleting ith row and kth column of A. Note that the
expansion in equation (3.2) can be carried out along any row i or column
j of the original matrix A.
Example
Consider the matrix derived in Chapter 1 for the recycle example, viz.
equation (1.8). Let us calculate the determinant of the matrix using the
Laplace expansion algorithm around the first row.

1 0 0.306

det(A) = 0 1 0.702

−2 1 0

1 0.702 0 0.702 0 1

= 1 + (−1)1+2 × 0 × + (−1)1+3 × 0.306 ×

1 0 −2 0 −2 1
= 1 × (−0.702) + 0 + 0.306 × 2 = −0.09
A MATLAB implementation of this will be done as follows:

3.4. DIRECT METHODS 60
»A=[1 0 0.306; 0 1 0.702; -2 1 0] % Define matrix A

»det(A) % calculate the determinant
3.4 Solving a system of linear equations
3.4.1 Cramers rule

Consider a 2 × 2 system of equations,
" #" # " #
a11 a12 x1 b1
=
a21 a22 x2 b2
Direct elimination of the variable x2 results in
(a11 a22 − a12 a22 ) x1 = a22 b1 − a12 b2
which can be written in an laternate form as,
det(A) x1 = det(A(1))
where the matrix A(1) is obtained from A after replacing the first column
with the vector b. i.e.,

b a
1 12
A(1) =
b2 a22
This generalizees to n × n system as follows,

det(A(1)) det(A(k)) det(A(n))
x1 = , ··· xk = , ··· xn = .
det(A) det(A) det(A)
where A(k) is an n × n matrix obtained from A by replacing the kth
column with the vector b. It should be clear from the above that, in order
to have a unique solution, the determinant of A should be non-zero. If
the determinant is zero, then such matrices are called singular.
Example
Continuing with the recycle problem (equation (1.8) of Chapter 1), solu-
tion using Cramer’s rule can be implemented with MATLAB as follows:
    
1 0 0.306 x1 101.48
    
A x = b ⇒  0 1 0.702   x2  =  225.78 
−2 1 0 x3 0
»A=[1 0 0.306; 0 1 0.702; -2 1 0]; % Define matrix A

»b=[101.48 225.78 0]’ % Define right hand side vector b
»A1=[b, A(:,[2 3])] % Define A(1)
»A2=[A(:,1),b, A(:, 3)] % Define A(2)
»A3=[A(:,[1 2]), b ] % Define A(3)
»x(1) = det(A1)/det(A) % solve for coponent x(1)
»norm(A*x’-b) % Check residual
3.4.2 Matrix inverse
We defined the inverse of a matrix A as that matrix B which, when mul-

tiplied by A produces the identity matrix - i.e., AB = I; but we did not
develop a scheme for finding B. We can do so now by combining Cramer’s
rule and Laplace expansion for a determinant as follows. Using Laplace
expansion of the determinant of A(k) around column k,
detA(k) = b1 A1k + b2 A2k + · · · + bn Ank k = 1, 2, · · · , n
where Aik are the cofactors of A. The components of the solution vector,
x are,
x1 = (b1 A11 + b2 A21 + · · · + bn An1 )/det(A)
xj = (b1 A1j + b2 A2j + · · · + bn Anj )/det(A)
xn = (b1 A1n + b2 A2n + · · · + bn Ann )/det(A)
The right hand side of this system of equations can be written as a vector
matrix product as follows,
    
x1 A11 A21 ··· An1 b1
 x2   A12 A22 ··· An2  b2 
  1   
 .. =  .. ..  .. 
  det(A)  ..  
 .   . . .  . 
xn A1n A2n · · · Ann bn
or
x =Bb
Premultiplying the original equation A x = b by A−1 we get
A−1 Ax = A−1 b or x = A−1 b

Comparing the last two equations, it is clear that,

 
A11 A21 · · · An1
 A A22 · · · An2 
1  12  adj(A)
B = A−1 =  . .. =
 .. 
det(A)  .. . .  det(A)
A1n A2n · · · Ann
The above equation can be thought of as the definition for the adjoint
of a matrix. It is obtained by simply replacing each element with its
cofactor and then transposing the resulting matrix.
Inverse of a diagonal matrix
Inverse of a diagonal matrix, D,

 
d11 0 ··· 0
 0 d ··· 0 
 22 
D= 
 .. .. 
 . 0 . 0 
0 0 · · · dnn
is given by,  
1
d11 0 ··· 0
 1 
 0 ··· 0 
 d22 
D−1 = . .. 
 .. 0 . 0 
 
1
0 0 ··· dnn
It is quite easy to verify using the definition of matrix multiplication that

DD−1 = I.
Inverse of a triangular matrix
Inverse of a triangular matrix is also triangular. Suppose U is a given

upper triangular matrix, then the elements of V = U −1 , can be found
sequentially in an efficient manner by simply using the definition UV =
I. This equation, in expanded form, is
    
u11 u12 · · · u1n v11 v12 · · · v1n 1 0 ··· 0
 0 u22 · · · u2n     
   v21 v22 · · · v2n   0 1 · · · 0 
 
..    
..  =  . 
 .. .. .. 
 0 0 . .  . .   . .. 
0 0 · · · unn vn1 vn2 · · · vnn 0 0 ··· 1
We can develop the algorithm ( i.e., find out the rules) by simply carry-
ing out the matrix multiplication on the left hand side and equating it
element-by-element to the right hand side. First let us convince ourself

that V is also upper triangular, i.e.,
vij = 0 i>j (3.3)
Consider the element (n, 1) which is obtianed by summing the product

of each element of n-th row of U (consisting mostly of zeros!) with the
corresponding element of the 1-st column of V . The only non-zero term
in this product is
unn vn1 = 0
Since unn 6= 0 it is clear that vn1 = 0. Carrying out similar arguments in
a sequential manner it is easy to verify equation 3.3 and thus establish
that V is also upper triangular.
The non-zero elements of V can also be found in a sequential manner
as follows. For each of the diagonal elements (i, i) summing the product
of each element of i-th row of U with the corresponding element of the
i-th column of V , the only non-zero term is,
1
vii = i = 1, · · · , n (3.4)
uii
Next, for each of the upper elements (i, j) summing the product of
each element of i-th row of U with the corresponding element of the j-th
column of V , we get,
j
X
uii vij + uir vr j = 0
r =i+1
and hence we get,
j
1 X
vij =− uir vr j j = 2, · · · , n; j > i; i = j − 1, 1 (3.5)
uii r =i+1
Note that equation (3.5) should be applied in a specific order, as oth-

erwise, it may involve unknown elements vr j on the right hand side.
First, all of the diagonal elements of V ( viz. vii ) must be calcualted from
equation (3.4) as they are needed on the right hand side of equation (3.5).
Next the order indicated in equation (3.5), viz. increasing j from 2 to n
and for each j decreasing i from (j − 1) to 1, sould be obeyed to avoid
having unknowns values appearing on the right hand side of (3.5).
A MATLAB implementation of this algorithm is shown in figure 3.1

to illustrate precisely the order of the calculations. Note that the built-
in, general purpose MATLAB inverse function ( viz. inv(U) ) does not
take into account the special structure of a triangular matrix and hence
is computationally more expensive than the invu funcion of figure 3.1.
This is illustrated with the following example.
Example
Consider the upper triangular matrix,

 
1 2 3 4
 0 2 3 1 
 
U = 
 0 0 1 2 
0 0 0 4
Let us find its inverse using both the built-in MATLAB function inv(U)
and the function invu(U) of figure 3.1 that is applicable sepcifically for
an upper triangular matrix. You can also compare the floating point
opeartion count for each of the algorithm. Work through the following
example using MATLAB. Make sure that the function invu of figure 3.1
in the search path of MATLAB.
»U=[1 2 3 4; % matrix entry continues on next three lines

»0 2 3 1;
»0 0 1 2;
»0 0 0 4]; % Definition of matrix A complete
»flops(0); % initialize flop count
»inv(U) % MATLAB built-in function
»flops % examine flops for MATLAB built-in function (ans: 131)
»flops(0); % initialize flop count
»invu(U) % inverse of upper triangular matrix
»flops % examine flops for gauss solver (ans: 51)
3.4.3 Gaussian elimination

Gaussian elimination is one of the most efficient algorithms for solving
a large system of linear algebraic equations. It is based on a systematic
generalization of a rather intuitive elimination process that we routinely
apply to a small, say, (2 × 2) systems. e.g.,
10x1 + 2x2 = 4
function v=invu(u)
% Invert upper triangular matrix
% u - (nxn) matrix
% v - inv(a)
n=size(u,1); % get number of rows in matrix a
for i=2:n
for j=1:i-1
v(i,j)=0;
end
end
for i=1:n
v(i,i)=1/u(i,i);
end
for j=2:n
for i=j-1:-1:1
v(i,j) = -1/u(i,i)*sum(u(i,i+1:j)*v(i+1:j,j));
end
end
Figure 3.1: MATLAB implementation of inverse of an upper triangular

matrix
x1 + 4x2 = 3
From the first equation we have x1 = (4 − 2x2 )/10 which is used to

eliminate the variable x1 from the second equation, viz. (4 − 2x2 )/10 +
4x2 = 3 which is solved to get x2 = 0.6842. In the second phase, the
value of x2 is back substituted into the first equation and we get x1 =
0.2632. We could have reversed the order and eliminated x1 from the
first equation after rearranging the second equation as x1 = (3 − 4x2 ).
Thus there are two phases to the algorithm: (a) forward elimination of
one variable at a time until the last equation contains only one unknown;
(b) back substitution of variables. Also, note that we have used two rules
during the elimination process: (i) two equations (or two rows) can be
interchanged as it is merely a matter of book keeping and it does not in
any way alter the problem formulation, (ii) we can replace any equation
with a linear combination of itself with another equation. A conceptual
description of a naive Gaussian elimination algorithm is shown in figure
3.2. All of the arithmetic operations needed to eliminate one variable at
a time are identified in the illustration. Study that carefully.
We call it a naive scheme as we have assumed that none of the diag-

onal elements are zero, although this is not a requirement for existence
of a solution. The reason for avoiding zeros on the diagoanls is to avoid
division by zeros in step 2 of the illustration 3.2. If there are zeros on
the diagonal, we can interchange two equations in such a way the di-
agonals do not contain zeros. This process is called pivoting. Even if
we organize the equations in such a way that there are no zeros on the
diagonal, we may end up with a zero on the diagonal during the elimina-
tion process (likely to occur in step 3 of illustration 3.2). If that situation
arises, we can continue to exchange that particular row with another one
to avoid division by zero. If the matrix is singular we wil eventually end
up with an unavoidable zero on the diagonal. This situation will arise if
the original set of equations is not linearly independent; in other words
the rank of the matrix A is less than n. Due to the finite precision of
computers, the floating point operation in step 3 of illustration 3.2 will
not result usually in an exact zero, but rather a very small number. Loss
of precision due to round off errors is a common problem with direct
methods involving large system of equations since any error introduced
at one step corrupts all subsequent calcualtions.
A MATLAB implementation of the algorithm is given in figure 3.3

through the function gauss.m. Note that it is
merely for
illustrating the
concepts involved
in the elimination
process; MATLAB
backslash, \
operator provides
a much more
elegant solution to
solve Ax = b in
STEP 1: Arrange A and b into an (n x n+1) matrix

ċ
ċ aa ċa b 
ċa b 
a11 a12 1n a1, n + 1  a11 a12 1n 1
 a21 a22 a2, n + 1  a
2
ċ a? b 
 2n 
21 a22 2n
  
 ?
 an1 an2 ċ a? nn an, n + 1 
  ?
 an1 an2 nn n

for i=1:n STEP 2: Make diagonal elements a(i,i) into 1.0

a12
 1 a11 ċ a1n a1, n + 1
a11 
for j=i+1:n+1;
a a ċa a11
a2, n + 1  a(i,j)=a(i,j)/a(i,i);

 21 22

2n
end
 ?
 an1 an2 ċ a? nn

an, n + 1 
STEP 3 : Make all elements in column i below diagonal into 0
1 a1×2 ċ a1×n a1×, n + 1 

for j=i+1:n  0 a22 − a21 a1×2 ċa − a21 a1×n a2, n + 1 − a21 a1×, n + 1  for k=i+1:n+1;

2n
  a(j,k)=a(j,k)- a(j,i)*a(i,k);
? ?

end  0 an2 − an1 a×
12 ċa nn
× ×
− an1 a1n an, n + 1 − an1 a1, n + 1 
 end
end End of forward elimination. Resulting matrix structure is:
×
 1 a12 ċa × a×1, n + 1 
ċa
1n
0 1 × a×2, n + 1 

? 
2n
 ?
0 0 ċ 1 a×n, n + 1 

STEP 4: Back substitution

×
 1 a12 ċa × a×1, n + 1 
ċa
1n
0 1 × a×2, n + 1  for j=n-1:-1:1;

? 
2n a(j,n+1) = a(j,n+1) - a(j,j+1:n)*a(j+1:n,n+1);
 ?
0 0 ċ 1 a×n, n + 1 
 end
a(n,n+1) = a(n,n+1)
Solution is returned in the last column a(1:n,n+1)
Figure 3.2: Naive Gaussian elimination shceme

function x=gauss(a,b)
% Naive Gaussian elimination. Cannot have zeros on diagonal
% a - (nxn) matrix
% b - column vector of length n
m=size(a,1); % get number of rows in matrix a

n=length(b); % get length of b
if (m ˜= n)
error(’a and b do not have the same number of rows’)
end
%Step 1: form (n,n+1) augmented matrix

a(:,n+1)=b;
for i=1:n
%Step 2: make diagonal elements into 1.0
a(i,i+1:n+1) = a(i,i+1:n+1)/a(i,i);
%Step 3: make all elements below diagonal into 0

for j=i+1:n
a(j,i+1:n+1) = a(j,i+1:n+1) - a(j,i)*a(i,i+1:n+1);
end
end
%Step 4: begin back substitution
for j=n-1:-1:1
a(j,n+1) = a(j,n+1) - a(j,j+1:n)*a(j+1:n,n+1);
end
%return solution
x=a(:,n+1)
Figure 3.3: MATLAB implementation of naive Gaussian elimination

Example
Let us continue with the recycle problem (equation (1.8) of Chapter 1).
First we obtain solution using the built-in MATLAB linear equation solver
( viz. x = A\b and record the floating point operations (flops). Then
we solve with the Gaussian elimination function gauss and compare the
flops. Note that in order to use the naive Gaussian elimination function,
we need to switch the 2nd and 3rd equations to avoid division by zero.
»A=[1 0 0.306; % matrix entry continues on next two lines

»-2 1 0;
»0 1 0.702]; % Definition of matrix A complete
»b=[101.48 0 225.78]’ ;% Define right hand side column vector b
»flops(0) ; % initialize flop count
»A\b % solution is : [23.8920 47.7840 253.5556]
»flops % examine flops for MATLAB internal solver (ans: 71)
»gauss(A,b) % solution is, of course : [23.8920 47.7840 253.5556]
»flops % examine flops for gauss solver (ans: 75)
»inv(A)*b % obtain solution using matrix inverse
»flops % examine flops for MATLAB internal solver (ans: 102)
Example - loss of accuracy & need for pivoting
The need for pivoting can be illustrated with the following simple exam-
ple.
x1 + x2 = 1
x1 + x2 = 2
where is a small number. In matrix form it will be,

" #" # " #
1 x1 1
=
1 1 x2 2
In using naive Gaussian elimination without rearranging the equations,

we first make the diagonal into unity, which results in
1 1
x1 + x2 =

Next we eliminate the variable x1 from the 2nd equation which resutls
in,
1 1
1− x2 = 2 −

Rearranging this and using back substitution we finally get x2 and x1 as,
1
2−
x2 = 1
1−
1 x2
x1 = −

The problem in computing x1 as → 0 should be clear now. As crosses
the threshold of finite precision of the computation (hardware or soft-
ware), taking the difference of two large numbers of comparable magni-
tude, can result in significant loss of precision. Let us solve the problem
once again after rearranging the equations as,
x1 + x2 = 2
x1 + x2 = 1
and apply Gaussian elimination once again. Since the diagonal element
in the first equation is already unity, we can eliminate x1 from the 2nd
equation to obtain,
1 − 2
(1 − )x2 = 1 − 2 or x2 =
1−
Back substitution yields,
x1 = 2 − x2
Both these computations are well behaved as → 0.
We can actually demonstrate this using the MATLAB function gauss
shown in figure 3.3 and compare it with the MATLAB built-in function
A\b which does use pivoting to rearrange the equations and minimize
the loss of precision. The results are compared in table 3.1 for in
the range of 10−15 to 10−17 . Since MATLAB uses double precision, this
range of is the threshold for loss of precision. Observe that the naive
Gaussian elimination produces incorrect results for < 10−16 .
3.4.4 Thomas algorithm

Many problems such as the stagewise separation problem we saw in sec-
tion §1.3.1 or the solution of differential equations that we will see in
Naive elimination Built-in

without pivoting MATLAB
gauss(A,b) A\b
1 × 10−15 [1 1] [1, 1]
1 × 10−16 [2 1] [1, 1]
1 × 10−17 [0 1] [1, 1]
Table 3.1: Loss of precision and need for pivoting
later chapters involve solving a system of linear equations T x = b with

a tridiagonal matrix structure.
     
d1 c1 0 ··· 0 x1 b1
 a d c · · · 0   x   b 
 1 2 2   2   2 
 ..   ..   .. 
T =  . 
 x = 
 .

 b = 
 .


     
 0 0 an−2 dn−1 cn−1   xn−1   bn−1 
0 ··· 0 an−1 dn xn bn
Since we know where the zero elements are, we do not have to carry out
the elimination steps on those entries of the matrix T ; but the essential
steps in the algorithm remain the same as in the Gaussian elimination
scheme and are illustrated in figure 3.4. MATLAB implementation is
shown in figure 3.5.
3.4.5 Gaussian elimination - Symbolic representaion

Given a square matrix A of dimension n × n it is possible to write it is
as the product of two matrices B and C, i.e., A = BC. This process is
called factorization and is in fact not at all unique - i.e., there are inifnitely
many possiblilities for B and C. This is clear with a simple counting of
the unknowns - viz. there are 2×n2 unknown elements in B and C while
only n2 equations can be obtained by equating each element of A with
the corresponding element from the product BC.
The extra degrees of freedom can be used to specify any specific
structure for B and C. For example we can require B = L be a lower
triangular matrix and C = U be an upper triangular matrix.This process
is called LU factorization or decomposition. Since each triangular matrix
has n × (n + 1)/2 unknowns, we still have a total of n2 + n unknowns.
The extra n degrees of freedom is often used in one of three ways:
• Doolitle method assigns the diagonal elements of L to be unity.

Given
 d1 c1 0 0 ? b1 
 a1 ? b2 
0 a d
d2 c2 0
0 0 ? 
0 n−2 n−1 cn − 1 bn − 1 
? ? ? an − 1 dn bn 
STEP 1: Eliminate lower diagonal elements
 d1 c1 0 0 ? b1  d(j) = d(j) - {a(j-1)/d(j-1)}*c(j-1)
for j=2:n 0 d 2∗ c2 ? b2∗
 0

b(j) = b(j) - {a(j-1)/d(j-1)}*b(j-1)
0 0 ? 
0 0 0 d∗n − 1 cn − 1 b∗n − 1 
end  ? ? ? 0 d∗n b∗n 
STEP 2: Back substitution
 d1 c1 0 0 ? b1 
d2∗ c2 ? b2∗
0 
0 for i=n-1:-1:1
0 0 ? 
b(i) = {b(i) - c(i)*b(i+1)}/d(i);
end
0 0 0 d∗n − 1 cn − 1 b∗n − 1 
? ? ? 0 d∗n b∗n  b(n) = b(n)/d(n)
Solution is stored in b
Figure 3.4: Thomas algorithm

function x=thomas(a,b,c,d)
% Thomas algorithm for tridiagonal systems
% d - diagonal elements, n
% b - right hand side forcing term, n
% a - lower diagonal elements, (n-1)
% c - upper diagonal elements, (n-1)
na=length(a); % get length of a

nb=length(b); % get length of b
nc=length(c); % get length of c
nd=length(d); % get length of d
if (nd ˜= nb | na ˜= nc | (nd-1) ˜= na)
error(’array dimensions not consistent’)
end
n=length(d);
%Step 1: forward elimination
for i=2:n
fctr=a(i-1)/d(i-1);
d(i) = d(i) - fctr*c(i-1);
b(i) = b(i) - fctr*b(i-1);
end
%Step 2: back substitution

b(n) = b(n)/d(n);
for j=n-1:-1:1
b(j) = (b(j) - c(j)*b(j+1))/d(j);
end
%return solution
x=b;
Figure 3.5: MATLAB implementation of Thomas algorithm

• Crout method assigns the diagonal elements of U to be unity.
• Cholesky method assigns the diagonal elements of L to be equal to

that ofU - i.e., lii = uii .
While a simple degree of freedom analysis, indicates that it is possible to

factorize a matrix into a product of lower and upper trinagular matrices,
it does not tell us how to find out the unknown elements.
Revisiting the Gaussian elimination method from a different perspec-
tive, will show the connection between LU factorization and Gaussian
elimination. Note that the algorithm outlined in section §3.4.3 is the
most computationally efficient scheme for implimenting Guassian elimi-
nation. The method to be outlined below is not computationally efficient,
but it is a useful conceptual aid in showing the connection between Guas-
sian elimination and LU factorization. Steps 2 and 3 of figure 3.2 that
involve making the diagonal into unity and all the elements below the di-
agonal into zero is equivalent to pre-multuplying A by L1 - i.e., L1 A = U1
or,
    (1) (1)

1
0 · · · 0 a11 a 12 · · · a1n 1 a12 · · · a1n
 aa1121   
 −  a21 a22 · · · a2n   
(1) (1) 
 a11 1 · · · 0      0 a22 · · · a2n 
 .. ..  .. ..  =  .. .. 
 . 0  . .   
 . 0   . . 
an1
− a11 0 0 1 an1 an2 · · · ann (1)
0 an2 · · · ann
(1)
Repeating this process for the 2nd row, we pre-multiply U1 by L2 - i.e.,

L2 U1 = U2 or, in expanded form,
 
1 0 0 ··· 0  (1) (1) (1) 
 1   1 a12 a13 · · · a1n
 0 0 · · · 0  (1) (1)  
 (1)
 1 a12 · · · a1n (2) (2)
 a22
 (1) (1)   0 1 a23 · · · a2n 

 (1)
  
 0 a22 · · · a2n   
a
 0 − 32 1 ··· 0  ..
 a22
(1)  .. ..  =
 0 0
(2)
a33 ··· . 

  . 
.   
 .. .. ..   .. .. 
 . . 0 . 0  (1) (1)  0 0 . ··· . 
 
 an2
(1)  0 an2 · · · ann (2) (2)
0 − (1) 0 0 1 0 0 an3 · · · ann
a22
Continuing this process, we obtain in succession,
L1 A = U1
L2 U1 = U2
L 3 U2 = U 3
L 4 U3 = U 4
Ln−1 Un−2 = Un−1

Note that each Lj is a lower triangular matrix with non-zero elements on
the j-th column and unity on other diagonal elements. Eliminating all of
the intermediate Uj we obtain,
(Ln−1 Ln−2 · · · L1 )A = Un−1
Since the product of all lower triangular matrices is yet another lower
triangular matrix, we can write the above equation as,
LA = U
Also, the inverse of a lower triangular matrix is also lower triangular

- i.e., L̂ = L−1 . Hence a given square matrix A can be factored into a
product of a lower and upper triangular matrix as,
A = L−1 U = L̂U
Although the development in this section provides us with an algorithm

for constructing both L̂ and U, it is quite inefficient. A more direct and
efficient algorithm is developed next in section §3.4.6.
3.4.6 LU decomposition
Consider the product of L and U as shown in the expanded form be-
low. All of the elements of L and U are unkonwn. By carrying out the
matrix product on the left hand side and equating element-by-element
to the right hand side, we can develop sufficient number of equations
to find out all of the unkown elements on the left hand side. The trick,
however is, (as we did with inverting a triangular matrix) to carry out the
calculations in a particualr sequence so that no more than one unknown
appears in each equation.
    
l11 0 0 ··· 0 1 u12 u13 · · · u1n a11 a12 ··· a1n
 l     
 21 l22 0 ··· 0  0 1 u23 · · · u2n   a21 a22 ··· a2n 
    
 l31 l32 l33 · · · 0   0 0 1 · · · u3n  =  a31 a32 ··· a3n 
 . ..   ..   
 . .. ..  . .. .. ..   . .. .. 
 . . . . 0   .. . . . .   .. . ··· . 
ln1 ln2 ln3 · · · lnn 0 0 0 ··· 1 an1 an2 · · · ann
Let us first consider elements in column 1 of L. Caryying out the

multiplication and equating we obtain,
li1 = ai1 i = 1, · · · , n (3.6)

Next focusing on the elements in the first row of U we get, It would be

inefficient to
proceed to the 2nd
u1j = a1j /l11 j = 2, · · · , n (3.7) column of L. Why?
Next we alternate between a column of L and a row of U. The general

expression for any element i in column j of L is,
j−1
X
lij = aij − lik ukj j = 2, · · · , n i = j, · · · , n (3.8)
k=1
Similarly the general expression for any element i in row j of U is,
h Pj−1 i
aji − k=1 ljk uki
uji = j = 2, · · · , n i = j + 1, · · · , n (3.9)
ljj
Equations (3.6-3.9) form the basic algorithm for LU decomposition. In

order to illustrate the implementation of equations (3.6-3.9) as an algo-
rithm, a MATLAB function called LU.m is shown in figure 3.6. Note that
MATLAB provides a built-in function for LU decomposition called lu(A).
Recognizing that A can be factored into the product LU, one can im-
plement an efficient scheme for solving a system of linear algebraic equa-
tions Ax = b repeatedly, particularly when the matrix A remains un-
changed, but different solutions are required for different forcing terms
on the right hand side, b. The equation
Ax = b
can be written as
LUx = b ⇒ Ux = L−1 b = b0
and hence
x = U −1 b0
The operations required for forward elimination and back substitution

are stored in the LU factored matrix and as we saw earlier it is rela-
tively efficient to invert triangular matrices. Hence two additional vector-
matrix products provide a solution for each new value of b.
function [L,U]=LU(a)
% Naive LU decomposition
% a - (nxn) matrix
% L,U - are (nxn) factored matrices
% Usage [L,U]=LU(A)
n=size(a,1); % get number of rows in matrix a
%Step 1: first column of L

L(:,1)=a(:,1);
%Step 2: first row of U

U(1,:)=a(1,:)/L(1,1);
%Step 3: Alternate between column of L and row of U

for j=2:n
for i = j:n
L(i,j) = a(i,j) - sum(L(i,1:j-1)’.*U(1:j-1,j));
end
U(j,j) = 1;
for i=j+1:n
U(j,i)=(a(j,i) - sum(L(j,1:j-1)’.*U(1:j-1,i) ) )/L(j,j);
end
end
Figure 3.6: MATLAB implementation of LU decomposition algorithm

3.5. ITERATIVE METHODS 78
Example
Work through the following exercise in MATLAB to get a feel for the
built-in MATLAB implementation of LU factorization with that given in
figure 3.6. Before you work through the exercise make sure that the file
LU.m that contains the function illustrated in figure 3.6 is in the MATLAB
path. Also, be aware of the upper case function LU of figure 3.6 and the
lower case lu which is the built-in function.
»A=[1 0 0.306; 0 1 0.702; -2 1 0]; % Define matrix A

»b=[101.48 225.78 0]’ % Define right hand vector b
»flops(0) % initialize flop count
»x=A\b % solve using built-in solver
»flops % flops = 74
»flops(0) % re-initialize flop count
»[l,u]=LU(A) %Use algorithm in figure 3.6
»flops % flops = 24
»x=inv(u)*(inv(l)*b) % Solve linear system
»flops % Cumulative flops = 213
»flops(0) % re-initialize flop count
»[L,U]=lu(A) %use built-in function
»flops % flops = 9
»x=inv(U)*(inv(L)*b) % Solve linear system
»flops % Cumulative flops = 183
3.5 Iterative algorithms for systems of linear equations
The direct methods discussed in section §3.4 have the advantage of pro-
ducing the solution in a finite number of calculations. They suffer, how-
ever, from loss of precision due to accumulated round off errors. This
problem is particulalry sever in large dimensional systems (more than
10,000 equations). Iterative methods, on the other hand, produce the
result in an asymptotic manner by repeated application of a simple al-
gorithm. Hence the number of floating point operations required to pro-
duce the final result cannot be known a priori. But they have the natural
ability to eliminate errors at every step of the iteration. For an author-
itative account of iterative methods for large linear systems see Young
(1971).
Iterative methods rely on the concepts developed in Chapter 2. They
are extended naturally from a single equation (one-dimensional system)
to a system of equations (n-dimensional system). The development par-

allels that of section §2.7 on fixed point iterations schemes. Given an
equation of the form, A x = b we can rearrange it into a form,
x (p+1) = G(x (p) ) p = 0, 1, · · · (3.10)
Here we can view the vector x as a point in a n-dimensional vector space

and the above equation as an iterative map that maps a point x (p) into
another point x (p+1) in the n-dimensional vector space. Starting with an
initial guess x (0) we calculate successive iterates x (1) , x (2) · · · until the
sequence converges. The only difference from chapter 2 is that the above
iteration is applied to a higher dimensional system of (n) equations. Note
that G(x) is also vector. Since we are dealing with a linear system, G will
be a linear function of x which is constructed from the given A matrix.
G can typically be represented as
x (p+1) = G(x (p) ) = T x (p) + c. (3.11)
In section §2.7 we saw that a given equation f (x) = 0 can be rearranged

into the form x = g(x) in several different ways. In a similar manner,
a given equation Ax = b can be rearranged into the form x (p+1) =
G(x (p) ) in more than one way. Different choices of G results in different
iterative methods. In section §2.7 we also saw that the condition for
convergence of the seuqence xi+1 = g(xi ) is g 0 (r ) < 1. Recognizing
that the derivative of G(x (p) ) with respect to x (p) is a matrix, G0 = T
a convergence condition similar to that found for the scalar case must
depend on the properties of the matrix T . Another way to demonstrate
this is as follows. Once the sequence x (1) , x (2) · · · converges to, say, r
equation (3.11) becomes,
r = T r + c.
Subtracting equation (3.11) from the above,
(x (p+1) − r) = T (x (p) − r).
Now, recognizing that (x (p) − r) = (p) is a measure of the error at it-

eration level p, we have
(p+1) = T (p) .
Thus, the error at step (p + 1) depend on the error at step (p). If the
matrix T has the property of amplifying the error at any step, then the
iterative sequence will diverge. The property of the matrix T that de-
termines this feature is called the spectral radius. The spectral radius is
defined as the largest eigenvalue in magnitude of T . For converence of

the iterative sequence the spectral radius of T should be less than unity,
ρ(T ) < 1 (3.12)
3.5.1 Jacobi iteration

The Jacobi iteration rearranges the given equations in the form,
(p+1) (p) (p) (p)
x1 = (b1 − a12 x2 − a13 x3 − · · · − a1n xn )/a11
 
j−1 n
(p+1)
X (p)
X (p)
xj = bj − ajk xk − ajk xk  /ajj (3.13)
k=1 k=j+1
(p+1) (p) (p) (p)
xn = (bn − an1 x1 − an2 x2 − · · · − an,n−1 xn−1 )/ann
where the variable xj has been extracted form the j − th equation and
expressed as a function of the remaining variables. The above set of
equations can be applied repetitively to update each component of the
unknown vector x=(x1 , x2 , · · · , xn ) provided an inital guess is known
for x. The above equation can be written in matrix form as, Note that MATLAB
functions
Lx (p) + Dx (p+1) + Ux (p) = b diag
tril
where the matrices D, L, U are defined in term of components of A as triu
follows. are useful in
 
a11 0 ··· 0 extracting parts of
 0 a  a given matrix A
 22 · · · 0 
D=  .. ..


 . 0 . 0 
0 0 · · · ann
   
0 0 ··· 0 0 a12 · · · a1n
 a21 0 ··· 0   0 0 · · · a2n 
   
L=
 .. ..

 U =  .. .. 

 . . 0   0 0 . . 
an1 an2 ··· 0 0 0 ··· 0
which can be rearranged as,
x (p+1) = D−1 (b − (L + U)x (p) ) (3.14)
and hence G(x (p) ) = −D−1 (L + U)x (p) + D−1 b and G0 = T = −D−1 (L +
U). This method has been shown to be convergent as long as the original
matrix A is diagonally dominant, i.e.,
n
X

|aii | ≥ aij ı = 1, · · · , n
j=1,j6=i
An examination of equation (3.13) reveals that none of the diagonal

elements can be zero. If any is found to be zero, one can easily exchange
the positions of any two equations to avoid this problem. Equation (3.13)
is used in actual computational implementation, while the matrix form
of the equation (3.14) is useful for conceptual description and conver-
gence analysis. Note that each element in the equation set (3.13) can be
updated independent of the others in any order because the right hand
side of equation (3.13) is evaluated at the p-th level of iteration. This
method requires that x (p) and x (p+1) be stored as two separate vectors
until all elements of x (p+1) have been updated using equation (3.13). A
minor variation of the algorithm which uses a new value of the element
in x (p+1) as soon as it is available is called the Gauss-Seidel method. It
has the dual advantage of faster convergence than the Jacobi iteration
as well as reduced storage requirement for only one array x.
3.5.2 Gauss-Seidel iteration
In the Gauss-Seidel iteration we rearrange the given equations in the

form,
(p+1) (p) (p) (p)

x1 = (b1 − a12 x2 − a13 x3 − · · · − a1n xn )/a11
 
j−1 n
(p+1)
X (p+1)
X (p)
xj = bj − ajk xk − ajk xk  /ajj (3.15)
k=1 k=j+1
(p+1) (p+1) (p+1) (p+1)
xn = (bn − an1 x1 − an2 x2 − · · · − an,n−1 xn−1 )/ann
Observe that known values of the elements in x (p+1) are used on the
right hand side of the above equations (3.15) as soon as they are available
within the same iteration. We have used the superscripts p and (p + 1)
explicitly in equation (3.15) to indicate where the newest values occur.
In a computer program there is no need to assign separate arrays for
p and (p + 1) levels of iteration. Using just a single array for x will
automatically propagate the newest values as soon as they are updated.
The above equation can be written symbolically in matrix form as,
Lx (p+1) + Dx (p+1) + Ux (p) = b

where the matrices D, L, U are defined as before. Factoring x (p+1) we

get,
x (p+1) = (L + D)−1 (b − Ux (p) ) (3.16)
and hence G(x (p) ) = −(L + D)−1 Ux (p) + (L + D)−1 b and G0 = T =
−(L + D)−1 U. Thus the convergence of this scheme depends on the
spectral radius of the matrix, T = −(L + D)−1 U. This method has
also been shown to be convergent as long as the original matrix A is
diagonally dominant.
MATLAB implementation of the Gauss-Seidel algorithm is shown in
figure 3.7.
3.5.3 Successive over-relaxation (SOR) scheme

The relaxation scheme can be thought of as a convergence acceleration
scheme that can be applied to any of the basic iterative methods like
Jacobi or Gauss-Seidel schemes. We introduce an extra parameter, ω
often called the relaxation parameter and choose its value in such a way
that we can either speed up convergence by using ω > 1 (called over-
relaxation ) or in some difficult problems with poor initial guess we can
attempt to enlarge the region of convergence using ω < 1 (called under-
relaxation). Let us illustrate the implementation with the Gauss-Seidel
scheme. The basic Gauss-Seidel scheme is:
 
j−1 n
(p+1)
X (p+1)
X (p)
t := xj = bj − ajk xk − ajk xk  /ajj (3.17)
k=1 k=j+1
(p+1)
Instead of accepting the value of xj computed from the above for-
mula as the current value, we store it in a temporary variable t and form
(p+1)
a better (or accelerated) estimate of xj from,
(p+1) (p) (p)

xj = xj + ω [t − xj ]
Observe that if ω = 1, the method remains the same as Gauss-Seidel

scheme. For ω > 1, then the difference between two successive iterates
(the term in the square bracketts) is amplified and added to the current
(p)
value xj .
The above opeartions can be written in symbollic matrix form as,
x (p+1) = x (p) + ω[{D−1 (b − Lx (p+1) − Ux (p) )} − x (p) ]
where the term in braces represent the Gauss-Seidel scheme. After ex-
tracting x (p+1) from the above equation, it can be cast in the standard
function x=GS(a,b,x,tol,max)
% Gauss-Seidel iteration
% a - (nxn) matrix
% b - column vector of length n
% x - initial guess vector x
% tol - convergence tolerance
% max - maximum number of iterations
% Usage x=GS(A,b,x)
m=size(a,1); % get number of rows in matrix a

n=length(b); % get length of b
if (m ˜= n)
error(’a and b do not have the same number of rows’)
end
if nargin < 5, max=100; end
if nargin < 4, max=100; tol=eps; end
if nargin == 2
error(’Initial guess is required’)
end
count=0;
while (norm(a*x-b) > tol & count < max),

x(1) = ( b(1) - a(1,2:n)*x(2:n) )/a(1,1);
for i=2:n-1
x(i) = (b(i) - a(i,1:i-1)*x(1:i-1) - ...
a(i,i+1:n)*x(i+1:n) )/a(i,i);
end
x(n) = ( b(n) - a(n,1:n-1)*x(1:n-1) )/a(n,n);
count=count+1;
end
if (count >= max)

fprintf(1,’Maximum iteration %3i exceeded\n’,count)
fprintf(1,’Residual is %12.5e\n ’,norm(a*x-b) )
end
Figure 3.7: MATLAB implementation of Gauss-Seidel algorithm

iterative from of equation (3.11) as,

x (p+1) = (D + ωL)−1 [(1 − ω)D − ωU]x (p) + ω(D + ωL)−1 b (3.18)
Thus the convergence of the relaxation method depends on the spectral
radius of the matrix T (ω) := (D + ωL)−1 [(1 − ω)D − ωU]. Since this
matrix is a function of ω we have gained a measure of control over the
convergence of the iterative scheme. It has been shown (Young, 1971)
that the SOR method is convergent for 0 < ω < 2 and that there is an
optimum value of ω which results in the maximum rate of convergence.
The optimum value of ω is very problem dependent and often difficult
to determine precisely. For linear problems, typical values in the range
of ω ≈ 1.7 ∼ 1.8 are used.
3.5.4 Iterative refinement of direct solutions

We have seen that solutions obtained with direct methods are prone
to accumulation of round-off errors, while iterative methods have the
natural ability to remove errors. In an attempt to combine the best of
both worlds, one might construct an algorithm that takes the error-prone
solution from a direct method as an initial guess to an iterative method
and thus improve the accuracy of the solution.
Let us illustrate this concept as applied to improving the accuracy of
a matrix inverse. Suppose B is an approximate (error-prone) inverse of a
given matrix A obtained by one of the direct methods outlined in section
§3.4. If B is the error in B then
A(B + B ) = I or AB = (I − AB)
We do not, of course, know B and our task is to attempt to estimate
it approximately. Premultiplying above equation by B, and recognizing
BA ≈ I, we have
B = B(I − AB)
Observe carefully that we have used the approximation BA ≈ I on the
left hand side where products of order unity are involved and not on the
right hand side where difference between numbers of order unity are
involved. Now we have an estimate of the error B which can be added
to the approximate result B to obtain,
B + B = B + B(I − AB) = B(2I − AB)
Hence the iterative sequence should be,
B(p+1) = B(p) (2I − AB(p) ) p = 0, 1 · · · (3.19)

3.6. GRAM-SCHMIDT ORTHOGONALIZATION PROCEDURE 85
3.6 Gram-Schmidt orthogonalization procedure
Given a set of n linearly independent vectors, {x i | i = 1, · · · , n}, the

objective is to produce an orthonormal set of vectors {ui | i = 1, · · · , n}.
√
We begin by normalizing the x 1 vector using the norm ||x 1 || = x 1 · x 1
and call it u1 .
x1
u1 =
||x 1 ||
Subsequently we construct other vectors orthogonal to u1 and normalize

each one. For example we construct u02 by subtracting u1 from x 2 in such
a way that u02 contains no components of u1 - i.e.,
u02 = x 2 − c0 u1
In the above c0 is to be found in such a way that u02 is orthogonal to u1 .
uT1 · u02 = 0 = uT1 · x 2 − c0 or c0 = uT1 · x 2
Similarly we have,
u03 = x 3 − c1 u1 − c2 u2
Requiring orthogonality with respect to both u1 and u2
uT1 · u03 = 0 = uT1 · x 3 − c1 or c1 = uT1 · x 3
uT2 · u03 = 0 = uT2 · x 3 − c2 or c2 = uT2 · x 3
u03 = x 3 − (uT1 · x 3 ) u1 − (uT2 · x 3 ) u2
In general we have,
s−1
X u0s
u0s = x s − (uTj · x s ) uj us = s = 2, · · · n; (3.20)
j=1
||u0s ||
3.7. THE EIGENVALUE PROBLEM 86
3 v35 5 p
5
1 v12 2
p1 v34
p2
v46 p6
4 6
Figure 3.8: Laminar flow in a pipe network
3.7 The eigenvalue problem
3.7.1 Wei-Prater analysis of a reaction system
3.8 Singular value decomposition
3.9 Genaralized inverse
3.10 Power iteration
3.11 Software tools
3.11.1 Lapack, Eispack library
3.11.2 MATLAB
3.11.3 Mathematica
3.12.1 Laminar flow through a pipeline network
Consider laminar flow through the network shown in figure 3.8. The gov-
erning equations are the pressure drop equations for each pipe element
i − j and the mass balance equation at each node.
The pressure drop between nodes i and j is given by,
32µlij
pi − pj = αij vij where αij = (3.21)
d2ij
The mass balance at node 2 is given, for example by,
d212 v12 = d223 v23 + d224 v24 (3.22)

Similar equations apply at nodes 3 and 4. Let the unknown vector be
x = [p2 p3 p4 v12 v23 v24 v34 v35 v46 ]

There will be six momentum balance equations, one for each pipe ele-
ment, and three mass balance (for incompressible fluids volume balance)
equations, one at each node. Arrange them as a system of nine equations
in nine unknowns and solve the resulting set of equations. Take the vis-
cosity of the fluid, µ = 0.1P a · s. The dimensions of the pipes are given
below.
Table 1
Element no 12 23 24 34 35 46
dij (m) 0.1 0.08 0.08 0.10 0.09 0.09

lij (m) 1000 800 800 900 1000 1000
a) Use MATLAB to solve this problem for the specified pressures of

p1 = 300kP a and p5 = p6 = 100kP a. You need to assemble the
system of equations in the form A x = b. Report flops. When
reporting flops, report only for that particular operation - i.e., ini-
tialize the counter using flops(0) before every operation.
• Compute the determinant of A. Report Flops.

• Compute the LU factor of A using built-in function lu. Report
flops. What is the structure of L? Explain. The function LU
provided in the lecture notes will fail on this matrix. Why?
• Compute the solution using inv(A)*b. Report flops.
• Compute the rank of A. Report Flops.
• Since A is sparse ( i.e., mostly zeros) we can avoid unnecessary
operations, by using sparse matrix solvers. MATLAB Ver4.0
(not 3.5) provides such a facility. Sparse matrices are stored
using triplets (i, j, s) where (i, j) identifies the non-zero en-

try in the matrix and s its corresponding value. The MATLAB
function find(A) examines A and returns the triplets. Use,
» [ii,jj,s]=find(A)
Then construct the sparse matrix and store it in S using
» S=sparse(ii,jj,s)
Then to solve using the sparse solver and keep flops count,
use
» flops(0); x = S\b; flops
Compare flops for solution by full and sparse matrix solution.

To graphically view the structure of the sparse matrix, use
» spy(S)
Remember that you should have started MATLAB under X-

windows for any graphics display of results!
• Compute the determinant of the sparse matrix, S (should be
the same as the full matrix!). Report and compare flops.
b) Find out the new velocity and pressure distributions when p6 is

changed to 150kP a.
c) Suppose the forcing (column) vector in part (a) is b1 and that in

part (b) is b2, report and explain the difference in flops for the fol-
lowing two ways of obtaining the two solutions using the sparse
matrix. Note that in the first case both solutions are obtained si-
multaneously.
» b = [b1, b2];flops(0);x = S\b,flops

» flops(0);x1 = S\b1, x2 = S\b2,flops
Repeat the above experiment with the full matrix, A and report
flops.
d) Comment on how you would adopt the above problem formulation

if a valve on line 34 is shut so that there is no flow in that line 34.
Let knowledge grow from more to more,
But more of reverence in us dwell;
That mind and soul, according well,
May make one music as before.
— ALFRED TENNYSON
Chapter 4
Systems of nonlinear algebraic

equations
In this chapter we extend the concepts developed in Chapter 2 - viz.

finding the roots of a system of nonlinear algebraic equations of the
form,
f (x) = 0 (4.1)
where f (x) is a vector function of x - i.e., there are n equations which
can be written in expanded or component form as,
f1 (x1 , x2 , · · · , xn ) = 0
f2 (x1 , x2 , · · · , xn ) = 0
···
fn (x1 , x2 , · · · , xn ) = 0
As with the scalar case, the equation is satisfied only at selected values of
x = r = [r1 , r2 , · · · , rn ], called the roots. The separation process model
discussed in section §1.3.1 (variations 2 and 3, in particular) and the
reaction sequence model of section §1.3.5 are two of the many exam-
ples in chemical engineering that give rise to such non-linear system of
equations. As with the scalar case, the equations often depend on other
parameters, and we will represent them as
f (x; p) = 0 (4.2)
89
where p represents a set of known parameter values. In such cases it

may be required to construct solution families for ranges of values of
p - i.e., x(p). This task is most efficiently achieved using continuation
methods. For some of the recent developments on algorithms for non-
linear equations see Ortega and Rheinboldt (1970), Rabinowitz (1970),
Scales (1985) and Drazin (1992).
4.1 Newton’s method
For a scalar equation a geometrical interpretation of the Newton scheme

is easy to develop as shown in figure 2.2d. This is difficult to visualize
for higher dimensional systems. The algorithm developed in section
§2.5 can, however, be generalized easily to higher dimensional systems.
The basic concept of linearizing a nonlinear function remains the same
as with a scalar case. We need to make use of the multivariate form
of the Taylor series expansion. We will illustrate the concepts with a
two-dimensional system of equations written in component form as,
f1 (x1 , x2 ) = 0
f2 (x1 , x2 ) = 0
Thus the vectors f (x) = [f1 (x1 , x2 ), f2 (x1 , x2 )] and x = [x1 , x2 ] con-
tain two elements. Let the roots be represented by r = [r1 , r2 ] - i.e.,
f (r) = 0.
Suppose x (0) be some known initial guess for the solution vector x
and let the root be at a small displacement δ from x (0) - i.e.,
r = x (0) + δ
If we can device a scheme to estimate δ then we can apply such a scheme

repeatedly to get closer to the root r. Variations in the function value
f1 (x1 , x2 ) can be caused by variations in either components x1 or x2 .
Recognizing this, a bi-variate Taylor series expansion around x (0) can be
written as,
(0) (0) (0) (0)

f1 (x1 + δ 1 , x2 + δ2 ) = f1 (x1 , x2 ) +

∂f1
∂f1

δ + δ2 +O(δ2 )
∂x1 [x1(0) ,x2(0) ] ∂x2 [x1(0) ,x2(0) ]
1
| {z } | {z }
variation due to x1 variation due to x2
(0) (0) (0) (0)

f2 (x1 + δ 1 , x2 + δ2 ) = f2 (x1 , x2 ) +

∂f2
∂f2

δ1 + δ2 +O(δ2 )
∂x1 [x1(0) ,x2(0) ] ∂x2 [x1(0) ,x2(0) ]
| {z } | {z }
variation due to x1 variation due to x2
Since δ is supposed to be small, we can neglect higher order terms O(δ2 )

in the above equations and this step is the essence of the linearization
process. Since x (0) + δ = r and f (r) = 0, the left hand sides of the
above equations are zero. Thus we get,

(0) (0) ∂f1
∂f1

0 = f1 (x1 , x2 ) + δ + δ2
∂x1 [x1(0) ,x2(0) ] ∂x2 [x1(0) ,x2(0) ]
1

(0) (0) ∂f2
∂f2

0 = f2 (x1 , x2 ) + δ + δ2
∂x1 [x1(0) ,x2(0) ] ∂x2 [x1(0) ,x2(0) ]
1
These are two linear equations in two unknowns [δ1 , δ2 ]. Note that the
two functions [f1 , f2 ] and their four partial derivatives are required to
(0) (0)
be evaluated at the guess value of [x1 , x2 ]. The above equations can
be arranged into matrix form as,
" # " #  " #
∂f1 ∂f1
0 f1 δ1
= + ∂x1
∂f2
∂x2
∂f2

0 f2 δ2
∂x1 ∂x2
or in symbolic form
0 = f (0) + J (0) δ
where J (0) is called the Jacobian matrix and the superscript is a reminder
that quantities are evaluated using the current guess value of x (0) . Thus,
the displacement vector δ is obtained by solving the linear system,
δ = −J −1 f
In general then, given x (0) , the algorithm consists of (i) evaluating the
function and the Jacobian at the current iterate x (k) , (ii) solving the linear
system for the displacement vector δ(k) and (iii) finding the new estimate
for the iterate x (k+1) from the equations
δ(k) = −[J (k) ]−1 f (k) x (k+1) = x (k) + δ(k) k = 0, 1, · · · (4.3)

Convergence check
The final step is to check if we are close enough to the desired root r
so that we can terminate the iteration. One test might be to check if the
absolute difference between two successive values of x is smaller than
a specified tolerance. This can be done by computing the norm of δ
|δ| ≤
Another test might be to check if the absolute value of the function f

at the end of every iteration is below a certain tolerance. Since we are
dealing with vectors, once again a norm of f must be calculated.
|f | ≤
The norm can be computed as

qP
n 2
i=1 fi
|f | =
n
where n is the dimension of the system.
In addition to the above convergence tests, we might wish to place a
limit on the number of times the iteration is repeated. A MATLAB func-
tion, constructed in the form of a m-file, is shown in figure 4.1. Note that
this MATLAB function requires an initial guess as well as two external
fucntions for computing the function values and the Jacobian.
Example of reactors in series
This example is from section §1.3.5 and consists of a system of nonlinear

equations that model a series of continuously stirred tank reactors. A
sketch is shown in figure 1.6 Recall that the model equations are given
by,
fi := βa2i + ai − ai−1 = 0 i = 1···n (4.4)
As discussed in section §1.3.5, there are n equations and (n + 2)
variables in total. Hence we have two degrees of freedom. We consider
a design situation where inlet and outlet concentrations are specified
as, say, a0 = 5.0, an = 0.5 mol/lit. The unknown vector consists of n
variable elements,
x = {a1 , a2 · · · an−1 , β}
We are required to determine the volume of each reactor for a given
number n. The volume is given by the expression β = kV /F . The rate
constant is k = 0.125 lit/(mol min) and the feed rate is F = 25 lit/min.
function x=newton(Fun,Jac,x,tol,trace)
% Newton method for a system of nonlinear equations
% Fun - name of the external function to compute f
% Jac - name of the externan function to compute J
% x - vector of initial guesses
%
% Usage newton(’Fun’,’Jac’,x)
%Check inputs
max=25;
n=length(x);
count=0;
f=1;
while (norm(f) > tol & count < max), %check convergence
f = feval(Fun,x); %evaluate the function
J = feval(Jac,x); %evaluate Jacobian
x = x -J\f; %update the guess
count=count+1;
if trace,
fprintf(1,’Iter.# = %3i Resid = %12.5e\n’, count,norm(f));
end
end
if (count >= max)

fprintf(1,’Maximum iteration %3i exceeded\n’,count)
fprintf(1,’Residual is %12.5e\n ’,norm(f) )
end
Figure 4.1: MATLAB implementation of Newton scheme

In solving the above equations we are primarily interested in β, but we

also get all of the intermediate concentrations. Before we can invoke the
algorithm of figure 4.1 we need to write two functions for evaluation the
function values and the Jacobian. The Jacobian is given by,
 
∂f1 ∂f1
 ∂x1 0 0 · · · ∂xn 
 ∂f2 ∂f2 ∂f2 
 ∂x1 ∂x2 0 ··· ∂xn 
 
 ∂f3 ∂f3 ∂f3 
J =  0 ∂x2 ∂x3 · · · ∂xn 
 
 .. .. .. .. 
 . 0 . . . 
 
∂fn ∂fn
0 ··· 0 ∂xn−1 ∂xn
 
2xn x1 + 1 0 0 · · · x12
 

 −1 2xn x2 + 1 0 · · · x22 

 0 −1 2xn x3 + 1 · · · x32 
=   (4.5)
 .. .. .. .. 
 . . 
 . 0 . 
0 ··· 0 −1 a2n
The MATLAB functions to compute the function in equation (4.4) and

the Jacobian in equation (4.5) are shown in figure 4.2. Work through
the following example after ensuring that the three m-files newton.m,
cstrF.m and cstrJ.m are the MATLAB search path.
»x=[1 .5 .2 .1 0]’ % Initial guess for n=5

»r=newton(’cstrF’,’cstrJ’,x,1.e-10,1) % call Newton scheme
Iter.# = 1 Resid = 4.06325e+00
Iter.# = 2 Resid = 1.25795e+01
Iter.# = 3 Resid = 2.79982e+00
Iter.# = 4 Resid = 4.69658e-01
Iter.# = 5 Resid = 2.41737e-01
Iter.# = 6 Resid = 4.74318e-03
Iter.# = 7 Resid = 1.61759e-06
Iter.# = 8 Resid = 1.25103e-12
r =
2.2262
1.2919
0.8691
0.6399
0.5597
function f=cstrF(x)
% Reactor in series model, the function
% x=[a(1),a(2), .., a(n-1),beta]
% f(i) = beta a(i)ˆ2 + a(i) - a(i-1)
n=length(x);
a0=5.0; an=0.5; %define parameters in equation
f(1) = x(n)*x(1)ˆ2 + x(1) - a0;

for i = 2:n-1
f(i)= x(n)*x(i)ˆ2 + x(i) - x(i-1);
end
f(n) = x(n)*anˆ2 + an - x(n-1);
f=f’;
function J=cstrJ(x)
% Reactor in series model, the Jacobian
% x=[a(1),a(2), .., a(n-1),beta]
n=length(x);
a0=5.0; an=0.5; %define parameters in equation
J(1,1) = x(n)*2*x(1) + 1;
J(1,n) = x(1)ˆ2;
for i = 2:n-1
J(i,i) = x(n)*2*x(i) + 1;
J(i,i-1) = -1;
J(i,n) = x(i)ˆ2;
end
J(n,n) = anˆ2;
J(n,n-1) = -1;
Figure 4.2: CSTR in series example - function & Jacobian evaluation

4.2. EULER-NEWTON CONTINUATION 96
»V=r(5)*25/.125 % compute V (ans:111.9427)

»x=[1:-.1:.1]’ % repeat solution for n=10
»r=newton(’cstrF’,’cstrJ’,x,1.e-10,1) % call Newton scheme
»V=r(10)*25/.125 % compute V (ans:44.9859)
In the above example, observe first that the number of reactors is defined
implicitly by the length of the vector x. Secondly observe the quadratic
convergence of the iteration sequence - viz. the residual goes down from
10−3 in iteration number 6 to 10−6 in iteration number 7 and 10−12 in
iteration number 8. In other words the number of significant digits of
accuracy doubles with every iteration once the iteration reaches close to
the root.
4.2 Euler-Newton continuation
4.3 Arc-length continuation
4.4 Quasi-Newton methods
4.4.1 Levenberg-Marquadt method
4.4.2 Steepest descent method
4.4.3 Broyden’s method
4.5.1 Turbulent flow through a pipeline network

Consider turbulent flow through the network shown in figure 3.8. The
governing equations are the pressure drop equations for each pipe ele-
ment i − j and the mass balance equation at each node. The pressure
drop between nodes i and j is given by,
2 2f ρlij
pi − pj = αij vij where αij = (4.6)
d2ij
In general the friction factor f is given by the Moody chart or its equiv-
alent Churchill correlation. In fully developed turbulent flow it is rel-
atively insensitive to changes in Re. Hence take it to be a constant
f = 0.005.
The unknown vector is,
x = [p2 p3 p4 v12 v23 v24 v34 v35 v46 ]
There will be six momentum balance equations, one for each pipe ele-
ment, and three mass balance (for incompressible fluids volume balance)
equations, one at each node. Arrange them as a system of nine equations
in nine unknowns and solve the resulting set of equations. Take the vis-
cosity of the fluid, µ = 0.1P a · s and the density as ρ = 1000kg/m3 . The
dimensions of the pipes are given below.
Table 1
Element no 12 23 24 34 35 46
dij (m) 0.1 0.08 0.08 0.10 0.09 0.09

lij (m) 1000 800 800 900 1000 1000
a) Use MATLAB to solve this problem using Newton method for the
specified pressures of p1 = 300kP a and p5 = p6 = 100kP a. Re-
port the number of iterations and the flops.
b) Implement the Jacobi iteration to solve this problem. Report the

number of iterations and the flops.
c) Implement the Gauss-Seidel iteration to solve this problem. Report

the number of iterations and the flops.
d) Implement the Successive-relaxation iteration to solve this prob-

lem. Report the number of iterations and the flops.
e) Find out the new velocity and pressure distributions when p6 is

changed to 150kP a.
f) Comment on how you would adopt the above problem formulation

if a valve on line 34 is shut so that there is no flow in that line 34.
The chess-board is the world; the pieces are the phe-
nomena of the universe; the rules of the game are
what we call the laws of Nature. The player on the
other side is hidden from us. We know that his play
is always fair, just, and patient. But also we know,
to our cost, that he never overlooks a mistake, or
makes the smallest allowance for ignorance.
— T.H. HUXLEY
Chapter 5
Functional approximations
In previous chapters we have developed algorithms for solving systems

of linear and nonlinear algebraic equations. Before we undertake the
development of algorithms for differential equations, we need to develop
some basic concepts of functional approximations. In this respect the
present chapter is a bridge between the realms of lumped parameter
models and distributed and/or dynamic models.
There are at least two kinds of functional approximation problems

that we encounter frequently. In the first class of problem, a known
function f (x) is approximated by another function, Pn (x) for reasons
of computational necessity or expediency. As modelers of physical phe-
nomena, we often encounter a second class of problem in which there
is a need to represent an experimentally observed, discrete set of data
of the form {xi , fi |i = 1, · · · n} as a function of the form f (x) over the
domain of the independent variable x.
98
5.1. APPROXIMATE REPRESENTATION OF FUNCTIONS 99
5.1 Approximate representation of functions
5.1.1 Series expansion

As an example of the first class of problem, consider the evaluation of
the error function given by,
Zx
2 2
er f (x) = √ e−ξ dξ
π 0
Since the integral does not have a closed form expression, we have to
use a series expansion for,
X∞
2 (−1)k ξ 2k
e−ξ =
k=0
k!
Note that this expansion is around ξ = 0. We can integrate the series

expansion term-by-term to obtain,
∞
2 X (−1)k x 2k+1
er f (x) = √
π k=0 (2k + 1)k!
We can now choose to approximate this function as,

n
2 X (−1)k x 2k+1
er f (x) ≈ P2n+1 (x) = √ + R(x)
π k=0 (2k + 1)k!
by truncating the infinite series to n terms. The error introduced by

truncating such a series is called the truncation error and the magnitude
of the residual function, R(x) represents the magnitude of the truncation
error. For x close to zero a few terms of the series (small n) are adequate.
The convergence of the series is demonstrated in Table 5.1. It is clear
that as we go farther away from x = 0, more terms are required for
P2n+1 (x) to represent er f (x) accurately. The error distribution, defined
as (x, n) := |er f (x) − P2n+1 (x)|, is shown in figure 5.1. It is clear
from figure 5.1a, that for a fixed number of terms, say n = 8, the error
increases with increasing values of x. For larger values of x, more terms
are required to keep the error small. For x = 2.0, more than 10 terms
are required to get the error under control.
5.1.2 Polynomial approximation

In the above example we chose to construct an approximate function to
represent f (x) = er f (x) by expanding f (x) in Taylor series around x =
n P2n+1 (x = 0.5) P2n+1 (x = 1.0) P2n+1 (x = 2.0)

2 0.5207 0.865091 2.85856
4 0.5205 0.843449 2.09437
6 0.5205 0.842714 1.33124
8 0.5205 0.842701 1.05793
10 0.5205 0.842701 1.00318
20 0.5205 0.842701 0.995322
Exact 0.5205 0.842701 0.995322
Table 5.1: Convergence of P2n+1 (x) to er f (x) at selected values of x
0.0007 0.000025
0.0006 (a)
0.00002 (b) x=0.5
0.0005
error
0.0004 0.000015
0.0003 n=2 4 6 8 0.00001
0.0002 -6
0.0001 5. 10
0 0
0 0.5 1 1.5 2 2 4 6 8 10
x n
1.75
0.01 (c) x=1.0 (d) x=2.0
1.5
0.008 1.25
error
0.006 1
0.75
0.004
0.5
0.002
0.25
0 0
2 4 6 8 10 2 4 6 8 10
n n
Figure 5.1: Error distribution of (x, n) := |er f (x) − P2n+1 (x)| for dif-
ferent levels of truncation
0. This required that all the higher order derivative be available at x = 0.

Also, since the expansion was around x = 0, the approximation fails
increasingly as x moves away from zero. In another kind of functional
approximation we can attempt to get a good representation of a given
function f (x) over a range x ∈ [a, b]. We do this by choosing a set of
n basis functions, {φi (x)|i = 1 · · · n} that are linearly independent and
representing the approximation as,
n
X
f (x) ≈ Pn (x) = ai φi (x)
i=1
Here the basis functions φi (x) are known functions, chosen with care
to form a linearly independent set and ai are unknown constants that
are to be determined in such a way that we can make Pn (x) as good an
approximation to f (x) as possible - i.e., we can define an error as the
difference between the exact function and the approximate representa-
tion,
(x; ai ) = |f (x) − Pn (x)|
and device a scheme to select ai such that the error is minimized.
Example
So far we have outlined certain general concepts, but left open the choice
of a specific basis functions φi (x), the definition of the norm |.| in the
error or the minimization procedure to get ai .
Let the basis functions be
φi (x) = x i−1 i = 1, · · · n
which, incidentally is a poor choice, but one that is easy to understand.

Hence the approximate function will be a polynomial of degree (n − 1)
of the form,
Xn
Pn−1 (x) = ai x i−1
i=1
Next, let us introduce the idea of collocation to evaluate the error at n

selected points in the range of interest x ∈ [a, b]. We choose n points
{xk |k = 1, · · · n} because we have introduced n degrees of freedom
(unknowns) in ai . A naive choice would be to space these collocation
points equally in the interval [a, b] - i.e.,
(b − a)
xk = a + (k − 1) k = 1, · · · , n
(n − 1)
Finally we can require the error at these points to be exactly equal to

zero - i.e.,
(xk ; ai ) = f (xk ) − Pn−1 (xk ) = 0
or
n
X
ai xki−1 = f (xk ) k = 1, · · · , n (5.1)
i=1
which yields n linear equations in n unknowns ai . This can be written

in matrix form
Pa = f
where the elements of matrix P are given by, Pk,i = xki−1 and the vectors
are a = [a1 , · · · , an ] and f = [f (x1 ), · · · , f (xn )]. Thus we have re-
duced the functional approximation problem to one of solving a system
of linear algebraic equations and tools of chapter 3 become useful!
Let us be even more specific now and focus on approximating the
error function f (x) = er f (x) over the interval x ∈ [0.1, 0.5]. Let us
also choose n = 5 - i.e., a quartic polynomial. This will allow us to write
out the final steps of the approximation problem explicitly. The equally
spaced collocation points are,
xk = {0.1, 0.2, 0.3, 0.4, 0.5}
and the error function values at the collocation points are
f = f (xk ) = [0.1125, 0.2227, 0.3286, 0.4284, 0.5205]
Thus, equation (5.1) yields the following system

   
1 x1 x12 x13 x14 1.0 0.10 0.010 0.0010 0.0001
   
 1 x2 x22 x23 x24   1.0 0.20 0.040 0.0080 0.0016 
   
P = 
 1 x3 x32 x33 x34 =
  1.0 0.30 0.090 0.0270 0.0081 
   
 1 x4 x42 x43 x44  1.0 0.40 0.160 0.0640 0.0256 
1 x5 x52 x53 x54 1.0 0.50 0.250 0.1250 0.0625
Solution of the linear system yields the unknown coefficients as
a = {0.0001, 1.1262, 0.0186, −0.4503, 0.1432}
A MATLAB function that shows the implementation of the above pro-

cedure for a specified degree of polynomial n is given in figure 5.2. Recall
that we had made a comment earlier that the basis function φi (x) =
x i−1 i = 1, · · · n is a poor choice. We can understand why this is so,
function a=erf_apprx(n)
% Illustration functional (polynomial) approximation
% fits error function in (0.1, 0.5) to a
% polynomial of degree n
%define interval
a = 0.1; b=0.5;
%pick collocation points

x=a + [0:(n-1)] *(b-a)/(n-1);
%Calculate the error function at collocation points

f=erf(x); %Note that erf is a MATLAB function
%Calculate the matrix

for k=1:n
P(k,:) = x(k).ˆ[0:n-1];
end
%Print the determinant of P

fprintf(1,’Det. of P for deg. %2i is = %12.5e\n’, n,det(P) );
%Determine the unknown coefficients a_i

a=P\f’;
Figure 5.2: MATLAB implementation illustrating steps of functional ap-

proximation
5.2. APPROXIMATE REPRESENTATION OF DATA 104
by using the function shown in figure 5.2 for increasing degree of poly-
nomials. The matrix P becomes poorly scaled and nearly singular with
increasing degree of polynomial as evidence by computing the determi-
nant of P. For example the determinant of P is 1.60000 × 10−2 for n = 3
and it goes down rapidly to 1.21597 × 10−12 for n = 6. Selecting cer-
tain orthogonal polynomials such as Chebyshev polynomials and using
the roots of such polynomials as the collocation points results in well
conditioned matrices and improved accuracy. More on this in section
§5.8.
Note that MATLAB has a function called polyfit(x,y,n) that will
accept a set of pairwise data {xk , yk = f (xk ) | k = 1, · · · , m} and pro-
duce a polynomial fit of degree n (which can be different from m) using a
least-squares minimization. Try using the function polyfit(x,y,n) for
the above example and compare the polynomial coefficients a produced
by the two approaches.
»x=[0.1:0.1:0.5] % Define Collocation Points

»y=erf(x) % Calculate the function at Collocation Points
»a=polyfit(x,y,4)% Fit 4th degree polynomial. Coefficients in a
»polyval(a,x) % Evaluate the polynomial at collocation pts.
»erf(x) % Compare with exact values at the same pts.
5.2 Approximate representation of data
The concepts of polynomial approximation were discussed in section

§5.1.2 in the context of constructing approximate representations of
complicated functions (such as the error function). We will develop and
apply these ideas further in later chapters for solving differential equa-
tions. Let us briefly explore the problem of constructing approximate
functions for representing a discrete set of m pairs of data points
{(xk , fk ) | k = 1, · · · , m}
gathered typically from experiments. As an example, let us consider the

saturation temperature vs. pressure data taken from steam tables and
shown in Table 5.2. Here the functional form that we wish to construct
is to represent pressure as a function of temperature, P (T ) over the tem-
perature range T ∈ [220, 232]. A number of choices present themselves.
T (o F ) P (psia)
220.0000 17.1860
224.0000 18.5560
228.0000 20.0150
232.0000 21.5670
Table 5.2: Saturation temperature vs. pressure from steam tables
• We can choose to fit a cubic polynomial, P3 (T ) that will pass through

each of the four data points over the temperature range T ∈ [220, 232].
This will be considered as a global polynomial as it covers the entire
range of interest in T .
• Alternately we can choose to construct piecewise polynomials of a

lower degree with a limited range of applicability. For example, we
can take the first three data points and fit a quadratic polynomial,
and the last three points and fit a different quadratic polynomial.
• As a third alternative, we can choose to fit a global polynomial of

degree less than three, that will not pass through any of the given
data points, but will produce a function that minimizes the error
over the entire range of T ∈ [220, 232].
The procedures developed in section §5.1.2 are directly applicable

to the first two choices and hence they warrant no further discussion.
Hence we develop the algorithm only the third choice dealing with the
least-squares minimization concept.
5.2.1 Least squares approximation
Suppose there are m independent experimental observations (m = 4 in

the above example) and we wish to fit a global polynomial of degree n
(n < m) we define the error at every observation point as,
k = (Pn−1 (xk ) − fk ) k = 1, · · · , m
The basis functions are still the set, {x i−1 | i = 1, · · · n} and the poly-
nomial is
Xn
Pn−1 (x) = ai x i−1
i=1
Here ai are the unknowns that we wish to determine. Next we construct

an objective function which is the sum of squares of the error at every
observation point - viz.
Pm 2 Pm Pm Pn i−1
k=1 k k=1 (Pn−1 (xk ) − fk )
2
k=1 ( i=1 ai xk − f k )2
J(a) = = =
m m m
The scalar objective function J(a) is a function of n unknowns ai . From
elemetary calculus, the condition for the function J(a) to have a mini-
mum is,
∂J(a)
=0
∂a
This condition provides n linear equations of the form Pa = b that can
be solved to obtain a. The expanded form of the equations are,
 Pm Pm Pm Pm    Pm 
k=1 1 k=1 xk k=1 xk2 ··· k=1 xkn−1 a 1 fk
 Pm Pm 2
Pm 3 Pm n    Pmk=1 
 k=1 xk xk xk ··· xk   a2   k=1 fk xk 
 Pm Pk=1 Pk=1 Pmk=1 n+1   Pm 
 2 m 3 m 4   a3   2 
 k=1 kx k=1 kx x
k=1 k k=1 kx  
 = k=1 fk xk 
 .. .. .. ..  .   .. 
 . . . .   ..   . 
   P 
Pm n−1 Pm n Pm n+1 Pm 2(n−1) an m n−1
k=1 x k k=1 x k=1 x
k ···
k k=1 x k k=1 fk xk
Observe that the equations are not only linear, but the matrix is also
symmetric. Work through the following example using MATLAB to gen-
erate a quadratic, least-squares fit for the data shown in Table 5.2. Make
sure that you understand what is being done at each stage of the cal-
cualtion. This example illustrates a cubic fit that passes through each of
the four data points, followed by use of the cubic fit to interpolate data
at intermediate temperatures of T = [222, 226, 230]. In the last part the
least squares solution is obtained using the procedure developed in this
section. Finally MATLAB’s polyfit is used to generate the same least
squares solution!
»x=[220,224,228,232] % Define temperatures

»f=[17.186,18.556,20.015,21.567] % Define pressures
»a3=polyfit(x,f,3) % Fit a cubic. Coefficients in a3
»polyval(a3,x) % Check cubic passes through pts.
»xi=[222,226,230] % Define interpolation points
»polyval(a3,xi) % Evaluate at interpolation pts.
»%get ready for least square solution!
»x2=x.ˆ2 % Evaluate x 2
5.3. DIFFERENCE OPERATORS 107
»P=[4,sum(x),sum(x2); ... % Define matrix P over next 3 lines

»sum(x), sum(x2), sum(x3); ...
»sum(x2), sum(x3), sum(x4) ]
»b=[sum(f), f*x’, f*x2’] % Define right hand side
»a = P\b0 % ans: (82.0202,-0.9203,0.0028)
»c=polyfit(x,f,2) % Let MATLAB do it! compare c & a
»norm(f-polyval(a3,x)) % error in cubic fit 3.3516 × 10−14
»norm(f-polyval(c,x)) % error in least squares fit 8.9443 × 10−4
5.3 Difference operators
In the previous sections we developed polynomial approximation schemes

in such a way that they required a solution of a system of linear alge-
braic equation. For uniformly spaced data, introduction of difference
operators and difference tables, allows us to solve the same polynomial
approximation problem in a more elegant manner without the need for
solving s system of algebraic equations. This difference operator ap-
proach also lends itself naturally to recursive construction of higher
degree polynomials with very little additional computation as well as
extension to numerical differentiation and integration of discrete set of
data.
Consider the set of data {(xi , fi ) | i = 1, · · · , m} where the indepen-
dent variable, x is varied uniformly generating an equally spaced data -
i.e.,
xi+1 = xi + h, i = 1, · · · m or xi = x1 + (i − 1)h
The forward difference operator, as introduced already in section §2.8,

is defined by,
Forward difference operator
∆fi = fi+1 − fi (5.2)
In a similar manner we can define a backward difference, central differ-

ence and shift operators as shown below.
Backward difference operator
∇fi = fi − fi−1 (5.3)

Central difference operator
δfi = fi+1/2 − fi−1/2 (5.4)
Shift operator
Efi = fi+1 (5.5)

We can also add the differential operator to the above list.
Differential operator
df (x)
Df (x) = = f 0 (x) (5.6)
dx
The difference operators are nothing but rules of calculations, just
like a differential operator defines a rule for differentiation. Clearly these
rules can be applied repeatedly to obtaind higher order differences. For
example a second order forward difference with respect to reference
point i is,
∆2 fi = ∆(∆fi ) = ∆(fi+1 − fi ) = fi+2 − 2fi+1 + fi
5.3.1 Operator algebra

Having introduced some new definitions of operators, we can discover
some interesting relationships between various operators such as the
following.
∆fi = fi+1 − fi and Efi = fi+1
Combining these two we can write,
∆fi = Efi − fi = (E − 1)fi
Since the operand fi is the same on both sides of the equation, the op-
erators (which define certain rules and hence have certain effects on the
operand fi ) have an equivalent effect given by,
∆ = (E − 1) or E = (1 + ∆) (5.7)
Equation (5.7) can then be applied on any other operand like fi+k . All
of the operators satisfy the distributive, commutative and associative
rules of algebra. Also, repeated application of the operation can be rep-
resented by,
E α = (1 + ∆)α
Note that E α f (x) simply implies that the function f is evaluated after
shifting the independent variable by α - i.e.,
E α f (x) = f (x + αh)
Hence α can be an integer or any real number. Similarly, we have
∇fi = fi − fi−1 and Efi−1 = fi and fi−1 = E −1 fi
where we have introduce the inverse of the shift operator E to shift back-
wards. Combining these we can write,
∇fi = fi − E −1 fi = (1 − E −1 )fi
Once again recognizing that the operand fi is the same on both sides of
the equation, the operators are related by,
∇ = (1 − E −1 ) or E −1 = (1 − ∇) or E = (1 − ∇)−1 (5.8)
Yet another relation between the shift operator E and the differential
operator D can be developed by considering the Taylor series expansion
of f (x + h),
h2 00
f (x + h) = f (x) + hf 0 (x) + f (x) + · · ·
2!
which can be written in operator notation as,
" #
h2 D 2
Ef (x) = 1 + hD + + · · · f (x)
2!
The term in square brackets is the exponential function and hence
E = ehD (5.9)
While such a game of discovering relationships between various oper-

ators can be played indefinitely, let us turn to developing some useful
algorithms from these.
5.3.2 Newton forward difference approximation

Our objective is to construct a polynomial representation for the discrete
set of data {(xi , fi ) | i = 1, · · · , m} using an alternate approach from
that of section §5.1.2. Assuming that there is a function f (x) represent- Is such an
assumption always
valid?
ing the given data, we can express such a function as,
f (x) = Pn (x) + R(x)
where Pn (x) is the polynomial approximation to f (x) and R(x) is the

residual error. Given a set of m data points we know at least one way
(section §5.1.2) to can construct a polynomial of degree (m − 1). Now
let us use the power of operator algebra to develop an alternate way to
construct such a polynomial and in the process, also learn something
about the residual function R(x). Applying equation (5.7) repeatedly α
time on f (x) we get,
E α f (x) = (1 + ∆)α f (x)
Now for integer values of α the right hand side is the binomial expansion
while for any real number, it yields an infinite series. Using such an
expansion the above equation can be written as,

α(α − 1) 2 α(α − 1)(α − 2) 3
f (x + αh) = 1 + α∆ + ∆ + ∆ + ···
2! 3!

α(α − 1)(α − 2) · · · (α − n + 1) n
∆ + · · · f (x)(5.10)
n!
Up to this point in our development we have merely used tricks of
operator algebra. We will now make the direct connection to the given,
discrete set of data {(xi , fi ) | i = 1, · · · , m}. Taking x1 as the reference
point, the transformation
x = x1 + αh
makes α the new independent variable and for integer values of α =

0, 1, · · · (m − 1) we retrieve the equally spaced data set {x1 , x2 , · · · xm }
and for non-integer (real) values of α we can reach the other values of
x ∈ (x1 , xm ). Splitting equation (5.10) into two parts,

α(α − 1) 2 α(α − 1)(α − 2) 3
f (x1 + αh) = 1 + α∆ + ∆ + ∆ + ···
2! 3!

α(α − 1)(α − 2) · · · (α − m + 2) m−1
∆ f (x1 ) + R(x)
(m − 1)!
we can recognize the terms in the square brackets as a polynomial of de-

gree (m − 1) in the transformed variable α. We still need to determine
the numbers {∆f (x1 ), ∆2 f (x1 ), · · · , ∆(m−1) f (x1 )}. These can be com-
puted and organized as a forward difference table shown in figure 5.3.
Since forward differences are needed for constructing the polynomial, it
x1 f1
∆f1
x2 f2 ∆2f1
∆f2 ∆3f1
x3 f3 ∆2f2 ∆4f1
∆f3 ∆3f2
x4 f4 ∆2f3
∆f4 ∆m-1f1
x5 f5
∆4fm-4
∆3fm-3
xm-1 fm-1 ∆2fm-2
∆fm-1
xm fm
Figure 5.3: Structure of Newton forward difference table for m equally

spaced data
is called the Newton forward difference polynomial and it is given by,

α(α − 1) 2 α(α − 1)(α − 2) 3
Pm−1 (x1 + αh) = 1 + α∆ + ∆ + ∆ + ···
2! 3!

α(α − 1)(α − 2) · · · (α − m + 2) m−1
∆ f (x1 ) + O(hm ) (5.11)
(m − 1)!
The polynomial in equation (5.11) will pass through the given data set
{(xi , fi ) | i = 1, · · · , m} - i.e., for integer values of α = 0, 1, · · · (m − 1)
it will return values of {f1 , f2 , · · · fm }. This implies that the residual
function R(x) will have roots at the data points {xi | i = 1, · · · , m}. For
a polynomial of degree (m − 1), shown in equation (5.11), the residual at
other values of x is typically represented as R(x) ≈ O(hm ) to suggest
that the leading term in the truncated part of the series is of order m.
Example
A set if five (m = 5) equally spaced data points and the forward differ-
ence table for the data are shown in figure 5.4. For this example, clearly
h = 1 and x = x1 + α. We can take the reference point as x1 = 2 and
construct the following linear, quadratic and cubic polynomials, respec-
tively. Note that
P4 (2 + α) =
P3 (2 + α) for this
case! Why?
x1 = 2 f1 = 8
∆f 1 = 19
x2 = 3 f 2 = 27 ∆2 f 1 = 18
∆f 2 = 37 ∆3 f 1 = 6
x3 = 4 f 3 = 64 ∆2 f 2 = 24 ∆4 f 1 = 0
∆f 3 = 61 ∆3 f 2 = 6
x4 = 5 f 4 = 125 ∆2 f 3 = 30
∆f 4 = 91
x5 = 6 f 5 = 216
Figure 5.4: Example of a Newton forward difference table
P1 (2 + α) = [1 + α∆] f (x1 ) = (8) + α(19) + O(h2 )

α(α − 1)
P2 (2 + α) = (8) + α(19) + (18) + O(h3 )
2!
α(α − 1) α(α − 1)(α − 2)
P3 (2 + α) = (8) + α(19) + (18) + (6) + O(h4 )
2! 3!
You can verify easily that P1 (2 + α) passes through {x1 , x2 }, P2 (2 + α)
passes through {x1 , x2 , x3 } and P3 (2+α) passes through {x1 , x2 , x3 , x4 }.
For finding the interpolated value of f (x = 3.5) for example, first deter-
mine the values of α at x = 3.5 from the equation x = x1 + αh. It is
α = (3.5 − 2)/1 = 1.5. Using this value in the cubic polynomial,
1.5(0.5) 1.5(0.5)(−0.5)
P3 (2 + 1.5) = (8) + 1.5(19) + (18) + (6) = 42.875
2! 3!
As another example, by taking x3 = 4 as the reference point we can
construct the following quadratic polynomial
α(α − 1)
P2 (4 + α) = (64) + α(61) + (30)
2!
which will pass through the data set {x3 , x4 , x5 }. This illustration should
show that once the difference talbe is constructed, a variety of polyno-
mials of varying degrees can be constructed quite easily.
5.3.3 Newton backward difference approximation

An equivalent class of polynomials using the backward difference oper-
ator based on equation (5.8) can be developed. Applying equation (5.8)
repeatedly α times on f (x) we get,
E α f (x) = (1 − ∇)−α f (x)
which can be expanded as before to yield,

α(α + 1) 2 α(α + 1)(α + 2) 3
f (x + αh) = 1 + α∇ + ∇ + ∇ + ···
2! 3!

α(α + 1)(α + 2) · · · (α + n − 1) n
∇ + · · · f (x)(5.12)
n!
As with the Newton forward formula, the above equation (5.12) termi-
nates at a finite number of terms for integer values of α and for non-
integer values, it will always be an infinite series which must be tran-
cated, thus sustaining a trunctaion error.
In making the precise connection to a given discrete data set {(xi , fi ) | i =
0, −1 · · · , −n}, typically the largest value of x (say, x0 ) is taken as the
reference point. The transformation
x = x0 + αh
makes α the new independent variable and for negative integer values
of α = −1, · · · − n we retrieve the equally spaced data set {x−1 , · · · x−n }
and for non-integer (real) values of α we can reach the other values of
x ∈ (x−n , x0 ). Splitting equation (5.12) into two parts,

α(α + 1) 2 α(α + 1)(α + 2) 3
f (x0 + αh) = 1 + α∇ + ∇ + ∇ + ···
2! 3!

α(α + 1)(α + 2) · · · (α + n − 1) n
∇ + · · · f (x0 ) + R(x)
n!
we can recognize the terms in the square brackets as a polynomial of
degree n in the transformed variable α. We still need to determine
the numbers {∇f (x0 ), ∇2 f (x0 ), · · · , ∇n f (x0 )}. These can be computed
and organized as a backward difference table shown in figure 5.5. Since
backward differences are needed for constructing the polynomial, it is
called the Newton backward difference polynomial and it is given by,

α(α + 1) 2 α(α + 1)(α + 2) 3
Pn (x0 + αh) = 1 + α∇ + ∇ + ∇ + ···
2! 3!

α(α + 1)(α + 2) · · · (α + n − 1) n
∇ f (x0 ) + O(hn+1 ) (5.13)
n!
The polynomial in equation (5.13) will pass through the given data set
{(xi , fi ) | i = 0, · · · , −n} - i.e., for integer values of α = 0, −1, · · · − n it
will return values of {f0 , f−1 , · · · f−n }. At other values of x the residual
will be of order O(hn+1 ).
x− 4 f − 4
∇f − 3
x− 3 f − 3 ∇2 f − 2
∇f − 2 ∇3 f − 1
x− 2 f − 2 ∇2 f − 1 ∇4 f 0
∇f − 1 ∇3 f 0
x− 1 f − 1 ∇2 f 0
∇f 0
x0 f 0
Figure 5.5: Structure of Newton backward difference table for 5 equally

spaced data
x− 4 = 2 f −4 = 8 ? ? ? ? ?
? ? ∇f − 3 = 19 ? ? ? ?
x− 3 = 3 f − 3 = 27 ? ∇2 f − 2 = 18 ? ? ?
x
? NBF
ar?
ound -2 ∇f − 2 = 37 ? ? ∇ f −1 = 6
3 ?
x− 2 = 4 f − 2 = 64 ? ∇ f − 1 = 24
2 ? ? ∇ f0 = 0
4
? NFF aro
? und x-2 ∇f − 1 = 61 ? ? ∇ f0 = 6
3 ?
x− 1 = 5 f − 1 = 125 ? ∇ f 0 = 30
2 ? ? ?
? ? ∇f 0 = 91 ? ? ? ?
x0 = 6 f 0 = 216 ound ?x 0 ? ? ? ?
ar
NBF
Figure 5.6: Example of a Newton backward difference table
Example
A set if five (n = 4) equally spaced data points and the backward differ-
ence table for the data are shown in figure 5.6. This is the same example
as used in the previous section! It is clear that h = 1 and x = x0 + α.
In the previous case we constructed a linear, quadratic and cubic poly-
nomials, with x3 = 4 as the reference point. In the present case let us
use the same reference point, but it is labelled as x−2 = 4. A quadratic
backward difference polynomial in α is,
α(α + 1)
P2 (4 + α) = (64) + α(37) + (18) + O(h3 )
2!
which passes through the points (x−2 , f−2 ), (x−3 , f−3 ) and (x−4 , f−4 )
for α = 0, −1, −2, respectively. Recall that the forward difference poly-
nomial around the same point was, Calculate the
interpolated value
α(α − 1) of f (4.5) from
P2 (4 + α) = (64) + α(61) + (30)
2! these two
polynomials
which passes through the three forward point for α = 0, 1, 2. Although
they are based on the same reference point, these are two different poly-
nomials passing through a different set of data points.
As a final example, let us construct a quadratic backward difference
polynomial around x0 = 6. It is, Is this polynomial
different from the
α(α + 1) NFF, P2 (4 + α)
P2 (6 + α) = (216) + α(91) + (30)
2! constructed above?
5.4. INVERSE INTERPOLATION 116
150
125
f(x=?)=100
100
f(x) 75
P2
(4+
(spurious root)
(desired root)
x=0.286962
x=4.64637
α)
50
25
f(x) = x3
0
-2 0 2 4 6
x
Figure 5.7: Example of a inverse interpolation
5.4 Inverse interpolation
One of the objectives in constructing an interpolating polynomial is to

be able to evaluate the function f (x) at values of x other than the ones
in the discrete set of given data points (xi , fi ). The objective of inverse
interpolation is to determine the independent variable x for a given value
of f using a given discrete data set (xi , fi ). If xi are equally spaced, we
can combined two of the tools (polynomial curve fitting and root finding)
to meet this objective, although this must be done with caution.
We illustrate this with the example data shown in figure 5.4. Suppose
we wish to find the value of x where f = 100. Using the three data points
in the neighbourhood of f = 100 in figure 5.4 viz. (x3 , x4 , x5 ), and using
a quadratic polynomial fit, we have,
α(α − 1)
P2 (4 + α) = (64) + α(61) + (30)
2!
A graph of this polynomial approximation P2 (4 + α) and the actual func-
tion f (x) = x 3 used to generate the data given in figure 5.4 are shown
in figure 5.7. It is clear that the polynomial approximation is quite good
in the range of x ∈ (4, 6), but becomes a poor approximation for lower
values of x. Note, in particular, that if we solve the inverse interpolation
problem by setting
α(α − 1)
P2 (4 + α) − 100 = 0 or (64) + α(61) + (30) − 100 = 0
2!
5.4. INVERSE INTERPOLATION 117
we will find two roots. One of them at α = 0.64637 or x = 4.64637 is

the desired root while the other at α = −3.71304 or x = 0.286962 is a
spurious one. This problem can become compounded as we use higher
degree polynomial in an effort to improve accuracy.
In order to achieve high accuracy, but stay close to the desired root,
we can generate an initial guess from a linear interpolation, followed by
constructing a fixed point iteration scheme on the polynomial approxi-
mation of the desired accuracy. Convergence is generally fast as shown
in Dahlquist and Björck (1974). Suppose we wish to find x corresponding
to f (x) = d, the desired function value. We first construct a polynomial
of degree (m − 1) to represent the tabular data.

α(α − 1) 2 α(α − 1)(α − 2) 3
Pm−1 (x1 + αh) = 1 + α∆ + ∆ + ∆ + ···
2! 3!

α(α − 1)(α − 2) · · · (α − m + 2) m−1
∆ f (x1 )
(m − 1)!
Then we let f (x1 + αh) = d and rearrange the polynomial in the form
αi+1 = g(αi ) i = 0, 1, 2 · · ·
where g(α) is obtained by rearranging the polynomial,

1 α(α − 1) 2 α(α − 1)(α − 2) 3
g(α) = d − f1 − ∆ f1 − ∆ f1 + · · ·
∆f1 2! 3!
and the initial guess obtained by truncating the polynomial after the
linear term,
d − f1
α0 =
∆f1
Example
Continuing with the task of finding x where f (x) = 100 for the data
shown in figure 5.4, the fixed point iterate is,
αi+1 = [100 − 64 − 15αi (αi − 1)] /61
and the initial guess is
d − f1 100 − 64
α0 = = = 0.5902
∆f1 61
The first ten iterates, produced from the m-file given below,
5.5. LAGRANGE POLYNOMIALS 118
i αi
1 .59016393
2 .64964028
3 .64613306
4 .64638815
5 .64636980
6 .64637112
7 .64637102
8 .64637103
9 .64637103
10 .64637103
Table 5.3: Inverse interpolation
function a=g(a)
for i=1:10
fprintf(1,’%2i %12.7e\n’,i,a);
a=(100 - 64 - 15*a*(a-1))/61;
end
are shown in Table 5.3.
5.5 Lagrange polynomials
So far we have examined ways to construct polynomial approximations

using equally spaced data in x. For a data set {xi , fi |i = 0, · · · n}, that
contains unequally spaced data in the independent variable x, we can
construct Lagrange interpolation formula as follows.
n
X
Pn (x) = fi δi (x) (5.14)
i=0
where
n
Y x − xj
δi (x) =
xi − x j
j=0,j6=i
Note that (
0 j 6= i
δi (xj ) =
1 j=i
and each δi (x) is a polynomial of degree n. It is also clear from equation

(5.14) that Pn (xj ) = fj - i.e., the polynomial passes through the data
points (xj , fj ).
An alternate way to construct the Lagrange polynomial is based on
introducing the divided difference and constructing a divided difference
table. The polynomial itself is written in the form
n
X i
Y
Pn (x) = ai (x − xj−1 ) (5.15)
i=0 j=0
= a0 + a1 (x − x0 ) + a2 (x − x0 )(x − x1 ) + · · ·
+an (x − x0 ) · · · (x − xn−1 )
The advantage of writing it the form shown in equation (5.15) is that the
unknown coefficients ai can be constructed recursively or found directly
from the divided difference table. The first divided difference is defined
by the equation,
f1 − f 0
f [x0 , x1 ] =
x1 − x 0
Similarly the second divided difference is defined as,
f [x1 , x2 ] − f [x0 , x1 ]
f [x0 , x1 , x2 ] =
x2 − x 0
With these definitions, we return to the task of finding the coefficients
ai in equation (5.15) For example, the first coefficient a0 is,
Pn (x0 ) = a0 = f [x0 ] = f0
The second coefficient, a1 , is obtained from,
Pn (x1 ) = a0 + a1 (x1 − x0 ) = f1
f1 − f0
a1 = = f [x0 , x1 ]
x1 − x 0
The third coefficient is obtained from,
Pn (x2 ) = a0 + a1 (x2 − x0 ) + a2 (x2 − x0 )(x2 − x1 ) = f2
The only unknown here is a2 , which after some rearrangement becomes,
f [x1 , x2 ] − f [x0 , x1 ]
a2 = = f [x0 , x1 , x2 ]
x2 − x 0
In general the n-th coefficient is the n-th divided difference.
an = f [x0 , x1 , · · · , xn ]
x0 = 1.0 f 0 = 1.000
f[ x0 , x1 ] = 3.6400
x1 = 1.2 f 1 = 1.728 f[ x0 , x1 , x2 ] = 3.700
f[ x1 , x2 ] = 5.4900 f[ x0 , x1 , x2 , x3 ] = 1.000
x2 = 1.5 f 2 = 3.375 f[ x1 , x2 , x3 ] = 4.300
f[ x2 , x3 ] = 7.2100
x3 = 1.6 f 3 = 4.096
Figure 5.8: Structure of divided difference table for 4 unequally spaced

data
Example
Consider the example data and the divided difference table shown in fig-
ure 5.8. If we wish to construct a quadratic polynomial passing through
(x0 , f0 ), (x1 , f1 ), (x2 , f2 ) for example using equation (5.14), it will be
(x − x1 )(x − x2 ) (x − x0 )(x − x2 ) (x − x0 )(x − x1 )

P2 (x) = f0 + f1 + f2
(x0 − x1)(x0 − x2 ) (x1 − x0 )(x1 − x2 ) (x2 − x0 )(x2 − x1 )
(x − 1.2)(x − 1.5) (x − 1)(x − 1.5) (x − 1)(x − 1.2)
= 1.00 + 1.728 + 3.375
(1 − 1.2)(1 − 1.5) (1.2 − 1)(1.2 − 1.5) (1.5 − 1)(1.5 − 1.2)
The same polynomial using equation (5.15) and the difference table shown
in figure 5.8 will be written as,
P2 (x) = f0 + f [x0 , x1 ](x − x0 ) + f [x0 , x1 , x2 ](x − x0 )(x − x1 )

= 1.000 + 3.64(x − 1) + 3.70(x − 1)(x − 1.2)
Observe that in order to construct a cubic polynomial by adding the addi-

tional data point (x3 , f3 ) Lagrange polynomial based on equation (5.14)
requires a complete reconstruction of the equation, while that based on
equation (5.15) is simply,
P2 (x) = f0 + f [x0 , x1 ](x − x0 ) + f [x0 , x1 , x2 ](x − x0 )(x − x1 ) +

f [x0 , x1 , x2 , x3 ](x − x0 )(x − x1 )(x − x2 )
= 1.000 + 3.64(x − 1) + 3.70(x − 1)(x − 1.2) +

1(x − 1)(x − 1.2)(x − 1.5)
A MATLAB function that implements that Lagrange interpolation for-

mula shown in equation (5.14) is given in figure 5.9. This function ac-
function f=LagrangeP(xt,ft,x)
% (xt,ft) are the table of unequally spaced values
% x is where interpolated values are required
% f the interpolated values are returned
m=length(x);
nx=length(xt);
ny=length(ft);
if (nx ˜= ny),
error(’ (xt,ft) do not have the same # values’)
end
for k=1:m
sum = 0;
for i=1:nx
delt(i)=1;
for j=1:nx
if (j ˜= i),
delt(i) = delt(i)*(x(k)-xt(j))/(xt(i)-xt(j));
end
end
sum = sum + ft(i) * delt(i) ;
end
f(k)=sum;
end
Figure 5.9: MATLAB implementation of Lagrange interpolation polyno-

mial
5.6. NUMERICAL DIFFERENTIATION 122
cepts a table of values (xt, f t), constructs the highest degree Lagrange
polynomial that is possible and finally evaluates and returns the inter-
polated values of the function y at specified values of x.
»xt=[1 1.2 1.5 1.6] % Define xt, unequally spaced

»ft=[1 1.728 3.375 4.096] % Define ft, the function values
»x=[1.0:0.1:1.6] % x locations for interpolation
»f=LagrangeP(xt,ft,x) % interpolated f values.
5.6 Numerical differentiation
Having obtained approximate functional representations as outlined in

sections §5.3.2 or §5.5, we can proceed to construct algorithms for ap-
proximate representations of derivatives.
5.6.1 Approximations for first order derivatives

Consider the Newton forward formula given in equation (5.11)

α(α − 1) 2
f (x) ≈ Pm−1 (x1 + αh) = 1 + α∆ + ∆ +
2!

α(α − 1)(α − 2) 3 α(α − 1) · · · (α − m + 2) m−1
∆ + ··· + ∆ f (x1 ) + O(hm )
3! (m − 1)!
that passes through the given data set {(xi , fi ) | i = 1, · · · , m}. Note
that the independent variable x has been transformed into α using x =
x1 + αh, hence dx/dα = h. Now, the first derivative is obtained as,

0 df dPm−1 dPm−1 dα 1 α + (α − 1) 2
f (x) = ≈ = = ∆+ ∆ +
dx dx dα dx h 2

{α(α − 1) + (α − 1)(α − 2) + α(α − 2)} 3
∆ + · · · f (x1 )(5.16)
6
Equation (5.16) forms the basis of deriving a class of approximations for
first derivatives from a tabular set of data. Note that the equation (5.16)
is still a function in α and hence it can be used to evaluate the derivative
at any value of x = x1 + αh. Also, the series can be truncated after
any number of terms. Thus, a whole class of successively more accurate
representations for the first derivative can be constructed from equa-
tion (5.16) by truncating the series at higher order terms. For example
evaluating the derivative at the reference point x1 , ( i.e., α = 0) equation

(5.16) reduces to,

0 1 1 2 1 3 1 4 1 m−1
f (x1 ) = ∆ − ∆ + ∆ − ∆ ··· ± ∆ f (x1 )+O(hm−1 )
h 2 3 4 m−1
This equation can also be obtained directly using equation (5.9) as,
E = ehD or hD = ln E = ln (1 + ∆)
Expanding the logarithmic term we obtain,
∆2 ∆3 ∆4
hD = ∆ − + − + ···
2 3 4
Operating both sides with f (x1 ) ( i.e., using x1 as the reference point),
we get,
" #
0 1 ∆2 ∆3 ∆4
Df (x1 ) = f (x1 ) = ∆− + − + · · · f (x1 ) (5.17)
h 2 3 4
Now, truncating the series after the first term (m = 2),
1
f 0 (x1 ) = [∆f (x1 )] + O(h)
h
1
= [f2 − f1 ] + O(h)
h
which is a 2-point, first order accurate, forward difference approximation
for first derivative at x1 . Truncating the series after the first two terms
(m = 3),

1 1
f 0 (x1 ) = ∆f (x1 ) − ∆2 f (x1 ) + O(h2 )
h 2

1 1
= (f2 − f1 ) − (f1 − 2f2 + f3 ) + O(h2 )
h 2
1
= [−3f1 + 4f2 − f3 ] + O(h2 )
2h
which is the 3-point, second order accurate, forward difference approx-
imation for the first derivative at x1 . Clearly both are approximate rep-
resentations of the first derivative at x1 , but the second one is more
accurate since the truncation error is of the order h2 .
Note that while, equation (5.17) is evaluated at the reference point
on both sides of the equation, the earlier equation (5.16) is a polynomial
that is constructed around the reference point x1 , but can be evaluated at
Derivative Difference approximation truncation

at xi error
f 0 (xi ) (fi+1 − fi )/h O(h)
f 0 (xi ) (fi − fi−1 )/h O(h)
f 0 (xi ) (−3fi + 4fi+1 − fi+2 )/2h O(h2 )
f 0 (xi ) (+3fi − 4fi−1 + fi−2 )/2h O(h2 )
f 0 (xi ) (fi+1 − fi−1 )/2h O(h2 )
f 0 (xi ) (fi−2 − 8fi−1 + 8fi+1 − fi+2 )/12h O(h4 )
f 00 (xi ) (fi+1 − 2fi + fi−1 )/h2 O(h2 )
f 00 (xi ) (fi+2 − 2fi+1 + fi )/h2 O(h)
f 00 (xi ) (−fi−3 + 4fi−2 − 5fi−1 + 2fi )/h2 O(h2 )
f 00 (xi ) (−fi+3 + 4fi+2 − 5fi+1 + 2fi )/h2 O(h2 )
Table 5.4: Summary of difference approximations for derivatives
any other point by choosing appropriate α values. For example, consider

the first derivative at x = x2 or α = 1. Two term truncation of equation
(5.16) yields,
0 1 1 2
f (x2 ) = ∆ + ∆ f (x2 ) + O(h2 )
h 2
or
1
f 0 (x2 ) = [f3 − f1 ] + O(h2 )
2h
which is a 3-point, second order accurate, central difference approxima-
tion for the first derivative at x2 .
Going through a similar exercise as above with the Newton backward
difference formula (5.13), truncating the series at various levels and us-
ing diffrent reference points, one can easily develop a whole class of
approximations for first order derivatives. Some of the useful ones are
summarized in Table 5.4.
5.6.2 Approximations for second order derivatives

The second derivative of the polynomial approximation is obtained by
taking the derivative of equation (5.16) one more time - viz.

00 d dPm−1 dα dα 1 h
f (x) ≈ = 2 ∆2 +
dα dα dx dx h
{α + (α − 1) + (α − 1) + (α − 2) + α + (α − 2)} 3
∆
6
+ · · ·] f (x1 ) + O(hm−2 ) (5.18)
Evaluating at α = 0 ( i.e., x = x1 ), we obtain,

1 11 4 5 5 137 6
f 00 (x1 ) = ∆ 2
− ∆ 3
+ ∆ − − ∆ + ∆ · · · f (x1 ) (5.19)
h2 12 6 180
This equation can also be obtained directly using equation (5.9) as,
(hD)2 = (ln E)2 = (ln (1 + ∆))2
Expanding the logarithmic term we obtain,

" #2
2 ∆2 ∆3 ∆4 ∆5
(hD) = ∆− + − + ···
2 3 4 5

2 3 11 4 5 5 137 6 7 7 363 8
= ∆ −∆ + ∆ − ∆ + ∆ − ∆ + ∆ ···
12 6 180 10 560
Operating both sides on f (x1 ) ( i.e., using x1 as the reference point), we

get,

2 00 1 2 3 11 4
D f (x1 ) = f (x1 ) = 2 ∆ − ∆ + ∆ − · · · f (x1 )
h 12
Truncating after one term,
1 h 2 i 1
f 00 (x1 ) = 2
∆ f (x 1 ) = 2 (f1 − 2f2 + f3 ) + O(h)
h h
Truncating after two terms,
1 h 2 i
f 00 (x1 ) = 2
∆ f (x1 ) − ∆3 f (x1 )
h
1
= (2f1 − 5f2 + 4f3 − f4 ) + O(h2 )
h2
Evaluating equation(5.18) at x2 (or α = 1), we get,

" #
00 1 2 δ3
f (x2 ) = 2 ∆ −0· f (x1 ) + O(h2 )
h 6
Note that the third order term turns out to be zero and hence this for-
mula turns out to be more accurate. This is a 3-point, second order ac-
curate central difference approximation for the second derivative given
as,
1 h i 1
f 00 (x2 ) = 2 ∆2 f (x1 ) = 2 (f1 − 2f2 + f3 ) + O(h2 )
h h
5.7. NUMERICAL INTEGRATION 126
5.6.3 Taylor series approach
One can derive finite difference approximations from Taylor series ex-
pansion also. Consider the following expansions around xi .
h2 00 h3 000
f (xi + h) = f (xi ) + hf 0 (xi ) + f (xi ) + f (xi ) + · · ·
2 3!
h2 00 h3 000
f (xi − h) = f (xi ) − hf 0 (xi ) + f (xi ) − f (xi ) + · · ·
2 3!
Subtracting the second from the first equation, and extracting f 0 (xi ), we
get,
f (xi + h) − f (xi − h) h2 00
f 0 (xi ) = − f (xi ) + · · ·
2h 6
or,
fi+1 − fi−1
f 0 (xi ) = + O(h2 )
2h
which is a central difference formula for the first derivative that we de-
rived in the last section §5.6. Adding the two Taylor series expansions
above, we get,
h4 0000
fi+1 + fi−1 = 2fi + h2 f 00 (xi ) + f (xi ) + · · ·
12
or,
1
f 00 (xi ) = [fi+1 + fi−1 − 2fi ] + O(h2 )
h2
which is a central difference formula for the second derivative that we
derived in the last section §5.6.
5.7 Numerical integration
The ability to evaluate definite integrals numerically is useful either (i)

when the function is complicated and hence is not easy to integrate ana-
lytically or (ii) the data is given in equally spaced, tabular from. In either
case the starting point is to use the functional approximation methods
seen in earlier sections followed by the integration of the approximate
function. By doing this formally, with the Newton forward polynomials,
we can develop a class of integration formulas. Consider the integral,
Zb
f (x)dx (5.20)
a
Since the function can be represented by a polynomial of degree n as,
f (x) = Pn (x0 + αh) + O(hn+1 )
with an error of order hn+1 , we can use this approximation to carry out
the integration. We first divide the interval x ∈ [a, b] into n subdivisions
as shown in the sketch below; hence there will be
f(x)
f1 f
2 fi
x=b
x=-a
1 2 i n-1
i=0 i=n
(n + 1) data points labelled as {x0 , x1 , · · · xn } and we have
h = (b − a)/n, x = x0 + αh dx = hdα
As an illustration let us take a first degree polynomial between x0 and

x1 . We have
Z x1 Z1 Z1
f (x)dx ≈ P1 (x0 + αh)hdα + O(h2 )hdα
x0 0 0
or,
Z x1 Z1
f (x)dx ≈ [1 + α∆] f0 h dα + O(h3 )
x0 0
which upon, completing the integration becomes,

Z x1
h
f (x)dx ≈ [f0 + f1 ] + O(h3 ) (5.21)
x0 2 | {z }
local error
This formula is the well known trapezoidal rule for numerical integra-
tion. The geometrical interpretation is that it represents the shaded area
under the curve. Note that while numerical differentiation, as developed
in equation (5.16), lowers the order of the truncation error by one due to
the term dα/dx = 1/h numerical integration increases the order of the
truncation error by one due to the term dx = hdα. In the above formula
the truncation error is of order O(h3 ). It is called the local truncation
error since it is the error in integrating over one interval x ∈ (x0 , x1 ).
To obtain the complete integral over the interval x ∈ [a, b] we apply
equation (5.21) repeatedly over each of the subdivisions as,
Zb n Z xi
X Xn X n
h
f (x)dx = f (x)dx = [fi−1 + fi ] + O(h3 )
a
i=1
xi−1
i=1
2 i=1
| {z }
global error
Recalling that n = (b − a)/h the global or accumulated error becomes

of order O(h2 ). Thus the trapezoidal rule has a local truncation error
of order O(h3 ) and a global truncation error of order O(h2 ) and the
equation is,
Zb n
hX
f (x)dx = [fi−1 + fi ] + O(h2 ) (5.22)
a 2 i=1 | {z }
global error
By taking an quadratic functional approximation and integrating over

the range of x ∈ [x0 , x2 ] we obtain the Simpson’s rule.
Z x2 Z2 Z2
f (x) dx ≈ P2 (x0 + αh) h dα + O(h3 )hdα
x0 0 0
or,
Z x2 Z2
α(α − 1) 2
f (x) dx ≈ 1 + α∆ + ∆ f0 h dα + O(h4 )
x0 0 2
which upon, completing the integration becomes,

Z x2
h
f (x) dx ≈ [f0 + 4f1 + f2 ] + O(h4 ) (5.23)
x0 3 | {z }
local error
Note that the next neglected term in the polynomial P2 (x0 + αh) that
corresponds to order O(h3 ) term viz.
Z2
α(α − 1)(α − 2)) 3
∆ f0 h dα
0 3!
turns out to be exactly zero, thus making the local truncation error in
the Simpson’s rule to be actually of order O(h5 ) Repeated application of
the Simpson’s rule results in,
Zb
h
f (x)dx = [f0 + 4f1 + 2f2 + 4f3 + 2f4 + · · · + fn ] + O(h4 )
a 3 | {z }
global error
(5.24)
Note that in applying Simpson’s rule repeatedly over the interval x ∈
[a, b], we must have an even number of intervals (n even) or equivalently
an odd number of points.
5.7.1 Romberg Extrapolation

An idea similar to that used in section §2.8 to accelerate convergence
is the notion of extrapolation to improve accuracy of numerical integra-
tion. The basic idea is to estimate the truncation error by evaluating the
integral on two different grid sizes, h1 and h2 . Let us apply this idea on
the trapezoidal rule which has a global truncation error of O(h2 ). Let
the exact integral be represented as,
I = I(h1 ) + E(h1 )
where I(h1 ) is the approximate estimate of the integral using grid size
of h1 and E(h1 ) is the error. Similarly we have,
I = I(h2 ) + E(h2 )
But, for trapezoidal rule we have E(hi ) ∝ h2i . Hence
E(h1 ) h2
= 12
E(h2 ) h2
We can combine these equations as,
I = I(h1 ) + E(h1 ) = I(h2 ) + E(h2 )
or,
h21
I(h1 ) + E(h2 ) = I(h2 ) + E(h2 )
h22
which can be solved to obtain E(h2 ) as,
I(h1 ) − I(h2 )
E(h2 ) =
1 − (h1 /h2 )2
Hence a better estimate for the integral is,
1
I = I(h2 ) + [I(h2 ) − I(h1 )]
(h1 /h2 )2 − 1
If h2 = h1 /2 then we have,
1
I = I(h2 ) + [I(h2 ) − I(h1 )]
22 − 1
Since we have estimated and eliminated the truncation error of order

O(h2 ) term, the above equation will have an error of order O(h4 ) which
is the next leading term in the Taylor series expansion. By repeated
application of the above approach to estimate and eliminate successively
higher order terms, we can arrive at the following general formula for
Romberg extrapolation.
4k−1 Ij+1,k−1 − Ij,k−1

Ij,k = (5.25)
4k−1 − 1
Example of Romberg extrapolation
Consider the following integral,

Z 0.8
(0.2 + 25x − 200x 2 + 675x 3 − 900x 4 + 400x 5 ) dx = 1.64053334
0
(5.26)
which has the exact value as shown. A sketch of the function f (x) and
the Romberg extrapolation results are shown in figure 5.10. It is clear
from this example that by combining three rather poor estimates of the
integral on grids of h = 0.8, 0.4 and 0.2, a result accurate to eight sig-
nificant digits has been obtained! For example, I2,2 is obtained by using
j = 2 and k = 2 which results in,
4 × I3,1 − I2,1 4 × 1.4848 − 1.0688

I2,2 = = = 1.6234667
4−1 3
Similarly, I1,3 is obtained by using j = 1 and k = 3 which results in,
42 × I2,2 − I1,2 42 × 1.6234667 − 1.3674667

I1,3 = = = 1.64053334
42 − 1 15
3.5
3
2.5
2
f(x)
1.5
1
0.5
0
0 0.2 0.4 0.6 0.8
x
O ( h2 ) O ( h4 ) O ( h6 )
(k = 1) (k = 2) (k = 3)
j = 1 h = 0.8 0.1728 1.3674667 1.64053334
j = 2 h = 0.4 1.0688 1.6234667 1.64053334
j = 3 h = 0.2 1.4848 1.6394667
j = 4 h = 0.1 1.6008
Figure 5.10: Illustration of Romberg extrapolation

5.8. ORTHOGONAL FUNCTIONS 132
function f=int_ex(x)
%defines a 5th degree polynomial
m=length(x);
for i=1:m
f(i) = 0.2 + 25*x(i) - 200*x(i)ˆ2 + ...
675*x(i)ˆ3 - 900*x(i)ˆ4 + 400*x(i)ˆ5;
end
Figure 5.11: MATLAB implementation of quadrature evaluation
MATLAB example
MATLAB provides a m-file to evaluate definite integrals using adaptive,

recursive Simpson’s quadrature. You must of course, define the func-
tion through a m-file which should accept a vector of input arguments
and return the corresponding function values. A m-file that implements
equation (5.26) is shown in figure 5.11. After creating such a file work
through the following example during an interactive session.
»quad(’int ex’,0,0.8,1e-5) % evaluate integral over (0,0.8)

»quad(’int ex’,0,0.8,1e-5,1) % Display results graphically
»quad(’int ex’,0,0.8,1e-8) % Note the warning messges!
5.7.2 Gaussian quadratures

5.7.3 Multiple integrals
5.8 Orthogonal functions
5.9 Piecewise continuous functions - splines

Education has produced a vast population able to
read but unable to distinguish what is worth read-
ing.
— G.M. TREVELYAN
Chapter 6
Ordinary differential equations

- Initial value problems
In this chapter we develop algorithms for solving systems of linear and

nonlinear ordinary differential equations of the initial value type. Such
models arise in describing lumped parameter, dynamic models. Entire
books (Lapidus & Seinfeld, 1971; Lambert, 1973) are devoted to the de-
velopment of algorithms for such problems. We will develop only ele-
mentary concepts of single and multistep methods, implicit and explicit
methods, and introduce concepts of numerical stability and stiffness.
General purpose routines such as LSODE that implement several of
advanced features such as automatic step size control, error control etc.
are available from NETLIB.
6.1 Model equations and initial conditions
Ordinary differential equations of the initial type are represented typi-

cally as a system of first order equations of the form,
dy
= f (y, t) (6.1)
dt
y(t = t0 ) = y 0 (6.2)
where y(t) is a vector containing elements y = {y1 (t), y2 (t), · · · , yn (t)}

and our objective is to construct an approximate representation of the
function y(t) over some interval of interest t ∈ [t0 , tf ] that satisfy the
133
6.1. MODEL EQUATIONS AND INITIAL CONDITIONS 134
initial conditions, given by another vector y 0 = {y1,0 , y2,0 , · · · , yn,0 }.

If the functions f (y, t) depend on t explicitly, then the equations are
called non-autonomous; otherwise they are called an autonomous sys-
tem of equations.
6.1.1 Higher order differential equations

A higher order differential equation (say of order n) can be converted into
an equivalent system of (n) first order equations. Consider the equation,
dn θ dn−1 θ dθ
an + a n−1 + · · · a1 + a0 θ = b (6.3)
dt n dt n−1 dt
subject to a set of n initial conditions at t0 of the form,

dn−1 θ

= cn−1
dt n−1 t0

dn−2 θ

= cn−2
dt n−2 t0
.. .
. = .. (6.4)

dθ
= c1
dt t0
θ|t0 = c0
Since all of these conditions are given at t0 , this remains an initial value
problem. Equation (6.3) can be recast into a system of n first order equa-
tions of the form (6.1) as follows. Let us define θ and all of its (n − 1)
successive higher derivatives as
dθ d2 θ dn−1 θ
y1 (t) = θ(t), y2 (t) = , y3 (t) = , ··· yn (t) =
dt dt 2 dt n−1
Then we have,
dy1
= y2 , y1 (t0 ) = c0
dt
dy2
= y3 , y2 (t0 ) = c1
dt
.. .
. = .. (6.5)
dyn−1
= yn , yn−1 (t0 ) = cn−2
dt
dyn 1
= [b − a0 y1 − a1 y2 − · · · − an−1 yn ], , yn (t0 ) = cn−1
dt an
6.2. TAYLOR SERIES EXPANSION 135
where the last equation has been obtained from the n-th order equation
(6.3). Also shown in equations (6.5), are the transformed initial condi-
tions from equation (6.4) in terms of the new variable set y.
Note that the coefficients {a0 , a1 , · · · an , b} in equation (6.3) can in
general be nonlinear functions of θ and its derivatives. This nonlinearity
will reflect in equations (6.5),since the coefficients {a0 , a1 , · · · an , b} will
be functions of the transformed variables {y1 , y2 · · · yn }.
6.2 Taylor series expansion
Consider the differential equation,
dy
= f (y), y(t0 ) = y0 (6.6)
dt
Our task is to construct a sequence {yn |n = 0, 1, · · ·} that represents

an approximate solution to y(t) at a discrete set of points {tn |n =
0, 1, · · ·}. We can achieve this by constructing a Taylor series expansion
for y(t) around tn with a step size of h as,
h2
y(tn + h) = y(tn ) + y 0 (tn )h + y 00 (tn ) + ···
2
Truncating after the linear term and recognizing that y 0 (tn ) = f (yn ),
we have the Euler scheme for generating the solution sequence,
yn+1 = yn + hf (yn ) + O(h2 ) n = 0, 1, · · · (6.7)

| {z }
local error
which is single-step, explicit scheme with a local truncation error of order

O(h2 ). It is called a single-step method because it requires only the value
at yn to predict the value at next step yn+1 . It is explicit because the right
hand side terms [yn + hf (yn )] can be computed explicitly using known
value of yn .
6.2.1 Alternate derivation using interpolation polynomials
Rewriting the differential equation (6.6) as,

Z yn+1 Z tn+1
dy = f (y)dt (6.8)
yn tn
and using Newton forward and backward interpolating polynomials to

approximate the function f (y) we can recover, not only the Euler scheme,
but develop a mechanism for obtaining a whole class of implicit and mul-
tistep methods. First let us use the m-th degree Newton forward poly-
nomial from equation (5.11), viz.

α(α − 1) 2
f (y) ≈ Pm (tn + αh) = 1 + α∆ + ∆ + ···
2!

α(α − 1)(α − 2) · · · (α − m + 1) m
∆ fn + O(hm+1 )
(m)!
where tn has been used as the reference point, fn means f (yn ) and h
is the step size. Since t = tn + αh we have dt = hdα. Using a one term
expansion in equation (6.8), ( i.e., m = 0) results in,
Z tn+1 Z tn+1
yn+1 − yn = P0 (tn + αh)dt + O(h)dt
tn t
| n {z }
local truncation error
Z1 Z1
= P0 (tn + αh) hdα + O(h)hdα
0 0
Z1
= fn hdα + O(h2 )
0
= fn h + O(h2 )
which is the same equation as (6.7). This approach, however, lends itself
naturally to further development of higher order methods. For example
a two-term expansion ( i.e., m = 1) results in,
Z tn+1 Z tn+1
yn+1 − yn = P1 (tn + αh)dt + O(h2 )dt
tn t
| n {z }
Z1 Z1
= P1 (tn + αh) hdα + O(h2 )hdα
0 0
Z1
= [fn + α∆fn ] hdα + O(h3 )
0
" #1
α2
= h α fn + ∆fn + O(h3 )
2 0

1
= h fn + (fn+1 − fn ) + O(h3 ).
2
Hence we have the final form of the modified Euler scheme as,
h
yn+1 = yn + [fn + fn+1 ] + O(h3 ) n = 0, 1, · · · (6.9)
2 | {z }
local error
Both the Euler method given in equation (6.7) and the modified Euler
scheme given by equation (6.9) are single-step methods since only yn
is required to predict yn+1 . The modified Euler method is an implicit
scheme since we need to compute fn+1 which depends on yn+1 . Note
that implicit schemes requires the solution of a nonlinear algebraic equa-
tion at every time step. Thus to calculate yn+1 from equation (6.9) we
need to use an iterative method that involves providing an initial guess
for yn+1 and using equation (6.9) as a fixed point iteration scheme until
yn+1 converges to desired accuracy. At a first glance, this might appear
to be a disadvantage of the implicit schemes. However, implicit schemes
have the ability to anticipate sharp changes in the solution between yn
and yn+1 and hence are suitable (in fact required) for solving the so
called stiff differential equations.
This initial guess could be provided by the Euler method ( viz. equa-
tion (6.7)). When an explicit scheme is combined with an implicit scheme
in this manner, we have the so called predictor-corrector scheme. The
Euler and modified Euler predictor-corrector pair is,
P C hh P
i
yn+1 = yn + hf (yn ) and yn+1 = yn + f (yn ) + f (yn+1 )
2
(6.10)
where the superscript P represents the predicted value from an explicit
scheme and C represents the corrected value from an implicit scheme.
It should be clear that extending the Newton forward polynomial to
a three-term expansion will not be fruitful, since that would involve not
only fn+1 , but also fn+2 . We can, however, use Newton backward poly-
nomials to develop higher order methods as will be done in section §6.3.
But, let us explore first the reason for and the circumstances under which
implicit schemes are useful.
6.2.2 Stability limits

Let us consider a model, linear equation,
dy
= λ y, y(t = 0) = 1
dt
which has the analytical solution,

y(t) = eλt
For λ < 0 the exact solution decreases monotonically to zero as t → ∞.
Let us examine the sequence {yn |n = 0, 1, · · ·} generated by the explicit,
Euler scheme and the implicit, modified Euler scheme. Note that in this
model problem the function f (y) = λy. The Euler equation is,
yn+1 = yn + hf (yn ) = yn + hλyn = [1 + hλ]yn
Thus the sequence is,
y1 = [1 + hλ]y0
y2 = [1 + hλ]y1 = [1 + hλ]2 y0
y3 = [1 + hλ]y2 = [1 + hλ]3 y0
.. .
. = ..
leading to the general solution,
yn = [1 + hλ]n y0
When the step size h is chosen to be too large (more specifically |hλ| > 2
in this case), the sequence will diverge, while the exact solution remains
bounded. This phenomenon is called numerical instability caused by the
discretization. Explicit methods in general have such a stability bound
on the step size h.
Let us examine the behavior of an implicit scheme - viz. the modified
Euler scheme.
h h
yn+1 = yn + [fn + fn+1 ] = yn + λyn + λyn+1
2 2
Note that yn+1 appears on both sides. Solving for yn+1 we get,

1 + hλ/2
yn+1 = yn
1 − hλ/2
Thus the sequence is,

1 + hλ/2
y1 = y0
1 − hλ/2

1 + hλ/2 1 + hλ/2 2
y2 = y1 = y0
1 − hλ/2 1 − hλ/2

1 + hλ/2 1 + hλ/2 3
y3 = y2 = y0
1 − hλ/2 1 − hλ/2
.. ..
. = .
slow time scale solution: e-ct/k

x
fast time scale solution: e-kt/m
c
k x
t
Figure 6.1: Spring and dash pot model
leading to the general solution,

n
1 + hλ/2
yn = y0
1 − hλ/2
h i
1+hλ/2
It is clear that for λ < 0, the ratio 1−hλ/2 < 1 for any choice of step
size h. Thus the implicit scheme is absolutely stable. Hence, for explicit
schemes, the choice of h is governed by both stability and truncation er-
ror considerations while for implicit schemes only truncation error con-
siderations dictate the choice of step size, h.
6.2.3 Stiff differential equations
The physical interpretation for λ in the above model problem is that it

represents the characteristic time scale of the problem. For a second
order equation (or equivalently a system of two first-order equations),
there will be two such time scales λ1 and λ2 . If the time scales are widely
separated in magnitude then we have a stiff system of differential equa-
tions. Consider the spring and dash pot model shown in figure 6.1. The
displacement x is modelled by the force balance equation,
d2 x dx
m +k + cx = 0
dt 2 dt
where c is the spring constant, k is the damping factor, m is the mass

and x is the displacement. Let us assume that it is subject to the initial
conditions x(t = 0) = 0 and x 0 (t = 0) = constant. We can write the
characteristic equation as,
m 2 c
λ +λ+ =0
k k
and hence the two roots are given by,
p
−1 ± 1 − 4mc/k2
λ=
(2m/k)
In the limit of m → 0 we can approximate these as,
k c
λ1 = − and λ2 = −
m k
where L’Hopitals rule is used to obtain the second root. Clearly as m → 0,

we have λ1 >> λ2 and this limit corresponds to the stiff behavior of the
solution. In general the stiffness ratio is defined as the ratio of the largest
to the smallest eigenvalues. In this example the stiffness ratio is (k2 /mc)
and it becomes large as m is made small. The solution satisfying the first
initial condition is,
−kt/m ct/k
x(t) = A1 [e
| {z } − |e {z }]
f ast slow
where the fast and slow response terms are as shown. The sketch in
figure 6.1 also shows the fast and slow response solutions. Note that if
m = 0, the order of the differential equation drops by one and λ1 is the
only time scale for the problem. This kind of phenomena also occurs
in a number of chemical reaction systems, where some of the reactions
can occur on a rapid time scale while others take place on a longer time
scale. The ozone decomposition model discussed in section §1.4.2 is
another example of stiff differential equations.
It should now be clear that stiffness phenomena corresponds to large
eigenvalues and fast response regions where the solution changes rapidly.
In a system of n first order equations there will be n characteristic roots
or eigenvalues. If λmax is the largest eigenvalue, then explicit schemes
will typically have a numerical stability limit of the form |hλmax | <
constant. Hence explicit schemes require that extremely small step
size h be used in regions where the system responds very rapidly; oth-
erwise the integration sequence will diverge. Implicit schemes that are
absolutely stable have no such restrictions. The integration sequence
using implicit schemes will remain bounded. The choice of step size is
determined only by the desired accuracy of the solution. Stability analy-
sis for a variety of explicit and implicit methods are discussed in greater
detail by Lapidus and Seinfeld (1971).
6.3. MULTISTEP METHODS 141
6.3 Multistep methods
6.3.1 Explicit schemes

Consider approximating the function f (y) in equation (6.8) by the fol-
lowing m-th degree Newton backward polynomial from equation (5.13),

α(α + 1) 2
f (y) ≈ Pm (tn + αh) = 1 + α∇ + ∇ + ···
2!

α(α + 1)(α + 2) · · · (α + m − 1) m
∇ fn + O(hm+1 )
m!
Here, tn has been used as the reference point. Since this polynomial in-
volves only points at earlier times such as {fn , fn−1 , fn−2 · · ·}, we can de-
velop a class of explicit schemes of high orders. These are called Adams-
Bashforth schemes. Consider a three-term expansion ( i.e., m = 2). Equa-
tion (6.8) becomes,
Z tn+1 Z tn+1
tn t
| n {z }
Z1 Z1
0 0
Z1
α(α + 1) 2
= fn + α∇fn + ∇ fn hdα + O(h4 )
0 2!
" ( ) #1
α2 1 α3 α2
= h α fn + ∇fn + + ∇ fn + O(h4 )
2
2 2! 3 2 0

1 5
= h fn + (fn − fn−1 ) + (fn − 2fn−1 + fn−2 ) + O(h4 ).
2 12
which can be rearranged into the form,
h
yn+1 = yn + [23 fn − 16 fn−1 + 5 fn−2 ] + O(h4 ) n = 2, 3, 4, · · ·
12 | {z }
local error
(6.11)
The following points should be observed on the above equation.
• This is a multistep scheme since it requires (yn , yn−1 , yn−2 ) to pre-

dict yn+1 .
• Hence it is not a self-starter! Typically, in a well posed initial value

problem, we know only y0 . Hence y1 and y2 must be generated
from other single-step methods before we can switch to the above

multistep scheme.
• It is an explicit method as yn+1 does not appear on the right hand

side.
• As a consequence it cannot anticipate sharp changes in y(t) - i.e.,

not suitable for stiff differential equations.
• Makes good use of previous calculations to give low truncation er-

ror.
• requires only one function evaluation per step.
The next higher order scheme can be developed from a four-term ex-
pansion ( i.e., m = 3) . This is called 5-th order Adams-Bashforth scheme.
viz.
Z tn+1 Z tn+1
tn t
| n {z }
Z1 Z1
0 0
Z1
α(α + 1) 2 α(α + 1)(α + 2) 3
= fn + α∇fn + ∇ fn + ∇ fn hdα + O(h5 )
0 2! 3!

1 5 2 3 3
= h fn + ∇fn + ∇ fn + ∇ fn + O(h5 )
2 12 8
which can be rearranged into the form,
h
yn+1 = yn + [55 fn − 59 fn−1 + 37 fn−2 − 9fn−3 ] + O(h5 ) n = 3, 4, · · ·
24 | {z }
local error
(6.12)
6.3.2 Implicit schemes

In order to construct implicit schemes we need to construct backward
polynomial approximations with tn+1 as the reference point. viz.

α(α + 1) 2
f (y) ≈ Pm (tn+1 + αh) = 1 + α∇ + ∇ + ···
2!

α(α + 1)(α + 2) · · · (α + m − 1) m
∇ fn+1 + O(hm+1 )
m!
In this manner fn+1 is introduced on the right hand side. This class
of implicit schemes are called Adams-Moulton schemes. We are still
integrating one step from tn to tn+1 . Since t = tn+1 + αh and the limits
of integration in α become (−1, 0). A four-term expansion results in,
Z tn+1 Z tn+1
yn+1 − yn = P3 (tn+1 + αh)dt + O(h4 )dt
tn t
| n {z }
Z0 Z0
−1 −1
Z0
α(α + 1) 2 α(α + 1)(α + 2) 3
= fn+1 + α∇fn+1 + ∇ fn+1 + ∇ fn+1 hdα + O(h5
−1 2! 3!

1 1 2 1 3
= h fn+1 − ∇fn+1 − ∇ fn+1 − ∇ fn+1 + O(h5 )
2 12 24
which can be expanded and rearranged into the form,
h
yn+1 = yn + [9 fn+1 + 19 fn − 5 fn−1 + fn−2 ] + O(h5 ) n = 2, 4, · · ·
24 | {z }
local error
(6.13)
The pair of explicit-implicit schemes given by (6.12,6.13) respectively can
be used as a predictor-corrector pair.
6.3.3 Automatic stepsize control

Some of the start up and step size control issues are illustrated in figure
6.2 for the 5-th order Adams schemes developed in the last section. Note
that only y0 is given and hence (y1 , y2 , y3 ) must be generated using
some other single step methods with a step size of h before the 5-th
order Adams scheme given by equations (6.12,6.13) can be used. In doing
so, it is important to realize that any error introduced in these three steps
are likely to be propagated during the transient phase of the simulation.
Hence if lower order schemes are used to generate (y1 , y2 , y3 ), then
smaller step sizes must be used. The difference between the predicted
and corrected values could be used as a measure of the truncation error.
If this error is below an acceptable tolerance, then we can choose to
double the next step size. But this can begin only after y6 has been
computed, because we need four previous values at equal intervals of
(2h) - i.e., (y0 , y2 , y4 , y6 ). If at any time during the intergration process,
the difference between the predicted and corrected values is above the
if you need to halve the step size

generate y4.5 and y5.5
possible to double
step size
y0 } y1 y2 y3 y4 y5 y6
given
begin using
generate using
Adams scheme
other methods
Figure 6.2: Stepsize control strategies for multistep methods
tolerance, then we must halve the step size and repeat the calculation
for that step. In so doing, we need to generate intermediate values at
intervals of (h/2). For example if the result for y7 does not meet the
tolerance, then we repeat the calculation from y6 with a step size of
h/2. We need to generate intermediate values at y4.5 and y5.5 . This can
be done using the Newton backward interpolation polynomials; but the
truncation errors in the interpolating polynomials should be of the same
order and the Adams scheme. Specifically the interpolation rules are:
1
yn− 1 = [35yn + 140yn−1 − 70yn−2 + 28yn−3 − 5yn−4 ]
2 128
1
yn− 3 = [−yn + 24yn−1 + 54yn−2 − 16yn−3 + 3yn−4 ]
2 64
Example
MATLAB has several built-in functions for solving initial value problems.
The functions named ADAMS and GEAR, use multistep methods. All of
the ODE solvers in MATLAB are part of the SIMULINK toolbox. Hence
the m-files that define the problem must have a special structure. In
this section we illustrate how to use these functions to solve the ozone
decomposition model. Recall that the equations are,
dy1
= f1 (y1 , y2 ) = −y1 − y1 y2 + κy2
dt
dy2
= f2 (y1 , y2 ) = (y1 − y1 y2 − κy2 )/
dt
The initial compositions are y(t = 0) = [1.0, 0.0]. The parameters are
= 1/98 and κ = 3.0. The equations are clearly nonlinear. The m-file
6.4. RUNGE-KUTTA METHODS 145
function [ydot,y0]=ozone(t,y,u,flag)
k=3.0;epsilon=1/98;
if abs(flag) == 1
ydot(1) = -y(1) - y(2)*y(1) + k*epsilon*y(2);
ydot(2) = (y(1)-y(1)*y(2)-epsilon*k*y(2))/epsilon;
elseif flag == 0
ydot=[2,0,0,0,0,0]; %first element=number of equations
y0=[1 0]; %initial conditions
else
ydot=[];
end
Figure 6.3: MATLAB implementation of ozone decomposition model
named ozone.m is shown in figure 6.3. This function should be written

in such a way that it should return the derivatives y 0 (t) in the variable
ydot when iflag==1 and it should return the number of equations and
the initial condition when iflag==0 as shown in the figure 6.3.
To use the GEAR and ADAMS functions and integrate the ozone model
to a final time of 3.0 do the following during an interactive MATLAB ses-
sion.
»tf=3.0; % Define final time

»type ozone % test that ozone.m exists
»[t,y]=gear(’ozone’,tf) % Integrate using GEAR
»[t,y]=adams(’ozone’,tf) % Integrate using ADAMS
The independent variable t and the solution y at the same time values
are returned in the corresponding varaibles. These results are shown
graphically in figure 6.4. Observe that y2 (t) increases rapidly from an
initial condition of zero and hence the system is very stiff during the
early times.
6.4 Runge-Kutta Methods
While multi-step methods achieve high order accuracy by making effi-

cient use of previously generated results, Runge-Kutta methods achieve
1.00
y1
0.75 y2
y1 or y2
0.50
0.25
0
0 1 2 3
time
Figure 6.4: Results of ozone decomposition model shows a stiff system

behavior
the same goal in a single step, but at the expense of requiring many
function evaluations per step. Being single-step schemes, they are self-
starters. They are also classified as explicit, semi-implicit and implicit
schemes. Implicit schemes require solution of a set on non-linear alge-
braic equations at every time step, but they are suitable for stiff differ-
ential equations.
6.4.1 Explicit schemes
Explicit schemes have the general form,
v
X
yn+1 = yn + w i ki (6.14)
i=1
 
i−1
X
ki = h f tn + ci , yn + aij kj  , c1 = 0, i = 1, 2, · · · v
j=1
In these equations, {ci , wi , aij } are all parameters. The development of

a specific scheme entails determining the best possibile values for these
constants by matching the expansion of this formula with a Taylor series
expansion. Often these parameter values are given in tabular form as,
0
c2 a21
c3 a31 a32
c4 a41 a42 a43
w1 w2 w3 w4
or
c A
w
For explicit methods, A is a lower triangular matrix.
6.4.2 Euler formula revisited

Let us consider v = 1 in equation (6.14) and write the equations explicitly
as,
yn+1 = yn + w1 k1 (6.15)
k1 = h f (tn , yn )
or
yn+1 = yn + w1 h f (tn , yn )
The procedure to determine w1 is to match the above equaiton with the
Taylor series expansion,
h2 00 h3 000
yn+1 = yn + hyn0 + y + y + ··· (6.16)
2 n 3! n
The first term on the right hand side is the same in both equations.
Recongnizing that y 0 = f the second term will also match if we make
w1 = 1 and this results in recovering the Euler scheme developed in
equation (6.7). We cannot match with any higher order terms and hence
the local truncation error is of order O(h2 ).
6.4.3 A two-stage (v = 2) Runge-Kutta scheme

Let us consider v = 2 in equation (6.14) and write the equations explicitly
as,
yn+1 = yn + w1 k1 + w2 k2 (6.17)
k1 = h f (tn , yn )
k2 = h f (tn + c2 h, yn + a21 k1 )
This scheme has four unknown paramters {w1 , w2 , c2 , a21 } which must
be determined by matching the Taylor series expansion of equation (6.17)
with the equation (6.16). Expanding equation (6.17) we get,
yn+1 = yn + w1 h f (tn , yn ) + w2 h f (tn + c2 h, yn + a21 k1 )
or,
" #
∂f ∂f
yn+1 = yn + w1 h fn + w2 h f + (c2 h) + a21 (hf ) + ···
∂t ∂y n
(6.18)
Substituting for y 0 and its higher derivatives in terms of f in equation
(6.16) and expanding, we get,
h2 df
yn+1 = yn + hfn + + O(h3 )
2 dt
or " #
h2 ∂f ∂f ∂y
yn+1 = yn + hfn + + + O(h3 ) (6.19)
2 ∂t ∂y ∂t
Now comparing the fn terms between equations (6.18) and (6.19) we
require that
w1 + w 2 = 1
∂f
for the two equations to match. Next comparing ∂t terms, we require,
w2 c2 = 1/2
∂f ∂y
Finally comparing ∂y ∂t we require,
w2 a21 = 1/2
Thus, we have matched all terms of order O(h2 ) leaving a truncation er-
ror or order O(h3 ). In that process we have developed 3 constraint equa-
tions on the four unknonws {w1 , w2 , c2 , a21 } appearing in the 2-stage
RUnge-Kutta scheme (6.17). Any choice of values for {w1 , w2 , c2 , a21 }
that satisfies the above three constraints will result in a 3-rd order, 2-
stage Runge-Kutta scheme. Since there are four variables and only three
equations, we have one extra degree of freedom. Hence the solution is
not unique. Two sets of results are:
w1 = 2/3, w2 = 1/3, c2 = 3/2, a21 = 3/2
and
w1 = 1/2, w2 = 1/2, c2 = 1, a21 = 1
The later is the equivalent of the predictor-corrector pair using Euler and
modified Euler schemes developed in equations (6.10). In summary this
scheme is a 2-stage RK method since it requires two function evaluations
per step. It is explicit and has a local truncation error of O(h3 ).
Using the first set of parameter values in equation (6.17), we have,
2 2
yn+1 = yn + k1 + k2
3 3
k1 = h f (tn , yn )
3 3
k2 = h f (tn + h, yn + k1 )
2 2
or in tabular form,
0
3/2 3/2
2/3 2/3
A fourth order Runge-Kutta scheme
Higher order Runge-Kutta methods can be developed by carrying out

the matching process with the Taylor series expansion to higher order
terms. An explicit fourth-order form that matches with the Taylor series
to h4 terms (and hence has a truncation error of O(h5 ) is,
0
1/2 1/2
1/2 0 1/2
1 0 0 1
1/6 2/6 2/6 1/6
Which in expanded form,
1
yn+1 = yn + [k1 + 2k2 + 2k3 + k4 ] (6.20)
6
k1 = h f (tn , yn )
h k1
k2 = h f (tn + , yn + )
2 2
h k2
k3 = h f (tn + , yn + )
2 2
k4 = h f (tn + h, yn + k3 )
As with all explicit schemes, it is good for non-stiff differential equations.

Embedded forms
The embedded forms of Runge-Kutta algorithms provide a pair of schemes

that use a common set of function evaluations to predict two estimates of
the solution at yn+1 . Typically a lower order scheme is embedded within
a higher order scheme. The motivation for developing such schemes is
to have a convenient estimate of the local truncation error at every time
step, which can then be used to develop a step size control stragegy. A
popular scheme, called RK45, is given below.
k1 0
1 1
k2 4 4
3 3 9
k3 8 32 32
12 1932 7200 7296
k4 13 2197 − 2197 2197
439 3680 −845
k5 1 216 −8 513 410
25 1408 2197 (4)

216 0 2565 4104 − 15 ← yn+1
1 8
− 3544
1859 11
k6 2 − 27 2 2565 4104 − 40
16 6656 28561 9 2 (5)
135 0 12825 56430 − 50 55 ← yn+1
MATLAB has an implimentation of this scheme both as a built-in func-

tion called rk45.m and a m-file called ode45.m.
6.4.4 Semi-implicit & implicit schemes

In the general form given in equation (6.14), we fixed c1 = 0 and A to be
lower triangular. These constraints on the paramters ensure that each
of the ki could be computed explicitly without the need for iterative
solution. Let us now relax these constraints and write the general form
as,
v
X
yn+1 = yn + w i ki (6.21)
i=1
 
v
X
ki = h f tn + ci , yn + aij kj  , i = 1, 2, · · · v
j=1
Thus, for a two-stage process (v = 2), we will have,
yn+1 = yn + w1 k1 + w2 k2 (6.22)
k1 = h f (tn + c1 h, yn + a11 k1 + a12 k2 )

k2 = h f (tn + c2 h, yn + a21 k1 + a22 k2 )
or in compact tabular form,
c1 a11 a12
c2 a21 a22
w1 w2
Note that in order to compute k1 and k2 in the above form, we need

to solve two sets of nonlinear algebraic equations simulataneously. On
the positive side of such schemes, fully implicit nature of the algorithm
results in numerically stable schemes making them suitable for stiff dif-
ferential equations. Also, a two-stage scheme (v = 2) has eight param-
eters, and hence we can match these equations with the Taylor series
expansion to higher order terms. Hence more accurate formulas can be
constructed. An example of a fully implicit, 2-stage, 4-th order accurate
scheme is the Gauss form given by,
√ √
(3 − √3)/6 1/4 √ (3 − 2 3)/12
(3 + 3)/6 (3 − 2 3)/12 1/4
1/2 1/2
For an extensive catalogue of such schemes see Lapidus and Seinfeld

(1971).
6.4.5 Semi-Implicit forms of Rosenbrock
While the fully-implicit forms of the last section §6.4.4, have desirable
stability properties, they are computationally demanding since a system
of non-linear algebraic equations must be solved iteratively at every time
step. In an effort to reduce the computational demand while retaining
the stability characteristics, Rosenbrock (1963) proposed a special form
of the algorithm. These are suitable for autonomous system of equations
of the form,
dy
= f (y) y(t = t0 ) = y 0
dt
A 2-stage, 3-rd order scheme is shown below.
y n+1 = y n + w1 k 1 + w2 k 2 (6.23)
−1
k1 = h [I − ha1 J(y n )] f (y n )
k2 = h [I − ha2 J(y n + c1 k1 )]−1 f (y n + b1 k1 )
6.5. DYNAMICAL SYSTEMS THEORY 152
The parameters are,

p p
a1 = 1 + 6/6, a2 = 1 − 6/6
w1 = −0.41315432, w2 = 1.41315432
√ p √
−6 − 6 + 58 + 20 6
b1 = c1 = √
6+2 6
∂f
Here J = ∂ y is the Jacobian, which must be evaluated at every time
step. Note that the main advantage in using equation (6.23) is that k1 , k2
could be computed without the need for iteration, although it requires
two matrix inverse computations per step.
6.5 Dynamical systems theory

He who can, does.
He who cannot teaches.
— GEORGE BERNARD SHAW
Chapter 7
Ordinary differential equations

- Boundary value problems
In the present chapter we develop algorithms for solving systems of (lin-

ear or nonlinear) ordinary differential equations of the boundary value
type. Such equations arise in describing distributed, steady state mod-
els in one spatial dimension. The differential equations are transformed
into systems of (linear and nonlinear) algebraic equations through a dis-
cretization process. In doing so, we use the tools and concepts devel-
oped in Chapter 5. In particular we will develop (i) finite difference meth-
ods using the difference approximations given in Table 5.4, (ii) shooting
methods based on methods for initial value problems seen in chapter 6
and (iii) the method of weighted residuals using notions of functional
approximation developed in Chapter 5.
We will conclude this chapter with an illustration of a powerful soft-
ware package called COLSYS that solves a system of multi-point bound-
ary value problems using the collocation method, cubic splines and adap-
tive mesh refinement. It is available from NETLIB.
7.1 Model equations and boundary conditions
Consider the model for heat transfer through a fin developed in section
§1.5.1. We will consider three specific variations on this model equation
(1.22) and the associated boundary conditions. First let us scale the
problem by introducing the following dimensionless temperature and
153
7.1. MODEL EQUATIONS AND BOUNDARY CONDITIONS 154
distance variables,
T − T∞ x
θ= ξ=
T0 − T ∞ L
Using these definitions, equation (1.22) can be rewritten as,
" #
d dθ
kA − hP L2 θ = 0 (7.1)
dξ dξ
θ(ξ = 0) = 1 θ(ξ = 1) = 0
where the objective is to find the continuous function θ(ξ) over the
domain of interest, viz. ξ ∈ [0, 1] for a prescribed set of parameters
{k, A, h, P , L}. All of these parameters can be constants, or some of them
might be dependent on the position ξ ( e.g., A(ξ)) or on the unknown
temperature itself ( e.g., k(θ)). We examine each case next.
For constant area, A and thermal conductivity, k, equation (7.1) re-
sults in a second order linear differential equation with constant coeffi-
cients for which an analytical solution is possible. But we focus only on
developing methodologies for obtaining a numerical solutions.
d2 θ hP L2
− θ = 0 (7.2)
dξ 2 kA
θ(ξ = 0) = 1, θ(ξ = 1) = 0
When the values of the dimensionless temperature, θ on the boundary

are specified, the boundary condition is called the Dirichlet boundary
conditions.
In a variation of the above model, if we let the area be a variable
A(ξ) ( i.e., tapered fin), but keep the thermal conductivity, k, constant,
we obtain a variable coefficient, linear boundary value problem, still of
second order.
d2 θ dA(ξ) dθ hP L2
A(ξ) 2
+ − θ = 0 (7.3)
dξ dξ dξ k

dθ

θ(ξ = 0) = 1, = 0
dξ ξ=1
7.1. MODEL EQUATIONS AND BOUNDARY CONDITIONS 155
In the above model, we have also introduced a variation on the bound-

ary condition at ξ = 1. Such boundary conditions, where the derivatives
are specified, are called Newman boundary conditions. The temperature
value at ξ = 1 is an unknown and must be found as part of the solution
procedure.
In yet another variation, consider the case where the thermal conduc-
tivity is a function of temperature, viz. k(θ) = α + βθ 2 where α and β
are experimentally determined constants. Let the area, A be a constant.
The equation is nonlinear and must be solved numerically.
" #2
d2 θ dk(θ) dθ hP L2
k(θ) 2 + − θ = 0 (7.4)
dξ dθ dξ A
" #
dθ
θ(ξ = 0) = 1, k + hθ = 0
dξ ξ=1
At this opportunity a third variation on the boundary condition, called

the mixed or Robin boundary condition has been used. Once again, the
temperature value at x = L is an unknown and must be found as part of
the solution procedure.
All of these problems can be represented symbolically as,
Dθ = f on Ω
Bθ = g on ∂Ω
where D and B are differential operators, Ω is the domain of interest

and ∂Ω represents its boundary. Our task is to obtain an approximate
solution, θ̃ to the above problem. The approximation consists in con-
structing a discrete version of the differential equations which results in
a system algebraic equations. If the differential equations are linear (as
in equations (7.2,7.3), then the resulting discrete, algebraic equations will
also be linear of the type Aθ̃ = b and methods of Chapter 3 can be used
to obtain the final approximate solution. If the differential equations
are nonlinear (as in equation (7.4)) then the resulting discrete, algebraic
equations will also be nonlinear of the type F(θ̃) = 0 and methods of
Chapter 4 can be used to obtain the final approximate solution.
In the following sections we develop various schemes for construct-
ing approximate solutions.
7.2. FINITE DIFFERENCE METHOD 156
θ(ξ)
θ1
θ2
θi
∆ξ
ξ=0
ξ=1
1 2 i n
i=0 i=n+1
Figure 7.1: One dimensional finite difference grid of equally spaced data
points
7.2 Finite difference method
7.2.1 Linear problem with constant coefficients

Let us consider equation (7.2) which is a linear problem subject to Dirich-
let boundary conditions. In solving equation (7.2) by the finite difference
method, we divide the domain Ω = ξ ∈ [0, 1] into (n + 1) equally spaced
subdivisions as shown in figure 7.1. The distance between the two grid
points is denoted by ∆ξ. The grid spacing ∆ξ and the value of the inde-
pendent variable ξ at the nodal point i are given by,
1−0
∆ξ = ξi = i ∆ξ i = 0, 1, · · · (n + 1)
n+1
Next, instead of attempting to find a solution θ(ξ) as a continuous func-
tion of ξ that satisfies the differential equation (7.2) exactly at every ξ, we
content ourselves with finding an approximate solution at the selected
nodal points shown in the figure - i.e., {θ(ξi ) = θ̃i |i = 1, 2, · · · n}, where
n is the number of interior grid points. Note that for the Dirichlet type
of boundary conditions θ0 and θn+1 are known. Hence, it remains to
determine only n unknowns at the interior points. We obtain n equa-
tions by evaluating the differential equation (7.2) at the interior nodal
points. In doing this we replace all the derivatives by the corresponding
finite difference approximation from Table 5.4. Clearly, we have several
choices; but it is important to match the truncation error in every term to
be of the same order. We illustrate this process using central difference
approximations.
Using the central difference approximation for the second derivative

in equation (7.2), we obtain,
" #
θ̃i−1 − 2θ̃i + θ̃i+1 2 hP L2
+ O((∆ξ) ) − θ̃i = 0 i = 1, 2, · · · n
(∆ξ)2 kA
The term in the square brackets is the central difference approximation
for the second derivative and O((∆ξ)2 ) is included merely to remind us
of the order of the truncation error. We now have a system of n linear
algebraic equations. Let us write these out explicitly for n = 4.
" #
hP L2 2
θ̃0 − 2θ̃1 + θ̃2 − (∆ξ) θ̃1 = 0
kA
" #
hP L2 2
θ̃1 − 2θ̃2 + θ̃3 − (∆ξ) θ̃2 = 0
kA
" #
hP L2 2
θ̃2 − 2θ̃3 + θ̃4 − (∆ξ) θ̃3 = 0
kA
" #
hP L2 2
θ̃3 − 2θ̃4 + θ̃5 − (∆ξ) θ̃4 = 0
kA
θ̃0 in the first equation and θ̃5 in the last equation are known from the
Dirichlet boundary conditions. The above equations can be expressed in
matrix notation as, T θ̃ = b,
    
−(2 + α) 1 0 0 θ̃1 −θ̃0
   
 1 −(2 + α) 1 0  θ̃2 
= 0  
  
  0 
 0 1 −(2 + α) 1  θ̃
 3  
0 0 1 −(2 + α) θ̃4 −θ̃5
(7.5)
hP L2 2
where α = kA (∆ξ) . Note that the boundary values θ̃0 and θ̃5 appear
as forcing terms on the right hand side. Equation (7.5) is the discrete
version of equation (7.2). Once the structure is apparent, we can increase
n to reduce (∆ξ) and hence reduce the truncation error. In the limit of
∆ξ → 0 the solution to equations (7.5) will approach that of equation
(7.2). The matrix size will increase with decreasing ∆ξ and increasing n.
The matrix T is tridiagonal and hence the Thomas algorithm developed
in section §3.4.4 can be used to get the solution.
7.2.2 Linear problem with variable coefficients

Next we consider equation (7.3) which is also a linear problem subject
to Dirichlet condition at ξ = 0 and Neuman boundary condition at ξ =
1. It also has variable coefficients, A(ξ) and A0 (ξ) - i.e., A is a known

function of ξ. The discretization procedure remains the same as seen in
the previous section §7.2.1, with the exception that θ̃n+1 is now included
in the unknown set
θ̃ = {θ̃1 , θ̃2 , · · · , θ̃n+1 }.
But we have the Neuman boundary condition as an extra condition that
will provide the additional equation. Using the central difference ap-
proximations for both the second and first derivatives in equation (7.3),
we obtain,
" #
θ̃i−1 − 2θ̃i + θ̃i+1 2
A(ξ
+ i) + O((∆ξ) )
(∆ξ)2
" #
θ̃ i+1 − θ̃ i−1 hP L2
A0 (ξi ) + O((∆ξ)2 ) − θ̃i = 0 i = 1, 2, · · · n + 1
2(∆ξ) k
Once again the truncation error term has been retained at this stage, only
to emphasize that it should preferably be of the same order for every
derivative that has been replaced by a difference approximation; other-
wise the effective error is of the same order as the term with the lowest
order truncation error. Multiplying throughout by (∆ξ)2 and collecting
like terms together, we get,
" #
0 (∆ξ) hP L2 (∆ξ)2
A(ξi ) − A (ξi ) θ̃i−1 + −2A(ξi ) − θ̃i +
2 k

0 (∆ξ)
A(ξi ) + A (ξi ) θ̃i+1 = 0, i = 1, 2, · · · n + 1
2
Letting

0 (∆ξ)
ai = A(ξi ) − A (ξi )
" 2 #
hP L (∆ξ)2
2
di = −2A(ξi ) − i = 1, · · · n + 1
k

0 (∆ξ)
ci = A(ξi ) + A (ξi )
2
we can rewrite the equation as,
ai θ̃i−1 + di θ̃i + ci θ̃i+1 = 0, i = 1, 2, · · · n + 1
Observe that the coefficients {ai , di , ci } in the above equations are known.
However, unlike in equation (7.5), they vary with the grid point location
i. Also, for i = 1, θ̃0 on the left boundary is known through the Dirichlet
boundary condition. The last equation for i = n + 1 needs special atten-

tion since it contains the unknown θ̃n+2 which lies outside the domain
of interest. So far we have not used the Neuman boundary condition at
the right boundary ξn+1 = 1. Using the central difference approximation
for the first derivative to discretize the Neuman boundary condition we
get, " #
dθ
θ̃n+2 − θ̃n
≈ =0
dξ ξn+1 =1 2∆ξ
which implies θ̃n+2 = θ̃n . This can be used to eliminate θ̃n+2 from the
last equation, which becomes,
(ai + ci )θ̃n + di θ̃n+1 = 0, for i=n+1
Thus we obtain a tridiagonal system of linear equation of the form T θ̃ =

b. For n = 4, as an example, we get the follwoing five equations.
    
d1 c1 0 0 0 θ̃1 −a1 θ̃0
    
 a2 d2 c2 0 0  θ̃2   0 
    
 0 a3 d3 c3 0  θ̃3 = 0  (7.6)
    
 0 0 a4 d4 c4 
 θ̃4
 
 0 
0 0 0 a5 + c5 d5 θ̃5 0
Equation (7.6) is the discrete version of equation (7.3). Once the structure
is apparent, we can increase n to reduce (∆ξ) and hence reduce the
truncation error. In the limit of ∆ξ → 0 the solution to equations (7.6)
will approach that of equation (7.3). The matrix size will increase with
decreasing ∆ξ and increasing n. The matrix T is tridiagonal and hence
the Thomas algorithm developed in section §3.4.4 can be used to get the
solution.
7.2.3 Nonlinear problem

Conceptually there is no difference in discretizing a linear or a non-linear
differential equation. The process of constructing a grid and replacing
the differential equations with the difference equations at each grid point
is the same. The main difference lies in the choice of solution technique
available for solving the nonlinear algebraic equations. Let us consider
the nonlinear model represented by equation (7.4). In this case we have
a Robin condition at ξ = 1 and hence θ̃n+1 is unknown. Thus the un-
knowns on the discrete grid consist of
θ̃ = {θ̃1 , θ̃2 , · · · , θ̃n+1 }.

and we need n + 1 equations.

Discretizing equation (7.4) at a typical grid point i, we obtain the
following (n + 1) nonlinear algebraic equations.
" #
θ̃i−1 − 2θ̃i + θ̃i+1
fi (θ̃) := k(θ̃i ) +
(∆ξ)2
" #2
0 θ̃i+1 − θ̃i−1 hP L2
k (θ̃i ) − θ̃i = 0 i = 1, 2, · · · n + 1
2(∆ξ) A
In this set of equations i = 1 and i = (n+1) require special consideration
to incorporate the boundary conditions. Thus, making use of the left
boundary condition, θ̃0 = 1, f1 (θ̃) becomes,
" # " #2
1 − 2θ̃1 + θ̃2 0 θ̃2 − 1 hP L2
f1 (θ̃1 , θ̃2 ) := k(θ̃1 ) + k ( θ̃ 1 ) − θ̃1 = 0
(∆ξ)2 2(∆ξ) A
At the right boundary, we discretize the Robin boundary condition as,
" #
θ̃n+2 − θ̃n
k(θ̃n+1 ) + hθ̃n+1 = 0
2(∆ξ)
" ! #
2(∆ξ)h h i
θ̃n+2 = θ̃n − θ̃n+1 = θ̃n − βθ̃n+1
k(θ̃n+1 )
This can be used in the equation fn+1 (θ̃) to eliminate θ̃n+2 .
" #
θ̃n − 2θ̃n+1 + [θ̃n − βθ̃n+1 ]
fn+1 (θ̃n , θ̃n+1 ) := k(θ̃n+1 ) +
(∆ξ)2
" #2
0 [θ̃n − βθ̃n+1 ] − θ̃n hP L2
k (θ̃n+1 ) − θ̃n+1 = 0
2(∆ξ) A
The above equations f1 = 0, f2 = 0, · · · fn+1 = 0 can be represented
symbolically as a system of (n + 1) nonlinear equations of the form,
F(θ̃) = 0. These can be solved most effectively by the Newton method
for the unknowns θ̃ = {θ̃1 , θ̃2 , · · · , θ̃n+1 }.
p+1 p p
θ̃ = θ̃ − J −1 f (θ̃ ) p = 0, 1, · · ·
∂F
The Jacobian matrix, J = has the following tridiagonal structure.
∂ θ̃
 ∂f1 ∂f1

0 0
 ∂ θ̃1 ∂ θ̃2 
 ∂f2 ∂f2 ∂f2
0 


J=

∂ θ̃1 ∂ θ̃2
∂f3
∂ θ̃3
∂f3 ∂f3 

 0 
 ∂ θ̃2 ∂ θ̃3 ∂ θ̃4 
∂f4 ∂f4
0 0 ∂ θ̃3 ∂ θ̃4
7.3. QUASILINEARIZATION OF NONLINEAR EQUATIONS 161
7.3 Quasilinearization of nonlinear equations
Recall the Newton method applied to a system of nonlinear algebraic

equations:
f (x) = 0 (7.7)
A solution is sought by linearizing the equations using
x = xa + δ (7.8)
where x a is an approximate solution and δ is the correction vector.

Hence the original equation becomes,

∂f
δ + O(δ2 )
f (x a + δ) = f (x a ) + (7.9)
∂x xa
Neglecting second and higher order terms, one gets the linearized equa-
tion,
∂f
δ = −f (x a ) (7.10)
∂x xa
Similar concepts can be applied to a (system of) differential equations
as well. Consider for example the equation given below
" #
d dθ
k(θ)ξ =0 (7.11)
dξ dξ
with the boundary conditions
θ(ξ = 1) = 1 inner wall; θ(ξ = ξ0 ) = 0 outer wall
The thermal conductivity is given by k(θ) = α + γθ 2 The equation can

be written as,
" #2
d2 θ dθ dk dθ
f (θ) := k(θ)ξ + k(θ) +ξ =0 (7.12)
dξ 2 dξ dθ dξ
The solution to this nonlinear problem can be obtained by successive

iteration from an initial guess θa (ξ). Let the actual solution be written
as
θ(ξ) = θa (ξ) + δ(ξ) (7.13)
where δ(ξ) is a small correction to be obtained by solving a linearized
form of equation (7.12). The linearized equation is derived by substitut-
ing equation (7.13) in equation (7.12), expanding the resulting expression
and neglecting quadratic and higher order terms in δ(ξ) in the expan-
sion.
7.3. QUASILINEARIZATION OF NONLINEAR EQUATIONS 162
" #2
d2 (θa + δ) d(θa + δ) dk
d(θ a + δ)
k(θa +δ)ξ +k(θa +δ) +ξ =0
dξ 2 dξ dθ θa +δ dξ
(7.14)
Note that each nonlinear term must now be expanded in Taylor series,
e.g.
dk
δ + ···
k(θa + δ) ≈ k(θa ) +
dθ θa

dk dk d 2k
+
≈ δ + ··· (7.15)
dθ θa +δ dθ θa dθ 2 θa
etc. Hence the expanded form of the equation is,

" #
d 2 θa d2 δ dk
δ d 2θ
a dk
δ d 2δ
ξ k(θa ) + k(θa ) 2 + + + ···
dξ 2 dξ dθ θa dξ 2 dθ θa dξ 2

dθa dδ dk
δ dθa + dk δ dδ + · · ·
k(θa ) + k(θa ) +
dξ dξ dθ θa dξ dθ θa dξ
  ( ) ( ) ( )
 dk
d 2k  dθ
2
dδ
2
dθ dδ
+ a a
ξ δ + ···  + +2  = (7.16)
0
 dθ θa dθ 2 θa  dξ dξ dξ dξ
Now, terms of order δ2 and higher are neglected since δ is small. The
terms that are evaluated at the current guess θa are moved to the right
hand side. Hence we get,

d2 δ dk
d 2 θa dδ dk
δ dθa +
k(θa )ξ +ξ δ + k(θa ) +
dξ 2 dθ θa dξ 2 dξ dθ θa dξ
( )2 ( )
d2 k
dθa dk
dθa dδ
ξ δ + 2ξ
dθ 2 θa dξ dθ θa dξ dξ
 )2 
(
2
d θa dθa
dk dθa 
= − k(θa )ξ + k(θa ) +ξ (7.17)
dξ 2 dξ dθ θa dξ
Note that the right hand side of equation (7.17) is the same as equation
(7.12) evaluated at θa and when it is zero the problem is solved! The
left hand side of equation (7.17) is linear in δ subject to homogeneous
boundary condition of δ(1) = 0 and δ(ξ0 ) = 0. Hence the solution (or
correction) δ(ξ) will be zero when the iteration is converged. Now one
can discretize equation (7.17) and solve for the correction iteratively.
Equation (7.17) can also be written in operator form as,
Dθ δ = −f (θ)
7.4. CONTROL VOLUME METHOD 163
θ(ξ)
θ1
θ2
θi
∆ξ
ξ=0
ξ=1
1 2 i n
i=0 i=n+1
Figure 7.2: One dimensional control volume discretization
where the linear operator Dθ , called the Fréchet derivative is given by,
" ( )#
d2 · dk
dθa d·
k(θa )ξ 2
+ k(θa ) + 2ξ +
dξ dθ θa dξ dξ
 ( )2 
2k
dk d2 θa dk dθa d dθ
ξ + +ξ
a ·
dθ θa dξ 2 dθ θa dξ dθ 2 θa dξ
7.4 Control volume method
7.5 Shooting method
7.6 Collocation methods
7.7 Method of weighted residuals

Where observation is concerned, chance favours
only the prepared mind.
— LOUIS PASTEUR
Appendix A
Networked computing
environment within the
chemical and materials
engineering department
A.1 Introduction
In these notes you will be introduced to the networked computing envi-

ronment within the department of chemical and materials engineering at
the University of Alberta. You can choose to work with a variety of work-
stations such as the IBM RS/6000 which use the AIX operating system
(a version of Unix), SUN workstations using SUNOS operating system, or
the personal computers (Prospec 486) which use OS/2 and/or Win95 with
DOS emulation. Hence forth the computers will be identified merely by
the operating system used to run them viz. AIX, SunOS, OS/2 or WIN95
machines. Computers that are available for student use are distributed
as follows in the Chemical and Mineral Engineering (CME) building.
Computers for teaching and research
AIX machines Nine IBM RS 6000 - model 220 in room CME 244 +
One IBM RS 6000 - model 550 as server
SUNOS machines Three SUN workstations in room CME 244.
A.1
A.1. INTRODUCTION A.2
OS/2 machines Twenty Prospec 486 computers in room CME 244
WIN95 machines a variety of PC’s in room CME 474.
Computers for graduate research
AIX machines A variety of IBM RS 6000 machines in room CME

475. These machines belong to individual researchers;
but all staff and graduate students can have access
provided the priority based on ownership is hon-
ored.
You are expected to be familiar with the basic concepts of operating sys-
tems, file systems, editors and structured programming and debugging
principles. I will focus on the concepts that are unique to the networked
environment and discuss briefly the software tools that are available for
various tasks. This document is intended merely as an introduction to
get you started and to identify the hardware and software resources that
are available on the network. Once you are aware of the art of the possi-
ble, you are on your own to make the proper choice of machine, software
tools and implement solutions strategies for solving a variety of prob-
lems. Advanced users should consult the original documentation on
each software. Online help is available for most of the software.
In a heterogeneous, networked environment, such as the one we have,
the range of services such as file space, printer, plotter capabilities, avail-
ability of specialized software etc. are distributed on a variety of ma-
chines, often matching the software needs with hardware capabilities.
For example, graphics packages and word processors on PC’s are ade-
quate for most document processing needs. Powerful editors are avail-
able on all of the machines; you may already have your own favorite
editor on one of these machines. Use it! Data processing, visualization,
and simulation of flow and chemical processes are numerically inten-
sive tasks and are best carried out on powerful workstations. MAPLE
(a symbolic computation package) and APSEN (a process simulator) are
made available only on one IBM RS 6000 machine for licensing reasons.
Hence it is essential to develop the skills to navigate through the network,
transfer data between computers and select the machine best suited for
a task.
Workstations using AIX or SUNOS operating system support multi-
tasking and multiple users. Multi-tasking implies that the computer can
handle several tasks (or processes) at the same time using time sharing
principles. In addition, AIX (and all other flavors of UNIX) can handle
multiple users by keeping track of user directories and accounting in-

formation. Thus, you need to get a userid and some disk space to store
your own files in your home directory.
On the other hand OS/2 supports multi-tasking, but not multiple
users. This implies that whenever you use an OS/2 machine, you can edit
a program in one window and compose a letter in another ( i.e., multi-
tasking feature), but you are responsible for keeping your personal files
separately in a floppy. If you leave them in a hard drive, there is no
guarantee that they will be available to you the next day! But you do not
need a userid to use OS/2 machines.
All of the machines (AIX, SunOS and OS/2) provide a friendly Graph-
ical User Interface (GUI) to interact with the operating system. The basic
element of a GUI is a window and a desktop. AIX uses Motif window man-
ager (a standard that is getting wide acceptance) and OS/2 uses the Pre-
sentation Manager (or PM). The figure A.1 illustrates the basic anatomy
of a window and how to manipulate its size and location. A typical AIX
window on IBM RS 6000 and an OS/2 window on Prospec 486, provide
a shell through which you can enter commands to the operating system.
Well designed applications avoid command line based interaction with
the computer; instead you have to merely (double) click on the icon to
start an application. These icons, of course, have to be made available
to the user and the concept of a desktop is used in both OS/2 and AIX
to organize and present file systems and groups of applications to the
user. Since AIX is a multiuser environment, each user can customize the
desktop. This information is saved for each user and during subsequent
logins you are presented with your own customized desktop. Under
OS/2 there is only one standard desktop administered by the computer
support staff - do not mess with it as it will make it difficult for subse-
quent users to access programs in a standard manner.
While illustrating a dialogue with the computer, following conventions
are adopted throughout this document.
courier font will indicate the prompt from the computer

bold font will indicate a command that you should enter
exactly as shown
italics font will indicate a parameter like a file name, directory
name etc. that should be substituted with the real name.
Unlike DOS, UNIX is case sensitive and most of the UNIX commands are
in lower case.
Click the left mouse

button here to maximize the
Click the left mouse Click the left mouse window size.
button here for a pull down button here to minimize the [Click again to recover the
menu. window (make into icon). original size]
[Click on the icon to recover
the window]
Click the left

mouse button on the
title bar and drag to move
the window
Click anywhere on the

window to make it the current window.
This window will then have the focus of the
keyboard. i.e. anything that you enter from
keyboard goes into this window.
X-windows
background screen.
Click the left Click the left When the cursor is in this
mouse button on a mouse button on a side region, clicking the left or
corner and drag to resize bar and drag to resize the right mouse button offers
the window window choice of additional
menus
Figure A.1: Anatomy of a Motif window

A.2. USERID A.5
A.2 Userid
All students get a userid on the General Purpose Unix servers (or GPU)
maintained by the Computing and Networking Services (CNS). If you also
purchase the Netsurf 97 CD-ROM from CNS (for about $10), you get a
powerful set of software for connecting your home computer to the IN-
TERNET through the University. For more information about computing
and network services see their Web site at http:\\www.ualberta.ca\CNS.
Once you get the userid on the machines gpu.srv.ualberta.ca you can en-
able access to the machines maintained by the department of chemical
and materials engineering by (a) signing on to gpu.srv.ualberta.ca and
(b) running the following script:
[userid@gpu]> /afs/ualberta.ca/labs/cmel244/bin/register-244
You need to do this only once at the begining of the year. If you
encounter any problems see one of the DACS center staff (Mr. Bob Barton
or Mr. Jack Gibeau).
A.3 Overview of the network
Figure A.2 provides a conceptual frame work for introducing the net-
work structure. A variety of workstations and personal computers are
connected by ethernet. This local area network (LAN) is a subnet of the
larger university wide ethernet network as well as the world wide INTER-
NET network. The underlying communication protocol is called TCP/IP
(Transmission Control Protocol/Internet Protocol) and has been widely
accepted as the standard.
In addition to providing some basic connectivity between machines,
a network enables sharing of hardware and software resources as well
as sharing of information and communication on a world wide basis.
We will focus on the departmental subnet to illustrate various concepts
which can then be easily extended to national and international level
networks. The figure A.2 illustrates the logical dependencies for the
purpose of sharing resources between various machines and not the
physical connections. As a user, we need not concern ourselves with
the network hardware connections. It is sufficient to realize that any
machine on the network can address any other machine, much like tele-
phone connections. This immediately requires that each machine on
the INTERNET have a unique IP number and a host name. For exam-
A.3. OVERVIEW OF THE NETWORK A.6
U of A servers provide
HOME NEWS, ftp, Gopher, e-mail
services
U of A modem pool dial 492-0024 ugrads.eche.ualberta.ca

DIAL>telnet machine.eche.ualberta.ca IBM RS/6000 550
AIX 3.2.3
NIS server, NFS server,
do p,
in ft
te
s
X-server/clients
w
w t,
x- lne
x- lne
MATLAB, MAPLE, ASPEN, w t,

te
in ft
FORTRAN, emacs, gnuplot do p,
w
Room: CME 244 s
cmel31.ucs.ualberta.ca
to Room: CME 244
cmel39.ucs.ualberta.ca cmel11.ucs.ualberta.ca
IBM RS/6000 220 to
AIX 3.2.3 cmel28.ucs.ualberta.ca
NIS client to ugrads,
X-server/clients Prospec 486
OS/2
MATLAB, FORTRAN, emacs, X-clients
gnuplot, xmgr, khoros, tex
DOS/Windows emulation,
Lotus 123, Wordperfect
Room: CME 475

ravana.eche.ualberta.ca
do p,
IBM RS/6000 320

in ft
s
w
w t,
x- lne
AIX 3.2.3
te
NIS server, NFS server,

do p,
in ft
s
w
X-server/clients
w t,
x- lne
FORTRAN, C++, emacs,

te
gnuplot, xmgr, khoros, tex,

FLOW3D
Room: CME 475
six Any
IBM RS/6000 machine on
AIX 3.2.3 INTERNET.
NIS client to ravana, Staff offices,
X-server/clients rooms CME 473, 475
FORTRAN, C++, emacs, various labs in CME
gnuplot, xmgr, khoros, tex, building
FLOW3D
Figure A.2: Logical dependencies between various machines in the net-

work
A.3. OVERVIEW OF THE NETWORK A.7
hostname IP number Machine type Operating

system
for undergraduate teaching and course related work
ugrads.labs.ualberta.ca 129.128.44.50 IBM RS 6000 - 550 AIX 4.1
ugrads1.labs.ualberta.ca 129.128.44.51 Sun Sparc Ultra2 SunOS 5.5.1
ugrads2.labs.ualberta.ca 129.128.44.52 Sun Sparc Ultra2 SunOS 5.5.1
cmel31 .ucs.ualberta.ca 129.128.44.31 IBM RS 6000 - 220 AIX 4.1
··· ··· ··· ···
cmel39 .ucs.ualberta.ca 129.128.44.39 IBM RS 6000 - 220 AIX 4.1
cmel11 .ucs.ualberta.ca 129.128.44.11 Prospec 486 OS 2
··· ··· ··· ···
cmel28 .ucs.ualberta.ca 129.128.44.29 Prospec 486 OS 2
for graduate research and faculty use
ravana.eche.ualberta.ca 129.128.56.102 IBM RS 6000 - 320 AIX 4.1
prancer.eche.ualberta.ca 129.128.56.23 IBM RS 6000 - 320 AIX 4.1
comet.eche.ualberta.ca 129.128.56.19 IBM RS 6000 - 375 AIX 4.1
dancer.eche.ualberta.ca 129.128.56.11 IBM RS 6000 - 375 AIX 4.1
cupid.eche.ualberta.ca 129.128.56.82 IBM RS 6000 - 355 AIX 4.1
dina.eche.ualberta.ca 129.128.56.70 IBM RS 6000 - 320 AIX 4.1
jhm2.eche.ualberta.ca 129.128.56.74 IBM RS 6000 - 375 AIX 4.1
chopin.eche.ualberta.ca 129.128.56.11 IBM RS 6000 - 320 AIX 4.1
poincare.eche.ualberta.ca 129.128.56.15 IBM RS 6000 - ??? AIX 4.1
handel.eche.ualberta.ca 129.128.56.54 IBM RS 6000 - 3AT AIX 4.1
brahms.eche.ualberta.ca 129.128.56.13 IBM RS 6000 - 320 AIX 4.1
Table A.1: List of machines in the chemical and materials engineering

network
ple the server within the department of chemical and materials engi-
neering for undergraduate student use has the IP number 129.128.44.50
and the host name ugrads.labs.ualberta.ca. The part of the IP number
"129.128.56." represents the department of chemical and materials engi-
neering subnet and we can connect up to 256 computers to this subnet.
Similarly the last part of the host name, viz. eche.ualberta.ca, also called
domain name represents the chemical and materials engineering do-
main. Different machines on this domain will have different names like,
prancer.eche.ualberta.ca, ravana.eche.ualberta.ca. Table A.1 lists com-
puters that are generally available for chemical and materials engineer-
ing students and staff. To determine your eligibility to have access to
any of these resources see one of the DACS center staff. Typically under-
graduate students can expect to have access to ugrads.eche.ualberta.ca
and the associated machines, while graduate students and staff will have
A.4. THE CLIENT-SERVER MODEL A.8
access to ravana.eche.ualberta.ca and its associated machines. You can

signon to these machines through a variety of means including remote
connections from home. These procedures will be outlined later in this
document.
A.4 The client-server model
In a networked, distributed computing environment, the client-server

model serves a very useful role in both providing a variety of services on
the network and in sharing the limited resources with maximum flexibil-
ity. From a users point of view understanding the client-server concept
helps in (i) navigating through the network and using the resources ef-
ficiently and (ii) diagnosing likely sources of problems when things do
not work as anticipated.
In simple terms a server provides a service to a client authorized to
request such a service. Typically, both the server and the client are pro-
grams that communicate over the network. The server is always running,
listening for requests over the network. Such programs that are running
all the time are also called daemons. When the client program initiates a
request, the server program checks for authentication and provides the
service. Some of the common services are explained below.
A.4.1 File servers

Let us take an example of Network File System, often called NFS. It is a set
of protocols developed by Sun Microsystems and it has been accepted
widely as a standard. It uses TCP/IP for communication, but provides a
higher level function in terms of sharing files between various machines
in a heterogeneous environment.
In figure A.3, the computer chopin.eche.ualberta.ca has a local file
system called /usr/local which contains a variety of very useful programs
like emacs, gnuplot, TeX, xmgr, pine etc. This file system is quite large
with about 450 Megabytes of programs and data that can be used by
other similar machines. So this file system is "exported" to group of
machines within the network. Thus chopin.eche.ualberta.ca acts as the
NFS server to this group of machines. The client machines then "mount"
this file system and make it appear as a local file system. In this way
the user sees the same file structure on each of the machine and the
administrator has to install and maintain only one copy of the software.
In a similar fashion, the authorization to signon to a group of ma-
chines ( i.e., user ids and passwords) is also maintained in a central loca-
ethernet
chopin.eche.ualberta.ca
ravana prancer comet
NFS clients
/usr/local
The file system chopin:/usr/local
NFS SERVER is mounted as /usr/local on each of
these machines. Logically it appears as a local
The file system /usr/local file system. But any directories and files under
resides physically in chopin /usr/local are actually fetched from chopin.
but has been exported to a group
of machines
Some useful Unix commands

df - Reports information about space on file systems
showmount - Displays a list of clients that have remotely mounted file systems
exportfs - Exports and unexports directories to NFS clients
mount - Makes a file system available for use
du - Summarizes disk usage
Figure A.3: Network File System server and clients
tion and served by the NIS (Network Information Service) server machine.
Figure A.4 shows a such a setup.
The Andrew File System or AFS
The University of Alberta supports sharing of the same file space from a
variety of machines on campus. There is a single home directory for
each userid from any UNIX machine on campus. CNS provides lim-
ited amount of disk space for each user (about 10Meg). In addition
the department of chemical and materials engineering provides addi-
tional space to undergraduate students under a directory called lab-
disk and for graduate students under a directory called chemeng. The
relevant directories will be found as a subdirectory under the home
directory of each user. To learn more about AFS go to the web site
http://www.ualberta.ca/HELP/afs/afs1.html.
A.4.2 License servers

The client-server model is also used to manage software licenses. Due
to financial constraints, unrestricted licenses are rarely purchased. Typ-
prancer brahms chopin
ethernet
ravana.eche.ualberta.ca
jhm2 dina comet
NIS SERVER
serves information on NIS clients
/home user ids and passwords When you attempt to signon to a client, it verifies
to other clients. with the NIS server that your user id is a valid one.
Your home directory is also made available on each client.
The file system /home
contains home directories of users.
It has been exported to other machines.

passwd - Changes a user’s password [it may take a while to propagate to
other clients]
rlogin - Connects the local host with a remote host.
rsh - Executes the specified command at the remote host
Figure A.4: Network Information Service

ically each software is licensed to run only on a specific group of ma-

chines. We also rely heavily on a large collection of freely available soft-
ware. Two of the common licensing arrangements are floating network
license and node locked license. The former allows a fixed number of
concurrent users of a software on any machine in the network, while the
later allows any number of users on a single machine.
Typically a license server (which is a small program) runs on one of
the machines in the network as a daemon. Let us use FLOW3D and ASPEN
PLUS as examples. FLOW3D is a commercial software package for use
in fluid dynamics simulations and ASPEN PLUS is a very powerful steady
state process flow sheet simulator. For managing FLOW3D, the license
server (two programs named lmgrd and CFDSd) runs all the time on blue-
jay.ucs.ualberta.ca. Any time FLOW3D is started from any of the clients
within the network, the client contacts the server to check whether it
has the permission to start a new copy of the program. The server can
either grant the permission or deny it because that particular client is
not allowed to run that software or all of the available licenses are in use
at that time. For FLOW3D we have license for 100 concurrent users.
For ASPEN PLUS, the department of chemical and materials engineer-
ing has a multiuser, node locked license. This implies that any number
of users can use the program at the same time, but the software can
run on only one machine. In our case ASPEN PLUS runs on ugrads1 or
ugrads2.
For MATLAB the university has a site license for up to 150 concurrent
users. The most recent version of MATLAB 5.0 is available on ugrads1
or ugrads2.
Table A.2 lists all of the software available on the undergraduate and
graduate research network and the licensing status of each software.
A.4.3 X servers
X-windows is a sophisticated communication/window management soft-
ware developed at MIT. It uses TCP/IP for communication. It allows you
to interact with the local workstation through windows, menus, dialog
boxes, icons etc., in an intuitive way and minimizes the need for interac-
tion through command lines. One can view this as the next logical step
in the historical progression of the way humans have interacted with
computers. Originally it was accomplished through batch processing,
followed by interactive computing using commands which is being re-
placed currently by more intuitive interaction through a Graphical User
Interface (GUI) using menus and icons. One can expect this to be fol-
lowed by voice level interaction, leading ultimately to interaction with
Software ugrad-net grad-net Windows Comments

ASPEN PLUS ugrads1 X process simulator
floating, multiuser license
MAPLE num.srv X symbolic computation package
floating, multiuser license
MATLAB ugrads1 prancer X powerful linear algebra package
cmel31 - comet with numerous tool boxes
cmel39 floating, multiuser license
FORTRAN all AIX machines IBM HESC license agreement
xde all AIX machines X powerful debugger
IBM HESC license agreement
emacs all AIX machines a powerful editor
GNU public license
gnuplot all AIX machines a simple 2-D graphics package
GNU public license
tgif all AIX machines X a powerful drawing package
free to use license
pine all AIX machines a standard e-mail program
free to use license
tin all AIX machines a standard news reader
free to use license
TeX all AIX machines a powerful typesetter
free to use license
123 OS/2 Windows 3.1 spread sheet
Freelance OS/2 Windows 3.1 Graphics
Wordperfect OS/2 Windows 3.1 word processor
HYSIM OS/2 DOS process simulator
Table A.2: List of software on the chemical and materials engineering

network
machines using natural language!

While Microsoft (Windows, Windows NT), Apple (MacIntosh), NeXT
(NeXTSTEP), IBM (OS2-PM) and others have their own Graphical User In-
terface to interact with their operating system and application programs,
X-Windows goes one step further, in making the GUI software indepen-
dent of the underlying hardware and operating system. How this is ac-
complished is of interest only to programmers. From a users point of
view we should understand the concepts and be able to reap the benefits
of these features. In particular, X-windows makes the following possible:
• Signon from a local machine to a remote machine. The two ma-

chines can be either of the same make or completely different, e.g.,
a 486 running under OS/2 can be the local machine and IBM RS 6000
can be the remote machine.
• Run an application software on the remote machine e.g., MATLAB,

ASPEN PLUS, FLOW3D etc. , but have the X-windows display routed
to your local machine. In other words you can interact with a re-
mote computer, located as far away as in another building or an-
other continent, with full window capabilities! (One needs a high
speed link to interact with a computer in another continent, of
course)
A conceptual description of the client-server interaction and the steps

required to establish this interaction between a local and a remote com-
puter using X-windows are illustrated in figure A.5. Note that under
AIX the server program is called xinit and the client program is called
aixterm. Under SunOS, the client program is called xterm.
A.4.4 News server
The INTERNET network provides a news service. Discussions and ex-

change of ideas take place through organized and moderated news groups.
It is more like a conference on thousands of topics. You can choose to
participate in topics of your choice. The topics are as wide ranging as
the ones on classical music (rec.music.classical), MATLAB (comp.soft-
sys.matlab), the AIX operating system (comp.unix.aix), dynamical sys-
tems theory (comp.theory.dynamic-sys) etc. The server is maintained
and administered by Computing and Network Services. The client pro-
gram is called a news reader and several versions of news readers exist.
On the AIX machines, the news reader can be invoked with the command
NOTE: The string
user@machine:dir
is a prompt from
the computer
Note: The client machine need not

even have a monitor attached to it.
It merely runs the applications and
local sends the display to the local
X -server (xinit) machine.
remote
This machine must have a monitor,
X client (aixterm)
keyboard and mouse.
ethernet
Command sequence What they do
1. login signon to the local machine with user id and password
2. xinit start the X-server
3. xhost remote permit remote machine to send display to local
4. rlogin remote signon to the remote machine from local
5. aixterm -display local:0 & start X-client on remote and redirect the display to local.
NOTE: Substitute the actual name of the machine for local and remote. Use fully qualified
domain names if the machines are on different subnet.

xinit - starts the X-server on the local machine
xhost - permit remote host to send display to local machine
aixterm - start a X-terminal (or X-client) and send the display to the
machine local.
Some useful Unix files to explore

If these files are found in your home directory, they are used in controling rlogin
between machines and in customizing your X-window environment.
.rhosts - controls rlogin between machines
.xinitrc - customizes the X-server.
.Xdefaults - customizes the Window appearance and functionality
.mwmrc - customizes the Motif window manager
Figure A.5: X-Windows client-server model

user@machine:dir> tin
This program maintains a local database of the news groups that you
have subscribed. It fetches the latest articles from the server and allows
you to read and, optionally respond to articles. If you post a response,
it goes to the whole internet community. This service is a privilege and
should be used responsibly. Any abuse of the system will result in denial
of access to the entire network services.
A.4.5 Gopher service
While the News service allows two-way communication and thus active
participation, Gopher is a worldwide one-way information retrieval ser-
vice. A number of Universities and organizations participate in provid-
ing this service. The University of Alberta is one of the participants. The
Gopher server is maintained by Computing and Network Services on one
of their workstations. Any client on the campus ethernet (or even from
a home computer with a modem) can connect to this server and browse
through a wealth of information on such topics as
1. What is CWIS?/
2. What’s New on CWIS?/
3. Student Information and Services/
4. Libraries on the Network/
5. University Faculties and Departments/
6. Administrative Policies and Procedures (MAPPS)/
7. Campus Phone Directory <CSO>
8. Computing Resources/
9. International CWIS Systems/
10. Magazines and Publications/
11. University Services Directory/
12. Weather Report.
The Student Information and Services includes job posting, and

the International CWIS Systems gives access to other Gopher ser-
vices around the world. There are also nodes that link you to the Library
of Congress and other university libraries. To connect to Gopher from
any computer on campus with telnet feature enter,
user@machine:dir> telnet gopher.srv.ualberta.ca
Use as your userid gopherc; no password is required. When you connect

A.5. COMMUNICATION A.16
to the campus network from home through a modem by dialing 492-

0024, CWIS will be provided as a menu option and selecting it will con-
nect you to the Gopher server. When you are logged on to ugrads.ucs.ualberta.ca,
you can also access gopher by entering
user@machine:dir> gopher
When you are logged on to gpu.srv.ualberta.ca, you can also access go-
pher or world wide web by entering
user@machine:dir> lynx
and selecting the appropriate menu.
A.4.6 World wide web service
This is a high end version of information sharing tool (like gopher) us-
ing windows and hypertext links established between computers on the
internet on a global level. It also operates on a client-server model.
Netscape is available on all of the AIX machines. Anyone with some
information to share can operate a server on one of their workstations
and there are thousands of such servers on the INTERNET. To access
information on WWW, start the client on your local workstation by en-
tering the following after you have started X-windows.
user@machine:dir> netscape
Your client should automatically know where the nearest server is lo-
cated. Typically there will be home page from where you can begin your
exploration. Have fun! Watch the time you spend on this one!! It is a
time sink!!!
A.5 Communication
A.5.1 Software tools
There are many communication software packages available to connect

from one computer to another. The following are available for use on
the Chemical And Materials Engineering network.
kermit Preferred for communication from home computer to Chemi-

cal And Materials Engineering network via a modem. It allows
for VT100 emulation ( i.e., full screen use of editors like emacs,
news reader tin, e-mail program pine and gopher) and tek-
tronix emulation ( i.e., plotting programs like gnuplot can dis-
play graphs on your home computer in color!) and file trans-
fer between home computer and the university computer. You
can get a copy of MS-KERMIT Ver 3.10 for your PC from DACS
center staff. It can be distributed freely without any licensing
problems. Additional details can be found in Gianone (1991).
telnet Preferred for communication between UNIX machines or from

OS/2 to UNIX machines over ethernet. It supports VT100 em-
ulation, but has no graphic or file transfer support. Only com-
mand line interaction with the remote machine is possible.
ftp Preferred for file transfer over ethernet. Both AIX and OS/2
provide support for ftp.
tn3270 Preferred for ethernet communication from DOS based PC’s

(in room CME 473, staff offices etc. ) to workstations. It pro-
vides VT100 and tektronix emulation as well as ftp support.
aixterm Preferred for remote login between AIX machines or from

OS/2 machines to AIX machines. It supports full GUI features
with the remote machine.
A.5.2 Connection to the AIX machines from OS/2 machines

If you are using a 486 PC under OS/2 in room CME 244, the following
steps will establish a X-windows connection to a remote IBM RS 6000
running AIX.
• Make sure that the X-server is running under OS/2. This server is
called presentation manager X or PMX server and, under normal
operating conditions, it should be active and running. To verify
that it is running, look for the PMX icon on the screen. If you do
not find one, the key sequence ctrl Esc will give a list of windows.
Check if PMX is listed. If you do not find it, you can start a X-server
as follows:
> Open a OS/2 Full Screen Window by double clicking on the

appropriate icon.
> To start the PMX server, enter If PMX server is
already running,
[C:\] xinit you will be
informed of that
fact.
• Open a OS/2 Full Screen Window, if one is not open already.
• Telnet to one of the IBM RS 6000 machines - e.g., cmel34.ucs.ualberta.ca
[C:\] telnet cmel34.ucs.ualberta.ca
Remember that you can choose any one of the 9 AIX machines in
the network. You will see the same home directory and other soft-
ware resources from each of the nine machines. Hence choose a
random number between 31 and 39, so that no one machine is over
loaded all the time.
• Signon to the remote AIX machine.
login: userid
Password: password
• Start the aixterm client on the remote machine, redirecting the dis-
play to the local machine by entering
user@machine:dir> open-Xpc nn
where nn is the local station number of the Prospec 486 running

OS/2. This effectively tells the AIX machine (in this example cmel34)
to open a X-Window and display the window on the local worksta-
tion (i.e. cmelnn). Alternatively, you can enter,
user@machine:dir> nohup aixterm -display cmelnn.ucs:0 & Observe that the

computer prompt
NOTE: open-Xpc is nothing but a script that executes the above user@machine:dir>
identifies your
command for you!
userid, the
• Exit from the AIX machine and from the OS/2 Full Screen Window. machine name and
i.e. exit twice! the present
directory.
user@machine:dir> exit
[C:\] exit
• After some delay, an AIX window will open on your local 486 sta-
tion! Now you are connected to the AIX machine and typically you
will be in a shell called “ksh”. You can start any application like
MATLAB, xmgr etc. by simply entering the name of the program
e.g.,
user@machine:dir> matlab
If an application, such as ASPEN PLUS, uses X-windows effectively,

it might open up other windows, dialogue boxes etc. and also pro-
vide online help windows. To start ASPEN PLUS enter
user@machine:dir> mmg
which is the graphical version of the model manager.
• Before you leave the OS/2 station, either logout from the remote
machine with the logout command as follows
user@machine:dir> logout You can also use

ctrl-D key to logout
or close the aixterm window.
Caution: If you leave the OS/2 station without completing this last
step, the next person using that OS/2 station will have access to
your account on the AIX machine!
A.5.3 Connection to the AIX machines from a home computer

Use of Kermit is recommended. You also need a modem. Kermit has an
initialization file which can be setup in such a way that it automatically
dials the telephone number. The available phone numbers to connect
to the University of Alberta network server are : 492-4811 or 492-0024
(2400 Baud) , 492-0096 (9600 Baud) or 492-3214 (high speed modem).
Sample initialization files can be obtained from DACS center staff. You
start the kermit program by entering
C:> kermit -finitfile.ini
where initfile.ini is the name of the initialization file. By default it is

mskermit.ini. If no file by name mskermit.ini exists in your directory and
you do not specify explicitly the name of an initialization file as shown
above with the -f flag, then the Kermit program will start on the PC, but
you will be left with the prompt
MS-KERMIT>
At this stage you can ask for additional help on kermit by entering the
kermit command help or ?. But the connection to the University network
must be done manually with the following commands
MSKERMIT>set speed 2400 ;this sets the modem speed

MSKERMIT>set port com2 ;this sets the port
MSKERMIT>OUTPUT ATDT4924811 ;this dials the number
MSKERMIT>connect ;this connects you to the Univ. network
These steps can be automated by putting the above kermit commands

in the initialization file. Note that anything following a semi-colon is
treated as a comment by MS-KERMIT. Whichever procedure you use, you
should get the following prompt if everything has gone well up to this
stage.
DIAL>
At this stage enter
telnet, 129.128.44.50
or
telnet, ugrads.eche.ualberta.ca
to get connected to the AIX machine or any other valid internet number
(or name) for which you have a valid userid.
A.5.4 Connection to the AIX machine from a DOS machine using

tn3270
All the PCs in room CME 473 and those in staff offices that have ethernet
cards are connected to INTERNET. The communication program tn3270
is also available in each of the machine. On these machines you can start
the connection by entering,
[C:\] tn3270 remote.machine Substitute the

actual machine
You should be connected directly to the remote AIX machine. This com- name.
munication program is quite powerful. It provides full VT 100 emulation
and hence allows full screen use of editors (emacs, vi), news reader (tin)
etc. It also supports tektronix emulation and has facilities for capturing
graphics screen images as Postscript files on the local PC. Use the Alt-H
key sequence to get brief online help.
A.5.5 File transfer with Kermit

A client-server concept is used once again to transfer files between the
remote IBM RS 6000 and the local home computer. The MS-KERMIT on
your PC is the client. You must start a server on the IBM RS 6000, called
the C-KERMIT. Once you are logged in to the IBM RS 6000, enter
user@machine:dir> kermit
to start the C-KERMIT program. The prompt string should change to
C-Kermit>
Enter,
C-Kermit> server This command is

for C-Kermit
to put C-KERMIT in server mode. Then escape back to the local MS-
KERMIT (Typically ctrl-[C is the escape sequence. The prompt should
be,
MS-Kermit>
Any command you enter now is acted up on by the local PC. To fetch a
file from the AIX machine (in your home directory, of course) to local PC,
enter,
MS-Kermit> get filename This command is

for MS-Kermit
To send a file from your local PC to the AIX machine (in your home
directory, of course) enter,
MS-Kermit> send filename
Once the file transfer is completed, enter Although this

command is for
MS-Kermit> fin MS-Kermit, it tells
C-Kermit to
to signal the server that file transfer session is finished. Then enter, terminate server
mode. The next
MS-Kermit> C command, C,
re-establishes
direct connection
to re-connect to the AIX machine. Note that the C-Kermit is still running to AIX
on the AIX; it has only terminated its server mode. Finally enter
FTP subcommand its function

help ask for help on FTP subcommands
get remotefname localfname get a file from the remote machine
mget remote-pattern get multiple files matching the pattern
put localfname remotefname put a file on the remote machine
mput local-pattern send multiple files matching the pattern
bin to enable file transfer in binary mode
ascii to enable file transfer in ascii mode (this is the default)
ls list files in the current directory on the remote machine
dir also list files in the current directory on the remote machine
cd change the directory on the remote machine
lcd change the directory on the local machine
pwd display present working directory on remote machine
quit to terminate the FTP session
Table A.3: List of ftp commands
C-Kermit> quit The quit command

is acted on by C-
to terminate the C-KERMIT program on the AIX machine and to return Kermit
to the shell level.
A.5.6 File transfer with ftp
The file transfer program called ftp allows transfer of files between ma-
chines on ethernet. This is also implemented on the client-server model.
From a local machine connect to the remote machine as follows:
user@machine:dir> ftp 129.128.44.50
On the remote machine the ftp daemon called ftpd acts as the server.
You will be prompted for userid and password. Once the connection is
established you can use the commands shown in Table A.3 to transfer
file.
The FTP procedure summarized above is essentially the same on
most of the machines. Be bold and try them out and observe how fast
the file transfer is compared to KERMIT through a serial line.
A.6. OPERATING SYSTEMS A.23
A.5.7 File transfer from AIX to OS/2 machines

ftp feature is also available under OS/2 machine. If you want to make a
copy of your personal files on the AIX machine to a floppy disk, do the
following.
• Make sure that you have a writable, formatted floppy disk in drive
A: of an OS/2 machine.
• Open a OS/2 Full Screen Window, if one is not open already.
• Enter,
[C:\] ftp remote.aix.machine Substitute the

actual machine
name.
• Signon as usual.
• Change the local directory to A: using
[C:\] lcd A:
• Transfer the file using,
[C:\] get remote.filename Substitute the

actual file name.
A.6 Operating systems
A.6.1 How to signon and logout

The first time you signon to an AIX machine you will be forced to change
your password. Subsequent times, you will be given informational mes-
sage regarding your previous singon time and date and any unsuccess-
ful attempts to singon to your id. Change your passwords periodically.
Make sure that you always logout before you leave your workstation.
The following commands summarize the syntax for loging in and out
of AIX systems and for changing passwords.
Login procedure
login: userid Substitute the

password: password actual userid and
password.
Logout procedure
user@machine:dir> logout You can also use

ctrl-D key to logout
Changing password
user@machine:dir> passwd Respond to the

prompts for the
old and new
passwords
A.6.2 Customizing your environment - the .profile file
Once you login successfully to an AIX machine, you will always start the
session in your home directory. There should be a file in your home
directory named .profile. Every time you login, the contents of this file
are executed once. Hence one can use this file to customize the work-
ing environment on an AIX machine with the help of this file. There is
also a system wide profile file, which is used to control such things as
the default search path for finding executable programs, controlling the
prompt string, identifying the terminal type, and to define a number of
environment variables that other application programs might need.
You can feel free to copy my version of this file in /u/kumar/.profile
and adopt it to your needs. My version of this file enables the command
line editing features - i.e., all the commands that you enter during a ses-
sion are stored in a buffer and you can scroll back and forth to retrieve
previous commands with Ctrl-p for previous and Ctrl-n for next. After
retrieving a previous command you can edit it with the cursor movement
keys Ctrl-b for backward, Ctrl-f forward and Ctrl-d for deleting the cur-
rent character. In effect it supports the same editing capabilities on the
command line that the editor emacs supports for a file. More on emacs
later. You can also set alias for the most frequently used commands.
For example you can set
alias dir=’ls -al’
so that when you enter dir you will get the directory listing. Another
useful alias is
alias rm=’rm -i’
which prompts you for confirmation before removing (deleting) files.
A.6.3 File management
If you are familiar with the DOS directory and file structure, it should be
equally easy to work with the file systems of OS/2 and AIX. While DOS file
names are restricted to 12 characters, AIX allows very long file names.
But the mechanisms for creating directories and navigating up and down
the directory tree structure are essentially the same. Frequently used
commands that relate to file management are tabulated below.
Task UNIX OS/2 DOS 5.0 or above

seek online help man command help command help command
list the directory ls -al dir dir
list contents of a file cat fname type fname type fname
list one page at a time more fname more fname more < fname
create a file touch fname
erase a file rm fname erase fname erase fname
copy a file cp fn1 fn2 copy fn1 fn2 copy fn1 fn2
append fn1 to fn2 cat fn1 >> fn2 copy fn2+fn1
rename a file mv fn1 fn2 rename fn1 fn2 rename fn1 fn2
create a directory mkdir dirname mkdir dirname mkdir dirname
change directory cd dirname cd dirname cd dirname
delete an empty directory rmdir dirname rmdir dirname rmdir dirname
present working directory pwd cd prompt pg
disk usage summary du -s chkdsk
status of system ps -l pstat
A.6.4 How to get online help

Each of these commands, particularly in UNIX, can take several optional
parameters, flags etc. that further identify any specific features of the
command that you want to enable. This command list is by no means
complete. They are the more frequently used commands. On UNIX addi-
tional details on each of the command can be found with the command
user@machine:dir> man command
These are called manual pages or man pages for short. Try
user@machine:dir> man man
to get started! Although the man pages provide only limited help, they
are always available from any type of terminal. Much more exhaustive
online help using hypertext is available when you are connected to an
AIX machine via X-windows. To access this enter
user@machine:dir> info
and follow the instructions. Extensive help based on hypertext is also

A.7. EDITORS A.26
available on OS/2 machines.
A.7 Editors
The functions that one expects from a good editor have several common
features. These can be broadly grouped into the following.
cursor movements by characters, words, lines, blocks, etc.

delete text by characters, words, lines, blocks, etc.
insert text by characters, words, lines, blocks, etc.
move text by characters, words, lines, blocks, etc.
copy text by characters, words, lines, blocks, etc.
locate text by strings perhaps with regular expressions
search & replace text1 by text2
A good editor must have a set of commands or keyboard sequences

that invoke functions to carry out the above tasks. On AIX there are
several editors. I prefer emacs, as I find it to be extremely powerful; it
also has built in online help and hence you can learn at your own pace
and grow with it. OS/2 also has a full screen visual editor. I am sure that
you will have your own favorite editor. If you have one use it and ignore
this section.
A.7.1 Emacs - the ultimate in editors

If you want to learn the ultimate in editors (my biased view of course!)
try emacs. To start this editor in AIX use the command,
user@machine:dir> emacs
The anatomy of the emacs editor screen is shown in Figure A.6. This
is also a full screen, visual editor. It works under both X-windows and
VT100 emulation.
Buffers
Emacs uses the concept of a buffer to keep a temporary copy of the file
that you are editing. You can edit any number of files at a time and each
A.7. EDITORS A.27
Indicates that the file is read only.

emacs mode relative file position
Buffer name
%%
C-p
Mini buffer where

commands to emacs are entered. cursor
C-b movements C-f
** in this position indicates
that the file has been modified.
C-n
Figure A.6: Anatomy of an emacs window

A.7. EDITORS A.28
file is kept in a separate buffer. There is a one line command buffer called
minibuffer at the bottom of the screen and a text buffer at the top. A
status line in reverse video separates the minibuffer from the text buffer.
The key sequence C-x b allows you to cycle through the buffers.
Key sequences
Being the most powerful editor, emacs provides a large number of edit-
ing functions. These functions are accessed either by entering the func-
tion name in the command minibuffer or by directly entering a key se-
quence. This process of attaching key sequences to functions is called
key bindings. Actually when you enter a key sequence like C-f, it invokes
a function called forward-char. Some of the frequently used functions
and their key bindings are listed in Table A.7.1. In these C-x means the
keeping the control key pressed, enter the character x, while M-x implies
a two-character key sequence with the Esc key as the first one followed
by the character x. If you get into trouble or become confused with this
editor at any time enter the key sequence C-g C-g to discontinue what
you started. This means that while keeping the control key down, enter
the letter g couple of times.
Command completion
If you find the use of key sequences difficult to remember,but would

rather remember only the function names (and that too only vaguely!),
then you can enter M-x followed by the first few letters of the function
name and the space-bar key. A list of all possible functions beginning
with those few letters will be displayed in a lower window. This process
is called function completion. Try
M-x sa space-bar
You will see that the function name will be completed till save- and
stop as an ambiguity exists at the point. Another space-bar will show all
functions that begin with save- and you can continue to identify the one
you want - e.g., b space-bar will identify the function as save-buffer.
Recall that this function is bound to the key C-x C-s.
This concept of completion is also used in selecting a file. For exam-
ple, the key sequence C-x C-f will get a new file into a buffer for editing.
Once you execute the key sequence, the current directory will be dis-
played in the minibuffer. Entering the first few characters of a file name
followed by space-bar will complete the file name till the next point of
ambiguity.
A.7. EDITORS A.29
task key sequence function-name

ask for help C-h C-h help-for-help
ask for tutorial C-h t help-with-tutorial
ask for information C-h i info
to end emacs session C-x C-c save-buffers-kill-emacs
abort if you get into trouble C-g keyboard-quit
edit a new file in a new buffer C-x C-f find-file

insert a new file at cursor position C-x i insert-file
switch to a different buffer C-x b switch-to-buffer
save the current buffer into the file C-x C-s save-buffer
save the current buffer as a new file C-x C-w write-file
move one character backward C-b backward-char

move one character forward C-f forward-char
move to previous line C-p previous-line
move to next line C-n next-line
delete the current character C-d delete-char
undo C-x u advertised-undo
move to beginning of line C-a beginning-of-line

move to end of line C-e end-of-line
open a new line for typing C-o open-line
kill from cursor to end of line C-k kill-line
yank it back (explore kill ring!) C-y yank
move to beginning of buffer C-x [ backward-page
move to end of buffer C-x ] forward-page
page forward to next screen C-v scroll-up

page backward to previous screen M-v scroll-down
redisplay current screen C-l recenter
incremental search forward C-s string isearch-forward

incremental search backward C-r isearch-backward
search & replace Mx % query-replace
to cancel search C-g C-g
start recording keystrokes C-x ( start-kbd-macro

end recording keystrokes C-x ) end-kbd-macro
replay/execute recorded keystrokes C-x e call-last-kbd-macro
NOTE: C-h means keeping the control key down, enter the key “h”.
M-v means press and release the Meta key or the Esc key, then press the
key “v”.
C-x [ means keeping the control key down, enter the key “x”, then press
the key “[”
Table A.4: List of frequently used emacs functions and their key bindings
A.8. FORTRAN COMPILERS A.30
Emacs contains many more functions than listed in Table A.7.1 and
not all functions have key bindings. You are encouraged to go through
the online tutorial which can be invoked with the keystrokes C-h t. Online
documentation is available via C-h C-h i. The key sequence C-h C-h b
will show all the key bindings within emacs in a buffer called *Help*
after splitting the screen into two windows. You can switch to the lower
window with the key sequence C-x o which is the same as invoking the
function other-window. Repeating the key sequence C-x o will cycle you
through the various windows. To expand the current buffer into full
screen use the key sequence C-x 1.
This is only a minimal introduction to emacs. This editor is not for
the uninitiated user. You need patience to master this editor, but if you
persist the rewards in terms of increased productivity are great. With
this introduction you should be able to explore emacs deeper and deeper!
If you are bored while using emacs, chat with the doctor - use M-x doctor
- have fun.
A.8 Fortran compilers
Fortran compilers and X-window based debuggers are available on all of

the AIX machines. The FORTRAN compiler on the AIX server is named
xlf. To compile your code use the command
user@machine:dir> xlf srcfile.f -O
This is adequate for any self contained FORTRAN program. The -O op-
tion invokes optimization of the code. The executable code from the
compiler will automatically be stored in a file named a.out. To execute
your program use,
user@machine:dir> a.out
If you read data from unit 5 and write data to unit 6 within your FOR-
TRAN program, these I/O will be redirected, by default, to your terminal.
If you want to use files connected to these units, then you can use
user@machine:dir> a.out <indata >outdata
Another way to use these files is to OPEN then explicitly within your
FORTRAN program.
If your program calls any library routines like the nswc math library,
then you must specify the location of the library as follows in the com-
A.9. DEBUGGERS A.31
pilation step.
user@machine:dir> xlf srcfile.f -O -lnswc -L/usr/local/lib
Here -l parameter identifies the name of the file containing the library
and the -L parameter identifies the directory where the library file re-
sides. Note that the library file will be named libnswc.a. By convention,
however, the prefix lib and the postfix .a need not be specified in the
parameter -l.
Use the man xlf to find about other parameters the xlf compiler can
accept.
A.9 Debuggers
A source level debugger allows one to step through the program, exe-
cuting one line at a time, set break points at pre-selected lines, display
values of variables and change values of variables. There is an excellent
X-Windows based front end called xde to the debugger (dbx) on the RISC
server that is quite easy to use. Find out more about it using man xde. If
you want to use this debugger, however, you should access the computer
through X-Windows.
In order to enable the debugging features you should compile the
program with the -g option as follows.
user@machine:dir> xlf srcfile.f -g
The a.out file so produced will contain all the symbol table information
which is needed by the debugger. To invoke the debugger use,
user@machine:dir> xde a.out
This will open up several windows, one of which will show the source
listing. Other windows provide menus and buttons that you can use
to set break points, begin execution, display variables etc. There is also
online help explaining the various features of xde.
For the real computer hacks, a command line version of the debugger
called dbx is available. You can use this to debug FORTRAN, C and
PASCAL programs. Venturing into this is recommended only if you know
the language C quite well. This can be accessed, however, without X-
windows.
A.10. APPLICATION PROGRAMS A.32
A.10 Application programs
There are a large number of application programs on the AIX machines.

They are all located under the directory "/usr/local". The procedure for
starting several useful applications is given in the next few sections.
Online help or online documentation is often adequate to learn more
about the software.
A.10.1 ASPEN PLUS

APSEN is a powerful steady state process flow sheet simulator. It is
useful for carrying out rigorous, steady state mass and energy balance
calculations. It has a fairly extensive thermodynamic data base. The X-
window based front end, called the model manager allows you to define
the problem interactively. The flow sheet can be constructed graphi-
cally by grabbing modules and connecting them up. Operating condi-
tions can be defined interactively by filling out forms. The expert sys-
tem interface ensures that all required input parameters are specified
before the simulation is started. ASPEN PLUS is licensed to run only on
ugrads.eche.ualberta.ca.
• Establish a X-window connection to ugrads from any OS/2 machine
(follow section A.5.2) or from other AIX machines (see figure A.5).
• From the command line enter,
user@machine:dir> mmg
to start the model manager. Explore more on your own!

Advanced users of ASPEN PLUS can use the simulator directly from the
command line using
user@machine:dir> aspen
The input file containing the commands to ASPEN PLUS must be pre-
pared by the user using any standard editor. One can think of the Model
Manager as a front end that enables you to build the command file for
ASPEN PLUS in a GUI environment.
A.10.2 Xmgr
This is a powerful 2-D plotting and data analysis package. You can con-
trol every facet of the graph with this program. It runs only under X-
A.10. APPLICATION PROGRAMS A.33
windows. This program is available on most of the AIX machines. To

start the program
• Establish a X-window connection to ugrads from any OS/2 machine
(follow section A.5.2) or from other AIX machines (see figure A.5).
• From the command line enter,
user@machine:dir> xmgr
to open up a main window. Explore more on your own!
A.10.3 TEX
TEX is a powerful typesetting package that is widely available on several
platforms and it has very few licensing restrictions. This document, in
fact, was typeset using TEX. It is available on most of the AIX machines.
The authoritative document on TEX is by Knuth (Knuth, 1984) and on
LATEX is by Lamport (Lamport, 1986). The steps for compiling, previewing
and printing a TEX document are outlined here. See the above references
on how to create a TEX document. To compile a TEX file use,
user@machine:dir> tex file.tex
To compile a LATEX file use,
user@machine:dir> latex file.tex
Either of these steps should produce a file named file.dvi, which is a

device independent file. The contents of this "dvi" file can then be used
to either preview the document on the screen or send it to a printer. To
preview use,
user@machine:dir> xdvi.sh file.dvi
Note that on the AIX machine, you must be using X-windows to use the
previewer. A program called dvips takes the "dvi" file and sends it to
a postscript printer. Since printer configurations vary, see one of the
support staff to find out the exact procedure for printing a document.
A.10.4 pine - the mail reader

Pine is the standard mail reader on our AIX network. To start it simply
enter
A.11. DISTRIBUTED QUEUEING SYSTEM A.34
user@machine:dir> pine
This program will work under VT100 emulation and offer full screen
support. Online help can be accessed with the question mark, ?. Note
that e-mail addresses are formed as userid@machine.eche.ualberta.ca.
Here userid is your signon name or id on the AIX machine, machine is the
name of the machine. The rest of the e-mail address, eche.ualberta.ca
refers to the chemical and materials engineering subnet domain name.
If you have accounts on several machines it is recommended that you
select and consistently use one machine as your primary e-mail system.
A.10.5 tin - the news reader
Tin is the standard news reader on our AIX network. To start it simply
enter
user@machine:dir> tin
This program will work under VT100 emulation. Online help can be
accessed with the "h" key.
A.11 Distributed Queueing System
The Distributed Queueing System (DQS) allows CPU intensive tasks to

be executed on a machine that is relatively free within the graduate net-
work, while at the same time allowing faster keyboard/terminal response
to interactive users. Any task that takes longer that 5 min of CPU time
should be submitted through the DQS. Otherwise the jobs will be termi-
nated automatically. The commands that you should be familiar with
are
• qsub - for submitting batch jobs.
• qstat - to monitor the status of jobs in the DQS
• qmon - under X-windows to monitor status of jobs.
• qdel - to delete your own (and only your) jobs.
For the experts, online documentation via man pages is available on each
of the above commands. The following examples illustrate how to con-
struct a script file that you would submit using qsub.
A.11. DISTRIBUTED QUEUEING SYSTEM A.35
A.11.1 Example of a CPU intensive MATLAB job
Construct a script file named, say test.bat, containing the following

lines.
#!/bin/csh
# make the current directory the CWD
#$ -cwd
# lets put STDOUT/STDERR in the file "gaga"
#$ -eo gaga
# i’d like to know when she fires
#$ -mu user@address.ualberta.ca
#$ -mb
# and when she finishes
#$ -me
matlab >out « ’eof’ #the shell starts matlab; output goes to "out"
secant(’ass3a’,[0,1],eps,1) %matlab acts on this line and runs secant
fzero(’ass3a’,0.5,eps,1) %matlab acts on this line and runs fzero
quit %matlab acts on this line and quits
eof
ps -ael #back in the shell; output goes to "gaga"
ls -al #The shell acts on this too!
In the above file any line beginning with "#$" is a command for the
queueing system and any line that begins with "#" is a comment line.
The command
#$ -eo gaga
send all of the output that normally appears on the screen during an in-
teractive session to a file named "gaga". [Change the file name to some-
thing different from "gaga"!] The command
sends mail to the user when the job starts. Use your correct e-mail ad-
dress here. Similarly the command
#$ -me
sends mail when the job finishes.
After the preamble you can put any sequence of commands that you
would normally enter during an interactive session and these will be ex-
ecuted in sequential order. In the above example the command
matlab >out « ’eof’
starts MATLAB, redirects MATLAB output to a file named out and takes
the input to MATLAB from the following lines until "eof", the end-of-file
marker. The lines following "eof" should make sense to the shell, as it
interprets these lines. If you understand these principles you can con-
struct any complicated script using the full programming capabilities of
A.12. PRINTING REPORTS, GRAPHS ETC. A.36
the Kron Shell! To submit the job to DQS, use,
user@machine:dir> qsub -G Medium 1 test.job
The machines are grouped into "High", "Medium" and "Low" groups based
on the hardware capabilities. Here "-G" identifies the group as "Medium"
and only one machine is requested for the job with "1". Note that DQS
supports "parallel virtual machine" for jobs that can be executed in par-
allel on more than a single machine. In such cases you should request
the number of machines by replacing "1" with "n" where "n" can be be-
tween 1 and 9 since we have only nine machines. The default group is
"Medium" and the default number of machines is "1". So you could have
simply entered,
user@machine:dir> qsub test.job
The command qstat shows the current status of your job.
user@machine:dir> qstat
A.11.2 Example of a FLOW3D job

Construct a script file named, say m01.bat, containing the following lines.
#!/bin/csh
#$ -cwd
#$ -eo gaga
#$ -mb
#$ -me
runf3d -fort m02.f -command m02.fc -release 3.2.1 -geom m01.geo
A.12 Printing reports, graphs etc.
In room CME 244 there are several old Epson printers connected directly
to the OS/2 machines. Each printer serves two OS/2 machines. In room
CME 473 each of the DOS machine is connected directly to a HP Laser Jet
printer.
Since AIX and OS/2 have network support, printing on these ma-
chines can be done using network printers. Hence there are no printers
connected directly to each of these machines. There are two network
LaserJet printers, one located in room CME 244 and the other in CME
473. The network printers also operate on the client-server model. It
is important that you understand how the print servers operate and
use this facility in a responsible manner. If the facility is abused this
service will be discontinued and you will have to take your print jobs to
the Micro DEMO center at the book store!
• You are responsible for providing your own paper.
• Do not send a print request to a remote printer in another location.
• If you send a print job to a remote location, the job can be deleted
by the system administrator if it interferes with other queued print
jobs.
• Be present in the room where the printer is located and be prepared

to attend to the printer immediately after you send a print job to
the printer - i.e., feed paper, watch for paper jams, collect your
output etc. .
• The printers are to be used only for course/research related work.

For personal needs take your print job to the Micro DEMO center.
Most applications running under OS/2, AIX or DOS can generate printed
output for a number of different types of printers. Each type of printer/plotter
understands a particular set of instructions. For example HP LaserJet
uses PCL (Print Control Language) format, while HP plotters use HPGL
(HP Graphics Language). Another widely used page description language
is called Postscript. Application programs use something called a printer
driver to generate the output suitable for a particular output device.
If the output is generated on a computer that is connected directly
to a printer (like most DOS machines to HP LaserJet in CME 473 or OS/2
machines to Epson in CME 244) then you can send the print job directly
to the printer, typically through the printer port LPT1:
If the output is generated on a computer that does not have a printer
connected directly to it or you want a LaerJet output from an OS/2 ma-
chine, then you must use one of the network printers. In this case you
must first save the output from the application program into a file. Most
well designed application programs will give you the option to select the
output device and to save the printed output into a file. Use the follow-
ing convention in naming the output files:
filename.ps postscript
filename.eps Encapsulated postscript
(Useful for merging graphics with text)
filename.pcl HP Laser Jet
filename.hgl HP plotters using HPGL instruction set
filename.dot Epson dot matrix printers
Note that such files contain specific instructions that are understood by
specific printers/plotters. Hence they must be sent to the appropriate
printers. If you send a PCL file to an Epson printer you will get garbled
output that will make no sense!
A.12.1 Using the network printer from AIX machines

The simplest approach is to use the script named prnt on AIX machines
available within the chemical and materials engineering department. It
is a script written by Bob Barton. Hence it is available only on machines
maintained by Bob Barton.
user@machine:dir> prnt options file name
Unix experts may want to try the basic set of unix commands lpr, lpq,
lprm to submit, monitor and manage a print task. On the AIX machines,
the command to submit a PCL file to a network printer is
user@machine:dir> lpr -Pprinter name file name.pcl
where the printer name is either LJ 244 (LaerJet printer in room CME
244), LJ 473 (LaserJet printer in room CME 473) or PS 475 (the postscript
printer in room CME 475). Your job is then queued and you can examine Note that PS 475 is
the status of the queue with the command, available only on
the graduate
user@machine:dir> lpq -Pprinter name network, while
LJ 244 and LJ 473
are available on the
You can remove your print jobs from the queue with the command,
undergraduate
network.
user@machine:dir> lprm -Pprinter name job number
where job number is the number returned by the lpq command.

If your application program does not support the HP Laser Jet ( i.e.,
PCL format) output device, you can select the postcript output device
and save the file as fname.ps. A public domain utility program called
A.13. ANONYMOUS FTP SERVICE A.39
"ghostscript" allows such postscript files to be either previewed on the

monitor of an AIX machine under X-windows or convert the file to PCL
format for printing on a Laser Jet printer. To use this conversion pro-
gram enter,
user@machine:dir> gs -sDEVICE=laserjet -sOutputFile=fname.pcl fname.ps
If you are using X-windows and want to preview the contents of the file
fname.ps enter,
user@machine:dir> DISPLAY=hostname:0
user@machine:dir> gs fname.ps
where hostname is the name of the X-client. To find out more about
"ghostscript" options use one of
user@machine:dir> gs -?
user@machine:dir> man gs
Important Note: Once you have printed the document delete the output
files from your home directory.
A.12.2 Using the network printer from OS/2 machines
The "lpr" command is also available on OS/2 machines. From an applica-

tion program you should select the output device to be HP LaserJet and
save the output into a file as discussed in the previous section. This file
will be on the local computer in the directory that you select. You can
print such an output file as follows:
On HP LaserJet in CME 473
[C:\] lpr -ssprint.ucs.ualbeta.ca -pHPLaserJ fname
A.13 Anonymous ftp service - Netlib and other archives
A number of computer sites on the INTERNET provide a useful software

distribution service called anonymous-ftp-service. You can signon to
their machines as an anonymous user and retrieve (or download) soft-
ware. You should use anonymous as the userid and your e-mail address
A.13. ANONYMOUS FTP SERVICE A.40
as the password. Recall from section §A.10.4, that the e-mail addresses
have the general form userid@machine.eche.ualberta.ca. You must be
courteous in using such services. Normally you are requested to down-
load software only during off-peak hours. You must also conform to all
the software licensing conditions. A sample ftp session is given below.
Sample ftp session
user@machine:dir> ftp ftp-os2.nmsu.edu

Connected to hobbes.NMSU.Edu.
220 hobbes FTP server (Version 5.1 (NeXT 1.0) Tue Jul 21, 1992)
ready.
Name (ftp-os2.nmsu.edu:kumar): anonymous
331 Guest login ok, send ident as password.
Password: userid@machine.eche.ualberta.ca
230 Guest login ok, access restrictions apply.
ftp>
Note that it is not possible to connect to INTERNET from the subnet

129.128.44.nnn ( i.e., from room CME 244). Some well known sites and
the type of software that they serve are listed below.
Type of Software internet address domain name

MS-DOS related software 192.88.110.20 WSMR-SIMTEL20.ARMY.MIL
MS-DOS related software 128.252.135.4 wuarchive.wustl.edu
AIX related software 128.97.2.211 Harpo.SEAS.UCLA.EDU
dynamical systems theory 132.239.86.10 lyapunov.ucsd.edu
TEXrelated software 192.92.115.8 Niord.SHSU.edu
TEXrelated software 134.173.4.23 ymir.claremont.edu
NeXT related software 128.210.15.30 sonata.cc.purdue.edu
OS/2 related software 128.123.35.151 ftp-os2.nmsu.edu
U of A ftp server 129.128.76.12 ftp.srv.ualberta.ca
numerical analysis software 192.20.225.2 research.att.com
The reasonable man adapts himself to the world:
the unreasonable one persists in trying to adapt the
world to himself. Therefore all progress depends on
the unreasonable man.
— GEORGE BERNARD SHAW
Appendix B
An introduction to MATLAB
B.1 Introduction
MATLAB is a powerful, interactive software tool for numerical compu-

tations normally encountered in engineering and science. It is available
on several platforms including personal computers using DOS, worksta-
tions, mainframes and supercomputers using UNIX. It brings together a
large collection of powerful numerical algorithms (from LINPACK, EIS-
PACK etc) for solving a variety of linear algebra problems and makes
them available to the user through an interactive and easy-to-use inter-
face. Very little programming effort is required to use many of its stan-
dard functions. Yet, an experienced programmer can write advanced
functions and even develop entire tool boxes for specific applications
like control system design and signal processing. In fact several such
tool boxes already exist.
MATLAB is available in the micro computer lab in room CME 244 on
nine of the IBM RS/6000-220 workstations and on the SunOS servers
ugrads1.labs and ugrads2.labs running under SunOS operating system.
These machines are on the internet and hence are accessible through a
variety of means. Note that a student edition of MATLAB is available
from the Book store. If you have a PC at home the software and the
manual is a great buy.
Since MATLAB is interactive, you are encouraged to try out the ex-
amples as you read this manual. After each step, observe the outcome
carefully. Since computers are programmed to respond in predictable
B.1
B.2. STARTING A MATLAB SESSION B.2
manner, the key to mastering them is to be very observant.

Familiarity with the basic concepts of the operating system and the
networked environment are assumed. In this notes you will be intro-
duced to some of the basic numerical and graphical capabilities of the
MATLAB. In particular the following will be explored.
• Starting a MATLAB session
• Using built in HELP, DEMO features
• data entry, line editing features of MATLAB.
• Summary of some of the built in functions in MATLAB for solving

problems in
> linear algebra

> root finding
> curve fitting
> numerical integration
> integration of initial value problems
> nonlinear equations and optimization
> basic plotting capabilities
> Writing MATLAB functions and scripts - the m-file
For the adventurous, here are some of its advanced features. Explore
them on your own! The package provides features (or tool boxes) for sig-
nal processing, control system design, identification and optimization
through what are called m-files. The graphic features support include
3D and contour plotting as well as device drivers for a variety of out-
put devices including Postscript and meta file capabilities for producing
high quality plots (not just screen dumps!). It also provides facilities
for developing ones own tool boxes as well as facilities for interfacing
with other high level languages such as FORTRAN or C and invoke such
routines from within MATLAB.
B.2 Starting a MATLAB session
B.2.1 Direct access on AIX machines

Find a free station in room CME 244 the range of numbers 31 to 39.
These are the AIX machines. Signon by entering your userid and passs-
word. i.e.,
B.2. STARTING A MATLAB SESSION B.3
Console login: userid

password: password
Next start the X-Window interface by entering
user@machine:dir> xinit
You may copy the files ".mwmrc" and ".Xdefaults" from the directory
"/afs/ualberta.ca/home/k/u/kumar/". These files customize the X-windows
environment when the X-server is started with the command "xinit". Sev-
eral windows will be started up and one of them will be named "Aixterm".
Normally this would be the shell "ksh" and all the paths to application
program will be setup correctly for you to start running application pro-
grams. To start MATLAB simply enter
If MATLAB does not start, seek help from the system administrator!
B.2.2 Remote access from OS/2 machines or home computer
The procedure for connecting from an OS/2 machine to AIX machine

via X-windows was outlined in section A.5.2. This will provide a full X-
window based access to MATLAB. If you want to use MATLAB from a
home computer, only VT100 based emulation support is available, un-
less you have X-windows client on your home computer. Follow the steps
outlined in section A.5.3 to connect to an AIX machine from home using
kermit. In either case, after a successful connection has been estab-
lished, enter,
to start a MATLAB session. The current version of MATLAB is 4.0a. It

has advanced 2-D and 3-D graphics capabilities. The graphics features,
however, rely heavily on X-windows. Hence you must invoke MATLAB
under X-windows in order to see the graphs and images on the screen.
If you start MATLAB from a home computer, or a VT100 terminal on
campus. you are limited to seeing the textual output only on the screen.
You can still generate graphs with appropriate MATLAB commands, save
them on to a file or print them, but you cannot see them on the screen.
B.3. MATLAB BASICS B.4
B.3 MATLAB basics
Once you start MATLAB successfully, you should see the following prompt
on your screen.
< M A T L A B (tm) >
(c) Copyright 1984-92 The MathWorks, Inc.
All Rights Reserved
Version 4.0a
Dec 11 1992
Commands to get started: intro, demo, help help

Commands for more information: help, whatsnew, info, subscribe
»
This provides you with an interactive workspace in which you can define
any number of variables and invoke any function. To exit MATLAB at
any time enter
» quit
The commands that you enter within MATLAB are acted upon immedi-
ately. As soon as you enter a line like,
» fname
MATLAB checks if "fname" is a valid MATLAB command or a built in func-

tion. If so it will be executed immediately. If not MATLAB searches the
path to look for a external function or a file by the name "fname.m". Such
a file is called a m-file, as its file extension is "m". If such a file is found
it will execute the contents of that file. If not, MATLAB will generate an
appropriate error message. m-files can be either scripts (i.e. a series of
valid MATLAB commands that are executed often and hence stored in
a file) or they can be used to define entirely new MATLAB functions of
your own. More on m-files later.
While in MATLAB, if you have the need to execute a UNIX shell com-
mand, you can do so with the escape character ! - e.g., try When you exit the
editor using
» !emacs ctrl-x ctrl-c,
you will return to
to invoke the emacs editor, or MATLAB
» !ls -al
to look at a list of all the files in your directory.
B.3.1 Using built in HELP, DEMO features
MATLAB provides extensive online help using commands like help, demo,
type, lookfor, whatsnew. They are not only useful for checking the syn-
tax of a particular function, but also for exploring and learning about new
topics. Since the help command often generates lots of text that tend to
scroll by very quickly, it is useful to enable a feature called “more” with
the command,
» more on
When this is enabled, you will be shown one screen full of information
at a time. Note that this is also UNIX feature that you can use with any
program that generates lots of scrolling text. To get started with the
online help, first get a list of help topics using
» help
Table B.1 provides a list of help topics which should give you some idea
about the broad scope of MATLAB. You can obtain a list of functions
under each topic (or directory) by entering help topic. For example to
get a listing of general purpose commands (the first item in the above
table) enter,
» help general
The list so produced is given in Table B.2 to serve as a reference mate-

rial. Many of the functions that will be useful in a numerical methods
course are listed in subsequent sections of this chapter. One way to
become proficient in MATLAB is to use this HELP feature liberally - i.e.,
when ever you are in doubt call on the HELP!
There is also a built in DEMO feature. To invoke this feature simply
enter Try some graphics
demos. In Version
» demo 4.0, this works only
under X-windows
It will provide you with a menu of items. Select the ones that interest
you most. You can also search by keywords using the command lookfor.
Try,
directory/topic Brief description

matlab/general General purpose commands
matlab/ops Operators and special characters
matlab/lang Language constructs and debugging
matlab/elmat Elementary matrices and matrix manipulation
matlab/specmat Specialized matrices
matlab/elfun Elementary math functions
matlab/specfun Specialized math functions
matlab/matfun Matrix functions & numerical linear algebra
matlab/datafun Data analysis and Fourier transform functions
matlab/polyfun Polynomial and interpolation functions
matlab/funfun Function functions & nonlinear numerical methods
matlab/sparfun Sparse matrix functions
matlab/plotxy Two dimensional graphics
matlab/plotxyz Three dimensional graphics
matlab/graphics General purpose graphics functions
matlab/color Color control and lighting model functions
matlab/sounds Sound processing functions
matlab/strfun Character string functions
matlab/iofun Low-level file I/O functions
matlab/demos Demonstrations and samples
toolbox/control Control System Toolbox
toolbox/ident System Identification Toolbox
toolbox/local Local function library
toolbox/optim Optimization Toolbox
toolbox/signal Signal Processing Toolbox
simulink/simulink SIMULINK model analysis and construction functions
simulink/blocks SIMULINK block library
simulink/simdemos SIMULINK demonstrations and samples
Table B.1: List of MATLAB Ver 4.0 help topics

Function Brief description

Managing commands and functions
help On-line documentation
what Directory listing of M-, MAT- and MEX-files
type List M-file
lookfor Keyword search through the HELP entries
which Locate functions and files
demo Run demos
path Control MATLAB’s search path
Managing variables and the workspace
who List current variables
whos List current variables, long form
load Retrieve variables from disk
save Save workspace variables to disk
clear Clear variables and functions from memory
pack Consolidate workspace memory
size Size of matrix
length Length of vector
disp Display matrix or text
Working with files and the operating system
cd Change current working directory
dir Directory listing
delete Delete file
getenv Get environment value
! Execute operating system command
unix Execute operating system command & return result
diary Save text of MATLAB session
Controlling the command window
cedit Set command line edit/recall facility parameters
clc Clear command window
home Send cursor home
format Set output format
echo Echo commands inside script files
more Control paged output in command window
Starting and quitting from MATLAB
quit Terminate MATLAB
startup M-file executed when MATLAB is invoked
matlabrc Master startup M-file
Table B.2: General purpose MATLAB Ver 4.0 commands

» lookfor inverse
which will scan for and print out the names of functions which have the
keyword "inverse" in their help information. The result is reproduced
below.
INVHILB Inverse Hilbert matrix.

ACOS Inverse cosine.
ACOSH Inverse hyperbolic cosine.
ASIN Inverse sine.
ASINH Inverse hyperbolic sine.
ATAN Inverse tangent.
ATAN2 Four quadrant inverse tangent.
ATANH Inverse hyperbolic tangent.
ERFINV Inverse of the error function.
INVERF Inverse Error function.
INV Matrix inverse.
PINV Pseudoinverse.
IFFT Inverse discrete Fourier transform.
IFFT2 Two-dimensional inverse discrete Fourier transform.
UPDHESS Performs the Inverse Hessian Update.
B.3.2 Data entry, line editing features of MATLAB
The basic variables in MATLAB are treated as matrices. Vectors and

scalar are special cases of a general matrix data structure. Similarly
MATLAB handles complex variables and numbers in a natural way. Real
variables, then are, special cases. Note that MATLAB is case sensitive.
• MATLAB remembers the previous command lines that you have en-
tered. You can recall them by simply using the up and down arrow
keys (or ctrl-p and ctrl-n key combinations) and then edit them
and reenter the edited command as a new command. Basically,
it supports the following emacs key definitions for command line
editing.
Function Key sequence

Previous line ctrl-p
Next line ctrl-n
One character left ctrl-b
One character right ctrl-f
One word left esc b, ctrl-l
One word right esc f, ctrl-r
Cursor to beginning of line ctrl-a
Cursor to end of line ctrl-e
Cancel line ctrl-u
Delete character ctrl-d
Insert toggle ctrl-t
Delete to end of line ctrl-k
• To assign a value to a variable use the assignment operator "=". For

example,
» A = [1 2 3; 4 5 6; 7 8 9]
will result in a 3x3 matrix. Note that there is no need to explicitly

declare the dimension of an array. Since MATLAB is case sensitive
you have defined only "A " and "a" remains undefined. Similarly
» x=[2+4*i, 3+5*i]
will generate a complex vector with two elements. If you want to

add another element enter what would the
value of x(3) be?
» x(4)=5+6*i
Note that the dimension of the vector x is now automatically in-

creased to 4. Observe that the square brackets are used in forming
vectors and matrices. Semicolon is used to separate rows in a ma-
trix. Comma is used to separate individual elements of a vector (or
matrix). Parentheses are used to identify individual array elements.
(Try help punct and help paren)
• After you have defined the variables A and x as above, go through

the following exercise and make sure you understand the result.
» A(2:3,1:2)
Observe the use of () and : to select a sub block of A. Next, try What might
happen if the size
of sub-blocks are
different?
» B(4:5,2:3)=A(2:3,1:2)
This demonstrates how to extract a sub-block matrix of A and as-

sign it to another sub-block of B. Next, try, What would be the
value of x after you
» x(4:-1:1) execute this
command? Why?
which reverses the order of elements of x. Next, try the command,
» p=[1 3]; x(p)
Observe that there are two commands, separated by semicolon.

This example also demonstrates a powerful way of selecting spe-
cific elements of a vector. This is easily extended to matrices also.
Well, try,
» q=[2 3]; A(p,q)
I hope you get the idea.
• To examine the value of a variable simply enter the name of the

variable. All the variables that you define during a MATLAB session
are stored in the workspace ( i.e., in computer memory) and they
remain available for all subsequent calculations during the entire
MATLAB session i.e., until you “quit” MATLAB.
• You can declare any variable to be global in nature using,
» global A
If the same variable is also declared as global in several functions,

then all those functions share the same value. To check if a variable
is global use,
» isglobal(A)
A value of 1 is returned if it is global.
• To examine the list of variables currently defined in your workspace

and the attributes of those variables, use one of the two commands
"who" and "whos".
» whos
• To generate a set of equally spaced values in a simple manner fol-

low the example below:
» x = 0 : 0.05 : 1.0
will generate x = [0 0.05 0.1 0.15 0.2 · · · 1.0]. (Try help colon).
• To suppress the automatic echoing of any line that you enter from
keyboard, terminate such a line with a semi-colon ";". For example
» x = 0 : 0.05 : 1.0;
will define x as before, but will not echo its value. (Try help punct).
• To continue the entry of a long statement onto the next line use an
ellipsis consisting of three or more dots at the end of a line to be
continued. For example
» s = 1 − 1/2 + 1/3 − 1/4 + 1/5 − 1/6 + 1/7 · · ·

» −1/8 + 1/9 − 1/10 + 1/11
• Anything that follows a % sign is treated as a comment. For exam-

ple the following is a valid command line.
» I = 1 : 1 : 20 %generating a set of integers from 1 to 20
• The numeric display format is controlled by the "format" command.

Use
» format long
for 14 digits display. (Try help format )
• You can save the contents of a workspace with the "save" command.
Try, Try the command
!ls jnk*
» save jnk Observe that the
extension .mat has
In the next few statements examine the currently defined variables, been added
clear the workspace and load a previously saved workspace.
» whos
» clear
» whos
» load jnk
» whos
• The following matrix operations are available in MATLAB. You can

use help on each of them to find out more precise information on
them.
+ addition, e.g., C = A + B ⇒ Cij = Aij + Bij

− subtraction, e.g., e.g., C = A − B ⇒ Cij = Aij − Bij
P
∗ matrix multiplication, e.g., C = A ∗ B ⇒ Cij = k Aik Bkj
ˆ Matrix power. Z = Xˆy is X to the y power if y is a scalar and
X is square. If y is an integer greater than one, the power
is computed by repeated multiplication. For other values of
y the calculation involves eigenvalues and eigenvectors. (try
help arith).
0 Matrix transpose. X 0 is the complex conjugate transpose of X.
X.0 is the non-conjugate transpose. (try help punct).
\ left division. A\B is the matrix division of A into B, which is
roughly the same as inv(A)*B , except it is computed in a dif-
ferent way. If A is an N-by-N matrix and B is a column vector
with N components, or a matrix with several such columns,
then X = A\B is the solution to the equation A ∗ X = B com-
puted by Gaussian elimination. (try help slash)
/ right division. B/A is the matrix division of A into B, which is
roughly the same as B*inv(A).
Note that the dimensions of the matrices must be compatible for

the above operations to be valid; if you attempt matrix operations
between incompatible matrices an appropriate error message is
generated.
• The following relational operators are available in MATLAB. Try

help relop for additional details.
< Less than relational operator

> Greater than relational operator
<= Less than or equal
>= Greater than or equal
== equal
˜= not equal
They are applied element-by-element between matrices of the same

size, producing a resultant matrix consisting of 0’s and 1’s.
• Element-by-element multiplicative operations are obtained as fol-

lows:
operator example index notation

.∗ C = A.*B Cij = Aij Bij
Bij
.ˆ C = A.ˆB Cij = Aij
./ C = A./B Cij = Aij /Bij
.\ C = A.\B Cij = Bij /Aij
B.3.3 Linear algebra related functions in MATLAB

A list of all advanced matrix related functions in MATLAB is given in
Table B.3 Use the help command on each of these functions to find out
more about the function and its exact syntax.
Work through the following exercise to become familiar with the us-
age of some of the linear algebra functions and refresh some of the re-
sults from a first year linear algebra course.
Exercise - review of 1st year linear algebra
• Define the matrix, A and a vector, b as
» A = [1 0 0.307; 0 1 0.702; -2 1 0] Why was "A"

» b = [0.369*275;0.821*275;0]; echoed on the
screen, while "b"
was not? Is "b" a
Observe the two different ways semicolon has been used here. What
row or a column
are they?
vector?
• Solve the equation Ax = b using,
» x = A\b
• Verify that x satisfies the equation by calculating,
» A*x - b
Function name Action

Matrix analysis
cond Matrix condition number
norm Matrix or vector norm
rcond LINPACK reciprocal condition estimator
rank Number of linearly independent rows or columns
det Determinant
trace Sum of diagonal elements
null Null space
orth Orthogonalization
rref Reduced row echelon form
Linear equations
\ and / Linear equation solution; use "help slash"
chol Cholesky factorization
lu Factors from Gaussian elimination
inv Matrix inverse
qr Orthogonal-triangular decomposition
qrdelete Delete a column from the QR factorization
qrinsert Insert a column in the QR factorization
nnls Non-negative least-squares
pinv Pseudoinverse
lscov Least squares in the presence of known covariance
Eigenvalues and singular values
eig Eigenvalues and eigenvectors
poly Characteristic polynomial
hess Hessenberg form
qz Generalized eigenvalues
rsf2csf Real block diagonal form to complex diagonal form
cdf2rdf Complex diagonal form to real block diagonal form
schur Schur decomposition
balance Diagonal scaling to improve eigenvalue accuracy
svd Singular value decomposition
Matrix functions
expm Matrix exponential
expm1 M-file implementation of expm
expm2 Matrix exponential via Taylor series
expm3 Matrix exponential via eigenvalues and eigenvectors
logm Matrix logarithm
sqrtm Matrix square root
funm Evaluate general matrix function
Table B.3: Linear algebra related functions in MATLAB

• Next calculate the norm of the residual,
» norm(A*x - b)
• Determine the rank of A using
» rank(A)
• Carry out the LU decomposition using
» [L,U]=lu(A) Why does "L" not

appear to be lower
triangular?
• Calculate the determinant of A using
» det(A)
» det(L)*det(U)
• Calculate the eigenvalues and eigenvectors of A.
» [v,d]=eig(A)
» prod(diag(d)) Can you explain
this result?
• Find out the characteristic polynomial of A and its roots.
» c1=poly(A)
» roots(c1) Can you explain
» prod(ans) these results?
B.3.4 Root finding
x=fsolve(’fun’,x0) solution to a system of nonlinear equations (or

zeros of a multivariable function).
fun(x) is a function that you should write to
evaluate f(x) - i.e., you define your problem in
an m-file.
x0 is the initial guess for the root. [There is
obviously more to it than I can describe here!
Read the manual or try help fsolve].
fzero(’fun’,x0,tol) finds the root of a single nonlinear algebraic

equation.
fun(x) is the external function describing your
problem that you should write in a m-file.
x0 is the initial guess.
poly(V) if V is a vector then it returns the coefficients

of the polynomial with roots determined by V.
i.e., roots and poly are inverse functions of each
other.
roots(c) computes all the roots of a polynomial whose

coefficients are in c. i.e., Pn (x) = (c1 x n +
c2 x n−1 + · · · + cn+1 ).
B.3.5 Curve fitting
c=polyfit(x,y,n) least-squares curve fitting of degree n. The co-

efficients in descending powers of x are returned
in c.
polyval(c,s) evaluates the polynomial whose coefficients are

in c at locations determined by s.
yi=spline(x,y,xi) Generates a cubic spline through the data vec-

tors (x, y) and then computes a vector of in-
terpolated values of yi at xi.
diff(x,n) computes the n forward differences from the

vector x.
The other functions of possible interest are fmin, fmins, residue, conv
, table1.
B.3.6 Numerical integration, ordinary differential equations
quad(’fun’,a,b,tol,trace) computes the definite integral over the limit

(a,b) using adaptive recursive Simpson’s rule.
fun(x) is an external function that you must
provide in a m-file.
tol is the acceptable global error. trace is an
optional flag to monitor the integration pro-
cess.
[t,y]=ode45(’fun’,t0,tf,y0,tol,trace) integrates a system of nonstiff differ-

ential equations of the form dy/dt = f (y) us-
ing 4 and 5 order Runge-Kutta methods.
fun(y)is the external function which defines your
problem. You must provide this via a m-file.
(t0,y0) is the initial condition.
tf is the final point at which you want to stop
the integration.
tol is the acceptable global error in the solu-
tion. trace trace is the optional flag to print
intermediate results.
The other functions of possible interest are ode23, quad8
B.3.7 Basic graphics capabilities

MATLAB ver 4.0 maintains separate graphics windows and a text win-
dow. Your interactive dialogue takes place on the text window. When
you enter any graphics command, MATLAB plots that graph immedi-
ately on the graphics window. It can open several graphics windows.
So, clearly commands are needed to select a specific window to be the
current one. The list of graphics related commands are given in Table
B.4. Work through the following exercise interactively and observe the
computer response in order to understand the basic graphic capabilities
of MATLAB. Make sure that you are running MATLAB under X-windows.
Text following the percent sign (%) are explanatory comments. You need
not enter them.
Exercise - producing a simple graph
»x=0:0.1:2*pi; % create a vector x in the range (0,2 Pi)

»figure(1) % open a graphics window labeled Figure 1
»figure(2) % open a graphics window labeled Figure 2
»plot(x,sin(x)) % plot sin(x)
»hold % keep the graph of sin(x)
»plot(x,cos(x),’go’) % add graph of cos(x) with line type ’go’
»title(’My first plot’) % put some title
»xlabel(’x-axis’) % label the x-axis
»ylabel(’y-axis’) % label the x-axis
»print -deps fig1.eps% produce a postscript copy in file fig1.eps
»!ls -al fig1.eps % verify that the figure is saved in a file
»!xpreview fig1.eps % use the postscript previewer of AIX (optional)

Figure window creation and control
figure Create Figure (graph window)
gcf Get handle to current figure
clf Clear current figure
close Close figure
Axis creation and control
subplot Create axes in tiled positions
axes Create axes in arbitrary positions
gca Get handle to current axes
cla Clear current axes
axis Control axis scaling and appearance
caxis Control pseudocolor axis scaling
hold Hold current graph
Handle Graphics objects
figure Create figure window
axes Create axes
line Create line
text Create text
patch Create patch
surface Create surface
image Create image
uicontrol Create user interface control
uimenu Create user interface menu
Handle Graphics operations
set Set object properties
get Get object properties
reset Reset object properties
delete Delete object
drawnow Flush pending graphics events
newplot M-file preamble for NextPlot property
Hardcopy and storage
print Print graph or save graph to file
printopt Configure local printer defaults
orient Set paper orientation
capture Screen capture of current figure
Movies and animation
moviein Initialize movie frame memory
getframe Get movie frame
movie Play recorded movie frames
Miscellaneous
ginput Graphical input from mouse
ishold Return hold state
whitebg Set graphics window defaults for white background
graymon Set graphics window defaults for gray-scale monitors
Table B.4: Graphics related function in MATLAB

»gcf % get current figure (should be 2)

»figure(1) % make figure 1 the current figure
»close(1) % close window 1
»gcf % get current figure (should be 2)
»close(2) % close window 1
In this exercise you produced the data from within MATLAB. If you have
columns of data in a file, you can read them into MATLAB and plot them
as above. The postscript file produced in the above example can be
merged with other documents or printed on a postscript printer. Use
help print to find out about support for other type of printers and plot-
ters.
B.3.8 Control System Toolbox

The Control system toolbox, which uses MATLAB matrix functions, was
built to provide specialized functions in control engineering. The Con-
trol system toolbox is a collection of algorithms, expressed in m-files,
that implement common control system design, analysis, and modeling
techniques.
Dynamic systems can be modeled as transfer functions or in state-
space form. Both continuous-time and discrete-time system are han-
dled. Conversions between various model representations are possible.
Time responses, frequency responses, and root-locus measures can be
computed and plotted. Other functions allow pole-placement, optimal
control, and estimation.
The following example shows the use of some of the control system
design and analysis tools available in MATLAB.
Example
The process transfer function, G is defined as:
1
G=
(s + 1)(s + 2)(s + 3)
The transfer function is entered into Matlab by entering the numerator

and the denominator coefficients separately as follows:
» num = 1;
» den1 = [1 1];
» den2 = [1 2];
» den3 = [1 3];
The denominator polynomial is the product of the three terms. Convo-

lution, conv, is used to obtain the polynomial product:
» den = conv(den1,conv(den2,den3));
To get an open-loop process response to a unit step change, the function

step can be used:
» t = 0:0.5:5; Define time in the

» y = step(num,den,t); range 0-5.
» plot(t,y,’*’); Generate step
response and plot.
The Bode plot can be obtained by first defining a vector of frequencies,

and then using the function bode:
» w = logspace(-1,1); Generate equally

» [mag,phase] = bode(num,den,w); spaced data in the
range 10−1 and 101
The bode plots for amplitude ratio and phase can be obtained by typing:
» loglog(w,mag)
» semilogx(w,phase)
The root-locus can be obtained by defining a vector of desired gains, and

then using the function rlocus:
» k = 20:5:70; Define gains in the

» y = rlocus(num,den,k); range 20-70.
» plot(y,’*’) Generate and plot
the root-locus.
The closed-loop transfer function can be represented by:
Y G c Gp
=
Ysp 1 + G c Gp
The closed-loop transfer function is calculated and entered into Matlab

for analysis using the same functions used in the open-loop system.
Discretization can only be done through the state-space model represen-
tation. Therefore, it is necessary to transform transfer function models
to state-space models. The transfer function model can easily be trans-
formed into the state-space model by using the function tf2ss:
» [A,B,C,D] = tf2ss(num,den);
dx
where A, B, C, D are matrices in the differential equations dt := Ax + Bu
and y = Cx + Du. To obtain a discretized model, the function c2d is
used:
» [ad,bd] = c2d(A,B,Ts); % Ts is the sample time
This function converts state-space models from continuous time to discrete-

time assuming a zero-order hold on the input. To obtain a step response
on the discretized model, the function dstep can be used:
» y = dstep(ad,bd,C,D,1,100);
» plot(y),title(’step response’);
Several additional control functions that are available in Matlab are

listed in Table B.5 The online help screen should be referred to for
information on how to use these tools. The function what can be used
to find out what other functions are available.
B.3.9 Producing printed output of a MATLAB session
If you want to produce a hard copy of your interactive MATLAB session,

you can log a record of the entire session in a file with the diary com-
mand. The command
» diary file
will start recording every keyboard entry and most of the computers tex-
tual response (not graphics) in file. To suspend the recording, use
» diary off
and to resume recording you can use,
» diary on
The file contains simple text (ASCII) and can be printed on the network
printer.
Function name Purpose

Functions for model conversion
[num, den] = ss2tf (a, b, c, d, iu) State-space to transfer function
[z, p, k] = sstzp(a, b, c, d, iu) State-space to zero-pole
[a, b, c, d] = tf 2ss(num, den) Transfer function to state-space
[z, p, k] = tf 2zp(num, den) Transfer function to zero-pole
[a, b, c, d] = zp2ss(z, p, k) Zero-pole to state-space
[num, den] = zp2tf (z, p, k) Zero-pole to transfer function
[ad, bd] = c2d(a, b, T s) Continuous to discrete
[a, b] = d2c(ad, bd, T s) Discrete to continuous
Functions for modeling
append Append system dynamics
connect System interconnection
parallel Parallel system connection
series Series system connection
ord2 Generate A,B,C,D for a second order system
Continuous time and frequency domain analysis
impulse impulse response
step Step response
lsim Simulation with arbitrary inputs
bode Bode and Nichols plots
nyquist Nyquist plots
Discrete time and frequency domain analysis
dimpulse Unit sample response
dstep Step response
dlsim Simulation with arbitrary inputs
dbode Discrete Bode plots
Table B.5: List of functions from control system tool box

B.3.10 What are m-files?

MATLAB derives its strength and wide popularity from being extensi-
ble through the m-file facility. Extensibility means that using a core set
of built-in functions, users can extend the capabilities of MATLAB by
writing their own functions. The functions are stored in files with the
extension “.m”. Any file with the extension “.m” in the search path of
MATLAB is treated as a MATLAB m-file. To find out the current path of
MATLAB enter,
» path
You can list the contents of a m-file with the type command. While the
help command produces only documentation on the function, the type
command produces a complete listing of the function. Try,
» type sin
» help sin
» type erf
» help erf
Note the “sin” is a built-in function and hence no code is listed. On the
other hand “erf” is the error function implemented as a m-file and hence
a complete listing is produced.
The m-files can take two forms - viz. (i) a script file and (ii) files that
define entirely new functions. Such files should be in the MATLAB search
path.
Example of a script file
In a script file, you can put any MATLAB commands that you would
normally enter in an interactive session. Simply entering the name of
the file would then execute the contents of that file. For example to
enter a large matrix, create a file called "A.m" in your home directory
using your favorite editor. This file should contain the following text.
B = [ 1 2 3 4 5 6 7 8 9;
2 3 4 5 6 7 8 9 0;
3 4 5 6 7 8 9 0 1;
4 5 6 7 8 9 0 1 2;
5 6 7 8 9 0 1 2 3]
b=sin(B)
To execute the contents of this file from within MATLAB enter,
»A
Note that a matrix variable "B" of size (5 × 9) has been defined in your
workspace and the variable "b" contains the values of sin(B).
In a script file you can include any such sequence of valid MATLAB
commands, including program flow control commands like for, if,
while loops etc. However a script file is not a function file and hence
you cannot pass any arguments to the script. Also, when you execute
a script file from the workspace, all of the variables defined in a script
file become global variables. In contrast any variable defined within a
function file is local to that function and only the results returned by the
function become global in nature.
Example of a function file
Let us take the example of an isothermal, multicomponent flash equa-

tion, given by,
XN
zi (1 − Ki )
f (ψ) := =0
i=1
1 + ψ(Ki − 1)
In this equation, (zi , Ki ) are known vectors of length N and ψ is the
unknown scalar variable. So we should like to write a function, say,
flash(psi) that would return the value of f (ψ). This function should be
saved in a file named "flash.m" in your home directory. Such a function
might look as follows:
function f=flash(psi)
% Calculates the flash equation f=flash(psi,K,z)
% K is a vector of any length of equilibrium ratios
% K, z are known.
% psi is the vapor fraction
% This is the last line of help. Notice the blank line below!
global K z
f=((1-K).*z) ./ (1+(K-1)*psi);
f=sum(f);
Let us understand the anatomy of this function. The first line should
always contain the keyword "function" in order to identify it as a func-
tion definition and not a script file. Then a list of values computed and
returned by the function should appear - in the present case only "f" is
being returned. If you have more variables being returned you would
list them as "[f1, f2, f3] etc. Next, the equal sign is followed by the
name of the function. Then the list of input variables are given in paren- Note that the file
thesis. The next several lines begin with the percent sign (%) and hence name is
are treated as comments. Here is the place to put the documentation constructed by
appending ".m" to
on what the function does and how to use it. This is also the part that
the function name.
is printed out when a user asks for help on this function. A blank line
In the above
signifies the end of the help message. The actual code follows the blank example the file
line. Notice the use of element-by-element multiplication of two vectors name will be
which avoids the use of do loops. How elegant! flash.m
Assuming that you have created a file called "flash.m" containing
the above lines, work through the following steps.
» help flash
» type flash
» global K z
» z=[.25 .25 .25 .25]
» K=[1 .5 .4 .1]
» whos
» flash(0.1)
As a challenge, take up the task of modifying the function flash such

that it will take in a vector of ψ values and return the corresponding
function values!
B.3.11 Programming features
If you know any one high level programming language such a FORTRAN,
C or even BASIC, you should have no difficulty in understanding the ele-
mentary program flow control features of MATLAB. A list of help topics
is given in Table B.6. Let us take the example of "flash.m" and illustrate
the use of "if" and "for" constructs. First we check if the length of vec-
tors K, z are the same; if not we generate an error message. Note the
length and error are built-in MATLAB functions. In the next section
we determine the length of input vector "x" and build a loop to calculate
the function for each element of "x" and store it in the corresponding
element of "f". Use "help relop" and "help lang" to find out more about
relational operators and programming language features.
Example
function f=flash(x)
% K is a vector of any length of equil ratios.
% K, z are defined as global in main
% x is the vapor fraction
% The following is the isothermal flash eqn.
global K z
if ( length(K) ˜= length(z) )
error(’Number of K values & compositions do not match’)
end
n=length(x); %Find the length of the input vector

for i = 1:n %setup a loop for each element of x
t=((K-1).*z) ./ (1+(K-1)*x(i));
t=sum(t);
f(i) = t;
end
B.3.12 Debugging tools

Version 4.0 of MATLAB provides for the first time some debugging tools.
If you are familiar with the debugging concepts, use of this facility should
be straight forward. Basically it is a tool for debugging new functions
that a user develops. It provides tools for the following:
• to setup break points, where the execution of a function will be

suspended and the control returned to the user, (dbstop function-
name)
• examine values of variable,
• execute one line at a time. (dbstep)
• trace the sequence of function calls (dbstack),
• resume execution. (dbcont)
• list a function with line numbers. (dbtype function-name)


MATLAB as a programming language
script About MATLAB scripts and M-files
function Add new function
eval Execute string with MATLAB expression
feval Execute function specified by string
global Define global variable
nargchk Validate number of input arguments
Control flow
if Conditionally execute statements
else Used with IF
elseif Used with IF
end Terminate the scope of FOR, WHILE and IF statements
for Repeat statements a specific number of times
while Repeat statements an indefinite number of times
break Terminate execution of loop
return Return to invoking function
error Display message and abort function
Interactive input
input Prompt for user input
keyboard Invoke keyboard as if it were a Script-file
menu Generate menu of choices for user input
pause Wait for user response
uimenu Create user interface menu
uicontrol Create user interface control
Debugging commands
dbstop Set breakpoint
dbclear Remove breakpoint
dbcont Resume execution
dbdown Change local workspace context
dbstack List who called whom
dbstatus List all breakpoints
dbstep Execute one or more lines
dbtype List M-file with line numbers
dbup Change local workspace context
dbquit Quit debug mode
Table B.6: Program control related help topics

• quit debugging session. (dbquit)
Since all the variables with in a function are treated as local variables,
their values are not available in the workspace. To examine their values
within a function, you have to be able to stop the execution at a spec-
ified line within a function and examine the values of local variables.
The following exercise illustrates the debugging process using the flash
function developed earlier. So we assume that a "flash.m" file exists in
your current directory.
Exercise - debugging a function
»K=[2 1 .5 .3]; % define K values

»z=[.25 .25 .25 .25]; % define z
»global K z % define K z to be global variables
»whos % examine current variables
»dbstop flash % set up break point in flash
»psi=[.2:.2:.8] % define psi (a vector)
»flash(psi) % begin execution of flash, it will stop at
8 global K z the first executable line (line 8 here)
K»dbtype flash % list the function with line numbers
1 function f=flash(x)
2 % K is a vector of any length of equil ratios.
3 % z is the feed composition (same length as K)
4 % K, z are defnied as global in main
5 % x is the vapor fraction
6 % The following is the isothermal flash eqn.
7
8 global K z
9 if ( length(K) = length(z) )
10 error(’Number of K values & compositions do not match’)
11 end
12 n=length(x)
13 for i = 1:n
14 t=((K-1).*z) ./ (1+(K-1)*x(i));
15 t=sum(t);
16 f(i) = t;
17 end
K»dbstop 12 % set up a new break point at line 12
K»dbcont % resume execution from current line 8
12 n=length(x) stops before executing line 12
K»x % examine the value of x
K»n % n - should be undefined at this stage

K»dbstep % execute one line and stop at line 13
13 for i = 1:n
K»n % now display value of n (should be 4)
K»dbstatus flash % display the break points.
K»K % display the value of K
K»z % display the value of z
K»dbcont % resume execution till the end of function.
K»dbquit % terminate debugging of flash
This simple example illustrates how to set up break points, examine

values of variables and step through execution one line at a time.
Angling may be said to be so like the mathematics,
that it can never be fully learnt.
— IZAAK WALTON
So is UNIX.
— K. Nandakumar
Appendix C
Some basic unix commands
C.1 Introduction to the shell and the desktop
In a command line oriented, interactive environment, a command shell

(which is a program or a process) accepts a command from the keyboard,
passes it to the operating system for execution, prints out any error or in-
formational messages generated by the command and displays a prompt
string indicating the readiness to accept another command. There are
several shells available under AIX. The Kron shell or ksh is one of the
most powerful shells and is the default shell on the AIX machines main-
tained by the department of chemical engineering.
In the GUI oriented environment, the equivalent of a command shell
is the desktop which organizes various tools and application programs
such as file manager, program manager, printer manager, etc. as objects
accessible via icons. The interaction takes place through dialogue boxes
and forms that must be completed. Program execution begins simply by
double clicking on the appropriate icons.
If you have a good reason to change your default shell to something
other that ksh, you can do so with the chsh command,
user@machine:dir> chsh
This command will display your current shell, and prompt you for the
C.1
C.1. INTRODUCTION TO THE SHELL AND THE DESKTOP C.2
name of the new shell. The change takes effect when you login the next
time. At any time you can invoke a new shell, different from the login
shell, e.g.,
user@machine:dir> csh
invokes a C-shell.
You can invoke a desktop at any time on AIX machines by entering
user@machine:dir> xdt3
Since the use of desktop is supposed to be rather intuitive, you are en-
couraged to explore its features on your own!
C.1.1 The ".profile" file

The ".profile" file is used to customize the shell environment for ksh.
The following is a typical example of a ".profile" file.
PATH=$PATH:$HOME/bin:$KHOROS_HOME/bin
HOSTNAME=‘hostname‘
PS1=’ $LOGNAME@$HOSTNAME:$PWD>’
EDITOR=emacs
#defines alias for commonly used commands

alias ls=’ls -al’
#export environment variables to other processes ...

export PATH HOSTNAME PS1 EDITOR TERM
Here, several environment variables such as PATH, HOSTNAME, PS1 etc.

have been defined. The concept of the environment is like a bulletin
board. You can post definitions of any number of variables there. Ap-
plication programs that expect specific variables can look for them and
use their values, if they are defined. Note that when you set the value of
a variable ( i.e., left hand side of an equal sign), the name is used without
any prefix. When you want to retrieve the value the $ prefix is used. For
example try,
user@machine:dir> echo $PATH
to look at the value of the current path.

C.1. INTRODUCTION TO THE SHELL AND THE DESKTOP C.3
In the first line of the example above, a system wide environment

variable $PATH, which is already defined, is redefined to add additional
paths, such as $HOME/bin separated by colon. Observe the ’$’ prefix to
the name of the environment variable. $HOME itself is an environment
variable, containing the value of the home directory. In the 2nd line the
variable HOSTNAME is defined to contain the name of the workstation.
This name is actually retrieved by the program ‘hostname‘. The 3rd line
redefines the prompt string PS1 using other variables such as LOGNAME,
HOSTNAME and PWD. These variables contain, respectively, the values of
the userid, machine name and present working directory. If the variable
EDITOR is set to emacs, then command line editing features using emacs
keys are enabled.
You can also define aliases for certain commands. In the example
above, the "ls" string is defined to be ’ls -al’ - so when you enter "ls" at
the prompt, the command ’ls -al’ is executed. To examine all the cur-
rently defined aliases, enter,
user@machine:dir> alias
By default, the "rm" command removes files without prompting you for
confirmation which could result in accidental deletion of files. The alias
defined above, assigns ’rm -i’ to "rm". The keyword "-i" stands for inter-
active mode and hence you will always be prompted before removing a
file.
The variables defined in a shell environment are available only to that
shell environment and not to other shells that you may start from the
current one. The export command is used to export the variables to all
subsequent shells. The last line in the above example exports several
environment variables.
To look at all of the environment variables defined in the current ksh
shell, enter,
user@machine:dir> set
To examine the value of an environment variable, enter,
user@machine:dir> echo $PS1
To set a new environment variable, use
user@machine:dir> DUMMY=junk
C.2. MANAGING FILES C.4
In addition to assigning values to environment variables, the shell

allows programming flow control features. Thus one can write quite
powerful scripts to execute a complex sequence of commands. A script
is nothing but a set of AIX instructions placed in a file. By enabling the
execute permission for this file, and entering its name from the command
line you can cause the instructions in that file to be executed.
C.2 Managing files
In managing your files and directories, you need to be able to list the
contents of a directory or file, copy and move files, compress and uncom-
press files, create and delete files and directories, control the ownership
and access to files etc. Commands to carryout these tasks are illustrated
below with specific examples. Try them out at a terminal. To get a com-
plete description of each command use the man pages i.e.,
C.2.1 Making sense of the directory listing - the "ls" command
The ls command produces a listing of all the files in the current direc-
tory. In its most useful form, you will use the “-al” keywords, i.e.,
user@machine:dir> ls -al dir
Typically, files that begin with the “.” ( e.g., .profile) are treated as hidden
files. They keyword “-a” however lists all of the files including the hid-
den ones. The keyword “-l” produces the long listing, a sample of which
is shown in figure C.1. This listing provides information on file access
control, ownership, size, time stamp etc. Each line contains information
for a file or directory. The first character identifies whether it is a file (-),
a directory (d) or a symbolic link (l). A symbolic link is a pointer to some
other file (think of it as an alias). The next set of nine characters iden-
tify the file access control, in groups of three. Since AIX is a multiuser
environment, users can control ownership and access of their files to
others. The possible access modes are: read (r), write (w) execute (x) or
none(-). These modes apply to (user, group, others). The groups are es-
tablished by the system administrator. The owner and group names are
listed next, followed by file size in bytes, the time stamp for last change
and the file name.
permission
time stamp File name
{
{
control Owner Group File size
{
{
{
{
drwxr-sr-x 27 kumar sys 1536 May 24 23:14 .
drwxr-sr-x 59 sys sys 1536 May 13 08:52 ..
-rw-r--r-- 1 kumar others 1937 Jan 07 11:47 .Xdefaults
drwx------ 2 kumar others 512 Jul 21 1992 .elm
-rw-r--r-- 1 kumar sys 2504 May 19 12:08 .mwmrc
-rwxr-xr-x 1 kumar sys 610 May 04 12:36 .profile
-rw------- 1 kumar sys 348 May 14 12:22 .rhosts
drwxr-xr-x 3 kumar others 512 Jul 21 1992 .tin
-rw-r--r-- 1 kumar sys 136 May 11 14:11 .xdt3
-rw-r----- 1 kumar others 1222 Jan 19 1992 Ass1.m
drwxr-xr-x 2 kumar others 512 May 19 13:12 CHEM2
drwx------ 2 kumar others 512 May 27 1992 Mail
{
r read permission
w write permission 1st set applies to owner
x execute permission 2nd set applies to group
- no permission 3rd set applies to all
d indicates a directory
l indicates a symbolic link
Examples:
The file .profile has (read,write,execute) permission for owner (kumar in this case) and
(read,execute) permission for both the group (sys in this case) and everyone.
The directory Mail has (read,write,execute) permission for owner only
The command
chmod g+r file
will give read access to group for file, while
chmod o-w file
takes away write access to all for file
Other related Unix commands
ls -al - detailed listing of directory such as the above
chmod - change permission on files and directories
chown - change ownership of files and directories
rm - remove or delete a file
rmdir - remove or delete a directory
mkdir - create a new directory
Figure C.1: Output of the "ls" command

C.2.2 Changing permission on files - the "chmod" command
The chmod command allows you to modify the access control of files and
directories.
Examples
• To give read permission to group for file use,
user@machine:dir> chmod g+r file
• To give write permission to everyone for all the files in a directory

use,
user@machine:dir> chmod -R a+w dir
Note the the "-R" flag stands for recursive use of the command for
all files in all subdirectories.
• Note that, in order to give read permission to a directory, the exe-

cute permission at the directory level must be set.
user@machine:dir> chmod a+x dir
C.2.3 Moving files
The mv (move) command moves files and directories from one directory
to another, or renames a file or directory. You cannot move a file onto
itself.
Warning: The mv command can overwrite many existing files unless you
specify the -i flag. The -i flag prompts you to confirm before it overwrites
a file.
Examples
• To rename a file, enter:
user@machine:dir> mv oldname newname
This renames file oldname to newname. If a file named newname

already exists, its contents are replaced with those of oldname.
• To move a directory, enter:
user@machine:dir> mv olddir newdir
This moves all files and directories under olddir to the directory
named newdir, if newdir exists. Otherwise, the directory olddir is
renamed to newdir.
• To move several files into another directory, enter:
user@machine:dir> mv file1 dir1/file2 newdir
This moves file1 to newdir and dir1/file2 to newdir/file2.
C.2.4 Copying files

The cp command creates a copy of the contents of the file or directory
from a source to a target. If the file specified as the target exists, the
copy writes over the original contents of the file. If you are coping more
than one source file, the target must be a directory.
Examples
• To make a copy of a file in the current directory, enter:
user@machine:dir> cp file.old file.new
If file.new does not already exist, then the cp command creates it.
If it does exist, then the cp command replaces it with a copy of the
file.old file.
• To copy a file in your current directory into another directory, enter:
user@machine:dir> cp file.old dir/sub.dir/
This copies file.old to dir/sub.dir/file.old.
• To copy all the files in a directory to a new directory, enter:
user@machine:dir> cp /home/user/dir1/* /home/user/dir2
This copies all the files in the directory /home/user/dir1/ to the

directory /home/user/dir2/. As a variant, explore the "-R" flag to
copy not only all of the files, but also all of the subdirectories.
C.2.5 Changing ownership of files - the "chown" command

The chown command changes the owner of the file specified by the File
parameter to the user specified by the Owner parameter. The Owner
parameter can be specified either as a user ID or as a login name found
in the /etc/passwd file. Optionally, a group can also be specified. The
group can be specified either as a group ID or as a group name found in
the /etc/group file. The syntax is,
user@machine:dir> chown -R owner:group file
Only the root user can change the owner of a file.
C.2.6 Compressing files - the "compress" command

The compress command reduces the size of files using adaptive Lempel-
Zev coding. Each original file specified by the file parameter is replaced
by a compressed file with a ".Z" appended to its name. The compressed
file retains the same ownership, modes, and access and modification
times of the original file. If compression does not reduce the size of a
file, a message is written to standard error and the original file is not
replaced. The syntax is,
user@machine:dir> compress file
To restore the file to its original state use the command,
user@machine:dir> uncompress file
Also try the GNU version of compress utility called gzip and gunzip -
they are more efficient in both speed and size.
C.2.7 Removing files - the "rm" command

The rm command removes the entries for the specified file or files from a
directory. If an entry is the last link to a file, the file is then deleted. You
do not need read or write permission for the file you want to remove.
However, you must have write permission for the directory containing
that file.
Examples
• To delete a file, enter:

C.3. MANAGING PROCESSES C.9
user@machine:dir> rm myfile
If there is another link to this file, then the file remains under that
name, but the name myfile is removed. If myfile is the only link,
the file itself is deleted. Caution: You are not asked for confir-
mation before deleting the file. It is useful to set an alias in your
".profile" file to redefine "rm" as
After each file name is displayed, enter "y" to delete the file, or
press the Enter key to keep it.
C.3 Managing processes
Since AIX is a multi-tasking operating system, several tasks (or processes)

can be running at the same time. So we need a set of tools to monitor the
currently running processes and the resources they consume, suspend
or terminate specific processes, set priority for certain tasks or schedule
some tasks for execution at specified times. Commands to accomplish
these tasks are illustrated next.
C.3.1 Examining jobs or processes - the "ps" command

The ps command displays a set of currently running tasks. In its sim-
plest and most useful form, the command is,
user@machine:dir> ps -ael
This provides a long listing of all the currently running processes in-
cluding all of the daemons started by the root at the time of booting the
computer. A typical sample output might look like,
F S UID PID PPID C PRI NI ADDR SZ WCHAN TTY TIME CMD

200001 R 21 15101 17095 13 66 20 196d 116 pts/0 0:00 ps
240801 S 21 17095 16070 3 61 20 1dce 108 pts/0 0:00 ksh
260801 S 0 3637 3112 0 60 20 b25 260 - 0:00 sendmail
260801 S 0 12169 1 0 60 20 8a5 152 hft/0 0:00 lmgrd
222801 S 0 12938 12169 0 60 20 16aa 352 hft/0 0:05 CFDSd
40801 S 0 10342 8542 0 60 20 357a 196 - 0:11 nfsd
The process name (or the command name) is shown in the last col-
umn. Other useful parameters are the process identification number
(PID), the nice value (NI) which determines the priority of the process,
and the cpu time (TIME) used up by the task. In the above example
listing, sendmail is the mail program, lmgrd is the license manager dae-
mon, CFDSd is the license server for FLOW3D program, nfsd is the NFS
daemon; all of these tasks are run by root with a user identification num-
ber (UID) of 0. Note that the ps command itself is a task.
C.3.2 Suspending jobs or processes

If you started a process like "emacs" or "matlab" and you want to sus-
pend that task and return to the shell you can do so with the key se-
quence
user@machine:dir> ctrl-z
The PID number is displayed at that time. Even if you did not note it
down, you can find a list of all suspended jobs with the command
user@machine:dir> jobs
To resume the job, enter
user@machine:dir> fg %n
where n is the job number produced by the jobs command (and not the
PID number!). The "fg" command brings a job to the foreground.
C.3.3 Terminating jobs or processes - the "kill" command

If you started a process in error and want to terminate it, you can use the
"kill" command. You need to find out the PID number of the process
using "ps command.
user@machine:dir> kill -9 PID
Except for the super user (or root), one can terminate only those pro-
cesses that belong to (or initiated by) individual users.
C.3.4 Initiating background jobs - the "nohup" command

The nohup command runs a command or a script ignoring all hangups
and QUIT signals. Use this command to run programs in the background
after logging off. To run a nohup command in the background, add an
& (ampersand) to the end of the command. Its syntax is:
user@machine:dir> nohup command or script &
When used in its simplest form as above, any output that would normally
appear on the screen will be saved in a file named nohup.out in the
current directory. Wait before logging off because the nohup command
takes a moment to start the command or script you specified. If you
log off too quickly, your command or script may not run at all. Once
your command or script starts, logging off does not affect it. Note that
in order to run a script, the script file must have execute permission.
C.3.5 Script files & scheduling jobs - the "at" command
If you wish to schedule a job to begin at a specified time (typically late

night), use the at command. The job should be constructed in the form
of a script file. For example a file named test.bat contains the following
lines and has execute permission enabled with the chmod command.
matlab >out « ’eof’

secant(’ass3a’,[0,1],eps,1)
fzero(’ass3a’,0.5,eps,1)
quit
eof
In the above script we start MATLAB in the first line and redirect any
output generated by MATLAB for the standard output ( i.e., screen during
an interactive session) to a file named out. During an interactive session,
MATLAB expects commands from the standard input ( i.e., the keyboard).
Such inputs are now taken from the script file itself as seen in the next
few lines where we execute some MATLAB functions and finally quit
MATLAB.
The contents of such a script file can be executed interactively while
logged in to a machine by simply entering the file name as
user@machine:dir> test.bat
The above example illustrates script programming in its simplest form.

It is possible to write very sophisticated scripts in the Kron shell or any
other shell. When you invoke MATLAB, for example, with the command
matlab, you are actually executing a powerful script. Browse through
the matlab script file using
user@machine:dir> pg /usr/local/matlab/bin/matlab
C.4. LIST OF OTHER USEFUL AIX COMMAND C.12
to appreciate the power of script programming.

Once a script file has been constructed, you can schedule it to be ex-
ecuted at a specified time using the at command as follows
user@machine:dir> at 21:00 scrit file
which will begin executing the script file at 21:00 hours. To examine a
listing of all the jobs scheduled use,
user@machine:dir> at -l
To remove a job that you have accidentally submitted, you can use,
user@machine:dir> at -r job number
C.4 List of other useful AIX command
A list of less frequently used AIX commands is given in Table C.1. You
can use either the man page feature with
or the
user@machine:dir> info
command which starts the InfoExplorer to find out about the syntax and
usage of these and other commands. The directory /usr/bin contains
all of the Unix commands.
C.4. LIST OF OTHER USEFUL AIX COMMAND C.13
command Function
at to schedule a task to start at a given time
cat to list a file
cd to change directory
diff compare two files
dosformat formats a floppy diskette using MS-DOS standards
dosread copies a DOS file from a floppy
doswrite copies a unix file to a DOS formatted floppy
find find a file
info InfoExplorer - online documentation
ksh start a Kron Shell
make a powerful UNIX make facility
mail read mail
mkdir create a directory
man display online manual pages
logout logout of current AIX session
lpq list the queue of print jobs
lpr send a print job to a network printer
lprm remove a print job from a queue
nice control job priority
nohup Don’t kill a process upon logout
pg display a file, one page at a time
ping to check if another machine is alive
pwd display present working directory
rlogin remote login to another machine
rcp remote copy files from one host to other
need to have ".rhosts" file setup
rm remove (delete) files
rmdir remove directories
rsh execute a command on a remote machine
need to have ".rhosts" file setup
rusers list remote users in the local area network
script logs a terminal session to a file
talk talk to another user currently signed on
tar archive files
telnet connect to remote hosts
whoami find out the current user
xinit start X-server
xlc c-compiler
xlC c++ compiler
xlf Fortran compiler
Table C.1: List of other useful AIX command

Métodos Numéricos en Ing. Química para Programar

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Métodos Numéricos en Ing. Química para Programar

Uploaded by

Copyright:

Available Formats

Contents

2 Single nonlinear algebraic equation 22

2.11 Software tools . . . . . . . . . . . . . . . . . . . . . . . . . . 43

3 Review of Linear Algebra 53

6 Ordinary differential equations - Initial value problems 133

6.3.1 Explicit schemes . . . . . . . . . . . . . . . . . . . . 141

7 Ordinary differential equations - Boundary value problems 153

A ChE Network A.1

A.5.6 File transfer with ftp . . . . . . . . . . . . . . . . . . A.22

B MATLAB primer B.1

B.3.11 Programming features . . . . . . . . . . . . . . . . . B.25

C Unix primer C.1

2.1 Example of fixed point iteration . . . . . . . . . . . . . . . 37

3.1 Loss of precision and need for pivoting . . . . . . . . . . . 71

5.1 Convergence of P2n+1 (x) to er f (x) at selected values of x 100

A.1 List of machines in the chemical and materials engineering

B.1 List of MATLAB Ver 4.0 help topics . . . . . . . . . . . . . B.6

C.1 List of other useful AIX command . . . . . . . . . . . . . . C.13

1.1 Scheme of things - a personal view . . . . . . . . . . . . . 2

2.1 Graphs of some simple functions . . . . . . . . . . . . . . 23

3.1 MATLAB implementation of inverse of an upper triangular

4.1 MATLAB implementation of Newton scheme . . . . . . . 93

5.1 Error distribution of (x, n) := |er f (x) − P2n+1 (x)| for

6.1 Spring and dash pot model . . . . . . . . . . . . . . . . . . 139

7.1 One dimensional finite difference grid of equally spaced

A.1 Anatomy of a Motif window . . . . . . . . . . . . . . . . . . A.4

C.1 Output of the "ls" command . . . . . . . . . . . . . . . . . C.5

The physical, mathematical and

In the Trekie language, the prime directive of a chemical engineer should

For Chemical engineers

Conceptual modelling Empirical observations

Construct Conservation laws Construct empirical

Figure 1.1: Scheme of things - a personal view

namic, lumped or distributed. Examples to illustrate each type of model

1.2 Classification of chemical process models

In modelling chemical processes, one is interested in tracking material

Figure 1.2: Three stage separation process

1.3 Lumped parameter, steady state models

1.3.1 Example of a stagewise separation process

trations of each stage and this relationship could be determined in a

where we have introduced the subscript n to indicate the equilibrium

Stage 1: V (y2 − y1 ) = L(x1 − x0 )

This, however, is not an independent equation since summing equations

Specifications: a 3-stage linear system

in the unknowns (x1 , x2 , x3 ). This can be represented in compact matrix

where T represents the tridiagonal matrix

First variation: n-stage separation sequence

Once we have expressed the mathematical model in a symbolic, matrix

y (a) linear model (b) nonlinear model

Figure 1.3: Linear and nonlinear equilibrium model

stages. Suppose there are n stages, then we merely have,

Second variation: nonlinear equilibrium model

Instead of assuming the equilibrium ratios, K in equation (1.1) to be

Third variation: alternate specification

In all of the previous cases we considered the inlet streams to be spec-

f1 (L, x2 ; V , y1 , x0 ) := V (Kx2 − y1 ) − L(y1 /K − x0 ) = 0

In spite of assuming a linear equilibrium model, the above set of equa-

1.3.2 Process flow sheet simulation

C-balance: x + R(0.302 + 0.004) = (0.274 + 0.095) × 275

loop 2 Find {x, y, R, P}

Figure 1.4: Example of material balance equations in a process flow

H2 -balance: y + R(0.694 + 0.008) = (0.631 + 2 × 0.095) × 275

5.1 Error distribution of (x, n) := |er f (x) − P2n+1 (x)| for