Professional Documents
Culture Documents
I Do Like CFD Vol1 - 2ed - v00
I Do Like CFD Vol1 - 2ed - v00
1
Governing Equations and Exact Solutions
Second Edition
PDF Version
Katate Masatsuka
ii
I do like CFD, VOL.1, Second Edition (PDF Version)
Copyright c 2013 by Katate Masatsuka
The PDF version is sponsored by
All rights reserved. This work is the intellectual property of the author.
All material in this book is legally protected by U.S. & International copyright laws.
Unless otherwise expressly stated, no part of this PDF le may be modied, changed, exploited,
reproduced, republished in any form by any means without a written permission of the author.
Author:
Katate Masatsuka
e-mail: [email protected]
URL: http://www.cfdbooks.com
Comments and suggestions are appreciated.
Please feel free to send messages to the author.
[ Note: Katate Masatsuka is a pen name. ]
Published and distributed by Katate Masatsuka.
This PDF le is hyperlinked:
All internal references such as table of contents, equations, gures, and citations with the back references are
hyperlinked (i.e., click and jump). There are some external links also (i.e., jump to a website).
Hard Copy Version:
Hard copy version is available with extra contents at Lulu.com.
Share "I do like CFD, VOL.1, Second Edition" with your friends who like CFD.
iii
In Memory of Professor Emeritus Haruo Oguro
iv
Contents
Preface xi
1 Basics 1
1.1 Dierential Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Vectors and Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Vectors and Tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.2 Dot/Inner Product and Tensor Product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.3 Vector/Outer Product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.4 Dyadic Tensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.5 Scalar Triple Product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.6 Vector Triple Product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.7 Identity Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Orthogonal Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Index Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4.1 Einsteins Summation Convention . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4.2 Kroneckers Delta . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4.3 Eddingtons Epsilon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5 Div, Grad, Curl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.5.1 Coordinate Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.5.2 Gradient of Scalar Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.5.3 Divergence of Vector Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.5.4 Curl of Vector Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.5.5 Gradient of Vector Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.5.6 Divergence of Tensor Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.5.7 Laplacian of Scalar Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.5.8 Laplacian of Vector Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.5.9 Material/Substantial/Lagrangian Derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.5.10 Complex Lamellar and Beltrami Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.6 Del Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.7 Vector Identities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.8 Coordinate Transformation of Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.9 Coordinate Transformation of Partial Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.10 Eigenvalues and Eigenvectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.11 Similar Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.12 Divergence Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.13 Stokes Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.14 Conservation Laws . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.15 Conservation Laws in Generalized Coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.16 Classication of Systems of Dierential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.16.1 General System of Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.16.2 One-Dimensional Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.16.3 Two-Dimensional Steady Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.17 Classication of Scalar Dierential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.18 Pseudo-Transient Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.19 Operator Splitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
v
vi CONTENTS
1.20 Rotational Invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
1.21 Plane Waves of Hyperbolic Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
1.22 Rankine-Hugoniot Relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
1.23 Linearly Degenerate and Genuinely Nonlinear Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2 Model Equations 45
2.1 Linear Advection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.2 Circular Advection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.3 Diusion Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.4 Diusion Equation: Mixed Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.5 Diusion Equation: First-Order Hyperbolic System . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.6 Advection-Diusion Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.7 Advection-Reaction Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.8 Poisson/Laplace Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.9 Cauchy-Riemann Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.9.1 Velocity Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.9.2 Stream Function and Velocity Potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.9.3 Vorticity and Pressure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.9.4 Time-Dependent Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.10 Biharmonic Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.11 Euler-Tricomi Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.12 Relaxation Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.13 Burgers Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.14 Trac Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.15 Viscous Burgers Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3 Euler Equations 59
3.1 Thermodynamic Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.2 Speed of Sound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.3 Euler Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.4 1D Euler Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.4.1 Conservative Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.4.2 Primitive Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.4.3 Characteristic Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.4.4 Parameter Vector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.5 2D Euler Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.5.1 Conservative Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.5.2 Primitive Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.5.3 Symmetric Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.5.4 Parameter Vector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.6 3D Euler Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.6.1 Conservative Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.6.2 Primitive Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.6.3 Symmetric Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.6.4 Parameter Vector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.6.5 Derivatives of Eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.7 Axisymmetric Euler Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
3.8 Rotational Invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.9 Convenient Ordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.10 Homogeneity Property of the Euler Fluxes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
3.11 Nondimensionalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.11.1 Nondimensionalized Euler Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.11.2 Free Stream Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
3.11.3 Speed of Sound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.11.4 Stagnation Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.12 Change of Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.12.1 One Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
CONTENTS vii
3.12.2 Two Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
3.12.3 Three Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
3.13 Incompressible/Pseudo-Compressible Euler Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
3.14 Homentropic/Isothermal Euler Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
3.14.1 1D Homentropic/Isothermal Euler Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.14.2 2D Homentropic/Isothermal Euler Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
3.14.3 3D Homentropic/Isothermal Euler Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
3.15 Linear Acoustics Equations (Linearized Euler Equations) . . . . . . . . . . . . . . . . . . . . . . . . . 113
3.15.1 1D Linear Acoustics Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
3.15.2 2D Linear Acoustics Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.15.3 3D Linear Acoustics Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
3.16 Quasi-1D Euler Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
3.17 2D Steady Euler Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
3.18 Bernoullis Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
3.19 Gas Dynamics Equation (Nonlinear Potential Equation) . . . . . . . . . . . . . . . . . . . . . . . . . . 120
3.20 Linear Potential Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
3.21 Small Perturbation Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
3.22 Incompressible Potential Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
3.23 Croccos Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
3.24 Kelvins Circulation Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
3.25 Vorticity Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
3.26 Shallow-Water (or Saint Venant) Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
3.26.1 1D Shallow-Water Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
3.26.2 2D Shallow-Water Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
3.26.3 3D Shallow-Water Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
4 Navier-Stokes Equations 131
4.1 Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4.2 Viscous Stresses and Heat Fluxes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
4.3 Articial Bulk Viscosity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
4.4 Viscosity and Heat Conductivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
4.5 Other Energy Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
4.6 Total Pressure Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
4.7 Zero Divergence and Viscous Force on Solid Body . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
4.8 1D Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
4.9 2D Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
4.10 3D Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
4.11 Axisymmetric Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
4.12 3D Normal Viscous Flux and Jacobian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
4.13 Expanded Form of Viscous Terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
4.14 Nondimensionalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
4.14.1 Nondimensionalized Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
4.14.2 Free Stream Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
4.14.3 Speed of Sound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
4.14.4 Stagnation Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
4.15 Reduced Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
4.16 Quasi-Linear Form of the Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
4.16.1 1D Quasi-Linear Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
4.16.2 2D Quasi-Linear Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
4.16.3 3D Quasi-Linear Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
4.17 Incompressible Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
4.18 2D Incompressible Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
4.19 3D Incompressible Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
4.20 Axisymmetric Incompressible Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
4.21 Pseudo-Compressible Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
4.22 Vorticity Transport Equation and Stream Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
4.23 Vorticity Transport Equation and Vector Potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
viii CONTENTS
4.24 Stream Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
4.24.1 Three-Dimensional Stream Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
4.24.2 Cartesian Coordinates (2D) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
4.24.3 Polar Coordinates (2D) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
4.24.4 Axisymmetric Cylindrical Coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
4.24.5 Axisymmetric Spherical Coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
4.24.6 Axisymmetric (x, y) Coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
4.24.7 Volume Flow and Stream Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
4.25 Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
4.26 2D Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
4.27 3D Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
4.28 Axisymmetric Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
4.29 Boundary Layer Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
5 Turbulence Equations 171
5.1 Averages and Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
5.1.1 Reynolds Averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
5.1.2 Favre Averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
5.1.3 Spatial Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
5.2 Reynolds-Averaged Incompressible Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . 174
5.3 Favre-Averaged Compressible Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . 175
5.4 Filtered Incompressible Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
5.5 Filtered Compressible Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
5.6 Turbulence Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
6 Methods for Exact Solutions 179
6.1 Exact Solutions with Separation of Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
6.1.1 Laplace Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
6.1.2 Diusion Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
6.1.3 Advection-Diusion Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
6.1.4 Advection-Diusion-Reaction Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
6.1.5 Oscillating Solutions with Complex Variables: Diusion . . . . . . . . . . . . . . . . . . . . . . 183
6.1.6 Oscillating Solutions with Complex Variables: Advection-Diusion . . . . . . . . . . . . . . . . 184
6.2 Exact Solutions with Complex Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
6.3 Superposition for Nonlinear Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
6.4 Exact Solutions with Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
6.4.1 Cole-Hopf Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
6.4.2 Hodograph Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
6.5 Exact Solutions for Linear Systems of Conservation Laws . . . . . . . . . . . . . . . . . . . . . . . . . 191
6.6 Simple Wave Solutions for Nonlinear Systems of Conservation Laws . . . . . . . . . . . . . . . . . . . 193
6.7 Some Exact Simple Wave Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
6.7.1 Case 1: r
j
= constant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
6.7.2 Case 2: r
j
U . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
6.7.3 Case 3: Other cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
6.8 Manufactured Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
6.9 Fine Grid Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
7 Exact Solutions 201
7.1 Linear Advection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
7.1.1 1D Linear Advection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
7.1.2 2D Linear Advection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
7.1.3 2D Circular Advection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
7.1.4 3D Circular (Helicoidal) Advection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
7.2 Diusion Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
7.2.1 1D Diusion Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
7.2.2 2D Diusion Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
7.2.3 3D Diusion Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
CONTENTS ix
7.3 Advection-Diusion Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
7.3.1 1D Advection-Diusion Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
7.3.2 2D Advection-Diusion Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
7.4 Advection-Reaction Equation(3D) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
7.5 Advection-Diusion-Reaction Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
7.6 Spherical Advection Equation(3D) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
7.7 Burgers Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
7.7.1 1D Burgers Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
7.7.2 1D Burgers Equation: Riemann Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
7.7.3 2D Burgers Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
7.8 Viscous Burgers Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
7.8.1 1D Viscous Burgers Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
7.8.2 1D Viscous Burgers Equation with Source Terms . . . . . . . . . . . . . . . . . . . . . . . . . 219
7.8.3 2D Viscous Burgers Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
7.9 Laplace Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
7.9.1 2D Laplace Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
7.9.2 3D Laplace Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
7.10 Poisson Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
7.10.1 2D Poisson Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
7.10.2 3D Poisson Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
7.11 Incompressible-Euler/Cauchy-Riemann/Laplace Equations . . . . . . . . . . . . . . . . . . . . . . . . . 224
7.11.1 Stagnation Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
7.11.2 Flow over a Rankine Half-Body . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
7.11.3 Cavity Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
7.11.4 Flow over a Circular Cylinder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
7.11.5 Flow over an Elliptic Cylinder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
7.11.6 Flow over an Airfoil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
7.11.7 Flow over Airfoils . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
7.11.8 Fraenkels Flow: Constant Vorticity Flow over a Cylinder . . . . . . . . . . . . . . . . . . . . . 235
7.12 3D Incompressible Euler Equations (Axisymmetric Flows) . . . . . . . . . . . . . . . . . . . . . . . . . 237
7.12.1 Stagnation Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
7.12.2 Flow over a Rankine Half-Body . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
7.12.3 Flow over a Sphere . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
7.13 Euler Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
7.13.1 One-Dimensional Simple Acoustic Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
7.13.2 Entropy Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
7.13.3 Two-Dimensional Unsteady Isentropic Vortex Convection . . . . . . . . . . . . . . . . . . . . . 241
7.13.4 Linearized Potential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
7.13.5 Nozzle Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
7.13.6 Ringlebs Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
7.13.7 Supersonic Flow over a Cylinder (Stagnation Pressure) . . . . . . . . . . . . . . . . . . . . . . 252
7.13.8 Normal/Oblique Shock Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
7.14 Riemann Problems for the Euler Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
7.14.1 Basic Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
7.14.2 Construction of a Riemann Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
7.14.3 Exact Riemann Solver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
7.15 Incompressible Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
7.15.1 Smooth 2D Time-Dependent Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
7.15.2 Smooth 3D Time-Dependent Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
7.15.3 Couette Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
7.15.4 Unsteady Couette Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
7.15.5 Couette-Poiseuille Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
7.15.6 Hagen-Poiseuille Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
7.15.7 Axially Moving Co-centric Cylinders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
7.15.8 Rotating Co-centric Cylinders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
7.15.9 Flat Plate Boundary Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
7.15.10Viscous Stagnation Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
x CONTENTS
7.15.11Stokes Flow over a Sphere . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
7.16 Compressible Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
7.16.1 Compressible Couette Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
7.16.2 Viscous Shock Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
7.16.3 Manufactured Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
Bibliography 277
Extra Contents 287
Preface
Second Edition (2013)
I still like CFD and found more topics that I like after the publication of the rst edition. Also, I received reports
on typos and errors from people who like CFD (Thank you so much!). So, I decided to revise the book. The contents
remain highly biased, but can still be a good reference book for students studying basics of CFD and researchers
developing fundamental CFD algorithms. I hope that the second edition will make them like CFD even more.
Katate Masatsuka
Yorktown, October 2013
First Edition (2009)
CFD: Computational Fluid Dynamics.
I like CFD. I like it so much that I decided to write how much I like CFD. Here, in this volume, I focus on basic
notations and formulas, governing equations, and exact solutions used in CFD. I hope that you like them too.
Basics:
Some basic stu such as notations, formulas, and theorems.
Model Equations:
Model equations commonly used in the algorithm development.
Euler Equations:
The Euler equations and related equations for inviscid ows.
Navier-Stokes Equations:
The Navier-Stokes equations and related equations for viscous ows.
Turbulence Equations:
Averaged Navier-Stokes equations and a little about turbulence models.
Exact Solutions I:
General solutions and techniques for deriving exact solutions.
Exact Solutions II:
Exact solutions for selected equations which can be used for accuracy study.
Naturally, the content is highly biased. This is everything about what I like in CFD. It does not contain any topics
that I dont like. So, this is certainly not a textbook, but can still be a good reference book. In particular, this volume
may be useful for students studying basics of CFD and researchers developing fundamental CFD algorithms (I hope
that they like CFD too).
Katate Masatsuka
Yorktown, February 2009
xi
xii CONTENTS
Chapter 1
Basics
1.1 Dierential Notations
The following notations for derivatives are widely used:
First derivatives:
u
x
=
x
u = u
x
. (1.1.1)
Second derivatives:
2
u
x
2
=
xx
u = u
xx
. (1.1.2)
n-th derivative:
n
u
x
n
=
xx
..
n
u = u
xx
..
n
. (1.1.3)
It is nice to have more than one way to express the same thing. In particular, I like the subscript notation such as
u
x
because it is very easy to write and also takes much less vertical space than others. Of course, it will be very
inconvenient for very high-order derivatives (because the expression will be too long), but such high-order derivatives
do not usually arise in CFD books. pa
Material/Substantial/Lagrangian derivative:
D
Dt
=
t
+u
x
+v
y
+w
z
. (1.1.4)
I like this derivative. It is very interesting. It represents the time rate of change of the quantity when it is being
convected in a ow with the velocity (u, v, w). So, when this derivative is zero, D/Dt = 0, the distribution (e.g., a
contour plot) of (x, y, z) in space moves with the velocity (u, v, w), preserving its initial shape. On the other hand,
when the time derivative is zero,
t
= 0, the distribution of (x, y, z) does not move and stays xed in space for all
times.
1
2 CHAPTER 1. BASICS
1.2 Vectors and Operators
1.2.1 Vectors and Tensors
Vectors and tensors are usually denoted by boldface letters, rather than by arrows put on top of them. It is simple
and very eective. I like it. By the way, in this book, the vector is almost always a column vector and a row vector is
expressed as a transpose of a column vector indicated by the superscript t.
a =
_
_
a
1
a
2
a
3
_
_
, b =
_
_
b
1
b
2
b
3
_
_
, a
t
= [a
1
, a
2
, a
3
], b
t
= [b
1
, b
2
, b
3
]. (1.2.1)
Examples of the second-rank tensors:
A =
_
_
A
11
A
12
A
13
A
21
A
22
A
23
A
31
A
32
A
33
_
_
, A
t
=
_
_
A
11
A
21
A
31
A
12
A
22
A
32
A
13
A
23
A
33
_
_
, (1.2.2)
where A
t
is a transpose of A that is dened as a matrix with rows and columns interchanged.
1.2.2 Dot/Inner Product and Tensor Product
The dot/inner product yields a scalar quantity by combining a pair of vectors as follows:
a b = a
t
b = a
1
b
1
+a
2
b
2
+a
3
b
3
, (1.2.3)
a
2
= a a = a
t
a, (1.2.4)
or a pair of tensors as follows:
A : B =
all i,j
A
ij
B
ij
= A
11
B
11
+A
12
B
12
+A
13
B
13
+A
21
B
21
+A
22
B
22
+A
23
B
23
+A
31
B
31
+A
32
B
32
+A
33
B
33
. (1.2.5)
The latter is often called the tensor product. Two dots are used for the tensor product because it is summed over two
indices, perhaps. In any case, it produces a single real number from vectors/tensors no matter how many components
they have. This is very nice. Of course, we have
a b = b a, A : B = B : A. (1.2.6)
1.2.3 Vector/Outer Product
The vector/outer product produces a vector:
a b =
e
1
e
2
e
3
a
1
a
2
a
3
b
1
b
2
b
3
= (a
2
b
3
a
3
b
2
)e
1
+ (a
3
b
1
a
1
b
3
)e
2
+ (a
1
b
2
a
2
b
1
)e
3
=
_
_
a
2
b
3
a
3
b
2
a
3
b
1
a
1
b
3
a
1
b
2
a
2
b
1
_
_
, (1.2.7)
where e
1
, e
2
, e
3
are the unit vectors, i.e., a = a
1
e
1
+ a
2
e
2
+ a
3
e
3
and b = b
1
e
1
+ b
2
e
2
+ b
3
e
3
. I like the fact that
a b and b a represent two dierent vectors, pointing opposite directions with the same magnitude. It is very
geometrical and easy to visualize.
1.2. VECTORS AND OPERATORS 3
1.2.4 Dyadic Tensor
The dyadic tensor, denoted by , is a second-rank tensor made out of two vectors:
ab = ab
t
=
_
_
a
1
a
2
a
3
_
_
[b
1
, b
2
, b
3
] =
_
_
a
1
b
1
a
1
b
2
a
1
b
3
a
2
b
1
a
2
b
2
a
2
b
3
a
3
b
1
a
3
b
2
a
3
b
3
_
_
, (1.2.8)
so, the (i, j)-th component is given by
(ab)
ij
= a
i
b
j
. (1.2.9)
In fact, the dyadic tensor is dened, with c = [c
1
, c
2
, c
3
]
t
, by
(ab)c = a(b c), (1.2.10)
from which you can gure out what ab should be, i.e., write out the right hand side in components and rearrange
the result in the form as in the left hand side to nd Equation (1.2.8). Of course, we have in general,
ab = ba. (1.2.11)
Unlike the vector product, the geometrical meaning of this is not immediately clear to me. But I like it. It is simple
enough.
1.2.5 Scalar Triple Product
I like the scalar triple product dened for three vectors a, b, and c by
a (b c). (1.2.12)
It is, in fact, equal to the volume of the parallelepiped formed by the vectors a, b, and c. I like it very much because
then I can easily understand various formulas associated with the scalar triple product. For example, we have
a (b c) = b (c a) = c (a b), (1.2.13)
and of course
(b c) a = (c a) b = (a b) c, (1.2.14)
which are trivial relations because all represent the same volume. It is also easy to understand that the volume
represented by the scalar triple product is a signed volume, meaning that switching the two vectors in the vector
product results in the negative of the scalar triple product (because the vector product points the opposite direction):
a (b c) = a (c b), (1.2.15)
b (c a) = b (a c), (1.2.16)
c (a b) = c (b a). (1.2.17)
As you might expect and can easily verify, the scalar triple product can be expressed by the determinant:
a (b c) =
a
1
a
2
a
3
b
1
b
2
b
3
c
1
c
2
c
3
. (1.2.18)
It is quite easy to see that the scalar triple product vanishes if the two vectors in the vector product are the same
(because then the vector product becomes the zero vector). It is also quite easy to understand that the scalar triple
product vanishes if any two vectors are identical, e.g.,
a (a c) = a (b a) = 0. (1.2.19)
4 CHAPTER 1. BASICS
It vanishes, of course, because the parallelepiped becomes at in that case and the volume ( = the scalar triple product)
becomes zero. I now tell you why I really like the scalar triple product. I like it because the volume V of a tetrahedron
formed by the three vectors, a, b, and c, is given by
V =
1
6
|a (b c)| , (1.2.20)
and a tetrahedron is one of my favorite computational cells in CFD. Note that the vectors a, b, and c are the edge
vectors centered at a common vertex of the tetrahedron. The volume is dened with the absolute value of the scalar
triple product to keep it positive in case any of the vectors points the opposite direction. Of course, then, it is easy to
understand that the scalar triple product does not change at all under rotation of coordinate system:
a (b c) = a
(b
), (1.2.21)
where a
, b
, and c
= Ra, b
= Rb, c
= Rc. (1.2.22)
This is natural because rotating a parallelepiped (or a tetrahedron) does not change its volume.
1.2.6 Vector Triple Product
Of course, there is a vector triple product. The vector triple product is dened by
a (b c). (1.2.23)
The following formula is known as Lagranges formula or vector product expansion:
a (b c) = b(a c) c(a b). (1.2.24)
It is known also as the BAC-CAB identity. That is useful for memorizing the formula. But thats not the reason that I
like the vector triple product. I like it because it is useful in CFD. To see how useful it is, consider the dot product of
the vector triple product and a:
[a (b c)] a = 0. (1.2.25)
Yes, it is zero because it is a scalar triple product with two identical vectors: a, bc, and a. It means that the vector
triple product of a, b, and c is perpendicular to a. That is interesting. In CFD, we sometimes need to nd vectors
tangent to a face of a computational cell from the face normal vector, or more generally, nd a vector perpendicular to
a given vector. For example, the situation arises in the rotated-hybrid Riemann solvers [112]. Usually in nite-volume
methods, a single numerical ux is computed in the direction normal to the face, n, where the hat indicates that
the vector is a unit vector. The rotated-hybrid method computes the face-normal numerical ux as a combination of
two numerical uxes by breaking n into two orthogonal directions and applying two dierent numerical uxes in the
two directions. The method applies a robust but dissipative numerical ux in a physically meaningful direction, n
1
,
e.g., normal to a shock wave, and a less robust but much less dissipative numerical ux in the direction n
2
that is
perpendicular to n
1
. The resulting ux has been shown to be very robust for strong shocks without adding too much
dissipation in smooth regions. Since n
1
is determined by a physical consideration, the problem is to nd n
2
that is
perpendicular to a given vector, n
1
. Then, we can nd n
2
by the vector triple product [112]:
n
2
=
n
1
( n n
1
)
| n
1
( n n
1
)|
, (1.2.26)
where the denominator is the magnitude of the numerator to make n
2
a unit vector. Note that the numerator is not
necessarily a unit vector. Clearly, we have n
2
n
1
= 0, i.e., n
2
is perpendicular to n
1
as desired. That is nice. I like
the vector triple product very much.
1.2.7 Identity Matrix
This is a matrix whose diagonal entries are all 1 and other entries are zero. Such a matrix is called the identity matrix.
For example, the 3 3 identity matrix is given by
I =
_
_
1 0 0
0 1 0
0 0 1
_
_
. (1.2.27)
1.3. ORTHOGONAL VECTORS 5
Note that the n n identity matrix has n
2
n zeroes. This means that the number of zeroes increases quadratically
with n. Of course, I like the identity matrix. In particular, I like its alternative notation which is very compact. For
example, the 3 3 identity matrix above is denoted by
I = diag(1, 1, 1). (1.2.28)
This is very useful when we write the identity matrix of a very large size. Why is it called the identity matrix? Well, I
think its because we have
IA = AI = A, (1.2.29)
for any matrix A in any dimension.
1.3 Orthogonal Vectors
Consider three orthogonal unit vectors, n, , m (see Figure 1.3.1):
n n = = m m = 1, (1.3.1)
n = m = m n = 0, (1.3.2)
n = m, (1.3.3)
= mn, (1.3.4)
m = n , (1.3.5)
nn + +mm = I. (1.3.6)
Suppose that n = [n
x
, n
y
, n
z
]
t
, = [
x
,
y
,
z
]
t
, m = [m
x
, m
y
, m
z
]
t
. Then, the above relations can be expanded as
follows:
n
2
x
+n
2
y
+n
2
z
= 1, (1.3.7)
2
x
+
2
y
+
2
z
= 1, (1.3.8)
m
2
x
+m
2
y
+m
2
z
= 1, (1.3.9)
n
x
x
+n
y
y
+n
z
z
= 0, (1.3.10)
x
m
x
+
y
m
y
+
z
m
z
= 0, (1.3.11)
m
x
n
x
+m
y
n
y
+m
z
n
z
= 0, (1.3.12)
y
m
z
z
m
y
= n
x
,
z
m
x
x
m
z
= n
y
,
x
m
y
y
m
x
= n
z
,
n
y
z
n
z
y
= m
x
,
n
z
x
n
x
z
= m
y
,
n
x
y
n
y
x
= m
z
,
m
y
n
z
m
z
n
y
=
x
,
m
z
n
x
m
x
n
z
=
y
,
m
x
n
y
m
y
n
x
=
z
,
(1.3.13)
n
2
x
+
2
x
+m
2
x
= 1, (1.3.14)
n
2
y
+
2
y
+m
2
y
= 1, (1.3.15)
n
2
z
+
2
z
+m
2
z
= 1, (1.3.16)
n
x
n
y
+
x
y
+m
x
m
y
= 0, (1.3.17)
n
x
n
z
+
x
z
+m
x
m
z
= 0, (1.3.18)
n
y
n
z
+
y
z
+m
y
m
z
= 0. (1.3.19)
Introduce an arbitrary vector v = [u, v, w]
t
, and dene the following quantities,
q
n
= v n, q
= v , q
m
= v m, (1.3.20)
6 CHAPTER 1. BASICS
m
n
+q
m
m = v, (1.3.21)
q
2
n
+q
2
+q
2
m
= v
2
. (1.3.22)
Equation (1.3.21) can be expanded as follows:
q
n
n
x
+q
x
+q
m
m
x
= u, (1.3.23)
q
n
n
y
+q
y
+q
m
m
y
= v, (1.3.24)
q
n
n
z
+q
z
+q
m
m
z
= w. (1.3.25)
These relationships can be dierentiated as follows:
dq
n
n
x
+dq
m
m
x
+dq
x
= du, (1.3.26)
dq
n
n
y
+dq
m
m
y
+dq
y
= dv, (1.3.27)
dq
n
n
z
+dq
m
m
z
+dq
z
= dw. (1.3.28)
q
n
dq
n
+q
dq
+q
m
dq
m
= udu +v dv +wdw. (1.3.29)
I like these relations because they can be very useful when I write a three-dimensional CFD code with v as a velocity
vector, and n, , m as a normal vector of a cell face and two tangent vectors respectively. In particular, these relations
can be used to eliminate tangent vectors from expressions such as the absolute value of a Jacobian matrix projected
along a normal vector which arises, for example, in the Roe solver [128]. See Sections 3.6.1 and 3.13.
Now consider a symmetric second-rank tensor :
=
_
_
xx
xy
xz
yx
yy
yz
zx
zy
zz
_
_
, (1.3.30)
1.4. INDEX NOTATION 7
where
yx
=
xy
,
zx
=
xz
, and
zy
=
yz
. We dene the following quantities:
nx
=
xx
n
x
+
xy
n
y
+
xz
n
z
, (1.3.31)
ny
=
yx
n
x
+
yy
n
y
+
yz
n
z
, (1.3.32)
nz
=
zx
n
x
+
zy
n
y
+
zz
n
z
, (1.3.33)
nn
=
nx
n
x
+
ny
n
y
+
nz
n
z
, (1.3.34)
n
=
nx
x
+
ny
y
+
nz
z
, (1.3.35)
nm
=
nx
m
x
+
ny
m
y
+
nz
m
z
, (1.3.36)
nv
=
nx
u +
ny
v +
nz
w. (1.3.37)
Then, we have the following identities:
nn
n
x
+
nm
m
x
+
n
x
=
nx
, (1.3.38)
nn
n
y
+
nm
m
y
+
n
y
=
ny
, (1.3.39)
nn
n
z
+
nm
m
z
+
n
z
=
nz
, (1.3.40)
nn
d
nn
+
nm
d
nm
+
n
d
n
=
nx
d
nx
+
ny
d
ny
+
nz
d
nz
, (1.3.41)
nn
dq
n
+
nm
dq
m
+
n
dq
=
nx
du +
ny
dv +
nz
dw, (1.3.42)
d
nn
n
x
+d
nm
m
x
+d
n
x
= d
nx
, (1.3.43)
d
nn
n
y
+d
nm
m
y
+d
n
y
= d
ny
, (1.3.44)
d
nn
n
z
+d
nm
m
z
+d
n
z
= d
nz
. (1.3.45)
I like these identities very much because they can be useful for eliminating tangent vectors from expressions such as
the absolute value of a Jacobian matrix projected along a normal vector that arises, for example, in the upwind viscous
ux [106].
1.4 Index Notation
I like index notation because it is compact and useful especially for proving vector identities. Also, I can save my
time and money because I dont need to write for summations any more.
1.4.1 Einsteins Summation Convention
Einsteins summation convention means that whenever two indices are repeated, summation is taken over that index:
for example,
a
i
b
i
= a
1
b
1
+a
2
b
2
+a
3
b
3
, (1.4.1)
which is nice because the dot product a b can be simply written as a
i
b
i
without using boldface letters or . The i-th
component of the matrix-vector product Ab can be expressed as
(Ab)
i
= A
ij
b
j
= A
i1
b
1
+A
i2
b
2
+A
i3
b
3
, (1.4.2)
where (Ab)
i
denotes the i-th component of Ab. Thus, we can write
Ab =
_
_
(Ab)
1
(Ab)
2
(Ab)
3
_
_
=
_
_
A
11
b
1
+A
12
b
2
+A
13
b
3
A
21
b
1
+A
22
b
2
+A
23
b
3
A
31
b
1
+A
32
b
2
+A
33
b
3
_
_
. (1.4.3)
Of course, the index used for the summation is a dummy, and so we can use any letter we like, e.g.,
a b = a
i
b
i
= a
j
b
j
= a
k
b
k
, (1.4.4)
which are all equal to a
1
b
1
+a
2
b
2
+a
3
b
3
.
8 CHAPTER 1. BASICS
1.4.2 Kroneckers Delta
Of course, I like the second-rank tensor called Kroneckers delta:
ij
=
_
1 if i = j,
0 if i = j.
(1.4.5)
This is nothing but the components of the identity matrix:
I =
_
_
11
12
13
21
22
23
31
32
33
_
_
=
_
_
1 0 0
0 1 0
0 0 1
_
_
. (1.4.6)
I like Kroneckers delta especially because it is sometimes called the substitution operator base on the following property:
ij
a
j
= a
i
, (1.4.7)
which looks like i is substituted into j. This is nice also because
ij
has disappeared on the right hand side, making
the right hand side very simple. By the substitution, we have
ij
a
i
b
j
= a
i
b
i
, (1.4.8)
which means
I : (ab) = a
t
I b = a b. (1.4.9)
In a way, we have just proved a vector identity. It is also easy to show that
ii
= 3, (1.4.10)
which is the trace of the identity matrix, and also that
ij
jk
=
ik
, (1.4.11)
i.e., the product of two identity matrices is the identity matrix:
II = I. (1.4.12)
1.4.3 Eddingtons Epsilon
The third-rank tensor called Eddingtons epsilon:
ijk
=
_
_
_
0 if any two of i, j, k are the same,
1 for even permutation,
1 for odd permutation,
(1.4.13)
reminds me of the Rubik cube that I like. By denition, we have
ijk
=
jik
=
ikj
, (1.4.14)
ijk
=
jki
=
kij
. (1.4.15)
And we have naturally
ij
ijk
= 0, (1.4.16)
which is because Kroneckers delta turns
ijk
into
iik
that is zero by denition. I would use E to denote the third-rank
tensor of Eddingtons epsilon. Then, I have
E
1jk
_
_
111
112
113
121
122
123
131
132
133
_
_
=
_
_
0 0 0
0 0 1
0 1 0
_
_
, (1.4.17)
E
2jk
_
_
211
212
213
221
222
223
231
232
233
_
_
=
_
_
0 0 1
0 0 0
1 0 0
_
_
, (1.4.18)
E
3jk
_
_
311
312
313
321
322
323
331
332
333
_
_
=
_
_
0 1 0
1 0 0
0 0 0
_
_
, (1.4.19)
1.4. INDEX NOTATION 9
which look to me like the slices of the Rubic cube. I like E especially because I can write the vector product as
a b = E(ab). (1.4.20)
As you might expect, then, Eddingtons epsilon can be used to express the determinant of A,
detA =
ijk
A
1i
A
2j
A
3k
, (1.4.21)
or of three vectors,
ijk
a
i
b
j
c
k
=
a
1
a
2
a
3
b
1
b
2
b
3
c
1
c
2
c
3
. (1.4.22)
Even more interestingly, it can be rewritten as
ijk
a
i
b
j
c
k
= a
i
i1
i2
i3
b
1
b
2
b
3
c
1
c
2
c
3
= a
i
b
j
i1
i2
i3
j1
j2
j3
c
1
c
2
c
3
= a
i
b
j
c
k
i1
i2
i3
j1
j2
j3
k1
k2
k3
, (1.4.23)
which means
ijk
=
i1
i2
i3
j1
j2
j3
k1
k2
k3
. (1.4.24)
I like this formula because I can use it to derive various identities. Consider
ijk
lmn
=
i1
i2
i3
j1
j2
j3
k1
k2
k3
l1
l2
l3
m1
m2
m3
n1
n2
n3
, (1.4.25)
which can be written as
ijk
lmn
=
i1
i2
i3
j1
j2
j3
k1
k2
k3
l1
m2
n3
l1
m2
n3
l1
m2
n3
ir
lr
ir
mr
ir
nr
jr
lr
jr
mr
jr
nr
kr
lr
kr
mr
kr
nr
. (1.4.26)
Therefore, we obtain by Equation (1.4.11)
ijk
lmn
=
il
im
in
jl
jm
jn
kl
km
kn
. (1.4.27)
This is a very useful formula. For example, with l = k, it becomes
ijk
kmn
=
kk
im
in
jm
jn
jk
im
in
km
kn
+
ik
jm
jn
km
kn
= 3
im
in
jm
jn
im
in
jk
km
jk
kn
jm
jn
ik
km
ik
kn
= 3
im
in
jm
jn
im
in
jm
jn
jm
jn
im
in
= 3
im
in
jm
jn
im
in
jm
jn
im
in
jm
jn
, (1.4.28)
10 CHAPTER 1. BASICS
and thus
ijk
kmn
=
im
in
jm
jn
=
im
jn
in
jm
, (1.4.29)
which is sometimes called the - identity. Furthermore, with l = i and m = j, Equation (1.4.27) becomes
ijk
ijn
= 3
kn
kn
= 2
kn
. (1.4.30)
Finally, by setting l = i, m = j, and n = k, we obtain
ijk
ijk
= 6. (1.4.31)
These relations can be used to prove vector identities. For example, the i-th component of the triple vector product
a (b c) can be expressed as
(a (b c))
i
=
ijk
a
j
(
klm
b
l
c
m
), (1.4.32)
which becomes by the formula (1.4.29)
(a (b c))
i
=
ijk
klm
a
j
b
l
c
m
=
kij
klm
a
j
b
l
c
m
= (
il
jm
jl
im
)a
j
b
l
c
m
=
il
jm
a
j
b
l
c
m
jl
im
a
j
b
l
c
m
= a
j
b
i
c
j
a
j
b
j
c
i
= b
i
(a
j
c
j
) c
i
(a
j
b
j
), (1.4.33)
and we have thus proved Lagranges formula,
a (b c) = b(a c) c(a b). (1.4.34)
1.5 Div, Grad, Curl
1.5.1 Coordinate Systems
I like vector dierential operators, especially in Cartesian coordinates (x, y, z). In cylindrical (r, , z) or spherical
coordinates (r, , ) (see Figures 1.5.1 and 1.5.2), I like them, of course, but they can be a little complicated. Note
that these coordinates are related: in cylindrical coordinates,
x = r cos , (1.5.1)
y = r sin , (1.5.2)
z = z, (1.5.3)
while in spherical coordinates,
x = (r sin ) cos , (1.5.4)
y = (r sin ) sin , (1.5.5)
z = r cos . (1.5.6)
Now, denote the unit vectors for these coordinates by
e
x
, e
y
, e
z
, in Cartesian coordinates, (1.5.7)
e
r
, e
, e
z
, in cylindrical coordinates, (1.5.8)
e
r
, e
, e
r
z
(r, , z)
Figure 1.5.1: Cylindrical coordinates.
and vector components as in
a = a
1
e
x
+a
2
e
y
+a
3
e
z
, in Cartesian coordinates, (1.5.10)
a = a
r
e
r
+a
+a
z
e
z
, in cylindrical coordinates, (1.5.11)
a = a
r
e
r
+a
+a
ndS, (1.5.13)
or simply by
d = grad dx, (1.5.14)
where
dx = dxe
x
+dy e
y
+dz e
z
, in Cartesian coordinates, (1.5.15)
dx = dr e
r
+rd e
+dz e
z
, in cylindrical coordinates, (1.5.16)
dx = dr e
r
+r sin d e
+r de
z
(r, , )
r
+
z
e
z
, in cylindrical coordinates, (1.5.19)
grad =
r
e
r
+
1
r sin
+
1
r
a ndS, (1.5.21)
and yields
div a =
a
1
x
+
a
2
y
+
a
3
z
in Cartesian coordinates, (1.5.22)
div a =
1
r
(ra
r
)
r
+
1
r
a
+
a
z
z
in cylindrical coordinates, (1.5.23)
div a =
1
r
2
(r
2
a
r
)
r
+
1
r sin
a
+
1
r sin
(a
sin )
a ndS, (1.5.25)
which becomes
curl a =
e
x
e
y
e
z
z
a
1
a
2
a
3
e
r
re
e
z
z
a
r
ra
a
z
e
r
re
r sin e
a
r
ra
sin a
(ra
)
z
_
e
r
+r
_
a
r
z
a
z
r
_
e
+
_
(ra
)
r
a
r
_
e
z
_
, (1.5.30)
curl a =
1
r
2
sin
__
(ra
sin )
(ra
_
e
r
+r sin
_
(ra
)
r
a
r
_
e
+ r
_
a
r
(ra
sin )
r
_
e
_
=
1
r sin
_
(a
sin )
_
e
r
+
1
r
_
(ra
)
r
a
r
_
e
+
1
r
_
1
sin
a
r
(ra
)
r
_
e
. (1.5.31)
1.5.5 Gradient of Vector Functions
The gradient of a vector function a(x, y, z) is dened, as in the scalar case, by
grad a = lim
V 0
1
V
andS, (1.5.32)
or simply by
da = (grad a) dx. (1.5.33)
14 CHAPTER 1. BASICS
In Cartesian coordinates:
grad a =
_
_
a
1
x
a
1
y
a
1
z
a
2
x
a
2
y
a
2
z
a
3
x
a
3
y
a
3
z
_
_
. (1.5.34)
In cylindrical coordinates:
grad a =
_
_
a
r
r
1
r
a
r
r
a
r
z
a
r
1
r
a
+
a
r
r
a
z
a
z
r
1
r
a
z
a
z
z
_
_
. (1.5.35)
In spherical coordinates:
grad a =
_
_
a
r
r
1
r sin
a
r
r
1
r
a
r
r
a
r
1
r sin
a
+
a
r
r
+
a
r tan
1
r
a
r
1
r sin
a
r tan
1
r
a
+
a
r
r
_
_
. (1.5.36)
I like the gradient of a vector, especially those extra terms arising in cylindrical and spherical coordinates. They
make the formulas look very special. These terms come from the fact that some of the unit vectors vary in space; their
derivatives are non-zero: in cylindrical coordinates,
e
r
= e
,
e
= e
r
, (1.5.37)
and in spherical coordinates,
e
r
= sin e
,
e
= cos e
sin e
r
,
e
= cos e
,
e
r
= e
,
e
= e
r
. (1.5.38)
For example, in cylindrical coordinates, the extra terms arise as follows. Consider the vector,
a = a
r
e
r
+a
+a
z
e
z
. (1.5.39)
Take the dierential of a,
da =
_
a
r
r
dr +
a
r
r
rd +
a
r
z
dz
_
e
r
+a
r
_
e
r
r
dr +
e
r
r
rd +
e
r
z
dz
_
+
_
a
r
dr +
a
r
rd +
a
z
dz
_
e
+a
r
dr +
e
r
rd +
z
dz
_
+
_
a
z
r
dr +
a
z
r
rd +
a
z
z
dz
_
e
z
+a
z
_
e
z
r
dr +
e
z
r
rd +
e
z
z
dz
_
,
1.5. DIV, GRAD, CURL 15
which by Equation(1.5.37) becomes
da =
_
a
r
r
dr +
a
r
r
rd +
a
r
z
dz
_
e
r
+
_
a
r
r
_
rd e
+
_
a
r
dr +
a
r
rd +
a
z
dz
_
e
_
a
r
_
rd e
r
+
_
a
z
r
dr +
a
z
r
rd +
a
z
z
dz
_
e
z
=
_
a
r
r
dr +
_
a
r
r
a
r
_
rd +
a
r
z
dz
_
e
r
+
_
a
r
dr +
_
a
r
+
a
r
r
_
rd +
a
z
dz
_
e
+
_
a
z
r
dr +
a
z
r
rd +
a
z
z
dz
_
e
z
. (1.5.40)
We nd the gradient by writing the result in the form, da = (grad a) dx, where dx = (dr, rd, dz):
da =
_
_
a
r
r
1
r
a
r
r
a
r
z
a
r
1
r
a
+
a
r
r
a
z
a
z
r
1
r
a
z
a
z
z
_
_
_
_
dr
rd
dz
_
_
. (1.5.41)
That is, the matrix above is the gradient of a in cylindrical coordinates. I like deriving, in the same way, the gradient
in spherical coordinates also, but Ill let you do it because it is simple and fun. Enjoy.
The divergence of a tensor will also have extra terms and the Laplacian of a vector will also have such in these
coordinates as we will see below.
1.5.6 Divergence of Tensor Functions
The divergence of a tensor function A(x, y, z) is dened, in the same way as in the scalar case, by
div A = lim
V 0
1
V
AndS. (1.5.42)
In Cartesian coordinates:
div A =
_
_
A
11
x
+
A
12
y
+
A
13
z
A
21
x
+
A
22
y
+
A
23
z
A
31
x
+
A
32
y
+
A
33
z
_
_
. (1.5.43)
In cylindrical coordinates:
A =
_
_
A
rr
A
r
A
rz
A
r
A
A
z
A
zr
A
z
A
zz
_
_
, (1.5.44)
16 CHAPTER 1. BASICS
div A =
_
_
1
r
(rA
rr
)
r
+
1
r
A
r
+
A
rz
z
A
r
1
r
(rA
r
)
r
+
1
r
A
+
A
z
z
+
A
r
r
1
r
(rA
zr
)
r
+
1
r
A
z
+
A
zz
z
_
_
. (1.5.45)
In spherical coordinates:
A =
_
_
A
rr
A
r
A
r
A
r
A
A
r
A
_
_
, (1.5.46)
div A =
_
_
1
r
2
(r
2
A
rr
)
r
+
1
r sin
A
r
+
1
r sin
(A
r
sin )
r
A
r
1
r
2
(r
2
A
r
)
r
+
1
r sin
A
+
1
r sin
(A
sin )
+
A
r
r
+
A
r tan
1
r
2
(r
2
A
r
)
r
+
1
r sin
A
+
1
r sin
(A
sin )
r tan
+
A
r
r
_
_
. (1.5.47)
Again, I like those extra terms arising in cylindrical and spherical coordinates. What is interesting is the way they
arise. OK, so let me show you how to derive the divergence of a tensor function in cylindrical coordinates. I would
begin with the denition (1.5.42). Consider a small volume dened by the increment in each coordinate direction:
V = dr (rd) dz = rdrddz. (1.5.48)
Since the volume is small, the surface integral on the right hand side of the denition (1.5.42) can be expressed as
AndS =
(An
r
e
r
dS)
r
dr +
(An
dS)
d +
(An
z
e
z
dS)
z
dz, (1.5.49)
where n = n
r
e
r
+n
+n
z
e
z
. It is actually more convenient to write by dS
r
= n
r
dS, dS
= n
dS, dS
z
= n
z
dS,
AndS =
(AdS
r
e
r
)
r
dr +
(AdS
d +
(AdS
z
e
z
)
z
dz, (1.5.50)
where
dS
r
= rddz, dS
= drdz, dS
z
= rddr. (1.5.51)
Note that the above expression has been obtained by Taylor expansion between the two opposite surface of the volume
and the surface vector has also been expanded together because it can change along a coordinate direction. Each term
can be computed in a straightforward manner:
(AdS
r
e
r
)
r
dr =
r
_
_
_
_
_
A
rr
A
r
A
rz
A
r
A
A
z
A
zr
A
z
A
zz
_
_
_
_
dS
r
0
0
_
_
_
_
_
dr
=
r
{dS
r
[A
rr
e
r
+A
r
e
+A
zr
e
z
]} dr
=
r
{r [A
rr
e
r
+A
r
e
+A
zr
e
z
]} drddz
=
__
(rA
rr
)
r
e
r
+
(rA
r
)
r
e
+
(rA
zr
)
r
e
z
_
+r
_
A
rr
e
r
r
+A
r
r
+A
zr
e
z
r
__
drddz
=
_
1
r
(rA
rr
)
r
e
r
+
1
r
(rA
r
)
r
e
+
1
r
(rA
zr
)
r
e
z
_
V. (1.5.52)
1.5. DIV, GRAD, CURL 17
Similarly, we obtain
(AdS
d =
r
{dS
[A
r
e
r
+A
+A
z
e
z
]} d
=
__
A
r
e
r
+
A
+
A
z
e
z
_
+
_
A
r
e
r
+A
+A
z
e
z
__
drddz,
which becomes by Equation (1.5.37)
(AdS
d =
__
A
r
e
r
+
A
+
A
z
e
z
_
+ (A
r
e
e
r
)
_
drddz
=
__
1
r
A
r
r
_
e
r
+
_
1
r
A
+
A
r
r
_
e
+
1
r
A
z
e
z
_
V. (1.5.53)
The last component is the simplest one:
(AdS
z
e
z
)
z
dz =
_
A
rz
z
e
r
+
A
z
z
e
+
A
zz
z
e
z
_
V. (1.5.54)
Finally, substituting these results into Equation (1.5.50) and subsequently substituting it into the denition (1.5.42),
where V is denoted here by the small volume V , we nd
div A = lim
V 0
1
V
AndS
=
_
1
r
(rA
rr
)
r
+
1
r
A
r
r
+
A
rz
z
_
e
r
+
_
1
r
(rA
r
)
r
+
1
r
A
+
A
r
r
+
A
z
z
_
e
+
_
1
r
(rA
zr
)
r
+
1
r
A
z
+
A
zz
z
_
e
z
, (1.5.55)
which is the same as Equation (1.5.45). We have thus derived the divergence in cylindrical coordinates. As we have
seen above, the extra terms are generated by the derivatives of the unit vectors, which vary in space. I nd it very
interesting. Do you?
1.5.7 Laplacian of Scalar Functions
The Laplacian of a scalar function (x, y, z) is dened as the divergence of the gradient:
div grad =
2
x
2
+
2
y
2
+
2
z
2
, (1.5.56)
div grad =
1
r
r
_
r
r
_
+
1
r
2
2
+
2
z
2
, (1.5.57)
div grad =
1
r
2
r
_
r
2
r
_
+
1
r
2
sin
2
2
+
1
r
2
sin
_
sin
_
, (1.5.58)
in Cartesian, cylindrical, and spherical coordinates respectively. Incidentally, the Laplacian operator is often denoted
by :
= div grad . (1.5.59)
This is a very simple and economical notation. I like it.
18 CHAPTER 1. BASICS
1.5.8 Laplacian of Vector Functions
The Laplacian of a vector function is dened as the divergence of the gradient just like in the scalar case. But it is a
bit more complicated than the scalar case in cylindrical and spherical coordinates.
In Cartesian coordinates:
div grad a =
_
2
a
1
x
2
+
2
a
1
y
2
+
2
a
1
z
2
2
a
2
x
2
+
2
a
2
y
2
+
2
a
2
z
2
2
a
3
x
2
+
2
a
3
y
2
+
2
a
3
z
2
_
_
. (1.5.60)
In cylindrical coordinates:
div grad a =
_
_
div grad a
r
a
r
r
2
r
2
a
div grad a
+
2
r
2
a
r
r
2
div grad a
z
_
_
, (1.5.61)
where the Laplacian for each scalar component is given by Equation (1.5.57).
In spherical coordinates:
div grad a =
_
_
div grad a
r
2a
r
r
2
2
r
2
a
2a
r
2
tan
2
r
2
sin
a
div grad a
+
2
r
2
a
r
r
2
sin
2
2
r
2
sin
a
div grad a
r
2
sin
2
+
2
r
2
sin
2
a
r
+
2 cos
r
2
sin
2
_
,
(1.5.62)
where the Laplacian for each scalar component is given by Equation (1.5.58).
1.5.9 Material/Substantial/Lagrangian Derivative
The material derivative of a scalar function: in Cartesian coordinates,
D
Dt
=
t
+ (grad ) v =
d
dt
+u
d
dx
+v
d
dy
+w
d
dz
, (1.5.63)
in cylindrical coordinates with the velocity (v
r
, v
, v
z
),
D
Dt
=
t
+ (grad ) v =
d
dt
+v
r
r
+
v
+v
z
z
, (1.5.64)
and in spherical coordinates with the velocity (v
r
, v
, v
),
D
Dt
=
t
+ (grad ) v =
d
dt
+v
r
r
+
v
r sin
+
v
. (1.5.65)
The material derivative of a vector function a(x, y, z, t) is simple in Cartesian coordinates:
Da
Dt
=
a
t
+ (grad a)v =
_
_
Da
1
Dt
Da
2
Dt
Da
3
Dt
_
_
, (1.5.66)
1.6. DEL OPERATOR 19
but of course it has some extra terms in cylindrical coordinates,
Da
Dt
=
a
t
+ (grad a)v =
_
_
Da
r
Dt
a
r
Da
Dt
+
a
r
v
r
Da
z
Dt
_
_
, (1.5.67)
where the material derivative for each scalar component is given by Equation (1.5.64), and also in spherical coordinates,
Da
Dt
=
a
t
+ (grad a)v =
_
_
Da
r
Dt
a
+a
r
Da
Dt
+
v
(a
r
+a
cot )
r
Da
Dt
+
a
r
v
cot
r
_
_
, (1.5.68)
where the material derivative for each scalar component is given by Equation (1.5.65).
1.5.10 Complex Lamellar and Beltrami Flows
A vector eld dened by a(x, y, z) which satises
a (curl a) = 0, (1.5.69)
is called a complex lamellar eld. If a is a uid velocity, this corresponds to a ow where the velocity and the vorticity
(the curl of the velocity) are orthogonal to each other. It is then called a complex lamellar ow. I like it because it
sounds like a ow over a lifting wing section. On the other hand, a vector eld that satises
a (curl a) = 0, (1.5.70)
is called a Beltrami eld. If a is a velocity, this corresponds to a ow where the velocity and vorticity are parallel to
each other. This is called a Beltrami ow. I like this one also. It reminds me of trailing vortices of an aircraft.
Consult other books such as Ref.[67] to nd out more details on the vector dierential operators.
1.6 Del Operator
The del operator (which is called nabla as a symbol) in Cartesian coordinates is dened by
=
_
x
,
y
,
z
_
. (1.6.1)
I like the del operator. But some people dont like it because they say it depends on the coordinate system and doesnt
carry much physical meaning while div, grad, and curl are independent of coordinate systems and do carry clear physical
meanings. I know what they mean, and actually like it. Nonetheless, it is a useful operator. Div, grad, and curl can
be expressed by using the del operator as follows:
grad = , (1.6.2)
div a = a, (1.6.3)
curl a = a, (1.6.4)
grad a = a, (1.6.5)
div A = A. (1.6.6)
20 CHAPTER 1. BASICS
What is nice about the del operator is that it can be treated as a vector and thus can be used as follows.
div grad = =
2
= , (1.6.7)
curl(grad ) = , (1.6.8)
grad(div a) = ( a), (1.6.9)
div(curl a) = a, (1.6.10)
curl(curl a) = a, (1.6.11)
(grad a)b = (b ) a. (1.6.12)
This is a great advantage of . But it must be kept in mind that it is actually an operator. That is, a b = b a, but
a = a.
1.7 Vector Identities
Do you like to memorize the following vector identities?
curl a = E(grad a)
t
, (1.7.1)
div curl a = 0, (1.7.2)
curl grad = 0, (1.7.3)
grad () = grad + grad , (1.7.4)
grad (a) = agrad +grad a, (1.7.5)
a curl b = (grad b)
t
a (grad b)a, (1.7.6)
(grad a)a = grad
a a
2
a curl a, (1.7.7)
a (grad a)a = a grad
a a
2
, (1.7.8)
(ab) : grad a = b grad
a a
2
, (1.7.9)
grad (a b) = (grad a)b + (grad b)a +a curl b +b curl a, (1.7.10)
div (a) = div a + grad a, (1.7.11)
div grad a = grad(div a) curl(curl a), (1.7.12)
div (a b) = b curl a a curl b, (1.7.13)
div(ab) = (grad a)b +a div b, (1.7.14)
a div(ab) = div
_
a a
2
b
_
+
a a
2
div b, (1.7.15)
a div(bc) = (a b) div c + (ac) : grad b, (1.7.16)
div (A) = Agrad +div A, (1.7.17)
div (Ab) = (div A
t
) b +A
t
: grad b, (1.7.18)
curl (a) = curl a + grad a, (1.7.19)
curl (a b) = a div b + (grad a)b (div a)b (grad b)a, (1.7.20)
a (Ab) = A : (ab), (1.7.21)
a (Ab) = (A
t
a) b, (1.7.22)
a (Ab) = (Aa) b if A is symmetric, (1.7.23)
1.8. COORDINATE TRANSFORMATION OF VECTORS 21
where E is Eddingtons epsilon; and are scalar functions; a, b, and c, are vector functions; and A is a second-rank
tensor function. To me, it is more interesting to prove these identities than to memorize them. In doing so, the
component notation is very useful. For example, the (i, j)-th component of div (a) can be written and expanded as
[div (a)]
ij
=
(a
j
)
x
j
=
a
j
x
j
+
x
j
a
j
, (1.7.24)
and so, we have just proved
div (a) = div a + grad a. (1.7.25)
Another example would be a curl b, whose i-th component is
(a curl b)
i
=
ijk
a
j
kmn
b
n
x
m
=
ijk
kmn
a
j
b
n
x
m
, (1.7.26)
which becomes by the formula (1.4.29),
(a curl b)
i
= (
im
jn
in
jm
)a
j
b
n
x
m
=
im
jn
a
j
b
n
x
m
in
jm
a
j
b
n
x
m
=
im
a
j
b
j
x
m
in
a
j
b
n
x
j
= a
j
b
j
x
i
a
j
b
i
x
j
= ((grad b)
t
a)
i
((grad b)a)
i
. (1.7.27)
Thus, Equation (1.7.6) has been proved. Try it for yourself. Its fun.
1.8 Coordinate Transformation of Vectors
I like to switch back and forth among the thee coordinate systems, and actually often need to transform a vector
such as a uid velocity from one coordinate system to another. Let v be the velocity vector whose components are
denoted by
v = [u, v, w]
t
in Cartesian coordinates, (1.8.1)
v = [u
r
, u
, u
z
]
t
in cylindrical coordinates, (1.8.2)
v = [u
r
, u
, u
]
t
in spherical coordinates. (1.8.3)
Then, the velocity vector can be transformed from cylindrical coordinates to the Cartesian coordinates by
_
_
u
v
w
_
_ =
_
_
cos sin 0
sin cos 0
0 0 1
_
_
_
_
u
r
u
u
z
_
_, (1.8.4)
and from the Cartesian coordinates to cylindrical coordinates by
_
_
u
r
u
u
z
_
_ =
_
_
cos sin 0
sin cos 0
0 0 1
_
_
_
_
u
v
w
_
_. (1.8.5)
22 CHAPTER 1. BASICS
Similarly, the velocity vector can be transformed from spherical coordinates to the Cartesian coordinates by
_
_
u
v
w
_
_ =
_
_
cos sin sin cos cos
sin sin cos sin cos
cos 0 sin
_
_
_
_
u
r
u
_, (1.8.6)
and from the Cartesian coordinates to spherical coordinates by
_
_
u
r
u
_ =
_
_
cos sin sin sin cos
sin cos 0
cos cos sin cos sin
_
_
_
_
u
v
w
_
_. (1.8.7)
1.9 Coordinate Transformation of Partial Derivatives
Sometime I feel like transforming partial derivatives form one coordinate system to another. For cylindrical coor-
dinates, we have
x = r cos , y = r sin , z = z, (1.9.1)
and
r =
x
2
+y
2
, (1.9.2)
cos =
x
x
2
+y
2
, (1.9.3)
sin =
y
x
2
+y
2
. (1.9.4)
Then, by the chain rule, we have
x
=
r
x
r
+
x
+
z
x
z
, (1.9.5)
y
=
r
y
r
+
y
+
z
y
z
, (1.9.6)
z
=
r
z
r
+
z
+
z
z
z
. (1.9.7)
Of course, we have
z
x
= 0,
z
y
= 0,
r
z
= 0,
z
= 0,
z
z
= 1, (1.9.8)
and we nd from Equations (1.9.2)
r
x
= cos ,
r
y
= sin , (1.9.9)
and from Equations (1.9.3) and (1.9.4)
x
=
sin
r
,
y
=
cos
r
. (1.9.10)
Therefore, we have the following transformation:
_
z
_
_
=
_
_
cos
sin
r
0
sin
cos
r
0
0 0 1
_
_
_
z
_
_
. (1.9.11)
1.9. COORDINATE TRANSFORMATION OF PARTIAL DERIVATIVES 23
The inverse transformation is given by
_
z
_
_
=
_
_
cos sin 0
r sin r cos 0
0 0 1
_
_
_
z
_
_
. (1.9.12)
For spherical coordinates, we have
x = (r sin ) cos , y = (r sin ) sin , z = r cos , (1.9.13)
and
r =
x
2
+y
2
+z
2
, (1.9.14)
cos =
x
x
2
+y
2
, (1.9.15)
sin =
y
x
2
+y
2
, (1.9.16)
cos =
z
x
2
+y
2
+z
2
, (1.9.17)
cos =
x
2
+y
2
x
2
+y
2
+z
2
. (1.9.18)
Then, by the chain rule, we have
x
=
r
x
r
+
x
+
x
, (1.9.19)
y
=
r
y
r
+
y
+
y
, (1.9.20)
z
=
r
z
r
+
z
+
z
. (1.9.21)
We nd from Equations (1.9.14)
r
x
= cos sin ,
r
y
= sin sin ,
r
z
= cos , (1.9.22)
from Equations (1.9.15) and (1.9.16)
x
=
sin
r sin
,
y
=
cos
r sin
,
z
= 0, (1.9.23)
and Equations (1.9.17) and (1.9.18)
x
=
cos cos
r
,
y
=
sin cos
r
,
z
=
sin
r
. (1.9.24)
Therefore, we obtain the following transformation:
_
z
_
_
=
_
_
sin cos
sin
r sin
cos cos
r
sin sin
cos
r sin
sin cos
r
cos 0
sin
r
_
_
_
_
. (1.9.25)
24 CHAPTER 1. BASICS
The inverse transformation is given by
_
_
=
_
_
sin cos sin sin cos
r sin sin r sin cos 0
r cos cos r sin cos r sin
_
_
_
z
_
_
. (1.9.26)
If I wish, I can use these transformations to derive the vector operators such as the gradient in cylindrical or spherical
coordinates from those in Cartesian coordinates. Anyway, it is always good to understand the relationship between
dierent coordinate systems. So, of course, I like these transformations.
1.10 Eigenvalues and Eigenvectors
A scalar quantity is called an eigenvalue of A if it satises
(AI)r = 0, (1.10.1)
for a square matrix A and a non-zero vector r. The vector r is called the right eigenvector of A associated with this
particular eigenvalue . A row vector
t
that satises,
t
(AI) = 0, (1.10.2)
is called the left eigenvector of A, again associated with this particular eigenvalue . I like these right and left
eigenvectors because they are not uniquely determined and any scalar multiple of an eigenvector is also an eigenvector,
i.e., still satises the above equation. So, I can scale the eigenvectors in any way I want. This is nice.
Note that there exist n eigenvalues (together with associated eigenvectors) for an n n matrix A:
1
,
2
,
3
, ,
n
. (1.10.3)
If these values are distinct, their associated eigenvectors are linearly independent. (But even if there are some repeated
eigenvalues, their eigenvectors may still be linearly independent.) If the eigenvectors are linearly independent, then we
can diagonalize the matrix A:
= R
1
AR, (1.10.4)
where R is the right eigenvector matrix,
R =
_
_
r
1
r
2
r
3
r
n
_
_
. (1.10.5)
is a diagonal matrix with the eigenvalues of A placed along its diagonal,
=
_
1
0 0 0 0
0
2
0 0 0
0 0
3
0 0
0 0 0
4
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 0
0 0 0 0
n
_
_
, (1.10.6)
1.11. SIMILAR MATRICES 25
and L = R
1
is in fact the left eigenvector matrix (whose rows are the left eigenvectors),
L =
_
t
1
t
2
t
3
.
.
.
t
n
_
_
. (1.10.7)
Although R is not unique (since the eigenvectors cannot be uniquely dened), L is uniquely dened this way for a
given R. By the way, if the matrix A is symmetric, its eigenvalues are all real and the eigenvectors corresponding to
dierent eigenvalues are orthogonal to one another. This is nice. I like symmetric matrices.
1.11 Similar Matrices
Consider square matrices, A, B, and M of the same size. If A and B are related through M in the form:
B = M
1
AM, (1.11.1)
then A and B are said to be similar and Equation (1.11.1) is called a similarity transformation. If A and B are similar,
they have the same eigenvalues. Moreover, their eigenvectors, say r
A
and r
B
for a particular eigenvalue, are related by
r
B
= M
1
r
A
. (1.11.2)
So, once I know the eigenvectors of A, I can nd eigenvectors of any matrices similar to A simply by using this relation.
In particular, this is very useful when we analyze conservation laws in various forms because change of variables always
results in a similarity transformation of the Jacobian matrices and so it suces to analyze just one form of the equation.
This is very nice. I like the similarity transformation.
1.12 Divergence Theorem
The divergence theorem is given by
V
div a dV =
S
a ndS, (1.12.1)
where n is the normal vector to the surface of the volume V . I like it, but sometimes I get confused because it is called
the Gauss theorem, or the Green-Gauss theorem, or Greens theorem in two dimensions (not quite right, though; see
Section 1.13),
V
_
a
x
x
+
a
y
y
_
dxdy =
S
(a
x
dy a
y
dx), (1.12.2)
where (dx, dy) denotes the innitesimal element along the boundary in the counter-clockwise direction. Also, some
people call the following equation the divergence theorem,
V
grad dV =
S
ndS, (1.12.3)
which should be called the gradient theorem, I think. That is probably because each component can be considered as
a part of the divergence theorem above. For example, in two dimensions, we have
x
dxdy =
S
dy, (1.12.4)
26 CHAPTER 1. BASICS
which is basically the rst term in Equation (1.12.2). Incidentally, the divergence theorem may be thought of as an
integration by parts. Look at this. We obtain from Equation (1.7.11),
V
div a dV =
V
div (a) dV
V
grad a dV
=
S
a ndS
V
grad a dV. (1.12.5)
Of course, the divergence theorem applies to second-rank tensors (denoted by A):
V
div AdV =
S
AndS. (1.12.6)
It is also interesting to derive another form [120] by replacing a by c b, where c is a constant vector in Equation
(1.12.1):
V
div (c b) dV =
S
(c b) ndS, (1.12.7)
which becomes by Equations (1.7.13) and (1.2.14)
V
(b
curl c c curl b) dV =
S
(b n) c dS, (1.12.8)
c
_
V
curl bdV =
S
(b n) dS
_
, (1.12.9)
and thus
V
curl bdV =
S
(b n) dS. (1.12.10)
Anyway, these formulas are often used in CFD, for example, to evaluate derivatives, to reduce the order of dierential
equations (e.g., the Galerkin formulation of the Laplace equation), or to simply compute the volume of a computational
cell. So, the divergence theorem is very useful in CFD, and I like it.
1.13 Stokes Theorem
If there is the divergence theorem, there is, of course, the curl theorem. But it is more widely known as Stokes
theorem. It is given, for an arbitrary vector a, by
S
curl a ndS =
S
a dl, (1.13.1)
where S is a bounded surface, dS and n are the innitesimal area and its unit outward vector, respectively, and
dl = (dx, dy, dz) denotes the innitesimal line-element in the counter-clockwise direction over the boundary of S
denoted by S. In Cartesian coordinates, we write a = (a
x
, a
y
, a
z
) and obtain the curl theorem as
S
__
a
z
y
a
y
z
_
dydz +
_
a
x
z
a
z
x
_
dzdx +
_
a
y
x
a
x
y
_
dxdy
_
=
S
(a
x
dx +a
y
dy +a
z
dz). (1.13.2)
I like it because Greens theorem is immediately obtained by setting a
z
= 0 and others to be constant in z:
S
_
a
y
x
a
x
y
_
dxdy =
S
(a
x
dx +a
y
dy). (1.13.3)
It is also possible to derive the following from the Stokes theorem [120]: replace a by c b, where c is a constant
vector in Equation (1.13.1):
S
curl (c b) ndS =
S
(c b) dl, (1.13.4)
1.14. CONSERVATION LAWS 27
which becomes by Equations (1.7.20) and (1.2.14)
S
(c div b + (
grad c)b (
S
c (b dl), (1.13.5)
c
_
S
(ndiv b (grad b)
t
n) dS =
S
(b dl)
_
, (1.13.6)
and thus
S
(ndiv b (grad b)
t
n) dS =
S
(b dl). (1.13.7)
In uid dynamics, Stokes theorem arises naturally in Kelvins circulation theorem as in Section 3.24 and also in
relation to the three-dimensional stream functions as in Section 4.24, for example. In fact, the curl theorem and the
divergence theorem are both special cases of Stokes theorem (also called the Kelvin-Stokes theorem). So, I actually
like to call them both Stokes theorem. See dierential geometry textbooks, e.g., Ref.[148], for more details.
1.14 Conservation Laws
Many ow equations come in the form of conservation laws. That is, the time rate of change of a quantity within
a control volume is balanced by the net change per unit time brought by the local ow. A typical example of such
a quantity is the uid mass. To express the conservation law mathematically, consider a control volume V with the
boundary S. Let U be a vector of conservative variables (quantities which are conserved), and F
n
be the ux vector
normal to the control volume boundary (e.g., mass ow per unit time). Then, we can write the conservation laws as
d
dt
V
UdV +
S
F
n
dS = 0. (1.14.1)
This is called the integral form of conservation laws. It simply states that the integral value of U changes in time only
due to the net eect of the ux across the control volume boundary. In Cartesian coordinates, the normal ux vector
can be written as
F
n
= Fn
x
+Gn
y
+Hn
z
, (1.14.2)
where n = (n
x
, n
y
, n
z
), and F, G, H are the ux vectors in the directions of x, y, z, respectively. Also, we may write
d
dt
V
UdV +
S
F ndS = 0, (1.14.3)
where F is a second-rank ux tensor dened by F = [F, G, H]. I like this form because I then immediately notice that
the surface integral can be written as a volume integral by the divergence theorem:
d
dt
V
UdV +
V
div F dV = 0, (1.14.4)
where, of course, every component of F must be dierentiable within the control volume. A nice thing about this is
that as long as the control volume is xed in space, we can write this as a single integral,
V
_
U
t
+ div F
_
dV = 0, (1.14.5)
and since this is valid for arbitrary control volumes, we can write
U
t
+ div F = 0. (1.14.6)
This is called the dierential form of conservation laws. In Cartesian coordinates, it is written as
U
t
+
F
x
+
G
y
+
H
z
= 0. (1.14.7)
28 CHAPTER 1. BASICS
I like it, but I prefer the integral form because the integral form allows discontinuous solutions (or weak solutions, more
generally) while the dierential form does not. The integral form is more fundamental, more general, and thus more
useful especially in developing shock-capturing methods. It is also useful even when the control volume deforms in
time, i.e., for moving mesh methods [52]. Moreover, the integral form is applicable to control volumes of any shape; it
is the basis of the nite-volume method. So, it is indeed very nice.
Having said that, I actually like the dierential form (1.14.6) very much because it can be written in a conservative
form by expressing the divergence in terms of the surface integral as suggested by the denition (1.5.21):
U
t
+ lim
V 0
1
V
F ndS = 0. (1.14.8)
Of course, it cannot be exact unless V 0, and so it will be an approximation in a nite (computational) volume:
U
t
+
1
V
F ndS 0. (1.14.9)
But it is actually widely used in node-centered nite-volume methods, which can be thought of as solving the dierential
form at nodes in a conservative manner. It can be a good approximation especially for second-order accurate methods.
It is also very useful to construct a conservative scheme. In fact, all advantages mentioned above about the integral form
are valid also for the conservative integral formulation of the dierential form. If you dont like such an approximation,
the exact conversion of the divergence to the surface integral is possible by the primitive function approach. Dene G
by
F =
1
V
V
G(, , ) d d d, (1.14.10)
where V is the control volume dened in the neighborhood of a point (x, y, z) with a xed volume V , and (, , )
is a set of dummy variables to perform the integral over the control volume. Then, F is the volume average of G, or
equivalently, G is the primitive function of F:
divF =
1
V
div
V
G(, , ) d d d, =
1
V
G ndS, (1.14.11)
which is exact for a xed V . Therefore, the dierential form (1.14.6) can be written exactly as
U
t
+
1
V
G ndS = 0. (1.14.12)
Note that the ux needed at the control volume boundary is the primitive function of the physical ux in this case.
This form is the basis of high-order nite-dierence methods typically developed for regular grids [181].
I like to compare the three forms above on a one-dimensional control volume. Consider a control volume of spacing
x, which we call j, dened by the left boundary x
j
x/2 and the right boundary x
j
+ x/2, where x
j
is the
center of the control volume.
Integral Form
Applying the integral form (1.14.1) over the control volume, we obtain
dU
j
dt
+
1
x
_
F
j+1/2
F
j1/2
= 0, (1.14.13)
where F
j1/2
denotes the ux at x
j
x/2, and U
j
is the averaged solution over the control volume,
U
j
=
1
x
xj+x/2
xjx/2
Udx. (1.14.14)
I like it very much because the uxes on the left and right boundaries can be computed from the solution at the
boundaries, i.e.,
F
j+1/2
= F(U
j+1/2
), F
j1/2
= F(U
j1/2
), (1.14.15)
1.15. CONSERVATION LAWS IN GENERALIZED COORDINATES 29
which is nice and simple. Note that the formulation (1.14.13) is exact and therefore higher-order numerical schemes
can be developed by improving the accuracy of the interface uxes (and the quadrature of the ux integral in two and
three dimensions). This is a very popular formulation in CFD.
Approximate Conservative Dierential Form
Next, applying the conservative approximation to the dierential form (1.14.9) at the center x = x
j
, we get
dU
j
dt
+
1
x
_
F
j+1/2
F
j1/2
= 0, (1.14.16)
where U
j
is the point value of the solution at x = x
j
. This is another popular formulation in practical CFD codes.
Obviously, it is a second-order accurate approximation to the dierential form of the conservation law because it is
equivalent to the central dierence formula to the derivative F/x. It means that higher-order approximation to the
uxes does not necessarily lead to higher-order accuracy in the whole equation. To achieve higher-order accuracy, a
more sophisticated discretization is required for the time derivative, e.g, the Galerkin discretization, rather than the
point evaluation as above. If it has source terms, they also need to be discretized carefully for higher-order accuracy
[102].
Exact Conservative Dierential Form
On the other hand, applying the exact form (1.14.12) at the center x = x
j
, we obtain
dU
j
dt
+
1
x
_
F
G
j+1/2
F
G
j1/2
_
= 0, (1.14.17)
where F
G
is the x-component of G, i.e., the primitive function of F,
F(x) =
1
x
x+x/2
xx/2
F
G
() d,
F
x
=
1
x
_
F
G
j+1/2
F
G
j1/2
_
. (1.14.18)
This form is exact. Therefore, high-order accuracy in F
G
j+1/2
and F
G
j1/2
does yield high-order accuracy in the whole
equation, which is very nice. Note that it is the primitive function of the ux, not the ux itself, that is required at
the control volume boundaries. Note also that the point value of the ux is now the volume average of the primitive
function:
F(x
j
) =
1
x
xj+x/2
xjx/2
F
G
(x) dx. (1.14.19)
As you might have noticed by now, thee forms are equivalent to one another for second-order accurate methods
because the volume-averaged solution diers from the point value by O(x
2
):
U
j
= U
j
+O(x
2
). (1.14.20)
which can be easily proved by substituting the Taylor expansion of U(x) around x = x
j
into the right hand side of
Equation (1.14.14). I like second-order methods because it doesnt really matter what quantity to compute (volume
averages or point values). That is, I can say that my second-order CFD code computes volume averages when I set
the initial solution by volume averages, or I can also say that it computes point values when I set the initial solution
by the point values. I like it very much.
1.15 Conservation Laws in Generalized Coordinates
Consider the coordinate transformation:
= (x, y, z), = (x, y, z) = (x, y, z), (1.15.1)
30 CHAPTER 1. BASICS
which denes a map between the physical coordinates (x, y, z) and the computational (or generalized) coordinates
(, , ). It maps a curvilinear grid in the physical space to a Cartesian grid in the computational space. Then, partial
derivatives are transformed by the chain rule,
x
=
x
+
x
+
x
, (1.15.2)
y
=
y
+
y
+
y
, (1.15.3)
z
=
z
+
z
+
z
. (1.15.4)
Metric coecients such as
x
,
x
,
x
, are given by
x
= J(y
),
y
= J(x
),
z
= J(x
), (1.15.5)
x
= J(y
),
y
= J(x
),
z
= J(x
), (1.15.6)
x
= J(y
),
y
= J(x
),
z
= J(x
), (1.15.7)
J =
1
x
(y
) x
(y
) x
(y
)
, (1.15.8)
which can be readily computed once the transformation is given in the following form,
x = x(, , ), y = y(, , ), z = z(, , ). (1.15.9)
For more details, see Refs.[36, 57, 155]. For example, the dierential form of the conservation law (1.14.7) can be
written as
U
t
+
F
+
G
+
H
= 0, (1.15.10)
where
U
=
U
J
, (1.15.11)
F
=
1
J
(F
x
+G
y
+H
z
) , (1.15.12)
G
=
1
J
(F
x
+G
y
+H
z
) , (1.15.13)
H
=
1
J
(F
x
+G
y
+H
z
) . (1.15.14)
If F, G, H depend on the derivatives of U, these derivatives are transformed by Equations (1.15.2), (1.15.3), and
(1.15.4). I like this form because now I can work with a regular Cartesian grid in the space (, , ), and apply the
straightforward nite-dierence formulas.
1.16 Classication of Systems of Dierential Equations
1.16.1 General System of Equations
Consider the system of n quasi-linear partial dierential equations in the form,
U
t
+AU
x
+BU
y
+CU
z
= 0, (1.16.1)
1.16. CLASSIFICATION OF SYSTEMS OF DIFFERENTIAL EQUATIONS 31
P
P
A B x
t
Figure 1.16.1: Domain of dependence (lower triangle) and range of inuence (upper triangle)
of a point P, for a linear 22 hyperbolic system. They are bounded by the characteristic lines
(wave paths) of right- and left-running waves. Domain of dependence of a point P
includes
the right boundary.
where U is a set of n solution variables, and A, B, and C are nn square matrices. I like such a system because it
may have a plane wave solution,
U = f(x n
n
t) U
0
, (1.16.2)
where f is a scalar function, n = (n
x
, n
y
, n
z
) is a unit vector normal to the plane wave,
n
is the plane wave speed
in the direction of n, and U
0
is a constant vector. In one dimension, this becomes
U = f(x t) U
0
, (1.16.3)
and the wave travels to the left if < 0 or to the right if > 0. In the space with (x, t) coordinates, the wave
is represented by a curve dened by a local slope dx/dt = , along which f is constant. Such a curve is called a
characteristic curve, and f is called the Riemann invariant. In two dimensions, we may use the polar angle to indicate
the normal direction n = (cos , sin ), and write the solution (1.16.2) as
U = f(xcos +y sin ()t) U
0
, (1.16.4)
where () denotes the wave speed in the direction of . Once () is known, we can determine the wave front by
constructing an envelope of plane waves. This is discussed in Section 1.21. In two dimensions, the wave may propagate
not only in a particular direction as in one dimension but also in all directions at varying speeds. In the space with
(x, y, t) coordinates, such a wave is generally represented by a surface. Such a surface is called characteristic surface,
and again f is the Riemann invariant.
In order to see if the system (1.16.1) has a plane wave solution, we substitute Equation (1.16.2) into the system
to get
(An
x
+Bn
y
+Cn
z
n
I) U
0
= 0. (1.16.5)
This means that the speed of the plane wave
n
corresponds to an eigenvalue of the projected Jacobian,
A
n
= An
x
+Bn
y
+Cn
z
, (1.16.6)
and U
0
is the corresponding eigenvector, and that there may be n such plane wave solutions. Therefore, if all eigenvalues
of A
n
are real and the corresponding eigenvectors are linearly independent, there will be n independent characteristic
surfaces and n Riemann invariants, and thus a solution at a point in space can be completely determined, in principle
by solving a set of n equations of constant Riemann invariants. In this case, the system (1.16.1) is called hyperbolic
in time. Moreover, if the eigenvalues are distinct, it is called strictly hyperbolic. In that case, the eigenvectors are
automatically linearly independent. If all eigenvalues are complex, then the system is called elliptic. Of course, the
system will be mixed if it has both real and complex eigenvalues. Also, if there are less than n real eigenvalues (i.e.,
A
n
is not full rank), then the system is said to be parabolic.
An important implication of the existence of the plane wave solution is that a solution at a point in space depends
only on the solution values in the region where the plane wave has traveled before reaching that point. Such a region
32 CHAPTER 1. BASICS
P
x
t
Figure 1.16.2: Domain of dependence of a point P for a parabolic problem.
is called the domain of dependence. In one dimension, a typical 22 hyperbolic system has two plane wave solutions:
one running to the left and the other to the right. Then, the domain of dependence of a certain point is the union of
the left and right intervals where the right- and left-running waves pass through until they reach that point. This grows
in time, and will look like a triangle bounded by characteristic lines (or curves for nonlinear waves) in (x, t)-space (see
Figure 1.16.1). Note that for a scalar hyperbolic equation the domain of dependence becomes a single point, such as
the point A in Figure 1.16.1 for a right-running wave. Also, it is interesting to note that it will take a nite time for
a solution at a point P to inuence other points in the domain. Of course, it cannot propagate faster than the plane
waves (characteristics), and therefore the range of inuence (the region inuenced by the solution at P) is bounded
by the characteristics passing through the point P (see Figure 1.16.1). Since the domain of dependence is bounded,
a hyperbolic problem can be solved for a given initial solution if the domain is innite or periodic (i.e., no boundary).
This type of problem is often called an initial-value problem. If the domain is bounded as in Figure 1.16.1, then a
part of the solution needs to be specied on the boundary because there might be a characteristic which cannot be
traced back to the initial solution (see P
k
= r
k
t
k
, (1.16.16)
where
t
k
is the k-th row of L. In this form, we see that the matrix
k
projects the solution change into the space of
the k-th eigenvector. For this reason, it is often called the projection matrix [99, 106]. It arises also in the context of
projected spatial operators in two dimensional equations [117]. Of course, we have
n
k=1
k
= I. (1.16.17)
By using the projection matrix, we can write the decomposition of the Jacobian matrix, A = RL, as
A =
n
k=1
k
. (1.16.18)
Then, the positive and negative parts of the Jacobian matrix, which are the parts corresponding to positive and negative
eigenvalues, respectively, can be easily identied:
A
k
0
k
, A
+
=
k
>0
k
, (1.16.19)
where A
and A
+
denote the positive and negative parts, respectively. Such a decomposition is often used to construct
upwind schemes: one-sided scheme towards the upwind direction. See Ref.[99, 106], for examples that explicitly utilize
the projection matrix.
It is also interesting to write the conservation law (1.16.7) as follows,
U
t
+RLARLU
x
= 0, (1.16.20)
U
t
+RV
x
= 0, (1.16.21)
U
t
+
n
k=1
k
v
k
x
r
k
= 0. (1.16.22)
This shows how each subproblem (advection term) aects the time rate of conservative variables, and again, each
contribution is proportional to the right eigenvector. It is also very clear from this that each subproblem, i.e., Equation
(1.16.11), is obtained simply by multiplying the system by the left eigenvector
t
k
from the left.
34 CHAPTER 1. BASICS
1.16.3 Two-Dimensional Steady Equations
The classication discussed in Section 1.16.1 can be extended to steady equations. Consider the steady two-dimensional
system,
AU
x
+BU
y
= 0. (1.16.23)
This admits a steady plane wave solution of the form,
U = f(xn
x
+yn
y
) U
0
, (1.16.24)
provided
(An
x
+Bn
y
) U
0
= 0. (1.16.25)
That is, if the generalized eigenvalue problem,
det (BA) = 0, (1.16.26)
has a real solution = n
x
/n
y
. For a system of n equations, there will be n eigenvalues. If all n eigenvalues are real
and the corresponding eigenvectors are linearly independent, the steady system is called hyperbolic. If all are complex,
it is called elliptic. It can, of course, be mixed if there are both real and complex eigenvalues. Again, if there are less
than n real eigenvalues, the system will be parabolic.
As an example, convert the linear scalar second-order partial dierential equation,
a
xx
+b
xy
+c
yy
+d
x
+e
y
+f = 0, (1.16.27)
where a, b, c, d, e, f are functions of (x, y, z), into a system of the form (1.16.23), by introducing new variables,
u =
x
, v =
y
, (1.16.28)
leading to
AU
x
+BU
y
= S, (1.16.29)
where
U =
_
u
v
_
, A =
_
a b/2
0 1
_
, B =
_
b/2 c
1 0
_
, S =
_
du ev f
0
_
, (1.16.30)
and the second equation (v
x
u
y
= 0) is a necessary condition that comes from the denition of the new variables
(1.16.28). Note also that we split b
xy
into two:
b
2
xy
+
b
2
xy
=
b
2
u
y
+
b
2
v
x
. Now, we have
det (BA) = 0, (1.16.31)
i.e.,
det
_
b/2 a c b/2
1
_
, (1.16.32)
or
a
2
b +c = 0, (1.16.33)
and therefore we see that the system is hyperbolic if b
2
4ac > 0 (and the eigenvectors are linearly independent),
and elliptic if b
2
4ac < 0. Of course, it is called parabolic if b
2
4ac = 0. In this case, there exists only a single
(degenerate) characteristic, and it is not sucient to describe the solution behavior. This type of equation requires
low-order terms in S to describe the solution. For example, the parabolic equation u
y
= u
xx
is basically the heat
equation with y interpreted as time. Obviously, its character cannot be described without the rst-order term, u
y
.
It is interesting to note that if A
1
exists, we can write the two-dimensional system (1.16.23) in the form,
U
x
+A
1
BU
y
= 0. (1.16.34)
This looks like a one-dimensional system with x taken as time. It will really behave like a time-dependent system if
x-axis is contained entirely in the range of inuence of any point (no backward characteristics). If this is true, then, x
is called time-like and y is called space-like. The two-dimensional steady Euler system is a good example. See Section
3.17 for details.
By the way, I like adding a time-derivative term to the steady system because then the whole system can be
hyperbolic in time: a real eigenvalue indicates a wave traveling in a particular direction while a complex eigenvalue
indicates a wave traveling isotropically (all directions). Finally, it is interesting to know that the eigenvalues and
eigenvectors of the steady system can be used to decompose the system into hyperbolic and elliptic parts. See
Ref.[117] for details.
1.17. CLASSIFICATION OF SCALAR DIFFERENTIAL EQUATIONS 35
1.17 Classication of Scalar Dierential Equations
Consider, again, the second-order partial dierential equation,
a
xx
+b
xy
+c
yy
+d
x
+e
y
+f = 0, (1.17.1)
where a, b, c, d, e, f are functions of (x, y, z). It is really nice that we can easily identify the type of the equation,
before attempting to solve it, by
b
2
4ac < 0, Elliptic,
b
2
4ac = 0, Parabolic,
b
2
4ac > 0, Hyperbolic,
(1.17.2)
as shown in the previous subsection. This is very important because the strategy for numerically solving the equation
can be quite dierent for dierent types. A typical elliptic equation is the Laplace equation,
u
xx
+u
yy
= 0. (1.17.3)
This equation requires a boundary condition everywhere on the domain boundary: specify the solution (Dirichlet
condition), specify the solution derivative normal to the boundary (Neumann condition), or specify the combination of
the solution and the normal derivative (Robin condition) [155]. The solution inside the domain is smooth. In particular,
its maximum/minimum occurs only on the boundary; this is often called the maximum principle.
A typical parabolic equation is the heat equation,
u
y
u
xx
= 0, (1.17.4)
where y-axis acts like time. See Figure 1.16.2 with t replaced by y. An initial distribution of u (temperature) along
x-axis (y = 0) will continue to diuse along x-direction as y increases. This equation requires boundary conditions
at y = 0 (initial temperature distribution specied) and along the left and right boundaries (temperature at two end
points on x-axis). The solution is again smooth. Even if an initial solution is not smooth, it will be smeared out as
time goes on. This is very nice.
A typical hyperbolic equation is the wave equation,
u
xx
u
yy
= 0, (1.17.5)
where again y may be considered as time. This describes a solution traveling in x with time y (recall that characteristic
surfaces exist for hyperbolic systems). It requires an initial solution to be specied at y = 0, but the solution at each
(x, y) depends only on a part of the initial data. Basically, it only needs to know where it comes from in the initial
data. Also, by taking a domain to be periodic, i.e., u(x
min
) = u(x
max
), we can avoid boundary conditions, thus
leading to a pure initial-value problem. The solution does not have to be smooth and can be even discontinuous. But
such a solution is not dierentiable and will not satisfy the wave equation above. In fact, there is an integral equation
associated with the wave equation which allows discontinuous solutions. See Section 2.1 for details.
I like the fact that only the rst three terms, i.e., the second derivative terms, determine the character of the
equation; we do not need to take into account all the terms to analyze the equations. In fact, generally, many
important features of solutions of partial dierential equations depend only on the highest-order terms (called the
principal part) of the equation [180]. This is really nice. Suppose that you have a partial dierential equation in the
form (1.16.27) with an extremely complicated source term f(x, y), e.g., those involving terms like
x
y
or
2
x
. Then,
just look at the three second-order terms only, a
xx
+ b
xy
+ c
yy
, and you will nd the type of your equation by
Equation (1.17.2), which then guides you to an appropriate numerical method for solving it.
1.18 Pseudo-Transient Equations
I like CFD because partial dierential equations are often manipulated for computational advantages or just for conve-
nience. They even change the type. For example, consider the following elliptic equation:
u
xx
+u
yy
= 0. (1.18.1)
36 CHAPTER 1. BASICS
It can be solved as a time-dependent equation,
u
= u
xx
+u
yy
, (1.18.2)
where is a time-like variable called pseudo time. The solution to the original problem is then obtained as a steady
state solution to the pseudo-time-dependent problem. The method is generally called the pseudo-transient method (or
continuation). It is very interesting and useful because methods developed for time-dependent initial-value problems
can be directly applied to steady problems (those without time derivatives). Note that the type of partial dierential
equations have been changed from elliptic to parabolic. This is very interesting. The elliptic equation can be solved as
a parabolic equation; the solution to the elliptic equation is obtained in the steady state. I like it because then I think I
can also dene a pseudo-transient equation by introducing the second derivative u
:
u
= u
xx
+u
yy
, (1.18.3)
which is the wave equation and thus hyperbolic. In this case, the solution to the elliptic equation is obtained by solving
a hyperbolic equation towards the steady state. Im actually not aware of any work in CFD that actually does so.
However, I do know a method that does it in the form of rst-order system:
u
= (p
x
+q
y
), p
=
1
T
r
(u
x
p), q
=
1
T
r
(u
y
q), (1.18.4)
where T
r
is a free parameter. This system is hyperbolic, but reduces to the elliptic equation (1.18.1) in the steady
state for any positive value of T
r
as described in Section 2.5. This is very interesting because we can apply methods
developed for hyperbolic systems to solve the elliptic problem. I like it very much. See Section 2.5 and Refs.[98, 99,
103, 104] for details on the method of rst-order hyperbolic systems.
The pseudo-transient method is, in a way, a method of manipulated governing equations as mentioned at the
beginning. Consider a hyperbolic conservation law in two dimensions:
U
t
+F
x
+G
y
= 0, (1.18.5)
where U is a solution vector, and F and G are the ux vectors. If we are only interested in the steady state, then the
time t can be thought of as pseudo time :
U
+F
x
+G
y
= 0. (1.18.6)
The system is still hyperbolic, but we can alter the character of wave propagations by multiplying the spatial part by
a matrix P:
U
+P(F
x
+G
y
) = 0, (1.18.7)
where P may be taken as a positive denite matrix so that the waves do not change the direction of propagation and
the same type of boundary conditions apply. The matrix P can be designed, for example, to optimize the condition
number of the equation, the ratio of the fastest wave speed to the slowest wave speed, in order to accelerate the
convergence to the steady state (i.e., no waves left behind) [77]. Note that the system reduces to the original system in
the steady state as long as P is non-singular (which is true for a positive denite matrix). That is, the transient solution
may be completely wrong, but it converges faster to the steady solution of the original equation. This technique is
called the local-preconditioning technique [46, 80, 117, 161, 168]. It has an added benet of improving the accuracy
of numerical schemes in the low-Mach-number limit [45, 163]. Furthermore, some preconditioning matrices can be
used to split the governing equations into hyperbolic and elliptic parts [117], enabling optimal discretization [94, 113]
as well as optimal multigrid convergence [116]. That is very interesting. I like it. I also nd it interesting that the
pseudo-compressible formulation [21] for the incompressible Euler and Navier-Stokes equations can also be thought of
as a kind of pseudo-transient method.
Another area where the pseudo-transient method is found useful and actually widely applied is the solution method
for nonlinear systems, e.g., Newtons method. Newtons method is a very powerful method for solving nonlinear
systems, but it can take a long time to achieve a rapid convergence unless the initial solution is close enough to the
solution. To enhance the convergence, the pseudo-transient method can be employed in its early stage (typically by
gradually increasing the pseudo time step). Such a technique is often employed in practical CFD codes. It is a simple
but very eective method and I like it very much.
Finally, time-dependent computations can be performed with the pseudo-transient method by the dual-time ap-
proach [59] in which the pseudo-transient method can be used to solve, at each physical time step, the system of
equations arising from an implicit-time stepping scheme. I like the fact that a fast steady solver is an essential building
block of implicit-time stepping.
1.19. OPERATOR SPLITTING 37
1.19 Operator Splitting
In CFD, it is sometimes convenient to solve a partial dierential equation term by term. Consider the two-dimensional
evolution equation of a solution u(x, y, t) in the form:
u
t
= (A+B) u, (1.19.1)
where A and B are two dierent linear dierential operators. For example, A may be the advective operator and B
can be the diusive operator, or A is an x-derivative operator and B is a y-derivative operator. I like that A and B
are linear because then I can easily generate as many time derivatives as I like:
u
tt
= (A+B) u
t
= (A+B)
2
u, (1.19.2)
u
ttt
= (A+B) u
tt
= (A+B)
3
u, (1.19.3)
u
tttt
= (A+B) u
ttt
= (A+B)
4
u, (1.19.4)
and so on. I would use them to compute the solution at t + t from the solution u at t as follows:
u(x, y, z, t + t) = u + t u
t
+
t
2
2
u
tt
+
t
3
6
u
ttt
+
= u + t(A+B) u +
t
2
2
(A+B)
2
u +
t
3
6
(A+B)
3
u +
=
_
1 + t(A+B) +
t
2
2
(A+B)
2
+
t
3
6
(A+B)
3
+
_
u, (1.19.5)
which can be expressed as
u(x, y, t + t) = exp ((A+B)t) u. (1.19.6)
Note that u and all its time derivatives on the right hand side are understood as values at t. In CFD, it is sometimes
easier to solve each operator independently in a sequential manner: e.g., solve
u
t
= Au, (1.19.7)
and then solve
u
t
= B u
, (1.19.8)
to compute the solution at the next time step. This is called the operator splitting method or the fractional step
method. The above splitting in particular is often called the Godunov splitting [5]. It is a simple and interesting idea.
There are many good reasons that I like it. In this way, a numerical method can be designed specically for each
term, and the coding can be made simple. Also, accuracy and stability of the numerical method may be controlled
independently at each step. For example, we may solve the rst step by an explicit high-order upwind scheme and the
second step by an implicit diusion scheme to allow large time steps. However, it is not immediately guaranteed that
the operator splitting formulation is equivalent to the original formulation. To see this, expand the solutions at the
two stages:
u
(x, y, t + t) =
_
1 + tA+
t
2
2
A
2
+
t
3
6
A
3
+
_
u, (1.19.9)
u(x, y, t + t) =
_
1 + tB +
t
2
2
B
2
+
t
3
6
B
3
+
_
u
t
= Au with t/2, (1.19.14)
u
t
= B u
with t, (1.19.15)
u
t
= Au
()
_
, (1.21.1)
where () is the plane wave speed in the direction of (0 < 2). See Figure 1.20.1. In particular, I like the
following examples.
40 CHAPTER 1. BASICS
O
()
F(x, y, ) = xcos + y sin () = 0
x
y
Figure 1.20.1: Envelope of plane waves dened by the equation of the plane wave (the tangent to the
envelope) F(x, y, ) = 0 and the equation of vanishing derivative F/ = 0.
(a) Linear Advection:
Consider the linear scalar hyperbolic equation,
u
t
+au
x
+bu
y
= 0, (1.21.2)
where a and b are constants. To see if this has a plane wave solution, we substitute a general plane wave solution,
u = f(xcos +y sin ()t), (1.21.3)
into Equation (1.21.3) and solve the resulting equation for (). The solution is
() = a cos +b sin , (1.21.4)
which is real, meaning that it has a plane wave solution. We then proceed to nd the wave front by
_
X()
Y ()
_
=
_
cos sin
sin cos
_ _
a cos +b sin
a sin +b cos
_
=
_
a
b
_
. (1.21.5)
Therefore, a point disturbance simply propagates in the direction (a, b).
(b) Isotropic Advection:
Consider the following equation,
u
t
+ (Rcos )u
x
+ (Rsin )u
y
= 0, (1.21.6)
In this case, we nd () = R and
() = 0, and so
_
X()
Y ()
_
=
_
cos sin
sin cos
_ _
R
0
_
=
_
Rcos
Rsin
_
. (1.21.7)
Therefore, we nd that the envelope is a circle dened by
X
2
+Y
2
= R
2
, (1.21.8)
i.e., a disturbance propagates in every direction at the same speed.
(c) System of Equations:
Consider a system of hyperbolic equations,
U
t
+AU
x
+BU
y
= 0. (1.21.9)
1.22. RANKINE-HUGONIOT RELATION 41
This has a plane wave solution of the form
U = f(xcos +y sin ()t)U
0
, (1.21.10)
if () is an eigenvalue of the matrix,
A
n
= Acos +Bsin . (1.21.11)
As an example, consider the two-dimensional Euler equations, for which we nd
() = q
n
a, q
n
, q
n
, (1.21.12)
where q
n
= (u, v) (cos , sin ) is the ow speed in the direction of and a is the speed of sound. So, two waves,
corresponding to those with the speed q
n
, simply propagate in the ow direction at the ow speed q =
u
2
+v
2
;
and the other two, corresponding to those with the speeds q
n
a, have the envelopes given by
_
X()
Y ()
_
=
_
cos sin
sin cos
_ _
ucos +v sin a
usin +v cos
_
=
_
u a cos
v a sin
_
, (1.21.13)
i.e., the envelope is a circle centered at (u, v),
(X u)
2
+ (Y v)
2
= a
2
. (1.21.14)
In general, I like wave fronts because they can be in quite interesting shapes; see Refs.[60, 117], for example, for
intriguing magnetohydrodynamic waves.
1.22 Rankine-Hugoniot Relation
Let n be the unit normal vector of a discontinuous surface (a shock wave), and V
D
be the normal speed of the
discontinuity. Then, apply the integral form of conservation laws (1.14.1) across the discontinuity, and let V 0 to
obtain the Rankine-Hugoniot relation,
[F
n
] = V
D
[U], (1.22.1)
where [ ] denotes the jump across the discontinuity. Derive this by yourself or see Refs.[50, 51, 60] for derivations.
Suppose that the jump is small, so that we can write
dF
n
= V
D
dU, (1.22.2)
and hence
F
n
U
dU = V
D
dU. (1.22.3)
This shows that the shock speed V
D
is an eigenvalue of
F
n
U
and the jump dU is the corresponding eigenvector. Note
that this is always true for linear equations, but not for nonlinear equations. For nonlinear equations, this is true only
for small jumps. I emphasize that the shock speed is NOT an eigenvalue of the Jacobian matrix for large jumps. The
one-dimensional Euler equations is a good example: the shock speed associated with one of the acoustic waves is given
by (derived from the Rankine-Hugoniot relation)
V
D
= u
L
+a
L
1 +
+ 1
2
_
p
R
p
L
p
L
_
, (1.22.4)
where = 1.4; u
L
, p
L
, and a
L
are the velocity, pressure, and the speed of sound on one side of the shock; and p
R
is
the pressure on the other side of the shock. For small change, i.e., p
R
p
L
0, it reduces to
V
D
u
L
+a
L
, (1.22.5)
which is nothing but the eigenvalue of the Jacobian, i.e., the one associated with the same acoustic wave (see Section
3.4.2). This is actually nice. There are two similar eigenvalues in the Euler equations, u + a and u a, which are
42 CHAPTER 1. BASICS
sometimes called two nonlinear elds associated with the acoustic wave. We then see from Equation (1.22.5) that the
shock speed given by Equation (1.22.4) is associated with the eld of u+a, not the one of ua. Yes, we can identify
the eld associated with a shock wave by taking the limit p
R
p
L
in the shock speed. In fact, if we look at the other
shock speed (again, derived from the Rankine-Hugoniot relation):
V
D
= u
L
a
L
1 +
+ 1
2
_
p
R
p
L
p
L
_
, (1.22.6)
we nd by expanding it for small change p
R
p
L
0,
V
D
u
L
a
L
. (1.22.7)
So, V
D
is the shock speed associated with the other eld. This is nice, isnt it? I like it very much.
1.23 Linearly Degenerate and Genuinely Nonlinear Fields
Consider a hyperbolic system of the form,
U
t
+AU
x
= 0. (1.23.1)
The k-th eigenvalue of the coecient matrix A,
k
(U) (or the eigenvector r
k
(U)) denes a characteristic eld called
the
k
-th eld (or r
k
-th eld), or simply k-th characteristic eld.
A k-th characteristic eld is said to be linearly degenerate if
k
(U)
U
r
k
(U) = 0, for all U. (1.23.2)
In the case of linear hyperbolic systems, all eigenvalues are constant and therefore all elds are linearly degenerate. The
contact discontinuity is a discontinuity in a linearly degenerate eld: the characteristics are all parallel across such a
discontinuity just like the scalar linear advection. On the other hand, a k-th characteristic eld is said to be genuinely
nonlinear if
k
(U)
U
r
k
(U) = 0, for all U. (1.23.3)
I like this genuinely nonlinear one because the characteristic speed can now change across waves, thus creating much
more sophisticated waves than the linearly degenerate eld. Specically, there can be nonlinear shocks (converging
characteristics) and nonlinear expansions (diverging characteristics) in this case. Note that any nonlinear scalar con-
servation laws with non-zero characteristic speed is genuinely nonlinear and such scalar conservations laws are said to
have a convex ux. See Refs.[81, 159] for details.
The one-dimensional Euler equations in the primitive variables U = [, u, p]
t
can be a good example: the eigenvalues
and eigenvectors are given by (see Section 3.4.2)
1
= u a,
2
= u,
3
= u +a, (1.23.4)
r
1
=
_
a
1
a
_
_
, r
2
=
_
_
1
0
0
_
_, r
3
=
_
a
1
a
_
_
, (1.23.5)
1.23. LINEARLY DEGENERATE AND GENUINELY NONLINEAR FIELDS 43
where a =
1
U
r
1
=
_
_
(u a)/
(u a)/u
(u a)/p
_
a
1
a
_
_
=
_
_
a
2
1
a
2p
_
a
1
a
_
_
=
+ 1
2
= 0, (1.23.6)
2
U
r
2
=
_
_
u/
u/u
u/p
_
_
1
0
0
_
_
=
_
_
0
1
0
_
_
1
0
0
_
_
= 0, (1.23.7)
3
U
r
3
=
_
_
(u +a)/
(u +a)/u
(u +a)/p
_
a
1
a
_
_
=
_
a
2
1
a
2p
_
a
1
a
_
_
=
+ 1
2
= 0, (1.23.8)
so that the
1
-eld and
3
-eld (acoustic waves) are genuinely nonlinear but the
2
-eld (entropy wave) is linearly
degenerate.
44 CHAPTER 1. BASICS
Chapter 2
Model Equations
2.1 Linear Advection
I like the linear advection equation:
u
t
+ (grad u) = 0, (2.1.1)
or
u
t
+ div (u) = 0, (2.1.2)
or simply
Du
Dt
= 0, (2.1.3)
where is a constant advection velocity at which the solution u travels. In Cartesian coordinates, the advection
equation becomes, with = (a, b, c),
1D: u
t
+au
x
= 0, (2.1.4)
2D: u
t
+au
x
+bu
y
= 0, (2.1.5)
3D: u
t
+au
x
+bu
y
+cu
z
= 0. (2.1.6)
The linear advection equation is a model equation for a general quasi-linear system of equations, e.g., in one dimension,
U
t
+AU
x
= 0, (2.1.7)
where U is a vector of solutions and A is a coecient matrix having real eigenvalues. This includes the Euler equations
of gas dynamics. To develop a method for solving such a system, it is a good idea to rst study the model equation
which is simpler but retains important features of the general system. Once we develop and study a method for the
model equation, we may move on to extend it to the system of equations. For example, since the system (2.1.7) can
be diagonalized, i.e., decomposed into a set of linear advection equations (see Section 1.16.2), a method developed for
the linear advection equation can be directly applied to each component.
In studying the model equation, I like its integral form,
d
dt
V
udV +
S
u ndS = 0, (2.1.8)
because this allows discontinuous solutions: any initial solutions, whether continuous or discontinuous, will be simply
advected at the velocity . This is, in fact, a good model for developing a method for computing ows with shock
waves. Many techniques for accurately and stably computing discontinuous solutions have been successfully extended
to more general equations such as the Euler equations.
Note that the linear advection equation is closely related to the wave equation in one dimension, but not so much
in higher dimensions. In one dimension, the wave equation is
u
tt
a
2
u
xx
= 0, (2.1.9)
45
46 CHAPTER 2. MODEL EQUATIONS
which can be written as
(
t
a
x
) (
t
u +a
x
u) = 0. (2.1.10)
Therefore, the wave equation has two general solutions: one traveling at the speed a and the other at the speed a.
The linear advection equation governs only the one traveling at the speed a. On the other hand, in two dimensions,
the wave equation is dened by
u
tt
a
2
(u
xx
+u
yy
) = 0, (2.1.11)
where a > 0. This has a plane wave solution in all directions at the same speed a (substituting a function of the
form f(xcos + y sin t) into the wave equation, and solving the resulting equation for the wave speed , we
obtain = a which is independent of the direction ). In contrast, the linear advection equation has a plane wave
solution only in a particular direction, i.e., specied by the constant advection vector (a, b). To describe an isotropic
wave propagation, the advection vector needs to be a function of space:
(a, b) = a(cos , sin ). (2.1.12)
This describes a wave traveling isotropically at speed a (see Section 1.21). The advection equation is then still linear,
but it might be nice to stay with constant coecients. Then, we may use a system description of the isotropic wave:
u
t
= a (p
x
+q
y
) , (2.1.13)
p
t
= au
x
, (2.1.14)
q
t
= au
y
. (2.1.15)
It is easy to show by straightforward dierentiations that this system is equivalent to the wave equation (2.1.11). Note
however that this system representation has introduced an additional constraint, (p
y
q
x
)
t
= 0, to the wave equation.
This condition implies that any zero or nonzero vorticity is preserved at all later times. This type of model system
was used in Ref.[95] for developing vorticity-preserving schemes. In three dimensions, basically the same is true, except
that the number of constraints now becomes three; the vorticity has three components in three dimensions.
Incidentally, some people use the term convection instead of advection. Personally, I like advection better simply
because Im used to it. I like not to discuss which is better.
2.2 Circular Advection
The advection equation will still be linear even if = (a, b, c) is a function of space. For example, the two-dimensional
advection equation,
u
t
+au
x
+bu
y
= 0, (2.2.1)
with the advection velocity,
(a, b) = (y, x), (2.2.2)
represents a ow eld that rotates clockwise around the origin. This is called the circular advection equation. A
three-dimensional circular advection equation would be
u
t
+au
x
+bu
y
+cu
z
= 0, (2.2.3)
with
(a, b, c) = (y, x, 0). (2.2.4)
This represents a circular ow eld around z-axis. Alternatively, we can take = (0, z, y) to create a ow around
x-axis, or = (z, 0, x) to create a ow around y-axis. Yet, we may take = (y, x, V ) where V is a constant. I
like this one because it describes a more 3D ow: rotating around z-axis and simultaneously moving in the z direction
at a constant speed V . Note that the circular advection equation can be written also in the conservative form,
u
t
+ div (u) = 0, (2.2.5)
= (y, x) in two dimensions, and = (y, x, 0), (z, 0, x), (z, 0, x) or (y, x, V ) in three dimensions.
2.3. DIFFUSION EQUATION 47
2.3 Diusion Equation
I like the diusion equation:
u
t
= div ( grad u) , (2.3.1)
where is a positive diusion coecient. As long as does not depend on the solution itself, this equation is linear.
In Cartesian coordinates, the diusion equation becomes
1D: u
t
= (u
x
)
x
, (2.3.2)
2D: u
t
= (u
x
)
x
+ (u
y
)
y
, (2.3.3)
3D: u
t
= (u
x
)
x
+ (u
y
)
y
+ (u
z
)
z
. (2.3.4)
The diusion equation is parabolic in time (prove it!). It requires both an initial solution and boundary conditions.
Generally, its solution is very smooth. Even if an initial solution is not smooth, it will be made smooth as time goes on
because of the smoothing eect of the second-derivative term on the right hand side, i.e., diusion term. The solution
will reach a steady state, unless we introduce a time-dependent boundary condition and/or source term. Then, the
steady solution will be equivalent to that of the Laplace equation (see Section 2.8).
Of course, I like its integral form also,
d
dt
V
udV +
S
grad u ndS = 0, (2.3.5)
or
d
dt
V
udV +
u
n
dS = 0. (2.3.6)
This is a good starting point for developing a nite-volume scheme for the diusion equation. In general, it is
relatively easy to develop a scheme for the diusion equation in the sense that it only deals with smooth solutions (no
discontinuities), especially for second-order methods on regular computational grids. The issue, if any, would be the
eciency of the numerical method. In particular, an explicit scheme can be very inecient due to a severely restricted
time step; implicit schemes are generally preferred. All these (smoothing eect and eciency) are characteristics of the
viscous part of the Navier-Stokes equations. Therefore, the diusion equation is a typical model for the viscous terms
of the Navier-Stokes equations. Numerical methods for the viscous terms are often developed based on the diusion
equations or equivalently on the Laplace equation for spatial discretizations. By the way, the integral form above shows
that the integral value of the solution depends only on the normal solution gradient over the control volume boundary,
S. This is very interesting. It is basically one-dimensional, and the normal gradient can be easily evaluated on regular
grids in which the normal direction often is perfectly aligned with the direction of two solution data points, i.e., a
simple nite-dierence formula gives the normal gradient. I like it because I think it is one of the reasons that it is
relatively easy to develop diusion schemes on regular grids. For irregular grids, on the other hand, the construction
of diusion schemes is not necessarily easy. Accuracy and convergence can be easily deteriorated unless the normal
gradient is carefully evaluated. See, for example, Refs.[101, 108] for diusion schemes on irregular grids.
Note that the viscous terms of the compressible Navier-Stokes equations are actually more than the diusion
equation because they involve not only the normal gradient but also the gradients tangential to the control volume
boundary as described in Section 4.12. Some numerical schemes ignore the tangential gradients, e.g., a thin-layer-
approximation [89, 111]. Such schemes are known to be robust, but accuracy is not obtained because they are not
consistent schemes. Nevertheless, I like such schemes because they can be useful for constructing an implicit scheme
for the compressible Navier-Stokes equations [111]. See Section 4.12.
2.4 Diusion Equation: Mixed Formulation
Often, the diusion equation is written and discretized in the following rst-order system:
u
t
= div p, (2.4.1)
p = grad u, (2.4.2)
48 CHAPTER 2. MODEL EQUATIONS
where p = (p, q, r) is the vector of variables that represents the gradients. In Cartesian coordinates, it reads in one
dimension,
u
t
= p
x
, p = u
x
, (2.4.3)
in two dimensions,
u
t
= (p
x
+q
y
), p = u
x
, q = u
y
, (2.4.4)
and in three dimensions,
u
t
= (p
x
+q
y
+r
z
), p = u
x
, q = u
y
, r = u
z
. (2.4.5)
This formulation is known as the mixed formulation in the nite-element-method community [182], and widely used,
for example, in high-order methods [25, 66, 115]. I like the mixed formulation because apparently it involves only
rst-order derivatives, which are easier to discretize, especially on unstructured grids, than second-order derivatives. In
fact, a variety of successful diusion schemes have been developed based on the mixed formulation [176, 182]. Note
however that we cannot really avoid the second-derivatives in this formulation. The system is perfectly equivalent
to the diusion equation of second derivatives and the numerical schemes developed for the system is subject to the
well-known severe limitation: the explicit time step of O(h
2
), where h is a mesh spacing. I actually like it, though,
because it shows that the system is indeed equivalent to the diusion equation. Notice that the rst-order system is
parabolic in time and thus does not alter the physics of the original diusion equation. The rst-order formulation is
used just for the sake of convenience in discretization. I like such an approach very much.
2.5 Diusion Equation: First-Order Hyperbolic System
There is an interesting rst-order hyperbolic system associated with the diusion equation:
u
t
= p
x
, p
t
=
1
T
r
(u
x
p), (2.5.1)
where T
r
is a constant. In two and three dimensions, it is given, respectively, by
u
t
= (p
x
+q
y
), p
t
=
1
T
r
(u
x
p), q
t
=
1
T
r
(u
y
q), (2.5.2)
and
u
t
= (p
x
+q
y
+r
z
), p
t
=
1
T
r
(u
x
p), q
t
=
1
T
r
(u
y
q), r
t
=
1
T
r
(u
z
r). (2.5.3)
This system is equivalent to the diusion equation in the limit T
r
0 [20], and often called the hyperbolic heat
equations. Consider the vector form:
t
U+
x
F +
y
G+
y
H = S, (2.5.4)
where
U =
_
_
u
p
q
r
_
_
, F =
_
_
p
u/T
r
0
0
_
_
, G =
_
_
q
0
u/T
r
0
_
_
, H =
_
_
r
0
0
u/T
r
_
_
, S =
_
_
0
p/T
r
q/T
r
r/T
r
_
_
. (2.5.5)
The ux Jacobians are given by
A=
F
U
=
_
_
0 0 0
1
T
r
0 0 0
0 0 0 0
0 0 0 0
_
_
, B=
G
U
=
_
_
0 0 0
0 0 0 0
1
T
r
0 0 0
0 0 0 0
_
_
, C=
H
U
=
_
_
0 0 0
0 0 0 0
0 0 0 0
1
T
r
0 0 0
_
_
. (2.5.6)
2.5. DIFFUSION EQUATION: FIRST-ORDER HYPERBOLIC SYSTEM 49
Consider the projected Jacobian along an arbitrary vector, n = [n
x
, n
y
, n
z
]
t
:
A
n
= An
x
+Bn
y
+Cn
z
=
_
_
0 n
x
n
y
n
z
n
x
/T
r
0 0 0
n
y
/T
r
0 0 0
n
z
/T
r
0 0 0
_
_
. (2.5.7)
The projected Jacobian has the following eigenvalues:
1
=
T
r
,
2
=
T
r
,
3,4
= 0. (2.5.8)
The eigenvalues are real, and the associated eigenvectors can be shown to be linearly independent. Hence the system
is hyperbolic. Note that the eigenvalues are independent of n, and therefore the system describes a wave propagating
isotropically. The third and fourth eigenvalues correspond to the inconsistency damping mode [98], acting on the
quantities, q
x
p
y
and r
x
p
z
that should be zero. Although it has a hyperbolic character, it is hyperbolic only for
a very short time of the order T
r
. After a long time compared with T
r
, it behaves like the diusion equation as the
waves will be damped by the eect of the source term. There have been eorts on solving this system instead of the
diusion equation [129] as it models in part a relaxation system for the Navier-Stokes equations. I like this approach
because it avoids the discretization of second derivatives. Also, it allows us to apply numerical schemes well developed
for hyperbolic systems to diusion problems. However, these advantages do not come free. The small relaxation time
T
r
makes the system sti and consequently makes it dicult to solve numerically if not impossible.
On the other hand, if we are interested only in the steady state, then, T
r
can take any positive value [98]. Observe
that when time derivatives vanish, the hyperbolic system becomes, for any positive T
r
, in one dimension,
0 = p
x
,
0 = (u
x
p)/T
r
,
0 = p
x
,
p = u
x
,
0 = u
xx
, (2.5.9)
in two dimensions,
0 = (p
x
+q
y
),
0 = (u
x
p)/T
r
,
0 = (u
y
q)/T
r
,
0 = (p
x
+q
y
),
p = u
x
,
q = u
y
,
0 = (u
xx
+u
yy
), (2.5.10)
and in three dimensions,
0 = (p
x
+q
y
+r
z
),
0 = (u
x
p)/T
r
,
0 = (u
y
q)/T
r
,
0 = (u
z
r)/T
r
,
0 = (p
x
+q
y
+r
z
),
p = u
x
,
q = u
y
,
r = u
z
,
0 = (u
xx
+u
yy
+u
zz
). (2.5.11)
Therefore, the system is equivalent to the steady diusion equation (the Laplace equation) for arbitrary T
r
( > 0) in
contrast to the hyperbolic heat equations that are equivalent to the diusion equation at any instant of time but require
extremely small T
r
. To distinguish the two, the hyperbolic system designed to solve steady problems with arbitrary T
r
is simply called the hyperbolic diusion system. The equivalence to the Laplace equation emerges naturally also in the
energy estimate [104]. Dene the energy associated with the hyperbolic diusion system as
E
u
2
+L
2
r
(p
2
+q
2
)
2
. (2.5.12)
The governing equation for the energy can be derived by multiplying the hyperbolic system (2.5.4) by the row vector,
E
= (u, L
2
r
p, L
2
r
q), from the left:
t
E +
x
f
E
+
y
g
E
+
y
h
E
= (p
2
+q
2
+r
2
), (2.5.13)
where
f
E
= up, g
E
= uq, h
E
= ur. (2.5.14)
50 CHAPTER 2. MODEL EQUATIONS
By integrating the energy equation over the domain , we obtain
d
dt
E dV =
f
E
ndA
(p
2
+q
2
+r
2
) dV, (2.5.15)
where denotes the boundary of the domain, f
E
= (f
E
, g
E
, h
E
), n is the unit outward normal vector, and dA
denotes the innitesimal boundary area. This integral form shows that the total energy is damped by the magnitude
of the solution gradient and changed only by the boundary ux. In the steady state, we have (p, q, r) = u, and the
energy estimate reduces to
u udV
u
u
n
dA = 0, (2.5.16)
which is the well-known energy estimate of the Laplace equation.
I really like that the stiness is no longer an issue in this approach. The free parameter, T
r
, can therefore be chosen
to make the system strongly hyperbolic for all times, and can also be dened to accelerate the convergence of numerical
methods [98, 99, 104]. In fact, it naturally leads to a very fast method for computing steady state solutions of the
diusion equation with O(h) explicit time step, not O(h
2
) [98, 99, 103, 104]. One might have noticed that the rst-
order hyperbolic system is equivalent to the mixed formulation in the steady state. However, the discretization is very
dierent in that methods for hyperbolic systems, e.g., upwind scheme, are directly applicable to the hyperbolic diusion
system but not to the mixed formulation. This is a great advantage of the hyperbolic formulation because various
schemes and techniques developed in the past decades for hyperbolic systems can be applied to the discretization of
diusion. Naturally, it is also possible to develop rst-order schemes for diusion [104]. In addition, in this approach, the
solution gradient (p, q, r) can be computed to the same order of accuracy to that of the solution u [98]. Furthermore,
the Neumann condition can be implemented as the Dirichlet condition by the extra variables that will be equivalent to
the solution gradients in the steady state. Do I like this approach? Oh, yes, I do like this approach very much.
2.6 Advection-Diusion Equation
I like the advection-diusion equation:
u
t
+ (grad u) = div ( grad u) , (2.6.1)
or in the conservative form,
u
t
+ div (u grad u) = 0. (2.6.2)
In Cartesian coordinates, it is written as
1D: u
t
+au
x
= (u
x
)
x
, (2.6.3)
2D: u
t
+au
x
+bu
y
= (u
x
)
x
+ (u
y
)
y
, (2.6.4)
3D: u
t
+au
x
+bu
y
+cu
z
= (u
x
)
x
+ (u
y
)
y
+ (u
z
)
z
, (2.6.5)
or
1D: u
t
+ (au u
x
)
x
= 0, (2.6.6)
2D: u
t
+ (au u
x
)
x
+ (bu u
y
)
y
= 0, (2.6.7)
3D: u
t
+ (au u
x
)
x
+ (bu u
y
)
y
+ (cu u
z
)
z
= 0. (2.6.8)
The advection-diusion equation is a model equation for a general quasi-linear system of equations of the form, e.g.,
in one dimension,
U
t
+AU
x
= DU
xx
, (2.6.9)
where U is a vector of solutions, A is a coecient matrix having real eigenvalues, and D is a positive coecient
matrix. This includes the Navier-Stokes equations written in the quasi-linear form (see Section 4.16). Basically, it is
simply the sum of the advection and diusion equations. But to solve this numerically, we should be very careful if
we want to simply add a diusion scheme to an advection scheme. Such a scheme may fail in the intermediate case
2.6. ADVECTION-DIFFUSION EQUATION 51
0 L
0
1
Pe=10.0
x
u
Figure 2.6.1: A typical steady solution of the advection-diusion equation. u(0) = 0 and u(L) = 1.
where the advective eect and the diusive eect are equally important. In fact, some high-order schemes are known
to lose the formal accuracy due to incompatible discretizations of the advection and diusion terms [24, 115, 125].
One important consideration in preserving accuracy is to treat the advection and diusion terms in a unied manner.
For this purpose, the conservative form (2.6.2) may be particularly useful in which the advection and diusion terms
are equally treated in the ux form. Naturally, then, the integral form,
d
dt
V
udV +
S
_
u n
u
n
_
dS = 0, (2.6.10)
will be useful also.
Another diculty, which may arise in advection-diusion problems, is a rather drastic change in the character of
the equation. Consider the one-dimensional advection-diusion equation,
u
t
+au
x
= u
xx
, (2.6.11)
where a > 0, in x (0, L). This is a typical parabolic problem, and therefore we must specify solution values at both
boundaries: u(0) and u(L) which must be dierent for the solution to be non-trivial. However, if = 0, this becomes
a simple advection equation; we are allowed to specify only one value at the left boundary, i.e., u(0). The solution at
the right boundary u(L) must be left unknown and determined by the wave coming all the way from the left boundary.
Then, what happens in general ( = 0) is that the wave attempts to propagate the left boundary value u(0) to the
right, but cannot continue to the right boundary because u(0) = u(L); the diusion term resolves the discrepancy by
smoothing the solution between u(L) and the one transported by the wave from the left (see Figure 2.6.1). Certainly,
if the advective eect dominates the diusive eect, the inuence of diusion will be conned within a narrow region
near the right boundary. There is a nice dimensionless parameter that characterizes this balance:
Pe =
aL
, (2.6.12)
which is called the Peclet number. As Pe 0, the diusion dominates; thus the narrow region gets broader and
broader. On the other hand, as Pe , the advection dominates; thus the narrow region gets extremely narrow (as
if making it look like a pure advection problem by ignoring the right boundary condition u(L)). The corresponding
parameter in the Navier-Stokes equations is the Reynolds number; the narrow region is observed near a body in a ow
and called the boundary layer. Of course, we may as well ignore such a layer by discarding the diusion term completely
(solve the Euler equations instead of the Navier-Stokes equations). But if we really need to resolve the layer (e.g., need
to compute the viscous drag or the heat transfer on a body), we will have to retain the diusion term and confront a
variety of issues in computing such a solution in terms of both accuracy and eciency.
I like the advection-diusion equation. It is a very simple linear scalar equation, but contains essential diculties
that we will encounter when solving the Navier-Stokes equations for practical problems. It is useful not only for
illustrating such diculties in a simpler setting, but also for developing a new method for the Navier-Stokes equations.
52 CHAPTER 2. MODEL EQUATIONS
2.7 Advection-Reaction Equation
The advection-reaction equation is given by
u
t
+ (grad u) = ku. (2.7.1)
In Cartesian coordinates, it is written as
1D: u
t
+au
x
= ku, (2.7.2)
2D: u
t
+au
x
+bu
y
= ku, (2.7.3)
3D: u
t
+au
x
+bu
y
+cu
z
= ku. (2.7.4)
This equation models the transport of a material in a ow with a constant speed , decaying at the rate of k. This is
a typical model equation for conservation laws with a source term:
u
t
+ div f = s, (2.7.5)
where f = u and s = ku. Basically I like it, but not with large k because it makes the equation sti and very
dicult to solve numerically.
2.8 Poisson/Laplace Equations
The Poisson equation:
div grad u = f, (2.8.1)
where f is a given function of space. If f = 0, it is called the Laplace equation. In Cartesian coordinates, the Poisson
equation is written as
1D u
xx
= f(x), (2.8.2)
2D u
xx
+u
yy
= f(x, y), (2.8.3)
3D u
xx
+u
yy
+u
zz
= f(x, y, z). (2.8.4)
I like the Laplace equation because its solution is very smooth. In particular, it is well known that the minimum
and the maximum of the solution can occur only on the boundary (maaaximum principle), implying a smooth solution
in the interior. On the other hand, solutions of the Poisson equation could be non-smooth, depending on the source
term. Anyway, in both cases, the operator is the Laplacian which is very easy to discretize on regular grids: a simple
central dierencing works ne. Although typically it takes a long time to obtain a numerical solution (expensive direct
methods or a large number of iterations), the multigrid technique works extremely well for this type of equation [17].
Well, if the grid is not regular and/or high-order accuracy is required, that is a dierent story as mentioned in Section
2.3.
2.9 Cauchy-Riemann Equations
2.9.1 Velocity Components
In incompressible irrotational ows, the velocity components (u, v) satisfy the Cauchy-Riemann equations:
u
x
+v
y
= 0, for incompressibility, (2.9.1)
v
x
u
y
= 0, for irrotationality. (2.9.2)
Write the system in the vector form,
AU
x
+BU
y
= 0, (2.9.3)
2.9. CAUCHY-RIEMANN EQUATIONS 53
where
U =
_
u
v
_
, A =
_
1 0
0 1
_
, B =
_
0 1
1 0
_
. (2.9.4)
Then, A is invertible, and we nd the eigenvalues and eigenvectors of A
1
B:
1
= i, r
1
=
_
1
i
_
,
2
= i, r
2
=
_
1
i
_
, (2.9.5)
where i =
1. Therefore, this system is elliptic. If, in addition, u and v are twice dierentiable, they satisfy the
Laplace equation independently,
u
xx
+u
yy
= 0, (2.9.6)
v
xx
+v
yy
= 0. (2.9.7)
Another interesting fact is that the contours of u and v are orthogonal to each other. To see this, observe that the
dot product of the gradients vanishes by the Cauchy-Riemann equations:
(gradu) (gradv) = u
x
v
x
+u
y
v
y
= (v
y
)v
x
+ (v
x
)v
y
= 0. (2.9.8)
I like the Cauchy-Riemann equations because they are basically the same as a pair of Laplaces equations. Discretize
the Cauchy-Riemann equations by any method, you will often nd yourself solving the associated Laplace equations in
eect [61, 107]. Note that solutions of the Laplace equations do not necessarily satisfy the Cauchy-Riemann equations.
For example, the solution, (u, v) = r(sin , cos ), which represents a sold body rotation, satises the Laplace
equations, but does not satisfy the Cauchy-Riemann equations (the vorticity does not vanish). So, if you really want to
solve the Cauchy-Riemann equations, you need a numerical method that really solves the Cauchy-Riemann equations
[105, 107].
The Cauchy-Riemann system is known as a model for the elliptic subsystem of the Euler equations [93, 123] (the
acoustic subsystem for subsonic ows). In fact, it is a model for all elliptic subsystems of a general two-dimensional
dierential system. That is, any two-dimensional system of partial dierential equations can be decomposed into a set
of advection equations and/or a set of Cauchy-Riemann systems. This is very interesting. See Ref.[117] for details.
2.9.2 Stream Function and Velocity Potential
The stream function and the velocity potential are dened by
u =
x
=
y
, (2.9.9)
v =
y
=
x
. (2.9.10)
Then, it is obvious that and satisfy the following Cauchy-Riemann equations,
x
+
y
= 0, (2.9.11)
y
= 0. (2.9.12)
If, in addition, and are twice dierentiable, they satisfy the Laplace equation independently,
xx
+
yy
= 0, (2.9.13)
xx
+
yy
= 0. (2.9.14)
I like the stream function because the boundary condition is made very simple, i.e., constant along a solid body. On the
other hand, the velocity potential can be tricky because it may not be continuous for multiply-connected regions such
as a ow around a lifting airfoil (see Ref.[68], for example). Of course, the contours of and , or equivalently the
streamlines and the equi-potential lines, are orthogonal at their points of intersections because (grad) (grad) = 0.
2.9.3 Vorticity and Pressure
In Stokes ows, where the viscous terms balance with the pressure gradient, the vorticity and the pressure satisfy the
Cauchy-Riemann equations (see Section 4.25):
p
x
+
y
= 0, (2.9.15)
p
y
x
= 0, (2.9.16)
54 CHAPTER 2. MODEL EQUATIONS
where p is the pressure, = v
x
u
y
is the vorticity, and is the viscosity. I like it very much because the theory
of complex variables is directly applicable and a variety of exact solutions can be found for viscous ows just like for
potential ows [54].
2.9.4 Time-Dependent Systems
Augment the Cauchy-Riemann equations with time-derivatives as follows:
u
t
+u
x
+v
y
= 0, (2.9.17)
v
t
v
x
+u
y
= 0, (2.9.18)
which may be written in the vector form,
U
t
+AU
x
+BU
y
= 0, (2.9.19)
where
U =
_
u
v
_
, A =
_
1 0
0 1
_
, B =
_
0 1
1 0
_
. (2.9.20)
I like this system because it is no longer the Cauchy-Riemann equations, but equivalent to the Cauchy-Riemann
equations in the steady state. So, I can compute a solution of the Cauchy-Riemann equations by integrating this in time
until the solution stops changing [93, 134]. Note that the time terms have been added in such a way that the resulting
system is stable [93]. Also note that the eigenvalues and eigenvectors of the projected Jacobian A
n
= An
x
+ Bn
y
,
where (n
x
, n
y
) is an arbitrary unit vector, are given by
1
= 1, r
1
=
_
n
x
+ 1
n
y
_
,
2
= 1, r
2
=
_
n
x
1
n
y
_
. (2.9.21)
Therefore, this system is hyperbolic in time (although elliptic in space). On the other hand, it is interesting that if we
add time-derivatives to the Laplace equations,
u
t
= u
xx
+u
yy
, (2.9.22)
v
t
= v
xx
+v
yy
, (2.9.23)
these are not hyperbolic but parabolic in time. If we wish to have hyperbolic equations, we must introduce the
second-derivatives, u
tt
and v
tt
, thus resulting in the wave equations:
u
tt
= u
xx
+u
yy
, (2.9.24)
v
tt
= v
xx
+v
yy
. (2.9.25)
These wave equations can be derived also from the Cauchy-Riemann system (2.9.17) and (2.9.18). In either case, i.e.,
parabolic or hyperbolic, these equations are equivalent to the Laplace equations in their steady states. But I denitely
like the hyperbolic equations better since a simple explicit scheme will work very well (with a large time step) to quickly
reach a steady state.
2.10 Biharmonic Equation
The biharmonic equation is a fourth-order partial dierential equation given by
div grad(div grad u) = f, (2.10.1)
or equivalently by
2
u = f, (2.10.2)
where f is a given function of space. In Cartesian coordinates, the biharmonic equation is written as
1D u
xxxx
= f(x), (2.10.3)
2D u
xxxx
+u
yyyy
+ 2u
xxyy
= f(x, y), (2.10.4)
3D u
xxxx
+u
yyyy
+u
zzzz
+ 2u
xxyy
+ 2u
yyzz
+ 2u
zzxx
= f(x, y, z). (2.10.5)
2.11. EULER-TRICOMI EQUATION 55
I like the biharmonic equation because it does arise in CFD. Stokes equations, which govern very low Reynolds number
ows, formulated by the stream function and the vorticity, involve the biharmonic equation for the stream function.
The biharmonic equation is typically solved with the following boundary conditions:
u = g
1
(x, y, z),
u
n
= g
2
(x, y, z), (2.10.6)
where g
1
and g
2
are given functions, and u/n is the derivative in the direction normal to the boundary. For two-
dimensional Stokes ow, u corresponds to the stream function. Then, the rst boundary condition may be specied
with g
1
= 0 to for a solid wall boundary, meaning that the wall is a streamline. The second boundary condition species
the velocity tangential to the boundary, and thus the no-slip condition can be specied with g
2
= 0 at the wall.
Discretization of fourth-order derivatives may be straightforward on a regular computational grid, but not so
straightforward on irregular grids. An interesting approach is to write the biharmonic equation as a pair of the Laplace
equations [160]:
div grad u = v, (2.10.7)
div grad v = f, (2.10.8)
where v is a new variable. It is easy to verify that these equations are equivalent to the biharmonic equation for u. Then,
the biharmonic equation can be solved by any well-known method for the Laplace equation, e.g., multigrid methods.
I like it. It is always nice to reduce the order of partial dierential equations. However, the boundary condition for
the new variable v is not known and may be dicult to apply in some applications [2]. Nevertheless, the number of
boundary conditions required is basically two [160]. For Stokes ows, the boundary conditions as described above will
suce and the boundary condition is not necessary for v. See Ref.[160] for details.
2.11 Euler-Tricomi Equation
Euler-Tricomi equation,
yu
xx
+u
yy
+u
zz
= 0, (2.11.1)
is an interesting model equation often used in the study of numerical methods for transonic ows. Some people call it
the Tricomi equation, and I dont know exactly why they ignore Euler. Anyway, it is an interesting equation because it
changes its type depending on y: it is hyperbolic where y < 0 and elliptic where y > 0. It thus models transonic ows,
which is hyperbolic when the local velocity exceeds the speed of sound and elliptic when the velocity is lower than the
speed of sound (see Section 3.17 and also Section 3.20). The Euler-Tricomi equation actually develops steep waves
in the hyperbolic region and yields a smooth solution in the elliptic region. This model has been employed for testing
a numerical algorithm designed for transonic ow calculations [58, 137]. To deal with the mixed behavior, a mixed
numerical algorithm (i.e., central-upwind) is employed in Ref.[58] whereas the equation is made hyperbolic in pseudo
time and solved by an upwind method in Ref.[137]. The latter seems to be a popular approach as typically done for
the Euler equations, which are physically hyperbolic in time but mixed hyperbolic-elliptic in space as shown in Section
3.17.
2.12 Relaxation Model
I like relaxation models but I dont know much about them. See Refs.[62, 88] for details. As far as I know, the
one-dimensional relaxation model is given by
u
t
+v
x
= 0, (2.12.1)
v
t
+a
2
u
x
= (cu v)/, (2.12.2)
where > 0, a
2
> c
2
. For 1, this system behaves like the advection-diusion equation,
u
t
+cu
x
= (a
2
c
2
)u
xx
. (2.12.3)
56 CHAPTER 2. MODEL EQUATIONS
This can be shown by the Chapman-Enskog expansion [81]. Note that this is a simple model for moment equations
(derived from the Boltzmann equation) which behave like the Navier-Stokes equations in the limit of small relaxation
time. The two-dimensional version is given by
u
t
+v
x
+w
y
= 0, (2.12.4)
v
t
+a
2
u
x
= (v ru)/, (2.12.5)
w
t
+b
2
u
y
= (w su)/, (2.12.6)
(2.12.7)
where > 0, a
2
> r
2
, b
2
> s
2
. For 1, this system behaves like the following advection-diusion equation,
u
t
+ru
x
+su
y
=
_
(a
2
r
2
)u
xx
2rsu
xy
+ (b
2
s
2
)u
yy
. (2.12.8)
Note that it is more advantageous to solve the rst-order relaxation system than the advection-diusion equations
because you can avoid any diculties in dealing with mixed derivatives [165].
We may generalize the relaxation model, and write a system of nonlinear conservation laws,
U
t
+F
x
= 0, (2.12.9)
as a linear system with a source term, by introducing a new set of variables V,
U
t
+V
x
= 0, (2.12.10)
V
t
+AU
x
=
F(U) V
, (2.12.11)
where A and are constants carefully chosen such that the solution to this system is equivalent to the original
conservation law: the eigenvalues of A must satisfy the so-called sub-characteristic condition and must be very small
(see Ref.[81] for details). I like to avoid nonlinearity in this way (A is now constant), but we now have a sti source
term instead ( must be very small). So, it is dicult anyway. Methods which use this approach to avoid nonlinearity
are called relaxation methods. See Ref.[81] for details.
2.13 Burgers Equation
I like Burgers equation (also known as inviscid Burgers equation):
u
t
+uu
x
= 0. (2.13.1)
This can be written in the conservative form,
u
t
+
_
u
2
/2
_
x
= 0. (2.13.2)
Then, we can derive a shock speed V
s
by integrating this over a discontinuity across two states, u
L
and u
R
, i.e., from
the Rankine-Hugoniot relation (1.22.1),
V
s
=
u
L
+u
R
2
. (2.13.3)
Also, if we multiply Equation (2.13.2) by 2u, then we get another conservation form,
(u
2
)
t
+
_
2
3
u
3/2
_
x
= 0, (2.13.4)
from which we nd
V
s
=
2
3
_
u
L
+u
R
u
L
u
R
u
L
+u
R
_
. (2.13.5)
So, the two conservation forms share the strong solutions (those that satisfy the dierential form), but have dierent
weak solutions (those that satisfy the integral form). In fact, it is natural that they have dierent weak solutions
2.14. TRAFFIC EQUATIONS 57
because the conserved quantities are dierent: Equation (2.13.2) conserves u while Equation (2.13.4) conserves u
2
. I
like this particular example because it tells us that it is very important to choose the right conservative variables so
that we obtain physically meaningful solutions.
Two-dimensional version is given by
u
t
+uu
x
+vu
y
= 0, (2.13.6)
v
t
+uv
x
+vv
y
= 0. (2.13.7)
The following version is also often used,
u
t
+uu
x
+u
y
= 0. (2.13.8)
The latter is basically the same, at a steady state, as the one-dimensional equation (2.13.1), with y taken as a time-like
axis. There is also the three-dimensional version:
u
t
+uu
x
+vu
y
+wu
z
= 0, (2.13.9)
v
t
+uv
x
+vv
y
+wv
z
= 0, (2.13.10)
w
t
+uw
x
+vw
y
+ww
z
= 0. (2.13.11)
Incidentally, some people think that it is Burgers equation. This is wrong. It is Burgers equation because it is the
equation of Burgers, not of Burger.
2.14 Trac Equations
The trac equation is given by
t
+ (v)
x
= 0, (2.14.1)
where is the density of cars traveling on the road (cars/distance) and v is a local speed of a car. It is reasonable to
assume that v is a function of , e.g.,
v
v
max
= 1
max
, (2.14.2)
which means that the speed goes up if there are less cars while the speed goes down if there are more cars. Then, the
characteristic speed is given by
f
=
(v)
= v
max
_
1
2
max
_
. (2.14.3)
The shock speed V
s
can also be found, by the Rankine-Hugoniot relation (1.22.1), as
V
s
=
f
R
f
L
L
= v
max
_
1
R
+
L
max
_
. (2.14.4)
I like the trac equation because the relation between the speed and the car density, such as Equation (2.14.2),
is required for the trac equation to be closed but can be determined in any way we want. Various forms may be
proposed for modeling a real trac ow. But remember that we would have to derive the characteristic and shock
speeds for each relation we choose. For example, why dont you take this relation,
v
v
max
= 1
2
2
max
, (2.14.5)
and derive the characteristic and shock speeds? See Ref.[81, 170] for more details on the trac equation.
58 CHAPTER 2. MODEL EQUATIONS
2.15 Viscous Burgers Equations
Viscous Burgers equation is given by
u
t
+uu
x
= u
xx
, (2.15.1)
where is a positive constant. I like this equation because it is a simple model for the Navier-Stokes equations. Of
course, I like the two-dimensional version also:
u
t
+uu
x
+vu
y
= (u
xx
+u
yy
) , (2.15.2)
v
t
+uv
x
+vv
y
= (v
xx
+v
yy
) , (2.15.3)
which is in fact the two-dimensional incompressible Navier-Stokes equations without the pressure gradient. I like this
one also,
u
t
+uu
x
+u
y
= u
xx
, (2.15.4)
which will behave like the one-dimensional equation (with time-like y) at a steady state. Finally, of course, I like the
three-dimensional version,
u
t
+uu
x
+vu
y
+wu
z
= (u
xx
+u
yy
+u
zz
) , (2.15.5)
v
t
+uv
x
+vv
y
+wv
z
= (v
xx
+v
yy
+v
zz
) , (2.15.6)
w
t
+uw
x
+vw
y
+ww
z
= (w
xx
+w
yy
+w
zz
) . (2.15.7)
One of the most interesting features of the viscous Burgers equation is that its solution can be constructed from a
solution of the diusion equation. This is very interesting, isnt it? See Section 6.4.1 for details.
Chapter 3
Euler Equations
3.1 Thermodynamic Relations
I like gases that obey the ideal gas law (also called the thermal equation of state):
p = RT, (3.1.1)
where p is the gas pressure, is the gas density, T is the gas temperature, and R is the gas constant which is 287
N m/kg for the sea-level air. Naturally, a gas that obeys this relation is called an ideal gas (or a thermally perfect gas).
The ideal gas law implies that the state of a gas at a point in a ow is determined by any two of the three variables:
the pressure p, the density , and the temperature T. These variables are then called state variables: the variables
which depend only on the state of gas and not on how it was changed from one state to another. It is interesting that
this relation (3.1.1) has nothing to do with time. In a ow, all these variables will generally change in time. But their
relation (3.1.1) holds at any instant of time at any location in the ow (no matter how violent the ow is). In other
words, the gas is always in equilibrium. Of course, this may not be true in reality, but this is the basic assumption of
thermodynamics, and in fact, it applies remarkably well to a wide range of ows, from low-speed to supersonic. That
is why I like thermodynamics so much. It is very useful for uid dynamics and CFD.
The internal energy of a gas is dened as the sum of all microscopic energies related to molecular motions (trans-
lation, rotation, vibration, etc). In this book, we use e to denote the specic internal energy (specic means per-unit-
mass). The internal energy is generally a state variable and depends on two other state variables, e.g, T and p. But it
is well known that for ideal gases it depends only on the temperature:
e = e(T). (3.1.2)
Note now that from the equation of state that p/ also depends only on T. This makes us want to introduce another
quantity called the specic enthalpy, denoted by h:
h e +
p
, (3.1.3)
so that the enthalpy also depends only on the temperature,
h = h(T). (3.1.4)
It is interesting that the rst law of thermodynamics, which is a simple energy conservation law, can be expressed in
terms of both e and h: with the specic volume v dened by v = 1/,
dq = de +pdv, (3.1.5)
or
dq = dh
1
dp, (3.1.6)
where dq is an added specic heat. I must mention here that both de and dh depend only on the state while dq
depends on the process (how the heat is added), and that the work done, say dw, depends on the process also but this
59
60 CHAPTER 3. EULER EQUATIONS
has been replaced by pdv, i.e., a reversible work done solely by the pressure (i.e., no viscous forces). Note that the
above equations (3.1.5) and (3.1.6) are true for non-ideal gases also. Now, what is even more interesting is that if we
dene the constant-volume specic heat c
v
and the constant-pressure specic heat c
p
as the amounts of heat required
to increase the temperature of unit mass by one degree at constant volume and pressure respectively,
c
v
=
_
q
T
_
v
, c
p
=
_
q
T
_
p
, (3.1.7)
then, it immediately follows from Equations (3.1.5) and (3.1.6) that
c
v
=
_
e
T
_
v
, c
p
=
_
h
T
_
p
. (3.1.8)
This is very interesting. It shows that c
v
represents a change in e per unit change in T at constant volume while c
p
represents a change in h per unit change in T at constant pressure. It is known from kinetic theory that for molecules
with degrees of freedom for translational and rotational energy, the specic heats are given by
c
v
=
2
R, (3.1.9)
c
p
=
_
1 +
2
_
R, (3.1.10)
and thus their ratio is independent of R,
=
c
p
c
v
=
+ 2
, (3.1.11)
where is called the ratio of specic heats (also called the adiabatic exponent). For monatomic gases, we have = 3
and hence = 5/3; for diatomic gases (e.g., air), we have = 5 and hence = 7/5. These results imply that is a
constant. But this is not a general result; vibrational energy of molecules has not been taken into account. If the gas
temperature is high enough to excite the vibrational energy, then it will generally depend on T.
For ideal gases, both e and h are functions of T only, and therefore, we obtain from Equation (3.1.5)
c
v
=
_
q
T
_
v
=
e
T
, (3.1.12)
c
p
=
_
q
T
_
p
=
e
T
+p
_
v
T
_
p
= c
v
+p
_
v
T
_
p
. (3.1.13)
Since the last term above becomes R/p by the ideal gas law (3.1.1), we have the important relation:
c
p
c
v
= R. (3.1.14)
This is very nice. Now we can solve this with Equations (3.1.9), (3.1.10), and (3.1.11) for c
v
and c
p
to get
c
v
=
1
1
R, (3.1.15)
c
p
=
1
R. (3.1.16)
For high temperature gases, may depend on T, and so are c
v
and c
p
. Then, by integrating Equation (3.1.8), which
are now ordinary derivatives, we obtain
e(T) =
c
v
dT + constant, (3.1.17)
h(T) =
c
p
dT + constant. (3.1.18)
Yet, for a calorically perfect gas, which is dened as a gas with constant specic heats, we can write
e = c
v
T =
R
1
T, (3.1.19)
h = c
p
T =
R
1
T. (3.1.20)
3.1. THERMODYNAMIC RELATIONS 61
This is very nice. For example, we can arrange the equation of state into the form,
p = ( 1)e = ( 1)
_
E
v
2
2
_
, (3.1.21)
where v is the velocity vector and E is the specic total energy,
E =
1
1
p
+
1
2
v
2
. (3.1.22)
Relation (3.1.21) is then called the equation of state for a calorically perfect gas (a calorically perfect gas is also
called a polytropic gas [76, 81]). By the way, some people use E and H to denote the total energy and enthalpy:
E =
p
1
+
1
2
v
2
and H = E +p. Others use E to denote the specic total energy as in Equation (3.1.22) and H to
denote the specic total enthalpy,
H = E +
p
. (3.1.23)
I like the latter, but the former is ne too. They are just dierent denitions. I like not to discuss which is better.
Throughout this book, I will use the latter simply because I like it.
There is another quantity that I like. Consider the following,
dq
T
=
de
T
+p
dv
T
, (3.1.24)
which is the rst law (3.1.5) divided by the temperature. This can be written as, by the ideal gas law,
dq
T
= c
v
dT
T
+Rd
_
1
_
, (3.1.25)
which further can be written as, for a calorically perfect gas, as
dq
T
=
d(c
v
T)
T
d(R)
= d
_
ln T
cv
ln
R
_
. (3.1.26)
Therefore,
dq
T
can be written as an exact dierential; it is thus a state variable (although dq is not a state variable).
This is dened as an entropy s:
ds
dq
T
= d
_
ln T
cv
ln
R
_
. (3.1.27)
This is nice but we must recall that this relation is subject to the assumption of reversible process. In general, e.g., for
viscous ows with heat transfer, this is not true, and the second law of thermodynamics states that the entropy must
increase:
ds =
dq
T
+ds
, (3.1.28)
where ds
0, (3.1.29)
and for an adiabatic and reversible process we have
ds = 0. (3.1.30)
That is, the entropy is constant for an adiabatic and reversible process. Such a process is called isentropic. For an
isentropic process, some very important relations can be derived. We have from Equation (3.1.27)
ds = c
v
dT
T
R
d
, (3.1.31)
62 CHAPTER 3. EULER EQUATIONS
and thus by the ideal gas law (dp/p = d/ +dT/T),
ds = c
v
dp
p
c
p
d
= c
p
dT
T
R
dp
p
. (3.1.32)
For a calorically perfect gas, we can write these as
1
R
ds =
dT
T
( 1)
d
=
dp
p
d
=
dT
T
( 1)
dp
p
, (3.1.33)
and thus
1
R
ds = d
_
ln
T
1
_
= d
_
ln
p
_
= d
_
ln
T
p
1
_
. (3.1.34)
Therefore, we obtain the following important relations for an isentropic process (ds = 0):
T
1
= constant, (3.1.35)
p
= constant, (3.1.36)
T
p
1
= constant, (3.1.37)
where constants on the right hand sides are, of course, not all the same. These relations are generally called isentropic
relations or adiabatic relations (dq = 0). They are very useful for a wide range of applications because many ows can
be considered as adiabatic and reversible. For example, a supersonic ow over an airfoil can be adiabatic and reversible
except inside possible shock waves or inside a thin viscous/thermal boundary layer on the airfoil. Even for a curved
shock wave across which the entropy increases by dierent amount for dierent streamlines, the ow is isentropic along
each streamline (except for the jump across the shock wave). Note that in such a case the constants in these relations
dier from one streamline to another.
Overall, I like a calorically perfect gas since it has actually been used quite successfully for a wide range of ows.
So, in the rest of this book, I will simply call it a perfect gas.
3.2 Speed of Sound
The speed of sound is a speed at which a disturbance produced by a sound wave travels in a gas. The distur-
bance is generally very small (a sound wave is not at all as strong as a shock wave), so that the process is considered
as isentropic. The speed of sound a is then given by
a
2
=
_
p
_
s
. (3.2.1)
For a perfect gas, we have the isentropic relation p/
. (3.2.2)
3.3. EULER EQUATIONS 63
Using the equation of state, p = RT, we can write this also in terms of T,
a
2
= RT. (3.2.3)
Generally, I like to use a to denote the speed of sound, but sometimes I will instead use c to avoid a possible conict
with other variables. But in the special case that the speed of sound is constant everywhere in a ow eld (e.g.,
isothermal ows), I think I would rather use c than a since it looks like standing for constant.
3.3 Euler Equations
Conservative Form
The conservative form of the Euler equations:
t
+ div(v) = 0, (3.3.1)
t
(v) + div(vv) + grad p = 0, (3.3.2)
t
(E) + div(vH) = 0. (3.3.3)
Some people claim that only the momentum equation should be called the Euler equation because Leonhard Euler
actually derived the momentum equation only. But usually all these equations (momentum + continuity + energy) are
called the Euler equations. Either way, I like the Euler equations.
Primitive Form
The primitive form of the Euler equations:
D
Dt
+ div v = 0, (3.3.4)
Dv
Dt
+ grad p = 0, (3.3.5)
Dp
Dt
+c
2
div v = 0. (3.3.6)
Other Energy Equations
I like energy equations because the energy equations, (3.3.3) and (3.3.6), may be replaced by any of the followings:
(H)
t
p
t
+ div(vH) = 0, (3.3.7)
DE
Dt
+ div (pv) = 0, (3.3.8)
De
Dt
+p div v = 0, (3.3.9)
Dh
Dt
Dp
Dt
= 0, (3.3.10)
DH
Dt
p
t
= 0, (3.3.11)
D
Dt
_
v
2
2
_
+ (grad p) v = 0, (3.3.12)
64 CHAPTER 3. EULER EQUATIONS
or even by the entropy equation. From Equation (3.1.27), we have
ds = c
v
dT
T
R
d
, (3.3.13)
which can be written for a perfect gas as
ds =
de
T
p
T
2
d. (3.3.14)
Therefore, we obtain from Equations (3.3.9) and (3.3.4)
Ds
Dt
= 0. (3.3.15)
According to the second law of thermodynamics, this should be interpreted as an inequality in general:
Ds
Dt
0. (3.3.16)
3.4 1D Euler Equations
3.4.1 Conservative Form
U
t
+F
x
= 0, (3.4.1)
where
U =
_
u
E
_
_, F =
_
_
u
u
2
+p
uH
_
_. (3.4.2)
Flux Jacobian
A =
F
U
=
_
_
0 1 0
( 3)
u
2
2
(3 )u 1
_
1
2
u
2
H
_
u H + (1 )u
2
u
_
_
. (3.4.3)
Eigenstructure
A = RL, (3.4.4)
where
=
_
_
u c 0 0
0 u 0
0 0 u +c
_
_
, (3.4.5)
R =
_
_
1 1 1
u c u u +c
H uc u
2
/2 H +uc
_
_
, (3.4.6)
3.4. 1D EULER EQUATIONS 65
L =
_
_
1
2
_
1
2c
2
u
2
+
u
c
_
1
2
_
1
c
2
u +
1
c
_
1
2c
2
1
1
2c
2
u
2
1
c
2
u
1
c
2
1
2
_
1
2c
2
u
2
u
c
_
1
2
_
1
c
2
u
1
c
_
1
2c
2
_
_
, (3.4.7)
LdU =
_
_
dp c du
2c
2
dp c
2
d
c
2
dp +c du
2c
2
_
_
. (3.4.8)
It is interesting to note that the quantity
dpc
2
d
c
2
corresponds to the entropy change ds and the corresponding right
eigenvector indicates (as discussed in Section 1.16.2) how the conservative variables change due to this entropy change,
_
(u)
(E)
_
_
_
_
1
u
u
2
/2
_
_. (3.4.9)
This is often called an entropy wave. The associated eigenvalue u gives the speed at which the entropy wave travels:
it moves with the local uid velocity. It is actually more interesting to look at it in the primitive form where we see
exactly what variable is aected by the entropy change.
3.4.2 Primitive Form
W
t
+A
w
W
x
= 0, (3.4.10)
where
W =
_
_
u
p
_
_
, A
w
=
_
_
u 0
0 u 1/
0 c
2
u
_
_
. (3.4.11)
Incidentally, I wouldnt call A
w
the primitive Jacobian. This is because A
w
is not a Jacobian [166]: there does not
exist a (ux) vector H such that H/W = A
w
. So, I simply call A
w
the coecient matrix of the primitive form of
the Euler equations.
Eigenstructure
A
w
= R
w
L
w
, (3.4.12)
where
=
_
_
u c 0 0
0 u 0
0 0 u +c
_
_
, R
w
=
_
2c
1
2c
1
2
0
1
2
c
2
0
c
2
_
_
, (3.4.13)
L
w
=
_
_
0 1
1
c
1 0
1
c
2
0 1
1
c
_
_
, L
w
dU =
_
_
du
dp
c
d
dp
c
2
du +
dp
c
_
_
. (3.4.14)
66 CHAPTER 3. EULER EQUATIONS
The second component above represents the entropy wave, and we thus nd
_
u
p
_
_
_
_
1
0
0
_
_. (3.4.15)
Therefore, it is the density that changes due to the entropy wave; the velocity and the pressure are not aected. A
typical example is a contact surface (or contact discontinuity) observed in shock tube problems: a surface across which
two dierent gases are initially in contact moves with the local uid speed carrying the density jump. In fact, the
entropy wave is sometimes called the contact wave.
Change of Variables
dU =
U
W
dW, dW =
W
U
dU, (3.4.16)
where
U
W
=
_
_
1 0 0
u 0
u
2
/2 u 1/( 1)
_
_
, (3.4.17)
W
U
=
_
U
W
_
1
=
_
_
1 0 0
0
( 1)u
2
/2 ( 1)u 1
_
_
. (3.4.18)
Note that the coecient matrix A
w
is similar to the conservative Jacobian A,
A
w
=
W
U
A
U
W
. (3.4.19)
Hence, the eigenvectors are related by
L = L
w
W
U
, R =
U
W
R
w
, (3.4.20)
and the eigenvalues are the same as those of the conservative Jacobian because similar matrices share the same
eigenvalues (see Section 1.11).
3.4.3 Characteristic Variables
Multiply the primitive form (3.4.10) by L
w
from the left to get
L
w
W
t
+L
w
A
w
W
x
= 0, (3.4.21)
L
w
W
t
+L
w
A
w
R
w
L
w
W
x
= 0, (3.4.22)
W
c
t
+W
c
x
= 0, (3.4.23)
where the new variables dened by
W
c
= L
w
W =
_
_
u
1
c
p
1
c
2
p
u +
1
c
p
_
_
, (3.4.24)
3.4. 1D EULER EQUATIONS 67
are called the characteristic variables, and the equation (3.4.23) is called the characteristic form. It is very nice that
the system (3.4.23) is a completely diagonalized form of the Euler equations, so that we have a set of independent
scalar nonlinear advection equations:
_
u
t
1
c
p
t
_
+ (u c)
_
u
x
1
c
p
x
_
= 0, (3.4.25)
_
t
1
c
2
p
t
_
+u
_
x
1
c
2
p
x
_
= 0, (3.4.26)
_
u
t
+
1
c
p
t
_
+ (u +c)
_
u
x
+
1
c
p
x
_
= 0. (3.4.27)
The second equation corresponds to an entropy advection, and the other two correspond to an acoustic wave propa-
gation. This shows that the entropy is simply convected along a streamline and not aected by the acoustic waves.
In other words, the entropy is constant through the acoustic waves (unless they are strong enough to become shock
waves). For adiabatic (isentropic) ows, the characteristic variables can be integrated:
W
c
=
_
_
u
1
c
p
1
c
2
p
u +
1
c
p
_
_
=
_
_
u
2
1
c
_
s
_
u +
2
1
c
_
_
_
, (3.4.28)
where the entropy s has been scaled by ( 1)/(R). These integrated quantities are called Riemann invariants.
We thus have
t
_
u
2
1
c
_
+ (u c)
x
_
u
2
1
c
_
= 0, (3.4.29)
s
t
+u
s
x
= 0, (3.4.30)
t
_
u +
2
1
c
_
+ (u +c)
x
_
u +
2
1
c
_
= 0. (3.4.31)
This system is often used to implement a characteristic boundary condition in CFD [51]. Now, if we assume = 3,
this system reduces to a set of Burgers equations.
t
(u c) + (u c)
x
(u c) = 0, (3.4.32)
s
t
+u
s
x
= 0, (3.4.33)
t
(u +c) + (u +c)
x
(u +c) = 0. (3.4.34)
On the other hand, if we assume a simple wave, i.e., for example,
u
2
1
c = constant, s = constant, (3.4.35)
everywhere (alternatively, u +
2
1
c = constant may be assumed), then it can be shown that the system of Equations
(3.4.29), (3.4.30), and (3.4.31) reduces to the following single Burgers equation:
t
(u +c) + (u +c)
x
(u +c) = 0. (3.4.36)
To show this, let
R
= u
2
1
c = constant, (3.4.37)
68 CHAPTER 3. EULER EQUATIONS
and solve for c,
c =
1
2
(u R
). (3.4.38)
Then, we nd
R
+
= u +
2
1
c = 2u R
, (3.4.39)
u +c =
+ 1
2
u
1
2
R
, (3.4.40)
and derive the relation,
R
+
=
4
+ 1
_
(u +c) +
3
4
R
_
. (3.4.41)
Finally, substitute this into Equation (3.4.31) and multiply the result by
+1
4
to arrive at Equation (3.4.36).
After all, I like the one-dimensional Euler system because it can be diagonalized completely and thus written as a
set of independent scalar advection equations as shown above. This is very nice but dicult in higher dimensions.
3.4.4 Parameter Vector
I like the parameter vector Z [128],
Z =
_
_
1
u
H
_
_
=
_
_
z
1
z
2
z
3
_
_
, (3.4.42)
because every component of conservative variables and uxes can be expressed as quadratics in the components of Z,
U =
_
u
E
_
_ =
_
_
z
2
1
z
1
z
2
z
1
z
3
+
1
z
2
2
2
_
_
, (3.4.43)
F =
_
_
u
u
2
+p
uH
_
_ =
_
_
z
1
z
2
z
2
2
+p
z
2
z
3
_
_, (3.4.44)
p =
1
_
z
1
z
3
z
2
2
2
_
. (3.4.45)
This is very useful, for instance, for linearizing the Euler equations [128].
3.5 2D Euler Equations
3.5.1 Conservative Form
U
t
+F
x
+G
y
= 0, (3.5.1)
where
U =
_
u
v
E
_
_
, F =
_
_
u
u
2
+p
uv
uH
_
_
, G =
_
_
v
uv
v
2
+p
vH
_
_
. (3.5.2)
3.5. 2D EULER EQUATIONS 69
Normal Flux
Projection of the ux in the direction of n = [n
x
, n
y
]
t
:
F
n
= [F, G] n = Fn
x
+Gn
y
=
_
_
q
n
q
n
u +pn
x
q
n
v +pn
y
q
n
H
_
_
, (3.5.3)
where
q
n
= un
x
+vn
y
. (3.5.4)
Jacobians
A =
F
U
=
_
_
0 1 0 0
( 1)
q
2
2
u
2
(3 )u (1 )v 1
uv v u 0
_
1
2
q
2
H
_
u H + (1 )u
2
(1 )uv u
_
_
, (3.5.5)
B =
G
U
=
_
_
0 0 1 0
uv v u 0
( 1)
q
2
2
v
2
(1 )u (3 )v 1
_
1
2
q
2
H
_
v (1 )uv H + (1 )v
2
v
_
_
, (3.5.6)
A
n
=
F
n
U
= An
x
+Bn
y
=
_
_
0 n
x
n
y
0
( 1)
q
2
2
n
x
uq
n
un
x
( 1)un
x
+q
n
un
y
( 1)vn
x
( 1)n
x
( 1)
q
2
2
n
y
vq
n
vn
x
( 1)un
y
vn
y
( 1)vn
y
+q
n
( 1)n
y
_
1
2
q
2
H
_
q
n
Hn
x
( 1)uq
n
Hn
y
( 1)vq
n
q
n
_
_
=
_
_
0 n
t
0
( 1)
q
2
2
n q
n
v vn + (1 )nv +q
n
I ( 1)n
_
1
2
q
2
H
_
q
n
Hn
t
(1 )v
t
q
n
q
n
_
_
, (3.5.7)
where q
2
= u
2
+v
2
, v = [u, v]
t
, and I is the 22 identify matrix. I like A
n
very much because it is very convenient:
simply by taking n = [1, 0]
t
and [0, 1]
t
, we obtain A and B respectively.
Eigenstructure
A
n
= R
n
n
L
n
, (3.5.8)
70 CHAPTER 3. EULER EQUATIONS
where
n
=
_
_
q
n
c 0 0 0
0 q
n
0 0
0 0 q
n
+c 0
0 0 0 q
n
_
_
, (3.5.9)
R
n
=
_
_
1 1 1 0
u cn
x
u u +cn
x
x
v cn
y
v v +cn
y
y
H q
n
c q
2
/2 H +q
n
c q
_
, (3.5.10)
L
n
=
_
_
1
2
_
1
2c
2
q
2
+
q
n
c
_
1
2
_
1
c
2
u +
n
x
c
_
1
2
_
1
c
2
v +
n
y
c
_
1
2c
2
1
1
2c
2
q
2
1
c
2
u
1
c
2
v
1
c
2
1
2
_
1
2c
2
q
2
q
n
c
_
1
2
_
1
c
2
u
n
x
c
_
1
2
_
1
c
2
v
n
y
c
_
1
2c
2
q
x
y
0
_
_
, (3.5.11)
L
n
dU =
_
_
dp c dq
n
2c
2
dp c
2
d
c
2
dp +c dq
n
2c
2
dq
_
, (3.5.12)
where [
x
,
y
]
t
is a tangent vector (perpendicular to n) such as
_
x
y
_
=
_
n
y
n
x
_
, (3.5.13)
and q
= u
x
+v
y
. (3.5.14)
I like R
n
and L
n
very much. Again, they are very convenient: simply by taking n = [1, 0]
t
and [0, 1]
t
, we obtain
eigenvectors of A and B. It is also nice that the normal Jacobian A
n
can be diagonalized as
n
= L
n
A
n
R
n
. (3.5.15)
It would be even nicer if A and B could be diagonalized simultaneously so that the Euler equations could be written as
a set of scalar advection equations. This was easy for the one-dimensional Euler equations, but it seems very dicult
in two (and also three) dimensions. But it is still nice that the steady two-dimensional Euler equations can be fully
diagonalized (see Section 3.17).
3.5.2 Primitive Form
W
t
+A
w
W
x
+B
w
W
y
= 0, (3.5.16)
where
W =
_
u
v
p
_
_
, A
w
=
_
_
u 0 0
0 u 0 1/
0 0 u 0
0 c
2
0 u
_
_
, B
w
=
_
_
v 0 0
0 v 0 0
0 0 v 1/
0 0 c
2
v
_
_
. (3.5.17)
Of course, as in Section 3.4.2, I wouldnt call A
w
and B
w
the primitive Jacobians because they are not Jacobians.
I do like to call them the coecient matrices of the primitive form of the Euler equations.
3.5. 2D EULER EQUATIONS 71
Eigenstructure
A
w
n
= R
w
n
n
L
w
n
, (3.5.18)
where
A
w
n
=
_
_
q
n
n
x
n
y
0
0 q
n
0 n
x
/
0 0 q
n
n
y
/
0 c
2
n
x
c
2
n
y
q
n
_
_
, (3.5.19)
n
=
_
_
q
n
c 0 0 0
0 q
n
0 0
0 0 q
n
+c 0
0 0 0 q
n
_
_
, (3.5.20)
R
w
n
=
_
_
1 1 1 0
n
x
c
0
n
x
c
n
y
c
0
n
y
c
c
2
0 c
2
0
_
_
, L
w
n
=
_
_
0
n
x
2c
n
y
2c
1
2c
2
1 0 0
1
c
2
0
n
x
2c
n
y
2c
1
2c
2
0
x
y
0
_
_
, (3.5.21)
L
w
n
dW =
_
_
dp c dq
n
2c
2
dp c
2
d
c
2
dp +c dq
n
2c
2
dq
_
. (3.5.22)
Note that the fourth component represents a wave which brings a change in the tangential velocity component. This
is often called a shear wave or a vorticity wave (because such a velocity jump can be represented by a vortex sheet).
Change of Variables
dU =
U
W
dW, dW =
W
U
dU, (3.5.23)
where
U
W
=
_
_
1 0 0 0
u 0 0
v 0 0
q
2
/2 u v 1/( 1)
_
_
, (3.5.24)
W
U
=
_
U
W
_
1
=
_
_
1 0 0 0
0 0
0
1
0
( 1)q
2
/2 ( 1)u ( 1)v 1
_
_
, (3.5.25)
A
w
n
=
W
U
A
n
U
W
, L
n
= L
w
n
W
U
, R
n
=
U
W
R
w
n
. (3.5.26)
72 CHAPTER 3. EULER EQUATIONS
3.5.3 Symmetric Form
I like the symmetric form of the Euler equations because the coecient matrices become very simple.
V
t
+A
v
V
x
+B
v
V
y
= 0, (3.5.27)
where
dV =
_
_
dp
c
du
dv
dp a
2
d
_
_
, A
v
=
_
_
u c 0 0
c u 0 0
0 0 u 0
0 0 0 u
_
_
, B
v
=
_
_
v 0 c 0
0 v 0 0
c 0 v 0
0 0 0 v
_
_
, (3.5.28)
and dV is called symmetric variables (symmetric variables are not unique; see Refs.[10, 40] for other choices). Look
at the matrices above. They are very sparse and simple. Moreover, I like the fourth equation which clearly shows how
the entropy evolves:
s
t
+us
x
+vs
y
= 0, (3.5.29)
where the entropy s has been scaled by ( 1)p/R. That is, it is convected along a streamline, thus implying the
isentropic ow.
Change of Variables
The symmetric variables and the conservative variables are transformed into each other as follows:
dU =
U
V
dV, dV =
V
U
dU, (3.5.30)
where the transformation matrices are given by
U
V
=
_
c
0 0
1
c
2
u
c
0
u
c
2
v
c
0
v
c
2
c
_
q
2
2c
2
+
1
1
_
u v
q
2
2c
2
_
_
, (3.5.31)
V
U
=
_
U
V
_
1
=
_
_
1
2
q
2
c
( 1)
u
c
( 1)
v
c
1
c
0 0
0
1
0
1
2
q
2
c
2
( 1)u ( 1)v 1
_
_
. (3.5.32)
Eigenstructure
A
v
n
= R
v
n
n
L
v
n
, (3.5.33)
3.5. 2D EULER EQUATIONS 73
where
A
v
n
=
_
_
q
n
c n
x
c n
y
0
c n
x
q
n
0 0
c n
y
0 q
n
0
0 0 0 q
n
_
_
, (3.5.34)
v
n
=
_
_
q
n
0 0 0
0 q
n
c 0 0
0 0 q
n
+c 0
0 0 0 q
n
_
_
, (3.5.35)
R
v
n
=
_
_
0 1 1 0
0 n
x
n
x
x
0 n
y
n
y
y
1 0 0 0
_
_
, (3.5.36)
L
w
n
=
_
_
0 0 0 1
1
2
1
2
n
x
1
2
n
y
0
1
2
1
2
n
x
1
2
n
y
0
0
x
y
0
_
_
. (3.5.37)
It is interesting to note that the change in the symmetric variables due to the acoustic and shear waves are given by
_
_
p
c
u
v
p a
2
_
1
n
x
n
y
0
_
_
,
_
_
1
n
x
n
y
0
_
_
,
_
_
0
y
0
_
_
, (3.5.38)
which shows that the entropy is constant through the acoustic and shear waves. Of course, the entropy may increase
across a shock wave (a strong acoustic wave); this change is given by the Rankine-Hugoniot relation.
Streamline/Natural Coordinates
I like the streamline coordinates, (, ) where taken along streamlines and perpendicular to streamlines. The
symmetric form (3.5.27) becomes even simpler in this coordinate system:
V
t
+
A
v
+
B
v
= 0, (3.5.39)
where
d
V =
_
_
dp
c
dq
dq
dp c
2
d
_
_
,
A
v
=
_
_
q c 0 0
c q 0 0
0 0 q 0
0 0 0 q
_
_
,
B
v
=
_
_
0 0 c 0
0 0 0 0
c 0 0 0
0 0 0 0
_
_
, (3.5.40)
q = ucos +v sin , (3.5.41)
q
A
v
= A
v
cos +B
v
sin , (3.5.43)
B
v
= A
v
sin +B
v
cos , (3.5.44)
d
V =
V
V
dV =
_
_
1 0 0 0
0 cos sin 0
0 sin cos 0
0 0 0 1
_
_
dV, (3.5.45)
74 CHAPTER 3. EULER EQUATIONS
and is a local ow angle dened by = tan
1
_
v
u
_
. Look at
B
v
. It is extremely simple.
We have obtained this form from Equation (3.5.27) by setting x axis to be a streamline direction and y axis to be
normal to the streamline. Or more formally, we invert this change of coordinates,
= cos
x
+ sin
y
, (3.5.46)
= sin
x
+ cos
y
, (3.5.47)
to get
x
= cos
sin
, (3.5.48)
y
= sin
+ cos
, (3.5.49)
and substitute these into Equation (3.5.27).
Note that we always have q
= 0 by denition (zero velocity across streamlines), but its rate of change is not
necessarily zero along the streamlines.
3.5.4 Parameter Vector
The parameter vector is dened by
Z =
_
1
u
v
H
_
_
=
_
_
z
1
z
2
z
3
z
4
_
_
. (3.5.50)
I like it because every component of conservative variables and uxes can be expressed as a quadratic in the components
of Z [128].
U =
_
u
v
E
_
_
=
_
_
z
2
1
z
1
z
2
z
1
z
3
z
1
z
4
+
1
z
2
2
+z
2
3
2
_
_
, (3.5.51)
F =
_
_
u
u
2
+p
uv
uH
_
_
=
_
_
z
1
z
2
z
2
2
+p
z
2
z
3
z
2
z
4
_
_
, (3.5.52)
G =
_
_
v
uv
v
2
+p
vH
_
_
=
_
_
z
1
z
3
z
2
z
3
z
2
3
+p
z
3
z
4
_
_
, (3.5.53)
p =
1
_
z
1
z
4
z
2
2
+z
2
3
2
_
. (3.5.54)
We can write the conservative form of the Euler equations in terms of the parameter vector variables:
U
t
+A
z
Z
x
+B
z
Z
y
= 0, (3.5.55)
3.5. 2D EULER EQUATIONS 75
A
z
=
F
Z
=
_
_
z
2
z
1
0 0
1
z
4
+ 1
z
2
1
z
3
1
z
1
0 z
3
z
2
0
0 z
4
0 z
2
_
_
, (3.5.56)
B
z
=
G
Z
=
_
_
z
3
0 z
1
0
0 z
3
z
2
0
1
z
4
1
z
2
+ 1
z
3
1
z
1
0 0 z
4
z
3
_
_
. (3.5.57)
Note that the Jacobian matrices are linear in the parameter vector variables, which can be useful to linearize the Euler
equations over a computational cell [31].
Perform change of variables as follows:
dZ =
Z
U
dU, dU =
U
Z
dZ, (3.5.58)
where
U
Z
=
_
_
2z
1
0 0 0
z
2
z
1
0 0
z
3
0 z
1
0
z
4
z
2
1
z
3
z
1
_
, (3.5.59)
Z
U
=
1
z
1
_
_
1
2
0 0 0
1
2
z
2
z
1
1 0 0
1
2
z
3
z
1
0 1 0
1
2
z
2
2
+z
2
3
z
2
1
1
2
z
4
z
1
( 1)
z
2
z
1
( 1)
z
3
z
1
_
,
=
1
_
1
2
0 0 0
1
2
u 1 0 0
1
2
v 0 1 0
1
2
_
u
2
+v
2
_
1
2
H ( 1)u ( 1)v
_
_
, (3.5.60)
from which we nd also
Z = 2
Z
U
U, U =
1
2
U
Z
Z. (3.5.61)
Compare these with Equation (3.5.58). It is very interesting, isnt it? Moreover, since we have
F = AU, G = BU, (3.5.62)
76 CHAPTER 3. EULER EQUATIONS
for the Euler equations (see Section 3.10), we nd from Equation (3.5.61) that
F = AU =
1
2
F
U
U
Z
Z =
1
2
A
z
Z, (3.5.63)
G = BU =
1
2
G
U
U
Z
Z =
1
2
B
z
Z. (3.5.64)
These show how we can compute the uxes by the parameter vector. I like these formulas because they would be useful
if we carry the parameter vector as working variables in a CFD code (i.e., we can directly compute the uxes without
converting the parameter vector into the primitive variables). I like them also because they show that the physical
uxes are quadratic in Z.
3.6 3D Euler Equations
3.6.1 Conservative Form
U
t
+F
x
+G
y
+H
z
= 0, (3.6.1)
where
U =
_
u
v
w
E
_
_
, F =
_
_
u
u
2
+p
uv
uw
uH
_
_
, G =
_
_
v
vu
v
2
+p
vw
vH
_
_
, H =
_
_
w
wu
wv
w
2
+p
wH
_
_
. (3.6.2)
Normal Flux
F
n
= [F, G, H] n = Fn
x
+Gn
y
+Hn
z
=
_
_
q
n
q
n
u +pn
x
q
n
v +pn
y
q
n
w +pn
z
q
n
H
_
_
, (3.6.3)
where n = [n
x
, n
y
, n
z
]
t
and
q
n
= un
x
+vn
y
+wn
z
. (3.6.4)
Jacobians
A =
_
_
0 1 0 0 0
1
2
q
2
u
2
(3 )u (1 )v (1 )w 1
uv v u 0 0
uw w 0 u 0
_
1
2
q
2
H
_
u H + (1 )u
2
(1 )uv (1 )uw u
_
_
, (3.6.5)
3.6. 3D EULER EQUATIONS 77
B =
_
_
0 0 1 0 0
vu v u 0 0
1
2
q
2
v
2
(1 )v (3 )u (1 )w 1
vw 0 w v 0
_
1
2
q
2
H
_
v (1 )uv H + (1 )v
2
(1 )vw v
_
_
, (3.6.6)
C =
_
_
0 0 0 1 0
uw w 0 u 0
vw 0 w v 0
1
2
q
2
w
2
(1 )u (1 )v (3 )w 1
_
1
2
q
2
H
_
w (1 )uw (1 )vw H + (1 )w
2
w
_
_
, (3.6.7)
where A =
F
U
, B =
G
U
, C =
H
U
, and q
2
= u
2
+v
2
+w
2
. Also, I like this one:
A
n
= An
x
+Bn
y
+Cn
z
=
_
_
0 n
x
n
y
n
z
0
K
2
q
2
n
x
uq
n
un
x
Kun
x
+q
n
un
y
Kvn
x
un
z
Kwn
x
Kn
x
K
2
q
2
n
y
vq
n
vn
x
Kun
y
vn
y
Kvn
y
+q
n
vn
z
Kwn
y
Kn
y
K
2
q
2
n
z
wq
n
wn
x
Kun
z
wn
y
Kvn
z
wn
z
Kwn
z
+q
n
Kn
z
_
K
2
q
2
H
_
q
n
Hn
x
Kuq
n
Hn
y
Kvq
n
Hn
z
Kwq
n
q
n
_
_
=
_
_
0 n
t
0
1
2
q
2
n vq
n
vn ( 1)nv +q
n
I ( 1)n
_
1
2
q
2
H
_
q
n
Hn
t
( 1)v
t
q
n
q
n
_
_
, (3.6.8)
where K = 1, v = [u, v, w]
t
, and I is the 33 identity matrix. We may take n = [1, 0, 0]
t
, [0, 1, 0]
t
, or [0, 0, 1]
t
,
to obtain A, B, or C, respectively.
Eigenstructure
A
n
= R
n
n
L
n
, (3.6.9)
78 CHAPTER 3. EULER EQUATIONS
where
n
=
_
_
q
n
c 0 0 0 0
0 q
n
0 0 0
0 0 q
n
+c 0 0
0 0 0 q
n
0
0 0 0 0 q
n
_
_
, (3.6.10)
R
n
=
_
_
1 1 1 0 0
u c n
x
u u +c n
x
x
m
x
v c n
y
v v +c n
y
y
m
y
w c n
z
w w +c n
z
z
m
z
H q
n
c q
2
/2 H +q
n
c q
q
m
_
_
, (3.6.11)
L
n
=
_
_
Kq
2
4c
2
+
q
n
2c
_
K
2c
2
u +
n
x
2c
_
_
K
2c
2
v +
n
y
2c
_
_
K
2c
2
w +
n
z
2c
_
K
2c
2
1
Kq
2
2c
2
Ku
c
2
Kv
c
2
Kw
c
2
K
c
2
Kq
2
4c
2
q
n
2c
_
K
2c
2
u
n
x
2c
_
_
K
2c
2
v
n
y
2c
_
_
K
2c
2
w
n
z
2c
_
K
2c
2
q
x
y
z
0
q
m
m
x
m
y
m
z
0
_
_
, (3.6.12)
L
n
dU =
_
_
dp c dq
n
2c
2
dp c
2
d
c
2
dp +c dq
n
2c
2
dq
dq
m
_
_
, (3.6.13)
where = [
x
,
y
,
z
]
t
, m = [m
x
, m
y
, m
z
]
t
, and n are mutually orthogonal unit vectors, and
q
= u
x
+v
y
+w
z
, q
m
= um
x
+vm
y
+wm
z
. (3.6.14)
See Section 1.3 for formulas and properties of mutually orthogonal unit vectors. Again, it is nice that the normal
Jacobian A
n
can be diagonalized as
n
= L
n
A
n
R
n
, (3.6.15)
but it is very dicult to diagonalize A, B, and C simultaneously. In three dimensions, even the steady equations
cannot be diagonalized for the same reason that it is hard to diagonalize the two-dimensional time-dependent Euler
system.
Absolute Normal Jacobian
Incidentally, I like to show that the absolute value of the normal Jacobian dened by
|A
n
| R
n
|
n
|L
n
, (3.6.16)
3.6. 3D EULER EQUATIONS 79
where
n
= diag(|
1
|, |
2
|, |
3
|, |
4
|, |
5
|) (3.6.17)
=
_
_
|q
n
c| 0 0 0 0
0 |q
n
| 0 0 0
0 0 |q
n
+c| 0 0
0 0 0 |q
n
| 0
0 0 0 0 |q
n
|
_
_
, (3.6.18)
can be expressed without ambiguous tangent vectors, and m. First, denote the k-th column of R
n
by r
k
and the
k-th row of L
n
by
t
k
, so that we can write
|A
n
| R
n
|
n
|L
n
=
5
k=1
|
k
| r
k
t
k
. (3.6.19)
(Note that is a tangent vector and
t
k
is the k-th left eigenvector.) Now, we have
r
4
t
4
=
_
_
0 0
t
0
q
0
q
2
t
0
_
_
, (3.6.20)
where 0
t
= [0, 0, 0], and also
r
5
t
5
=
_
_
0 0
t
0
q
m
m mm 0
q
2
m
q
m
m
t
0
_
_
. (3.6.21)
Then, we can eliminate the tangent vectors by adding these two and using the results from Section 1.3:
r
4
t
4
+r
5
t
5
=
_
_
0 0
t
0
q
n
n v I nn 0
q
2
n
q
2
v
t
q
n
n
t
0
_
_
, (3.6.22)
where I is the 3x3 identity matrix. Here, we notice that
2
=
4
=
5
= |q
n
| and r
2
t
2
does not depend on tangent
vectors. Then, we would want to add r
2
t
2
,
r
2
t
2
=
_
_
1
1
2
M
2
1
c
2
v
t
1
c
2
1
2
M
2
v
1
c
2
vv
1
c
2
v
q
2
2
_
1
1
2
M
2
_
1
2
M
2
v
t
1
2
M
2
_
_
, (3.6.23)
where M
2
= q
2
/c
2
, to Equation (3.6.22) to get
|
2
| r
2
t
2
+
5
k=4
|
k
| r
k
t
k
= |q
n
|
_
r
2
t
2
+r
4
t
4
+r
5
t
5
_
= |q
n
|
_
_
1
1
2
M
2
1
c
2
v
t
1
c
2
_
1 +
1
2
M
2
_
v +q
n
n
1
c
2
vv +I nn
1
c
2
v
q
2
n
q
2
2
_
1 +
1
2
M
2
_ _
1 +
1
2
M
2
_
v
t
q
n
n
t
1
2
M
2
_
_
. (3.6.24)
80 CHAPTER 3. EULER EQUATIONS
Next, consider r
1
t
1
and r
3
t
3
:
r
1
t
1
=
_
_
1
4
M
2
+
M
n
2
1
2c
2
v
n
2c
1
2c
2
(v c n)
_
1
4
M
2
+
M
n
2
_
(v c n)
_
1
2c
2
v
n
2c
_
(v c n)
1
2c
2
(H q
n
c)
_
1
4
M
2
+
M
n
2
_
(H q
n
c)
_
1
2c
2
v
n
2c
_
(H q
n
c)
1
2c
2
_
_
, (3.6.25)
r
3
t
3
=
_
_
1
4
M
2
M
n
2
1
2c
2
v +
n
2c
1
2c
2
(v +c n)
_
1
4
M
2
M
n
2
_
(v +c n)
_
1
2c
2
v +
n
2c
_
(v +c n)
1
2c
2
(H +q
n
c)
_
1
4
M
2
M
n
2
_
(H +q
n
c)
_
1
2c
2
v +
n
2c
_
(H +q
n
c)
1
2c
2
_
_
, (3.6.26)
where M
n
= q
n
/c. It may be possible to simplify the sum, |
1
| r
1
t
1
+|
3
| r
3
t
3
, but I stop here because they are already
independent of the tangent vectors. I like this very much because I can now compute the dissipation term (3.6.19),
using the normal vector n only. This is very useful because tangent vectors are not uniquely dened for a given normal
vector n, especially in three dimensions. By the way, the matrix |A
n
| is required, for example, for implementing a class
of residual-distribution schemes [15, 19, 30, 164].
I also like to show that the product of the absolute normal Jacobian and the dierential of the conservative variables,
|A
n
| dU R
n
|
n
|L
n
dU, (3.6.27)
is independent of the tangent vectors. It goes like this:
|A
n
| dU = R
n
|
n
|L
n
dU
=
5
k=1
|
k
| r
k
t
k
dU
=
3
k=1
|
k
| r
k
t
k
dU+|q
n
|
_
_
0
(
x
dq
+m
x
dq
m
)
(
y
dq
+m
y
dq
m
)
(
z
dq
+m
z
dq
m
)
(q
dq
+q
m
dq
m
)
_
_
=
3
k=1
|
k
| r
k
t
k
dU+|q
n
|
_
_
0
(du dq
n
n
x
)
(dv dq
n
n
y
)
(dw dq
n
n
z
)
(udu +vdv +wdw q
n
dq
n
)
_
_
, (3.6.28)
where the results in Section 1.3 have been used in the last step. This is very nice. I like it. Basically, the tangent
vectors should not appear at all because obviously |A
n
| depends only on the normal vector. It should always be possible
to eliminate tangent vectors, and m, from |A
n
| for any equations. By the way, the vector |A
n
| dU is often used, for
example, for implementing Roes approximate Riemann solver [128].
3.6. 3D EULER EQUATIONS 81
Expanded Form
Sometimes, it is convenient to expand the conservative form of the Euler equations in terms of the derivatives of the
primitive variables:
F
x
+G
y
+H
z
=
_
_
(u)
x
+ (v)
y
+ (w)
z
(u
2
)
x
+ (uv)
y
+ (uw)
z
+p
x
(vu)
x
+ (v
2
)
y
+ (vw)
z
+p
y
(wu)
x
+ (wv)
y
+ (w
2
)
z
+p
z
(uH)
x
+ (vH)
y
+ (wH)
x
_
_
=
_
_
(u
x
+v
y
+w
z
) +u
x
+v
y
+w
z
u{(u
x
+v
y
+w
z
) +u
x
+v
y
+w
z
} +uu
x
+vu
y
+wu
z
+p
x
v {(u
x
+v
y
+w
z
) +u
x
+v
y
+w
z
} +uv
x
+vv
y
+wv
z
+p
y
w{(u
x
+v
y
+w
z
) +u
x
+v
y
+w
z
} +uw
x
+vw
y
+ww
z
+p
z
H(u
x
+v
y
+w
z
) +u(H)
x
+v(H)
y
+w(H)
z
_
_
, (3.6.29)
where
(H)
x
=
1
p
x
+
1
2
x
(u
2
+v
2
+w
2
) +(uu
x
+vv
x
+ww
x
), (3.6.30)
(H)
y
=
1
p
y
+
1
2
y
(u
2
+v
2
+w
2
) +(uu
y
+vv
y
+ww
y
), (3.6.31)
(H)
z
=
1
p
z
+
1
2
z
(u
2
+v
2
+w
2
) +(uu
z
+vv
z
+ww
z
). (3.6.32)
I like it because this form can be useful for numerically computing a source term arising from the method of manufactured
solutions where we dene simple functions for the primitive variables and substitute them into the Euler equations to
derive a source term, so that the simple functions are the exact solutions for the Euler equations with the source term.
3.6.2 Primitive Form
W
t
+A
w
W
x
+B
w
W
y
+C
w
W
z
= 0, (3.6.33)
where
W =
_
u
v
w
p
_
_
, A
w
=
_
_
u 0 0 0
0 u 0 0 1/
0 0 u 0 0
0 0 0 u 0
0 c
2
0 0 u
_
_
, (3.6.34)
B
w
=
_
_
v 0 0 0
0 v 0 0 0
0 0 v 0 1/
0 0 0 v 0
0 0 c
2
0 v
_
_
, C
w
=
_
_
w 0 0 0
0 w 0 0 0
0 0 w 0 0
0 0 0 w 1/
0 0 0 c
2
w
_
_
. (3.6.35)
82 CHAPTER 3. EULER EQUATIONS
Eigenstructure
A
w
n
= R
w
n
n
L
w
n
, (3.6.36)
where
A
w
n
=
_
_
q
n
n
x
n
y
n
z
0
0 q
n
0 0 n
x
/
0 0 q
n
0 n
y
/
0 0 0 q
n
n
z
/
0 c
2
n
x
c
2
n
y
c
2
n
z
q
n
_
_
, (3.6.37)
n
=
_
_
q
n
c 0 0 0 0
0 q
n
0 0 0
0 0 q
n
+c 0 0
0 0 0 q
n
0
0 0 0 0 q
n
_
_
, (3.6.38)
R
w
n
=
_
_
1 1 1 0 0
n
x
c
0
n
x
c
m
x
n
y
c
0
n
y
c
m
y
n
z
c
0
n
z
c
m
z
c
2
0 c
2
0 0
_
_
, (3.6.39)
L
w
n
=
_
_
0
n
x
2c
n
y
2c
n
z
2c
1
2c
2
1 0 0 0
1
c
2
0
n
x
2c
n
y
2c
n
z
2c
1
2c
2
0
x
y
z
0
0 m
x
m
y
m
z
0
_
_
. (3.6.40)
Again, as in Section 3.4.2, there are no such things as primitive Jacobians. Note that we have two shear/vorticity
waves (fourth and fth components above) in three dimensions.
Derivatives of Conservative Fluxes
We can write the spatial part of the conservative form of the Euler equations in terms of the primitive variables:
U
t
+
F
W
W
x
+
G
W
W
y
+
H
W
W
z
= 0, (3.6.41)
3.6. 3D EULER EQUATIONS 83
F
W
=
_
_
u 0 0 0
u
2
2u 0 0 1
uv v u 0 0
uw w 0 u 0
1
2
uq
2
H +u
2
uv uw
1
u
_
_
, (3.6.42)
G
W
=
_
_
v 0 0 0
vu v u 0 0
v
2
0 2v 0 1
vw 0 w v 0
1
2
vq
2
vu H +v
2
vw
1
v
_
_
, (3.6.43)
H
W
=
_
_
w 0 0 0
wu w 0 u 0
wv 0 w v 0
w
2
0 0 2w 1
1
2
wq
2
wu wv H +w
2
1
w
_
_
. (3.6.44)
I like these derivatives of the conservative uxes because they can be useful in computing the gradient of the uxes in
terms of the gradient of the primitive variables:
F
x
=
F
W
W
x
,
F
y
=
F
W
W
y
,
F
z
=
F
W
W
z
, (3.6.45)
and similarly for G and H. Also, it is possible in the same way to compute the gradient of the normal ux:
F
n
x
=
F
n
W
W
x
,
F
n
y
=
F
n
W
W
y
,
F
n
z
=
F
n
W
W
z
, (3.6.46)
where
F
n
W
=
F
W
n
x
+
G
W
n
y
+
H
W
n
z
=
_
_
q
n
n
x
n
y
n
z
0
q
n
u (q
n
+un
x
) un
y
un
z
n
x
q
n
v vn
x
(q
n
+vn
y
) vn
z
n
y
q
n
w wn
x
wn
y
(q
n
+wn
z
) n
z
1
2
q
n
q
2
Hn
x
+q
n
u Hn
y
+q
n
v Hn
z
+q
n
w
1
q
n
_
_
. (3.6.47)
Change of Variables
dU =
U
W
dW, dW =
W
U
dU, (3.6.48)
84 CHAPTER 3. EULER EQUATIONS
where
U
W
=
_
_
1 0 0 0 0
u 0 0 0
v 0 0 0
w 0 0 0
q
2
/2 u v w 1/( 1)
_
_
, (3.6.49)
W
U
=
_
_
1 0 0 0 0
u/
1
0 0 0
v/ 0
1
0 0
w/ 0 0
1
0
( 1)q
2
/2 ( 1)u ( 1)v ( 1)w 1
_
_
, (3.6.50)
W
U
=
_
U
W
_
1
, A
w
n
=
W
U
A
n
U
W
, L
n
= L
w
n
W
U
, R
n
=
U
W
R
w
n
. (3.6.51)
3.6.3 Symmetric Form
V
t
+A
v
V
x
+B
v
V
y
+C
v
V
z
= 0, (3.6.52)
where
dV =
_
_
dp
c
du
dv
dw
dp a
2
d
_
_
, A
v
=
_
_
u c 0 0 0
c u 0 0 0
0 0 u 0 0
0 0 0 u 0
0 0 0 0 u
_
_
, (3.6.53)
B
v
=
_
_
v 0 c 0 0
0 v 0 0 0
c 0 v 0 0
0 0 0 v 0
0 0 0 0 v
_
_
, C
v
=
_
_
w 0 0 c 0
0 w 0 0 0
0 0 w 0 0
c 0 0 w 0
0 0 0 0 w
_
_
. (3.6.54)
Note that the fth equation is an advection equation for the entropy:
s
t
+us
x
+vs
y
+ws
z
= 0, (3.6.55)
where the entropy s has been scaled by ( 1)p/R. So, in all dimensions, basically the Euler equations describe
isentropic ows. It is the Rankine-Hugoniot relation (the integral form of the Euler equations) that describes non-
isentropic ows (shock waves).
Change of Variables
dU =
U
V
dV, dV =
V
U
dU, (3.6.56)
where
U
V
=
_
c
0 0 0
1
c
2
u
c
0 0
u
c
2
v
c
0 0
v
c
2
w
c
0 0
w
c
2
c
_
M
2
2
+
1
1
_
u v w
M
2
2
_
_
, (3.6.57)
3.6. 3D EULER EQUATIONS 85
V
U
=
_
U
V
_
1
=
_
_
1
2
q
2
c
( 1)
u
c
( 1)
v
c
( 1)
w
c
1
c
0 0 0
0
1
0 0
0 0
1
0
1
2
q
2
c
2
( 1)u ( 1)v ( 1)w 1
_
_
. (3.6.58)
Eigenstructure
A
v
n
= R
v
n
n
L
v
n
, (3.6.59)
where
A
v
n
=
_
_
q
n
c n
x
c n
y
c n
z
0
c n
x
q
n
0 0 0
c n
y
0 q
n
0 0
c n
z
0 0 q
n
0
0 0 0 0 q
n
_
_
, (3.6.60)
v
n
=
_
_
q
n
0 0 0 0
0 q
n
c 0 0 0
0 0 q
n
+c 0 0
0 0 0 q
n
0
0 0 0 0 q
n
_
_
, (3.6.61)
R
v
n
=
_
_
0 1 1 0 0
0 n
x
n
x
x
m
x
0 n
y
n
y
y
m
y
0 n
z
n
z
z
m
z
1 0 0 0 0
_
_
, (3.6.62)
L
v
n
=
_
_
0 0 0 0 1
1
2
1
2
n
x
1
2
n
y
1
2
n
z
0
1
2
1
2
n
x
1
2
n
y
1
2
n
z
0
0
x
y
z
0
0 m
x
m
y
m
z
0
_
_
. (3.6.63)
Again, as in two dimensions, it can be seen from the right eigenvectors that the acoustic and shear waves do not aect
the entropy.
Absolute Normal Jacobian
I like the absolute value of the normal Jacobian very very much for the symmetric form,
|A
v
n
| R
v
n
|
v
n
|L
v
n
, (3.6.64)
86 CHAPTER 3. EULER EQUATIONS
because it can be much simpler than that of the conservative form (see Section 3.6.1). If we denote the k-th column
of R
v
n
by r
k
and the k-th row of L
v
n
by
t
k
, then we can write
|A
v
n
| =
5
k=1
|
k
| r
k
t
k
. (3.6.65)
First, consider the following components:
r
1
t
1
=
_
_
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 0 1
_
_
, (3.6.66)
r
4
t
4
=
_
_
0 0
t
0
0 0
0 0
t
0
_
_
, (3.6.67)
r
5
t
5
=
_
_
0 0
t
0
0 mm 0
0 0
t
0
_
_
, (3.6.68)
where 0
t
= [0, 0, 0], and so
|
1
| r
1
t
1
+|
4
| r
4
t
4
+|
5
| r
5
t
5
= |q
n
|
_
_
0 0
t
0
0 +mm 0
0 0
t
1
_
_
= |q
n
|
_
_
0 0
t
0
0 I nn 0
0 0
t
1
_
_
, (3.6.69)
where I is the 3x3 identity matrix. Second, consider the other two components:
r
2
t
2
=
1
2
_
_
1 n
t
0
n nn 0
0 0
t
0
_
_
, (3.6.70)
r
3
t
3
=
1
2
_
_
1 n
t
0
n nn 0
0 0
t
0
_
_
, (3.6.71)
and so,
|
2
| r
2
t
2
+|
3
| r
3
t
3
=
_
_
||
|| n
t
0
|| n || nn 0
0 0
t
0
_
_
, (3.6.72)
where
|| =
|
3
| +|
2
|
2
=
|q
n
+c| +|q
n
c|
2
, (3.6.73)
|| =
|
3
| |
2
|
2
=
|q
n
+c| |q
n
c|
2
. (3.6.74)
3.6. 3D EULER EQUATIONS 87
Finally, we obtain
|A
v
n
| =
5
k=1
|
k
| r
k
t
k
=
_
_
||
|| n
t
0
|| n |q
n
| I + (|| |q
n
|) nn 0
0 0
t
|q
n
|
_
_
. (3.6.75)
Comparing this with |A
n
| in Section 3.6.1, we immediately see that this is so much simpler. I really like it. Note also
that
|A
n
| = R
n
|
n
|L
n
=
U
V
R
v
n
|
n
|L
v
n
V
U
=
U
V
|A
v
n
|
V
U
, (3.6.76)
which is possible because the conservative form and the symmetric form are related through a similarity transformation.
Hence, the absolute Jacobian |A
n
| can be calculated via a much simpler matrix (3.6.75). This is very nice.
Streamline Coordinates
Take x-axis as a streamline direction denoted by , and y-axis and z-axis as mutually orthogonal normal directions
denoted by
and
m
, then I get the following.
V
t
+
A
v
+
B
v
+
C
v
V
m
= 0, (3.6.77)
where
d
V =
_
_
dp
c
dq
dq
dq
m
dp a
2
d
_
_
,
A
v
=
_
_
q c 0 0 0
c q 0 0 0
0 0 q 0 0
0 0 0 q 0
0 0 0 0 q
_
_
, (3.6.78)
B
v
=
_
_
0 0 c 0 0
0 0 0 0 0
c 0 0 0 0
0 0 0 0 0
0 0 0 0 0
_
_
,
C
v
=
_
_
0 0 0 c 0
0 0 0 0 0
0 0 0 0 0
c 0 0 0 0
0 0 0 0 0
_
_
, (3.6.79)
and we have set the unit vector in direction to be n = u/q, and the unit vectors in the other two directions,
and
m
, to be and m, and dened the followings:
q = un
x
+vn
y
+wn
z
, (3.6.80)
q
= u
x
+v
y
+w
z
, (3.6.81)
q
m
= um
x
+vm
y
+wm
z
, (3.6.82)
A
v
= A
v
n
x
+B
v
n
y
+C
v
n
z
, (3.6.83)
B
v
= A
v
x
+B
v
y
+C
v
z
, (3.6.84)
C
v
= A
v
m
x
+B
v
m
y
+C
v
m
z
, (3.6.85)
d
V =
V
V
dV =
_
_
1 0 0 0 0
0 n
x
n
y
n
z
0
0
x
y
z
0
0 m
x
m
y
m
z
0
0 0 0 0 1
_
_
dV. (3.6.86)
88 CHAPTER 3. EULER EQUATIONS
3.6.4 Parameter Vector
The parameter vector is dened by
Z =
_
1
u
v
w
H
_
_
=
_
_
z
1
z
2
z
3
z
4
z
5
_
_
. (3.6.87)
I like it as in two dimensions because again every component of conservative variables and uxes can be expressed as
a quadratic in the components of Z [128].
U =
_
u
v
w
E
_
_
=
_
_
z
2
1
z
1
z
2
z
1
z
3
z
1
z
4
z
1
z
5
+
1
z
2
2
+z
2
3
+z
2
4
2
_
_
, (3.6.88)
F =
_
_
u
u
2
+p
uv
uw
uH
_
_
=
_
_
z
1
z
2
z
2
2
+p
z
2
z
3
z
2
z
4
z
2
z
5
_
_
, (3.6.89)
G =
_
_
v
vu
v
2
+p
vw
vH
_
_
=
_
_
z
1
z
3
z
3
z
2
z
2
3
+p
z
3
z
4
z
3
z
5
_
_
, (3.6.90)
H =
_
_
w
wu
wv
w
2
+p
wH
_
_
=
_
_
z
1
z
4
z
4
z
2
z
4
z
3
z
2
4
+p
z
4
z
5
_
_
, (3.6.91)
p =
1
_
z
1
z
5
z
2
2
+z
2
3
+z
2
4
2
_
. (3.6.92)
We can write the conservative form of the Euler equations in terms of the parameter vector variables:
U
t
+A
z
Z
x
+B
z
Z
y
+C
z
Z
z
= 0, (3.6.93)
3.6. 3D EULER EQUATIONS 89
A
z
=
F
Z
=
_
_
z
2
z
1
0 0 0
1
z
5
+ 1
z
2
1
z
3
1
z
4
1
z
1
0 z
3
z
2
0 0
0 z
4
0 z
2
0
0 z
5
0 0 z
2
_
_
, (3.6.94)
B
z
=
G
Z
=
_
_
z
3
0 z
1
0 0
0 z
3
z
2
0 0
1
z
5
1
z
2
+ 1
z
3
1
z
4
1
z
1
0 0 z
4
z
3
0
0 0 z
5
0 z
3
_
_
, (3.6.95)
C
z
=
H
Z
=
_
_
z
4
0 0 z
1
0
0 z
4
0 z
2
0
0 0 z
4
z
3
0
1
z
5
1
z
2
1
z
3
+ 1
z
4
1
z
1
0 0 0 z
5
z
4
_
_
. (3.6.96)
Again, the Jacobian matrices are linear in the parameter vector variables, and this can be useful to linearize the Euler
equations over a computational cell [31]. Of course, I like to perform change of variables:
dZ =
Z
U
dU, dU =
U
Z
dZ, (3.6.97)
where
U
Z
=
_
_
2z
1
0 0 0 0
z
2
z
1
0 0 0
z
3
0 z
1
0 0
z
4
0 0 z
1
0
z
5
z
2
1
z
3
1
z
4
z
1
_
, (3.6.98)
Z
U
=
1
z
1
_
_
1
2
0 0 0 0
1
2
z
2
z
1
1 0 0 0
1
2
z
3
z
1
0 1 0 0
1
2
z
4
z
1
0 0 1 0
1
2
z
2
2
+z
2
3
+z
2
4
z
2
1
1
2
z
5
z
1
( 1)z
2
z
1
( 1)z
3
z
1
( 1)z
4
z
1
_
(3.6.99)
90 CHAPTER 3. EULER EQUATIONS
=
1
_
1
2
0 0 0 0
u
2
1 0 0 0
v
2
0 1 0 0
w
2
0 0 1 0
( 1)q
2
2
H
2
( 1)u ( 1)v ( 1)w
_
_
, (3.6.100)
from which we nd
Z = 2
Z
U
U, U =
1
2
U
Z
Z. (3.6.101)
As in the two-dimensional case, I nd it very interesting. Of course, since we have
F = AU, G = BU, H = CU, (3.6.102)
for the Euler equations even in three dimensions (see Section 3.10), we nd from Equation (3.6.101) as in two dimensions
that
F = AU =
1
2
F
U
U
Z
Z =
1
2
A
z
Z, (3.6.103)
G = BU =
1
2
G
U
U
Z
Z =
1
2
B
z
Z, (3.6.104)
H = CU =
1
2
H
U
U
Z
Z =
1
2
C
z
Z. (3.6.105)
Yes, we can directly compute the uxes by the parameter vector variables without converting them into the primitive
variables. I really like the fact that the physical uxes are quadratic in Z as seen in the above equations.
3.6.5 Derivatives of Eigenvalues
Sometimes, it is necessary to dierentiate a numerical ux function, for example, for constructing implicit schemes.
Then, we would need to dierentiate various quantities, e.g., the eigenvalues with respect to the conservative variables.
I have done it before, and I like it. Suppose we wish to compute q
n
/U. It can be computed as follows.
q
n
U
=
q
n
W
W
U
= [0, n
x
, n
y
, n
z
, 0]
_
_
1 0 0 0 0
0 0 0
0
1
0 0
0 0
1
0
1
2
q
2
( 1)u ( 1)v ( 1)w 1
_
_
=
_
q
n
,
n
x
,
n
y
,
n
z
, 0
_
, (3.6.106)
3.6. 3D EULER EQUATIONS 91
which is the derivative of the eigenvalue, q
n
. How about the speed of sound? In the same way, we obtain
c
U
=
c
W
W
U
=
_
c
2
, 0, 0, 0,
2c
_
_
_
1 0 0 0 0
0 0 0
0
1
0 0
0 0
1
0
1
2
q
2
( 1)u ( 1)v ( 1)w 1
_
_
=
_
c
2
_
1
( 1)
2
q
2
c
2
_
,
( 1)u
2c
,
( 1)v
2c
,
( 1)w
2c
,
( 1)
2c
_
. (3.6.107)
So, it is now easy to get the derivatives of the remaining eigenvalues, q
n
c and q
n
+c.
(q
n
c)
U
=
q
n
U
c
U
=
_
q
n
+
c
2
_
1
( 1)
2
q
2
c
2
_
n
x
+
( 1)u
2c
n
y
+
( 1)v
2c
n
z
+
( 1)w
2c
( 1)
2c
_
_
t
, (3.6.108)
(q
n
+c)
U
=
q
n
U
+
c
U
=
_
q
n
c
2
_
1
( 1)
2
q
2
c
2
_
n
x
( 1)u
2c
n
y
( 1)v
2c
n
z
( 1)w
2c
( 1)
2c
_
_
t
. (3.6.109)
92 CHAPTER 3. EULER EQUATIONS
Some numerical ux functions involve the maximum eigenvalue, |q
n
| + c. Its derivative can be obtained in the same
way. First compute
|q
n
|
U
=
|q
n
|
W
W
U
= [0, sign(q
n
) n
x
, sign(q
n
) n
y
, sign(q
n
) n
z
, 0]
_
_
1 0 0 0 0
0 0 0
0
1
0 0
0 0
1
0
1
2
q
2
( 1)u ( 1)v ( 1)w 1
_
_
=
_
|q
n
|
,
sign(q
n
) n
x
,
sign(q
n
) n
y
,
sign(q
n
) n
z
, 0
_
, (3.6.110)
and then we obtain
(|q
n
| +c)
U
=
_
|q
n
|
c
2
_
1
( 1)
2
q
2
c
2
_
sign(q
n
) n
x
( 1)u
2c
sign(q
n
) n
y
( 1)v
2c
sign(q
n
) n
z
( 1)w
2c
( 1)
2c
_
_
t
. (3.6.111)
If we wish to dierentiate Roes approximate Riemann solver [128], which is a very popular numerical ux function
for the Euler equations, we need to dierentiate the eigenvalues evaluated at the Roe-averaged quantities:
=
R
, (3.6.112)
u =
L
u
L
+
R
u
R
L
+
R
, (3.6.113)
v =
L
v
L
+
R
v
R
L
+
R
, (3.6.114)
w =
L
w
L
+
R
w
R
L
+
R
, (3.6.115)
3.6. 3D EULER EQUATIONS 93
H =
L
H
L
+
R
H
R
L
+
R
, (3.6.116)
c =
( 1)
_
H
q
2
2
_
, (3.6.117)
q
2
= u
2
+ v
2
+ w
2
, (3.6.118)
where the subscripts L and R indicate the two states that are averaged. See Ref.[128] for details and Ref.[109] for the
Roe-averaged density, . Here, Im interested to compute the derivatives of the eigenvalues, q
n
, q
n
+ c, and q
n
c,
where q
n
= un
x
+ vn
y
+ wn
z
. We can compute the derivative of q
n
as follows.
q
n
U
=
q
n
W
H
W
H
W
H
W
H
W
W
U
, (3.6.119)
where
W
H
=
_
, u, v, w,
H
_
t
, W
H
= [, u, v, w, H]
t
. (3.6.120)
Note that we are interested in q
n
/U
L
and q
n
/U
R
, but the subscripts L and R have been dropped because the
above expression is valid for both L and R. From now on, the quantities with no subscripts are understood as either
the state L or the state R. The rst derivative on the right hand side, i.e., q
n
/
W
H
, is given by
q
n
W
H
= [0, n
x
, n
y
, n
z
, 0] . (3.6.121)
The transformation matrix
W
H
/W
H
is given by
W
H
W
H
=
_
_
2
0 0 0 0
u u
2( + )
+
0 0 0
v v
2( + )
0
+
0 0
w w
2( + )
0 0
+
0
H
H
2( + )
0 0 0
+
_
_
. (3.6.122)
I like this matrix very much because it has been arranged such that it can be used to compute both
W
H
/W
H
L
and
W
H
/W
H
R
. That is,
W
H
W
H
L
=
_
_
2
L
0 0 0 0
u
L
u
2(
L
+ )
L
+
0 0 0
v
L
v
2(
L
+ )
0
L
L
+
0 0
w
L
w
2(
L
+ )
0 0
L
L
+
0
H
L
H
2(
L
+ )
0 0 0
L
L
+
_
_
, (3.6.123)
94 CHAPTER 3. EULER EQUATIONS
and
W
H
W
H
R
=
_
_
2
R
0 0 0 0
u
R
u
2(
R
+ )
R
+
0 0 0
v
R
v
2(
R
+ )
0
R
R
+
0 0
w
R
w
2(
R
+ )
0 0
R
R
+
0
H
R
H
2(
R
+ )
0 0 0
R
R
+
_
_
. (3.6.124)
It is nice. I like it. So, I continue to drop the subscripts below. Next, the matrix W
H
/W is given by
W
H
W
=
_
_
1 0 0 0 0
0 1 0 0 0
0 0 1 0 0
0 0 0 1 0
a
2
( 1)
u v w
( 1)
_
_
, (3.6.125)
which also can be used to compute both
W
H
L
/W
H
L
and
W
H
R
/W
H
R
. We are now ready to compute the derivative.
First compute
W
H
W
=
W
H
W
H
W
H
W
=
_
_
2
0 0 0 0
u u
2( + )
+
0 0 0
v v
2( + )
0
+
0 0
w w
2( + )
0 0
+
0
H
H
2( + )
0 0 0
+
_
_
_
_
1 0 0 0 0
0 1 0 0 0
0 0 1 0 0
0 0 0 1 0
a
2
( 1)
u v w
( 1)
_
_
=
_
_
2
0 0 0 0
u u
2( + )
+
0 0 0
v v
2( + )
0
+
0 0
w w
2( + )
0 0
+
0
H
H
2( + )
a
2
( 1)( + )
u
+
v
+
w
+
( 1)( + )
_
_
, (3.6.126)
3.6. 3D EULER EQUATIONS 95
and then
W
H
U
=
W
H
W
W
U
=
1
+
_
_
( + )
2
0 0 0 0
u u
2
0 0 0
v v
2
0 0 0
w w
2
0 0 0
H
H
2
a
2
1
u v w
1
_
_
_
_
1 0 0 0 0
u
0 0 0
v
0
1
0 0
w
0 0
1
0
( 1)q
2
2
(1 )u (1 )v (1 )w 1
_
_
=
1
+
_
_
( + )
2
0 0 0 0
u + u
2
1 0 0 0
v + v
2
0 1 0 0
w + w
2
0 0 1 0
H
H
2
a
2
1
+
( 2)q
2
2
(1 )u (1 )v (1 )w
_
_
, (3.6.127)
and we nally obtain
q
n
U
=
q
n
W
H
W
H
W
H
W
H
W
W
U
=
q
n
W
H
W
H
W
W
U
=
q
n
W
H
W
H
U
=
1
+
[0, n
x
, n
y
, n
z
, 0]
_
_
( + )
2
0 0 0 0
u + u
2
1 0 0 0
v + v
2
0 1 0 0
w + w
2
0 0 1 0
H
H
2
a
2
1
+
( 2)q
2
2
(1 )u (1 )v (1 )w
_
_
=
1
+
_
q
n
+ q
n
2
n
x
, n
y
, n
z
, 0
_
. (3.6.128)
In the same way, we can derive also
| q
n
|
U
=
1
+
_
sign( q
n
)(q
n
+ q
n
)
2
, sign( q
n
)n
x
, sign( q
n
)n
y
, sign( q
n
)n
z
, 0
_
. (3.6.129)
96 CHAPTER 3. EULER EQUATIONS
Of course, the derivative of the speed of sound can be obtained in the same way. First compute,
c
W
H
=
_
0,
( 1) u
2 c
,
( 1) v
2 c
,
( 1) w
2 c
,
1
2 c
_
, (3.6.130)
and then
c
U
=
c
W
H
W
H
U
=
1
+
_
_
1
2 c
_
v
2
+ v v
2
+
H
H
2
c
2
1
+
( 2)q
2
2
_
( 1)( u + ( 1)u)
2 c
( 1)( v + ( 1)v)
2 c
( 1)( w + ( 1)w)
2 c
( 1)
2 c
_
_
t
, (3.6.131)
where v = [ u, v, w]
t
and v = [u, v, w]
t
. Now we can easily compute the following derivatives.
( q
n
c)
U
=
1
+
_
q
n
+ q
n
2
1
2 c
_
v
2
+ v v
2
+
H
H
2
c
2
1
+
( 2)q
2
2
_
n
x
+
( 1)( u + ( 1)u)
2 c
n
y
+
( 1)( v + ( 1)v)
2 c
n
z
+
( 1)( w + ( 1)w)
2 c
( 1)
2 c
_
_
t
, (3.6.132)
( q
n
+ c)
U
=
1
+
_
q
n
+ q
n
2
+
1
2 c
_
v
2
+ v v
2
+
H
H
2
c
2
1
+
( 2)q
2
2
_
n
x
( 1)( u + ( 1)u)
2 c
n
y
( 1)( v + ( 1)v)
2 c
n
z
( 1)( w + ( 1)w)
2 c
( 1)
2 c
_
_
t
, (3.6.133)
(| q
n
| + c)
U
=
1
+
_
sign( q
n
)(q
n
+ q
n
)
2
+
1
2 c
_
v
2
+ v v
2
+
H
H
2
c
2
1
+
( 2)q
2
2
_
sign( q
n
)n
x
( 1)( u + ( 1)u)
2 c
sign( q
n
)n
y
( 1)( v + ( 1)v)
2 c
sign( q
n
)n
z
( 1)( w + ( 1)w)
2 c
( 1)
2 c
_
_
t
. (3.6.134)
3.7. AXISYMMETRIC EULER EQUATIONS 97
Other derivatives can be obtained in the same way. It is so easy because we have
W
H
/U. That is, we rst
dierentiate any Roe-averaged quantity with respect to
W
H
and then multiply it by the matrix
W
H
/U from the
right. I like it. I hope you like it, too. Note that we have already obtained the following derivatives from the matrix
(3.6.127):
_
_
U
u
U
v
U
w
U
H
U
_
_
=
1
+
_
_
( + )
2
0 0 0 0
u + u
2
1 0 0 0
v + v
2
0 1 0 0
w + w
2
0 0 1 0
H
H
2
a
2
1
+
( 2)q
2
2
(1 )u (1 )v (1 )w
_
_
. (3.6.135)
It is, of course, possible to make use of these derivatives for dierentiating Roe-averaged quantities. For example, we
can express c/U as a combination of u/U, v/U, w/U, and
H/U, and then evaluate each term by the
corresponding row vector in the above matrix. Finally, I say again that all the expressions above are valid for both L
and R.
_
_
U
L
u
U
L
v
U
L
w
U
L
H
U
L
_
_
=
1
L
+
_
_
(
L
+ )
2
L
0 0 0 0
u
L
+ u
2
1 0 0 0
v
L
+ v
2
0 1 0 0
w
L
+ w
2
0 0 1 0
H
L
H
2
a
2
L
1
+
( 2)q
2
L
2
(1 )u
L
(1 )v
L
(1 )w
L
_
_
, (3.6.136)
_
_
U
R
u
U
R
v
U
R
w
U
R
H
U
R
_
_
=
1
R
+
_
_
(
R
+ )
2
R
0 0 0 0
u
R
+ u
2
1 0 0 0
v
R
+ v
2
0 1 0 0
w
R
+ w
2
0 0 1 0
H
R
H
2
a
2
R
1
+
( 2)q
2
R
2
(1 )u
R
(1 )v
R
(1 )w
R
_
_
. (3.6.137)
To illustrate how these formulas are used in an implicit solver, I wrote a 2D implicit Euler solver and have made it
available at http://www.cfdbooks.com. I hope you like it.
3.7 Axisymmetric Euler Equations
I like the axisymmetric Euler equations very much. These equations formally two-dimensional equations but de-
scribe three-dimensional ows. That is very ecient. In fact, the axisymmetric Euler equations have been often
employed in CFD, e.g., Refs.[145, 173, 179].
98 CHAPTER 3. EULER EQUATIONS
Consider cylindrical coordinates in Figure 1.5.1, where the velocity is denoted by v = [u
r
, u
, u
z
]
t
. An axisymmetric
ow is dened as a ow with no variation in with the z-axis as the axis of symmetry. Hence, u
t
_
u
z
u
r
E
_
_
+
_
_
1
r
(ru
r
)
r
+
u
z
z
1
r
(ru
z
u
r
)
r
+
(u
2
z
)
z
+
p
z
1
r
(ru
2
r
)
r
+
(u
r
u
z
)
z
+
p
r
1
r
(ru
r
H)
r
+
u
z
H
z
_
_
= 0. (3.7.1)
Note that the continuity equation comes from Equation (3.3.1) with Equation (1.5.23), the momentum equations from
Equations (3.3.2) with Equation (1.5.45), and the energy equation from Equation (3.3.3) with Equation (1.5.23). The
system can be written in the conservative form (multiply the both sides by r):
t
_
_
r
ru
z
ru
r
rE
_
_
+
_
_
(ru
r
)
r
+
(ru
z
)
z
(ru
z
u
r
)
r
+
(ru
2
z
)
z
+
(rp)
z
(ru
2
r
)
r
+
(ru
r
u
z
)
z
+
(rp)
r
(ru
r
H)
r
+
(ru
z
H)
z
_
_
=
_
_
0
0
p
0
_
_
. (3.7.2)
The system is very similar to the two-dimensional Euler system. To see this more clearly, let us replace (z, r) by (x, y)
and (u
z
, u
r
) by (u, v) to get
t
_
_
y
yu
yv
yE
_
_
+
_
_
(yu)
x
+
(yv)
y
(yu
2
)
x
+
(yuv)
y
+
(yp)
x
(yuv)
x
+
(yv
2
)
y
+
(yp)
y
(yuH)
x
+
(yvH)
y
_
_
=
_
_
0
0
p
0
_
_
. (3.7.3)
In the vector form, it can be written as
(yU)
t
+ (yF)
x
+ (yG)
y
= S, (3.7.4)
where
U =
_
u
v
E
_
_
, F =
_
_
u
u
2
+p
uv
uH
_
_
, G =
_
_
v
uv
v
2
+p
vH
_
_
, S =
_
_
0
0
p
0
_
_
. (3.7.5)
It is clear now that the system diers from the two-dimensional Euler system only by the factor y multiplied by the
conservative variables and the ux vectors, and the pressure source term. In Ref.[173], the axisymmetric form and the
two-dimensional form are expressed in a single form with a logical switch, = 0 for the two-dimensional form and
= 1 for the axisymmetric form:
(
a
U)
t
+ (
a
F)
x
+ (
a
G)
y
= S, (3.7.6)
where
a
= (1 ) + y. (3.7.7)
This is nice. If you have a two-dimensional Euler code, you almost have an axisymmetric Euler code, which is exciting
because the axisymmetric code is essentially a three-dimensional CFD code. I like it very much.
3.8. ROTATIONAL INVARIANCE 99
3.8 Rotational Invariance
The Euler equations have the rotational invariance property (see Section 1.20)
F
n
= n
x
F +n
y
G+n
z
H = T
1
F(TU), (3.8.1)
where
T =
_
_
1 0 0 0 0
0 n
x
n
y
n
z
0
0
x
y
z
0
0 m
x
m
y
m
z
0
0 0 0 0 1
_
_
, (3.8.2)
T
1
= T
t
=
_
_
1 0 0 0 0
0 n
x
x
m
x
0
0 n
y
y
m
y
0
0 n
z
z
m
z
0
0 0 0 0 1
_
_
, (3.8.3)
and n, , m are three orthogonal unit vectors (see Section 1.3). So, we can do the following. First compute TU,
TU =
_
q
n
q
q
m
E
_
_
, (3.8.4)
and evaluate F(TU),
F(TU) =
_
_
q
n
q
2
n
+p
q
n
q
q
n
q
m
q
n
H
_
_
, (3.8.5)
and then rotate it back to obtain a ux vector projected in an arbitrary direction n,
F
n
= T
1
F(TU) =
_
_
q
n
q
n
u +pn
x
q
n
v +pn
y
q
n
w +pn
z
q
n
H
_
_
, (3.8.6)
where the relation (1.3.21) has been used to get u, v, and w.
So, basically, we need only F and dont really need G and H. In fact, some people take advantage of this property
when they write a nite-volume code: set up one subroutine to evaluate the nite-volume interface ux based on F
only, and then perform the rotation as described above to get a numerical ux in arbitrary directions. I like this way
of computing a numerical ux, but I also like to use the form (3.8.6) directly to implement a numerical ux since it
seems more intuitive.
3.9 Convenient Ordering
I like to order the components as follows:
U =
_
E
u
v
w
_
_
, F =
_
_
u
uH
u
2
+p
uv
uw
_
_
, G =
_
_
v
vH
vu
v
2
+p
vw
_
_
, H =
_
_
w
wH
wu
wv
w
2
+p
_
_
. (3.9.1)
100 CHAPTER 3. EULER EQUATIONS
Then, simply by ignoring the fth row and H, we obtain the two-dimensional Euler equations, and by ignoring further
the fourth row and G, we obtain the one-dimensional Euler equations. The same is true for the ux, Jacobians, and
eigenvectors:
Normal Flux
F
n
= Fn
x
+Gn
y
+Hn
z
=
_
_
q
n
q
n
H
q
n
u +pn
x
q
n
v +pn
y
q
n
w +pn
z
_
_
. (3.9.2)
Jacobians
A
n
=
_
_
0 0 n
t
_
1
2
q
2
H
_
q
n
q
n
Hn
t
( 1)v
t
q
n
( 1)
q
2
2
n vq
n
( 1)n vn ( 1)nv +q
n
I
_
_
. (3.9.3)
Note that by taking n = [1, 0, 0]
t
, [0, 1, 0]
t
, or [0, 0, 1]
t
, we obtain A , B , or C.
Right-Eigenvector Matrix
R
n
=
_
_
1 1 1 0 0
H q
n
c q
2
/2 H +q
n
c q
q
m
u c n
x
u u +c n
x
x
m
x
v c n
y
v v +c n
y
y
m
y
w c n
z
w w +c n
z
z
m
z
_
_
. (3.9.4)
This is convenient in the case that we want to perform one- or two-dimensional calculations with a 3D code (simply
limit the size of arrays). I like this technique. Its really simple and useful. The CFD code called EulFS [15] actually
takes advantage of this ordering. Of course, alternatively, if we have a 3D code, we can perform one-dimensional
calculations by using a string of 3D cells aligned in x-axis. You can also do two-dimensional calculations by using a
two-dimensional grid with a unit width added in the third direction (this creates a 3D grid of prismatic cells). But I
think that the ordering technique is still useful for other purposes also. For example, it can be used also for extended
systems such as magnetohydrodynamic equations: add magnetic eld components to the end of the hydrodynamic
state vector so that one can run purely hydrodynamic cases with an MHD code simply by limiting the size of the
solution array.
3.10 Homogeneity Property of the Euler Fluxes
If we assume that p/ is a function of e only (i.e., if p = f(e)), then the Euler uxes are homogeneous of de-
gree 1 in terms of the conservative variables. That is, for any scalar , we have
F(U) = F(U), G(U) = G(U), H(U) = H(U). (3.10.1)
Dierentiating both sides with respect to , and then setting = 1, we obtain
F
U
U = AU = F,
G
U
U = BU = G,
H
U
U = CU = H. (3.10.2)
Therefore, the conservative form,
U
t
+F
x
+G
y
+H
z
= 0, (3.10.3)
can be written not only as
U
t
+AU
x
+BU
y
+CU
z
= 0, (3.10.4)
3.11. NONDIMENSIONALIZATION 101
but also as
U
t
+ (AU)
x
+ (BU)
y
+ (CU)
z
= 0. (3.10.5)
I like this because the Jacobian matrices look like as if they are constant even though they are actually not. In fact,
this can be useful. The Steger-Warming ux vector splitting [149] takes advantage of this property.
Also, note that we have
dF = d(AU) = AdU+dAU, (3.10.6)
dF = AdU, (3.10.7)
from which we nd
dAU = 0. (3.10.8)
I think that I like it. Im not sure, but it must be useful.
3.11 Nondimensionalization
3.11.1 Nondimensionalized Euler Equations
Nondimensionalized quantities are nice. No matter how large the actual free stream speed is (which depends on units
also), we can set it equal to 1 by scaling the velocity by the free stream speed. So, a CFD code does not have to deal
with too large (or too small) numbers. To be even nicer, the Euler equations stay in exactly the same form for various
nondimensionalizations. Consider the following nondimensionalization:
r
, p
=
p
p
r
, v
=
v
V
r
, x
=
x
L
, t
=
t
t
r
, (3.11.1)
where L is a reference length, the variables with are the nondimensional variables, and those with the subscript r are
some reference values. Substituting these into the continuity equation (3.3.1), we obtain
+
t
r
V
r
L
div
) = 0, (3.11.2)
where div
+ div
) = 0. (3.11.4)
For the momentum equations (3.3.2), we nd
(
)
t
+ div
) +
p
r
r
V
2
r
grad
= 0, (3.11.5)
which suggests that we set
p
r
=
r
V
2
r
, (3.11.6)
and therefore the momentum equations are also identical to those for dimensional variables:
(
)
t
+ div
) + grad
= 0. (3.11.7)
102 CHAPTER 3. EULER EQUATIONS
For the energy equation (3.3.3), we rst nondimensionalize the internal energy by V
2
r
:
e
=
e
V
2
r
, (3.11.8)
so that we have
E
=
E
V
2
r
= e
+
v
2
2
, (3.11.9)
and also
H
=
H
V
2
r
= e
+
v
2
2
+
p
. (3.11.10)
Hence, the total specic energy and enthalpy can be calculated in the same forms for the nondimensional variables.
Furthermore, it follows that the equation of state for a calorically perfect gas (3.1.21) also holds for the nondimensional
variables:
p
= ( 1)
. (3.11.11)
Finally, with the above nondimensionalization, the energy equation (3.3.3) becomes
(
)
t
+ div
) = 0, (3.11.12)
which is again the same form as that for the dimensional variables. So, all equations are invariant for the nondimen-
sionalization. This is very nice. So, Im happy to summarize the result below.
Euler Equations for Nondimensionalized Variables
+ div
) = 0, (3.11.13)
(
)
t
+ div
) + grad
= 0, (3.11.14)
(
)
t
+ div
) = 0, (3.11.15)
where E
= e
+
v
2
2
, H
= E
+
p
, and
r
, p
=
p
r
V
2
r
, e
=
e
V
2
r
, v
=
v
V
r
, x
=
x
L
, t
=
t
L/V
r
. (3.11.16)
The reference length can be taken as anything you like. For example, if the diameter of a cylinder is chosen as the
reference length, you generate a computational grid around the cylinder with the diameter 1. If you are given a grid,
whatever having the length 1 can be thought of as the part chosen as the reference length. Then, it is now a matter
of what values to choose for the reference density and velocity. There can be various choices. I especially like those I
describe below.
3.11.2 Free Stream Values
We may use the free stream values (indicated by the subscript ) as reference, i.e., set
r
=
and V
r
= V
, which
give the following nondimensionalization:
, p
=
p
V
2
, e
=
e
V
2
, v
=
v
V
, x
=
x
L
, t
=
t
L/V
. (3.11.17)
Then, we can set the free stream values as follows:
= 1, p
=
1
M
2
, e
=
p
( 1)
=
M
2
( 1)
, v
=
v
= n
, (3.11.18)
where n
is a unit vector in the direction of the free stream. This means that the free stream speed is 1 and only
the direction, i.e., n
, needs to be specied at a free stream inow. Yes, also the free stream Mach number M
needs to be given, in order to set the free stream pressure. Note that the pressure is inversely proportional to M
2
and
therefore it will become extremely large for ows with a very small Mach number. In fact, this is not a popular choice
for compressible ow simulations (popular for incompressible ow simulations, though).
3.12. CHANGE OF VARIABLES 103
3.11.3 Speed of Sound
We may use the free stream speed of sound a
, p
=
p
a
2
, e
=
e
a
2
, v
=
v
a
, x
=
x
L
, t
=
t
L/a
. (3.11.19)
Then, we can set the free stream values as follows:
= 1, p
=
1
, e
=
1
( 1)
, v
=
v
= M
. (3.11.20)
Again, we only need to specify the Mach number and its direction at the free stream inow. I like this nondimension-
alization particularly because the velocity becomes equivalent to the Mach number at the free stream inow and also
that the free stream pressure is constant (independent of M
0
=
_
1 +
1
2
M
2
_
1
1
, p
=
p
0
a
2
0
=
p
p
0
=
1
_
1 +
1
2
M
2
_
1
, (3.11.21)
(you can derive these relations from Bernoullis equation for compressible ows in Section 3.18), and
e
=
e
a
2
0
, v
=
v
a
0
=
v
a
a
a
0
=
v
a
=
v
a
a
, x
=
x
L
, t
=
t
L/a
0
. (3.11.22)
Then, we can set the free stream values as follows:
=
_
1 +
1
2
M
2
1
1
, p
=
1
_
1 +
1
2
M
2
1
, (3.11.23)
e
=
p
( 1)
, v
=
v
= M
. (3.11.24)
Again, only the Mach number and the free stream direction need to be specied at the free stream inow. This will
be convenient when you want to compare a computed pressure at a stagnation point with a known value. See Section
7.13.7 for example.
3.12 Change of Variables
I like converting variables, from conservative to primitive, for example. In particular, I like those I describe in this
section since these are valid for any set of nondimensionalized variables discussed in Section 3.11.
3.12.1 One Dimension
Let U = [, u, E]
t
= [u
1
, u
2
, u
3
]
t
, W = [, u, p]
t
= [w
1
, w
2
, w
3
]
t
, and Z =
[1, u, H]
t
= [z
1
, z
2
, z
3
]
t
. Then we
have the following relations.
From conservative variables to others
= u
1
, (3.12.1)
u = u
2
/u
1
, (3.12.2)
p = ( 1)
_
u
3
1
2
u
2
2
u
1
_
, (3.12.3)
104 CHAPTER 3. EULER EQUATIONS
z
1
=
u
1
, (3.12.4)
z
2
= u
2
/
u
1
, (3.12.5)
z
3
=
u
1
_
u
3
u
1
1
2
_
u
2
u
1
_
2
_
, (3.12.6)
H =
u
3
u
1
1
2
_
u
2
u
1
_
2
. (3.12.7)
From primitive variables to others
= w
1
, (3.12.8)
u = w
1
w
2
, (3.12.9)
E =
w
3
1
+
w
1
2
w
2
2
, (3.12.10)
z
1
=
w
1
, (3.12.11)
z
2
=
w
1
w
2
, (3.12.12)
z
3
=
1
w
3
w
1
+
w
1
2
w
2
2
, (3.12.13)
H =
1
w
3
w
1
+
1
2
w
2
2
. (3.12.14)
From parameter vector variables to others
= z
2
1
, (3.12.15)
u = z
2
/z
1
, (3.12.16)
p =
1
_
z
1
z
3
z
2
2
2
_
, (3.12.17)
u = z
1
z
2
, (3.12.18)
E =
z
1
z
3
+
1
2
z
2
2
, (3.12.19)
H = z
3
/z
1
. (3.12.20)
3.12.2 Two Dimensions
Let U = [, u, v, E]
t
= [u
1
, u
2
, u
3
, u
4
]
t
, W = [, u, v, p]
t
= [w
1
, w
2
, w
3
, w
4
]
t
, and Z =
[1, u, v, H]
t
=
[z
1
, z
2
, z
3
, z
4
]
t
. Then we have the following relations.
From conservative variables to others
= u
1
, (3.12.21)
u = u
2
/u
1
, (3.12.22)
v = u
3
/u
1
, (3.12.23)
p = ( 1)
_
u
4
1
2
u
2
2
+u
2
3
u
1
_
, (3.12.24)
3.12. CHANGE OF VARIABLES 105
z
1
=
u
1
, (3.12.25)
z
2
= u
2
/
u
1
, (3.12.26)
z
3
= u
3
/
u
1
, (3.12.27)
z
4
=
u
1
_
u
4
u
1
1
2
u
2
2
+u
2
3
u
2
1
_
, (3.12.28)
H =
u
4
u
1
1
2
u
2
2
+u
2
3
u
2
1
. (3.12.29)
From primitive variables to others
= w
1
, (3.12.30)
u = w
1
w
2
, (3.12.31)
v = w
1
w
3
, (3.12.32)
E =
w
4
1
+
w
1
2
_
w
2
2
+w
2
3
_
, (3.12.33)
z
1
=
w
1
, (3.12.34)
z
2
=
w
1
w
2
, (3.12.35)
z
3
=
w
1
w
3
, (3.12.36)
z
4
=
1
w
4
w
1
+
w
1
2
_
w
2
2
+w
2
3
_
, (3.12.37)
H =
1
w
4
w
1
+
1
2
(w
2
2
+w
2
3
). (3.12.38)
From parameter vector variables to others
= z
2
1
, (3.12.39)
u = z
2
/z
1
, (3.12.40)
v = z
3
/z
1
, (3.12.41)
p =
1
_
z
1
z
4
z
2
2
+z
2
3
2
_
, (3.12.42)
u = z
1
z
2
, (3.12.43)
v = z
1
z
3
, (3.12.44)
E =
z
1
z
4
+
1
2
_
z
2
2
+z
2
3
_
, (3.12.45)
H = z
4
/z
1
. (3.12.46)
3.12.3 Three Dimensions
Let U = [, u, v, w, E]
t
= [u
1
, u
2
, u
3
, u
4
, u
5
]
t
, W = [, u, v, w, p]
t
= [w
1
, w
2
, w
3
, w
4
, w
5
]
t
, and Z =
[1, u, v, w, H]
t
=
[z
1
, z
2
, z
3
, z
4
, z
5
]
t
. Then we have the following relations.
From conservative variables to others
= u
1
, (3.12.47)
u = u
2
/u
1
, (3.12.48)
v = u
3
/u
1
, (3.12.49)
w = u
4
/u
1
, (3.12.50)
106 CHAPTER 3. EULER EQUATIONS
p = ( 1)
_
u
5
1
2
u
2
2
+u
2
3
+u
2
4
u
1
_
, (3.12.51)
z
1
=
u
1
, (3.12.52)
z
2
= u
2
/
u
1
, (3.12.53)
z
3
= u
3
/
u
1
, (3.12.54)
z
4
= u
4
/
u
1
, (3.12.55)
z
5
=
u
1
_
u
5
u
1
1
2
u
2
2
+u
2
3
+u
2
4
u
2
1
_
, (3.12.56)
H =
u
5
u
1
1
2
u
2
2
+u
2
3
+u
2
4
u
2
1
. (3.12.57)
From primitive variables to others
= w
1
, (3.12.58)
u = w
1
w
2
, (3.12.59)
v = w
1
w
3
, (3.12.60)
w = w
1
w
4
, (3.12.61)
E =
w
5
1
+
w
1
2
_
w
2
2
+w
2
3
+w
2
4
_
, (3.12.62)
z
1
=
w
1
, (3.12.63)
z
2
=
w
1
w
2
, (3.12.64)
z
3
=
w
1
w
3
, (3.12.65)
z
4
=
w
1
w
4
, (3.12.66)
z
5
=
1
w
5
w
1
+
w
1
2
_
w
2
2
+w
2
3
+w
2
4
_
, (3.12.67)
H =
1
w
5
w
1
+
1
2
(w
2
2
+w
2
3
+w
2
4
). (3.12.68)
From parameter vector variables to others
= z
2
1
, (3.12.69)
u = z
2
/z
1
, (3.12.70)
v = z
3
/z
1
, (3.12.71)
w = z
4
/z
1
, (3.12.72)
p =
1
_
z
1
z
5
z
2
2
+z
2
3
+z
2
4
2
_
, (3.12.73)
u = z
1
z
2
, (3.12.74)
v = z
1
z
3
, (3.12.75)
w = z
1
z
4
, (3.12.76)
E =
z
1
z
5
+
1
2
_
z
2
2
+z
2
3
+z
2
4
_
, (3.12.77)
H = z
5
/z
1
. (3.12.78)
3.13. INCOMPRESSIBLE/PSEUDO-COMPRESSIBLE EULER EQUATIONS 107
3.13 Incompressible/Pseudo-Compressible Euler Equations
Incompressible uids are dened as uids with negligible density variation (e.g., water). This is equivalent to negligible
Mach number, and in fact, a ow with a Mach number below 0.3 may be considered as incompressible. If we simplify
the Euler equations by using the constant density assumption, = constant, the results will be called the incompressible
Euler equations.
Conservative Form of the Incompressible Euler Equations:
div (v) = 0, (3.13.1)
t
(v) + div(vv) + grad p = 0. (3.13.2)
Primitive Form of the Incompressible Euler Equations:
div v = 0, (3.13.3)
Dv
Dt
+ grad p = 0. (3.13.4)
I like these incompressible Euler equations because they are closed without the energy equation (which may be solved
a posteriori) and also because there is no time derivative in the continuity equation even for unsteady ows (because
the density is constant). But in fact the latter can be a disadvantage: a simple time-marching method is not applicable
because the whole system cannot be written as a time-dependent system. Now, Chorin proposed a pseudo-compressible
method [21], in which a constant articial speed of sound a
= p/a
2
, (3.13.5)
and a time derivative is brought back to the continuity equation in the form,
+ div (a
2
v) = 0, (3.13.7)
where t
is an articial time and P = p/ which is called the kinematic pressure. This helps the incompressible Euler
equations recover the hyperbolic character, so that a time-marching method is now directly applicable. The resulting
equations are called the pseudo-compressible Euler equations. This is very interesting. I like it. Note however that
only when
reaches a steady state in the articial time, is the original continuity equation (div v = 0) satised. That
is, the solution will be accurate only at a steady state. The value of the articial speed of sound a
depends on the
problem and it is usually assigned a value between 0.1 and 10.0. See Refs.[21, 75, 97, 131, 132] for details. Note that
some people formulate the pseudo-compressible equation with a slightly dierent notation for the articial parameter,
e.g., with = a
2
[97, 131, 132]. I like both notations, but in this book I will use a
throughout.
The pseudo-compressible Euler equations:
P
t
+ div (a
2
v) = 0, (3.13.8)
v
t
_
P
u
v
w
_
_
, F =
_
_
a
2
u
u
2
+P
uv
uw
_
_
, G =
_
_
a
2
v
vu
v
2
+P
vw
_
_
, H =
_
_
a
2
w
wu
wv
w
2
+P
_
_
. (3.13.11)
108 CHAPTER 3. EULER EQUATIONS
The normal ux:
F
n
= Fn
x
+Gn
y
+Hn
z
=
_
_
a
2
q
n
q
n
u +Pn
x
q
n
v +Pn
y
q
n
w +Pn
z
_
_
. (3.13.12)
The Jacobians:
A =
F
U
=
_
_
0 a
2
0 0
1 2u 0 0
0 v u 0
0 w 0 u
_
_
, (3.13.13)
B =
G
U
=
_
_
0 0 a
2
0
0 v u 0
1 0 2v 0
0 0 w v
_
_
, (3.13.14)
C =
H
U
=
_
_
0 0 0 a
2
0 w 0 u
0 0 w v
1 0 0 2w
_
_
, (3.13.15)
A
n
= An
x
+Bn
y
+Cn
z
=
_
_
0 a
2
n
x
a
2
n
y
a
2
n
z
n
x
q
n
+un
x
un
y
un
z
n
y
vn
x
q
n
+vn
y
vn
z
n
y
wn
x
wn
y
q
n
+wn
z
_
_
. (3.13.16)
The eigenstructure of the Jacobian matrix A
n
:
A
n
= R
n
n
L
n
, (3.13.17)
where
n
=
_
_
q
n
c 0 0 0
0 q
n
0 0
0 0 q
n
0
0 0 0 q
n
+c
_
_
, (3.13.18)
c =
q
2
n
+a
2
, (3.13.19)
R
n
=
_
_
(q
n
+c)c 0 0 (q
n
c)c
u (q
n
+c)n
x
x
m
x
u (q
n
c)n
x
v (q
n
+c)n
y
y
m
y
v (q
n
c)n
y
w (q
n
+c)n
z
z
m
z
w (q
n
c)n
z
_
_
, (3.13.20)
L
n
=
_
_
1
2c
2
(q
n
c)n
x
2c
2
(q
n
c)n
y
2c
2
(q
n
c)n
z
2c
2
c
2
x
q
n
q
c
2
n
x
y
q
n
q
c
2
n
y
z
q
n
q
c
2
n
z
q
m
c
2
m
x
q
n
q
m
c
2
n
x
m
y
q
n
q
m
c
2
n
y
m
z
q
n
q
m
c
2
n
z
1
2c
2
(q
n
+c)n
x
2c
2
(q
n
+c)n
y
2c
2
(q
n
+c)n
z
2c
2
_
_
, (3.13.21)
3.14. HOMENTROPIC/ISOTHERMAL EULER EQUATIONS 109
= [
x
,
y
,
z
]
t
, m = [m
x
, m
y
, m
z
]
t
, n are mutually orthogonal unit vectors, and
q
= u
x
+v
y
+w
z
, q
m
= um
x
+vm
y
+wm
z
. (3.13.22)
The characteristic variables (amplitudes of the waves) are given by
L
n
dU =
_
_
dP + (q
n
c)dq
n
2c
2
dq
(dP +q
n
dq
n
)
c
2
dq
m
q
m
(dP +q
n
dq
n
)
c
2
dP + (q
n
+c)dq
n
2c
2
_
_
. (3.13.23)
The product of the absolute normal Jacobian and the dierential of the variables,
|A
n
| dU R
n
|
n
|L
n
dU, (3.13.24)
which can be used to construct the upwind ux, is, of course, independent of the tangent vectors as it should:
|A
n
| dU = R
n
|
n
|L
n
dU
=
4
k=1
|
k
| r
k
t
k
dU
= |
1
| r
1
t
1
dU+|
4
| r
4
t
4
dU+|q
n
|
_
_
0
(
x
dq
+m
x
dq
m
)
(q
x
+q
m
m
x
)(dP +q
n
dq
n
)
c
2
(
y
dq
+m
y
dq
m
)
(q
y
+q
m
m
y
)(dP +q
n
dq
n
)
c
2
(
z
dq
+m
z
dq
m
)
(q
z
+q
m
m
z
)(dP +q
n
dq
n
)
c
2
_
_
= |
1
| r
1
t
1
dU+|
4
| r
4
t
4
dU+|q
n
|
_
_
0
(du dq
n
n
x
)
(u q
n
n
x
)(dP +q
n
dq
n
)
c
2
(dv dq
n
n
y
)
(v q
n
n
y
)(dP +q
n
dq
n
)
c
2
(dw dq
n
n
z
)
(w q
n
n
z
)(dP +q
n
dq
n
)
c
2
_
_
, (3.13.25)
where the k-th column of R
n
is denoted by r
k
, the k-th row of L
n
is denoted by
t
k
, and the identities in Section 1.3
have been used in the last step.
Note that this system is in conservative form and it is equivalent to the primitive form only in the steady state since
the continuity equation will be satised only in the steady state. Note also that the time derivative in the momentum
equation is actually real but has been written as articial in the above. This is because it is meaningless to be real if
the momentum equation is solved simultaneously with the pseudo-compressible continuity equation.
3.14 Homentropic/Isothermal Euler Equations
As we have seen in the previous sections, in all dimensions, the Euler equations involve a scalar advection equa-
tion for the entropy s:
s
t
+us
x
+vs
y
+ws
z
= 0, (3.14.1)
110 CHAPTER 3. EULER EQUATIONS
where (u, v, w) is a local ow velocity. This means that the entropy is constant along a streamline. But at the same
time, it means also that the entropy is not necessarily uniform for the entire ow eld and can dier from one streamline
to another. Such a ow is called an isentropic ow. So, the Euler equations basically describe isentropic ows, with
a capability of accounting for non-isentropic features such as shocks by the Rankine-Hugoniot relation (the integral
form). A typical example of the isentropic ow is a ow behind a curved shock where the entropy is constant along
each streamline, but not uniform because its increase across the shock depends on the shock intensity which varies in
space. On the other hand, it would be very nice if the entropy were constant everywhere. It would be the case, for
example, when the entropy is initially constant everywhere and no shock waves are created anywhere. Such a ow is
called a homentropic ow (a special case of an isentropic ow). What is nice about a homentropic ow is that the
Euler equations close without the energy equation. For homentropic ows, the adiabatic relation,
p/
= K, (3.14.2)
is valid everywhere, i.e., K is a global constant (not only along a streamline), say K = p
1
+
1
2
v
2
, (3.14.3)
that the total energy can be explicitly computed everywhere once the density and the velocity are determined by the
continuity and momentum equations, and therefore the energy equation is no longer required. This implies also that
more generally, if the pressure is a function of the density only, i.e., p = p() (such a ow or a process is called
barotropic [18]), the Euler equations will close without the energy equation. Then, the speed of sound c is given by
c =
(), (3.14.4)
which becomes, in homentropic ows,
c =
K
1
. (3.14.5)
On the other hand, a ow with = 1 is called an isothermal ow. Here, by Equation (3.1.11), this is equivalent
to a gas with innite degrees of freedom. Also, from Equations (3.1.9) and (3.1.10), we see that the specic heats
become innite. This means that it is extremely dicult to change the temperature of the gas, thus leading to the term
isothermal. I like isothermal ows because in this case also the Euler equations close without the energy equation. In
particular, the speed of sound is given by
c
2
= RT = p/, (3.14.6)
which is now a constant. Isothermal ows are mysterious because = 1 is articial. There are no such gases in reality.
In either homentropic or isothermal ow, the Euler equations close without the energy equation, and the pressure
and the speed of sound are given by
_
_
_
p = K
, c =
K
1
, for homentropic ows,
p = c
2
, c =
_
u +c
2c
1
2c
u c
2c
1
2c
_
_, LdU =
_
_
1
2
d
2c
du
1
2
d +
2c
du
_
_. (3.14.13)
Of course, this system is hyperbolic only if c is real, which is true if p
u
v
_
_
, F =
_
_
u
p +u
2
uv
_
_
, G =
_
_
v
uv
p +v
2
_
_
. (3.14.15)
Jacobians
A =
F
U
=
_
_
0 1 0
c
2
u
2
2u 0
uv v u
_
_
, B =
G
U
=
_
_
0 0 1
uv v u
c
2
v
2
0 2v
_
_
, (3.14.16)
A
n
= An
x
+Bn
y
=
_
_
0 n
x
n
y
c
2
n
x
uq
n
q
n
+un
x
un
y
c
2
n
y
vq
n
vn
x
q
n
+vn
y
_
_
, (3.14.17)
where
q
n
= un
x
+vn
y
. (3.14.18)
Eigenstructure
A
n
= R
n
n
L
n
, (3.14.19)
where
n
=
_
_
q
n
c 0 0
0 q
n
0
0 0 q
n
+c
_
_
, (3.14.20)
R
n
=
_
_
1 0 1
u c n
x
n
y
u +c n
x
v c n
y
n
x
v +c n
y
_
_
, L
n
=
_
_
c +q
n
2c
n
x
2c
n
y
2c
q
n
y
n
x
c q
n
2c
n
x
2c
n
y
2c
_
_
, (3.14.21)
and
q
= un
y
+vn
x
. (3.14.22)
112 CHAPTER 3. EULER EQUATIONS
3.14.3 3D Homentropic/Isothermal Euler Equations
U
t
+F
x
+G
y
+H
z
= 0, (3.14.23)
where
U =
_
u
v
w
_
_
, F =
_
_
u
p +u
2
uv
uw
_
_
, G =
_
_
v
vu
p +v
2
vw
_
_
, H =
_
_
w
wu
wv
p +w
2
_
_
. (3.14.24)
Jacobians
A =
F
U
=
_
_
0 1 0 0
c
2
u
2
2u 0 0
uv v u 0
uw w 0 u
_
_
, (3.14.25)
B =
G
U
=
_
_
0 0 1 0
vu v u 0
c
2
v
2
0 2v 0
vw 0 w v
_
_
, (3.14.26)
C =
H
U
=
_
_
0 0 0 1
wu w 0 u
wv 0 w v
c
2
w
2
0 0 2w
_
_
, (3.14.27)
A
n
= An
x
+Bn
y
+Cn
z
=
_
_
0 n
x
n
y
n
z
c
2
n
x
uq
n
un
x
+q
n
un
y
un
z
c
2
n
y
vq
n
vn
x
vn
y
+q
n
vn
z
c
2
n
z
wq
n
wn
x
wn
y
wn
z
+q
n
_
_
. (3.14.28)
Eigenstructure
A
n
= R
n
n
L
n
, (3.14.29)
where
n
=
_
_
q
n
c 0 0 0
0 q
n
0 0
0 0 q
n
0
0 0 0 q
n
+c
_
_
, (3.14.30)
R
n
=
_
_
1 0 0 1
u c n
x
x
m
x
u +c n
x
v c n
y
y
m
y
v +c n
x
w c n
z
z
m
z
w +c n
y
_
_
, L
n
=
_
_
c +q
n
2c
n
x
2c
n
y
2c
n
z
2c
q
x
y
z
q
m
m
x
m
y
m
z
c q
n
2c
n
x
2c
n
y
2c
n
z
2c
_
_
, (3.14.31)
= [
x
,
y
,
z
]
t
, m = [m
x
, m
y
, m
z
]
t
, n are mutually orthogonal unit vectors, and
q
= u
x
+v
y
+w
z
, q
m
= um
x
+vm
y
+wm
z
. (3.14.32)
3.15. LINEAR ACOUSTICS EQUATIONS (LINEARIZED EULER EQUATIONS) 113
3.15 Linear Acoustics Equations (Linearized Euler Equations)
Decompose variables as follows:
=
0
+
, (3.15.1)
p = p
0
+p
, (3.15.2)
(u, v, w) = (u
0
, v
0
, w
0
) + (u
, v
, w
), (3.15.3)
where the variables with the subscript 0 are uniform stream values and those with a prime are small perturbations.
Substitute these into the Euler equations, and make approximations (ignoring products of perturbations), then we
obtain the so-called linear acoustics equations.
Flows with small perturbations can be considered as homentropic (because shock waves will not appear). Then,
the equation of state can be written as
p = P(), (3.15.4)
e.g., p = K
=
_
dP
d
_
=0
= c
2
0
, (3.15.5)
where c
0
is the speed of sound which is a constant. This will close the linear acoustics equations. Note that
c
0
=
_
dP
d
_
=0
, (3.15.6)
and therefore
dP
d
must be positive. Dont worry. It is positive in homentropic ows:
c
0
=
K
1
0
=
_
p
0
0
_
1
0
=
p
0
0
> 0, (3.15.7)
for physically realistic pressure and density.
So, basically, the linear acoustics equations are valid only for ows with small perturbations, such as a sound wave
propagation or a ow over a slender body. The governing equations are now linear, and therefore nonlinear waves such
as shocks or expansions all reduce to linear waves. Also, simple wave solutions can be superposed to produce more
general solutions. This can be very useful. I like it.
3.15.1 1D Linear Acoustics Equations
W
t
+A
w
x
= 0, (3.15.8)
where
W
=
_
p
_
, A
w
=
_
u
0
0
c
2
0
1/
0
u
0
_
. (3.15.9)
Eigenstructure
A
w
= R
w
L
w
, (3.15.10)
where
=
_
u
0
c 0
0 u
0
+c
_
, R
w
=
_
_
0
c
2
0
0
c
2
0
c
0
c
0
_
_
, L
w
=
_
_
1
2
0
c
2
0
1
2c
0
1
2
0
c
2
0
1
2c
0
_
_
. (3.15.11)
114 CHAPTER 3. EULER EQUATIONS
3.15.2 2D Linear Acoustics Equations
W
t
+A
w
x
+B
w
y
= 0, (3.15.12)
where
W
=
_
_
p
_
_
, A
w
=
_
_
u
0
0
c
2
0
0
1/
0
u
0
0
0 0 u
0
_
_
, B
w
=
_
_
v
0
0
0
c
2
0
0 v
0
0
1/
0
0 v
0
_
_
. (3.15.13)
Of course, these matrices are not Jacobians (see Section 3.4.2).
Eigenstructure
A
w
n
= R
w
n
n
L
w
n
, (3.15.14)
where
A
w
n
=
_
_
q
n0
0
c
2
0
n
x
0
c
2
0
n
y
n
x
/
0
q
n0
0
n
y
/
0
0 q
n0
_
_
,
w
n
=
_
_
q
n0
c 0 0
0 q
n0
0
0 0 q
n0
+c
_
_
, (3.15.15)
R
w
n
=
_
0
c
2
0
0
0
c
2
0
c
0
n
x
x
c
0
n
x
c
0
n
y
y
c
0
n
y
_
_
, L
w
n
=
_
_
1
2
0
c
2
0
n
x
2c
0
n
y
2c
0
0
x
y
1
2
0
c
2
0
n
x
2c
0
n
y
2c
0
_
_
, (3.15.16)
and (
x
,
y
) = (n
y
, n
x
).
3.15.3 3D Linear Acoustics Equations
W
t
+A
w
x
+B
w
y
+C
w
z
= 0, (3.15.17)
where
W
=
_
_
p
_
, A
w
=
_
_
u
0
0
c
2
0
0 0
1/
0
u
0
0 0
0 0 u
0
0
0 0 0 u
0
_
_
, (3.15.18)
B
w
=
_
_
v
0
0
0
c
2
0
0
0 v
0
0 0
1/
0
0 v
0
0
0 0 0 v
0
_
_
, C
w
=
_
_
w
0
0 0
0
c
2
0
0 w
0
0 0
0 0 w
0
0
1/
0
0 0 w
0
_
_
. (3.15.19)
Eigenstructure
A
w
n
= R
w
n
n
L
w
n
, (3.15.20)
3.16. QUASI-1D EULER EQUATIONS 115
where
A
w
n
=
_
_
q
n0
0
c
2
0
n
x
0
c
2
0
n
y
0
c
2
0
n
z
n
x
/
0
q
n0
0 0
n
y
/
0
0 q
n0
0
n
z
/
0
0 0 q
n0
_
_
, (3.15.21)
n
=
_
_
q
n0
c 0 0 0
0 q
n0
0 0
0 0 q
n0
0
0 0 0 q
n0
+c
_
_
, (3.15.22)
R
w
n
=
_
0
c
2
0
0 0
0
c
2
0
c
0
n
x
x
m
x
c
0
n
x
c
0
n
y
y
m
y
c
0
n
y
c
0
n
z
z
m
z
c
0
n
z
_
_
, (3.15.23)
L
w
n
=
_
_
1
2
0
c
2
0
n
x
2c
0
n
y
2c
0
n
z
2c
0
0
x
y
z
0 m
x
m
y
m
z
1
2
0
c
2
0
n
x
2c
0
n
y
2c
0
n
z
2c
0
_
_
, (3.15.24)
= [
x
,
y
,
z
]
t
, m = [m
x
, m
y
, m
z
]
t
, n are mutually orthogonal unit vectors, and
q
= u
x
+v
y
+w
z
, q
m
= um
x
+vm
y
+wm
z
. (3.15.25)
3.16 Quasi-1D Euler Equations
The quasi-1D Euler system is a model for a nozzle ow. The nozzle itself is two-dimensional, i.e., its section area A(x)
varies along the axial direction x, but ow variables are assumed to depend only on x. Here is the quasi-1D Euler
system:
U
t
+
F
x
= S, (3.16.1)
where
U =
_
_
A
uA
EA
_
_
, F =
_
_
uA
u
2
A
u(E +p/)A
_
_
, S =
_
_
0
A
p
x
0
_
_
, (3.16.2)
and E = e +u
2
/2. I like this system because it can be written in the following form (A(x) only in the source term),
U =
_
_
u
E
_
_
, F =
_
_
u
u
2
+p
u(E +p/)
_
_
, S =
1
A
DA
Dt
_
_
u
E +p
_
_
, (3.16.3)
and also in the following form (no pressure gradient in the source term),
U =
_
_
A
uA
EA
_
_
, F =
_
_
uA
(u
2
+p)A
u(E +p/)A
_
_
, S =
_
_
0
p
A
x
0
_
_
. (3.16.4)
116 CHAPTER 3. EULER EQUATIONS
See Ref.[57] for derivation, and Ref.[159] for the case that A changes in time. A really nice thing about the quasi-1D
Euler system is that if I set A , then it is in the same form as the 1D Euler system (except for the source term),
and therefore the eigenstructure is the same as the 1D Euler system (which we already know). This is very nice.
On the other hand, if I set A(x) = 2x, then I obtain the Euler equations for cylindrically symmetric wave motion,
i.e., the Euler equations in cylindrical coordinates (r, , z) with no variations in the z direction. For example, I obtain
from Equation (3.16.3),
U
t
+
F
r
= S, (3.16.5)
where
U =
_
_
u
E
_
_
, F =
_
_
u
u
2
+p
u(E +p/)
_
_
, S =
u
r
_
_
u
E +p
_
_
, (3.16.6)
and u is now the velocity in the radial direction. Note also that I have replaced x by r to emphasize the cylindrical
symmetry. Similarly, if I set A(x) = 4x
2
, then I obtain the Euler equations for spherically symmetric wave motion:
U =
_
_
u
E
_
_
, F =
_
_
u
u
2
+p
u(E +p/)
_
_
, S =
2u
r
_
_
u
E +p
_
_
. (3.16.7)
This is very interesting, isnt it?
3.17 2D Steady Euler Equations
The 2D steady Euler equations are obtained by removing the time derivative from the 2D Euler equations (3.5.16),
A
w
W
x
+B
w
W
y
= 0, (3.17.1)
where W = [, u, v, p]
t
. Since A
w
is invertible (for u
2
= c
2
), we can write this as
W
x
+ (A
w
)
1
B
w
W
y
= 0, (3.17.2)
where
(A
w
)
1
B
w
=
1
u
2
c
2
_
_
v
u
(u
2
c
2
) v u
v
u
0 uv c
2
0 0
v
u
(u
2
c
2
)
u
2
c
2
u
0 vc
2
uc
2
uv
_
_
. (3.17.3)
Therefore, x-axis may be time-like (depending on ow directions). The eigenvalues and eigenvectors of (A
w
)
1
B
w
are
given by
1,2
=
v
u
, (3.17.4)
3
=
uv c
q
2
c
2
u
2
c
2
=
v u
u +v
, (3.17.5)
4
=
uv +c
q
2
c
2
u
2
c
2
=
u +v
u v
, (3.17.6)
3.17. 2D STEADY EULER EQUATIONS 117
R
w
=
_
_
1 0
u +v
u v
0
u
q
2
v u
M
2
(u +v)
v +u
M
2
(u v)
0
v
q
2
1
M
2
1
M
2
0 0
c
2
u +v
c
2
u v
_
_
, (3.17.7)
L
w
=
_
_
1 0 0
1
c
2
0 u v 1
0
v
2c
2
(u +v)
u
2c
2
(u +v)
1
2c
2
(u +v)
0
v
2c
2
(u v)
u
2c
2
(u v)
1
2c
2
(u v)
_
_
, (3.17.8)
where =
M
2
1 =
q
2
/c
2
1 =
(u
2
+v
2
)/c
2
1 and c =
p/. Clearly, the character of the 2D Steady
Euler Equations depends on the Mach number:
Supersonic M > 1 Real eigenvalues Hyperbolic,
Subsonic M < 1 Complex eigenvalues Elliptic.
(3.17.9)
Of course, it can be mixed: hyperbolic (supersonic) in some region while elliptic (subsonic) in other regions. In
particular, I like supersonic ows because then the system (3.17.2) can be fully diagonalized, with the characteristic
variables dened by V = L
w
W:
v
1
x
+
u
v
v
1
y
= 0, (3.17.10)
v
2
x
+
u
v
v
2
y
= 0, (3.17.11)
v
3
x
+
v u
u +v
v
3
y
= 0, (3.17.12)
v
4
x
+
u +v
u v
v
4
y
= 0, (3.17.13)
where
V =
_
_
v
1
v
2
v
3
v
4
_
_
=
_
_
p
c
2
(uu +vv) +p
u +v
2c
2
{ (uv vu) p}
u v
2c
2
_
. (3.17.14)
Note that the eigenvalues represent the slope dy/dx of the characteristics. Therefore, they can be expressed as vectors:
1,2
= [u, v]
t
,
3
= [u +v, v u]
t
,
4
= [u v, v +u]
t
. (3.17.15)
Then, we nd
div
3
= div(v) +, div
4
= div(v) , (3.17.16)
where = v
x
u
y
. The left hand side (the divergence of the characteristic speed vector) represents the diver-
gence/convergence of characteristics, implying the existence of expansion/shock waves. I like these relations very
118 CHAPTER 3. EULER EQUATIONS
much because they indicate connections between the existence of the nonlinear waves and the physical quantities such
as the vorticity and divergence of the velocity eld. Interestingly, it is not just the vorticity or the divergence but their
combinations (the sum or the dierence) that tell us whether a shock or expansion wave exists (and also of which
family it is). So, if the sum div(v) + is negative/positive, there is a shock/expansion from 3-family; if the dierence
div(v) is negative/positive, there is a shock/expansion from 4-family. Or you might want to combine them into
one,
_
div
3
_
2
+
_
div
4
_
2
= {div(v)}
2
+
2
. (3.17.17)
This can also be a useful scalar (non-negative) quantity for detecting nonlinear waves (a shocks and/or expansions of
any family). Actually, these relations were successfully used for adaptive Euler schemes in Ref.[100].
I like the steady Euler equations because they are of mixed type. There have been attempts to construct optimal
discretization schemes for the Euler equations, i.e., central scheme for the elliptic part and upwind scheme for the
hyperbolic part, (see Refs.[94, 113] and references therein). That sounds like an optimal strategy, but typically in
most cases numerical schemes for the Euler equations are constructed based on the unsteady character, i.e., hyperbolic
character. Yes, the Euler equations are hyperbolic in time, and can never be elliptic in time, which in fact makes it
easier to construct a numerical scheme. So, I actually like both approaches.
3.18 Bernoullis Equation
Bernoullis equation is basically an integral of the Euler equation (the momentum equation),
v
t
+(gradv)v + grad p = 0, (3.18.1)
which can be written, by Equation (1.7.7), as
v
t
+ grad
_
v
2
2
_
v curlv + grad p = 0. (3.18.2)
This can be integrated for various ows; thus there are various forms of Bernoullis equation.
Unsteady, Incompressible, Irrotational Flows
For incompressible irrotational ows, we have = constant and curlv = 0. The latter implies the existence of the
velocity potential : v = grad. Then, the Euler equations (3.18.2) can be written as
grad
_
t
+
v
2
2
+
p
_
= 0. (3.18.3)
This means that the quantity in the parenthesis is spatially constant but may change in time. We therefore obtain
t
+
v
2
2
+
p
= f(t), (3.18.4)
where f(t) is an arbitrary function of time. This is called the unsteady Bernoullis equation and its steady version,
v
2
/2 +p/ = constant, is the famous Bernoullis equation obtained by Daniel Bernoulli.
Steady, Incompressible, Rotational Flows
For steady incompressible rotational ows, the Euler equations (3.18.2) becomes
grad
_
v
2
2
+
p
_
v curlv = 0. (3.18.5)
Because of the second term, there can be no spatially constant quantities. However, noting that the vector (vcurlv)
is perpendicular to the velocity vector v, we take the dot product of this equation and the velocity,
v grad
_
v
2
2
+
p
_
v (v curlv) = 0, (3.18.6)
3.18. BERNOULLIS EQUATION 119
so that the second term vanishes and we obtain
v grad
_
v
2
2
+
p
_
= 0. (3.18.7)
This shows that we do not have a global constant quantity but a streamwise constant quantity:
v
2
2
+
p
= constant, (3.18.8)
along a streamline, i.e., the stagnation pressure p
0
= v
2
/2+p takes dierent constant values for dierent streamlines.
To see how it changes across streamlines, we take the dot product of Equation (3.18.5) and a unit vector n in the
direction normal to the streamline,
n grad
_
v
2
2
+
p
_
n (v curlv) = 0, (3.18.9)
which gives the spatial rate of change of the stagnation pressure normal to the streamline,
p
0
n
=
n
_
p +
v
2
2
_
= n (v curlv) . (3.18.10)
This shows that the vorticity (curlv) is the source of the spatial variation of p
0
. However, it is also clear that in order
for p
0
to be constant everywhere, the vorticity does not have to be zero, or equivalently the ow does not have to be
irrotational. If the ow satises
v curlv = 0, (3.18.11)
everywhere, meaning that it is a Beltrami ow (see Section 1.5), p
0
will be a global constant. This is nice. I also like
two-dimensional ows in which we always have v curlv, and so by choosing n in the same direction as (v curlv),
we can write
n
_
p +
v
2
2
_
= q, (3.18.12)
q is the ow speed and is the magnitude of the vorticity.
Unsteady, Compressible, Irrotational Flows
For compressible ows, the density is not constant. This makes it non-trivial to integrate the pressure gradient term
1
grad p in the Euler equation. But it can still be integrated if we can express it as a gradient of some quantity. This
is possible for an adiabatic ow of a calorically perfect gas:
dp
= dh Tds = dh =
R
1
dT =
1
1
d(a
2
), (3.18.13)
where s is the entropy which is uniform for adiabatic ows and a is the speed of sound. This gives
1
grad p =
1
1
grad (a
2
). (3.18.14)
Now, introducing the velocity potential , we can write the the Euler equations (3.18.2) as
grad
_
t
+
v
2
2
+
a
2
1
_
= 0. (3.18.15)
Therefore we obtain the Bernoullis equation for unsteady compressible irrotational ows:
t
+
v
2
2
+
a
2
1
= g(t), (3.18.16)
where g(t) is an arbitrary function of time.
120 CHAPTER 3. EULER EQUATIONS
Steady, Compressible, Rotational Flows
In the same way as in the previous case, we can derive Bernoullis equations for steady compressible rotational ows.
Basically, we just need to replace p/ by
a
2
1
. Along a streamline, we have
v
2
2
+
a
2
1
= constant. (3.18.17)
Normal to the streamline, we have
n
_
v
2
2
+
a
2
1
_
= n (v curlv) , (3.18.18)
where n is a unit vector normal to the streamline.
3.19 Gas Dynamics Equation (Nonlinear Potential Equation)
Assume that the vorticity is everywhere zero: curl v = 0. Then, we can dene the velocity potential as
v = grad , (3.19.1)
and substitute this into the continuity equation to get the so-called potential equation:
div ( grad ) = 0. (3.19.2)
In Cartesian coordinates, the potential is dened as
u =
x
, v =
y
, w =
z
, (3.19.3)
and the potential equation becomes
(
x
)
x
+ (
y
)
y
+ (
z
)
z
= 0. (3.19.4)
This nonlinear equation requires an additional equation for the density which can be obtained from the energy equation
(Bernoullis equation) for an adiabatic ow of a calorically perfect gas (see Section 3.18):
a
2
1
+
u
2
+v
2
+w
2
2
= H = constant, (3.19.5)
where a =
1
+
u
2
+v
2
+w
2
2
, (3.19.6)
and the adiabatic relation, p/
= constant, as
=
_
1
1
2
M
2
_
u
2
+v
2
+w
2
u
2
+v
2
+w
2
1
__
1/(1)
. (3.19.7)
This formulation is valid for a steady adiabatic irrotational ow. It is also worth noting that in a steady adiabatic
irrotational ow the entropy is constant everywhere, i.e., the ow is homentropic (see Section 3.23). It follows then
that it is not valid for ows with shocks. But this conservative form is still approximately valid if the Mach number is
close to 1; it has actually been used successfully for various such ows. See Refs.[51, 155] for more details.
In two dimensions, the nonlinear potential ow can be described by the following Cauchy-Riemann type system:
(u)
x
+
(v)
y
= 0, (3.19.8)
v
x
u
y
= 0. (3.19.9)
3.20. LINEAR POTENTIAL EQUATION 121
See Ref.[134] for numerical methods for solving this system. I like the Cauchy-Riemann system, but in this case I like
the scalar potential equation (3.19.14) better simply because it is a scalar equation. It is so much easier to develop a
numerical scheme for a scalar equation than for a system.
I also like the potential equation written in a nonconservative form. To derive it, expand the continuity equation,
div v +v grad = 0, (3.19.10)
and take a dot product of the velocity vector and the momentum equation,
v [(gradv)v] +
a
2
v grad = 0, (3.19.11)
where the pressure gradient has been replaced by the density gradient by using the relation, grad p = a
2
grad . Now
it is clear that this and the continuity equation can be combined to give
div v
1
a
2
v [(gradv)v] = 0. (3.19.12)
This equation is known as the gas dynamics equation. I like it very much, especially the name, Gas Dynamics Equation.
It sounds exciting. In Cartesian coordinates, the gas dynamics equation is written as
_
1
u
2
a
2
_
u
x
+
_
1
v
2
a
2
_
v
y
+
_
1
w
2
a
2
_
w
z
uv
a
2
(u
y
+v
x
)
vw
a
2
(v
z
+w
y
)
wu
a
2
(w
x
+u
z
) = 0. (3.19.13)
Of course, this cannot be solved for the three velocity components (u, v, w) because it is a scalar equation. But if the
ow is irrotational and thus the velocity is given by the velocity potential : u =
x
, v =
y
, w =
z
, then we can
write the gas dynamics equation in terms of :
_
1
2
x
a
2
_
xx
+
_
1
2
y
a
2
_
yy
+
_
1
2
z
a
2
_
zz
2
x
y
a
2
xy
2
y
z
a
2
yz
2
z
x
a
2
zx
= 0. (3.19.14)
This is a scalar equation for , and so can be solved for with Equation (3.19.7). This equation is perhaps more
widely known as the full potential equation rather than as the gas dynamics equation.
Incidentally, the velocity potential exists even in viscous ows as long as the vorticity is zero. The viscous terms
actually vanish from the governing equations for irrotational ows (curl v = 0), but the viscous stresses do not
necessarily vanish and can be expressed in terms of the velocity potential. See Ref.[65] for an interesting theory of
viscous potential ows. Also, the potential formulation can be extended to rotational ows by decomposing the velocity
into the irrotational (potential) part and the rotational part. See Refs.[43, 44, 96] for details.
3.20 Linear Potential Equation
In general, I like to linearize nonlinear equations. I would linearize the nonlinear potential equation as follows.
In two dimensions, a ow over a slender body with the free stream velocity U
, v
( u
, v
<< U
):
u = U
+u
, v = v
. (3.20.1)
Then, introducing the velocity potential and its perturbation as
u =
x
= U
x
, v =
y
=
y
, (3.20.2)
we nd from Equation (3.19.14) that, for two-dimensional ows,
(1 M
2
xx
+
yy
= M
2
( + 1)
xx
U
. (3.20.3)
Note that the right hand side must be retained for transonic ows M
xx
will be very small for transonic ows and the equation will become nearly one-
dimensional without the right hand side. However, the right hand side can be ignored for subsonic and supersonic
ows, resulting in
(1 M
2
xx
+
yy
= 0. (3.20.4)
122 CHAPTER 3. EULER EQUATIONS
This is the two-dimensional linear potential equation. In three dimensions, we introduce a perturbation velocity w
in
z direction as
w = w
=
z
=
z
<< U
, (3.20.5)
and obtain, from Equation (3.19.14), the following three-dimensional linear potential equation:
(1 M
2
xx
+
yy
+
zz
= 0. (3.20.6)
In either two or three dimensions, once the perturbation potential
1
2
U
2
=
2
M
2
(p/p
1) =
_
2u
+ (1 M
2
)
u
2
U
2
+
v
2
+w
2
U
2
_
. (3.20.7)
I like the linear potential equation because it can be solved easily for small Mach number (it is almost the Laplace
equation), but not so much for a large Mach number (it is now hyperbolic. See Section 1.17 for classication). See
Refs.[53, 84] for more details on the linearized potential equations, and Refs.[51, 155] for numerical methods for solving
them.
3.21 Small Perturbation Equations
We can write the linear potential equation in terms of the velocity components, which are called the small per-
turbation equations (or the Prandtl-Glauert equations). For supersonic ows (M
2
u
x
v
y
= 0, (3.21.1)
v
x
u
y
= 0, (3.21.2)
where
2
= M
2
2
1 0
_
_
. (3.21.4)
Eigenstructure
A = RL, (3.21.5)
where
=
_
_
1
0
0
1
_
, R =
_
_
1
2
1
2
1
2
1
2
_
_
, L =
_
_
1
1
_
_
. (3.21.6)
Characteristic Form
Lw
x
+LAw
y
= 0, (3.21.7)
(Lw)
x
+(Lw)
y
= 0, (3.21.8)
that is,
x
(u v) +
1
y
(u v) = 0, (3.21.9)
x
(u +v)
1
y
(u +v) = 0. (3.21.10)
3.22. INCOMPRESSIBLE POTENTIAL EQUATION 123
This shows that u v is constant along dy/dx = 1/ while u +v is constant along dy/dx = 1/. Note that this
diagonalization is valid only for supersonic ows (M
2 because then
= 1 and everything looks very simple.
For subsonic ows (M
) u
x
+v
y
= 0, (3.21.11)
v
x
u
y
= 0. (3.21.12)
Then, the eigenvalues become complex, and therefore the system is elliptic. As for the Mach number, I like to set
M
= 0 (incompressible ows) because the system reduces to the Cauchy-Riemann system. As a matter of fact, I
would use the Prandtl-Glauert transformation:
= x, = y
1 M
2
, U =
1 M
2
u, V = v, (3.21.13)
so that the system, (3.21.11) and (3.21.12), becomes the Cauchy-Riemann system,
U
+V
= 0, (3.21.14)
V
= 0, (3.21.15)
for any Mach number (|M
| < 1). This is very nice. Any subsonic small perturbation ow can be transformed into
an incompressible ow. See Refs.[74, 84] for more details on the Prandtl-Glauert transformation.
3.22 Incompressible Potential Equation
As we have seen in the previous sections, the velocity potential is very useful in computing steady compressible
irrotational ows. Here, we consider steady incompressible ows. For steady incompressible ows, the continuity
equation is div v = 0. Therefore, by substituting the velocity potential (3.19.1) into the continuity equation, we obtain
div grad = 0. (3.22.1)
That is, the velocity potential satises the Laplace Equation in all dimensions. I like it because various exact solutions
are available for the Laplace equation. Two-dimensional ows are even nicer because a potential ow can be described
by the velocity potential and the stream function as described in Section 2.9. It is nice especially because the velocity
potential and the stream function satisfy the Cauchy-Riemann equations (see Section 2.9.2), and we can generate
various exact solutions by the theory of complex variables. Note that the applicability of the theory of complex
variables is limited to two-dimensional ows; it doesnt apply to general three-dimensional ows including axisymmetric
ows because the stream function does not satisfy the Laplace equation. See Section 4.24 for the governing equation
of the stream function.
3.23 Croccos Equation
Croccos equation (or Croccos theorem) is a modied form of the momentum equation. Consider the momentum
equation written in the primitive form,
v
t
+ (gradv)v +
1
grad p = 0. (3.23.1)
Using the following relations,
(gradv)v = grad
_
v
2
2
_
v curl v, (3.23.2)
Tds = dh
1
grad p, (3.23.3)
H = h +
v
2
2
, (3.23.4)
124 CHAPTER 3. EULER EQUATIONS
the momentum equation can be written as
v
t
v = Tgrad s grad H, (3.23.5)
where = curl v. This is Croccos equation. It is a quite interesting equation. For example, we can use it to show
that the entropy is constant along a streamline for a steady adiabatic ow. To do so, rst, note that for a steady
adiabatic ow, the equation for the total enthalpy (3.3.7)
DH
Dt
=
1
p
t
, (3.23.6)
reduces to
v grad H = 0, (3.23.7)
which implies, incidentally, that the total enthalpy is preserved along a streamline. Now, take a dot product of the
velocity and Croccos equation to get
v (v ) = v grad H Tv grad s. (3.23.8)
Observe that the left hand side vanishes because v (v ), and also that the rst term on the right vanishes by
Equation (3.23.7). Then, we are left with
0 = v grad s. (3.23.9)
This indicates that the entropy is constant along a streamline. This completes the proof for constant entropy along a
streamline in a steady adiabatic ow. On the other hand, if we assume that = 0 and H is uniform, then it follows
immediately from Equation (3.23.5) that the entropy is constant everywhere, i.e., the ow is homentropic. That is, zero
vorticity implies uniform entropy if H is uniform. Basically, Croccos equation is another form of the Euler equation
(i.e., the momentum equation). So, it is perfectly possible to replace the momentum equation by Croccos equation
in the Euler system (then we would have to call it the Crocco system). I like Croccos equation. It is a quite insightful
equation.
3.24 Kelvins Circulation Theorem
The circulation is dened for a closed curve C by
=
C
v dl, (3.24.1)
where dl is an element taken counterclockwise along C, and it is related to the vorticity = curl v by
=
C
v dl =
S
curl v ndS =
S
ndS, (3.24.2)
i.e., the circulation around C is equal to the in/out-ow of the vorticity across any surface S whose boundary is C.
The rate of change of the circulation of a closed curve C (moving with the uid) can be obtained easily from the
primitive form of the Euler equations:
D
Dt
=
C
grad p
dl. (3.24.3)
If derived from the Navier-Stokes equations, by the way, it will have a contribution from the viscous term on the right
hand side:
D
Dt
=
C
1
div dl
C
grad p
dl, (3.24.4)
where is the viscous stress tensor (4.2.1). For inviscid incompressible ows, is constant and the right hand side of
Equation (3.24.3) vanishes:
C
grad p
dl =
1
C
grad p dl =
1
C
dp = 0. (3.24.5)
3.25. VORTICITY EQUATION 125
Therefore, we have from Equation (3.24.3)
D
Dt
= 0, (3.24.6)
i.e., the circulation is conserved. This is interesting. Even for compressible ows, where Equation (3.24.3) can be
written as
D
Dt
=
C
grad p
dl =
C
(grad h T grad s) dl =
C
T grad s dl, (3.24.7)
if the entropy is constant, then again the circulation is conserved:
D
Dt
= 0. (3.24.8)
This shows that the circulation around a closed curve remains constant at all times as the curve moves with the uid.
This is called Kelvins circulation theorem. I like it because it gives interesting and important explanations about
vortical ows. The vorticity denes a vector eld, and thus integral curves (i.e., the curves tangent to the vorticity
vector) may be dened. These curves are generally called the vortex lines. The surface formed by the vortex lines
passing through a closed curve is called a vortex tube (or laments). The strength of the vortex tube is dened as the
integral of the vorticity over a cross-section of a tube, which is equivalent to the line integral of the velocity around the
closed curve, i.e., the circulation . Now I begin to see the connection between the vortex tube and Kelvins circulation
theorem. Below are the theorems, known as Helmholzs theorems, obtained as direct consequences of the circulation
theorem:
1. A ow that is initially irrotational remains irrotational if no rotational forces exist.
2. The strength of a vortex tube is constant along its length.
3. A vortex tube cannot end in the ow, and must extend to the boundary or form a closed loop.
One of the most interesting behaviors that I like, which is implied by the second theorem, is that the vortex strength
increases/decreases as the cross-section of the vortex tube shrinks/expands. This is called the vortex stretching
mechanism as the cross-sectional changes are associated with the stretching of the tube.
3.25 Vorticity Equation
Evolution of the vorticity is described by the vorticity equation, which can be derived by taking the curl of the
momentum equations. Consider the primitive form of the momentum equations:
v
t
+ (grad v)v +
grad p
= 0, (3.25.1)
which can be written by Equation (1.7.7) as
v
t
+ grad
v v
2
v curl v +
grad p
= 0. (3.25.2)
Taking the curl, we get
t
+
curl
_
grad
v v
2
_
curl(v ) + curl
_
grad p
_
= 0, (3.25.3)
where = curl v is the vorticity, which can be written by Equations (1.7.19) and (1.7.20) as
t
v
_
grad p = 0. (3.25.4)
Therefore, we have
D
Dt
=
t
+ (grad )v = (grad v) (div v) +
grad grad p
2
. (3.25.5)
126 CHAPTER 3. EULER EQUATIONS
The left hand side is the rate of change of the vorticity along the streamlines. The rst terms on the right hand
side describes the vortex stretching and twisting (or tilting), and the second term describes stretching due to the
compressibility. The third term is called the baroclinic term, describing the change in the vorticity by the misalignment
of the gradient of the density and the gradient the pressure. The baloclinic term vanishes for barotropic gases, i.e., if
the pressure is a function of the density only, because then the density and pressure gradients become parallel to each
other and the cross product vanishes. An example of such ows is the homentropic ow discussed in Section 3.14. Of
course, the barocliic term vanishes also for incompressible ows because the density gradient vanishes.
The vorticity equation can be arranged further by using the following relations:
t
_
_
=
1
t
=
1
t
+
_
_
div v +
2
(grad ) v, (3.25.6)
where the continuity equation (3.3.1) has been substituted, and
_
grad
_
__
v =
_
grad
_
1
_
+
1
(grad )
_
v
=
1
2
(grad ) v +
1
(grad )v
=
2
(grad ) v +
1
_
=
t
_
_
+
_
grad
_
__
v =
1
t
+
_
_
div v +
1
_
= (grad v)
_
_
+
grad grad p
3
. (3.25.9)
Comparing this form with Equation (3.25.5), we see that the stretching term due to compressibility is gone. In fact,
I like this form very much because the compressibility eect is incorporated into the stretching-twisting term and it
shows that it is the quantity /, not , that behaves like the vorticity in the same way as in incompressible ows.
That is, / satises the equation that satises in incompressible ows.
I like the vorticity equation because it describes very interestingly how the vorticity evolves in a ow and it is one
of the basic tools to understand turbulence as well as the basic equation for vortex methods [29].
3.26 Shallow-Water (or Saint Venant) Equations
Like a smooth water surface, when the vertical (z-direction) velocity and the vertical displacement of the surface
is suciently small (compared with the frequency of water waves; see Figure 3.26.1), the water depth h and the
velocity v = [u, v]
t
are governed by the following shallow-water equations (also called the Saint Venant equations):
t
h + div(hv) = 0, (3.26.1)
t
(hv) + div(hvv) + grad(gh
2
/2) = ghgrad z
0
(x, y), (3.26.2)
where g = 9.8m/s
2
and z
0
(x, y) is a function that denes the shape of the bottom. Note that all quantities are
independent of z. It is interesting that the shallow-water equations can be derived from the Euler equations: add
a gravitational force term in the z-component, ignore the z-velocity and any variation in z, and then integrate all
equations in z from the bottom to the water surface. In doing so, we obtain from the z-momentum equation,
p
z
= g p = p
0
+g(h +z
0
z). (3.26.3)
This is a familiar relationship, called the hydrostatic pressure relation. Conversely, if we write p =
1
2
gh
2
, rewrite h as
, and assume z
0
= constant, then, the shallow-water equations become
t
+ div(v) = 0, (3.26.4)
t
(v) + div(vv) + gradp = 0, (3.26.5)
3.26. SHALLOW-WATER (OR SAINT VENANT) EQUATIONS 127
u
h
z
0
(x)
x
z
O
Figure 3.26.1: One-dimensional free surface ow.
which are nothing but the Euler equations without the energy equation. In fact, if we look at the eigenstructure, we
nd that the quantity
gh corresponds to the speed of sound:
gh =
2p/h =
2p/. (3.26.6)
This implies that we have = 2. Therefore, the Euler equations above are in the form of the homentropic/isothermal
Euler equations with = 2. Basically, the shallow-water equations are a subset of the Euler equations. This is very
interesting. I like it.
Incidentally, it is easy to derive a conservation form for = gh, but it requires dierentiations and therefore it is
not really valid for ows with shocks. See Ref.[81] for details.
3.26.1 1D Shallow-Water Equations
h
t
+ (hu)
x
= 0, (3.26.7)
(hu)
t
+
_
hu
2
+
1
2
gh
2
_
x
= gh(z
0
)
x
, (3.26.8)
or in the vector form
U
t
+F
x
= S, (3.26.9)
where
U =
_
h
hu
_
, F =
_
hu
hu
2
+
1
2
gh
2
_
, S =
_
0
gh(z
0
)
x
_
. (3.26.10)
Jacobian
A =
F
U
=
_
0 1
gh u
2
2u
_
. (3.26.11)
128 CHAPTER 3. EULER EQUATIONS
Eigenstructure
A = RL, (3.26.12)
where
=
_
u
gh 0
0 u +
gh
_
, R =
_
_
1 1
u
gh u +
gh
_
_
, (3.26.13)
L =
_
_
u +
gh
2
gh
1
2
gh
gh
2
gh
1
2
gh
_
_
, LdU =
_
_
1
2
dh
h
2
gh
du
1
2
dh +
h
2
gh
du
_
_
. (3.26.14)
3.26.2 2D Shallow-Water Equations
h
t
+ (hu)
x
+ (hv)
y
= 0, (3.26.15)
(hu)
t
+
_
hu
2
+
1
2
gh
2
_
x
+ (huv)
y
= gh(z
0
)
x
, (3.26.16)
(hv)
t
+ (huv)
x
+
_
hv
2
+
1
2
gh
2
_
y
= gh(z
0
)
y
, (3.26.17)
or in the vector form
U
t
+F
x
+G
y
= S, (3.26.18)
where
U =
_
_
h
hu
hv
_
_
, F =
_
_
hu
p +hu
2
huv
_
_
, G =
_
_
hv
huv
p +hv
2
_
_
, S =
_
_
0
gh(z
0
)
x
gh(z
0
)
y
_
_
. (3.26.19)
Jacobians
A =
F
U
=
_
_
0 1 0
gh u
2
2u 0
uv v u
_
_
, B =
G
U
=
_
_
0 0 1
uv v u
gh v
2
0 2v
_
_
, (3.26.20)
A
n
= An
x
+Bn
y
=
_
_
0 n
x
n
y
ghn
x
uq
n
q
n
+un
x
un
y
ghn
y
vq
n
vn
x
q
n
+vn
y
_
_
, (3.26.21)
where
q
n
= un
x
+vn
y
. (3.26.22)
Eigenstructure
A
n
= R
n
n
L
n
, (3.26.23)
3.26. SHALLOW-WATER (OR SAINT VENANT) EQUATIONS 129
where
n
=
_
_
q
n
gh 0 0
0 q
n
0
0 0 q
n
+
gh
_
_
, (3.26.24)
R
n
=
_
_
1 0 1
u
ghn
x
x
u +
ghn
x
v
ghn
y
y
v +
ghn
y
_
_
, (3.26.25)
L
n
=
_
gh +q
n
2
gh
n
x
2
gh
n
y
2
gh
q
x
y
gh q
n
2
gh
n
x
2
gh
n
y
2
gh
_
_
, (3.26.26)
and
q
= u
x
+v
y
, (3.26.27)
[
x
,
y
] = [n
y
, n
x
]. (3.26.28)
3.26.3 3D Shallow-Water Equations
Unfortunately, I cannot write them because I dont know them. I would denitely like them if the shallow-water
equations existed in three dimensions.
130 CHAPTER 3. EULER EQUATIONS
Chapter 4
Navier-Stokes Equations
4.1 Navier-Stokes Equations
I like a little bit more realistic model for describing a uid ow than the Euler equations. So, I will add some missing
pieces to the Euler equations: the viscous stress tensor and the heat ux vector q. Yes, these are the Navier-Stokes
equations.
Conservative form of the Navier-Stokes equations
t
+ div(v) = 0, (4.1.1)
t
(v) + div(vv) = div , (4.1.2)
t
(E) + div(vH) = div (v) div q, (4.1.3)
where is the total surface stress tensor which consists of the static pressure p and the viscous stress tensor ,
= pI +, (4.1.4)
where I is the identity matrix.
Primitive form of the Navier-Stokes equations
D
Dt
+ div v = 0, (4.1.5)
Dv
Dt
+ grad p = div , (4.1.6)
Dp
Dt
+p div v = ( 1) [ : grad v div q] . (4.1.7)
The continuity equation (4.1.5) can be derived from the conservative form (4.1.1) simply by expanding the divergence
term. The momentum equation (4.1.6) can be derived from the conservative form (4.1.2) by expanding the left hand
side and substituting the continuity equation (4.1.5). Note also that
div = div (pI) + div = grad p + div . (4.1.8)
To derive the pressure equation (4.1.7), rst derive the kinetic energy equation by taking the dot product of the velocity
vector and the momentum equation (4.1.6),
D
Dt
_
v
2
2
_
+ (grad p) v = : grad v + div (v), (4.1.9)
131
132 CHAPTER 4. NAVIER-STOKES EQUATIONS
then subtract the kinetic equation from the energy equation (4.1.3), and nally use the equation of state for a perfect
gas, i.e., p = ( 1)e, and the continuity equation to convert the time derivative of the internal energy into that of
the pressure.
Which form do you like? I think I like the conservative form better because conservation is a very important property
(to ensure a correct shock speed in particular), and it is easier to develop a conservative scheme by directly discretizing
the conservative form.
4.2 Viscous Stresses and Heat Fluxes
In Newtonian uids (e.g., the sea-level air or water), the viscous stress tensor depends linearly on the strain and it
is given by
= (div v)I +
_
grad v + (grad v)
t
, (4.2.1)
where grad v + (grad v)
t
is the rate of strain tensor (also called the deformation tensor), is the viscosity coecient,
and is the second viscosity coecient. The bulk viscosity
B
dened by
B
= +
2
3
, (4.2.2)
is usually very small, and therefore we may set
B
= 0, i.e.,
=
2
3
. (4.2.3)
This is called Stokes hypothesis. I like this because it is a reasonable assumption. Consider the average normal stress
P (also called the mean pressure),
P
ii
3
= p
_
1
3
3 div v +
2
3
div v
_
= p
_
+
2
3
_
div v
= p
B
div v, (4.2.4)
where it has been dened as negative because
ii
is negative when acting onto the uid element. Basically, this gives,
independent of coordinates, an average normal force acting on a uid element; such a quantity is generally considered
as the thermodynamic pressure. Then, it is a state variable. That is, under the assumption of the thermodynamic
equilibrium, it should not depend on the rate of the volume change, div v. So, it would be reasonable to accept Stokes
hypothesis
B
= 0, so that we have a reasonable relation, P = p. Of course, this is only a hypothesis, but it works very
well for a wide range of conditions. See Refs.[136, 169] for more details. Note that the viscosity coecient generally
depends on the temperature. On the other hand, the heat ux is given by Fouriers law,
q = grad T, (4.2.5)
where is the heat conductivity which is a function of the temperature. The negative sign is put to dene q as positive
when the temperature gradient is negative. That is, the heat ows from higher to lower temperature. In fact, this is
the basic assumption (or observation) of this empirical law. I think that it is indeed very reasonable.
I like both Newtonian uids and Fouriers law for the same reason: both the viscous stress and heat ux depend
linearly on the derivatives of the primitive variables. This is nice.
4.3 Articial Bulk Viscosity
As mentioned in the previous section, the bulk viscosity
B
is typically set to be zero, but there are methods in
CFD that bring it back in the governing equations as an articial viscosity to improve numerical algorithms:
=
_
articial
B
2
3
_
(div v)I +
_
grad v + (grad v)
t
, (4.3.1)
4.4. VISCOSITY AND HEAT CONDUCTIVITY 133
where
articial
B
(> 0) is the articial bulk viscosity. In Ref.[124], the articial bulk viscosity is introduced to accelerate
the convergence of an incompressible Navier-Stokes solver. The solver is based on the pseudo-compressible formulation
[21]. This formulation generates pseudo-acoustic-waves (see Section 3.13), and the numerical convergence towards the
steady state depends on how quickly the pseudo-acoustic-waves are eliminated via propagation and damping. The
articial bulk viscosity is then introduced to enhance the damping eect. I like the idea very much, especially the fact
that the bulk viscosity term has no eect on the steady solution because it is proportional to the divergence of the
velocity div v that vanishes in the steady state for incompressible ows. Therefore, it only aects the convergence of
the numerical method. The idea has been extended to the compressible Euler [90] and Navier-Stokes equations [3],
where the articial bulk viscosity term is formulated by div (v) so that it vanishes in the steady state for compressible
ows. Of course, these methods are designed for steady problems, but time-accurate computations are possible by
implicit time-integration schemes where the fast steady solver can be employed to solve the implicit-residual equations
over each physical time step.
Another example is the localized articial diusivity method [27], where the articial bulk viscosity is introduced as
a means to eectively capture shock waves without deteriorating the resolution of turbulence structure. To eectively
accomplish the task, the articial bulk viscosity is designed to get activated locally at shocks, thus the term "localized".
The localized articial diusivity method has been applied to compressible shock-turbulent simulations with very high-
order numerical schemes. See Ref.[69] and references therein for details.
It is interesting to observe the impact of the articial bulk viscosity term on the divergence = div v and the
vorticity = curl v:
t
+ = div
_
1
grad (
articial
B
)
_
+ , (4.3.2)
t
+ =
1
2
grad grad (
articial
B
) + , (4.3.3)
which have been obtained by taking the divergence and the curl of the momentum equations (4.1.6) divided by the
density. It is clear that the articial bulk viscosity plays a role of diusion for the divergence and is expected to provide
damping for faster convergence and for shock capturing. On the other hand, the second equation shows apparently
that the articial bulk viscosity can actually generate the vorticity. In the convergence acceleration methods mentioned
above, it actually does not aect the vorticity at all because it vanishes in the steady state. In the localized articial
diusivity method, it is expected to be small at numerically captured shocks because the cross product of the density
gradient and the gradient of the articial bulk viscosity is expected to be small (they are both dominant locally in the
direction normal to a shock front) [70]. It is nice to be small because accurate prediction of vorticity production across
a shock can be very important in turbulent simulations [78].
I like these ideas in that they attempt to improve numerical methods by improving the governing equations rather
than the discretization. In a way, it makes the construction of numerical schemes easier: given such improved governing
equations (with desired features built in), you just have to discretize them accurately.
4.4 Viscosity and Heat Conductivity
The heat conductivity can be evaluated through the viscosity coecient by utilizing the denition of the Prandtl
number Pr,
Pr =
c
p
=
R
( 1)
=
R
Pr( 1)
. (4.4.1)
This is very nice because the heat ux vector can then be expressed in any of the following forms:
q = grad T =
Pr( 1)
grad (a
2
) =
Pr
grad e =
Pr
grad h, (4.4.2)
where the relations, a
2
= p/ = RT, e =
R
1
T, and h =
R
1
T have been used. It is nice to know then that
the Prandtl number for the sea-level air is nearly constant for a wide range of temperature; Pr = 0.72 is a valid
approximation for the temperature between 200K and 1000K. In general, it is known that the Prandtl number is
related to by
Pr =
4
9 5
. (4.4.3)
134 CHAPTER 4. NAVIER-STOKES EQUATIONS
This can be useful for non-ideal gases. See Ref.[159] for more details. Therefore all we need to evaluate the heat ux
is the viscosity coecient . The viscosity coecient can be computed by Sutherlands law,
=
0
T
0
+C
T +C
_
T
T
0
_3
2
, C = 110.5 [K], T
0
= 273.1 [K],
0
= 1.716 10
5
[Pa s], (4.4.4)
for air. This is also a good approximation for a wide range of temperature, between 170K and 1900K. See Ref.[169]
for more details.
4.5 Other Energy Equations
We obtain the equation for the internal energy e by subtracting the kinetic energy equation (4.1.9) from the total
energy equation (4.1.3):
De
Dt
= p div v div q + , (4.5.1)
where is the dissipation term given by
= : grad v =
_
grad v + (grad v)
t
: grad v +(div v)
2
. (4.5.2)
It can be shown that the dissipation term can be written as
=
2
_
grad v + (grad v)
t
2
3
(div v)I
_
:
_
grad v + (grad v)
t
2
3
(div v)I
_
+
_
+
2
3
_
(div v)
2
=
2
__
u
i
x
j
+
u
j
x
i
_
2
3
ij
u
k
x
k
_
2
+
_
+
2
3
__
u
k
x
k
_
2
. (4.5.3)
It follows immediately from this that the dissipation term is positive if
0, +
2
3
0. (4.5.4)
This is very nice. It shows that Stokes hypothesis ( =
2
3
) guarantees the positivity of the dissipation term. If you
are interested to derive Equation (4.5.3) from Equation (4.5.2), rst you need to understand these identities,
_
grad v + (grad v)
t
:
_
grad v (grad v)
t
= 0, (4.5.5)
(div v)I :
_
grad v (grad v)
t
= 0. (4.5.6)
Then, you write
= : grad v
= :
_
1
2
_
grad v + (grad v)
t
_
+
1
2
_
grad v (grad v)
t
_
_
, (4.5.7)
and expand this with = (div v)I +[grad v + (grad v)
t
] and the identities above to get
=
2
_
grad v + (grad v)
t
:
_
grad v + (grad v)
t
+
2
(div v)I :
_
grad v + (grad v)
t
=
2
_
grad v + (grad v)
t
:
_
grad v + (grad v)
t
+(div v)
2
. (4.5.8)
Finally, you somehow arrange this into the nal form (4.5.3), e.g., by adding
_
2
3
(div v)
2
2
3
(div v)
2
and refactoring.
4.6. TOTAL PRESSURE EQUATION 135
Basically, the energy equation in the Navier-Stokes equations can be replaced by any of the followings:
(H)
t
p
t
+ div(vH) = div (v) div q, (4.5.9)
DH
Dt
p
t
= div (v) div q, (4.5.10)
DE
Dt
+ div (pv) = div (v) div q, (4.5.11)
De
Dt
+p div v = div q, (4.5.12)
Dh
Dt
Dp
Dt
= div q, (4.5.13)
D
Dt
_
v
2
2
_
+ (grad p) v = div (v) . (4.5.14)
It is very nice that we have so many options for the energy equation. Naturally, the entropy equation/inequality can
also be used:
Ds
Dt
+ div
_
q
T
_
=
T
+
T
2
(gradT) (gradT) 0. (4.5.15)
4.6 Total Pressure Equation
I nd it interesting that the governing equation for the total pressure, p
t
= p + v
2
/2, can be derived quite eas-
ily for steady incompressible ows. Take the dot product of the velocity and the steady version of the momentum
equation (4.1.2) to get
v div(vv) = v grad p +v div , (4.6.1)
which becomes by the vector identities (1.7.15) and (1.7.11), and the continuity equation
div
_
v
v v
2
_
+
v v
2
div(v) = div(pv) +p
(div v)I +
_
grad v + (grad v)
t
=
_
grad v + (grad v)
t
, (4.7.4)
which may be integrated over the body to calculate the viscous force. It is even more interesting to attempt to calculate
the viscous force. Consider the viscous force acting on an innitesimal surface element on the body:
|
body
ndS =
_
grad v + (grad v)
t
ndS, (4.7.5)
where dS is the innitesimal surface element and n is the unit outward normal vector. We write it as
|
body
ndS =
_
grad v (grad v)
t
ndS + 2(grad v)
t
ndS. (4.7.6)
Then, we can show that the last term vanishes on the body. First, I can easily show that
(grad v)
t
ndS = grad q
n
dS, (4.7.7)
where q
n
= v n. The gradient of the normal velocity can be decomposed into three orthogonal directions: the normal
direction and two tangential directions.
grad q
n
=
q
n
n
n +
q
n
+
q
n
m
m, (4.7.8)
where the tangential directions are denoted by and m, and their unit vectors are denoted by and m, respectively.
We then immediately nd that the tangential derivatives vanish by the no-slip condition (i.e., no suction or injection):
(grad v)
t
ndS =
q
n
n
ndS. (4.7.9)
Does this vanish also? Yes, it does. Recall that the velocity divergence vanishes on the body, which can be expressed
in the orthogonal coordinates taken with respect to the normal direction:
div v =
q
n
n
+
q
+
q
m
m
= 0. (4.7.10)
The no-slip condition implies that the tangential velocities are uniformly zero in the tangential directions on the body.
Therefore, their derivatives vanish, and we are left with
q
n
n
= 0, (4.7.11)
on the body, and thus we do have
(grad v)
t
ndS = 0. (4.7.12)
We therefore conclude that
|
body
ndS =
_
grad v (grad v)
t
ndS. (4.7.13)
I like it because it can be written by Equation (1.7.6) as
|
body
ndS = (curl v n) dS = ( n) dS. (4.7.14)
This is interesting. The viscous force can be calculated by integrating the vorticity over the body. The result is, of
course, valid also for incompressible ows [11, 122]. So, if your CFD code produces accurate vorticity on the body, the
viscous force can be calculated by integrating it over the body for steady incompressible/compressible ows.
4.8. 1D NAVIER-STOKES EQUATIONS 137
4.8 1D Navier-Stokes Equations
The 1D Navier-Stokes equations in conservative form:
U
t
+
F
x
= 0, (4.8.1)
where
U =
_
u
E
_
_, F =
_
_
u
u
2
+p
xx
uH
xx
u +q
x
_
_, (4.8.2)
xx
= (2 +)
u
x
, (4.8.3)
q
x
= k
T
x
=
Pr( 1)
a
2
x
=
Pr( 1)
x
_
p
_
. (4.8.4)
I like that Stokes hypothesis ( =
2
3
) simplies the viscous stress to
xx
=
4
3
u
x
. (4.8.5)
Incidentally, the dissipation term (4.5.2) becomes
= 2
_
u
x
_
2
+
_
u
x
_
2
, (4.8.6)
which is always positive, of course. I like this dissipation term, but it is not needed in the conservative form (4.8.1).
4.9 2D Navier-Stokes Equations
The 2D Navier-Stokes equations in conservative form:
U
t
+
F
x
+
G
y
= 0, (4.9.1)
where
U =
_
u
v
E
_
_
, (4.9.2)
F =
_
_
u
u
2
+p
xx
uv
yx
uH
xx
u
xy
v +q
x
_
_
, (4.9.3)
G =
_
_
v
uv
xy
v
2
+p
yy
vH
yx
u
yy
v +q
y
_
_
, (4.9.4)
138 CHAPTER 4. NAVIER-STOKES EQUATIONS
xx
= 2
u
x
+
_
u
x
+
v
y
_
, (4.9.5)
yy
= 2
v
y
+
_
u
x
+
v
y
_
, (4.9.6)
xy
=
yx
=
_
u
y
+
v
x
_
, (4.9.7)
q
x
= k
T
x
=
Pr( 1)
a
2
x
=
Pr( 1)
x
_
p
_
, (4.9.8)
q
y
= k
T
y
=
Pr( 1)
a
2
y
=
Pr( 1)
y
_
p
_
. (4.9.9)
I like again that Stokes hypothesis ( =
2
3
) simplies the viscous stresses to
xx
=
2
3
_
2
u
x
v
y
_
, (4.9.10)
yy
=
2
3
_
2
v
y
u
x
_
. (4.9.11)
Again, the dissipation term (4.5.2) is always positive:
=
_
2
_
u
x
_
2
+ 2
_
v
y
_
2
+
_
v
x
+
u
y
_
2
_
+
_
u
x
+
v
y
_
2
. (4.9.12)
I like this dissipation term, but again it has nothing to do with the conservative form (4.9.1).
4.10 3D Navier-Stokes Equations
The 3D Navier-Stokes equations in conservative form:
U
t
+
F
x
+
G
y
+
H
z
= 0, (4.10.1)
where
U =
_
u
v
w
E
_
_
, (4.10.2)
F =
_
_
u
u
2
+p
xx
uv
yx
uw
zx
uH
xx
u
xy
v
xz
w +q
x
_
_
, (4.10.3)
G =
_
_
v
vu
xy
v
2
+p
yy
vw
zy
vH
yx
u
yy
v
yz
w +q
y
_
_
, (4.10.4)
4.11. AXISYMMETRIC NAVIER-STOKES EQUATIONS 139
H =
_
_
w
wu
xz
wv
yz
w
2
+p
zz
wH
zx
u
zy
v
zz
w +q
z
_
_
, (4.10.5)
xx
= 2
u
x
+
_
u
x
+
v
y
+
w
z
_
, (4.10.6)
yy
= 2
v
y
+
_
u
x
+
v
y
+
w
z
_
, (4.10.7)
zz
= 2
w
z
+
_
u
x
+
v
y
+
w
z
_
, (4.10.8)
xy
=
yx
=
_
u
y
+
v
x
_
, (4.10.9)
xz
=
zx
=
_
w
x
+
u
z
_
, (4.10.10)
yz
=
zy
=
_
v
z
+
w
y
_
, (4.10.11)
q
x
= k
T
x
=
Pr( 1)
a
2
x
=
Pr( 1)
x
_
p
_
, (4.10.12)
q
y
= k
T
y
=
Pr( 1)
a
2
y
=
Pr( 1)
y
_
p
_
, (4.10.13)
q
z
= k
T
z
=
Pr( 1)
a
2
z
=
Pr( 1)
z
_
p
_
. (4.10.14)
Of course, I like Stokes hypothesis ( =
2
3
) which simplies the viscous stresses to
xx
=
2
3
_
2
u
x
v
y
w
z
_
, (4.10.15)
yy
=
2
3
_
2
v
y
u
x
w
z
_
, (4.10.16)
zz
=
2
3
_
2
w
z
u
x
v
y
_
. (4.10.17)
Yes, I like the dissipation term (4.5.2) that is always positive:
=
_
2
_
u
x
_
2
+ 2
_
v
y
_
2
+ 2
_
w
z
_
2
+
_
v
x
+
u
y
_
2
+
_
w
y
+
v
z
_
2
+
_
u
z
+
w
x
_
2
_
+
_
u
x
+
v
y
+
w
z
_
2
, (4.10.18)
which is again irrelevant to the conservative form (4.10.1).
4.11 Axisymmetric Navier-Stokes Equations
140 CHAPTER 4. NAVIER-STOKES EQUATIONS
I like the axisymmetric Navier-Stokes equations as much as the axisymmetric Euler equations in Section 3.7. Again,
these equations describe a three-dimensional ow but formally two-dimensional equations. Since the inviscid part is
already described in Section 3.7, here we focus on the viscous terms.
Consider cylindrical coordinates in Figure 1.5.1, where the velocity is denoted by v = [u
r
, u
, u
z
]
t
and the heat ux
is denoted by q = [q
r
, q
, q
z
]
t
. By taking the z-axis as the axis of symmetry, we assume an axisymmetric ow with
u
= 0 and / = 0. First, the divergence of the heat ux can be written by Equation (1.5.23) as
div q =
1
r
(rq
r
)
r
1
r
q
q
z
z
. (4.11.1)
Next, the divergence of the viscous stresses in the momentum equations is given by (see Equation (1.5.45))
div = div
_
rr
r
rz
zr
z
zz
_
_
=
_
_
1
r
(r
rr
)
r
+
1
r
+
rz
z
1
r
(r
r
)
r
+
1
r
z
z
+
r
r
1
r
(r
zr
)
r
+
1
r
+
zz
z
_
_
. (4.11.2)
where the viscous stress tensor is given by
= (div v)I +
_
grad v + (grad v)
t
=
_
1
r
(ru
r
)
r
+
1
r
u
+
u
z
z
_
I +
_
_
_
_
u
r
r
1
r
u
r
r
u
r
z
1
r
u
+
u
r
r
z
u
z
r
1
r
u
z
u
z
z
_
_
+
_
_
u
r
r
r
u
z
r
1
r
u
r
1
r
u
+
u
r
r
1
r
u
z
u
r
z
z
u
z
z
_
_
_
_
=
_
u
r
r
+
u
z
z
+
u
r
r
_
I +
_
_
2
u
r
r
0
u
r
z
+
u
z
r
0 0 0
u
z
r
+
u
r
z
0 2
u
z
z
_
_
+
_
_
0 0 0
0 2
u
r
r
0
0 0 0
_
_
. (4.11.3)
Therefore, we have
div = div
+s, (4.11.4)
where
div
= div
_
rr
rz
zr
zz
_
_
=
_
_
1
r
(r
rr
)
r
+
rz
z
0
1
r
(r
zr
)
r
+
zz
z
_
_
, s =
_
2u
r
r
2
r
_
u
r
r
+
u
z
z
+
u
r
r
_
0
0
_
_
, (4.11.5)
=
_
u
r
r
+
u
z
z
+
u
r
r
_
I +
_
_
2
u
r
r
0
u
r
z
+
u
z
r
0 0 0
u
z
r
+
u
r
z
0 2
u
z
z
_
_
. (4.11.6)
4.11. AXISYMMETRIC NAVIER-STOKES EQUATIONS 141
It is clear that the -component is zero. Hence, we have only two equations: r- and z-components. Finally, the viscous
work term in the energy equation can be written as
v =
_
u
r
r
+
u
z
z
+
u
r
r
_
I
_
_
u
r
u
z
_
_
+
_
_
2
u
r
r
0
u
r
z
+
u
z
r
0 0 0
u
z
r
+
u
r
z
0 2
u
z
z
_
_
_
_
u
r
u
z
_
_
+
_
_
0 0 0
0 2
u
r
r
0
0 0 0
_
_
_
_
u
r
u
z
_
_
.
(4.11.7)
Clearly, the -component is zero, and we have v =
v, and thus
div(v) = div(
v). (4.11.8)
We are now ready to write the viscous terms in the two-dimensional form. Switch the z-equation and the r-equation
just for convenience, and replace (z, r) by (x, y), (q
z
, q
r
) by (q
x
, q
y
), and (u
z
, u
r
) by (u, v) to write
div q =
1
y
_
(yq
x
)
x
+
(yq
y
)
y
_
, (4.11.9)
div =
1
y
_
_
(y
xx
)
x
+
(y
xy
)
y
(y
yx
)
x
+
(y
yy
)
y
_
_
+
_
_
0
2v
y
2
y
_
u
x
+
v
y
+
v
y
_
_
_
, (4.11.10)
=
_
xx
xy
yx
yy
_
=
_
u
x
+
v
y
+
v
y
_
I +
_
_
2
u
x
u
y
+
v
x
v
x
+
u
y
2
v
y
_
_
, (4.11.11)
div(v) =
1
y
_
_
(y
xx
u)
x
+
(y
xy
v)
y
(y
yx
u)
x
+
(y
yy
v)
y
_
_
, (4.11.12)
where v and q are now two-dimensional vectors, and is a two-dimensional tensor. It is important to note that the
velocity divergence in the viscous stress tensor (4.11.11) has an extra term v/y, and the same for the second component
of the source term in the viscous-stress divergence (4.11.10). Finally, the axisymmetric Navier-Stokes equations can
be written as
(yU)
t
+ (yF)
x
+ (yG)
y
= (yF
v
)
x
+ (yG
v
)
y
+S, (4.11.13)
where
U =
_
u
v
E
_
_
, F =
_
_
u
u
2
+p
uv
uH
_
_
, G =
_
_
v
uv
v
2
+p
vH
_
_
, (4.11.14)
F
v
=
_
_
0
xx
yx
xx
u +
xy
v q
x
_
_
, G
v
=
_
_
0
xy
yy
yx
u +
yy
v q
y
_
_
, S =
_
_
0
0
p
2v
y
_
u
x
+
v
y
+
v
y
_
0
_
_
.(4.11.15)
142 CHAPTER 4. NAVIER-STOKES EQUATIONS
As in the inviscid case, following Ref.[173], we can express the axisymmetric form and the two-dimensional form in a
single form with a logical switch, = 0 for the two-dimensional form and = 1 for the axisymmetric form:
(
a
U)
t
+ (
a
F)
x
+ (
a
G)
y
= (
a
F
v
)
x
+ (
a
G
v
)
y
+ S, (4.11.16)
where
a
= (1 ) + y. (4.11.17)
This is really nice. If you have a two-dimensional Navier-Stokes code, you almost have an axisymmetric Navier-Stokes
code, which is essentially a three-dimensional CFD code. I really like it very much.
4.12 3D Normal Viscous Flux and Jacobian
Often, we consider the inviscid and viscous uxes separately. The viscous ux of the Navier-Stokes equations,
denoted by F
v
, G
v
, H
v
, are given by
F
v
=
_
_
0
xx
yx
zx
vx
+q
x
_
_
, G
v
=
_
_
0
xy
yy
zy
vy
+q
y
_
_
, H
v
=
_
_
0
xz
yz
zz
vz
+q
z
_
_
, (4.12.1)
where
vx
=
xx
u +
xy
v +
xz
w, (4.12.2)
vy
=
yx
u +
yy
v +
yz
w, (4.12.3)
vz
=
zx
u +
zy
v +
zz
w. (4.12.4)
It is so much simpler than the inviscid ux. I like it very much. In CFD, we often deal with the ux projected along
a particular direction, e.g., normal to a control volume boundary. Let n = [n
x
, n
y
, n
z
]
t
be a unit vector. Then, the
viscous ux along n, denoted by F
n
, is given by
F
v
n
= F
v
n
x
+G
v
n
y
+H
v
n
z
=
_
_
0
nx
ny
nz
nv
+q
n
_
_
, (4.12.5)
where
nx
=
xx
n
x
+
xy
n
y
+
xz
n
z
, (4.12.6)
ny
=
yx
n
x
+
yy
n
y
+
yz
n
z
, (4.12.7)
nz
=
zx
n
x
+
zy
n
y
+
zz
n
z
, (4.12.8)
nv
=
nx
u +
ny
v +
nz
w, (4.12.9)
q
n
= q
x
n
x
+q
y
n
y
+q
z
n
z
. (4.12.10)
All quantities above have been assumed to have been evaluated at some average state, which is indicated by the over
bar. For example, in nite-volume methods, the normal viscous ux F
v
n
needs to be computed at control-volume
boundary faces, and n denotes the normal vector. I like the normal viscous ux because all we need for the heat ux
is the normal gradient:
q
n
= n grad T =
Pr( 1)
n grad(a
2
) =
Pr( 1)
n grad
_
p
_
, (4.12.11)
4.12. 3D NORMAL VISCOUS FLUX AND JACOBIAN 143
which we may write simply as
q
n
=
T
n
=
Pr( 1)
a
2
n
=
Pr( 1)
n
_
p
_
=
Pr( 1)
_
p
n
p
n
_
, (4.12.12)
where /n denotes the derivative along n, i.e., /n = n
x
/x + n
y
/y + n
z
/z. So, we need only one
derivative, i.e., the one along n. This is nice. I like it. On the other hand, the viscous stresses require the derivatives
in the other directions. To see this, recall that the viscous stress tensor is given by
= (div v)I +
_
grad v + (grad v)
t
. (4.12.13)
Then, the viscous stress tensor projected along n is given by
n =
_
nx
ny
nz
_
_ = (div v)n +
_
(grad v)n + (grad v)
t
n
= (div v)n +
_
v
n
+ grad u
n
_
=
_
_
u
x
+
v
y
+
w
z
_
n
x
+
_
u
n
+
u
n
x
_
_
u
x
+
v
y
+
w
z
_
n
y
+
_
v
n
+
u
n
y
_
_
u
x
+
v
y
+
w
z
_
n
z
+
_
w
n
+
u
n
z
_
_
_
, (4.12.14)
where u
n
= un
x
+ v n
y
+ wn
z
. It is clear from the above that it cannot be written in terms of only /n, and
therefore we need all derivatives of the velocity components.
It is often necessary to dierentiate the viscous ux, for example, when implicit schemes are constructed. Consider
the normal viscous Jacobian:
F
v
n
U
=
F
v
n
W
W
U
=
_
_
0
t
nx
W
ny
W
nz
W
nv
W
+
q
n
W
_
_
_
_
1 0 0 0 0
0 0 0
0
1
0 0
0 0
1
0
1
2
q
2
( 1)u ( 1)v ( 1)w 1
_
_
,(4.12.15)
where U = [, u, v, w, E]
t
and W = [, u, v, w, p]
t
. Let us assume Stokes hypothesis and write the viscous
stresses as
nx
= s
nx
,
ny
= s
ny
,
nz
= s
nz
,
nv
= s
nv
, (4.12.16)
where s
nv
= s
nx
u +s
ny
v +s
nz
w, and
s
nx
= s
xx
n
x
+s
xy
n
y
+s
xz
n
z
, s
ny
= s
yx
n
x
+s
yy
n
y
+s
yz
n
z
, s
nz
= s
zx
n
x
+s
zy
n
y
+s
zz
n
z
, (4.12.17)
s
xx
=
2
3
_
2
u
x
v
y
w
z
_
, s
yy
=
2
3
_
2
v
y
u
x
w
z
_
, s
zz
=
2
3
_
2
w
z
u
x
v
y
_
, (4.12.18)
s
xy
= s
yx
=
u
y
+
v
x
, s
xz
= s
zx
=
w
x
+
u
z
, s
yz
= s
zy
=
v
z
+
w
y
. (4.12.19)
144 CHAPTER 4. NAVIER-STOKES EQUATIONS
The derivative of
nx
can be obtained as follows.
nx
W
=
W
s
nx
+
s
nx
W
=
_
s
nx
,
s
nx
u
,
s
nx
v
,
s
nx
w
,
p
s
nx
_
=
_
T
T
s
nx
u
_
1
3
u
x
n
x
+
u
n
_
v
_
2
3
v
y
n
x
+
v
x
n
y
_
w
_
2
3
w
z
n
x
+
w
x
n
z
_
T
T
p
s
nx
_
_
t
. (4.12.20)
Similarly, we have
ny
W
=
_
T
T
s
ny
u
_
2
3
u
x
n
y
+
u
y
n
x
_
u
_
1
3
v
y
n
y
+
v
n
_
w
_
2
3
w
z
n
y
+
w
y
n
z
_
T
T
p
s
ny
_
_
t
,
nz
W
=
_
T
T
s
nz
u
_
2
3
u
x
n
z
+
u
z
n
x
_
v
_
2
3
v
y
n
x
+
v
z
n
y
_
w
_
1
3
w
z
n
z
+
w
n
_
T
T
p
s
nz
_
_
t
, (4.12.21)
nv
W
=
W
s
nv
+
s
nv
W
=
_
s
nv
,
u
u
nx
+u
nx
u
,
v
v
ny
+v
ny
v
,
w
w
nz
+w
nz
w
,
p
s
nv
_
=
_
T
T
s
nv
s
nx
u
u
+
_
un
x
2
3
u
n
_
u
u
x
+u
u
u
n
s
ny
v
v
+
_
vn
y
2
3
u
n
_
v
v
y
+v
v
v
n
s
nz
w
w
+
_
wn
z
2
3
u
n
_
w
w
z
+w
w
w
n
T
T
p
s
nv
_
_
t
. (4.12.22)
4.13. EXPANDED FORM OF VISCOUS TERMS 145
On the other hand, the heat ux can be dierentiated as follows.
q
n
W
=
1
Pr( 1)
_
T
n
+
T
n
, 0, 0, 0,
p
T
n
+
p
T
n
_
=
1
Pr( 1)
_
T
T
T
n
+
T
n
, 0, 0, 0,
T
T
p
T
n
+
p
T
n
_
. (4.12.23)
For Sutherlands law (4.4.4), the derivative /T is given by
T
=
T
_
0
T
0
+C
T +C
_
T
T
0
_
3
2
_
T
T
=
0
2(T +C/T
0
)
_
1 + 3
C
T
0
T
_
T
T
. (4.12.24)
That is all we need to compute the viscous Jacobian (4.12.15). I like it because it is a general expression; the specic
form depends on how the interface values, i.e., the quantities and derivatives with the over bar, are dened. For
example, if the temperature T is evaluated as the average of two states, T
L
and T
R
, and the Jacobian is sought with
respect to the left state, then
T
T
L
=
T
L
_
T
L
+T
R
2
_
=
1
2
. (4.12.25)
Also, if the normal derivative is evaluated by the dierence between the two states,
u
n
=
u
R
u
L
s
, (4.12.26)
where s is the distance between the two states, then the derivative with respect to the left state is given by
u
L
u
n
=
1
s
. (4.12.27)
It really depends on the discretization method, but the derivatives not in the normal direction such as
u
u
x
can often
become a source of diculty. For example, on unstructured grids, these derivatives (as well as a part of the normal
gradient) lead to a non-compact stencil [49] and the Jacobian matrix can become prohibitively large. In such a case,
these derivatives are often ignored in the construction of the Jacobian matrix. The corresponding discretization is then
not consistent, but it is perhaps OK for the Jacobian matrix as it acts only as a driver in implicit solvers. As long as
the residual (i.e., the right hand side of the implicit formulation) is constructed by a consistent scheme, accuracy is
obtained at convergence although the convergence may not be optimal [111, 157].
4.13 Expanded Form of Viscous Terms
Sometimes, it is convenient to expand the Navier-Stokes equations in terms of the derivatives of the primitive
variables. As we have shown the expanded form already for the Euler equations in Section 3.6.1, we here focus on the
viscous part only:
F
v
x
+
G
v
y
+
H
v
z
=
_
xx
x
xy
y
xz
z
yx
x
yy
y
yz
z
zx
x
zy
y
zz
z
vx
x
vy
y
vz
z
+
q
x
x
+
q
y
y
+
q
z
z
_
_
, (4.13.1)
146 CHAPTER 4. NAVIER-STOKES EQUATIONS
where
xx
x
= 2
x
u
x
+ 2
2
u
x
2
+
x
_
u
x
+
v
y
+
w
z
_
+
x
_
u
x
+
v
y
+
w
z
_
, (4.13.2)
yy
y
= 2
y
v
y
+ 2
2
v
y
2
+
y
_
u
x
+
v
y
+
w
z
_
+
y
_
u
x
+
v
y
+
w
z
_
, (4.13.3)
zz
z
= 2
z
w
z
+ 2
2
w
z
2
+
z
_
u
x
+
v
y
+
w
z
_
+
z
_
u
x
+
v
y
+
w
z
_
, (4.13.4)
xy
x
=
yx
x
=
x
_
u
y
+
v
x
_
+
x
_
u
y
+
v
x
_
, (4.13.5)
xy
y
=
yx
y
=
y
_
u
y
+
v
x
_
+
y
_
u
y
+
v
x
_
, (4.13.6)
xz
x
=
zx
x
=
x
_
w
x
+
u
z
_
+
x
_
w
x
+
u
z
_
, (4.13.7)
xz
z
=
zx
z
=
z
_
w
x
+
u
z
_
+
z
_
w
x
+
u
z
_
, (4.13.8)
yz
y
=
zy
y
=
y
_
v
z
+
w
y
_
+
y
_
v
z
+
w
y
_
, (4.13.9)
yz
z
=
zy
z
=
z
_
v
z
+
w
y
_
+
z
_
v
z
+
w
y
_
, (4.13.10)
vx
x
= u
xx
x
+v
xy
x
+w
xz
x
+
xx
u
x
+
xy
v
x
+
xz
w
x
, (4.13.11)
vy
y
= u
yx
y
+v
yy
y
+w
yz
y
+
yx
u
y
+
yy
v
y
+
yz
w
y
, (4.13.12)
vz
z
= u
zx
z
+v
zy
z
+w
zz
z
+
zx
u
z
+
zy
v
z
+
zz
w
z
, (4.13.13)
q
x
x
=
Pr( 1)
_
x
_
1
p
x
p
x
_
+
_
x
p
x
+
1
2
p
x
2
+
p
x
p
x
2
__
, (4.13.14)
q
y
y
=
Pr( 1)
_
y
_
1
p
y
p
y
_
+
_
y
p
y
+
1
2
p
y
2
+
p
y
p
y
2
__
, (4.13.15)
q
z
z
=
Pr( 1)
_
z
_
1
p
z
p
z
_
+
_
z
p
z
+
1
2
p
z
2
+
p
z
p
z
2
__
. (4.13.16)
4.14. NONDIMENSIONALIZATION 147
The derivatives of the viscosity can be computed for Sutherlands law (4.4.4) by
x
=
T
_
0
T
0
+C
T +C
_
T
T
0
_3
2
_
T
x
=
0
2(T +C/T
0
)
_
1 + 3
C
T
0
T
_
T
x
, (4.13.17)
y
=
T
_
0
T
0
+C
T +C
_
T
T
0
_3
2
_
T
y
=
0
2(T +C/T
0
)
_
1 + 3
C
T
0
T
_
T
y
, (4.13.18)
z
=
T
_
0
T
0
+C
T +C
_
T
T
0
_3
2
_
T
z
=
0
2(T +C/T
0
)
_
1 + 3
C
T
0
T
_
T
z
. (4.13.19)
The derivatives of can be easily computed from the above by Stokes hypothesis, =
2
3
. The derivatives of the
temperature can be computed from the derivatives of the density and the pressure. Again, I like this expanded form
because it can be useful in numerically computing the source term arising from the method of manufactured solutions.
Note that the form of the derivatives depends on the nondimensionalization for the viscosity and the temperature,
which is the subject of the next section.
4.14 Nondimensionalization
4.14.1 Nondimensionalized Navier-Stokes Equations
It is always nice to nondimensionalize the governing equations. For the Navier-Stokes equations, in addition to the
quantities introduced for the Euler equations in Section 3.11:
r
, p
=
p
r
V
2
r
, e
=
e
V
2
r
, v
=
v
V
r
, x
=
x
L
, t
=
t
L/V
r
, (4.14.1)
we need to introduce nondimensionalized the temperature and the viscosity,
T
=
T
T
r
,
r
. (4.14.2)
Substituting these quantities into the Navier-Stokes equations, (4.1.1), (4.1.2), (4.1.3), we obtain
+ div
) = 0, (4.14.3)
(
)
t
+ div
) + grad
= div
, (4.14.4)
(
)
t
+ div
) = div
) div
, (4.14.5)
where E
= e
+
v
2
2
, H
= E
+
p
. These are exactly the same as in the dimensional version, but the viscous stress
tensor and the heat ux vector have some dimensionless parameters,
=
r
r
V
r
L
2
3
(div
)I + grad
+ (grad
)
t
_
, (4.14.6)
q
=
RT
r
r
V
3
r
L( 1)Pr
grad
. (4.14.7)
Also, note that the equation of state (3.1.1) is now written as
p
=
_
RT
r
V
2
r
_
. (4.14.8)
148 CHAPTER 4. NAVIER-STOKES EQUATIONS
Therefore, a specic form of the equation of state as well as the viscous and heat uxes depend on the choice of the
reference values:
r
, V
r
, T
r
, and
r
. Note that because of the forms of the nondimensional pressure and specic
internal energy, the equation of state for the calorically perfect gas holds exactly as in the dimensional form:
p
= ( 1)
, (4.14.9)
for any choice of the reference states (see Section 3.11), but the one with the temperature, i.e., p = RT, depends on
the reference states. Note that the governing equations are virtually in the same form as the dimensional form for an
appropriately dened scaled viscosity as shown below.
4.14.2 Free Stream Values
We may use the free stream values for the reference quantities,
r
=
, V
r
= V
, T
r
= T
,
r
=
, p
=
p
V
2
, e
=
e
V
2
, v
=
v
V
, x
=
x
L
, t
=
t
L/V
, T
=
T
T
. (4.14.10)
Then, the equation of state is given by
p
M
2
, (4.14.11)
and the viscous stress and the heat ux are given by,
2
3
(div
)I + grad
+ (grad
)
t
_
, (4.14.12)
q
( 1)M
2
Pr
grad
( 1)Pr
grad
_
p
_
, (4.14.13)
where the scaled viscosity
is dened by
Re
, Re
. (4.14.14)
The free stream values are set as follows:
= 1, p
=
1
M
2
, v
=
v
= n
, e
=
1
( 1)M
2
, T
= 1,
= 1. (4.14.15)
The nondimensionalized viscosity coecient
is given by
=
0
T
0
+C
T +C
_
T
T
0
_3
2
_
0
T
0
+C
T
+C
_
T
T
0
_3
2
_
1
=
1 +C/T
+C/T
(T
)
3
2
, (4.14.16)
with C = 110.5 [K] for air. The derivative of the nondimensionalized viscosity coecient is given by
=
1
2
+C/T
_
1 + 3
C/T
_
. (4.14.17)
Note that in order to set a free stream inow condition, we need to specify not only the Mach number M
and the
free stream ow direction n
, which are required for the Euler equations (see Section 3.11), but also the free stream
temperature T
, e.g., T
= 293.15[K] for the sea-level condition. Of course, the free stream Reynolds number Re
must be specied, and also the Prandtl number Pr needs to be given (Pr = 0.72 for the sea-level air).
The Reynolds number Re
must be set carefully in relation to the reference length scale L because it is dened
based on L as in Equation (4.14.14). It is important to note that if you solve the nondimensionalized system on a
computational grid, then Re
is the Reynolds number based on the part having the length of 1 in the grid (i.e., in
the nondimensionalized coordinates, x
). So, for example, if you want to simulate a ow around a cylinder where the
Reynolds number based on the diameter of the cylinder is 100, then you generate a grid around the cylinder with the
diameter 1 and set Re
= 100. On the other hand, if you are given a grid where the diameter of the cylinder is 2 in
the grid, then Re
is the Reynolds number based on the radius of the cylinder (which is 1 in the grid). Then, you
need to set Re
= 50 to simulate a ow with the Reynolds number 100 based on the diameter. Or if the diameter is
0.1 in the grid, then you need to set Re
, we may use the free stream speed of sound for the reference velocity, V
r
= a
, p
=
p
a
2
, e
=
e
a
2
, v
=
v
a
, x
=
x
L
, t
=
t
L/a
, (4.14.18)
T
=
T
T
. (4.14.19)
Then, the equation of state is given by
p
, (4.14.20)
and the viscous stress and the heat ux are given by,
2
3
(div
)I + grad
+ (grad
)
t
_
, (4.14.21)
q
( 1)Pr
grad
( 1)Pr
grad
_
p
_
, (4.14.22)
where the scaled viscosity
is dened by
=
M
Re
. (4.14.23)
The free stream values are set as follows:
= 1, p
=
1
, v
=
v
= M
, e
=
1
( 1)
, T
= 1,
= 1. (4.14.24)
The nondimensionalized viscosity coecient
(not to mention Re
0
=
_
1 +
1
2
M
2
_
1
1
, p
=
p
0
a
2
0
=
1
_
1 +
1
2
M
2
_
1
, (4.14.25)
e
=
e
a
2
0
, v
=
v
a
0
, x
=
x
L
, t
=
t
L/a
0
, (4.14.26)
T
=
T
T
0
=
_
1 +
1
2
M
2
_
1
,
. (4.14.27)
Then, the equation of state is given by
p
, (4.14.28)
and the viscous stress and the heat ux are given by,
2
3
(div
)I + grad
+ (grad
)
t
_
, (4.14.29)
q
( 1)Pr
grad
( 1)Pr
grad
_
p
_
, (4.14.30)
150 CHAPTER 4. NAVIER-STOKES EQUATIONS
where the scaled viscosity
is dened by
=
m
Re
, m
=
_
1 +
1
2
M
2
+1
2(1)
. (4.14.31)
In this case, the free stream values are set as follows:
=
_
1 +
1
2
M
2
1
1
, p
=
1
_
1 +
1
2
M
2
1
, (4.14.32)
e
=
T
( 1)
, v
=
v
a
0
=
v
= M
, (4.14.33)
T
=
_
1 +
1
2
M
2
_
1
,
= 1. (4.14.34)
The nondimensionalized viscosity coecient
,
and the free stream temperature T
, in addition to Re
and Pr.
4.15 Reduced Navier-Stokes Equations
I like reduced Navier-Stokes equations such as thin-layer, parabolized, or conical Navier-Stokes equations. They are ap-
proximated forms of the full Navier-Stokes equations and hence simpler and cheaper to solve than the full Navier-Stokes
equations. If you like them too, see Refs.[51, 155].
4.16 Quasi-Linear Form of the Navier-Stokes Equations
Often for various purposes (e.g., analysis, scheme development, etc.), it is nice to write the Navier-Stokes equa-
tions in the quasi-linear form,
V
t
+A
v
V
x
+B
v
V
y
+C
v
V
z
= D
v
V
xx
+E
v
V
yy
+F
v
V
zz
+G
v
V
xy
+H
v
V
yz
+I
v
V
zx
, (4.16.1)
for some set of variables V, where the superscript v merely indicates that the coecient matrices are associated with
V. To be able to do so, it is necessary to assume that the viscosity and the heat conductivity are constant. Also,
it is necessary to ignore non-linear terms which involve a product of derivatives. It will be, therefore, no longer the
full Navier-Stokes system. However, recall that not all terms are necessary to identify the type of partial dierential
equations as discussed in Section 1.17. In fact, the quasi-linear form of the Navier-Stokes system retains the essential
feature of the full Navier-Stokes system; this is the form modeled by the scalar advection-diusion or the viscous
Burgers equation. For example, a symmetric form of the quasi-linear Navier-Stokes system has been successfully used
in the development of convergence acceleration techniques [1, 72, 77, 167].
Consider the primitive form of the Navier-Stokes equations,
D
Dt
+ div v = 0, (4.16.2)
Dv
Dt
+ grad p = div , (4.16.3)
Dp
Dt
+p div v = ( 1) [ : grad v div q] . (4.16.4)
Obviously, the dissipation term ( grad v) involves products of derivatives and needs to be ignored. Also, a part of
the heat transfer term needs to be ignored as well. To see this, we rst express T in terms of p and ,
T =
1
R
p
=
( 1)c
p
p
=
1
(c
p
/)( 1)
p
, (4.16.5)
4.16. QUASI-LINEAR FORM OF THE NAVIER-STOKES EQUATIONS 151
thus
T =
Pr( 1)
p
. (4.16.6)
Then, the heat transfer term can be written as
div q = div (gradT)
=
Pr( 1)
div grad
_
p
_
=
Pr( 1)
div
_
1
gradp
p
2
grad
_
=
Pr( 1)
_
2
grad gradp +
1
2
_
grad
p
2
div grad
_
. (4.16.7)
Hence, the rst and the third terms involve products of derivatives and need to be ignored. The pressure equation
then becomes
Dp
Dt
+p div v =
Pr
_
div gradp
p
div grad
_
. (4.16.8)
In this sense, therefore, the quasi-linear Navier-Stokes equations may be called linearized Navier-Stokes equations, or to
be precise the Navier-Stokes equations with the linearized energy equation. Clearly, the continuity equation is already in
the quasi-linear form. The momentum equation is also in the quasi-linear form, which can be clearly seen by expanding
the viscous term as follows,
div = div
_
(div v)I +grad v +(grad v)
t
grad p =
1
3
grad (div v) + div grad v, (4.16.11)
Dp
Dt
+p div v =
Pr
div gradp
a
2
Pr
div grad, (4.16.12)
are now ready to be written out in the quasi-linear form (4.16.1) in one, two, and three dimensions.
152 CHAPTER 4. NAVIER-STOKES EQUATIONS
4.16.1 1D Quasi-Linear Navier-Stokes Equations
Primitive Form
The primitive form in one dimension is given by
t
+u
x
+u
x
= 0, (4.16.13)
u
t
+uu
x
+
1
p
x
=
4
3
u
xx
, (4.16.14)
p
t
+up
x
+pu
x
=
Pr
p
xx
a
2
Pr
xx
. (4.16.15)
This can be written in the form:
W
t
+A
w
W
x
= D
w
W
xx
, (4.16.16)
where
W =
_
u
p
_
_, A
w
=
_
_
u 0
0 u 1/
0 p u
_
_, D
w
=
_
_
0 0 0
0 4/3 0
a
2
/Pr 0 /Pr
_
_. (4.16.17)
Symmetric Form
I like the following variables,
dV =
_
a d
, du,
a dT
T
( 1)
_
t
, (4.16.18)
where a =
, since the quasi-linear Navier-Stokes system is symmetrized by these variables [1]. Naturally, these
variables are called symmetrizing variables, and often called parabolic symmetrizing variables to distinguish them
from the symmetrizing variables for the Euler equations (3.5.28) which are in contrast called hyperbolic symmetrizing
variables.
The primitive variables W and the symmetrizing variables V are transformed into each other as follows.
dV =
V
W
dW, dW =
W
V
dV, (4.16.19)
where
V
W
=
_
_
a
0 0
0 1 0
( 1)
0
a
p
( 1)
_
_
, (4.16.20)
W
V
=
_
V
W
_
1
=
_
a
0 0
0 1 0
a
0
a
_
. (4.16.21)
The primitive form is now transformed, with
V
W
multiplied from the left, into the symmetric form:
V
t
+A
v
V
x
= D
v
V
xx
, (4.16.22)
4.16. QUASI-LINEAR FORM OF THE NAVIER-STOKES EQUATIONS 153
where
A
v
=
V
W
A
w
W
V
=
_
_
u
a
0
a
u
a
0
a
u
_
_
, D
v
=
V
W
D
w
W
V
=
_
_
0 0 0
0
4
3
0
0 0
Pr
_
_
. (4.16.23)
It is clear from these matrices that the system is indeed symmetric. Note that in transforming the viscous term we
have ignored again non-linear terms involving products of derivatives:
V
W
D
w
W
xx
=
_
V
W
D
w
W
V
_
V
W
W
xx
= D
v
_
x
_
V
W
W
x
_
x
_
V
W
_
W
x
_
D
v
V
xx
. (4.16.24)
Or equivalently, it can be thought of as the transformation matrix
V
W
being frozen at a certain constant state (so
that its derivative is zero).
4.16.2 2D Quasi-Linear Navier-Stokes Equations
Primitive Form
The primitive form in two dimensions is given by
t
+u
x
+u
x
+v
y
+v
y
= 0, (4.16.25)
u
t
+uu
x
+vu
y
+
1
p
x
=
4
3
u
xx
+
3
v
yx
+u
yy
, (4.16.26)
v
t
+uv
x
+vv
y
+
1
p
y
= v
xx
+
3
u
xy
+
4
3
v
yy
, (4.16.27)
p
t
+up
x
+vp
y
+pu
x
+pv
y
=
Pr
(p
xx
+p
yy
)
a
2
Pr
(
xx
+
yy
). (4.16.28)
This can be written in the form:
W
t
+A
w
W
x
+B
w
W
y
= D
w
W
xx
+G
w
W
xy
+E
w
W
yy
, (4.16.29)
where W = [, u, v, p]
t
,
A
w
=
_
_
u 0 0
0 u 0 1/
0 0 u 0
0 p 0 u
_
_
, B
w
=
_
_
v 0 0
0 v 0 0
0 0 v 1/
0 0 p v
_
_
, (4.16.30)
D
w
=
_
_
0 0 0 0
0 4/3 0 0
0 0 0
a
2
/Pr 0 0 /Pr
_
_
, G
w
=
_
_
0 0 0 0
0 0 /3 0
0 /3 0 0
0 0 0 0
_
_
, (4.16.31)
E
w
=
_
_
0 0 0 0
0 0 0
0 0 4/3 0
a
2
/Pr 0 0 /Pr
_
_
. (4.16.32)
154 CHAPTER 4. NAVIER-STOKES EQUATIONS
Symmetric Form
The symmetrizing variables,
dV =
_
a d
, du, dv,
a dT
T
( 1)
_
t
, (4.16.33)
and the primitive variables W = [, u, v, p]
t
are transformed to each other by the following matrices:
V
W
=
_
_
a
0 0 0
0 1 0 0
0 0 1 0
( 1)
0 0
a
p
( 1)
_
_
, (4.16.34)
W
V
=
_
a
0 0 0
0 1 0 0
0 0 1 0
a
0 0
a
_
. (4.16.35)
The primitive form is now transformed, with
V
W
multiplied from the left, into the symmetric form:
V
t
+A
v
V
x
+B
v
V
y
= D
v
V
xx
+G
v
V
xy
+E
v
V
yy
, (4.16.36)
where
A
v
=
_
_
u
a
0 0
a
u 0 a
0 0 u 0
0 a
0 u
_
_
, B
v
=
_
_
v 0
a
0
0 v 0 0
a
0 v a
0 0 a
v
_
_
, (4.16.37)
D
v
=
_
_
0 0 0 0
0
4
3
0 0
0 0 0
0 0 0
Pr
_
_
, G
v
=
_
_
0 0 0 0
0 0
3
0
0
3
0 0
0 0 0 0
_
_
, (4.16.38)
E
v
=
_
_
0 0 0 0
0 0 0
0 0
4
3
0
0 0 0
Pr
_
_
. (4.16.39)
4.16. QUASI-LINEAR FORM OF THE NAVIER-STOKES EQUATIONS 155
4.16.3 3D Quasi-Linear Navier-Stokes Equations
Primitive Form
The primitive form in three dimensions is given by
t
+u
x
+u
x
+v
y
+v
y
+w
z
+w
z
= 0, (4.16.40)
u
t
+uu
x
+vu
y
+wu
z
+
1
p
x
=
4
3
u
xx
+
3
v
yx
+
3
w
zx
+u
yy
+u
zz
, (4.16.41)
v
t
+uv
x
+vv
y
+wv
z
+
1
p
y
= v
xx
+
3
u
xy
+
3
w
zy
+
4
3
v
yy
+v
zz
, (4.16.42)
w
t
+uw
x
+vw
y
+ww
z
+
1
p
z
= w
xx
+
3
u
xz
+
3
v
yz
+w
yy
+
4
3
w
zz
, (4.16.43)
p
t
+up
x
+vp
y
+wp
z
+pu
x
+pv
y
+pw
z
=
Pr
(p
xx
+p
yy
+p
zz
)
a
2
Pr
(
xx
+
yy
+
zz
). (4.16.44)
This can be written in the form:
W
t
+A
w
W
x
+B
w
W
y
+C
w
W
z
= D
w
W
xx
+E
w
W
yy
+F
w
W
zz
+G
w
W
xy
+H
w
W
yz
+I
w
W
zx
, (4.16.45)
where W = [, u, v, w, p]
t
,
A
w
=
_
_
u 0 0 0
0 u 0 0 1/
0 0 u 0 0
0 0 0 u 0
0 p 0 0 u
_
_
, B
w
=
_
_
v 0 0 0
0 v 0 0 0
0 0 v 0 1/
0 0 0 v 0
0 0 p 0 v
_
_
, C
w
=
_
_
w 0 0 0
0 w 0 0 0
0 0 w 0 0
0 0 0 w 1/
0 0 0 p w
_
_
, (4.16.46)
D
w
=
_
_
0 0 0 0 0
0 4/3 0 0 0
0 0 0 0
0 0 0 0
a
2
/Pr 0 0 0 /Pr
_
_
, E
w
=
_
_
0 0 0 0 0
0 0 0 0
0 0 4/3 0 0
0 0 0 0
a
2
/Pr 0 0 0 /Pr
_
_
, (4.16.47)
F
w
=
_
_
0 0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 4/3 0
a
2
/Pr 0 0 0 /Pr
_
_
, G
w
=
_
_
0 0 0 0 0
0 0 /3 0 0
0 /3 0 0 0
0 0 0 0 0
0 0 0 0 0
_
_
, (4.16.48)
H
w
=
_
_
0 0 0 0 0
0 0 0 0 0
0 0 0 /3 0
0 0 /3 0 0
0 0 0 0 0
_
_
, I
w
=
_
_
0 0 0 0 0
0 0 0 /3 0
0 0 0 0 0
0 /3 0 0 0
0 0 0 0 0
_
_
. (4.16.49)
Symmetric Form
The symmetrizing variables,
dV =
_
a d
( 1)
_
t
, (4.16.50)
156 CHAPTER 4. NAVIER-STOKES EQUATIONS
and the primitive variables W = [, u, v, w, p]
t
are transformed to each other by the following matrices:
V
W
=
_
_
a
0 0 0 0
0 1 0 0 0
0 0 1 0 0
0 0 0 1 0
( 1)
0 0 0
a
p
( 1)
_
_
, (4.16.51)
W
V
=
_
a
0 0 0 0
0 1 0 0 0
0 0 1 0 0
0 0 0 1 0
a
0 0 0
a
_
. (4.16.52)
The primitive form is now transformed, with
V
W
multiplied from the left, into the symmetric form:
V
t
+A
v
V
x
+B
v
V
y
+C
v
V
z
= D
v
V
xx
+E
v
V
yy
+F
v
V
zz
+G
v
V
xy
+H
v
V
yz
+I
v
V
zx
, (4.16.53)
where
A
v
=
_
_
u
a
0 0 0
a
u 0 0 a
0 0 u 0 0
0 0 0 u 0
0 a
0 0 u
_
_
, B
v
=
_
_
v 0
a
0 0
0 v 0 0 0
a
0 v 0 a
0 0 0 v 0
0 0 a
0 v
_
_
, (4.16.54)
C
v
=
_
_
w 0 0
a
0
0 w 0 0 0
0 0 w 0 0
a
0 0 w a
0 0 0 a
w
_
_
, D
v
=
_
_
0 0 0 0 0
0
4
3
0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
Pr
_
_
, (4.16.55)
E
v
=
_
_
0 0 0 0 0
0 0 0 0
0 0
4
3
0 0
0 0 0 0
0 0 0 0
Pr
_
_
, F
v
=
_
_
0 0 0 0 0
0 0 0 0
0 0 0 0
0 0 0
4
3
0
0 0 0 0
Pr
_
_
, (4.16.56)
4.17. INCOMPRESSIBLE NAVIER-STOKES EQUATIONS 157
G
v
=
_
_
0 0 0 0 0
0 0
3
0 0
0
3
0 0 0
0 0 0 0 0
0 0 0 0 0
_
_
, H
v
=
_
_
0 0 0 0 0
0 0 0 0 0
0 0 0
3
0
0 0
3
0 0
0 0 0 0 0
_
_
, (4.16.57)
I
v
=
_
_
0 0 0 0 0
0 0 0
3
0
0 0 0 0 0
0
3
0 0 0
0 0 0 0 0
_
_
. (4.16.58)
4.17 Incompressible Navier-Stokes Equations
As I said before, I like incompressible uids: constant density or small Mach number. So, of course, I like the incom-
pressible Navier-Stokes equations, which are obtained from the full Navier-Stokes equations by assuming = constant.
Conservative Form of the Incompressible Navier-Stokes Equations
div v = 0, (4.17.1)
t
(v) + div(vv) + grad p div = 0, (4.17.2)
c
v
DT
Dt
div (k grad T) = 0. (4.17.3)
where
=
_
grad v + (grad v)
t
, (4.17.4)
= : grad v =
_
grad v + (grad v)
t
: grad v. (4.17.5)
Primitive Form of the Incompressible Navier-Stokes Equations
div v = 0, (4.17.6)
Dv
Dt
+ grad p div = 0, (4.17.7)
c
v
DT
Dt
div (k grad T) = 0. (4.17.8)
If the viscosity coecient is constant, the viscous term can be written as the Laplacian,
div = div
_
_
grad v + (grad v)
t
_
= div grad v +div(grad v)
t
= div grad v +grad(div v)
= div grad v, (4.17.9)
158 CHAPTER 4. NAVIER-STOKES EQUATIONS
which is nice. Then, assuming also that the heat conductivity is constant, we have
div v = 0, (4.17.10)
Dv
Dt
+
1
. (4.17.13)
I like the fact that the energy equation (the temperature equation) is completely independent from other equations.
So, we can solve the continuity and momentum equations rst (to determine the velocity eld), and then, if necessary,
solve the energy equation separately for a given velocity eld. Also, I like the idea of pseudo-compressibility. See
Section 4.21.
4.18 2D Incompressible Navier-Stokes Equations
I like the 2D incompressible Navier-Stokes equations in any of the following forms.
Conservative Form
(u)
x
+
(v)
y
= 0, (4.18.1)
(u)
t
+
(u
2
)
x
+
(uv)
y
=
p
x
+
x
_
u
x
_
+
y
_
u
y
_
+
x
_
u
x
_
+
y
_
v
x
_
, (4.18.2)
(v)
t
+
(uv)
x
+
(v
2
)
y
=
p
y
+
x
_
v
x
_
+
y
_
v
y
_
+
x
_
u
y
_
+
y
_
v
y
_
, (4.18.3)
c
v
_
T
t
+u
T
x
+v
T
y
_
=
x
_
k
T
x
_
+
y
_
k
T
y
_
+ , (4.18.4)
where
=
_
2
_
u
x
_
2
+ 2
_
v
y
_
2
+
_
v
x
+
u
y
_
2
_
. (4.18.5)
Primitive Form
u
x
+
v
y
= 0, (4.18.6)
u
t
+u
u
x
+v
u
y
=
p
x
+
x
_
u
x
_
+
y
_
u
y
_
+
x
_
u
x
_
+
y
_
v
x
_
, (4.18.7)
4.19. 3D INCOMPRESSIBLE NAVIER-STOKES EQUATIONS 159
v
t
+u
v
x
+v
v
y
=
p
y
+
x
_
v
x
_
+
y
_
v
y
_
+
x
_
u
y
_
+
y
_
v
y
_
, (4.18.8)
c
v
_
T
t
+u
T
x
+v
T
y
_
=
x
_
k
T
x
_
+
y
_
k
T
y
_
+ . (4.18.9)
Case of Constant Viscosity and Heat Conductivity
u
x
+
v
y
= 0, (4.18.10)
u
t
+u
u
x
+v
u
y
=
1
p
x
+
_
2
u
x
2
+
2
u
y
2
_
, (4.18.11)
v
t
+u
v
x
+v
v
y
=
1
p
y
+
_
2
v
x
2
+
2
v
y
2
_
, (4.18.12)
c
v
_
T
t
+u
T
x
+v
T
y
_
= k
_
2
T
x
2
+
2
T
y
2
_
+ . (4.18.13)
4.19 3D Incompressible Navier-Stokes Equations
Of course, I like the 3D incompressible Navier-Stokes equations also in any of the following forms.
Conservative Form
(u)
x
+
(v)
y
+
(w)
z
= 0, (4.19.1)
(u)
t
+
(u
2
)
x
+
(uv)
y
+
(uw)
z
=
p
x
+
x
_
u
x
_
+
y
_
u
y
_
+
z
_
u
z
_
+
x
_
u
x
_
+
y
_
v
x
_
+
z
_
w
x
_
,
(4.19.2)
(v)
t
+
(uv)
x
+
(v
2
)
y
+
(vw)
z
=
p
y
+
x
_
v
x
_
+
y
_
v
y
_
+
z
_
v
z
_
+
x
_
u
y
_
+
y
_
v
y
_
+
z
_
w
y
_
,
(4.19.3)
(w)
t
+
(uw)
x
+
(vw)
y
+
(w
2
)
z
=
p
z
+
x
_
w
x
_
+
y
_
w
y
_
+
z
_
w
z
_
+
x
_
u
z
_
+
y
_
v
z
_
+
z
_
w
z
_
,
(4.19.4)
c
v
_
T
t
+u
T
x
+v
T
y
+w
T
z
_
=
x
_
k
T
x
_
+
y
_
k
T
y
_
+
z
_
k
T
z
_
+ ,
(4.19.5)
160 CHAPTER 4. NAVIER-STOKES EQUATIONS
where
=
_
2
_
u
x
_
2
+ 2
_
v
y
_
2
+ 2
_
w
z
_
2
+
_
v
x
+
u
y
_
2
+
_
w
y
+
v
z
_
2
+
_
u
z
+
w
x
_
2
_
. (4.19.6)
Primitive Form
u
x
+
v
y
+
w
z
= 0, (4.19.7)
u
t
+u
u
x
+v
u
y
+w
u
z
=
p
x
+
x
_
u
x
_
+
y
_
u
y
_
+
z
_
u
z
_
+
x
_
u
x
_
+
y
_
v
x
_
+
z
_
w
x
_
, (4.19.8)
v
t
+u
v
x
+v
v
y
+w
v
z
=
p
y
+
x
_
v
x
_
+
y
_
v
y
_
+
z
_
v
z
_
+
x
_
u
y
_
+
y
_
v
y
_
+
z
_
w
y
_
, (4.19.9)
w
t
+u
w
x
+v
w
y
+w
w
z
=
p
z
+
x
_
w
x
_
+
y
_
w
y
_
+
z
_
w
z
_
+
x
_
u
z
_
+
y
_
v
z
_
+
z
_
w
z
_
, (4.19.10)
c
v
_
T
t
+u
T
x
+v
T
y
+w
T
z
_
=
x
_
k
T
x
_
+
y
_
k
T
y
_
+
z
_
k
T
z
_
+ .
(4.19.11)
Case of Constant Viscosity and Heat Conductivity
u
x
+
v
y
+
w
z
= 0, (4.19.12)
u
t
+u
u
x
+v
u
y
+w
u
z
=
1
p
x
+
_
2
u
x
2
+
2
u
y
2
+
2
u
z
2
_
, (4.19.13)
v
t
+u
v
x
+v
v
y
+w
v
z
=
1
p
y
+
_
2
v
x
2
+
2
v
y
2
+
2
v
z
2
_
, (4.19.14)
w
t
+u
w
x
+v
w
y
+w
w
z
=
1
p
z
+
_
2
w
x
2
+
2
w
y
2
+
2
w
z
2
_
, (4.19.15)
c
v
_
T
t
+u
T
x
+v
T
y
+w
T
z
_
= k
_
2
T
x
2
+
2
T
y
2
+
2
T
z
2
_
+ . (4.19.16)
4.20 Axisymmetric Incompressible Navier-Stokes Equations
I like the axisymmetric incompressible Navier-Stokes equations because they are much simpler than the compressible
version in Section 4.11. I really like the assumptions that and are constant, and thus = / is also constant.
Then, we write the continuity equation (4.17.1) and the conservative momentum equations (4.17.2) in cylindrical
4.20. AXISYMMETRIC INCOMPRESSIBLE NAVIER-STOKES EQUATIONS 161
coordinates in Figure 1.5.1, where the velocity is denoted by v = (u
r
, u
, u
z
):
1
r
(ru
r
)
r
+
u
z
z
= 0, (4.20.1)
u
r
t
+
1
r
(ru
2
r
)
r
+
(u
r
u
z
)
z
=
(p/)
r
+
_
1
r
r
_
r
u
r
r
_
+
2
u
r
z
2
u
r
r
_
, (4.20.2)
u
z
t
+
1
r
(ru
z
u
r
)
r
+
(u
2
z
)
z
=
(p/)
z
+
_
1
r
r
_
r
u
z
r
_
+
2
u
z
z
2
_
, (4.20.3)
where the viscous terms are obtained by using the Laplacian form (4.17.9). Note that u
vanishes and therefore it is ignored in the above. Note also that the viscous terms
have been obtained by Equation (1.5.61), i.e., the Laplacian of a vector. The momentum equations in the primitive
form can be obtained from the primitive form also by directly writing out (gradv)v or by inserting the continuity
equation into the conservative momentum equations:
u
r
t
+u
r
u
r
r
+u
z
u
r
z
=
(p/)
r
+
_
1
r
r
_
r
u
r
r
_
+
2
u
r
z
2
u
r
r
_
, (4.20.4)
u
z
t
+u
r
u
z
r
+u
z
u
z
z
=
(p/)
z
+
_
1
r
r
_
r
u
z
r
_
+
2
u
z
z
2
_
. (4.20.5)
Now, as we have done in Section 4.11, we multiply all the equations by r, replace (r, z) by (y, x) and (u
r
, u
z
) by (v, u),
and reverse the order of the momentum equations to obtain the axisymmetric incompressible Navier-Stokes equations
in a form similar to the two-dimensional equations:
(yu)
x
+
(yv)
y
= 0, (4.20.6)
(yu)
t
+
(yu
2
)
x
+
(yuv)
y
=
(yp/)
x
+
_
x
_
y
u
x
_
+
y
_
y
u
y
__
, (4.20.7)
(yv)
t
+
(yvu)
x
+
(yv
2
)
y
=
(yp/)
y
+
_
x
_
y
v
x
_
+
y
_
y
v
y
__
+
p
v
y
. (4.20.8)
In the vector form, it can be written as
(yU)
t
+ (yF)
x
+ (yG)
y
= (yF
v
)
x
+ (yG
v
)
y
+S, (4.20.9)
where
U =
_
_
0
u
v
_
_
, F =
_
_
u
u
2
+p/
uv
_
_
, G =
_
_
v
uv
v
2
+p/
_
_
, (4.20.10)
F
v
=
_
_
0
u/x
v/x
_
_
, G
v
=
_
_
0
u/y
v/y
_
_
, S =
_
_
0
0
p/
v
y
_
_
. (4.20.11)
Of course, in the same way as described in Section 4.11, we can express the axisymmetric form and the two-dimensional
form in a single form with a logical switch, = 0 for the two-dimensional form and = 1 for the axisymmetric form:
(
a
U)
t
+ (
a
F)
x
+ (
a
G)
y
= (
a
F
v
)
x
+ (
a
G
v
)
y
+ S, (4.20.12)
where
a
= (1 ) + y. (4.20.13)
162 CHAPTER 4. NAVIER-STOKES EQUATIONS
This is really nice.
One of the reasons that I like the axisymmetric incompressible Navier-Stokes equations very much is that the stream
function for axisymmetric ows can be easily found from the continuity equation (4.20.6):
u =
1
y
y
, v =
1
y
x
. (4.20.14)
Clearly, then, the continuity equation is automatically satised. Its so simple. I like it.
4.21 Pseudo-Compressible Navier-Stokes Equations
Yes, I like the pseudo-compressible Navier-Stokes equations. Introduce the articial density,
= p/a
2
, (4.21.1)
and add its time derivative to the continuity equation,
+ div v = 0, (4.21.2)
which can be written also as
P
t
+ div (a
2
v) = 0, (4.21.3)
where P = p/ which is the kinematic pressure. The incompressible Navier-Stokes equations with this modied
continuity equation are the pseudo-compressible Navier-Stokes equations [21, 75, 97, 132]:
P
t
+ div (a
2
v) = 0, (4.21.4)
v
t
. (4.21.6)
The time derivative in the momentum equation is a real time derivative but not realistic when solved with the pseudo-
compressible continuity equation (so t
instead of t).
In Cartesian coordinates:
U
t
+F
x
+G
y
+H
z
= F
v
x
+G
v
y
+H
v
z
, (4.21.7)
U =
_
_
P
u
v
w
_
_
, F =
_
_
a
2
u
u
2
+P
uv
uw
_
_
, G =
_
_
a
2
v
uv
v
2
+P
uw
_
_
, H =
_
_
a
2
w
uv
vw
w
2
+P
_
_
, (4.21.8)
F
v
=
_
_
0
2u
x
u
y
+v
x
u
z
+w
x
_
_
, G
v
=
_
_
0
v
x
+u
y
2v
y
v
z
+w
y
_
_
, H
v
=
_
_
0
w
x
+u
z
w
y
+v
z
2w
z
_
_
. (4.21.9)
If we assume that the viscosity is constant, then the viscous term can be simplied to the Laplacian as shown in
Equation (4.17.9),
v
t
_
0
u
x
v
x
w
x
_
_
, G
v
=
_
_
0
u
y
v
y
w
y
_
_
, H
v
=
_
_
0
u
z
v
z
w
z
_
_
. (4.21.11)
This is nice. I like it.
4.22. VORTICITY TRANSPORT EQUATION AND STREAM FUNCTION 163
4.22 Vorticity Transport Equation and Stream Function
The vorticity transport equation (the curl of the momentum equation (4.17.11)) is given by
t
+u
x
+v
y
= (
xx
+
yy
) , (4.22.1)
where = v
x
u
y
is the vorticity. To solve this for , the velocity components (u, v) must be given or solved
simultaneously. A nice way to compute the velocity is to introduce the stream function as
u =
y
, v =
x
, (4.22.2)
so that the continuity equation is automatically satised, and that we see from = v
x
u
y
that the stream function
satises the following:
xx
+
yy
= . (4.22.3)
That is, we now have two equations (4.22.1) and (4.22.3) for two unknowns (, ). So, we only need to solve these
two equations (4.22.1) and (4.22.3) to get and , and determine the velocity eld.
Once the velocity is obtained (via ), the pressure can be computed as a solution of the following Poisson equation
(the divergence of the momentum equations):
div grad p = 2 (u
x
v
y
u
y
v
x
) = 2
_
xx
yy
(
xy
)
2
. (4.22.4)
These are the fundamental equations for the famous vorticity-stream function approach (see Ref.[155] for details). It
is nice that instead of solving three equations simultaneously (continuity and two momentum equations), we solve two
equations simultaneously and then solve one equation independently afterwards. I surely like these equations.
4.23 Vorticity Transport Equation and Vector Potential
For a three-dimensional incompressible ow, we write the velocity as follows,
v = grad + curl , (4.23.1)
where is a scalar potential and is a vector potential. Note that the rst term on the right is curl-free while the
second term is divergence-free. This is called the Helmholtz decomposition. Substitute Equation (4.23.1) into the
continuity equation to nd the governing equation of ,
div grad = 0. (4.23.2)
Substitute Equation (4.23.1) into the vorticity, = curl v, to get the governing equation of ,
div grad = . (4.23.3)
Also, take the curl of the momentum equations (4.17.11) to get the governing equation of ,
D
Dt
= (grad v) + div grad . (4.23.4)
Now, we can determine the velocity eld by solving the above seven equations, (4.23.2), (4.23.3), and (4.23.4). Once
the velocity eld is determined, we can compute the pressure by solving the Poisson equation (the divergence of the
momentum equations):
div grad p = div [div (vv)] , (4.23.5)
whose components are
x
i
_
p
x
i
_
=
x
i
_
(v
i
v
j
)
x
j
_
. (4.23.6)
164 CHAPTER 4. NAVIER-STOKES EQUATIONS
We may use the continuity equation to write this as
x
i
_
p
x
i
_
=
v
j
x
i
v
i
x
j
. (4.23.7)
In Cartesian coordinates, we can write the Poisson equation as
p
xx
+p
yy
+p
zz
= (u
2
)
xx
(uv)
xy
(uw)
xz
(vu)
yx
(v
2
)
yy
(vw)
yz
(wu)
zx
(wv)
zy
(w
2
)
zz
, (4.23.8)
or we obtain directly from Equation (4.23.7)
p
xx
+p
yy
+p
zz
=
_
(u
x
)
2
+ (v
y
)
2
+ (w
z
)
2
+ 2u
y
v
x
+ 2u
z
w
x
+ 2w
y
v
z
, (4.23.9)
which we can rewrite as (by the continuity equation)
p
xx
+p
yy
+p
zz
= 2 [u
x
v
y
+v
y
w
z
+w
z
u
x
u
y
v
x
u
z
w
x
w
y
v
z
] . (4.23.10)
These equations are fundamental equations for vortex methods [29].
4.24 Stream Functions
I like stream functions. They are useful in generating as well as visualizing ows. In 2D, we have only one stream
function, and we get two velocity components from a scalar function. That is nice. However, in 3D, we need two
stream functions. Thats one more scalar function. But it is still nice compared with the vorticity, which has one
component in 2D but three components in 3D.
4.24.1 Three-Dimensional Stream Functions
I like to set = grad and = 0 because then I nd from Equation (4.23.1) that
v = curl
= curl ( grad )
= curl(grad) + grad grad
= grad grad . (4.24.1)
This means that v is tangent to the two surfaces dened by =constant and = constant, and so that these two
families of surfaces are composed of streamlines. Therefore, and can be thought of as three-dimensional stream
functions. Of course, the continuity equation, div v = 0, is automatically satised because div(curl ) = 0. This is very
interesting. In fact, various well-known two-dimensional stream functions can be derived from the three-dimensional
stream functions.
4.24.2 Cartesian Coordinates (2D)
Consider Cartesian coordinates (x, y, z), and denote the velocity components by v = [u, v, w]
t
. We can derive a
well-known two-dimensional stream function by setting = (x, y) and = z, i.e., = grad = e
z
. Equation
(4.24.1) becomes
u =
y
, v =
x
. (4.24.2)
The governing equation of the stream function is given by curl v = , which becomes
x
2
+
2
y
2
=
z
, (4.24.3)
where = (0, 0,
z
). For irrotational ows, the stream function satises the Laplace equation:
x
2
+
2
y
2
= 0. (4.24.4)
That is nice.
4.24. STREAM FUNCTIONS 165
4.24.3 Polar Coordinates (2D)
Consider cylindrical coordinates (r, , z), and denote the velocity components by v = [u
r
, v
, u
z
]
t
. We can derive
a well-known two-dimensional stream function in polar coordinates by setting = (r, ) and = z, i.e., =
grad = e
z
. Equation (4.24.1) becomes
u
r
=
1
r
, u
r
. (4.24.5)
The governing equation of the stream function is given by curl v = , which becomes in cylindrical coordinates
1
r
r
_
r
r
_
+
1
r
2
2
=
z
. (4.24.6)
Hence, for irrotational ows, the stream function satises, again, the Laplace equation:
1
r
r
_
r
r
_
+
1
r
2
2
= 0. (4.24.7)
4.24.4 Axisymmetric Cylindrical Coordinates
We can derive the stream function for axisymmetric ows (i.e., u
z
, u
z
=
1
r
r
, (4.24.8)
where the axis of symmetry has been taken along the z-axis (see Figure 1.5.1). Axisymmetric ows do not have
variations in by denition, and therefore they are essentially two-dimensional. Here, the two-dimensional plane is
dened by the (z, r)-plane. The governing equation of the stream function is given by curl v = , which becomes in
cylindrical coordinates for axisymmetric ows
z
_
1
r
z
_
+
r
_
1
r
r
_
=
, (4.24.9)
or equivalently
1
r
_
z
2
+
2
r
2
1
r
r
_
=
. (4.24.10)
Note that the left hand side is not the Laplacian. The Laplacian for axisymmetric ows in cylindrical coordinates is
given by
div grad = div
_
r
,
1
r
,
z
_
=
1
r
r
_
r
r
_
+
2
z
2
=
2
r
2
+
2
z
2
+
1
r
r
, (4.24.11)
which does not precisely match the operator on the left hand side of Equation 4.24.10. Therefore, the stream function
does not satisfy the Laplace equation in axisymmetric irrotational ows.
4.24.5 Axisymmetric Spherical Coordinates
The stream function for axisymmetric ows can be derived also in spherical coordinates (r, , ), where the velocity
components are denoted by v = [u
r
, u
, u
]
t
. Note that here is the angle, not the scalar potential. By setting
= (r, ) and = , i.e., = grad = e
, u
=
1
rsin
r
, (4.24.12)
166 CHAPTER 4. NAVIER-STOKES EQUATIONS
where again the axis of symmetry has been taken along the z-axis (see Figure 1.5.2). The governing equation of the
stream function is given by curl v = , which becomes in spherical coordinates for axisymmetric ows
1
r
_
r
_
1
sin
r
_
+
_
1
r
2
sin
__
=
. (4.24.13)
Again, the left hand side is not the Laplacian. The stream function does not satisfy the Laplace equation in axisymmetric
irrotational ows.
4.24.6 Axisymmetric (x, y) Coordinates
It is straightforward to show that Equation (4.24.8) and Equation (4.24.12) are equivalent, and both can be written as
u =
1
y
y
, v =
1
y
x
. (4.24.14)
Note that here the x-axis is the horizontal axis of symmetry and the y-axis is the vertical axis. Of course, the above
expression is equivalent to the one obtained directly from the axisymmetric continuity equation, i.e., Equation (4.20.14)
in Section 4.20. The governing equation of the stream function is
x
2
+
2
y
2
1
y
y
= y
, (4.24.15)
which will not reduce to the Laplace equation for irrotational ows as shown in Section 4.24.4. For irrotational ows,
we can dene the velocity potential, :
u =
x
, v =
y
. (4.24.16)
The system (4.24.14) and (4.24.16) looks like the Cauchy-Riemann system for and as described in Section 2.9.2,
but not really because of the non-constant coecient 1/y. Note also that although does not satisfy the Laplace
equation, the velocity potential satises the Laplace equation. Anyway, I like axisymmetric ows especially expressed
in the (x, y)-plane because the equations are very similar to those in two-dimensional ows and it makes it easy to
understand the dierences between two-dimensional and axisymmetric ows as we have just seen.
4.24.7 Volume Flow and Stream Functions
Of course, the stream functions are associated with the volume ow. Consider a volume ow Q through a section S,
Q =
S
v ndS, (4.24.17)
where dS and n are the innitesimal area and the unit outward normal vector of the section, respectively. By Stokes
theorem in Section 1.13, it reduces to the line integral over the boundary of S,
Q =
S
curl dl =
S
grad dl, (4.24.18)
where dl denotes the innitesimal length vector in the counterclockwise direction. Assume that the section S is bounded
by four stream surfaces, =
1
,
2
and =
1
,
2
:
S = S
1
+S
1
+S
2
+S
2
, (4.24.19)
where
1
,
2
,
1
, and
2
are constants, S
1
denotes the boundary where =
1
, and similarly for others. Then, on
each boundary portion, one of the functions is constant. Noting that grad = 0 on S
1
and S
2
, we obtain
Q =
1
1
grad dl +
2
2
grad dl. (4.24.20)
By denition, d = grad dl. We thus nally obtain
Q =
1
(
1
2
) +
2
(
2
1
) = (
2
1
)(
2
1
). (4.24.21)
4.25. STOKES EQUATIONS 167
Therefore, the volume ow through the section S is given by the product of the dierences of the stream functions.
I like the stream functions because they are applicable to a wide range of ows: irrotational, rotational, inviscid,
and viscous ows. Note also that the stream functions can be dened for compressible ows as v = (1/)curl ,
where is the density, so that it automatically satises the steady continuity equation div(v) = 0. See, for exam-
ple, Refs.[18, 67, 120] for more details on three-dimensional stream functions. See Ref.[140] for an application of
three-dimensional stream functions to CFD.
4.25 Stokes Equations
I like Stokes equations because they are linear. In fact, we obtain Stokes equations by ignoring the nonlinear convective
term from the incompressible Navier-Stokes equations. This is a valid approximation for small Reynolds number ows.
These ows are often called Stokes ows or creeping ows.
Stokes Equations
Stokes equations are given by
div v = 0, (4.25.1)
v
t
+ grad p div grad v = 0, (4.25.2)
where and are constants. Note that the momentum equation has been obtained by ignoring the convective term
(grad v)v from Equation (4.17.7). I like the fact that the pressure satises the Laplace equation, which can be easily
shown by taking the divergence of the momentum equation:
(div v)
t
+ div grad p div div grad v = 0, (4.25.3)
which becomes by Equation (1.7.12)
div v)
t
+ div grad p div [grad(div v) curl(curl v)] = 0, (4.25.4)
div grad p div grad(
div v) +
(
(
(
(
(
((
div curl(curl v) = 0, (4.25.5)
and thus
div grad p = 0. (4.25.6)
This is the Laplace equation for the pressure. Thats nice. Recall that the pressure satises the Poisson equation
(4.22.4) when the convective term is present.
Stokes Equations (pressure and vorticity)
Stokes equations can be written as a rst-order system in terms of the pressure and the vorticity. By Equation (1.7.12),
we have
div grad v = grad(div v) curl curl v = curl , (4.25.7)
where = curl v = [
x
,
y
,
z
]
t
is the vorticity. Then, we obtain
div v = 0, (4.25.8)
curl v = , (4.25.9)
v
t
+ grad p +curl = 0. (4.25.10)
168 CHAPTER 4. NAVIER-STOKES EQUATIONS
It is nice to be a rst-order system because often the discretization is relatively simpler compared with second derivatives.
Nevertheless, I also nd it interesting that the vorticity obeys the diusion equation. Take the curl of the momentum
equation:
t
+
t
= div grad , (4.25.12)
where = /. More interestingly, for steady ows, the vorticity satises the Laplace equation:
div grad = 0. (4.25.13)
4.26 2D Stokes Equations
I like the 2D Stokes equations. Especially I like the steady version because they reduce to two sets of Cauchy-Riemann
type equations.
2D Stokes Equations
u
x
+
v
y
= 0, (4.26.1)
u
t
+
p
x
_
2
u
x
2
+
2
u
y
2
_
= 0, (4.26.2)
v
t
+
p
y
_
2
v
x
2
+
2
v
y
2
_
= 0. (4.26.3)
2D Stokes Equations (pressure and vorticity)
u
x
+
v
y
= 0, (4.26.4)
v
x
u
y
=
z
, (4.26.5)
u
t
+
p
x
+
z
y
= 0, (4.26.6)
v
t
+
p
y
z
x
= 0. (4.26.7)
For steady ows, these are basically two sets of the Cauchy-Riemann equations; they are coupled through the vorticity
z
. It is nice to be the Cauchy-Riemann system because exact solutions can be found by using the theory of complex
variables [54]. I nd it interesting that the Stokes equations cannot accurately describe a ow around an object in
two dimensions: a far-eld free stream condition cannot be satised simultaneously with the no-slip condition on the
object. This is called Stokes paradox. To resolve this, we must retain a leading term in a linearized advection term;
this is called Oseens approximation. See Ref.[54] for more details.
4.27 3D Stokes Equations
I like the 3D Stokes equations because Stokes paradox does not apply and a ow over a body can be properly
described. It is also exciting that the vorticity is now a vector having three components (it has only one component in
two dimensions).
4.28. AXISYMMETRIC STOKES EQUATIONS 169
3D Stokes Equations
u
x
+
v
y
+
w
z
= 0, (4.27.1)
u
t
+
p
x
_
2
u
x
2
+
2
u
y
2
+
2
u
z
2
_
= 0, (4.27.2)
v
t
+
p
y
_
2
v
x
2
+
2
v
y
2
+
2
v
z
2
_
= 0, (4.27.3)
w
t
+
p
z
_
2
w
x
2
+
2
w
y
2
+
2
w
z
2
_
= 0. (4.27.4)
3D Stokes Equations (pressure and vorticity)
u
x
+
v
y
+
w
z
= 0, (4.27.5)
w
y
v
z
=
x
, (4.27.6)
u
z
w
x
=
y
, (4.27.7)
v
x
u
y
=
z
, (4.27.8)
p
x
+
_
z
y
y
z
_
= 0, (4.27.9)
u
t
+
p
y
+
_
x
z
z
x
_
= 0, (4.27.10)
v
t
+
p
z
+
_
y
x
x
y
_
= 0. (4.27.11)
It is really nice to be a rst-order system, but now I dont see the Cauchy-Riemann equations for steady ows. The
theory of complex variables cannot be applied, and exact solutions need to be found by some other methods [54].
"How about axisymmetric ows?" Thats a good point. As we have seen before, axisymmetric equations are almost
like two-dimensional equations. It is worth looking at.
4.28 Axisymmetric Stokes Equations
Axisymmetric Stokes Equations
Axisymmetric Stokes equations can be obtained from the axisymmetric incompressible Navier-Stokes equations (4.20.8)
by ignoring the convective terms:
(yu)
x
+
(yv)
y
= 0, (4.28.1)
(yu)
t
+
(yp)
x
_
x
_
y
u
x
_
+
y
_
y
u
y
__
= 0, (4.28.2)
(yv)
t
+
(yp)
y
_
x
_
y
v
x
_
+
y
_
y
v
y
__
= p
v
y
, (4.28.3)
where as before the x-axis has been taken as the axis of symmetry.
170 CHAPTER 4. NAVIER-STOKES EQUATIONS
Axisymmetric Stokes Equations (pressure and vorticity)
For axisymmetric ows, the rst-order system of Equations (4.25.8), (4.25.9), and (4.25.10) becomes
(yu)
x
+
(yv)
y
= 0, (4.28.4)
v
x
u
y
= , (4.28.5)
(yu)
t
+
(yp)
x
+
(y)
y
= 0, (4.28.6)
(yv)
t
+
(yp)
y
(y)
x
= p, (4.28.7)
where is the -component of the vorticity, which is the only nonzero component. As expected, the axisymmetric
system is very similar to the two-dimensional system, but not similar enough. Unfortunately, the momentum equations
do not become purely the Cauchy-Riemann system for steady ows because of the presence of the pressure source term.
Nevertheless, the stream function can be dened for the continuity equation as in the axisymmetric incompressible
Navier-Stokes equations (see Section 4.20), which can be useful for deriving exact solutions [54].
4.29 Boundary Layer Equations
The boundary layer equations for two-dimensional incompressible ows are given by
u
x
+
v
y
= 0, (4.29.1)
u
u
x
+v
u
y
= U
U
x
+
2
u
y
2
, (4.29.2)
where U = U(x) is the velocity outside the boundary layer (constant for a ow over a at plate). This can be solved
numerically, for example, as follows [158]. First, we compute v by numerically integrating the continuity equation in
y direction with the boundary condition u(x, 0) = v(x, 0) = 0, then solve the momentum equation for given v, and
repeat this process until convergence. In this case, the momentum equation behaves like the viscous Burgers equation
(2.15.4). It is nice that a scalar scheme can be used for solving such a nonlinear system. To be even nicer, there exists
an exact solution for this system (see Section 7.15.9), so that the accuracy of the numerical solution can be directly
measured.
Chapter 5
Turbulence Equations
5.1 Averages and Filters
Express a physical quantity a as the sum of the average a and the uctuation a
: for example,
v = v +v
, = +
, p = p +p
, h = h +h
, T = T +T
, H = H +H
. (5.1.1)
This is called the Reynolds decomposition. Here we must be careful about the fact that the uctuation is not necessarily
small (compared with the average), and also that usually uctuations of the viscosity, the heat conductivity, and the
specic heats are small and therefore ignored.
In turbulence calculations, typically we average the Navier-Stokes equations and solve for averaged quantities. After
averaging, as we will see later, we nd extra terms created which involve uctuations and cannot be evaluated from
averaged quantities. So, we derive governing equations for these extra terms, but then we will nd similar extra terms
again that cannot be evaluated, and this process goes on forever. Therefore, at some point, we must stop and model
these terms, i.e., evaluate them somehow in terms of averaged quantities.
The most fundamental in turbulence calculations is the averaging. There are many dierent ways of averaging.
Properties of averages and uctuations depend on their denitions. Also, terms that need to be modeled, in averaged
governing equations, depend on them.
5.1.1 Reynolds Averaging
The Reynolds averaging includes ensemble averaging, time averaging, and space averaging. These are very interesting.
For example, they are identical to one another for steady homogeneous turbulence (which is called Ergodic hypothesis).
The simulation based on the Reynolds-averaged Navier-Stokes equations is called RANS.
Ensemble Averaging
The ensemble averaging:
f(x, t) = lim
N
1
N
N
i=1,N
f
i
(x, t), (5.1.2)
which is the average of measurements, f
i
(x, t) where i = 1, 2, 3, N, obtained by N experiments. I like it, but it
is unrealistic to perform this averaging in practice. You can, of course, formally (or theoretically or mathematically,
whatever) use this averaging to write down averaged governing equations.
Time Averaging
The time averaging:
f(x, t) =
1
T
t+T
t
f(x, t) dt, (5.1.3)
171
172 CHAPTER 5. TURBULENCE EQUATIONS
where T must be larger than the uctuation time scale but smaller than the average time scale. I like this one also,
but in real ows, these two time scales are comparable, so such T may not exist. Of course, you can still formally
derive averaged governing equations based on this assumption. For a steady ow, by taking T , we may dene
f(x) = lim
T
1
T
t+T
t
f(x, t) dt. (5.1.4)
Space Averaging
The space averaging:
f(t) = lim
V
1
V
V
f(x, t) dV, (5.1.5)
which can be used for homogeneous turbulence in which the mean ow quantities are uniform in space. Of course,
many of us are probably not interested in uniform mean ows. However, for any ows, if we look close enough, there
could be a limit beyond which the ow looks locally uniform. Then, the homogeneous turbulence does make sense,
and the space averaging can be useful. I like it.
Properties of Reynolds Averaging
The Reynolds averaging has the following properties:
f = f, (5.1.6)
f
= 0, (5.1.7)
fg
= 0, (5.1.8)
fg = f g, (5.1.9)
f +g = f +g. (5.1.10)
They all make sense and thus easy to remember. I like them.
5.1.2 Favre Averaging
Favre Averaging
The Favre averaging (denoted by a tilde):
f =
f
, (5.1.11)
which may be called the mass-weighted averaging or the mass-averaging. In fact, I like it very much because we can
greatly simplify the averaged compressible Navier-Stokes equations using this particular averaging, much simpler than
those obtained by using the Reynolds averaging.
Usually, only the velocity and thermodynamic variables are Favre-averaged and the density and the pressure are
Reynolds-averaged:
v =
v
,
h =
h
,
T =
T
,
H =
H
. (5.1.12)
Also note that the uctuation associated with Favre averaging
f is denoted by f
:
v = v +v
, = +
, p = p +p
, h =
h +h
, T =
T +T
, H =
H +H
. (5.1.13)
5.1. AVERAGES AND FILTERS 173
Properties of Favre Averaging
Favre averaging has the following properties:
f
= f f
, (5.1.14)
f
, (5.1.15)
f
= 0, (5.1.16)
f
= 0, (5.1.17)
f =
f. (5.1.18)
These are more complicated than those of the Reynolds Averaging, but I like them. These are very important, anyway.
5.1.3 Spatial Filter
Spatial Filter
The spatial lter is a local space (or a local wave-number space) averaging. A typical spatial lter is the volume-average
box lter (also called the top-hat lter):
f(x, t) =
1
x+x/2
xx/2
y+y/2
yy/2
z+z/2
zz/2
f(, t) ddd, (5.1.19)
where
x = (x, y, z)
t
= (x
1
, x
2
, x
3
)
t
, = (, , )
t
= (
1
,
2
,
3
)
t
, = (xyz)
1/3
. (5.1.20)
We may think of it as the integral over a computational cell: f(x, t) can be thought of as a smooth (low-frequency)
component that can be resolved on that grid. This smooth component is called a grid scale (GS) or a resolvable scale,
and the rest (high-frequency component that cannot be resolved) is called a sub-grid scale (SGS), i.e.,
f = f +f
, (5.1.21)
in which f
is the sub-grid scale. Note that in principle the lter spacing, (x, y, z), does not have to be the grid
spacing (as long as larger than the grid spacing). So, the terms like GS and SGS may be confusing. This ltering
operation can be thought of as a decomposition of a ow into large eddies(GS) and small eddies(SGS). The smaller the
lter spacing is, the more universal the properties of small eddies can be (Kolmogorovs law: a theory by Kolmogorov,
which is sometimes called K41 theory because it came out in 1941 [73]. See Refs.[12, 156] for details). If the lter
spacing is small enough, then, a universal model can be developed: methods applicable to many dierent ows can
be developed. Yes, a smaller lter spacing implies ner computational grids, and hence such simulations are generally
very expensive.
We can generalize the ltering operation by writing it in the form,
f(x, t) =
i
) =
_
1/
3
|x
i
i
| < x
i
/2,
0 otherwise.
(5.1.23)
There are other types of lters such as the Gaussian lter or the Fourier cut-o lter. See Ref.[171] for details.
The large eddy simulation (LES) refers to a simulation based on ltered governing equations with suciently small
lter spacing such that the sub-grid scale will be homogeneous. I like the idea, but it requires a very ne computational
grid (very expensive). If I use a grid ne enough to resolve all scales, then no turbulence models are needed, and the
simulation is called the direct numerical simulations (DNS). I like this one. Its very simple. It is particularly nice that
174 CHAPTER 5. TURBULENCE EQUATIONS
we do not need turbulence modeling. However it is even more expensive than LES: the grid must be extremely or often
prohibitively ne especially for high-Reynolds-number ows encountered in many practical applications. There is also
a class of methods called detached eddy simulations (DES) [147, 151]. In DES, the near-wall region (too expensive to
resolve by LES) is solved by RANS and the rest of the domain (where relatively large eddies dominate) is solved by
LES. I like the idea of DES, but I really think that it would be very very nice if DNS would become feasible for practical
ows. It is so simple.
Properties of Spatial Filter
Generally, ltering once and twice, we obtain dierent values:
f = f, (5.1.24)
which is not the case in the Reynolds averaging. Therefore, ltered governing equations and averaged governing
equations can be quite dierent. We must be careful about the following also:
f
= 0, (5.1.25)
fg
= 0. (5.1.26)
Do I like it? Well, sure, of course, I like it.
5.2 Reynolds-Averaged Incompressible Navier-Stokes Equations
Take the Reynolds average of the incompressible Navier-Stokes equations to get the Reynolds-averaged incompressible
Navier-Stokes equations ( = constant):
div (v) = 0, (5.2.1)
t
(v) + div(vv) + grad p div = 0, (5.2.2)
where
= v
=
_
grad v + (grad v)
t
. (5.2.3)
The extra term,
v
, (5.2.4)
has been generated by the averaging operation. This is usually interpreted as a stress and called the Reynolds stress.
Without the Reynolds stress, the system would have been closed and solved for the average quantities (just like the
laminar equations). It is possible to derive its own governing equations; but it is so much simpler to model this
term (i.e., to relate somehow to the average quantities) by the Boussinesq approximation (or the gradient transport
hypothesis) [16]:
v
=
T
_
grad v + (grad v)
t
2
3
kI, (5.2.5)
where
T
is called the eddy viscosity or the turbulence viscosity, and
k =
1
2
v
, (5.2.6)
is called the turbulent energy. Basically, we assume that this uctuation stress is proportional to the gradient of the
average quantities (similarly to Newtonian uids). The second term
2
3
kI is needed in order for the Boussinesq
approximation to be valid when traced: the trace of the right hand side must be equal to that of the left hand side
v
i
v
i
= 2k. But this term is usually ignored if we have no ways to evaluate k, e.g., in zero-equation or one-
equation models. To evaluate k, we would probably have to derive and solve the governing equation for k. Typically,
two-equation models include such an option (see Section 5.6).
5.3. FAVRE-AVERAGED COMPRESSIBLE NAVIER-STOKES EQUATIONS 175
I like ignoring k and employing the Boussinesq approximation because then I can write the viscous stress term as
= ( +
T
)
_
grad v + (grad v)
t
, (5.2.7)
so that practically the form of the Reynolds-averaged incompressible Navier-Stokes equations is the same as that of
the original (laminar) incompressible Navier-Stokes equations except that it has one extra viscosity coecient,
T
,
+
T
. (5.2.8)
So, the focus is now on estimating the turbulent viscosity,
T
. You might want to evaluate it algebraically (algebraic
models), by deriving and solving a governing equation for
T
(one-equation models), or by deriving and solving
governing equations for
T
as well as k (or any relevant two quantities: two-equation models).
5.3 Favre-Averaged Compressible Navier-Stokes Equations
Apply the Favre averaging (instead of the Reynolds averaging) to the compressible Navier-Stokes equations to ob-
tain the following relatively simple form of the averaged Navier-Stokes Equations:
t
+ div( v) = 0, (5.3.1)
t
( v) + div( v v) + grad p = div , (5.3.2)
t
(
E) + div( v
T, (5.3.4)
E = e +
v v
2
+k, (5.3.5)
H =
h +
v v
2
+k, (5.3.6)
k =
v
2
, (5.3.7)
= v
=
_
grad v + (grad v)
t
2
3
(div v)I v
, (5.3.8)
q =
P
r
grad h v
, (5.3.9)
and also
K = v
(v
)/2, (5.3.10)
but this one is considered to be small and often ignored. Also, as a matter of fact, because v
= 0, we actually have
=
_
grad v + (grad v)
t
2
3
(div v)I +
_
grad v
+ (grad v
)
t
2
3
(div v
)I,
(5.3.11)
but the last two terms are considered to be small and often ignored. So, I have done so in Equation (5.3.8). Also,
the turbulent energy k in
E and
H are often ignored. Now, with these terms all ignored, we apply the Boussinesq
approximation:
v
=
T
_
grad v + (grad v)
t
2
3
T
(div v)I
2
3
kI, (5.3.12)
v
=
T
Pr
T
grad
h, (5.3.13)
where of course, k will have to be ignored if there are no ways to estimate it; Pr
T
is called the turbulent Prandtl number
and Pr
T
= 0.9 is a good approximation for the sea-level air. So, we ignore k and obtain the following Favre-averaged
176 CHAPTER 5. TURBULENCE EQUATIONS
Navier-Stokes equations:
t
+ div( v) = 0, (5.3.14)
t
( v) + div( v v) + grad p = div , (5.3.15)
t
(
E) + div( v
T, (5.3.17)
E = e +
v v
2
, (5.3.18)
H =
h +
v v
2
, (5.3.19)
= ( +
T
)
_
grad v + (grad v)
t
2
3
( +
T
) (div v)I, (5.3.20)
q =
_
Pr
+
T
Pr
T
_
grad
h. (5.3.21)
Therefore, the dierences between these equations and the laminar equations are the turbulent viscosity and the
turbulent Prandtl number:
+
T
, (5.3.22)
Pr
Pr
+
T
Pr
T
, (5.3.23)
But since we may take Pr
T
= 0.9, it is again a matter of how to estimate
T
just like in the Reynolds-averaged
incompressible Navier-Stokes equations. This is nice and simple. I like it.
5.4 Filtered Incompressible Navier-Stokes Equations
Apply ltering to the incompressible Navier-Stokes equations to get the ltered incompressible Navier-Stokes equations:
div (v) = 0, (5.4.1)
t
(v) + div(vv) + grad p div = 0, (5.4.2)
where
SGS
= (vv vv), (5.4.3)
is called the SGS stress, which is usually expanded as
SGS
= (vv vv) + (v
v +vv
) +v
, (5.4.4)
and on the right hand side the rst term is called the Leonard stress, the second the cross-term stress, the third the
SGS Reynolds stress. It may look a little bit more complicated than the Reynolds stress in the Reynolds-averaged
incompressible Navier-Stokes equations (5.2.4), but I like it.
5.5 Filtered Compressible Navier-Stokes Equations
Apply mass-averaged ltering to the compressible Navier-Stokes Equations to obtain the simple ltered compress-
ible Navier-Stokes Equations:
t
+ div( v) = 0, (5.5.1)
t
( v) + div( v v) + grad p = div , (5.5.2)
t
(
E) + div( v
T,
E = e +
v v
2
+k
SGS
,
H =
h +
v v
2
+k
SGS
, (5.5.4)
=
_
grad v + (grad v)
t
2
3
(div v)I
SGS
, q =
Pr
grad
h +q
SGS
, (5.5.5)
k
SGS
=
1
2
( vv v v), (5.5.6)
SGS
= (
vv v v), (5.5.7)
q
SGS
= (
Ev
E v), (5.5.8)
K
SGS
= ( v v) ( pv p v). (5.5.9)
Usually, if the Mach number is not large, k
SGS
and K
SGS
are ignored and only
SGS
and q
SGS
are modeled. These
equations may look more complicated than the Favre-Averaged Navier-Stokes Equations, but essentially the same in
that extra terms are introduced in the viscous stress and the heat ux. I like these equations.
5.6 Turbulence Models
In many cases, the Boussinesq type approximation is made, and it leads us to the problem of estimating the eddy
viscosity,
T
. There are various ways to evaluate the eddy viscosity. The simplest would be the algebraic models where
the eddy viscosity is computed algebraically in terms of the mean ow quantities. This includes, for examples, Prandtls
mixing-length theory [121], the Cebeci-Smith model [143], the Baldwin-Lomax model [8], and their numerous variants.
These methods can give a good estimate for the eddy viscosity in some cases, but its success essentially depends on
the ow because the eddy viscosity is not an intrinsic property of the uid (unlike the molecular viscosity, ). In other
words, the models need to be tuned (e.g., constants somehow adjusted) for each application. Johnson-King model [63]
is a little bit more sophisticated model with an ordinary dierential equation incorporated (often called 1/2 equation
model), but still its applicability is limited because it does not directly estimate
T
. I like these algebraic models
because they are relatively easy to implement, but I must admit that we denitely need more sophisticated methods for
successful and general turbulence calculations. Beyond the algebraic model, the simplest method would be to directly
solve a governing equation for
T
. This is good because
T
will be solved along with the ow equations and thus
it now depend on the ow. This includes, for examples, the Baldwin-Barth model [7, 6], the Spalart-Allmaras model
[146], and their variants. I like these models. They are still simple. It is possible to couple the turbulence equation
with the ow equations, but it is also possible to solve them separately. Personally, I like the latter. In fact, these
models work successfully for many dierent ows, but not so much for separated ows. It looks like that we need
more sophisticated models. Then, the simplest thing to do is to solve another equation for yet another turbulence
quantity. This leads to two-equation models. This includes, for example, the Menter shear stress transport (SST)
model [92], and its variants. These models are widely used; so people must like them. Anyway, I know there are many
other turbulence models, e.g., Reynolds stress model, Smagorinsky model for LES, etc., but honestly speaking, I am
not really an expert in turbulence modeling (although I like it very much). So, I recommend readers to consult other
books such as Ref.[171] for more details.
178 CHAPTER 5. TURBULENCE EQUATIONS
Chapter 6
Methods for Exact Solutions
I like studying methods for deriving exact solutions. Various exact solutions have been used for accuracy verication
in CFD, but it is sometimes dicult to nd detailed descriptions on how they were derived. It is nice to know various
methods for deriving exact solutions so that I can generate an exact solution myself whenever I need one.
6.1 Exact Solutions with Separation of Variables
I like separation of variables because it makes it possible to obtain exact solutions for some partial dierential equations.
It is a very powerful technique.
6.1.1 Laplace Equation
Consider the Laplace equation:
u
xx
+u
yy
= 0. (6.1.1)
In separation of variables, we rst assume the solution of the form,
u = X(x)Y (y), (6.1.2)
and substitute this into the Laplace equation to get
1
X
d
2
X
dx
2
+
1
Y
d
2
Y
dy
2
= 0, (6.1.3)
from which we nd
1
X
d
2
X
dx
2
=
1
Y
d
2
Y
dy
2
= constant. (6.1.4)
So, now the problem is to solve the two ordinary dierential equations. If we want the solution to be periodic in
x-direction, then set the constant to be negative, say, k
2
, and we obtain
X(x) = Asin(kx) +Bcos(kx), (6.1.5)
Y (y) = Ce
ky
+De
ky
, (6.1.6)
where A, B, C, and D are arbitrary constants which are determined by boundary conditions. Therefore, we obtain an
exact solution as
u(x, y) = X(x)Y (y) = [Asin(kx) +Bcos(kx)]
_
Ce
ky
+De
ky
. (6.1.7)
It is, of course, true that any linear combination of the solution with dierent k, i.e.,the Fourier series:
u(x, y) =
k=1
X
k
(x)Y
k
(y) =
k=1
[A
k
sin(kx) +B
k
cos(kx)]
_
C
k
e
ky
+D
k
e
ky
, (6.1.8)
179
180 CHAPTER 6. METHODS FOR EXACT SOLUTIONS
is also an exact solution because the equation is linear. This is called the principle of superposition. It may be useful
for deriving a solution which will t a given domain and boundary conditions. But just one term would be enough for
the purpose of accuracy verication of numerical schemes.
As an example, we consider deriving a solution from the fundamental solution (6.1.7) in a square domain [x, y] =
[0, 1] [0, 1] with the following boundary conditions:
u(x, 0) = u(0, y) = u(1, y) = 0, (6.1.9)
u(x, 1) = sin(x). (6.1.10)
These boundary conditions require that we set
B = 0, C = D, AC =
1
sinh()
, (6.1.11)
in Equation (6.1.7), and thus the exact solution is given by
u(x, y) =
sin(x) sinh(y)
sinh()
. (6.1.12)
Now, consider another set of boundary conditions,
u(0, y) = u(x, 0) = u(x, 1) = 0, (6.1.13)
u(1, y) = sin(y), (6.1.14)
to obtain the exact solution,
u(x, y) =
sin(y) sinh(x)
sinh()
, (6.1.15)
and add this to the previous one to generate a new exact solution,
u(x, y) =
sinh(x) sin(y) + sinh(y) sin(x)
sinh()
, (6.1.16)
which is the exact solution to the following boundary condition:
u(0, y) = u(x, 0) = 0, (6.1.17)
u(x, 1) = sin(x), (6.1.18)
u(1, y) = sin(y). (6.1.19)
This superposition is possible because the Laplace equation is linear. It is very nice. I like linear partial dierential
equations.
6.1.2 Diusion Equation
I like the two-dimensional diusion equation also,
u
t
= (u
xx
+u
yy
) . (6.1.20)
This can also be easily solved by separation of variables. First, we assume
u(x, y, t) = T(t)X(x)Y (y), (6.1.21)
substitute it into the diusion equation to nd
1
T(t)
dT
dt
=
X(x)
d
2
X
dx
2
+
Y (y)
d
2
Y
dy
2
, (6.1.22)
and so by setting
1
T(t)
dT
dt
=
2
(a
2
+b
2
),
1
X(x)
d
2
X
dx
2
= (a)
2
,
1
Y (y)
d
2
Y
dy
2
= (b)
2
, (6.1.23)
6.1. EXACT SOLUTIONS WITH SEPARATION OF VARIABLES 181
where a and b are arbitrary constants, we nd
T(t) = Ce
2
(a
2
+b
2
)t
, (6.1.24)
X(x) = A
1
sin(ax) +A
2
cos(ax), (6.1.25)
Y (y) = B
1
sin(by) +B
2
cos(by), (6.1.26)
where C, A
1
, A
2
, B
1
, and B
2
are arbitrary constants. Therefore, we obtain an exact solution as
u(x, y, t) = T(t)X(x)Y (y) = Ce
2
(a
2
+b
2
)t
[A
1
sin(ax) +A
2
cos(ax)] [B
1
sin(by) +B
2
cos(by)] , (6.1.27)
where C, A
1
, A
2
, B
1
, and B
2
are arbitrary constants which can be determined by specifying boundary conditions. Of
course, any linear combination of the solution with dierent a and b can also be an exact solution. Constants may
be determined to produce a desired solution for given domain and boundary conditions. But I like a simple one, even
simpler than the solution above. For example, I would set A
1
= 1, A
2
= 0, B
1
= 1, and B
2
= 0:
u(x, y, t) = T(t)X(x)Y (y) = Ce
2
(a
2
+b
2
)t
sin(ax) sin(by). (6.1.28)
This solution in a square domain with the Dirichlet boundary condition is enough for verifying the accuracy of a diusion
scheme.
6.1.3 Advection-Diusion Equation
Consider the two-dimensional advection-diusion equation,
u
t
+au
x
+bu
y
= (u
xx
+u
yy
), (6.1.29)
where a, b, and are constants, and a
2
+b
2
= 0. This can be transformed into the following form,
u
t
a
2
+b
2
+u
= (u
+u
), (6.1.30)
where = ax + by and = bx ay (see Section 1.15 for the transformation). Basically, the coordinate system has
been chosen along and normal to the advection direction. So, there is only one advection term present, which is a
derivative along the advection direction. It is always nice to have less terms. I like it. To solve this, I rst assume
u(, , t) = T(t)X()Y (), (6.1.31)
substitute it into the transformed advection-diusion equation (6.1.30) to nd
1
(a
2
+b
2
)T(t)
dT
dt
=
1
X()
dX
d
+
X()
d
2
X
d
2
+
Y ()
d
2
Y
d
2
, (6.1.32)
and so by setting,
1
(a
2
+b
2
)T(t)
dT
dt
=
2
(B
2
A
2
), (6.1.33)
1
X()
dX
d
+
1
X()
d
2
X
d
2
= (A)
2
,
1
Y ()
d
2
Y
d
2
= (B)
2
, (6.1.34)
where A and B are arbitrary constants (the constants on the right hand sides have been chosen to make the solution
periodic in ), we nd
T(t) = C
0
e
2
(a
2
+b
2
)(B
2
A
2
)t
, (6.1.35)
X() = C
1
exp(
1
) +C
2
exp(
2
), Y () = C
3
sin(B) +C
4
cos(B), (6.1.36)
where C
0
, C
1
, C
2
, C
3
, C
4
are arbitrary constants, and
1
=
1 +
1 + 4A
2
2
2
,
2
=
1
1 + 4A
2
2
2
. (6.1.37)
182 CHAPTER 6. METHODS FOR EXACT SOLUTIONS
Therefore, we obtain an exact solution as
u = T(t)X()Y () = C
0
e
2
(a
2
+b
2
)(B
2
A
2
)t
[C
1
exp(
1
) +C
2
exp(
2
)] [C
3
sin(B) +C
4
cos(B)] . (6.1.38)
Note that it is easy to obtain a steady solution by taking A = B (see Section 7.3). This is very nice. But this solution
looks complicated to me. I cannot even think about taking a linear combination of this solution (which is also exact,
though). I like a simple one such as
u = C
0
cos(B) exp
_
1
1 + 4A
2
2
2
_
, (6.1.39)
which has been obtained by taking C
1
= 0, C
2
= 1, C
3
= 0, C
4
= 1. I can now imagine what this solution looks like.
Can you?
6.1.4 Advection-Diusion-Reaction Equation
I like the one-dimensional advection-diusion-reaction equation,
u
t
+au
x
= u
xx
ku, (6.1.40)
where a, , and k are positive constants. I like it because I can easily derive an exact solution. Assume
u(x, t) = T(t)X(x), (6.1.41)
and substitute it into the advection-diusion-reaction equation to nd
1
T(t)
dT
dt
=
a
X(x)
dX
dx
+
X(x)
d
2
X
dx
2
k. (6.1.42)
I like solutions that decay in time, and so I set
1
T(t)
dT
dt
= A
2
a
X(x)
dX
dx
+
X(x)
d
2
X
dx
2
k = A
2
, (6.1.43)
where A is an arbitrary constant. Then, I obtain
T(t) = C
0
e
A
2
t
, (6.1.44)
X(x) = C
1
exp(
1
x) +C
2
exp(
2
x), (6.1.45)
where C
0
, C
1
, C
2
are arbitrary constants, and
1
=
a +
a
2
+ 4(k +A
2
)
2
,
2
=
a
a
2
+ 4(k +A
2
)
2
. (6.1.46)
Therefore, we obtain
u = T(t)X(x) = C
0
e
A
2
t
_
C
1
exp
_
a +
a
2
+ 4(k +A
2
)
2
x
_
+C
2
exp
_
a
a
2
+ 4(k +A
2
)
2
x
__
. (6.1.47)
I would set A
2
= k so that the solution decays at the reaction rate k:
u = C
0
e
kt
_
C
1
exp
_
a +
a
2
+ 8k
2
x
_
+C
2
exp
_
a
a
2
+ 8k
2
x
__
. (6.1.48)
I like this exact solution because a steady solution is easily obtained by taking A = 0 in Equation (6.1.47):
u = C
1
exp
_
a +
a
2
+ 4k
2
x
_
+C
2
exp
_
a
a
2
+ 4k
2
x
_
. (6.1.49)
6.1. EXACT SOLUTIONS WITH SEPARATION OF VARIABLES 183
6.1.5 Oscillating Solutions with Complex Variables: Diusion
Notice that the exact solutions in the previous subsections will decay as time goes on and eventually vanish (because
of the exponential term). It would be nice also to have solutions which keep changing and never vanish. Then, I like
the idea of using the complex variables in separation of variables because it makes it possible to obtain such solutions
[54]. Consider the one-dimensional diusion equation,
u
t
= u
yy
, in y (0, d), (6.1.50)
with the time-dependent boundary conditions,
u(0) = 0, (6.1.51)
u(d) = U cos(t), (6.1.52)
where U and are arbitrary constants. Note that it is very hard to apply the solution obtained in Section 6.1.2 to this
problem because it is not compatible with the boundary condition at y = d. Now, an interesting strategy for deriving
an exact solution to this problem is to rst write the boundary condition as
u(d) = U exp(it), (6.1.53)
where i =
1, so that its real part describes the original condition; then, derive a solution by separation of variables,
which possibly involves complex variables; and nally extract the real part of the solution, which is the exact solution
for the original boundary condition. So, assume u(y, t) = Y (y)T(t) and substitute into the diusion equation to get
1
T(t)
dT
dt
=
Y (y)
d
2
Y
dy
2
= constant. (6.1.54)
The boundary condition suggests that the constant should be i, which leads to
T(t) = C exp(it), (6.1.55)
Y (y) = Aexp(y) +Bexp(y), (6.1.56)
where C, A, B are arbitrary constants and
=
= (1 +i)k, k =
2
, (6.1.57)
(note that
i = [exp(i/2)]
1/2
= exp(i/4) = (1+i)/
1
=
a +
a
2
+ 4i
2
,
2
=
a
a
2
+ 4i
2
. (6.1.68)
Therefore, by applying the boundary conditions, we obtain the following complex solution:
u(y, t) = T(t)Y (y) =
exp(
1
y) exp(
2
y)
exp(
1
d) exp(
2
d)
U exp(it), (6.1.69)
whose real part is the exact solution to the original problem. I havent been able to analytically extract the real part of
the complex solution, but I like it because I can easily compute it numerically and it is useful for testing time-accurate
advection-diusion schemes.
6.2 Exact Solutions with Complex Variables
I like the Cauchy-Riemann system for the velocity potential and the stream function ,
x
+
y
= 0, (6.2.1)
y
= 0, (6.2.2)
because this implies that and are the real and imaginary parts of an analytic function F of a complex variable
Z = x +iy:
F(Z) = (z, y) +i(z, y), (6.2.3)
where i =
1. F is called the complex potential. The velocity is obtained by the derivative of the complex potential:
W(Z) = F
Z
/1
. (6.2.6)
If = , this is a uniform stream with the velocity C which is a constant. If = /2, it represents a ow against
a horizontal wall (stagnation ow).
6.2. EXACT SOLUTIONS WITH COMPLEX VARIABLES 185
(b) Uniform Stream at an angle :
F = U
Ze
i
, (6.2.7)
W = U
e
i
, (6.2.8)
u = U
cos , (6.2.9)
v = U
sin . (6.2.10)
(c) Source/Sink with a strength located at Z = Z
0
= x
0
+iy
0
:
F =
2
ln(Z Z
0
), (6.2.11)
W =
2(Z Z
0
)
, (6.2.12)
u =
2
x x
0
(x x
0
)
2
+ (y y
0
)
2
, (6.2.13)
v =
2
y y
0
(x x
0
)
2
+ (y y
0
)
2
. (6.2.14)
(d) Doublet with a strength m at an angle located at Z = Z
0
= x
0
+iy
0
:
F =
me
i
2(Z Z
0
)
, (6.2.15)
W =
me
i
2(Z Z
0
)
2
, (6.2.16)
u =
m
2
_
(x x
0
)
2
(y y
0
)
2
cos + 2(x x
0
)(y y
0
) sin
[(x x
0
)
2
+ (y y
0
)
2
]
2
, (6.2.17)
v =
m
2
2(x x
0
)(y y
0
) cos
_
(x x
0
)
2
(y y
0
)
2
sin
[(x x
0
)
2
+ (y y
0
)
2
]
2
. (6.2.18)
(e) Clockwise Vortex with a positive circulation located at Z = Z
0
= x
0
+iy
0
:
F =
i
2
ln(Z Z
0
), (6.2.19)
W =
i
2(Z Z
0
)
, (6.2.20)
u =
2
y y
0
(x x
0
)
2
+ (y y
0
)
2
, (6.2.21)
v =
2
x x
0
(x x
0
)
2
+ (y y
0
)
2
. (6.2.22)
(f) Superposition:
Above solutions can be superposed to create a new ow: e.g., a uniform ow around a cylinder (of unit radius) at
the origin is obtained by the sum of the uniform stream and the doublet.
F = U
Z +
U
Z
, (6.2.23)
W = U
Z
2
, (6.2.24)
186 CHAPTER 6. METHODS FOR EXACT SOLUTIONS
u = U
cos 2
r
2
, (6.2.25)
v =
U
sin 2
r
2
. (6.2.26)
Here, the doublet strength m has been set m = 2U
:
v(x, t) = v
+v
(x x
0
, t), (6.3.1)
where x
0
indicates some reference center location of the perturbation eld (which may be the origin). More specically,
here we consider a type of solution in which the perturbation is convected at the mean ow velocity, i.e.,
v(x, t) = v
+v
(x x
0
v
t). (6.3.2)
If such a solution exists, it will be very interesting. The perturbation ow eld will not aect the mean ow at all; it
will be simply convected in the uniform stream. Of course, the density, the pressure, and the temperature will also be
convected together at the mean ow velocity:
(x, t) = (x x
0
v
t), (6.3.3)
p(x, t) = p(x x
0
v
t), (6.3.4)
T(x, t) = T
+T
(x x
0
v
t), (6.3.5)
where only the temperature has been decomposed into the mean T
=
p
=
_
T
T
_ 1
1
,
p
p
=
_
T
T
_
1
, (6.3.6)
which can be used to calculate p and from T. Then, we seek such a solution for the Euler equations:
t
+ div(v) = 0, (6.3.7)
v
t
+ (grad v)v +
1
grad p = 0. (6.3.8)
Note that the energy equation is not needed for homentropic ows. Now, substituting the velocity (6.3.2), the density
(6.3.3), the pressure (6.3.4), and the temperature (6.3.5) into the Euler equations, we obtain
div(v
) = 0, (6.3.9)
(grad v
)v
+
R
1
grad T
= 0, (6.3.10)
6.3. SUPERPOSITION FOR NONLINEAR EQUATIONS 187
where the isentropic relation
dp
=
R
1
dT has been used to replace the pressure gradient by the temperature gradient.
This is interesting. This means that our desired unsteady solution exists if the perturbations satisfy the above steady
Euler equations. A good example is a vortex-type perturbation in two-dimensions:
u
r
= 0, (6.3.11)
u
= f(r), (6.3.12)
where (u
r
, u
) is the perturbation velocity in polar coordinates and f(r) is a function to be determined later. Likewise,
the density, the pressure, and the temperature depend only on r. Then, the continuity equation and the -momentum
equation are identically satised, and we are left with the r-momentum equation,
(u
)
2
r
=
R
1
dT
dr
. (6.3.13)
Integrating this to the far-eld (r ), where the perturbation is assumed to vanish, we get
T
=
1
R
r
f(r)
2
r
dr. (6.3.14)
Now, the form of f(r) needs to be chosen so that this can be fully integrated. A good choice for f(r) is
f(r) =
K
2
re
(1r
2
)/2
, (6.3.15)
which describes a vortex perturbation with the strength K and the rate of decay (it quickly decays to the far-eld
for a large value of ). Integrating Equation (6.3.14) with this f(r), we obtain
T
=
K
2
( 1)
8
2
R
e
(1r
2
)
. (6.3.16)
Using this result, we nally obtain the exact solution of a vortex, initially centered at (x
0
, y
0
), convected in a uniform
ow:
u(x, y, t)
a
=
u
+
u
(x, y)
a
=
u
K
2a
ye
(1r
2
)/2
, (6.3.17)
v(x, y, t)
a
=
v
+
v
(x, y)
a
=
v
+
K
2a
xe
(1r
2
)/2
, (6.3.18)
T(x, y, t)
T
=
T
+
T
(x, y)
T
= 1
K
2
( 1)
8
2
a
2
e
(1r
2
)
, (6.3.19)
(x, y, t)
=
_
T(x, y, t)
T
_ 1
1
, (6.3.20)
p(x, y, t)
p
=
_
T(x, y, t)
T
_
1
, (6.3.21)
where a
,
x = x x
0
u
t, (6.3.22)
y = y y
0
v
t, (6.3.23)
r =
x
2
+y
2
. (6.3.24)
I like this solution very much. It is really interesting that it is possible to superpose a vortex on a uniform ow and
that such a simple ow is an exact solution to the Euler equations. Moreover, other exact solutions could be obtained
in a similar way, i.e., by choosing some other steady solution as a perturbation.
188 CHAPTER 6. METHODS FOR EXACT SOLUTIONS
6.4 Exact Solutions with Transformations
6.4.1 Cole-Hopf Transformations
One-Dimensional Viscous Burgers Equation
Consider the one-dimensional viscous Burgers equation,
u
t
+uu
x
= u
xx
. (6.4.1)
It is well known that the transformation
u = 2
(ln )
x
= 2
, (6.4.2)
called the Cole-Hopf transformation, relates u(x, t) and (x, t) > 0 in such a way that if is a positive solution of the
linear diusion equation,
t
=
xx
, (6.4.3)
then, u is a solution of the Burgers equation (6.4.1). This means that we can generate various exact solutions of the
viscous Burgers equation simply by substituting exact solutions of the linear diusion equation into Equation (6.4.2).
This is nice. I like it. See Ref.[13] for various exact solutions, and also see Section 7.8.1 for a few examples.
Two-Dimensional Viscous Burgers Equation
Consider the two-dimensional viscous Burgers equations,
u
t
+uu
x
+vu
y
(u
xx
+u
yy
) = 0, (6.4.4)
v
t
+uv
x
+vv
y
(v
xx
+v
yy
) = 0. (6.4.5)
Again, we have a nice transformation [37],
u = 2
(ln )
x
= 2
, v = 2
(ln )
y
= 2
, (6.4.6)
which relate u(x, t), v(x, t) and (x, t) > 0 in such a way that if is a positive solution of the linear diusion equation,
t
= (
xx
+
yy
), (6.4.7)
then, u and v are solutions of the Burgers equations (6.4.4) and (6.4.5). See Ref.[37] for examples, or Section 7.8.3
for an example.
Vector Form of the Viscous Burgers Equation (3D)
Consider the vector form of the viscous Burgers equation,
U
t
+ (grad U)U = div grad U, (6.4.8)
where U = [u, v, w]
t
. It is nice to know that there is in fact a vector version of the Cole-Hopf transformation. That
is, the transformation
U = 2 grad (ln ) = 2
grad
, (6.4.9)
relates the solution vector U and a scalar function (x, y, z, t) > 0 in such a way that if is a positive solution of the
linear diusion equation,
t
= div grad , (6.4.10)
then, U is a solution of the vector form of the viscous Burgers equation (6.4.8). This is very nice. I like it very much.
6.4. EXACT SOLUTIONS WITH TRANSFORMATIONS 189
Viscous Burgers Equation with Source Terms
Consider the one-dimensional viscous Burgers equation with various source terms,
u
t
+uu
x
= u
xx
+u
2
ku +d, (6.4.11)
where , , k, and d are constants. It is nice to be able to nd exact solutions to this equation. I like the following
derivation. The equation is transformed by
u = p
(ln )
x
= p
, (6.4.12)
into
[(
t
xx
)
x
+k
x
d/p]
2
+ [
t
(p + 3)
xx
+p
x
]
x
[p + 2](
x
)
3
= 0, (6.4.13)
which is true for any if
(
t
xx
)
x
+k
x
d/p = 0, (6.4.14)
t
(p + 3)
xx
+p
x
= 0, (6.4.15)
p + 2 = 0, (6.4.16)
i.e., p = 2, and
(
t
xx
)
x
+k
x
+
d
2
= 0, (6.4.17)
xx
2
x
= 0, (6.4.18)
thus resulting
2
xx
+k
x
+
d
2
= 0. (6.4.19)
Insert a trial solution = C(t) e
x
, where C(t) is an arbitrary function of time, into this to nd the characteristic
equation,
2
2
+k +
d
2
= 0. (6.4.20)
Then, this can be solved easily, i.e.,
=
k
k
2
4d
4
. (6.4.21)
We now consider three cases: k
2
4d > 0, k
2
4d < 0, and k
2
4d = 0.
(a) Two distinct real roots: k
2
4d > 0
Denote the two distinct real roots by
1
and
2
, i.e.,
1
=
k
k
2
4d
4
,
2
=
k +
k
2
4d
4
, (6.4.22)
then one solution
1
is given by
1
= C
1
(t) e
1x
, (6.4.23)
where C
1
(t) depends only on time, and it can be determined by inserting the solution into Equation (6.4.15) and
solving the resulting ordinary dierential equation:
C
1
(t) = K
1
e
1(1+2)t
, (6.4.24)
where K
1
is an arbitrary constant. Therefore, the solution is given by
1
= K
1
exp [
1
x +
1
(
1
+ 2)t] . (6.4.25)
190 CHAPTER 6. METHODS FOR EXACT SOLUTIONS
Now, we obtain a similar solution for the other root
2
, and superpose the two to get the general solution,
= K
1
exp [
1
x +
1
(
1
+ 2)t] +K
2
exp [
2
x +
2
(
2
+ 2)t] , (6.4.26)
where K
2
is another arbitrary constant. Finally, we obtain an exact solution to the Burgers equation (6.4.11) from
the solution (6.4.12),
u = 2
1
K
1
exp [
1
x +
1
(
1
+ 2)t] +
2
K
2
exp [
2
x +
2
(
2
+ 2)t]
K
1
exp [
1
x +
1
(
1
+ 2)t] +K
2
exp [
2
x +
2
(
2
+ 2)t]
. (6.4.27)
(b) Complex conjugate roots: k
2
4d < 0
Express the complex roots as
1,2
=
r
i
i
, (6.4.28)
where
i =
1,
r
=
k
4
,
i
=
4d k
2
4
, (6.4.29)
then insert these roots directly into Equation (6.4.26), split it into the real and imaginary parts, and form a linear
combination of them to get the general solution,
= e
rx+(
2
r
2
i
+2r)t
[K
1
cos +K
2
sin ] , (6.4.30)
where K
1
and K
2
are arbitrary constants, and
=
i
x + 2(
r
i
+
i
)t. (6.4.31)
Therefore an exact solution is given by
u = 2
= 2
(
r
K
1
+
i
K
2
) cos + (
r
K
i
K
1
) sin
K
1
cos +K
2
sin
. (6.4.32)
(c) Multiple roots: k
2
4d = 0
Denote the single root by
s
,
s
=
k
4
, (6.4.33)
and nd the general solution of Equation (6.4.19),
= e
sx
[C
1
(t) +xC
2
(t)] . (6.4.34)
Insert this into Equation (6.4.15) and solve the resulting system of ordinary dierential equations for C
1
(t) and C
2
(t)
straightforwardly to get
C
1
(t) = 2K
1
(
s
+)t e
s(s+2)t
, (6.4.35)
C
2
(t) = K
2
e
s(s+2)t
, (6.4.36)
where K
1
and K
2
are arbitrary constants, and thus obtain
= e
1x+s(s+2)t
[K
2
x + 2K
1
(
s
+)t] . (6.4.37)
Finally, we obtain an exact solution as follows,
u = 2
= 2
K
2
(1 +
s
x) + 2K
s
(
s
+)t
K
2
x + 2K
1
(
s
+)t
. (6.4.38)
Finally, I remark that the method described above is originally studied and exact solutions of a more general equation
is discussed in Ref.[9].
6.5. EXACT SOLUTIONS FOR LINEAR SYSTEMS OF CONSERVATION LAWS 191
6.4.2 Hodograph Transformation
Consider the compressible continuity equation in two dimensions,
(u)
x
+
(v)
y
= 0. (6.4.39)
This is automatically satised by the stream function dened by
u =
0
y
=
0
y
, (6.4.40)
v =
x
=
x
, (6.4.41)
where the subscript 0 denotes the stagnation state. This stream function is governed by the nonlinear equation,
_
1
_
_
2
2
y
c
2
_
xx
+
_
1
_
_
2
2
x
c
2
_
yy
+ 2
_
_
2
y
c
2
xy
= 0, (6.4.42)
where c is the speed of sound given by
c
2
= c
2
0
1
2
_
_
2
(
2
x
+
2
y
), (6.4.43)
and
=
_
1 +
1
2
_
_
2
2
x
+
2
y
c
2
_ 1
1
. (6.4.44)
These are basically derived from the irrotationality condition v
x
u
y
= 0 and the momentum equations (see Ref.[138]).
Obviously, it is not very easy to solve this nonlinear equation; it looks even more complicated than the nonlinear potential
equation (3.19.14). However, there is a nice way to get around it. That is, we exchange the independent variables
(x, y) and the dependent variables (V, ), where V is the ow speed and is the ow angle, and transform the equation
(6.4.42) into
V
2
2
V
2
+V
_
1 +
V
2
c
2
_
V
+
_
1
V
2
c
2
_
2
= 0, (6.4.45)
(see Ref.[138] for derivation). This is a linear dierential equation in and therefore can be solved relatively easily.
In fact, to nd analytical solutions, we dont really need to solve this in a formal way. Pick any function and see if
it satises the equation. If it does, then it is a solution. Moreover, any linear combination of such solutions (i.e.,
superposition) is also a solution because the equation is linear. This is very nice. Especially, I like the solution called
Ringlebs ow described in Section 7.13.6. See Ref.[138] for other examples.
6.5 Exact Solutions for Linear Systems of Conservation Laws
Consider the linear hyperbolic system of one-dimensional conservation laws of the form,
U
t
+A
U
x
= 0, (6.5.1)
where U = U(x, t) and Ais a constant coecient matrix having real eigenvalues and linearly independent eigenvectors.
If we are given an initial solution as
U(x, 0) = U
0
(x), (6.5.2)
then, we can nd the exact solution as follows. First, we multiply Equation (6.5.1) by the left-eigenvector matrix L
(of A) from the left,
L
U
t
+LARL
U
x
= 0, (6.5.3)
192 CHAPTER 6. METHODS FOR EXACT SOLUTIONS
where R is the right-eigenvector matrix (the inverse of L). Because the matrix L is constant, we can write
(LU)
t
+LAR
(LU)
x
= 0. (6.5.4)
Dening the characteristic variables (or the Riemann invariants) by
W = LU, (6.5.5)
and the diagonal matrix by = LAR, we arrive at the diagonalized form of the system (6.5.1).
W
t
+
W
x
= 0. (6.5.6)
We point out that each component is now a linear scalar advection equation
w
k
t
+
k
w
k
x
= 0, (6.5.7)
where w
k
denotes the k-th component of W and
k
is the k-th eigenvalue. Therefore, the exact solution of the k-th
component is given simply by
w
k
(x, t) = w
k
(x
k
t, 0). (6.5.8)
Then, by denition (6.5.5), we obtain
U(x, t) = RW(x, t) =
k
w
k
(x, t)r
k
=
k
w
k
(x
k
t, 0)r
k
, (6.5.9)
or equivalently, (by Equation (6.5.5) again),
U(x, t) =
k
(
k
U
0
(x
k
t))r
k
, (6.5.10)
where
k
is the k-th row of L (the k-th left-eigenvector) and r
k
is the k-th column of R (the k-th right-eigenvector).
This is the exact solution for the initial solution (6.5.2).
Now we consider a simple wave. A simple wave is dened as a solution whose variation is conned in a one-
dimensional subspace spanned by a single eigenvector. For example, if the initial solution is a j-th simple wave, we
project it onto the j-th eigenspace to write
U
0
(x) =
k
(
k
U
0
(x))r
k
= (
j
U
0
(x))r
j
+
k:k=j
c
k
r
k
, (6.5.11)
where c
k
are constant. Then, the exact solution (6.5.10) can be written as
U(x, t) =
k
(
k
U
0
(x
j
t))r
k
= (
j
U
0
(x
j
t))r
j
+
k:k=j
c
k
r
k
= U
0
(x
j
t). (6.5.12)
Therefore, the initial solution is preserved perfectly for a simple wave. This is natural because only one advection speed
is relevant in a simple wave. Finally, we give an example: set the following initial solution,
U
0
(x) = U+ sin(x) r
j
, (6.5.13)
where U is a constant state and is an amplitude of the sine wave, then the exact solution is given by
U(x, t) = U
0
(x
j
t) = U+ sin(x
j
t) r
j
. (6.5.14)
It is indeed simple and also general enough to be applicable to arbitrary linear hyperbolic systems of the form Equation
(6.5.1). Particularly, I like the fact that the result (and all the above discussion) is valid even for steady systems, i.e.,
Equation (6.5.1) with (x, t) replaced by (x, y) or (y, x). This is very nice.
6.6. SIMPLE WAVE SOLUTIONS FOR NONLINEAR SYSTEMS OF CONSERVATION LAWS 193
6.6 Simple Wave Solutions for Nonlinear Systems of Conservation Laws
Consider the nonlinear hyperbolic system of one-dimensional conservation laws of the form,
U
t
+
F(U)
x
= 0, (6.6.1)
or
U
t
+A(U)
U
x
= 0, (6.6.2)
where U = U(x, t) and A(U) is a coecient matrix which is diagonalizable but no longer constant. Unlike the linear
case, for a given initial solution such as
U(x, 0) = U
0
(x), (6.6.3)
we cannot nd the exact solution generally in a closed form. To see this, as in the linear case, we multiply Equation
(6.6.2) by the left-eigenvector matrix L from the left,
L
U
t
+LARL
U
x
= 0. (6.6.4)
Here is a diculty. We can no longer dene the characteristic variables as in Equation (6.5.5) because L is not constant
any more. Instead, I dene the characteristic variables by
W = LU, (6.6.5)
and arrive at the diagonalized form of the system.
W
t
+
W
x
= 0. (6.6.6)
Each component is now a nonlinear scalar advection equation with the exact solution,
w
k
(x, t) = w
k
(x
k
t, 0), (6.6.7)
which is valid up to the time when a shock forms. Note that this solution is given implicitly in general because
k
depends on the solution. Then, we use Equation (6.6.5) to get the solution in the conservative variables, (6.6.5)
U(x, t) = RW(x, t) =
k
w
k
(x, t)r
k
=
k
w
k
(x
k
t, 0)r
k
, (6.6.8)
and so
U(x, t) =
k
(
k
U
0
(x
k
t))r
k
. (6.6.9)
I like it, but it is too bad that we cannot get U(x, t) unless this is integrable. We cannot project even the initial
solution onto the space of an eigenvector exactly:
U
0
(x) =
k
(
k
U
0
(x))r
k
. (6.6.10)
Then, we consider a simple wave solution because it gives us a hope to get exact solutions. Let h(x, t) be any
smooth function that satises the j-th component of Equation (6.6.6),
h
t
+
j
(U)
h
x
= 0, (6.6.11)
so that h(x, t) is the j-th characteristic variable w
j
(x, t) = h(x
j
(U) t) = h(), and assume that other characteristic
variables are constant. Then, we have from Equation (6.6.5)
U(x, t) = r
j
(U) h(), (6.6.12)
or because U(x, t) = U(), we may write
dU()
d
= r
j
(U)
dh()
d
. (6.6.13)
If this is integrable, we can nd an exact solution U for the j-th simple wave. This is possible in some cases. Note
that the variable U can be taken to be any variable (e.g., conservative or primitive) as long as the eigenvector r
j
is
dened consistently. Some examples are given in the next section.
194 CHAPTER 6. METHODS FOR EXACT SOLUTIONS
6.7 Some Exact Simple Wave Solutions
6.7.1 Case 1: r
j
= constant
I like eigenvectors that are independent of the solution because the ordinary dierential equation (6.6.13) is then trivial.
We readily obtain a simple wave solution as
U() = h() r
j
+ constant. (6.7.1)
The entropy wave in the Euler equations is a good example: the eigenvector, based on the primitive variables V =
[, u, p]
t
, is given by
r
j
=
_
_
1
0
0
_
_
, (6.7.2)
with eigenvalue
j
= u. So, it follows immediately from Equation (6.7.1) that
= h(x ut) + constant, (6.7.3)
u = constant, (6.7.4)
p = constant. (6.7.5)
This shows that the wave does not deform because the speed u is constant (the entropy wave is linearly degenerate).
Any function (even a discontinuous one) can be chosen for h(x). It will be simply convected at the velocity u.
6.7.2 Case 2: r
j
U
Consider the case where the eigenvector is proportional to the solution vector,
r
j
= MU, (6.7.6)
where M is a constant matrix. In this case, we have from Equation (6.6.13),
dU()
d
= MU
dh()
d
= MUh
(). (6.7.7)
This is a linear system of ordinary dierential equations and therefore can be solved analytically by a standard technique.
This is nice. A good example is the Alfvn wave of the ideal magnetohydrodynamic equations. For the primitive variables
U = [, v
x
, v
y
, v
z
, B
x
, B
y
, B
z
, p]
t
, where (v
x
, v
y
, v
z
) is the velocity vector and (B
x
, B
y
, B
z
) is the magnetic eld, with
the coecient matrix,
A =
_
_
v
x
0 0 0 0 0
0 v
x
0 0
B
y
4
B
z
4
1
0 0 v
x
0
B
x
4
0 0
0 0 0 v
x
0
B
x
4
0
0 B
y
B
x
0 v
x
0 0
0 B
z
0 B
x
0 v
x
0
0 a
2
0 0 0 0 v
x
_
_
, (6.7.8)
6.7. SOME EXACT SIMPLE WAVE SOLUTIONS 195
where a is the speed of sound (a
2
= p/), the eigenvector for the Alfvn waves is given by
r
j
=
_
_
0
0
B
z
B
y
4 B
z
4 B
y
0
_
_
, (6.7.9)
with the eigenvalues
j
= v
x
B
x
4
. (6.7.10)
Note that zeroes in Equation (6.7.9) correspond to the density, the velocity component v
x
, and the pressure. This
means that these variables are constant through the Alfvn waves (see Section 1.16.2). Therefore, the eigenvalues
are constant; this wave is linearly degenerate. Now, because the density is constant, the eigenvector is linear in the
primitive variable,
r
j
= MU, (6.7.11)
where U is taken to be the primitive variable and
M=
_
_
0 0 0 0 0 0 0
0 0 0 0 0 0 0
0 0 0 0 0 0 1
0 0 0 0 0 1 0
0 0 0 0 0 0 0
0 0 0 0 0 0
4
0 0 0 0 0
4 0
_
_
. (6.7.12)
Since only four variables, v
y
, v
z
, p, B
y
, B
z
, are relevant, we ignore other components in the following. Now, we derive
an exact solution by solving Equation (6.6.13):
d
d
_
_
v
y
v
z
B
y
B
z
_
_
=
_
_
0 0 0 1
0 0 1 0
0 0 0
4
0 0
4 0
_
_
_
_
v
y
v
z
B
y
B
z
_
_
h
(). (6.7.13)
The standard diagonalization yields the decoupled system,
d(v
y
B
y
/
4)
d
= 0, (6.7.14)
d(v
z
B
z
/
4)
d
= 0, (6.7.15)
d(B
z
+i B
y
)
d
= i a (B
z
+i B
y
) h
(), (6.7.16)
d(B
z
i B
y
)
d
= i a (B
z
i B
y
) h
(), (6.7.17)
whose general solution is
v
y
() = B
y
()/
4 +c
vy
, (6.7.18)
v
z
() = B
z
()/
4 +c
vz
, (6.7.19)
B
y
() = K
R
cos(
4 h()) +K
I
sin(
4 h()), (6.7.20)
B
z
() = K
I
cos(
4 h()) K
R
sin(
4 h()), (6.7.21)
196 CHAPTER 6. METHODS FOR EXACT SOLUTIONS
where c
vy
, c
vz
, K
R
and K
I
are arbitrary constants. This is the exact simple wave solution for the Alfvn wave in the
most general form. Note that we can use only one solution, i.e., we can take only either positive or negative sign in
the formula; otherwise it will not be a simple wave. Here is an example: take h() = /
4 and K
I
= 0, and obtain
the simple wave associated with
+
j
,
v
y
(x
+
j
t) = K
R
cos(x
+
j
t) +c
vy
, (6.7.22)
v
z
(x
+
j
t) = K
R
sin(x
+
j
t) +c
vz
, (6.7.23)
B
y
(x
+
j
t) =
4 K
R
cos(x
+
j
t), (6.7.24)
B
z
(x
+
j
t) =
4 K
R
sin(x
+
j
t), (6.7.25)
where c
vy
, c
vz
, and K
R
are arbitrary constants, and K
R
=
4 K
R
. I like this simple Alfvn wave solution. It is
linearly degenerate, and therefore the wave must be preserved at all times. This can be a good test case for code
verication.
6.7.3 Case 3: Other cases
The acoustic wave in the Euler equations is a good example. In terms of the primitive variables, the acoustic eigenvectors
are given by
r
j
=
_
_
/a
1
a
_
_
, (6.7.26)
where a is the speed of sound. The associated eigenvalues are
j
= u a. (6.7.27)
It must be kept in mind that there are two acoustic waves, + and , but we can take only either + or for this
solution to be a simple wave. Now, we insert these into Equation (6.6.13) to get
d
d
_
_
u
p
_
_
=
_
_
/a
1
a
_
_
dh()
d
. (6.7.28)
This is nonlinear, except for the velocity which can be integrated easily,
u
u
du =
h
h
dh, (6.7.29)
to give
u() = h() +u
, (6.7.30)
where we have set h
= 0 so that u
= p
. (6.7.31)
We take advantage of this adiabatic relation to write the speed of sound in terms of the density as
a
a
=
_
_
1
2
, (6.7.32)
which is substituted into the rst component of Equation (6.7.28) to yield
d
d
=
_
_
3
2
dh()
d
. (6.7.33)
This can be integrated easily,
/
1
_
_
3
2
d
_
_
=
1
a
h
h=0
dh, (6.7.34)
6.7. SOME EXACT SIMPLE WAVE SOLUTIONS 197
resulting
()
=
_
1
1
2
h()
a
_ 2
1
, (6.7.35)
or by Equation (6.7.30)
()
=
_
1
1
2
_
u()
a
__ 2
1
, (6.7.36)
where M
= u
/a
=
_
()
=
_
1
1
2
_
u()
a
__
2
1
. (6.7.37)
This completes the derivation of the exact simple wave solution for the acoustic wave. Now, we summarize the exact
acoustic simple wave solution:
u(x
j
t)
a
= M
+
h(x
j
t)
a
, (6.7.38)
(x
j
t)
=
_
1
1
2
_
u(x
j
t)
a
__ 2
1
, (6.7.39)
p(x
j
t)
p
=
_
1
1
2
_
u(x
j
t)
a
__
2
1
, (6.7.40)
where
j
= u a and h is an arbitrary function which must be chosen to keep the density and the pressure positive.
This is a very nice solution, but it is given implicitly, i.e.,
j
depends on the solution itself, and also it is valid only
when there are no shocks.
We point out here that the simple wave solution satises Burgers equation as shown in Section 3.4.3. For example,
the simple wave solution corresponding to
+
j
= u +a satises Equation (3.4.36) which is repeated here,
V
t
+V V
x
= 0, (6.7.41)
where V = u +a. I like this because V = u +a can be written in terms of u only, by using the fact that u
2
1
a =
u
2
1
a
= constant:
V =
+ 1
2
u +
_
a
1
2
u
_
, (6.7.42)
so that, by solving this for u, a solution V of Burgers equation (6.7.41) can be translated into the simple wave solution
of the Euler equation by
u()
a
=
2
( + 1)
V ()
a
2
+ 1
_
1
1
2
M
_
, (6.7.43)
where = x V t, and the density and the pressure follow from Equations (6.7.39) and (6.7.40). Again, V must be
chosen such that the density and the pressure stay positive. I like this solution because basically I can add any constant
to u (and still V satises Burgers equation), and so we may set up a solution as
u()
a
= M
+
2
( + 1)
V ()
a
, (6.7.44)
i.e., we use the solution V of Burgers equation as a perturbation to the mean ow. Compare this with Equation
(6.7.38) and see how you can determine h() by a solution V () of Burgers equation. See Section 7.13.1 for some
examples.
The bottom-line is that the exact solution can be obtained if Equation (6.6.5) is integrable so that the Riemann
invariants can be dened explicitly. For example, the shallow-water equations and the isothermal Euler equations have
this property. I like such equations.
198 CHAPTER 6. METHODS FOR EXACT SOLUTIONS
6.8 Manufactured Solutions
Pick any function, substitute it into a governing equation you wish to solve, and dene whatever left as a source
term. This way, you can cook up an exact solution to any equations (by introducing a source term). It is so simple. I
like it very much. Many people call such solutions manufactured solutions [127]. Here are some simple examples.
(a) 1D Linear Advection:
Substitute u = sin x into the advection equation (plus a possible source term f), au
x
= f, to get
a cos x = f, (6.8.1)
which suggests to dene f = a cos x. Therefore, we can say that the advection equation,
au
x
= f with f = a cos x, (6.8.2)
has the following exact solution,
u = sin x. (6.8.3)
(b) Poisson Equation:
Substitute u = sin x + cos y into the Poisson equation, u
xx
+u
yy
= f, to get
sin x cos y = f, (6.8.4)
which suggests to dene f = sin x cos y. Therefore, the Poisson equation,
u
xx
+u
yy
= f with f = sin x cos y, (6.8.5)
has the following exact solution,
u = sin x + cos y. (6.8.6)
I like cooking up exact solutions this way, but it always introduces a source term, i.e., an additional complication
to the original equation. Honestly speaking, I dont really like source terms because a scheme could lose its formal
accuracy unless the source term discretization is carefully designed [47, 82, 174].
(c) 2D Linear Advection:
In this example, we manufacture a solution by choosing appropriate coecients rather than introducing a source
term. Substitute u = x
2
+y
2
into the two-dimensional advection equation, au
x
+bu
y
= 0, to get
2xa + 2yb = 0, (6.8.7)
which can be true if we take a = y and b = x. Hence, the advection equation,
yu
x
xu
y
= 0, (6.8.8)
has the following exact solution,
u = x
2
+y
2
. (6.8.9)
After all, I like the method of manufactured solutions very much because it is very convenient and useful especially
for complicated equations, such as the Navier-Stokes equations [83, 133] and turbulence models [34, 135]. Also note that
exact solutions for compressible ow equations can be manufactured by substituting exact incompressible ow solutions
into compressible ow equations and generating appropriate source terms. This will be useful for manufacturing a
physically realistic exact solution, which does look like a ow. It is very important to have a physically meaningful
manufactured solution because a solution will not be useful if it gives a negative viscosity, for example.
6.9. FINE GRID SOLUTIONS 199
6.9 Fine Grid Solutions
Some people use a numerical solution on a very ne grid as an exact solution: for example, compute the error of
a numerical solution against the ne grid solution, and estimate the order of accuracy of a scheme. I like the idea, but
this is wrong because the ne grid solution is not an exact solution. Such an error convergence study shows only how
quickly the solution converge to this particular ne grid solution, not to the exact solution. Note that the error or the
order of accuracy would not make sense at all if the scheme turned out to be inconsistent. You would then only know
how close your solution is to a wrong solution or how quickly your solution converges to the wrong solution. Anyway,
always remember that a ne grid solution is not an exact solution.
200 CHAPTER 6. METHODS FOR EXACT SOLUTIONS
Chapter 7
Exact Solutions
I like exact solutions because I can directly measure the errors of numerical solutions. In doing so, in many cases, the
domain can be in any shape, and the boundary condition can be specied directly by using the exact solutions (the
Dirichlet condition). Also, a regular domain may be used with a periodic boundary condition so that any boundary
issues can be avoided. In either way, exact solutions are very convenient and useful for code verication. Note that the
verication is to check if a numerical method is correctly implemented, i.e., to see if it is solving the target equation
with a design accuracy. Even if a code has been veried, it still needs to be validated, for example, with experimental
data, i.e., to see if it is solving the right equation. Exact solutions are useful generally for verication, not for validation.
See Ref.[127] for more details on the dierence between verication and validation.
7.1 Linear Advection
7.1.1 1D Linear Advection
I like the linear advection equation,
u
t
+au
x
= 0, (7.1.1)
where a is a constant, because it simply means that any initial solution is merely convected at the speed a, preserving
its initial prole. The exact solution in the general form is therefore given by
u = f(x at), (7.1.2)
for the initial solution u(x, 0) = f(x). Note that we can avoid, in many cases, any boundary issues by using a periodic
boundary condition: u(1) = u(0) in the domain x [0, 1]. I like this periodic boundary condition very much. It is very
interesting that it can be interpreted as transforming the plane domain into a cylinder surface by rolling up the plane
with respect to the vertical axis and connecting the left and right boundaries. Here are some examples.
(a) Smooth solution (Figure 7.1.1):
u = sin[2(x at)], (7.1.3)
for the initial solution u(x, 0) = sin(2x).
(b) Steep solution (Figure 7.1.2):
u = exp
_
(x x
0
at)
2
d
_
, (7.1.4)
for the initial solution u(x, 0) = exp
_
(xx0)
2
d
_
where d is a constant, e.g., d = 0.01.
201
202 CHAPTER 7. EXACT SOLUTIONS
0 1
-1
0
1
dx/dt = a
x
u
Figure 7.1.1: Smooth solution.
0 1
0
1
dx/dt = a
x
u
Figure 7.1.2: Steep solution.
0 1
0
1
dx/dt = a
x
u
Figure 7.1.3: Discontinuous Solu-
tion.
(c) Discontinuous solution (Figure 7.1.3):
u =
_
1 if x
L
+at < x < x
R
+at,
0 elsewhere,
(7.1.5)
where x
L
and x
R
are arbitrary constants that specify the width of the non-zero part of the solution. Remember
that discontinuous solutions do not satisfy the dierential form of the advection equation but do satisfy the integral
form (2.1.8).
7.1.2 2D Linear Advection
Of course, I like the linear advection equation in two dimensions also,
u
t
+au
x
+bu
y
= 0, (7.1.6)
where (a, b) is a constant advection vector. Any initial solution is simply convected at the velocity (a, b), preserving
its initial prole. The exact solution in the general form is given by
u(x, y, t) = f(x at, y bt), (7.1.7)
for the initial solution of the form, u(x, y, 0) = f(x, y). Here are some examples.
(a) One-dimensional wave solution: (a, b) = (a, 0),
u(x, y, t) = U
x
=
x
+
x
= a
+b
, (7.1.13)
y
=
y
+
y
= b
, (7.1.14)
and
a
x
+b
y
= (a
2
+b
2
)
, (7.1.15)
because in these coordinates, the advection equation (7.1.6) is written as
u
t
+ (a
2
+b
2
)u
= 0, (7.1.16)
so that the steady equation is simply
u
= 0. (7.1.17)
Hence, any function of is an exact solution to the two-dimensional linear advection equation: e.g.,
u =
3
+
2
+ + 1, u = sin(), u = exp(). (7.1.18)
This is very nice, isnt it? I like it so much. Note that boundary values must be specied on inow boundaries,
according to the ow direction (a, b).
7.1.3 2D Circular Advection
The circular advection equation,
u
t
+au
x
+bu
y
= 0, (7.1.19)
where (a, b) = (y, x), describes a clockwise circular advection around the origin at the angular speed rad/sec.
I like it because it is a very simple and good test case to see how long a numerical scheme can accurately preserve an
initial prole. Here are some examples.
(a) Rotating Smooth Cone:
In the domain [1, 1] [1, 1], start with the initial solution,
u(x, y, 0) =
_
exp
r
2
/d
for r 0.25,
0 otherwise,
(7.1.20)
where r =
(x + 0.25)
2
+y
2
and (a, b) = (2y, 2x). The cone returns to its initial position at t = 1.
204 CHAPTER 7. EXACT SOLUTIONS
(b) Rotating Cylinder:
In the same setting as above, we may set instead
u(x, y, 0) =
_
1 for r 0.25,
0 otherwise,
(7.1.21)
to rotate a cylinder which is discontinuous and more dicult compute accurately than a cone.
(c) Steady solutions:
Here is a steady test case where the initial solution is set up as
u(x, 0, 0) =
_
f(x) for 0.65 x 0.35,
0 elsewhere,
(7.1.22)
in the domain [1, 1] [0, 1]. The initial prole f(x) is then convected in a clockwise circular direction; the exact
steady solution is given, in the polar coordinates (r, ), by
u(r, ) =
_
f(r) for 0.65 r 0.35,
0 elsewhere.
(7.1.23)
At a steady state, for example, we may compare the solution at the outow boundary where x > 0 and y = 0 with
the initial prole to see how well its initial shape has been preserved. Here are examples of f(x):
(c-1) Smooth solution:
f(x) = exp
(x+0.5)
2
/0.01
. (7.1.24)
(c-2) Discontinuous solution:
f(x) =
_
1 for 0.65 x 0.35,
0 elsewhere.
(7.1.25)
I like the smooth one better because the numerical error should converge to zero as I rene the mesh whereas
it will not converge generally if the solution is discontinuous. But, of course, if I am interested to study the
non-oscillatory behavior of numerical solutions, I would use the discontinuous one for which numerical oscillations
are more likely to occur.
7.1.4 3D Circular (Helicoidal) Advection
I like the 3D circular advection equation,
u
t
+au
x
+bu
y
+cu
z
= 0, (7.1.26)
where
(a, b, c) = (z, V, x), (7.1.27)
and > 0 and V > 0. This describes a ow rotating clockwise around (at the angular speed rad/sec) and
simultaneously transported (at the speed V ) along the y-axis. So, any initial solution imposed on the (x, y)-plane at
z = 0 is rotated around and transported along the y-axis. For example, in the domain dened by
(x, y, z) [1, 0] [0, 1] [0, 1], (7.1.28)
if we specify the initial solution,
u(x, y, 0) = exp
_
50
_
(x + 0.5)
2
+ (y 0.5)
2
_
, (7.1.29)
in the (x, y)-plane at z = 0, then the exact steady solution in the (y, z)-plane at x = 0 is given by
u(0, y, z) = exp
_
50
_
(y 0.5 V )
2
+ (z 0.5)
2
_
. (7.1.30)
It is a very simple and useful solution for evaluating the accuracy of a three-dimensional advection scheme.
7.2. DIFFUSION EQUATION 205
0 1
-1
0
1
t=0
t=0.2
t=0.6
x
u
Figure 7.1.5: Unsteady solution of the 1D diusion
equation. n = 2.
0 10
1
0
1
x
u
t=0.0
t=0.25
t=0.5
t=0.75
Figure 7.1.6: Unsteady solution that oscillates forever.
U
0
= 1 and = 2.
7.2 Diusion Equation
I like the diusion equation because it is easy to obtain exact solutions by separation of variables (see Section 6.1).
7.2.1 1D Diusion Equation
The one-dimensional diusion equation,
u
t
= u
xx
, in (0, 1), (7.2.1)
has the following exact solutions.
(a) Time dependent solution I (Figure 7.1.5):
u = e
n
2
2
t
sin(nx), (7.2.2)
for the initial solution u(x, 0) = sin(nx) with the boundary conditions u(0, t) = u(1, t) = 0. where n is an
integer. I like this simple solution, but if you are looking for something more involved, you can generate it yourself
by superposition (see Section 6.1).
(b) Time dependent solution II:
u(x, t) =
A
t +B
e
(x1)
2
4(t+B)
, (7.2.3)
where A, B and x
0
are arbitrary constants. for the initial solution and the boundary conditions:
u(x, 0) =
A
B
e
(x1)
2
4B
, u(0, t) =
A
t +B
e
1
4(t+B)
, u(1, t) =
A
t +B
. (7.2.4)
Note that the boundary condition changes in time at both ends. I like it. This type of solution is used for accuracy
verication of high-order schemes [119].
(c) Time dependent solution that will never decay to zero:
As derived in Section 6.1.5, there exists a time-dependent solution that oscillates forever:
u(x, t) = Q[sinh(kx) cos(kx) sinh(k) cos(k) + cosh(kx) sin(kx) cosh(k) sin(k)] cos(t)
+Q[sinh(kx) cos(kx) cosh(k) sin(k) cosh(kx) sin(kx) sinh(k) cos(k)] sin(t), (7.2.5)
206 CHAPTER 7. EXACT SOLUTIONS
where k and are arbitrary positive constants, and
Q =
U
cosh
2
(k) cos
2
(k)
. (7.2.6)
It satises the diusion equation for the following boundary conditions:
u(0) = 0, (7.2.7)
u(1) = U cos(t). (7.2.8)
I like this one since it can be a good test case for studying a long time behavior of a diusion scheme.
(d) Another time dependent solution that will never decay to zero:
There exists another time-dependent solution that oscillates forever:
u(x, t) = U exp
_
2
x
_
cos
_
_
t
1
2
x
__
, (7.2.9)
where U and are arbitrary constants. This exact solution can be easily derived by a trial solution of the form:
u = Ue
i(t+x)
, (7.2.10)
where is determined so as to satisfy the diusion equation
1
. That is, you substitute it into the diusion equation
and solve the resulting equation for . I like this solution, especially, the way it is derived. You pick a function
and make it satisfy the target equation. Thats a wonderful way of deriving exact solutions. Its like you are
creating a solution rather than solving the equation. Of course, the solution itself is also very interesting in that it
represents a traveling wave entering the domain at x = 0 and quickly damped towards x . See Figure 7.1.6.
The dierence between the previous solution (7.2.5) and this solution (7.2.9) lies in the boundary condition: the
former has u = 0 precisely at x = 0 whereas the latter has u = 0 theoretically at x . At the other end, both
solutions have the same boundary condition: u = U cos(t).
(e) Steady solution (a simple linear function):
u(x) = (1 x)u
L
+xu
R
, (7.2.11)
with the boundary conditions u(0) = u
L
and u(1) = u
R
. This is too simple. Any diusion scheme should be able
to compute this exactly.
7.2.2 2D Diusion Equation
The two-dimensional diusion equation,
u
t
= (u
xx
+u
yy
), (7.2.12)
has exact solutions of the form,
u(x, y, t) = e
2
(n
2
+m
2
)t
sin(nx) sin(my), (7.2.13)
where n and m are integers so that u = 0 on the boundary. It is straightforward to set up the problem in a square
domain. But sometimes it is too simple. Then, I would consider the following highly anisotropic problem,
u
t
= (u
xx
+u
yy
), (x, y) (0, 1) (0, 0.005), (7.2.14)
with the initial solution,
u(x, y, 0) = 5 sin(x) sin(4000y), (7.2.15)
and the boundary condition, u = 0 on the boundary. The exact solution is given by
u(x, y, t) = 5 exp(16000001
2
t) sin(x) sin(4000y). (7.2.16)
1
This derivation is due to Professor P. L. Roe.
7.3. ADVECTION-DIFFUSION EQUATION 207
Figure 7.2.1: Anisotropic solution of the unsteady diusion equation (t = 0).
The solution at t = 0 is shown in Figure 7.2.1. This is a very good test case for diusion schemes because it involves
high-aspect-ratio and possibly highly-skewed grids as typically encountered in practical CFD computations for viscous
problems. Many diusion schemes work very well for isotropic problems, but not necessarily for this problem. In
Refs.[101, 108], this problem is used to illustrate numerical diculties such as oscillations and instability with highly
stretched and skewed grids.
Various other exact solutions can be obtained by superposition as mentioned earlier in Section 6.1. I like the two-
dimensional diusion equation especially because it is nothing but the Laplace equation at a steady state and therefore
it has many many exact steady solutions (see Section 7.9).
7.2.3 3D Diusion Equation
The three-dimensional diusion equation,
u
t
= (u
xx
+u
yy
+u
zz
), in (0, 1) (0, 1) (0, 1), (7.2.17)
has exact solutions such as
u(x, y, z, t) = e
2
(n
2
+m
2
+l
2
)t
sin(nx) sin(my) sin(lz), (7.2.18)
where n, m, and l are integers so that u = 0 on the boundary. Other exact solutions can be obtained by superposition
as mentioned earlier in Section 6.1. Of course, I like this equation because it is again nothing but the Laplace equation
at a steady state and it has many many many exact solutions (see Section 7.9). Also, I like highly anisotropic problems,
which can be easily constructed from the above solution as in the two-dimensional case.
7.3 Advection-Diusion Equation
7.3.1 1D Advection-Diusion Equation
The one-dimensional advection-diusion equation,
u
t
+au
x
= u
xx
, (7.3.1)
where a and (> 0) are constants, has the following exact solutions.
(a) Time-dependent solution:
u(x, t) = exp(k
2
t) sin k(x at) in [0, 1], (7.3.2)
where k is an arbitrary constant, for the initial condition u(x) = sin(kx) with a periodic boundary condition.
208 CHAPTER 7. EXACT SOLUTIONS
0 1
0
1
Re=10.0
x
u
Figure 7.3.1: Boundary Layer.
0 1
0
1
Re=25
x
u
Figure 7.3.2: Friedrichs model.
c = 0.3.
0 1
0
1
Re=25
x
u
Figure 7.3.3: Manufactured solu-
tion. C = 1/R
e
.
(b) Time-dependent solution that will never decay to zero:
As derived in Section 6.1.6, there exists an exact solution that oscillates forever:
u(x, t) = R
_
exp(
1
x) exp(
2
x)
exp(
1
) exp(
2
)
U exp(it)
_
, (7.3.3)
where i =
1, R indicates the real part of the complex solution (the expression in the square brackets), U and
are arbitrary constants, and
1
=
a +
a
2
+ 4i
2
,
2
=
a
a
2
+ 4i
2
. (7.3.4)
The boundary conditions are
u(0) = 0, (7.3.5)
u(1) = U cos(t). (7.3.6)
I like this one because it can be useful for studying a long-time behavior of time-accurate advection-diusion
schemes. I would like it better if it were possible to extract the real part of the solution analytically.
(c) Steady boundary-layer type solution (Figure 7.3.1):
u(x) =
1 e
xRe
1 e
Re
in [0, 1], (7.3.7)
where Re = a/, with the boundary conditions u(0) = 0 and u(1) = 1.
(d) Steady boundary-layer type solution (Figure 7.3.2):
The Friedrichs model is given by
u
xx
+u
x
c = 0 in (0, 1), (7.3.8)
where c is an arbitrary constant (modeling the pressure gradient). It has the exact solution,
u(x) = cx + (1 c)
1 e
xRe
1 e
Re
, (7.3.9)
where Re = 1/, with the boundary conditions u(0) = 0 and u(1) = 1. Particularly I like this one because it
has an interesting geometrical interpretation. The solution curve in the three-dimensional space with coordinates
(u, x, p = u
x
) is an intersection of two surfaces dened by
p +u cx = constant, (7.3.10)
x + ln(p c) = constant, (7.3.11)
where the constants are determined by the boundary condition. See Ref.[107] for details. The solution shown in
Figure 7.3.2 is a projection of this curve onto (x, u)-plane.
7.3. ADVECTION-DIFFUSION EQUATION 209
(a) t = 0.0 (b) t = 0.05 (c) t = 0.1
Figure 7.3.4: Unsteady solution (7.3.16): C = 0.01, = 0.5, a = 1, b = 0, A = 0.5, B = 2.0.
(a) t = 0.0 (b) t = 0.5 (c) t = 1.0
Figure 7.3.5: Unsteady solution (7.3.17): C = 50.0, = 0.15, a = 2.5, b = 2.5, c
1
= 1.0, c
2
= 1.0.
(e) Manufactured solution (Figure 7.3.3):
u(x) =
1 e
xRe
1 e
Re
+C sin(x) in [0, 1], (7.3.12)
where C is an arbitrary constant, is the exact solution of the following steady equation,
au
x
= u
xx
+C [a cos(x) + sin(x)] . (7.3.13)
I like this solution because it is non-trivial in the diusion limit (Re 0).
7.3.2 2D Advection-Diusion Equation
The two-dimensional advection-diusion equation,
u
t
+au
x
+bu
y
= (u
xx
+u
yy
), (7.3.14)
where a, b, and (> 0) are constants, has the following exact solutions.
(a) Unsteady solution 1 (Figure 7.3.4):
As derived in Section 6.1.3, we have the following exact unsteady solution,
u(x, y, t) = Ce
2
(a
2
+b
2
)(B
2
A
2
)t
cos(B) exp
_
1
1 + 4A
2
2
2
_
, (7.3.15)
where = ax + by, = bx ay, and A, B, and C are arbitrary constants. If we rationalize the exponent, we
obtain
u(x, y, t) = Ce
2
(a
2
+b
2
)(B
2
A
2
)t
cos(B) exp
_
2A
2
1 +
1 + 4A
2
2
_
, (7.3.16)
which can be used safely in the diusion limit, i.e., as 0. So, I like this one better.
210 CHAPTER 7. EXACT SOLUTIONS
Figure 7.3.6: Exponential steady solution (7.3.21): C =
0.009, = 0.1, a = 1.0, b = 0.0, A = 2.0.
Figure 7.3.7: Boundary-layer type solution (7.3.22): =
0.05, a = 1.0, b = 1.0.
(b) Unsteady solution 2 (Figure 7.3.5):
Here is another exact unsteady solution to the advection-diusion equation (7.3.14):
u(x, y, t) =
C
4t +c
1
exp
_
(x at c
2
)
2
(4t +c
1
)
(y bt c
2
)
2
(4t +c
1
)
_
, (7.3.17)
where C, c
1
, and c
2
are arbitrary constants. I like the fact that this will reduce to an exact solution of the
two-dimensional diusion equation if we set a = b = 0.
(c) Polynomial solutions:
Quadratic: u(x, y) = u
0
+C
1
+C
2
_
2 +
2
_
, (7.3.18)
Cubic: u(x, y) = u
0
+C
3
+C
4
_
+
2
2
_
+C
5
_
+
3
6
_
, (7.3.19)
Quartic: u(x, y) = u
0
+C
6
+C
7
_
+
2
2
_
+C
8
_
+
3
6
_
+C
9
_
3
+
2
2
2
+
2
2
+
4
24
_
, (7.3.20)
where u
0
, C
1
, C
2
, C
3
, C
4
, C
5
, C
6
, C
7
, C
8
, C
9
are all arbitrary constants. I like these polynomial solutions
because I can directly check, for example, whether a numerical scheme can preserve a quadratic/cubic/quartic
exact solution. Such a polynomial-preserving property can be a useful measure of the accuracy of a scheme,
especially on unstructured grids.
(d) Exponential solution (Figure 7.3.6):
Simply by setting A = B in the solution (7.3.16), we obtain the following steady exact solution:
u(x, y) = C cos(A) exp
_
2A
2
1 +
1 + 4A
2
2
_
, (7.3.21)
where C and A are arbitrary constants. I like this solution because it is very smooth and thus can be easily used
to verify the accuracy of numerical schemes.
7.3. ADVECTION-DIFFUSION EQUATION 211
Figure 7.3.8: Boundary-layer type manufactured solution (7.3.24): Re = 10
4
, x
0
= 0.2.
(e) Boundary-layer type solution I (Figure 7.3.7):
u(x, y) =
_
1 exp
_
(x 1)
a
_ _
1 exp
_
(y 1)
b
_
_
1 exp
_
_ _
1 exp
_
Re
x +x
0
, Re =
a
2
+b
2
. (7.3.25)
The reference location x
0
can be chosen to avoid the singularity, e.g., x
0
= 0.2. It resembles the x-component of
the velocity for the laminar at-plate boundary layer ow in Section 7.15.9. It is exact for the steady advection-
diusion equation with a source term, f():
au
x
+bu
y
= (u
xx
+u
yy
) +f(), (7.3.26)
where
f() = exp()
_
a
x
+b
y
(
xx
+
yy
(
2
x
+
2
y
))
. (7.3.27)
Look at Figure 7.3.8. It is just like a boundary layer in a very thin domain. It can be a good test case for advection-
diusion schemes. I like it very much. Of course, the same solution can be made exact for the compressible
Navier-Stokes equations with a suitable source term.
212 CHAPTER 7. EXACT SOLUTIONS
7.4 Advection-Reaction Equation(3D)
The advection-reaction equation,
u
t
+au
x
+bu
y
+cu
z
= ku, (7.4.1)
has the exact solution of the form,
u(x, y, z, t) = e
kt
f(x at, y bt, z ct), (7.4.2)
where u(x, y, z, 0) = f(x, y, z); and its two-dimensional version,
u
t
+au
x
+bu
y
= ku, (7.4.3)
has the exact solution of the form,
u(x, y, t) = e
kt
g(x at, y bt), (7.4.4)
where u(x, y, 0) = g(x, y); and its one-dimensional version,
u
t
+au
x
= ku, (7.4.5)
has the exact solution of the form,
u(x, t) = e
kt
h(x at), (7.4.6)
where u(x, 0) = h(x). So, as soon as you choose the initial solution, you immediately have the exact solution. I like
it. Basically, these exact solutions can be derived very easily. All are derived from the fact that
du
dt
= ku, (7.4.7)
along the characteristic (dx/dt, dy/dt, dz/dt) = (a, b, c). Or equivalently, we may apply the change of variables,
u = e
kx
a
v, in one dimension, (7.4.8)
u = e
k
2
(
x
a
+
y
b
)
v, in two dimensions, (7.4.9)
u = e
k
3
(
x
a
+
y
b
+
z
c
)
v, in three dimensions, (7.4.10)
to transform the advection-reaction equation into a pure advection equation of v, which can be solved easily. Note
that t = x/a = y/b = z/c along the characteristic. Either way, it is very simple to derive the exact solution.
7.5 Advection-Diusion-Reaction Equation
(a) Steady Solution:
Consider the steady advection-diusion-reaction equation:
au
x
= u
xx
ku +s(x), (7.5.1)
where a, (> 0), and k(> 0) are constants, and s(x) is an arbitrary function of x. I like this equation because
the exact solution can be easily derived. Substitute u(x) = Aexp(
1
x) + Bexp(
2
x) + s(x)/k, A and B are
constants, into Equation (7.5.1), and nd that the solution is exact if
2
1,2
a
1,2
k = 0, (7.5.2)
7.6. SPHERICAL ADVECTION EQUATION(3D) 213
which yields
1,2
=
a
a
2
+ 4k
2
. (7.5.3)
Therefore, the exact solution is given by
u(x) = Aexp
_
a +
a
2
+ 4k
2
x
_
+Bexp
_
a
a
2
+ 4k
2
x
_
+
s(x)
k
, (7.5.4)
where A and B can be determined from the boundary conditions, e.g., by solving u(x = 0) = 0 and u(x = 1) = 1
for A and B. Interestingly, this exact solution is used to construct a nite-dierence scheme in Ref.[141].
(b) Unsteady Solution I:
As derived in Section 6.1.4, the following solution,
u = C
0
e
kt
_
C
1
exp
_
a +
a
2
+ 8k
2
x
_
+C
2
exp
_
a
a
2
+ 8k
2
x
__
, (7.5.5)
where C
0
, C
1
, C
2
are arbitrary constants, is the exact solution to the unsteady advection-diusion-reaction equa-
tion,
u
t
+au
x
= u
xx
ku. (7.5.6)
(c) Unsteady Solution II (manufactured solution):
The exact solution to the unsteady version of the advection-diusion-reaction equation can be constructed by the
method of manufactured solutions. For example, the solution [141]
u(x, t) = x
2
e
t
, (7.5.7)
satises the advection-diusion-reaction equation,
u
t
+au
x
= u
xx
ku +s(x, t), (7.5.8)
where
s(x, t) = e
t
_
(k 1)x
2
+ 2ax 2
. (7.5.9)
Of course, various other exact solutions can be constructed by the method of manufactured solutions.
7.6 Spherical Advection Equation(3D)
Consider the spherical advection equation:
u
t
+u
r
=
u
r
, (7.6.1)
where r is the distance from the origin. This equation was used in Ref.[71] as a model for aeroacoustics for studying
the long-time behavior of advection schemes in a large domain r [5, 400]. I like this equation because it can be
written as
(ru)
t
+ (ru)
r
= 0, (7.6.2)
so that the exact solution is simply ru(r, t) = f(r t), i.e.,
u(r, t) =
1
r
f(r t), (7.6.3)
where f is determined by the initial condition, say u(r, 0) =
1
r
sin(2r), meaning f(r) = sin(2r). This is very nice.
214 CHAPTER 7. EXACT SOLUTIONS
7.7 Burgers Equation
7.7.1 1D Burgers Equation
For the one-dimensional Burgers equation,
t
u +
x
_
u
2
/2
_
= 0, (7.7.1)
or
t
u +u
x
u = 0, (7.7.2)
basically a shock is developed if the initial solution is monotonically decreasing (characteristics converging) while a
rarefaction is developed if monotonically increasing (characteristics diverging). What is nice about this equation is that
u is constant along the characteristics dx/dt = u, which in turn means that the characteristics are straight lines. So,
it is very easy to nd the exact solution (simply trace back along a characteristic line) until characteristics collide with
each other to form a shock. In particular, I like the following exact solutions.
(a) Smooth solution turning into a shock (Figure 7.7.1):
A smooth initial solution such as
u =
1
2t
s
sin(2x), x [0, 1], (7.7.3)
with the periodic boundary condition u(0, t) = u(1, t) = 0, leads to the following exact solution,
u =
1
2t
s
sin[2(x ut)], (7.7.4)
where t
s
is a free parameter that species the time when a shock is formed, i.e.,
x
u at x = 0.5. Note that
it is given implicitly and therefore must be solved numerically at each point and time (x, t): for example, for a
given location x
i
at time t
n
, iterate on u
k
i
by a simple xed-point iteration,
u
k+1
i
=
1
2t
s
sin[2(x
i
u
k
i
t
n
)], (7.7.5)
until |u
k+1
i
u
k
i
| < 10
15
. I like this solution because it can be used continuously to measure the numerical
accuracy up to the shock formation as well as to test the shock-capturing capability after the shock formation.
(b) Smooth rarefaction (Figure 7.7.2):
A smooth initial solution such as
u = Atanh(k(x 0.5)), (7.7.6)
leads to the following exact solution,
u = Atanh(k(x 0.5 ut)), (7.7.7)
where A and k are arbitrary constants. Of course, it is given implicitly again, but can be computed numerically as
in the previous one.
(c) Compression turning to a shock (Figure 7.7.3):
A linear initial condition such as
u(x, 0) =
_
_
3
2
if x 0,
3
2
2x if 0 < x < 1,
1
2
if x 1,
(7.7.8)
7.7. BURGERS EQUATION 215
" #
$
"
$
x
u
Figure 7.7.1: Smooth solution turning into a shock:
solid line at t = 0, dashed line at t = 0.5, and dotted
line at t = 0.9. A =
1
2ts
with t
s
= 1.
" #
$
"
$
x
u
Figure 7.7.2: Smooth solution spreading out: solid
line at t = 0, dashed line at t = 0.5, and dotted line
at t = 0.9. A = 0.2 and k = 20.
leads to the following exact solution: the solution before a shock is formed,
u(x, t < 1/2) =
_
_
3
2
if x
3t
2
,
3 4x
2 4t
if
3t
2
< x < 1
t
2
,
1
2
if x 1
t
2
,
(7.7.9)
and the solution after the shock is formed at t = 1/2,
u(x, t 1/2) =
_
_
3
2
if x
1+t
2
,
1
2
if x
1+t
2
.
(7.7.10)
It is quite simple. I like it.
(d) Accelerating shock (Figure 7.7.4):
Consider a linear initial condition with a shock,
u(x, 0) =
_
_
_
u
L
if x 0,
u
R
(x 1) if x > 0,
(7.7.11)
where u
L
and u
R
are arbitrary constants. Assuming u
L
> 0, we would like a solution where the shock travels
to the right keeping the left side of the shock uniform u = u
L
. On the right side of the shock, the solution is
continuous, and therefore the exact solution is obtained by tracing back the characteristics:
u(x, t) = u(x ut, 0) = u
R
((x ut) 1), (7.7.12)
which can be solved for u to yield
u(x, t) =
u
R
(x 1)
1 +u
R
t
. (7.7.13)
216 CHAPTER 7. EXACT SOLUTIONS
0.5
0.75
u = 3/2
u =
3 4x
2 4t
u =
1
2
1
1
x
t
O
Figure 7.7.3: Compression wave turning into a straight
shock at (x, t) = (0.5, 0.75).
1
1
u = u
L
u =
u
R
(x 1)
1 +u
R
t
Shock
x
t
O
Figure 7.7.4: Accelerating shock wave solution with the
parameters: u
R
= 2u
L
and u
L
=
2
3
(1 +
3).
Therefore, the exact solution to the problem can be written as
u(x, t) =
_
_
u
L
if x x
s
(t),
u
R
(x 1)
1 +u
R
t
if x x
s
(t),
(7.7.14)
where x
s
(t) is the location of the shock. See Figure 7.7.4; observe that the shock is accelerating. The shock path
can be obtained by solving the ordinary dierential equation for the shock speed (2.13.3),
V
s
(t) =
dx
s
(t)
dt
=
1
2
_
u
L
+
u
R
(x
s
1)
1 +u
R
t
_
, (7.7.15)
with the initial condition x
s
(0) = 0. The solution is
x
s
(t) =
_
u
L
u
R
+ 1
_
_
1
1 +u
R
t
_
+u
L
t. (7.7.16)
In order for the shock to travel to the right, we must choose the parameters such that
u
R
2u
L
. (7.7.17)
And I personally like the following values:
u
L
=
2
3
(1 +
3), u
R
= 2u
L
, (7.7.18)
because it will lead to x
s
(1) = 1, i.e., the shock will reach x = 1 at t = 1 as in Figure 7.7.4. In this case, the
shock path can be expressed by
t =
x +
3x
1 +
3
. (7.7.19)
I like this solution very much. This is a very good test case to check if a numerical scheme can compute a shock
wave with the correct traveling speed. It is well-known that conservation is very important to ensure a correct
shock speed; the shock computed by a non-conservative scheme may not reach x = 1 at t = 1.
7.7. BURGERS EQUATION 217
V
s
=
u
L
+u
R
2
u = u
L
(> u
R
) u = u
R
x
t
O
Figure 7.7.5: Isolated Shock for the 1D Burgers equation.
u = u
L
(< u
R
) u = u
R
dx
dt
= u
R
dx
dt
= u
L
x
t
O
Figure 7.7.6: Rarefaction for the 1D Burgers equation.
7.7.2 1D Burgers Equation: Riemann Problems
The Riemann problem consists of a hyperbolic equation and piecewise constant initial data with a single discontinuity.
I like it especially because I can cook up exact solutions easily. Consider the Riemann problem for Burgers equation,
t
u +
x
_
u
2
/2
_
= 0, (7.7.20)
u(x, 0) =
_
u
L
if x < 0,
u
R
if x > 0.
(7.7.21)
This has the following exact solutions.
(a) Shock (Figure 7.7.5):
Choose u
L
and u
R
such that
u
L
> u
R
, (7.7.22)
then a shock is created and it will travel at the speed,
V
s
=
u
L
+u
R
2
, (7.7.23)
and so, the exact solution is given by
u(x, t) =
_
u
L
if x/t < V
s
,
u
R
if x/t > V
s
.
(7.7.24)
(b) Rarefaction (Figure 7.7.6):
Choose u
L
and u
R
such that
u
L
< u
R
, (7.7.25)
then a rarefaction wave is created, and the exact solution is given by
u(x, t) =
_
_
u
L
if x/t < u
L
,
x/t u
R
u
L
u
R
u
L
+
x/t u
L
u
R
u
L
u
R
if u
L
< x/t < u
R
,
u
R
if x/t > u
R
.
(7.7.26)
(c) Sonic Rarefaction (Figure 7.7.7):
Choose u
L
and u
R
such that
u
L
< 0, u
R
> 0, and u
L
= u
R
, (7.7.27)
218 CHAPTER 7. EXACT SOLUTIONS
u = u
L
= u
R
u = u
R
(> 0)
x
t
O
Figure 7.7.7: Sonic Rarefaction for the 1D Burgers equation.
then a rarefaction wave is created at the center, expanding to the left and the right at the same speed, which is
called a sonic rarefaction. The exact solution is given by
u(x, t) =
_
_
u
L
if x/t < u
L
,
x/t if u
L
< x/t < u
R
,
u
R
if x/t > u
R
.
(7.7.28)
Some numerical schemes fail to compute this solution, preserving the initial discontinuity which must break up
immediately.
7.7.3 2D Burgers Equation
The two-dimensional Burgers equation of the form,
uu
x
+u
y
= 0, (7.7.29)
is basically a one-dimensional equation with y considered as time. So, any time-dependent one-dimensional solutions
can be exact solutions of this equation if t is replaced by y. Namely, any exact solutions for u
t
+uu
x
= 0 in (x, t)-space,
such as those in Sections 7.7.1 and 7.7.2, are exact solutions for u
y
+uu
x
= 0 in (x, y)-space. For example, those in
Figures 7.7.3, 7.7.4, 7.7.5, 7.7.6, and 7.7.7 can be directly interpreted, with t replaced by y, as the exact solutions of
the two-dimensional equation (7.7.29). This is very convenient. I love it. Actually, the curved shock solution in Figure
7.7.4 is my favorite, and I have used it as a test case for mesh adaptation [114].
7.8 Viscous Burgers Equation
7.8.1 1D Viscous Burgers Equation
The one-dimensional viscous Burgers equation,
u
t
+uu
x
= u
xx
, (7.8.1)
has the following exact solutions.
(a) Unsteady/Steady solution (Figure 7.8.1):
u =
1 + (2a 1) exp((1 a)/)
1 + exp((1 a)/)
, (7.8.2)
= x at x
0
, (7.8.3)
7.8. VISCOUS BURGERS EQUATION 219
0 1
0
1
Vs=0.5
x
u
Figure 7.8.1: Unsteady solution of the 1D viscous Burg-
ers equation. u
L
= 1, u
R
= 0, = 0.01.
0 1
0
1
x
u
Figure 7.8.2: Steady solution of the 1D viscous Burgers
equation. x
0
= 0.5, b = 1.0, c = 0.5, = 0.01.
x
0
, a, and are constants, e.g., x
0
= 0.1, a = 0.5, = 0.001 (see Ref.[64]). This solution is nice because it can
be expressed also in the following form,
u = u
R
+
1
2
(u
L
u
R
)
_
1 tanh
_
(u
L
u
R
)(x V
s
t)
4
__
, (7.8.4)
where
V
s
=
u
L
+u
R
2
, u
L
= lim
x
u, u
R
= lim
x
u, u
L
> u
R
. (7.8.5)
Personally, I like this form better. It is very simple and intuitive. Moreover, if we take u
R
= u
L
, then the shock
speed V
s
vanishes and thus we obtain a steady solution. This is also very nice.
(b) Steady solution (Figure 7.8.2):
A slightly modied version of the viscous Burgers equation [155],
u
t
+ (bu c)u
x
= u
xx
, (7.8.6)
has the exact steady solution,
u =
c
b
_
1 tanh
_
c(x x
0
)
2
__
, (7.8.7)
where b(= 0), c(= 0), and x
0
are arbitrary constants.
I like this version also [154],
u
t
+uu
x
=
2
(uu
x
)
x
, x (0, 1), (7.8.8)
because it has the following very smooth steady exact solution:
u(x) = e
x
. (7.8.9)
7.8.2 1D Viscous Burgers Equation with Source Terms
As shown in Section 6.4.1, the one-dimensional Burgers equation with source terms,
u
t
+uu
x
= u
xx
+u
2
ku +d, (7.8.10)
where (> 0), , k, and d are arbitrary constants, has the following analytical solutions, depending on the value of
D = k
2
4d.
220 CHAPTER 7. EXACT SOLUTIONS
0 1
2
3
x
u
Figure 7.8.3: Exact solution of the 1D viscous Burgers equation with source terms: Solid t=0.0,
Dash t=0.125, Dash-Dot t=0.25.
(a) D > 0
u = 2
1
K
1
exp [
1
x +
1
(
1
+ 2)t] +
2
K
2
exp [
2
x +
2
(
2
+ 2)t]
K
1
exp [
1
x +
1
(
1
+ 2)t] +K
2
exp [
2
x +
2
(
2
+ 2)t]
, (7.8.11)
where K
1
and K
2
are arbitrary constants, and
1
=
k
D
4
,
2
=
k +
D
4
. (7.8.12)
(b) D < 0
u = 2
(
r
K
1
+
i
K
2
) cos + (
r
K
i
K
1
) sin
K
1
cos +K
2
sin
, (7.8.13)
where K
1
and K
2
are arbitrary constants, and
=
i
x + 2(
r
i
+
i
)t,
r
=
k
4
,
i
=
D
4
. (7.8.14)
(c) D = 0
u = 2
K
2
(1 +
s
x) + 2K
s
(
s
+)t
K
2
x + 2K
1
(
s
+)t
, (7.8.15)
where K
1
and K
2
are arbitrary constants, and
s
=
k
4
. (7.8.16)
In particular, I like the case (a) because the following set of parameters generates a nice and smooth traveling wave
solution shown in Figure 7.8.3:
= 0.01, K
1
= 100.0, K
2
= 0.01, = 0.1, k = 0.5, d = 0.6. (7.8.17)
I really like this case because in other cases I have not been able to discover a set of parameters which produces a
solution with no singularities.
7.9. LAPLACE EQUATION 221
Figure 7.8.4: Steady solution for the 2D viscous Burgers equation, (7.8.24) with = 0.05.
7.8.3 2D Viscous Burgers Equations
The two-dimensional viscous Burgers equations,
u
t
+uu
x
+vu
y
(u
xx
+u
yy
) = 0, (7.8.18)
v
t
+uv
x
+vv
y
(v
xx
+v
yy
) = 0, (7.8.19)
have an exact steady solution in the form [37],
u =
2
_
a
2
+a
4
y +a
5
_
e
(xx0)
e
(xx0)
_
cos(y)
a
1
+a
2
x +a
3
y +a
4
xy +a
5
_
e
(xx0)
+e
(xx0)
_
cos(y)
, (7.8.20)
v =
2
_
a
3
+a
4
x a
5
_
e
(xx0)
e
(xx0)
_
sin(y)
a
1
+a
2
x +a
3
y +a
4
xy +a
5
_
e
(xx0)
+e
(xx0)
_
cos(y)
, (7.8.21)
where a
1
, a
2
, a
3
, a
4
, a
5
, x
0
are arbitrary constants. In Ref.[37], the following set of parameters,
a
1
= a
2
= 1.3 10
13
, a
3
= a
4
= 0, a
5
= 1, = 25, x
0
= 1, = 0.04, (7.8.22)
are used to generate a smooth boundary-layer type solutions. Well, in fact, my favorite is the following two-dimensional
viscous Burgers equation,
u
t
+uu
x
+u
y
u
xx
= 0. (7.8.23)
This is nice because it is a scalar equation and also it has a simple and interesting exact steady solution:
u(x, y, t = ) = 4 tanh
_
y + 2x
1
2
. (7.8.24)
As shown in Figure 7.8.4, this solution has a very simple boundary-layer type feature. A similar solution was used in
Ref.[23] to verify the accuracy of high-order schemes.
7.9 Laplace Equation
I like the Laplace equation very much because it has a variety of solutions in two and three dimensions although
it has only a trivial linear solution in one dimension (7.2.11).
222 CHAPTER 7. EXACT SOLUTIONS
7.9.1 2D Laplace Equation
I like the two-dimensional Laplace equation:
u
xx
+u
yy
= 0. (7.9.1)
It has various smooth exact solutions:
(a) u = xy, (7.9.2)
(b) u = x
2
y
2
, (7.9.3)
(c) u =
2y
(1 +x)
2
+y
2
, (7.9.4)
(d) u = exp(kx) sin(ky), (7.9.5)
(e) u =
sinh(x) sin(y) + sinh(y) sin(x)
sinh()
, (7.9.6)
where k is an arbitrary constant. We may take the domain to be [0, 1] [0, 1] for all, and specify the boundary condition
directly by using the exact solutions (the Dirichlet condition).
7.9.2 3D Laplace Equation
The three-dimensional Laplace equation,
u
xx
+u
yy
+u
zz
= 0, (7.9.7)
or in spherical coordinates,
1
r
2
r
_
r
2
r
_
+
1
r
2
sin
2
2
+
1
r
2
sin
_
sin
_
= 0, (7.9.8)
has the following exact solutions.
(a) u =
k
r
, (7.9.9)
(b) u = exp(
2k
1
x) sin(k
1
y) cos(k
1
z) +A
2
exp(
2k
2
z) sin(k
2
x) cos(k
2
y)
+A
3
exp(
2k
3
y) sin(k
3
z) cos(k
3
x), (7.9.11)
where A
i
and k
i
with i = 1, 2, 3 are arbitrary constants. This is nice. I like it.
Note also that any two dimensional solutions in the previous subsection and their linear combinations are exact
solutions to the three-dimensional Laplace equation. For example, these solutions,
(d) u = A
1
xy +A
2
yz +A
3
zx, (7.9.12)
(e) u = A
1
(x
2
y
2
) +A
2
(y
2
z
2
) +A
3
(z
2
x
2
), (7.9.13)
(f) u = A
1
exp(k
1
x) sin(k
1
y) +A
2
exp(k
2
y) sin(k
2
z) +A
3
exp(k
3
z) sin(k
3
x), (7.9.14)
are all exact solutions to the three-dimensional Laplace equation. There are many many many possible exact solutions
for the Laplace equation.
7.10. POISSON EQUATIONS 223
7.10 Poisson Equations
I like the Poisson equation because exact solutions can be easily manufactured (see Section 6.8). Basically, you
can easily cook up an exact solution simply by picking up a function u(x, y, z), substituting it into the left hand side
of the Poisson equation (i.e., the Laplacian part), and dening whatever left as the source term f(x, y, z). Just like
the Laplace equation, the Poisson equation is often used as a model equation for developing numerical schemes for the
viscous terms.
7.10.1 2D Poisson Equations
The two-dimensional Poisson equation,
u
xx
+u
yy
= f(x, y), (7.10.1)
has the following manufactured solutions:
(a) u = e
xy
with f = (x
2
+y
2
)e
xy
, (7.10.2)
(b) u = exp(kx) sin(ky) +
A
4
(x
2
+y
2
) with f = A, (7.10.3)
(c) u = sinh x with f = sinh x, (7.10.4)
(d) u = exp
_
x
2
+y
2
d
_
with f =
4(x
2
+y
2
+d)
d
2
exp
_
x
2
+y
2
d
_
, (7.10.5)
where k, A and d are arbitrary constants. Of course, as in two dimensions, we may take the domain to be [0, 1] [0, 1]
for all, and specify the boundary conditions directly by using the exact solutions. Also, we can superpose some of these
solutions to generate a new exact solution such as
(e) u = Ae
xy
+Bsinh x with f = A(x
2
+y
2
)e
xy
+Bsinh x, (7.10.6)
where A and B are arbitrary constants.
Smooth isotropic solutions are nice, but if numerical schemes are to be applied eventually to practical high-Reynolds-
number viscous simulations involving boundary layers, they must be tested for anisotropic solutions. I like the following
anisotropic solution:
(f) u = sin(x) sin(2000y) with f = (1 + 2000
2
)
2
sin(x) sin(2000y), (7.10.7)
where the domain is taken to be (x, y) (0, 1) (0, 0.005), so that u = 0 on the boundary. This solution is very
similar to the anisotropic solution to the diusion equation (7.2.16).
7.10.2 3D Poisson Equations
The three-dimensional Poisson equation,
u
xx
+u
yy
+u
zz
= f(x, y, z), (7.10.8)
has the following manufactured solutions:
(a) u = e
kxyz
with f = k(x
2
+y
2
+z
2
)e
xyz
, (7.10.9)
(b) u =
A
6
(x
2
+y
2
+z
2
) with f = A, (7.10.10)
(c) u = C
x
sinh x +C
y
sinh y +C
z
sinh z with f = C
x
sinh x +C
y
sinh y +C
z
sinh z, (7.10.11)
(d) u = exp
_
r
2
d
_
with f =
2(2r
2
+ 3d)
d
2
exp
_
r
2
d
_
, (7.10.12)
where k, A, C
x
, C
y
, C
z
, and d are arbitrary constants, and r
2
= x
2
+y
2
+z
2
.
224 CHAPTER 7. EXACT SOLUTIONS
7.11 Incompressible-Euler/Cauchy-Riemann/Laplace Equations
Two-dimensional incompressible irrotational ows are governed by the Cauchy-Riemann equations, in terms of the
stream-function and the velocity potential ,
x
+
y
= 0, (7.11.1)
y
= 0, (7.11.2)
or in terms of the velocity components u and v,
u
x
+v
y
= 0, (7.11.3)
v
x
u
y
= 0, (7.11.4)
or the Laplace equations:
xx
+
yy
= 0, (7.11.5)
xx
+
yy
= 0, (7.11.6)
or
u
xx
+u
yy
= 0, (7.11.7)
v
xx
+v
yy
= 0. (7.11.8)
Note that exact solutions to these equations are exact for the Euler equations in the incompressible limit, and therefore
can be used to test an Euler code. As mentioned in Section 6.2, I nd it very interesting that a ow is obtained
simply by picking up an analytic function F of a complex variable Z = x +iy = re
i
. The velocity potential and the
stream function are immediately obtained as the real and imaginary parts, respectively: F(Z) = (z, y) + i(z, y).
The velocity components (u, v) are obtained as the real and imaginary parts of the complex velocity, respectively:
W(Z) = F
= dF/dZ = u iv. Some basic ows are given in Section 6.2. More practical ows such as a ow over
an airfoil can also be obtained by a mapping technique (the conformal mapping) as shown in this section. We also
consider a rotational ow for which the velocity potential does not exist but the stream function exists.
7.11.1 Stagnation Flow
Stagnation ow is obtained by the following complex potential:
F =
Z
2
2
. (7.11.9)
The velocity potential and the stream function are given by
=
1
2
(y
2
x
2
), = xy. (7.11.10)
The complex velocity is given by
W =
dF
dZ
= Z, (7.11.11)
which thus yields by W = u iv
u = x, v = y. (7.11.12)
Streamlines are plotted in Figure 7.11.1 where the ow goes from the right to the left towards the wall represented by
the y-axis. I like the stagnation ow because the velocity components are linear function of space. So, a second-order
Euler code should be able to produce the exact solution. It sounds like an easy ow to compute. However, for practical
problems, the stagnation region is often a major source of convergence diculty due to the vanishing velocity. See
Refs.[32, 162] for details.
7.11. INCOMPRESSIBLE-EULER/CAUCHY-RIEMANN/LAPLACE EQUATIONS 225
x
y
0 0.5 1
-0.5
0
0.5
Figure 7.11.1: Inviscid Stagnation Flow in 2D.
x
y
-1 0 1
-1
0
1
Figure 7.11.2: Flow over a Rankine Half-Body in 2D.
7.11.2 Flow over a Rankine Half-Body
Superposing a source and a free stream, we can generate a ow over a half-body:
F(Z) = V
Z +
2
ln Z, (7.11.13)
where is the source strength. The velocity potential and the stream function are given by
= V
r cos +
2
ln r = V
x +
4
ln(x
2
+y
2
), (7.11.14)
= V
r sin +
2
= V
y +
2
tan
1
(y/x). (7.11.15)
The complex velocity is given by
W =
dF
dZ
= V
+
2Z
=
_
V
+
2
cos
r
_
i
2
sin
r
, (7.11.16)
which thus yields by W = u iv
u = V
+
2
cos
r
= V
+
2
x
x
2
+y
2
, (7.11.17)
v =
2
sin
r
=
2
y
x
2
+y
2
. (7.11.18)
Streamlines are plotted in Figure 7.11.2. Clearly, we see a ow over a half-body, which is called the Rankine half-body.
The equation of the half-body is given by = /2:
y =
2V
( ), (7.11.19)
which can be used to discretize the boundary of the half-body. The above expression represents the thickness of the
half-body. It shows that the thickness converges to
y
max
=
2V
, (7.11.20)
226 CHAPTER 7. EXACT SOLUTIONS
x
y
0 0.5 1
-1
-0.5
0
Figure 7.11.3: Cavity Flow: Velocity potential.
x
y
0 0.5 1
-1
-0.5
0
Figure 7.11.4: Cavity Flow: Streamlines.
as x (i.e., 0). The location of the stagnation point can be easily found by solving u = 0 at y = 0:
x
stagnation
=
2V
. (7.11.21)
Since we also have v = 0 at y = 0, it is a stagnation point.
A ow over a Rankine half-body can be useful for testing a CFD code and checking the performance for a ow
with a stagnation point. You may want to take only the upper half of the domain as the ow is symmetric anyway.
I like the upper half. It may be called a Rankine quater-body of which Ive never heard. Also, it is possible to create
a closed body by placing a sink behind the source. The resulting ow is known as a Rankine oval. See your favorite
uid dynamics textbook for details.
7.11.3 Cavity Flow
In a square domain [0, 1] [1, 0], a ow that comes out at the upper left corner and sinks in the upper right corner
is given by the following complex potential
2
,
F(Z) = +i = mln
_
g(Z)
g(
a
2
)
_
, (7.11.22)
g(Z) = sinh
_
Z
2a
_
k=even
_
1
sinh
2 Z
2
sinh
2 k
2
_
k=odd
_
1
sinh
2 Z
2
sinh
2 k
2
_, (7.11.23)
where Z = x+iy, k = 1, 2, 3, , and m and a are arbitrary constants. This solution converges very fast: say, k = 20
is sucient in practice. For example, we take
m =
2
, (7.11.24)
a = 1, (7.11.25)
2
This solution was derived by Professor P. L. Roe at the University of Michigan.
7.11. INCOMPRESSIBLE-EULER/CAUCHY-RIEMANN/LAPLACE EQUATIONS 227
so that (x, y) and (x, y) take the following boundary values,
(x, 0) = 0, (7.11.26)
(x, 1) = (0, y) = (1, y) = 1, (7.11.27)
(0.5, 0) = 0, (7.11.28)
i.e., the boundary becomes a wall (see Figures 7.11.3 and 7.11.4: the solutions with k = 22.).
The solution is singular at the upper corners: obviously, the stream function is multi-valued and the potential goes
positive/negative innity. But I like this solution. I can easily avoid the singularity, if I really wish, by restricting the
domain with the upper boundary taken somewhere a bit below y = 0.
7.11.4 Flow over a Circular Cylinder
A ow over a circular cylinder is obtained by superposing a uniform stream, a doublet, and possibly a vortex, in the
form of the complex potential,
F(Z) = +i = V
_
Ze
i
+
R
2
e
i
Z
_
+
i
2
ln Z, (7.11.29)
i.e., with Z = re
i
,
= V
cos( )
_
r +
R
2
r
_
2
, (7.11.30)
= V
sin( )
_
r
R
2
r
_
+
2
ln r, (7.11.31)
which give
u
r
=
r
=
1
r
= V
cos( )
_
1
R
2
r
2
_
, (7.11.32)
u
=
1
r
r
= V
sin( )
_
1 +
R
2
r
2
_
2r
. (7.11.33)
Alternatively, either from the complex velocity W(Z),
W(Z) =
dF
dZ
= u iv = V
_
e
i
R
2
e
i
Z
2
_
+
i
2Z
, (7.11.34)
or by the transformation (1.8.4), we obtain the Cartesian velocity components,
u = V
cos
(x
2
y
2
) cos + 2xy sin
r
4
V
R
2
+
y
2r
2
, (7.11.35)
v = V
sin +
(x
2
y
2
) sin 2xy cos
r
4
V
R
2
x
2r
2
. (7.11.36)
This describes a ow past a circular cylinder of radius R with the free stream V
1
2
V
2
= 1
u
2
r
+v
2
V
2
= 1
u
2
+v
2
V
2
. (7.11.37)
The ow is symmetric with respect to both x-axis and y-axis if is zero. If it is not zero, the symmetry breaks
down (see Figures 7.11.5 and 7.11.6). The stagnation points (zero tangential velocity at r = R) are given by
sin
stag
=
4RV
, (7.11.38)
228 CHAPTER 7. EXACT SOLUTIONS
Figure 7.11.5: Pressure coecient C
p
of the cylinder
ow: V
= 1, = 0, = .
Figure 7.11.6: Streamlines of the cylinder ow: V
= 1,
= 0, = .
where
stag
gives the location of a stagnation point. Note that the stagnation points do not exist on the cylinder
surface if
|| > 4RV
. (7.11.39)
Note also that a force F = (F
x
, F
y
) acting on the cylinder can be obtained by the Blasius formula,
F
x
iF
y
=
i
2
Cylinder
[W(Z)]
2
dZ, (7.11.40)
giving
F
x
iF
y
= V
e
i(/2+)
. (7.11.41)
Therefore, the force acts perpendicular to the free stream, and so the drag D and the lift L are given by
L = V
, (7.11.42)
D = 0. (7.11.43)
This result is called the Kutta-Joukowsky theorem. Of course, I like it. It is very nice that the drag is zero. This is
actually called dAlemberts paradox since the drag will be non-zero in reality. But an Euler code (being without any
viscous eects) should be able to produce this paradoxical result. Can your Euler code predict a zero-drag potential
ow?
Finally, I nd it interesting that the line integral over a closed curve C of the complex velocity gives the volume
ow Q across C and the circulation around C:
+iQ =
C
W dz, (7.11.44)
which is very easy to prove. So, Ill leave it to you. Have fun.
7.11.5 Flow over an Elliptic Cylinder
Analytic function g that maps a point = +i into a point Z = x +iy (see Figure 7.11.7):
Z = g(), (7.11.45)
7.11. INCOMPRESSIBLE-EULER/CAUCHY-RIEMANN/LAPLACE EQUATIONS 229
-plane
f(), w()
O R R
R
R
Z = g()
x
y
Z-plane
F(Z), W(Z)
O
A
A
B
B
Figure 7.11.7: Conformal mapping from a circle to an ellipse. A = R +
2
/R
2
and B = R
2
/R
2
.
is called a conformal mapping. Consider a ow in Z-plane dened by the complex potential F(Z). This ow can be
mapped onto -plane easily by Z = g(). The corresponding ow in -plane is then described by the complex potential
f() given by
f() = F(g()), (7.11.46)
and the complex velocity w() is given by
w() =
df()
d
=
dF
dZ
dZ
d
= W(g()) g
(). (7.11.47)
Now, reverse the viewpoint. As a matter of fact, the mapping is useful in obtaining a complicated ow in Z-plane
from a simple ow in -plane: f() F(Z) and w() W(Z). For example, the Joukowsky transformation,
Z = +
2
, (7.11.48)
maps a circle in -plane onto an ellipse in Z-plane with the foci at Z = 2, and a circle with radius onto a line
segment of length 4. I like the Joukowsky transformation, but I like its inverse better because it is more useful.
Suppose we wish to compute a velocity at a point Z
p
in Z-plane, but we know only f(), i.e., the complex velocity in
-plane (the simpler ow). Then, rst, we must map a point Z
p
back to the corresponding point
p
in -plane,
p
= g
1
(Z
p
), (7.11.49)
and then compute the desired velocity by
W(Z
p
) = W(g(
p
)) = w(
p
)/g
(
p
) = f
(
p
)/g
(
p
). (7.11.50)
This way, we can compute the velocity at a point Z
p
in Z-plane. The key is to have the inverse transformation g
1
at
hand. It is generally dicult to obtain the inverse transformation, but it is often possible to do so. For the Joukowsky
transformation (7.11.48), since it can be written in the form,
Z 2
Z + 2
=
_
+
_
2
, (7.11.51)
an inverse is found to be
=
(Z + 2)
1
2
+ (Z 2)
1
2
(Z + 2)
1
2
(Z 2)
1
2
. (7.11.52)
230 CHAPTER 7. EXACT SOLUTIONS
Figure 7.11.8: Pressure coecient C
p
of the ow over an
ellipse: V
= 1, R = 3/4, = 10
, = 0, =
3/4.
Figure 7.11.9: Streamlines of the ow over an ellipse:
V
= 1, R = 3/4, = 10
, = 0, =
3/4.
With this, it is now possible to obtain the velocity in a ow over an elliptic cylinder (or a at plate) in Z-plane from a
ow over a circular cylinder in -plane, by setting up the complex potential,
f() = V
_
e
i
+
R
2
e
i
_
+
i
2
ln , (7.11.53)
where R > (or R = for a at plate), leading to the velocity
W(Z) = u iv = f
()/g
()
=
V
_
e
i
R
2
e
i
2
_
+
i
2
1
2
2
. (7.11.54)
Note that we would have to use LHopitals rule to evaluate it at = .
The surface of the elliptic cylinder is dened by
x = Acos , (7.11.55)
y = Bsin , (7.11.56)
where
A = R +
2
R
, (7.11.57)
B = R
2
R
, (7.11.58)
which can be inverted to get
R =
A+B
2
, (7.11.59)
=
A
2
B
2
4
. (7.11.60)
So, now we can do the following.
7.11. INCOMPRESSIBLE-EULER/CAUCHY-RIEMANN/LAPLACE EQUATIONS 231
TE
=
= +i
O
Figure 7.11.10: Circle that is mapped onto an airfoil.
Construction of an Ellipse:
1. Choose the major axis A and the minor axis B of the ellipse.
2. Generate an ellipse by Equations (7.11.55) and (7.11.56) in Z-plane.
Computation of Exact Solution:
1. Determine R and by
R =
A+B
2
, =
A
2
B
2
4
. (7.11.61)
2. Choose V
, , and .
3. Pick any point Z
p
= x
p
+iy
p
(on or outside the ellipse), and get the corresponding
p
by Equation
(7.11.52).
4. Compute the exact velocity at the point Z
p
by Equation (7.11.54), using
p
, and the pressure by
C
p
=
p p
1
2
V
2
= 1
u
2
+v
2
V
2
. (7.11.62)
An example is shown in Figures 7.11.8 and 7.11.9.
7.11.6 Flow over an Airfoil
To generate a ow over an airfoil, rst set up a ow over a shifted circular cylinder in -plane as follows. Consider a
circle of radius a, centered at
= , (7.11.63)
and passing through the point,
= , (7.11.64)
where is a real value (nearly a half chord of the resulting airfoil) while is a complex value (see Figure 7.11.10).
This circle is represented by the formula,
= +ae
i
, (7.11.65)
232 CHAPTER 7. EXACT SOLUTIONS
where
0 2, (7.11.66)
= ( +i), (7.11.67)
a =
(1 +)
2
+
2
, (7.11.68)
and and are positive constants that are related to the thickness and the camber respectively of the resulting airfoil.
In fact, the point
TE
= corresponds to the trailing edge of the resulting airfoil and if we denote the angle that
indicates this point by , i.e.,
TE
= = +ae
i
, (7.11.69)
then we nd
= sin
1
_
a
_
= sin
1
_
(1 +)
2
+
2
_
. (7.11.70)
Note that the point here is to shift the circle o from the origin. Recall that a circle centered at the origin -plane is
Joukowsky-transformed onto an ellipse in Z-plane. Now imagine that a circle o from the origin will be Joukowsky-
transformed onto something dierent from an ellipse, which turns out to be an airfoil-like shape in Z-plane.
The complex potential that describes a ow over the circle dened above is given by
f()
V
= e
i
+i 2a sin( +) ln( ) +
a
2
e
i
, (7.11.71)
which leads to the complex velocity,
w()
V
= e
i
+i
2a sin( +)
a
2
e
i
( )
2
. (7.11.72)
Note that the circulation has been determined such that the velocity be nite at the trailing edge (the Kutta
condition). In fact, the velocity at the trailing edge is given by
w(TE)
dg/d
, but dg/d vanishes for the Joukowsky
transformation. Therefore, in order for the velocity to be nite, it is required that w(
TE
) = 0. This gives =
4aV
sin( + ). For other types of transformations, even if dg/d does not vanish, we must have w(
TE
) = 0
(i.e., the trailing edge must be a stagnation point) because otherwise a ow is not uniquely determined and there are
a plenty of experimental evidences that support this.
Now, we are ready to transform the circle (and a ow around it) onto something like an airfoil (and a ow around
it). Specically, we create a ow around an airfoil from a ow around the circle by a mapping Z = g(), and compute
a velocity at any point in the airfoil ow by transforming the point back to the -plane by = g
1
(Z) and then
evaluate the velocity by
W(Z)
V
=
u iv
V
=
w()/g
()
V
. (7.11.73)
For this purpose, I give two examples for the mapping Z = g(): the Joukowsky transformation and the Krmn-
Tretz transformation. Airfoils generated by these transformations are called the Joukowsky and Krmn-Tretz airfoils
respectively.
Joukowsky Airfoils
The Joukowsky transformation:
Z = g
1
() = +
2
. (7.11.74)
An inverse is given by
= g
1
1
() =
(Z + 2)
1
2
+ (Z 2)
1
2
(Z + 2)
1
2
(Z 2)
1
2
. (7.11.75)
7.11. INCOMPRESSIBLE-EULER/CAUCHY-RIEMANN/LAPLACE EQUATIONS 233
The derivative of the transformation,
g
1
() = 1
2
2
, (7.11.76)
is needed to compute the velocity by Equation (7.11.73).
I like the Joukowsky airfoils because they look very smooth, especially near the trailing edge. In fact, the trailing
edge is cusped, i.e., zero thickness, and so the velocity there is not zero but nite (not a stagnation point). This can
be nice because the ow is then smooth everywhere.
Note that if is not zero (cambered airfoils), there may be a region (below the airfoil but above the x axis) where
the inverse (7.11.75) does not work. I would suggest the use of the Krmn-Tretz transformation to avoid this trouble.
Krmn-Tretz Airfoil
The Krmn-Tretz transformation:
Z n
Z +n
=
_
+
_
n
, n = 2
, (7.11.77)
where is a free parameter to specify the trailing edge angle. This transformation can be written as
Z = g
2
() =
( +)
n
+ ( )
n
( +)
n
( )
n
n. (7.11.78)
An inverse transformation is then given by
= g
1
2
=
(Z +n)
1
n
+ (Z n)
1
n
(Z +n)
1
n
(Z n)
1
n
. (7.11.79)
Note that the inverse transformation does not work if is too large for the same reason as in the inverse Joukowsky
transformation (increasing the thickness parameter may help in keeping a valid inverse transformation with a relatively
large ). The derivative of the transformation is given by
g
2
() =
4n
2
2
[( +)( )]
n1
[( +)
n
( )
n
]
2
, (7.11.80)
which is needed to compute the velocity by Equation (7.11.73).
I like Krmn-Tretz airfoils because these airfoils have a trailing edge with a nite angle. This is more realistic
and practical than the cusped trailing edge. Note also that the Joukowsky transformation is in fact a special case of
the Krmn-Tretz transformation: the Krmn-Tretz transformation with = 0 is the Joukowsky transformation.
Anyway, so, I focus on the Krmn-Tretz transformation in the rest of the section.
234 CHAPTER 7. EXACT SOLUTIONS
Figure 7.11.11: Pressure coecient C
p
of the ow over a
Krmn-Tretz airfoil: = 0.5, = 10
, = 0.1, = 0,
= 0.
Figure 7.11.12: Streamlines of the ow over a Krmn-
Tretz airfoil: = 0.5, = 10
, = 0.1, = 0, = 0.
Construction of a Krmn-Tretz Airfoil:
1. Specify the parameters:
: Chord length 2, (7.11.81)
: Angle of attack (radian) , (7.11.82)
: Thickness , (7.11.83)
: Camber , (7.11.84)
: Trailing edge angle (radian) . (7.11.85)
2. Compute the following quantities:
a =
(1 +)
2
+
2
: Radius of the circle, (7.11.86)
= sin
1
_
a
_
: Trailing edge location, (7.11.87)
(
c
,
c
) = (, ) : Center of the circle. (7.11.88)
This completely denes the circle by Equation (7.11.65).
3. Substitute points on the circle,
i
= (
i
,
i
), i = 1, 2, 3, . . . , N, represented by
i
= a cos(
i
) +
c
, (7.11.89)
i
= a sin(
i
) +
c
, (7.11.90)
with
i
in [0, 2] into the transformation (7.11.78) to obtain the points Z that form an airfoil (NB:
= 0 corresponds to the trailing edge).
4. (Optional) Compute the velocity at the airfoil points by substituting the points generated in the
previous step into Equation (7.11.73). Note that the inverse transformation g
1
2
is not needed here,
so that the velocity on the surface of the airfoil can be safely computed for any values of camber,
thickness, and trailing edge angle.
7.11. INCOMPRESSIBLE-EULER/CAUCHY-RIEMANN/LAPLACE EQUATIONS 235
Computation of Exact Solution (the whole ow eld):
1. Pick a point Z
p
= x
p
+ iy
p
(on and anywhere outside the airfoil!) where you want to compute the
velocity.
2. Transform the point Z
p
by Equation (7.11.79) onto the corresponding point
p
in -plane.
3. Compute the velocity by substituting
p
into Equation (7.11.73).
An example, which was generated by the sample program available at http://www.cfdbooks.com, is given in Figures
7.11.11 and 7.11.12; it is in fact a Joukowsky airfoil since = 0. A ow over a cylinder can be generated by taking
= 180
. Note that the ow has a stagnation point is at the rear of the cylinder because the circulation in in Equation
(7.11.72) has been determined by the Kutta condition. It is interesting to compute the lift force, which is given by
L = V
= 4V
2
_
1
1
r
2
_
y, (7.11.92)
where r =
x
2
+y
2
, is the constant vorticity, and U
x
, (7.11.93)
236 CHAPTER 7. EXACT SOLUTIONS
Figure 7.11.13: Streamlines of Fraenkels ow.
but the result is very complicated:
u = y +
2
_
2y
2
r
4
+
_
1
1
r
2
_
+
x
2
_
1
1
r
4
+
4y
2
r
6
_
ln
_
r
2
2x + 1
r
2
+ 2x + 1
_
+
4x
2
y
2
_
1
1
r
4
_
(r
2
2x + 1)(r
2
+ 2x + 1)
_
2y(x
2
y
2
)
r
6
+
_
1 +
1
r
4
_
y
_
arctan
_
2y
r
2
1
_
+
__
1 +
1
r
4
_
(x
2
y
2
) 2
_
(x
2
y
2
1)
(r
2
1)
2
+ 4y
2
_
+U
_
1
x
2
y
2
r
4
_
, (7.11.94)
v =
2
_
2xy
r
4
+
y
2
_
1
1
r
4
+
4x
2
r
6
_
ln
_
r
2
2x + 1
r
2
+ 2x + 1
_
+
2xy(x
2
y
2
1)
_
1
1
r
4
_
(r
2
2x + 1)(r
2
+ 2x + 1)
_
2x(x
2
y
2
)
r
6
_
1 +
1
r
4
_
x
_
arctan
_
2y
r
2
1
_
2xy
__
1 +
1
r
4
_
(x
2
y
2
) 2
_
(r
2
1)
2
+ 4y
2
_
U
2xy
r
4
. (7.11.95)
These functions were used to generate the streamline plot in Figure 7.11.13. Note that the above exact solution is
valid only on the upper half of the domain, i.e., y 0, and to make it valid on the lower half, I have to change the
sign of y, i.e., y y. Also, Note also that the above exact solution is singular at (x, y) = (1, 0) and (1, 0) where
it should be set to be a stagnation point u = v = 0. By the way, it is very interesting that such a complicated ow
(as in Figure 7.11.13) can be represented by an analytical formula.
7.12. 3D INCOMPRESSIBLE EULER EQUATIONS (AXISYMMETRIC FLOWS) 237
x
y
0 0.5 1
-0.5
0
0.5
Figure 7.12.1: Inviscid Stagnation Flow in 3D.
x
y
-1 0 1
-1
0
1
Figure 7.12.2: Flow over a Rankine Half-Body in 3D.
7.12 3D Incompressible Euler Equations (Axisymmetric Flows)
I like 3D ows because they are much more realistic than 1D and 2D ows. 3D solutions can be used to verify
practical CFD codes, which solve 3D problems. Although the theory of complex variables cannot be applied, various
exact solutions can be derived for 3D irrotational ows because the velocity potential satises the Laplace equation.
Especially, I like axisymmetric ows because these ows can be described by a single stream function just like 2D ows.
Below, axisymmetric ows are described in (x, y)-plane with the x-axis as the axis of symmetry. Note that these are
3D solutions and therefore should not be used for 2D CFD codes that solve 2D equations.
7.12.1 Stagnation Flow
A three-dimensional stagnation ow is an axisymmetric ow. It is described by the following velocity potential:
=
1
4
(y
2
2x
2
). (7.12.1)
It is easy to verify that the above velocity potential satises the Laplace equation for axisymmetric ows:
x
2
+
2
y
2
+
1
y
y
= 0. (7.12.2)
The velocity components (u, v) are given by
u =
x
= x, v =
y
=
1
2
y, (7.12.3)
which dier from the two-dimensional case in Section 7.11.1 by the factor 1/2 in v. As in the two-dimensional case,
the velocity components are linear function of space. I like that. A second-order scheme should still give the exact
result in 3D. The stream function can be derived by integrating Equation (4.24.14):
=
1
2
xy
2
, (7.12.4)
which is dierent from the two-dimensional version (7.11.10). Streamlines are plotted in Figure 7.12.1, which should
be compared with Figure 7.11.1.
238 CHAPTER 7. EXACT SOLUTIONS
7.12.2 Flow over a Rankine Half-Body
A ow over a Rankine half-body can be obtained in 3D. As in 2D, we superpose a source and a free stream:
= V
x
4r
, (7.12.5)
where r =
x
2
+y
2
and the second term represents the three-dimensional source with the strength . It is straight-
forward to verify that the above potential satises the Laplace equation (7.12.2). The velocity components (u, v) are
given by
u =
x
= V
+
x
4r
3
, v =
y
=
y
4r
3
. (7.12.6)
The stream function can be derived as
=
1
2
V
y
2
+
4
_
1
x
r
_
. (7.12.7)
Streamlines are plotted in Figure 7.12.2. This is the Rankine half-body in three dimensions. Compare the streamlines
and the half-body with those in two dimensions, Figure 7.11.2. They are similar but dierent. In 3D, the equation of
the half-body is given by = /(2):
y =
2
V
, (7.12.9)
as x (i.e., 0), which is slightly thicker than the two-dimensional body. The location of the stagnation point
can be easily found by solving u = 0 at y = 0:
x
stagnation
= 2
. (7.12.10)
Of course, we have v = 0 at y = 0, and thus it is indeed a stagnation point.
7.12.3 Flow over a Sphere
A ow over a sphere at zero angle of attack is a valid axisymmetric ow. The following velocity potential, which
consists of a uniform ow and a doublet, is an exact solution that represents a ow over a sphere of radius R,
= V
x
_
1 +
1
2
_
R
r
_
3
_
. (7.12.11)
It is quite straightforward to verify that the above potential satises the Laplace equation (7.12.2). The velocity
components are given by
u =
x
= V
_
1 +
R
3
2r
3
_
1
3x
2
r
2
__
, (7.12.12)
v =
y
=
3xyV
R
3
2r
5
. (7.12.13)
The stream function can be derived in the form:
=
1
2
V
y
2
_
1
R
3
r
3
_
. (7.12.14)
7.13. EULER EQUATIONS 239
Figure 7.12.3: Flow over a Sphere (3D).
Streamlines are plotted in Figure 7.12.3. Note that the maximum speed on the sphere is
|u|
x=0,y=R
=
3
2
V
, (7.12.15)
which is smaller than that for a ow over a cylinder obtained from Equation (7.11.35),
|u| = 2V
. (7.12.16)
On the sphere (x
2
+y
2
= R
2
), we have
u
2
+v
2
=
9V
2
4
y
2
R
2
. (7.12.17)
Therefore, the pressure coecient over the sphere is given by
C
p
=
p p
1
2
V
2
= 1
_
u
2
+v
2
V
2
r=R
= 1
9
4
y
2
R
2
. (7.12.18)
I like the ow over a sphere. It is very simple and smooth. But it is not really simple to compute numerically. First,
it is not trivial to generate a computational grid around a sphere (unlike generating a grid around a two-dimensional
cylinder): structured grids need some eort to avoid possible geometric singularities. Second, it is often hard to obtain
a clean symmetric solution with respect to the (y, z)-plane (i.e., fore-aft symmetry) by an Euler code, particularly with
unstructured grids. How accurately can your Euler code compute this ow?
7.13 Euler Equations
I like the compressible Euler Equations. Basically, all the incompressible ows described in the previous section are the
exact solutions to the compressible Euler Equations with constant density. Here, we consider exact solutions for the
compressible Euler Equations with variable density.
240 CHAPTER 7. EXACT SOLUTIONS
7.13.1 One-Dimensional Simple Acoustic Waves
Consider the simple wave solution derived in Section 6.7.3:
u(x, t)
a
= M
+
2
( + 1)
V ()
a
, (7.13.1)
(x, t)
=
_
1 +
1
2
_
u()
a
__ 2
1
, (7.13.2)
p(x, t)
p
=
_
1 +
1
2
_
u()
a
__
2
1
, (7.13.3)
where = x V t, V = u +a, and V is a solution of Burgers equation. This is valid for smooth ows and the exact
solutions of Burgers equation given in Section 7.7.1 can be directly employed for V .
(a) Smooth solution turning into a shock:
Set an initial solution with
V (x) =
1
2t
s
sin(2x), (7.13.4)
then it gives the exact solution with
V () =
1
2t
s
sin(2(x V t)), (7.13.5)
where t
s
is a free parameter that species the time when a shock is formed. This solution is given implicitly and
therefore must be computed numerically (see Section 7.7.1).
(b) Smooth rarefaction:
Set an initial solution with
V (x) = Atanh(k(x 0.5)), (7.13.6)
then it gives the exact solution with
V () = Atanh(k(x 0.5 V t)), (7.13.7)
where A and k are arbitrary constants. Again, this is given implicitly and therefore must be computed numerically.
These solutions are very similar to those in Figures 7.7.1 and 7.7.2. These types of solutions were actually used in
Ref.[87] for verication of higher-order schemes. It is nice to have smooth exact solutions for nonlinear systems such
as the Euler equations. In particular, it is very nice that we can generate various exact solutions through the solutions
of Burgers equation. I really like it.
7.13.2 Entropy Waves
The entropy wave is an exact simple wave solution of the compressible Euler equations. The one-dimensional entropy
wave is given by
=
+Asin [(x U
t)] , (7.13.8)
u = U
, (7.13.9)
p = P
, (7.13.10)
where
, U
and P
+Asin [(x +y Q
t)] , (7.13.11)
u = U
, (7.13.12)
v = V
, (7.13.13)
p = P
, (7.13.14)
7.13. EULER EQUATIONS 241
!" #" " #" !"
!"
#"
"
#"
!"
x
y
t = 5.0
t = 0.0
t = 10.0
Figure 7.13.1: Density contours for the two-dimensional isentropic vortex convection. = 1, K = 5.0,
= 1,
p
= 1/, u
= v
= 2.0, (x
0
, y
0
) = (10, 10).
where V
is a constant and Q
= U
+V
+Asin [(x +y +z Q
t)] , (7.13.15)
u = U
, (7.13.16)
v = V
, (7.13.17)
w = W
, (7.13.18)
p = P
, (7.13.19)
where W
is a constant and Q
= U
+ V
+ W
=
u
K
2a
ye
(1r
2
)/2
, (7.13.20)
v(x, y, t)
a
=
v
+
K
2a
xe
(1r
2
)/2
, (7.13.21)
T(x, y, t)
T
= 1
K
2
( 1)
8
2
a
2
e
(1r
2
)
, (7.13.22)
(x, y, t)
=
_
T(x, y, t)
T
_ 1
1
, (7.13.23)
p(x, y, t)
p
=
_
T(x, y, t)
T
_
1
, (7.13.24)
242 CHAPTER 7. EXACT SOLUTIONS
0 1
0
1
x
y
Figure 7.13.2: Subsonic solution. Streamlines calculated by the total velocity (U
+ u, v) with U
= 1.0,
k = /
1 M
2
and M
= 0.5.
where
x = x x
0
u
t, (7.13.25)
y = y y
0
v
t, (7.13.26)
r =
x
2
+y
2
. (7.13.27)
This exact solution describes a vortex simply convected isentropically with the free stream velocity (u
, v
). It is
often employed to measure the accuracy of an Euler code [28, 142, 177]. For example, I like the following set of
parameters: = 1,
= 1, p
= 1/, u
= v
= 2, (x
0
, y
0
) = (10, 10), K = 5, because then the vortex starts
at (x, y) = (10, 10), moves to the origin at t = 5, and then will nally reach at (x, y) = (10, 10) precisely at t = 10
as shown in Figure 7.13.1. You may measure the accuracy of your Euler code at the nal location.
7.13.4 Linearized Potential Equations
The linearized potential equation (see Section 3.20),
(1 M
2
xx
+
yy
= 0, (7.13.28)
or the equivalent system for the perturbation velocity components (see Section 3.21),
(1 M
2
) u
x
+v
y
= 0, (7.13.29)
v
x
u
y
= 0, (7.13.30)
has various exact solutions.
Subsonic solution
As discussed in Section 3.21, the Prandtl-Glauert transformation
= x, = y
1 M
2
1 M
2
, U =
1 M
2
u, V = v, (7.13.31)
turns the governing equations into the Laplace equation for the potential,
= 0, (7.13.32)
and the Cauchy-Riemann system for the velocity components,
U
+V
= 0, (7.13.33)
V
= 0, (7.13.34)
7.13. EULER EQUATIONS 243
M
2
(u, v) = (0, 0.1)
(u, v) = (0, 0)
(u, v) = (0, 0)
(u, v) = (0, 0.1)
(u, v) = (0, 0)
(u, v) = (0, 0)
Figure 7.13.3: Supersonic solution. Inow perturbation velocity components are u = 0 and v = 0.1, with the
mean-ow Mach number M
(x, y) =
(x, y
1 M
2
1 M
2
. (7.13.35)
Alternatively, we may start from a set of solutions for the Cauchy-Riemann system, U(, ) and V (, ), and then
obtain the velocity components in the equivalent compressible ow as follows,
u(x, y) =
U(x, y
1 M
2
1 M
2
, (7.13.36)
v(x, y) = V (x, y
1 M
2
). (7.13.37)
For example, the exact solution (7.9.5) from Section 7.9.1,
(x, y) =
exp(kx) sin
_
ky
1 M
2
1 M
2
. (7.13.39)
To see what ow this represents, we derive the perturbation velocity components,
u(x, y) =
x
(x, y) =
k exp(kx) sin
_
ky
1 M
2
1 M
2
, (7.13.40)
v(x, y) =
y
(x, y) = k exp(kx) cos
_
ky
1 M
2
_
. (7.13.41)
The streamlines calculated from these are plotted in Figure 7.13.2. It does look like a ow. This is simple and smooth
enough for accuracy verication, possibly with the Dirichlet condition. I like this very much. It is really nice that any
solution of Laplaces equation can be transformed into a subsonic ow solution.
Supersonic solution
As shown in Section 3.21, for supersonic ows (M
M
2
M
2
1, (7.13.42)
u
M
2
M
2
1, (7.13.43)
244 CHAPTER 7. EXACT SOLUTIONS
with appropriate boundary conditions, we can construct a ow eld, starting from a supersonic free stream of the
Mach number M
> 1 (see Ref.[155] for more details if needed). An example is a supersonic ow through a dust: a
mean-ow coming from the left at Mach number M
= constant. (7.13.46)
We then obtain from these (see Ref.[84] for derivation) the fundamental relations for generating exact solutions:
_
A
A
_
2
=
1
M
_
2
+ 1
_
1 +
1
2
M
2
__
(+1)/(1)
, (7.13.47)
0
=
_
1 +
1
2
M
2
_
1
1
, (7.13.48)
p
p
0
=
_
1 +
1
2
M
2
_
1
, (7.13.49)
T
T
0
=
_
1 +
1
2
M
2
_
1
, (7.13.50)
where = 1.4 for air, A
d
dx
+
1
u
du
dx
+
1
A
dA
dx
= 0, (7.13.51)
and substitute the momentum equation,
u
du
dx
=
dp
dx
= a
2
d
dx
, (7.13.52)
to obtain
1
u
du
dx
=
1
1 M
2
1
A
dA
dx
. (7.13.53)
This shows that a subsonic ow is accelerated (du > 0) through a converging nozzle (dA < 0); and a supersonic ow
accelerated (du > 0) through a diverging nozzle (dA > 0). This is very interesting. I like it.
Now, suppose that a nozzle shape A(x) is given in x [0, 1], having the minimum (or throat) area at x = x
throat
.
Then, there exist three types of solutions: subsonic-supersonic, fully subsonic, and subsonic-supersonic with a shock.
7.13. EULER EQUATIONS 245
0 0.4 1
0
1
2.5
x
M
Figure 7.13.4: Mach number: the unique subsonic-
supersonic ow.
0 0.4 1
0
1
x
p
Figure 7.13.5: Pressure: the unique subsonic-
supersonic ow.
Subsonic-Supersonic Flow
This is a unique solution for a given nozzle shape: a subsonic inow achieves M = 1 precisely at the throat x = x
throat
so that A
= A(x
throat
), and then smoothly becomes supersonic and accelerates towards the exit. It should be noted
that the inow Mach number is not zero but a small value. Theoretically, we see from Equation (7.13.47) that M 0
when A 0. That is, a reservoir exists at x . In a way, we are looking at only a small part of a big nozzle
coming out from a reservoir. Or it may be interpreted as a ow through the nozzle in [0, 1] caused by a pressure
dierence applied to the inlet and outlet of the nozzle with a nite length.
Computation of Exact Solution:
We can numerically compute this unique ow solution at a set of N points {x
i
} = {x
1
=
0, x
2
, x
3
, , x
N
= 1} as follows:
1. Set A
= A(x
throat
).
2. Compute M(x
i
) by solving Equation (7.13.47) for given A(x
i
) and A
for all i.
(subsonic for x
i
[0, x
throat
]; supersonic for x
i
[x
throat
, 1].)
3. Compute (x
i
), p(x
i
), and T(x
i
) by Equations (7.13.48), (7.13.49), and (7.13.50).
Note that Equation (7.13.47) needs to be solved for the Mach number numerically. For this purpose, we may
employ the bisection method for the following equation,
f(M) =
_
A
A
_
2
1
M
_
2
+ 1
_
1 +
1
2
M
2
__
(+1)/(1)
= 0, (7.13.54)
with adaptive initial bounds: [0, 1] to nd a subsonic solution or [1, 100] to nd a supersonic solution. Of course, other
methods can be used, but I like the bisection method because it is very simple. In fact, I actually used it to obtain the
solution shown in Figures 7.13.4 and 7.13.4.
Fully Subsonic Flow
In this case, there exist an innite number of solutions for a given nozzle shape; the exit pressure p
e
must be specied
to make the solution unique. Observe also that in this case a ow can never reach M = 1 (A
= A(x) anywhere in
the nozzle). Therefore, A
1
, (7.13.55)
and then solving
_
A
e
A
_
2
1
M
e
_
2
+ 1
_
1 +
1
2
M
2
e
__
(+1)/(1)
= 0, (7.13.56)
246 CHAPTER 7. EXACT SOLUTIONS
0 0.4 1
0
1
2.5
x
M
Figure 7.13.6: Mach number: a fully subsonic ow.
0 0.4 1
0
1
x
p
Figure 7.13.7: Pressure: a fully subsonic ow.
for A
. Here, obviously, A
e
= A(x
N
) = A(1.0), M
e
= M(x
N
) = M(1.0), and p
e
= p(x
N
) = p(1.0).
Note that the exit pressure p
e
must satisfy
p
crit
e
p
e
< p
0
, (7.13.57)
where p
crit
e
is the pressure corresponding to the ideal subsonic solution, i.e., the one with A
= A(x
throat
) of Equation
(7.13.56), at the exit This p
crit
e
can be computed by rst solving Equation (7.13.56) with A
= A(x
throat
) to obtain
M
crit
e
, and then substituting this Mach number M
crit
e
into Equation (7.13.49).
Computation of Exact Solution:
A fully subsonic ow solution can be computed numerically as follows:
1. Compute M
crit
e
by solving Equation (7.13.56) and A
= A(x
throat
).
(Assume that the ow is subsonic.)
2. Compute p
crit
e
(7.13.48).
3. Specify the exit pressure p
e
under the condition (7.13.57).
4. Compute M
e
from p
e
by Equation (7.13.49).
5. Re-compute A
for all i.
(Note that the ow is subsonic everywhere.)
7. Compute (x
i
), p(x
i
), and T(x
i
) by Equations (7.13.48), (7.13.49), and (7.13.50).
An example is given in Figures 7.13.6 and 7.13.7.
Shocked Flow
In shocked ows also, the exit pressure p
e
must be specied to make the solution unique. This time, the exit pressure,
p
e
, must satisfy
p
ps
p
exit
< p
crit
e
, (7.13.58)
where p
ps
is a post-shock pressure given by
p
ps
=
_
1 +
1
2
M
2
ps
_
1
, (7.13.59)
M
ps
=
_
1 +
1
2
_
M
crit
e
_
2
_
M
crit
e
_
2
1
2
. (7.13.60)
7.13. EULER EQUATIONS 247
0 0.4 1
0
1
2.5
x
M
Figure 7.13.8: Mach number: a shocked ow.
0 0.4 1
0
1
x
p
Figure 7.13.9: Pressure: a shocked ow.
This is a pressure obtained based on the assumption that there exists a shock exactly at the exit in the ideal subsonic-
supersonic ow, i.e., the ow with A
= A(x
throat
). The condition (7.13.58) ensures the presence of a shock somewhere
between the throat (x = x
throat
) and the exit.
Computation of Exact Solution:
A shocked ow solution can be computed numerically as follows:
1. Specify the exit pressure p
e
under the condition (7.13.58).
2. Compute M
e
from p
e
by Equation (7.13.49).
3. Compute and store the ideal solution M
ideal
(x
i
) by solving Equation (7.13.47) with A
= A(x
throat
)
for all i.
(subsonic for x
i
[0, x
throat
]; supersonic for x
i
[x
throat
, 1.0].)
4. Re-compute A
_
1 +
1
2
[M
ideal
(x
i
)]
2
[M
ideal
(x
i
)]
2
1
2
, (7.13.62)
is satised. If it is, there is a shock between x
i1
and x
i
, and you dene i = i
shock
and stop
calculating M(x
i
).
7. The solution is then given by, for i = 1, 2, 3, , N,
M(x
i
) =
_
M
ideal
(x
i
) if i < i
shock
,
M
sub
(x
i
) otherwise,
(7.13.63)
and (x
i
), p(x
i
), and T(x
i
) are computed by Equations (7.13.48), (7.13.49), and (7.13.50).
An example is given in Figures 7.13.8 and 7.13.9.
248 CHAPTER 7. EXACT SOLUTIONS
Entrance Exit
M 0.4517 2.1972
p 0.8694 0.0939
Table 7.13.1: Data for the subsonic-supersonic ow in Figures 7.13.4 and 7.13.4.
Entrance Exit
M 0.2972 0.2091
p 0.9405 0.9700
Table 7.13.2: Data for the fully subsonic ow in Figures 7.13.6 and 7.13.7. p
e
= 0.97 was specied.
Entrance Exit
M 0.4517 0.3909
p 0.8694 0.9000
Table 7.13.3: Data for the shocked ow in Figures 7.13.8 and 7.13.9. p
e
= 0.9 was specied. The shock was found to
be at x = 0.8.
Figure 7.13.10: Upper half of a quadratic nozzle (7.13.64) with A
e
= 0.4 and A
t
= 0.2.
Example
Consider a quadratic shape nozzle,
A(x) =
25
9
(A
e
A
t
)
_
x
2
5
_
2
+A
t
, (7.13.64)
where A
e
and A
t
are the exit area and the throat area respectively (Figure 7.13.10). Note that x
throat
= 2/5. Applying
the algorithms described above, I obtained the results in Figures 7.13.6 to 7.13.9, and the data in Tables 7.13.1 to
7.13.3.
These exact solutions can be used to verify a numerical code for the quasi-1D Euler equations in Section 3.16. But
it is also possible to use these exact solutions as references in two-dimensional calculations. Generate a two-dimensional
grid based on the area function (only the upper half of the nozzle will suce because of the symmetry), and perform
a two-dimensional CFD calculation on it for one of the exact solutions described above. Then, we may compare the
numerical solution with the exact solution at the centerline of the nozzle. In this case, the exact solution is no longer
strictly exact. But it should be close for a smooth nozzle with A
e
A
t
.
I like the nozzle ow very much because it is useful also in magnetohydrodynamic ows: the solution can be used
without any modication for aligned ows (the magnetic eld parallel to the stream line) [42], and also with minor
modications for transverse ows (the magnetic eld perpendicular to the stream line) [85].
7.13.6 Ringlebs Flow
Ringlebs ow is an exact solution to the Euler equations for = 1.4 obtained by Ringleb in 1940 [126]. It is often used
to measure the accuracy of an Euler code (see Ref.[26] and references therein). This ow is obtained as a solution (a
stream function ) of the hodograph equation (6.4.45), so that the physical location (x, y) is given in terms of the
7.13. EULER EQUATIONS 249
x
y
-1 0 1
0
1
2
Figure 7.13.11: A computational grid for Ringlebs ow,
generated with a uniform spacing in rather than V , with
parameters
min
= 0.69,
max
= 1.2, V
min
= 0.5.
V=Vmin
Inflow
!
!
Outflow
y = 0 (symmetric line, u=0)
min
max
M < 1
M
=
1
M > 1
Figure 7.13.12: Ringlebs ow: a ow around the edge
of a at plate or a ow in a channel, being smoothly
accelerated from subsonic to supersonic. The supersonic
region can be avoided by taking suciently large values
of .
physical variables (V, ), i.e., the ow speed and angle (see Ref.[139] for details):
=
sin
V
, (7.13.65)
x(, V ) =
1
_
1
2V
2
2
_
+
L
2
, (7.13.66)
y(, V ) =
V
1 V
2
2
, (7.13.67)
where and L are functions of V :
= b
5
, (7.13.68)
L =
1
b
+
1
3b
3
+
1
5b
5
1
2
ln
_
1 +b
1 b
_
, (7.13.69)
b =
1 0.2V
2
, (7.13.70)
and the pressure p is given by p = b
7
. The isotachs (lines of constant speed) are circles given by
_
x
L
2
_
2
+y
2
=
_
1
2V
2
_
2
. (7.13.71)
Note that the variables have been nondimensionalized by the stagnation values; and so they should be understood
actually as
x
l
,
y
l
,
c
0
l
,
V
c
0
,
0
,
p
p
0
, (7.13.72)
where l is a reference length,
0
, p
0
, and c
0
are the stagnation density, pressure, and the speed of sound respectively.
Now, choose some values for
min
,
max
, V
min
, and allow V to increase up to a limit of the domain, e.g., y = 0,
thus Equation (7.13.67) giving V
max
= 1/. Then, we will have a domain like those in Figures 7.13.11 and 7.13.12;
we can generate its boundary points very easily.
250 CHAPTER 7. EXACT SOLUTIONS
Figure 7.13.13: Flow speed, density, and pressure have
the same contours .
Figure 7.13.14: Flow angle .
Generation of Boundary Points:
The following generates N
R
points on the right boundary, N
T
points on the top boundary, N
L
points
on the left boundary, N
B
points on the bottom boundary (with corner points being shared, i.e., the total
number of points on the whole boundary is N
L
+N
R
+N
B
+N
T
4).
1. Right Boundary:
Generate (x
i
, y
i
) by Equations (7.13.66) and (7.13.67) with =
min
and
V
i
[V
min
, 1/
min
], i = 1, 2, 3, , N
R
.
2. Top Boundary:
Generate (x
i
, y
i
) by Equations (7.13.66) and (7.13.67) with V = V
min
and
i
[
min
,
max
], i = 1, 2, 3, , N
T
.
3. Left Boundary:
Generate (x
i
, y
i
) by Equations (7.13.66) and (7.13.67) with =
max
and
V
i
[V
min
, 1/
max
], i = 1, 2, 3, , N
L
.
4. Bottom Boundary:
Generate (x
i
, y
i
) by Equations (7.13.66) and (7.13.67) with V = 1/
i
and
i
[
min
,
max
], i = 1, 2, 3, , N
B
. (y
i
should be zero for all i here.)
Note that we might want to reverse the node ordering in the step 2, so that boundary points are generated and
ordered counterclockwise over the entire boundary. Also, note that a uniform distribution of points in [V
min
, 1/
min
]
does not lead to the uniform distribution in the physical plane (points will be clustered near the midpoint). First,
generate a uniform grid in [sin
1
(
min
V
min
), /2], and then convert it to V
i
by V
i
=
sin i
min
. This will generate a
smoother grid in the physical plane such as the one shown in Figure 7.13.11.
Once we generate boundary points, we can generate a computational grid over the domain by connecting them in
some way (structured or unstructured). Or you can directly generate a structured grid by applying the above boundary-
point generation algorithm to the interior points with varying stream function values over the interval [
min
,
max
].
Whatever kind of grid you generate, you would want to run an Euler code on it with the appropriate boundary
conditions: uniform inow at the top boundary; wall boundary conditions on the left and right boundaries; outow
condition on the bottom boundary. Then, of course, you would certainly want to compute the error by comparing the
numerical solution with the exact solution.
7.13. EULER EQUATIONS 251
Figure 7.13.15: Velocity component, u = V cos . Figure 7.13.16: Velocity component, v = V sin .
Computation of Exact Solution:
The exact solution at a given location (x
p
, y
p
) can be computed as follows.
1. Numerically solve Equation (7.13.71) for V with the bounds [V
min
, 1/
min
].
2. Compute b by Equation (7.13.70), and then , p, and L.
3. Compute by Equation (7.13.66) in the form,
=
1
2V
2
_
x
p
L
2
_
. (7.13.73)
4. Compute by Equation (7.13.65) in the form,
= sin
1
(V ), [/2, /2]. (7.13.74)
5. Compute the velocity (u, v) = V (cos , sin ) where the negative sign is put to make the upper
boundary the inow).
6. Compute the actual Mach number (if you wish) by
M =
1.4
V
1.4 p/
=
V
p/
, (7.13.75)
where the factor
1.4 is needed because V , p, and are in fact V/c
0
, p/p
0
, and /
0
.
For solving Equation (7.13.71), we may employ the xed-point iteration: write the equation as
V
k+1
=
1
2
k
(x L
k
/2)
2
+y
2
, (7.13.76)
where
k
and L
k
are evaluated with the previous value V
k
, then iterate on V
k
starting with V
1
= (V
min
+1/
min
)/2
until |V
k+1
V
k
| < 10
15
, for example. I like this method because it is very simple.
I like Ringlebs ow very much because in particular the ow accelerates from subsonic to supersonic very smoothly
without shocks. That is, it can be very useful to verify the accuracy of an Euler code simultaneously for both subsonic
and supersonic ows. A sample program, which was used to generate the grid in Figure 7.13.11 and the solution in
Figures 7.13.13, 7.13.14, 7.13.15, and 7.13.16, is available at http://www.cfdbooks.com.
252 CHAPTER 7. EXACT SOLUTIONS
Shock Wave
M
> 1
Figure 7.13.17: Supersonic ow over a cylinder: the shock
is locally normal to the streamline along the centerline in-
dicated by the dashed line.
7.13.7 Supersonic Flow over a Cylinder (Stagnation Pressure)
The formula that relates the upstream Mach number M
=
_
1 +
2
+ 1
(M
2
1)
_
1/(1)
_
( + 1)M
2
( 1)M
2
+ 2
_
/(1)
, (7.13.77)
(see Ref.[84] for details). This will give the exact pressure at the stagnation point of a cylinder behind a bow shock
created by a supersonic ow (the bow shock is a normal shock at the line of symmetry). This is useful, for example,
when I want to see how accurately my Euler code can compute a supersonic ow over a cylinder (so that there will be
a bow shock in front of it) by comparing a computed pressure at the front stagnation point of the cylinder with the
exact value. For this purpose, I like to use the nondimensionalization by the stagnation values in Section 3.11 because
it allows me to directly compare the computed pressure value with the exact value.
7.13.8 Normal/Oblique Shock Waves
Propagating Normal Shock Wave
I like propagating normal shock waves, especially the one running to the right at the speed V
s
into the gas at rest (this
state is denoted by R),
W
R
= [
R
, u
R
, p
R
]
t
= [1, 0, 1/]
t
. (7.13.78)
It is nice that once I specify the shock speed V
s
(> 0), then all upstream quantities, W
L
= [
L
, u
L
, p
L
]
t
, are automat-
ically determined (by the Rankine-Hugoniot relation) as follows:
p
L
=
_
1 +
2
+ 1
_
V
2
s
1
_
_
1
, (7.13.79)
L
=
1 +
+ 1
1
p
L
+ 1
1
+p
L
, (7.13.80)
u
L
=
p
L
1
1 +
+1
2
(p
L
1)
. (7.13.81)
This can be used in the same form also in two and three dimensions with v = w = 0. It is very useful. Note that it is
very important to be able to compute a moving shock because the predicted shock speed could be wrong if a method
is not conservative.
7.13. EULER EQUATIONS 253
Shock Wave
V
s
L R
Figure 7.13.18: Normal shock wave: propagating V
s
> 0,
or stationary V
s
= 0.
Steady Normal Shock Wave
I like steady normal shock waves. In fact, it is very easy to cook up one: specify the upstream state, W
L
,
W
L
= [
L
, u
L
, p
L
]
t
=
_
1, 1,
1
M
2
L
_
t
, (7.13.82)
where M
L
> 1, then the downstream state, W
R
= [
R
, u
R
, p
R
]
t
is automatically determined in terms of M
L
,
R
=
( + 1)M
2
L
( 1)M
2
L
+ 2
, (7.13.83)
u
R
=
1
R
=
( 1)M
2
L
+ 2
( + 1)M
2
L
, (7.13.84)
p
R
=
_
1 +
2
+ 1
(M
2
L
1)
_
1
M
2
L
. (7.13.85)
Besides, if we express the two states in terms of the conservative variables, U
L
and U
R
, then they look rather simple:
U
L
= [
L
,
L
u
L
,
L
E
L
]
t
=
_
1, 1,
1
( 1)M
2
L
+
1
2
_
t
, (7.13.86)
U
R
= [
R
,
R
u
R
,
R
E
R
]
t
=
_
R
, 1,
p
R
1
+
u
2
R
2
_
t
, (7.13.87)
where
R
, p
R
, and u
R
are given by Equations (7.13.83), (7.13.84), and (7.13.85). Of course, this steady shock solution
can be used in the same form also in two and three dimensions with v = w = 0.
Oblique Shock Waves
We can easily create an oblique shock solution in a square domain by setting a free stream on the left at an angle
(< 0) coming towards the bottom wall, thus creating a oblique shock at the bottom-left corner. Specify the upstream
angle and a desired oblique shock angle
s
(see Figure 7.13.19), and then we can determine the corresponding
upstream Mach number by
M
2
sin(
s
+) cos(
s
+) [( + 1) tan(
s
) ( 1) tan(
s
+)]
, (7.13.88)
where it is safe to use the values that satisfy
s
+ < 60
> 1
Shock Wave
s
Figure 7.13.19: Oblique shock wave.
= 1 and p
_
1 +
1
2
M
2
sin
2
(
s
+)
M
2
sin
2
(
s
+)
1
2
, (7.13.89)
2
=
( + 1)M
2
sin
2
(
s
+)
( 1)M
2
sin
2
(
s
+) + 2
, (7.13.90)
p
2
= 1 +
2
+ 1
(M
2
sin
2
(
s
+) 1). (7.13.91)
(see Ref.[84] for details). It is very nice that the domain is a square; it will be very easy to generate a grid and also to
implement a numerical scheme.
7.14 Riemann Problems for the Euler Equations
The Riemann problem is an initial value problem with a single discontinuity that initially separates two constant
states, the left and right states. In particular, the Riemann problem with zero velocity on both states describes a shock
tube problem in gas dynamics: interactions of gases at two dierent states suddenly made to interact in a tube, e.g.,
by a diaphragm raptured at t = 0. I like Riemann problems for two reasons. First, it can be solved exactly even for
nonlinear equations such as the Euler equations. I nd it very interesting that the problem can be solved exactly for
highly nonlinear equations with discontinuous initial data, which sounds very dicult in general. Second, it can be
thought of as a building block of nite-volume methods. For example, two adjacent control volumes (computational
cells) dene a Riemann problem at the control-volume boundary, and the solution to the Riemann problem gives the
ux at the control-volume boundary that is required to compute the residual in nite-volume methods.
7.14.1 Basic Equations
The 1D Euler equations have three characteristic elds:
1
= u c,
2
= u,
3
= u +c. (7.14.1)
1
-eld and
3
-eld are genuinely nonlinear and therefore have nonlinear waves (shocks/rarefactions), but
2
-eld is
linearly degenerate and so has a contact wave only. See Section 1.23.
In the following equations, whenever they have double signs, the negative sign is associated with waves in
1
-eld
and the positive sign is associated with waves in
3
-eld. Also note that the subscript 0 indicates an undisturbed state.
7.14. RIEMANN PROBLEMS FOR THE EULER EQUATIONS 255
Shocks
The Rankine-Hugoniot relation across a shock is given by
p p
0
=
0
c
0
1 +
+ 1
2
_
p
p
0
1
_
(u u
0
), (7.14.2)
0
=
1 +
+ 1
1
p
p
0
+ 1
1
+
p
p
0
. (7.14.3)
These are obtained directly from the Rankine-Hugoniot relation in Section 1.22 (See Ref.[159] for derivation). Equation
(7.14.2) denes a curve in (p, u)-space, and it is called the Hugoniot curve. The shock speed V
s
is given by
V
s
= u
0
c
0
1 +
+ 1
2
_
p
p
0
1
_
. (7.14.4)
Note that Equation (7.14.2) denes a curve that passes through a point (p
0
, u
0
) and only the branch where p > p
0
is physically admissible. This is typically called the entropy condition: the entropy cannot decrease across the shock.
You will actually see that p > p
0
for s > s
0
by deriving the entropy jump, s s
0
= c
v
[ln(p/
) ln(p
0
/
0
)] from
the above equations. The other branch, where p < p
0
, is often called the non-physical branch of the Hugoniot curve.
Such a change is caused by a rarefaction wave.
Rarefactions
Integral relation across a rarefaction is given by
u
2
1
c = u
0
2
1
c
0
, (7.14.5)
p p
0
=
0
c
0
1
2
1 (p/p
0
)
1 (p/p
0
)
1
2
(u u
0
). (7.14.6)
These equations are obtained from the isentropic relations and the Riemann invariant u+
2c
1
= constant (See Ref.[159]
for derivation). Equation (7.14.6) denes a curve in (p, u)-space, and it is called the Poisson curve. Of course, only the
branch where p < p
0
is physically admissible. Then, a curve describing all admissible jumps is obtained by connecting
the Poisson and Hugoniot curves at (p
0
, u
0
), which will be the main tool for solving a Riemann problem. Across the
rarefaction, the solution does not jump but varies smoothly: u and c vary linearly, and p and are given by
p = p
0
_
c
c
0
_
2
1
, (7.14.7)
=
0
_
p
p
0
_1
. (7.14.8)
7.14.2 Construction of a Riemann Problem
I like the Riemann problem because it is simple. In fact, it is very easy to construct one. For example, if I want to
create a ow where a shock travels to the right, a contact wave follows to the right, and a rarefaction wave travels to
the left (see Figure 7.14.1), I can create it as follows.
256 CHAPTER 7. EXACT SOLUTIONS
O
W
R
W
L
W
R
W
L
S
3
C
2
R
1
tail
R
1
head
x
y
Figure 7.14.1: The Riemann problem I want to create:
3
-shock (S
3
), the contact
(C
2
), and
1
-rarefaction (R
1
).
Construction of a Riemann Problem:
1. Specify the right state W
R
= [
R
, u
R
, p
R
]
t
.
2. Specify p
(> p
R
), and determine u
= u
R
+
p
p
R
R
c
R
1 +
+1
2
_
p
pR
1
_
, (7.14.9)
and determine
R
by Equation (7.14.3), i.e.,
R
=
R
1 +
+ 1
1
p
p
R
+ 1
1
+
p
p
R
. (7.14.10)
3. Specify
L
. Across the contact, the velocity and the pressure do not change: u
L
= u
R
= u
and
p
L
= p
R
= p
.
4. Specify p
L
(> p
), and determine
L
by Equation (7.14.8), i.e.,
L
=
L
_
p
L
p
_1
, (7.14.11)
and the velocity by Equation (7.14.6), i.e.,
u
L
= u
+
2
( 1)
L
c
L
1 (p
/p
L
)
1
2
1 (p
/p
L
)
(p
p
L
). (7.14.12)
This completes the construction of a Riemann problem.
For the Riemann problem created above, the shock speed is given by
V
S3
= u
R
+c
R
1 +
+ 1
2
_
p
p
R
1
_
, (7.14.13)
the contact speed by
V
C2
= u
, (7.14.14)
and the tail and head speeds of the rarefaction by
V
tail
R1
= u
L
c
L
, V
head
R1
= u
L
c
L
. (7.14.15)
7.14. RIEMANN PROBLEMS FOR THE EULER EQUATIONS 257
These wave speeds may be useful when we choose the quantities like p
L
u
L
+
L
L
u
, (7.14.17)
where
L
= u
L
c
L
and
L
= u
L
. Note that you can construct any Riemann problem you want. For example, if
I stop after determine W
R
and set W
L
= W
R
, then I end up with a Riemann problem with a single shock; or if I set
L
=
R
, then there will be no contact waves.
7.14.3 Exact Riemann Solver
I like solving a given Riemann problem also. Sometimes the solution doesnt even exist, and thats fun. Given a set of
initial data, W
L
= [
L
, u
L
, p
L
]
t
and W
R
= [
R
, u
R
, p
R
]
t
, I will try to nd a solution as follows.
Exact Riemann Solver:
1. Find the pressure p
=
m
R
p
L
+m
L
p
R
+m
L
m
R
(u
L
u
R
)
m
L
+m
R
, (7.14.18)
where for Q = L or R, and
m
Q
=
_
Q
c
Q
1 +
+ 1
2
_
p
p
Q
1
_
if
p
p
Q
> 1 : shock,
Q
c
Q
1
2
1
p
p
Q
1
_
p
p
Q
_
1
2
if
p
p
Q
< 1 + : rarefaction.
(7.14.19)
Note that , say = 10
15
for double precision, is necessary to avoid the zero division. This must
be solved iteratively as the right hand side depends on p
.
2. Compute u
by the equation,
u
=
m
L
u
L
+m
R
u
R
+ (p
L
p
R
)
m
L
+m
R
. (7.14.20)
3. Compute
L
and
R
by
Q
=
_
Q
1 +
+ 1
1
p
p
Q
+ 1
1
+
p
p
Q
if
p
p
Q
> 1 : shock,
Q
_
p
p
Q
_1
if
p
p
Q
< 1 : rarefaction,
(7.14.21)
where Q = L or R. Now, the middle states have been completely determined, and this completes
the solution of a Riemann problem.
258 CHAPTER 7. EXACT SOLUTIONS
Some remarks are in order. First, if a rarefaction wave is detected, the solution inside the wave can be computed
by Equations (7.14.16) and (7.14.17). Second, it may happen that
u
L
+
2
1
c
L
< u
R
+
2
1
c
R
, (7.14.22)
then the pressure p
will be negative and thus a solution does not exist. We need to check this condition and make
sure that a solution exists, before solving a Riemann problem. Third, the intermediate pressure equation (7.14.18) is
obtained by solving the following system for p
,
p
p
L
= m
L
(u
u
L
),
1
-shock/rarefaction, (7.14.23)
p
p
R
= m
R
(u
u
R
),
3
-shock/rarefaction. (7.14.24)
These are the shock/rarefaction relations across the right-running and left-running waves respectively. Solving this for
u
for a given p
:
Upper bound: p
=
_
1
2
(u
L
u
R
) + (c
L
+c
R
)
c
L
(p
L
)
1
2
+c
R
(p
R
)
1
2
_
2
1
, (7.14.25)
Lower bound: p
=
m
R
p
L
+m
L
p
R
+m
L
m
R
(u
L
u
R
)
m
L
+m
R
, (7.14.26)
where m
L
=
L
c
L
and m
R
=
R
c
R
. The upper bound comes from a solution for two rarefaction waves (the intersection
of the two Poisson curves); the lower bound comes from a linearization, i.e., the solution of Equations (7.14.23) and
(7.14.24) with m
L
=
L
c
L
and m
R
=
R
c
R
. Fifth, the condition (7.14.22) indicates that the intersection between
the two Poisson curves (i.e., the upper bound) goes down to the negative part of the p-axis in the (p, u)-space.
Here are some examples where I assume a domain [5, 5] for all problems and denote the nal time by t
f
.
(a) Sods Problem 1 (Figures 7.14.2):
I like this problem. This is a famous shock tube problem by Sod [144].
W
L
= [
L
, u
L
, p
L
]
t
= [1.0, 0, 1.0]
t
, (7.14.27)
W
R
= [
R
, u
R
, p
R
]
t
= [0.125, 0, 0.1]
t
, (7.14.28)
t
f
= 1.7. (7.14.29)
(b) Sods Problem 2 (Figures 7.14.3):
I like this one also. This is another famous shock tube problem by Sod [144].
W
L
= [
L
, u
L
, p
L
]
t
= [1.0, 0, 1.0]
t
, (7.14.30)
W
R
= [
R
, u
R
, p
R
]
t
= [0.01, 0, 0.01]
t
, (7.14.31)
t
f
= 1.4. (7.14.32)
I like this one particularly because it contains a sonic point in the rarefaction wave and some schemes create a
small gap there.
(c) Sonic Rarefaction (Figures 7.14.4):
This is a sonic rarefaction and some schemes cannot break the initial discontinuity at all.
W
L
= [
L
, u
L
, p
L
]
t
= [3.857, 0.92, 10.333]
t
, (7.14.33)
W
R
= [
R
, u
R
, p
R
]
t
= [1.000, 3.55, 1.000]
t
, (7.14.34)
t
f
= 0.7. (7.14.35)
7.14. RIEMANN PROBLEMS FOR THE EULER EQUATIONS 259
! " !
"#$%
$
x
D
e
n
s
i
t
y
(a) Density, .
! " !
"#$
$
x
P
r
e
s
s
u
r
e
(b) Pressure, p.
! " !
"
"#$%
x
V
e
l
o
c
i
t
y
(c) Velocity, u.
! " !
"
#$%%
x
E
n
t
r
o
p
y
(d) Entropy, s/R =
1
1
ln(p/
).
Figure 7.14.2: Solutions for Sods problem 1.
(d) Slowly Moving Shock (Figures 7.14.5):
For this one, many schemes produce oscillations behind the shock [4].
W
L
= [
L
, u
L
, p
L
]
t
= [3.86, 0.81, 10.33]
t
, (7.14.36)
W
R
= [
R
, u
R
, p
R
]
t
= [1.00, 3.44, 1.00]
t
, (7.14.37)
t
f
= 1.7. (7.14.38)
(e) Left-Moving Waves in a Right-Moving Supersonic Flow:
I like this one,
W
L
= [
L
, u
L
, p
L
]
t
= [1.4, 1.1, 1]
t
, (7.14.39)
W
R
= [
R
, u
R
, p
R
]
t
= [1.4, 110, 100]
t
, (7.14.40)
thus M
L
= M
R
= 1.1. This is an interesting case where both states are supersonic to the right but some waves
will propagate to the left. This is indeed confusing because we tend to assume that all waves propagate to the
right if both states are supersonic to the right. That is true for scalar conservation laws but not for systems. See
it for yourself using your Euler code, or see Ref.[86] for details.
260 CHAPTER 7. EXACT SOLUTIONS
! " !
"#"$
$
x
D
e
n
s
i
t
y
(a) Density, .
! " !
"#"$
$
x
P
r
e
s
s
u
r
e
(b) Pressure, p.
! " !
"
#$%&
x
V
e
l
o
c
i
t
y
(c) Velocity, u.
! " !
"
!#$
x
E
n
t
r
o
p
y
(d) Entropy, s/R =
1
1
ln(p/
).
Figure 7.14.3: Solutions for Sods problem 2.
7.14. RIEMANN PROBLEMS FOR THE EULER EQUATIONS 261
! " !
"#$%
&#'%
x
D
e
n
s
i
t
y
(a) Density, .
! " !
#
#"$%%
x
P
r
e
s
s
u
r
e
(b) Pressure, p.
! " !
"#$%
&#'
x
V
e
l
o
c
i
t
y
(c) Velocity, u.
! " !
"
#$##
x
E
n
t
r
o
p
y
(d) Entropy, s/R =
1
1
ln(p/
).
Figure 7.14.4: Solutions for a sonic rarefaction.
262 CHAPTER 7. EXACT SOLUTIONS
! " !
#
$%&'
x
D
e
n
s
i
t
y
(a) Density, .
! " !
#
#"$%%
x
P
r
e
s
s
u
r
e
(b) Pressure, p.
! " !
#$%%
"$&'
x
V
e
l
o
c
i
t
y
(c) Velocity, u.
! " !
"
#$##
x
E
n
t
r
o
p
y
(d) Entropy, s/R =
1
1
ln(p/
).
Figure 7.14.5: Solutions for a slowly moving shock.
7.15. INCOMPRESSIBLE NAVIER-STOKES EQUATIONS 263
-1 0 1 2 3
-1
0
1
2
3
x
y
Figure 7.15.1: Unsteady solution for the 2D incompressible Navier-Stokes equations.
The velocity eld at t = 0.0 for k = 0.5 and = 0.01. The velocity will be damped
out as time goes on.
7.15 Incompressible Navier-Stokes Equations
I like incompressible Navier-Stokes equations because it does have exact solutions. Although often limited to sim-
ple geometries, these exact solutions can be useful for CFD code verication. Of course, I like Stokes equations also
because there exist a variety of exact solutions as mentioned in Section 4.25. Especially in two dimensions, the theory
of complex variables can be applied to generate various exact solutions. There are also various exact solutions available
for the axisymmetric Stokes ows as we will see in this section. Note that I like those exact solutions of the Stokes
equations where the convective terms vanish because then the solutions are also exact for the Navier-Stokes equations.
7.15.1 Smooth 2D Time-Dependent Solution
The two-dimensional incompressible Navier-Stokes equations,
u
x
+v
y
= 0, (7.15.1)
u
t
+uu
x
+vu
y
= p
x
+(u
xx
+u
yy
), (7.15.2)
v
t
+uv
x
+vv
y
= p
y
+(v
xx
+v
yy
), (7.15.3)
have the exact solution given by
u(x, y, t) = cos(kx) sin(ky) e
2k
2
t
, (7.15.4)
v(x, y, t) = sin(kx) cos(ky) e
2k
2
t
, (7.15.5)
p(x, y, t) =
1
4
[cos(2kx) + cos(2ky)] e
4k
2
t
, (7.15.6)
where k is an arbitrary constant [22, 33]. This solution can also be expressed in terms of the vorticity = v
x
u
y
and
the stream function ,
(x, y, t) = 2k cos(kx) cos(ky) e
2k
2
t
, (7.15.7)
(x, y, t) =
1
k
cos(kx) cos(ky) e
2k
2
t
, (7.15.8)
which is exact for the equation set (4.22.1) and (4.22.3), i.e.,
t
+u
x
+v
y
= (
xx
+
yy
) , (7.15.9)
xx
+
yy
= , (7.15.10)
where u =
y
and v =
x
. I like this solution very much because it is a rather simple solution for such a complicated
nonlinear system. See Figure 7.15.1 for the velocity eld generated by this solution. It is very interesting, isnt it?
264 CHAPTER 7. EXACT SOLUTIONS
O
h
U
x
y
Figure 7.15.2: Couette or Couette-Poiseuille ow: dp/dx = 0 or dp/dx = 0.
7.15.2 Smooth 3D Time-Dependent Solution
The three-dimensional incompressible Navier-Stokes equations,
u
x
+v
y
+w
z
= 0, (7.15.11)
u
t
+uu
x
+vu
y
+wu
z
= p
x
+(u
xx
+u
yy
+u
zz
), (7.15.12)
v
t
+uv
x
+vv
y
+wv
z
= p
y
+(v
xx
+v
yy
+v
zz
), (7.15.13)
w
t
+uw
x
+vw
y
+ww
z
= p
z
+(w
xx
+w
yy
+w
zz
), (7.15.14)
have the exact solution given by
u(x, y, t) = a [e
ax
sin(ay +dz) +e
az
sin(ax +dy)] e
d
2
t
, (7.15.15)
v(x, y, t) = a [e
ay
sin(az +dx) +e
ax
sin(ay +dz)] e
d
2
t
, (7.15.16)
w(x, y, t) = a [e
az
sin(ax +dy) +e
ay
sin(az +dx)] e
d
2
t
, (7.15.17)
p(x, y, t) =
a
2
2
_
e
2ax
+e
2ay
+e
2az
+ 2 sin(az +dx) cos(ay +dz)
+2 sin(ay +dz) cos(ax +dy) + 2 sin(ax +dy) cos(az +dx)] e
2d
2
t
,
(7.15.18)
where a and d are arbitrary constants [35]. I like this solution very much because it is, again, a very simple solution for
such a complicated nonlinear system of equations. I just want to point out that there is a typo in the original paper
[35].
7.15.3 Couette Flow
Consider a constant-density ow between two innite plates that are h apart with the upper plate moving at speed U
relative to the lower. Also, we assume that the upper plate is held at temperature T
1
and the lower plate at T
0
. I like
this setting because it brings a dramatic simplication to the incompressible Navier-Stokes equations. In this situation,
naturally, x-derivative of any quantity is zero as well as the y-component of the velocity is zero everywhere. Also the
pressure is constant everywhere since the ow does not accelerate in any direction. Then, the continuity equation is
trivially satised (all terms drop out), and we have the momentum and energy equations as follows,
0 = u
yy
, (7.15.19)
0 = T
yy
+(u
y
)
2
. (7.15.20)
where is the viscosity coecient and is the heat conductivity. This is very nice. We can easily obtain the exact
solutions: integrate the momentum equation to obtain the velocity, and then integrate the energy equation to obtain
7.15. INCOMPRESSIBLE NAVIER-STOKES EQUATIONS 265
the temperature as follows,
u(y) = U
y
h
, (7.15.21)
T(y) = T
0
+ (T
1
T
0
)
y
h
U
2
2
_
_
y
h
_
2
y
h
_
. (7.15.22)
This simple ow is called the Couette ow. It can be a nice and simple test case for a two-dimensional Navier-Stokes
code in a rectangular domain with a periodic boundary condition in x-direction, i.e., no in-ow and out-ow. Note
that the velocity is a linear function of y. Most practical incompressible Navier-Stokes codes should be able to produce
the exact velocity prole since they are typically designed to be exact for linear solutions.
7.15.4 Unsteady Couette Flow
The Couette ow can be made unsteady by oscillating the upper plate with the following time-dependent speed,
u(h) = U cos(t), (7.15.23)
where is a constant which denes the frequency. In this case, the momentum equation is given by
u
t
= u
yy
. (7.15.24)
This problem was solved in Section 6.1.5 by the separation of variable with complex variables. The solution is
u(y, t) = Q[sinh(ky) cos(ky) sinh(kd) cos(kd) + cosh(ky) sin(ky) cosh(kd) sin(kd)] cos(t)
+Q[sinh(ky) cos(ky) cosh(kd) sin(kd) cosh(ky) sin(ky) sinh(kd) cos(kd)] sin(t), (7.15.25)
where
Q =
U
cosh
2
(kd) cos
2
(kd)
, k =
2
. (7.15.26)
I like this one because the solution does not decay in time and oscillates forever, preserving the same amplitude. It
would be interesting to see how long a numerical scheme can simulate this solution accurately, of course with a periodic
boundary condition in x-direction.
7.15.5 Couette-Poiseuille Flow
I like ows with constant pressure gradient because it is often possible to obtain the exact solution. For example, I like
the Couette-Poiseuille ow, which is the Couette ow with an added constant pressure gradient, dp/dx = constant.
The momentum equation and the energy are then given by
0 = p
x
+u
yy
, (7.15.27)
0 = kT
yy
+(u
y
)
2
. (7.15.28)
We can integrate these equations straightforwardly and obtain
u(y) = U
y
h
+
h
2
2
_
dp
dx
__
y
h
_
y
h
_
2
_
, (7.15.29)
T(y) = T
0
+ (T
1
T
0
)
y
h
3
12k
_
dp
dx
_
2 _
V (y) V (0) {V (h) V (0)}
y
h
_
, (7.15.30)
where
V (y) = (u
y
)
4
=
_
U
h
+
h
2
_
dp
dx
_
_
1 2
y
h
_
_
4
. (7.15.31)
It is more complicated than the solution of the Couette ow, but the ow itself is very simple. If your incompressible
Navier-Stokes code is accurate enough to preserve quadratic solutions, it should be able to produce the exact velocity
prole.
266 CHAPTER 7. EXACT SOLUTIONS
dp
dz
< 0
u
z
O
z
O
r
R
Figure 7.15.3: Hagen-Poiseuille ow.
7.15.6 Hagen-Poiseuille Flow
The Hagen-Poiseuille ow is an incompressible ow through a pipe of radius R driven by a constant pressure gradient
in the z-direction in cylindrical coordinates. The only non-zero velocity component is u
z
and all variables are functions
of r only, i.e., u
z
= u
z
(r). Then, the continuity equation is satised identically, and the momentum equation is given
by
0 =
dp
dz
+
1
r
d
dr
_
r
du
z
dr
_
. (7.15.32)
Here, I like the fact that dp/dz is constant because it makes it very easy to integrate this equation. Simply integrating
twice and applying the boundary conditions, i.e., u
z
(R) = 0 and u
z
(0) is nite, we obtain
u
z
(r) =
1
4
_
dp
dz
_
(R
2
r
2
). (7.15.33)
The velocity distribution is therefore parabolic, just like that of the Couette-Poiseuille ow. Oh, I forgot to work out a
temperature distribution. Did you say that you wanted to derive it by yourself? OK, Ill leave it to you.
7.15.7 Axially Moving Co-centric Cylinders
Consider an incompressible ow between axially moving cylinders at constant speeds (see Figure 7.15.4). For this ow,
clearly, the velocity components (u
r
, u
, u
z
) are such that u
z
= u
z
(r) and u
r
= u
0
, and temperature T
0
, while the outer cylinder has r
1
,
1
, and T
1
respectively. Note that the only non-zero velocity
component is u
, and all variables are functions of r only. Then, the continuity equation is trivially satised, and the
momentum equations and the energy equation are simplied as follows,
0 =
dp
dr
+
u
2
r
, (7.15.38)
0 =
d
2
u
dr
2
+
d
dr
_
u
r
_
, (7.15.39)
0 =
k
r
d
dr
_
r
dT
dr
_
+
_
du
dr
u
r
_
2
. (7.15.40)
These are so simple that we can integrate them analytically to get the exact solutions:
u
= r
0
0
r
1
/r r/r
1
r
1
/r
0
r
0
/r
1
+r
1
1
r/r
0
r
0
/r
r
1
/r
0
r
0
/r
1
, (7.15.41)
T T
0
T
1
T
0
=
r
2
0
2
0
k(T
1
T
0
)
r
4
1
(1
1
/
0
)
2
r
4
1
r
4
0
_
1
r
2
0
r
2
__
1
ln(r/r
0
)
ln(r
1
/r
0
)
_
+
ln(r/r
0
)
ln(r
1
/r
0
)
. (7.15.42)
Oh, I forgot to calculate the pressure. Well, you can get the pressure easily (if you wish) by integrating the momentum
equation (7.15.38). I think that it can be a good exercise. By the way, unlike the previous cases, in this ow the
pressure varies across the shear layer. This is because the domain is curved: the pressure gradient is needed to balance
the centrifugal force, or equivalently, to create the centripetal force.
7.15.9 Flat Plate Boundary Layer
The at plate boundary layer solution is a solution to the boundary layer equations,
u
x
+
v
y
= 0, (7.15.43)
u
u
x
+v
u
y
=
2
u
y
2
, (7.15.44)
for a ow over a semi-innite at plate that extends x (Figure 7.15.6). First, to satisfy the continuity equation,
we introduce the stream function ,
=
U
Re
x
f(), (7.15.45)
268 CHAPTER 7. EXACT SOLUTIONS
r
1
r
0
r
r
1
1
r
0
0
Figure 7.15.5: Rotating co-centric cylinders.
where U
Re
x
, Re
x
=
U
, (7.15.46)
so that the velocity components are given by
u
U
= f
, (7.15.47)
v
U
=
1
2
Re
x
(f
f). (7.15.48)
Therefore, the velocity component u depends only on , i.e., self-similar. This is very nice. The solution is valid at any
location on the at plate (possibly far enough from the leading edge). On the other hand, v depends both on and
x. This may not sound very nice, but it makes sense. The boundary layer grows rapidly from the leading edge, but it
will not grow so much in the downstream (x ) simply because there are no reasons for that. In order to compute
the velocity components, we have to nd f() by solving the momentum equation which is written now in terms of f,
f
+
1
2
ff
= 0, (7.15.49)
with the boundary conditions,
u(x, 0) = v(x, 0) = 0 f(0) = f
(0) = 0, (7.15.50)
u(x, ) = U
() = 1. (7.15.51)
To solve this numerically, we rewrite this as a system of ordinary dierential equations (ODE):
dF
d
= G(F), (7.15.52)
where
F =
_
_
f
f
1
f
2
_
_
, G(F)
_
_
f
1
f
2
1
2
ff
2
_
_
, (7.15.53)
with
f(0) = f
1
(0) = 0, (7.15.54)
f
1
() = 1. (7.15.55)
7.15. INCOMPRESSIBLE NAVIER-STOKES EQUATIONS 269
U
+
1
2
gg
= 0, with g(0) = g
(0) = 0 and g
(0) = C
g2
for an arbitrary constant C
g2
, then
f = g(), =
1
()
, (7.15.56)
is the solution of Equation (7.15.49) that satises the desired boundary condition f
(0) =
3
g
(0) =
3
C
g2
. (7.15.57)
That is, integrating with this initial condition, you get what you want in one shot. So, here is a two step method. The
following computes the velocity (u, v) at a given location (x, y) = (x
p
, y
p
), =
p
.
Computation of Exact Solution:
To really call this exact, you have to solve the ODE by a very high-order method with extremely small
steps (e.g., the 4th-order Runge-Kutta method with = 10
5
). Note also that we set C
g2
= 1.
1. Numerically solve Equation (7.15.52) with f(0) = f
1
(0) = 0 and f
2
(0) = 1, say, up to = 20; call
this solution g() and set = 1/
(20).
2. Numerically solve Equation (7.15.52) with f(0) = f
1
(0) = 0 and f
2
(0) =
3
, up to the desired
location =
p
.
3. Compute the velocity there by
u
U
= f
1
(
p
), (7.15.58)
v
U
=
1
2
Re
xp
(f
1
(
p
) f(
p
)). (7.15.59)
Honestly speaking, I like to skip the rst step. So, I have computed (by using the 4th-order Runge-Kutta method
with = 5.0 10
6
and integrating the ODE up to = 100) the value of g
():
g
(100)
_
3
= 0.3320573362151946. (7.15.61)
So, I can skip the rst step from now on, and just perform the second step using this value. This is a very nice
and simple one step method. I like it so much that I coded the algorithm and have made the code available at
http://www.cfdbooks.com. I hope that you like it.
270 CHAPTER 7. EXACT SOLUTIONS
0 1
0
15
u / U
y
= x
2
, v =
1
y
x
= xy, (7.15.63)
which shows that the velocity vanishes at the wall (x = 0). It thus satises the no-slip condition in contrast to the
inviscid stagnation ow where v = 0 at x = 0. So, it is a viscous ow. For the viscous stagnation ow, it can be easily
shown that the vorticity is not zero:
=
v
x
u
y
=
y
2
, (7.15.64)
except along the line of symmetry (y = 0). This is the reason that the velocity potential cannot be dened. I like the
fact that the velocity components are quadratic function of space. I would expect that a third-order accurate scheme
will produce exact results for the viscous stagnation ow. How about the pressure? You can nd the pressure by
integrating the momentum equations (4.28.2) and (4.28.3):
p = p
w
2x, (7.15.65)
where p
w
denotes the pressure at the wall. So, the pressure is a linear function of space. This is interesting. Recall
that in the inviscid stagnation ow, the velocity is linear and the pressure is quadratic.
If you compare the solution above with the inviscid stagnation ow solution in Section 7.12.1, you nd that the
viscous solution does not match the inviscid one as x . In order to obtain the viscous solution that matches the
inviscid one, it is necessary to solve ordinary dierential equations. See Ref.[120, 169] for details.
By the way, if you asked "Where is the stagnation point?", that would be a good question. I like it. If the stagnation
point is dened as a point where the velocity is zero, then the entire wall will be the stagnation point in the viscous
stagnation ow. But if the stagnation point is dened as a point from which the streamlines separate apart (the
7.15. INCOMPRESSIBLE NAVIER-STOKES EQUATIONS 271
direction of surface streamlines cannot be dened), then the stagnation point will be at the origin (x, y) = (0, 0),
which matches the stagnation point in the inviscid stagnation ow. I like that. In uid dynamics, I think we actually
almost always take the latter denition; otherwise the word "stagnation point" will not be very useful in practical
applications.
7.15.11 Stokes Flow over a Sphere
Stokes ow over a sphere is an exact solution to the steady Stokes equations (see Section 4.25), which may be
expressed by
div v = 0, (7.15.66)
div grad = 0, (7.15.67)
where = curl v. We may introduce spherical coordinates (r, , ) with axisymmetry with respect to (i.e.,
= 0),
and write the continuity equation as
1
r
2
(r
2
u
r
)
r
+
1
r sin
(u
sin )
= 0. (7.15.68)
Then, we can dene the stream function as
u
r
=
1
r
2
sin
, u
=
1
r sin
r
, (7.15.69)
to automatically satisfy the continuity equation. In terms of the stream function, the other equation (7.15.67) becomes
_
2
r
2
+
1
r
2
2
1
r
2
tan
_
= 0. (7.15.70)
Now consider a uniform ow with the speed U (in the direction of z) over a sphere of radius R at the origin, i.e.,
r
=
= 0, on the sphere, r = R,
1
2
Ur
2
sin
2
+ constant, as r .
(7.15.71)
Stokes discovered the following exact solution,
=
1
4
UR
2
sin
2
_
R
r
3r
R
+
2r
2
R
2
_
, (7.15.72)
so that the velocity components are
u
r
= U cos
_
1 +
R
3
2r
3
3R
2r
_
, (7.15.73)
u
= U sin
_
1
R
3
4r
3
3R
4r
_
, (7.15.74)
the pressure is found from the momentum equation, grad p = div grad v, to be
p(r, ) p
=
3RU
2r
2
cos , (7.15.75)
(see Figures 7.15.9 and 7.15.10), the viscous stress on the sphere is given by
r
(r, ) =
_
1
r
u
r
+
u
r
u
r
3R
3
2r
3
U sin
r
_
, (7.15.76)
and therefore the drag force F is found by combining the viscous drag and the pressure drag,
F =
r
(R, ) sin (2R
2
sin d)
0
p(R, ) cos (2R
2
sin d)
= 4UR + 4UR
= 6UR. (7.15.77)
272 CHAPTER 7. EXACT SOLUTIONS
Figure 7.15.9: Pressure p p
du
dy
_
, (7.16.1)
0 =
d
dy
_
dT
dy
_
+
_
du
dy
_
2
. (7.16.2)
Note that the pressure is constant everywhere (no acceleration) but the density and the temperature change in the
y-direction. We also assume
= (T), = (T), (7.16.3)
i.e., the viscosity and the heat conductivity are functions of T only. Moreover, we assume that the Prandtl number
Pr is constant,
Pr =
c
p
= constant. (7.16.4)
Now, rst, we integrate the momentum equation (7.16.1) once to get,
du
dy
=
du
dy
y=0
=
w
, (7.16.5)
7.16. COMPRESSIBLE NAVIER-STOKES EQUATIONS 273
where
w
denotes the shear stress at the bottom wall. Then, integrating once more, we obtain
u(y) =
w
y
0
dy
(T)
. (7.16.6)
We will nd that this is a very nice relation when we use it later to eliminate the same integral arising in an energy
integral. Also, remember that we will integrate Equation (7.16.5) (in a slightly dierent way) to nally obtain the
velocity distribution.
We now consider integrating the energy equation. First, we note that
d
dy
_
u
du
dy
_
= u
d
dy
_
du
dy
_
+
_
du
dy
_
2
=
_
du
dy
_
2
, (7.16.7)
where the momentum equation has been used in the last step. Using this, we can write the energy equation (7.16.2)
as
0 =
d
dy
_
dT
dy
+u
du
dy
_
, (7.16.8)
so that it can be easily integrated once,
dT
dy
+u
du
dy
=
dT
dy
y=0
+ u
du
dy
y=0
= q
w
, (7.16.9)
where q
w
is the heat ux at the bottom wall (into the wall when q
w
< 0) and the no-slip condition u(0) = 0 has been
applied. This can be integrated once more by introducing the Prandtl number which is constant,
d
dy
_
c
p
T
Pr
+
u
2
2
_
= q
w
, (7.16.10)
thus,
c
p
(T T
w
)
Pr
+
u
2
2
= q
w
y
0
dy
(T)
. (7.16.11)
It is nice here that the integral on the right hand side can be eliminated by Equation (7.16.6):
c
p
(T T
w
)
Pr
+
u
2
2
=
q
w
u
w
. (7.16.12)
It is very nice that the integral is gone. Just for convenience, we eliminate also T
w
from above as follows: evaluate it
at y = d (the upper plate),
c
p
(T
T
w
)
Pr
+
U
2
2
=
q
w
U
w
, (7.16.13)
where the subscript denotes the state at the upper plate, then subtract Equation (7.16.12) from this to get
c
p
(T T
)
Pr
+
u
2
U
2
2
=
q
w
(u U
w
, (7.16.14)
which can be arranged into the following form,
T
T
= 1 +
q
w
w
U
( 1)PrM
2
_
1
u
U
_
+
1
2
PrM
2
_
1
u
2
U
2
_
, (7.16.15)
where we have used c
p
= R/( 1). Finally, in order to obtain the velocity distribution, we assume a power low
=
_
T
T
, (7.16.16)
274 CHAPTER 7. EXACT SOLUTIONS
0 1
0
1
u/U
y
/
d
Figure 7.16.1: Velocity prole for the compressible Couette ow with Pr = 0.73, = 1.4; M
= 20 (thick
solid line), M
= 3 (dashed line), M
w
y
u
0
du =
u
0
_
T
T
du
=
u
0
_
1 +
q
w
w
U
( 1)PrM
2
_
1
u
U
_
+
1
2
PrM
2
_
1
u
2
U
2
__
du. (7.16.17)
This can be integrated analytically (using the beta function), but the result will be very complicated. For code
verication purpose, it suces to take = 1 and q
w
= 0. Then, nally, we obtain the following implicit expression for
the velocity distribution,
y
d
_
1 +
1
3
PrM
2
_
=
u
U
_
1 +
1
2
PrM
2
_
1
1
3
u
2
U
2
__
, (7.16.18)
where
w
has been determined by the boundary condition u(d) = U
1 +
1
2
PrM
2
_
1
1
3
(V
k
)
2
, (7.16.19)
until |V
k+1
V
k
| < 10
15
, where V
k
= u
k
/U
_
u
u
2
+p
uH u +q
_
_ = 0, (7.16.20)
where
=
4
3
du
dx
, q =
Pr( 1)
d(p/)
dx
. (7.16.21)
Integrating the Navier-Stokes equations (7.16.20), we obtain
_
_
u
u
2
+p
uH u +q
_
_ = constant. (7.16.22)
Then, we apply this equation inside and outside the viscous shock to get
u = (u)
o
, (7.16.23)
u
2
+p = (u
2
+p)
o
, (7.16.24)
uH u +q = (uH)
o
, (7.16.25)
where the subscript o denotes the outside state and the inside state has no subscript. Note that = q = 0 outside
the viscous shock where the solution is constant. Note also that the terms on the right hand side are all constant; we
dene them as
m = (u)
o
, C
2
= (u
2
+p)
o
, C
3
= (uH)
o
, (7.16.26)
so that Equation (7.16.23) becomes u = m and can be substituted into other equations. We therefore obtain from
Equations (7.16.24) and (7.16.25)
= mu +p C
2
, (7.16.27)
q =
1
2
mu
2
m
1
p
C
2
u +C
3
. (7.16.28)
By Equations (7.16.21), we obtain
du
dx
=
1
K
u
_
mu +
m
u
p
C
2
_
, (7.16.29)
d(p/)
dx
=
1
K
T
_
1
2
mu
2
m
1
p
C
2
u +C
3
_
, (7.16.30)
where
K
u
=
4
3
, K
T
=
Pr( 1)
. (7.16.31)
If we assume that all variables have been nondimensionalized as in Section 4.14.3, then p/ = T/, and therefore
these equations can be written as
du
dx
=
1
C
u
_
mu +
mT
u
C
2
_
, (7.16.32)
dT
dx
=
C
T
_
1
2
mu
2
m
1
T
C
2
u +C
3
_
, (7.16.33)
where
C
u
=
4
3
Re
, C
T
=
Pr( 1)
M
Re
. (7.16.34)
276 CHAPTER 7. EXACT SOLUTIONS
1 0 1
1
2
3
4
5
x
D
e
n
s
i
t
y
1 0 1
0
1
2
3
4
u
x
V
e
l
o
c
i
t
y
u
1 0 1
0
5
10
15
p
x
P
r
e
s
s
u
r
e
p
1 0 1
3
2
1
0
T
x
T
e
m
p
e
r
a
t
u
r
e
T
1 0 1
6
4
2
0
x
V
i
s
c
o
u
s
s
t
r
e
s
s
1 0 1
1
1.5
2
2.5
3
3.5
q
x
H
e
a
t
f
l
u
x
q
Figure 7.16.2: Exact viscous shock solution: Pr = 3/4, Re
= 25, M
= 3.5, = 1.4, T
= 400(K).
This is a system of nonlinear ordinary dierential equations for u and T (the viscosity depends on T). It can be
integrated numerically in a straightforward manner. In what follows, we employ this nondimensionalized system.
To integrate the system, we rst choose the Mach number on the left side of the shock M
L
, and set up the left
and right states as follows:
L
= 1,
u
L
= M
L
,
p
L
= 1/,
T
L
= p
L
/
L
,
R
=
( + 1)M
2
L
( 1)M
2
L
+ 2
,
u
R
=
L
M
L
/
R
,
p
R
=
1
_
1 +
2
+ 1
(M
2
L
1)
_
,
T
R
= p
R
/
R
.
(7.16.35)
Note again that all variables are assumed to have been nondimensionalized as in Section 4.14.3. The constants are
computed by the left (or the right) state:
m =
L
u
L
, C
2
= mu
L
+p
L
, C
3
= u
L
_
p
L
1
+
1
2
mu
L
_
. (7.16.36)
We are now ready to integrate the system. Suppose we wish to compute the solution at every node in a one-dimensional
grid dened by the set of coordinates {x
k
} = x
1
, x
2
, x
3
, . . . , x
N
(x
1
< x
2
< . . . < x
N
), where N is the number of
nodes in the grid. We consider integrating the system from the right state. First, we numerically integrate the system
from x = x
N
to x = x
N1
with the initial values: u(x
N
) = u
R
+ where is a small number (e.g., = 1.0E-10)
7.16. COMPRESSIBLE NAVIER-STOKES EQUATIONS 277
and T(x
N
) = T
R
with a very small step x compared with x
N
x
N1
. Then, we do the same between two adjacent
nodes, k and k 1, where k = N 1, . . . , 3, 2, with the solutions at x = x
k
as the initial values. For example, we may
employ the classical fourth-order Runge-Kutta scheme with x =
x
k
x
k1
1000
between the two nodes, k and k 1. Note
that the last step in the integration between two nodes needs to be adjusted to nish exactly at x = x
k1
. Figure
7.16.2 shows the exact solution computed as described for N = 200. A sample program, which was used to generate
the solution in Figure 7.16.2, is available at http://www.cfdbooks.com. I hope you like it.
I like the viscous shock structure solution very much for two reasons. First, the solution is not really physical. The
ow inside the shock wave is unlikely to be in equilibrium, and therefore the Navier-Stokes system is not a valid model.
However, it is still useful for verifying the accuracy of a Navier-Stokes solver. Second, the small number is essential
to obtaining a meaningful solution. In fact, if = 0, the computation will fail. Nothing will happen because the right
hand side of the system (7.16.32) and (7.16.33) are exactly zero. We will not be able to reach the correct left state
of the shock. This is very interesting. The location of the shock depends on the magnitude of . You might want
to experiment with varying to see how it aects the shock location (or the existence). This solution has been used
successfully for accuracy verication by many researchers [106, 110, 175].
7.16.3 Manufactured Solutions
We can manufacture exact solutions for the compressible Navier-Stokes equations. In fact, many people do. See
Refs.[83, 91, 133] for examples. In particular, I like the idea of using exact solutions of the incompressible Navier-
Stokes Equations as manufactured solutions for the compressible Navier-Stokes Equations: substitute an incompressible
solution into the compressible equations and generate an appropriate source term. This way, we can have an exact
solution which looks like a physical ow. For example, in Ref.[118], the Couette-Poiseuille solution described in
Section 7.15.5 was used for verifying the accuracy of a high-order compressible Navier-Stokes code. You want to try it
now? Then, pick any incompressible ow solution from Section 7.15, and generate an appropriate source term for the
compressible Navier-Stokes equations (possibly by using a math software or by computing the source term numerically).
278 CHAPTER 7. EXACT SOLUTIONS
Bibliography
[1] S. Abarbanel and D. Gottlieb. Optimal Time Splitting for Two and Three Dimensional Navier-Stokes Equations
with Mixed Derivatives. ICASE Report 80-6 (1980) (see pp. 150, 152).
[2] I. Altas, J. Erhel, and M. M. Gupta. High Accuracy Solution of Three-Dimensional Biharmonic Equations.
Numerical Algorithms 29 (2002), pp. 119 (see p. 55).
[3] J. S. L. Alzaeili and K. Mazaheri. Bulk Viscosity Damping for Accelerating Convergence of Compressible Viscous
Flow Solvers. In: Proceedings of ECCOMAS CFD 2006. 2006 (see p. 133).
[4] M. Arora and P. L. Roe. On Postshock Oscillations due to Shock Capturing Schemes in Unsteady Flow. J.
Comput. Phys. 130 (1997), pp. 124 (see p. 259).
[5] K. A. Bagrinovskii and S. K. Godunov. Dierence Schemes for Multidimensional Problems. Dokl. Akad. Nauk
SSSR (NS) 115 (1957), pp. 431433 (see p. 37).
[6] B. S. Baldwin and T. J. Barth. A One-Equation Turbulence Transport Model for High Reynolds Number Wall-
Bounded Flows. NASA TM-102847. 1990 (see p. 177).
[7] B. S. Baldwin and T. J. Barth. A One-Equation Turbulence Transport Model for High Reynolds Number Wall-
Bounded Flows. In: AIAA Paper 91-0610. 1991 (see p. 177).
[8] B. S. Baldwin and H. Lomax. Thin-Layer Approximation and Algebraic Model for Separated Turbulent Flows.
AIAA Paper 78257. 1978 (see p. 177).
[9] A. F. Barannyk and I. I. Yuryk. Construction of Exact Solutions of Diusion Equation. In: Proceedings of
Institute of Mathematics of NAS of Ukraine. Vol. 50, Part 1. Institute of Mathematics, 2004, pp. 2933 (see
p. 190).
[10] T. J. Barth. Numerical Methods for Gasdynamic Systems on Unstructured Meshes. In: An Introduction to
Recent Developments in Theory and Numerics for Conservation Laws. Ed. by D. Kroner, M. Ohlberger, and
M. Rohde. Springer, 1997, pp. 195285 (see pp. 32, 72).
[11] G. K. Batchelor. An Introduction to Fluid Dynamics. Cambridge University Press, 1967 (see p. 136).
[12] G. K. Batchelor. The Theory of Homogeneous Turbulence. Cambridge University Press, 1970 (see p. 173).
[13] E. R. Benton and G. W. Platzman. A Table of Solutions of the One-Dimensional Burgers Equation. Quart.
Appl. Mafh. 30 (1972), pp. 195212 (see p. 188).
[14] S. Blanes and P. C. Moan. Practical Symplectic Partitioned Runge-Kutta and Runge-Kutta-Nystrm Methods.
J. Comp. Appl. Math. 142 (2002), pp. 313330 (see p. 38).
[15] A. Bonglioli. Fluctuation Splitting Schemes for the Compressible and Incompressible Euler and Navier-Stokes
Equations. Int. J. Comput. Fluid Dyn. 14 (2000), pp. 2139 (see pp. 80, 100).
[16] J. Boussinesq. Essai Sur Lar Thorie Des Eaux Courantes. Mem. Prsents Acad. Sci. 23 (1877), p. 46 (see
p. 174).
[17] W. L. Briggs, V. E. Henson, and S. F. McCormick. A Multigrid Tutorial. Second. SIAM, 2000 (see p. 52).
[18] R. G. Campbell. Foundations of Fluid Flow Theory. Addison-Wesley Publishing Company, 1973 (see pp. 110,
167).
[19] D. Caraeni and L. Fuchs. Compact Third-Order Multidimensional Upwind Scheme for Navier-Stokes Simulations.
Appl. Math. 15 (2002), pp. 373401 (see p. 80).
[20] C. Cattaneo. A Form of Heat-Conduction Equations which Eliminates the Paradox of Instantaneous Propagation.
Ct. R. Acad. Sci., Paris 247 (1958), pp. 431433 (see p. 48).
279
280 BIBLIOGRAPHY
[21] A. J. Chorin. A Numerical Method for Solving Incompressible Viscous Flow Problems. J. Comput. Phys. 2
(1967), pp. 1226 (see pp. 36, 107, 133, 162).
[22] A. J. Chorin. Numerical Solution of the Navier-Stokes Equations. Mathematics of Computation 22.104 (1968),
pp. 745762 (see p. 263).
[23] C.-S. Chou and C.-W. Shu. High Order Residual Distribution Conservative Finite Dierence WENO Schemes
for Steady State Problems on Non-Smooth Meshes. J. Comput. Phys. 214 (2006), pp. 698724 (see p. 221).
[24] B. Cockburn and C.-W. Shu. Runge-Kutta Discontinuous Galerkin Methods for Convection-Dominated Prob-
lems. J. Sci.Comput. 16.3 (2001), pp. 173261 (see p. 51).
[25] B. Cockburn and C.-W. Shu. The Local Discontinuous Galerkin Method for Time-Dependent Convection-
Diusion Systems. SIAM J. Numer. Anal. 35 (1998), pp. 24402463 (see p. 48).
[26] W. J. Coirier and K. G. Powell. An Accuracy Assessment of Cartesian-Mesh Approaches for the Euler Equations.
J. Comput. Phys. 117 (1995), pp. 121131 (see p. 248).
[27] A. W. Cook and W. H. Cabot. Hyperviscosity for Shock-Turbulence Interactions. J. Comput. Phys. 203 (2005),
pp. 379385 (see p. 133).
[28] C. Corre, G. Hanss, and A. Lerat. A Residual-Based Compact Schemes for the Unsteady Compressible Navier-
Stokes Equations. Comput. Fluids 34 (2005), pp. 561580 (see p. 242).
[29] G.-H. Cottet and P. D. Koumoutsakos. Vortex Methods: Theory and Practice. Cambridge University Press, 2000
(see pp. 126, 164).
[30] A. Csik, M. Ricchiuto, and H. Deconinck. A Conservative Formulation of the Multidimensional Upwind Residual
Distribution Schemes for General Nonlinear Conservation Laws. J. Comput. Phys. 179 (2002), pp. 286312 (see
p. 80).
[31] H. Deconinck, P. L. Roe, and R. Struijs. A Multi-Dimensional Generalization of Roes Flux Dierence Splitter
for the Euler Equations. Comput. Fluids 22 (1993), pp. 215222 (see pp. 75, 89).
[32] B. Diskin and J. T. Thomas. Convergence of defect-correction and multigrid iterations for inviscid ows. In:
20th AIAA Computational Fluid Dynamics Conference. AIAA Paper 2011-3235. Hawaii, 2011 (see p. 224).
[33] Weinan E and C.-W Shu. A Numerical Resolution Study of High Order Essentially Non-oscillatory Schemes
Applied to Incompressible Flow. J. Comput. Phys. 110 (1994), pp. 3946 (see p. 263).
[34] L. Eca et al. On the construction of Manufactured Solutions for One and Two-Equation Eddy-Viscosity Models.
Int. J. Numer. Meth. Fluids. 54 (2007), pp. 119154 (see p. 198).
[35] C. R. Ethier and D. A. Steinman. Exact Fully 3D Navier-Stokes Solutions for Benchmarking. Int. J. Numer.
Meth. Fluids. 19 (1994), pp. 369375 (see p. 264).
[36] C. A. J. Fletcher. Computational Techniques for Fluid Dynamics. Vol. 2. Springer, 1991 (see p. 30).
[37] C. A. J. Fletcher. Generating Exact Solutions of the Two-Dimensional Burgers Equations. Int. J. Numer. Meth.
Fluids. 3 (1983), pp. 213216 (see pp. 188, 221).
[38] L. E. Fraenkel. On Corner Eddies in Plane Inviscid Shear Flow. J. Fluid Mech. 11 (1961), pp. 400406 (see
p. 235).
[39] E. Godlewski and P.-A. Raviart. Numerical Approximation of Hyperbolic Systems of Conservation Laws. Springer,
1996 (see p. 32).
[40] S. K. Godunov. An interesting Class of Quasilinear Systems. Dokl. Akad. Nauk. SSSR 139 (1961), pp. 521523
(see p. 72).
[41] A. R. Gourlay and J. LL. Morris. On the Comparison of Multistep Formulations of the Optimized Lax-Wendro
Method for Nonlinear Hyperbolic Systems in Two Space Variables. J. Comput. Phys. 5 (1970), pp. 229243
(see p. 38).
[42] H. Grad. Reducible Problems in Magneto-Fluid Dynamic Steady Flows. Rev. Mod. Phys. 32.4 (1960), pp. 830
847 (see p. 248).
[43] B. Grossman. The Computation of Inviscid Rotational Gasdynamic Flows Using an Alternate Velocity Decom-
position. AIAA Paper 83-1900. 1983 (see p. 121).
[44] B. Grossman and S. K. Choi. The Computation of Rotational Conical Flows. AIAA Paper 84-0258. 1984 (see
p. 121).
BIBLIOGRAPHY 281
[45] H. Guillard and C. Viozat. On the Behaviour of Upwind Schemes in the Low Mach Number Limit. J. Comput.
Phys. 28 (1999), pp. 6386 (see p. 36).
[46] Y. H. and C. L. Merkle. The Applicaiton of Preconditioning in Viscous Flows. J. Comput. Phys. 105 (1993),
pp. 207223 (see p. 36).
[47] M. E. Habbard and P. Garcia Navarro. Flux Dierence Splitting and the Balancing of Source Terms and Flux
Gradients. J. Comput. Phys. 165 (2000), pp. 89125 (see p. 198).
[48] N. D. Halsey. Potential Flow Analysis of Multielement Airfoils Using Conformal Mapping. AIAA J. 17 (1979),
pp. 12811288 (see p. 235).
[49] A. Haselbacher and J. Blazek. Accurate and Ecient Discretization of Navier-Stokes Equations on Mixed Grids.
AIAA J. 38.11 (2000), pp. 20942102 (see p. 145).
[50] C. Hirsch. Numerical Computation of Internal and External Flows. Vol. 1. A Wiley - Interscience Publications,
1990 (see p. 41).
[51] C. Hirsch. Numerical Computation of Internal and External Flows. Vol. 2. A Wiley - Interscience Publications,
1990 (see pp. 41, 67, 120, 122, 150).
[52] C. W. Hirt, A. A. Amsden, and J. L. Cook. An Arbitrary Lagrangian-Eulerian Computing Method for All Flow
Speeds. J. Comput. Phys. 14.3 (1974), pp. 227257 (see p. 28).
[53] W. F. Hughes and J. A. Brighton. Schaums Outline Series, Fluid Dynamics. McGraw-Hill, 1995 (see p. 122).
[54] I. Imai. Fluid Dynamics. Vol. 1. Shokabo, 1973 (see pp. 54, 168170, 183).
[55] R. I. Issa. Rise of Total Pressure in Frictional Flow. AIAA J. 33.4 (1995), pp. 772774 (see p. 135).
[56] D. C. Ives. A Modern Look at Conformal Mapping Including Multiple Connected Region. AIAA J. 14 (1976),
pp. 10061011 (see p. 235).
[57] Jr. J. D. Anderson. Computational Fluid Dynamics: The Basic with Applications. McGraw-Hill, 1995 (see pp. 30,
116).
[58] A. Jameson. Numerical Solution of Nonlinear Partial Dierential Equations of Mixed Type. In: Third Symposium
on Numerical Solution of Partial Dierential Equations. 1975 (see p. 55).
[59] A. Jameson. Time Dependent Calculations Using Multigrid, with Applications to Unsteady Flows Past Airfoils
and Wings. In: AIAA Paper 91-1596. 1991 (see p. 36).
[60] A. Jerey and T. Taniuti. Nonlinear Wave Propagation. Academic Press, 1964 (see p. 41).
[61] B.-N. Jiang. The Least-Squares Finite Element Method. Springer, 1998 (see p. 53).
[62] S. Jin and C. D. Levermore. Numerical Schemes for Hyperbolic Conservation Laws with Sti Relaxation Terms.
J. Comput. Phys. 126 (1996), pp. 449467 (see p. 55).
[63] D. A. Johnson and L. S. King. A Mathematically Simple Turbulence Closure Model for Attached and Separated
Turbulent Boundary Layers. AIAA J. 23.11 (1985), pp. 16841692 (see p. 177).
[64] I. W. Johnson, A. J. Wathen, and M. J. Baines. Moving Finite Element Methods for Evolutionary Problems II:
Applications. J. Comput. Phys. 79 (1988), pp. 270297 (see p. 219).
[65] Daniel D. Joseph. Potential ow of viscous uids: Historical notes. Int. J. Multiphas. Flow 32 (2006), pp. 285
310 (see p. 121).
[66] R. Kannan, Y. Sun, and Z. J. Wang. A Study of Viscous Flux Formulations for an Implicit P-Multigrid Spectral
Volume Navier Stokes Solver. In: 46th AIAA Aerospace Sciences Meeting. AIAA Paper 2008-783. 2008 (see
p. 48).
[67] K. Karamcheti. Principles of Ideal-Fluid Aerodynamics. Reprinted. Krieger Publishing Company, 1980 (see
pp. 19, 167).
[68] J. Katz and A. Plotkin. Low-Speed Aerodynamics. Second. Cambridge University Press, 2001 (see pp. 53, 184).
[69] S. Kawai, S. K. Shankar, and S. K. Lele. Assesment of Localized Articial Diusivity Scheme for Large-Eddy
Simulation of Compressible Turbulent Flows. J. Comput. Phys. 229 (2010), pp. 17391762 (see p. 133).
[70] Soshi Kawai. Private Communication. 2013 (see p. 133).
[71] C. Kim. Multi-Dimensional Upwind Leapfrog Schemes and Their Applications. PhD thesis. Ann Arbor, Michi-
gan: University of Michigan, 1997 (see p. 213).
282 BIBLIOGRAPHY
[72] B. L. Kleb. Optimizing Runge-Kutta Schemes for Viscous Flow. PhD thesis. Ann Arbor, Michigan: University
of Michigan, 2004 (see p. 150).
[73] A. N. Kolmogorov. The Local Structure of Turbulence in Incompressible Viscous Fluid for Very Large Reynolds
Numbers. Doklady Akademiya Nauk SSSR 30 (1941), pp. 299303 (see p. 173).
[74] A. M. Kuethe and C.-Y. Chow. Foundations of Aerodynamics. Fifth. John Wiley & Son Ltd., 1998 (see p. 123).
[75] D. Kwak et al. A Three-Dimensional Incompressible Navier-Stokes Solver Using Primitive Variables. AIAA J.
24 (1986), pp. 390396 (see pp. 107, 162).
[76] L. D. Landau and E. M. Lifshitz. Fluid Mechanics. Second. Butterworth-Heinemann, 1987 (see p. 61).
[77] Dohyung Lee. Local Preconditioning of the Euler and Navier-Stokes Equations. PhD thesis. Ann Arbor,
Michigan: University of Michigan, 1996 (see pp. 36, 150).
[78] S. Lee, S. K. Lele, and P. Moin. Direct Numerical Simulation of Isotropic Turbulence Interacting with a Weak
Shock Wave. J. Fluid Mech. 251 (1993), pp. 533562 (see p. 133).
[79] W.-T. Lee. Local Preconditioning of the Euler Equations. PhD thesis. Ann Arbor, Michigan: University of
Michigan, 1991 (see p. 39).
[80] B. van Leer, W.-T. Lee, and P. L. Roe. Characteristic Time-Stepping or Local Preconditioning of the Euler
Equations. In: 10th AIAA Computational Fluid Dynamics Conference. AIAA Paper 91-1552. Hawaii, 1991 (see
p. 36).
[81] R. J. LeVeque. Finite Volume Methods for Hyperbolic Problems. Cambridge University Press, 2002 (see pp. 42,
56, 57, 61, 127).
[82] R. L. LeVeque. Balancing source terms and ux gradients in high-resolution Godunov methods - The quasi-
steady wave-propagation algorithm. J. Comput. Phys. 146 (1998), pp. 346365 (see p. 198).
[83] W. Liao et al. Textbook-Eciency Multigrid Solver for Three-Dimensional Unsteady Compressible Navier-Stokes
Equations. J. Comput. Phys. 227 (2008), pp. 71607177 (see pp. 198, 277).
[84] H. W. Liepmann and A. Roshko. Elements of Gasdynamics. Dover, 2002 (see pp. 122, 123, 244, 252, 254, 272).
[85] K. Liman and A. Siora. Magnetosonic Jet Flow. Mon. Not. R. Astron. Soc. 290 (1997), pp. 629635 (see
p. 248).
[86] T. Linde and P. Roe. On a Mistaken Notion of of "Proper Upwinding". J. Comput. Phys. 142 (1998), pp. 611
614 (see p. 259).
[87] R. B. Lowrie. Compact Higher-Order Numerical Methods for Hyperbolic Conservation Laws. PhD thesis. Ann
Arbor, Michigan: University of Michigan, 1996 (see p. 240).
[88] R. B. Lowrie and J. E. Morel. Methods for Hyperbolic Systems with Sti Relaxation. Int. J. Numer. Meth.
Fluids. 40 (2002), pp. 413423 (see p. 55).
[89] D. J. Mavriplis and S. Pirzadeh. Large-Scale Parallel Unstructured Mesh Computations for 3D High-Lift Analysis.
J. Aircraft 36.6 (1999), pp. 987998 (see p. 47).
[90] K. Mazaheri and P. L. Roe. Bulk Viscosity Damping for Accelerating Convergence of low Mach number Euler
Solvers. Int. J. Numer. Meth. Fluids. 41 (2003), pp. 633652 (see p. 133).
[91] J. M. McDonough. A Class of Model Problems for Testing Compressible Navier-Stokes Solvers. AIAA Paper
883646. 1988 (see p. 277).
[92] F. R. Menter. Two-Equation Eddy-Viscosity Turbulence Models for Engineering Applications. AIAA J. 32.8
(1994), pp. 15981605 (see p. 177).
[93] L. M. Mesaros. Multi-Dimensional Fluctuation Splitting Schemes for The Euler Equations on Unstructured
Grids. PhD thesis. Ann Arbor, Michigan: University of Michigan, 1995 (see pp. 53, 54).
[94] L. M. Mesaros and P. L. Roe. Multidimensional Fluctuation Splitting Schemes Based on Decomposition Meth-
ods. In: 12th AIAA Computational Fluid Dynamics Conference. AIAA Paper 95-1699. San Diego, 1995 (see
pp. 36, 118).
[95] K. W. Morton and P. L. Roe. Vorticity-Preserving Lax-Wendro-Type Schemes for the System Wave Equation.
SIAM J. Sci. Comput. 23 (2001), pp. 170192 (see p. 46).
[96] E. M. Murman and P. M. Stremel. A Vortex Wake Capturing Method for Potential Flow Calculations. AIAA
Paper 82-0947. 1982 (see p. 121).
BIBLIOGRAPHY 283
[97] R. E. Neel, A. G. Godfrey, and W. D. McGrory. Low-Speed, Time-Accurate Validation of GASP Version 4. In:
43rd AIAA Aerospace Sciences Meeting and Exhibit. AIAA Paper 2005-686. Reno, Nevada, 2005 (see pp. 107,
162).
[98] H. Nishikawa. A First-Order System Approach for Diusion Equation. I: Second-Order Residual-Distribution
Schemes. J. Comput. Phys. 227 (2007), pp. 315352. url: http://www.hiroakinishikawa.com/My_paper
s/nishikawa_jcp2007v227pp315-352_preprint.pdf (see pp. 36, 49, 50).
[99] H. Nishikawa. A First-Order System Approach for Diusion Equation. II: Unication of Advection and Diusion.
J. Comput. Phys. 229 (2010), pp. 39894016. url: http://www.hiroakinishikawa.com/My_papers/nish
ikawa_jcp2010v227pp3989-4016_preprint.pdf (see pp. 33, 36, 50).
[100] H. Nishikawa. Adaptive Quadrature Fluctuation-Splitting Schemes for the Euler Equations. Int. J. Numer. Meth.
Fluids. 57 (2008), pp. 112. url: http://www.hiroakinishikawa.com/My_papers/nishikawa_ijnmf200
8v57pp1-12_preprint.pdf (see p. 118).
[101] H. Nishikawa. Beyond Interface Gradient: A General Principle for Constructing Diusion Schemes. In: 40th
AIAA Fluid Dynamics Conference and Exhibit. AIAA Paper 2010-5093. Chicago, 2010. url: http://www.hir
oakinishikawa.com/My_papers/nishikawa_AIAA-2010-5093.pdf (see pp. 47, 207).
[102] H. Nishikawa. Divergence Formulation of Source Term. J. Comput. Phys. 231 (2012), pp. 63936400. url:
http://www.hiroakinishikawa.com/My_papers/nishikawa_JCP2012_Preprint.pdf (see p. 29).
[103] H. Nishikawa. First, Second, and Third Order Finite-Volume Schemes for Advection-Diusion. In: 21st AIAA
Computational Fluid Dynamics Conference. AIAA Paper 2011-2568. San Diego, California, 2013. url: http:
//www.hiroakinishikawa.com/My_papers/nishikawa_AIAA2013-2568.pdf (see pp. 36, 50).
[104] H. Nishikawa. First, Second, and Third Order Finite-Volume Schemes for Diusion. In: 51st AIAA Aerospace
Sciences Meeting. AIAA Paper 2011-1125. Grapevine, Texas, 2013. url: http://www.hiroakinishikawa.c
om/My_papers/nishikawa_AIAA2013-1125.pdf (see pp. 36, 49, 50).
[105] H. Nishikawa. Multigrid Third-Order Least-Squares Solution of Cauchy-Riemann Equations on Unstructured
Triangular Grids. Int. J. Numer. Meth. Fluids. 53 (2007), pp. 443454. url: http://www.hiroakinishikaw
a.com/My_papers/nishikawa_ijnmf2007v53pp443-454_preprint.pdf (see p. 53).
[106] H. Nishikawa. New-Generation Hyperbolic Navier-Stokes Schemes: O(1/h) Speed-Up and Accurate Viscous/Heat
Fluxes. In: 20th AIAA Computational Fluid Dynamics Conference. AIAA Paper 2011-3043. Hawaii, 2011. url:
http://www.hiroakinishikawa.com/My_papers/nishikawa_AIAA-2011-3043.pdf (see pp. 7, 33, 277).
[107] H. Nishikawa. On Grids and Solutions from Residual Minimization. PhD thesis. Ann Arbor, Michigan: Uni-
versity of Michigan, Aug. 2001. url: http://www.hiroakinishikawa.com/My_papers/hiroaki_nishika
wa_thesis2001.pdf (see pp. 53, 208).
[108] H. Nishikawa. Robust and Accurate Viscous Discretization via Upwind Scheme - I: Basic Principle. Comput.
Fluids 49.1 (2011), pp. 6286. url: http://www.hiroakinishikawa.com/My_papers/nishikawa_cf201
1_preprint.pdf (see pp. 47, 207).
[109] H. Nishikawa. The Roe-Averaged density. http://www.cfdnotes.com/cfdnotes_roe_averaged_density.
html. March, 2011 (see p. 93).
[110] H. Nishikawa. Two Ways to Extend Diusion Schemes to Navier-Stokes Schemes: Gradients or Upwinding.
In: 20th AIAA Computational Fluid Dynamics Conference. AIAA Paper 2011-3044. Hawaii, 2011. url: http:
//www.hiroakinishikawa.com/My_papers/nishikawa_AIAA-2011-3044.pdf (see p. 277).
[111] H. Nishikawa, B. Diskin, and J. L. Thomas. Critical Study of Agglomerated Multigrid Methods for Diusion.
AIAA J. 48.4 (2010), pp. 839847. url: http://www.hiroakinishikawa.com/My_papers/nishikawa_dis
kin_thomas_aiaaj2010v48pp839-847.pdf (see pp. 47, 145).
[112] H. Nishikawa and K. Kitamura. Very Simple, Carbuncle-Free, Boundary-Layer Resolving, Rotated-Hybrid Rie-
mann Solvers. J. Comput. Phys. 227 (2007), pp. 25602581. url: http://www.hiroakinishikawa.com/
My_papers/nishikawa_kitamura_jcp2008v227pp2560-2581_preprint.pdf (see p. 4).
[113] H. Nishikawa, M. Rad, and P. Roe. A Third-Order Fluctuation-Splitting Scheme That Preserves Potential
Flow. In: 15th AIAA Computational Fluid Dynamics Conference. AIAA Paper 01-2595. Anaheim, 2001. url:
http://www.hiroakinishikawa.com/My_papers/nishikawa_rad_roe_aiaa-2001-2595-400.pdf (see
pp. 36, 118, 235).
284 BIBLIOGRAPHY
[114] H. Nishikawa, M. Rad, and P. L. Roe. Grids and Solutions from Residual Minimisation. In: Computational Fluid
Dynamics 2000. Ed. by N. Satofuka. Springer-Verlag, 2000, pp. 119124. url: http://www.hiroakinishik
awa.com/My_papers/iccfd1_2000_kyoto_01.pdf (see p. 218).
[115] H. Nishikawa and P. L. Roe. On High-Order Fluctuation-Splitting Schemes for Navier-Stokes Equations. In:
Computational Fluid Dynamics 2004. Ed. by C. Groth and D. W. Zingg. Springer-Verlag, 2004, pp. 799804.
url: http://www.hiroakinishikawa.com/My_papers/iccfd3_2004_nishikawa_roe.pdf (see pp. 48,
51).
[116] H. Nishikawa and B. van Leer. Optimal Multigrid Convergence by Elliptic/Hyperbolic Splitting. J. Comput. Phys.
190 (2003), pp. 5263. url: http://www.hiroakinishikawa.com/My_papers/nishikawa_van_leer_jcp
2003v190pp52-63_preprint.pdf (see p. 36).
[117] H. Nishikawa et al. A General Theory of Local Preconditioning and Its Application to the 2D Ideal MHD
Equations. In: 16th AIAA Computational Fluid Dynamics Conference. AIAA Paper 2003-3704. Orlando, 2003.
url: http://www.hiroakinishikawa.com/My_papers/nishikawa_roe_suzuki_vanleer_aiaa2003-370
4 (see pp. 33, 34, 36, 41, 53).
[118] T. A. Oliver. Multigrid solution for high-order discontinuous Galerkin discretizations of the compressible Navier-
Stokes equations. MA thesis. Massachusetts Institute of Technology, Aug. 2004 (see p. 277).
[119] K. Ou, P. Vincent, and A. Jameson. High-Order Methods for Diusion Equation with Energy Stable Flux
Reconstruction Scheme. In: 49th AIAA Aerospace Sciences Meeting. AIAA Paper 2011-46. Orlando, 2011 (see
p. 205).
[120] C. Pozrikidis. Introduction to Theoretical and Computational Fluid Dynamics. Oxford University Press, 1997
(see pp. 26, 167, 269, 270).
[121] L. Prandtl. Uber die ausgebildete Turbulenz. Z.A.M.M. 5 (1925), pp. 136139 (see p. 177).
[122] L. Quartapelle and M. Napolitano. Force and Moment in Incompressible Flows. AIAA J. 21.6 (1983), pp. 911
913 (see p. 136).
[123] M. Rad. Residual Distribution Approach to the Euler Equations that Preserves Potential Flow. PhD thesis.
Ann Arbor, Michigan: University of Michigan, Aug. 2001 (see pp. 53, 235).
[124] J. D. Ramshaw and V. A. Mousseau. Accelerated Articial Compressibility Method for Steady-State Incom-
pressible Flow Calculations. Comput. Fluids 18 (1990), pp. 361367 (see p. 133).
[125] M. Ricchiuto et al. On Uniformly High-Order Accurate Residual Distribution Schemes for Advection-Diusion.
J. Comput. Appl. Math. 215 (2008), pp. 547556 (see p. 51).
[126] F. Ringleb. Exakte Losungen der Dierentialgleichungen einer adiabatischen Gasstromung. Z.A.M.M. 20.4
(1940), pp. 18598 (see p. 248).
[127] P. J. Roache. Verication and Validation in Computational Science and Engineering. Hermosa, 1998 (see pp. 198,
201).
[128] P. L. Roe. Approximate Riemann Solvers, Parameter Vectors, and Dierence Schemes. J. Comput. Phys. 43
(1981), pp. 357372 (see pp. 6, 68, 74, 80, 88, 92, 93).
[129] P. L. Roe and M. Arora. Characteristic-Based Schemes for Dispersive Waves I. The Method of Characteristics
for Smooth Solutions. Numer. Meth. Part. D. E. 9 (1993), pp. 459505 (see p. 49).
[130] P. L. Roe and H. Nishikawa. Adaptive Grid Generation by Minimising Residuals. Int. J. Numer. Meth. Fluids.
40 (2002), pp. 121136 (see p. 244).
[131] S. E. Rogers and D. Kwak. Upwind Dierencing Scheme for the Time-Accurate Incompressible Navier-Stokes
Equations. AIAA J. 28.2 (1990), pp. 253262 (see p. 107).
[132] S. E. Rogers, D. Kwak, and C. Kiris. Steady and Unsteady Solutions of the Incompressible Navier-Stokes
Equations. AIAA J. 29.4 (1991), pp. 603610 (see pp. 107, 162).
[133] C. J. Roy et al. Verication of Euler/Navier-Stokes Codes Using the Method of Manufactured Solutions. Int.
J. Numer. Meth. Fluids. 44 (2004), pp. 599620 (see pp. 198, 277).
[134] J. Roy, M. Hafez, and J. J. Chattot. Explicit Methods for the Solution of the Generalized Cauchy-Riemann
Equations and Simulation of Inviscid Rotational Flows. Comput. Fluids 31 (2002), pp. 769786 (see pp. 54,
121).
[135] C. L. Rumsey and J. L. Thomas. Application of FUN3D and CFL3D to the Third Workshop on CFD Uncertainty
Analysis. NASA TM-2008-215537. 2008 (see p. 198).
BIBLIOGRAPHY 285
[136] H. Schlichting and K. Gersten. Boundary Layer Theory. Eighth. Springer, 2000 (see p. 132).
[137] A. Sescu, A. A. Afjeh, and C. Sescu. Numerical Solution to Nonlinear Tricomi Equation Using WENO Schemes.
In: Eighth Mississippi State - UAB Conference on Dierential Equations and Computational Simulations. 2010,
pp. 235244 (see p. 55).
[138] A. H. Shapiro. The Dynamics and Thermodynamics of Compressible Fluid Flow. Vol. 1. The Ronald Press
Company, 1953 (see p. 191).
[139] A. H. Shapiro. The Dynamics and Thermodynamics of Compressible Fluid Flow. Vol. 2. The Ronald Press
Company, 1954 (see p. 249).
[140] A. Sherif and M. Hafez. Computation of Three-Dimensional Transonic Flows Using Two Stream Functions. Int.
J. Numer. Meth. Fluids. 8 (1988), pp. 1729 (see p. 167).
[141] T. W. H. Sheu, S. K. Wang, and R. K. Lin. An Implicit Scheme for Solving the Convection-Diusion-Reaction
Equation in Two Dimensions. J. Comput. Phys. 164 (2000), pp. 123142 (see p. 213).
[142] C.-W. Shu. Essentially Non-Oscillatory and Weighted Essentially Non-Oscillatory Schemes for Hyperbolic Con-
servation Laws. In: Lecture Notes in Mathematics. Ed. by A. Quarteroni. Vol. 1697. Springer, 1998, pp. 325
432 (see p. 242).
[143] A. M. O. Smith and T. Cebeci. Numerical Solution of the Turbulent Boundary-Layer Equations. Douglas Aircraft
Division Report DAC 33735, 1967 (see p. 177).
[144] G. A. Sod. A Survey of Several Finite Dierence Methods for Systems of Nonlinear Hyperbolic Conservation
Laws. J. Comput. Phys. 27 (1978), pp. 131 (see p. 258).
[145] P. Solin and K. Segeth. Description of Multi-Dimensional Finite Volume Solver EULER. Appl. Math. 47.2
(2002), pp. 169185315 (see p. 97).
[146] P. R. Spalart and S. R. Allmaras. A One-Equation Turbulence Model for Aerodynamic Flows. AIAA Paper
92-0439. 1992 (see p. 177).
[147] P. R. Spalart et al. Comments on the Feasibility of LES for Wings and on the Hybrid RANS/LES Approach.
Advances in DNS/LES, Proceedings of the First AFOSR International Conference on DNS/LES. 1997 (see
p. 174).
[148] Michael Spivak. Calculus on Manifolds: A Modern Approach To Classical Theorems Of Advanced Calculus.
Addison-Wesley Publishing Company, 1965 (see p. 27).
[149] J. L. Steger and R. F. Warming. Flux Vector Splitting of the Inviscid Gas-Dynamic Equations with Applications
to Finite Dierence Methods. J. Comput. Phys. 40 (1981), pp. 263293 (see p. 101).
[150] G. Strang. On the Construction and Comparison of Dierent Splitting Schemes. SIAM J. Numer. Anal. 5 (1968),
pp. 506517 (see p. 38).
[151] M. Strelets. Detached Eddy Simulation of Massively Separated Flows. In: 2001 (see p. 174).
[152] A. Suddhoo and I. M. Hall. Test Case for the Plane Potential Flow Past Multi-Element Aerofoils. Aeronaut. J.
(Dec. 1985), pp. 403414 (see p. 235).
[153] H. Sun, D. L. Darmofal, and R. Haimes. Flux Vector Splitting of the Inviscid Gas-Dynamic Equations with
Applications to Finite Dierence Methods. J. Comput. Phys. 231 (2012), pp. 541557 (see p. 211).
[154] Y. Sun, Z. J. Wang, and Y. Liu. Spectral (nite) Volume Method for Conservation Laws on Unstructured Grids
VI: Extension to Viscous Flow. J. Comput. Phys. 215 (2006), pp. 4158 (see p. 219).
[155] J. C. Tannehill, D. A. Anderson, and R. H. Pletcher. Computational Fluid Mechanics and Heat Transfer. Second.
Taylor & Francis, 1997 (see pp. 30, 35, 120, 122, 150, 163, 211, 219, 244).
[156] H. Tennekes and J. L. Lumley. A First Course in Turbulence. The MIT Press, 1972 (see p. 173).
[157] J. L. Thomas, B. Diskin, and H. Nishikawa. A Critical Study of Agglomerated Multigrid Methods for Diusion
on Highly-Stretched Grids. Comput. Fluids 41.1 (2011), pp. 8293 (see p. 145).
[158] G. T. Tomaich. A Genuinely Multi-Dimensional Upwinding Algorithm for for the Navier-Stokes Equations on
Unstructured Grids Using A Compact, Highly-Parallelizable Spatial Discretization. PhD thesis. Ann Arbor,
Michigan: University of Michigan, 1995 (see p. 170).
[159] E. F. Toro. Riemann Solvers and Numerical Methods for Fluid Dynamics: A Practical Introduction. Second.
Springer, 1999 (see pp. 42, 116, 134, 255).
286 BIBLIOGRAPHY
[160] U. Trottenberg, C. W. Oosterlee, and A. Schller. Multigrid. Academic Press, 2000 (see p. 55).
[161] E. Turkel. Preconditioning Methods for Solving the Incompressible and Low-Speed Compressible Equations. J.
Comput. Phys. 72 (1987), pp. 277298 (see p. 36).
[162] E. Turkel. Preconditioning Techniques in Computational Fluid Dynamics. Annu. Rev. Fluid Mech. 31 (1999),
pp. 385416 (see p. 224).
[163] E. Turkel. Review of Preconditioning Methods for Fluid Dynamics. ICASE Report 92-47 (1992) (see p. 36).
[164] E. van der Weide et al. A Parallel, Implicit, Multi-Dimensional Upwind, Residual Distribution Method for the
Navier-Stokes Equations on Unstructured Grids. Comput. Mech. 23 (1999), pp. 199208 (see p. 80).
[165] B. van Leer. Computational Fluid Dynamics: Science or Toolbox? In: 15th AIAA Computational Fluid Dynamics
Conference. AIAA Paper 2001-2520. Anaheim, 2001 (see p. 56).
[166] B. van Leer. Upwind and High-Resolution Methods for Compressible Flow: From Doner Cell to Residual-
Distribution Schemes. Commun. Comput. Phys. 1.2 (2006), pp. 192206 (see p. 65).
[167] S. Venkateswaran and C. L. Merkle. Analysis of time-derivative preconditioning for the Navier-Stokes equations.
In: Fifth International Symposium on Computational Fluid Dynamics. 1995 (see p. 150).
[168] J. M. Weiss and W. A. Smith. Preconditioning Applied to Variable and Constant Density Flows. AIAA J. 33.11
(1995), pp. 20502057 (see p. 36).
[169] F. M. White. Viscous Fluid Flow. Second. McGraw-Hill, 1991 (see pp. 132, 134, 270).
[170] G. Whitham. Linear and Nonlinear Waves. Wiley-Interscience, 1974 (see p. 57).
[171] D. C. Wilcox. Turbulence Modeling for CFD. DCW Industries, Inc., 1998 (see pp. 173, 177).
[172] B. R. Williams. An Exact Test Case for the Plane Potential Flow about Two Adjacent Lifting Aerofoils. Royal
Aeronautical Establishment, R&M No. 3713 (1971) (see p. 235).
[173] W. A. Wood and W. L. Kleb. 2-D/Axisymmetric Formulation of Multi-Dimensional Upwind Scheme. In: 15th
AIAA Computational Fluid Dynamics Conference. AIAA Paper 2001-2630. Anaheim, 2001 (see pp. 97, 98, 142).
[174] Y. Xing and C.-W. Shu. High Order Well-Balanced Finite Volume WENO schemes and Discontinuous Galerkin
Methods for a Class of Hyperbolic Systems with Source Terms. J. Comput. Phys. 214 (2006), pp. 567598 (see
p. 198).
[175] K. Xu. A Gas-Kinetic BGK Scheme for the Navier-Stokes Equations and Its Connection with Articial Dissipation
and Godunov Method. J. Comput. Phys. 171 (2001), pp. 289335 (see p. 277).
[176] Yan Xu and Chi-Wang Shu. Local Discontinuous Galerkin Methods for High-Order Time-Dependent Partial
Dierential Equations. Commun. Comput. Phys. 7.1 (2010), pp. 146 (see p. 48).
[177] H. C. Yee, M. Vinokur, and M. J. Djomehri. Entropy Splitting and Numerical Dissipation. J. Comput. Phys.
162 (2000), pp. 3381 (see p. 242).
[178] H. Yoshida. Construction of Higher Order Symplectic Integrators. Phys. Lett. A 150 (1990), pp. 262268 (see
p. 38).
[179] S. T. Yu. Convenient Method to Convert Two-Dimensional CFD Codes into Axisymmetric Ones. J. Propul.
Power 9.3 (1993), pp. 493495 (see p. 97).
[180] E. C. Zachmanoglou and D. W. Thoe. Introduction to Partial Dierential Equations with Applications. Dover,
1986 (see p. 35).
[181] X. Zhang and C.-W. Shu. Positivity-Preserving High Order Finite Dierence WENO Schemes for Compressible
Euler Equations. J. Comput. Phys. 231 (2012), pp. 22452258 (see p. 28).
[182] O. C. Zienkiewicz and R. L. Taylor. The Finite Element Method. Volume 1. McGraw-Hill Company, 1994 (see
p. 48).
Extra Contents
[ Available in the printed version. ]
287