Fundamentals of CFD

Fundamentals of Computational Fluid
Dynamics
Harvard Lomax and Thomas H. Pulliam

NASA Ames Research Center
David W. Zingg
University of Toronto Institute for Aerospace Studies
August 26, 1999

Contents
1 INTRODUCTION 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Problem Specification and Geometry Preparation . . . . . . . 2
1.2.2 Selection of Governing Equations and Boundary Conditions . 3
1.2.3 Selection of Gridding Strategy and Numerical Method . . . . 3
1.2.4 Assessment and Interpretation of Results . . . . . . . . . . . . 4
1.3 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 CONSERVATION LAWS AND THE MODEL EQUATIONS 7

2.1 Conservation Laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 The Navier-Stokes and Euler Equations . . . . . . . . . . . . . . . . . 8
2.3 The Linear Convection Equation . . . . . . . . . . . . . . . . . . . . 12
2.3.1 Differential Form . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.2 Solution in Wave Space . . . . . . . . . . . . . . . . . . . . . . 13
2.4 The Diffusion Equation . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4.1 Differential Form . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4.2 Solution in Wave Space . . . . . . . . . . . . . . . . . . . . . . 15
2.5 Linear Hyperbolic Systems . . . . . . . . . . . . . . . . . . . . . . . . 16
2.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3 FINITE-DIFFERENCE APPROXIMATIONS 21
3.1 Meshes and Finite-Difference Notation . . . . . . . . . . . . . . . . . 21
3.2 Space Derivative Approximations . . . . . . . . . . . . . . . . . . . . 24
3.3 Finite-Difference Operators . . . . . . . . . . . . . . . . . . . . . . . 25
3.3.1 Point Difference Operators . . . . . . . . . . . . . . . . . . . . 25
3.3.2 Matrix Difference Operators . . . . . . . . . . . . . . . . . . . 25
3.3.3 Periodic Matrices . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3.4 Circulant Matrices . . . . . . . . . . . . . . . . . . . . . . . . 30
iii
3.4 Constructing Differencing Schemes of Any Order . . . . . . . . . . . . 31
3.4.1 Taylor Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.4.2 Generalization of Difference Formulas . . . . . . . . . . . . . . 34
3.4.3 Lagrange and Hermite Interpolation Polynomials . . . . . . . 35
3.4.4 Practical Application of Padé Formulas . . . . . . . . . . . . . 37
3.4.5 Other Higher-Order Schemes . . . . . . . . . . . . . . . . . . . 38
3.5 Fourier Error Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.5.1 Application to a Spatial Operator . . . . . . . . . . . . . . . . 39
3.6 Difference Operators at Boundaries . . . . . . . . . . . . . . . . . . . 43
3.6.1 The Linear Convection Equation . . . . . . . . . . . . . . . . 44
3.6.2 The Diffusion Equation . . . . . . . . . . . . . . . . . . . . . . 46
3.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4 THE SEMI-DISCRETE APPROACH 51

4.1 Reduction of PDE’s to ODE’s . . . . . . . . . . . . . . . . . . . . . . 52
4.1.1 The Model ODE’s . . . . . . . . . . . . . . . . . . . . . . . . 52
4.1.2 The Generic Matrix Form . . . . . . . . . . . . . . . . . . . . 53
4.2 Exact Solutions of Linear ODE’s . . . . . . . . . . . . . . . . . . . . 54
4.2.1 Eigensystems of Semi-Discrete Linear Forms . . . . . . . . . . 54
4.2.2 Single ODE’s of First- and Second-Order . . . . . . . . . . . . 55
4.2.3 Coupled First-Order ODE’s . . . . . . . . . . . . . . . . . . . 57
4.2.4 General Solution of Coupled ODE’s with Complete Eigensystems 59
4.3 Real Space and Eigenspace . . . . . . . . . . . . . . . . . . . . . . . . 61
4.3.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.3.2 Eigenvalue Spectrums for Model ODE’s . . . . . . . . . . . . . 62
4.3.3 Eigenvectors of the Model Equations . . . . . . . . . . . . . . 63
4.3.4 Solutions of the Model ODE’s . . . . . . . . . . . . . . . . . . 65
4.4 The Representative Equation . . . . . . . . . . . . . . . . . . . . . . 67
4.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5 FINITE-VOLUME METHODS 71
5.1 Basic Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.2 Model Equations in Integral Form . . . . . . . . . . . . . . . . . . . . 73
5.2.1 The Linear Convection Equation . . . . . . . . . . . . . . . . 73
5.2.2 The Diffusion Equation . . . . . . . . . . . . . . . . . . . . . . 74
5.3 One-Dimensional Examples . . . . . . . . . . . . . . . . . . . . . . . 74
5.3.1 A Second-Order Approximation to the Convection Equation . 75
5.3.2 A Fourth-Order Approximation to the Convection Equation . 77
5.3.3 A Second-Order Approximation to the Diffusion Equation . . 78
5.4 A Two-Dimensional Example . . . . . . . . . . . . . . . . . . . . . . 80
5.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6 TIME-MARCHING METHODS FOR ODE’S 85

6.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.2 Converting Time-Marching Methods to O∆E’s . . . . . . . . . . . . . 87
6.3 Solution of Linear O∆E’s With Constant Coefficients . . . . . . . . . 88
6.3.1 First- and Second-Order Difference Equations . . . . . . . . . 89
6.3.2 Special Cases of Coupled First-Order Equations . . . . . . . . 90
6.4 Solution of the Representative O∆E’s . . . . . . . . . . . . . . . . . 91
6.4.1 The Operational Form and its Solution . . . . . . . . . . . . . 91
6.4.2 Examples of Solutions to Time-Marching O∆E’s . . . . . . . . 92
6.5 The λ − σ Relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.5.1 Establishing the Relation . . . . . . . . . . . . . . . . . . . . . 93
6.5.2 The Principal σ-Root . . . . . . . . . . . . . . . . . . . . . . . 95
6.5.3 Spurious σ-Roots . . . . . . . . . . . . . . . . . . . . . . . . . 95
6.5.4 One-Root Time-Marching Methods . . . . . . . . . . . . . . . 96
6.6 Accuracy Measures of Time-Marching Methods . . . . . . . . . . . . 97
6.6.1 Local and Global Error Measures . . . . . . . . . . . . . . . . 97
6.6.2 Local Accuracy of the Transient Solution (erλ , |σ| , erω ) . . . . 98
6.6.3 Local Accuracy of the Particular Solution (erµ ) . . . . . . . . 99
6.6.4 Time Accuracy For Nonlinear Applications . . . . . . . . . . . 100
6.6.5 Global Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.7 Linear Multistep Methods . . . . . . . . . . . . . . . . . . . . . . . . 102
6.7.1 The General Formulation . . . . . . . . . . . . . . . . . . . . . 102
6.7.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.7.3 Two-Step Linear Multistep Methods . . . . . . . . . . . . . . 105
6.8 Predictor-Corrector Methods . . . . . . . . . . . . . . . . . . . . . . . 106
6.9 Runge-Kutta Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.10 Implementation of Implicit Methods . . . . . . . . . . . . . . . . . . . 110
6.10.1 Application to Systems of Equations . . . . . . . . . . . . . . 110
6.10.2 Application to Nonlinear Equations . . . . . . . . . . . . . . . 111
6.10.3 Local Linearization for Scalar Equations . . . . . . . . . . . . 112
6.10.4 Local Linearization for Coupled Sets of Nonlinear Equations . 115
6.11 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7 STABILITY OF LINEAR SYSTEMS 121

7.1 Dependence on the Eigensystem . . . . . . . . . . . . . . . . . . . . . 122
7.2 Inherent Stability of ODE’s . . . . . . . . . . . . . . . . . . . . . . . 123
7.2.1 The Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.2.2 Complete Eigensystems . . . . . . . . . . . . . . . . . . . . . . 123
7.2.3 Defective Eigensystems . . . . . . . . . . . . . . . . . . . . . . 123
7.3 Numerical Stability of O∆E ’s . . . . . . . . . . . . . . . . . . . . . . 124
7.3.1 The Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7.3.2 Complete Eigensystems . . . . . . . . . . . . . . . . . . . . . . 125
7.3.3 Defective Eigensystems . . . . . . . . . . . . . . . . . . . . . . 125
7.4 Time-Space Stability and Convergence of O∆E’s . . . . . . . . . . . . 125
7.5 Numerical Stability Concepts in the Complex σ-Plane . . . . . . . . . 128
7.5.1 σ-Root Traces Relative to the Unit Circle . . . . . . . . . . . 128
7.5.2 Stability for Small ∆t . . . . . . . . . . . . . . . . . . . . . . . 132
7.6 Numerical Stability Concepts in the Complex λh Plane . . . . . . . . 135
7.6.1 Stability for Large h. . . . . . . . . . . . . . . . . . . . . . . . 135
7.6.2 Unconditional Stability, A-Stable Methods . . . . . . . . . . . 136
7.6.3 Stability Contours in the Complex λh Plane. . . . . . . . . . . 137
7.7 Fourier Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . 141
7.7.1 The Basic Procedure . . . . . . . . . . . . . . . . . . . . . . . 141
7.7.2 Some Examples . . . . . . . . . . . . . . . . . . . . . . . . . . 142
7.7.3 Relation to Circulant Matrices . . . . . . . . . . . . . . . . . . 143
7.8 Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
7.9 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
8 CHOICE OF TIME-MARCHING METHODS 149

8.1 Stiffness Definition for ODE’s . . . . . . . . . . . . . . . . . . . . . . 149
8.1.1 Relation to λ-Eigenvalues . . . . . . . . . . . . . . . . . . . . 149
8.1.2 Driving and Parasitic Eigenvalues . . . . . . . . . . . . . . . . 151
8.1.3 Stiffness Classifications . . . . . . . . . . . . . . . . . . . . . . 151
8.2 Relation of Stiffness to Space Mesh Size . . . . . . . . . . . . . . . . 152
8.3 Practical Considerations for Comparing Methods . . . . . . . . . . . 153
8.4 Comparing the Efficiency of Explicit Methods . . . . . . . . . . . . . 154
8.4.1 Imposed Constraints . . . . . . . . . . . . . . . . . . . . . . . 154
8.4.2 An Example Involving Diffusion . . . . . . . . . . . . . . . . . 154
8.4.3 An Example Involving Periodic Convection . . . . . . . . . . . 155
8.5 Coping With Stiffness . . . . . . . . . . . . . . . . . . . . . . . . . . 158
8.5.1 Explicit Methods . . . . . . . . . . . . . . . . . . . . . . . . . 158
8.5.2 Implicit Methods . . . . . . . . . . . . . . . . . . . . . . . . . 159
8.5.3 A Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
8.6 Steady Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
8.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
9 RELAXATION METHODS 163
9.1 Formulation of the Model Problem . . . . . . . . . . . . . . . . . . . 164
9.1.1 Preconditioning the Basic Matrix . . . . . . . . . . . . . . . . 164
9.1.2 The Model Equations . . . . . . . . . . . . . . . . . . . . . . . 166
9.2 Classical Relaxation . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
9.2.1 The Delta Form of an Iterative Scheme . . . . . . . . . . . . . 168
9.2.2 The Converged Solution, the Residual, and the Error . . . . . 168
9.2.3 The Classical Methods . . . . . . . . . . . . . . . . . . . . . . 169
9.3 The ODE Approach to Classical Relaxation . . . . . . . . . . . . . . 170
9.3.1 The Ordinary Differential Equation Formulation . . . . . . . . 170
9.3.2 ODE Form of the Classical Methods . . . . . . . . . . . . . . 172
9.4 Eigensystems of the Classical Methods . . . . . . . . . . . . . . . . . 173
9.4.1 The Point-Jacobi System . . . . . . . . . . . . . . . . . . . . . 174
9.4.2 The Gauss-Seidel System . . . . . . . . . . . . . . . . . . . . . 176
9.4.3 The SOR System . . . . . . . . . . . . . . . . . . . . . . . . . 180
9.5 Nonstationary Processes . . . . . . . . . . . . . . . . . . . . . . . . . 182
9.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
10 MULTIGRID 191
10.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
10.1.1 Eigenvector and Eigenvalue Identification with Space Frequencies191
10.1.2 Properties of the Iterative Method . . . . . . . . . . . . . . . 192
10.2 The Basic Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
10.3 A Two-Grid Process . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
10.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
11 NUMERICAL DISSIPATION 203

11.1 One-Sided First-Derivative Space Differencing . . . . . . . . . . . . . 204
11.2 The Modified Partial Differential Equation . . . . . . . . . . . . . . . 205
11.3 The Lax-Wendroff Method . . . . . . . . . . . . . . . . . . . . . . . . 207
11.4 Upwind Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
11.4.1 Flux-Vector Splitting . . . . . . . . . . . . . . . . . . . . . . . 210
11.4.2 Flux-Difference Splitting . . . . . . . . . . . . . . . . . . . . . 212
11.5 Artificial Dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
11.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
12 SPLIT AND FACTORED FORMS 217

12.1 The Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
12.2 Factoring Physical Representations — Time Splitting . . . . . . . . . 218
12.3 Factoring Space Matrix Operators in 2–D . . . . . . . . . . . . . . . . 220
12.3.1 Mesh Indexing Convention . . . . . . . . . . . . . . . . . . . . 220
12.3.2 Data Bases and Space Vectors . . . . . . . . . . . . . . . . . . 221
12.3.3 Data Base Permutations . . . . . . . . . . . . . . . . . . . . . 221
12.3.4 Space Splitting and Factoring . . . . . . . . . . . . . . . . . . 223
12.4 Second-Order Factored Implicit Methods . . . . . . . . . . . . . . . . 226
12.5 Importance of Factored Forms in 2 and 3 Dimensions . . . . . . . . . 226
12.6 The Delta Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
12.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
13 LINEAR ANALYSIS OF SPLIT AND FACTORED FORMS 233

13.1 The Representative Equation for Circulant Operators . . . . . . . . . 233
13.2 Example Analysis of Circulant Systems . . . . . . . . . . . . . . . . . 234
13.2.1 Stability Comparisons of Time-Split Methods . . . . . . . . . 234
13.2.2 Analysis of a Second-Order Time-Split Method . . . . . . . . 237
13.3 The Representative Equation for Space-Split Operators . . . . . . . . 238
13.4 Example Analysis of 2-D Model Equations . . . . . . . . . . . . . . . 242
13.4.1 The Unfactored Implicit Euler Method . . . . . . . . . . . . . 242
13.4.2 The Factored Nondelta Form of the Implicit Euler Method . . 243
13.4.3 The Factored Delta Form of the Implicit Euler Method . . . . 243
13.4.4 The Factored Delta Form of the Trapezoidal Method . . . . . 244
13.5 Example Analysis of the 3-D Model Equation . . . . . . . . . . . . . 245
13.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
A USEFUL RELATIONS AND DEFINITIONS FROM LINEAR AL-

GEBRA 249
A.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
A.2 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
A.3 Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
A.4 Eigensystems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
A.5 Vector and Matrix Norms . . . . . . . . . . . . . . . . . . . . . . . . 254
B SOME PROPERTIES OF TRIDIAGONAL MATRICES 257

B.1 Standard Eigensystem for Simple Tridiagonals . . . . . . . . . . . . . 257
B.2 Generalized Eigensystem for Simple Tridiagonals . . . . . . . . . . . . 258
B.3 The Inverse of a Simple Tridiagonal . . . . . . . . . . . . . . . . . . . 259
B.4 Eigensystems of Circulant Matrices . . . . . . . . . . . . . . . . . . . 260
B.4.1 Standard Tridiagonals . . . . . . . . . . . . . . . . . . . . . . 260
B.4.2 General Circulant Systems . . . . . . . . . . . . . . . . . . . . 261
B.5 Special Cases Found From Symmetries . . . . . . . . . . . . . . . . . 262
B.6 Special Cases Involving Boundary Conditions . . . . . . . . . . . . . 263
C THE HOMOGENEOUS PROPERTY OF THE EULER EQUATIONS265
Chapter 1
INTRODUCTION
1.1 Motivation
The material in this book originated from attempts to understand and systemize nu-
merical solution techniques for the partial differential equations governing the physics
of fluid flow. As time went on and these attempts began to crystallize, underlying
constraints on the nature of the material began to form. The principal such constraint
was the demand for unification. Was there one mathematical structure which could
be used to describe the behavior and results of most numerical methods in common
use in the field of fluid dynamics? Perhaps the answer is arguable, but the authors
believe the answer is affirmative and present this book as justification for that be-
lief. The mathematical structure is the theory of linear algebra and the attendant
eigenanalysis of linear systems.
The ultimate goal of the field of computational fluid dynamics (CFD) is to under-
stand the physical events that occur in the flow of fluids around and within designated
objects. These events are related to the action and interaction of phenomena such
as dissipation, diffusion, convection, shock waves, slip surfaces, boundary layers, and
turbulence. In the field of aerodynamics, all of these phenomena are governed by
the compressible Navier-Stokes equations. Many of the most important aspects of
these relations are nonlinear and, as a consequence, often have no analytic solution.
This, of course, motivates the numerical solution of the associated partial differential
equations. At the same time it would seem to invalidate the use of linear algebra for
the classification of the numerical methods. Experience has shown that such is not
the case.
As we shall see in a later chapter, the use of numerical methods to solve partial
differential equations introduces an approximation that, in effect, can change the
form of the basic partial differential equations themselves. The new equations, which
1
2 CHAPTER 1. INTRODUCTION
are the ones actually being solved by the numerical process, are often referred to as
the modified partial differential equations. Since they are not precisely the same as
the original equations, they can, and probably will, simulate the physical phenomena
listed above in ways that are not exactly the same as an exact solution to the basic
partial differential equation. Mathematically, these differences are usually referred to
as truncation errors. However, the theory associated with the numerical analysis of
fluid mechanics was developed predominantly by scientists deeply interested in the
physics of fluid flow and, as a consequence, these errors are often identified with a
particular physical phenomenon on which they have a strong effect. Thus methods are
said to have a lot of “artificial viscosity” or said to be highly dispersive. This means
that the errors caused by the numerical approximation result in a modified partial
differential equation having additional terms that can be identified with the physics
of dissipation in the first case and dispersion in the second. There is nothing wrong,
of course, with identifying an error with a physical process, nor with deliberately
directing an error to a specific physical process, as long as the error remains in some
engineering sense “small”. It is safe to say, for example, that most numerical methods
in practical use for solving the nondissipative Euler equations create a modified partial
differential equation that produces some form of dissipation. However, if used and
interpreted properly, these methods give very useful information.
Regardless of what the numerical errors are called, if their effects are not thor-
oughly understood and controlled, they can lead to serious difficulties, producing
answers that represent little, if any, physical reality. This motivates studying the
concepts of stability, convergence, and consistency. On the other hand, even if the
errors are kept small enough that they can be neglected (for engineering purposes),
the resulting simulation can still be of little practical use if inefficient or inappropriate
algorithms are used. This motivates studying the concepts of stiffness, factorization,
and algorithm development in general. All of these concepts we hope to clarify in
this book.
1.2 Background
The field of computational fluid dynamics has a broad range of applicability. Indepen-
dent of the specific application under study, the following sequence of steps generally
must be followed in order to obtain a satisfactory solution.
1.2.1 Problem Specification and Geometry Preparation

The first step involves the specification of the problem, including the geometry, flow
conditions, and the requirements of the simulation. The geometry may result from
1.2. BACKGROUND 3
measurements of an existing configuration or may be associated with a design study.

Alternatively, in a design context, no geometry need be supplied. Instead, a set
of objectives and constraints must be specified. Flow conditions might include, for
example, the Reynolds number and Mach number for the flow over an airfoil. The
requirements of the simulation include issues such as the level of accuracy needed, the
turnaround time required, and the solution parameters of interest. The first two of
these requirements are often in conflict and compromise is necessary. As an example
of solution parameters of interest in computing the flowfield about an airfoil, one may
be interested in i) the lift and pitching moment only, ii) the drag as well as the lift
and pitching moment, or iii) the details of the flow at some specific location.
1.2.2 Selection of Governing Equations and Boundary Con-

ditions
Once the problem has been specified, an appropriate set of governing equations and
boundary conditions must be selected. It is generally accepted that the phenomena of
importance to the field of continuum fluid dynamics are governed by the conservation
of mass, momentum, and energy. The partial differential equations resulting from
these conservation laws are referred to as the Navier-Stokes equations. However, in
the interest of efficiency, it is always prudent to consider solving simplified forms
of the Navier-Stokes equations when the simplifications retain the physics which are
essential to the goals of the simulation. Possible simplified governing equations include
the potential-flow equations, the Euler equations, and the thin-layer Navier-Stokes
equations. These may be steady or unsteady and compressible or incompressible.
Boundary types which may be encountered include solid walls, inflow and outflow
boundaries, periodic boundaries, symmetry boundaries, etc. The boundary conditions
which must be specified depend upon the governing equations. For example, at a solid
wall, the Euler equations require flow tangency to be enforced, while the Navier-Stokes
equations require the no-slip condition. If necessary, physical models must be chosen
for processes which cannot be simulated within the specified constraints. Turbulence
is an example of a physical process which is rarely simulated in a practical context (at
the time of writing) and thus is often modelled. The success of a simulation depends
greatly on the engineering insight involved in selecting the governing equations and
physical models based on the problem specification.
1.2.3 Selection of Gridding Strategy and Numerical Method

Next a numerical method and a strategy for dividing the flow domain into cells, or
elements, must be selected. We concern ourselves here only with numerical meth-
ods requiring such a tessellation of the domain, which is known as a grid, or mesh.
Many different gridding strategies exist, including structured, unstructured, hybrid,

composite, and overlapping grids. Furthermore, the grid can be altered based on
the solution in an approach known as solution-adaptive gridding. The numerical
methods generally used in CFD can be classified as finite-difference, finite-volume,
finite-element, or spectral methods. The choices of a numerical method and a grid-
ding strategy are strongly interdependent. For example, the use of finite-difference
methods is typically restricted to structured grids. Here again, the success of a sim-
ulation can depend on appropriate choices for the problem or class of problems of
interest.
1.2.4 Assessment and Interpretation of Results

Finally, the results of the simulation must be assessed and interpreted. This step can
require post-processing of the data, for example calculation of forces and moments,
and can be aided by sophisticated flow visualization tools and error estimation tech-
niques. It is critical that the magnitude of both numerical and physical-model errors
be well understood.
1.3 Overview
It should be clear that successful simulation of fluid flows can involve a wide range of
issues from grid generation to turbulence modelling to the applicability of various sim-
plified forms of the Navier-Stokes equations. Many of these issues are not addressed
in this book. Some of them are presented in the books by Anderson, Tannehill, and
Pletcher [1] and Hirsch [2]. Instead we focus on numerical methods, with emphasis
on finite-difference and finite-volume methods for the Euler and Navier-Stokes equa-
tions. Rather than presenting the details of the most advanced methods, which are
still evolving, we present a foundation for developing, analyzing, and understanding
such methods.
Fortunately, to develop, analyze, and understand most numerical methods used to
find solutions for the complete compressible Navier-Stokes equations, we can make use
of much simpler expressions, the so-called “model” equations. These model equations
isolate certain aspects of the physics contained in the complete set of equations. Hence
their numerical solution can illustrate the properties of a given numerical method
when applied to a more complicated system of equations which governs similar phys-
ical phenomena. Although the model equations are extremely simple and easy to
solve, they have been carefully selected to be representative, when used intelligently,
of difficulties and complexities that arise in realistic two- and three-dimensional fluid
flow simulations. We believe that a thorough understanding of what happens when
1.4. NOTATION 5
numerical approximations are applied to the model equations is a major first step in
making confident and competent use of numerical approximations to the Euler and
Navier-Stokes equations. As a word of caution, however, it should be noted that,
although we can learn a great deal by studying numerical methods as applied to the
model equations and can use that information in the design and application of nu-
merical methods to practical problems, there are many aspects of practical problems
which can only be understood in the context of the complete physical systems.
1.4 Notation
The notation is generally explained as it is introduced. Bold type is reserved for real
physical vectors, such as velocity. The vector symbol is used for the vectors (or
column matrices) which contain the values of the dependent variable at the nodes
of a grid. Otherwise, the use of a vector consisting of a collection of scalars should
be apparent from the context and is not identified by any special notation. For
example, the variable u can denote a scalar Cartesian velocity component in the Euler
and Navier-Stokes equations, a scalar quantity in the linear convection and diffusion
equations, and a vector consisting of a collection of scalars in our presentation of
hyperbolic systems. Some of the abbreviations used throughout the text are listed
and defined below.
PDE Partial differential equation

ODE Ordinary differential equation
O∆E Ordinary difference equation
RHS Right-hand side
P.S. Particular solution of an ODE or system of ODE’s
S.S. Fixed (time-invariant) steady-state solution
k-D
k-dimensional space

bc Boundary conditions, usually a vector
O(α) A term of order (i.e., proportional to) α
Chapter 2
CONSERVATION LAWS AND

THE MODEL EQUATIONS
We start out by casting our equations in the most general form, the integral conserva-
tion-law form, which is useful in understanding the concepts involved in finite-volume
schemes. The equations are then recast into divergence form, which is natural for
finite-difference schemes. The Euler and Navier-Stokes equations are briefly discussed
in this Chapter. The main focus, though, will be on representative model equations,
in particular, the convection and diffusion equations. These equations contain many
of the salient mathematical and physical features of the full Navier-Stokes equations.
The concepts of convection and diffusion are prevalent in our development of nu-
merical methods for computational fluid dynamics, and the recurring use of these
model equations allows us to develop a consistent framework of analysis for consis-
tency, accuracy, stability, and convergence. The model equations we study have two
properties in common. They are linear partial differential equations (PDE’s) with
coefficients that are constant in both space and time, and they represent phenomena
of importance to the analysis of certain aspects of fluid dynamic problems.
2.1 Conservation Laws

Conservation laws, such as the Euler and Navier-Stokes equations and our model
equations, can be written in the following integral form:
t2 t2
QdV − QdV + n.FdSdt = P dV dt (2.1)
V (t2 ) V (t1 ) t1 S(t) t1 V (t)
In this equation, Q is a vector containing the set of variables which are conserved,
e.g., mass, momentum, and energy, per unit volume. The equation is a statement of
7
8 CHAPTER 2. CONSERVATION LAWS AND THE MODEL EQUATIONS
the conservation of these quantities in a finite region of space with volume V (t) and
surface area S(t) over a finite interval of time t2 − t1 . In two dimensions, the region
of space, or cell, is an area A(t) bounded by a closed contour C(t). The vector n is
a unit vector normal to the surface pointing outward, F is a set of vectors, or tensor,
containing the flux of Q per unit area per unit time, and P is the rate of production
of Q per unit volume per unit time. If all variables are continuous in time, then Eq.
2.1 can be rewritten as

d
QdV + n.FdS = P dV (2.2)
dt V (t) S(t) V (t)
Those methods which make various numerical approximations of the integrals in Eqs.
2.1 and 2.2 and find a solution for Q on that basis are referred to as finite-volume
methods. Many of the advanced codes written for CFD applications are based on the
finite-volume concept.
On the other hand, a partial derivative form of a conservation law can also be
derived. The divergence form of Eq. 2.2 is obtained by applying Gauss’s theorem to
the flux integral, leading to
∂Q
+ ∇.F = P (2.3)
∂t
where ∇. is the well-known divergence operator given, in Cartesian coordinates, by

∂ ∂ ∂
∇. ≡ i +j +k . (2.4)
∂x ∂y ∂z
and i, j, and k are unit vectors in the x, y, and z coordinate directions, respectively.
Those methods which make various approximations of the derivatives in Eq. 2.3 and
find a solution for Q on that basis are referred to as finite-difference methods.
2.2 The Navier-Stokes and Euler Equations

The Navier-Stokes equations form a coupled system of nonlinear PDE’s describing
the conservation of mass, momentum and energy for a fluid. For a Newtonian fluid
in one dimension, they can be written as
∂Q ∂E
+ =0 (2.5)
∂t ∂x
with      
ρ ρu 0
     
     
     
Q=  ρu  , E=  2
ρu + p − 4 ∂u
µ  (2.6)
     3 ∂x 
     
4
e u(e + p) 3
µu ∂u
∂x
+ κ ∂T
∂x
2.2. THE NAVIER-STOKES AND EULER EQUATIONS 9
where ρ is the fluid density, u is the velocity, e is the total energy per unit volume, p is
the pressure, T is the temperature, µ is the coefficient of viscosity, and κ is the thermal
conductivity. The total energy e includes internal energy per unit volume ρ (where
is the internal energy per unit mass) and kinetic energy per unit volume ρu2 /2.
These equations must be supplemented by relations between µ and κ and the fluid
state as well as an equation of state, such as the ideal gas law. Details can be found
in Anderson, Tannehill, and Pletcher [1] and Hirsch [2]. Note that the convective
fluxes lead to first derivatives in space, while the viscous and heat conduction terms
involve second derivatives. This form of the equations is called conservation-law or
conservative form. Non-conservative forms can be obtained by expanding derivatives
of products using the product rule or by introducing different dependent variables,
such as u and p. Although non-conservative forms of the equations are analytically
the same as the above form, they can lead to quite different numerical solutions in
terms of shock strength and shock speed, for example. Thus the conservative form is
appropriate for solving flows with features such as shock waves.
Many flows of engineering interest are steady (time-invariant), or at least may be
treated as such. For such flows, we are often interested in the steady-state solution of
the Navier-Stokes equations, with no interest in the transient portion of the solution.
The steady solution to the one-dimensional Navier-Stokes equations must satisfy
∂E
=0 (2.7)
∂x
If we neglect viscosity and heat conduction, the Euler equations are obtained. In
two-dimensional Cartesian coordinates, these can be written as
∂Q ∂E ∂F
+ + =0 (2.8)
∂t ∂x ∂y
with        
q1 ρ ρu ρv
q   ρu   ρu2 + p   ρuv 
Q= = 
 2     
, E = , F = 2  (2.9)
 q3   ρv   ρuv   ρv + p 
q4 e u(e + p) v(e + p)
where u and v are the Cartesian velocity components. Later on we will make use of
the following form of the Euler equations as well:
∂Q ∂Q ∂Q
+A +B =0 (2.10)
∂t ∂x ∂y
The matrices A = ∂∂Q

E ∂F
and B = ∂Q are known as the flux Jacobians. The flux vectors
given above are written in terms of the primitive variables, ρ, u, v, and p. In order
to derive the flux Jacobian matrices, we must first write the flux vectors E and F in
terms of the conservative variables, q1 , q2 , q3 , and q4 , as follows:
 
  q2
E1  
   
   
   3−γ q22 γ−1 q32

   (γ − 1)q4 + − 
 E2   2 q1 2 q1 
   
E=   =   (2.11)
   
   q3 q2 
 E3   q1 
   
   
 
E4  
q23 q32 q2
γ q4q1q2 − γ−1
2 q12
+ q12
 
 
F1 q3
 
   
   
   q3 q2 
 F2   
   q1 
   
F =   =   (2.12)
   3−γ q32 γ−1 q22

 F3   (γ − 1)q4 + − 
   2 q1 2 q1 
   
   
 
F4 q22 q3 q33
γ q4q1q3 − γ−1
2 q12
+ q12
We have assumed that the pressure satisfies p = (γ − 1)[e − ρ(u2 + v 2 )/2] from the
ideal gas law, where γ is the ratio of specific heats, cp /cv . From this it follows that
the flux Jacobian of E can be written in terms of the conservative variables as
 
0 1 0 0
 
 
 

 a21 (3 − γ) qq21 (1 − γ) qq31 γ−1 
∂Ei  
 
A= =   (2.13)
∂qj  q3
q2 

 − qq21 qq31 q1 q1 0 
 
 
 
a41 a42 a43 q
γ q21
where
2 2
γ − 1 q3 3 − γ q2
a21 = −
2 q1 2 q1
2.2. THE NAVIER-STOKES AND EULER EQUATIONS 11
 3 2 
q2 q3 q2  q4 q2
a41 = (γ − 1) + −γ
q1 q1 q1 q1 q1
 2 2 
q4 γ − 1  q2 q3 
a42 = γ − 3 +
q1 2 q1 q1

q2 q3
a43 = −(γ − 1) (2.14)
q1 q1
and in terms of the primitive variables as
 
0 1 0 0
 
 
 
 a21 (3 − γ)u (1 − γ)v (γ − 1) 
 
 
A=   (2.15)
 
 −uv v u 0 
 
 
 
a41 a42 a43 γu
where
γ−1 2 3−γ 2
a21 = v − u
2 2
ue
a41 = (γ − 1)u(u2 + v 2 ) − γ
ρ
e γ−1
a42 = γ − (3u2 + v 2 )
ρ 2
a43 = (1 − γ)uv (2.16)

Derivation of the two forms of B = ∂F/∂Q is similar. The eigenvalues of the flux
Jacobian matrices are purely real. This is the defining feature of hyperbolic systems
of PDE’s, which are further discussed in Section 2.5. The homogeneous property of
the Euler equations is discussed in Appendix C.
The Navier-Stokes equations include both convective and diffusive fluxes. This
motivates the choice of our two scalar model equations associated with the physics
of convection and diffusion. Furthermore, aspects of convective phenomena associ-
ated with coupled systems of equations such as the Euler equations are important in
developing numerical methods and boundary conditions. Thus we also study linear
hyperbolic systems of PDE’s.
2.3 The Linear Convection Equation

2.3.1 Differential Form
The simplest linear model for convection and wave propagation is the linear convection
equation given by the following PDE:
∂u ∂u
+a =0 (2.17)
∂t ∂x
Here u(x, t) is a scalar quantity propagating with speed a, a real constant which may
be positive or negative. The manner in which the boundary conditions are specified
separates the following two phenomena for which this equation is a model:
(1) In one type, the scalar quantity u is given on one boundary, corresponding
to a wave entering the domain through this “inflow” boundary. No bound-
ary condition is specified at the opposite side, the “outflow” boundary. This
is consistent in terms of the well-posedness of a 1st -order PDE. Hence the
wave leaves the domain through the outflow boundary without distortion or
reflection. This type of phenomenon is referred to, simply, as the convection
problem. It represents most of the “usual” situations encountered in convect-
ing systems. Note that the left-hand boundary is the inflow boundary when
a is positive, while the right-hand boundary is the inflow boundary when a is
negative.
(2) In the other type, the flow being simulated is periodic. At any given time,
what enters on one side of the domain must be the same as that which is
leaving on the other. This is referred to as the biconvection problem. It is
the simplest to study and serves to illustrate many of the basic properties of
numerical methods applied to problems involving convection, without special
consideration of boundaries. Hence, we pay a great deal of attention to it in
the initial chapters.
Now let us consider a situation in which the initial condition is given by u(x, 0) =
u0 (x), and the domain is infinite. It is easy to show by substitution that the exact
solution to the linear convection equation is then
u(x, t) = u0 (x − at) (2.18)
The initial waveform propagates unaltered with speed |a| to the right if a is positive
and to the left if a is negative. With periodic boundary conditions, the waveform
travels through one boundary and reappears at the other boundary, eventually re-
turning to its initial position. In this case, the process continues forever without any
2.3. THE LINEAR CONVECTION EQUATION 13
change in the shape of the solution. Preserving the shape of the initial condition
u0 (x) can be a difficult challenge for a numerical method.
2.3.2 Solution in Wave Space

We now examine the biconvection problem in more detail. Let the domain be given
by 0 ≤ x ≤ 2π. We restrict our attention to initial conditions in the form
u(x, 0) = f (0)eiκx (2.19)
where f (0) is a complex constant, and κ is the wavenumber. In order to satisfy the
periodic boundary conditions, κ must be an integer. It is a measure of the number of
wavelengths within the domain. With such an initial condition, the solution can be
written as
u(x, t) = f (t)eiκx (2.20)
where the time dependence is contained in the complex function f (t). Substituting
this solution into the linear convection equation, Eq. 2.17, we find that f (t) satisfies
the following ordinary differential equation (ODE)
df
= −iaκf (2.21)
dt
which has the solution
f (t) = f (0)e−iaκt (2.22)
Substituting f (t) into Eq. 2.20 gives the following solution
u(x, t) = f (0)eiκ(x−at) = f (0)ei(κx−ωt) (2.23)
where the frequency, ω, the wavenumber, κ, and the phase speed, a, are related by
ω = κa (2.24)
The relation between the frequency and the wavenumber is known as the dispersion
relation. The linear relation given by Eq. 2.24 is characteristic of wave propagation
in a nondispersive medium. This means that the phase speed is the same for all
wavenumbers. As we shall see later, most numerical methods introduce some disper-
sion; that is, in a simulation, waves with different wavenumbers travel at different
speeds.
An arbitrary initial waveform can be produced by summing initial conditions of
the form of Eq. 2.19. For M modes, one obtains

M
u(x, 0) = fm (0)eiκm x (2.25)
m=1
where the wavenumbers are often ordered such that κ1 ≤ κ2 ≤ · · · ≤ κM . Since the
wave equation is linear, the solution is obtained by summing solutions of the form of
Eq. 2.23, giving

M
u(x, t) = fm (0)eiκm (x−at) (2.26)
m=1
Dispersion and dissipation resulting from a numerical approximation will cause the
shape of the solution to change from that of the original waveform.
2.4 The Diffusion Equation

2.4.1 Differential Form
Diffusive fluxes are associated with molecular motion in a continuum fluid. A simple
linear model equation for a diffusive process is
∂u ∂2u
=ν 2 (2.27)
∂t ∂x
where ν is a positive real constant. For example, with u representing the tempera-
ture, this parabolic PDE governs the diffusion of heat in one dimension. Boundary
conditions can be periodic, Dirichlet (specified u), Neumann (specified ∂u/∂x), or
mixed Dirichlet/Neumann.
In contrast to the linear convection equation, the diffusion equation has a nontrivial
steady-state solution, which is one that satisfies the governing PDE with the partial
derivative in time equal to zero. In the case of Eq. 2.27, the steady-state solution
must satisfy
∂2u
=0 (2.28)
∂x2
Therefore, u must vary linearly with x at steady state such that the boundary con-
ditions are satisfied. Other steady-state solutions are obtained if a source term g(x)
is added to Eq. 2.27, as follows:

∂u ∂2u
=ν − g(x) (2.29)
∂t ∂x2
giving a steady state-solution which satisfies
∂2 u
− g(x) = 0 (2.30)
∂x2
2.4. THE DIFFUSION EQUATION 15
In two dimensions, the diffusion equation becomes

∂u ∂2 u ∂2 u
=ν + − g(x, y) (2.31)
∂t ∂x2 ∂y 2
where g(x, y) is again a source term. The corresponding steady equation is
∂2u ∂2u
+ − g(x, y) = 0 (2.32)
∂x2 ∂y 2
While Eq. 2.31 is parabolic, Eq. 2.32 is elliptic. The latter is known as the Poisson
equation for nonzero g, and as Laplace’s equation for zero g.
2.4.2 Solution in Wave Space

We now consider a series solution to Eq. 2.27. Let the domain be given by 0 ≤ x ≤ π
with boundary conditions u(0) = ua , u(π) = ub . It is clear that the steady-state
solution is given by a linear function which satisfies the boundary conditions, i.e.,
h(x) = ua + (ub − ua )x/π. Let the initial condition be

M
u(x, 0) = fm (0) sin κm x + h(x) (2.33)
m=1
where κ must be an integer in order to satisfy the boundary conditions. A solution

of the form

M
u(x, t) = fm (t) sin κm x + h(x) (2.34)
m=1
satisfies the initial and boundary conditions. Substituting this form into Eq. 2.27
gives the following ODE for fm :
dfm
= −κ2m νfm (2.35)
dt
and we find
fm (t) = fm (0)e−κm νt
2
(2.36)
Substituting fm (t) into equation 2.34, we obtain

M
fm (0)e−κm νt sin κm x + h(x)
2
u(x, t) = (2.37)
m=1
The steady-state solution (t → ∞) is simply h(x). Eq. 2.37 shows that high wavenum-
ber components (large κm ) of the solution decay more rapidly than low wavenumber
components, consistent with the physics of diffusion.
2.5 Linear Hyperbolic Systems

The Euler equations, Eq. 2.8, form a hyperbolic system of partial differential equa-
tions. Other systems of equations governing convection and wave propagation phe-
nomena, such as the Maxwell equations describing the propagation of electromagnetic
waves, are also of hyperbolic type. Many aspects of numerical methods for such sys-
tems can be understood by studying a one-dimensional constant-coefficient linear
system of the form
∂u ∂u
+A =0 (2.38)
∂t ∂x
where u = u(x, t) is a vector of length m and A is a real m × m matrix. For
conservation laws, this equation can also be written in the form
∂u ∂f
+ =0 (2.39)
∂t ∂x
where f is the flux vector and A = ∂∂uf is the flux Jacobian matrix. The entries in the
flux Jacobian are
∂fi
aij = (2.40)
∂uj
The flux Jacobian for the Euler equations is derived in Section 2.2.
Such a system is hyperbolic if A is diagonalizable with real eigenvalues.1 Thus
Λ = X −1 AX (2.41)
where Λ is a diagonal matrix containing the eigenvalues of A, and X is the matrix
of right eigenvectors. Premultiplying Eq. 2.38 by X−1 , postmultiplying A by the
product XX −1 , and noting that X and X −1 are constants, we obtain
Λ

∂X −1 u ∂ X −1 AX X −1 u
+ =0 (2.42)
∂t ∂x
With w = X −1 u, this can be rewritten as
∂w ∂w
+Λ =0 (2.43)
∂t ∂x
When written in this manner, the equations have been decoupled into m scalar equa-
tions of the form
∂wi ∂wi
+ λi =0 (2.44)
∂t ∂x
1
See Appendix A for a brief review of some basic relations and definitions from linear algebra.
2.6. PROBLEMS 17
The elements of w are known as characteristic variables. Each characteristic variable

satisfies the linear convection equation with the speed given by the corresponding
eigenvalue of A.
Based on the above, we see that a hyperbolic system in the form of Eq. 2.38 has a
solution given by the superposition of waves which can travel in either the positive or
negative directions and at varying speeds. While the scalar linear convection equation
is clearly an excellent model equation for hyperbolic systems, we must ensure that
our numerical methods are appropriate for wave speeds of arbitrary sign and possibly
widely varying magnitudes.
The one-dimensional Euler equations can also be diagonalized, leading to three
equations in the form of the linear convection equation, although they remain non-
linear, of course. The eigenvalues of the flux Jacobian matrix, or wave speeds, are
u, u + c, and u − c, where u is the local fluid velocity, and c = γp/ρ is the local
speed of sound. The speed u is associated with convection of the fluid, while u + c
and u − c are associated with sound waves. Therefore, in a supersonic flow, where
|u| > c, all of the wave speeds have the same sign. In a subsonic flow, where |u| < c,
wave speeds of both positive and negative sign are present, corresponding to the fact
that sound waves can travel upstream in a subsonic flow.
The signs of the eigenvalues of the matrix A are also important in determining
suitable boundary conditions. The characteristic variables each satisfy the linear con-
vection equation with the wave speed given by the corresponding eigenvalue. There-
fore, the boundary conditions can be specified accordingly. That is, characteristic
variables associated with positive eigenvalues can be specified at the left boundary,
which corresponds to inflow for these variables. Characteristic variables associated
with negative eigenvalues can be specified at the right boundary, which is the in-
flow boundary for these variables. While other boundary condition treatments are
possible, they must be consistent with this approach.
2.6 Problems
1. Show that the 1-D Euler equations can be written in terms of the primitive
variables R = [ρ, u, p]T as follows:
∂R ∂R
+M =0
∂t ∂x
where
 
u ρ 0
M = 
 0 u ρ−1 
0 γp u
Assume an ideal gas, p = (γ − 1)(e − ρu2 /2).
2. Find the eigenvalues and eigenvectors of the matrix M derived in question 1.
3. Derive the flux Jacobian matrix A = ∂E/∂Q for the 1-D Euler equations result-
ing from the conservative variable formulation (Eq. 2.5). Find its eigenvalues
and compare with those obtained in question 2.
4. Show that the two matrices M and A derived in questions 1 and 3, respectively,
are related by a similarity transform. (Hint: make use of the matrix S =
∂Q/∂R.)
5. Write the 2-D diffusion equation, Eq. 2.31, in the form of Eq. 2.2.
6. Given the initial condition u(x, 0) = sin x defined on 0 ≤ x ≤ 2π, write it in the
form of Eq. 2.25, that is, find the necessary values of fm (0). (Hint: use M = 2
with κ1 = 1 and κ2 = −1.) Next consider the same initial condition defined
only at x = 2πj/4, j = 0, 1, 2, 3. Find the values of fm (0) required to reproduce
the initial condition at these discrete points using M = 4 with κm = m − 1.
7. Plot the first three basis functions used in constructing the exact solution to
the diffusion equation in Section 2.4.2. Next consider a solution with boundary
conditions ua = ub = 0, and initial conditions from Eq. 2.33 with fm (0) = 1
for 1 ≤ m ≤ 3, fm (0) = 0 for m > 3. Plot the initial condition on the domain
0 ≤ x ≤ π. Plot the solution at t = 1 with ν = 1.
8. Write the classical wave equation ∂2 u/∂t2 = c2 ∂ 2 u/∂x2 as a first-order system,

i.e., in the form
∂U ∂U
+A =0
∂t ∂x
where U = [∂u/∂x, ∂u/∂t]T . Find the eigenvalues and eigenvectors of A.
9. The Cauchy-Riemann equations are formed from the coupling of the steady
compressible continuity (conservation of mass) equation
∂ρu ∂ρv
+ =0
∂x ∂y
and the vorticity definition
∂v ∂u
ω=− + =0
∂x ∂y
2.6. PROBLEMS 19
where ω = 0 for irrotational flow. For isentropic and homenthalpic flow, the
system is closed by the relation

γ−1 2 1
γ−1
ρ= 1− u + v2 − 1
2
Note that the variables have been nondimensionalized. Combining the two
PDE’s, we have
∂f (q) ∂g(q)
+ =0
∂x ∂y
where

u −ρu −ρv
q= , f= , g=
v v −u
One approach to solving these equations is to add a time-dependent term and

find the steady solution of the following equation:
∂q ∂f ∂g
+ + =0
∂t ∂x ∂y
(a) Find the flux Jacobians of f and g with respect to q.

(b) Determine the eigenvalues of the flux Jacobians.
(c) Determine the conditions (in terms of ρ and u) under which the system is
hyperbolic, i.e., has real eigenvalues.
(d) Are the above fluxes homogeneous? (See Appendix C.)
Chapter 3
FINITE-DIFFERENCE
APPROXIMATIONS
In common with the equations governing unsteady fluid flow, our model equations
contain partial derivatives with respect to both space and time. One can approxi-
mate these simultaneously and then solve the resulting difference equations. Alterna-
tively, one can approximate the spatial derivatives first, thereby producing a system
of ordinary differential equations. The time derivatives are approximated next, lead-
ing to a time-marching method which produces a set of difference equations. This
is the approach emphasized here. In this chapter, the concept of finite-difference
approximations to partial derivatives is presented. These can be applied either to
spatial derivatives or time derivatives. Our emphasis in this chapter is on spatial
derivatives; time derivatives are treated in Chapter 6. Strategies for applying these
finite-difference approximations will be discussed in Chapter 4.
All of the material below is presented in a Cartesian system. We emphasize the
fact that quite general classes of meshes expressed in general curvilinear coordinates
in physical space can be transformed to a uniform Cartesian mesh with equispaced
intervals in a so-called computational space, as shown in Figure 3.1. The computational
space is uniform; all the geometric variation is absorbed into variable coefficients of the
transformed equations. For this reason, in much of the following accuracy analysis,
we use an equispaced Cartesian system without being unduly restrictive or losing
practical application.
3.1 Meshes and Finite-Difference Notation

The simplest mesh involving both time and space is shown in Figure 3.2. Inspection
of this figure permits us to define the terms and notation needed to describe finite-
21
22 CHAPTER 3. FINITE-DIFFERENCE APPROXIMATIONS
Figure 3.1: Physical and computational spaces.
difference approximations. In general, the dependent variables, u, for example, are

functions of the independent variables t, and x, y, z. For the first several chapters
we consider primarily the 1-D case u = u(x, t). When only one variable is denoted,
dependence on the other is assumed. The mesh index for x is always j, and that for
t is always n. Then on an equispaced grid
x = xj = j∆x (3.1)
t = tn = n∆t = nh (3.2)
where ∆x is the spacing in x and ∆t the spacing in t, as shown in Figure 3.2. Note
that h = ∆t throughout. Later k and l are used for y and z in a similar way. When
n, j, k, l are used for other purposes (which is sometimes necessary), local context
should make the meaning obvious.
The convention for subscript and superscript indexing is as follows:
u(t + kh) = u([n + k]h) = un+k
u(x + m∆x) = u([j + m]∆x) = uj+m (3.3)
(n+k)
u(x + m∆x, t + kh) = uj+m
Notice that when used alone, both the time and space indices appear as a subscript,
but when used together, time is always a superscript and is usually enclosed with
parentheses to distinguish it from an exponent.
3.1. MESHES AND FINITE-DIFFERENCE NOTATION 23
t
Grid or
Node
Points
n+1
∆x
n
∆t
n-1
j-2 j-1 j j+1 j+2 x
Figure 3.2: Space-time grid arrangement.
Derivatives are expressed according to the usual conventions. Thus for partial
derivatives in space or time we use interchangeably
∂u ∂u ∂2 u
∂x u = , ∂t u = , ∂xx u = , etc. (3.4)
∂x ∂t ∂x2
For the ordinary time derivative in the study of ODE’s we use
du
u = (3.5)
dt
In this text, subscripts on dependent variables are never used to express derivatives.
Thus ux will not be used to represent the first derivative of u with respect to x.
The notation for difference approximations follows the same philosophy, but (with
one exception) it is not unique. By this we mean that the symbol δ is used to represent
a difference approximation to a derivative such that, for example,
δx ≈ ∂x , δxx ≈ ∂xx (3.6)
but the precise nature (and order) of the approximation is not carried in the symbol
δ. Other ways are used to determine its precise meaning. The one exception is the
symbol ∆, which is defined such that
∆tn = tn+1 − tn , ∆xj = xj+1 − xj , ∆un = un+1 − un, etc. (3.7)
When there is no subscript on ∆t or ∆x, the spacing is uniform.

3.2 Space Derivative Approximations

A difference approximation can be generated or evaluated by means of a simple Taylor
series expansion. For example, consider u(x, t) with t fixed. Then, following the
notation convention given in Eqs. 3.1 to 3.3, x = j∆x and u(x + k∆x) = u(j∆x +
k∆x) = uj+k . Expanding the latter term about x gives1

∂u 1 ∂2u 1 ∂nu
uj+k = uj + (k∆x) + (k∆x)2 + · · · + (k∆x)n + · · · (3.8)
∂x j
2 ∂x2 j
n! ∂xn j
Local difference approximations to a given partial derivative can be formed from linear
combinations of uj and uj+k for k = ±1, ±2, · · ·.
For example, consider the Taylor series expansion for uj+1:

∂u 1 ∂2u 1 ∂nu
uj+1 = uj + (∆x) + (∆x)2 + · · · + (∆x)n +··· (3.9)
∂x j
2 ∂x2 j
n! ∂xn j
Now subtract uj and divide by ∆x to obtain

uj+1 − uj ∂u 1 ∂2u
= + (∆x) +··· (3.10)
∆x ∂x j
2 ∂x2 j

Thus the expression (uj+1 −uj )/∆x is a reasonable approximation for ∂∂xu as long as
j
∆x is small relative to some pertinent length scale. Next consider the space difference
approximation (uj+1 − uj−1 )/(2∆x). Expand the terms in the numerator about j and
regroup the result to form the following equation

uj+1 − uj−1 ∂u 1 ∂3 u 1 ∂5 u
− = ∆x2 + ∆x4 ... (3.11)
2∆x ∂x j
6 ∂x3 j
120 ∂x5 j
When expressed in this manner, it is clear that the discrete terms on the left side of
the equation represent a first derivative with a certain amount of error which appears
on the right side of the equal sign. It is also clear that the error depends on the
grid spacing to a certain order. The error term containing the grid spacing to the
lowest power gives the order of the method. From Eq.
3.10, we see that the expression
(uj+1 −uj )/∆x is a first-order approximation to ∂x . Similarly, Eq. 3.11 shows that
∂u
j
(uj+1 − uj−1)/(2∆x) is a second-order approximation to a first derivative. The latter
is referred to as the three-point centered difference approximation, and one often sees
the summary result presented in the form

∂u uj+1 − uj−1
= + O(∆x2 ) (3.12)
∂x j
2∆x
1
We assume that u(x, t) is continuously differentiable.
3.3. FINITE-DIFFERENCE OPERATORS 25
3.3 Finite-Difference Operators

3.3.1 Point Difference Operators
Perhaps the most common examples of finite-difference formulas are the three-point
centered-difference approximations for the first and second derivatives:2

∂u 1
= (uj+1 − uj−1) + O(∆x2 ) (3.13)
∂x j
2∆x

∂2u 1
= (uj+1 − 2uj + uj−1) + O(∆x2 ) (3.14)
∂x2 j ∆x2
These are the basis for point difference operators since they give an approximation to
a derivative at one discrete point in a mesh in terms of surrounding points. However,
neither of these expressions tells us how other points in the mesh are differenced or
how boundary conditions are enforced. Such additional information requires a more
sophisticated formulation.
3.3.2 Matrix Difference Operators

Consider the relation
1
(δxx u)j = (uj+1 − 2uj + uj−1 ) (3.15)
∆x2
which is a point difference approximation to a second derivative. Now let us derive a
matrix operator representation for the same approximation. Consider the four point
mesh with boundary points at a and b shown below. Notice that when we speak of
“the number of points in a mesh”, we mean the number of interior points excluding
the boundaries.
a 1 2 3 4 b
x= 0 − − − − π
j= 1 · · M
Four point mesh. ∆x = π/(M + 1)
Now impose Dirichlet boundary conditions, u(0) = ua , u(π) = ub and use the
centered difference approximation given by Eq. 3.15 at every point in the mesh. We
2
We will derive the second derivative operator shortly.
arrive at the four equations:
1
(δxx u)1 = (ua − 2u1 + u2 )
∆x2
1
(δxx u)2 = (u1 − 2u2 + u3)
∆x2
1
(δxx u)3 = (u2 − 2u3 + u4)
∆x2
1
(δxx u)4 = (u3 − 2u4 + ub) (3.16)
∆x2
Writing these equations in the more suggestive form
(δxx u)1 = ( ua −2u1 +u2 )/∆x2

(δxx u)2 = ( u1 −2u2 +u3 )/∆x2
(3.17)
(δxx u)3 = ( u2 −2u3 +u4 )/∆x2
(δxx u)4 = ( u3 −2u4 +ub )/∆x2
it is clear that we can express them in a vector-matrix form, and further, that the
resulting matrix has a very special form. Introducing
   
u1 ua
 u2  1  0 
u =   =  
 , bc   (3.18)
 u3  ∆x2  0 
u4 ub
and
 
−2 1
1  1 −2 1 

A= 2  1 −2
 (3.19)
∆x  1 
1 −2
we can rewrite Eq. 3.17 as

δxxu = Au + bc (3.20)
This example illustrates a matrix difference operator. Each line of a matrix differ-
ence operator is based on a point difference operator, but the point operators used
from line to line are not necessarily the same. For example, boundary conditions may
dictate that the lines at or near the bottom or top of the matrix be modified. In the
extreme case of the matrix difference operator representing a spectral method, none
of the lines is the same. The matrix operators representing the three-point central-
difference approximations for a first and second derivative with Dirichlet boundary
conditions on a four-point mesh are
   
0 1 −2 1
1  −1 0 1 
 1  1 −2 1 

δx =  , δxx = 2   (3.21)
2∆x  −1 0 1  ∆x  1 −2 1 
−1 0 1 −2
As a further example, replace the fourth line in Eq. 3.16 by the following point
operator for a Neumann boundary condition (See Section 3.6.):

2 1 ∂u 2 1
(δxx u)4 = − (u4 − u3 ) (3.22)
3 ∆x ∂x b
3 ∆x2
where the boundary condition is

∂u ∂u
= (3.23)
∂x x=π
∂x b
Then the matrix operator for a three-point central-differencing scheme at interior

points and a second-order approximation for a Neumann condition on the right is
given by
 
−2 1

1  1 −2 1 

δxx =   (3.24)
∆x2  1 −2 1 
2/3 −2/3
Each of these matrix difference operators is a square matrix with elements that are
all zeros except for those along bands which are clustered around the central diagonal.
We call such a matrix a banded matrix and introduce the notation
 
b c 1
 
 a b c 
 
B(M : a, b, c) =
 ... 
 (3.25)
  ..
 a b c 
  .
a b M
where the matrix dimensions are M × M . Use of M in the argument is optional,
and the illustration is given for a simple tridiagonal matrix although any number of
bands is a possibility. A tridiagonal matrix without constants along the bands can be
expressed as B(a, b, c). The arguments for a banded matrix are always odd in number
and the central one always refers to the central diagonal.
We can now generalize our previous examples. Defining u as3
 
u1
 
 u2 
 
 u3 
u =   (3.26)
 .. 
 . 
 
uM
we can approximate the second derivative of u by

1
δxxu = B(1, −2, 1)
u + bc (3.27)
∆x2

where bc stands for the vector holding the Dirichlet boundary conditions on the
left and right sides:
1
=
2 [ua , 0, · · · , 0, ub ]
T
bc (3.28)
∆x
If we prescribe Neumann boundary conditions on the right side, as in Eqs. 3.24 and
3.22, we find
1
δxxu = B(a, b, 1)u + bc
(3.29)
∆x2
where
a = [1, 1, · · · , 2/3]T
b = [−2, −2, −2, · · · , −2/3]T
T
1 2∆x ∂u

bc = 2 ua , 0, 0, · · · ,
∆x 3 ∂x b
Notice that the matrix operators given by Eqs. 3.27 and 3.29 carry more informa-
tion than the point operator given by Eq. 3.15. In Eqs. 3.27 and 3.29, the boundary
conditions have been uniquely specified and it is clear that the same point operator
has been applied at every point in the field except at the boundaries. The ability to
specify in the matrix derivative operator the exact nature of the approximation at the
3
Note that u is a function of time only since each element corresponds to one specific spatial
location.
various points in the field including the boundaries permits the use of quite general
constructions which will be useful later in considerations of stability.
Since we make considerable use of both matrix and point operators, it is important
to establish a relation between them. A point operator is generally written for some
derivative at the reference point j in terms of neighboring values of the function. For
example
(δx u)j = a2 uj−2 + a1 uj−1 + buj + c1 uj+1 (3.30)
might be the point operator for a first derivative. The corresponding matrix operator
has for its arguments the coefficients giving the weights to the values of the function
at the various locations. A j-shift in the point operator corresponds to a diagonal
shift in the matrix operator. Thus the matrix equivalent of Eq. 3.30 is
δxu = B(a2 , a1 , b, c1 , 0)u (3.31)
Note the addition of a zero in the fifth element which makes it clear that b is the
coefficient of uj .
3.3.3 Periodic Matrices

The above illustrated cases in which the boundary conditions are fixed. If the bound-
ary conditions are periodic, the form of the matrix operator changes. Consider the
eight-point periodic mesh shown below. This can either be presented on a linear mesh
with repeated entries, or more suggestively on a circular mesh as in Figure 3.3. When
the mesh is laid out on the perimeter of a circle, it doesn’t really matter where the
numbering starts as long as it “ends” at the point just preceding its starting location.
··· 7 8 1 2 3 4 5 6 7 8 1 2 ···
x = − − 0 − − − − − − − 2π −
j= 0 1 · · · · · · M
Eight points on a linear periodic mesh. ∆x = 2π/M
The matrix that represents differencing schemes for scalar equations on a periodic
mesh is referred to as a periodic matrix. A typical periodic tridiagonal matrix operator
8 2
7 3
6 4
5
Figure 3.3: Eight points on a circular mesh.
with nonuniform entries is given for a 6-point mesh by

 
b1 c2 a6
 
 a1 b2 c3 
 
 a2 b3 c4 

Bp (6 : a, b, c) =   (3.32)
a3 b4 c5 
 
 
 a4 b5 c6 
c1 a5 b6
3.3.4 Circulant Matrices

In general, as shown in the example, the elements along the diagonals of a periodic
matrix are not constant. However, a special subset of a periodic matrix is the circulant
matrix, formed when the elements along the various bands are constant. Circulant
matrices play a vital role in our analysis. We will have much more to say about them
later. The most general circulant matrix of order 4 is
 
b0 b1 b2 b3
 b3 b0 b1 b2 
 
  (3.33)
 b2 b3 b0 b1 
b1 b2 b3 b0
Notice that each row of a circulant matrix is shifted (see Figure 3.3) one element to
the right of the one above it. The special case of a tridiagonal circulant matrix is
3.4. CONSTRUCTING DIFFERENCING SCHEMES OF ANY ORDER 31
given by
 
b c a 1
 
 a b c 
 
= ... 
Bp (M : a, b, c)   (3.34)
  ..
 a b c 
  .
c a b M
When the standard three-point central-differencing approximations for a first and

second derivative, see Eq. 3.21, are used with periodic boundary conditions, they take
the form
 
0 1 −1
1  −1 0 1 
 1
(δx )p =  = B (−1, 0, 1)
2∆x  −1 0 1  2∆x p
1 −1 0
and
 
−2 1 1
1  1 −2 1 
 1
(δxx )p =   = Bp (1, −2, 1) (3.35)
∆x2  1 −2 1  ∆x2
1 1 −2
Clearly, these special cases of periodic operators are also circulant operators. Later
on we take advantage of this special property. Notice that there are no boundary
condition vectors since this information is all interior to the matrices themselves.
3.4 Constructing Differencing Schemes of Any Or-

der
3.4.1 Taylor Tables
The Taylor series expansion of functions about a fixed point provides a means for con-
structing finite-difference point-operators of any order. A simple and straightforward
way to carry this out is to construct a “Taylor table,” which makes extensive use of
the expansion given by Eq. 3.8. As an example, consider Table 3.1, which represents
a Taylor table for an approximation of a second derivative using three values of the
function centered about the point at which the derivative is to be evaluated.

∂2u 1
− (a uj−1 + b uj + c uj+1) = ?
∂x2 j ∆x2
· · ·
2 3 4
∆x·
∆x
2
∆x
3
∆x
4
∂u ∂ u ∂ u ∂ u
uj ∂x j ∂x2 j ∂x3 j ∂x4 j

∂2u
∆x2 · ∂x2 j
1
−a · uj−1 −a −a · (-1) · 11 −a · (-1)2 · 21 −a · (-1)3 · 16 1

−a · (-1)4 · 24
−b · uj −b
−c · uj+1 −c −c · (1) · 11 −c · (1)2 · 12 −c · (1)3 · 16 1
−c · (1)4 · 24
Table 3.1. Taylor table for centered 3-point Lagrangian approximation to a second
derivative.
The table is constructed so that some of the algebra is simplified. At the top of the
table we see an expression with a question mark. This represents one of the questions
that a study of this table can answer; namely, what is the local error caused by the
use of this approximation? Notice that all of the terms in the equation appear in
a column at the left of the table (although, in this case, ∆x2 has been multiplied
into each term in order to simplify the terms to be put into the table). Then notice
that at the head of each column there appears the common factor that occurs in the
expansion of each term about the point j, that is,

∂k u
∆x · k
k = 0, 1, 2, · · ·
∂xk j
The columns to the right of the leftmost one, under the headings, make up the Taylor
table. Each entry is the coefficient of the term at the top of the corresponding column
in the Taylor series expansion of the term to the left of the corresponding row. For
example, the last row in the table corresponds to the Taylor series expansion of
−c uj+1:

1 ∂u 1 ∂2 u
−c uj+1 = −c uj − c · (1) · ∆x · − c · (1)2 · ∆x2 ·
1 ∂x j
2 ∂x2 j

3 4
1 ∂ u 1 ∂ u
−c · (1)3 · ∆x3 · − c · (1)4 · ∆x4 · − · · · (3.36)
6 ∂x3 j
24 ∂x4 j
A Taylor table is simply a convenient way of forming linear combinations of Taylor

series on a term by term basis.
Consider the sum of each of these columns. To maximize the order of accuracy
of the method, we proceed from left to right and force, by the proper choice of a, b,
and c, these sums to be zero. One can easily show that the sums of the first three
columns are zero if we satisfy the equation
    
−1 −1 −1 a 0

0 −1     
 1  b  =  0
−1 0 −1 c −2
The solution is given by [a, b, c] = [1, −2, 1].

The columns that do not sum to zero constitute the error.
We designate the first nonvanishing sum to be ert , and refer to

it as the Taylor series error.
In this case ert occurs at the fifth column in the table (for this example all even
columns will vanish by symmetry) and one finds

1 −a −c 4
4 ∂ u −∆x2 ∂ 4 u
ert = + ∆x = (3.37)
∆x2 24 24 ∂x4 j
12 ∂x4 j
Note that ∆x2 has been divided through to make the error term consistent. We
have just derived the familiar 3-point central-differencing point operator for a second
derivative

∂2u 1
− (uj−1 − 2uj + uj+1) = O(∆x2 ) (3.38)
∂x2 j ∆x2
The Taylor table for a 3-point backward-differencing operator representing a first

derivative is shown in Table 3.2.

∂u 1
− (a2 uj−2 + a1 uj−1 + b uj ) = ?
∂x j
∆x
· · ·
2 3 4
∆x·
∆x
2
∆x
3
∆x
4
∂u ∂ u ∂ u ∂ u

∆x · ∂u
∂x j
1
−a2 · uj−2 −a2 −a2 · (-2) · 11 −a2 · (-2)2 · 12 −a2 · (-2)3 · 16 1

−a2 · (-2)4 · 24
−a1 · uj−1 −a1 −a1 · (-1) · 11 −a1 · (-1)2 · 12 −a1 · (-1)3 · 16 1
−a1 · (-1)4 · 24
−b · uj −b
Table 3.2. Taylor table for backward 3-point Lagrangian approximation to a first
derivative.
This time the first three columns sum to zero if

    
−1 −1 −1 a2 0

 2 1 0   a1  =  −1 
  

−4 −1 0 b 0
which gives [a2 , a1 , b] = 12 [1, −4, 3]. In this case the fourth column provides the leading
truncation error term:

1 8a2 a1 ∂3u ∆x2 ∂ 3 u
ert = + ∆x3 = (3.39)
∆x 6 6 ∂x3 j
3 ∂x3 j
Thus we have derived a second-order backward-difference approximation of a first

derivative:

∂u 1
− (uj−2 − 4uj−1 + 3uj ) = O(∆x2 ) (3.40)
∂x j
2∆x
3.4.2 Generalization of Difference Formulas

In general, a difference approximation to the mth derivative at grid point j can be
cast in terms of q + p + 1 neighboring points as

∂mu
q
− ai uj+i = ert (3.41)
∂xm j i=−p
where the ai are coefficients to be determined through the use of Taylor tables to
produce approximations of a given order. Clearly this process can be used to find
forward, backward, skewed, or central point operators of any order for any derivative.
It could be computer automated and extended to higher dimensions. More important,
however, is the fact that it can be further generalized. In order to do this, let us
approach the subject in a slightly different way, that is from the point of view of
interpolation formulas. These formulas are discussed in many texts on numerical
analysis.
3.4.3 Lagrange and Hermite Interpolation Polynomials

The Lagrangian interpolation polynomial is given by

K
u(x) = ak (x)uk (3.42)
k=0
where ak (x) are polynomials in x of degree K. The construction of the ak (x) can be
taken from the simple Lagrangian formula for quadratic interpolation (or extrapola-
tion) with non-equispaced points
(x1 − x)(x2 − x) (x0 − x)(x2 − x)

u(x) = u0 + u1
(x1 − x0 )(x2 − x0 ) (x0 − x1 )(x2 − x1 )
(x0 − x)(x1 − x)
+u2 (3.43)
(x0 − x2 )(x1 − x2 )
Notice that the coefficient of each uk is one when x = xk , and zero when x takes
any other discrete value in the set. If we take the first or second derivative of u(x),
impose an equispaced mesh, and evaluate these derivatives at the appropriate dis-
crete point, we rederive the finite-difference approximations just presented. Finite-
difference schemes that can be derived from Eq. 3.42 are referred to as Lagrangian
approximations.
A generalization of the Lagrangian approach is brought about by using Hermitian
interpolation. To construct a polynomial for u(x), Hermite formulas use values of the
function and its derivative(s) at given points in space. Our illustration is for the case
in which discrete values of the function and its first derivative are used, producing
the expression

∂u
u(x) = ak (x)uk + bk (x) (3.44)
∂x k
Obviously higher-order derivatives could be included as the problems dictate. A com-

plete discussion of these polynomials can be found in many references on numerical
methods, but here we need only the concept.
The previous examples of a Taylor table constructed explicit point difference op-
erators from Lagrangian interpolation formulas. Consider next the Taylor table for
an implicit space differencing scheme for a first derivative arising from the use of a
Hermite interpolation formula. A generalization of Eq. 3.41 can include derivatives
at neighboring points, i.e.,

s
∂mu
q
bi + ai uj+i = ert (3.45)
i=−r ∂xm j+i i=−p
analogous to Eq. 3.44. An example formula is illustrated at the top of Table 3.3. Here
not only is the derivative at point j represented, but also included are derivatives at
points j − 1 and j + 1, which also must be expanded using Taylor series about point
j. This requires the following generalization of the Taylor series expansion given in
Eq. 3.8:
∞

∂mu 1 n
∂mu
n ∂
= (k∆x) (3.46)
∂xm j+k n=0 n! ∂xn ∂xm j
The derivative terms now have coefficients (the coefficient on the j point is taken
as one to simplify the algebra) which must be determined using the Taylor table
approach as outlined below.

∂u ∂u ∂u 1
d + +e − (auj−1 + buj + cuj+1 ) = ?
∂x j−1
∂x j
∂x j+1
∆x
· · · ·
2 3 4 5
∆x·
∆x
2
∆x
3
∆x
4
∆x
5
∂u ∂ u ∂ u ∂ u ∂ u
uj ∂x j ∂x2 j ∂x3 j ∂x4 j ∂x5 j

∆x · d ∂u
∂x j−1
d d · (-1) · 11 d · (-1)2 · 12 d · (-1)3 · 16 1
d · (-1)4 · 24

∆x · ∂u
∂x
1
j
∆x · ∂u
e ∂x e e · (1) · 11 e · (1)2 · 12 e · (1)3 · 16 1
e · (1)4 · 24
j+1
−a · uj−1 −a −a · (-1) · 11 −a · (-1)2 · 12 −a · (-1)3 · 16 −a · (-1)4 · 241 1
−a · (-1)5 · 120
−b · uj −b
−c · uj+1 −c −c · (1) · 11 −c · (1)2 · 12 −c · (1)3 · 16 1
−c · (1)4 · 24 1
−c · (1)5 · 120
Table 3.3. Taylor table for central 3-point Hermitian approximation to a first derivative.
To maximize the order of accuracy, we must satisfy the relation

    
−1 −1 −1 0 0 a 0
 1 0 −1 1 1  b   −1 
    

 −1 0 −1 −2 2    
 c =  0 

    
 1 0 −1 3 3   d   0 
−1 0 −1 −4 4 e 0
having the solution [a, b, c, d, e] = 41 [−3, 0, 3, 1, 1]. Under these conditions the sixth
column sums to

∆x4 ∂ 5 u
ert = (3.47)
120 ∂x5 j
and the method can be expressed as

∂u ∂u ∂u 3
+4 + − (−uj−1 + uj+1) = O(∆x4 ) (3.48)
∂x j−1
∂x j
∂x j+1
∆x
This is also referred to as a Padé formula.
3.4.4 Practical Application of Padé Formulas

It is one thing to construct methods using the Hermitian concept and quite another
to implement them in a computer code. In the form of a point operator it is probably
not evident at first just how Eq. 3.48 can be applied. However, the situation is quite
easy to comprehend if we express the same method in the form of a matrix operator.
A banded matrix notation for Eq. 3.48 is
1 1
B(1, 4, 1)δxu =
B(−1, 0, 1)u + bc (3.49)
6 2∆x
in which Dirichlet boundary conditions have been imposed.4 Mathematically this is
equivalent to

1
δxu = 6[B(1, 4, 1)]−1
B(−1, 0, 1)u + bc (3.50)
2∆x
which can be reexpressed by the “predictor-corrector” sequence
1
ũ =
B(−1, 0, 1)u + bc
2∆x
δxu = 6 [B(1, 4, 1)]−1ũ (3.51)
4
In this case the vector containing the boundary conditions would include values of both u and
∂u/∂x at both boundaries.
With respect to practical implementation, the meaning of the predictor in this

sequence should be clear. It simply says – take the vector array u, difference it,
add the boundary conditions, and store the result in the intermediate array ũ. The
meaning of the second row is more subtle, since it is demanding the evaluation of an
inverse operator, but it still can be given a simple interpretation. An inverse matrix
operator implies the solution of a coupled set of linear equations. These operators
are very common in finite difference applications. They appear in the form of banded
matrices having a small bandwidth, in this case a tridiagonal. The evaluation of
[B(1, 4, 1)]−1 is found by means of a tridiagonal “solver”, which is simple to code,
efficient to run, and widely used. In general, Hermitian or Padé approximations can
be practical when they can be implemented by the use of efficient banded solvers.
3.4.5 Other Higher-Order Schemes

Hermitian forms of the second derivative can also be easily derived by means of a
Taylor table. For example

1
δxxu = 12 [B(1, 10, 1)]−1 B(1, −2, 1) u + bc (3.52)
∆x2
is O(∆x4 ) and makes use of only tridiagonal operations. It should be mentioned that
the spline approximation is one form of a Padé matrix difference operator. It is given
by

1
δxxu = 6 [B(1, 4, 1)]−1 B(1, −2, 1)u + bc (3.53)
∆x2
but its order of accuracy is only O(∆x2 ). How much this reduction in accuracy is
offset by the increased global continuity built into a spline fit is not known. We note
that the spline fit of a first derivative is identical to any of the expressions in Eqs.
3.48 to 3.51.
A final word on Hermitian approximations. Clearly they have an advantage over
3-point Lagrangian schemes because of their increased accuracy. However, a more
subtle point is that they get this increase in accuracy using information that is still
local to the point where the derivatives are being evaluated. In application, this can
be advantageous at boundaries and in the vicinity of steep gradients. It is obvious, of
course, that five point schemes using Lagrangian approximations can be derived that
have the same order of accuracy as the methods given in Eqs. 3.48 and 3.52, but they
will have a wider spread of space indices. In particular, two Lagrangian schemes with
the same order of accuracy are (here we ignore the problem created by the boundary
conditions, although this is one of the principal issues in applying these schemes):
∂ u 1
− Bp (1, −8, 0, 8, −1)u = O ∆x4 (3.54)
∂x 12∆x
3.5. FOURIER ERROR ANALYSIS 39
∂ 2u 1
− Bp (−1, 16, −30, 16, −1)
u = O ∆x4
(3.55)
∂x2 12∆x2
3.5 Fourier Error Analysis

In order to select a finite-difference scheme for a given application one must be able
to assess the accuracy of the candidate schemes. The accuracy of an operator is often
expressed in terms of the order of the leading error term determined from a Taylor
table. While this is a useful measure, it provides a fairly limited description. Further
information about the error behavior of a finite-difference scheme can be obtained
using Fourier error analysis.
3.5.1 Application to a Spatial Operator

An arbitrary periodic function can be decomposed into its Fourier components, which
are in the form eiκx , where κ is the wavenumber. It is therefore of interest to examine
how well a given finite-difference operator approximates derivatives of eiκx . We will
concentrate here on first derivative approximations, although the analysis is equally
applicable to higher derivatives.
The exact first derivative of eiκx is
∂eiκx
= iκeiκx (3.56)
∂x
If we apply, for example, a second-order centered difference operator to uj = eiκxj ,

where xj = j∆x, we get
uj+1 − uj−1
(δx u)j =
2∆x
eiκ∆x(j+1) − eiκ∆x(j−1)
=
2∆x
(e iκ∆x
− e−iκ∆x )eiκxj
=
2∆x
1
= [(cos κ∆x + i sin κ∆x) − (cos κ∆x − i sin κ∆x)]eiκxj
2∆x
sin κ∆x iκxj
= i e
∆x
= iκ∗ eiκxj (3.57)
2.5
2
κ*∆ x
th
1.5 4 Pade
th
4 Central
1
nd
2 Central
0.5
0
0 0.5 1 1.5 2 2.5 3
κ∆ x
Figure 3.4: Modified wavenumber for various schemes.
where κ∗ is the modified wavenumber. The modified wavenumber is so named because

it appears where the wavenumber, κ, appears in the exact expression. Thus the degree
to which the modified wavenumber approximates the actual wavenumber is a measure
of the accuracy of the approximation.
For the second-order centered difference operator the modified wavenumber is given
by
sin κ∆x
κ∗ = (3.58)
∆x
Note that κ∗ approximates κ to second-order accuracy, as is to be expected, since
sin κ∆x κ3 ∆x2

=κ− +...
∆x 6
Equation 3.58 is plotted in Figure 3.4, along with similar relations for the stan-
dard fourth-order centered difference scheme and the fourth-order Padé scheme. The
expression for the modified wavenumber provides the accuracy with which a given
wavenumber component of the solution is resolved for the entire wavenumber range
available in a mesh of a given size, 0 ≤ κ∆x ≤ π.
In general, finite-difference operators can be written in the form
(δx )j = (δxa )j + (δxs )j

3.5. FOURIER ERROR ANALYSIS 41
where (δxa )j is an antisymmetric operator and (δxs )j is a symmetric operator.5 If we

restrict our interest to schemes extending from j − 3 to j + 3, then
1
(δxa u)j = [a1 (uj+1 − uj−1) + a2 (uj+2 − uj−2 ) + a3 (uj+3 − uj−3)]
∆x
and
1
(δxs u)j = [d0 uj + d1 (uj+1 + uj−1) + d2 (uj+2 + uj−2) + d3 (uj+3 + uj−3 )]
∆x
The corresponding modified wavenumber is
1
iκ∗ = [d0 + 2(d1 cos κ∆x + d2 cos 2κ∆x + d3 cos 3κ∆x)
∆x
+ 2i(a1 sin κ∆x + a2 sin 2κ∆x + a3 sin 3κ∆x) (3.59)
When the finite-difference operator is antisymmetric (centered), the modified wavenum-

ber is purely real. When the operator includes a symmetric component, the modified
wavenumber is complex, with the imaginary component being entirely error. The
fourth-order Padé scheme is given by
3
(δx u)j−1 + 4(δx u)j + (δx u)j+1 = (uj+1 − uj−1)
∆x
The modified wavenumber for this scheme satisfies6
3 iκ∆x
iκ∗ e−iκ∆x + 4iκ∗ + iκ∗ eiκ∆x = (e − e−iκ∆x )
∆x
which gives
3i sin κ∆x
iκ∗ =
(2 + cos κ∆x)∆x
The modified wavenumber provides a useful tool for assessing difference approx-
imations. In the context of the linear convection equation, the errors can be given
a physical interpretation. Consider once again the linear convection equation in the
form
∂u ∂u
+a =0
∂t ∂x
5
In terms of a circulant matrix operator A, the antisymmetric part is obtained from (A − AT )/2
and the symmetric part from (A + AT )/2.
6
Note that terms such as (δx u)j−1 are handled by letting (δx u)j = ik ∗ eiκj∆x and evaluating the
shift in j.
on a domain extending from −∞ to ∞. Recall from Section 2.3.2 that a solution

initiated by a harmonic function with wavenumber κ is
u(x, t) = f (t)eiκx (3.60)
where f (t) satisfies the ODE

df
= −iaκf
dt
Solving for f (t) and substituting into Eq. 3.60 gives the exact solution as
u(x, t) = f (0)eiκ(x−at)
If second-order centered differences are applied to the spatial term, the following
ODE is obtained for f (t):

df sin κ∆x
= −ia f = −iaκ∗ f (3.61)
dt ∆x
Solving this ODE exactly (since we are considering the error from the spatial approx-
imation only) and substituting into Eq. 3.60, we obtain
∗ t)
unumerical (x, t) = f (0)eiκ(x−a (3.62)
where a∗ is the numerical (or modified) phase speed, which is related to the modified
wavenumber by
a∗ κ∗
=
a κ
For the above example,
a∗ sin κ∆x
=
a κ∆x
The numerical phase speed is the speed at which a harmonic function is propagated
numerically. Since a∗ /a ≤ 1 for this example, the numerical solution propagates
too slowly. Since a∗ is a function of the wavenumber, the numerical approximation
introduces dispersion, although the original PDE is nondispersive. As a result, a
waveform consisting of many different wavenumber components eventually loses its
original form.
Figure 3.5 shows the numerical phase speed for the schemes considered previously.
The number of points per wavelength (P P W ) by which a given wave is resolved is
given by 2π/κ∆x. The resolving efficiency of a scheme can be expressed in terms
of the P P W required to produce errors below a specified level. For example, the
second-order centered difference scheme requires 80 P P W to produce an error in
phase speed of less than 0.1 percent. The 5-point fourth-order centered scheme and
3.6. DIFFERENCE OPERATORS AT BOUNDARIES 43
1.2
0.8 nd
* 2 Central
a
_
a
0.6 th
4 Central
th
0.4 4 Pade
0.2
0
0 0.5 1 1.5 2 2.5 3
κ ∆x
Figure 3.5: Numerical phase speed for various schemes.
the fourth-order Padé scheme require 15 and 10 P P W respectively to achieve the

same error level.
For our example using second-order centered differences, the modified wavenumber
is purely real, but in the general case it can include an imaginary component as well,
as shown in Eq. 3.59. In that case, the error in the phase speed is determined from
the real part of the modified wavenumber, while the imaginary part leads to an error
in the amplitude of the solution, as can be seen by inspecting Eq. 3.61. Thus the
antisymmetric portion of the spatial difference operator determines the error in speed
and the symmetric portion the error in amplitude. This will be further discussed in
Section 11.2.
3.6 Difference Operators at Boundaries
As discussed in Section 3.3.2, a matrix difference operator incorporates both the

difference approximation in the interior of the domain and that at the boundaries.
In this section, we consider the boundary operators needed for our model equations
for convection and diffusion. In the case of periodic boundary conditions, no special
boundary operators are required.
3.6.1 The Linear Convection Equation

Referring to Section 2.3, the boundary conditions for the linear convection equation
can be either periodic or of inflow-outflow type. In the latter case, a Dirichlet bound-
ary condition is given at the inflow boundary, while no condition is specified at the
outflow boundary. Here we assume that the wave speed a is positive; thus the left-
hand boundary is the inflow boundary, and the right-hand boundary is the outflow
boundary. Thus the vector of unknowns is
u = [u1 , u2 , . . . , uM ]T (3.63)
and u0 is specified.
Consider first the inflow boundary. It is clear that as long as the interior difference
approximation does not extend beyond uj−1, then no special treatment is required
for this boundary. For example, with second-order centered differences we obtain

δx u = Au + bc (3.64)
with
   
0 1 −u0
   
1  −1 0 1  1  0 
A=  , =
bc   (3.65)
2∆x 
 −1 0 1 
 2∆x 
 0 

... ..
.
However, if we use the fourth-order interior operator given in Eq. 3.54, then the
approximation at j = 1 requires a value of uj−2, which is outside the domain. Hence,
a different operator is required at j = 1 which extends only to j − 1, while having
the appropriate order of accuracy. Such an operator, known as a numerical boundary
scheme, can have an order of accuracy which is one order lower than that of the
interior scheme, and the global accuracy will equal that of the interior scheme.7 For
example, with fourth-order centered differences, we can use the following third-order
operator at j = 1:
1
(δx u)1 = (−2u0 − 3u1 + 6u2 − u3 ) (3.66)
6∆x
which is easily derived using a Taylor table. The resulting difference operator has the
form of Eq. 3.64 with
   
−6 12 −2 −4u0
   
1  −8 0 8 −1  1  u0 
A=  , =
bc   (3.67)
12∆x 
 1 −8 0 8 −1 
 12∆x 
 0 

... ..
.
7
Proof of this theorem is beyond the scope of this book; the interested reader should consult the
literature for further details.
3.6. DIFFERENCE OPERATORS AT BOUNDARIES 45
This approximation is globally fourth-order accurate.

At the outflow boundary, no boundary condition is specified. We must approx-
imate ∂u/∂x at node M with no information about uM +1 . Thus the second-order
centered-difference operator, which requires uj+1, cannot be used at j = M . A
backward-difference formula must be used. With a second-order interior operator,
the following first-order backward formula can be used:
1
(δx u)M = (uM − uM −1 ) (3.68)
∆x
This produces a difference operator with
 
0 1  
  −u0
 −1 0 1   
   0 
1  −1 0 1  1  
  =  0 
A=  ... , bc   (3.69)
2∆x   2∆x  .. 
   
   . 
 −1 0 1 
0
−2 2
In the case of a fourth-order centered interior operator the last two rows of A require
modification.
Another approach to the development of boundary schemes is in terms of space
extrapolation. The following formula allows uM +1 to be extrapolated from the interior
data to arbitrary order on an equispaced grid:
(1 − E −1 )p uM +1 = 0 (3.70)
where E is the shift operator defined by Euj = uj+1 and the order of the approxima-
tion is p − 1. For example, with p = 2 we obtain
(1 − 2E −1 + E −2 )uM +1 = uM +1 − 2uM + uM −1 = 0 (3.71)
which gives the following first-order approximation to uM +1 :
uM +1 = 2uM − uM −1 (3.72)
Substituting this into the second-order centered-difference operator applied at node
M gives
1 1
(δx u)M = (uM +1 − uM −1 ) = (2uM − uM −1 − uM −1)
2∆x 2∆x
1
= (uM − uM −1) (3.73)
∆x
which is identical to Eq. 3.68.
3.6.2 The Diffusion Equation

In solving the diffusion equation, we must consider Dirichlet and Neumann boundary
conditions. The treatment of a Dirichlet boundary condition proceeds along the same
lines as the inflow boundary for the convection equation discussed above. With the
second-order centered interior operator
1
(uj+1 − 2uj + uj−1 )
(δxx u)j = (3.74)
∆x2
no modifications are required near boundaries, leading to the difference operator given
in Eq. 3.27.
For a Neumann boundary condition, we assume that ∂u/∂x is specified at j =
M + 1, that is

∂u ∂u
= (3.75)
∂x M +1
∂x b
Thus we design an operator at node M which is in the following form:

1 c ∂u
(δxx u)M = 2
(auM −1 + buM ) + (3.76)
∆x ∆x ∂x M +1
where a, b, and c are constants which can easily be determined using a Taylor table,
as shown in Table 3.4.
 
∂2u 1 c ∂u
−  2 (auj−1 + buj ) +  =?
∂x2 j ∆x ∆x ∂x j+1
· · ·
2 3 4
∆x·
∆x
2
∆x
3
∆x
4
∂u ∂ u ∂ u ∂ u

∂2u
∆x2 · ∂x2 j
1
−a · uj−1 −a −a · (-1) · 11 −a · (-1)2 · 12 −a · (-1)3 · 16 1

−a · (-1)4 · 24
−b · uj −b
−∆x · c · ∂u
∂x
−c −c · (1) · 11 −c · (1)2 · 12 −c · (1)3 · 16
j+1
Table 3.4. Taylor table for Neumann boundary condition.
Solving for a, b, and c, we obtain the following operator:

1 2 ∂u
(δxx u)M = 2
(2uM −1 − 2uM ) + (3.77)
3∆x 3∆x ∂x M +1
3.7. PROBLEMS 47
which produces the difference operator given in Eq. 3.29. Notice that this operator
is second-order accurate. In the case of a numerical approximation to a Neumann
boundary condition, this is necessary to obtain a globally second-order accurate for-
mulation. This contrasts with the numerical boundary schemes described previously
which can be one order lower than the interior scheme.
We can also obtain the operator in Eq. 3.77 using the space extrapolation idea.
Consider a second-order backward-difference approximation applied at node M + 1:

∂u 1
= (uM −1 − 4uM + 3uM +1 ) + O(∆x2 ) (3.78)
∂x M +1
2∆x
Solving for uM +1 gives

1 ∂u
uM +1 = 4uM − uM −1 + 2∆x + O(∆x3 ) (3.79)
3 ∂x M +1
Substituting this into the second-order centered difference operator for a second
derivative applied at node M gives
1
(δxx u)M = (uM +1 − 2uM + uM −1 ) (3.80)
∆x2
1 ∂u
= 2
3uM −1 − 6uM + 4uM − uM −1 + 2∆x
3∆x ∂x M +1

1 2 ∂u
= (2u M −1 − 2uM ) + (3.81)
3∆x2 3∆x ∂x M +1
3.7 Problems
1. Derive a third-order finite-difference approximation to a first derivative in the
form
1
(δx u)j = (auj−2 + buj−1 + cuj + duj+1)
∆x
Find the leading error term.
2. Derive a finite-difference approximation to a first derivative in the form
1
a(δx u)j−1 + (δx u)j = (buj−1 + cuj + duj+1)
∆x
3. Using a 4 (interior) point mesh, write out the 4×4 matrices and the boundary-
condition vector formed by using the scheme derived in question 2 when both
u and ∂u/∂x are given at j = 0 and u is given at j = 5.
4. Repeat question 2 with d = 0.
5. Derive a finite-difference approximation to a third derivative in the form
1
(δxxx u)j = (auj−2 + buj−1 + cuj + duj+1 + euj+2 )
∆x3
6. Derive a compact (or Padé) finite-difference approximation to a second deriva-

tive in the form
1
d(δxx u)j−1 + (δxx u)j + e(δxx u)j+1 = (auj−1 + buj + cuj+1 )
∆x2
7. Find the modified wavenumber for the operator derived in question 1. Plot the
real and imaginary parts of κ∗ ∆x vs. κ∆x for 0 ≤ κ∆x ≤ π. Compare the real
part with that obtained from the fourth-order centered operator (Eq. 3.54).
8. Application of the second-derivative operator to the function eiκx gives
∂ 2 eiκx
2
= −κ2 eiκx
∂x
Application of a difference operator for the second derivative gives
(δxx eiκj∆x )j = −κ∗2 eiκx
thus defining the modified wavenumber κ∗ for a second derivative approxima-

tion. Find the modified wavenumber for the second-order centered difference op-
erator for a second derivative, the noncompact fourth-order operator (Eq. 3.55),
and the compact fourth-order operator derived in question 6. Plot (κ∗ ∆x)2 vs.
(κ∆x)2 for 0 ≤ κ∆x ≤ π.
9. Find the grid-points-per-wavelength (P P W ) requirement to achieve a phase

speed error less than 0.1 percent for sixth-order noncompact and compact cen-
tered approximations to a first derivative.
3.7. PROBLEMS 49
10. Consider the following one-sided differencing schemes, which are first-, second-,
and third-order, respectively:
(δx u)j = (uj − uj−1) /∆x
(δx u)j = (3uj − 4uj−1 + uj−2 ) /(2∆x)
(δx u)j = (11uj − 18uj−1 + 9uj−2 − 2uj−3) /(6∆x)
Find the modified wavenumber for each of these schemes. Plot the real and
imaginary parts of κ∗ ∆x vs. κ∆x for 0 ≤ κ∆x ≤ π. Derive the two leading
terms in the truncation error for each scheme.
Chapter 4
THE SEMI-DISCRETE
APPROACH
One strategy for obtaining finite-difference approximations to a PDE is to start by

differencing the space derivatives only, without approximating the time derivative.
In the following chapters, we proceed with an analysis making considerable use of
this concept, which we refer to as the semi-discrete approach. Differencing the space
derivatives converts the basic PDE into a set of coupled ODE’s. In the most general
notation, these ODE’s would be expressed in the form
du
= F (u, t) (4.1)
dt
which includes all manner of nonlinear and time-dependent possibilities. On occasion,
we use this form, but the rest of this chapter is devoted to a more specialized matrix
notation described below.
Another strategy for constructing a finite-difference approximation to a PDE is
to approximate all the partial derivatives at once. This generally leads to a point
difference operator (see Section 3.3.1) which, in turn, can be used for the time advance
of the solution at any given point in the mesh. As an example let us consider the
model equation for diffusion
∂u ∂2 u
=ν 2
∂t ∂x
Using three-point central-differencing schemes for both the time and space derivatives,
we find  
(n+1) (n−1) (n) (n) (n)
uj − uj uj+1 − 2uj + uj−1
= ν 
2h ∆x2
or
(n+1) (n−1) 2hν (n) (n) (n)

uj = uj + u − 2u + u (4.2)
∆x2 j+1 j j−1
51
52 CHAPTER 4. THE SEMI-DISCRETE APPROACH
Clearly Eq. 4.2 is a difference equation which can be used at the space point j to
advance the value of u from the previous time levels n and n − 1 to the level n + 1.
It is a full discretization of the PDE. Note, however, that the spatial and temporal
discretizations are separable. Thus, this method has an intermediate semi-discrete
form and can be analyzed by the methods discussed in the next few chapters.
(n)
Another possibility is to replace the value of uj in the right hand side of Eq. 4.2
(n+1) (n−1)
with the time average of u at that point, namely (uj + uj )/2. This results in
the formula
   
(n+1) (n−1)
2hν  (n) u + uj
(n+1) (n−1)  j + uj−1
(n)
uj = uj + 2 uj+1 − 2 (4.3)
∆x 2
which can be solved for u(n+1) and time advanced at the point j. In this case, the
spatial and temporal discretizations are not separable, and no semi-discrete form
exists.
Equation 4.2 is sometimes called Richardson’s method of overlapping steps and
Eq. 4.3 is referred to as the DuFort-Frankel method. As we shall see later on, there
are subtle points to be made about using these methods to find a numerical solution
to the diffusion equation. There are a number of issues concerning the accuracy,
stability, and convergence of Eqs. 4.2 and 4.3 which we cannot comment on until we
develop a framework for such investigations. We introduce these methods here only to
distinguish between methods in which the temporal and spatial terms are discretized
separately and those for which no such separation is possible. For the time being, we
shall separate the space difference approximations from the time differencing. In this
approach, we reduce the governing PDE’s to ODE’s by discretizing the spatial terms
and use the well-developed theory of ODE solutions to aid us in the development of
an analysis of accuracy and stability.
4.1 Reduction of PDE’s to ODE’s
4.1.1 The Model ODE’s

First let us consider the model PDE’s for diffusion and biconvection described in
Chapter 2. In these simple cases, we can approximate the space derivatives with
difference operators and express the resulting ODE’s with a matrix formulation. This
is a simple and natural formulation when the ODE’s are linear.
4.1. REDUCTION OF PDE’S TO ODE’S 53
Model ODE for Diffusion

For example, using the 3-point central-differencing scheme to represent the second
derivative in the scalar PDE governing diffusion leads to the following ODE diffusion
model
du ν
= B(1, −2, 1)u + (bc) (4.4)
dt ∆x2
vector.
with Dirichlet boundary conditions folded into the (bc)
Model ODE for Biconvection

The term biconvection was introduced in Section 2.3. It is used for the scalar con-
vection model when the boundary conditions are periodic. In this case, the 3-point
central-differencing approximation produces the ODE model given by
du a
=− Bp (−1, 0, 1)u (4.5)
dt 2∆x
where the boundary condition vector is absent because the flow is periodic.
Eqs. 4.4 and 4.5 are the model ODE’s for diffusion and biconvection of a scalar in
one dimension. They are linear with coefficient matrices which are independent of x
and t.
4.1.2 The Generic Matrix Form

The generic matrix form of a semi-discrete approximation is expressed by the equation
du
= Au − f (t) (4.6)
dt
Note that the elements in the matrix A depend upon both the PDE and the type of
differencing scheme chosen for the space terms. The vector f (t) is usually determined
by the boundary conditions and possibly source terms. In general, even the Euler
and Navier-Stokes equations can be expressed in the form of Eq. 4.6. In such cases
the equations are nonlinear, that is, the elements of A depend on the solution u and
are usually derived by finding the Jacobian of a flux vector. Although the equations
are nonlinear, the linear analysis presented in this book leads to diagnostics that are
surprisingly accurate when used to evaluate many aspects of numerical methods as
they apply to the Euler and Navier-Stokes equations.
4.2 Exact Solutions of Linear ODE’s

In order to advance Eq. 4.1 in time, the system of ODE’s must be integrated using a
time-marching method. In order to analyze time-marching methods, we will make use
of exact solutions of coupled systems of ODE’s, which exist under certain conditions.
The ODE’s represented by Eq. 4.1 are said to be linear if F is linearly dependent on
u (i.e., if ∂F/∂u = A where A is independent of u). As we have already pointed out,
when the ODE’s are linear they can be expressed in a matrix notation as Eq. 4.6 in
which the coefficient matrix, A, is independent of u. If A does depend explicitly on
t, the general solution cannot be written; whereas, if A does not depend explicitly on
t, the general solution to Eq. 4.6 can be written. This holds regardless of whether or
not the forcing function, f, depends explicitly on t.
As we shall soon see, the exact solution of Eq. 4.6 can be written in terms of
the eigenvalues and eigenvectors of A. This will lead us to a representative scalar
equation for use in analyzing time-marching methods. These ideas are developed in
the following sections.
4.2.1 Eigensystems of Semi-Discrete Linear Forms

Complete Systems
An M × M matrix is represented by a complete eigensystem if it has a complete set
of linearly independent eigenvectors (see Appendix A) . An eigenvector, xm , and its
corresponding eigenvalue, λm , have the property that
Axm = λmxm or
[A − λm I]xm = 0 (4.7)
The eigenvalues are the roots of the equation
det[A − λI] = 0
We form the right-hand eigenvector matrix of a complete system by filling its columns
with the eigenvectors xm :
X = x1 , x2 . . . , xM
The inverse is the left-hand eigenvector matrix, and together they have the property
that
X −1 AX = Λ (4.8)
where Λ is a diagonal matrix whose elements are the eigenvalues of A.

4.2. EXACT SOLUTIONS OF LINEAR ODE’S 55
Defective Systems
If an M ×M matrix does not have a complete set of linearly independent eigenvectors,
it cannot be transformed to a diagonal matrix of scalars, and it is said to be defective.
It can, however, be transformed to a diagonal set of blocks, some of which may be
scalars (see Appendix A). In general, there exists some S which transforms any
matrix A such that
S −1 AS = J
where  
J1
 J2 
 
 .. 
J=  . 
 
 
 Jm 
...
and  
λm 1 1
 ..  ..
 λm .  .
(n)
Jm =   ..
 ... 
 1  .
λm n
The matrix J is said to be in Jordan canonical form, and an eigenvalue with multi-
plicity n within a Jordan block is said to be a defective eigenvalue. Defective systems
play a role in numerical stability analysis.
4.2.2 Single ODE’s of First- and Second-Order

First-Order Equations
The simplest nonhomogeneous ODE of interest is given by the single, first-order
equation
du
= λu + aeµt (4.9)
dt
where λ, a, and µ are scalars, all of which can be complex numbers. The equation
is linear because λ does not depend on u, and has a general solution because λ does
not depend on t. It has a steady-state solution if the right-hand side is independent
of t, i.e., if µ = 0, and is homogeneous if the forcing function is zero, i.e., if a = 0.
Although it is quite simple, the numerical analysis of Eq. 4.9 displays many of the
fundamental properties and issues involved in the construction and study of most
popular time-marching methods. This theme will be developed as we proceed.
The exact solution of Eq. 4.9 is, for µ = λ,
aeµt
u(t) = c1 eλt +
µ−λ
where c1 is a constant determined by the initial conditions. In terms of the initial

value of u, it can be written
λt eµt − eλt
u(t) = u(0)e + a
µ−λ
The interesting question can arise: What happens to the solution of Eq. 4.9 when
µ = λ? This is easily found by setting µ = λ + , solving, and then taking the limit
as → 0. Using this limiting device, we find that the solution to
du
= λu + aeλt (4.10)
dt
is given by
u(t) = [u(0) + at]eλt
As we shall soon see, this solution is required for the analysis of defective systems.
Second-Order Equations
The homogeneous form of a second-order equation is given by
d2 u du
2
+ a1 + a0 u = 0 (4.11)
dt dt
where a1 and a0 are complex constants. Now we introduce the differential operator
D such that
d
D≡
dt
and factor u(t) out of Eq. 4.11, giving
(D2 + a1 D + a0 ) u(t) = 0
The polynomial in D is referred to as a characteristic polynomial and designated

P (D). Characteristic polynomials are fundamental to the analysis of both ODE’s and
O∆E’s, since the roots of these polynomials determine the solutions of the equations.
For ODE’s, we often label these roots in order of increasing modulus as λ1 , λ2 , · · ·, λm ,
· · ·, λM . They are found by solving the equation P (λ) = 0. In our simple example,
there would be two roots, λ1 and λ2 , determined from
P (λ) = λ2 + a1 λ + a0 = 0 (4.12)
and the solution to Eq. 4.11 is given by
u(t) = c1 eλ1 t + c2 eλ2 t (4.13)
where c1 and c2 are constants determined from initial conditions. The proof of this
is simple and is found by substituting Eq. 4.13 into Eq. 4.11. One finds the result
c1 eλ1 t (λ21 + a1 λ1 + a0 ) + c2 eλ2 t (λ22 + a1 λ2 + a0 )
which is identically zero for all c1 , c2 , and t if and only if the λ’s satisfy Eq. 4.12.
4.2.3 Coupled First-Order ODE’s

A Complete System
A set of coupled, first-order, homogeneous equations is given by
u1 = a11 u1 + a12 u2
u2 = a21 u1 + a22 u2 (4.14)
which can be written

a11 a12
u = Au , u = [u1 , u2 ] T
, A = (aij ) =
a21 a22
Consider the possibility that a solution is represented by
u1 = c1 x11 eλ1 t + c2 x12 eλ2 t
u2 = c1 x21 eλ1 t + c2 x22 eλ2 t (4.15)
By substitution, these are indeed solutions to Eq. 4.14 if and only if

a11 a12 x11 x11 a11 a12 x12 x12
= λ1 , = λ2 (4.16)
a21 a22 x21 x21 a21 a22 x22 x22
Notice that a higher-order equation can be reduced to a coupled set of first-order
equations by introducing a new set of dependent variables. Thus, by setting
u 1 = u , u2 = u ,
we find Eq. 4.11 can be written
u1 = −a1 u1 − a0 u2
u2 = u1 (4.17)
which is a subset of Eq. 4.14.
A Derogatory System
Eq. 4.15 is still a solution to Eq. 4.14 if λ1 = λ2 = λ, provided two linearly independent
vectors exist to satisfy Eq. 4.16 with A = Λ. In this case

λ 0 1 1 λ 0 0 0
=λ and =λ
0 λ 0 0 0 λ 1 1
provide such a solution. This is the case where A has a complete set of eigenvectors
and is not defective.
A Defective System
If A is defective, then it can be represented by the Jordan canonical form

u1 λ 0 u1
= (4.18)
u2 1 λ u2
whose solution is not obvious. However, in this case, one can solve the top equation
first, giving u1 (t) = u1 (0)eλt . Then, substituting this result into the second equation,
one finds
du2
= λu2 + u1 (0)eλt
dt
which is identical in form to Eq. 4.10 and has the solution
u2 (t) = [u2 (0) + u1 (0)t]eλt
From this the reader should be able to verify that

u3 (t) = a + bt + ct2 eλt
is a solution to     
u1 λ u1
    
 u2  =  1 λ   u2 
u3 1 λ u3
if
1
a = u3 (0) , b = u2 (0) , c = u1 (0) (4.19)
2
The general solution to such defective systems is left as an exercise.
4.2.4 General Solution of Coupled ODE’s with Complete Eigen-

systems
Let us consider a set of coupled, nonhomogeneous, linear, first-order ODE’s with
constant coefficients which might have been derived by space differencing a set of
PDE’s. Represent them by the equation
du
= Au − f (t) (4.20)
dt
Our assumption is that the M × M matrix A has a complete eigensystem1 and can
be transformed by the left and right eigenvector matrices, X−1 and X, to a diagonal
matrix Λ having diagonal elements which are the eigenvalues of A, see Section 4.2.1.
Now let us multiply Eq. 4.20 from the left by X −1 and insert the identity combination
XX −1 = I between A and u. There results
du
X −1 = X −1 AX · X −1u − X −1f(t) (4.21)
dt
Since A is independent of both u and t, the elements in X−1 and X are also indepen-
dent of both u and t, and Eq. 4.21 can be modified to
d −1
X u = ΛX −1u − X −1f(t)
dt
and g such that
Finally, by introducing the new variables w
= X −1u ,
w g(t) = X −1f (t) (4.22)
we reduce Eq. 4.20 to a new algebraic form

dw − g(t)
= Λw (4.23)
dt
It is important at this point to review the results of the previous paragraph. Notice
that Eqs. 4.20 and 4.23 are expressing exactly the same equality. The only difference
between them was brought about by algebraic manipulations which regrouped the
variables. However, this regrouping is crucial for the solution process because Eqs.
1
In the following, we exclude defective systems, not because they cannot be analyzed (the example
at the conclusion of the previous section proves otherwise), but because they are only of limited
interest in the general development of our theory.
4.23 are no longer coupled. They can be written line by line as a set of independent,
single, first-order equations, thus
w1 = λ1 w1 − g1 (t)
..
.

wm = λm wm − gm (t)
..
.

wM = λM wM − gM (t) (4.24)
For any given set of gm (t) each of these equations can be solved separately and then
recoupled, using the inverse of the relations given in Eqs. 4.22:
u(t) = X w(t)

M
= wm (t)xm (4.25)
m=1
where xm is the m’th column of X, i.e., the eigenvector corresponding to λm .

We next focus on the very important subset of Eq. 4.20 when neither A nor f has
any explicit dependence on t. In such a case, the gm in Eqs. 4.23 and 4.24 are also
time invariant and the solution to any line in Eq. 4.24 is
1
wm (t) = cm eλm t + gm
λm
where the cm are constants that depend on the initial conditions. Transforming back
to the u-system gives
u(t) = X w(t)

M
= wm (t)xm
m=1
M
M
1
= cm eλm t xm + gmxm
m=1 m=1 λm
M
= cm eλm t xm + XΛ−1 X −1f
m=1
M
= cm eλm t xm + A−1f
m=1
(4.26)
Transient Steady-state
4.3. REAL SPACE AND EIGENSPACE 61
Note that the steady-state solution is A−1f , as might be expected.

The first group of terms on the right side of this equation is referred to classically
as the complementary solution or the solution of the homogeneous equations. The
second group is referred to classically as the particular solution or the particular
integral. In our application to fluid dynamics, it is more instructive to refer to these
groups as the transient and steady-state solutions, respectively. An alternative, but
entirely equivalent, form of the solution is
u(t) = c1 eλ1 t x1 + · · · + cm eλm t xm + · · · + cM eλM t xM + A−1f (4.27)
4.3 Real Space and Eigenspace

4.3.1 Definition
Following the semi-discrete approach discussed in Section 4.1, we reduce the partial
differential equations to a set of ordinary differential equations represented by the
generic form
du
= Au − f (4.28)
dt
The dependent variable u represents some physical quantity or quantities which relate
to the problem of interest. For the model problems on which we are focusing most
of our attention, the elements of A are independent of both u and t. This permits
us to say a great deal about such problems and serves as the basis for this section.
In particular, we can develop some very important and fundamental concepts that
underly the global properties of the numerical solutions to the model problems. How
these relate to the numerical solutions of more practical problems depends upon the
problem and, to a much greater extent, on the cleverness of the relator.
We begin by developing the concept of “spaces”. That is, we identify different
mathematical reference frames (spaces) and view our solutions from within each.
In this way, we get different perspectives of the same solution, and this can add
significantly to our understanding.
The most natural reference frame is the physical one. We say
If a solution is expressed in terms of u, it is said

to be in real space.
There is, however, another very useful frame. We saw in Sections 4.2.1 and 4.2 that
pre- and post-multiplication of A by the appropriate similarity matrices transforms A
into a diagonal matrix, composed, in the most general case, of Jordan blocks or, in the
simplest nondefective case, of scalars. Following Section 4.2 and, for simplicity, using
only complete systems for our examples, we found that Eq. 4.28 had the alternative
form

dw
= Λw − g
dt
which is an uncoupled set of first-order ODE’s that can be solved independently for
the dependent variable vector w. We say
it is said to
If a solution is expressed in terms of w,
be in eigenspace (often referred to as wave space).
The relations that transfer from one space to the other are:
= X −1u
w u = X w

g = X −1f f = Xg
The elements of u relate directly to the local physics of the problem. However, the
are linear combinations of all of the elements of u, and individually
elements of w
they have no direct local physical interpretation.
When the forcing function f is independent of t, the solutions in the two spaces
are represented by
u(t) = cm eλm t xm + XΛ−1 X −1f
and
1
wm (t) = cm eλm t + gm ; m = 1, 2, · · · , M
λm
for real space and eigenspace, respectively. At this point we make two observations:
1. the transient portion of the solution in real space consists of a linear combination
of contributions from each eigenvector, and
2. the transient portion of the solution in eigenspace provides the weighting of

each eigenvector component of the transient solution in real space.
4.3.2 Eigenvalue Spectrums for Model ODE’s

It is instructive to consider the eigenvalue spectrums of the ODE’s formulated by
central differencing the model equations for diffusion and biconvection. These model
ODE’s are presented in Section 4.1. Equations for the eigenvalues of the simple
tridiagonals, B(M : a, b, c) and Bp (M : a, b, c), are given in Appendix B. From Section

B.1, we find for the model diffusion equation with Dirichlet boundary conditions

ν mπ
λm = −2 + 2 cos
∆x2
M + 1
−4ν mπ
= sin2 ; m = 1, 2, · · · , M (4.29)
∆x2 2(M + 1)
and, from Section B.4, for the model biconvection equation

−ia 2mπ
λm = sin m = 0, 1, · · · , M − 1
∆x M
= −iκ∗m a m = 0, 1, · · · , M − 1 (4.30)
where
sin κm ∆x
κ∗m = m = 0, 1, · · · , M − 1 (4.31)
∆x
is the modified wavenumber from Section 3.5, κm = m, and ∆x = 2π/M . Notice
that the diffusion eigenvalues are real and negative while those representing periodic
convection are all pure imaginary. The interpretation of this result plays a very
important role later in our stability analysis.
4.3.3 Eigenvectors of the Model Equations

Next we consider the eigenvectors of the two model equations. These follow as special
cases from the results given in Appendix B.
The Diffusion Model

Consider Eq. 4.4, the model ODE’s for diffusion. First, to help visualize the matrix
structure, we present results for a simple 4-point mesh and then we give the general
case. The right-hand eigenvector matrix X is given by
 
sin (x1 ) sin (2x1 ) sin (3x1 ) sin (4x1 )
 sin (x ) sin (2x2 ) sin (3x2 ) sin (4x2 ) 
 2 
 
 sin (x3 ) sin (2x3 ) sin (3x3 ) sin (4x3 ) 
sin (x4 ) sin (2x4 ) sin (3x4 ) sin (4x4 )
The columns of the matrix are proportional to the eigenvectors. Recall that xj =
j∆x = jπ/(M + 1), so in general the relation u = X w
can be written as

M
uj = wm sin mxj ; j = 1, 2, · · · , M (4.32)
m=1
For the inverse, or left-hand eigenvector matrix X−1 , we find

 
sin (x1 ) sin (x2 ) sin (x3 ) sin (x4 )
 sin (2x ) sin (2x ) sin (2x ) sin (2x ) 
 1 2 3 4 
 
 sin (3x1 ) sin (3x2 ) sin (3x3 ) sin (3x4 ) 
sin (4x1 ) sin (4x2 ) sin (4x3 ) sin (4x4 )
= X−1u
The rows of the matrix are proportional to the eigenvectors. In general w
gives

M
wm = uj sin mxj ; m = 1, 2, · · · , M (4.33)
j=1
In the field of harmonic analysis, Eq. 4.33 represents a sine transform of the func-
tion u(x) for an M -point sample between the boundaries x = 0 and x = π with the
condition u(0) = u(π) = 0. Similarly, Eq. 4.32 represents the sine synthesis that
companions the sine transform given by Eq. 4.33. In summary,
For the model diffusion equation:
= X −1u
w is a sine transform from real space to (sine) wave
space.
u = X w
is a sine synthesis from wave space back to real
space.
The Biconvection Model

Next consider the model ODE’s for periodic convection, Eq. 4.5. The coefficient
matrices for these ODE’s are always circulant. For our model ODE, the right-hand
eigenvectors are given by
xm = ei j (2πm/M ) , j = 0, 1, · · · , M − 1
m = 0, 1, · · · , M − 1
With xj = j · ∆x = j · 2π/M , we can write u = X w

as

M −1
uj = wm eimxj ; j = 0, 1, · · · , M − 1 (4.34)
m=0
For a 4-point periodic mesh, we find the following left-hand eigenvector matrix
from Appendix B.4:
    
w1 1 1 1 1 u1
w  
1 1 e −2iπ/4
e−4iπ/4 e −6iπ/4   
 2   u2 
 =  −12iπ/4    = X −1
u
 w3  4  1 e−4iπ/4 e−8iπ/4 e   u3 
w4 1 e−6iπ/4 e−12iπ/4 e−18iπ/4 u4
In general
−1
1 M
wm = uj e−imxj ; m = 0, 1, · · · , M − 1
M j=0
This equation is identical to a discrete Fourier transform of the periodic dependent
variable u using an M -point sample between and including x = 0 and x = 2π − ∆x.
For circulant matrices, it is straightforward to establish the fact that the relation
u = X w represents the Fourier synthesis of the variable w back to u. In summary,
For any circulant system:
= X −1u
w is a complex Fourier transform from real space to
wave space.
u = X w
is a complex Fourier synthesis from wave space
back to real space.
4.3.4 Solutions of the Model ODE’s

We can now combine the results of the previous sections to write the solutions of our
model ODE’s.
The Diffusion Equation

For the diffusion equation, Eq. 4.27 becomes

M
uj (t) = cm eλm t sin mxj + (A−1 f )j , j = 1, 2, · · · , M (4.35)
m=1
where

−4ν 2 mπ
λm = 2 sin (4.36)
∆x 2(M + 1)
With the modified wavenumber defined as

2 κm ∆x
κ∗m = sin (4.37)
∆x 2
and using κm = m, we can write the ODE solution as

M
∗ 2
uj (t) = cm e−νκm t sin κm xj + (A−1 f )j , j = 1, 2, · · · , M (4.38)
m=1
This can be compared with the exact solution to the PDE, Eq. 2.37, evaluated at the
nodes of the grid:

M
cm e−νκm t sin κm xj + h(xj ),
2
uj (t) = j = 1, 2, · · · , M (4.39)
m=1
We see that the solutions are identical except for the steady solution and the
modified wavenumber in the transient term. The modified wavenumber is an approx-
imation to the actual wavenumber. The difference between the modified wavenumber
and the actual wavenumber depends on the differencing scheme and the grid resolu-
tion. This difference causes the various modes (or eigenvector components) to decay
at rates which differ from the exact solution. With conventional differencing schemes,
low wavenumber modes are accurately represented, while high wavenumber modes (if
they have significant amplitudes) can have large errors.
The Convection Equation

For the biconvection equation, we obtain

M −1
uj (t) = cm eλm t eiκm xj , j = 0, 1, · · · , M − 1 (4.40)
m=0
where
λm = −iκ∗m a (4.41)
with the modified wavenumber defined in Eq. 4.31. We can write this ODE solution
as

M −1
∗
uj (t) = cm e−iκm at eiκm xj , j = 0, 1, · · · , M − 1 (4.42)
m=0
4.4. THE REPRESENTATIVE EQUATION 67
and compare it to the exact solution of the PDE, Eq. 2.26, evaluated at the nodes of
the grid:

M −1
uj (t) = fm (0)e−iκm at eiκm xj , j = 0, 1, · · · , M − 1 (4.43)
m=0
Once again the difference appears through the modified wavenumber contained in
λm . As discussed in Section 3.5, this leads to an error in the speed with which various
modes are convected, since κ∗ is real. Since the error in the phase speed depends on
the wavenumber, while the actual phase speed is independent of the wavenumber,
the result is erroneous numerical dispersion. In the case of non-centered differencing,
discussed in Chapter 11, the modified wavenumber is complex. The form of Eq. 4.42
shows that the imaginary portion of the modified wavenumber produces nonphysical
decay or growth in the numerical solution.
4.4 The Representative Equation

In Section 4.3, we pointed out that Eqs. 4.20 and 4.23 express identical results but in
terms of different groupings of the dependent variables, which are related by algebraic
manipulation. This leads to the following important concept:
The numerical solution to a set of linear ODE’s (in which A is

not a function of t) is entirely equivalent to the solution obtained
if the equations are transformed to eigenspace, solved there in
their uncoupled form, and then returned as a coupled set to real
space.
The importance of this concept resides in its message that we can analyze time-
marching methods by applying them to a single, uncoupled equation and our con-
clusions will apply in general. This is helpful both in analyzing the accuracy of
time-marching methods and in studying their stability, topics which are covered in
Chapters 6 and 7.
Our next objective is to find a “typical” single ODE to analyze. We found the
uncoupled solution to a set of ODE’s in Section 4.2. A typical member of the family
is
dwm
= λm wm − gm (t) (4.44)
dt
The goal in our analysis is to study typical behavior of general situations, not partic-
ular problems. For such a purpose Eq. 4.44 is not quite satisfactory. The role of λm is
clear; it stands for some representative eigenvalue in the original A matrix. However,
the question is: What should we use for gm (t) when the time dependence cannot be
ignored? To answer this question, we note that, in principle, one can express any one
of the forcing terms gm (t) as a finite Fourier series. For example

−g(t) = ak eikt
k
for which Eq. 4.44 has the exact solution:
λt
ak eikt
w(t) = ce +
k ik − λ
From this we can extract the k’th term and replace ik with µ. This leads to
The Representative ODE

(4.45)
dw
= λw + aeµt
dt
which can be used to evaluate all manner of time-marching methods. In such evalua-
tions the parameters λ and µ must be allowed to take the worst possible combination
of values that might occur in the ODE eigensystem. The exact solution of the repre-
sentative ODE is (for µ = λ):
aeµt
w(t) = ceλt + (4.46)
µ−λ
4.5 Problems
1. Consider the finite-difference operator derived in question 1 of Chapter 3. Using
this operator to approximate the spatial derivative in the linear convection equa-
tion, write the semi-discrete form obtained with periodic boundary conditions
on a 5-point grid (M = 5).
2. Consider the application of the operator given in Eq. 3.52 to the 1-D diffusion
equation with Dirichlet boundary conditions. Write the resulting semi-discrete
ODE form. Find the entries in the boundary-condition vector.
3. Write the semi-discrete form resulting from the application of second-order cen-
tered differences to the following equation on the domain 0 ≤ x ≤ 1 with
boundary conditions u(0) = 0, u(1) = 1:
∂u ∂2u
= 2 − 6x
∂t ∂x
4.5. PROBLEMS 69
4. Consider a grid with 10 interior points spanning the domain 0 ≤ x ≤ π. For

initial conditions u(x, 0) = sin(mx) and boundary conditions u(0, t) = u(π, t) =
0, plot the exact solution of the diffusion equation with ν = 1 at t = 1 with
m = 1 and m = 3. (Plot the solution at the grid nodes only.) Calculate the
corresponding modified wavenumbers for the second-order centered operator
from Eq. 4.37. Calculate and plot the corresponding ODE solutions.
5. Consider the matrix
A = −Bp (10, −1, 0, 1)/(2∆x)
corresponding to the ODE form of the biconvection equation resulting from the
application of second-order central differencing on a 10-point grid. Note that
the domain is 0 ≤ x ≤ 2π and ∆x = 2π/10. The grid nodes are given by
xj = j∆x, j = 0, 1, . . . 9. The eigenvalues of the above matrix A, as well as the
matrices X and X −1 , can be found from Appendix B.4. Using these, compute
and plot the ODE solution at t = 2π for the initial condition u(x, 0) = sin x.
Compare with the exact solution of the PDE. Calculate the numerical phase
speed from the modified wavenumber corresponding to this initial condition
and show that it is consistent with the ODE solution. Repeat for the initial
condition u(x, 0) = sin 2x.
Chapter 5
FINITE-VOLUME METHODS
In Chapter 3, we saw how to derive finite-difference approximations to arbitrary

derivatives. In Chapter 4, we saw that the application of a finite-difference approxi-
mation to the spatial derivatives in our model PDE’s produces a coupled set of ODE’s.
In this Chapter, we will show how similar semi-discrete forms can be derived using
finite-volume approximations in space. Finite-volume methods have become popu-
lar in CFD as a result, primarily, of two advantages. First, they ensure that the
discretization is conservative, i.e., mass, momentum, and energy are conserved in a
discrete sense. While this property can usually be obtained using a finite-difference
formulation, it is obtained naturally from a finite-volume formulation. Second, finite-
volume methods do not require a coordinate transformation in order to be applied on
irregular meshes. As a result, they can be applied on unstructured meshes consisting
of arbitrary polyhedra in three dimensions or arbitrary polygons in two dimensions.
This increased flexibility can be used to great advantage in generating grids about
arbitrary geometries.
Finite-volume methods are applied to the integral form of the governing equations,
either in the form of Eq. 2.1 or Eq. 2.2. Consistent with our emphasis on semi-discrete
methods, we will study the latter form, which is
d
QdV + n.FdS = P dV (5.1)
dt V (t) S(t) V (t)
We will begin by presenting the basic concepts which apply to finite-volume strategies.
Next we will give our model equations in the form of Eq. 5.1. This will be followed
by several examples which hopefully make these concepts clear.
71
72 CHAPTER 5. FINITE-VOLUME METHODS
5.1 Basic Concepts

The basic idea of a finite-volume method is to satisfy the integral form of the con-
servation law to some degree of approximation for each of many contiguous control
volumes which cover the domain of interest. Thus the volume V in Eq. 5.1 is that of a
control volume whose shape is dependent on the nature of the grid. In our examples,
we will consider only control volumes which do not vary with time. Examining Eq.
5.1, we see that several approximations must be made. The flux is required at the
boundary of the control volume, which is a closed surface in three dimensions and a
closed contour in two dimensions. This flux must then be integrated to find the net
flux through the boundary. Similarly, the source term P must be integrated over the
control volume. Next a time-marching method1 can be applied to find the value of

QdV (5.2)
V
at the next time step.

Let us consider these approximations in more detail. First, we note that the
average value of Q in a cell with volume V is
1
Q̄ ≡ QdV (5.3)
V V
and Eq. 5.1 can be written as

d
V Q̄ + n.FdS = P dV (5.4)
dt S V
for a control volume which does not vary with time. Thus after applying a time-
marching method, we have updated values of the cell-averaged quantities Q̄. In order
to evaluate the fluxes, which are a function of Q, at the control-volume boundary, Q
can be represented within the cell by some piecewise approximation which produces
the correct value of Q̄. This is a form of interpolation often referred to as recon-
struction. As we shall see in our examples, each cell will have a different piecewise
approximation to Q. When these are used to calculate F(Q), they will generally
produce different approximations to the flux at the boundary between two control
volumes, that is, the flux will be discontinuous. A nondissipative scheme analogous
to centered differencing is obtained by taking the average of these two fluxes. Another
approach known as flux-difference splitting is described in Chapter 11.
The basic elements of a finite-volume method are thus the following:
1
Time-marching methods will be discussed in the next chapter.
5.2. MODEL EQUATIONS IN INTEGRAL FORM 73
1. Given the value of Q̄ for each control volume, construct an approximation to

Q(x, y, z) in each control volume. Using this approximation, find Q at the
control-volume boundary. Evaluate F(Q) at the boundary. Since there is a
distinct approximation to Q(x, y, z) in each control volume, two distinct values
of the flux will generally be obtained at any point on the boundary between two
control volumes.
2. Apply some strategy for resolving the discontinuity in the flux at the control-
volume boundary to produce a single value of F(Q) at any point on the bound-
ary. This issue is discussed in Section 11.4.2.
3. Integrate the flux to find the net flux through the control-volume boundary
using some sort of quadrature.
4. Advance the solution in time to obtain new values of Q̄.
The order of accuracy of the method is dependent on each of the approximations.

These ideas should be clarified by the examples in the remainder of this chapter.
In order to include diffusive fluxes, the following relation between ∇Q and Q is
sometimes used:
∇QdV = nQdS (5.5)
V S
or, in two dimensions,

∇QdA = nQdl (5.6)
A C
where the unit vector n points outward from the surface or contour.
5.2 Model Equations in Integral Form

5.2.1 The Linear Convection Equation
A two-dimensional form of the linear convection equation can be written as
∂u ∂u ∂u
+ a cos θ + a sin θ =0 (5.7)
∂t ∂x ∂y
This PDE governs a simple plane wave convecting the scalar quantity, u(x, y, t) with
speed a along a straight line making an angle θ with respect to the x-axis. The
one-dimensional form is recovered with θ = 0.
For unit speed a, the two-dimensional linear convection equation is obtained from
the general divergence form, Eq. 2.3, with
Q = u (5.8)
F = iu cos θ + ju sin θ (5.9)
P = 0 (5.10)
Since Q is a scalar, F is simply a vector. Substituting these expressions into a two-

dimensional form of Eq. 2.2 gives the following integral form
d
udA + n.(iu cos θ + ju sin θ)ds = 0 (5.11)
dt A C
where A is the area of the cell which is bounded by the closed contour C.
5.2.2 The Diffusion Equation

The integral form of the two-dimensional diffusion equation with no source term and
unit diffusion coefficient ν is obtained from the general divergence form, Eq. 2.3, with
Q = u (5.12)
F = −∇u (5.13)

∂u ∂u
= − i +j (5.14)
∂x ∂y
P = 0 (5.15)
Using these, we find

d
∂u ∂u
udA = n. i +j ds (5.16)
dt A C ∂x ∂y
to be the integral form of the two-dimensional diffusion equation.
5.3 One-Dimensional Examples

We restrict our attention to a scalar dependent variable u and a scalar flux f , as in
the model equations. We consider an equispaced grid with spacing ∆x. The nodes of
the grid are located at xj = j∆x as usual. Control volume j extends from xj − ∆x/2
to xj + ∆x/2, as shown in Fig. 5.1. We will use the following notation:
xj−1/2 = xj − ∆x/2, xj+1/2 = xj + ∆x/2 (5.17)

5.3. ONE-DIMENSIONAL EXAMPLES 75
j-1/2 j+1/2
∆ x
j-2 j-1 j j+1 j+2

L R L R
Figure 5.1: Control volume in one dimension.
uj±1/2 = u(xj±1/2 ), fj±1/2 = f (uj±1/2) (5.18)

With these definitions, the cell-average value becomes
1 xj+1/2
ūj (t) ≡ u(x, t)dx (5.19)
∆x xj−1/2
and the integral form becomes
xj+1/2
d
(∆xūj ) + fj+1/2 − fj−1/2 = P dx (5.20)
dt xj−1/2
Now with ξ = x − xj , we can expand u(x) in Eq. 5.19 in a Taylor series about xj
(with t fixed) to get
 
1 ∆x/2  ∂u ξ2 ∂2 u ξ3 ∂3u
ūj ≡ uj + ξ + + + . . . dξ
∆x −∆x/2 ∂x j
2 ∂x2 j
6 ∂x3 j

2 2 4 4
∆x ∂ u ∆x ∂ u
= uj + + + O(∆x6 ) (5.21)
24 ∂x2 j
1920 ∂x4 j
or
ūj = uj + O(∆x2 ) (5.22)
where uj is the value at the center of the cell. Hence the cell-average value and the
value at the center of the cell differ by a term of second order.
5.3.1 A Second-Order Approximation to the Convection Equa-

tion
In one dimension, the integral form of the linear convection equation, Eq. 5.11, be-
comes
dūj
∆x + fj+1/2 − fj−1/2 = 0 (5.23)
dt
with f = u. We choose a piecewise constant approximation to u(x) in each cell such

that
u(x) = ūj xj−1/2 ≤ x ≤ xj+1/2 (5.24)
Evaluating this at j + 1/2 gives

L
fj+1/2 = f (uLj+1/2) = uLj+1/2 = ūj (5.25)
where the L indicates that this approximation to fj+1/2 is obtained from the approx-
imation to u(x) in the cell to the left of xj+1/2 , as shown in Fig. 5.1. The cell to the
right of xj+1/2 , which is cell j + 1, gives
R
fj+1/2 = ūj+1 (5.26)
Similarly, cell j is the cell to the right of xj−1/2 , giving

R
fj−1/2 = ūj (5.27)
and cell j − 1 is the cell to the left of xj−1/2 , giving

L
fj−1/2 = ūj−1 (5.28)
We have now accomplished the first step from the list in Section 5.1; we have defined
the fluxes at the cell boundaries in terms of the cell-average data. In this example,
the discontinuity in the flux at the cell boundary is resolved by taking the average of
the fluxes on either side of the boundary. Thus
1 L 1
fˆj+1/2 = (fj+1/2 R
+ fj+1/2 ) = (ūj + ūj+1) (5.29)
2 2
and
1 L 1
fˆj−1/2 = (fj−1/2 R
+ fj−1/2 ) = (ūj−1 + ūj ) (5.30)
2 2
where fˆ denotes a numerical flux which is an approximation to the exact flux.
Substituting Eqs. 5.29 and 5.30 into the integral form, Eq. 5.23, we obtain
dūj 1 1 dūj 1
∆x + (ūj + ūj+1) − (ūj−1 + ūj ) = ∆x + (ūj+1 − ūj−1 ) = 0 (5.31)
dt 2 2 dt 2
With periodic boundary conditions, this point operator produces the following semi-
discrete form:
dū 1
=− Bp (−1, 0, 1)ū (5.32)
dt 2∆x
This is identical to the expression obtained using second-order centered differences,

except it is written in terms of the cell average ū, rather than the nodal values, u.
Hence our analysis and understanding of the eigensystem of the matrix Bp (−1, 0, 1)
is relevant to finite-volume methods as well as finite-difference methods. Since the
eigenvalues of Bp (−1, 0, 1) are pure imaginary, we can conclude that the use of the
average of the fluxes on either side of the cell boundary, as in Eqs. 5.29 and 5.30, can
lead to a nondissipative finite-volume method.
5.3.2 A Fourth-Order Approximation to the Convection Equa-

tion
Let us replace the piecewise constant approximation in Section 5.3.1 with a piecewise
quadratic approximation as follows
u(ξ) = aξ 2 + bξ + c (5.33)
where ξ is again equal to x − xj . The three parameters a, b, and c are chosen to
satisfy the following constraints:
1 −∆x/2
u(ξ)dξ = ūj−1
∆x −3∆x/2
1 ∆x/2
u(ξ)dξ = ūj (5.34)
∆x −∆x/2
1 3∆x/2
u(ξ)dξ = ūj+1
∆x ∆x/2
These constraints lead to
ūj+1 − 2ūj + ūj−1
a=
2∆x2
ūj+1 − ūj−1
b= (5.35)
2∆x
−ūj−1 + 26ūj − ūj+1

c=
24
With these values of a, b, and c, the piecewise quadratic approximation produces
the following values at the cell boundaries:
1
uLj+1/2 = (2ūj+1 + 5ūj − ūj−1) (5.36)
6
1
uR
j−1/2 = (−ūj+1 + 5ūj + 2ūj−1 ) (5.37)
6
1
uR
j+1/2 = (−ūj+2 + 5ūj+1 + 2ūj ) (5.38)
6
1
uLj−1/2 = (2ūj + 5ūj−1 − ūj−2) (5.39)
6
using the notation defined in Section 5.3.1. Recalling that f = u, we again use the
average of the fluxes on either side of the boundary to obtain
1
fˆj+1/2 = [f (uLj+1/2) + f (uR
j+1/2 )]
2
1
= (−ūj+2 + 7ūj+1 + 7ūj − ūj−1) (5.40)
12
and
1
fˆj−1/2 = [f (uLj−1/2) + f (uR
j−1/2 )]
2
1
= (−ūj+1 + 7ūj + 7ūj−1 − ūj−2) (5.41)
12
Substituting these expressions into the integral form, Eq. 5.23, gives
dūj 1
∆x + (−ūj+2 + 8ūj+1 − 8ūj−1 + ūj−2 ) = 0 (5.42)
dt 12
This is a fourth-order approximation to the integral form of the equation, as can be
verified using Taylor series expansions (see question 1 at the end of this chapter).
With periodic boundary conditions, the following semi-discrete form is obtained:
dū 1
=− Bp (1, −8, 0, 8, −1)ū (5.43)
dt 12∆x
This is a system of ODE’s governing the evolution of the cell-average data.
5.3.3 A Second-Order Approximation to the Diffusion Equa-

tion
In this section, we describe two approaches to deriving a finite-volume approximation
to the diffusion equation. The first approach is simpler to extend to multidimensions,
while the second approach is more suited to extension to higher order accuracy.
In one dimension, the integral form of the diffusion equation, Eq. 5.16, becomes
dūj
∆x + fj+1/2 − fj−1/2 = 0 (5.44)
dt
with f = −∇u = −∂u/∂x. Also, Eq. 5.6 becomes
b ∂u
dx = u(b) − u(a) (5.45)
a ∂x
We can thus write the following expression for the average value of the gradient of u
over the interval xj ≤ x ≤ xj+1 :
1 xj+1 ∂u 1
dx = (uj+1 − uj ) (5.46)
∆x xj ∂x ∆x
From Eq. 5.22, we know that the value of a continuous function at the center of a given
interval is equal to the average value of the function over the interval to second-order
accuracy. Hence, to second-order, we can write

∂u 1
fˆj+1/2 = − =− (ūj+1 − ūj ) (5.47)
∂x j+1/2
∆x
Similarly,
1
fˆj−1/2 = − (ūj − ūj−1) (5.48)
∆x
Substituting these into the integral form, Eq. 5.44, we obtain
dūj 1
∆x = (ūj−1 − 2ūj + ūj+1) (5.49)
dt ∆x
or, with Dirichlet boundary conditions,
dū 1
= B(1, −2, 1)
ū + bc (5.50)
dt ∆x2
This provides a semi-discrete finite-volume approximation to the diffusion equation,
and we see that the properties of the matrix B(1, −2, 1) are relevant to the study of
finite-volume methods as well as finite-difference methods.
For our second approach, we use a piecewise quadratic approximation as in Section
5.3.2. From Eq. 5.33 we have
∂u ∂u
= = 2aξ + b (5.51)
∂x ∂ξ
with a and b given in Eq. 5.35. With f = −∂u/∂x, this gives

1
R
fj+1/2 L
= fj+1/2 =− (ūj+1 − ūj ) (5.52)
∆x
1
R
fj−1/2 L
= fj−1/2 =− (ūj − ūj−1 ) (5.53)
∆x
Notice that there is no discontinuity in the flux at the cell boundary. This produces
dūj 1
= (ūj−1 − 2ūj + ūj+1 ) (5.54)
dt ∆x2
which is identical to Eq. 5.49. The resulting semi-discrete form with periodic bound-
ary conditions is
dū 1
= Bp (1, −2, 1)ū (5.55)
dt ∆x2
which is written entirely in terms of cell-average data.
5.4 A Two-Dimensional Example

The above one-dimensional examples of finite-volume approximations obscure some
of the practical aspects of such methods. Thus our final example is a finite-volume
approximation to the two-dimensional linear convection equation on a grid consisting
of regular triangles, as shown in Figure 5.2. As in Section 5.3.1, we use a piecewise
constant approximation in each control volume and the flux at the control volume
boundary is the average of the fluxes obtained on either side of the boundary. The
nodal data are stored at the vertices of the triangles formed by the grid. The control
volumes are regular hexagons with area A, ∆ is the length of the sides of the triangles,
and is the length of the sides of the hexagons. The following relations hold between
, ∆, and A.
1
= √ ∆
√3
3 3 2
A =
2
2
= (5.56)
A 3∆
The two-dimensional form of the conservation law is
d
QdA + n.Fdl = 0 (5.57)
dt A C
5.4. A TWO-DIMENSIONAL EXAMPLE 81
c b
p 1
d l a
4 ∆
0
5
e f
Figure 5.2: Triangular grid.
where we have ignored the source term. The contour in the line integral is composed
of the sides of the hexagon. Since these sides are all straight, the unit normals can
be taken outside the integral and the flux balance is given by
d 5
Q dA + nν · Fdl = 0
dt A ν=0 ν
where ν indexes a side of the hexagon, as shown in Figure 5.2. A list of the normals
for the mesh orientation shown is given in Table 5.1.
Side, ν Outward Normal, n

√
0 (i −
3 j)/2
1 √i
2 (i + √3 j)/2
3 (−i + 3 j)/2
4 −i
√
5 (−i − 3 j)/2
Table 5.1. Outward normals, see Fig. 5.2.

i and j are unit normals along x and y, respectively.
For Eq. 5.11, the two-dimensional linear convection equation, we have for side ν
/2
nν · Fdl = nν · (i cos θ + j sin θ) uν (ξ)dξ (5.58)
ν −/2
where ξ is a length measured from the middle of a side ν. Making the change of
variable z = ξ/, one has the expression
/2 1/2
u(ξ)dξ = u(z)dz (5.59)
−/2 −1/2
Then, in terms of u and the hexagon area A, we have

d 5 1/2
u dA + nν · (i cos θ + j sin θ) u(z)dz = 0 (5.60)
dt A ν=0 −1/2 ν
The values of nν · (i cos θ + j sin θ) are given by the expressions in Table 5.2. There
are no numerical approximations in Eq. 5.60. That is, if the integrals in the equation
are evaluated exactly, the integrated time rate of change of the integral of u over the
area of the hexagon is known exactly.
Side, ν nν · (i cos θ + j sin θ)

√
0 (cos θ − 3 sin θ)/2
1 cos
√θ
2 (cos θ + √3 sin θ)/2
3 (− cos θ + 3 sin θ)/2
4 − cos
√θ
5 (− cos θ − 3 sin θ)/2
Table 5.2. Weights of flux integrals, see Eq. 5.60.

Introducing the cell average,

u dA = Aūp (5.61)
A
and the piecewise-constant approximation u = ūp over the entire hexagon, the ap-
proximation to the flux integral becomes trivial. Taking the average of the flux on
either side of each edge of the hexagon gives for edge 1:

ūp + ūa 1/2 ūp + ūa
u(z)dz = dz = (5.62)
1 2 −1/2 2
5.5. PROBLEMS 83
Similarly, we have for the other five edges:

ūp + ūb
u(z)dz = (5.63)
2 2

ūp + ūc
u(z)dz = (5.64)
3 2

ūp + ūd
u(z)dz = (5.65)
4 2

ūp + ūe
u(z)dz = (5.66)
5 2

ūp + ūf
u(z)dz = (5.67)
0 2
Substituting these into Eq. 5.60, along with the expressions in Table 5.2, we obtain
dūp √
A + [(2 cos θ)(ūa − ūd ) + (cos θ + 3 sin θ)(ūb − ūe )
dt 2 √
+(− cos θ + 3 sin θ)(ūc − ūf )] = 0 (5.68)
or
dūp 1 √
+ [(2 cos θ)(ūa − ūd ) + (cos θ + 3 sin θ)(ūb − ūe )
dt 3∆ √
+(− cos θ + 3 sin θ)(ūc − ūf )] = 0 (5.69)
The reader can verify, using Taylor series expansions, that this is a second-order
approximation to the integral form of the two-dimensional linear convection equation.
5.5 Problems
1. Use Taylor series to verify that Eq. 5.42 is a fourth-order approximation to Eq.
5.23.
2. Find the semi-discrete ODE form governing the cell-average data resulting from
the use of a linear approximation in developing a finite-volume method for the
linear convection equation. Use the following linear approximation:
u(ξ) = aξ + b
where b = ūj and

ūj+1 − ūj−1
a=
2∆x
and use the average flux at the cell interface.
3. Using the first approach given in Section 5.3.3, derive a finite-volume approx-
imation to the spatial terms in the two-dimensional diffusion equation on a
square grid.
4. Repeat question 3 for a grid consisting of equilateral triangles.

Chapter 6
TIME-MARCHING METHODS
FOR ODE’S
After discretizing the spatial derivatives in the governing PDE’s (such as the Navier-
Stokes equations), we obtain a coupled system of nonlinear ODE’s in the form
du
= F (u, t) (6.1)
dt
These can be integrated in time using a time-marching method to obtain a time-
accurate solution to an unsteady flow problem. For a steady flow problem, spatial
discretization leads to a coupled system of nonlinear algebraic equations in the form
F (u) = 0 (6.2)
As a result of the nonlinearity of these equations, some sort of iterative method is
required to obtain a solution. For example, one can consider the use of Newton’s
method, which is widely used for nonlinear algebraic equations (See Section 6.10.3.).
This produces an iterative method in which a coupled system of linear algebraic
equations must be solved at each iteration. These can be solved iteratively using
relaxation methods, which will be discussed in Chapter 9, or directly using Gaussian
elimination or some variation thereof.
Alternatively, one can consider a time-dependent path to the steady state and use
a time-marching method to integrate the unsteady form of the equations until the
solution is sufficiently close to the steady solution. The subject of the present chapter,
time-marching methods for ODE’s, is thus relevant to both steady and unsteady flow
problems. When using a time-marching method to compute steady flows, the goal is
simply to remove the transient portion of the solution as quickly as possible; time-
accuracy is not required. This motivates the study of stability and stiffness, topics
which are covered in the next two chapters.
85
86 CHAPTER 6. TIME-MARCHING METHODS FOR ODE’S
Application of a spatial discretization to a PDE produces a coupled system of

ODE’s. Application of a time-marching method to an ODE produces an ordinary
difference equation (O∆E ). In earlier chapters, we developed exact solutions to our
model PDE’s and ODE’s. In this chapter we will present some basic theory of linear
O∆E’s which closely parallels that for linear ODE’s, and, using this theory, we will
develop exact solutions for the model O∆E’s arising from the application of time-
marching methods to the model ODE’s.
6.1 Notation
Using the semi-discrete approach, we reduce our PDE to a set of coupled ODE’s
represented in general by Eq. 4.1. However, for the purpose of this chapter, we need
only consider the scalar case
du
= u = F (u, t) (6.3)
dt
Although we use u to represent the dependent variable, rather than w, the reader
should recall the arguments made in Chapter 4 to justify the study of a scalar ODE.
Our first task is to find numerical approximations that can be used to carry out the
time integration of Eq. 6.3 to some given accuracy, where accuracy can be measured
either in a local or a global sense. We then face a further task concerning the numerical
stability of the resulting methods, but we postpone such considerations to the next
chapter.
In Chapter 2, we introduced the convention that the n subscript, or the (n) su-
perscript, always points to a discrete time value, and h represents the time interval
∆t. Combining this notation with Eq. 6.3 gives
un = Fn = F (un, tn ) ; tn = nh
Often we need a more sophisticated notation for intermediate time steps involving
temporary calculations denoted by ũ, ū, etc. For these we use the notation
ũn+α = F̃n+α = F (ũn+α, tn + αh)
The choice of u or F to express the derivative in a scheme is arbitrary. They are

both commonly used in the literature on ODE’s.
The methods we study are to be applied to linear or nonlinear ODE’s, but the
methods themselves are formed by linear combinations of the dependent variable and
its derivative at various time intervals. They are represented conceptually by

un+1 = f β1 un+1, β0 un , β−1 un−1 , · · · , α0 un , α−1 un−1 , · · · (6.4)
6.2. CONVERTING TIME-MARCHING METHODS TO O∆E’S 87
With an appropriate choice of the α s and β s, these methods can be constructed

to give a local Taylor series accuracy of any order. The methods are said to be
explicit if β1 = 0 and implicit otherwise. An explicit method is one in which the new
predicted solution is only a function of known data, for example, un, un−1, un , and
un−1 for a method using two previous time levels, and therefore the time advance is
simple. For an implicit method, the new predicted solution is also a function of the
time derivative at the new time level, that is, un+1 . As we shall see, for systems of
ODE’s and nonlinear problems, implicit methods require more complicated strategies
to solve for un+1 than explicit methods.
6.2 Converting Time-Marching Methods to O∆E’s

Examples of some very common forms of methods used for time-marching general
ODE’s are:
un+1 = un + hun (6.5)
un+1 = un + hun+1 (6.6)
and
ũn+1 = un + hun
1
un+1 = [un + ũn+1 + hũn+1 ] (6.7)
2
According to the conditions presented under Eq. 6.4, the first and third of these are
examples of explicit methods. We refer to them as the explicit Euler method and the
MacCormack predictor-corrector method,1 respectively. The second is implicit and
referred to as the implicit (or backward) Euler method.
These methods are simple recipes for the time advance of a function in terms of its
value and the value of its derivative, at given time intervals. The material presented
in Chapter 4 develops a basis for evaluating such methods by introducing the concept
of the representative equation
du
= u = λu + aeµt (6.8)
dt
written here in terms of the dependent variable, u. The value of this equation arises
from the fact that, by applying a time-marching method, we can analytically convert
1
Here we give only MacCormack’s time-marching method. The method commonly referred to as
MacCormack’s method, which is a fully-discrete method, will be presented in Section 11.3
such a linear ODE into a linear O∆E . The latter are subject to a whole body of
analysis that is similar in many respects to, and just as powerful as, the theory of
ODE’s. We next consider examples of this conversion process and then go into the
general theory on solving O∆E’s.
Apply the simple explicit Euler scheme, Eq. 6.5, to Eq. 6.8. There results
un+1 = un + h(λun + aeµhn )
or
un+1 − (1 + λh)un = haeµhn (6.9)
Eq. 6.9 is a linear O∆E , with constant coefficients, expressed in terms of the depen-
dent variable un and the independent variable n. As another example, applying the
implicit Euler method, Eq. 6.6, to Eq. 6.8, we find

un+1 = un + h λun+1 + aeµh(n+1)
or
(1 − λh)un+1 − un = heµh · aeµhn (6.10)
As a final example, the predictor-corrector sequence, Eq. 6.7, gives
ũn+1 − (1 + λh)un = aheµhn
1 1 1
− (1 + λh)ũn+1 + un+1 − un = aheµh(n+1) (6.11)
2 2 2
which is a coupled set of linear O∆E’s with constant coefficients. Note that the first
line in Eq. 6.11 is identical to Eq. 6.9, since the predictor step in Eq. 6.7 is simply
the explicit Euler method. The second line in Eq. 6.11 is obtained by noting that
ũn+1 = F (ũn+1, tn + h)
= λũn+1 + aeµh(n+1) (6.12)
Now we need to develop techniques for analyzing these difference equations so that
we can compare the merits of the time-marching methods that generated them.
6.3 Solution of Linear O∆E’s With Constant Co-

efficients
The techniques for solving linear difference equations with constant coefficients is as
well developed as that for ODE’s and the theory follows a remarkably parallel path.
This is demonstrated by repeating some of the developments in Section 4.2, but for
difference rather than differential equations.
6.3. SOLUTION OF LINEAR O∆E’S WITH CONSTANT COEFFICIENTS 89
6.3.1 First- and Second-Order Difference Equations

First-Order Equations
The simplest nonhomogeneous O∆E of interest is given by the single first-order equa-
tion
un+1 = σun + abn (6.13)
where σ, a, and b are, in general, complex parameters. The independent variable is

now n rather than t, and since the equations are linear and have constant coefficients,
σ is not a function of either n or u. The exact solution of Eq. 6.13 is
abn
un = c1 σ n +
b−σ
where c1 is a constant determined by the initial conditions. In terms of the initial
value of u it can be written
bn − σn
u n = u0 σ n + a
b−σ
Just as in the development of Eq. 4.10, one can readily show that the solution of the
defective case, (b = σ),
un+1 = σun + aσn
is
un = u0 + anσ−1 σ n
This can all be easily verified by substitution.
Second-Order Equations
The homogeneous form of a second-order difference equation is given by
un+2 + a1 un+1 + a0 un = 0 (6.14)
Instead of the differential operator D ≡ dtd used for ODE’s, we use for O∆E’s the
difference operator E (commonly referred to as the displacement or shift operator)
and defined formally by the relations
un+1 = Eun , un+k = E k un
Further notice that the displacement operator also applies to exponents, thus
bα · bn = bn+α = E α · bn
where α can be any fraction or irrational number.

The roles of D and E are the same insofar as once they have been introduced to
the basic equations the value of u(t) or un can be factored out. Thus Eq. 6.14 can
now be re-expressed in an operational notion as
(E 2 + a1 E + a0 )un = 0 (6.15)
which must be zero for all un . Eq. 6.15 is known as the operational form of Eq. 6.14.
The operational form contains a characteristic polynomial P (E) which plays the same
role for difference equations that P (D) played for differential equations; that is, its
roots determine the solution to the O∆E. In the analysis of O∆E’s, we label these
roots σ1 , σ2 , · · ·, etc, and refer to them as the σ-roots. They are found by solving the
equation P (σ) = 0. In the simple example given above, there are just two σ roots
and in terms of them the solution can be written
un = c1 (σ1 )n + c2 (σ2 )n (6.16)
where c1 and c2 depend upon the initial conditions. The fact that Eq. 6.16 is a
solution to Eq. 6.14 for all c1 , c2 and n should be verified by substitution.
6.3.2 Special Cases of Coupled First-Order Equations

A Complete System
Coupled, first-order, linear homogeneous difference equations have the form
(n+1) (n) (n)
u1 = c11 u1 + c12 u2
(n+1) (n) (n)
u2 = c21 u1 + c22 u2 (6.17)
which can also be written

T
un+1 = Cun un = u(n) (n) c c12
, 1 , u2 , C = 11
c21 c22
The operational form of Eq. 6.17 can be written
(n)
(c11 − E) c12 u1
= [ C − E I ]un = 0
c21 (c22 − E) u2
which must be zero for all u1 and u2 . Again we are led to a characteristic polynomial,
this time having the form P (E) = det [ C − E I ]. The σ-roots are found from

(c11 − σ) c12
P (σ) = det =0
c21 (c22 − σ)
6.4. SOLUTION OF THE REPRESENTATIVE O∆E’S 91
Obviously the σk are the eigenvalues of C and, following the logic of Section 4.2,

if x are its eigenvectors, the solution of Eq. 6.17 is

2
un = ck (σk )nxk
k=1
where ck are constants determined by the initial conditions.
A Defective System
The solution of O∆E’s with defective eigensystems follows closely the logic in Section
4.2.2 for defective ODE’s. For example, one can show that the solution to
    
ūn+1 σ ūn
    
 ûn+1  =  1 σ   ûn 
un+1 1 σ un
is
ūn = ū0 σ n
ûn = û0 + ū0 nσ−1 σ n

−1 n(n − 1) −2 n
un = u0 + û0 nσ + ū0 σ σ (6.18)
2
6.4 Solution of the Representative O∆E’s

6.4.1 The Operational Form and its Solution
Examples of the nonhomogeneous, linear, first-order ordinary difference equations,
produced by applying a time-marching method to the representative equation, are
given by Eqs. 6.9 to 6.11. Using the displacement operator, E, these equations can
be written
[E − (1 + λh)]un = h · aeµhn (6.19)
[(1 − λh)E − 1]un = h · E · aeµhn (6.20)

E −(1 + λh) ũ 1
1 =h· · aeµhn (6.21)
− 2 (1 + λh)E E − 12 u n
1E
2
All three of these equations are subsets of the operational form of the representative
O∆E
P (E)un = Q(E) · aeµhn (6.22)
which is produced by applying time-marching methods to the representative ODE, Eq.

4.45. We can express in terms of Eq. 6.22 all manner of standard time-marching meth-
ods having multiple time steps and various types of intermediate predictor-corrector
families. The terms P (E) and Q(E) are polynomials in E referred to as the charac-
teristic polynomial and the particular polynomial, respectively.
The general solution of Eq. 6.22 can be expressed as

K
Q(eµh )
un = n
ck (σk ) + ae µhn
· (6.23)
k=1 P (eµh )
where σk are the K roots of the characteristic polynomial, P (σ) = 0. When determi-
nants are involved in the construction of P (E) and Q(E), as would be the case for
Eq. 6.21, the ratio Q(E)/P (E) can be found by Kramer’s rule. Keep in mind that
for methods such as in Eq. 6.21 there are multiple (two in this case) solutions, one
for un and ũn and we are usually only interested in the final solution un. Notice also,
the important subset of this solution which occurs when µ = 0, representing a time
invariant particular solution, or a steady state. In such a case

K
Q(1)
un = ck (σk )n + a ·
k=1 P (1)
6.4.2 Examples of Solutions to Time-Marching O∆E’s

As examples of the use of Eqs. 6.22 and 6.23, we derive the solutions of Eqs. 6.19 to
6.21. For the explicit Euler method, Eq. 6.19, we have
P (E) = E − 1 − λh
Q(E) = h (6.24)
and the solution of its representative O∆E follows immediately from Eq. 6.23:
h
un = c1 (1 + λh)n + aeµhn ·
e µh
− 1 − λh
For the implicit Euler method, Eq. 6.20, we have
P (E) = (1 − λh)E − 1
Q(E) = hE (6.25)
6.5. THE λ − σ RELATION 93
so n
1 heµh
un = c1 + ae µhn
·
1 − λh (1 − λh)eµh − 1
In the case of the coupled predictor-corrector equations, Eq. 6.21, one solves for the
final family un (one can also find a solution for the intermediate family ũ), and there
results

E −(1 + λh) 1
P (E) = det 1 1 = E E − 1 − λh − λ2 h2
− 2 (1 + λh)E E−2 2

E h 1
Q(E) = det 1 hE = 2 hE(E + 1 + λh)
− 12 (1 + λh)E 2
The σ-root is found from

1
P (σ) = σ σ − 1 − λh − λ2 h2 = 0
2
which has only one nontrivial root (σ = 0 is simply a shift in the reference index).
The complete solution can therefore be written
1 µh
n h e + 1 + λh
1
un = c1 1 + λh + λ2 h2 + aeµhn · 2 (6.26)
2 1
eµh − 1 − λh − λ2 h2
2
6.5 The λ − σ Relation

6.5.1 Establishing the Relation
We have now been introduced to two basic kinds of roots, the λ-roots and the σ-roots.
The former are the eigenvalues of the A matrix in the ODE’s found by space differ-
encing the original PDE, and the latter are the roots of the characteristic polynomial
in a representative O∆E found by applying a time-marching method to the repre-
sentative ODE. There is a fundamental relation between the two which can be used
to identify many of the essential properties of a time-march method. This relation is
first demonstrated by developing it for the explicit Euler method.
First we make use of the semi-discrete approach to find a system of ODE’s and
then express its solution in the form of Eq. 4.27. Remembering that t = nh, one can
write
n n n
u(t) = c1 eλ1 h x1 + · · · + cm eλm h xm + · · · + cM eλM h xM + P.S. (6.27)
where for the present we are not interested in the form of the particular solution
(P.S.). Now the explicit Euler method produces for each λ-root, one σ-root, which
is given by σ = 1 + λh. So if we use the Euler method for the time advance of the
ODE’s, the solution2 of the resulting O∆E is
un = c1 (σ1 )n x1 + · · · + cm (σm )n xm + · · · + cM (σM )n xM + P.S. (6.28)
where the cm and the xm in the two equations are identical and σm = (1 + λm h).
Comparing Eq. 6.27 and Eq. 6.28, we see a correspondence between σm and eλm h .
Since the value of eλh can be expressed in terms of the series
1 1 1
eλh = 1 + λh + λ2 h2 + λ3 h3 + · · · + λn hn + · · ·
2 6 n!
the truncated expansion σ = 1 + λh is a reasonable3 approximation for small enough

λh.
Suppose, instead of the Euler method, we use the leapfrog method for the time
advance, which is defined by
un+1 = un−1 + 2hun (6.29)
Applying Eq. 6.8 to Eq. 6.29, we have the characteristic polynomial P (E) = E2 −
2λhE − 1, so that for every λ the σ must satisfy the relation
2
σm − 2λm hσm − 1 = 0 (6.30)
Now we notice that each λ produces two σ-roots. For one of these we find

σm = λm h + 1 + λ2m h2 (6.31)
1 1
= 1 + λm h + λ2m h2 − λ4m h4 + · · · (6.32)
2 8
This
is an approximation to eλm h with an error O(λ3 h3 ). The other root, λm h −
1 + λ2m h2 , will be discussed in Section 6.5.3.
2
Based on Section 4.4.
3
The error is O(λ2 h2 ).
6.5. THE λ − σ RELATION 95
6.5.2 The Principal σ-Root

Based on the above we make the following observation:
Application of the same time-marching method to all of the equations in

a coupled system linear ODE’s in the form of Eq. 4.6, always produces
one σ-root for every λ-root that satisfies the relation

(6.33)
1 1
σ = 1 + λh + λ2 h2 + · · · + λk hk + O hk+1
2 k!
where k is the order of the time-marching method.
We refer to the root that has the above property as the principal σ-root, and designate
it (σm )1 . The above property can be stated regardless of the details of the time-
marching method, knowing only that its leading error is O hk+1 . Thus the principal

root is an approximation to eλh up to O hk .
Note that a second-order approximation to a derivative written in the form
1
(δt u)n = (un+1 − un−1 ) (6.34)
2h
has a leading truncation error which is O(h2 ), while the second-order time-marching
method which results from this approximation, which is the leapfrog method:
un+1 = un−1 + 2hun (6.35)
has a leading truncation error O(h3 ). This arises simply because of our notation
for the time-marching method in which we have multiplied through by h to get an
approximation for the function un+1 rather than the derivative as in Eq. 6.34. The
following example makes this clear. Consider a solution obtained at a given time T
using a second-order time-marching method with a time step h. Now consider the
solution obtained using the same method with a time step h/2. Since the error per
time step is O(h3 ), this is reduced by a factor of eight (considering the leading term
only). However, twice as many time steps are required to reach the time T . Therefore
the error at the end of the simulation is reduced by a factor of four, consistent with
a second-order approximation.
6.5.3 Spurious σ-Roots

We saw from Eq. 6.30 that the λ − σ relation for the leapfrog method produces two
σ-roots for each λ. One of these we identified as the principal root which always
has the property given in 6.33. The other is referred to as a spurious σ-root and
designated (σm )2 . In general, the λ − σ relation produced by a time-marching scheme
can result in multiple σ-roots all of which, except for the principal one, are spurious.
All spurious roots are designated (σm )k where k = 2, 3, · · ·. No matter whether a
σ-root is principal or spurious, it is always some algebraic function of the product λh.
To express this fact we use the notation σ = σ(λh).
If a time-marching method produces spurious σ-roots, the solution for the O∆E in
the form shown in Eq. 6.28 must be modified. Following again the message of Section
4.4, we have
un = c11 (σ1 )n1 x1 + · · · + cm1(σm )n1 xm + · · · + cM 1 (σM )n1 xM + P.S.
+c12 (σ1 )n x1 + · · · + cm2 (σm )n xm + · · · + cM 2 (σM )n xM
2 2 2
+c13 (σ1 )n3
x1 + · · · + cm3 (σm )n xm + · · · + cM 3 (σM )n xM
3 3
+etc., if there are more spurious roots (6.36)
Spurious roots arise if a method uses data from time level n − 1 or earlier to
advance the solution from time level n to n + 1. Such roots originate entirely from
the numerical approximation of the time-marching method and have nothing to do
with the ODE being solved. However, generation of spurious roots does not, in itself,
make a method inferior. In fact, many very accurate methods in practical use for
integrating some forms of ODE’s have spurious roots.
It should be mentioned that methods with spurious roots are not self starting.
For example, if there is one spurious root to a method, all of the coefficients (cm )2
in Eq. 6.36 must be initialized by some starting procedure. The initial vector u0
does not provide enough data to initialize all of the coefficients. This results because
methods which produce spurious roots require data from time level n − 1 or earlier.
For example, the leapfrog method requires un−1 and thus cannot be started using
only un .
Presumably (i.e., if one starts the method properly) the spurious coefficients are
all initialized with very small magnitudes, and presumably the magnitudes of the
spurious roots themselves are all less than one (see Chapter 7). Then the presence of
spurious roots does not contaminate the answer. That is, after some finite time the
amplitude of the error associated with the spurious roots is even smaller then when
it was initialized. Thus while spurious roots must be considered in stability analysis,
they play virtually no role in accuracy analysis.
6.5.4 One-Root Time-Marching Methods

There are a number of time-marching methods that produce only one σ-root for each
λ-root. We refer to them as one-root methods. They are also called one-step methods.
6.6. ACCURACY MEASURES OF TIME-MARCHING METHODS 97
They have the significant advantage of being self-starting which carries with it the
very useful property that the time-step interval can be changed at will throughout
the marching process. Three one-root methods were analyzed in Section 6.4.2. A
popular method having this property, the so-called θ-method, is given by the formula

un+1 = un + h (1 − θ)un + θun+1
The θ-method represents the explicit Euler (θ = 0), the trapezoidal (θ = 12 ), and the
implicit Euler methods (θ = 1), respectively. Its λ − σ relation is
1 + (1 − θ)λh
σ=
1 − θλh
It is instructive to compare the exact solution to a set of ODE’s (with a complete
eigensystem) having time-invariant forcing terms with the exact solution to the O∆E’s
for one-root methods. These are
n n n
u(t) = c1 eλ1 h x1 + · · · + cm eλm h xm + · · · + cM eλM h xM + A−1f
un = c1 (σ1 )n x1 + · · · + cm (σm )n xm + · · · + cM (σM )n xM + A−1f (6.37)
respectively. Notice that when t and n = 0, these equations are identical, so that all
the constants, vectors, and matrices are identical except the u and the terms inside
the parentheses on the right hand sides. The only error made by introducing the time
marching is the error that σ makes in approximating eλh .
6.6 Accuracy Measures of Time-Marching Meth-

ods
6.6.1 Local and Global Error Measures
There are two broad categories of errors that can be used to derive and evaluate time-
marching methods. One is the error made in each time step. This is a local error such
as that found from a Taylor table analysis, see Section 3.4. It is usually used as the
basis for establishing the order of a method. The other is the error determined at the
end of a given event which has covered a specific interval of time composed of many
time steps. This is a global error. It is useful for comparing methods, as we shall see
in Chapter 8.
It is quite common to judge a time-marching method on the basis of results found
from a Taylor table. However, a Taylor series analysis is a very limited tool for finding
the more subtle properties of a numerical time-marching method. For example, it is
of no use in:
• finding spurious roots.
• evaluating numerical stability and separating the errors in phase and amplitude.
• analyzing the particular solution of predictor-corrector combinations.
• finding the global error.
The latter three of these are of concern to us here, and to study them we make use of
the material developed in the previous sections of this chapter. Our error measures
are based on the difference between the exact solution to the representative ODE,
given by
aeµt
u(t) = ceλt + (6.38)
µ−λ
and the solution to the representative O∆E’s, including only the contribution from
the principal root, which can be written as
n Q(eµh )
un = c1 (σ1 ) + ae µhn
· (6.39)
P (eµh )
6.6.2 Local Accuracy of the Transient Solution (erλ, |σ| , erω )

Transient error
The particular choice of an error measure, either local or global, is to some extent
arbitrary. However, a necessary condition for the choice should be that the measure
can be used consistently for all methods. In the discussion of the λ-σ relation we
saw that all time-marching methods produce a principal σ-root for every λ-root that
exists in a set of linear ODE’s. Therefore, a very natural local error measure for the
transient solution is the value of the difference between solutions based on these two
roots. We designate this by erλ and make the following definition
erλ ≡ eλh − σ1
The leading error term can be found by expanding in a Taylor series and choosing
the first nonvanishing term. This is similar to the error found from a Taylor table.
The order of the method is the last power of λh matched exactly.
Amplitude and Phase Error

Suppose a λ eigenvalue is imaginary. Such can indeed be the case when we study the
equations governing periodic convection which produces harmonic motion. For such
cases it is more meaningful to express the error in terms of amplitude and phase.
Let λ = iω where ω is a real number representing a frequency. Then the numerical
method must produce a principal σ-root that is complex and expressible in the form
σ1 = σr + iσi ≈ eiωh (6.40)
From this it follows that the local error in amplitude is measured by the deviation of
|σ1 | from unity, that is

era = 1 − |σ1 | = 1 − (σ1 )2r + (σ1 )2i
and the local error in phase can be defined as
erω ≡ ωh − tan−1 [(σ1 )i /(σ1 )r )] (6.41)
Amplitude and phase errors are important measures of the suitability of time-marching
methods for convection and wave propagation phenomena.
The approach to error analysis described in Section 3.5 can be extended to the
combination of a spatial discretization and a time-marching method applied to the
linear convection equation. The principal root, σ1 (λh), is found using λ = −iaκ∗ ,
where κ∗ is the modified wavenumber of the spatial discretization. Introducing the
Courant number, Cn = ah/∆x, we have λh = −iCn κ∗ ∆x. Thus one can obtain
values of the principal root over the range 0 ≤ κ∆x ≤ π for a given value of the
Courant number. The above expression for erω can be normalized to give the error
in the phase speed, as follows
erω tan−1 [(σ1 )i /(σ1 )r )]
erp = =1+ (6.42)
ωh Cn κ∆x
where ω = −aκ. A positive value of erp corresponds to phase lag (the numerical phase
speed is too small), while a negative value corresponds to phase lead (the numerical
phase speed is too large).
6.6.3 Local Accuracy of the Particular Solution (erµ )

The numerical error in the particular solution is found by comparing the particular
solution of the ODE with that for the O∆E. We have found these to be given by
1
P.S.(ODE) = aeµt ·
(µ − λ)
and
Q(eµh )
P.S.(O∆E) = ae ·
µt
P (eµh )
respectively. For a measure of the local error in the particular solution we introduce
the definition

P.S.(O∆E)
erµ ≡ h −1 (6.43)
P.S.(ODE)
The multiplication by h converts the error from a global measure to a local one, so
that the order of erλ and erµ are consistent. In order to determine the leading error
term, Eq. 6.43 can be written in terms of the characteristic and particular polynomials
as
co ! "
erµ = · (µ − λ)Q eµh − P eµh (6.44)
µ−λ
where
h(µ − λ)
co = lim
h→0 P eµh
The value of co is a method-dependent constant that is often equal to one. If the
forcing function is independent of time, µ is equal to zero, and for this case, many
numerical methods generate an erµ that is also zero.
The algebra involved in finding the order of erµ can be quite tedious. However,
this order is quite important in determining the true order of a time-marching method
by the process that has been outlined. An illustration of this is given in the section
on Runge-Kutta methods.
6.6.4 Time Accuracy For Nonlinear Applications

In practice, time-marching methods are usually applied to nonlinear ODE’s, and it
is necessary that the advertised order of accuracy be valid for the nonlinear cases as
well as for the linear ones. A necessary condition for this to occur is that the local
accuracies of both the transient and the particular solutions be of the same order.
More precisely, a time-marching method is said to be of order k if
erλ = c1 · (λh)k1 +1 (6.45)

erµ = c2 · (λh)k2 +1 (6.46)
where k = smallest of(k1 , k2 ) (6.47)
The reader should be aware that this is not sufficient. For example, to derive all of
the necessary conditions for the fourth-order Runge-Kutta method presented later
in this chapter the derivation must be performed for a nonlinear ODE. However, the
analysis based on a linear nonhomogeneous ODE produces the appropriate conditions
for the majority of time-marching methods used in CFD.
6.6.5 Global Accuracy

In contrast to the local error measures which have just been discussed, we can also
define global error measures. These are useful when we come to the evaluation of
time-marching methods for specific purposes. This subject is covered in Chapter 8
after our introduction to stability in Chapter 7.
Suppose we wish to compute some time-accurate phenomenon over a fixed interval
of time using a constant time step. We refer to such a computation as an “event”.
Let T be the fixed time of the event and h be the chosen step size. Then the required
number of time steps, is N, given by the relation
T = Nh
Global error in the transient
A natural extension of erλ to cover the error in an entire event is given by
Erλ ≡ eλT − (σ1 (λh))N (6.48)
Global error in amplitude and phase
If the event is periodic, we are more concerned with the global error in amplitude and
phase. These are given by
N
Era = 1 − (σ1 )2r + (σ1 )2i (6.49)
and

(σ1 )i
−1
Erω ≡ N ωh − tan
(σ1 )r
= ωT − N tan−1 [(σ1 )i /(σ1 )r ] (6.50)
Global error in the particular solution

Finally, the global error in the particular solution follows naturally by comparing the
solutions to the ODE and the O∆E. It can be measured by

Q eµh
Erµ ≡ (µ − λ) µh − 1
P e
6.7 Linear Multistep Methods

In the previous sections, we have developed the framework of error analysis for time
advance methods and have randomly introduced a few methods without addressing
motivational, developmental or design issues. In the subsequent sections, we introduce
classes of methods along with their associated error analysis. We shall not spend much
time on development or design of these methods, since most of them have historic
origins from a wide variety of disciplines and applications. The Linear Multistep
Methods (LMM’s) are probably the most natural extension to time marching of the
space differencing schemes introduced in Chapter 3 and can be analyzed for accuracy
or designed using the Taylor table approach of Section 3.4.
6.7.1 The General Formulation

When applied to the nonlinear ODE
du
= u = F(u, t)
dt
all linear multistep methods can be expressed in the general form

1
1
αk un+k = h βk Fn+k (6.51)
k=1−K k=1−K
where the notation for F is defined in Section 6.1. The methods are said to be linear
because the α’s and β’s are independent of u and n, and they are said to be K-step
because K time-levels of data are required to marching the solution one time-step, h.
They are explicit if β1 = 0 and implicit otherwise.
When Eq. 6.51 is applied to the representative equation, Eq. 6.8, and the result is
expressed in operational form, one finds
   

1
1
 αk E k un = h βk E k (λun + aeµhn ) (6.52)
k=1−K k=1−K
6.7. LINEAR MULTISTEP METHODS 103
We recall from Section 6.5.2 that a time-marching method when applied to the repre-
sentative equation must provide a σ-root, labeled σ1 , that approximates eλh through
the order of the method. The condition referred to as consistency simply means that
σ → 1 as h → 0, and it is certainly a necessary condition for the accuracy of any
time marching method. We can also agree that, to be of any value in time accuracy,
a method should at least be first-order accurate, that is σ → (1 + λh) as h → 0. One
can show that these conditions are met by any method represented by Eq. 6.51 if

αk = 0 and βk = (K + k − 1)αk
k k k
Since both sides of Eq. 6.51 can be multiplied by an arbitrary constant, these methods
are often “normalized” by requiring

βk = 1
k
Under this condition co = 1 in Eq. 6.44.
6.7.2 Examples
There are many special explicit and implicit forms of linear multistep methods. Two
well-known families of them, referred to as Adams-Bashforth (explicit) and Adams-
Moulton (implicit), can be designed using the Taylor table approach of Section 3.4.
The Adams-Moulton family is obtained from Eq. 6.51 with
α1 = 1, α0 = −1, αk = 0, k = −1, −2, · · · (6.53)
The Adams-Bashforth family has the same α’s with the additional constraint that
β1 = 0. The three-step Adams-Moulton method can be written in the following form
un+1 = un + h(β1 un+1 + β0 un + β−1 un−1 + β−2 un−2 ) (6.54)
A Taylor table for Eq. 6.54 can be generated as
un h · un h2 · un h3 · u

n h4 · u
n
un+1 1 1 1 1 1
2 6 24
−un −1
−hβ1 un+1 −β1 −β1 −β1 12 −β1 16
−hβ0 un −β0
−hβ−1 un−1 −β−1 β−1 −β−1 12 β−1 16
−hβ−2 un−2 −(−2)0 β−2 −(−2)1 β−2 −(−2)2 β−2 12 −(−2)3 β−2 16
Table 6.1. Taylor table for the Adams-Moulton three-step linear multistep method.
This leads to the linear system

    
1 1 1 1 β1 1
 2 0 −2 −4  β0   1 
    
   =  (6.55)
 3 0 3 12  β−1   1 
4 0 −4 −32 β−2 1
to solve for the β’s, resulting in
β1 = 9/24, β0 = 19/24, β−1 = −5/24, β−2 = 1/24 (6.56)
which produces a method which is fourth-order accurate.4 With β1 = 0 one obtains
    
1 1 1 β0 1
    
 0 −2 −4   −1 
β =  1  (6.57)
0 3 12 β−2 1
giving
β0 = 23/12, β−1 = −16/12, β−2 = 5/12 (6.58)
This is the third-order Adams-Bashforth method.
A list of simple methods, some of which are very common in CFD applications,
is given below together with identifying names that are sometimes associated with
them. In the following material AB(n) and AM(n) are used as abbreviations for the
(n)th order Adams-Bashforth and (n)th order Adams-Moulton methods. One can
verify that the Adams type schemes given below satisfy Eqs. 6.55 and 6.57 up to the
order of the method.
Explicit Methods
un+1 = un + hun Euler

un+1 = un−1 + 2hu
n Leapfrog
un+1 = un + 21 h 3un − un−1 AB2

un+1 h 23u − 16u + 5u
= un + 12 AB3
n n−1 n−2
Implicit Methods
un+1 = un + hun+1

Implicit Euler
un+1 1
= un + 2 h un + un+1 Trapezoidal (AM2)

un+1 = 13 4un − un−1 + 2hun+1 2nd-order Backward

un+1 h 5u + 8u − u
= un + 12 AM3
n+1 n n−1
4
Recall from Section 6.5.2 that a kth-order time-marching method has a leading truncation error
term which is O(hk+1 ).
6.7. LINEAR MULTISTEP METHODS 105
6.7.3 Two-Step Linear Multistep Methods

High resolution CFD problems usually require very large data sets to store the spatial
information from which the time derivative is calculated. This limits the interest
in multistep methods to about two time levels. The most general two-step linear
multistep method (i.e., K=2 in Eq. 6.51), that is at least first-order accurate, can be
written as

(1 + ξ)un+1 = [(1 + 2ξ)un − ξun−1] + h θun+1 + (1 − θ + ϕ)un − ϕun−1 (6.59)
Clearly the methods are explicit if θ = 0 and implicit otherwise. A list of methods
contained in Eq. 6.59 is given in Table 6.2. Notice that the Adams methods have
ξ = 0, which corresponds to α−1 = 0 in Eq. 6.51. Methods with ξ = −1/2, which
corresponds to α0 = 0 in Eq. 6.51, are known as Milne methods.
θ ξ ϕ Method Order
0 0 0 Euler 1
1 0 0 Implicit Euler 1
1/2 0 0 Trapezoidal or AM2 2
1 1/2 0 2nd Order Backward 2
3/4 0 −1/4 Adams type 2
1/3 −1/2 −1/3 Lees Type 2
1/2 −1/2 −1/2 Two–step trapezoidal 2
5/9 −1/6 −2/9 A–contractive 2
0 −1/2 0 Leapfrog 2
0 0 1/2 AB2 2
0 −5/6 −1/3 Most accurate explicit 3
1/3 −1/6 0 Third–order implicit 3
5/12 0 1/12 AM3 3
1/6 −1/2 −1/6 Milne 4
Table 6.2. Some linear one- and two-step methods, see Eq. 6.59.
One can show after a little algebra that both erµ and erλ are reduced to 0(h3 )
(i.e., the methods are 2nd-order accurate) if
1
ϕ=ξ−θ+
2
The class of all 3rd-order methods is determined by imposing the additional constraint
5
ξ = 2θ −
6
Finally a unique fourth-order method is found by setting θ = −ϕ = −ξ/3 = 16 .
6.8 Predictor-Corrector Methods

There are a wide variety of predictor-corrector schemes created and used for a variety
of purposes. Their use in solving ODE’s is relatively easy to illustrate and understand.
Their use in solving PDE’s can be much more subtle and demands concepts5 which
have no counterpart in the analysis of ODE’s.
Predictor-corrector methods constructed to time-march linear or nonlinear ODE’s
are composed of sequences of linear multistep methods, each of which is referred to
as a family in the solution process. There may be many families in the sequence, and
usually the final family has a higher Taylor-series order of accuracy than the inter-
mediate ones. Their use is motivated by ease of application and increased efficiency,
where measures of efficiency are discussed in the next two chapters.
A simple one-predictor, one-corrector example is given by
ũn+α = un + αhun

un+1 = un + h β ũn+α + γun (6.60)
where the parameters α, β and γ are arbitrary parameters to be determined. One

can analyze this sequence by applying it to the representative equation and using
the operational techniques outlined in Section 6.4. It is easy to show, following the
example leading to Eq. 6.26, that

P (E) = E α · E − 1 − (γ + β)λh − αβλ2h2 (6.61)
Q(E) = E α · h · [βE α + γ + αβλh] (6.62)
Considering only local accuracy, one is led, by following the discussion in Section 6.6,
to the following observations. For the method to be second-order accurate both erλ
and erµ must be O(h3 ). For this to hold for erλ , it is obvious from Eq. 6.61 that
1
γ+β =1 ; αβ =
2
which provides two equations for three unknowns. The situation for erµ requires some
algebra, but it is not difficult to show using Eq. 6.44 that the same conditions also
make it O(h3 ). One concludes, therefore, that the predictor-corrector sequence
ũn+α = un + αhun

1 1 2α − 1
un+1 = un + h ũ + un (6.63)
2 α n+α α
is a second-order accurate method for any α.
5
Such as alternating direction, fractional-step, and hybrid methods.
6.9. RUNGE-KUTTA METHODS 107
A classical predictor-corrector sequence is formed by following an Adams-Bashforth

predictor of any order with an Adams-Moulton corrector having an order one higher.
The order of the combination is then equal to the order of the corrector. If the order
of the corrector is (k), we refer to these as ABM(k) methods. The Adams-Bashforth-
Moulton sequence for k = 3 is
1
ũn+1 = un + h 3un − un−1
2
h
un+1 = un + 5ũn+1 + 8un − un−1 (6.64)
12
Some simple, specific, second-order accurate methods are given below. The Gazdag
method, which we discuss in Chapter 8, is
1
ũn+1 = un + h 3ũn − ũn−1
2
1
un+1 = un + h ũn + ũn+1 (6.65)
2
The Burstein method, obtained from Eq. 6.63 with α = 1/2 is
1
ũn+1/2 = un + hun
2
un+1 = un + hũn+1/2 (6.66)
and, finally, MacCormack’s method, presented earlier in this chapter, is
ũn+1 = un + hun
1
un+1 = [un + ũn+1 + hũn+1 ] (6.67)
2
Note that MacCormack’s method can also be written as
ũn+1 = un + hun
1
un+1 = un + h[un + ũn+1 ] (6.68)
2
from which it is clear that it is obtained from Eq. 6.63 with α = 1.
6.9 Runge-Kutta Methods

There is a special subset of predictor-corrector methods, referred to as Runge-Kutta
methods,6 that produce just one σ-root for each λ-root such that σ(λh) corresponds
6
Although implicit and multi-step Runge-Kutta methods exist, we will consider only single-step,
explicit Runge-Kutta methods here.
to the Taylor series expansion of eλh out through the order of the method and then
truncates. Thus for a Runge-Kutta method of order k (up to 4th order), the principal
(and only) σ-root is given by
1 1
σ = 1 + λh + λ2 h2 + · · · + λk hk (6.69)
2 k!
It is not particularly difficult to build this property into a method, but, as we pointed
out in Section 6.6.4, it is not sufficient to guarantee k’th order accuracy for the solution
of u = F (u, t) or for the representative equation. To ensure k’th order accuracy, the
method must further satisfy the constraint that
erµ = O(hk+1 ) (6.70)
and this is much more difficult.
The most widely publicized Runge-Kutta process is the one that leads to the
fourth-order method. We present it below in some detail. It is usually introduced in
the form
k1 = hF (un , tn )
k2 = hF (un + βk1 , tn + αh)
k3 = hF (un + β1 k1 + γ1 k2 , tn + α1 h)
k4 = hF (un + β2 k1 + γ2 k2 + δ2 k3 , tn + α2 h)
followed by
u(tn + h) − u(tn ) = µ1 k1 + µ2 k2 + µ3 k3 + µ4 k4 (6.71)
However, we prefer to present it using predictor-corrector notation. Thus, a scheme
entirely equivalent to 6.71 is
u#n+α = un + βhun
ũn+α1 = un + β1 hun + γ1 hu#n+α
un+α2 = un + β2 hun + γ2 hu#n+α + δ2 hũn+α1
un+1 = un + µ1 hun + µ2 hu#n+α + µ3 hũn+α1 + µ4 hun+α2 (6.72)
Appearing in Eqs. 6.71 and 6.72 are a total of 13 parameters which are to be
determined such that the method is fourth-order according to the requirements in
Eqs. 6.69 and 6.70. First of all, the choices for the time samplings, α, α1 , and α2 , are
not arbitrary. They must satisfy the relations
α = β
α1 = β1 + γ1
α2 = β2 + γ2 + δ2 (6.73)
6.9. RUNGE-KUTTA METHODS 109
The algebra involved in finding algebraic equations for the remaining 10 parameters
is not trivial, but the equations follow directly from finding P (E) and Q(E) and then
satisfying the conditions in Eqs. 6.69 and 6.70. Using Eq. 6.73 to eliminate the β’s
we find from Eq. 6.69 the four conditions
µ1 + µ2 + µ3 + µ4 = 1 (1)
µ2 α + µ3 α1 + µ4 α2 = 1/2 (2)
(6.74)
µ3 αγ1 + µ4 (αγ2 + α1 δ2 ) = 1/6 (3)
µ4 αγ1 δ2 = 1/24 (4)
These four relations guarantee that the five terms in σ exactly match the first 5 terms
in the expansion of eλh . To satisfy the condition that erµ = O(k 5 ), we have to fulfill
four more conditions
µ2 α2 + µ3 α12 + µ4 α22 = 1/3 (3)
µ2 α3 + µ3 α13 + µ4 α23 = 1/4 (4)
(6.75)
µ3 α2 γ1 + µ4 (α2 γ2 + α12 δ2 ) = 1/12 (4)
µ3 αα1 γ1 + µ4 α2 (αγ2 + α1 δ2 ) = 1/8 (4)
The number in parentheses at the end of each equation indicates the order that
is the basis for the equation. Thus if the first 3 equations in 6.74 and the first
equation in 6.75 are all satisfied, the resulting method would be third-order accurate.
As discussed in Section 6.6.4, the fourth condition in Eq. 6.75 cannot be derived
using the methodology presented here, which is based on a linear nonhomogenous
representative ODE. A more general derivation based on a nonlinear ODE can be
found in several books.7
There are eight equations in 6.74 and 6.75 which must be satisfied by the 10
unknowns. Since the equations are overdetermined, two parameters can be set arbi-
trarily. Several choices for the parameters have been proposed, but the most popular
one is due to Runge. It results in the “standard” fourth-order Runge-Kutta method
expressed in predictor-corrector form as
1
u#n+1/2 = un + hun
2
1
ũn+1/2 = un + hu#n+1/2
2
un+1 = un + hũn+1/2
1
un+1 = un + h un + 2 u#n+1/2 + ũn+1/2 + un+1 (6.76)
6
7
The present approach based on a linear inhomogeneous equation provides all of the necessary
conditions for Runge-Kutta methods of up to third order.
Notice that this represents the simple sequence of conventional linear multistep meth-
ods referred to, respectively, as

Euler Predictor 



Euler Corrector

≡ RK4
Leapfrog Predictor 


Milne Corrector
One can easily show that both the Burstein and the MacCormack methods given by
Eqs. 6.66 and 6.67 are second-order Runge-Kutta methods, and third-order methods
can be derived from Eqs. 6.72 by setting µ4 = 0 and satisfying only Eqs. 6.74 and the
first equation in 6.75. It is clear that for orders one through four, RK methods of order
k require k evaluations of the derivative function to advance the solution one time
step. We shall discuss the consequences of this in Chapter 8. Higher-order Runge-
Kutta methods can be developed, but they require more derivative evaluations than
their order. For example, a fifth-order method requires six evaluations to advance
the solution one step. In any event, storage requirements reduce the usefulness of
Runge-Kutta methods of order higher than four for CFD applications.
6.10 Implementation of Implicit Methods

We have presented a wide variety of time-marching methods and shown how to derive
their λ − σ relations. In the next chapter, we will see that these methods can have
widely different properties with respect to stability. This leads to various trade-
offs which must be considered in selecting a method for a specific application. Our
presentation of the time-marching methods in the context of a linear scalar equation
obscures some of the issues involved in implementing an implicit method for systems
of equations and nonlinear equations. These are covered in this Section.
6.10.1 Application to Systems of Equations

Consider first the numerical solution of our representative ODE
u = λu + aeµt (6.77)
using the implicit Euler method. Following the steps outlined in Section 6.2, we
obtained
(1 − λh)un+1 − un = heµh · aeµhn (6.78)

6.10. IMPLEMENTATION OF IMPLICIT METHODS 111
Solving for un+1 gives

1
un+1 = (un + heµh · aeµhn ) (6.79)
1 − λh
This calculation does not seem particularly onerous in comparison with the applica-
tion of an explicit method to this ODE, requiring only an additional division.
Now let us apply the implicit Euler method to our generic system of equations
given by
u = Au − f(t) (6.80)
where u and f are vectors and we still assume that A is not a function of u or t. Now
the equivalent to Eq. 6.78 is
(I − hA)un+1 − un = −hf(t + h) (6.81)
or
un+1 = (I − hA)−1 [un − hf(t + h)] (6.82)
The inverse is not actually performed, but rather we solve Eq. 6.81 as a linear system
of equations. For our one-dimensional examples, the system of equations which must
be solved is tridiagonal (e.g., for biconvection, A = −aBp (−1, 0, 1)/2∆x), and hence
its solution is inexpensive, but in multidimensions the bandwidth can be very large. In
general, the cost per time step of an implicit method is larger than that of an explicit
method. The primary area of application of implicit methods is in the solution of
stiff ODE’s, as we shall see in Chapter 8.
6.10.2 Application to Nonlinear Equations

Now consider the general nonlinear scalar ODE given by
du
= F (u, t) (6.83)
dt
Application of the implicit Euler method gives
un+1 = un + hF (un+1, tn+1 ) (6.84)
This is a nonlinear difference equation. As an example, consider the nonlinear ODE

du 1 2
+ u =0 (6.85)
dt 2
solved using implicit Euler time marching, which gives

1
un+1 + h u2n+1 = un (6.86)
2
which requires a nontrivial method to solve for un+1 . There are several different
approaches one can take to solving this nonlinear difference equation. An iterative
method, such as Newton’s method (see below), can be used. In practice, the “initial
guess” for this nonlinear problem can be quite close to the solution, since the “initial
guess” is simply the solution at the previous time step, which implies that a lineariza-
tion approach may be quite successful. Such an approach is described in the next
Section.
6.10.3 Local Linearization for Scalar Equations

General Development
Let us start the process of local linearization by considering Eq. 6.83. In order to
implement the linearization, we expand F (u, t) about some reference point in time.
Designate the reference value by tn and the corresponding value of the dependent
variable by un . A Taylor series expansion about these reference quantities gives

∂F ∂F
F (u, t) = F (un , tn ) + (u − un ) + (t − tn )
∂u n
∂t n

1 ∂2 F ∂2F
+ (u − un ) +
2
(u − un )(t − tn )
2 ∂u2 n
∂u∂t n

1 ∂2 F
+ (t − tn )2 + · · · (6.87)
2 ∂t2 n
On the other hand, the expansion of u(t) in terms of the independent variable t is

∂u 1 ∂2 u
u(t) = un + (t − tn ) + (t − tn )2 +··· (6.88)
∂t n
2 ∂t2 n
If t is within h of tn , both (t − tn )k and (u − un)k are O(hk ), and Eq. 6.87 can be
written

∂F ∂F
F (u, t) = Fn + (u − un) + (t − tn ) + O(h2 ) (6.89)
∂u n
∂t n
Notice that this is an expansion of the derivative of the function. Thus, relative to the
order of expansion of the function, it represents a second-order-accurate, locally-linear
approximation to F (u, t) that is valid in the vicinity of the reference station tn and
the corresponding un = u(tn ). With this we obtain the locally (in the neighborhood
of tn ) time-linear representation of Eq. 6.83, namely

du ∂F ∂F ∂F
= u + Fn − un + (t − tn ) + O(h2 ) (6.90)
dt ∂u n
∂u n
∂t n
Implementation of the Trapezoidal Method

As an example of how such an expansion can be used, consider the mechanics of
applying the trapezoidal method for the time integration of Eq. 6.83. The trapezoidal
method is given by
1
un+1 = un + h[Fn+1 + Fn ] + hO(h2 ) (6.91)
2
where we write hO(h2 ) to emphasize that the method is second order accurate. Using
Eq. 6.89 to evaluate Fn+1 = F (un+1 , tn+1 ), one finds

1 ∂F ∂F
un+1 = un + h Fn + (un+1 − un ) + h 2
+ O(h ) + Fn
2 ∂u n
∂t n
+hO(h2 ) (6.92)
Note that the O(h2 ) term within the brackets (which is due to the local linearization)
is multiplied by h and therefore is the same order as the hO(h2 ) error from the
Trapezoidal Method. The use of local time linearization updated at the end of each
time step, and the trapezoidal time march, combine to make a second-order-accurate
numerical integration process. There are, of course, other second-order implicit time-
marching methods that can be used. The important point to be made here is that
local linearization updated at each time step has not reduced the order of accuracy of
a second-order time-marching process.
A very useful reordering of the terms in Eq. 6.92 results in the expression

1 ∂F 1 ∂F
1− h ∆un = hFn + h2 (6.93)
2 ∂u n
2 ∂t n
which is now in the delta form which will be formally introduced in Section 12.6. In
many fluid mechanic applications the nonlinear function F is not an explicit function
of t. In such cases the partial derivative of F (u) with respect to t is zero and Eq. 6.93
simplifies to the second-order accurate expression

1 ∂F
1− h ∆un = hFn (6.94)
2 ∂u n
Notice that the RHS is extremely simple. It is the product of h and the RHS of
the basic equation evaluated at the previous time step. In this example, the basic
equation was the simple scalar equation 6.83, but for our applications, it is generally
the space-differenced form of the steady-state equation of some fluid flow problem.
A numerical time-marching procedure using Eq. 6.94 is usually implemented as
follows:
and save un.
1. Solve for the elements of hFn , store them in an array say R,
2. Solve for the elements of the matrix multiplying ∆un and store in some appro-
priate manner making use of sparseness or bandedness of the matrix if possible.
Let this storage area be referred to as B.
3. Solve the coupled set of linear equations

B∆un = R
for ∆un . (Very seldom does one find B−1 in carrying out this step).
4. Find un+1 by adding ∆un to un , thus
un+1 = ∆un + un
The solution for un+1 is generally stored such that it overwrites the value of un
and the process is repeated.
Implementation of the Implicit Euler Method

We have seen that the first-order implicit Euler method can be written
un+1 = un + hFn+1 (6.95)
if we introduce Eq. 6.90 into this method, rearrange terms, and remove the explicit
dependence on time, we arrive at the form

∂F
1−h ∆un = hFn (6.96)
∂u n
We see that the only difference between the implementation of the trapezoidal method
and the implicit Euler method is the factor of 21 in the brackets of the left side of
Eqs. 6.94 and 6.96. Omission of this factor degrades the method in time accuracy by
one order of h. We shall see later that this method is an excellent choice for steady
problems.
Newton’s Method
Consider the limit h → ∞ of Eq. 6.96 obtained by dividing both sides by h and
setting 1/h = 0. There results

∂F
− ∆un = Fn (6.97)
∂u n
or
−1
∂F
un+1 = un − Fn (6.98)
∂u n
This is the well-known Newton method for finding the roots of a nonlinear equation
F (u) = 0. The fact that it has quadratic convergence is verified by a glance at Eqs.
6.87 and 6.88 (remember the dependence on t has been eliminated for this case). By
quadratic convergence, we mean that the error after a given iteration is proportional
to the square of the error at the previous iteration, where the error is the difference
between the current solution and the converged solution. Quadratic convergence is
thus a very powerful property. Use of a finite value of h in Eq. 6.96 leads to linear
convergence, i.e., the error at a given iteration is some multiple of the error at the
previous iteration. The reader should ponder the meaning of letting h → ∞ for the
trapezoidal method, given by Eq. 6.94.
6.10.4 Local Linearization for Coupled Sets of Nonlinear Equa-

tions
In order to present this concept, let us consider an example involving some sim-
ple boundary-layer equations. We choose the Falkner-Skan equations from classical
boundary-layer theory. Our task is to apply the implicit trapezoidal method to the
equations
 2 
d3 f d2 f df
3
+ f 2
+ β 1 −  =0 (6.99)
dt dt dt
Here f represents a dimensionless stream function, and β is a scaling factor.

First of all we reduce Eq. 6.99 to a set of first-order nonlinear equations by the
transformations
d2 f df
2
u1 =
, u2 = , u3 = f (6.100)
dt dt
This gives the coupled set of three nonlinear equations

u1 = F1 = −u1 u3 − β 1 − u22
u2 = F2 = u1
u3 = F3 = u2 (6.101)
and these can be represented in vector notation as
du
= F (u) (6.102)
dt
Now we seek to make the same local expansion that derived Eq. 6.90, except that
this time we are faced with a nonlinear vector function, rather than a simple nonlinear
scalar function. The required extension requires the evaluation of a matrix, called
the Jacobian matrix.8 Let us refer to this matrix as A. It is derived from Eq. 6.102
by the following process
A = (aij ) = ∂Fi /∂uj (6.103)

For the general case involving a third order matrix this is
 
∂F1 ∂F1 ∂F1
 ∂u1 ∂u2 ∂u3 
 
 
 
 ∂F2 ∂F2 ∂F2 
A=   (6.104)
 ∂u1 ∂u2 ∂u3 
 
 
 
∂F3 ∂F3 ∂F3
∂u1 ∂u2 ∂u3
The expansion of F (u) about some reference state u can be expressed in a way
n
similar to the scalar expansion given by eq 6.87. Omitting the explicit dependency
on the independent variable t, and defining F as F (u ), one has 9
n n
8
Recall that we derived the Jacobian matrices for the two-dimensional Euler equations in Section
2.2
9
The Taylor series expansion of a vector contains a vector for the first term, a matrix times a
vector for the second term, and tensor products for the terms of higher order.
6.11. PROBLEMS 117

(u) = F
F + A u − u + O(h2 ) (6.105)
n n n
where t − tn and the argument for O(h2 ) is the same as in the derivation of Eq. 6.88.
Using this we can write the local linearization of Eq. 6.102 as

du − A u +O(h2 )
= Anu + F n n n (6.106)
dt
“constant”
which is a locally-linear, second-order-accurate approximation to a set of coupled
nonlinear ordinary differential equations that is valid for t ≤ tn + h. Any first- or
second-order time-marching method, explicit or implicit, could be used to integrate
the equations without loss in accuracy with respect to order. The number of times,
and the manner in which, the terms in the Jacobian matrix are updated as the solution
proceeds depends, of course, on the nature of the problem.
Returning to our simple boundary-layer example, which is given by Eq. 6.101, we
find the Jacobian matrix to be
 
− u3 2βu2 −u1

A= 1 0 0  (6.107)
0 1 0
The student should be able to derive results for this example that are equivalent to
those given for the scalar case in Eq. 6.93. Thus for the Falkner-Skan equations the
trapezoidal method results in
 
   

1+ h h
2 (u3 )n −βh(u2 )n 2 (u1)n  (∆u1 )n − (u1 u3 )n − β(1 − u22 )n
    


−h 2 1 0  (∆u2 )n  = h 

(u1 )n 
0 −h 1 (∆u3 )n (u2 )n
2
We find un+1 from ∆un + un , and the solution is now advanced one step. Re-evaluate
the elements using un+1 and continue. Without any iterating within a step advance,
the solution will be second-order-accurate in time.
6.11 Problems
1. Find an expression for the nth term in the Fibonacci series, which is given
by 1, 1, 2, 3, 5, 8, . . . Note that the series can be expressed as the solution to a
difference equation of the form un+1 = un + un−1 . What is u25 ? (Let the first
term given above be u0 .)
2. The trapezoidal method un+1 = un + 12 h(un+1 + un ) is used to solve the repre-
sentative ODE.
(a) What is the resulting O∆E?

(b) What is its exact solution?
(c) How does the exact steady-state solution of the O∆E compare with the
exact steady-state solution of the ODE if µ = 0?
3. The 2nd-order backward method is given by

1
un+1 = 4un − un−1 + 2hun+1
3
(a) Write the O∆E for the representative equation. Identify the polynomials
P (E) and Q(E).
(b) Derive the λ-σ relation. Solve for the σ-roots and identify them as principal
or spurious.
(c) Find erλ and the first two nonvanishing terms in a Taylor series expansion
of the spurious root.
(d) Perform a σ-root trace relative to the unit circle for both diffusion and
convection.
4. Consider the time-marching scheme given by

2h
un+1 = un−1 + (un+1 + un + un−1 )
3
(a) Write the O∆E for the representative equation. Identify the polynomials
P (E) and Q(E).
(b) Derive the λ − σ relation.
(c) Find erλ .
5. Find the difference equation which results from applying the Gazdag predictor-
corrector method (Eq. 6.65) to the representative equation. Find the λ-σ rela-
tion.
6. Consider the following time-marching method:
ũn+1/3 = un + hun /3
ūn+1/2 = un + hũn+1/3 /2
un+1 = un + hūn+1/2
6.11. PROBLEMS 119
Find the difference equation which results from applying this method to the
representative equation. Find the λ-σ relation. Find the solution to the differ-
ence equation, including the homogeneous and particular solutions. Find erλ
and erµ . What order is the homogeneous solution? What order is the particular
solution? Find the particular solution if the forcing term is fixed.
7. Write a computer program to solve the one-dimensional linear convection equa-

tion with periodic boundary conditions and a = 1 on the domain 0 ≤ x ≤ 1.
Use 2nd-order centered differences in space and a grid of 50 points. For the
initial condition, use
u(x, 0) = e−0.5[(x−0.5)/σ]
2
with σ = 0.08. Use the explicit Euler, 2nd-order Adams-Bashforth (AB2), im-
plicit Euler, trapezoidal, and 4th-order Runge-Kutta methods. For the explicit
Euler and AB2 methods, use a Courant number, ah/∆x, of 0.1; for the other
methods, use a Courant number of unity. Plot the solutions obtained at t = 1
compared to the exact solution (which is identical to the initial condition).
8. Repeat problem 7 using 4th-order (noncompact) differences in space. Use only

4th-order Runge-Kutta time marching at a Courant number of unity. Show
solutions at t = 1 and t = 10 compared to the exact solution.
9. Using the computer program written for problem 7, compute the solution at
t = 1 using 2nd-order centered differences in space coupled with the 4th-order
Runge-Kutta method for grids of 100, 200, and 400 nodes. On a log-log scale,
plot the error given by
(
)
) M
(uj − uexact )2
) j
*
j=1 M
where M is the number of grid nodes and uexact is the exact solution. Find the
global order of accuracy from the plot.
10. Using the computer program written for problem 8, repeat problem 9 using
4th-order (noncompact) differences in space.
11. Write a computer program to solve the one-dimensional linear convection equa-
tion with inflow-outflow boundary conditions and a = 1 on the domain 0 ≤
x ≤ 1. Let u(0, t) = sin ωt with ω = 10π. Run until a periodic steady state
is reached which is independent of the initial condition and plot your solution
compared with the exact solution. Use 2nd-order centered differences in space
with a 1st-order backward difference at the outflow boundary (as in Eq. 3.69)
together with 4th-order Runge-Kutta time marching. Use grids with 100, 200,
and 400 nodes and plot the error vs. the number of grid nodes, as described in
problem 9. Find the global order of accuracy.
12. Repeat problem 11 using 4th-order (noncompact) centered differences. Use a

third-order forward-biased operator at the inflow boundary (as in Eq. 3.67). At
the last grid node, derive and use a 3rd-order backward operator (using nodes
j − 3, j − 2, j − 1, and j) and at the second last node, use a 3rd-order backward-
biased operator (using nodes j − 2, j − 1, j, and j + 1; see problem 1 in Chapter
3).
13. Using the approach described in Section 6.6.2, find the phase speed error, erp ,
and the amplitude error, era , for the combination of second-order centered dif-
ferences and 1st, 2nd, 3rd, and 4th-order Runge-Kutta time-marching at a
Courant number of unity. Also plot the phase speed error obtained using exact
integration in time, i.e., that obtained using the spatial discretization alone.
Note that the required σ-roots for the various Runge-Kutta methods can be
deduced from Eq. 6.69, without actually deriving the methods. Explain your
results.
Chapter 7
STABILITY OF LINEAR
SYSTEMS
A general definition of stability is neither simple nor universal and depends on the
particular phenomenon being considered. In the case of the nonlinear ODE’s of
interest in fluid dynamics, stability is often discussed in terms of fixed points and
attractors. In these terms a system is said to be stable in a certain domain if, from
within that domain, some norm of its solution is always attracted to the same fixed
point. These are important and interesting concepts, but we do not dwell on them in
this work. Our basic concern is with time-dependent ODE’s and O∆E ’s in which the
coefficient matrices are independent of both u and t; see Section 4.2. We will refer
to such matrices as stationary. Chapters 4 and 6 developed the representative forms
of ODE’s generated from the basic PDE’s by the semidiscrete approach, and then
the O∆E’s generated from the representative ODE’s by application of time-marching
methods. These equations are represented by
du
= Au − f (t) (7.1)
dt
and
un+1 = Cun − g n (7.2)
respectively. For a one-step method, the latter form is obtained by applying a time-
marching method to the generic ODE form in a fairly straightforward manner. For

example, the explicit Euler method leads to C = I + hA, and gn = f(nh). Methods
involving two or more steps can always be written in the form of Eq. 7.2 by introducing
new dependent variables. Note also that only methods in which the time and space
discretizations are treated separately can be written in an intermediate semi-discrete
form such as Eq. 7.1. The fully-discrete form, Eq. 7.2 and the associated stability
definitions and analysis are applicable to all methods.
121
122 CHAPTER 7. STABILITY OF LINEAR SYSTEMS
7.1 Dependence on the Eigensystem
Our definitions of stability are based entirely on the behavior of the homogeneous
parts of Eqs. 7.1 and 7.2. The stability of Eq. 7.1 depends entirely on the eigensys-
tem1 of A. The stability of Eq. 7.2 can often also be related to the eigensystem of
its matrix. However, in this case the situation is not quite so simple since, in our
applications to partial differential equations (especially hyperbolic ones), a stability
definition can depend on both the time and space differencing. This is discussed in
Section 7.4. Analysis of these eigensystems has the important added advantage that
it gives an estimate of the rate at which a solution approaches a steady-state if a
system is stable. Consideration will be given to matrices that have both complete
and defective eigensystems, see Section 4.2.3, with a reminder that a complete system
can be arbitrarily close to a defective one, in which case practical applications can
make the properties of the latter appear to dominate.
If A and C are stationary, we can, in theory at least, estimate their fundamental
properties. For example, in Section 4.3.2, we found from our model ODE’s for dif-
fusion and periodic convection what could be expected for the eigenvalue spectrums
of practical physical problems containing these phenomena. These expectations are
referred to many times in the following analysis of stability properties. They are
important enough to be summarized by the following:
• For diffusion dominated flows the λ-eigenvalues tend to lie along the negative
real axis.
• For periodic convection-dominated flows the λ-eigenvalues tend to lie along the
imaginary axis.
In many interesting cases, the eigenvalues of the matrices in Eqs. 7.1 and 7.2
are sufficient to determine the stability. In previous chapters, we designated these
eigenvalues as λm and σm for Eqs. 7.1 and 7.2, respectively, and we will find it
convenient to examine the stability of various methods in both the complex λ and
complex σ planes.
1
This is not the case if the coefficient matrix depends on t even if it is linear.
7.2. INHERENT STABILITY OF ODE’S 123
7.2 Inherent Stability of ODE’s

7.2.1 The Criterion
Here we state the standard stability criterion used for ordinary differential equa-
tions.
For a stationary matrix A, Eq. 7.1 is inherently stable if, when f is (7.3)
constant, u remains bounded as t → ∞.
Note that inherent stability depends only on the transient solution of the ODE’s.
7.2.2 Complete Eigensystems

If a matrix has a complete eigensystem, all of its eigenvectors are linearly independent,
and the matrix can be diagonalized by a similarity transformation. In such a case it
follows at once from Eq. 6.27, for example, that the ODE’s are inherently stable if
and only if
(λm ) ≤ 0 for all m (7.4)
This states that, for inherent stability, all of the λ eigenvalues must lie on, or to the
left of, the imaginary axis in the complex λ plane. This criterion is satisfied for the
model ODE’s representing both diffusion and biconvection. It should be emphasized
(as it is an important practical consideration in convection-dominated systems) that
the special case for which λ = ±i is included in the domain of stability. In this case
it is true that u does not decay as t → ∞, but neither does it grow, so the above
condition is met. Finally we note that for ODE’s with complete eigensystems the
eigenvectors play no role in the inherent stability criterion.
7.2.3 Defective Eigensystems

In order to understand the stability of ODE’s that have defective eigensystems, we
inspect the nature of their solutions in eigenspace. For this we draw on the results in
Sections 4.2.3 and especially on Eqs. 4.18 to 4.19 in that section. In an eigenspace
related to defective systems the form of the representative equation changes from a
single equation to a Jordan block. For example, instead of Eq. 4.45 a typical form of
the homogeneous part might be
    
u1 λ u1
    
 u2  =  1 λ   u2 
u3 1 λ u3
for which one finds the solution
u1 (t) = u1 (0)eλt
u2 (t) = [u2 (0) + u1 (0)t]eλt

1
u3 (t) = u3 (0) + u2 (0)t + u1 (0)t2 eλt (7.5)
2
Inspecting this solution, we see that for such cases condition 7.4 must be modified
to the form
(λm ) < 0 for all m (7.6)
since for pure imaginary λ, u2 and u3 would grow without bound (linearly or quadrat-
ically) if u2 (0) = 0 or u1(0) = 0. Theoretically this condition is sufficient for stability
in the sense of Statement 7.3 since tk e−||t → 0 as t → ∞ for all non-zero . However,
in practical applications the criterion may be worthless since there may be a very
large growth of the polynomial before the exponential “takes over” and brings about
the decay. Furthermore, on a computer such a growth might destroy the solution
process before it could be terminated.
Note that the stability condition 7.6 excludes the imaginary axis which tends to be
occupied by the eigenvalues related to biconvection problems. However, condition 7.6
is of little or no practical importance if significant amounts of dissipation are present.
7.3 Numerical Stability of O∆E ’s

7.3.1 The Criterion
The O∆E companion to Statement 7.3 is
For a stationary matrix C, Eq. 7.2 is numerically stable if, when g is (7.7)
constant, un remains bounded as n → ∞.
We see that numerical stability depends only on the transient solution of the O∆E ’s.
This definition of stability is sometimes referred to as asymptotic or time stability.
As we stated at the beginning of this chapter, stability definitions are not unique. A
definition often used in CFD literature stems from the development of PDE solutions
that do not necessarily follow the semidiscrete route. In such cases it is appropriate
to consider simultaneously the effects of both the time and space approximations. A
time-space domain is fixed and stability is defined in terms of what happens to some
norm of the solution within this domain as the mesh intervals go to zero at some
constant ratio. We discuss this point of view in Section 7.4.
7.4. TIME-SPACE STABILITY AND CONVERGENCE OF O∆E’S 125
7.3.2 Complete Eigensystems

Consider a set of O∆E ’s governed by a complete eigensystem. The stability criterion,
according to the condition set in Eq. 7.7, follows at once from a study of Eq. 6.28
and its companion for multiple σ-roots, Eq. 6.36. Clearly, for such systems a time-
marching method is numerically stable if and only if
|(σm )k | ≤ 1 for all m and k (7.8)
This condition states that, for numerical stability, all of the σ eigenvalues (both
principal and spurious, if there are any) must lie on or inside the unit circle in the
complex σ-plane.
This definition of stability for O∆E ’s is consistent with the stability definition for
ODE’s. Again the sensitive case occurs for the periodic-convection model which places
the “correct” location of the principal σ–root precisely on the unit circle where the
solution is only neutrally stable. Further, for a complete eigensystem, the eigenvectors
play no role in the numerical stability assessment.
7.3.3 Defective Eigensystems

The discussion for these systems parallels the discussion for defective ODE’s. Examine
Eq. 6.18 and note its similarity with Eq. 7.5. We see that for defective O∆E’s the
required modification to 7.8 is
|(σm )k | < 1 for all m and k (7.9)
since defective systems do not guarantee boundedness for |σ| = 1, for example in Eq.
7.5 if |σ| = 1 and either u2(0) or u1 (0) = 0 we get linear or quadratic growth.
7.4 Time-Space Stability and Convergence of O∆E’s

Let us now examine the concept of stability in a different way. In the previous
discussion we considered in some detail the following approach:
1. The PDE’s are converted to ODE’s by approximating the space derivatives on

a finite mesh.
2. Inherent stability of the ODE’s is established by guaranteeing that (λ) ≤ 0.
3. Time-march methods are developed which guarantee that |σ(λh)| ≤ 1 and this
is taken to be the condition for numerical stability.
This does guarantee that a stationary system, generated from a PDE on some fixed
space mesh, will have a numerical solution that is bounded as t = nh → ∞. This
does not guarantee that desirable solutions are generated in the time march process
as both the time and space mesh intervals approach zero.
Now let us define stability in the time-space sense. First construct a finite time-
space domain lying within 0 ≤ x ≤ L and 0 ≤ t ≤ T . Cover this domain with a grid
that is equispaced in both time and space and fix the mesh ratio by the equation2
∆t
cn =
∆x
Next reduce our O∆E approximation of the PDE to a two-level (i.e., two time-planes)
formula in the form of Eq. 7.2. The homogeneous part of this formula is
un+1 = Cun (7.10)
Eq. 7.10 is said to be stable if any bounded initial vector, u0 , produces a bounded
solution vector, un , as the mesh shrinks to zero for a fixed cn . This is the classical
definition of stability. It is often referred to as Lax or Lax-Richtmyer stability. Clearly
as the mesh intervals go to zero, the number of time steps, N , must go to infinity in
order to cover the entire fixed domain, so the criterion in 7.7 is a necessary condition
for this stability criterion.
The significance of this definition of stability arises through Lax’s Theorem, which
states that, if a numerical method is stable (in the sense of Lax) and consistent then
it is convergent. A method is consistent if it produces no error (in the Taylor series
sense) in the limit as the mesh spacing and the time step go to zero (with cn fixed, in
the hyperbolic case). This is further discussed in Section 7.8. A method is convergent
if it converges to the exact solution as the mesh spacing and time step go to zero in
this manner.3 Clearly, this is an important property.
Applying simple recursion to Eq. 7.10, we find
un = Cnu0
and using vector and matrix p-norms (see Appendix A) and their inequality relations,
we have
||un|| = ||Cnu0 || ≤ ||Cn || · ||u0|| ≤ ||C||n · ||u0|| (7.11)

2
This ratio is appropriate for hyperbolic problems; a different ratio may be needed for parabolic
problems such as the model diffusion equation.
3
In the CFD literature, the word converge is used with two entirely different meanings. Here we
refer to a numerical solution converging to the exact solution of the PDE. Later we will refer to the
numerical solution converging to a steady state solution.
7.4. TIME-SPACE STABILITY AND CONVERGENCE OF O∆E’S 127
Since the initial data vector is bounded, the solution vector is bounded if
||C|| ≤ 1 (7.12)
where ||C|| represents any p-norm of C. This is often used as a sufficient condition
for stability.
Now we need to relate the stability definitions given in Eqs. 7.8 and 7.9 with that
given in Eq. 7.12. In Eqs. 7.8 and 7.9, stability is related to the spectral radius of
C, i.e., its eigenvalue of maximum magnitude. In Eq. 7.12, stability is related to a
p-norm of C. It is clear that the criteria are the same when the spectral radius is a
true p-norm.
Two facts about the relation between spectral radii and matrix norms are well
known:
1. The spectral radius of a matrix is its L2 norm when the matrix is normal, i.e.,
it commutes with its transpose.
2. The spectral radius is the lower bound of all norms.
Furthermore, when C is normal, the second inequality in Eq. 7.11 becomes an equal-
ity. In this case, Eq. 7.12 becomes both necessary and sufficient for stability. From
these relations we draw two important conclusions about the numerical stability of
methods used to solve PDE’s.
• The stability criteria in Eqs. 7.8 and 7.12 are identical for stationary systems
when the governing matrix is normal. This includes symmetric, asymmetric,
and circulant matrices. These criteria are both necessary and sufficient for
methods that generate such matrices and depend solely upon the eigenvalues of
the matrices.
• If the spectral radius of any governing matrix is greater than one, the method
is unstable by any criterion. Thus for general matrices, the spectral radius
condition is necessary4 but not sufficient for stability.
4
Actually the necessary condition is that the spectral radius of C be less than or equal to 1 +
O(∆t), but this distinction is not critical for our purposes here.
7.5 Numerical Stability Concepts in the Complex

σ-Plane
7.5.1 σ-Root Traces Relative to the Unit Circle
Whether or not the semi-discrete approach was taken to find the differencing approx-
imation of a set of PDE’s, the final difference equations can be represented by
un+1 = Cun − gn
Furthermore if C has a complete5 eigensystem, the solution to the homogeneous part

can always be expressed as
un = c1 σ1n x1 + · · · + cm σm

n
xm + · · · + cM σM
n
xM
where the σm are the eigenvalues of C. If the semi-discrete approach is used, we can
find a relation between the σ and the λ eigenvalues. This serves as a very convenient
guide as to where we might expect the σ-roots to lie relative to the unit circle in the
complex σ-plane. For this reason we will proceed to trace the locus of the σ-roots
as a function of the parameter λh for the equations modeling diffusion and periodic
convection6 .
Locus of the exact trace
Figure 7.1 shows the exact trace of the σ-root if it is generated by eλh representing
either diffusion or biconvection. In both cases the • represents the starting value
where h = 0 and σ = 1. For the diffusion model, λh is real and negative. As the
magnitude of λh increases, the trace representing the dissipation model heads towards
the origin as λh =→ −∞. On the other hand, for the biconvection model, λh = iωh
is always imaginary. As the magnitude of ωh increases, the trace representing the
biconvection model travels around the circumference of the unit circle, which it never
leaves. We must be careful in interpreting σ when it is representing eiωh . The fact
that it lies on the unit circle means only that the amplitude of the representation is
correct, it tells us nothing of the phase error (see Eq. 6.41). The phase error relates
to the position on the unit circle.
5
The subject of defective eigensystems has been addressed. From now on we will omit further
discussion of this special case.
6
Or, if you like, the parameter h for fixed values of λ equal to -1 and i for the diffusion and
biconvection cases, respectively.
7.5. NUMERICAL STABILITY CONCEPTS IN THE COMPLEX σ-PLANE 129
Examples of some methods
Now let us compare the exact σ-root traces with some that are produced by actual
time-marching methods. Table 7.1 shows the λ-σ relations for a variety of methods.
Figures 7.2 and 7.3 illustrate the results produced by various methods when they are
applied to the model ODE’s for diffusion and periodic-convection, Eqs. 4.4 and 4.5.
It is implied that the behavior shown is typical of what will happen if the methods
are applied to diffusion- (or dissipation-) dominated or periodic convection-dominated
problems as well as what does happen in the model cases. Most of the important
possibilities are covered by the illustrations.
1 σ − 1 − λh = 0 Explicit Euler
2 σ 2 − 2λhσ − 1 = 0 Leapfrog
3 2 3 1
σ − (1 + 2 λh)σ + 2 λh = 0 AB2
4 σ 3 − (1 + 23 16 5
12 λh)σ + 12 λhσ − 12 λh = 0
2
AB3
5 σ 1 − λh) − 1 = 0
(
Implicit Euler
6 1 1
σ(1 − 2 λh) − (1 + 2 λh) = 0 Trapezoidal
7 σ 2 (1 − 23 λh) − 43 σ + 31 = 0 2nd O Backward
8 5 8 1
σ (1 − 12 λh) − (1 + 12 λh)σ + 12 λh = 0
2
AM3
9 σ 2 − (1 + 13 15 2 2 1 5
12 λh + 24 λ h )σ + 12 λh(1 + 2 λh) = 0 ABM3
10 σ 3 − (1 + 2λh)σ2 + 23 λhσ − 12 λh = 0 Gazdag
11 1
σ − 1 − λh − 2 λ h = 0
2 2
RK2
12 1 2 2 1 3 3 1
σ − 1 − λh − 2 λ h − 6 λ h − 24 λ h = 0 4 4
RK4
13 σ 2 (1 − 13 λh) − 43 λhσ − (1 + 13 λh) = 0 Milne 4th
Table 7.1. Some λ − σ Relations
a. Explicit Euler Method
Figure 7.2 shows results for the explicit Euler method. When used for dissipation-
dominated cases it is stable for the range -2≤ λh ≤0. (Usually the magnitude of λ has
to be estimated and often it is found by trial and error). When used for biconvection
the σ- trace falls outside the unit circle for all finite h, and the method has no range
of stability in this case.
Ι(σ) Ι(σ)
λ h= 0 ωh= 0
R (σ) R (σ)
λ h= οο
i ωh
σ = e λ h , λh - oo σ=e , ωh oo
a) Dissipation b) Convection
Figure 7.1: Exact traces of σ-roots for model equations.
b. Leapfrog Method
This is a two-root method, since there are two σ’s produced by every λ. When
applied to dissipation dominated problems we see from Fig. 7.2 that the principal
root is stable for a range of λh, but the spurious root is not. In fact, the spurious
root starts on the unit circle and falls outside of it for all (λh) < 0. However, for
biconvection cases, when λ is pure imaginary, the method is not only stable, but it
also produces a σ that falls precisely on the unit circle in the range 0 ≤ ωh ≤ 1.
As was pointed out above, this does not mean, that the method is without error.
Although the figure shows that there is a range of ωh in which the leapfrog method
produces no error in amplitude, it says nothing about the error in phase. More is said
about this in Chapter 8.
c. Second-Order Adams-Bashforth Method

This is also a two-root method but, unlike the leapfrog scheme, the spurious root
starts at the origin, rather than on the unit circle, see Fig. 7.2. Therefore, there is
a range of real negative λh for which the method will be stable. The figure shows
that the range ends when λh < −1.0 since at that point the spurious root leaves the
circle and |σ2 | becomes greater than one. The situation is quite different when the
σ1
σ1
λ h=-2
a) Euler Explicit
ωh = 1
σ2 σ1
σ2 σ1
b) Leapfrog
σ1
σ2 σ1 σ2
λ h=-1
c) AB2
Diffusion Convection
Figure 7.2: Traces of σ-roots for various methods.
λ-root is pure imaginary. In that case as ωh increases away from zero the spurious
root remains inside the circle and remains stable for a range of ωh. However, the
principal root falls outside the unit circle for all ωh > 0, and for the biconvection
model equation the method is unstable for all h.
d. Trapezoidal Method
The trapezoidal method is a very popular one for reasons that are partially illustrated
in Fig. 7.3. Its σ-roots fall on or inside the unit circle for both the dissipating and
the periodic convecting case and, in fact, it is stable for all values of λh for which
λ itself is inherently stable. Just like the leapfrog method it has the capability of
producing only phase error for the periodic convecting case, but there is a major
difference between the two since the trapezoidal method produces no amplitude error
for any ωh — not just a limited range between 0 ≤ ωh ≤ 1.
e. Gazdag Method
The Gazdag method was designed to produce low phase error. Since its characteristic
polynomial for σ is a cubic (Table 7.1, no. 10), it must have two spurious roots in
addition to the principal one. These are shown in Fig. 7.3. In both the dissipation
and biconvection cases, a spurious root limits the stability. For the dissipating case,
a spurious root leaves the unit circle when λh < −21 , and for the biconvecting case,
when ωh > 23 . Note that both spurious roots are located at the origin when λ = 0.
f,g. Second- and Fourth-Order Runge-Kutta Methods, RK2 and RK4

Traces of the σ-roots for the second- and fourth-order Runge-Kutta methods are
shown in Fig. 7.3. The figures show that both methods are stable for a range of
λh when λh is real and negative, but that the range of stability for RK4 is greater,
going almost all the way to -2.8, whereas RK2 is limited to -2. On the other hand for
biconvection RK2
√ is unstable for all ωh, whereas RK4 remains inside the unit circle
for 0 ≤ ωh ≤ 2 2. One can show that the RK4 stability limit is about |λh| < 2.8 for
all complex λh for which (λ) ≤ 0.
7.5.2 Stability for Small ∆t

It is interesting to pursue the question of stability when the time step size, h, is small
so that accuracy of all the λ-roots is of importance. Situations for which this is not
the case are considered in Chapter 8.
Mild instability
All conventional time-marching methods produce a principal root that is very close
to eλh for small values of λh. Therefore, on the basis of the principal root, the
stability of a method that is required to resolve a transient solution over a relatively
short time span may be a moot issue. Such cases are typified by the AB2 and RK2
methods when they are applied to a biconvection problem. Figs. 7.2c and 7.3f show
that for both methods the principal root falls outside the unit circle and is unstable
for all ωh. However, if the transient solution of interest can be resolved in a limited
number of time steps that are small in the sense of the figure, the error caused by this
instability may be relatively unimportant. If the root had fallen inside the circle the
σ1
σ1
λ h=- οο d) Trapezoidal ω h = 2/3

σ2
σ3 σ1
σ1
σ2 σ3
e) Gazdag
σ1 σ1
λ h=-2 f) RK2
σ1 σ1
g) RK4
λ h=-2.8 ωh= 2 2
Diffusion Convection
Figure 7.3: Traces of σ-roots for various methods (cont’d).

method would have been declared stable but an error of the same magnitude would
have been committed, just in the opposite direction. For this reason the AB2 and the
RK2 methods have both been used in serious quantitative studies involving periodic
convection. This kind of instability is referred to as mild instability and is not a
serious problem under the circumstances discussed.
Catastrophic instability
There is a much more serious stability problem for small h that can be brought about
by the existence of certain types of spurious roots. One of the best illustrations of this
kind of problem stems from a critical study of the most accurate, explicit, two-step,
linear multistep method (see Table 7.1):

un+1 = −4un + 5un−1 + 2h 2un + un−1 (7.13)
One can show, using the methods given in Section 6.6, that this method is third-
order accurate both in terms of erλ and erµ , so from an accuracy point of view it is
attractive. However, let us inspect its stability even for very small values of λh. This
can easily be accomplished by studying its characteristic polynomial when λh → 0.
From Eq. 7.13 it follows that for λh = 0, P (E) = E 2 + 4E − 5. Factoring P (σ) = 0
we find P (σ) = (σ − 1)(σ + 5) = 0. There are two σ-roots; σ1 , the principal one,
equal to 1, and σ2 , a spurious one, equal to -5!!
In order to evaluate the consequences of this result, one must understand how
methods with spurious roots work in practice. We know that they are not self start-
ing, and the special procedures chosen to start them initialize the coefficients of the
spurious roots, the cmk for k > 1 in Eq. 6.36. If the starting process is well designed
these coefficients are forced to be very small, and if the method is stable, they get
smaller with increasing n. However, if the magnitude of one of the spurious σ is
equal to 5, one can see disaster is imminent because (5)10 ≈ 107 . Even a very small
initial value of cmk is quickly overwhelmed. Such methods are called catastrophically
unstable and are worthless for most, if not all, computations.
Milne and Adams type methods

If we inspect the σ-root traces of the multiple root methods in Figs. 7.2 and 7.3, we
find them to be of two types. One type is typified by the leapfrog method. In this
case a spurious root falls on the unit circle when h → 0. The other type is exemplified
by the 2nd-order Adams-Bashforth and Gazdag methods. In this case all spurious
roots fall on the origin when h → 0.
The former type is referred to as a Milne Method. Since at least one spurious root
for a Milne method always starts on the unit circle, the method is likely to become
7.6. NUMERICAL STABILITY CONCEPTS IN THE COMPLEX λH PLANE135
unstable for some complex λ as h proceeds away from zero. On the basis of a Taylor
series expansion, however, these methods are generally the most accurate insofar as
they minimize the coefficient in the leading term for ert .
The latter type is referred to as an Adams Method. Since for these methods
all spurious methods start at the origin for h = 0, they have a guaranteed range of
stability for small enough h. However, on the basis of the magnitude of the coefficient
in the leading Taylor series error term, they suffer, relatively speaking, from accuracy.
For a given amount of computational work, the order of accuracy of the two
types is generally equivalent, and stability requirements in CFD applications generally
override the (usually small) increase in accuracy provided by a coefficient with lower
magnitude.
7.6 Numerical Stability Concepts in the Complex

λh Plane
7.6.1 Stability for Large h.
The reason to study stability for small values of h is fairly easy to comprehend.
Presumably we are seeking to resolve some transient and, since the accuracy of the
transient solution for all of our methods depends on the smallness of the time-step,
we seek to make the size of this step as small as possible. On the other hand, the
cost of the computation generally depends on the number of steps taken to compute
a solution, and to minimize this we wish to make the step size as large as possible.
In the compromise, stability can play a part. Aside from ruling out catastrophically
unstable methods, however, the situation in which all of the transient terms are
resolved constitutes a rather minor role in stability considerations.
By far the most important aspect of numerical stability occurs under conditions
when:
• One has inherently stable, coupled systems with λ–eigenvalues having widely
separated magnitudes.
or
• We seek only to find a steady-state solution using a path that includes the
unwanted transient.
In both of these cases there exist in the eigensystems relatively large values of
|λh| associated with eigenvectors that we wish to drive through the solution process
without any regard for their individual accuracy in eigenspace. This situation is
the major motivation for the study of numerical stability. It leads to the subject of
stiffness discussed in the next chapter.
7.6.2 Unconditional Stability, A-Stable Methods

Inherent stability of a set of ODE’s was defined in Section 7.2 and, for coupled sets
with a complete eigensystem, it amounted to the requirement that the real parts of all
λ eigenvalues must lie on, or to the left of, the imaginary axis in the complex λ plane.
This serves as an excellent reference frame to discuss and define the general stability
features of time-marching methods. For example, we start with the definition:
A numerical method is unconditionally stable if it is stable for all ODE’s

that are inherently stable.
A method with this property is said to be A-stable. A method is Ao -stable if the

region of stability contains the negative real axis in the complex λh plane, and I-stable
if it contains the entire imaginary axis. By applying a fairly simple test for A-stability
in terms of positive real functions to the class of two-step LMM’s given in Section
6.7.3, one finds these methods to be A-stable if and only if
1
θ ≥ϕ+ (7.14)
2
1
ξ≥− (7.15)
2
1
ξ ≤θ+ϕ− (7.16)
2
A set of A-stable implicit methods is shown in Table 7.2.
θ ξ ϕ Method Order
1 0 0 Implicit Euler 1
1/2 0 0 Trapezoidal 2
1 1/2 0 2nd O Backward 2
3/4 0 −1/4 Adams type 2
1/3 −1/2 −1/3 Lees 2
1/2 −1/2 −1/2 Two-step trapezoidal 2
5/8 −1/6 −2/9 A-contractive 2
Table 7.2. Some unconditionally stable (A-stable) implicit methods.
Notice that none of these methods has an accuracy higher than second-order. It can
be proved that the order of an A-stable LMM cannot exceed two, and, furthermore
that of all 2nd-order A-stable methods, the trapezoidal method has the smallest
truncation error.
Returning to the stability test using positive real functions one can show that a
two-step LMM is Ao -stable if and only if
1
θ ≥ϕ+ (7.17)
2
1
ξ≥− (7.18)
2
0≤θ−ϕ (7.19)
For first-order accuracy, the inequalities 7.17 to 7.19 are less stringent than 7.14 to
7.16. For second-order accuracy, however, the parameters (θ, ξ, ϕ) are related by the
condition
1
ϕ=ξ−θ+
2
and the two sets of inequalities reduce to the same set which is
ξ ≤ 2θ − 1 (7.20)
1
ξ≥− (7.21)
2
Hence, two-step, second-order accurate LMM’s that are A-stable and Ao -stable share
the same (ϕ, ξ, θ) parameter space. Although the order of accuracy of an A-stable
method cannot exceed two, Ao -stable LMM methods exist which have an accuracy of
arbitrarily high order.
It has been shown that for a method to be I-stable it must also be A-stable.
Therefore, no further discussion is necessary for the special case of I-stability.
It is not difficult to prove that methods having a characteristic polynomial for
which the coefficient of the highest order term in E is unity7 can never be uncondi-
tionally stable. This includes all explicit methods and predictor-corrector methods
made up of explicit sequences. Such methods are referred to, therefore, as condition-
ally stable methods.
7.6.3 Stability Contours in the Complex λh Plane.

A very convenient way to present the stability properties of a time-marching method
is to plot the locus of the complex λh for which |σ| = 1, such that the resulting
7
Or can be made equal to unity by a trivial normalization (division by a constant independent of
λh). The proof follows from the fact that the coefficients of such a polynomial are sums of various
combinations of products of all its roots.
I( λ h) I(λ h) I(λ h)
Stable
Unstable
1 1
Stable Unstable
Stable
R(λ h) R(λ h) Unstable R( λ h)
a) Euler Explicit θ=0 b) Trapezoid Implicit θ = 1/2 c) Euler Implicit θ=1
Figure 7.4: Stability contours for the θ-method.
contour goes through the point λh = 0. Here |σ| refers to the maximum absolute
value of any σ, principal or spurious, that is a root to the characteristic polynomial for
a given λh. It follows from Section 7.3 that on one side of this contour the numerical
method is stable while on the other, it is unstable. We refer to it, therefore, as a
stability contour.
Typical stability contours for both explicit and implicit methods are illustrated in
Fig. 7.4, which is derived from the one-root θ-method given in Section 6.5.4.
Contours for explicit methods

Fig. 7.4a shows the stability contour for the explicit Euler method. In the following
two ways it is typical of all stability contours for explicit methods:
1. The contour encloses a finite portion of the left-half complex λh-plane.
2. The region of stability is inside the boundary, and therefore, it is conditional.
However, this method includes no part of the imaginary axis (except for the origin)
and so it is unstable for the model biconvection problem. Although several explicit
methods share this deficiency (e.g., AB2, RK2), several others do not (e.g., leapfrog,
Gazdag, RK3, RK4), see Figs. 7.5 and 7.6. Notice in particular that the third- and
fourth-order Runge-Kutta
√ methods, Fig. 7.6, include a portion of the imaginary axis
out to ±1.9i and ±2 2i, respectively.
Contours for unconditionally stable implicit methods

Fig. 7.4c shows the stability contour for the implicit Euler method. It is typical of
many stability contours for unconditionally stable implicit methods. Notice that the
Figure 7.5: Stability contours for some explicit methods.
I( λh)
3.0
Stable RK4
Regions
RK3
2.0
RK2
1.0
RK1
R(λ h)
-3.0 -2.0 -1.0
Figure 7.6: Stability contours for Runge-Kutta methods.

I(λh)
I(λh)
Stable Outside Stable Outside
2.0 2.0
1.0 1.0 Unstable

Unstable
R(λ h) 1.0 R( λh)

1.0 2.0 3.0 4.0 2.0
a) 2nd Order Backward Implicit b) Adams Type
Figure 7.7: Stability contours for the 2 unconditionally stable implicit methods.
method is stable for the entire range of complex λh that fall outside the boundary.
This means that the method is numerically stable even when the ODE’s that it is
being used to integrate are inherently unstable. Some other implicit unconditionally
stable methods with the same property are shown in Fig. 7.7. In all of these cases
the imaginary axis is part of the stable region.
Not all unconditionally stable methods are stable in some regions where the ODE’s
they are integrating are inherently unstable. The classic example of a method that
is stable only when the generating ODE’s are themselves inherently stable is the
trapezoidal method, i.e., the special case of the θ-method for which θ = 12 . The
stability boundary for this case is shown in Fig. 7.4b. The boundary is the imaginary
axis and the numerical method is stable for λh lying on or to the left of this axis.
Two other methods that have this property are the two-step trapezoidal method

un+1 = un−1 + h un+1 + un−1
and a method due to Lees
2
un+1 = un−1 + h un+1 + un + un−1
3
Notice that both of these methods are of the Milne type.
Contours for conditionally stable implicit methods

Just because a method is implicit does not mean that it is unconditionally stable.
Two illustrations of this are shown in Fig. 7.8. One of these is the Adams-Moulton
3rd-order method (no. 8, Table 7.1). Another is the 4th-order Milne method given
by the point operator
1
un+1 = un−1 + h un+1 + 4un + un−1
3
√
and shown in Table 7.1 as no. 13. It is stable only for λ = ±iω when 0 ≤ ω ≤ 3.
Its stability boundary is very similar to that for the leapfrog method (see Fig. 7.5b).
7.7. FOURIER STABILITY ANALYSIS 141
I(λh)
I(λh)
3.0
3.0
Stable
only on 2.0
2.0 Imag Axis
Stable 1.0
1.0
-3.0 -2.0 -1.0 R(λ h) R(λ h)
a) AM3 b) Milne 4th Order
Figure 7.8: Stability contours for the 2 conditionally stable implicit methods.
7.7 Fourier Stability Analysis

By far the most popular form of stability analysis for numerical schemes is the Fourier
or von Neumann approach. This analysis is usually carried out on point operators and
it does not depend on an intermediate stage of ODE’s. Strictly speaking it applies
only to difference approximations of PDE’s that produce O∆E’s which are linear,
have no space or time varying coefficients, and have periodic boundary conditions.8
In practical application it is often used as a guide for estimating the worthiness of
a method for more general problems. It serves as a fairly reliable necessary stability
condition, but it is by no means a sufficient one.
7.7.1 The Basic Procedure

One takes data from a “typical” point in the flow field and uses this as constant
throughout time and space according to the assumptions given above. Then one
imposes a spatial harmonic as an initial value on the mesh and asks the question:
Will its amplitude grow or decay in time? The answer is determined by finding the
conditions under which
u(x, t) = eαt · eiκx (7.22)
is a solution to the difference equation, where κ is real and κ∆x lies in the range
0 ≤ κ∆x ≤ π. Since, for the general term,
(n+) (n)
uj+m = eα(t+∆t) · eiκ(x+m∆x) = eα∆t · eiκm∆x · uj
8
Another way of viewing this is to consider it as an initial value problem on an infinite space
domain.
(n)
the quantity uj is common to every term and can be factored out. In the remaining
expressions, we find the term eα∆t which we represent by σ, thus:
σ ≡ eα∆t
n
Then, since eαt = eα∆t = σ n , it is clear that
For numerical stability |σ| ≤ 1 (7.23)
and the problem is to solve for the σ’s produced by any given method and, as a
necessary condition for stability, make sure that, in the worst possible combination
of parameters, condition 7.23 is satisfied9 .
7.7.2 Some Examples

The procedure can best be explained by examples. Consider as a first example the
finite difference approximation to the model diffusion equation known as Richardson’s
method of overlapping steps. This was mentioned in Section 4.1 and given as Eq.
4.1:
(n+1) (n−1) 2∆t (n) (n) (n)

uj = uj +ν u − 2u + u (7.24)
∆x2 j+1 j j−1
Substitution of Eq. 7.22 into Eq. 7.24 gives the relation

2∆t iκ∆x
σ = σ−1 + ν e − 2 + e−iκ∆x
∆x2
or

4ν∆t
2 (1 − cos κ∆x) σ − 1 = 0
2
σ + (7.25)
∆x
2b
Thus Eq. 7.22 is a solution of Eq. 7.24 if σ is a root of Eq. 7.25. The two roots of
7.25 are √
σ1,2 = −b ± b2 + 1
from which it is clear that one |σ| is always > 1. We find, therefore, that by the
Fourier stability test, Richardson’s method of overlapping steps is unstable for all ν,
κ and ∆t.
If boundedness is required in a finite time domain, the condition is often presented as |σ| ≤
9
1 + O(∆t).
7.8. CONSISTENCY 143
As another example consider the finite-difference approximation for the model

biconvection equation
(n+1) (n) a∆t (n) (n)

uj = uj − uj+1 − uj−1 (7.26)
2∆x
In this case
a∆t
σ =1− · i · sin κ∆x
∆x
from which it is clear that |σ| > 1 for all nonzero a and κ. Thus we have another
finite-difference approximation that, by the Fourier stability test, is unstable for any
choice of the free parameters.
7.7.3 Relation to Circulant Matrices

The underlying assumption in a Fourier stability analysis is that the C matrix, deter-
mined when the differencing scheme is put in the form of Eq. 7.2, is circulant. Such
being the case, the eiκx in Eq. 7.22 represents an eigenvector of the system, and the
two examples just presented outline a simple procedure for finding the eigenvalues of
the circulant matrices formed by application of the two methods to the model prob-
lems. The choice of σ for the stability parameter in the Fourier analysis, therefore,
is not an accident. It is exactly the same σ we have been using in all of our previous
discussions, but arrived at from a different perspective.
If we examine the preceding examples from the viewpoint of circulant matrices
and the semi-discrete approach, the results present rather obvious conclusions. The
space differencing in Richardson’s method produces the matrix s · Bp (1, −2, 1) where
s is a positive scalar coefficient. From Appendix B we find that the eigenvalues of this
matrix are real negative numbers. Clearly, the time-marching is being carried out by
the leapfrog method and, from Fig. 7.5, this method is unstable for all eigenvalues with
negative real parts. On the other hand, the space matrix in Eq. 7.26 is Bp (−1, 0, 1),
and according to Appendix B, this matrix has pure imaginary eigenvalues. However,
in this case the explicit Euler method is being used for the time-march and, according
to Fig. 7.4, this method is always unstable for such conditions.
7.8 Consistency
Consider the model equation for diffusion analysis
∂u ∂2u
=ν 2 (7.27)
∂t ∂x
Many years before computers became available (1910, in fact), Lewis F. Richardson
proposed a method for integrating equations of this type. We presented his method
in Eq. 4.2 and analyzed its stability by the Fourier method in Section 7.7.
In Richardson’s time, the concept of numerical instability was not known. How-
ever, the concept is quite clear today and we now know immediately that his approach
would be unstable. As a semi-discrete method it can be expressed in matrix notation
as the system of ODE’s:
du ν
= B(1, −2, 1)u + (bc) (7.28)
dt ∆x2
with the leapfrog method used for the time march. Our analysis in this Chapter
revealed that this is numerically unstable since the λ-roots of B(1,-2,1) are all real
and negative and the spurious σ-root in the leapfrog method is unstable for all such
cases, see Fig. 7.5b.
The method was used by Richardson for weather prediction, and this fact can now
be a source of some levity. In all probability, however, the hand calculations (the
only approach available at the time) were not carried far enough to exhibit strange
phenomena. We could, of course, use the 2nd-order Runge-Kutta method to integrate
Eq. 7.28 since it is stable for real negative λ’s. It is, however, conditionally stable
and for this case we are rather severely limited in time step size by the requirement
∆t ≤ ∆x2 /(2ν).
There are many ways to manipulate the numerical stability of algorithms. One of
them is to introduce mixed time and space differencing, a possibility we have not yet
considered. For example, we introduced the DuFort-Frankel method in Chapter 4:
   
(n+1) (n−1)
2ν∆t  (n) u + uj
(n+1) (n−1)  j + uj+1
(n)
uj = uj + 2 uj−1 − 2 (7.29)
∆x 2
in which the central term in the space derivative in Eq. 4.2 has been replaced by its
average value at two different time levels. Now let
2ν∆t
α≡
∆x2
and rearrange terms

(n+1) (n−1) (n) (n)
(1 + α)uj = (1 − α)uj + α uj−1 + uj+1
There is no obvious ODE between the basic PDE and this final O∆E. Hence, there
is no intermediate λ-root structure to inspect. Instead one proceeds immediately to
the σ-roots.
7.8. CONSISTENCY 145
The simplest way to carry this out is by means of the Fourier stability analysis
introduced in Section 7.7. This leads at once to

(1 + α)σ = (1 − α)σ−1 + α eiκ∆x + e−iκ∆x
or
(1 + α)σ2 − 2ασ cos(k∆x) − (1 − α) = 0
The solution of the quadratic is

α cos κ∆x ± 1 − α2 sin2 κ∆x
σ=
1+α
There are 2M σ-roots all of which are ≤ 1 for any real α in the range 0 ≤ α ≤ ∞.
This means that the method is unconditionally stable!
The above result seems too good to be true, since we have found an unconditionally
stable method using an explicit combination of Lagrangian interpolation polynomials.
The price we have paid for this is the loss of con-

sistency with the original PDE.
To prove this, we expand the terms in Eq. 7.30 in a Taylor series and reconstruct
the partial differential equation we are actually solving as the mesh size becomes very
small. For the time derivative we have
1 (n+1) (n−1)
1 2
uj − uj = (∂t u)(n) (n)
j + ∆t (∂ttt u)j + · · ·
2∆t 6
and for the mixed time and space differences
(n) (n+1) (n−1) (n) 2
uj−1 − uj − uj + uj+1 ∆t
= (∂xx u)(n)
j − (∂tt u)(n)
j +
∆x2 ∆x
1
∆x2 (∂xxxx u)(n)
j −
12

1 ∆t 2
∆t2 j +···
(∂tttt u)(n) (7.30)
12 ∆x
Replace the terms in Eq. 7.30 with the above expansions and take the limit as
∆t , ∆x → 0. We find
∂u ∂2u ∂2 u
= ν 2 − νr2 2 (7.31)
∂t ∂x ∂t
where
∆t
r≡
∆x
Eq. 7.27 is parabolic. Eq. 7.31 is hyperbolic. Thus if ∆t → 0 and ∆x → 0 in such

a way that ∆x ∆t remains constant, the equation we actually solve by the method in
Eq. 7.30 is a wave equation, not a diffusion equation. In such a case Eq. 7.30 is not
uniformly consistent with the equation we set out to solve even for vanishingly small
step sizes. The situation is summarized in Table 7.3. If ∆t → 0 and ∆x → 0 in such
a way that ∆t2 remains constant, then r2 is O(∆t), and the method is consistent.
∆x
However, this leads to a constraint on the time step which is just as severe as that
imposed by the stability condition of the second-order Runge-Kutta method shown
in the table.
2nd-order Runge-Kutta DuFort–Frankel

For Stability
2
∆t ≤ ∆x
2ν ∆t ≤ ∞
Conditionally Stable Unconditionally Stable

For Consistency
Uniformly Consistent Conditionally Consistent,

2
with Approximates ∂∂tu = ν ∂∂xu2
∂u 2
∂t
= ν ∂∂xu2 only if

∆t 2 <
ν ∆x
Therefore

∆t < ∆x ν
Table 7.3: Summary of accuracy and consistency conditions for RK2 and
Du Fort-Frankel methods. = an arbitrary error bound.
7.9 Problems
1. Consider the ODE
du
u = = Au + f
dt
with    
−10 −0.1 −0.1 −1

A= 1 −1 1 
 , f = 0 


10 1 −1 0
7.9. PROBLEMS 147
(a) Find the eigenvalues of A using a numerical package. What is the steady-
state solution? How does the ODE solution behave in time?
(b) Write a code to integrate from the initial condition u(0) = [1, 1, 1]T from
time t = 0 using the explicit Euler, implicit Euler, and MacCormack meth-
ods. In all three cases, use h = 0.1 for 1000 time steps, h = 0.2 for 500
time steps, h = 0.4 for 250 time steps and h = 1.0 for 100 time steps.
Compare the computed solution with the exact steady solution.
(c) Using the λ-σ relations for these three methods, what are the expected
bounds on h for stability? Are your results consistent with these bounds?
2. (a) Compute a table of the numerical values of the σ-roots of the 2nd-order
Adams-Bashforth method when λ = i. Take h in intervals of 0.05 from 0
to 0.80 and compute the absolute values of the roots to at least 6 places.
(b) Plot the trace of the roots in the complex σ-plane and draw the unit circle
on the same plot.
(c) Repeat the above for the RK2 method.
3. When applied to the linear convection equation, the widely known Lax–Wendroff
method gives:
1 1
un+1
j = unj − Cn (unj+1 − unj−1) + Cn2 (unj+1 − 2unj + unj−1)
2 2
where Cn , known as the Courant (or CFL) number, is ah/∆x. Using Fourier
stability analysis, find the range of Cn for which the method is stable.
4. Determine and plot the stability contours for the Adams-Bashforth methods of
order 1 through 4. Compare with the Runge-Kutta methods of order 1 through
4.
5. Recall the compact centered 3-point approximation for a first derivative:
3
(δx u)j−1 + 4(δx u)j + (δx u)j+1 = (uj+1 − uj−1)
∆x
By replacing the spatial index j by the temporal index n, obtain a time-marching
method using this formula. What order is the method? Is it explicit or implicit?
Is it a two-step LMM? If so, to what values of ξ, θ, and φ (in Eq. 6.59) does it
correspond? Derive the λ-σ relation for the method. Is it A-stable, Ao -stable,
or I-stable?
6. Write the ODE system obtained by applying the 2nd-order centered difference
approximation to the spatial derivative in the model diffusion equation with
periodic boundary conditions. Using Appendix B.4, find the eigenvalues of the
spatial operator matrix. Given that the λ-σ relation for the 2nd-order Adams-
Bashforth method is
σ 2 − (1 + 3λh/2)σ + λh/2 = 0
show that the maximum stable value of |λh| for real negative λ, i.e., the point
where the stability contour intersects the negative real axis, is obtained with
λh = −1. Using the eigenvalues of the spatial operator matrix, find the maxi-
mum stable time step for the combination of 2nd-order centered differences and
the 2nd-order Adams-Bashforth method applied to the model diffusion equa-
tion. Repeat using Fourier analysis.
7. Consider the following PDE:

∂u ∂2 u
=i 2
∂t ∂x
√
where i = −1. Which explicit time-marching methods would be suitable for
integrating this PDE, if 2nd-order centered differencing is used for the spa-
tial differences and the boundary conditions are periodic? Find the stability
condition for one of these methods.
8. Using Fourier analysis, analyze the stability of first-order backward differenc-

ing coupled with explicit Euler time marching applied to the linear convection
equation with positive a. Find the maximum Courant number for stability.
9. Consider the linear convection with a positive wave speed as in problem 8. Apply
a Dirichlet boundary condition at the left boundary. No boundary condition is
permitted at the right boundary. Write the system of ODE’s which results from
first-order backward spatial differencing in matrix-vector form. Using Appendix
B.1, find the λ-eigenvalues. Write the O∆E which results from the application
of explicit Euler time marching in matrix-vector form, i.e.,
un+1 = Cun − gn
Write C in banded matrix notation and give the entries of g. Using the λ-
σ relation for the explicit Euler method, find the σ-eigenvalues. Based on
these, what is the maximum Courant number allowed for asymptotic stability?
Explain why this differs from the answer to problem 8. Hint: is C normal?
Chapter 8
CHOICE OF TIME-MARCHING
METHODS
In this chapter we discuss considerations involved in selecting a time-marching method

for a specific application. Examples are given showing how time-marching methods
can be compared in a given context. An important concept underlying much of this
discussion is stiffness, which is defined in the next section.
8.1 Stiffness Definition for ODE’s

8.1.1 Relation to λ-Eigenvalues
The introduction of the concept referred to as “stiffness” comes about from the nu-
merical analysis of mathematical models constructed to simulate dynamic phenom-
ena containing widely different time scales. Definitions given in the literature are
not unique, but fortunately we now have the background material to construct a
definition which is entirely sufficient for our purposes.
We start with the assumption that our CFD problem is modeled with sufficient
accuracy by a coupled set of ODE’s producing an A matrix typified by Eq. 7.1.
Any definition of stiffness requires a coupled system with at least two eigenvalues,
and the decision to use some numerical time-marching or iterative method to solve
it. The difference between the dynamic scales in physical space is represented by
the difference in the magnitude of the eigenvalues in eigenspace. In the following
discussion we concentrate on the transient part of the solution. The forcing function
may also be time varying in which case it would also have a time scale. However,
we assume that this scale would be adequately resolved by the chosen time-marching
method, and, since this part of the ODE has no effect on the numerical stability of
149
150 CHAPTER 8. CHOICE OF TIME-MARCHING METHODS
I(λ h)
Stable
Region
λh
1 R(λ h)
Accurate
Region
Figure 8.1: Stable and accurate regions for the explicit Euler method.
the homogeneous part, we exclude the forcing function from further discussion in this
section.
Consider now the form of the exact solution of a system of ODE’s with a com-
plete eigensystem. This is given by Eq. 6.27 and its solution using a one-root, time-
marching method is represented by Eq. 6.28. For a given time step, the time integra-
tion is an approximation in eigenspace that is different for every eigenvector xm . In
many numerical applications the eigenvectors associated with the small |λm | are well
resolved and those associated with the large |λm | are resolved much less accurately,
if at all. The situation is represented in the complex λh plane in Fig. 8.1. In this
figure the time step has been chosen so that time accuracy is given to the eigenvectors
associated with the eigenvalues lying in the small circle and stability without time
accuracy is given to those associated with the eigenvalues lying outside of the small
circle but still inside the large circle.
The whole concept of stiffness in CFD arises from the fact that we often do
not need the time resolution of eigenvectors associated with the large |λm | in
the transient solution, although these eigenvectors must remain coupled into
the system to maintain a high accuracy of the spatial resolution.
8.1. STIFFNESS DEFINITION FOR ODE’S 151
8.1.2 Driving and Parasitic Eigenvalues

For the above reason it is convenient to subdivide the transient solution into two
parts. First we order the eigenvalues by their magnitudes, thus
|λ1 | ≤ |λ2 | ≤ · · · ≤ |λM | (8.1)

Then we write
Transient p
M
= cm eλm t xm + cm eλm t xm (8.2)
Solution m=1 m=p+1

Driving Parasitic
This concept is crucial to our discussion. Rephrased, it states that we can separate
our eigenvalue spectrum into two groups; one [λ1 → λp ] called the driving eigenvalues
(our choice of a time-step and marching method must accurately approximate the time
variation of the eigenvectors associated with these), and the other, [λp+1 → λM ], called
the parasitic eigenvalues (no time accuracy whatsoever is required for the eigenvectors
associated with these, but their presence must not contaminate the accuracy of the
complete solution). Unfortunately, we find that, although time accuracy requirements
are dictated by the driving eigenvalues, numerical stability requirements are dictated
by the parasitic ones.
8.1.3 Stiffness Classifications

The following definitions are somewhat useful. An inherently stable set of ODE’s is
stiff if
|λp | |λM |
In particular we define the ratio
Cr = |λM | / |λp |
and form the categories
Mildly-stiff Cr < 102
Strongly-stiff 103 < Cr < 105
Extremely-stiff 106 < Cr < 108
Pathologically-stiff 109 < Cr
It should be mentioned that the gaps in the stiff category definitions are intentional
because the bounds are arbitrary. It is important to notice that these definitions
make no distinction between real, complex, and imaginary eigenvalues.
8.2 Relation of Stiffness to Space Mesh Size

Many flow fields are characterized by a few regions having high spatial gradients of
the dependent variables and other domains having relatively low gradient phenomena.
As a result it is quite common to cluster mesh points in certain regions of space and
spread them out otherwise. Examples of where this clustering might occur are at a
shock wave, near an airfoil leading or trailing edge, and in a boundary layer.
One quickly finds that this grid clustering can strongly affect the eigensystem of
the resulting A matrix. In order to demonstrate this, let us examine the eigensystems
of the model problems given in Section 4.3.2. The simplest example to discuss relates
to the model diffusion equation. In this case the eigenvalues are all real, negative
numbers that automatically obey the ordering given in Eq. 8.1. Consider the case
when all of the eigenvalues are parasitic, i.e., we are interested only in the converged
steady-state solution. Under these conditions, the stiffness is determined by the ratio
λM /λ1 . A simple calculation shows that
2
4ν π 4ν ∆x
λ1 = − 2 sin2 ≈− = −ν
∆x 2(M + 1) ∆x2 2

4ν 2 π 4ν
λM ≈ − 2 sin =−
∆x 2 ∆x2
and the ratio is
4 M +1 2
λM /λ1 ≈ = 4
∆x2 π
The most important information found from this example is the fact that the
stiffness of the transient solution is directly related to the grid spacing. Furthermore,
in diffusion problems this stiffness is proportional to the reciprocal of the space mesh
size squared. For a mesh size M = 40, this ratio is about 680. Even for a mesh of
this moderate size the problem is already approaching the category of strongly stiff.
For the biconvection model a similar analysis shows that
1
|λM | / |λ1 | ≈
∆x
Here the stiffness parameter is still space-mesh dependent, but much less so than for
diffusion-dominated problems.
We see that in both cases we are faced with the rather annoying fact that the more
we try to increase the resolution of our spatial gradients, the stiffer our equations tend
to become. Typical CFD problems without chemistry vary between the mildly and
strongly stiff categories, and are greatly affected by the resolution of a boundary layer
since it is a diffusion process. Our brief analysis has been limited to equispaced
8.3. PRACTICAL CONSIDERATIONS FOR COMPARING METHODS 153
problems, but in general the stiffness of CFD problems is proportional to the mesh
intervals in the manner shown above where the critical interval is the smallest one in
the physical domain.
8.3 Practical Considerations for Comparing Meth-

ods
We have presented relatively simple and reliable measures of stability and both the
local and global accuracy of time-marching methods. Since there are an endless
number of these methods to choose from, one can wonder how this information is
to be used to pick a “best” choice for a particular problem. There is no unique
answer to such a question. For example, it is, among other things, highly dependent
upon the speed, capacity, and architecture of the available computer, and technology
influencing this is undergoing rapid and dramatic changes as this is being written.
Nevertheless, if certain ground rules are agreed upon, relevant conclusions can be
reached. Let us now examine some ground rules that might be appropriate. It should
then be clear how the analysis can be extended to other cases.
Let us consider the problem of measuring the efficiency of a time–marching method
for computing, over a fixed interval of time, an accurate transient solution of a coupled
set of ODE’s. The length of the time interval, T , and the accuracy required of the
solution are dictated by the physics of the particular problem involved. For example,
in calculating the amount of turbulence in a homogeneous flow, the time interval
would be that required to extract a reliable statistical sample, and the accuracy
would be related to how much the energy of certain harmonics would be permitted
to distort from a given level. Such a computation we refer to as an event.
The appropriate error measures to be used in comparing methods for calculating
an event are the global ones, Era , Erλ and Erω , discussed in Section 6.6.5, rather
than the local ones erλ , era , and erp discussed earlier.
The actual form of the coupled ODE’s that are produced by the semi-discrete
approach is
du
= F (u, t)
dt
At every time step we must evaluate the function F (u, t) at least once. This function
is usually nonlinear, and its computation usually consumes the major portion of the
computer time required to make the simulation. We refer to a single calculation of the
vector F (u, t) as a function evaluation and denote the total number of such evaluations
by Fev .
8.4 Comparing the Efficiency of Explicit Methods

8.4.1 Imposed Constraints
As mentioned above, the efficiency of methods can be compared only if one accepts a
set of limiting constraints within which the comparisons are carried out. The follow
assumptions bound the considerations made in this Section:
1. The time-march method is explicit.
2. Implications of computer storage capacity and access time are ignored. In some
contexts, this can be an important consideration.
3. The calculation is to be time-accurate, must simulate an entire event which

takes a total time T , and must use a constant time step size, h, so that
T = Nh
where N is the total number of time steps.
8.4.2 An Example Involving Diffusion

Let the event be the numerical solution of
du
= −u (8.3)
dt
from t = 0 to T = − ln(0.25) with u(0) = 1. Eq. 8.3 is obtained from our representa-
tive ODE with λ = −1, a = 0. Since the exact solution is u(t) = u(0)e−t , this makes
the exact value of u at the end of the event equal to 0.25, i.e., u(T ) = 0.25. To the
constraints imposed above, let us set the additional requirement
• The error in u at the end of the event, i.e., the global error, must be < 0.5%.
We judge the most efficient method as the one that satisfies these conditions and
has the fewest number of evaluations, Fev . Three methods are compared — explicit
Euler, AB2, and RK4.
First of all, the allowable error constraint means that the global error in the am-
plitude, see Eq. 6.48, must have the property:
+ +
+ Erλ +
+ + < 0.005
+ λT +
e
8.4. COMPARING THE EFFICIENCY OF EXPLICIT METHODS 155
Then, since h = T /N = − ln(0.25)/N, it follows that

+ +
+ N +
+1 − (σ1 (ln(.25)/N)) /.25+ < 0.005
where σ1 is found from the characteristic polynomials given in Table 7.1. The results
shown in Table 8.1 were computed using a simple iterative procedure.
Method N h σ1 Fev Erλ

Euler 193 .00718 .99282 193 .001248 worst
AB2 16 .0866 .9172 16 .001137
RK4 2 .6931 .5012 8 .001195 best
Table 8.1: Comparison of time-marching methods for a simple dissipation problem.
In this example we see that, for a given global accuracy, the method with the
highest local accuracy is the most efficient on the basis of the expense in evaluating
Fev . Thus the second-order Adams-Bashforth method is much better than the first-
order Euler method, and the fourth-order Runge-Kutta method is the best of all. The
main purpose of this exercise is to show the (usually) great superiority of second-order
over first-order time-marching methods.
8.4.3 An Example Involving Periodic Convection

Let us use as a basis for this example the study of homogeneous turbulence simulated
by the numerical solution of the incompressible Navier-Stokes equations inside a cube
with periodic boundary conditions on all sides. In this numerical experiment the
function evaluations contribute overwhelmingly to the CPU time, and the number of
these evaluations must be kept to an absolute minimum because of the magnitude of
the problem. On the other hand, a complete event must be established in order to
obtain meaningful statistical samples which are the essence of the solution. In this
case, in addition to the constraints given in Section 8.4.1, we add the following:
• The number of evaluations of F (u, t) is fixed.
Under these conditions a method is judged as best when it has the highest global
accuracy for resolving eigenvectors with imaginary eigenvalues. The above constraint
has led to the invention of schemes that omit the function evaluation in the cor-
rector step of a predictor-corrector combination, leading to the so-called incomplete
predictor-corrector methods. The presumption is, of course, that more efficient meth-
ods will result from the omission of the second function evaluation. An example is
the method of Gazdag, given in Section 6.8. Basically this is composed of an AB2
predictor and a trapezoidal corrector. However, the derivative of the fundamental
family is never found so there is only one evaluation required to complete each cycle.
The λ-σ relation for the method is shown as entry 10 in Table 7.1.
In order to discuss our comparisions we introduce the following definitions:
• Let a k-evaluation method be defined as one that requires k evaluations of
F (u, t) to advance one step using that method’s time interval, h.
• Let K represent the total number of allowable Fev .
• Let h1 be the time interval advanced in one step of a one-evaluation method.
The Gazdag, leapfrog, and AB2 schemes are all 1-evaluation methods. The second
and fourth order RK methods are 2- and 4-evaluation methods, respectively. For a 1-
evaluation method the total number of time steps, N , and the number of evaluations,
K, are the same, one evaluation being used for each step, so that for these methods
h = h1 . For a 2-evaluation method N = K/2 since two evaluations are used for
each step. However, in this case, in order to arrive at the same time T after K
evaluations, the time step must be twice that of a one–evaluation method so h = 2h1 .
For a 4-evaluation method the time interval must be h = 4h1 , etc. Notice that
as k increases, the time span required for one application of the method increases.
However, notice also that as k increases, the power to which σ1 is raised to arrive
at the final destination decreases; see the Figure below. This is the key to the true
comparison of time-march methods for this type of problem.
0 T uN
k=1 | • • • • • • • | [σ(λh1 )]8
k=2 | 2h1 • • • | [σ(2λh1 )]4
k=4 | 4h1 • | [σ(4λh1 )]2
Step sizes and powers of σ for k-evaluation methods used to get to the same value
of T if 8 evaluations are allowed.
In general, after K evaluations, the global amplitude and phase error for k-
evaluation methods applied to systems with pure imaginary λ-roots can be written1
Era = 1 − |σ1 (ikωh1 )|K/k (8.4)
1
See Eqs. 6.38 and 6.39.
8.4. COMPARING THE EFFICIENCY OF EXPLICIT METHODS 157

K [σ1 (ikωh1 )]imaginary
Erω = ωT − tan−1 (8.5)
k [σ1 (ikωh1 )]real
Consider a convection-dominated event for which the function evaluation is very

time consuming. We idealize to the case where λ = iω and set ω equal to one. The
event must proceed to the time t = T = 10. We consider two maximum evaluation
limits K = 50 and K = 100 and choose from four possible methods, leapfrog, AB2,
Gazdag, and RK4. The first three of these are one-evaluation methods and the last
one is a four-evaluation method. It is not difficult to show that on the basis of local
error (made in a single step) the Gazdag method is superior to the RK4 method in
both amplitude and phase. For example, for ωh = 0.2 the Gazdag method produces
a |σ1 | = 0.9992276 whereas for ωh = 0.8 (which must be used to keep the number of
evaluations the same) the RK4 method produces a |σ1 | = 0.998324. However, we are
making our comparisons on the basis of global error for a fixed number of evaluations.
First of all we see that for a one-evaluation method h1 = T /K. Using this, and the
fact that ω = 1, we find, by some rather simple calculations2 made using Eqs. 8.4 and
8.5, the results shown in Table 8.2. Notice that to find global error the Gazdag root
must be raised to the power of 50 while the RK4 root is raised only to the power of
50/4. On the basis of global error the Gazdag method is not superior to RK4 in either
amplitude or phase, although, in terms of phase error (for which it was designed) it
is superior to the other two methods shown.
K leapfrog AB2 Gazdag RK4

ωh1 = .1 100 1.0 1.003 .995 .999
ωh1 = .2 50 1.0 1.022 .962 .979
a. Amplitude, exact = 1.0.

K leapfrog AB2 Gazdag RK4
ωh1 = .1 100 −.96 −2.4 .45 .12
ωh1 = .2 50 −3.8 −9.8 1.5 1.5
b. Phase error in degrees.
Table 8.2: Comparison of global amplitude and phase errors for four methods.
2
The σ1 root for the Gazdag method can be found using a numerical root finding routine to trace
the three roots in the σ-plane, see Fig. 7.3e.
Using analysis such as this (and also considering the stability boundaries) the
RK4 method is recommended as a basic first choice for any explicit time-accurate
calculation of a convection-dominated problem.
8.5 Coping With Stiffness

8.5.1 Explicit Methods
The ability of a numerical method to cope with stiffness can be illustrated quite nicely
in the complex λh plane. A good example of the concept is produced by studying
the Euler method applied to the representative equation. The transient solution is
un = (1 + λh)n and the trace of the complex value of λh which makes |1 + λh| = 1
gives the whole story. In this case the trace forms a circle of unit radius centered at
(−1, 0) as shown in Fig. 8.1. If h is chosen so that all λh in the ODE eigensystem
fall inside this circle the integration will be numerically stable. Also shown by the
small circle centered at the origin is the region of Taylor series accuracy. If some λh
fall outside the small circle but stay within the stable region, these λh are stiff, but
stable. We have defined these λh as parasitic eigenvalues. Stability boundaries for
some explicit methods are shown in Figs. 7.5 and 7.6.
For a specific example, consider the mildly stiff system composed of a coupled
two-equation set having the two eigenvalues λ1 = −100 and λ2 = −1. If uncoupled
and evaluated in wave space, the time histories of the two solutions would appear as a
rapidly decaying function in one case, and a relatively slowly decaying function in the
other. Analytical evaluation of the time histories poses no problem since e−100t quickly
becomes very small and can be neglected in the expressions when time becomes large.
Numerical evaluation is altogether different. Numerical solutions, of course, depend
upon [σ(λm h)]n and no |σm | can exceed one for any λm in the coupled system or else
the process is numerically unstable.
Let us choose the simple explicit Euler method for the time march. The coupled
equations in real space are represented by
u1 (n) = c1 (1 − 100h)n x11 + c2 (1 − h)n x12 + (P S)1

u2 (n) = c1 (1 − 100h)n x21 + c2 (1 − h)n x22 + (P S)2 (8.6)
We will assume that our accuracy requirements are such that sufficient accuracy is
obtained as long as |λh| ≤ 0.1. This defines a time step limit based on accuracy
considerations of h = 0.001 for λ1 and h = 0.1 for λ2 . The time step limit based
on stability, which is determined from λ1 , is h = 0.02. We will also assume that
c1 = c2 = 1 and that an amplitude less than 0.001 is negligible. We first run 66
time steps with h = 0.001 in order to resolve the λ1 term. With this time step the
8.5. COPING WITH STIFFNESS 159
λ2 term is resolved exceedingly well. After 66 steps, the amplitude of the λ1 term
(i.e., (1 − 100h)n ) is less than 0.001 and that of the λ2 term (i.e., (1 − h)n ) is 0.9361.
Hence the λ1 term can now be considered negligible. To drive the (1 − h)n term to
zero (i.e., below 0.001), we would like to change the step size to h = 0.1 and continue.
We would then have a well resolved answer to the problem throughout the entire
relevant time interval. However, this is not possible because of the coupled presence
of (1 − 100h)n , which in just 10 steps at h = 0.1 amplifies those terms by ≈ 109 , far
outweighing the initial decrease obtained with the smaller time step. In fact, with
h = 0.02, the maximum step size that can be taken in order to maintain stability,
about 339 time steps have to be computed in order to drive e−t to below 0.001. Thus
the total simulation requires 405 time steps.
8.5.2 Implicit Methods

Now let us re-examine the problem that produced Eq. 8.6 but this time using an
unconditionally stable implicit method for the time march. We choose the trapezoidal
method. Its behavior in the λh plane is shown in Fig. 7.4b. Since this is also a one-
root method, we simply replace the Euler σ with the trapezoidal one and analyze the
result. It follows that the final numerical solution to the ODE is now represented in
real space by
n n
1 − 50h 1 − 0.5h
u1 (n) = c1 x11 + c2 x12 + (P S)1
1 + 50h 1 + 0.5h
n n
1 − 50h 1 − 0.5h
u2 (n) = c1 x21 + c2 x22 + (P S)2 (8.7)
1 + 50h 1 + 0.5h
In order to resolve the initial transient of the term e−100t , we need to use a step size
of about h = 0.001. This is the same step size used in applying the explicit Euler
method because here accuracy is the only consideration and a very small step size
must be chosen to get the desired resolution. (It is true that for the same accuracy
we could in this case use a larger step size because this is a second-order method,
but that is not the point of this exercise). After 70 time steps the λ1 term has
amplitude less than 0.001 and can be neglected. Now with the implicit method we
can proceed to calculate the remaining part of the event using our desired step size
h = 0.1 without any problem of instability, with 69 steps required to reduce the
amplitude of the second term to below 0.001. In both intervals the desired solution
is second-order accurate and well resolved. It is true that in the final 69 steps one
σ-root is [1 − 50(0.1)]/[1 + 50(0.1)] = 0.666 · · ·, and this has no physical meaning
whatsoever. However, its influence on the coupled solution is negligible at the end
of the first 70 steps, and, since (0.666 · · ·)n < 1, its influence in the remaining 70
steps is even less. Actually, although this root is one of the principal roots in the
system, its behavior for t > 0.07 is identical to that of a stable spurious root. The
total simulation requires 139 time steps.
8.5.3 A Perspective
It is important to retain a proper perspective on a problem represented by the above
example. It is clear that an unconditionally stable method can always be called upon
to solve stiff problems with a minimum number of time steps. In the example, the
conditionally stable Euler method required 405 time steps, as compared to about
139 for the trapezoidal method, about three times as many. However, the Euler
method is extremely easy to program and requires very little arithmetic per step. For
preliminary investigations it is often the best method to use for mildly-stiff diffusion
dominated problems. For refined investigations of such problems an explicit method
of second order or higher, such as Adams-Bashforth or Runge-Kutta methods, is
recommended. These explicit methods can be considered as effective mildly stiff-
stable methods. However, it should be clear that as the degree of stiffness of the
problem increases, the advantage begins to tilt towards implicit methods, as the
reduced number of time steps begins to outweigh the increased cost per time step.
The reader can repeat the above example with λ1 = −10, 000, λ2 = −1, which is in
the strongly-stiff category.
There is yet another technique for coping with certain stiff systems in fluid dynamic
applications. This is known as the multigrid method. It has enjoyed remarkable
success in many practical problems; however, we need an introduction to the theory
of relaxation before it can be presented.
8.6 Steady Problems

In Chapter 6 we wrote the O∆E solution in terms of the principal and spurious roots
as follows:
un = c11 (σ1 )n1 x1 + · · · + cm1(σm )n1 xm + · · · + cM 1 (σM )n1 xM + P.S.
+c12 (σ1 )n x1 + · · · + cm2 (σm )n xm + · · · + cM 2 (σM )n xM
2 2 2
x1 + · · · + cm3 (σm )n xm + · · · + cM 3 (σM )n xM
+c13 (σ1 )n3 3 3
+etc., if there are more spurious roots (8.8)
When solving a steady problem, we have no interest whatsoever in the transient por-
tion of the solution. Our sole goal is to eliminate it as quickly as possible. Therefore,
8.7. PROBLEMS 161
the choice of a time-marching method for a steady problem is similar to that for a
stiff problem, the difference being that the order of accuracy is irrelevant. Hence the
explicit Euler method is a candidate for steady diffusion dominated problems, and
the fourth-order Runge-Kutta method is a candidate for steady convection dominated
problems, because of their stability properties. Among implicit methods, the implicit
Euler method is the obvious choice for steady problems.
When we seek only the steady solution, all of the eigenvalues can be considered to
be parasitic. Referring to Fig. 8.1, none of the eigenvalues are required to fall in the
accurate region of the time-marching method. Therefore the time step can be chosen
to eliminate the transient as quickly as possible with no regard for time accuracy.
For example, when using the implicit Euler method with local time linearization, Eq.
6.96, one would like to take the limit h → ∞, which leads to Newton’s method, Eq.
6.98. However, a finite time step may be required until the solution is somewhat close
to the steady solution.
8.7 Problems
1. Repeat the time-march comparisons for diffusion (Section 8.4.2) and periodic
convection (Section 8.4.3) using 2nd- and 3rd-order Runge-Kutta methods.
2. Repeat the time-march comparisons for diffusion (Section 8.4.2) and periodic
convection (Section 8.4.3) using the 3rd- and 4th-order Adams-Bashforth meth-
ods. Considering the stability bounds for these methods (see problem 4 in
Chapter 7) as well as their memory requirements, compare and contrast them
with the 3rd- and 4th-order Runge-Kutta methods.
3. Consider the diffusion equation (with ν = 1) discretized using 2nd-order central

differences on a grid with 10 (interior) points. Find and plot the eigenvalues
and the corresponding modified wavenumbers. If we use the explicit Euler time-
marching method what is the maximum allowable time step if all but the first
two eigenvectors are considered parasitic? Assume that sufficient accuracy is
obtained as long as |λh| ≤ 0.1. What is the maximum allowable time step if all
but the first eigenvector are considered parasitic?
Chapter 9
RELAXATION METHODS
In the past three chapters, we developed a methodology for designing, analyzing,

and choosing time-marching methods. These methods can be used to compute the
time-accurate solution to linear and nonlinear systems of ODE’s in the general form
du
= F (u, t) (9.1)
dt
which arise after spatial discretization of a PDE. Alternatively, they can be used to
solve for the steady solution of Eq. 9.1, which satisfies the following coupled system
of nonlinear algebraic equations:
F (u) = 0 (9.2)
In the latter case, the unsteady equations are integrated until the solution converges
to a steady solution. The same approach permits a time-marching method to be used
to solve a linear system of algebraic equations in the form
Ax = b (9.3)
To solve this system using a time-marching method, a time derivative is introduced

as follows
dx
= Ax − b (9.4)
dt
and the system is integrated in time until the transient has decayed to a sufficiently
low level. Following a time-dependent path to steady state is possible only if all of
the eigenvalues of the matrix A (or −A) have real parts lying in the left half-plane.
Although the solution x = A−1b exists as long as A is nonsingular, the ODE given by
Eq. 9.4 has a stable steady solution only if A meets the above condition.
163
164 CHAPTER 9. RELAXATION METHODS
The common feature of all time-marching methods is that they are at least first-
order accurate. In this chapter, we consider iterative methods which are not time
accurate at all. Such methods are known as relaxation methods. While they are
applicable to coupled systems of nonlinear algebraic equations in the form of Eq. 9.2,
our analysis will focus on their application to large sparse linear systems of equations
in the form
Abu − f b = 0 (9.5)
where Ab is nonsingular, and the use of the subscript b will become clear shortly. Such
systems of equations arise, for example, at each time step of an implicit time-marching
method or at each iteration of Newton’s method. Using an iterative method, we seek
to obtain rapidly a solution which is arbitrarily close to the exact solution of Eq. 9.5,
which is given by
u∞ = A−1f b (9.6)
b
9.1 Formulation of the Model Problem

9.1.1 Preconditioning the Basic Matrix
It is standard practice in applying relaxation procedures to precondition the basic
equation. This preconditioning has the effect of multiplying Eq. 9.5 from the left
by some nonsingular matrix. In the simplest possible case the conditioning matrix
is a diagonal matrix composed of a constant D(b). If we designate the conditioning
matrix by C, the problem becomes one of solving for u in
CAbu − C f b = 0 (9.7)
Notice that the solution of Eq. 9.7 is
u = [CAb ]−1 C f b = A−1 −1

C f b = A−1
b C b fb (9.8)
which is identical to the solution of Eq. 9.5, provided C−1 exists.

In the following we will see that our approach to the iterative solution of Eq. 9.7
depends crucially on the eigenvalue and eigenvector structure of the matrix CAb , and,
equally important, does not depend at all on the eigensystem of the basic matrix Ab .
For example, there are well-known techniques for accelerating relaxation schemes if
the eigenvalues of CAb are all real and of the same sign. To use these schemes, the
9.1. FORMULATION OF THE MODEL PROBLEM 165
conditioning matrix C must be chosen such that this requirement is satisfied. A

choice of C which ensures this condition is the negative transpose of Ab .
For example, consider a spatial discretization of the linear convection equation
using centered differences with a Dirichlet condition on the left side and no constraint
on the right side. Using a first-order backward difference on the right side (as in
Section 3.6), this leads to the approximation
   



0 1 −ua 


   

 −1 0 1   0 
1    
δxu =  −1 0 1   0 
 u+   (9.9)
2∆x 
   

 −1 0 1   0 


 

−2 2 0
The matrix in Eq. 9.9 has eigenvalues whose imaginary parts are much larger than
their real parts. It can first be conditioned so that the modulus of each element is 1.
This is accomplished using a diagonal preconditioning matrix
 
1 0 0 0 0
 
 0 1 0 0 0 
 
D= 2∆x 
 0 0 1 0 0 
 (9.10)
 
 0 0 0 1 0 
1
0 0 0 0 2
which scales each row. We then further condition with multiplication by the negative
transpose. The result is
A2 = −AT1 A1 =
   
0 1 0 1
   
 −1 0 1   −1 0 1 
   
 −1 0 1   −1 0 1 
   
   
 −1 0 1   −1 0 1 
−1 −1 −1 1
 
−1 0 1
 


0 −2 0 1 

=  1 0 −2 0 1  (9.11)
 
 
 1 0 −2 1 
1 1 −2
If we define a permutation matrix P 1 and carry out the process P T [−AT1 A1 ]P (which
just reorders the elements of A1 and doesn’t change the eigenvalues) we find
    
0 1 0 0 0 −1 0 1 0 0 0 0 1
     


0 0 0 1 0 



0 −2 0 1 



1 0 0 0 0 

 0 0 0 0 1   1 0 −2 0 1   0 0 0 1 0 
     
     
 0 0 1 0 0   1 0 −2 1   0 1 0 0 0 
1 0 0 0 0 1 1 −2 0 0 1 0 0
 
−2 1
 
 1 −2 1 
 
=
 1 −2 1 
 (9.12)
 
 1 −2 1 
1 −1
which has all negative real eigenvalues, as given in Appendix B. Thus even when
the basic matrix Ab has nearly imaginary eigenvalues, the conditioned matrix −ATb Ab
is nevertheless symmetric negative definite (i.e., symmetric with negative real eigen-
values), and the classical relaxation methods can be applied. We do not necessarily
recommend the use of −ATb as a preconditioner; we simply wish to show that a broad
range of matrices can be preconditioned into a form suitable for our analysis.
9.1.2 The Model Equations

Preconditioning processes such as those described in the last section allow us to
prepare our algebraic equations in advance so that certain eigenstructures are guar-
anteed. In the remainder of this chapter, we will thoroughly investigate some simple
equations which model these structures. We will consider the preconditioned system
of equations having the form
Aφ − f = 0 (9.13)
where A is symmetric negative definite.2 The symbol for the dependent variable has
been changed to φ as a reminder that the physics being modeled is no longer time
1
A permutation matrix (defined as a matrix with exactly one 1 in each row and column and has
the property that P T = P −1 ) just rearranges the rows and columns of a matrix.
2
We use a symmetric negative definite matrix to simplify certain aspects of our analysis. Relax-
ation methods are applicable to more general matrices. The classical methods will usually converge
if Ab is diagonally dominant, as defined in Appendix A.
9.1. FORMULATION OF THE MODEL PROBLEM 167
accurate when we later deal with ODE formulations. Note that the solution of Eq.
9.13, φ = A−1f , is guaranteed to exist because A is nonsingular. In the notation of
Eqs. 9.5 and 9.7,
A = CAb and f = C f (9.14)

b
The above was written to treat the general case. It is instructive in formulating the
concepts to consider the special case given by the diffusion equation in one dimension
with unit diffusion coefficient ν:
∂u ∂2 u
= 2 − g(x) (9.15)
∂t ∂x
This has the steady-state solution
∂2u
= g(x) (9.16)
∂x2
which is the one-dimensional form of the Poisson equation. Introducing the three-
point central differencing scheme for the second derivative with Dirichlet boundary
conditions, we find
du 1 − g
= B(1, −2, 1)u + (bc) (9.17)
dt ∆x2
contains the boundary conditions and g contains the values of the source
where (bc)
term at the grid nodes. In this case
1
Ab = B(1, −2, 1)
∆x2
fb
= g − (bc) (9.18)
Choosing C = ∆x2 I, we obtain
B(1, −2, 1)φ = f (9.19)
where f = ∆x2 fb . If we consider a Dirichlet boundary condition on the left side and
either a Dirichlet or a Neumann condition on the right side, then A has the form

A = B 1, b, 1
b = [−2, −2, · · · ·, −2, s]T
s = −2 or −1 (9.20)
Note that s = −1 is easily obtained from the matrix resulting from the Neumann
boundary condition given in Eq. 3.24 using a diagonal conditioning matrix. A tremen-
dous amount of insight to the basic features of relaxation is gained by an appropriate
study of the one-dimensional case, and much of the remaining material is devoted to
this case. We attempt to do this in such a way, however, that it is directly applicable
to two- and three-dimensional problems.
9.2 Classical Relaxation

9.2.1 The Delta Form of an Iterative Scheme
We will consider relaxation methods which can be expressed in the following delta
form:

H φn+1 − φn = Aφn − f (9.21)
where H is some nonsingular matrix which depends upon the iterative method. The
matrix H is independent of n for stationary methods and is a function of n for
nonstationary ones. The iteration count is designated by the subscript n or the
superscript (n). The converged solution is designated φ∞ so that
φ∞ = A−1f (9.22)
9.2.2 The Converged Solution, the Residual, and the Error

Solving Eq. 9.21 for φn+1 gives
φn+1 = [I + H −1 A]φn − H −1f = Gφn − H −1f (9.23)
where
G ≡ I + H −1 A (9.24)
Hence it is clear that H should lead to a system of equations which is easy to solve,
or at least easier to solve than the original system. The error at the nth iteration is
defined as
en ≡ φn − φ∞ = φn − A−1f (9.25)
9.2. CLASSICAL RELAXATION 169
where φ∞ was defined in Eq. 9.22. The residual at the nth iteration is defined as
rn ≡ Aφn − f (9.26)
Multiply Eq. 9.25 by A from the left, and use the definition in Eq. 9.26. There results
the relation between the error and the residual
Aen − rn = 0 (9.27)
Finally, it is not difficult to show that

en+1 = Gen (9.28)
Consequently, G is referred to as the basic iteration matrix, and its eigenvalues, which
we designate as σm , determine the convergence rate of a method.
In all of the above, we have considered only what are usually referred to as sta-
tionary processes in which H is constant throughout the iterations. Nonstationary
processes in which H (and possibly C) is varied at each iteration are discussed in
Section 9.5.
9.2.3 The Classical Methods

Point Operator Schemes in One Dimension
Let us consider three classical relaxation procedures for our model equation
B(1, −2, 1)φ = f (9.29)
as given in Section 9.1.2. The Point-Jacobi method is expressed in point operator

form for the one-dimensional case as
1 (n)
φ(n+1)
j = φj−1 + φ(n)
j+1 − fj (9.30)
2
(n+1)
This operator comes about by choosing the value of φj such that together with
the old values of φj−1 and φj+1, the jth row of Eq. 9.29 is satisfied. The Gauss-Seidel
method is
(n+1) 1 (n+1) (n)

φj = φj−1 + φj+1 − fj (9.31)
2
This operator is a simple extension of the point-Jacobi method which uses the most
recent update of φj−1 . Hence the jth row of Eq. 9.29 is satisfied using the new values of
φj and φj−1 and the old value of φj+1. The method of successive overrelaxation (SOR)
is based on the idea that if the correction produced by the Gauss-Seidel method tends
to move the solution toward φ∞ , then perhaps it would be better to move further in
this direction. It is usually expressed in two steps as
1 (n+1) (n)

φ̃j = φj−1 + φj+1 − fj
2
(n+1) (n) (n)
φj = φj + ω φ̃j − φj (9.32)
where ω generally lies between 1 and 2, but it can also be written in the single line
(n+1) ω (n+1) (n) ω (n) ω
φj = φj−1 + (1 − ω)φj + φj+1 − fj (9.33)
2 2 2
The General Form

The general form of the classical methods is obtained by splitting the matrix A in
Eq. 9.13 into its diagonal, D, the portion of the matrix below the diagonal, L, and
the portion above the diagonal, U , such that
A=L+D+U (9.34)
Then the point-Jacobi method is obtained with H = −D, which certainly meets
the criterion that it is easy to solve. The Gauss-Seidel method is obtained with
H = −(L + D), which is also easy to solve, being lower triangular.
9.3 The ODE Approach to Classical Relaxation

9.3.1 The Ordinary Differential Equation Formulation
The particular type of delta form given by Eq. 9.21 leads to an interpretation of
relaxation methods in terms of solution techniques for coupled first-order ODE’s,
about which we have already learned a great deal. One can easily see that Eq. 9.21
results from the application of the explicit Euler time-marching method (with h = 1)
to the following system of ODE’s:
dφ
H = Aφ − f (9.35)
dt
This is equivalent to

dφ
= H −1 C Abφ − f b = H −1 [Aφ − f] (9.36)
dt
9.3. THE ODE APPROACH TO CLASSICAL RELAXATION 171
In the special case where H−1 A depends on neither u nor t, H −1f is also independent
of t, and the eigenvectors of H −1 A are linearly independent, the solution can be
written as
φ = c1 eλ1 tx1 + · · · + cM eλM txM +φ∞ (9.37)

error
where what is referred to in time-accurate analysis as the transient solution, is now

referred to in relaxation analysis as the error. It is clear that, if all of the eigenvalues
of H −1 A have negative real parts (which implies that H −1 A is nonsingular), then the
system of ODE’s has a steady-state solution which is approached as t → ∞, given by
φ∞ = A−1f (9.38)
which is the solution of Eq. 9.13. We see that the goal of a relaxation method is to
remove the transient solution from the general solution in the most efficient way pos-
sible. The λ eigenvalues are fixed by the basic matrix in Eq. 9.36, the preconditioning
matrix in 9.7, and the secondary conditioning matrix in 9.35. The σ eigenvalues are
fixed for a given λh by the choice of time-marching method. Throughout the remain-
ing discussion we will refer to the independent variable t as “time”, even though no
true time accuracy is involved.
In a stationary method, H and C in Eq. 9.36 are independent of t, that is, they
are not changed throughout the iteration process. The generalization of this in our
approach is to make h, the “time” step, a constant for the entire iteration.
Suppose the explicit Euler method is used for the time integration. For this method
σm = 1 + λm h. Hence the numerical solution after n steps of a stationary relaxation
method can be expressed as (see Eq. 6.28)
φn = c1x1 (1 + λ1 h)n + · · · + cmxm (1 + λm h)n + · · · + cM xM (1 + λM h)n +φ∞ (9.39)

error
The initial amplitudes of the eigenvectors are given by the magnitudes of the cm .
These are fixed by the initial guess. In general it is assumed that any or all of the
eigenvectors could have been given an equally “bad” excitation by the initial guess,
so that we must devise a way to remove them all from the general solution on an
equal basis. Assuming that H −1 A has been chosen (that is, an iteration process has
been decided upon), the only free choice remaining to accelerate the removal of the
error terms is the choice of h. As we shall see, the three classical methods have all
been conditioned by the choice of H to have an optimum h equal to 1 for a stationary
iteration process.
9.3.2 ODE Form of the Classical Methods

The three iterative procedures defined by Eqs. 9.30, 9.31 and 9.32 obey no apparent
pattern except that they are easy to implement in a computer code since all of the
data required to update the value of one point are explicitly available at the time
of the update. Now let us study these methods as subsets of ODE as formulated in
Section 9.3.1. Insert the model equation 9.29 into the ODE form 9.35. Then
dφ
H = B(1, −2, 1)φ − f (9.40)
dt
As a start, let us use for the numerical integration the explicit Euler method
φn+1 = φn + hφn (9.41)
with a step size, h, equal to 1. We arrive at
H(φn+1 − φn ) = B(1, −2, 1)φn − f (9.42)
It is clear that the best choice of H from the point of view of matrix algebra is
−B(1, −2, 1) since then multiplication from the left by −B−1 (1, −2, 1) gives the cor-
rect answer in one step. However, this is not in the spirit of our study, since multi-
plication by the inverse amounts to solving the problem by a direct method without
iteration. The constraint on H that is in keeping with the formulation of the three
methods described in Section 9.2.3 is that all the elements above the diagonal (or
below the diagonal if the sweeps are from right to left) are zero. If we impose this
constraint and further restrict ourselves to banded tridiagonals with a single constant
in each band, we are led to
2
B(−β, , 0)(φn+1 − φn ) = B(1, −2, 1)φn − f (9.43)
ω
where β and ω are arbitrary. With this choice of notation the three methods presented
in Section 9.2.3 can be identified using the entries in Table 9.1.
TABLE 9.1: VALUES OF β and ω IN EQ. 9.43 THAT

LEAD TO CLASSICAL RELAXATION METHODS
β ω Method Equation
0 1 Point-Jacobi 6.2.3
1 1 Gauss-Seidel 6.2.4
1 2/ 1 + sin M π+ 1 Optimum SOR 6.2.5
9.4. EIGENSYSTEMS OF THE CLASSICAL METHODS 173
The fact that the values in the tables lead to the methods indicated can be verified
by simple algebraic manipulation. However, our purpose is to examine the whole
procedure as a special subset of the theory of ordinary differential equations. In this
light, the three methods are all contained in the following set of ODE’s

dφ 2
= B −1 (−β, , 0) B(1, −2, 1)φ − f (9.44)
dt ω
and appear from it in the special case when the explicit Euler method is used for its
numerical integration. The point operator that results from the use of the explicit
Euler scheme is

ωβ (n+1) ω ωh (n) ωh
(n+1) (n) (n)
φj = φj−1 + (h − β)φj−1 − (ωh − 1)φj + φj+1 − fj(9.45)
2 2 2 2
This represents a generalization of the classical relaxation techniques.
9.4 Eigensystems of the Classical Methods

The ODE approach to relaxation can be summarized as follows. The basic equation
to be solved came from some time-accurate derivation
Abu − f b = 0 (9.46)
This equation is preconditioned in some manner which has the effect of multiplication
by a conditioning matrix C giving
Aφ − f = 0 (9.47)
An iterative scheme is developed to find the converged, or steady-state, solution of

the set of ODE’s
dφ
H = Aφ − f (9.48)
dt
This solution has the analytical form
φn = en + φ∞ (9.49)
where en is the transient, or error, and φ∞ ≡ A−1f is the steady-state solution. The
three classical methods, Point-Jacobi, Gauss-Seidel, and SOR, are identified for the
one-dimensional case by Eq. 9.44 and Table 9.1.
Given our assumption that the component of the error associated with each eigen-
vector is equally likely to be excited, the asymptotic convergence rate is determined
by the eigenvalue σm of G (≡ I + H −1 A) having maximum absolute value. Thus
Convergence rate ∼ |σm |max , m = 1, 2, · · · , M (9.50)
In this section, we use the ODE analysis to find the convergence rates of the three
classical methods represented by Eqs. 9.30, 9.31, and 9.32. It is also instructive to
inspect the eigenvectors and eigenvalues in the H−1 A matrix for the three methods.
This amounts to solving the generalized eigenvalue problem
Axm = λm Hxm (9.51)
for the special case
2
B(1, −2, 1)xm = λm B(−β, , 0)xm (9.52)
ω
The generalized eigensystem for simple tridigonals is given in Appendix B.2. The
three special cases considered below are obtained with a = 1, b = −2, c = 1, d = −β,
e = 2/ω, and f = 0. To illustrate the behavior, we take M = 5 for the matrix order.
This special case makes the general result quite clear.
9.4.1 The Point-Jacobi System

If β = 0 and ω = 1 in Eq. 9.44, the ODE matrix H −1 A reduces to simply B( 12 , −1, 12 ).
The eigensystem can be determined from Appendix B.1 since both d and f are zero.
The eigenvalues are given by the equation

mπ
λm = −1 + cos , m = 1, 2, . . . , M (9.53)
M +1
The λ-σ relation for the explicit Euler method is σm = 1 + λm h. This relation can
be plotted for any h. The plot for h = 1, the optimum stationary case, is shown in
Fig. 9.1. For h < 1, the maximum |σm | is obtained with m = 1, Fig. 9.2 and for
h > 1, the maximum |σm | is obtained with m = M , Fig. 9.3. Note that for h > 1.0
(depending on M ) there is the possibility of instability, i.e. |σm | > 1.0. To obtain the
optimal scheme we wish to minimize the maximum |σm | which occurs when h = 1,
|σ1 | = |σM | and the best possible convergence rate is achieved:

π
|σm |max = cos (9.54)
M +1
For M = 40, we obtain |σm |max = 0.9971. Thus after 500 iterations the error content
associated with each eigenvector is reduced to no more than 0.23 times its initial level.
Again from Appendix B.1, the eigenvectors of H−1 A are given by

xj = (xj )m = sin j mπ
, j = 1, 2, . . . , M (9.55)
M +1
This is a very “well-behaved” eigensystem with linearly independent eigenvectors and
distinct eigenvalues. The first 5 eigenvectors are simple sine waves. For M = 5, the
eigenvectors can be written as
  √   
 √1/2   √
3/2
 
1

 3/2   3/2   0 
     
  ,   ,  
 x3 =  −1  ,
x1 = 
 √1  x2 =  √0
    
 3/2   − 3/2   0 
√
1/2 − 3/2 1
 √   
3/2 1/2
 √   √ 
 − 3/2   − 3/2 
   
x4 =   ,  
 √0  x5 =  √1  (9.56)
   
 3/2   − 3/2 
√
− 3/2 1/2
The corresponding eigenvalues are, from Eq. 9.53

√
λ1 = −1 + 23 = −0.134 · · ·
λ2 = −1 + 12 = −0.5
λ3 = −1 = −1.0 (9.57)
λ4 = −1 − 12 = −1.5
√
λ5 = −1 − 23 = −1.866 · · ·
From Eq. 9.39, the numerical solution written in full is

√
φn − φ∞ = c1 [1 − (1 − 3 )h]nx1
2
1 n
+ c2 [1 − (1 − )h] x2
2
1.0
h = 1.0
m=5 m=1
Values are equal
σm
m=4 m=2
Μ=5
m=3
0.0
-2.0 -1.0 0.0
Figure 9.1: The σ, λ relation for Point-Jacobi, h = 1, M = 5.
+ c3 [1 − (1 )h]nx3
1
+ c4 [1 − (1 + )h]nx4
2√
3 n
+ c5 [1 − (1 + )h] x5 (9.58)
2
9.4.2 The Gauss-Seidel System

If β and ω are equal to 1 in Eq. 9.44, the matrix eigensystem evolves from the relation
B(1, −2, 1)xm = λm B(−1, 2, 0)xm (9.59)
which can be studied using the results in Appendix B.2. One can show that the H−1 A
matrix for the Gauss-Seidel method, AGS , is
AGS ≡ B −1 (−1, 2, 0)B(1, −2, 1) =

1.0 Approaches 1.0
h < 1.0
h
σm
De
ng
cr
si
ea
ea
si
cr
ng
De
h
0.0
-2.0 -1.0 0.0
λ mh
Figure 9.2: The σ, λ relation for Point-Jacobi, h = 0.9, M = 5.
Exceeds 1.0 at h ∼
∼ 1.072 / Ustable h > 1.072
1.0
h > 1.0
In
cr
ea
h
si
σm
ng
ng
si
ea
h
cr
In
M = 5
0.0
-2.0 -1.0 0.0
λmh
Figure 9.3: The σ, λ relation for Point-Jacobi, h = 1.1, M = 5.

 
−1 1/2 1
 0 −3/4 1/2  2
 
 


0 1/8 −3/4 1/2 

3
 0 1/16 1/8 −3/4 1/2  4
  (9.60)
 
 0 1/32 1/16 1/8 −3/4 1/2  5
 
 .. .. ... ... ... ... ...  ..
 . .  .
0 1/2M · · · M
The eigenvector structure of the Gauss-Seidel ODE matrix is quite interesting. If M
is odd there are (M + 1)/2 distinct eigenvalues with corresponding linearly indepen-
dent eigenvectors, and there are (M − 1)/2 defective eigenvalues with corresponding
principal vectors. The equation for the nondefective eigenvalues in the ODE matrix
is (for odd M )
mπ M +1
λm = −1 + cos2 ( ) , m = 1, 2, . . . , (9.61)
M +1 2
and the corresponding eigenvectors are given by
j−1
xm = cos mπ mπ M +1
sin j , m = 1, 2, . . . , (9.62)
M +1 M +1 2
The λ-σ relation for h = 1, the optimum stationary case, is shown in Fig. 9.4. The
σm with the largest amplitude is obtained with m = 1. Hence the convergence rate is
2
π
|σm |max = cos (9.63)
M +1
Since this is the square of that obtained for the Point-Jacobi method, the error associ-
ated with the “worst” eigenvector is removed in half as many iterations. For M = 40,
|σm |max = 0.9942. 250 iterations are required to reduce the error component of the
worst eigenvector by a factor of roughly 0.23.
The eigenvectors are quite unlike the Point-Jacobi set. They are no longer sym-
metrical, producing waves that are higher in amplitude on one side (the updated side)
than they are on the other. Furthermore, they do not represent a common family for
different values of M .
The Jordan canonical form for M = 5 is
 
φ1
 
 φ2 
 
   
X −1 AGS X = JGS = 
 φ3 1 
 (9.64)
   
  φ3 1  
   
φ3
1.0
h = 1.0
σm
Μ=5
2 defective λ m
0.0
-2.0 -1.0 0.0
λmh
Figure 9.4: The σ, λ relation for Gauss-Seidel, h = 1.0, M = 5.
The eigenvectors and principal vectors are all real. For M = 5 they can be written
   √       
1/2 √ 3/2 1 0 0
         
 3/4   3/4   0   2   0 
         
x1 =  3/4  , x2 =   ,      
   √0  x3 =  0  , x4 =  −1  , x5 =  4  (9.65)
         
 9/16   −√3/16   0   0   −4 
9/32 − 3/32 0 0 1
The corresponding eigenvalues are
λ1 = −1/4
λ2 = −3/4
λ3 = −1

(4) Defective, linked to (9.66)
(5) λ3 Jordan block
The numerical solution written in full is thus
φn − φ∞ = c1 (1 − h )nx1

4
3h n
+ c2 (1 − ) x2
4

n n(n − 1)
+ c3 (1 − h) + c4 h (1 − h)n−1 + c5 h2
n
(1 − h)n−2 x3
1! 2!

n
+ c4 (1 − h) + c5 h (1 − h)
n n−1
x4
1!
+ c5 (1 − h)nx5 (9.67)
9.4.3 The SOR System

If β = 1 and 2/ω = x in Eq. 9.44, the ODE matrix is B−1 (−1, x, 0)B(1, −2, 1).
One can show that this can be written in the form given below for M = 5. The
generalization to any M is fairly clear. The H −1 A matrix for the SOR method,
ASOR ≡ B −1 (−1, x, 0)B(1, −2, 1), is
 
−2x4 x4 0 0 0
 


−2x3 + x4 x3 − 2x4 x4 0 0 

1  
 −2x2 + x3 x2 − 2x3 + x4 x3 − 2x4 x4 0  (9.68)
x5 
−2x + x 2
x − 2x + x x − 2x + x
2 3 2 3 4
x − 2x4
3
x4 
 
−2 + x 1 − 2x + x2 x − 2x2 + x3 x2 − 2x3 + x4 x3 − 2x4
Eigenvalues of the system are given by

2
ωpm + zm
λm = −1 + , m = 1, 2, . . . M (9.69)
2
where
zm = [4(1 − ω) + ω2 p2m ]1/2

pm = cos[mπ/(M + 1)]
If ω = 1, the system is Gauss-Seidel. If 4(1 − ω) + ω2 pm < 0, zm and λm are complex.

If ω is chosen such that 4(1 − ω) + ω2p21 = 0, ω is optimum for the stationary case,
and the following conditions hold:
1. Two eigenvalues are real, equal and defective.
2. If M is even, the remaining eigenvalues are complex and occur in conjugate

pairs.
3. If M is odd, one of the remaining eigenvalues is real and the others are complex
occurring in conjugate pairs.
One can easily show that the optimum ω for the stationary case is

π
ωopt = 2/ 1 + sin (9.70)
M +1
and for ω = ωopt
2
λm = ζm −1

xm = ζ j−1 sin j mπ
m (9.71)
M +1
where

ωopt
ζm = pm + i p21 − p2m
2
Using the explicit Euler method to integrate the ODE’s, σm = 1 − h + hζm 2
, and if
h = 1, the optimum value for the stationary case, the λ-σ relation reduces to that
shown in Fig. 9.5. This illustrates the fact that for optimum stationary SOR all the
|σm | are identical and equal to ωopt − 1. Hence the convergence rate is
|σm |max = ωopt − 1 (9.72)

π
ωopt = 2/ 1 + sin
M +1
For M = 40, |σm |max = 0.8578. Hence the worst error component is reduced to less
than 0.23 times its initial value in only 10 iterations, much faster than both Gauss-
Seidel and Point-Jacobi. In practical applications, the optimum value of ω may have
to be determined by trial and error, and the benefit may not be as great.
For odd M , there are two real eigenvectors and one real principal vector. The
remaining linearly independent eigenvectors are all complex. For M = 5 they can be
written
     √   
1/2 −6 3(1 )/2 1
     √ √   
 1/2 
 
 9 
 


3(1 ± i 2)/6 
 0 
 
x1 =     
 1/3  , x2 =  16  , x3,4 =  √ 0√  ,  
 x5 =  1/3  (9.73)
       
 1/6   13   √ 3(5 ± i√ 2)/54   0 
1/18 6 3(7 ± 4i 2)/162 1/9
The corresponding eigenvalues are
λ1 = −2/3
(2) Defective linked to λ1
√
λ3 = −(10 − 2 2i)/9
√
λ4 = −(10 + 2 2i)/9
λ5 = −4/3 (9.74)
1.0
h = 1.0
2 real defective λ m
σm 2 complex λ m
1 real λ m
Μ=5
0.0
-2.0 -1.0 0.0
λmh
Figure 9.5: The σ, λ relation for optimum stationary SOR, M = 5, h = 1.
The numerical solution written in full is
φn − φ∞ = [c1 (1 − 2h/3)n + c2 nh(1 − 2h/3)n−1 ]x1

+ c2 (1 − 2h/3)nx2
√
+ c3 [1 − (10 − 2 2i)h/9]nx3
√
+ c4 [1 − (10 + 2 2i)h/9]nx4
+ c5 (1 − 4h/3)nx5 (9.75)
9.5 Nonstationary Processes

In classical terminology a method is said to be nonstationary if the conditioning
matrices, H and C, are varied at each time step. This does not change the steady-
state solution A−1
b f b , but it can greatly affect the convergence rate. In our ODE
approach this could also be considered and would lead to a study of equations with
nonconstant coefficients. It is much simpler, however, to study the case of fixed
H and C but variable step size, h. This process changes the Point-Jacobi method
to Richardson’s method in standard terminology. For the Gauss-Seidel and SOR
methods it leads to processes that can be superior to the stationary methods.
The nonstationary form of Eq. 9.39 is
9.5. NONSTATIONARY PROCESSES 183
/
N /
N
φ = c1x1 (1 + λ1 hn ) + · · · + cmxm (1 + λm hn )
N
n=1 n=1
/
N
+ · · · + cM xM (1 + λM hn ) + φ∞ (9.76)
n=1
where the symbol Π stands for product. Since hn can now be changed at each step,
the error term can theoretically be completely eliminated in M steps by taking hm =
−1/λm , for m = 1, 2, · · · , M . However, the eigenvalues λm are generally unknown and
costly to compute. It is therefore unnecessary and impractical to set hm = −1/λm
for m = 1, 2, . . . , M . We will see that a few well chosen h’s can reduce whole clusters
of eigenvectors associated with nearby λ’s in the λm spectrum. This leads to the
concept of selectively annihilating clusters of eigenvectors from the error terms as
part of a total iteration process. This is the basis for the multigrid methods discussed
in Chapter 10.
Let us consider the very important case when all of the λm are real and nega-
tive (remember that they arise from a conditioned matrix so this constraint is not
unrealistic for quite practical cases). Consider one of the error terms taken from

M /
N
eN ≡ φN − φ∞ = cmxm (1 + λm hn ) (9.77)
m=1 n=1
and write it in the form

/
N
cmxm Pe (λm ) ≡ cmxm (1 + λm hn ) (9.78)
n=1
where Pe signifies an “Euler” polynomial. Now focus attention on the polynomial
(Pe )N (λ) = (1 + h1 λ)(1 + h2 λ) · · · (1 + hN λ) (9.79)
treating it as a continuous function of the independent variable λ. In the annihilation

process mentioned after Eq. 9.76, we considered making the error exactly zero by
taking advantage of some knowledge about the discrete values of λm for a particular
case. Now we pose a less demanding problem. Let us choose the hn so that the
maximum value of (Pe )N (λ) is as small as possible for all λ lying between λa and λb
such that λb ≤ λ ≤ λa ≤ 0. Mathematically stated, we seek
max |(Pe )N (λ)| = minimum , with(Pe )N (0) = 1 (9.80)

λb ≤λ≤λa
This problem has a well known solution due to Markov. It is

2λ − λa − λb
TN
λa − λb
(Pe )N (λ) = (9.81)
−λa − λb
TN
λa − λb
where
TN (y) = cos(N arccos y) (9.82)
are the Chebyshev polynomials along the interval −1 ≤ y ≤ 1 and

N N
1 1
TN (y) = y + y2 − 1 + y − y2 − 1 (9.83)
2 2
are the Chebyshev polynomials for |y| > 1. In relaxation terminology this is generally
referred to as Richardson’s method, and it leads to the nonstationary step size choice
given by

1 1 (2n − 1)π
= −λb − λa + (λb − λa ) cos , n = 1, 2, . . . N (9.84)
hn 2 2N
Remember that all λ are negative real numbers representing the magnitudes of λm in
an eigenvalue spectrum.
The error in the relaxation process represented by Eq. 9.76 is expressed in terms
0
of a set of eigenvectors, xm , amplified by the coefficients cm (1 + λm hn ). With each
eigenvector there is a corresponding eigenvalue. Eq. 9.84 gives us the best choice of a
series of hn that will minimize the amplitude of the error carried in the eigenvectors
associated with the eigenvalues between λb and λa .
As an example for the use of Eq. 9.84, let us consider the following problem:
Minimize the maximum error asso-

ciated with the λ eigenvalues in the (9.85)
interval −2 ≤ λ ≤ −1 using only 3
iterations.
The three values of h which satisfy this problem are

(2n − 1)π
hn = 2/ 3 − cos (9.86)
6
9.5. NONSTATIONARY PROCESSES 185
and the amplitude of the eigenvector is reduced to
(Pe )3 (λ) = T3 (2λ + 3)/T3 (3) (9.87)
where
! √ √ "
T3 (3) = [3 + 8]3 + [3 − 8]3 /2 ≈ 99 (9.88)
A plot of Eq. 9.87 is given in Fig. 9.6 and we see that the amplitudes of all the
eigenvectors associated with the eigenvalues in the range −2 ≤ λ ≤ −1 have been
reduced to less than about 1% of their initial values. The values of h used in Fig. 9.6
are
√
h1 = 4/(6 − 3)
h2 = 4/(6 − 0)
√
h3 = 4/(6 + 3)
Return now to Eq. 9.76. This was derived from Eq. 9.37 on the condition that the
explicit Euler method, Eq. 9.41, was used to integrate the basic ODE’s. If instead
the implicit trapezoidal rule
1
φn+1 = φn + h(φn+1 + φn ) (9.89)
2
is used, the nonstationary formula
 
1

M /
N 1 + h λ
n m
φN = cmxm  2 +
 1  φ∞ (9.90)
m=1 n=1 1 − h λ
n m
2
would result. This calls for a study of the rational “trapezoidal” polynomial, Pt :
 
1
/N  1 + hn λ 
(Pt )N (λ) =  2  (9.91)
 1 
n=1 1 − h λ
n
2
under the same constraints as before, namely that
max |(Pt )N (λ)| = minimum , (9.92)

λb ≤λ≤λa
with (Pt )N (0) = 1
0.9
0.8
0.7
0.6
(Pe )3 (λ)
0.5
0.4
0.3
0.2
0.1
−2 −1.8 −1.6 −1.4 −1.2 −1 −0.8 −0.6 −0.4 −0.2 0

λ
0.02
0.015
0.01
0.005
(P ) (λ)
e 3
−0.005
−0.01
−0.015
−0.02
−2 −1.8 −1.6 −1.4 −1.2 −1 −0.8 −0.6 −0.4 −0.2 0
λ
Figure 9.6: Richardson’s method for 3 steps, minimization over −2 ≤ λ ≤ −1.

9.6. PROBLEMS 187
The optimum values of h can also be found for this problem, but we settle here for
the approximation suggested by Wachspress
(n−1)/(N −1)
2 λa
= −λb , n = 1, 2, · · · , N (9.93)
hn λb
This process is also applied to problem 9.85. The results for (Pt )3 (λ) are shown in
Fig. 9.7. The error amplitude is about 1/5 of that found for (Pe )3 (λ) in the same
interval of λ. The values of h used in Fig. 9.7 are
h1 = 1
√
h2 = 2
h3 = 2
9.6 Problems
1. Given a relaxation method in the form
n = Aφ
H∆φ n − f
show that
0 + (I − Gn )A−1 f
n = Gn φ
φ
where G = I + H −1 A.
2. For a linear system of the form (A1 + A2 )x = b, consider the iterative method
(I + µA1 )x̃ = (I − µA2 )xn + µb

(I + µA2 )xn+1 = (I − µA1 )x̃ + µb
where µ is a parameter. Show that this iterative method can be written in the
form
H(xk+1 − xk ) = (A1 + A2 )xk − b
Determine the iteration matrix G if µ = −1/2.

3. Using Appendix B.2, find the eigenvalues of H−1 A for the SOR method with
A = B(4 : 1, −2, 1) and ω = ωopt . (You do not have to find H −1 . Recall that
the eigenvalues of H −1 A satisfy Axm = λm Hxm .) Find the numerical values,
not just the expressions. Then find the corresponding |σm | values.
0.9
0.8
0.7
0.6
(Pt )3 (λ)
0.5
0.4
0.3
0.2
0.1
−2 −1.8 −1.6 −1.4 −1.2 −1 −0.8 −0.6 −0.4 −0.2 0

λ
−3
x 10
2
1.5
0.5
(Pt )3 (λ)
−0.5
−1
−1.5
−2
−2 −1.8 −1.6 −1.4 −1.2 −1 −0.8 −0.6 −0.4 −0.2 0
λ
Figure 9.7: Wachspress method for 3 steps, minimization over −2 ≤ λ ≤ −1.

9.6. PROBLEMS 189
4. Solve the following equation on the domain 0 ≤ x ≤ 1 with boundary conditions

u(0) = 0, u(1) = 1:
∂2u
− 6x = 0
∂x2
For the initial condition, use u(x) = 0. Use second-order centered differences
on a grid with 40 cells (M = 39). Iterate to steady state using
(a) the point-Jacobi method,

(b) the Gauss-Seidel method,
(c) the SOR method with the optimum value of ω, and
(d) the 3-step Richardson method derived in Section 9.5.
Plot the solution after the residual is reduced by 2, 3, and 4 orders of mag-
nitude. Plot the logarithm of the L2 -norm of the residual vs. the number of
iterations. Determine the asymptotic convergence rate. Compare with the the-
oretical asymptotic convergence rate.
Chapter 10
MULTIGRID
The idea of systematically using sets of coarser grids to accelerate the convergence of
iterative schemes that arise from the numerical solution to partial differential equa-
tions was made popular by the work of Brandt. There are many variations of the
process and many viewpoints of the underlying theory. The viewpoint presented here
is a natural extension of the concepts discussed in Chapter 9.
10.1 Motivation
10.1.1 Eigenvector and Eigenvalue Identification with Space
Frequencies
Consider the eigensystem of the model matrix B(1, −2, 1). The eigenvalues and
eigenvectors are given in Sections 4.3.2 and 4.3.3, respectively. Notice that as the
magnitudes of the eigenvalues increase, the space-frequency (or wavenumber) of the
corresponding eigenvectors also increase. That is, if the eigenvalues are ordered such
that
|λ1 | ≤ |λ2 | ≤ · · · ≤ |λM | (10.1)
then the corresponding eigenvectors are ordered from low to high space frequencies.
This has a rational explanation from the origin of the banded matrix. Note that
∂2
sin(mx) = −m2 sin(mx) (10.2)
∂x2
and recall that

1 1
δxxφ =
2 B(1, −2, 1)φ = X D(λ) X −1φ (10.3)
∆x ∆x2
191
192 CHAPTER 10. MULTIGRID
where D(λ) is a diagonal matrix containing the eigenvalues. We have seen that
X −1φ represents a sine transform, and Xφ, a sine synthesis. Therefore, the opera-
tion 1 2 D(λ) represents the numerical approximation of the multiplication of the
∆x
appropriate sine wave by the negative square of its wavenumber, −m2 . One finds
that
2
1 M +1 mπ
2 λm = −2 + 2 cos ≈ −m2 , m << M (10.4)
∆x π M +1
Hence, the correlation of large magnitudes of λm with high space-frequencies is to be
expected for these particular matrix operators. This is consistent with the physics of
diffusion as well. However, this correlation is not necessary in general. In fact, the
complete counterexample of the above association is contained in the eigensystem
for B( 12 , 1, 21 ). For this matrix one finds, from Appendix B, exactly the opposite
behavior.
10.1.2 Properties of the Iterative Method

The second key motivation for multigrid is the following:
• Many iterative methods reduce error components corresponding to eigenvalues

of large amplitude more effectively than those corresponding to eigenvalues of
small amplitude.
This is to be expected of an iterative method which is time accurate. It is also

true, for example, of the Gauss-Seidel method and, by design, of the Richardson
method described in Section 9.5. The classical point-Jacobi method does not share
this property. As we saw in Section 9.4.1, this method produces the same value of |σ|
for λmin and λmax . However, the property can be restored by using h < 1, as shown
in Fig. 9.2.
When an iterative method with this property is applied to a matrix with the
above correlation between the modulus of the eigenvalues and the space frequency of
the eigenvectors, error components corresponding to high space frequencies will be
reduced more quickly than those corresponding to low space frequencies. This is the
key concept underlying the multigrid process.
10.2 The Basic Process

First of all we assume that the difference equations representing the basic partial
differential equations are in a form that can be related to a matrix which has certain
10.2. THE BASIC PROCESS 193
basic properties. This form can be arrived at “naturally” by simply replacing the
derivatives in the PDE with difference schemes, as in the example given by Eq. 3.27,
or it can be “contrived” by further conditioning, as in the examples given by Eq. 9.11.
The basic assumptions required for our description of the multigrid process are:
1. The problem is linear.
2. The eigenvalues, λm , of the matrix are all real and negative.
3. The λm are fairly evenly distributed between their maximum and minimum
values.
4. The eigenvectors associated with the eigenvalues having largest magnitudes can
be correlated with high frequencies on the differencing mesh.
5. The iterative procedure used greatly reduces the amplitudes of the eigenvectors
associated with eigenvalues in the range between 21 |λ|max and |λ|max.
These conditions are sufficient to ensure the validity of the process described next.
Having preconditioned (if necessary) the basic finite differencing scheme by a pro-
cedure equivalent to the multiplication by a matrix C, we are led to the starting
formulation
C[Abφ∞ − f b ] = 0 (10.5)
where the matrix formed by the product CAb has the properties given above. In Eq.
10.5, the vector f b represents the boundary conditions and the forcing function, if
any, and φ∞ is a vector representing the desired exact solution. We start with some
initial guess for φ∞ and proceed through n iterations making use of some iterative
process that satisfies property 5 above. We do not attempt to develop an optimum
procedure here, but for clarity we suppose that the three-step Richardson method
illustrated in Fig. 9.6 is used. At the end of the three steps we find r, the residual,
where
r = C[Abφ − f b ] (10.6)
Recall that the φ used to compute r is composed of the exact solution φ∞ and the
error e in such a way that
Ae − r = 0 (10.7)
where
A ≡ CAb (10.8)
If one could solve Eq. 10.7 for e then

φ = φ − e (10.9)
∞
Thus our goal now is to solve for e. We can write the exact solution for e in terms of
the eigenvectors of A, and the σ eigenvalues of the Richardson process in the form:
M/2
/
3
M /
3
e = cmxm [σ(λm hn )] + cmxm [σ(λm hn )] (10.10)
m=1 n=1 m=M/2+1 n=1

very low amplitude
Combining our basic assumptions, we can be sure that the high frequency content of
e has been greatly reduced (about 1% or less of its original value in the initial guess).
In addition, assumption 4 ensures that the error has been smoothed.
Next we construct a permutation matrix which separates a vector into two parts,
one containing the odd entries, and the other the even entries of the original vector
(or any other appropriate sorting which is consistent with the interpolation approxi-
mation to be discussed below). For a 7-point example
    
e2 0 1 0 0 0 0 0 e1
    
 e4   0 0 0 1 0 0 0  e2 
    
 e6   0 0 0 0 0 1 0  e3 
     ee
    
 e1  =  1 0 0 0 0 0 0  e4  ; = P e (10.11)
     eo
 e3   0 0 1 0 0 0 0  e5 
    
 e5   0 0 0 0 1 0 0  e6 
    
e7 0 0 0 0 0 0 1 e7
Multiply Eq. 10.7 from the left by P and, since a permutation matrix has an inverse
which is its transpose, we can write
P A[P −1P ]e = P r (10.12)
The operation P AP −1 partitions the A matrix to form

 
A1 A2
  ee re
  = (10.13)
eo ro
A3 A4
Notice that
A1ee + A2eo = re (10.14)

is an exact expression.
At this point we make our one crucial assumption. It is that there is some connec-
tion between ee and eo brought about by the smoothing property of the Richardson
relaxation procedure. Since the top half of the frequency spectrum has been removed,
it is reasonable to suppose that the odd points are the average of the even points.
For example
1
e1 ≈ (ea + e2 )
2
1
e3 ≈ (e2 + e4 )
2
1
e5 ≈ (e4 + e6 ) or eo = A2ee (10.15)
2
1
e7 ≈ (e6 + eb )
2
It is important to notice that ea and eb represent errors on the boundaries where the
error is zero if the boundary conditions are given. It is also important to notice that we
are dealing with the relation between e and r so the original boundary conditions and
forcing function (which are contained in f in the basic formulation) no longer appear
in the problem. Hence, no aliasing of these functions can occur in subsequent steps.
Finally, notice that, in this formulation, the averaging of e is our only approximation,
no operations on r are required or justified.
If the boundary conditions are Dirichlet, ea and eb are zero, and one can write for
the example case
 
1 0 0
1
 1 1 0 

A2 =   (10.16)
2 0 1 1 
0 0 1
With this approximation Eq. 10.14 reduces to
A1ee + A2 A2ee = re (10.17)
or
Acee − re = 0 (10.18)
where
Ac = [A1 + A2 A2 ] (10.19)

The form of Ac , the matrix on the coarse mesh, is completely determined by the
choice of the permutation matrix and the interpolation approximation. If the original
A had been B(7 : 1, −2, 1), our 7-point example would produce
 
−2 1 1
 
 −2 1 1 
 
 −2 1 1   
  A1 A2
 
   
P AP −1 =   =  (10.20)

 1 −2 

  A3 A4
 1 1 −2 
 
 1 1 −2 
1 −2
and Eq. 10.18 gives
A2
A1 A2 Ac
  
 1    
−2 1 1   −1 1/2
    1 1 1   
 −2 + 1 1 ·   =  1/2 −1 1/2  (10.21)
2 1 1 
−2 1 1 1/2 −1
1
If the boundary conditions are mixed Dirichlet-Neumann, A in the 1-D model

equation is B(1, b, 1) where b = [−2, −2, ..., −2, −1]T . The eigensystem is given by
Eq. B.19. It is easy to show that the high space-frequencies still correspond to the
eigenvalues with high magnitudes, and, in fact, all of the properties given in Section
10.1 are met. However, the eigenvector structure is different from that given in
Eq. 9.55 for Dirichlet conditions. In the present case they are given by

(2m − 1)π
xjm = sin j ; m = 1, 2, · · · , M (10.22)
2M + 1
and are illustrated in Fig. 10.1. All of them go through zero on the left (Dirichlet)
side, and all of them reflect on the right (Neumann) side.
For Neumann conditions, the interpolation formula in Eq. 10.15 must be changed.
In the particular case illustrated in Fig. 10.1, eb is equal to eM . If Neumann conditions
are on the left, ea = e1 . When eb = eM , the example in Eq. 10.16 changes to
 
1 0 0
1
 1 1 0 

A2 =   (10.23)
2 0 1 1 
0 0 2
Figure 10.1: Eigenvectors for the mixed Dirichlet–Neumann case.
The permutation matrix remains the same and both A1 and A2 in the partitioned
matrix P AP −1 are unchanged (only A4 is modified by putting −1 in the lower right
element). Therefore, we can construct the coarse matrix from
A2
A1 A2 Ac
  
 
1   
−2 1 1  1 1  −1 1/2
    1  
 −2 + 1 1 ·   =
 1/2 −1 1/2 
 (10.24)
2 1 1 
−2 1 1 1/2 −1/2
2
which gives us what we might have “expected.”
We will continue with Dirichlet boundary conditions for the remainder of this
Section. At this stage, we have reduced the problem from B(1, −2, 1)e = r on the
fine mesh to 12 B(1, −2, 1)ee = re on the next coarser mesh. Recall that our goal is
to solve for e, which will provide us with the solution φ∞ using Eq. 10.9. Given ee
computed on the coarse grid (possibly using even coarser grids), we can compute eo
using Eq. 10.15, and thus e. In order to complete the process, we must now determine
the relationship between ee and e.
In order to examine this relationship, we need to consider the eigensystems of A
and Ac :
A = XΛX −1, Ac = Xc Λc Xc−1 (10.25)
For A = B(M : 1, −2, 1) the eigenvalues and eigenvectors are

mπ mπ j = 1, 2, · · · , M
λm = −2 1 − cos , xm = sin j , (10.26)
M +1 M +1 m = 1, 2, · · · , M
Based on our assumptions, the most difficult error mode to eliminate is that with
m = 1, corresponding to

π π
λ1 = −2 1 − cos , x1 = sin j , j = 1, 2, · · · , M (10.27)
M +1 M +1
For example, with M = 51, λ1 = −0.003649. If we restrict our attention to odd M ,
then Mc = (M − 1)/2 is the size of Ac . The eigenvalue and eigenvector corresponding
to m = 1 for the matrix Ac = 12 B(Mc , 1, −2, 1) are

2π 2π
(λc )1 = − 1 − cos , (xc )1 = sin j , j = 1, 2, · · · , Mc (10.28)
M +1 M +1
For M = 51 (Mc = 25), we obtain (λc )1 = −0.007291 = 1.998λ1 . As M increases,
(λc )1 approaches 2λ1 . In addition, one can easily see that (xc )1 coincides with x1 at
every second point of the latter vector, that is, it contains the even elements of x1 .
Now let us consider the case in which all of the error consists of the eigenvector
component x1 , i.e., e = x1 . Then the residual is
r = Ax1 = λ1x1 (10.29)
and the residual on the coarse grid is

re = λ1 (xc )1 (10.30)
since (xc )1 contains the even elements of x1 . The exact solution on the coarse grid
satisfies
ee = A−1 −1 −1
c re = Xc Λc Xc λ1 (xc )1
(10.31)
 
1
 
 0 
= λ1 Xc Λ−1
c 
 .. 

 . 
0
(10.32)
 
1/(λc )1
 
 0 
= λ1 Xc 
 .. 

 . 
0
(10.33)
λ1
= (xc )1
(λc )1
(10.34)
1
≈ (xc )1 (10.35)
2
Since our goal is to compute e = x1 , in addition to interpolating ee to the fine grid
(using Eq. 10.15), we must multiply the result by 2. This is equivalent to solving
1
Ac ee = re (10.36)
2
or
1
B(Mc : 1, −2, 1)ee = re (10.37)
4
In our case, the matrix A = B(M : 1, −2, 1) comes from a discretization of the
diffusion equation, which gives
ν
Ab = B(M : 1, −2, 1) (10.38)
∆x2
and the preconditioning matrix C is simply
∆x2
C= I (10.39)
ν
Applying the discretization on the coarse grid with the same preconditioning matrix
as used on the fine grid gives, since ∆xc = 2∆x,
ν ∆x2 1
C 2 B(Mc : 1, −2, 1) = 2
B(Mc : 1, −2, 1) = B(Mc : 1, −2, 1) (10.40)
∆xc ∆xc 4
which is precisely the matrix appearing in Eq. 10.37. Thus we see that the process is
recursive. The problem to be solved on the coarse grid is the same as that solved on
the fine grid.
The remaining steps required to complete an entire multigrid process are relatively
straightforward, but they vary depending on the problem and the user. The reduction
can be, and usually is, carried to even coarser grids before returning to the finest level.
However, in each case the appropriate permutation matrix and the interpolation
approximation define both the down- and up-going paths. The details of finding
optimum techniques are, obviously, quite important but they are not discussed here.
10.3 A Two-Grid Process

We now describe a two-grid process for the linear problem Aφ = f, which can be easily
generalized to a process with an arbitrary number of grids due to the recursive nature
of multigrid. Extension to nonlinear problems requires that both the solution and the
residual be transferred to the coarse grid in a process known as full approximation
storage multigrid.
1. Perform n1 iterations of the selected relaxation method on the fine grid, starting
=φ
with φ (1) . This gives1
n . Call the result φ
φ n + (I − Gn1 ) A−1 f
(1) = Gn1 φ (10.41)
1 1
where
G1 = I + H1−1A1 (10.42)
and H1 is defined as in Chapter 9 (e.g., Eq. 9.21). Next compute the residual based
(1) :
on φ
n + A (I − Gn1 ) A−1 f − f
(1) − f = AGn1 φ
r(1) = Aφ 1 1
n1 n1 −1
= AG φn − AG A f
1 1 (10.43)
2. Transfer (or restrict) r(1) to the coarse grid:

r(2) = R12r(1) (10.44)
In our example in the preceding section, the restriction matrix is
 
0 1 0 0 0 0 0
R1 =  0 0 0 1 0 0 0 
2 
 (10.45)
0 0 0 0 0 1 0
that is, the first three rows of the permutation matrix P in Eq. 10.11. This type of
restriction is known as “simple injection.” Some form of weighted restriction can also
be used.
3. Solve the problem A2e(2) = r(2) on the coarse grid exactly:2

e(2) = A−1
2 r(2) (10.46)
1
See problem 1 of Chapter 9.
2
Note that the coarse grid matrix denoted A2 here was denoted Ac in the preceding section.
10.3. A TWO-GRID PROCESS 201
Here A2 can be formed by applying the discretization on the coarse grid. In the
preceding example (eq. 10.40), A2 = 41 B(Mc : 1, −2, 1). It is at this stage that the
generalization to a multigrid procedure with more than two grids occurs. If this is the
coarsest grid in the sequence, solve exactly. Otherwise, apply the two-grid process
recursively.
4. Transfer (or prolong) the error back to the fine grid and update the solution:
φ (1) − I 1e(2)
n+1 = φ (10.47)
2
In our example, the prolongation matrix is

 
1/2 0 0
 
 1 0 0 
 
 1/2 1/2 0 
 

I21 =  0 1 0  (10.48)
 
 0 1/2 1/2 
 
 0 0 1 
 
0 0 1/2
which follows from Eq. 10.15.
Combining these steps, one obtains
n+1 = [I − I 1 A−1 R2 A]Gn1 φ

φ n − [I − I 1 A−1 R2 A]Gn1 A−1 f + A−1 f (10.49)
2 2 1 1 2 2 1 1
Thus the basic iteration matrix is
[I − I21 A−1 2 n1
2 R1 A]G1 (10.50)
The eigenvalues of this matrix determine the convergence rate of the two-grid process.
The basic iteration matrix for a three-grid process is found from Eq. 10.50 by
replacing A−1 −1
2 with (I − G3 )A2 , where
2
G23 = [I − I32 A−1 3 n2

3 R2 A2 ]G2 (10.51)
In this expression n2 is the number of relaxation steps on grid 2, I32 and R23 are the
transfer operators between grids 2 and 3, and A3 is obtained by discretizing on grid
3. Extension to four or more grids proceeds in similar fashion.
10.4 Problems
1. Derive Eq. 10.51.
2. Repeat problem 4 of Chapter 9 using a four-grid multigrid method together

with
(a) the Gauss-Seidel method,

(b) the 3-step Richardson method derived in Section 9.5.
Solve exactly on the coarsest grid. Plot the solution after the residual is reduced
by 2, 3, and 4 orders of magnitude. Plot the logarithm of the L2 -norm of the
residual vs. the number of iterations. Determine the asymptotic convergence
rate. Calculate the theoretical asymptotic convergence rate and compare.
Chapter 11
NUMERICAL DISSIPATION
Up to this point, we have emphasized the second-order centered-difference approxima-

tions to the spatial derivatives in our model equations. We have seen that a centered
approximation to a first derivative is nondissipative, i.e., the eigenvalues of the as-
sociated circulant matrix (with periodic boundary conditions) are pure imaginary.
In processes governed by nonlinear equations, such as the Euler and Navier-Stokes
equations, there can be a continual production of high-frequency components of the
solution, leading, for example, to the production of shock waves. In a real phys-
ical problem, the production of high frequencies is eventually limited by viscosity.
However, when we solve the Euler equations numerically, we have neglected viscous
effects. Thus the numerical approximation must contain some inherent dissipation to
limit the production of high-frequency modes. Although numerical approximations
to the Navier-Stokes equations contain dissipation through the viscous terms, this
can be insufficient, especially at high Reynolds numbers, due to the limited grid res-
olution which is practical. Therefore, unless the relevant length scales are resolved,
some form of added numerical dissipation is required in the numerical solution of
the Navier-Stokes equations as well. Since the addition of numerical dissipation is
tantamount to intentionally introducing nonphysical behavior, it must be carefully
controlled such that the error introduced is not excessive. In this Chapter, we discuss
some different ways of adding numerical dissipation to the spatial derivatives in the
linear convection equation and hyperbolic systems of PDE’s.
203
204 CHAPTER 11. NUMERICAL DISSIPATION
11.1 One-Sided First-Derivative Space Differenc-

ing
We investigate the properties of one-sided spatial difference operators in the context
of the biconvection model equation given by
∂u ∂u
= −a (11.1)
∂t ∂x
with periodic boundary conditions. Consider the following point operator for the
spatial derivative term
−a
−a(δx u)j = [−(1 + β)uj−1 + 2βuj + (1 − β)uj+1]
2∆x
−a
= [(−uj−1 + uj+1) + β(−uj−1 + 2uj − uj+1)] (11.2)
2∆x
The second form shown divides the operator into an antisymmetric component (−uj−1+
uj+1)/2∆x and a symmetric component β(−uj−1 + 2uj − uj+1)/2∆x. The antisym-
metric component is the second-order centered difference operator. With β = 0, the
operator is only first-order accurate. A backward difference operator is given by β = 1
and a forward difference operator is given by β = −1.
For periodic boundary conditions the corresponding matrix operator is
−a
−aδx = Bp (−1 − β, 2β, 1 − β)
2∆x
The eigenvalues of this matrix are

3 4
−a 2πm 2πm
λm = β 1 − cos + i sin for m = 0, 1, . . . , M − 1
∆x M M
If a is positive, the forward difference operator (β = −1) produces Re(λm ) > 0,

the centered difference operator (β = 0) produces Re(λm ) = 0, and the backward
difference operator produces Re(λm ) < 0. Hence the forward difference operator is
inherently unstable while the centered and backward operators are inherently stable.
If a is negative, the roles are reversed. When Re(λm ) = 0, the solution will either
grow or decay with time. In either case, our choice of differencing scheme produces
nonphysical behavior. We proceed next to show why this occurs.
11.2. THE MODIFIED PARTIAL DIFFERENTIAL EQUATION 205
11.2 The Modified Partial Differential Equation

First carry out a Taylor series expansion of the terms in Eq. 11.2. We are lead to the
expression
 
1  ∂u ∂2u ∆x3 ∂3 u β∆x4 ∂4 u
(δx u)j = 2∆x − β∆x2 + − + . . .
2∆x ∂x j
∂x2 j
3 ∂x3 j
12 ∂x4 j
We see that the antisymmetric portion of the operator introduces odd derivative
terms in the truncation error while the symmetric portion introduces even derivatives.
Substituting this into Eq. 11.1 gives
∂u ∂u aβ∆x ∂ 2 u a∆x2 ∂ 3 u aβ∆x3 ∂ 4 u

= −a + − + +... (11.3)
∂t ∂x 2 ∂x2 6 ∂x3 24 ∂x4
This is the partial differential equation we are really solving when we apply the
approximation given by Eq. 11.2 to Eq. 11.1. Notice that Eq. 11.3 is consistent with
Eq. 11.1, since the two equations are identical when ∆x → 0. However, when we use
a computer to find a numerical solution of the problem, ∆x can be small but it is
not zero. This means that each term in the expansion given by Eq. 11.3 is excited to
some degree. We refer to Eq. 11.3 as the modified partial differential equation. We
proceed next to investigate the implications of this concept.
Consider the simple linear partial differential equation
∂u ∂u ∂2u ∂3u ∂4u
= −a +ν 2 +γ 3 +τ 4 (11.4)
∂t ∂x ∂x ∂x ∂x
Choose periodic boundary conditions and impose an initial condition u = eiκx . Under
these conditions there is a wave-like solution to Eq. 11.4 of the form
u(x, t) = eiκx e(r+is)t
provided r and s satisfy the condition
r + is = −iaκ − νκ2 − iγκ3 + τ κ4
or
r = −κ2 (ν − τ κ2 ), s = −κ(a + γκ2 )
The solution is composed of both amplitude and phase terms. Thus
u = e −κ (ν−τ
2 2
κ ) iκ[x−(a+γκ )t] 2
e (11.5)
amplitude phase
It is important to notice that the amplitude of the solution depends only upon ν and
τ , the coefficients of the even derivatives in Eq. 11.4, and the phase depends only on
a and γ, the coefficients of the odd derivatives.
If the wave speed a is positive, the choice of a backward difference scheme (β = 1)
produces a modified PDE with ν − τ κ2 > 0 and hence the amplitude of the solution
decays. This is tantamount to deliberately adding dissipation to the PDE. Under the
same condition, the choice of a forward difference scheme (β = −1) is equivalent to
deliberately adding a destabilizing term to the PDE.
By examining the term governing the phase of the solution in Eq. 11.5, we see
that the speed of propagation is a + γκ2 . Referring to the modified PDE, Eq. 11.3
we have γ = −a∆x2 /6. Therefore, the phase speed of the numerical solution is less
than the actual phase speed. Furthermore, the numerical phase speed is dependent
upon the wavenumber κ. This we refer to as dispersion.
Our purpose here is to investigate the properties of one-sided spatial differencing
operators relative to centered difference operators. We have seen that the three-
point centered difference approximation of the spatial derivative produces a modified
PDE that has no dissipation (or amplification). One can easily show, by using the
antisymmetry of the matrix difference operators, that the same is true for any cen-
tered difference approximation of a first derivative. As a corollary, any departure
from antisymmetry in the matrix difference operator must introduce dissipation (or
amplification) into the modified PDE.
Note that the use of one-sided differencing schemes is not the only way to in-
troduce dissipation. Any symmetric component in the spatial operator introduces
dissipation (or amplification). Therefore, one could choose β = 1/2 in Eq. 11.2. The
resulting spatial operator is not one-sided but it is dissipative. Biased schemes use
more information on one side of the node than the other. For example, a third-order
backward-biased scheme is given by
1
(δx u)j = (uj−2 − 6uj−1 + 3uj + 2uj+1)
6∆x
1
= [(uj−2 − 8uj−1 + 8uj+1 − uj+2)
12∆x
+ (uj−2 − 4uj−1 + 6uj − 4uj+1 + uj+2)] (11.6)
The antisymmetric component of this operator is the fourth-order centered difference

operator. The symmetric component approximates ∆x3 uxxxx/12. Therefore, this
operator produces fourth-order accuracy in phase with a third-order dissipative term.
11.3. THE LAX-WENDROFF METHOD 207
11.3 The Lax-Wendroff Method

In order to introduce numerical dissipation using one-sided differencing, backward
differencing must be used if the wave speed is positive, and forward differencing must
be used if the wave speed is negative. Next we consider a method which introduces
dissipation independent of the sign of the wave speed, known as the Lax-Wendroff
method. This explicit method differs conceptually from the methods considered pre-
viously in which spatial differencing and time-marching are treated separately.
Consider the following Taylor-series expansion in time:
∂u 1 2 ∂ 2 u
u(x, t + h) = u + h + h 2
+ O(h3) (11.7)
∂t 2 ∂t
First replace the time derivatives with space derivatives according to the PDE (in
this case, the linear convection equation ∂∂tu + a ∂∂xu = 0). Thus
∂u ∂u ∂2u 2
2∂ u
= −a , = a (11.8)
∂t ∂x ∂t2 ∂x2
Now replace the space derivatives with three-point centered difference operators, giv-
ing
2
(n+1) (n) 1 ah (n) (n) 1 ah (n) (n) (n)
uj = uj − (uj+1 − uj−1) + (uj+1 − 2uj + uj−1 ) (11.9)
2 ∆x 2 ∆x
This is the Lax-Wendroff method applied to the linear convection equation. It is a

fully-discrete finite-difference scheme. There is no intermediate semi-discrete stage.
For periodic boundary conditions, the corresponding fully-discrete matrix operator
is   2  2  2 
1 ah ah ah 1 ah ah
un+1 = Bp   + ,1 − , − + 
un
2 ∆x ∆x ∆x 2 ∆x ∆x
The eigenvalues of this matrix are

2
ah 2πm ah 2πm
σm = 1 − 1 − cos −i sin for m = 0, 1, . . . , M − 1
∆x M ∆x M
For | ∆x
ah
| ≤ 1 all of the eigenvalues have modulus less than or equal to unity and hence
the method is stable independent of the sign of a. The quantity |∆x ah
| is known as the
Courant (or CFL) number. It is equal to the ratio of the distance travelled by a wave
in one time step to the mesh spacing.
The nature of the dissipative properties of the Lax-Wendroff scheme can be seen
by examining the modified partial differential equation, which is given by
∂u ∂u a ∂ 3 u a2 h ∂4u
+a = − (∆x2 − a2 h2 ) 3 − (∆x2 − a2 h2 ) 4 + . . .
∂t ∂x 6 ∂x 8 ∂x
This is derived by substituting Taylor series expansions for all terms in Eq. 11.9 and
converting the time derivatives to space derivatives using Eq. 11.8. The two leading
error terms appear on the right side of the equation. Recall that the odd derivatives on
the right side lead to unwanted dispersion and the even derivatives lead to dissipation
(or amplification, depending on the sign). Therefore, the leading error term in the
Lax-Wendroff method is dispersive and proportional to
a ∂3u a∆x2 ∂3u
− (∆x2 − a2 h2 ) 3 = − (1 − Cn2 ) 3
6 ∂x 6 ∂x
The dissipative term is proportional to
a2 h ∂4u a2 h∆x2 ∂4u
− (∆x2 − a2 h2 ) 4 = − (1 − Cn2 ) 4
8 ∂x 8 ∂x
This term has the appropriate sign and hence the scheme is truly dissipative as long
as Cn ≤ 1.
A closely related method is that of MacCormack. Recall MacCormack’s time-
marching method, presented in Chapter 6:
ũn+1 = un + hun
1
un+1 = [un + ũn+1 + hũn+1 ] (11.10)
2
If we use first-order backward differencing in the first stage and first-order forward
differencing in the second stage,1 a dissipative second-order method is obtained. For
the linear convection equation, this approach leads to
(n+1) (n)ah (n) (n)

ũj = uj − (uj − uj−1)
∆x
(n+1) 1 (n) (n+1) ah (n+1) (n+1)
uj = [uj + ũj − (ũj+1 − ũj )] (11.11)
2 ∆x
which can be shown to be identical to the Lax-Wendroff method. Hence MacCor-
mack’s method has the same dissipative and dispersive properties as the Lax-Wendroff
method. The two methods differ when applied to nonlinear hyperbolic systems, how-
ever.
1
Or vice-versa; for nonlinear problems, these should be applied alternately.
11.4. UPWIND SCHEMES 209
11.4 Upwind Schemes

In Section 11.1, we saw that numerical dissipation can be introduced in the spatial
difference operator using one-sided difference schemes or, more generally, by adding
a symmetric component to the spatial operator. With this approach, the direction
of the one-sided operator (i.e., whether it is a forward or a backward difference)
or the sign of the symmetric component depends on the sign of the wave speed.
When a hyperbolic system of equations is being solved, the wave speeds can be both
positive and negative. For example, the eigenvalues of the flux Jacobian for the one-
dimensional Euler equations are u, u + a, u − a. When the flow is subsonic, these are
of mixed sign. In order to apply one-sided differencing schemes to such systems, some
form of splitting is required. This is avoided in the Lax-Wendroff scheme. However,
as a result of their superior flexibility, schemes in which the numerical dissipation
is introduced in the spatial operator are generally preferred over the Lax-Wendroff
approach.
Consider again the linear convection equation:
∂u ∂u
+a =0 (11.12)
∂t ∂x
where we do not make any assumptions as to the sign of a. We can rewrite Eq. 11.12
as
∂u ∂u a ± |a|
+ (a+ + a− ) = 0 ; a± =
∂t ∂x 2
−
If a ≥ 0, then a = a ≥ 0 and a = 0. Alternatively, if a ≤ 0, then a+ = 0 and
+
a− = a ≤ 0. Now for the a+ (≥ 0) term we can safely backward difference and for the
a− (≤ 0) term forward difference. This is the basic concept behind upwind methods,
that is, some decomposition or splitting of the fluxes into terms which have positive
and negative characteristic speeds so that appropriate differencing schemes can be
chosen. In the next two sections, we present two splitting techniques commonly used
with upwind methods. These are by no means unique.
The above approach to obtaining a stable discretization independent of the sign
of a can be written in a different, but entirely equivalent, manner. From Eq. 11.2, we
see that a stable discretization is obtained with β = 1 if a ≥ 0 and with β = −1 if
a ≤ 0. This is achieved by the following point operator:
−1
−a(δx u)j = [a(−uj−1 + uj+1) + |a|(−uj−1 + 2uj − uj+1)] (11.13)
2∆x
This approach is extended to systems of equations in Section 11.5.
In this section, we present the basic ideas of flux-vector and flux-difference splitting.
For more subtle aspects of implementation and application of such techniques to
nonlinear hyperbolic systems such as the Euler equations, the reader is referred to
the literature on this subject.
11.4.1 Flux-Vector Splitting

Recall from Section 2.5 that a linear, constant-coefficient, hyperbolic system of partial
differential equations given by
∂u ∂f ∂u ∂u
+ = +A =0 (11.14)
∂t ∂x ∂t ∂x
can be decoupled into characteristic equations of the form
∂wi ∂wi
+ λi =0 (11.15)
∂t ∂x
where the wave speeds, λi , are the eigenvalues of the Jacobian matrix, A, and the
wi ’s are the characteristic variables. In order to apply a one-sided (or biased) spatial
differencing scheme, we need to apply a backward difference if the wave speed, λi , is
positive, and a forward difference if the wave speed is negative. To accomplish this,
let us split the matrix of eigenvalues, Λ, into two components such that
Λ = Λ+ + Λ− (11.16)
where
Λ + |Λ| Λ − |Λ|
Λ+ = , Λ− = (11.17)
2 2
With these definitions, Λ+ contains the positive eigenvalues and Λ− contains the neg-
ative eigenvalues. We can now rewrite the system in terms of characteristic variables
as
∂w ∂w ∂w ∂w ∂w
+Λ = + Λ+ + Λ− =0 (11.18)
∂t ∂x ∂t ∂x ∂x
The spatial terms have been split into two components according to the sign of the
wave speeds. We can use backward differencing for the Λ+ ∂∂xw term and forward
differencing for the Λ− ∂∂xw term. Premultiplying by X and inserting the product
X −1 X in the spatial terms gives
∂Xw ∂XΛ+ X −1 Xw ∂XΛ− X −1 Xw

+ + =0 (11.19)
∂t ∂x ∂x
11.4. UPWIND SCHEMES 211
With the definitions2
A+ = XΛ+X −1 , A− = XΛ−X −1 (11.20)
and recalling that u = Xw, we obtain
∂u ∂A+ u ∂A− u
+ + =0 (11.21)
∂t ∂x ∂x
Finally the split flux vectors are defined as
f + = A+ u, f − = A− u (11.22)
and we can write
∂u ∂f + ∂f −
+ + =0 (11.23)
∂t ∂x ∂x
In the linear case, the definition of the split fluxes follows directly from the defini-
tion of the flux, f = Au. For the Euler equations, f is also equal to Au as a result of
their homogeneous property, as discussed in Appendix C. Note that
f = f+ + f− (11.24)
Thus by applying backward differences to the f+ term and forward differences to the
f − term, we are in effect solving the characteristic equations in the desired manner.
This approach is known as flux-vector splitting.
When an implicit time-marching method is used, the Jacobians of the split flux
vectors are required. In the nonlinear case,
∂f + ∂f −
= A+ , = A− (11.25)
∂u ∂u
Therefore, one must find and use the new Jacobians given by
∂f + ∂f −
A++ = , A−− = (11.26)
∂u ∂u
For the Euler equations, A++ has eigenvalues which are all positive, and A−− has all
negative eigenvalues.
2
With these definitions A+ has all positive eigenvalues, and A− has all negative eigenvalues.
11.4.2 Flux-Difference Splitting

Another approach, more suited to finite-volume methods, is known as flux-difference
splitting. In a finite-volume method, the fluxes must be evaluated at cell bound-
aries. We again begin with the diagonalized form of the linear, constant-coefficient,
hyperbolic system of equations
∂w ∂w
+Λ =0 (11.27)
∂t ∂x
The flux vector associated with this form is g = Λw. Now, as in Chapter 5, we
consider the numerical flux at the interface between nodes j and j + 1, ĝj+1/2 , as a
function of the states to the left and right of the interface, wL and wR , respectively.
The centered approximation to gj+1/2 , which is nondissipative, is given by
1
ĝj+1/2 = (g(wL) + g(wR)) (11.28)
2
In order to obtain a one-sided upwind approximation, we require

λi (wi)L if λi > 0
(ĝi )j+1/2 = (11.29)
λi (wi )R if λi < 0
where the subscript i indicates individual components of w and g. This is achieved

with
1 1
(ĝi)j+1/2 = λi [(wi )L + (wi )R ] + |λi | [(wi )L − (wi )R ] (11.30)
2 2
or
1 1
ĝj+1/2 = Λ (wL + wR ) + |Λ| (wL − wR ) (11.31)
2 2
Now, as in Eq. 11.19, we premultiply by X to return to the original variables and
insert the product X −1 X after Λ and |Λ| to obtain
1 1
X ĝj+1/2 = XΛX −1 X (wL + wR ) + X|Λ|X −1X (wL − wR ) (11.32)
2 2
and thus
1 1
fˆj+1/2 = (fL + fR ) + |A| (uL − uR ) (11.33)
2 2
where
|A| = X|Λ|X −1 (11.34)

11.5. ARTIFICIAL DISSIPATION 213
and we have also used the relations f = Xg, u = Xw, and A = XΛX−1 .
In the linear, constant-coefficient case, this leads to an upwind operator which is
identical to that obtained using flux-vector splitting. However, in the nonlinear case,
there is some ambiguity regarding the definition of |A| at the cell interface j + 1/2.
In order to resolve this, consider a situation in which the eigenvalues of A are all of
the same sign. In this case, we would like our definition of fˆj+1/2 to satisfy

fL if all λi s > 0
fˆj+1/2 = (11.35)
fR if all λi s < 0
giving pure upwinding. If the eigenvalues of A are all positive, |A| = A; if they are
all negative, |A| = −A. Hence satisfaction of Eq. 11.35 is obtained by the definition
1 1
fˆj+1/2 = (fL + fR ) + |Aj+1/2 | (uL − uR ) (11.36)
2 2
if Aj+1/2 satisfies
fL − fR = Aj+1/2 (uL − uR ) (11.37)
For the Euler equations for a perfect gas, Eq. 11.37 is satisfied by the flux Jacobian
evaluated at the Roe-average state given by
√ √
ρL uL + ρR uR
uj+1/2 = √ √ (11.38)
ρL + ρR
√ √
ρL HL + ρR HR
Hj+1/2 = √ √ (11.39)
ρL + ρR
where u and H = (e + p)/ρ are the velocity and the total enthalpy per unit mass,
respectively.3
11.5 Artificial Dissipation

We have seen that numerical dissipation can be introduced by using one-sided dif-
ferencing schemes together with some form of flux splitting. We have also seen that
such dissipation can be introduced by adding a symmetric component to an antisym-
metric (dissipation-free) operator. Thus we can generalize the concept of upwinding
to include any scheme in which the symmetric portion of the operator is treated in
such a manner as to be truly dissipative.
3
Note that the flux Jacobian can be written in terms of u and H only; see problem 6 at the end
of this chapter.
For example, let

uj+1 − uj−1 −uj+1 + 2uj − uj−1
(δxa u)j = , (δxs u)j = (11.40)
2∆x 2∆x
Applying δx = δxa + δxs to the spatial derivative in Eq. 11.15 is stable if λi ≥ 0 and
unstable if λi < 0. Similarly, applying δx = δxa − δxs is stable if λi ≤ 0 and unstable if
λi > 0. The appropriate implementation is thus
λi δx = λi δxa + |λi |δxs (11.41)
Extension to a hyperbolic system by applying the above approach to the characteristic

variables, as in the previous two sections, gives
δx (Au) = δxa (Au) + δxs (|A|u) (11.42)
or
δx f = δxa f + δxs (|A|u) (11.43)
where |A| is defined in Eq. 11.34. The second spatial term is known as artificial
dissipation. It is also sometimes referred to as artificial diffusion or artificial viscosity.
With appropriate choices of δxa and δxs , this approach can be related to the upwind
approach. This is particularly evident from a comparison of Eqs. 11.36 and 11.43.
It is common to use the following operator for δxs

(δxs u)j = (uj−2 − 4uj−1 + 6uj − 4uj+1 + uj+2) (11.44)
∆x
where is a problem-dependent coefficient. This symmetric operator approximates
∆x3 uxxxx and thus introduces a third-order dissipative term. With an appropriate
value of , this often provides sufficient damping of high frequency modes without
greatly affecting the low frequency modes. For details of how this can be implemented
for nonlinear hyperbolic systems, the reader should consult the literature. A more
complicated treatment of the numerical dissipation is also required near shock waves
and other discontinuities, but is beyond the scope of this book.
11.6 Problems
1. A second-order backward difference approximation to a 1st derivative is given
as a point operator by
1
(δx u)j = (uj−2 − 4uj−1 + 3uj )
2∆x
11.6. PROBLEMS 215
(a) Express this operator in banded matrix form (for periodic boundary condi-
tions), then derive the symmetric and skew-symmetric matrices that have
the matrix operator as their sum. (See Appendix A.3 to see how to con-
struct the symmetric and skew-symmetric components of a matrix.)
(b) Using a Taylor table, find the derivative which is approximated by the
corresponding symmetric and skew-symmetric operators and the leading
error term for each.
2. Find the modified wavenumber for the first-order backward difference operator.
Plot the real and imaginary parts of κ∗ ∆x vs. κ∆x for 0 ≤ κ∆x ≤ π. Using
Fourier analysis as in Section 6.6.2, find |σ| for the combination of this spatial
operator with 4th-order Runge-Kutta time marching at a Courant number of
unity and plot vs. κ∆x for 0 ≤ κ∆x ≤ π.
3. Find the modified wavenumber for the operator given in Eq. 11.6. Plot the real
and imaginary parts of κ∗ ∆x vs. κ∆x for 0 ≤ κ∆x ≤ π. Using Fourier analysis
as in Section 6.6.2, find |σ| for the combination of this spatial operator with
4th-order Runge-Kutta time marching at a Courant number of unity and plot
vs. κ∆x for 0 ≤ κ∆x ≤ π.
4. Consider the spatial operator obtained by combining second-order centered dif-

ferences with the symmetric operator given in Eq. 11.44. Find the modified
wavenumber for this operator with = 0, 1/12, 1/24, and 1/48. Plot the real
and imaginary parts of κ∗ ∆x vs. κ∆x for 0 ≤ κ∆x ≤ π. Using Fourier analysis
as in Section 6.6.2, find |σ| for the combination of this spatial operator with
4th-order Runge-Kutta time marching at a Courant number of unity and plot
vs. κ∆x for 0 ≤ κ∆x ≤ π.
5. Consider the hyperbolic system derived in problem 8 of Chapter 2. Find the

matrix |A|. Form the plus-minus split flux vectors as in Section 11.4.1.
6. Show that the flux Jacobian for the 1-D Euler equations can be written in terms
of u and H. Show that the use of the Roe average state given in Eqs. 11.38 and
11.39 leads to satisfaction of Eq. 11.37.
Chapter 12
SPLIT AND FACTORED FORMS
In the next two chapters, we present and analyze split and factored algorithms. This
gives the reader a feel for some of the modifications which can be made to the basic
algorithms in order to obtain efficient solvers for practical multidimensional applica-
tions, and a means for analyzing such modified forms.
12.1 The Concept

Factored forms of numerical operators are used extensively in constructing and ap-
plying numerical methods to problems in fluid mechanics. They are the basis for a
wide variety of methods variously known by the labels “hybrid”, “time split”, and
“fractional step”. Factored forms are especially useful for the derivation of practical
algorithms that use implicit methods. When we approach numerical analysis in the
light of matrix derivative operators, the concept of factoring is quite simple to present
and grasp. Let us start with the following observations:
1. Matrices can be split in quite arbitrary ways.
2. Advancing to the next time level always requires some reference to a previous
one.
3. Time marching methods are valid only to some order of accuracy in the step
size, h.
Now recall the generic ODE’s produced by the semi-discrete approach
du
= Au − f (12.1)
dt
217
218 CHAPTER 12. SPLIT AND FACTORED FORMS
and consider the above observations. From observation 1 (arbitrary splitting of A) :
du
= [A1 + A2 ]u − f (12.2)
dt
where A = [A1 + A2 ] but A1 and A2 are not unique. For the time march let us choose
the simple, first-order,1 explicit Euler method. Then, from observation 2 (new data
un+1 in terms of old un ):
un+1 = [ I + hA1 + hA2 ]un − hf + O(h2 ) (12.3)
or its equivalent

un+1 = [ I + hA1 ][ I + hA2 ] − h2 A1 A2 un − hf + O(h2 )
Finally, from observation 3 (allowing us to drop higher order terms −h2 A1 A2un ):
un+1 = [ I + hA1 ][ I + hA2 ]un − hf + O(h2 ) (12.4)
Notice that Eqs. 12.3 and 12.4 have the same formal order of accuracy and, in
this sense, neither one is to be preferred over the other. However, their numerical
stability can be quite different, and techniques to carry out their numerical evaluation
can have arithmetic operation counts that vary by orders of magnitude. Both of these
considerations are investigated later. Here we seek only to apply to some simple cases
the concept of factoring.
12.2 Factoring Physical Representations — Time

Splitting
Suppose we have a PDE that represents both the processes of convection and dissi-
pation. The semi-discrete approach to its solution might be put in the form
du
= Acu + Adu + (bc) (12.5)
dt
where Ac and Ad are matrices representing the convection and dissipation terms,
respectively; and their sum forms the A matrix we have considered in the previous
sections. Choose again the explicit Euler time march so that
+ O(h2 )
un+1 = [ I + hAd + hAc ]un + h(bc) (12.6)
1
Second-order time-marching methods are considered later.
12.2. FACTORING PHYSICAL REPRESENTATIONS — TIME SPLITTING 219
Now consider the factored form

un+1 = [ I + hAd ] [ I + hAc ]un + h(bc)

+ h2 Ad Acun + (bc)
= [ I + hAd + hAc ]un + h(bc) +O(h2 ) (12.7)

Original Unfactored Terms Higher Order Terms
and we see that Eq. 12.7 and the original unfactored form Eq. 12.6 have identical
orders of accuracy in the time approximation. Therefore, on this basis, their selection
is arbitrary. In practical applications2 equations such as 12.7 are often applied in a
predictor-corrector sequence. In this case one could write

ũn+1 = [ I + hAc ]un + h(bc)
un+1 = [ I + hAd ]ũn+1 (12.8)
Factoring can also be useful to form split combinations of implicit and explicit
techniques. For example, another way to approximate Eq. 12.6 with the same order
of accuracy is given by the expression

un+1 = [ I − hAd ]−1 [ I + hAc ]un + h(bc)

+O(h2 )
= [ I + hAd + hAc ]un + h(bc) (12.9)

Original Unfactored Terms
where in this approximation we have used the fact that
[ I − hAd ]−1 = I + hAd + h2 A2d + · · ·
if h · ||Ad || < 1, where ||Ad || is some norm of [Ad ]. This time a predictor-corrector
interpretation leads to the sequence

ũn+1 = [ I + hAc ]un + h(bc)
[ I − hAd ]un+1 = ũn+1 (12.10)
The convection operator is applied explicitly, as before, but the diffusion operator is
now implicit, requiring a tridiagonal solver if the diffusion term is central differenced.
Since numerical stiffness is generally much more severe for the diffusion process, this
factored form would appear to be superior to that provided by Eq. 12.8. However,
the important aspect of stability has yet to be discussed.
2
We do not suggest that this particular method is suitable for use. We have yet to determine its
stability, and a first-order time-march method is usually unsatisfactory.
We should mention here that Eq. 12.9 can be derived for a different point of view
by writing Eq. 12.6 in the form
un+1 − un + O(h2 )
= Ac un + Ad un+1 + (bc)
h
Then

[ I − hAd ]un+1 = [ I + hAc ]un + h(bc)
12.3 Factoring Space Matrix Operators in 2–D

12.3.1 Mesh Indexing Convention
Factoring is widely used in codes designed for the numerical solution of equations
governing unsteady two- and three-dimensional flows. Let us study the basic concept
of factoring by inspecting its use on the linear 2-D scalar PDE that models diffusion:
∂u ∂2u ∂2u
= 2+ 2 (12.11)
∂t ∂x ∂y
We begin by reducing this PDE to a coupled set of ODE’s by differencing the space
derivatives and inspecting the resulting matrix operator.
A clear description of a matrix finite-difference operator in 2- and 3-D requires some
reference to a mesh. We choose the 3 × 4 point mesh3 shown in the Sketch 12.12.
In this example Mx , the number of (interior) x points, is 4 and My , the number of
(interior) y points is 3. The numbers 11 , 12 , · · · , 43 represent the location in the
mesh of the dependent variable bearing that index. Thus u32 represents the value of
u at j = 3 and k = 2.

My 13 23 33 43
k 12 22 32 42
1 11 21 31 41 (12.12)

1 j · · · Mx
Mesh indexing in 2-D.
3
This could also be called a 5 × 6 point mesh if the boundary points (labeled in the sketch)
were included, but in these notes we describe the size of a mesh by the number of interior points.
12.3. FACTORING SPACE MATRIX OPERATORS IN 2–D 221
12.3.2 Data Bases and Space Vectors

The dimensioned array in a computer code that allots the storage locations of the
dependent variable(s) is referred to as a data-base. There are many ways to lay out
a data-base. Of these, we consider only two: (1), consecutively along rows that are
themselves consecutive from k = 1 to My , and (2), consecutively along columns that
are consecutive from j = 1 to Mx . We refer to each row or column group as a
space vector (they represent data along lines that are continuous in space) and label
their sum with the symbol U . In particular, (1) and (2) above are referred to as x-
vectors and y-vectors, respectively. The symbol U by itself is not enough to identify
the structure of the data-base and is used only when the structure is immaterial or
understood.
To be specific about the structure, we label a data–base composed of x-vectors with
U , and one composed of y-vectors with U (y) . Examples of the order of indexing
(x)
for these space vectors are given in Eq. 12.16 part a and b.
12.3.3 Data Base Permutations

The two vectors (arrays) are related by a permutation matrix P such that
U (x) = Pxy U (y) and U (y) = Pyx U (x) (12.13)
where
Pyx = PTxy = P−1
xy
Now consider the structure of a matrix finite-difference operator representing 3-

point central-differencing schemes for both space derivatives in two dimensions. When
the matrix is multiplying a space vector U , the usual (but ambiguous) representation
is given by Ax+y . In this notation the ODE form of Eq. 12.11 can be written 4
dU
= Ax+y U + (bc) (12.14)
dt
If it is important to be specific about the data-base structure, we use the notation
A(x) (y)
x+y or Ax+y , depending on the data–base chosen for the U it multiplies. Examples
are in Eq. 12.16 part a and b. Notice that the matrices are not the same although
they represent the same derivative operation. Their structures are similar, however,
and they are related by the same permutation matrix that relates U(x) to U (y) . Thus
(x) (y)
Ax+y = Pxy · Ax+y · Pyx (12.15)
4
Notice that Ax+y and U , which are notations used in the special case of space vectors, are
subsets of A and u, used in the previous sections.
 
• x | o | 11
x • x | o |  21
 
 


x • x | o | 

31
 x • | o |  41
 
 
  −−
 
o | • x | o  12
 
 o | x • x | o  22
(x)  
Ax+y · U (x) =  ·

 o | x • x | o 
 32
 o | x • | o 42
 
 




−−
 | o | • x  13
 
 
 | o | x • x  23
 
 | o | x • x 33
| o | x • 43
a:Elements in 2-dimensional, central-difference, matrix
operator, Ax+y , for 3×4 mesh shown in Sketch 12.12.
Data base composed of My x–vectors stored in U (x) .
Entries for x → x, for y → o, for both → •.
(12.16)
 
• o | x | | 11
o • o | x | |  12
 
 
 o • | x | |  13
 
  −−
 
x | • o | x |  21
 
 

 x | o • o | x | 
 22
 x | o • | x |  23
 
(y)  
Ax+y · U (y) =  · −−
 
 | x | • o | x  31
 
 | | • | 
 x o o x  32
 

 | x | o • | x
 33
  −−
 
 


| | x | • o 

41
 | | x | o • o 42
| | x | o • 43
b: Elements in 2-dimensional, central-difference, matrix

operator, Ax+y , for 3×4 mesh shown in Sketch 12.12.
Data base composed of Mx y–vectors stored in U (y) .
Entries for x → x, for y → o, for both → •.
12.3.4 Space Splitting and Factoring

We are now prepared to discuss splitting in two dimensions. It should be clear that
(x)
the matrix Ax+y can be split into two matrices such that
A(x) (x) (x)

x+y = Ax + Ay (12.17)
where A(x) (x)

x and Ay are shown in Eq. 12.22. Similarily
(y)
Ax+y = A(y) (y)
x + Ay (12.18)
where the split matrices are shown in Eq. 12.23.

The permutation relation also holds for the split matrices so
A(x) (y)
y = Pxy Ay Pyx
and
A(x) (y)
x = Pxy Ax Pyx
The splittings in Eqs. 12.17 and 12.18 can be combined with factoring in the
manner described in Section 12.2. As an example (first-order in time), applying the
implicit Euler method to Eq. 12.14 gives

(x)
Un+1 = Un(x) + h A(x) (x) (x)
x + Ay Un+1 + h(bc)
or

I − hA(x)
(x) + O(h2 )
x − hAy
(x)
Un+1 = Un(x) + h(bc) (12.19)
As in Section 12.2, we retain the same first order accuracy with the alternative

I − hA(x) I − hA(x)
(x) + O(h2)
Un+1 = Un(x) + h(bc) (12.20)
x y
Write this in predictor-corrector form and permute the data base of the second row.
There results

I − hA(x) (x)
= Un(x) + h(bc)
x Ũ

(y)
I − hA(y)
y Un+1 = Ũ (y) (12.21)
 
x x | |
x x x | | 
 
 


x x x | | 

 x x | | 
 
 
 
 
 | x x | 
 
 | x x x | 
 
x ·U
A(x)  · U (x)
(x)
= 

 | x x x | 

 | x x | 
 
 
 
 
 | | x x 
 
 
 | | x x x 
 
 | | x x x
| | x x
  (12.22)
o | o |
 o | o | 
 
 


o | o | 

 o | o | 
 
 
 
 
o | o | o 
 
 o | o | o 
 
y ·U
A(x)  · U (x)
(x)
= 

 o | o | o 

 o | o | o
 
 
 
 
 | o | o 
 
 
 | o | o 
 
 | o | o 
| o | o
The splitting of A(x)

x+y .
 
x | x | |
 x | x | | 
 
 
 x | x | | 
 
 
 
x | x | x | 
 
 

 x | x | x | 

 x | x | x | 
 
 
x ·U
A(y)  · U (y)
(y)
= 
 
 | x | x | x 
 
 | | | 
 x x x 
 

 | x | x | x

 
 
 


| | x | x 

 | | x | x 
| | x | x
  (12.23)
o o | | |
o o o | | | 
 
 
 o o | | | 
 
 
 
 | o o | | 
 
 


| o o o | | 

 | o o | | 
 
 
y ·U
A(y)  · U (y)
(y)
= 
 
 | | o o | 
 
 | | o o o | 
 
 

 | | o o | 

 
 
 


| | | o o 

 | | | o o o
| | | o o
(y)
The splitting of Ax+y .
12.4 Second-Order Factored Implicit Methods

Second-order accuracy in time can be maintained in a certain factored implicit meth-
ods. For example, apply the trapezoidal method to Eq. 12.14 where the derivative
operators have been split as in Eq. 12.17 or 12.18. Let the data base be immaterial
be time invariant. There results
and the (bc)

1 1 1 1 + O(h3 )
I − hAx − hAy Un+1 = I + hAx + hAy Un + h(bc) (12.24)
2 2 2 2
Factor both sides giving

1 1 1
I − hAx I − hAy − h2 Ax Ay Un+1
2 2 4

1 1 1 2 + O(h3 )
= I + hAx I + hAy − h Ax Ay Un + h(bc) (12.25)
2 2 4
Then notice that the combination 14 h2 [Ax Ay ](Un+1 − Un ) is proportional to h3 since
the leading term in the expansion of (Un+1 − Un ) is proportional to h. Therefore, we
can write

1 1 1 1 + O(h3)(12.26)
I − hAx I − hAy Un+1 = I + hAx I + hAy Un + h(bc)
2 2 2 2
and both the factored and unfactored form of the trapezoidal method are second-order
accurate in the time march.
An alternative form of this kind of factorization is the classical ADI (alternating
direction implicit) method5 usually written

1 1 1
I − hAx Ũ = I + hAy Un + hFn

2
2
2
1 1 1
I − hAy Un+1 = I + hAx Ũ + hFn+1 + O(h3 ) (12.27)
2 2 2
For idealized commuting systems the methods given by Eqs. 12.26 and 12.27 differ
only in their evaluation of a time-dependent forcing term.
12.5 Importance of Factored Forms in 2 and 3 Di-

mensions
When the time-march equations are stiff and implicit methods are required to permit
reasonably large time steps, the use of factored forms becomes a very valuable tool
5
A form of the Douglas or Peaceman-Rachford methods.
12.5. IMPORTANCE OF FACTORED FORMS IN 2 AND 3 DIMENSIONS 227
for realistic problems. Consider, for example, the problem of computing the time
advance in the unfactored form of the trapezoidal method given by Eq. 12.24

1 1
I − hAx+y Un+1 = I + hAx+y Un + h(bc)
2 2
Forming the right hand side poses no problem, but finding Un+1 requires the solution
of a sparse, but very large, set of coupled simultaneous equations having the matrix
form shown in Eq. 12.16 part a and b. Furthermore, in real cases involving the Euler
or Navier-Stokes equations, each symbol (o, x, •) represents a 4 × 4 block matrix with
entries that depend on the pressure, density and velocity field. Suppose we were to
solve the equations directly. The forward sweep of a simple Gaussian elimination fills6
all of the 4 × 4 blocks between the main and outermost diagonal7 (e.g. between •
and o in Eq. 12.16 part b.). This must be stored in computer memory to be used to
find the final solution in the backward sweep. If Ne represents the order of the small
block matrix (4 in the 2-D Euler case), the approximate memory requirement is
(Ne × My ) · (Ne × My ) · Mx
floating point words. Here it is assumed that My < Mx . If My > Mx , My and Mx
would be interchanged. A moderate mesh of 60 × 200 points would require over 11
million words to find the solution. Actually current computer power is able to cope
rather easily with storage requirements of this order of magnitude. With computing
speeds of over one gigaflop,8 direct solvers may become useful for finding steady-state
solutions of practical problems in two dimensions. However, a three-dimensional
solver would require a memory of approximatly
Ne2 · My2 · Mz2 · Mx
words and, for well resolved flow fields, this probably exceeds memory availability for
some time to come.
On the other hand, consider computing a solution using the factored implicit equa-
tion 12.25. Again computing the right hand side poses no problem. Accumulate the
result of such a computation in the array (RHS). One can then write the remaining
terms in the two-step predictor-corrector form

1 (x) (x)
I − hAx Ũ = (RHS)(x)

2
1 (y) (y)
I − hAy Un+1 = Ũ (y) (12.28)
2
6
For matrices as small as those shown there are many gaps in this “fill”, but for meshes of
practical size the fill is mostly dense.
7
The lower band is also computed but does not have to be saved unless the solution is to be
repeated for another vector.
8
One billion floating-point operations per second.
which has the same appearance as Eq. 12.21 but is second-order time accurate. The
first step would be solved using My uncoupled block tridiagonal solvers9 . Inspecting
the top of Eq. 12.22, we see that this is equivalent to solving My one-dimensional
problems, each with Mx blocks of order Ne . The temporary solution Ũ (x) would then
be permuted to Ũ (y) and an inspection of the bottom of Eq. 12.23 shows that the final
step consists of solving Mx one-dimensional implicit problems each with dimension
My .
12.6 The Delta Form

Clearly many ways can be devised to split the matrices and generate factored forms.
One way that is especially useful, for ensuring a correct steady-state solution in a
converged time-march, is referred to as the “delta form” and we develop it next.
Consider the unfactored form of the trapezoidal method given by Eq. 12.24, and
be time invariant:
let the (bc)

1 1 1 1 + O(h3 )
I − hAx − hAy Un+1 = I + hAx + hAy Un + h(bc)
2 2 2 2
From both sides subtract
1 1
I − hAx − hAy Un
2 2
leaving the equality unchanged. Then, using the standard definition of the difference
operator ∆,
∆Un = Un+1 − Un
one finds

1 1 + O(h3 )
I − hAx − hAy ∆Un = h Ax+y Un + (bc) (12.29)
2 2
Notice that the right side of this equation is the product of h and a term that is
identical to the right side of Eq. 12.14, our original ODE. Thus, if Eq. 12.29 converges,
it is guaranteed to converge to the correct steady-state solution of the ODE. Now we
can factor Eq. 12.29 and maintain O(h2 ) accuracy. We arrive at the expression

1 1 + O(h3)
I − hAx I − hAy ∆Un = h Ax+y Un + (bc) (12.30)
2 2
This is the delta form of a factored, 2nd-order, 2-D equation.
9
A block tridiagonal solver is similar to a scalar solver except that small block matrix operations
replace the scalar ones, and matrix multiplications do not commute.
12.7. PROBLEMS 229
The point at which the factoring is made may not affect the order of time-accuracy,
but it can have a profound effect on the stability and convergence properties of a
method. For example, the unfactored form of a first-order method derived from the
implicit Euler time march is given by Eq. 12.19, and if it is immediately factored,
the factored form is presented in Eq. 12.20. On the other hand, the delta form of the
unfactored Eq. 12.19 is

[I − hAx − hAy ]∆Un = h Ax+y Un + (bc)
and its factored form becomes10

[I − hAx ][I − hAy ]∆Un = h Ax+y Un + (bc) (12.31)
In spite of the similarities in derivation, we will see in the next chapter that the
convergence properties of Eq. 12.20 and Eq. 12.31 are vastly different.
12.7 Problems
1. Consider the 1-D heat equation:
∂u ∂2u
=ν 2 0≤x≤9
∂t ∂x
Let u(0, t) = 0 and u(9, t) = 0, so that we can simplify the boundary conditions.
Assume that second order central differencing is used, i.e.,
1
(δxx u)j = (uj−1 − 2uj + uj+1)
∆x2
The uniform grid has ∆x = 1 and 8 interior points.
(a) Space vector definition

i. What is the space vector for the natural ordering (monotonically in-
creasing in index), u(1) ? Only include the interior points.
ii. If we reorder the points with the odd points first and then the even
points, write the space vector, u(2) ?
iii. Write down the permutation matrices,(P12 ,P21 ).
10
Notice that the only difference between the O(h2 ) method given by Eq. 12.30 and the O(h)
method given by Eq. 12.31 is the appearance of the factor 12 on the left side of the O(h2 ) method.
iv. The generic ODE representing the discrete form of the heat equation
is
du(1)
= A1 u(1) + f
dt
Write down the matrix A1 . (Note f = 0, due to the boundary condi-
tions) Next find the matrix A2 such that
du(2)
= A2 u(2)
dt
Note that A2 can be written as
 
D UT
A2 = 



U D
Define D and U .
v. Applying implicit Euler time marching, write the delta form of the
implicit algorithm. Comment on the form of the resulting implicit
matrix operator.
(b) System definition
In problem 1a, we defined u(1) , u(2) , A1 , A2 , P12 , and P21 which partition
the odd points from the even points. We can put such a partitioning to
use. First define extraction operators
 
1 0 0 0 0 0 0 0
 0 1 0 0 0 0 0 0 
 
   
 0 0 1 0 0 0 0 0 
  I4 04
 0 0 0 1 0 0 0 0 
=
  
I (o) =    
 0 0 0 0 0 0 0 0 
  04 04
 0 0 0 0 0 0 0 0 
 
 0 0 0 0 0 0 0 0 
0 0 0 0 0 0 0 0
 
0 0 0 0 0 0 0 0
 0 0 0 0 0 0 0 0 
 
   
 0 0 0 0 0 0 0 0 
  04 04
 0 0 0 0 0 0 0 0 
=
  
I (e) =    
 0 0 0 0 1 0 0 0 
  04 I4
 0 0 0 0 0 1 0 0 
 
 0 0 0 0 0 0 1 0 
0 0 0 0 0 0 0 1
12.7. PROBLEMS 231
which extract the odd even points from u(2) as follows: u(o) = I (o) u(2) and
u(e) = I (e) u(2).
i. Beginning with the ODE written in terms of u(2), define a splitting
A2 = Ao + Ae , such that Ao operates only on the odd terms, and Ae
operates only on the even terms. Write out the matrices Ao and Ae .
Also, write them in terms of D and U defined above.
ii. Apply implicit Euler time marching to the split ODE. Write down the
delta form of the algorithm and the factored delta form. Comment on
the order of the error terms.
iii. Examine the implicit operators for the factored delta form. Comment
on their form. You should be able to argue that these are now trangu-
lar matrices (a lower and an upper). Comment on the solution process
this gives us relative to the direct inversion of the original system.
Chapter 13
LINEAR ANALYSIS OF SPLIT

AND FACTORED FORMS
In Section 4.4 we introduced the concept of the representative equation, and used
it in Chapter 7 to study the stability, accuracy, and convergence properties of time-
marching schemes. The question is: Can we find a similar equation that will allow
us to evaluate the stability and convergence properties of split and factored schemes?
The answer is yes — for certain forms of linear model equations.
The analysis in this chapter is useful for estimating the stability and steady-state
properties of a wide variety of time-marching schemes that are variously referred
to as time-split, fractional-step, hybrid, and (approximately) factored. When these
methods are applied to practical problems, the results found from this analysis are
neither necessary nor sufficient to guarantee stability. However, if the results indicate
that a method has an instability, the method is probably not suitable for practical
use.
13.1 The Representative Equation for Circulant

Operators
Consider linear PDE’s with coefficients that are fixed in both space and time and with
boundary conditions that are periodic. We have seen that under these conditions
a semi-discrete approach can lead to circulant matrix difference operators, and we
discussed circulant eigensystems1 in Section 4.3. In this and the following section
we assume circulant systems and our analysis depends critically on the fact that all
circulant matrices commute and have a common set of eigenvectors.
1
See also the discussion on Fourier stability analysis in Section 7.7.
233
234 CHAPTER 13. LINEAR ANALYSIS OF SPLIT AND FACTORED FORMS
Suppose, as a result of space differencing the PDE, we arrive at a set of ODE’s

that can be written
du
= Aa pu + Ab pu − f(t) (13.1)
dt
where the subscript p denotes a circulant matrix. Since both matrices have the same
set of eigenvectors, we can use the arguments made in Section 4.2.3 to uncouple the
set and form the M set of independent equations
w1 = (λa + λb )1 w1 − g1 (t)

..
.

wm = (λa + λb )m wm − gm (t)
..
.

wM = (λa + λb )M wM − gM (t) (13.2)
The analytic solution of the m’th line is
wm (t) = cm e(λa +λb )m t + P.S.
Note that each λa pairs with one, and only one2 , λb since they must share a common
eigenvector. This suggests (see Section 4.4:
The representative equation for split, circulant systems is

du
= [λa + λb + λc + · · ·]u + aeµt (13.3)
dt
where λa + λb + λc + · · · are the sum of the eigenvalues in Aa , Ab , Ac , · · · that
share the same eigenvector.
13.2 Example Analysis of Circulant Systems

13.2.1 Stability Comparisons of Time-Split Methods
Consider as an example the linear convection-diffusion equation:
∂u ∂u ∂2u
+a =ν 2 (13.4)
∂t ∂x ∂x
2
This is to be contrasted to the developments found later in the analysis of 2-D equations.
13.2. EXAMPLE ANALYSIS OF CIRCULANT SYSTEMS 235
If the space differencing takes the form

du a ν
=− Bp (−1, 0, 1)u + Bp (1, −2, 1)u (13.5)
dt 2∆x ∆x2
the convection matrix operator and the diffusion matrix operator, can be represented
by the eigenvalues λc and λd , respectively, where (see Section 4.3.2):
ia
(λc )m = sin θm
∆x
4ν θm
(λd )m = − 2 sin2 (13.6)
∆x 2
In these equations θm = 2mπ/M , m = 0 , 1 , · · · , M − 1 , so that 0 ≤ θm ≤ 2π.
Using these values and the representative equation 13.4, we can analyze the stability
of the two forms of simple time-splitting discussed in Section 12.2. In this section we
refer to these as
1. the explicit-implicit Euler method, Eq. 12.10.
2. the explicit-explicit Euler method, Eq. 12.8.
1. The Explicit-Implicit Method

When applied to Eq. 13.4, the characteristic polynomial of this method is
P (E) = (1 − hλd )E − (1 + hλc )
This leads to the principal σ root
ah
1+i sin θm
σ= ∆x
hν θm
1 + 4 2 sin2
∆x 2
where we have made use of Eq. 13.6 to quantify the eigenvalues. Now introduce the
dimensionless numbers
ah
Cn = , Courant number
∆x
a∆x
R∆ = , mesh Reynolds number
ν
and we can write for the absolute value of σ

1 + Cn2 sin2 θm
|σ| = , 0 ≤ θm ≤ 2π (13.7)
Cn θm
1+4 sin2
R∆ 2
1.2 1.2
C = 2/ R
n ∆ C = R /2
n ∆
C = 2/ R
0.8 0.8 n ∆
C C
n n
0.4 0.4
0 0
0 4 8 0 4 8
R∆ R∆
Explicit-Implicit Explicit-Explicit
Figure 13.1: Ṡtability regions for two simple time-split methods.
A simple numerical parametric study of Eq. 13.7 shows that the critical range of
θm for any combination of Cn and R∆ occurs when θm is near 0 (or 2π). From this
we find that the condition on Cn and R∆ that make |σ| ≈ 1 is
2
Cn
1 + Cn2 sin2 = 1 + 4 sin2
R∆ 2
As → 0 this gives the stability region
2
Cn <
R∆
which is bounded by a hyperbola and shown in Fig. 13.1.
2. The Explicit-Explicit Method

An analysis similar to the one given above shows that this method produces

Cn θm
|σ| = 1+ Cn2 2
sin θm 1−4 sin2 , 0 ≤ θm ≤ 2π
R∆ 2
Again a simple numerical parametric study shows that this has two critical ranges
of θm , one near 0, which yields the same result as in the previous example, and the
13.2. EXAMPLE ANALYSIS OF CIRCULANT SYSTEMS 237
other near 180o , which produces the constraint that

1
Cn < R∆ for R∆ ≤ 2
2
The resulting stability boundary is also shown in Fig. 13.1. The totaly explicit,
factored method has a much smaller region of stability when R∆ is small, as we
should have expected.
13.2.2 Analysis of a Second-Order Time-Split Method

Next let us analyze a more practical method that has been used in serious compu-
tational analysis of turbulent flows. This method applies to a flow in which there is
a combination of diffusion and periodic convection. The convection term is treated
explicitly using the second-order Adams-Bashforth method. The diffusion term is
integrated implicitly using the trapezoidal method. Our model equation is again the
linear convection-diffusion equation 13.4 which we split in the fashion of Eq. 13.5. In
order to evaluate the accuracy, as well as the stability, we include the forcing func-
tion in the representative equation and study the effect of our hybrid, time-marching
method on the equation
u = λc u + λd u + aeµt
First let us find expressions for the two polynomials, P (E) and Q(E). The char-
acteristic polynomial follows from the application of the method to the homogeneous
equation, thus
1 1
un+1 = un + hλc (3un − un−1 ) + hλd (un+1 + un)
2 2
This produces
1 3 1 1
P (E) = (1 − hλd )E 2 − (1 + hλc + hλd )E + hλc
2 2 2 2
The form of the particular polynomial depends upon whether the forcing function is
carried by the AB2 method or by the trapezoidal method. In the former case it is
1
Q(E) = h(3E − 1) (13.8)
2
and in the latter
1
Q(E) = h(E 2 + E) (13.9)
2
Accuracy
From the characteristic polynomial we see that there are two σ–roots and they are
given by the equation
5
2
3 1 3 1 1
1 + hλc + hλd ± 1 + hλc + hλd − 2hλc 1 − hλd
σ= 2 2
2 2 2 (13.10)
1
2 1 − hλd
2
The principal σ-root follows from the plus sign and one can show
1 1
σ1 = 1 + (λc + λd )h + (λc + λd )2 h2 + λ3d + λc λ2d − λ2c λd − λ3c h3
2 4
From this equation it is clear that 16 λ3 = 16 (λc + λd )3 does not match the coefficient
of h3 in σ1 , so
erλ = O(h3 )
Using P (eµh ) and Q(eµh ) to evaluate erµ in Section 6.6.3, one can show
erµ = O(h3 )
using either Eq. 13.8 or Eq. 13.9. These results show that, for the model equation,
the hybrid method retains the second-order accuracy of its individual components.
Stability
The stability of the method can be found from Eq. 13.10 by a parametric study of cn
and R∆ defined in Eq. 13.7. This was carried out in a manner similar to that used
to find the stability boundary of the first-order explicit-implicit method in Section
13.2.1. The results are plotted in Fig. 13.2. For values of R∆ ≥ 2 this second-order
method has a much greater region of stability than the first-order explicit-implicit
method given by Eq. 12.10 and shown in Fig. 13.1.
13.3 The Representative Equation for Space-Split

Operators
Consider the 2-D model3 equations
∂u ∂2u ∂2u
= 2+ 2 (13.11)
∂t ∂x ∂y
3
The extension of the following to 3-D is simple and straightforward.
13.3. THE REPRESENTATIVE EQUATION FOR SPACE-SPLIT OPERATORS239
1.2
0.8
Cn
Stable
0.4
0 4 8
R
∆
Figure 13.2: Ṡtability regions for the second-order time-split method.
and
∂u ∂u ∂u
+ ax + ay =0 (13.12)
∂t ∂x ∂y
Reduce either of these, by means of spatial differencing approximations, to the coupled

set of ODE’s:
dU
= [Ax + Ay ]U + (bc) (13.13)
dt
for the space vector U . The form of the Ax and Ay matrices for three-point central
differencing schemes are shown in Eqs. 12.22 and 12.23 for the 3 × 4 mesh shown
in Sketch 12.12. Let us inspect the structure of these matrices closely to see how we
can diagonalize [Ax + Ay ] in terms of the individual eigenvalues of the two matrices
considered separately.
First we write these matrices in the form
   
B b̃0 · I b̃1 · I
 
A(x)
x =  B  A(x)
y =
 b̃−1 · I b̃0 · I b̃1 · I 

B b̃−1 · I b̃0 · I
where B is a banded matrix of the form B(b−1 , b0 , b1 ). Now find the block eigenvector
matrix that diagonalizes B and use it to diagonalize A(x)

x . Thus
     
X −1 B X Λ
   
 X −1  B  X  =
 Λ 

X −1 B X Λ
where  
λ1
 λ2 
Λ= 
 
 λ3 
λ4
Notice that the matrix A(x)
y is transparent to this transformation. That is, if we
set X ≡ diag(X)
   
b̃0 · I b̃1 · I b̃0 · I b̃1 · I
−1   
X  b̃−1 · I b̃0 · I b̃1 · I  X =  b̃−1 · I b̃0 · I b̃1 · I 

b̃−1 · I b̃0 · I b̃−1 · I b̃0 · I
One now permutes the transformed system to the y-vector data-base using the per-
mutation matrix defined by Eq. 12.13. There results

Pyx · X −1 A(x) (x)
x + Ay X · Pxy =
   
λ1 · I B̃
 λ2 · I   
   B̃ 
 +  (13.14)
 λ3 · I  
 B̃ 

λ4 · I B̃
where B̃ is the banded tridiagonal matrix B(b̃−1 , b̃0 , b̃1 ), see the bottom of Eq. 12.23.
Next find the eigenvectors X̃ that diagonalize the B̃ blocks. Let B̃ ≡ diag(B̃) and
X̃ ≡ diag(X̃) and form the second transformation
 
Λ̃  
  λ̃1
 Λ̃ 
Λ̃ =  
X̃ −1 B̃ X̃ =   ,  λ̃2 
 Λ̃ 
λ̃3
Λ̃
This time, by the same argument as before, the first matrix on the right side of
Eq. 13.14 is transparent to the transformation, so the final result is the complete
diagonalization of the matrix Ax+y :

X̃ −1 · Pyx · X −1 Ax+y X · Pxy · X̃ =
(x)
 
λ1 I + Λ̃
 
 λ2 I + Λ̃ 
  (13.15)
 
 λ3 I + Λ̃ 
λ4 I + Λ̃
13.3. THE REPRESENTATIVE EQUATION FOR SPACE-SPLIT OPERATORS241
It is important to notice that:
• The diagonal matrix on the right side of Eq. 13.15 contains every possible com-
bination of the individual eigenvalues of B and B̃.
Now we are ready to present the representative equation for two dimensional sys-
tems. First reduce the PDE to ODE by some choice4 of space differencing. This
results in a spatially split A matrix formed from the subsets
A(x)
x = diag(B) , A(y)
y = diag(B̃) (13.16)
where B and B̃ are any two matrices that have linearly independent eigenvectors (this
puts some constraints on the choice of differencing schemes).
Although Ax and Ay do commute, this fact, by itself, does not ensure the prop-
erty of “all possible combinations”. To obtain the latter property the structure of
the matrices is important. The block matrices B and B̃ can be either circulant or
noncirculant; in both cases we are led to the final result:
The 2–D representative equation for model linear systems is

du
= [λx + λy ]u + aeµt
dt
where λx and λy are any combination of eigenvalues from Ax and Ay , a and µ are
(possibly complex) constants, and where Ax and Ay satisfy the conditions in 13.16.
Often we are interested in finding the value of, and the convergence rate to, the
steady-state solution of the representative equation. In that case we set µ = 0 and
use the simpler form
du
= [λx + λy ]u + a (13.17)
dt
which has the exact solution
a
u(t) = ce(λx +λy )t − (13.18)
λx + λy
4
We have used 3-point central differencing in our example, but this choice was for convenience
only, and its use is not necessary to arrive at Eq. 13.15.
13.4 Example Analysis of 2-D Model Equations

In the following we analyze four different methods for finding a fixed, steady-state
solution to the 2-D representative equation 13.16. In each case we examine
1. The stability.
2. The accuracy of the fixed, steady-state solution.
3. The convergence rate to reach the steady-state.
13.4.1 The Unfactored Implicit Euler Method

Consider first this unfactored, first-order scheme which can then be used as a reference
case for comparison with the various factored ones. The form of the method is given
by Eq. 12.19, and when it is applied to the representative equation, we find
(1 − h λx − h λy )un+1 = un + ha
from which
P (E) = (1 − h λx − h λy )E − 1
Q(E) = h (13.19)
giving the solution

n
1 a
un = c −
1 − h λx − h λy λx + λy
Like its counterpart in the 1-D case, this method:
1. Is unconditionally stable.
2. Produces the exact (see Eq. 13.18) steady-state solution (of the ODE) for any
h.
3. Converges very rapidly to the steady-state when h is large.
Unfortunately, however, use of this method for 2-D problems is generally impractical
for reasons discussed in Section 12.5.
13.4. EXAMPLE ANALYSIS OF 2-D MODEL EQUATIONS 243
13.4.2 The Factored Nondelta Form of the Implicit Euler

Method
Now apply the factored Euler method given by Eq. 12.20 to the 2-D representative
equation. There results
(1 − h λx )(1 − h λy )un+1 = un + ha
from which
P (E) = (1 − h λx )(1 − h λy )E − 1
Q(E) = h (13.20)
giving the solution
n
1 a
un = c −
(1 − h λx )(1 − h λy ) λx + λy − hλx λy
We see that this method:
2. Produces a steady state solution that depends on the choice of h.
3. Converges rapidly to a steady-state for large h, but the converged solution is
completely wrong.
The method requires far less storage then the unfactored form. However, it is not
very useful since its transient solution is only first-order accurate and, if one tries to
take advantage of its rapid convergence rate, the converged value is meaningless.
13.4.3 The Factored Delta Form of the Implicit Euler Method

Next apply Eq. 12.31 to the 2-D representative equation. One finds
(1 − h λx )(1 − h λy )(un+1 − un ) = h(λx un + λy un + a)
which reduces to

(1 − h λx )(1 − h λy )un+1 = 1 + h2 λx λy un + ha
and this has the solution
n
1 + h2 λx λy a
un = c −
(1 − h λx )(1 − h λy ) λx + λy
This method:
2. Produces the exact steady-state solution for any choice of h.
3. Converges very slowly to the steady–state solution for large values of h, since
|σ| → 1 as h → ∞.
Like the factored nondelta form, this method demands far less storage than the un-
factored form, as discussed in Section 12.5. The correct steady solution is obtained,
but convergence is not nearly as rapid as that of the unfactored form.
13.4.4 The Factored Delta Form of the Trapezoidal Method

Finally consider the delta form of a second-order time-accurate method. Apply Eq.
12.30 to the representative equation and one finds

1 1
1 − hλx 1 − hλy (un+1 − un) = h(λx un + λy un + a)
2 2
which reduces to

1 1 1 1
1 − hλx 1 − hλy un+1 = 1 + hλx 1 + hλy un + ha
2 2 2 2
and this has the solution
 n
1 1
 1 + 2 hλx 1 + hλy 
2 a
un = c
  −
1 1 λx + λy
1 − hλx 1 − hλy
2 2
This method:
2. Produces the exact steady–state solution for any choice of h.
3. Converges very slowly to the steady–state solution for large values of h, since
|σ| → 1 as h → ∞.
All of these properties are identical to those found for the factored delta form of the
implicit Euler method. Since it is second order in time, it can be used when time
accuracy is desired, and the factored delta form of the implicit Euler method can be
used when a converged steady-state is all that is required.5 A brief inspection of eqs.
12.26 and 12.27 should be enough to convince the reader that the σ’s produced by
those methods are identical to the σ produced by this method.
5
In practical codes, the value of h on the left side of the implicit equation is literally switched
from h to 12 h.
13.5. EXAMPLE ANALYSIS OF THE 3-D MODEL EQUATION 245
13.5 Example Analysis of the 3-D Model Equation

The arguments in Section 13.3 generalize to three dimensions and, under the condi-
tions given in 13.16 with an A(z) 6
z included, the model 3-D cases have the following
representative equation (with µ = 0):
du
= [λx + λy + λz ]u + a (13.21)
dt
Let us analyze a 2nd-order accurate, factored, delta form using this equation. First
apply the trapezoidal method:
1
un+1 = un + h[(λx + λy + λz )un+1 + (λx + λy + λz )un + 2a]
2
Rearrange terms:

1 1
1 − h(λx + λy + λz ) un+1 = 1 + h(λx + λy + λz ) un + ha
2 2
Put this in delta form:

1
1 − h(λx + λy + λz ) ∆un = h[(λx + λy + λz )un + a]
2
Now factor the left side:

1 1 1
1 − hλx 1 − hλy 1 − hλz ∆un = h[(λx + λy + λz )un + a] (13.22)
2 2 2
This preserves second order accuracy since the error terms
1 2 1 3
h (λx λy + λx λz + λy λz )∆un and h λx λy λz
4 8
are both O(h3). One can derive the characteristic polynomial for Eq. 13.22, find the
σ root, and write the solution either in the form
 n
1 1 2 1 3
1 + h(λx + λy + λz ) + h (λx λy + λx λz + λy λz ) − h λx λy λz 
un = c 2 4 8 
 1 1 2 1 3 
1− h(λx + λy + λz ) + h (λx λy + λx λz + λy λz ) − h λx λy λz
2 4 8
a
− (13.23)
λx + λy + λz
6
Eqs. 13.11 and 13.12, each with an additional term.
or in the form
 n
1 1 1 1
 1 + 2 hλx 1 + hλy 1 + hλz − h3 λx λy λz  a
un = c 2
2
4  − (13.24)
 1 1 1  λx + λy + λz
1 − hλx 1 − hλy 1 − hλz
2 2 2
It is interesting to notice that a Taylor series expansion of Eq. 13.24 results in
1
σ = 1 + h(λx + λy + λz ) + h2 (λx + λy + λz )2 (13.25)
2
1 3 3
+ h λz + (2λy + 2λx ) + 2λ2y + 3λx λy + 2λ2y + λ3y + 2λx λ2y + 2λ2x λy + λ3x + · · ·
4
which verifies the second order accuracy of the factored form. Furthermore, clearly,
if the method converges, it converges to the proper steady-state.7
With regards to stability, it follows from Eq. 13.23 that, if all the λ’s are real and
negative, the method is stable for all h. This makes the method unconditionally stable
for the 3-D diffusion model when it is centrally differenced in space.
Now consider what happens when we apply this method to the biconvection model,
the 3-D form of Eq. 13.12 with periodic boundary conditions. In this case, central
differencing causes all of the λ’s to be imaginary with spectrums that include both
positive and negative values. Remember that in our analysis we must consider every
possible combination of these eigenvalues. First write the σ root in Eq. 13.23 in the
form
1 + iα − β + iγ
σ=
1 − iα − β + iγ
where α, β and γ are real numbers that can have any sign. Now we can always find
one combination of the λ’s for which α, and γ are both positive. In that case since
the absolute value of the product is the product of the absolute values
(1 − β)2 + (α + γ)2
|σ|2 = >1
(1 − β)2 + (α − γ)2
and the method is unconditionally unstable for the model convection problem.
From the above analysis one would come to the conclusion that the method rep-
resented by Eq. 13.22 should not be used for the 3-D Euler equations. In practical
cases, however, some form of dissipation is almost always added to methods that are
used to solve the Euler equations and our experience to date is that, in the presence
of this dissipation, the instability disclosed above is too weak to cause trouble.
7
However, we already knew this because we chose the delta form.
13.6. PROBLEMS 247
13.6 Problems
1. Starting with the generic ODE,
du
= Au + f
dt
we can split A as follows: A = A1 + A2 + A3 + A4 . Applying implicit Euler time
marching gives
un+1 − un
= A1 un+1 + A2 un+1 + A3 un+1 + A4 un+1 + f
h
(a) Write the factored delta form. What is the error term?
(b) Instead of making all of the split terms implicit, leave two explicit:
un+1 − un
= A1 un+1 + A2 un + A3 un+1 + A4 un + f
h
Write the resulting factored delta form and define the error terms.
(c) The scalar representative equation is
du
= (λ1 + λ2 + λ3 + λ4 )u + a
dt
For the fully implicit scheme of problem 1a, find the exact solution to the
resulting scalar difference equation and comment on the stability, conver-
gence, and accuracy of the converged steady-state solution.
(d) Repeat 1c for the explicit-implicit scheme of problem 1b.
Appendix A
USEFUL RELATIONS AND

DEFINITIONS FROM LINEAR
ALGEBRA
A basic understanding of the fundamentals of linear algebra is crucial to our develop-

ment of numerical methods and it is assumed that the reader is at least familar with
this subject area. Given below is some notation and some of the important relations
between matrices and vectors.
A.1 Notation
1. In the present context a vector is a vertical column or string. Thus
 
v1
 v2 
v =  
 .. 
 . 
vm
T
and its transpose v is the horizontal row
v T = [v1 , v2 , v3 , . . . , vm ] , v = [v1 , v2 , v3 , . . . , vm ]T
2. A general m × m matrix A can be written

 
a11 a12 · · · a1m
 a21 a22 · · · a2m 
 
A = (aij ) =  ... 
 
am1 am2 · · · amm
249
250APPENDIX A. USEFUL RELATIONS AND DEFINITIONS FROM LINEAR ALGEBRA
3. An alternative notation for A is

A = a1 , a2 , . . . , am
and its transpose AT is  

aT
 1T 
 
AT = 
a 
 .2 
 . 
 . 
aTm
4. The inverse of a matrix (if it exists) is written A−1 and has the property that
A−1 A = AA−1 = I, where I is the identity matrix.
A.2 Definitions
1. A is symmetric if AT = A.
2. A is skew-symmetric or antisymmetric if AT = −A.
6 6
3. A is diagonally dominant if aii ≥ j=i |aij | , i = 1, 2, . . . , m and aii > j=i |aij |
for at least one i.
4. A is orthogonal if aij are real and AT A = AAT = I
5. Ā is the complex conjugate of A.
6. P is a permutation matrix if P v is a simple reordering of v.
6
7. The trace of a matrix is i aii .
8. A is normal if AT A = AAT .
9. det[A] is the determinant of A.
10. AH is the conjugate transpose of A, (Hermitian).
11. If
a b
A=
c d
then
det[A] = ad − bc
and
−1 1 d −b
A =
det[A] −c a
A.3. ALGEBRA 251
A.3 Algebra
We consider only square matrices of the same dimension.
1. A and B are equal if aij = bij for all i, j = 1, 2, . . . , m.
2. A + (B + C) = (C + A) + B, etc.
3. sA = (saij ) where s is a scalar.
4. In general AB = BA.
5. Transpose equalities:
(A + B)T = AT + B T
(AT )T = A
(AB)T = B T AT
6. Inverse equalities (if the inverse exists):
(A−1 )−1 = A
(AB)−1 = B −1 A−1
(AT )−1 = (A−1 )T
7. Any matrix A can be expressed as the sum of a symmetric and a skew-symmetric

matrix. Thus:
1 1
A = A + AT + A − AT
2 2
A.4 Eigensystems
1. The eigenvalue problem for a matrix A is defined as
Ax = λx or [A − λI]x = 0
and the generalized eigenvalue problem, including the matrix B, as
Ax = λBx or [A − λB]x = 0
2. If a square matrix with real elements is symmetric, its eigenvalues are all real.
If it is asymmetric, they are all imaginary.
3. Gershgorin’s theorem: The eigenvalues of a matrix lie in the complex plane in

the union of circles having centers located by the diagonals with radii equal to
the sum of the absolute values of the corresponding off-diagonal row elements.
4. In general, an m × m matrix A has n x linearly independent eigenvectors with

n x ≤ m and nλ distinct eigenvalues (λi ) with nλ ≤ n x ≤ m.
5. A set of eigenvectors is said to be linearly independent if
a · xm + b · xn = xk , m = n = k
for any complex a and b and for all combinations of vectors in the set.
6. If A posseses m linearly independent eigenvectors then A is diagonalizable, i.e.,
X −1 AX = Λ
where X is a matrix whose columns are the eigenvectors,

X = x1 , x2 , . . . , xm
and Λ is the diagonal matrix

 
λ1 0 ··· 0
 ... .. 
 0 λ2 . 
Λ=
 .. ... ...


 . 0 
0 · · · 0 λm
If A can be diagonalized, its eigenvectors completely span the space, and A is

said to have a complete eigensystem.
7. If A has m distinct eigenvalues, then A is always diagonalizable, and with

each distinct eigenvalue there is one associated eigenvector, and this eigenvector
cannot be formed from a linear combination of any of the other eigenvectors.
8. In general, the eigenvalues of a matrix may not be distinct, in which case the
possibility exists that it cannot be diagonalized. If the eigenvalues of a matrix
are not distinct, but all of the eigenvectors are linearly independent, the matrix
is said to be derogatory, but it can still be diagonalized.
9. If a matrix does not have a complete set of linearly independent eigenvectors,

it cannot be diagonalized. The eigenvectors of such a matrix cannot span the
space and the matrix is said to have a defective eigensystem.
A.4. EIGENSYSTEMS 253
10. Defective matrices cannot be diagonalized but they can still be put into a com-
pact form by a similarity transform, S, such that
 
J1 0 ··· 0
 . . . .. 
 0 J2 . 
J = S −1 AS = 
 .. ... ...


 . 0 
0 · · · 0 Jk
where there are k linearly independent eigenvectors and Ji is either a Jordan

subblock or λi .
11. A Jordan submatrix has the form

 
λi 1 0 ··· 0
 ... .. 
 0 λi 1 . 
 
Ji =  ... 
 0 0 λi 0
 
 .. ... ... 
 . 1
0 ··· 0 0 λi
12. Use of the transform S is known as putting A into its Jordan Canonical form.
A repeated root in a Jordan block is referred to as a defective eigenvalue. For
each Jordan submatrix with an eigenvalue λi of multiplicity r, there exists one
eigenvector. The other r − 1 vectors associated with this eigenvalue are referred
to as principal vectors. The complete set of principal vectors and eigenvectors
are all linearly independent.
13. Note that if P is the permutation matrix

 
0 0 1
P = 0 1 0

 , P T = P −1 = P
1 0 0
then    
λ 1 0 λ 0 0
P 0 λ 1P = 1 λ 0
−1   

0 0 λ 0 1 λ
14. Some of the Jordan subblocks may have the same eigenvalue. For example, the
matrix   
λ1 1
  
 λ1 1  
 λ1 
 
 
 λ1 
 
 λ1 1 
 
 λ1 
 
 λ2 1 
 
 λ2 
λ3
is both defective and derogatory, having:
• 9 eigenvalues
• 3 distinct eigenvalues
• 3 Jordan blocks
• 5 linearly independent eigenvectors
• 3 principal vectors with λ1
• 1 principal vector with λ2
A.5 Vector and Matrix Norms

1. The spectral radius of a matrix A is symbolized by σ(A) such that
σ(A) = |σm |max
where σm are the eigenvalues of the matrix A.
2. A p-norm of the vector v is defined as

 1/p

M
||v||p =  |vj |p 
j=1
3. A p-norm of a matrix A is defined as
||Av||p
||A||p = max
x=0 ||v||p
A.5. VECTOR AND MATRIX NORMS 255
4. Let A and B be square matrices of the same order. All matrix norms must have
the properties
||A|| ≥ 0, ||A|| = 0 implies A = 0

||c · A|| = |c| · ||A||
||A + B|| ≤ ||A|| + ||B||
||A · B|| ≤ ||A|| · ||B||
5. Special p-norms are

6M
||A||1 = maxj=1,···,M i=1 |aij | maximum column sum

T
||A||2 = σ(A · A)
6M
||A||∞ = maxi=1,2,···,M j=1 |aij | maximum row sum
where ||A||p is referred to as the Lp norm of A.
6. In general σ(A) does not satisfy the conditions in 4, so in general σ(A) is not a
true norm.
7. When A is normal, σ(A) is a true norm, in fact, in this case it is the L2 norm.
8. The spectral radius of A, σ(A), is the lower bound of all the norms of A.
Appendix B
SOME PROPERTIES OF
TRIDIAGONAL MATRICES
B.1 Standard Eigensystem for Simple Tridiagonals

In this work tridiagonal banded matrices are prevalent. It is useful to list some of
their properties. Many of these can be derived by solving the simple linear difference
equations that arise in deriving recursion relations.
Let us consider a simple tridiagonal matrix, i.e., a tridiagonal with constant scalar
elements a,b, and c, see Section 3.4. If we examine the conditions under which the
determinant of this matrix is zero, we find (by a recursion exercise)
det[B(M : a, b, c)] = 0
if

√ mπ
b + 2 ac cos =0 , m = 1, 2, · · · , M
M +1
From this it follows at once that the eigenvalues of B(a, b, c) are

√ mπ
λm = b + 2 ac cos , m = 1, 2, · · · , M (B.1)
M +1
The right-hand eigenvector of B(a, b, c) that is associated with the eigenvalue λm
satisfies the equation
B(a, b, c)xm = λmxm (B.2)
and is given by
j −1
xm = (xj )m = a 2 sin j mπ , m = 1, 2, · · · , M (B.3)
c M +1
257
258 APPENDIX B. SOME PROPERTIES OF TRIDIAGONAL MATRICES
These vectors are the columns of the right-hand eigenvector matrix, the elements of
which are
j −1
a 2 sin jmπ j = 1, 2, · · · , M
X = (xjm ) = , (B.4)
c M +1 m = 1, 2, · · · , M
Notice that if a = −1 and c = 1,
j −1
a 2 = ei(j−1) π
2 (B.5)
c
The left-hand eigenvector matrix of B(a, b, c) can be written
m − 1
−1 2 c 2 mjπ m = 1, 2, · · · , M
X = sin ,
M +1 a M +1 j = 1, 2, · · · , M
In this case notice that if a = −1 and c = 1
m − 1
c 2 −i(m−1) π
=e 2 (B.6)
a
B.2 Generalized Eigensystem for Simple Tridiag-

onals
This system is defined as follows
     
b c x1 e f x1
  x    x 
a b c  2  d e f  2 
     
 a b   x3  = λ  d e   x3 
     
 ...   ..   ...  . 
 c .   f   .. 
a b xM d e xM
In this case one can show after some algebra that
det[B(a − λd, b − λe, c − λf ] = 0 (B.7)
if

mπ
b − λm e + 2 (a − λm d)(c − λm f ) cos =0 , m = 1, 2, · · · , M (B.8)
M +1
If we define
mπ
θm = , pm = cos θm
M +1
B.3. THE INVERSE OF A SIMPLE TRIDIAGONAL 259

eb − 2(cd + af )p2m + 2pm (ec − f b)(ea − bd) + [(cd − af )pm ]2
λm =
e2 − 4f dp2m
The right-hand eigenvectors are
j −1
xm = a − λm d 2
sin [jθm ] ,
m = 1, 2, · · · , M
c − λm f j = 1, 2, · · · , M
These relations are useful in studying relaxation methods.
B.3 The Inverse of a Simple Tridiagonal

The inverse of B(a, b, c) can also be written in analytic form. Let DM represent the
determinant of B(M : a, b, c)
DM ≡ det[B(M : a, b, c)]
Defining D0 to be 1, it is simple to derive the first few determinants, thus
D0 = 1
D1 = b
D2 = b2 − ac
D3 = b3 − 2abc (B.9)
One can also find the recursion relation
DM = bDM −1 − acDM −2 (B.10)
Eq. B.10 is a linear O∆E the solution of which was discussed in Section 4.2. Its
characteristic polynomial P (E) is P (E 2 − bE + ac) and the two roots to P (σ) = 0
result in the solution
 √ M +1 √ M +1 
1  b + b − 4ac
2 b − b − 4ac
2 
DM = √ −
b2 − 4ac  2 2 
M = 0, 1, 2, · · · (B.11)
where we have made use of the initial conditions D0 = 1 and D1 = b. In the limiting
case when b2 − 4ac = 0, one can show that
M
b
DM = (M + 1) ; b2 = 4ac
2
Then for M = 4
 
D3 −cD2 c2 D1 −c3 D0

1  −aD2 D1 D2 −cD1 D1 c2 D1 

B −1 =  
D4  a2 D1 −aD1 D1 D2 D1 −cD2 
−a3 D0 a2 D1 −aD2 D3
and for M = 5
 
D4 −cD3 c2 D2 −c3 D1 c4 D0
 −aD D1 D3 −cD1 D2 c2 D1 D1 −c3 D1 
1  2
3 

B −1 =  a D2 −aD1 D2 D2 D2 −cD2 D1 c2 D2 
D5 
 −a3 D1

a2 D1 D1 −aD2 D1 D3 D1 −cD3 
a4 D0 −a3 D1 a2 D2 −aD3 D4
The general element dmn is
Upper triangle:
m = 1, 2, · · · , M − 1 ; n = m + 1, m + 2, · · · , M
dmn = Dm−1DM −n (−c)n−m /DM
Diagonal:
n = m = 1, 2, · · · , M
dmm = DM −1 DM −m /DM
Lower triangle:
m = n + 1, n + 2, · · · , M ; n = 1, 2, · · · , M − 1
dmn = DM −m Dn−1 (−a)m−n /DM
B.4 Eigensystems of Circulant Matrices

B.4.1 Standard Tridiagonals
Consider the circulant (see Section 3.4.4) tridiagonal matrix
Bp (M : a, b, c, ) (B.12)
B.4. EIGENSYSTEMS OF CIRCULANT MATRICES 261
The eigenvalues are

2πm 2πm
λm = b + (a + c) cos − i(a − c) sin , m = 0, 1, 2, · · · , M − 1
M M
(B.13)
The right-hand eigenvector that satisfies Bp (a, b, c)xm = λmxm is
xm = (xj )m = ei j (2πm/M ) , j = 0, 1, · · · , M − 1 (B.14)
√
where i ≡ −1, and the right-hand eigenvector matrix has the form

ij 2πm j = 0, 1, · · · , M − 1
X = (xjm ) = e M ,
m = 0, 1, · · · , M − 1
The left-hand eigenvector matrix with elements x is

−im
2πj
1 M m = 0, 1, · · · , M − 1
X −1 = (xmj ) = e ,
M j = 0, 1, · · · , M − 1
1 X ∗ , where X∗ is the
Note that both X and X −1 are symmetric and that X −1 = M
conjugate transpose of X.
B.4.2 General Circulant Systems

Notice the remarkable fact that the elements of the eigenvector matrices X and X−1
for the tridiagonal circulant matrix given by eq. B.12 do not depend on the elements
a, b, c in the matrix. In fact, all circulant matrices of order M have the same set of
linearly independent eigenvectors, even if they are completely dense. An example of
a dense circulant matrix of order M = 4 is
 
b0 b1 b2 b3
 b b b b 
 3 0 1 2 
  (B.15)
 b2 b3 b0 b1 
b1 b2 b3 b0
The eigenvectors are always given by eq. B.14, and further examination shows that
the elements in these eigenvectors correspond to the elements in a complex harmonic
analysis or complex discrete Fourier series.
Although the eigenvectors of a circulant matrix are independent of its elements,
the eigenvalues are not. For the element indexing shown in eq. B.15 they have the
general form

M −1
λm = bj ei(2πjm/M )
j=0
of which eq. B.13 is a special case.
B.5 Special Cases Found From Symmetries
Consider a mesh with an even number of interior points such as that shown in Fig.
B.1. One can seek from the tridiagonal matrix B(2M : a, b, a, ) the eigenvector subset
that has even symmetry when spanning the interval 0 ≤ x ≤ π. For example, we seek
the set of eigenvectors xm for which
    
b a x1 x1
a b  x2   
 a    x2 
 ...  .   . 
 a  . 
 .  . 
  .  = λm 
. 
 ...  .   . 
 a  .  
 .. 
 
 
 a b a 
 x2 
 
 x2 
a b x1 x1
This leads to the subsystem of order M which has the form
 
b a
 
 a b a 
 
 .. 
 a . 
B(M : a, b, a)xm = 
 ...

 xm = λmxm (B.16)
 a 
 
 
 a b a 
a b+a
By folding the known eigenvectors of B(2M : a, b, a) about the center, one can show
from previous results that the eigenvalues of eq. B.16 are

(2m − 1)π
λm = b + 2a cos , m = 1, 2, · · · , M (B.17)
2M + 1
B.6. SPECIAL CASES INVOLVING BOUNDARY CONDITIONS 263
and the corresponding eigenvectors are

Line of Symmetry

xm = sin j(2m − 1)π , x =0 x=π
2M + 1
j = 1, 2, · · · , M • ◦ ◦ ◦ ◦ ◦ ◦ •

j = 1 2 3 4 5 6
M
Imposing symmetry about the same interval
j= 1 2 3
but for a mesh with an odd number of points,
M
see Fig. B.1, leads to the matrix
  a. An even-numbered mesh
b a
a b a 
  Line of Symmetry
 ... 
 a 

B(M : a, b, a) =  
..  x =0 x=π
 . a 
  • ◦ ◦ ◦ ◦ ◦ •
 a b a
2a b j = 1 2 3 4 5
M
j= 1 2 3
By folding the known eigenvalues of B(2M −
M
1 : a, b, a) about the center, one can show
from previous results that the eigenvalues of b. An odd–numbered mesh
eq. B.17 are Figure B.1 – Symmetrical folds for
special cases

(2m − 1)π
λm = b + 2a cos , m = 1, 2, · · · , M
2M
and the corresponding eigenvectors are

xm = sin j(2m − 1)π , j = 1, 2, · · · , M
2M
B.6 Special Cases Involving Boundary Conditions

We consider two special cases for the matrix operator representing the 3-point central
difference approximation for the second derivative ∂2 /∂x2 at all points away from the
boundaries, combined with special conditions imposed at the boundaries.
Note: In both cases

m = 1, 2, · · · , M
j = 1, 2, · · · , M
−2 + 2 cos(α) = −4 sin2 (α/2)
When the boundary conditions are Dirichlet on both sides,
 
−2 1
 
 1 −2 1 
  λm = −2 + 2 cos Mmπ
 1 −2 1  + 1 (B.18)
  xm = sin j mπ
 
 1 −2 1  M +1
1 −2
When one boundary condition is Dirichlet and the other is Neumann (and a diagonal
preconditioner is applied to scale the last equation),
 
−2 1
  (2m − 1)π
 1 −2 1  λm = −2 + 2 cos 2M + 1
 
 1 −2 1  (B.19)
 

 1 −2 1 

xm = sin j (2m − 1)π
2M + 1
1 −1
Appendix C
THE HOMOGENEOUS
PROPERTY OF THE EULER
EQUATIONS
The Euler equations have a special property that is sometimes useful in constructing
numerical methods. In order to examine this property, let us first inspect Euler’s
theorem on homogeneous functions. Consider first the scalar case. If F (u, v) satisfies
the identity
F (αu, αv) = αn F (u, v) (C.1)
for a fixed n, F is called homogeneous of degree n. Differentiating both sides with

respect to α and setting α = 1 (since the identity holds for all α), we find
∂F ∂F
u +v = nF (u, v) (C.2)
∂u ∂v
Consider next the theorem as it applies to systems of equations. If the vector
F (Q) satisfies the identity
F (αQ) = αn F (Q) (C.3)
for a fixed n, F is said to be homogeneous of degree n and we find

∂F
Q = nF (Q) (C.4)
∂q
265
266APPENDIX C. THE HOMOGENEOUS PROPERTY OF THE EULER EQUATIONS
Now it is easy to show, by direct use of eq. C.3, that both E and F in eqs. 2.11 and
2.12 are homogeneous of degree 1, and their Jacobians, A and B, are homogeneous
of degree 0 (actually the latter is a direct consequence of the former).1 This being
the case, we notice that the expansion of the flux vector in the vicinity of tn which,
according to eq. 6.105 can be written in general as,
E = En + An (Q − Qn ) + O(h2 )
F = Fn + Bn (Q − Qn ) + O(h2 ) (C.5)
can be written
E = An Q + O(h2 )
F = Bn Q + O(h2 ) (C.6)
since the terms En − An Qn and Fn − Bn Qn are identically zero for homogeneous

vectors of degree 1, see eq. C.4. Notice also that, under this condition, the constant
term drops out of eq. 6.106.
As a final remark, we notice from the chain rule that for any vectors F and Q

∂F (Q) ∂F ∂Q ∂Q
= =A (C.7)
∂x ∂Q ∂x ∂x
We notice also that for a homogeneous F of degree 1, F = AQ and

∂F ∂Q ∂A
=A + Q (C.8)
∂x ∂x ∂x
Therefore, if F is homogeneous of degree 1,

∂A
Q=0 (C.9)
∂x
in spite of the fact that individually [∂A/∂x] and Q are not equal to zero.
1
Note that this depends on the form of the equation of state. The Euler equations are homoge-
neous if the equation of state can be written in the form p = ρf (), where is the internal energy
per unit mass.
Bibliography
[1] D. Anderson, J. Tannehill, and R. Pletcher. Computational Fluid Mechanics and

Heat Transfer. McGraw-Hill Book Company, New York, 1984.
[2] C. Hirsch. Numerical Computation of Internal and External Flows, volume 1,2.
John Wiley & Sons, New York, 1988.
267

Fundamentals of CFD

Uploaded by

Copyright:

Available Formats

Fundamentals of CFD

Uploaded by

Document Information

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

Fundamentals of CFD

Uploaded by

Copyright:

Available Formats

Fundamentals of Computational Fluid

Harvard Lomax and Thomas H. Pulliam

August 26, 1999

2 CONSERVATION LAWS AND THE MODEL EQUATIONS 7

4 THE SEMI-DISCRETE APPROACH 51

6 TIME-MARCHING METHODS FOR ODE’S 85

7 STABILITY OF LINEAR SYSTEMS 121

8 CHOICE OF TIME-MARCHING METHODS 149

11 NUMERICAL DISSIPATION 203

12 SPLIT AND FACTORED FORMS 217

13 LINEAR ANALYSIS OF SPLIT AND FACTORED FORMS 233

A USEFUL RELATIONS AND DEFINITIONS FROM LINEAR AL-

B SOME PROPERTIES OF TRIDIAGONAL MATRICES 257

1.2.1 Problem Speciﬁcation and Geometry Preparation

measurements of an existing conﬁguration or may be associated with a design study.

1.2.2 Selection of Governing Equations and Boundary Con-

1.2.3 Selection of Gridding Strategy and Numerical Method

Many diﬀerent gridding strategies exist, including structured, unstructured, hybrid,

1.2.4 Assessment and Interpretation of Results

PDE Partial diﬀerential equation

CONSERVATION LAWS AND

2.1 Conservation Laws

2.2 The Navier-Stokes and Euler Equations

The matrices A = ∂∂Q

a43 = (1 − γ)uv (2.16)

2.3 The Linear Convection Equation

2.3.2 Solution in Wave Space

2.4 The Diﬀusion Equation

giving a steady state-solution which satisﬁes

In two dimensions, the diﬀusion equation becomes

2.4.2 Solution in Wave Space

where κ must be an integer in order to satisfy the boundary conditions. A solution

2.5 Linear Hyperbolic Systems

The elements of w are known as characteristic variables. Each characteristic variable

Assume an ideal gas, p = (γ − 1)(e − ρu2 /2).

2. Find the eigenvalues and eigenvectors of the matrix M derived in question 1.

8. Write the classical wave equation ∂2 u/∂t2 = c2 ∂ 2 u/∂x2 as a ﬁrst-order system,

One approach to solving these equations is to add a time-dependent term and

(a) Find the ﬂux Jacobians of f and g with respect to q.

3.1 Meshes and Finite-Diﬀerence Notation

Figure 3.1: Physical and computational spaces.

diﬀerence approximations. In general, the dependent variables, u, for example, are

j-2 j-1 j j+1 j+2 x

Figure 3.2: Space-time grid arrangement.

For the ordinary time derivative in the study of ODE’s we use

δx ≈ ∂x , δxx ≈ ∂xx (3.6)

∆tn = tn+1 − tn , ∆xj = xj+1 − xj , ∆un = un+1 − un, etc. (3.7)

When there is no subscript on ∆t or ∆x, the spacing is uniform.

3.2 Space Derivative Approximations

Now subtract uj and divide by ∆x to obtain

3.3 Finite-Diﬀerence Operators

3.3.2 Matrix Diﬀerence Operators

Four point mesh. ∆x = π/(M + 1)

arrive at the four equations:

(δxx u)1 = ( ua −2u1 +u2 )/∆x2

we can rewrite Eq. 3.17 as

where the boundary condition is

Then the matrix operator for a three-point central-diﬀerencing scheme at interior

we can approximate the second derivative of u by

(δx u)j = a2 uj−2 + a1 uj−1 + buj + c1 uj+1 (3.30)

The exact solution of Eq. 4.9 is, for µ = λ,