Lecture Notes in Mathematics 1985

Editors:
J.-M. Morel, Cachan
F. Takens, Groningen
B. Teissier, Paris
Torsten Linß

Layer-Adapted Meshes
for Reaction-Convection-Diffusion
Problems

123
Torsten Linß
Technical University of Dresden
Institute for Numerical Mathematics
Zellescher Weg 12-14
01062 Dresden
Germany
[email protected]

ISBN: 978-3-642-05133-3 e-ISBN: 978-3-642-05134-0

DOI: 10.1007/978-3-642-05134-0
Springer Heidelberg Dordrecht London New York

Lecture Notes in Mathematics ISSN print edition: 0075-8434

ISSN electronic edition: 1617-9692

Library of Congress Control Number: 2009940926

Mathematics Subject Classification (2000): 65L11, 65L12, 65L20, 65L50, 65L60, 65N06, 65N08,
65N12, 65N30, 65N50

c Springer-Verlag Berlin Heidelberg 2010

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current version, and permission for use must always be obtained from Springer. Violations
are liable to prosecution under the German Copyright Law.

The use of general descriptive names, registered names, trademarks, etc. in this publication does not
imply, even in the absence of a specific statement, that such names are exempt from the relevant protective
laws and regulations and therefore free for general use.

Cover design: SPi Publisher Services

Printed on acid-free paper

springer.com
Preface

This is a book on numerical methods for singular perturbation problems – in partic-

ular, stationary reaction-convection-diffusion problems exhibiting layer behaviour.
More precisely, it is devoted to the construction and analysis of layer-adapted
meshes underlying these numerical methods.
Numerical methods for singularly perturbed differential equations have been
studied since the early 1970s and the research frontier has been constantly ex-
panding since. A comprehensive exposition of the state of the art in the analysis
of numerical methods for singular perturbation problems is [141] which was pub-
lished in 2008. As that monograph covers a big variety of numerical methods, it
only contains a rather short introduction to layer-adapted meshes, while the present
book is exclusively dedicated to that subject.
An early important contribution towards the optimisation of numerical methods
by means of special meshes was made by N.S. Bakhvalov [18] in 1969. His paper
spawned a lively discussion in the literature with a number of further meshes be-
ing proposed and applied to various singular perturbation problems. However, in
the mid 1980s, this development stalled, but was enlivened again by G.I. Shishkin’s
proposal of piecewise-equidistant meshes in the early 1990s [121, 150]. Because of
their very simple structure, they are often much easier to analyse than other meshes,
although they give numerical approximations that are inferior to solutions on com-
peting meshes. Shishkin meshes for numerous problems and numerical methods
have been studied since and they are still very much in vogue.
With this contribution we try to counter this development and lay the emphasis
on more general meshes that – apart from performing better than piecewise-uniform
meshes – provide a deeper insight in the course of their analysis.
In this monograph, a classification and a survey are given of layer-adapted
meshes for reaction-convection-diffusion problems. The monograph aims at giving
a structured and comprehensive account of current ideas in the numerical analysis
for various methods on layer-adapted meshes. Both finite differences, finite elements
and finite volumes will be covered.
While for finite difference schemes applied to one-dimensional problems, a
rather complete convergence theory for arbitrary meshes is developed, the theory
is more fragmentary for other methods and problems. They still require the restric-
tion to certain classes of meshes.

v
vi Preface

The roots of this monograph are a survey lecture presented at the Oberwolfach
seminar Numerical Methods for Singular Perturbation Problems, 8–14 April 2001
organised by Pieter W. Hemker, Hans-Görg Roos and Martin Stynes, and a review
article [91] invited by Thomas J.R. Hughes. I am indebted to their invitations and
their continued encouragement.
My thanks also go to a series of colleagues I had the pleasure of working with
over the years and who consequently influenced this monograph: Sebastian Franz,
Anja Fröhner, R. Bruce Kellogg, Natalia Kopteva, Niall Madden, Hans-Görg Roos,
Martin Stynes and Relja Vulanović.
The finishing work on this monograph was supported by the Science Foundation
Ireland during a visit to the University of Limerick and by the Czech Academy of
Science through a visiting scholarship.

Dresden and Prague Torsten Linß

July 2009
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Layer-Adapted Meshes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1 Convection-Diffusion Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.1 Bakhvalov Meshes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.2 Shishkin Meshes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1.3 Shishkin-Type Meshes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.4 Turning-Point Boundary Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.1.5 Interior Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.1.6 Overlapping Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 Reaction-Convection-Diffusion Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2.1 Interior Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.2 Overlapping Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3 Two-Dimensional Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.1 Reaction-Diffusion Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3.2 Convection-Diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

Part I One Dimensional Problems

3 The Analytical Behaviour of Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

3.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.1.1 Stability of Differential Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.1.2 Green’s Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.1.3 M -Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2 Reaction-Convection-Diffusion Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.2.1 Stability and Green’s Function Estimates . . . . . . . . . . . . . . . . . . . . . . 39
3.2.2 Derivative Bounds and Solution Decomposition . . . . . . . . . . . . . . 45
3.3 Reaction-Diffusion Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.3.1 Scalar Reaction-Diffusion Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.3.2 Systems of Reaction-Diffusion Equations . . . . . . . . . . . . . . . . . . . . . 52
3.4 Convection-Diffusion Problems with Regular Layers . . . . . . . . . . . . . . . . . 57
3.4.1 Scalar Convection-Diffusion Problems . . . . . . . . . . . . . . . . . . . . . . . . 57
3.4.2 Weakly Coupled Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.4.3 Strongly Coupled Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

vii
viii Contents

3.5 Convection-Diffusion Problems with Turning-Point Layers . . . . . . . . . . 69

3.5.1 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.5.2 Derivative Bounds and Solution Decomposition . . . . . . . . . . . . . . 71

4 Finite Difference Schemes for Convection-Diffusion Problems . . . . . . . . . 77

4.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.2 A Simple Upwind Difference Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.2.1 Stability of the Discrete Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.2.2 A Priori Error Bounds. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.2.3 Error Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.2.4 A Posteriori Error Estimation and Adaptivity . . . . . . . . . . . . . . . . . 92
4.2.5 An Alternative Convergence Proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.2.6 The Truncation Error and Barrier Function Technique . . . . . . . . 100
4.2.7 Discontinuous Coefficients and Point Sources. . . . . . . . . . . . . . . . . 103
4.2.8 Quasilinear Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.2.9 Derivative Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.3 Second-Order Difference Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.3.1 Second-Order Upwind Schemes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.3.2 Central Differencing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
4.3.3 Convergence Acceleration Techniques . . . . . . . . . . . . . . . . . . . . . . . . . 121
4.3.4 A Numerical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
4.4 Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
4.4.1 Weakly Coupled Systems in One Dimension . . . . . . . . . . . . . . . . . . 134
4.4.2 Strongly Coupled Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
4.5 Problems with Turning Point Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
4.5.1 A First-Order Upwind Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
4.5.2 Convergence on Shishkin Meshes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
4.5.3 A Numerical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

5 Finite Element and Finite Volume Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

5.1 The Interpolation Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
5.2 Linear Galerkin FEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
5.2.1 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
5.2.2 Supercloseness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
5.2.3 Gradient Recovery and a Posteriori Error Estimation . . . . . . . . . 160
5.2.4 A Numerical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
5.3 Stabilised FEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
5.3.1 Artificial Viscosity Stabilisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
5.3.2 Streamline-Diffusion Stabilisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
Contents ix

5.4 An Upwind Finite Volume Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

5.4.1 Stability of the FVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
5.4.2 Convergence in the Energy Norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
5.4.3 Convergence in the Maximum Norm. . . . . . . . . . . . . . . . . . . . . . . . . . . 180
5.4.4 A Numerical Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182

6 Discretisations of Reaction-Convection-Diffusion Problems. . . . . . . . . . . . . 183

6.1 Reaction-Diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
6.1.1 Linear Finite Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
6.1.2 Central Differencing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
6.1.3 A Non-Monotone Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
6.1.4 A Compact Fourth-Order Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
6.2 Systems of Reaction-Diffusion Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
6.2.1 The Interpolation Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
6.2.2 Linear Finite Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
6.2.3 Central Differencing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
6.3 Reaction-Convection-Diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
6.3.1 The Interpolation Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
6.3.2 Simple Upwinding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223

Part II Two Dimensional Problems

7 The Analytical Behaviour of Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235

7.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
7.1.1 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
7.1.2 Regularity of Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
7.2 Reaction-Diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
7.2.1 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
7.2.2 Derivative Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
7.3 Convection-Diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
7.3.1 Regular Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
7.3.2 Characteristic Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245

8 Reaction-Diffusion Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247

8.1 Central Differencing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
8.1.1 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
8.1.2 Convergence on Layer-Adapted Meshes . . . . . . . . . . . . . . . . . . . . . . . 249
8.1.3 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
8.2 Arbitrary Bounded Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254

9 Convection-Diffusion Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257

9.1 Upwind Difference Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
9.1.1 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
9.1.2 Pointwise Error Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
9.1.3 Error Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
x Contents

9.2 Finite Element Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263

9.2.1 The Interpolation Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
9.2.2 Galerkin FEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
9.2.3 Artificial Viscosity Stabilisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
9.2.4 Streamline-Diffusion FEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
9.2.5 Characteristic Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
9.3 Finite Volume Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
9.3.1 Coercivity of the Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
9.3.2 Inverse Monotonicity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
9.3.3 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306

Conclusions and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
Notation

u solution of boundary value problems

uN numerical approximations of u
ε, εd , εc perturbation parameter(s)
L, L∗ differential operator, its adjoint
L, L∗ discrete operator (discretisation of L), its adjoint
G, G continuous and discrete Green’s functions
Ω domain
∂Ω = Γ boundary of Ω
n outward pointing unit vector normal to ∂Ω
N number of mesh intervals (in each coordinate direction)
C generic constant, independent of ε and N
ω̄, ω̄x × ω̄y sets of mesh points
h, hi , h̄i , k, kj , k̄ j mesh step sizes
vx , vx̄ , vx̆ , vx̂ , vx̃ , vx̌ difference operators
uI nodal interpolant of u
 · ∞ supremum norm
 · 1 L1 norm
 · −1,∞ W −1,∞ norm
||| · |||ε,∞ ε-weighted W 1,∞ norm
 · ∗,ω discrete version of the norm  · ∗
Vω finite element space on the mesh ω
a(·, ·) bilinear form
(·, ·),  · 0 scalar product and norm in L2 (Ω)
| · |1 , ||| · |||ε semi norm and ε-weighted energy norm in H 1 (Ω)
||| · |||SD , ||| · |||κ , ||| · |||ρ various method-dependent energy norms
(·, ·)D , | · |∗,D ,  · ∗,D scalar product and (semi) norm restricted to D ⊂ Ω
C l , C l,α (Hölder) function spaces

xi
Chapter 1
Introduction

Stationary linear reaction-convection-diffusion problems form the subject of this

monograph:

−εu − bu + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1

and its two-dimensional analogue

−εΔu − b · ∇u + cu = f in Ω ⊂ IR2 , u|∂Ω = g

with a small positive parameter ε.

Such problems arise in various models of fluid flow [52,53,73]; they appear in the
(linearised) Navier-Stokes and in the Oseen equations, in the equations modelling
oil extraction from underground reservoirs [32], flows in chemical reactors [3] and
convective heat transport with large Péclet number [56]. Other applications include
the simulation of semiconductor devices [130].

An Example

Consider the boundary-value problem of finding u ∈ C 2 (0, 1) ∩ C[0, 1] such that

−εu (x) − u (x) = 1 for x ∈ (0, 1), u(0) = u(1) = 0 (1.1)

with 0 < ε  1. Formally setting ε = 0, yields

−u (x) = 1 for x ∈ (0, 1), u(0) = u(1) = 0.

Unlike (1.1), this problem does not possess a solution in C 2 (0, 1) ∩ C[0, 1]. Conse-
quently, when ε approaches zero, the solution of (1.1) is badly behaved in some way.
The solution of (1.1) is

e−1/ε − e−x/ε
u(x, ε) = + 1 − x.
1 − e−1/ε

T. Linß, Layer-Adapted Meshes for Reaction-Convection-Diffusion Problems, Lecture 1

Notes in Mathematics 1985, DOI 10.1007/978-3-642-05134-0 1,
c Springer-Verlag Berlin Heidelberg 2010
2 1 Introduction

Due to the presence of the exponential e−x/ε , the solution u and its derivatives
change rapidly near x = 0 for small values of ε. Regions where this happens are
referred to as layers. Singularly perturbed problems are typically characterised by
the presence of such layers. The term boundary layer was introduced by Lud-
wig Prandtl at the Third International Congress of Mathematicians in Heidelberg
in 1904.
The solution of (1.1) may be regarded as a function of two variables:

u : [0, 1] × (0, 1] : (x, ε) → u(x, ε).

Taking limits of u for (x, ε) → (0, 0), we see that

lim lim u(x, ε) = 1 = 0 = lim lim . (1.2)

x→0 ε→0 ε→0 x→0

Thus, u as a function of two variables possesses a classical singularity at the point

(0, 0) in the (x, ε)-plane. For this reason we may call (1.1) a singularly perturbed
boundary-value problem.

What Is a Singularly Perturbed Problem?

Miller et al. [121] give the following characterisation:

The justification for the name ‘singular perturbation’ is that the nature of the
differential equations changes completely in the limit case, when the singular
perturbation parameter is equal to zero. For example, . . . equations change
from being nonlinear parabolic equations to nonlinear hyperbolic equations.
This describes a phenomenon that can lead to the formation of boundary layers and
typically will—if appropriate boundary conditions are imposed. Roos et al. [141]
describe singularly perturbed problems as follows.
They are differential equations (ordinary or partial) that depend on a small
positive parameter ε and whose solutions (or their derivatives) approach a
discontinuous limit as ε approaches zero. Such problems are said to be singu-
larly perturbed, where we regard ε as a perturbation parameter.
Both sources avoid a formal definition:
In the present monograph we propose the following definition.
Definition 1.1. Let B be a function space with norm  · B . Let D ⊂ IRd be a
parameter domain. The continuous function u : D → B, ε → u(ε) is said to be
regular for ε → ε∗ ∈ ∂D if there exists a function u∗ ∈ B such that:

lim uε − u∗ B = 0,
ε→ε∗

otherwise uε is said to be singular for ε → ε∗ .

1 Introduction 3

Let (Pε ) be a problem with solution u(ε) ∈ B for all ε ∈ D. We say (Pε ) is
singularly perturbed for ε → ε∗ ∈ ∂D in the norm  · B if u is singular for
ε → ε∗ . ♥

Remark 1.2. The definition is norm dependent. For example (1.1), is singularly
perturbed in the C 0 norm and the L∞ norm because of (1.2). However, it is not
singularly perturbed in the L2 norm. There exists a function u∗ : x → 1 − x with
 
uε − u∗ 0 = O ε1/2 .

The L2 norm fails to capture the boundary layer in u. ♣

Remark 1.3. Boundary conditions play an important role. Consider the boundary-
value problem

−εu (x) − u (x) = 1 for x ∈ (0, 1), u (0) = u(1) = 0.

This problem is singularly perturbed in the C 1 norm, but it is not perturbed in the
C 0 norm. The Neumann boundary condition at x = 0 leads to the formation of a
weak layer only. The first-order derivative remains bounded when ε → 0. ♣

Uniform Convergence

Classical convergence results for numerical methods for boundary-value problems

have the structure
 
u − uh  ≤ Khk ,

with the maximum mesh size h. The constant K depends on certain derivatives of u
and typically tends to infinity as the perturbation parameter ε approaches zero. This
means that the maximal step size h has to be chosen proportional to some positive
power of ε which is impractical. Therefore, we are looking for so-called uniform
or robust methods where the numerical costs are independent of the perturbation
parameter ε. More precisely, we are looking for robust methods in the sense of the
following definition:
Definition 1.4. Let uε be the solution of a singularly perturbed problem, and let
uN
ε be a numerical approximation of uε obtained by a numerical method with N
degrees of freedom. The numerical method is said to be uniformly convergent or
robust with respect to the perturbation parameter ε in the norm  ·  if
 
uε − uN
ε
 ≤ ϑ(N ) for N ≥ N0
4 1 Introduction

with a function ϑ satisfying

lim ϑ(N ) = 0 and ∂ε ϑ ≡ 0,

N →∞

and with some threshold value N0 > 0 that is independent of ε. ♥

Scope of the Monograph

Well-developed techniques are available for the computation of solutions outside

layers [123, 141], but the problem of resolving layers—which is of great practical
importance—is still under investigation. This field has witnessed a stormy devel-
opment. Layer-adapted meshes have first been proposed by Bakhvalov [18] in the
context of reaction-diffusion problems. In the late 1970s and early 1980s, special
meshes for convection-diffusion problems were investigated by Gartland [45], Li-
seikin [113, 114, 116], Vulanović [163–166] and others in order to achieve uniform
convergence. The discussion has been livened up by the introduction of special
piecewise-uniform meshes by Shishkin [150]. They will be described in more de-
tail in Section 2.1.3. Because of their simple structure, they have attracted much
attention and are now widely referred to as Shishkin meshes. A small survey of
these meshes can be found in the monograph [141], while [109, 121] and [134] are
devoted exclusively to them.
The performance of Shishkin meshes is however inferior to that of Bakhvalov
meshes, which has prompted efforts to improve them while retaining some of their
simplicity, in particular, the mesh uniformity outside the layers and the choice of
mesh transition point where the mesh changes from fine to coarse. For instance,
Vulanović [169] uses a piecewise-uniform mesh with more than one transition point.
Linß [81, 82] combines the ideas of Bakhvalov and Shishkin, while Beckett and
Mackenzie [20] combine an equidistribution idea [31] with a Shishkin-type transi-
tion point. With all these various mesh-construction ideas a natural question is:

Can a general theory be derived that allows one to immediately deduce the
robust convergence of standard schemes on special meshes and a guaranteed
rate of convergence?

A first attempt towards this can be found in [137], where a first-order upwind scheme
and a Galerkin FEM are studied on a class of so-called Shishkin-type meshes. A
more general criterion was derived in [84, 85] for an upwind-difference scheme in
one dimension.
The main purpose of this monograph is to give a survey of recent developments
and present the state of the art in the analysis of layer-adapted meshes for a wide
range of reaction-convection-diffusion problems.
Chapter 2
Layer-Adapted Meshes

Before surveying a few of the most important ideas from the literature for construct-
ing layer-adapted meshes, we shall introduce some basic concepts for describing
layer-adapted meshes.
Throughout ω̄ : 0 = x0 < x1 < · · · < xN = 1 denotes a generic mesh with N
subintervals on [0, 1], while ω is the set of inner mesh nodes. Set Ii := [xi−1 , xi ].
The local mesh sizes are hi := xi − xi−1 , i = 1, . . . , N , while the maximum step
size is h := max hi .
i=1,...,N

Definition 2.1. A strictly monotone function ϕ : [0, 1] → [0, 1] that maps a uniform
mesh ti = i/N , i = 0, . . . , N , onto a layer-adapted mesh by xi = ϕ(ti ), i =
0, . . . , N , is called a mesh generating function. ♥
A related approach is that of stretching functions or layer-damping transforma-
tion [49,114,115], which are used to transform a problem with layers into a problem
whose derivatives are bounded.
For a given mesh generating function ϕ ∈ W 1,1 (0, 1), the local mesh step sizes
can be computed using the formula
 ti
hi = ϕ(ti ) − ϕ(ti−1 ) = ϕ (t)dt. (2.1)
ti−1

Another important concept is that of mesh equidistribution.

Definition 2.2 (Equidistribution principle). Let M : [0, 1] → IR be a positive
function a.e. A mesh ω̄ is said to equidistribute the monitor function M if
  1
1
M (t)dt = M (t)dt for i = 1, . . . , N.
Ii N 0

♥
Given a monitor function M the associated mesh generating function is implicitly
defined by
 ϕ(t)  1
M (s)ds = ξ M (s)ds for t ∈ [0, 1]
0 0

T. Linß, Layer-Adapted Meshes for Reaction-Convection-Diffusion Problems, Lecture 5

Notes in Mathematics 1985, DOI 10.1007/978-3-642-05134-0 2,
c Springer-Verlag Berlin Heidelberg 2010
6 2 Layer-Adapted Meshes

and its derivative by

 1
 1
ϕ (t) = M (s)ds for t ∈ [0, 1].
M (ϕ(t)) 0

2.1 Convection-Diffusion Problems

Consider the boundary-value problem

−εu − bu + cu = f in (0, 1), u(0) = u(1) = 0, (2.2)

where ε is a small positive parameter, b ≥ β > 0 on [0, 1]. The boundary value
problem (2.2) has a unique solution that typically has an exponential boundary layer
at x = 0 which behaves like e−βx/ε . Figure 2.1 gives a plot of a typical solution
A quantity that will appear frequently in the error estimates later and which char-
acterises the convergence is
  
1 + ε−1 e−βs/pε ds.
[p]
ϑcd (ω̄) := max (2.3)
i=1,...,N Ii

For example, in Section 4.2 we shall establish for the maximum-norm error of a
first-order upwind difference scheme that

 
u − uN  ≤ Cϑ[1] (ω̄) (2.4)
∞ cd

0.5

0.4

0.3

0.2

0.1

Fig. 2.1 Typical solution 0

of (2.2) 0.2 0.4 0.6 0.8 1
2.1 Convection-Diffusion Problems 7

on an arbitrary mesh ω̄. Noting that

 1  
1 + ε−1 e−βs/pε ds ≤ C,
0

we see that an optimal mesh—optimal with respect to the order of convergence—

equidistributes the monitor function M : s → 1 + ε−1 e−βs/pε .

2.1.1 Bakhvalov Meshes

Bakhvalov’s idea [18] is to use an equidistant t-grid near x = 0, then to map this
grid back onto the x-axis by means of the (scaled) boundary layer function. That is,
grid points xi near x = 0 are defined by
  i
q 1 − e−βxi /σε = ti = for i = 0, 1, . . . ,
N

where the scaling parameters q ∈ (0, 1) and σ > 0 are user chosen: q is roughly the
portion of mesh points used to resolve the layer, while σ determines the grading of
the mesh inside the layer. Away from the layer, a uniform mesh in x is used with the
transition point τ such that, the resulting mesh generating function is C 1 [0, 1], i. e.,
⎧
⎨χ(t) := − σε ln q − t for t ∈ [0, τ ],
ϕ(t) = β q
⎩
π(t) := χ(τ ) + χ (τ )(t − τ ) otherwise,

where the point τ satisfies

1 − χ(τ )
χ (τ ) = . (2.5)
1−τ

Geometrically this means that (τ, χ(τ )) is the contact point of the tangent π to χ
that passes through the point (1, 1); see Fig. 2.2. When σε ≥ ρq, the equation (2.5)
does not possess a solution. In this case the Bakhvalov mesh is uniform with mesh
size N −1 .
The nonlinear equation (2.5) cannot be solved explicitly. However, the iteration

1 − χ(τi )
τ0 = 0, χ (τi+1 ) = , i = 0, 1, 2 . . .
1 − τi
8 2 Layer-Adapted Meshes

t=1 t=1

π(t) ϕ(t)

t=τ
q
χ(t)

x=χ(τ ) x=1 x=1

Fig. 2.2 Bakhvalov mesh: Construction of the mesh generating function (left) and the mesh gen-
erated (right)

converges fast. Moreover, the mesh obtained when the exact τ is replaced by the
first iterate has very similar properties; see, e.g., [11, 23]. In that case
σε σε βq
τ1 = q − and χ(τ1 ) = ln ,
β β σε

are the mesh transition points in the t and x coordinates.

Alternatively, the Bakhvalov mesh can be generated by equidistributing the mon-
itor function

MBa (s) := max 1, Kβε−1 e−βs/σε ; see [85].

Clearly, for p ≤ σ and arbitrary K > 0 there exists a constant C = C(σ, K) with

1 + ε−1 e−βs/pε ≤ C max 1, Kβε−1 e−βs/σε = CMBa (s).

Thus,
 1
[p] C C
ϑcd (ω̄) ≤ MBa (s)ds ≤ , if σ ≥ p, (2.6)
N 0 N
1
for a Bakhvalov mesh since 0 MBa (s)ds ≤ C.
Because (2.5) cannot be solved explicitly, Vulanović [163] proposed to replace
the exponential in the above construction by its (0, 1)-Padé approximation. Thus,
in (2.5) we would take
σε t
χ(t) = .
β t−q
2.1 Convection-Diffusion Problems 9

Meshes that arise from an approximation of Bakhvalov’s mesh generating func-

tion are called meshes of Bakhvalov type (B-type meshes). The following meshes
belong to this class: the meshes proposed by Liseikin and Yanenko [116] (quadratic
function outside layer) and meshes generated by equidistribution of monitor func-
tions which, have been extensively studied by the group of Sloan and Macken-
zie [20, 117, 132, 133], Gartland’s graded mesh [45] and its modification by Roos
and Skalický [140]. The estimate (2.6) holds for these meshes too.
These considerations and (2.4) give the typical convergence result for simple
upwinding on B-type meshes:
 
u − uN  ≤ CN −1 .
∞

I. e., uniform first-order convergence in the discrete maximum norm.

2.1.2 Shishkin Meshes

Another frequently-studied mesh is the so-called Shishkin mesh [121, 150]. This
is because of its simplicity—it is piecewise uniform. We describe this mesh for
problem (2.2). Let q ∈ (0, 1) and σ > 0 be two mesh parameters. We define a mesh
transition point τ by
 
σε
τ = min q, ln N .
β

Then the intervals [0, τ ] and [τ, 1] are divided into qN and (1 − q)N equidistant
subintervals (assuming that qN is an integer). This mesh (if q = τ ) may be regarded
as generated by the mesh generating function
⎧
⎪ σε t
⎪ ϕ̃(t) with ϕ̃(t) = ln N
⎨ for t ∈ [0, q],
β q
ϕ(t) =   (2.7)
⎪
⎪ σε 1−t
⎩1 − 1 − ln N otherwise
β 1−q

see Fig. 2.3. Again the parameter q is the amount of mesh points used to resolve
the layer. The mesh transition point τ has been chosen such that the layer term
exp(−βx/ε) is smaller than N −σ on [τ, 1]. Typically σ will be chosen equal to the
formal order of the method or sufficiently large to accommodate the error analysis.
Note that unlike the Bakhvalov mesh (and Vulanović’s modification of it), the
underlying mesh generating function is only piecewise C 1 [0, 1] and depends on N ,
the number of mesh points. For simplicity we shall assume throughout that q ≥
τ as otherwise N is exponentially large compared to 1/ε, and a uniform mesh is
sufficient to cope with the problem.
10 2 Layer-Adapted Meshes

ξ=1 ξ=1

ξ=q

x=λ x=1 x=1

Fig. 2.3 Shishkin mesh: mesh generating function (left) and the mesh generated (right)

0 1

0 τ 1

Fig. 2.4 Bakhvalov mesh (top) and Shishkin mesh (below) for a convection-diffusion equation

Although Shishkin meshes have a simple structure and many numerical methods
are easier to analyse on a Shishkin mesh than on B-type meshes, they give numerical
results that are inferior to those obtained by B-type meshes. For example,
 
u − uN  ≤ CN −1 ln N,
∞

for the aforementioned simple upwind scheme.

2.1.3 Shishkin-Type Meshes

The convergence on Shishkin meshes is spoilt by the logarithmic factor. This draw-
back prompted some work on improving Shishkin meshes. Roos and Linß [137]
attempt at a general description of these improved meshes. They introduce the
concept of a Shishkin-type mesh (S-type mesh) which is characterised by a
“Shishkin”-transition point τ = σεβ −1 ln N and a uniform submesh on [τ, 1]. Let
the mesh be generated by (2.7) with a monotone function ϕ̃ satisfying

ϕ̃(0) = 0 and ϕ̃(q) = ln N.

We introduce the mesh characterising function ψ(t) = exp(−ϕ̃(t)) for t ∈ [0, q].
This function is monotonically decreasing with ψ(0) = 1 and ψ(q) = N −1 .
2.1 Convection-Diffusion Problems 11

Lemma 2.3. Let ω̄ be a Shishkin-type mesh with σ ≥ p > 0. Assume there exists a
constant κ such that

max ϕ̃ (t) ≤ κN (2.8)

t∈[0,q]

Then
 
[p]  −1
ϑcd (ω̄) ≤ C h + max |ψ (t)| N , (2.9)
t∈[0,q]

Proof. For k = qN + 1, . . . , N we have

   x
−1 −βs/pε p −sβ/pε  k p
1+ε e ds = hk − e  ≤ h + N −1 ,
Ik β xk−1 β

by the choice of the transition point τ and because σ ≥ p.

The argument for k = 1, . . . , qN is slightly more laborious. First, (2.1) gives
 tk
σε
hk = ϕ̃ (t)dt. (2.10)
β tk−1

Therefore,

βhk
≤κ and eβhk /σε ≤ eκ , (2.11)
σε
by (2.8). Furthermore, the above integral representation gives

βhk −βxk /σε

e ≤ N −1 max |ψ  |, (2.12)
σε
because

ψ
ϕ̃ = − and min ψ(t) = ψ(tk ) = e−βxk /σε .
ψ t∈[ti−k ,tk ]

Next, use (2.11), (2.12) and σ ≥ p to obtain

   hk −xk−1 β/pε
1 + ε−1 e−βs/pε ds ≤ hk + e
Ik ε
eκ hk −xk β/σε σ
≤ hk + e ≤ h + N −1 max |ψ  |.
ε β
 
Finally, note that max |ψ  | ≥ 1 − N −1 /q. We are finished. 

We shall survey some of the S-type meshes proposed in the literature now. All
meshes satisfy (2.8).
12 2 Layer-Adapted Meshes

Shishkin mesh [121, 150]:

 
t ln N ln N
ψ(t) = exp − with max |ψ  | = and h ≤ CN −1 .
q q

Thus,

ϑcd (ω̄) ≤ CN −1 ln N.
[p]

Bakhvalov-Shishkin mesh [81, 82]:

This mesh uses Bakhvalov’s idea of inverting exp(−βx/σε) on [0, τ ], while using
a uniform mesh on [τ, 1]. Its mesh generating function is given by
   
1 t
ϕ̃(t) = − ln 1 − 1 − .
N q

The corresponding mesh characterising function is

 
1 t
ψ(t) = 1 − 1 − .
N q

It satisfies
 
1 1 1  
max |ψ  | = 1− ≤ and h ≤ C ε + N −1 .
q N q

Thus,
 
ϑcd (ω̄) ≤ C ε + N −1 .
[p]

In experiments it is observed that the convergence stalls when ε ≥ N −1 , but in

practise one typically has ε  N −1 .
A similar idea has been used by van Veldhuizen
 [161], but he chooses the tran-
sition point in a different manner: τ = 2εβ −1 2 + k ln N , where k is the formal
order of the underlying scheme.

Vulanović-Shishkin mesh [137]:

Again Vulanović’s idea of replacing the exponential by its (0, 1)-Padé approxima-
tion can be used. Then,
 
t ln N t ln N
ϕ̃(t) = , ψ(t) = exp −
q + (q − t) ln N q + (q − t) ln N
2.1 Convection-Diffusion Problems 13

and

4  
max |ψ  | ≤ , h ≤ C 1 + ε ln2 N N −1 .
q

For this mesh:

 
ϑcd (ω̄) ≤ C 1 + ε ln2 N N −1 .
[p]

Polynomial Shishkin mesh [137]:

Choose
 m
t
ϕ̃(t) = ln N, m ≥ 1. (2.13)
q

It can be shown that

max |ψ  | ≤ C (ln N ) h ≤ CN −1 .
1/m
and

Hence,

ϑcd (ω̄) ≤ CN −1 (ln N )

[p] 1/m
.

For m = 1 we recover the original Shishkin mesh, while for m > 1 the accuracy
will be improved.

A mesh with a rational ψ [137].

Let m > 1 and
  1
ϕ̃(t) = ln 1 + (N − 1)(t/q)m , ψ(t) = .
1 + (N − 1)(t/q)m

For this mesh |ψ  | ≤ N 1/m on [0, q]. Thus,

ϑcd (ω̄) ≤ CN −1+1/m .

[p]

Vulanović’s improved Shishkin mesh [169]:

Introduce additional mesh transition points
σε σε σε
τ1 = ln N, τ2 = ln ln N, ..., τ = ln ln · · · ln N.
β β β   
times
14 2 Layer-Adapted Meshes

By formally setting τ0 = 1 and τ +1 = 0, we have τ0 ≥ τ1 ≥ τ2 ≥ · · · ≥ τ +1 .

Then each of the intervals [λi+1 , λi ], i = 0, . . . , , is dissected uniformly.
This mesh can be characterised as a Shishkin-type mesh with

max |ψ  | ≤ C ln · · · ln N
 ln and h ≤ CN −1 .
times

Hence,

ϑcd (ω̄) ≤ CN −1 ln
[p]
· · · ln N.
 ln
times

Remark 2.4. All S-type meshes just mentioned also satisfy

 q
ϕ̃ (t)2 dt ≤ CN (2.14)
0

with some constant C. Then by (2.10) and the Cauchy-Schwarz inequality

qN 
 2  q
βhk
≤ σ 2 N −1 ϕ̃ (t)2 dt ≤ C.
ε 0
k=1

This property will be used later. ♣

Modified Shishkin meshes

Last but not least, two further modifications of the Shishkin mesh are worth men-
tioning. Both are due to Vulanović [168].
The first modification concerns the choice of transition point τ . It is chosen to
be τ = σεβ −1 L(N ), where L = L(N ) solves e−L = LN −1 . This change is
motivated by the appearance of two terms in the error analysis,
 σ
τ
e−βτ /σε and ,
εqN

that should be balanced. For a standard Shishkin mesh they are N −1 and N −1 ln N .
Although L(N ) behaves asymptotically like ln N , the modified transition point typ-
ically yields smaller error constants.
The second modification is a smoothing of the mesh generating function:
⎧
⎨ τq t for t ∈ [0, q],
ϕ(t) =
⎩
τ + τq (t − q) + k(t − q)3 otherwise,
2.1 Convection-Diffusion Problems 15

where k is chosen such that ϕ(1) = 1. This construction ensures ϕ ∈ C 2 [0, 1] with
ϕ ∞ ≤ C. A consequence is

|hi+1 − hi | ≤ CN −2 for i = 1, . . . , N − 1,

a property that is not enjoyed by the Shishkin mesh. It helps to simplify the error
analysis for certain difference schemes.

2.1.4 Turning-Point Boundary Layers

Turning-point layers are associated with zeros of the convection term. Consider the
boundary-value problem of finding u such that

−εx−κ u (x) − b(x)u (x) + c(x)u(x) = f (x) for x ∈ (0, 1),
u(0) = u(1) = 0,

with κ ≥ 0 and b ≥ β > 0 on [0, 1]. The solution will exhibit a boundary whose
k th -order derivatives behave like
 
−k/(κ+1) βxκ+1
ε exp − .
ε(κ + 1)

Bakhvalov meshes
Bakhvalov meshes for turning point layers equidistribute
  
βxκ+1
max 1, Kε−1/(κ+1) exp −
σε(κ + 1)

with user chosen parameters K > 0 and σ > 0.

Shishkin meshes
Choose mesh parameters q ∈ (0, 1) and σ > 0. Determine a mesh transition point
by solving
   1/(κ+1)
−σ βτ κ+1 σε(κ + 1)
N = exp − , i.e., τ= ln N .
ε(κ + 1) β

Then the interval [0, τ ] is divided into qN equidistant subintervals (assuming that
qN is an integer), while [τ, 1] is divided into (1 − q)N subintervals.
16 2 Layer-Adapted Meshes

2.1.5 Interior Layers

Interior layers may occur in convection-diffusion equations with a point source or

when the convection coefficient is discontinuous. In order to achieve uniform con-
vergence, these interior layers have to be resolved by the mesh.
Consider the following convection-diffusion problem:

−εu − (bu) + cu = f + αδd , in (0, 1), u(0) = u(1) = 0, (2.15)

where δd is the shifted Dirac-delta function. Also the convection coefficient b may
have a discontinuity at d ∈ (0, 1). Suppose b ≥ β1 > 0 on (0, d) and b ≥ β2 > 0
on (d, 1). The solution will exhibit two layers: a boundary layer e−β1 x/ε at x = 0
and an interior layer e−β2 (x−d)/ε to the right of x = d. Figure 2.5 depicts a typical
solution.
Uniform convergence for (2.15) will be established in terms of
  
1 + ε−1 e−β1 s/pε + Hd (s)ε−1 e−β2 (s−d)/pε ds,
[p]
ϑcdi (ω̄) := max
i=1,...,N Ii

where Hd is the shifted Heaviside function.

Bakhvalov meshes
Choosing mesh parameters K1 , K2 > 0 and σ1 , σ2 > 0, Bakhvalov meshes
for (2.15) are constructed by equidistributing the monitor function
 
max 1, K0 ε−1 e−β1 s/σ1 ε , Hd (s)K1 ε−1 e−β2 (s−d)/σ2 ε ;

A plot of the resulting mesh is found in Fig. 2.6. If σ1 ≥ p and σ2 ≥ p then

ϑcdi (ω̄) ≤ CN −1 .
[p]

0.5

0.4

0.3

0.2

0.1
Fig. 2.5 Typical solution
of (2.15). The point source is 0
at x = 1/3 0.2 0.4 0.6 0.8 1
2.1 Convection-Diffusion Problems 17

0 d 1

0 τ1 d d+τ2 1

Fig. 2.6 Bakhvalov mesh (top) and Shishkin mesh (below) for a convection-diffusion problem
with an internal layer right of x = d

Shishkin meshes

Let qi ∈ (0, 1), i = 1, . . . , 4 with qi = 1 and σ1 , σ2 > 0 be mesh parameters.
We set
   
q1 d σ 1 ε q3 (1 − d) σ2 ε
τ1 = min , ln N and τ2 = min , ln N .
q1 + q2 β1 q3 + q4 β2

Then the subintervals J1 = [0, τ1 ], J2 = [τ1 , d], J3 = [d, d+τ2 ] and J4 = [d+τ2 , 1]
are divided into qi N equidistant subintervals (assuming that qi N are integers). The
simplest choice is to take qi = 1/4, i = 1, . . . , 4, and N divisible by 4. Figure 2.6
depicts a Shishkin mesh with 16 mesh intervals for (2.15). If σ1 ≥ p and σ2 ≥ p
then ϑcdi (ω̄) ≤ CN −1 ln N .
[p]

2.1.6 Overlapping Layers

Multiple layers can occur, for example, in systems of convection-diffusion

equations:

− diag(ε)u − (Bu) + Au = f in (0, 1), u(0) = u(1) = 0,

where u and f are vectors with  components and A and B are  ×  matrices,
while the  ×  matrix diag(ε) is diagonal with k th entry εk .
We expect layers e−βm x/εm to form at x = 0 if bmm ≥ βm > 0, and lay-
ers e−βm (1−x)/εm at x = 1, if −bmm ≥ βm > 0; see Fig. 2.7.
Let J denote the set of indices m ∈ {1, . . . , } for which bmm is negative and
J ∗ its complement. Then set
   

[p] −1 −βm s/pεm −1 −βm (1−s)/pεm
ϑcd, (ω̄):= max 1+ εm e + εm e ds.
i=1,...,N Ii m∈I m∈I ∗
18 2 Layer-Adapted Meshes

2 2
1.8 1.8
1.6 1.6
1.4 1.4
1.2 1.2
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2

0 0
0.2 0.4 0.6 0.8 1 0.02 0.04 0.06 0.08 0.1

Fig. 2.7 Layers in a system of three convection-diffusion equations. The right plot zooms into the
neighbourhood of x = 0 and reveals the presence of two layers there

Bakhvalov meshes
Bakhvalov meshes for multiple layers can be constructed by generalising the
equidistribution principle for a single layer presented in Sect. 2.1.1. Define

MBa (s) := max 1, max Km βm ε−1
m e
−βm s/σm εm
,
m∈I

max Km βm ε−1
m e
−βm (1−s)/σm εm
for s ∈ [0, 1]
m∈I ∗

with mesh parameters σm > 0 and Km > 0. Choose the mesh points xi such
that the mesh equidistributes this monitor function. Clearly ϑcd, (ω̄) ≤ CN −1 if
[p]

σm ≥ p for all m.

Shishkin meshes
A Shishkin mesh for problem (4.81) is still piecewise equidistant, but now each
layer in u requires its own fine mesh. The mesh is constructed as follows: Let N ,
the number of mesh intervals, be divisible by  + 1. Let σ > 0 be arbitrary. Suppose
the equations in the system are arranged such that bk is positive for k = 1, . . . , i and
negative for k = i + 1, . . . ,  and that
ε1 ε2 εi ε ε −1 εi+1
0< ≤ ≤ ··· ≤ and 0< ≤ ≤ ··· ≤ .
β1 β2 βi β β −1 βi+1

Fix the mesh transition points τk as follows. Set τ0 = 0 and τ = 1. Then, if i > 0 set
 
i σεi
τi = min , ln N ,
 + 1 βi
 
kτk+1 σεk
τk = min , ln N for k = i − 1, . . . , 1;
k + 1 βk
2.2 Reaction-Convection-Diffusion Problems 19

0 1

0 τ1 τ2 1−τ3 1

Fig. 2.8 Bakhvalov mesh (top) and Shishkin mesh (below) for a system of three convection-
diffusion equations

and, if i < ,
 
 − i σεi+1
τi+1 = 1 − min , ln N ,
 + 1 βi+1
 
( − k)τk σεk+1
τk+1 = 1 − min , ln N for k = i + 1, . . . ,  − 1.
 − k + 1 βk+1

Then the mesh is obtained by dividing each of the intervals [τk , τk+1 ], for
k = 0, . . . , , into N/( + 1) subintervals of equal length. Figure 2.8 depicts a
Shishkin mesh with 16 mesh intervals for a system of three equations.
For Shishkin meshes ϑcd, (ω̄) ≤ CN −1 ln N if σm ≥ 1 for all m.
[p]

2.2 Reaction-Convection-Diffusion Problems

Now consider the boundary-value problem

−εd u − εc bu + cu = f in (0, 1), u(0) = u(1) = 0, (2.16)

with two small parameter εd and εc . Its solution typically has two exponential
boundary layers – one at either end of the domain. These behave like eμ0 x and
e−μ1 (1−x) , where the characteristic exponents μ0 < 0 and μ1 > 0 can be computed
from the coefficients in the differential equation. See Fig. 2.9 for a plot of typical
solution and Sect. 3.2 for details of the analysis.
The significant difference to convection-diffusion problems is the presence of
two layers. The construction of adapted meshes has to provide for this.
For this class of problem convergence can be described in terms of
  
1 + |μ0 | eμ0 s/p + μ1 e−μ1 (1−s)/p ds.
[p]
ϑrcd (ω̄) := max
i=1,...,N Ii

A special case of (2.16) is pure reaction-diffusion:

−ε2 u + cu = f in (0, 1), u(0) = u(1) = 0, (2.17)

20 2 Layer-Adapted Meshes

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1

Fig. 2.9 Layers in reaction-convection-diffusion problems (left, different widths) and in reaction-
diffusion problems (right, same widths)

0 1

0 τ1 1−τ2 1

Fig. 2.10 Bakhvalov mesh (top) and Shishkin mesh (below) for a reaction-convection-diffusion
equation. The layers at x = 0 and x = 1 have different widths

with c ≥ ρ2 on [0, 1], ρ > 0. Here μ0 = −μ1 = −ρ/ε and the characteristic
quantity is
  
1 + ε−1 e−ρs/pε + ε−1 e−ρ(1−s)/pε ds.
[p]
ϑrd (ω̄) := max
i=1,...,N Ii

Bakhvalov meshes
Choosing mesh parameters K0 , K1 > 0 and σ0 , σ1 > 0, Bakhvalov meshes
for (2.16) can be generated by equidistributing the monitor function
 
max 1, K0 |μ0 | eμ0 s/σ0 , K1 μ1 e−μ1 (1−s)/σ1 .

In case of a reaction-diffusion equation (2.17) this becomes

 
max 1, K0 ρε−1 e−ρs/σ0 ε , K1 ρε−1 e−ρ(1−s)/σ1 ε .

Figures 2.10 and 2.11 depict Bakhvalov meshes with 16 mesh intervals for (2.16)
and (2.17).
We have

ϑrcd (ω̄) ≤ CN −1 ϑrd (ω̄) ≤ CN −1

[p] [p]
and if σ0 , σ1 ≥ p.
2.2 Reaction-Convection-Diffusion Problems 21

0 1

0 τ 1−τ 1

Fig. 2.11 Bakhvalov mesh (top) and Shishkin mesh (below, with τ1 = τ2 = τ ) for a reaction-
diffusion problem. The layers at x = 0 and x = 1 have the same width

Shishkin meshes
Shishkin meshes are again piecewise uniform. Fixing mesh parameters q0 , q1 > 0
and σ0 , σ1 > 0 with q0 + q1 < 1, we define the mesh transition points
   
σ0 σ1
τ0 = min q0 , ln N and τ1 = min q1 , ln N .
|μ0 | μ1

Then the intervals [0, τ0 ] and [1−τ1 , 1] are dissected into q0 N and q1 N subintervals,
while [τ0 , 1 − τ1 ] is divided into (1 − q0 − q1 )N subintervals. Usually σ0 = σ1 and
q0 = q1 = 1/4 are considered in the literature.
For pure reaction-diffusion the transition points are
   
σ0 ε σ1 ε
τ0 = min q0 , ln N and τ1 = min q1 , ln N .
ρ ρ

For these piecewise uniform meshes

ϑrcd (ω̄) ≤ CN −1 ln N ϑrd (ω̄) ≤ CN −1 ln N

[p] [p]
and if σ0 , σ1 ≥ p.

Modifications of the above construction are conceivable by using ideas from

Sect. 2.1.3 in order to improve the accuracy.

2.2.1 Interior Layers

Interior layers in reaction-diffusion problem occur when the reaction coefficient or

 d in the interior
the right-hand side have a discontinuity at a point  of the domain.
Eq. (2.17) should be read as follows: Find u ∈ C 2 (0, d)∪(d, 1) ∩C 1 [0, 1] such that

−ε2 u + cu = f in (0, d) ∪ (d, 1), u(0) = u(1) = 0, (2.18)

c ≥ ρ2 on [0, 1], ρ > 0. In addition to the two boundary layers e−ρx/ε and
e−ρ(1−x)/ε , an internal layer e−ρ|x−d|/ε will form because of the discontinuity in
the data; see Fig. 2.12.
22 2 Layer-Adapted Meshes

Fig. 2.12 Reaction-diffusion 0.5

problems with an interior
layer at x = 1/3
0.4

0.3

0.2

0.1

0
0.2 0.4 0.6 0.8 1

0 d 1

0 d 1

Fig. 2.13 Bakhvalov mesh (top) and Shishkin mesh (below) for a reaction-diffusion problem with
an internal layer on both sides of x = d

Convergence of discretisations for (2.18) will be characterised by

  
1 + ε−1 e−ρs/pε + e−ρ|s−d|/pε
[p]
ϑrdi (ω̄) := max
i=1,...,N Ii

+ e−ρ(1−s)/pε ds.

Bakhvalov meshes
Choosing mesh parameters K > 0 and σ > 0, we construct Bakhvalov meshes
for (2.15) by equidistributing the monitor function

max 1, Kε−1 e−ρs/σε , Kε−1 e−ρ|s−d|/σε , Kε−1 e−ρ(1−s)/σε .

A plot of the resulting mesh is found in Fig. 2.13. If σ ≥ p then ϑcdi (ω̄) ≤ CN −1 .
[p]

The construction can be modified by choosing different K and σ for different layers.

Shishkin meshes
A possible construction for a piecewise uniform mesh adapted to (2.18) is as fol-
lows: Let N be divisible by 8. Choose a mesh parameter σ > 0 and set
   
d σε 1 − d σε
τ1 = min , ln N and τ2 = min , ln N .
4 ρ 4 ρ
2.2 Reaction-Convection-Diffusion Problems 23

Then the subintervals [0, τ1 ], [d − τ1 , d], [d, d + τ2 ] and [1 − τ2 , 1] are each divided
into N/8 equidistant subintervals, while [τ1 , d − τ1 ] and [d + τ2 , 1 − τ2 are divided
into N/4 subintervals. Figure 2.13 contains of a plot a Shishkin mesh with 16 mesh
intervals for (2.18). If σ ≥ p then ϑrdi (ω̄) ≤ CN −1 ln N .
[p]

2.2.2 Overlapping Layers

In systems of reaction-diffusion equations pairs of layers of different width occur

that overlap. Consider the following system of  equations: Find u such that

−E 2 u + Au = f in (0, 1), u(0) = u(1) = 0. (2.19)

where E = diag(ε1 , . . . , ε ) and the small positive parameters εm are distinct.

Each give rise to a pair of boundary layers e−κx/εm and e−κ(1−x)/εm at the two end
points of the domain, where the constant κ > 0 depends on the coupling matrix A
only. Consequently, layers from different perturbation parameters will overlap; see
Fig. 2.14.
For the above system convergence is described in terms of
    

[p] −1 −κs/pεm −κ(1−s)/pεm
ϑrd, (ω̄) := max 1+ εm e +e ds
i=1,...,N Ii m=1

The presence of multiple layers in the solution u at each end of the interval
[0, 1], forces us to generalise the layer-adapted mesh constructions by refining the
mesh separately for each layer. That is, when approaching an end-point of [0, 1],
one requires a fine mesh that undergoes a further refinement as one enters each new
layer in u.

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0.2 0.4 0.6 0.8 1 0.02 0.04 0.06 0.08 0.1

Fig. 2.14 Layers in a system of two reaction-diffusion equations. The right plot zooms into the
neighbourhood of x = 0 and unveils a thinner sublayer
24 2 Layer-Adapted Meshes

0 1
τ0 τ1 τ2 τ3

0 1

Fig. 2.15 Bakhvalov mesh (top) and Shishkin mesh (below) for a reaction-diffusion system with
multiple layers

Bakhvalov meshes
Bakhvalov meshes for a system of reaction-diffusion equations can be constructed
by equidistributing the monitor function

MBa (s) := max 1, K1 ε−1
1 e
−κs/σ1 ε1
, K1 ε−1
1 e
−κ(1−s)/σ1 ε1
,

−1 −κs/σ ε −1 −κ(1−s)/σ ε
...,K ε e ,K ε e

with positive user chosen constants σm and Km .

For these meshes, if σm ≥ p, m = 1, . . . , , then ϑrd, (ω̄) ≤ CN −1 .
[p]

Shishkin meshes
Piecewise equidistant meshes for (2.19) are constructed as follows: Let N , the num-
ber of mesh intervals, be divisible by 2( + 1). Let σ > 0 be arbitrary. Let the
perturbation parameters be sorted by magnitude: ε1 ≥ ε2 ≥ · · · ≥ ε . Fix the mesh
transition points τk by setting
 
kτk+1 σε +1−k
τ +1 = 1/2, τk = min , ln N for k = , . . . , 1,
k+1 κ

and τ0 = 0. Then the mesh is obtained by dividing each of the intervals [τk , τk+1 ]
and [1 − τk+1 , 1 − τk ], for k = 0, . . . , , into N/(2 + 2) subintervals of equal
length. Figure 2.15 depicts a Shishkin mesh with 24 mesh intervals for a system of
two equations.
If σ ≥ p, then ϑrd, (ω̄) ≤ CN −1 ln N for a Shishkin mesh.
[p]

2.3 Two-Dimensional Problems

The second part of this book (Chap. 7–9) is concerned with two-dimensional reac-
tion-convection-diffusion problems posed on the square Ω = (0, 1)2 :

−εΔu − bT ∇u + cu = f in Ω, u|∂Ω = g.
2.3 Two-Dimensional Problems 25

We shall only consider situations where the layers form along the four edges of
the domain. Therefore, we can restrict ourselves to discretisations on tensor product
meshes, i.e., meshes ω̄ = ω̄x × ω̄y with ω̄x : 0 = x0 < x1 < · · · < xN = 1 and
ω̄y : 0 = y0 < y1 < · · · < yM = 1. For the moment, we will allow meshes with
different numbers of mesh intervals in each coordinate direction. Later, we shall
take N = M for simplicity in the presentation.

2.3.1 Reaction-Diffusion Problems

We will start by considering the boundary-value problem

−ε2 Δu + cu = f in Ω = (0, 1)2 , u|∂Ω = g, (2.20)

where 0 < ε  1 and c > ρ2 on Ω̄ with a positive constant ρ. The convective field
b vanishes identically. Its solution typically exhibits exponential boundary layers
along all four edges of the domain. These layers behave like

e−xρ/ε , e−(1−x)ρ/ε , e−yρ/ε and e−(1−y)ρ/ε .

Figure 2.16 depicts a typical solution to (2.20).

3.5

2.5

1.5

0.5

0
1
0.8
0.6
0.4
1
0.2 0.6 0.8
y 0.4
0 0 0.2
x

Fig. 2.16 Reaction-diffusion problem on the square. Layers of width O (ε ln(1/ε)) at all four
boundaries
26 2 Layer-Adapted Meshes

Fig. 2.17 Bakhvalov mesh (left) and Shishkin mesh (right) for a reaction-diffusion problem on
the square

Appropriate meshes are constructed by taking tensor products of meshes for

one-dimensional reaction-diffusion problems from Sect. 2.2. Figure 2.17 displays
a tensor-product Bakhvalov mesh and a Shishkin mesh. One can also use tensor
products of any of the other meshes proposed for one dimension. It is also possible
to use different meshes in x- and y-direction.

2.3.2 Convection-Diffusion

The general convection-diffusion problem in two dimensions is

−εΔu − bT ∇u + cu = f in Ω ⊂ IR2 , u|∂Ω = g, (2.21)

where 0 < ε  1. Its solution may typically exhibit three different types of lay-
ers: interior layers, parabolic boundary layers and regular boundary layers. Let us
assume that Ω is a domain with a regular boundary that has a uniquely defined out-
ward normal n almost everywhere. Then the boundary can be divided into three
parts:
!
Γ − := x ∈ Γ : bT n < 0 inflow boundary,
!
Γ 0 := x ∈ Γ : bT n = 0 characteristic boundary and
!
Γ + := x ∈ Γ : bT n > 0 outflow boundary.

With this notation the layers can be classified as follows:

Regular Boundary Layers occur at the outflow boundary Γ + and have a width
of O (ε ln(1/ε)). They are often also called exponential boundary layers.
2.3 Two-Dimensional Problems 27

Parabolic Boundary Layers occur at characteristic boundaries Γ 0 where the

boundary is parallel to the characteristics of the vector field b. They are there-
fore also called characteristic boundary layers. In the non-degenerate case,
i.e., when
√ the convective field does not vanish at the boundary, their width
is O ( ε ln(1/ε)).
Interior Layers arise, e. g., from discontinuities in the boundary data at the in-
flow boundary Γ − and are propagated across the domain along the characteristics
of the vector field b. They are similar in nature to parabolic boundary layers and
therefore
√ also called characteristic or parabolic interior layers. Their thickness
is O ( ε ln(1/ε)). They will not be the subject of the theoretical investigations
in this monograph.

2.3.2.1 Regular Boundary Layers

Consider (2.21) on the square Ω = (0, 1)2 with b1 ≥ β1 and b2 ≥ β2 on Ω̄ with

positive constants β1 and β2 . The outflow boundary consists of the two edges x = 0
and y = 0, where boundary layers form that behave like

e−β1 x/ε and e−β2 y/ε .

Figure 2.18 depicts a typical solution.

1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1
0.8
0.6
0.4
1
y 0.2 0.6 0.8
0 0.2 0.4
0 x

Fig. 2.18 Convection-diffusion problem on the square with two exponential layers at the outflow
boundary
28 2 Layer-Adapted Meshes

Fig. 2.19 Bakhvalov mesh (left) and Shishkin mesh (right) for a convection-diffusion problem
with regular layers on the square

Appropriate meshes are constructed by taking tensor products of meshes for

one-dimensional convection-diffusion problems from Sect. 2.1. For example, for
a Shishkin mesh we take transition points
   
σε σε
τx = min q, ln N and τy = min q, ln N
β1 β2

for the meshes in x- and in y-direction, with mesh parameter q ∈ (0, 1) and σ > 0.
Figure 2.19 displays both a Bakhvalov mesh and a Shishkin mesh.

2.3.2.2 Characteristic Boundary Layers

Again consider (2.21) on the square Ω = (0, 1)2 , but with b1 ≥ β and b2 ≡ 0 on Ω̄
with a positive constant β. The outflow boundary consists of the edge x = 0 only,
but there are two characteristic boundaries at y = 0 and y = 1. Three boundary
layers form behaving like

e−βx/ε , e−y e−(1−y)

2 2
/ε /ε
and .

Figure 2.20 depicts a typical solution.

Note that for any m > 0 there exists a constant C > 0 such that
√ √
e−y ≤ e−my/ e−(1−y) ≤ e−m(1−y)/
2 2
/ε ε
and /ε ε
for all y ∈ [0, 1].
2.3 Two-Dimensional Problems 29

2
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.8 1
y 0.2 0.6
0 0.2 0.4
0 x

Fig. 2.20 Convection-diffusion problem on the square with an exponential layer (left) and two
characteristic layers (front and back)

Fig. 2.21 Bakhvalov mesh (left) and Shishkin mesh (right) for a convection-diffusion problem
with a regular layer and two characteristic layers

Therefore, an adapted mesh for a one-dimensional reaction-diffusion equation

can be used in y-direction to resolve the characteristic layers. In x-direction a mesh
for a one-dimensional convection-diffusion equation is used; see Fig 2.21.
 Part I
One Dimensional Problems
Chapter 3
The Analytical Behaviour of Solutions

In this chapter, we gather a number of analytical properties for singularly perturbed

boundary-value problems for second-order ordinary differential equations of the
general type

−εu − bu + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1 ,

with a small positive parameter ε and functions b, c, f : [0, 1] → IR, and of its
vector-valued counterpart

−Eu − Bu + Au = f in (0, 1), u(0) = γ 0 , u(1) = γ 1 ,

with E = diag(ε), ε = (ε1 , . . . , ε )T and small positive constants εi , i = 1, . . . , ,

with matrix-valued functions A, B : [0, 1] → IR , , and vector-valued functions
f , u : [0, 1] → IR .
We shall study stability properties of the differential operators, their Green’s
functions and the behaviour of derivatives of the solutions of various boundary-
value problems. Initial general considerations will be followed by specific results
for various classes of singularly perturbed problems:
• reaction-convection-diffusion problems,
• reaction-diffusion problems,
• convection-diffusion problems with regular layers and
• convection-diffusion problems with turning-point layers.
These results form the basis for our analysis of numerical methods in later chapters.

Notation. For functions v, w : D → IR we write “v ≤ w in/on D” if v(x) ≤ w(x)

for all x ∈ D and “v ≤ 0 in/on D” if v(x) ≤ 0 for all x ∈ D. Similarly, for vectors
v, w ∈ IRn we write “v ≤ w” if vi ≤ wi for i = 1, . . . , n. The notation A ≤ 0
will be used if all components of the matrix A are non-positive. We shall also use
“v < w” and v ≥ w” etc. with the obvious meaning.

T. Linß, Layer-Adapted Meshes for Reaction-Convection-Diffusion Problems, Lecture 33

Notes in Mathematics 1985, DOI 10.1007/978-3-642-05134-0 3,
c Springer-Verlag Berlin Heidelberg 2010
34 3 The Analytical Behaviour of Solutions

3.1 Preliminaries

In this section we consider basic stability properties of differential operators and of

matrices that will be employed throughout this monograph.

3.1.1 Stability of Differential Operators

Maximum and comparison principles are important tools for studying the stability
of differential operators. Consider the general second-order differential operator

Lu := − (au ) + bu + cu

with a, b, c : [0, 1] → IR and a > 0 on [0, 1].

Definition 3.1. The differential operator L is called inverse monotone or of pos-
itive type, if it obeys a comparison principle. That is, for any two functions
v, w ∈ C 2 (0, 1) ∩ C[0, 1]
⎫
Lv ≤ Lw in (0, 1), ⎬
v(0) ≤ w(0), =⇒ v ≤ w on [0, 1].
⎭
v(1) ≤ w(1)
♥
Remark 3.2. The inverse monotonicity can be equivalently characterised as follows:
For any function v ∈ C 2 (0, 1) ∩ C[0, 1]
⎫
Lv ≤ 0 in (0, 1), ⎬
v(0) ≤ 0, =⇒ v ≤ 0 on [0, 1].
⎭
v(1) ≤ 0
♣
Lemma 3.3. Let there exist a function ψ ∈ C 2 (0, 1) ∩ C[0, 1] with ψ > 0 on [0, 1]
and Lψ > 0 in (0, 1). Then the operator L is inverse monotone.
Proof. The proof of the comparison principle is by contradiction; see, e.g. [131].


Corollary 3.4. Let the assumptions of Lemma 3.3 be satisfied. Then for any two
functions v, w ∈ C 2 (0, 1) ∩ C[0, 1]
⎫
|Lv| ≤ Lw in (0, 1), ⎬
|v(0)| ≤ w(0), =⇒ |v| ≤ w on [0, 1]
⎭
|v(1)| ≤ w(1)

and we call w a barrier function of v.

3.1 Preliminaries 35

Lemma 3.5. Suppose there exists a function ψ ∈ C 2 (0, 1) ∩ C[0, 1] with ψ > 0
on [0, 1] and Lψ > 0 in (0, 1). Then for any function v ∈ C 2 (0, 1) ∩ C[0, 1] with
v(0) = v(1) = 0
 
 Lv 
|v| ≤ ψ  
 Lψ  in [0, 1].
∞

Proof. Set L̃v := Lv/Lψ. Clearly L̃ψ ≡ 1. Therefore, L̃ satisfies a comparison

principle. Next,

    (Lv) (x)   





  
 
 L̃v (x) =   ≤  Lv  =  Lv  L̃ψ (x) for x ∈ (0, 1).
(Lψ) (x)   
Lψ ∞  Lψ ∞

 % 
Thus, ψ Lv Lψ ∞ is a barrier function for v. The proposition of the lemma
follows. 


Definition 3.6. Let A and B be two normed spaces. The operator Λ : A → B is

said to be (A, B)-stable if there exists a constant K such that

vA ≤ KΛvB for all v ∈ A.

K is called the stability constant. ♥

Remark 3.7. In the case of discrete (difference) operators one has a family of spaces
Aω and B ω and a family of operators Λω . In this case, the stability inequality reads

vAω ≤ KΛω vB ω for all v ∈ Aω

and for all admissible meshes ω. ♣

Corollary 3.8. Let the assumptions of Lemma 3.5 be satisfied. Then for any function
v ∈ C 2 (0, 1) ∩ C[0, 1] with v(0) = v(1) = 0
 
 Lv 
v∞ 
≤ ψ∞   .
Lψ  ∞

Hence, the operator L is (L∞ , L∞ )-stable, or maximum-norm stable, i.e.,

v∞ ≤ K Lv∞

&
with the stability constant K = ψ∞ min |(Lψ) (x)|.
x∈[0,1]
36 3 The Analytical Behaviour of Solutions

3.1.2 Green’s Functions

Green’s functions form a very powerful tool in the analysis of differential equations.
For example, given the Green’s function G associated with L and Dirichlet boundary
conditions, any function v ∈ C 2 (0, 1) ∩ C[0, 1] with v(0) = v(1) = 0 can be
represented as
 1
v(x) = G(x, ξ) (Lv) (ξ)dξ for all x ∈ (0, 1). (3.1)
0

Remark 3.9. By (3.1) the operator L is inverse monotone if and only if G(x, ξ) ≥ 0
for all x, ξ ∈ (0, 1). ♣
Lemma 3.10. Suppose there exists a function ψ ∈ C 2 (0, 1) ∩ C[0, 1] with ψ > 0
on [0, 1] and Lψ > 0 in (0, 1). Then G ≥ on [0, 1] and we have the following
representation of the (Lψ)-weighted L1 -norm of the Green’s function:
 1
(Lψ) (ξ)G(x, ξ)dξ = ψ(x) for all x ∈ (0, 1).
0

If furthermore Lψ > 0 on [0, 1] then

ψ∞
G(x, ·)1 ≤ for all x ∈ (0, 1).
minξ∈[0,1] (Lψ) (ξ)

Proof. This is an immediate consequence of (3.1) and Lemma 3.3. 



Characterisation of the Green’s function.

For fixed ξ ∈ (0, 1) it solves

(LG(·, ξ)) (x) = δ(x − ξ) for x ∈ (0, 1), G(0, ξ) = G(1, ξ) = 0, (3.2)

with the Dirac-δ distribution. For fixed x ∈ (0, 1) we have

(L∗ G(x, ·)) (ξ) = δ(ξ − x) for ξ ∈ (0, 1), G(x, 0) = G(x, 1) = 0, (3.3)

with the adjoint operator

 
L∗ v = − (av  ) + (bv) + cv.

Eqs. (3.2) and (3.3) have to be read in the context of distributions. Alternatively,
for fixed ξ ∈ (0, 1), one may seek G(·, ξ) ∈ C 2 ((0, ξ) ∪ (ξ, 1)) ∩ C[0, 1] satisfying
' (
LG(·, ξ) = 0 in (0, 1) \ {ξ}, G(0, ξ) = G(1, ξ) = 0, − a∂x G(·, ξ) (ξ) = 1,
3.1 Preliminaries 37

where [v](ξ) := v(ξ + 0) − v(ξ − 0) denote the jump of v at ξ ∈ (0, 1). Similarly,
we can interpret the adjoint problem as finding, for any fixed x ∈ (0, 1), a function
G(x, ·) ∈ C 2 ((0, x) ∪ (x, 1)) ∩ C[0, 1] with
' (
L∗ G(x, ·) = 0 in (0, 1) \ {x}, G(x, 0) = G(x, 1) = 0, − a∂ξ G(x, ·) (x) = 1.

When studying Green’s functions for particular problems, we shall use the fol-
lowing comparison principle for functions with a derivative having a discontinuity
at an interior point of the domain.
Lemma 3.11. Suppose there exists a function ψ ∈ C 2 (0, 1) ∩ C[0, 1] with ψ > 0
on [0, 1] and Lψ > 0 in (0, 1). Let d ∈ (0, 1) be arbitrary, but fixed. Then for any
two functions v, w ∈ C[0, 1] ∩ C 2 ((0, ξ) ∪ (ξ, 1))
⎫
Lv ≤ Lw in (0, 1) \ {ξ} ⎪ ⎪
⎬
v(0) ≤ w(0)
=⇒ v ≤ w on [0, 1].
v(1) ≤ w(1) ⎪
⎪
⎭
−[av  ](ξ) ≤ −[aw ](ξ)

Proof. The proof of the comparison principle is by contradiction, cf. [131]. 



3.1.3 M -matrices

M -matrices are the discrete counterparts of inverse-monotone differential opera-

tors. They are of utmost importance in the convergence analysis of finite difference
methods and when studying systems of differential equations.
Definition 3.12. A matrix A ∈ IRn,n is called
• an L0 -matrix if aij ≤ 0 for all i, j = 1, . . . , n with i = j;
• inverse monotone if for any two vectors v, w ∈ IRn

Av ≤ Aw =⇒ v ≤ w;

• an M -matrix if it is an inverse monotone L0 -matrix. ♥

Remark 3.13. The inverse monotonicity of a matrix A can also be characterised
by A−1 ≥ 0, where it is implicitly assumed that A−1 exists. This is the discrete
counterpart of G ≥ 0 for the Green’s function of a differential operator. ♣
Lemma 3.14 (M -criterion). Let A ∈ IRn,n be an L0 -matrix. Then A is an M -
matrix if and only if there exists a vector e ∈ IRn with e > 0 and Ae > 0.
Moreover,
 −1  &
A  ≤ e min |(Ae)i | =: K  .
∞ ∞ i=1,...,n
38 3 The Analytical Behaviour of Solutions

Remark 3.15. The assumption that A is an L0 -matrix corresponds to the coefficient

of the second-order derivative in the differential operator L being negative, while
the existence of a particular vector e resembles the existence of a positive function
ψ with Lψ > 0. ♣

Remark 3.16. The estimate for A−1 implies

v∞ ≤ K  Av∞ for all v ∈ IRn ,

i. e., A is (∞ , i nf ty) stable, or maximum-norm stable. The stability constant K 

is similar in nature to the constant K in Cor. 3.8. ♣

Lemma 3.17. Let A be an L0 -matrix. Suppose there exists a vector e ∈ IRn with
e > 0 and Ae > 0. Then for any vector v ∈ IRn
 % 
 
|vi | ≤ ei max (Av)j (Ae)j  for i = 1, . . . , n.
j=1,...,n ∞

−1
Proof. Set à := diag (Ae) A. Clearly à is a L0 -matrix and Ãe ≡ 1. Therefore,
it satisfies a comparison principle. Imitating the proof of Lemma 3.5, we obtain the
proposition. 


3.2 Reaction-Convection-Diffusion Problems

In this section we consider the stationary linear reaction-convection-diffusion

problem

Lu := −εd u − εc bu + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1 , (3.4)

with two small parameters 0 < εd  1 and 0 ≤ εc  1. The functions b, c and f

are assumed to be sufficiently smooth with b(x) ≥ 1 and c(x) ≥ 1 for x ∈ [0, 1].
The solution of (3.4) can be described by the two roots of the characteristic
equation

−εd λ(x)2 − εc b(x)λ(x) + c(x) = 0. (3.5)

This quadratic equation defines two continuous functions λi : [0, 1] → IR with

) 2
εc b(x) εc b(x) c(x)
λ0 (x) = − − + ≤ −λ1 (x),
2εd 2εd εd
) 2
εc b(x) εc b(x) c(x)
λ1 (x) = − + + > 0.
2εd 2εd εd
3.2 Reaction-Convection-Diffusion Problems 39

Set
εc
μ0 := max λ0 (x) < ≤ 0 and μ1 := min λ1 (x) > 0.
x∈[0,1] εd x∈[0,1]

The decay of the boundary layers is determined by μ0 and μ1 . At x = 0 there is a

layer behaving like eμ0 x , while the layer at x = 1 is characterised by e−μ1 (1−x) .
There are essentially three regimes:

|μ0 | μ1
convection-diffusion εd  εc = 1 1/εd 1
reaction-convection-diffusion εd  ε2c  1 εc /εd 1/εc
√ √
reaction-diffusion ε2c  εd  1 1/ εd 1/ εd

Figure 2.9 depicts typical solutions of reaction-convection-diffusion problems.

There are two layers of different width at the two end point of the interval.

Remark 3.18. The values of λ0 and λ1 do not vary significantly on [0, 1] because

λ1 (ξ) c(ξ)λ0 (η) b−1 −1
∞ c∞ if εc > 0,
= ≥ −1/2
(3.6)
λ1 (η) c(η)λ0 (ξ) c∞ if εc = 0,

for all ξ, η ∈ [0, 1]. ♣

3.2.1 Stability and Green’s Function Estimates

Using the test function ψ ≡ 1, we see that for the operator L in (3.4) satisfies the
assumptions of and Lemma 3.3 because Lψ = c ≥ 1. Therefore, Corollary 3.8
applies and yields
  
 Lv 
v∞ ≤ max 
 c 
 , |v(0)|, |v(1)| for all v ∈ C[0, 1] ∩ C 2 (0, 1). (3.7)
∞

Deeper insight into the stability properties of (3.4) is gained by studying the
Green’s function associated with L. Note the adjoint operator is given by

L∗ v = −εd v  + εc (bv) + cv.

Figure 3.1 depicts a typical plot of G(x, ·). It is nonnegative, monotonically in-
creasing for ξ < x, but decreasing for ξ > x, and has a maximum of order O (μ1 )
at x = ξ. These properties will be rigorously proved now.
40 3 The Analytical Behaviour of Solutions

1
7

0.8 6

5
0.6
4

0.4 3

2
0.2
1

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

10 10

8 8

6 6

4 4

2 2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Fig. 3.1 Green’s function G(x, ·) associated with L and x = 1/4; εd = 10−3 , εc = 1,
10−1 , 10−2 , 0 (left to right, top to bottom)

Proposition 3.19. Let p ∈ [0, 1] be fixed. Then

−εd p2 μ2i − εc b(x)pμi + c(x) ≥ (1 − p)c(x) for i = 0, 1 and all x ∈ [0, 1].

Proof. We consider i = 0 first. Let x ∈ [0, 1] be arbitrary. Then

2
−εd (pμ0 ) − εc b(x)pμ0 ≥ pμ0 (−εd λ0 (x) − εc b(x))

because p ≤ 1 and 0 > μ0 ≥ λ0 (x) for all x ∈ [0, 1]. The characteristic equa-
tion (3.5) yields
2 μ0
−εd (pμ0 ) − εc b(x)pμ0 ≥ − pc(x)
λ0 (x)

Using c(x) ≥ 1 and λ0 (x) ≤ μ0 < 0 again, completes the proof for μ0 .
Now study i = 1.
2  
−εd (pμ1 ) − εc b(x)pμ1 ≥ p −εd λ1 (x)2 − εc b(x)λ1 (x)

because p ≤ 1 and 0 < μ1 ≤ λ1 (x) and μ21 ≤ λ1 (x)2 for all x ∈ [0, 1]. Using (3.5),
we are finished. 

3.2 Reaction-Convection-Diffusion Problems 41

Theorem 3.20. The Green’s function G associated with L satisfies


1 eμ1 (x−ξ) for 0 ≤ x ≤ ξ ≤ 1,
0 ≤ G(x, ξ) ≤ Ḡ(x, ξ) := ∗
β eμ0 (x−ξ) for 0 ≤ ξ ≤ x ≤ 1,

where β ∗ := εd (μ1 − μ0 ).
If furthermore

εc b + c ≥ 0 on [0, 1] (3.8)

then G is piecewise monotone with

∂x G(x, ξ) ≤ 0, ∂ξ G(x, ξ) ≥ 0 for 0 ≤ ξ < x ≤ 1,

∂x G(x, ξ) ≥ 0, ∂ξ G(x, ξ) ≤ 0 for 0 ≤ x < ξ ≤ 1.

and
∂x ∂ξ G(x, ξ) ≤ 0 for x, ξ ∈ [0, 1], x = ξ.

Proof. The inverse monotonicity of L implies 0 ≤ G on [0, 1]2 .

Proposition 3.19 gives LḠ(·, ξ) ≥ 0 on (0,' 1) \ {ξ}.(Clearly Ḡ(0, ξ) > 0 and
Ḡ(1, ξ) > 0. The jump of ∂x Ḡ satisfies −εd ∂x Ḡ(·, ξ) (ξ) = 1. Application of
Lemma 3.11 establishes the upper bound on G.
Next we prove the monotonicity of G. Because G(x, 0) = G(x, 1) = 0 for x ∈
[0, 1] and G ≥ 0 on [0, 1]2 we have at the boundary of the domain

∂ξ G(x, 0) ≥ 0 and Gξ (x, 1) ≤ 0 for x ∈ [0, 1].

Integrating (3.3) over [0, ξ], we get

   ξ
−εd ∂ξ G(x, ξ) − ∂ξ G(x, 0) + εc b(ξ)G(x, ξ) = − c(s)G(x, s)ds ≤ 0
0

for ξ < x. Thus,

εd ∂ξ G(x, ξ) ≥ εd ∂ξ G(x, 0) + εc b(ξ)G(x, ξ) ≥ 0 for ξ < x

because G(x, ξ) ≥ 0 and Gξ (x, 0) ≥ 0.

On the other hand, inspecting the differential equation (3.3), we see that v =
∂ξ G(x, ·) satisfies

−εd v  + εc bv = −(εc b + c)G ≤ 0 in (ξ, 1) and v(1) ≤ 0,

by (3.8). Application of a comparison principle for first-order operators yields v ≤ 0

on [x, 1].
42 3 The Analytical Behaviour of Solutions

Similarly, one proves that ∂x G(x, ξ) ≥ 0 for 0 ≤ x < ξ ≤ 1 and ∂x G(x, ξ) ≤ 0

for 0 ≤ ξ < x ≤ 1. Thus,

∂x ∂ξ G(x, 0) ≤ 0 and ∂x ∂ξ G(x, 1) ≤ 0 for x ∈ [0, 1].

because ∂x G(x, 0) = ∂x G(x, 1) = 0 for x ∈ [0, 1]. Differentiating (3.3) with respect
to x and integrating with respect to ξ, we get

− εd ∂x ∂ξ G(x, ξ) + εd ∂x ∂ξ G(x, 0) + εc b(ξ)∂x G(x, ξ) − εc b(0)∂x G(x, 0)

 ξ
=− c(s)∂x G(x, s)ds for ξ < x.
0

Hence,

∂x ∂ξ G(x, ξ) ≤ 0 for 0 ≤ ξ < x ≤ 1

because ∂x G(x, ξ) ≤ 0, ∂x ∂ξ G(x, 0) ≤ 0 and ∂x G(x, 0) = 0. For x < ξ, differenti-

ate (3.3) to see that v = ∂x ∂ξ G(x, ·) satisfies

−εv  + bv = −(b + c)∂x G ≤ 0 for x ∈ (0, ξ), v(1) ≤ 0,

by (3.8) and because ∂x G(x, ξ) ≥ 0 for x ≤ ξ. A comparison principle for first-

order operators gives ∂x ∂ξ G(x, ·) ≤ 0 on (x, 1]. 


Remark 3.21. The function Ḡ attains its maximum for x = ξ. Because λ0 is contin-
uous there exists a x∗ ∈ [0, 1] with λ0 (x∗ ) = μ0 . Therefore,

1 1 λ1 (x∗ )
≤ − = ≤ Cμ1 , by (3.6).
β∗ εd λ0 (x∗ ) c(x∗ )

If εc = 0 this estimate can be sharpened to

1 μ1 1
≤ ≤ √
β∗ 2 2 εd
√
since μ1 = −μ0 ≤ 1/ εd . ♣

For our further investigations, let us introduce the L∞ norm

v∞ := ess sup |v(x)|,

x∈[0,1]

the L1 norm
 1
v1 := |v(x)|dx
0
3.2 Reaction-Convection-Diffusion Problems 43
 
1,1 
and the W −1,∞ norm  · −1,∞ . Since W −1,∞ = W̊ the latter is defined by

v−1,∞ := sup u, v,

u∈W̊ 1,1 :|u|1,1 =1

where ·, · is the duality pairing. This norm can also be characterised by
 
 1 
v−1,∞ = min V  = min  v(s)ds + c , (3.9)

∞
V :V =v  c∈IR 
· ∞

see [5] for a more detailed discussion.

Remark 3.22. By (3.9)

  
 1
1 1 
v−1,∞ ≤
 v(s)ds − v(s)ds
 .
· 2 0 ∞

Furthermore
     x 
 1
1 1  1 1 
 v(s)ds −  
v(s)ds =  v(s)ds − v(s)ds .
 2 2 x
x 0 0

Hence,

2 v−1,∞ ≤ v1 ≤ v∞ .

Theorem 3.23. Let x and ξ ∈ (0, 1) be arbitrary. Then the Green’s function G
associated with L satisfies

cG(x, ·)1 ≤ 1. (3.10a)

If additionally (3.8) holds then

2 2
∂ξ G(x, ·)1 ≤ , ∂x G(·, ξ)1 ≤ , (3.10b)
β∗ β∗
2
∂x ∂ξ G(x, ·)1 ≤ (3.10c)
εd

and
   
  2b∞  
εd ∂ξ2 G(x, ·)1 ≤ εc + b  + 2.
β ∗  c ∞
44 3 The Analytical Behaviour of Solutions

Proof. The first estimate, in the weighted L1 -norm, follows from Lemma 3.10.
Next, for fixed x ∈ (0, 1),
 x  1
2
∂ξ G(x, ·)1 = ∂ξ G(x, ξ)dξ − ∂ξ G(x, ξ)dξ = 2G(x, x) ≤ ,
0 x β∗

by Theorem 3.20. Similarly, the bound for ∂x G(·, ξ)1 is obtained.
When calculating ∂x ∂ξ G(x, ·)1 note that
 1
∂x ∂ξ G(x, ξ)dξ = ∂x G(x, 1) − ∂x G(x, 0) = 0,
0

∂x ∂ξ G ≤ 0 for x = ξ and [∂x G(x, ·)](x) = −ε−1 d .

Finally, to get the last inequality of the theorem, integrate (3.3) and apply the
bounds for G and Gξ just established. 

Remark 3.24. Because c ≥ 1 on [0, 1], assuming (3.8) to hold true, provides an
upper threshold value for εc , for which the analysis is valid in the case that b attains
negative values. If b is non-negative everywhere then (3.8) is always satisfied. ♣
The bounds of Theorems 3.20 and 3.23 will be used to establish stability prop-
erties of the differential operator L now. To this end also introduce the weighted
W 1,∞ norm
 
εd  β∗
|||v|||∞ := max v ∞ , v∞ .
2 2

Theorem 3.25. Set β ∗ = εd (μ1 − μ0 ). The operator L satisfies

v∞ ≤ (Lv) /c∞ for all v ∈ W̊ 1,∞ (0, 1) ∩ W 2,∞ (0, 1), (3.11a)
2
v∞ ≤ ∗ Lv1 for all v ∈ W̊ 1,1 (0, 1) ∩ W 2,1 (0, 1) (3.11b)
β
2
v  1 ≤ ∗ Lv1 for all v ∈ W̊ 1,1 (0, 1) ∩ W 2,1 (0, 1) (3.11c)
β

and

|||v|||∞ ≤ Lv−1,∞ for all v ∈ W̊ 1,∞ (0, 1). (3.11d)

Proof. Recall the representation (3.1). Then the Hölder inequality, Theorem 3.20
and (3.10a) give (3.11a) and (3.11b).
Next, let V ∈ W 0,∞ (0, 1) be an arbitrary function with V  = Lv. Integrat-
ing (3.1) by parts, we obtain
 1
v(x) = − ∂ξ G(x, ξ)V (ξ)dξ for x ∈ (0, 1)
0
3.2 Reaction-Convection-Diffusion Problems 45

and
 1
v  (x) = − ∂x ∂ξ G(x, ξ)V (ξ)dξ for x ∈ (0, 1).
0

The Hölder inequality, (3.10b) and (3.10c) yield (3.11d).

Finally,
  
1 1 1   
|v  (x)| ≤ |∂x G(x, ξ)|  Lv (ξ) dξdx.
0 0 0

Changing the order of integration and using (3.10a), we obtain (3.11c). 



Remark 3.26. By (3.11a) the operator L is (L∞ , L∞ )-stable with a stability con-
stant independent of εc and εd . It is also (L∞ , L1 )-stable and (L∞ , W −1,∞ )-stable
because of (3.11b) and (3.11d), but the stability constants depend on εd and εc . ♣

3.2.2 Derivative Bounds and Solution Decomposition

We derive bounds on the derivative of the solution of (3.4) now. The argument fol-
lows [105]. Define

w0,α (x) := eαμ0 x and w1,α (x) := e−αμ1 (1−x) .

They will be used to describe the layers at x = 0 and x = 1, resp.

Proposition 3.27. For any p ∈ [0, 1] and for i = 0, 1

Lwi,p ≥ (1 − p)wi,p in (0, 1).

Proof. The result follows readily from Proposition 3.19. 



Proposition 3.28. For any ϑ ≥ b∞ and i = 0, 1

 
ϑ
−Lwi,ϑ ≥ − c∞ wi,ϑ in (0, 1).
b∞

Proof. We give the proof for i = 0 only because the case i = 1 is analogous. A
direct calculation gives
!
− (Lw0,ϑ ) (x) = εd ϑ2 μ20 − εc ϑb(x)μ0 − c(x) w0,ϑ (x).
46 3 The Analytical Behaviour of Solutions

The function λ0 is continuous. Therefore, it attains its maximum μ0 on [0, 1] and

there exists a ξ ∈ [0, 1] such that λ0 (ξ) = μ0 .
   
b(x) ϑc(ξ)
− (Lw0,ϑ ) (x) = εd ϑλ0 (ξ) ϑ −
2
+ − c(x) w0,ϑ (x)
b(ξ) b(ξ)
 
ϑ
≥ εd ϑλ0 (ξ)2 (ϑ − b∞ ) + − c∞ w0,ϑ (x),
b∞

by (3.5) and because b ≥ 1 and c ≥ 1. The result follows. 



Theorem 3.29. Let b, c, f ∈ C q [0, 1] for some q ∈ IN + . Let p, κ ∈ (0, 1) be

arbitrary, but fixed. Assume

qb ∞ εc ≤ κ(1 − p). (3.12)

Then
 
 (k) 
u (x) ≤ C 1 + (−μ0 )k epμ0 x + μk1 e−pμ1 (1−x) for x ∈ (0, 1) (3.13)

and k = 0, . . . , q.

Proof. The proof is by induction. For k = 0 the result follows from (3.7).
Now assume that (3.13) holds for k = 0, . . . , m < q. Then differentiating (3.4)
k-times, we get

Lk u(k) := Lu(k) − εc kb u = gk in (0, 1), (3.14)

where

g0 = f, c0 = c, ck = c − kεc b and 
gk = gk−1 − ck−1 u(k−1) .

We prove that (3.13) holds for k = m + 1. Eqs. (3.13) and (3.14) give
  
 
 Lm+1 u(m+1) (x) ≤ C 1 + (−μ0 )m w0,p (x) + μm 1 w1,p (x) , x ∈ (0, 1).

By Lemma 3.3 the operator Lm+1 obeys a comparison principle, because for ψ ≡ 1

Lm+1 ψ = c + εc (m + 1)b ≥ 1 − ε2 qb ∞ ≥ 1 − κ(1 − p) > 0.

Thus, if we had
 (m+1)   
u (0) ≤ C(−μ0 )m+1 and u(m+1) (1) ≤ Cμm+1 , (3.15)
1
3.2 Reaction-Convection-Diffusion Problems 47

then application of the comparison principle with the barrier function

w(x) = C1 1 + (−μ0 )m+1 w0,p (x) + μm+1

1 w1,p (x)

with C1 chosen sufficiently large, independently of εd and εc , would yield (3.13)

for k = m + 1.
Therefore, we are left with proving (3.15). Introduce

um (x) := u(m) (x) − u(m) (0)(1 − x).

This function satisfies

k k−1
|(Lm um ) (x)| ≤ C (−μ0 ) + (−μ0 ) w0,p (x) + μk−1
1 w1,p (x) ,

x ∈ (0, 1), and

um (0) = 0, |um (1)| ≤ Cμk1 .

Now consider
   
w(x) = C2 (−μ0 )k 1 + w0,p (x) − 2w0,ϑ (x) + C3 μk1 w1,p (x) − w1,p (0)

***as a possible barrier function for um . Let ϑ ≥ b∞ . Then Propositions 3.27
and 3.28 yield

(Lw) (x) ≥ C1 (−μ0 )k 1 − εc qb ∞ + (1 − p − εc qb ∞ ) w0,p (x)
* + 
ϑ
+2 − c∞ − εc qb ∞ w0,ϑ (x)
b∞

+ C2 μk1 − (c∞ + εc qb ∞ ) w1,p (0)

+ (1 − p − εc qb ∞ ) w1,p (x) .

Fix ϑ ≥ b∞ (1 + c∞ ) and recall (3.12). Then there exist positive constants
C4 = C4 (C2 , C3 ), C5 = C5 (C2 ) and C6 = C6 (C3 ) with

(Lw) (x) ≥ C4 (−μ0 )k + (−μ0 )k w0,p (x) + μk1 w1,p (x) + C5 μk0 − C6 ,

% , 
because μk1 w1,p (0) ≤ C. Note μ0 ≥ 1 b∞ + c∞ . Thus, choosing C2 and
C3 sufficiently large, independently of εd and εc , we see that w is a barrier function
for um .
48 3 The Analytical Behaviour of Solutions

By the definition of the derivative we have

w(x)
|um (0)| ≤ lim ≤ C2 (−μ0 )m+1 (2ϑ − p) + C3 μm+1
1 w1,p (0).
x→+0 x

Hence,

|um (0)| ≤ C(−μ0 )m+1 .

We get |u(m+1) (0)| ≤ C(−μ0 )m+1 , which is the first bound of (3.15). Analogously
one estimates u(m+1) (1). 


3.3 Reaction-Diffusion Problems

This section is concerned with linear reaction-diffusion problems. First, we shall

consider the scalar problem

Lu := −ε2 u + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1 ,

while the second part of the section is devoted to vector-valued problems

Lu := −E 2 u + Au = f in (0, 1), u(0) = γ 0 , u(1) = γ 1 ,

where E = diag(ε1 , . . . , ε ) and the small parameters εk are in (0, 1].

3.3.1 Scalar Reaction-Diffusion Problems

Consider the boundary-value problem: Given c, f ∈ C[0, 1], find u ∈ C 2 (0, 1) ∩

C[0, 1] such that

Lu := −ε2 u + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1 , (3.16)

with the perturbation parameter ε > 0. Furthermore, we assume c ≥ 1 on [0, 1].

We shall apply some of the results of Sect. 3.2. Note that εd = ε2 and εc = 0.
The characteristic roots satisfy
,
λ1 (x) = −λ0 (x) = ε−1 c(x) for x ∈ [0, 1] and μ1 = −μ0 ≥ ε−1 .

Therefore, two layers of equal widths will form at both ends of the domain.
3.3 Reaction-Diffusion Problems 49

Remark 3.30. The focus of our analysis is on linear problems. Using standard
linearisation techniques, all results can be generalised to semilinear problems

T u := −ε2 u + c(u, ·) = 0 in (0, 1), u(0) = γ0 , u(1) = γ1 ,

with ∂u c ≥ 1 in IR × (0, 1). ♣

3.3.1.1 Stability Properties

By (3.7) we have
  
 Lv 
v∞ ≤ max 
 c 
 , |v(0)|, |v(1)| (3.17)
∞

for all functions v ∈ C[0, 1] ∩ C 2 (0, 1).

When deriving bounds on the Green’s function we shall avail of Theorems 3.20
and 3.23. To this end, note that (3.8) is always satisfied and that β ∗ ≥ 2/ε, see
Remark 3.21. Also the Green’s function is symmetric, i.e., G(x, ξ) = G(ξ, x) for all
x, ξ ∈ [0, 1], because the operator L is self-adjoint.
Theorem 3.31. The Green’s function G associated with (3.16) satisfies the point-
wise bounds

e−|x−ξ|/ε
0 ≤ G(x, ξ) ≤ for x, ξ ∈ [0, 1],
2ε

and the (weighted) L1 -norm bounds

cG(x, ·)1 ≤ 1, ∂ξ G(x, ·)1 ≤ ε−1 and ∂ξξ G(x, ·)1 ≤ 2ε−2 .

Remark 3.32. The bounds of the theorem are slight improvements over results first
given by Kopteva [66]. ♣

Next stability inequalities for the operator will be derived. We shall make use of
the following:
Lemma 3.33. For any function g ∈ W 2,∞ [a, a + μ] with μ > 0 there holds
2 μ
g  ∞,[a,a+μ] ≤ g∞,[a,a+μ] + g  ∞,[a,a+μ] .
μ 2

Proof. This is Lemma 1 from [18]. 


50 3 The Analytical Behaviour of Solutions

Theorem 3.34. For any function v ∈ W̊ 1,∞ (0, 1) ∩ W 2,∞ (0, 1) the operator L
satisfies

5 + c∞
v∞ ≤ (Lv) /c∞ , ε v  ∞ ≤ Lv∞
2

and

ε2 v  ∞ ≤ (1 + c∞ ) Lv∞ .

Proof. The first inequality can be concluded from (3.17), Theorem 3.25 or Theo-
rem 3.31.
Then, ε2 v  = cv + Lv, a triangle inequality, the bound on v and c ≥ 1 yield the
third estimate of the theorem.
Finally, Lemma 3.33 with μ = ε applied to g = v gives the bound on v  . 


3.3.1.2 Bounds on Derivatives

Theorem 3.29 provides first bounds for the derivatives. Let c, f ∈ C q [0, 1] for some
q ∈ IN + . Let p ∈ (0, 1) be arbitrary, but fixed. Then

 (k) 
u (x) ≤ C 1 + ε−k e−px/ε + ε−k e−p(1−x)/ε , for x ∈ (0, 1) (3.18)

and k = 0, . . . , q.
Note, in order to establish bounds for derivatives of order q, the data is as-
sumed to lie in C q [0, 1], while general theory for second-order differential equations
guarantees the existence of a unique solution in C q+2 [0, 1]. The following theory
sharpens (3.18), gives bounds for higher-order derivatives and provides a solution
decomposition, which is useful for the analysis of some numerical methods.

Theorem 3.35. Suppose c, f ∈ C q [0, 1], q ∈ IN with c ≥ 1 on [0, 1]. Then (3.16)
possesses a unique solution u ∈ C q+2 [0, 1]. It can be decomposed as

u = v + w0 + w1

with

Lv = f, Lw0 = 0 and Lw1 = 0 in (0, 1).

The regular part v satisfies

 (m) 
v  ≤ C(1 + εq−m ),
∞
3.3 Reaction-Diffusion Problems 51

while for the layer pars w0 and w1 we have

 (m)   (m) 
w (x) ≤ Cε−m e−x/ε and w (x) ≤ Cε−m e−(1−x)/ε (3.19)
0 1

for x ∈ [0, 1] and m = 0, 1, . . . , q.

Proof. The functions
 c and f can be extended to functions c̄, f¯ ∈ C q [−1, 2] with
¯
c̄|[0,1] = c, f [0,1] = f and c̄ ≥ 1/2 on [−1, 2]. Let v̄ be the solution of

−ε2 v̄  + r̄v̄ = f¯, in (−1, 2), v̄(−1) = v̄(2) = 0 and set v := v̄|[0,1]
 
An affine transformation and (3.18) yield v (m) ∞ ≤ C for m = 0, . . . , q. Clearly
 
Lv = f on (0, 1). Therefore, v (q+2) ∞ ≤ Cε−2 . Lemma 3.33 with g = v (q) and
 (q+1) 
μ = ε gives v  ≤ Cε−1 .
∞
The layer components w0 and w1 solve

Lw0 = 0 in (0, 1), w0 (0) = u(0) − v(0), w0 (1) = 0

and

Lw1 = 0 in (0, 1), w1 (0) = 0, w1 (1) = u(1) − v(1).

Consider w0 . Application of Lemma 3.3 yields (3.19) for m = 0. From Lw0 = 0

we get (3.19) for m = 2. Lemma 3.33 with μ = ε and g = w0 , gives (3.19) for m =
1. The bounds on the higher-order derivatives on w0 follow upon differentiating
Lw0 = 0. The same argument is used for w1 . 


3.3.1.3 Discontinuous Data

The data, i.e. the reaction coefficient and the right-hand side, may have discon-
tinuities at a number of points in the domain. For simplicity, let us consider a
single point
 of discontinuity
 at x = d ∈ (0, 1). Then (3.16) takes the form: Given
c, f ∈ C (0, d) ∪ (d, 1) find u ∈ C 2 (0, d) ∪ (d, 1) ∩ C 1 [0, 1] such that

Lu := −ε2 u + cu = f in (0, d) ∪ (d, 1), u(0) = γ0 , u(1) = γ1 . (3.20)

The consequence of the discontinuity in the data is the formation of an interior layer
at x = d.
The modified
 operatorobeys a comparison principle too: For any two functions
v, w ∈ C 2 (0, d) ∪ (d, 1) ∩ C 1 [0, 1]
⎫
Lv ≤ Lw in (0, 1), ⎬
v(0) ≤ w(0), =⇒ v ≤ w on [0, 1].
⎭
v(1) ≤ w(1)
52 3 The Analytical Behaviour of Solutions

By means of this comparison principle, stability of the operator L can be

established. In particular, we have
  
 Lv 
v∞ ≤ max   c 
 , |v(0)|, |v(1)|
∞
 
for any function v ∈ C 2 (0, d)∪(d, 1) ∩C 1 [0, 1]. This is a generalisation of (3.17).
The bounds on the Green’s function associated with L remain unaffected by the
discontinuity of the reaction coefficient c.
Estimates for the derivatives are obtained as follows. First note that the
maximum-norm stability implies |v(d)| ≤ C. Then the solution of (3.20) is anal-
ysed on the two subdomains (0, d) and (d, 1) separately. On each of these domains
we have a problem with continuous data and the results of Sect. 3.3.1.2 apply.
We get
 (k)   
u (x) ≤ C 1 + ε−k e−x/ε + e−|d−x|/ε + e−(1−x)/ε ,

for x ∈ (0, d) ∪ (d, 1).

3.3.2 Systems of Reaction-Diffusion Equations

We now leave the scalar equation (3.16) and move on to systems of equations of this
 
type: Find u ∈ C 2 (0, 1) ∩ C[0, 1] such that

Lu := −E 2 u + Au = f in (0, 1), u(0) = u(1) = 0, (3.21)

where E = diag(ε1 , . . . , ε ) and the small parameter εk is in (0, 1] for k = 1, . . . , .

We set A = (aij ) and f = (fi ). Written out in full, (3.21) is

−ε21 u1 + a11 u1 + a12 u2 + · · · + a1 u = f1 in (0, 1), u1 (0) = u1 (1) = 0,

−ε22 u2 + a21 u1 + a22 u2 + · · · + a2 u = f2 in (0, 1), u2 (0) = u2 (1) = 0,
..
.
−ε2 u + a 1 u1 + a 2 u2 + · · · + a u = f in (0, 1), u (0) = u (1) = 0.

3.3.2.1 Stability

Assume that all entries aij of the coupling matrix A lie in C[0, 1] and that A has
positive diagonal entries. Assume likewise that all fi lie in C[0, 1]. Our analysis
follows that of [104] and is based on the stability properties of Sect. 3.3.1.1 for
scalar operators.
3.3 Reaction-Diffusion Problems 53

For each k, the k th equation of the system (3.21) can be written as


−ε2k uk + akk uk = fk − akm um .
m=1
m=k

The stability inequality (3.17) and a triangle inequality then yield

   
  akm   fk 
uk ∞ −   um ∞ 
≤  (3.22)
 akk  akk 
m=1 ∞ ∞
m=k

Define the  ×  constant matrix Γ = Γ (A) = (γkm ) by

 
 akm 
γkk = 1 and γkm = −    for k = m.
akk ∞

Suppose that Γ is inverse-monotone, i.e., that Γ is invertible and

Γ −1 ≥ 0; (3.23)

this can be verified by using Lemma 3.14. Then (3.22) immediately gives a bound
on u∞ in terms of the data A and f . We obtain the following stability result for
the operator L.

Theorem 3.36. Assume the matrix A has positive diagonal entries. Suppose also
that all entries of A lie in C[0, 1]. Assume that Γ (A) is inverse-monotone. Then for
k = 1, . . . ,  one has

  
  (Lv)m 
vk ∞ ≤ Γ −1  
km  a 
m=1 mm ∞

 
for any function v = (v1 , . . . , v )T ∈ C 2 (0, 1) ∩ C[0, 1] with v(0) = v(1) = 0.

Corollary 3.37. Under the hypotheses of Theorem 3.36, the boundary value prob-
lem (3.21) has a unique solution u, and u∞ ≤ C f ∞ for some constant C.
Thus, the operator L is (L∞ , L∞ ) stable, or maximum-norm stable, although
in general it is not inverse-monotone — the hypotheses of Theorem 3.36 do not in
general imply that (3.21) obeys a comparison principle.

Remark 3.38. The obvious analogue of the stability inequality (3.17) is also valid
for scalar reaction-diffusion problems posed in domains Ω lying in IRd for d > 1.
Consequently, Theorem 3.36 holds true also for reaction-diffusion systems posed
on Ω ⊂ IRd with d > 1. ♣
54 3 The Analytical Behaviour of Solutions

Remark 3.39. The above argument is closely related to Ostrovski’s concept of

H-matrices [129]. He calls a matrix A = (aij ) ∈ C , an H-matrix if its com-
panion matrix
⎛ ⎞
|a11 | −|a12 | · · · −|a1 |
⎜−|a21 | |a22 | · · · −|a2 |⎟
⎜ ⎟
⎜ . .. .. .. ⎟
⎝ .. . . . ⎠
−|a 1 | |a 2 | · · · −|a |

is an M -matrix. ♣
In [18, 58] the coupling matrix A is assumed to be coercive, viz.,

v T A(x)v ≥ μ2 v T v for all v ∈ IR and x ∈ [0, 1], (3.24)

where μ is some positive constant. The following result from [111], which slightly
generalises [159], establishes a connection between (3.23) and (3.24).
Lemma 3.40. Assume that A has positive diagonal entries and that Γ is inverse-
monotone. Then there exists a constant diagonal matrix D and a constant α > 0
such that

v T DA(x)v ≥ αv T v for all v ∈ IRd , x ∈ [0, 1],

i.e., the matrix DA is coercive uniformly in x.

Proof. As Γ −1 exists, one can define y, z ∈ IR by Γ y = 1 and Γ T z = 1. Then
Γ −1 ≥ 0 implies that yi > 0 and zi > 0 for i = 1, . . . , . Define the matrix-valued
function G = (gij ) by gij (x) = zi aij (x)yj for all i and j. Observe that both G and
GT are strictly diagonally dominant:
 
gii (x) − |gij (x)| ≥ aii (x)zi γij yj = aii (x)zi > 0,
j=i j
 
gii (x) − |gji (x)| ≥ aii (x)yi γji zj = aii (x)yi > 0.
j=i j

Thus (G + GT )/2 is strictly diagonally dominant and symmetric. Hence, there

exists a constant β > 0 such that

G + GT
v T Gv = v T GT v = v T v ≥ βv T v for all v ∈ IR and x ∈ [0, 1].
2

Define the diagonal matrix D = (dii ) by dii = zi /yi for all i. Then
  vi vj   vi  2 
T
v DA(x)v = dii aij vi vj = gij ≥β ≥α vi 2 .
i,j i,j
y i y j i
y i i



3.3 Reaction-Diffusion Problems 55

Remark 3.41. Lemma 3.40 remains valid for reaction-diffusion problems posed in
Ω ⊂ IRd with d > 1. ♣

Remark 3.42. As multiplication on the left by a positive diagonal matrix neither

changes the structure of (3.21) nor alters Γ (A), Lemma 3.40 implies that, without
loss of generality, if A has positive diagonal entries, then whenever (3.23) is sat-
isfied, one can assume that (3.24) holds true also. Thus, the hypothesis that (3.24)
alone holds true is more general than an assumption that (3.23) is valid, but the only
analyses [18, 58] that are based solely on (3.24), are restricted to the special case
ε1 = ε2 = · · · = ε ; see Sect. 7.2. ♣

3.3.2.2 Derivative Bounds

Let the coupling matrix A(x) be strictly diagonally dominant for all x ∈ [0, 1].
Then A has positive diagonal entries and there exists a constant β such that

 
  akm 
 
 akk  ≤ β < 1 for k = 1, . . . , . (3.25)
m=1 ∞
m=k

An application of the M-criterion (Lemma 3.14) with a constant test vector e shows
that Γ −1 ≥ 0. Define κ = κ(β) > 0 by

κ2 := (1 − β) min min akk (x).

k=1,..., x∈[0,1]

For arbitrary ε ∈ (0, 1] and 0 ≤ x ≤ 1, set

Bε (x) := e−κx/ε + e−κ(1−x)/ε .

For simplicity in our presentation we assume that

κ
ε 1 ≥ ε2 ≥ · · · ≥ ε and ε1 ≤ ;
4
the first chain of inequalities can always be achieved by renumbering the equations,
while the last inequality provides a threshold value for the validity of our analysis.
The next result generalises Theorem 3.35.
Theorem 3.43. Let A and f be twice continuously differentiable. Assume (3.25)
holds true. Then the solution u of (3.21) can be decomposed as u = v + w, where
v and w are defined by

−E 2 v  + Av = f in (0, 1), v(0) = A(0)−1 f (0), v(1) = A(1)−1 f (1),

56 3 The Analytical Behaviour of Solutions

and

−E 2 w + Aw = 0 in (0, 1), w(0) = −v(0), w(1) = −v(1).

For all x ∈ [0, 1] and k = 1, . . . , , the derivatives of v and w satisfy the bounds
 (ν)   
v  ≤ C 1 + ε2−ν for ν = 0, 1, . . . , 4,
k ∞ k

 (ν)  
k
w (x) ≤ C ε−ν
m Bεm (x) for ν = 0, 1, 2
k
m=1

and

 (ν)  
w (x) ≤ Cε2−ν ε−2
m Bεm (x) for ν = 3, 4.
k k
m=1

Proof. The proof in [104] is involved and full of technical details that will not be
discussed here. For the layer components the argument proceeds via induction for
the components w1 , w2 , . . . , w . First, bounds for w1 are established. Then w1 is
eliminated from the system and considered as an inhomogeneity for the system of
the remaining components. Next, w2 is bounded and subsequently eliminated; etc.



The bounds of Theorem 3.43 say that each component uk of the solution u can
be written as a sum of a smooth part (whose low-order derivatives are bounded
independently of the small parameters) and  overlapping layers, though the full
effect of these layers is manifested only in derivatives of order at least 3.
Figure 3.2 displays a typical solution in the case  = 2. The first plot shows
the two components on the entire domain [0, 1]; all that is apparent is that each
component has layers at x = 0 and x = 1. The second plot is a blow-up of the layer

4 4
u2 u2
3.5 u1 3.5 u1
3 3

2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.01 0.02 0.03 0.04 0.05 0.06

Fig. 3.2 Overlapping layers in a system of two reaction-diffusion equations

3.4 Convection-Diffusion Problems with Regular Layers 57

at x = 0 (the layer at x = 1 is similar) and we observe that while u1 has a standard

layer, in u2 there are two separate layers—exactly as predicted by Theorem 3.43.
This theorem also forecasts that u1 has no such visible behaviour, since interactions
between layers in u1 appear only in the third-order and higher derivatives, and these
are not easily noticed on graphs.

3.4 Convection-Diffusion Problems with Regular Layers

Now we consider the linear scalar convection-diffusion problem

Lu := −εu − bu + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1 ,

with 0 < ε  1, and its vector-valued counterpart

Lu := − diag(ε)u − Bu + Au = f in (0, 1), u(0) = u(1) = 0,

where ε = (ε1 , . . . , ε )T and the εk are small parameters.

3.4.1 Scalar Convection-Diffusion Problems

Given b, c, f ∈ C[0, 1], find u ∈ C 2 (0, 1) ∩ C[0, 1] such that

Lu := −εu − bu + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1 , (3.26)

where ε is a small positive parameter, b ≥ β > 0 on [0, 1].

For the mere sake of simplicity we shall also assume that

c ≥ 0 and b + c ≥ 0 on [0, 1]. (3.27)

The results hold without these restrictions too, see [5], but the arguments become
more complicated and the constants will be slightly different. Note that (3.27) can
always be ensured for ε smaller than a certain threshold value ε0 by a simple trans-
formation u = ûeδx with δ chosen appropriately. This is because b ≥ β > 0.
We shall apply some of the results of Sect. 3.2. This time εd = ε and εc = 1. For
the characteristic exponents we have
β
μ0 ≤ − and μ1 = O (1) for ε → 0.
ε
Therefore, only one layer at x = 0 will be present in the solution.
58 3 The Analytical Behaviour of Solutions

3.4.1.1 Stability

Using the test function ψ : x → ψ(x) = 1 − x, we see that the differential operator
L in (3.26) satisfies the hypotheses of Lemma 3.3, because Lψ ≥ β > 0. Conse-
quently, it obeys a comparison principle and in association with Lemma 3.5 we get
!
|u(x)| ≤ max |γ0 |, |γ1 | + (1 − x) f /b∞ for x ∈ [0, 1].

This implies

v∞ ≤ Lv/b∞ for all v ∈ C 2 [0, 1] with v(0) = v(1) = 0. (3.28a)

Alternatively, if c > 0 on [0, 1], then Lemma 3.5 with ψ ≡ 1 yields

v∞ ≤ Lv/c∞ for all v ∈ C 2 [0, 1] with v(0) = v(1) = 0, (3.28b)

see also Theorem 3.25.

Remark 3.44. One can try to derive stability results for L that generalise both
(3.28b) and (3.28a). This can be done, for example, by using as test function ψ
a general linear or quadratic function ϕ. But the resulting stability equality will be
more complicated and seem to be difficult to use. ♣

Green’s function estimates

Theorem 3.20 applies to (3.26) because the general assumption c ≥ 1 in Sect. 3.2
is not used in the argument leading to Theorem 3.20, but only c ≥ 0. Note that
β ∗ ≥ β, μ0 ≤ −β/ε and μ1 ≥ 0. We get

1 1 for 0 ≤ x ≤ ξ ≤ 1,
0 ≤ G(x, ξ) ≤
−β(x−ξ)/ε
β e for 0 ≤ ξ ≤ x ≤ 1.

These lower and upper bounds can also be verified by appealing to Lemma 3.3.
The first plot in Fig. 3.1 depicts the typical behaviour of the Green’s function for a
convection-diffusion problem.
Clearly (3.27) implies that (3.8) is satisfied. Therefore, the derivative of G possess
the sign pattern described in Theorem 3.20. Consequently, (3.10b,c) hold true and
can be used to obtain stability estimates for L.
We summarise the results.
Theorem 3.45. Suppose (3.27) holds true. Then the operator L in (3.26) satisfies
    
 Lv   Lv 
v∞ ≤ min 
 b 
 ,
 c 

∞ ∞ (3.29a)
for all v ∈ W̊ 1,∞ (0, 1) ∩ W 2,∞ (0, 1),
3.4 Convection-Diffusion Problems with Regular Layers 59

1 2
v∞ ≤ Lv1 and v  1 ≤ Lv1
β β (3.29b)
for all v ∈ W̊ 1,1
(0, 1) ∩ W 2,1
(0, 1)

and

|||v|||ε,∞ ≤ 2Lv−1,∞ for all v ∈ W̊ 1,∞ (0, 1), (3.29c)

with
!
|||v|||ε,∞ := max εv  ∞ , βv∞ .

Remark 3.46. Note that in view of Remark 3.22, the (L∞ , W −1,∞ )-stability (3.29c)
is the strongest of the three stability inequalities of Theorem 3.45. It was first given
by Andreev and Kopteva [11] and later analysed in more detail by Andreev [6]. It
implies that the operator L is also uniformly (L∞ , L1 )-stable and (L∞ , L∞ )-stable,
i. e. with stability constants that are independent of ε. ♣

Remark 3.47. The same stability results hold true for the differential operator in
conservative form, i. e., Lc u := −εu − (bu) + cu. ♣

3.4.1.2 Derivative Bounds and Solution Decomposition

The boundary value problem (3.26) has a unique solution that typically has an expo-
nential boundary layer at x = 0: u and its derivatives up to an arbitrary prescribed
order q can be bounded by
 (k) 
u (x) ≤ C 1 + ε−k e−βx/ε for k = 0, 1, . . . , q and x ∈ [0, 1], (3.30)

where the maximal order q depends on the smoothness of the data, see [61].
On a number of occasions, e. g. for the error analysis of a finite difference scheme
in Sect. 4.2.6 or of the FEM in Sect. 5.2, we need more detailed information on u and
its derivatives. In particular, a splitting of u into a regular solution component and
a boundary layer component will be required. This decomposition will be derived
now.
Defining L0 v := −bv  + cv, we follow [89] and construct the decomposition as
follows. Let v and w be the solution of the boundary-value problems

Lv = f in (0, 1), (L0 v) (0) = f (0), v(1) = γ1 (3.31a)

and

Lw = 0 in (0, 1), w(0) = γ0 − v(0), w(1) = 0. (3.31b)

60 3 The Analytical Behaviour of Solutions

First we study the regular solution component v. The operator L equipped with
mixed Robin and Dirichlet boundary conditions satisfies the following comparison
principle [131]. For any two functions z, z̄ ∈ C[0, 1] ∩ C 1 [0, 1) ∩ C 2 (0, 1)
⎫
Lz ≤ Lz̄ in (0, 1), ⎬
(L0 z) (0) ≤ (L0 z̄) (0), =⇒ z ≤ z̄ on [0, 1].
⎭
z(1) ≤ z̄(1)

Using this comparison principle with the barrier functions v̄ and −v̄ defined by

v̄(x) := β −1 f ∞ (1 − x) + |γ1 | for x ∈ [0, 1],

we get

|v(x)| ≤ v̄ ≤ C for x ∈ [0, 1].

To derive bounds on the derivatives of v, set h := f − cu and write v as

 1  1
h(0)
v(x) = Hv (s)ds + e−B(s) ds + γ1 ,
x b(0) x

where
 x  x
1 1
B(x) := b(s)ds and Hv (x) := h(s)eB(s)−B(x) ds.
ε 0 ε 0

Differentiating once, we get

h(0) −B(x)
v  (x) = −Hv (x) − e
b(0)

which gives

|v  (x)| ≤ C for x ∈ [0, 1]

because

C x
C 
Hv (x) ≤ eβ(s−x)/ε ds = 1 − e−βx/ε ≤ C. (3.32)
ε 0 β

Invoking the differential equation we get

|v  (x)| ≤ Cε−1 for x ∈ [0, 1].

3.4 Convection-Diffusion Problems with Regular Layers 61

However, if b, f ∈ C 1 (0, 1) then integration by parts and the boundary condition

imposed on v at x = 0 yield
  
 b(x) x h
v (x) = − (s)eB(s)−B(x) ds,
ε 0 b

from which the sharper estimate

|v  (x)| ≤ C for x ∈ [0, 1]

can be derived using (3.32). A bound for the third-order derivative is readily ob-
tained from the differential equation and the bounds on v  and v  :

|v  (x)| ≤ Cε−1 for x ∈ [0, 1].

This completes our analysis of the regular component v of u.

Now consider the boundary-layer term w. Corollary 3.4 with the barrier function

w̄ : [0, 1] → IR : x → w̄(x) := |γ0 − v(0)| e−βx/ε

yields

|w(x)| ≤ Ce−βx/ε for x ∈ [0, 1]. (3.33)

To bound the derivatives of w we use the fact that

 1  1
w(x) = Hw (s)ds + κ e−B(s) ds
x x

with
 x
1
Hw (x) = − (bw)(s)eB(s)−B(x) .
ε 0

Estimates for Hw are obtained using (3.33)


C x −βs/ε B(s)−B(x) C
|Hw (x)| ≤ e e ds ≤ exp(−βx/ε).
ε 0 ε

The coefficient κ is determined by the boundary condition for w(0):

  1 
1
κ= γ0 − v(0) − ϑw (s)ds ,
α 0
62 3 The Analytical Behaviour of Solutions

where
 1  1
ε
α= e−B(s) ds ≥ e− b ∞ s/ε
ds ≥ .
0 0 b∞

Thus

|κ| ≤ Cε−1 .

For w we have

w (x) = −Hw (x) − κe−B(x)

and therefore

|w (x)| ≤ Cε−1 e−βx/ε for x ∈ [0, 1],

by the above bounds for κ and Hw .

Use the differential equation and the estimates for w and w , to get

|w (x)| ≤ Cε−2 e−βx/ε for x ∈ [0, 1].

If b, c ∈ C 1 (0, 1) then differentiate (3.31b) once and apply the bounds for w, w
and w . Thus

|w (x)| ≤ Cε−3 e−βx/ε for x ∈ [0, 1].

We summarise the results.

Theorem 3.48. Let b, c, f ∈ C k [0, 1] with k ∈ {0, 1}. Then u ∈ C k+2 [0, 1] can be
decomposed as u = v +w, where the regular solution component v satisfies Lv = f
in (0, 1) and
 (i)   
v (x) ≤ C 1 + εk+1−i for i = 0, 1, . . . , k + 2, x ∈ [0, 1], (3.34a)

while the boundary layer component w satisfies Lw = 0 in (0, 1) and

 (i) 
w (x) ≤ Cε−i e−βx/ε for i = 0, 1, . . . , k + 2, x ∈ [0, 1]. (3.34b)

Remark 3.49. A similar decomposition is given in [30,121]. However, the construc-

tion there requires more smoothness of the data of the problem because the regular
solution component v is defined via solutions of first-order problems. ♣

Remark 3.50. Some applications, e. g., the analysis of higher-order schemes in

Sect. 4.3.1.3 or [153] or of extrapolation schemes [124], require decompositions
3.4 Convection-Diffusion Problems with Regular Layers 63

with bounds for higher-order derivatives. To derive those, note that the boundary
condition (−bv  + cv)(0) = f (0) imposed on v corresponds to v  (0) = 0. In order
to prove Theorem 3.48 for k = 2, we would impose the boundary condition
 
− (b − ε(b − c)) v  + (c − εc ) v (0) = (f − εf  ) (0)

instead. This corresponds to setting v  (0) = 0. The operator L with this boundary
condition satisfies a comparison principle too, provided that ε is smaller than a cer-
tain threshold value ε0 . This principle implies the boundedness of v. Then proceed
as above to get bounds for the derivatives. ♣

3.4.1.3 Discontinuous Coefficients and Point Sources

Consider the convection-diffusion problem in conservative form with a point source:

Lc u := −εu − (bu) + cu = f + αδd , in (0, 1),

(3.35)
u(0) = γ0 , u(1) = γ1 ,

where δd is the shifted Dirac-delta function δd (x) = δ(x − d) with d ∈ (0, 1). The
coefficient b may also have a discontinuity at x = d. Assume that b ≥ β1 > 0 in
(0, d) and b ≥ β2 > 0 in (d, 1). The argument follows [88].
Problem (3.35) has to be read in a distributional context. Alternatively, one may
seek a solution u ∈ C[0, 1] ∩ C 2 ((0, d) ∪ (d, 1)) with

Lc u = f in (0, 1) \ {d}, u(0) = γ0 , u(1) = γ1 , − (ε[u ] + [b]u) (d) = α,

where [v](d) := v(d + 0) − v(d − 0) is the jump of v at d.

The solution of (3.35) typically has an exponential boundary layer at the outflow
boundary x = 0 and an internal layer at x = d caused by the concentrated source or
the discontinuity of the convective field. Figure 2.5 depicts a typical solution.
Using the negative-norm stability (3.29c), we obtain u∞ ≤ C. Then, u solves

Lu = f in (0, d), u(0) = 0, u(d) = ρ,

and

Lu = f in (d, 1), u(d) = ρ, u(1) = 0

where ρ is uniformly bounded in ε, i.e., |ρ| ≤ C.

Next apply the results from Sect. 3.4.1.2 separately on each of the two subdo-
mains. We obtain
 (k)   
u (x) ≤ C 1 + ε−k e−β1 x/ε + Hd (x)e−β2 (x−d)/ε
(3.36)
for x ∈ (0, d) ∪ (d, 1)
64 3 The Analytical Behaviour of Solutions

where Hd denotes the shifted Heaviside function, i.e.,


0 for x < d,
Hd (x) =
1 for x > d.

Note that (3.36) holds for the one-sided derivatives at x = d too.

3.4.2 Weakly Coupled Systems of Convection-Diffusion Equations

We now leave the scalar convection-diffusion equation and move on to systems of

equations this type.
The system is said to be weakly coupled if the convective coupling matrix B is
diagonal, so the system is coupled only through the lower-order reaction terms. We
follow [96]. In one dimension such systems can be written as

Lu := − diag(ε)u − diag(b)u + Au = f on (0, 1),

(3.37)
u(0) = u(1) = 0,

where ε = (ε1 , . . . , ε )T and the small parameter εk is in (0, 1] for k = 1, . . . , .

For k = 1, . . . , , the k th equation of (3.37) is


−εk uk − bk uk + akj uj = fk on (0, 1), uk (0) = uk (1) = 0. (3.38)
j=1

Assume that for each k one has akk ≥ 0 and either bk ≥ βk or bk ≤ −βk
on [0, 1] with positive constants βk .

3.4.2.1 Stability

We follow the argument of [96, 111]. Rewrite (3.38) as


−εk uk − bk uk + akk uk = − akm um + fk . (3.39)
m=1
m=k

Then (3.28) yields

    
  fk   fk 
uk ∞ + γ̃km um ∞ 
≤ min   ,   for i = 1, . . . , ,
m=1
akk  ∞
 bk 
∞
m=k
3.4 Convection-Diffusion Problems with Regular Layers 65

where the  ×  constant matrix Γ̃ = Γ̃ (A, b) = (γ̃km ) is defined by

    
 akm   
γ̃kk = 1 and γ̃km = − min   ,  akm  for k = m.
 akk   bk 
∞ ∞

Repeating the analysis of Sect. 3.3.2.1, we reach the following stability result.

Theorem 3.51. Assume that the matrix A has non-negative diagonal entries. Sup-
pose also that all entries of A lie in C[0, 1]. Assume that Γ̃ (A, b) is inverse-
monotone. Then for k = 1, . . . ,  one has

     
−1   (Lv)m   (Lv)m 
vk ∞ ≤ Γ̃ 
min   
,  
m=1
km amm ∞ bm  ∞

 
for any function v = (v1 , . . . , v )T ∈ C 2 (0, 1) ∩ C[0, 1] with v(0) = v(1) = 0.
In this inequality the first term in min{. . . } should be omitted if it does not exist.

Corollary 3.52. Under the hypotheses of Theorem 3.51 the boundary value prob-
lem (3.37) has a unique solution u, and u∞ ≤ C f ∞ for some constant C.

Remark 3.53. When bounding the source term, one can use (3.29c) to establish that

u∞ ≤ C max fm −1,∞ .

m=1,...,

This inequality allows the right-hand side to be a generalised function like the
δ-distribution. ♣

3.4.2.2 Bounds on Derivatives

Applying the scalar-equation analysis of Sect. 3.4.1.2 to (3.39), it is shown in [96]

that for k = 1, . . . , , x ∈ [0, 1] and ν = 0, 1 one has
⎧
  ⎨1 + ε−ν
k e
−βk (1−x)/εk
if bk < 0,
 (ν) 
uk (x) ≤ C
⎩1 + ε−ν e−βk x/εk if b > 0.
k k

Thus, there are boundary layers in the solution at x = 0 and/or x = 1, but no

strong interaction between the layers in different components uk is apparent when
first-order derivatives are considered, which is quite unlike the reaction-diffusion
case of Theorem 3.43. These bounds on the uk also reveal a lack of sharpness in the
results of some papers published previously to [96] that gave derivative bounds for
this problem with stronger interactions between different components.
66 3 The Analytical Behaviour of Solutions

3.4.3 Systems of Strongly Coupled Convection-Diffusion

Equations

We now return to the general problem with strong coupling: this means that for
each k one has bkm ≡ 0 for some k = m. The analysis follows [100]. The boundary-
 
value problem under consideration is: Find u ∈ C 2 (0, 1) ∩ C[0, 1] such that

Lu := − diag(ε)u − Bu + Au = f in (0, 1),

(3.40)
u(0) = u(1) = 0,

where as before f : [0, 1] → IR is vector-valued, while A, B : [0, 1] → IR ,

are matrix-valued functions. The  ×  matrix diag(ε) is diagonal with k th entry
εk ∈ (0, 1] for all k.
Assume for each k there exists a positive constant βk such that

bkk ≥ βk or bkk ≤ −βk , on [0, 1]. (3.41a)

Furthermore let

akk ≥ 0 and bkk + akk ≥ 0 on [0, 1]. (3.41b)

Rewrite the k th equation of the system (3.40) as

Lk uk := −εk uk − bkk uk + akk uk

'  ( (3.42a)
= fk + (bkm um ) − (bkm + akm )um
m=1
m=k

with boundary conditions

uk (0) = uk (1) = 0. (3.42b)

Apply the stability bound of (3.29c) and Remark 3.22 to obtain


|||uk |||εk ,∞ ≤ {2 bkm ∞ + bkm + akm 1 } um ∞ + 2 fk −1,∞ ,
m=1
m=k

where
!
|||v|||εk ,∞ := max εk v  ∞ , βk v∞
3.4 Convection-Diffusion Problems with Regular Layers 67

Then, after some minor manipulation, gather the uj terms to the left-hand side in
the manner of Sect. 3.3.2.1. Define the  ×  matrix Υ = Υ (A, B) = (γkm ) by

2 bkm ∞ + bkm + akm 1

γkk = 1 and γkm = − for k = m.
βk

We get


|||uk |||εk ,∞ + γkm |||uj |||εm ,∞ ≤ 2 fk −1,∞ .
m=1
m=k

Theorem 3.54. Let B and A satisfy (3.41). Suppose Υ is inverse-monotone. Then

the operator L is (L∞ , W −1,∞ )-stable with

 
|||uk |||εk ,∞ ≤ 2 Υ −1 km
(Lu)m −1,∞ for k = 1, . . . , .
m=1

Corollary 3.55. Under the hypotheses of Theorem 3.54 the boundary value prob-
lem (3.40) possesses a unique solution u, and

|||u|||ε,∞ := max |||uk |||εk ,∞ ≤ C max fk −1,∞

k=1,..., k=1,...,

for some constant C. In view of Remark (3.22) this implies the bounds

u∞ ≤ Cf 1 and u∞ ≤ Cf ∞ .

3.4.3.1 Derivative Bounds

Application of (3.29b) to (3.42) gives

*  +
βk uk 1 ≤ 2 fk 1 + 
bkm ∞ um 1 + akm 1 um ∞ .
m=1
k=k

Recalling the definition of the matrix Υ , we get

 *  +
βk uk 1 + γkm βm um 1 ≤ 2 fk 1 + akm 1 um ∞
m=1 m=1
m=k m=k

≤ 2 fk 1 + C u∞ ,

by Theorem 3.54. Thus, if Υ is inverse monotone we get a uniform bound for the
L1 norm of u .
68 3 The Analytical Behaviour of Solutions

Theorem 3.56. Let B and A satisfy (3.41). Suppose Υ is inverse monotone. Then
the solution u of (3.40) satisfies u 1 ≤ C f 1 .
The L1 -norm bounds of Theorem 3.56 are sufficient to prove the existence of
a mesh that yields robust uniform convergence for a first-order upwind difference
scheme, see Sect. 4.4.2. However, in order to design layer-adapted meshes, a priori
pointwise bounds are required. These are typically obtained by a splitting of the
solution into layer parts and a regular part, that is independent of the perturbation
parameter and captures the behaviour of the solution away from any layers; see e.g.
Theorem 3.48.
To illustrate the difficulties in obtaining such a decomposition let us consider the
following example of (3.40):

−ε1 u1 + u1 − u2 = 1, u1 (0) = u1 (1) = 0,

(3.43)
−ε2 u2 − 3u2 = 2, u2 (0) = u2 (1) = 0.

Figure 3.3 displays its solution for fixed ε2 = 10−4 and three different values of ε1 .
Obviously there does not exist a regular solution component that is independent of
the perturbation parameters, nor can there (in general) be anything like the solution
of the reduced problem for strongly coupled convection-diffusion.
The special case where all diffusion coefficients εk are the same is studied
in [127]. A lengthy analysis with some further assumptions on the data of the prob-
lem leads to the following result:
Theorem 3.57. Suppose that εk = ε for k = 1, . . . , . Then the solution u of (3.40)
can be decomposed as u = v + w where
 (ν)   
v  ≤ C 1 + ε2−ν for ν = 0, 1, 2, 3,
∞

and for x ∈ [0, 1] one has

 (ν) 
w (x) ≤ Cε−ν e−βx/ε for ν = 0, 1, 2, 3 and k = 1, . . . , .
k
!
with β := min β1 , . . . , β .

1 1 1
0.9 0.9 0.9
0.8 0.8 0.8
0.7 0.7 0.7
0.6 0.6 0.6
0.5 0.5 0.5
0.4 0.4 0.4
0.3 0.3 0.3
0.2 0.2 0.2
0.1 0.1 0.1
0 0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

ε1 = 10−2 , ε2 = 10−4 ε1 = 10−4 , ε2 = 10−4 ε1 = 10−6 , ε2 = 10−4

Fig. 3.3 Solution of (3.43) for different values of ε (u2 dashed)

3.5 Convection-Diffusion Problems with Turning-Point Layers 69

3.5 Convection-Diffusion Problems with Turning-Point Layers

Turning point layers are associated with zeros of the convection coefficient. Let us
consider the boundary-value problem

−εu − pbu + c(·, u) = 0 in (q, 1), u(q) = γq , u(1) = γ1

with q ∈ {−1, 0}. We assume that b(x) ≥ β > 0, ∂u c ≥ 0 and sign p(x) = sign x
for x ∈ (q, 1). The assumption on p implies that the point x = 0 is a turning point.
If q = 0, then the turning point coincides with a boundary; we call this a boundary
turning point problem. When q = −1 we have an interior layer.
Turning point problems with ∂u c(0, ·) > 0 were considered in a number of pa-
pers in the 1980s. For interior turning point problems this additional assumption
implies that the solution of the reduced problem is continuous and therefore, no
strong layer is present. This means the problem is not singularly perturbed in the
maximum norm. For boundary turning points, the situation is different since the
solution of the reduced problem will in general not match the boundary condition
prescribed at the outflow boundary. However, if ∂u c(0, ·) > 0, then the dominat-
ing feature of the problem is the relation between the diffusion and reaction terms
and the problem has the character of a reaction-diffusion problem. Here we restrict
ourselves to the case ∂u c(0, ·) = 0.
We shall study the linear convection-diffusion problem

Lu := −εu − pbu + pcu = pf in (0, 1), u(0) = γ0 , u(1) = γ1 , (3.44)

where p, b, c, f ∈ C 1 [0, 1] with

p > 0, p > 0, b ≥ β > 0 and c ≥ 0 in (0, 1). (3.45)

Particular focus will be on the case p(x) = xκ , κ > 0.

3.5.1 Stability

Suppose (3.45) holds. Then using the test function ψ : x → ψ(x) = 1 − x, we

see that the differential operator L in (3.44) satisfies the hypotheses of Lemma 3.3,
because Lψ ≥ pb > 0 in (0, 1). Therefore, L obeys a comparison principle and
we have

v∞ ≤ Lv/pb∞ for all v ∈ C 2 [0, 1] with v(0) = v(1) = 0,

70 3 The Analytical Behaviour of Solutions

by Lemma 3.5. If c > 0 on [0, 1], then Lemma 3.5 with ψ ≡ 1 yields also

v∞ ≤ Lv/pc∞ for all v ∈ C 2 [0, 1] with v(0) = v(1) = 0.

Thus, L is (L∞ , L∞,1/p )-stable with a 1/p-weighted maximum norm.

Furthermore, we have for the solution of (3.44)
    
f  f 
|u|∞ ≤ max |γ0 |, |γ1 | + min    
b ,c .
∞ ∞

Green’s function estimates

Lemma 3.58. Assume (3.45) holds true. Then
⎧
⎪
1 ⎨1    for 0 ≤ x ≤ ξ ≤ 1,
0 ≤ G(x, ξ) ≤ Ĝ(x, ξ) := β x
p(ξ)β ⎪
⎩exp − ε p(s)ds for 0 ≤ ξ ≤ x ≤ 1.
ξ

Proof. First note that the operator L satisfies the assumptions of Lemma 3.11 which
we like to apply.
Let ξ ∈ (0, 1) be arbitrary, but fixed. Clearly Ĝ(0, ξ) ≥ 0 and Ĝ(1, ξ) ≥ 0.
Furthermore,

0 for 0 < x < ξ,
∂x Ĝ(x, ξ) = βp(x)
− ε Ĝ(x, ξ) for ξ < x < 1.

and
⎧
⎨0 for 0 ≤ x < ξ
∂x2 Ĝ(x, ξ) =  
⎩− βε βp(x)
ε − p 
(x) Ĝ(x, ξ) for ξ < x ≤ 1.

Thus,

LĜ(·, ξ) ≥ pcĜ(·, ξ) ≥ 0 in (0, ξ) ∪ (ξ, 1),

because p ≥ 0 on [0, 1]. The jump of ∂x Ĝ satisfies

 
−ε ∂x Ĝ(·, ξ) (ξ) = 1.

Thus Ĝ is a barrier function for G. Applying Lemma 3.11, we are finished. 


Remark 3.59. By Lemma 3.58
1
G(x, ξ) ≤ for all x, ξ ∈ (0, 1).
βp(ξ)
3.5 Convection-Diffusion Problems with Turning-Point Layers 71

However, numerical experiments indicate that when p(x) = xκ , κ ≥ 0, one has the
sharper bound
 −κ
G(x, ξ) ≤ C ε1/(κ+1) + ξ for all x, ξ ∈ (0, 1).

But, we do not have a rigorous proof for this. ♣

Theorem 3.60. Suppose (3.45) holds true. Then the operator L in (3.44) satisfies

v∞ ≤ β −1 (Lv)/p1 for all v ∈ W̊ 1,1 (0, 1) ∩ W 2,1 (0, 1).

Thus, the operator L is (L∞ , L1;1/p ) with a 1/p-weighted L1 norm.

Proof. Use the representation (3.1) and Lemma 3.58 to obtain the assertion of the
theorem. 


Corollary 3.61. Suppose (3.45) holds true. Then the solution u of (3.44) satisfies

v∞ ≤ max |γ0 |, |γ1 | + β −1 f 1 .

Remark 3.62. In view of the a posteriori error analysis for problems with regular
layers (see Sect. 4.2.4) stronger negative-norm stability is desirable. Unfortunately,
this is not available yet. ♣

3.5.2 Derivative Bounds and Solution Decomposition

We follow [92] to derive a decomposition of u into a regular solution component v

and a layer part w for p(x) = xκ with κ ≥ 0. In doing so, we generalise the analysis
of Sect. 3.4.1.2 for problems with regular boundary layers.

Theorem 3.63. Let b, c, f ∈ C 1 [0, 1] and p(x) = xκ , κ ≥ 0. Assume b > β̃ > 0

and c ≥ 0 on [0, 1]. Then (3.44) has a unique solution u ∈ C 3 [0, 1] and this solution
can be decomposed as u = v + w, where the regular solution component v satisfies
Lv = f in (0, 1),
       
v  + v   + v   ≤ C and εv  (x) ≤ Cxκ for x ∈ [0, 1],
∞ ∞ ∞

while the boundary layer component w satisfies Lw = 0 in (0, 1) and

 
 (i)  κ+1
w (x) ≤ Cμ−i exp − β̃x for i = 0, 1, 2, 3, x ∈ (0, 1),
ε(κ + 1)

where μ = ε1/(κ+1) .
72 3 The Analytical Behaviour of Solutions

Remark 3.64. For κ = 0, i.e. in the case of a regular boundary layer, this result is
known from Theorem 3.48. ♣
Remark 3.65. In [92] the semilinear problem

−εu − pbu + pc(·, u) = 0 in (0, 1), u(0) = γ0 , u(1) = γ1 ,

with p(x) = xκ and ∂u c ≥ 0 on [0, 1] × IR is analysed. The same bounds on

derivatives are obtained. ♣
Proof (of Theorem 3.63). The construction of the decomposition is similar to the
one in Sect. 3.4.1.2. Set Bz := −bz  + cz, Let v and w be the solutions of the
boundary-value problems

Lv = pf in (0, 1), (Bv) (0) = f (0), v(1) = γ1

and
Lw = 0 in (0, 1), w(0) = γ0 − v(0), w(1) = 0.

Preliminaries
Before starting the main argument, let us provide some auxiliary results. Set

1 x p
B(x) := s b(s)ds
ε 0

and let β ∗ with b(x) ≥ β ∗ > 0 be arbitrary.

In the analysis bounds for certain integral expressions involving B are required.
First,

β ∗ sκ+1 − xκ+1
B(s) − B(x) ≤ for 0 ≤ s ≤ x ≤ 1. (3.46)
ε κ+1

From this, for arbitrary ν ≥ 0 we get


β ∗ x (κ+ν)
s exp(B(s) − B(x))ds
ε 0
  ∗ κ+1  (3.47)
β∗ x κ β s − xκ+1
≤ s exp ds ≤ 1.
ε 0 ε κ+1

We shall also use

 1   
1
b∞ sκ+1
exp(−B(s))ds ≥ exp − ds
0 0 (κ + 1)ε
 1/μ    1  
b∞ tκ+1 a∞ tκ+1
=μ exp − dt ≥ μ exp − dt = Cμ.
0 (κ + 1) 0 (κ + 1)
(3.48)
3.5 Convection-Diffusion Problems with Turning-Point Layers 73

Proposition 3.66. For arbitrary ν > 0 there exists a constant C = C(ν) such that
  ∗ ν+1 
xν x β s − xν+1
exp ds ≤ C for all x ≥ 0, ε > 0.
ε 0 ε ν+1

Proof. Using the transformations

 1/(ν+1)  1/(ν+1)
x = ε(ν + 1)t/β ∗ and s = ε(ν + 1)σ/β ∗ ,

we see that
  
β ∗ xν x
β ∗ sν+1 − xν+1
exp ds
ε 0 ε ν+1
 t
−t ν/(ν+1)
=e t eσ σ −ν/(ν+1) dσ =: Fν (t).
0

Clearly Fν ∈ C 0 [0, ∞) and Fν (0) = 0 for ν > 0. Moreover, limt→∞ Fν (t) = 1.

Thus, there exists a constant C(ν) > 0 such that Fν ∞ ≤ C(ν). 


The regular solution component

The operator L satisfies the following comparison principle: For any two functions
z, z̄ ∈ C[0, 1] ∩ C 1 [0, 1) ∩ C 2 (0, 1)
⎫
Lz ≤ Lz̄ in (0, 1), ⎬
(Bz) (0) ≤ (Bz̄) (0), =⇒ z ≤ z̄ on [0, 1].
⎭
z(1) ≤ z̄(1)

Application of this principle with

 
v ± (x) := ± β −1 (1 − x) f ∞ + γ1

yields

v∞ ≤ C.

Now let us bound the derivatives of v. The function v can be written as

 1  1
hv (0)
v(x) = Hv (s)ds + exp(−B(s))ds + γ1 ,
x b(0) x

where hv := f − cv and

1 x  
Hv (x) = sκ hv (s) exp B(s) − B(x) ds for x ∈ [0, 1].
ε 0
74 3 The Analytical Behaviour of Solutions

From this representation we immediately get


 1 x   hv (0)  
v (x) = − sκ hv (s) exp B(s) − B(x) ds − exp −B(x) . (3.49)
ε 0 b(0)

Hence,

v   ≤ C,

by (3.47).
Differentiating (3.49) once and using integration by parts, we get
  
xκ b(x) x hv  
v  (x) = − (s) exp B(s) − B(x) ds.
ε 0 b

Therefore,

   κ x  
v (x) ≤ C x exp B(s) − B(x) ds
ε 0
  
xκ x
β sκ+1 − xκ+1
≤C exp ds
ε 0 ε κ+1

and

|v  | ≤ C.

by Prop. 3.66.
A bound for the third-order derivative is obtained from the differential equation
and the bounds on v  and v  :
   
−εv  = xκ bv + hv + κxκ−1 bv  + hv .

Let F (x) := bv  + hv . Eq. (3.31a) implies F (0) = 0. On the other hand, we have
  
|F  (x)| =  bv  + hv (x) ≤ C,
 
by our earlier bounds for v, v  and v  . Thus, F (x) ≤ Cx. We get
 
εv  (x) ≤ Cxκ for x ∈ (0, 1).

This completes our analysis of the regular part of u.

3.5 Convection-Diffusion Problems with Turning-Point Layers 75

The boundary layer component

Let βi be arbitrary but fixed constants with

min b(x) = β1 > β2 > β3 > β̃ > 0.

x∈[0,1]

The operator L with Dirichlet boundary conditions satisfies a comparison principle.

With the barrier functions
 
β1 xp+1
w± (x) := ±|γ0 − v(0)| exp − ,
ε p+1

it gives
 
  κ+1
w(x) ≤ C exp − β1 x for x ∈ (0, 1). (3.50)
ε κ+1

To bound the derivatives of w use the representation

 1
 v(0) − γ + Hw (s)ds  1
1 0  
w(x) = Hw (s)ds −  1 0
exp −B(s) ds,
x x
exp(−B(s))ds
0

where

1 x    
Hw (x) = − sκ cw (s) exp B(s) − B(x) ds.
ε 0

Thus,
 1
v(0) − γ0 + Hw (s)ds  
w (x) = −Hw (x) +  1 0
exp −B(x) . (3.51)
 
exp −B(s) ds
0

Using by (3.50) and then (3.46) with β ∗ = β1 , we obtain

   
  κ+1
β1 xκ+1 β2 xκ+1
Hw (x) ≤ C x exp − ≤ C exp − . (3.52)
ε ε κ+1 ε κ+1

From (3.47), (3.48), (3.51) and (3.52) we get

 
   κ+1
w (x) ≤ Cμ−1 exp − β2 x for x ∈ (0, 1).
ε(κ + 1)
76 3 The Analytical Behaviour of Solutions

Use the differential equation, the estimates for w and w and Prop. 3.66 to get
 
   κ+1
w (x) ≤ Cμ−2 exp − β3 x for x ∈ (0, 1).
ε(κ + 1)

Differentiate (3.31b), apply the bounds for w, w and w and use Prop. 3.66 again
in order to get
 
   β̃xκ+1
w (x) ≤ Cμ exp −
−3
for x ∈ (0, 1).
ε(κ + 1)

This completes the proof of Theorem 3.63. 


Chapter 4
Finite Difference Schemes
for Convection-Diffusion Problems

This chapter is concerned with finite-difference discretisations of the stationary lin-

ear convection-diffusion problem

Lu := −εu − bu + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1 , (4.1)

with b ≥ β > 0 on [0, 1]. For the sake of simplicity we shall assume that

c ≥ 0 and b ≥ 0 on [0, 1]. (4.2)

Using (4.1) as a model problem, a general convergence theory for certain first-
and second-order upwinded difference schemes on arbitrary and on layer-adapted
meshes is derived. The close relationship between the differential operator and its
upwinded discretisations is highlighted.

4.1 Notation

Meshes and mesh functions

Throughout this chapter let ω̄ : 0 = x0 < x1 < · · · < xN = 1 be an arbitrary
partition of [0, 1] with mesh intervals Ii := [xi−1 , xi ]. The set of inner mesh points
is denoted by ω. The midpoint of Ii is xi−1/2 := (xi − xi+1 )/2 and its length
hi := xi − xi−1 . Let h := max hi be the maximal mesh size.
i=1,...,N
We shall identify mesh functions v : ω̄ → IR : xi → vi with vectors v ∈ IRN +1
and with spline functions

v ∈ V ω := S10 (ω̄) := w ∈ C 0 [0, 1] : w|Ii ∈ Π1 for i = 1, . . . , N .

!
Let IR0N +1 := v ∈ IRN +1 : v0 = vN = 0 be the space of mesh functions that
vanish at the boundary. Furthermore, V0ω := S10 (ω̄) ∩ H01 (0, 1).

T. Linß, Layer-Adapted Meshes for Reaction-Convection-Diffusion Problems, Lecture 77

Notes in Mathematics 1985, DOI 10.1007/978-3-642-05134-0 4,
c Springer-Verlag Berlin Heidelberg 2010
78 4 Finite Difference Schemes for Convection-Diffusion Problems

Difference operators
In our notation of difference operators we follow Samarski’s text book [146]. For
any mesh function v ∈ IRN +1 set
vi+1 − vi vi − vi−1 vi − vi−1
vx;i := , vx̄;i := vx;i−1 = , vx̆;i :=
hi+1 hi hi+1
vi+1 − vi vi − vi−1 vi+1 − vi−1
vx̂;i := , vx̌;i := , vx̊;i :=
h̄i h̄i 2h̄i

with the weighted mesh increment h̄ defined by

h1 hi + hi+1 hN
h̄0 := , h̄i := , i = 1, . . . , N − 1, and h̄N := .
2 2 2

To simplify the notation we set gi := g(xi ) for any g ∈ C[0, 1].

Further, less frequently used, difference operators will be introduced when
needed.

Discrete norms and inner products

For any mesh function v ∈ IRN +1 define the ∞ (semi-)norms

v∞,ω := max |vi | , v∞,ω̄ := max |vi | ,

i=1,...,N −1 i=0,...,N

|[v||∞,ω := max |vi | , |||v|||ε,∞,ω := max ε |[vx ||∞,ω , βv∞,ω̄ ,

i=0,...,N −1

the 1 norm


N −1
v1,ω := hj+1 |vj |
j=0

and the w−1,∞ norm

N −1 
 
v−1,∞,ω := min V ∞,ω̄ = min  hj+1 vj + c .
V :Vx =v c∈IR 
j=· ∞,ω̄

We shall also use the following discrete inner products:


N −1 
N −1
[w, v)ω := hi+1 wi vi and (w, v)ω := hi+1 wi vi .
i=0 i=1
4.2 A Simple Upwind Difference Scheme 79

4.2 A Simple Upwind Difference Scheme

In this section we study a first-order difference scheme for the discretisation of (4.1)
on arbitrary meshes. Find uN ∈ IRN +1 such that
' (
LuN i
= fi for i = 1, . . . , N − 1, uN N
0 = γ0 , uN = γ1 (4.3)

with

[Lv]i := −εvx̄x;i − bi vx;i + ci vi for v ∈ IRN +1 .

At first glance the discretisation of the second-order derivative is a bit non-

standard, because on non-uniform meshes it is not consistent in the maximum norm,
but it has advantages that become clearer in the course of our analysis. More fre-
quently used is the central difference approximation ui ≈ ux̄x̂;i . An upwind
scheme based on this discretisation of the second-order derivative will be studied
in Sect. 4.2.6, because the technique used there becomes more important in 2D, see
Sect. 9.1.
The difference scheme (4.3) can be generated by a finite-element approach. To
this end consider (4.1) with homogeneous boundary conditions. Its weak formula-
tion is: Find u ∈ H01 (0, 1) such that

a(u, v) = f (v) for all v ∈ H01 (0, 1)

with

a(w, v) := ε (w , v  ) − (bw − cw, v) , f (v) := (f, v)

1 
and the L2 (0, 1)-scalar product (w, v) := 0 wv (s)ds.
A standard FEM approximation is: Find uN ∈ V0ω such that

a(uN , v) = f (v) for all v ∈ V0ω .

The integrals in the bilinear form and the linear functional have to approximated.
Use the left-sided rectangle rule Ii g(s)ds ≈ hi gi−1 , to arrive at: Find uN ∈ V0ω
such that
   
au uN , v = fu v for all v ∈ V0ω , (4.4)

where

au (w, v) := ε [wx , vx )ω − (bwx − cw, v)ω and fu (v) := (f, v)ω .

80 4 Finite Difference Schemes for Convection-Diffusion Problems

Taking as test functions v the standard hat-function basis in V0ω , we see that (4.3)
and (4.4) are equivalent. In particular,

au (w, v) = (Lw, v)ω = (w, L∗ v)ω for all w, v ∈ V0ω ,

with the adjoint operator L∗ given by

[L∗ v]j = −εvξ̄ξ;j + (bv)ξ̆;j + cj vj .

This is verified using summation by parts; cf. [146].

4.2.1 Stability of the Discrete Operator

The matrix associated with the difference operator L is a L0 -matrix because all
off-diagonal entries are non-positive. Application of the M -criterion (Lemma 3.14)
with a test vector with components ei = 2−xi , i = 0, . . . , N establishes the inverse
monotonicity of L. Thus, L satisfies a comparison principle: For any mesh functions
v, w ∈ IRN +1
⎫
Lv ≤ Lw on ω, ⎬
v0 ≤ w0 , =⇒ v ≤ w on ω̄. (4.5)
⎭
vN ≤ wN

This comparison principle and Lemma 3.17 give the stability inequality
 N !
ui  ≤ max |γ0 |, |γ1 | + (1 − xi ) f /b for i = 0, . . . , N.
∞,ω

Thus, the operator L is (∞ , ∞ )-stable with

v∞,ω̄ ≤ Lv/b∞,ω for all v ∈ IR0N +1 .

Alternatively, if c > 0 on [0, 1], then Lemma 3.17 with e ≡ 1 yields

v∞,ω̄ ≤ Lv/c∞,ω for all v ∈ IR0N +1 .

Note the analogy with (3.28).

Green’s function estimates

Using the discrete Green’s function G : ω̄ 2 → IR : (xi , ξj ) → Gi,j = G(xi , ξj )
associated with L and Dirichlet boundary conditions, any mesh function v ∈ IR0N +1
can be represented as

vi = au (v, Gi,· ) = (Lv, Gi,· )ω = (v, L∗ Gi,· )ω for i = 1, . . . , N − 1.

4.2 A Simple Upwind Difference Scheme 81

Taking for v the standard basis in V0ω , we see that for fixed i = 1, . . . , N − 1

[L∗ Gi,· ]j = δi,j for j = 1, . . . , N − 1, Gi,0 = Gi,N = 0, (4.6)

where

h−1
i+1 if i = j,
δi,j :=
0 otherwise,

is a discrete equivalent of the Dirac-δ distribution. As function of the first argument

G solves, for fixed j = 1, . . . , N − 1,

[LG·,j ]i = δi,j for i = 1, . . . , N − 1, G0,j = GN,j = 0.

Theorem 4.1. Suppose (4.2) holds true. Then the Green’s function G associated
with the discrete operator L and Dirichlet boundary conditions satisfies
⎧
⎪
⎪1 for 0 ≤ i ≤ j ≤ N,
1⎨ 3 i  −1
0 ≤ Gi,j ≤ βhk+1
β⎪⎪ 1+ for 0 ≤ j < i ≤ N,
⎩ ε
k=j+1

Gx;i,j ≤ 0, Gξ;i,j ≥ 0 for 0 ≤ j < i < N,

Gx;i,j ≥ 0, Gξ;i,j ≤ 0 for 0 ≤ i ≤ j < N,
Gxξ;ij ≤ 0 for 0 ≤ i, j < N, i = j

and

1
0 ≤ Gxξ;ii ≤ for i = 0, . . . , N − 1.
εhi+1

Proof. The upper and lower bounds on G are verified using (4.5).
Since G ≥ 0 on ω̄ 2 and Gi,0 = 0 for i = 0, . . . , N , we have Gξ;i,0 ≥ 0 for
i = 0, . . . , N . By multiplying (4.6) by hj+1 and summing over j, we get

j
−εGξ;i,j + εGξ;i,0 + bj Gi,j = − hk+1 ck Gi,k for j = 1, . . . , i − 1.
k=1

Hence,

εGξ;i,j ≥ εGξ;i,0 + bj Gi,j ≥ 0 for j = 1, . . . , i − 1,

82 4 Finite Difference Schemes for Convection-Diffusion Problems

since Gi,j ≥ 0 and Gξ;i,0 ≥ 0. On the other hand, Gξ;i,N −1 ≤ 0 for i = 0, . . . , N

because G ≥ 0 on ω̄ 2 and Gi,N = 0 for i = 0, . . . , N . By inspecting the difference
equation (4.6), we see that vj := Gξ;i,j satisfies, for i < j < N ,

ε hj bj−1  
− (vj − vj−1 ) + vj−1 = − bξ̆;j + cj Gi,j ≤ 0, (4.7)
hj+1 hj+1

by (4.2). Since vN −1 ≤ 0, induction for decreasing j yields Gξ;i,j = vj ≤ 0 for

i ≤ j < N.
Similarly, one can prove that Gx;i,j ≥ 0 for 0 ≤ i < j − 1 and Gx;i,j ≤ 0 for
j ≤ j < N . Thus,

Gxξ;i,0 ≤ 0 and Gxξ;i,N −1 ≤ 0 for i = 0, . . . , N − 1.

because Gx;i,0 = Gx;i,N = 0 for 0 ≤ i < N . Taking differences of (4.6) with

respect to i and summing over j, we get


j
−εGxξ;i,j + εGxξ;i,0 + bj Gx;i,j + hk+1 ck Gx;i,k = −δi,j for 0 < j ≤ i.
k=1

Therefore,
1
Gxξ;i,j ≤ 0 for 0≤j<i<N and Gxξ;i,i ≤ for 0 ≤ i < N
εhi+1

because Gx;i,j ≥ 0, Gxξ;i,0 ≤ 0 and Gx;i,0 = 0.

For i < j, take differences of (4.7) to see that vj = Gxξ;i,j satisfies

ε hj bj−1  
− (vj − vj−1 ) + vj−1 = − bξ̆;j + cj Gx;i,j ≤ 0
hj+1 hj+1
for j = i + 2, . . . , N − 1.

because b , c ≥ 0 and Gx;i,j ≥ 0 for i < j. We get Gxξ;i,j ≤ 0 for 0 ≤ i < j < N .
Finally, for i = j, use


N −1
hj+1 Gxξ;i,j = Gx;i,N − Gx;i,0 = 0
j=0

in order to obtain hi+1 Gxξ;i,i ≥ 0. 


Mimicking the arguments of Theorem 3.23 we obtain its discrete counterpart.
Theorem 4.2. Suppose (4.2) holds true. Then the Green’s function G associated
with the discrete operator L satisfies
1 2 2
Gi,· 1,ω ≤ , Gξ;i,· 1,ω ≤ , Gx;·,j 1,ω ≤
β β β
4.2 A Simple Upwind Difference Scheme 83

and

2
Gxξ;i,· 1,ω ≤ .
ε
for all i, j = 1, . . . , N − 1.
The 1 -norms bounds are used to establish stability properties for L that resemble
those of Theorem 3.45 for the differential operator L.
Theorem 4.3. Suppose (4.2) holds true. Then the operator L satisfies

v∞,ω ≤ min Lv/b∞,ω , Lv/c∞,ω , (4.8a)

−1 −1
v∞,ω ≤ β Lv1,ω , vx 1,ω ≤ 2β Lv1,ω (4.8b)

and

|||v|||ε,∞,ω ≤ 2Lv−1,∞,ω (4.8c)

for all v ∈ IR0N +1 .

Remark 4.4. Similar to Remark 3.22 we have

2 v−1,∞,ω ≤ v1,ω ≤ v∞,ω .

Therefore, the (∞ , w−1,∞ )-stability (4.8c) is the strongest of the three stability in-
equalities of Theorem 4.3. It was first derived by Andreev and Kopteva [11], though
their derivation is different. A systematic approach can be found in [5], where stabil-
ity of both the continuous operator L and of its discrete counterpart L is investigated.
So far the (∞ , w−1,∞ )-stability inequality gives the sharpest error bounds for one-
dimensional problems. But unlike the (∞ , 1 ) stability, it is unclear whether it can
be generalised to higher dimensions. ♣

Remark 4.5. The same stability results hold true if the convection-diffusion problem
in conservative form

Lc u := −εu − (bu) + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1 , (4.9)

is discretised by
' c N(  N
L u i := −εuN
x̄x;i − bu x;i
+ ci uN
i = fi for i = 1, . . . , N − 1,
(4.10)
uN
0 = γ0 , uN
N = γ1 .

♣
84 4 Finite Difference Schemes for Convection-Diffusion Problems

Remark 4.6. The (∞ , 1 ) stability (4.8b) was first given by Andreev and Savin
[12] for a modification of Samarskii’s scheme [144]. It has been used in a number
of publications to establish uniform convergence on S-type and B-type meshes; see,
e. g., [10,12,106,154]. Details of a convergence analysis can be found in Sect. 4.2.5.
This stability result can be generalised to study two-dimensional problems; see
Sect. 9.3.2. ♣

Corollary 4.7. By Theorem (4.3) there exists a unique solution of (4.3) and it
satisfies
 N
u  ≤ min f /b∞,ω , f /c∞,ω .
∞,ω̄

4.2.2 A Priori Error Bounds

Let us consider the approximation error of the simple upwind scheme (4.3) applied
to the boundary value problem (4.1). We give a convergence analysis based on the
negative-norm stability of Theorem 4.3.
Introduce the continuous and discrete operators and functions
 
    1  1
Av (x) := εv  (x) + bv (x) + (b + c)v (s)ds, F(x) := f (s)ds
x x

and


N −1 
N −1
[Av]i := εvx̄;i + bi vi + hk+1 (bx;k vk+1 + ck vk ) , Fi := hk+1 fk .
k=i k=i

Note that Lv = −(Av) and f = −F  on (0, 1), and Lv = −(Av)x and f = −Fx
on ω. Thus,

Au − F ≡ α on (0, 1) and AuN − F ≡ a on ω (4.11)

with constants α and a.

In view of the stability inequality (4.8c) we have
        
u − uN  ≤ 2 L u − uN −1,∞,ω = 2 min A u − uN + c∞,ω .
ε,∞,ω c∈IR

Taking c = a − α, where a and α are the constants from (4.11), we get

 
u − uN  ≤ 2Au − Au − F + F∞,ω . (4.12)
ε,∞,ω
4.2 A Simple Upwind Difference Scheme 85

Furthermore,

(Au − Au − F + F)i

N −1  xN

= ε (ux̄ − u )i + hk+1 bx;k uk+1 − (b u) (x)dx
k=i xi (4.13)

N −1  xN
+ hk+1 (ck uk − fk ) − (cu − f ) (x)dx.
k=i xi

Taylor expansions with the integral form of the remainder give

   xk
 
hk+1 (ck uk − fk ) − (cu − f ) (x)dx = cu − f (s)dsdx,
Ik+1 Ik+1 x
   xk+1
 
hk+1 bx;k uk+1 − (b u) (x)dx = b (x) u (s)dsdx
Ik+1 Ik+1 x

and
    x
 ε xk
 1 
ε (ux̄ − u )k = u (s)dsdx = bu − cu + f (s)dsdx,
hk Ik x hk Ik xk

by (4.1). Combining these representations with (4.12) and (4.13) we get the follow-
ing general convergence result.
Theorem 4.8. Let u be the solution of (4.1) and uN that of (4.3). Then

 
u − uN  ≤ 2 max (C1 |u (x)| + C2 |u(x)| + C3 ) dx
ε,∞,ω k=1,...,N Ik

with the constants

C1 := c∞ + b ∞ + b∞ , C2 := c∞ + c ∞

and

C3 := f ∞ + f  ∞ .

Remark 4.9. A similar result is given in [85] for the discretisation of the conser-
vative form (4.9). When using the conservative form, the last two terms in (4.13)
which involve bx and b disappear. ♣

Corollary 4.10. Theorem 4.8 and the a priori bounds (3.30) yield
 
u − uN  ≤ Cϑcd (ω̄),
[1]
ε,∞,ω
86 4 Finite Difference Schemes for Convection-Diffusion Problems

[p]
where the characteristic quantity ϑcd (ω̄) has been defined on p. 6:
  
1 + ε−1 e−βs/pε ds.
[p]
ϑcd (ω̄) := max
i=1,...,N Ii

Remark 4.11. The mesh function uN can be extended to a piecewise linear function
on the mesh ω̄. For convenience we denote this extended function by uN also. Then
 
u − uN  [1]
≤ Cϑcd (ω̄)
ε,∞

follows from a triangle inequality and our bounds for the interpolation error. ♣

Remark 4.12. Corollary 4.10 allows to immediate deduce (almost) first-order uni-
form convergence for particular meshes. Suppose the mesh parameter σ in the
definition of the meshes (see Sect. 2.1) satisfies σ ≥ 1. Then
⎧
⎪
⎪CN −1
⎪
⎨   for Bakhvalov meshes,
 
u − uN  ≤ C h + N −1 max |ψ  (ξ)| for S-type meshes and
ε,∞ ⎪
⎪ ξ∈[0,q]
⎪
⎩
CN −1 ln N for Shishkin meshes,

by (2.6) and (2.9). ♣

A numerical example
Table 4.1 displays numerical results for the upwind scheme (4.3) on a Bakhvalov
mesh applied to the test problem

−εu − u + 2u = ex−1 , u(0) = u(1) = 0. (4.14)

Table 4.1 Simple upwinding on a Bakhvalov mesh (q = 1/2)

σ = 0.2 σ = 0.4 σ = 0.6 σ = 0.8 σ = 1.0
N error rate error rate error rate error rate error rate
27 2.246e-2 0.23 1.173e-2 0.47 6.856e-3 0.69 4.658e-3 0.87 3.995e-3 0.97
28 1.913e-2 0.22 8.482e-3 0.45 4.258e-3 0.68 2.547e-3 0.88 2.036e-3 0.98
29 1.641e-2 0.21 6.201e-3 0.44 2.662e-3 0.67 1.388e-3 0.87 1.030e-3 0.99
210 1.416e-2 0.21 4.576e-3 0.43 1.675e-3 0.66 7.586e-4 0.87 5.193e-4 0.99
211 1.224e-2 0.20 3.403e-3 0.42 1.062e-3 0.65 4.155e-4 0.87 2.611e-4 0.99
212 1.063e-2 0.20 2.545e-3 0.41 6.784e-4 0.64 2.281e-4 0.86 1.310e-4 1.00
213 9.226e-3 0.20 1.911e-3 0.41 4.361e-4 0.63 1.256e-4 0.86 6.568e-5 1.00
214 8.030e-3 0.20 1.439e-3 0.41 2.819e-4 0.62 6.937e-5 0.85 3.290e-5 1.00
215 6.969e-3 0.21 1.086e-3 0.40 1.830e-4 0.62 3.846e-5 0.85 1.647e-5 1.00
216 6.026e-3 — 8.207e-4 — 1.193e-4 — 2.139e-5 — 8.245e-6 —
4.2 A Simple Upwind Difference Scheme 87

In our computations we have fixed the parameter q and varied σ to illustrate the
sharpness of our theoretical results. The errors are measured in the discrete max-
imum norm  · ∞,ω̄ . Apparently, choosing σ < 1 adversely affects the order of
convergence. Similar observations can be made for the Shishkin mesh and other
meshes.

4.2.3 Error Expansion

In the previous section we have seen that the error of the simple upwind scheme (4.3)
satisfies
 
u − uN  [1]
≤ Cϑcd (ω̄).
ε,∞,ω

Now an expansion of the error of this scheme is constructed. We shall show there
exists a function ψ, the leading term of the error, such that

u − uN = ψ + second order terms.

This result can be applied to analyse, e.g., derivative approximations, defect correc-
tion and Richardson extrapolation, see Sect. 4.2.9 and 4.3.3.
For the sake of simplicity, we will study the conservative form (4.9), i. e.,

Lc u := −εu − (bu) + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1 ,

and its discretisation by (4.10):

' c N(  N
L u i := −εuN x̄x;i − bu x;i
+ ci uN
i = fi for i = 1, . . . , N − 1,
uN
0 = γ0 , uN
N = γ1 .

Corresponding to Sect. 4.2.2 we introduce


  1 
(Ac v) (x) := εv  (x) + bv (x) + cv (s)ds
x

and


N −1
[Ac v]i := εvx̄;i + bi vi + hk+1 ck vk .
k=i

Note that Lc v = −(Ac v) on (0, 1) and that Lc v = −(Ac v)x on ω.

88 4 Finite Difference Schemes for Convection-Diffusion Problems

4.2.3.1 Construction of the Error Expansion

We define the leading term of the error expansion as the solution of

Lc ψ = Ψ  in (0, 1), ψ(0) = ψ(1) = 0, (4.15)

where
 1
h(x)  
Ψ (x) = ε u (x) − hg  (s)ds,
2 x

with
h(x) = x − xk−1 for x ∈ (xk−1 , xk ) and g = f − cu.

Note that Ψ is discontinuous at the mesh nodes. Hence, Lc ψ is a generalised

function. Therefore, (4.15) has to be interpreted in the context of distributions. Al-
ternatively, one may seek a solution ψ ∈ C 2 ((0, 1) \ ω) ∩ C[0, 1] such that

Lc ψ = Ψ  in (0, 1) \ ω, ψ(0) = ψ(1) = 0,

and

hi 
−ε[ψ  ](xi ) = [Ψ ](xi ) = −ε u (xi ) for xi ∈ ω.
2

Since Ac ψ = −Ψ on (0, 1) \ ω, we have

−1 
' c (  
 N 1
A ψ i = ε ψx̄;i − ψi−0 + hk+1 ck ψk − (cψ)(s)ds + Ψi−0 .
k=i xi

Thus,
 
' c ( hi  
A u − ψ − uN i = ε ux̄;i − ui + ui − ε ψx̄;i − ψi−0

2
 1  −1
  N
+ g − hg  (x)dx − hk+1 gk (4.16)
xi k=i

N −1  1
− hk+1 ck ψk + (cψ)(s)ds.
k=i xi

The function ψ has been designed such that the terms on the right-hand side that
involve u are of second order. Those involving ψ are formally only first-order terms,
but second order is gained since ψ itself is first order.
4.2 A Simple Upwind Difference Scheme 89

In order to bound the terms on the right-hand side, bounds for the derivatives of
u up to order three are needed. These are provided by (3.30). The following theorem
gives bounds for the leading term ψ of the error expansion and its derivatives up to
order two, which are also required. Because of the number of technical details, its
proof is deferred to the end of this section.
Lemma 4.13. Let ψ be the solution of the boundary-value problem (4.15). Assume
that b, c ∈ C 2 [0, 1] and f ∈ C 1 [0, 1]. Then
   
 (k) 
ψ (x) ≤ Cϑcd (ω̄) 1 + ε−k e−βx/2ε
[2]
for x ∈ (0, 1) \ ω, k = 0, 1,
(4.17a)

and
 
ε |ψ  (x)| ≤ Cϑcd (ω̄) 1 + ε−1 e−βx/2ε
[2]
for x ∈ (0, 1) \ ω. (4.17b)

Later we shall also show that (3.30), (4.16) and Lemma 4.13 yield

 c   2
A (u − ψ − uN ) ≤ C ϑ
[2]
(ω̄) . (4.18)
∞,ω cd

Then Theorem 4.3 yields our main result of this section.

Theorem 4.14. Let u, uN and ψ be the solutions of (4.9), (4.10) and (4.15), re-
spectively. Assume that b, c, f ∈ C 2 [0, 1]. Then
   2
u − ψ − uN  [2]
≤ C ϑcd (ω̄) .
ε,∞,ω

4.2.3.2 Detailed Proofs

Proof of Lemma 4.13

Now we derive bounds for the derivatives of the leading term ψ in the error
expansion. The following auxiliary result will be used several times in the subse-
quent analysis.
Proposition 4.15. Let x ∈ (xk−1 , xk ) and σ > 0 be arbitrary. Then
   x  
h(x) 1 + ε−1 e−βx/σε ≤ 1 + ε−1 e−βs/σε ds
xk−1

Proof. Let
   x  
−1 −βx/σε
F (x) := h(x) 1 + ε e and G(x) := 1 + ε−1 e−βs/σε ds.
xk−1
90 4 Finite Difference Schemes for Convection-Diffusion Problems

Clearly F (xk−1 ) = G(xk−1 ) = 0 and

h(x)β −βx/σε
F  (x) = 1 + ε−1 e−βx/ε − e ≤ 1 + ε−1 e−βx/σε = G (x)
σε2

for x ∈ (xk−1 , xk ). The result follows. 


First (3.30) implies
   1  
|Ψ (x)| ≤ Cεh(x) 1 + ε−2 e−βx/ε + C h(s) 1 + ε−1 e−βs/ε ds.
x

This inequality, (3.29c) and Prop. 4.15 yield (4.17a) for k = 0.

Next, we derive bounds on ψ  . Set

1 x
B(x) := b(s)ds, a(x) := Ψ  (x) + (c − b )(x)ψ(x)
ε 0

and
 x
1
χ(x) := a(s)eB(s)−B(x) ds.
ε 0

Then ψ can be written as

 1  1 1
χ(s)ds
ψ(x) = χ(s)ds + κ e−B(s) ds with κ = − 0
1 −B(s) .
x x 0
e ds

For ψ  we get

ψ  (x) = −χ(x) − κe−B(x) . (4.19)

Apparently the critical point is to derive bounds on χ. Integration by parts and the
definition of Ψ yield

2χ(x) = (hu ) (x) − ζ(x) (4.20)

with
 x
1 
ζ(x) := hbu − 2h(f − cu) − 2(c − b )ψ (s)eB(s)−B(x) ds.
ε 0

For the first term on the right-hand side of (4.20) we have by (3.30) and Prop. 4.15

   2  
(hu ) (x) ≤ Ch(x) 1 + ε−1 e−βx/2ε ≤ Cϑ[2] (ω̄) 1 + ε−1 e−βx/2ε .
cd
(4.21)
4.2 A Simple Upwind Difference Scheme 91

To bound ζ(x), the second term in (4.20), we use (3.30), (4.17a) for k = 0
and (4.21):
 x   
C
h(s) 1 + ε−2 e−βs/ε + ϑcd (ω̄) eβ(s−x)/ε ds
[2]
|ζ(x)| ≤
ε0
 x 
1 + ε−1 eβs/2ε eβ(s−x)/ε ds
[2]
≤ Cϑcd (ω̄)
0 
≤ Cϑcd (ω̄) 1 + ε−1 e−βx/2ε .
[2]

This, eq. (4.20) and inequality (4.21) give

 
|χ(x)| ≤ Cϑcd (ω̄) 1 + ε−1 e−βx/2ε .
[2]
(4.22)

Integrating (4.22), we obtain

|κ| ≤ Cε−1 ϑcd (ω̄),

[2]
(4.23)

1
since 0 e−B(s) ds ≥ ε/b∞ . Combining (4.19)-(4.23), we get (4.17a) for k = 1.
Finally, the bound (4.17b) for the second-order derivative of ψ follows from
(4.15), (3.30), (4.17a) and Prop. 4.15.

Proof of (4.18)
We now bound the terms on the right-hand side of (4.16). For the first two terms a
Taylor expansion with the integral form of the remainder yields
    
 hi 
ε ux̄;i − ui + ui  ≤ C (x − xi−1 ) 1 + ε−2 e−βx/ε dx
2 Ii

by (3.30). To estimate the right-hand side we use the following result from [24].

Lemma 4.16. Let g be a positive monotonically decreasing function on [a, b]. Let
p ∈ IN + . Then
 b  p
b
p−1 1
g(ξ) (ξ − a) dξ ≤ g(ξ)1/p dξ .
a p a

Proof. Consider the two integrals as functions of the upper integration limit. 


We get
     2  2
 hi  
 
ε ux̄;i − ui + ui  ≤ C −1 −βx/2ε
1+ε e
[2]
dx ≤ C ϑcd (ω̄) .
2 Ii
(4.24)
92 4 Finite Difference Schemes for Convection-Diffusion Problems

Next we bound the third term in (4.16). Assuming c, f ∈ C 2 [0, 1], we have
   x
 
g(x) − (x − xk )g  (x) dx − hk+1 gk = (s − xk )g  (s)ds.
Ik+1 Ik+1 xk

Thus,
 
 
  
 (g(x) − (x − xk )g ) dx − hk+1 gk 
 Ik+1 
    2
(s − xk ) 1 + ε−2 e−βs/ε ds ≤ Chk+1 ϑcd (ω̄) ,
[2]
≤ Chk+1
Ik+1

by (3.30) and Lemma 4.16. Hence

 
 1  
N −1   2
   [2]
 g − hg (x)dx − hk+1 gk  ≤ C ϑcd (ω̄) . (4.25)
 xi 
k=i

To bound the remaining terms we use the bounds on ψ and its derivatives from
Lemma 4.13. A Taylor expansion and (4.17b) yield
  2
  xk
ε ψx̄;k − ψk−0
 ≤ε |ψ  (x)| dx ≤ C ϑcd (ω̄) .
[2]
(4.26)
xk−1

Finally,
 
 xk+1 
 (cψ)(s)ds − hk+1 (cψ)k 

xk
  2
|(cψ) (s)| dξds ≤ Chk+1 ϑcd (ω̄) ,
[2]
≤ hk+1
Ik+1

by (4.17a). Therefore,
N −1  1 
   2
  [2]
 hk+1 ck ψk − (cψ)(s)ds ≤ C ϑcd (ω̄) . (4.27)
 xi 
k=i

Applying (4.24)–(4.27) to (4.16) and taking the maximum over i = 0, . . . , N −1,

we get (4.18).

4.2.4 A Posteriori Error Estimation and Adaptivity

In Sect. 4.2.2 the stability of the discrete operator L was used to bound the error
in the discrete maximum norm in terms of the derivative of the exact solution.
4.2 A Simple Upwind Difference Scheme 93

Now, in the first part of this section, roles are interchanged and the stability of the
continuous operator L is used to bound the error in the continuous maximum norm
in terms of finite differences of the numerical solution. We follow [64].

4.2.4.1 A Posteriori Error Bounds

Let uN be the piecewise-linear function that solves (4.3). Then (3.29c) yields
        
u − uN  ≤ 2 L u − uN −1,∞ = 2 min A u − uN + c∞ .
ε,∞ c∈IR

Clearly
       
min A u − uN + c∞ ≤ A u − uN + a − α∞ , (4.28)
c∈IR

where a and α are the constants from (4.11). Furthermore, for any x ∈ (xi−1 , xi ),
   
A u − uN + a − α = [AuN ]i − AuN (x) − Fi + F(x).

We bound thetwo terms on the right-hand side.


Since uN = uN x̄,i for all x ∈ (0, 1) \ ω, we have

 
[AuN ]i − AuN (x)

 −1  
N 1   N xi  
= k+1 −
hk+1 bx;k uN b u (s)ds + b(s) uN (s)ds
k=i xk x

   −1
xi   1   N
− cuN (s)ds − cuN (s)ds + hk+1 ck uN
k ,
x xi k=i

by the definitions of A and A and by integration by parts. For the terms on the
right-hand side, Taylor expansions give
  
   N   
 
hk+1 bx;k uk+1 −
N
b u (s)ds ≤ hk+1 b ∞ uN N
k+1 − uk ,
 Ik+1 
 xk 
     
 b(s) uN (s)ds ≤ b∞ uN N 
k − uk−1 ,

x
  
  N 
 
hk+1 ck uk −
N
cu (s)ds
 Ik+1 
!  N 
≤ h2k+1 c ∞ max |uN  N
k+1 |, |uk | + hk+1 c∞ uk+1 − uk
N
94 4 Finite Difference Schemes for Convection-Diffusion Problems

and
 
 xk       N !
 cu N
(s)ds ≤ hk c∞ max uN  
k , uk−1
 .

x

Thus,
   
[AuN ]i − AuN (x) ≤ C1 max |uN
k+1 − uk | + C2 hu ∞,ω
N N
(4.29)
k=0,...,N −1

with the constants C1 and C2 form Theorem 4.8.

Next bound F − F.
 xk 
 
 f (s)ds ≤ hk f ∞
 
x

and
  
 
 
hk+1 fk − f (s)ds ≤ h2k+1 f  ∞
 Ik+1 

yield

|Fk − F(x)| ≤ C3 h.

Combining this with (4.28) and (4.29), then taking the supremum over all x ∈
(0, 1) \ ω, we get
    
L u − uN  ≤ C1 max |uN
i+1 − ui | + h C2 u ∞,ω + C3
N N
−1,∞ i=0,...,N −1

with the constants C1 , C2 and C3 from Theorem 4.8.

Finally, use (3.29c) in order to obtain the main result of this section.
Theorem 4.17. Let u be the solution of (4.1) and uN that of (4.3). Then
   
u − uN  ≤ 2C1 max |uN
i+1 − ui | + 2h C2 u ∞,ω + C3 .
N N
ε,∞ i=0,...,N −1

Corollary 4.18. Theorem 4.17 and Corollary 4.7 yield

   
u − uN  ≤C max hi+1 1 + |uN
x;i | .
ε,∞ i=0,...,N −1

Note the analogy of these results to Theorem 4.8 and to Corollary 4.10.
4.2 A Simple Upwind Difference Scheme 95

4.2.4.2 An Adaptive Method

From Theorem 4.8 it is easily concluded that the error of our upwind scheme
satisfies
 ,
 
u − uN  ≤ C max 1 + u (x)2 dx.
ε,∞,ω i=1,...,N Ii

On the other hand,

 1 ,
1 + u (x)2 dx ≤ C,
0

by (3.30). Thus, if the mesh is designed so that

 ,  ,
1 + u (x)2 dx = 1 + u (x)2 dx (4.30)
Ii−1 Ii

for i = 1, . . . , N − 1, i. e., if the mesh equidistributes the arc length of the exact
solution, then
 
u − uN  ≤ CN −1 . (4.31)
ε,∞,ω

However, u is not available. An idea that leads to an adaptive method is to approx-

imate the integrals in (4.30) by the mid-point quadrature rule, and u (xk−1/2 ) by
a central difference quotient and finally to replace u by the numerical solution uN .
We get
 , 4  2
1 + u (x)2 dx ≈ hi 1 + uN x̄;i .
Ii

Thus setting
  4  2
Qi = Qi uN , ω := 1 + uN
x̄;i ,

we can replace (4.30) by

1 
N
hi Qi = hj Qj for i = 1, . . . , N − 1. (4.32)
N j=1

Now solving the difference equation (4.3) and the discretised equidistribution prin-
ciple (4.32) simultaneously, we get an adaptive method.
Kopteva and Stynes [69] proved that the nonlinear system of equations (4.3)
and (4.32) possesses a solution and the error of the solution uN obtained satis-
fies (4.31). An essential ingredient in the analysis is the a posteriori error bound of
96 4 Finite Difference Schemes for Convection-Diffusion Problems

Theorem 4.17. They proceed by considering a mesh movement algorithm, originally

due to de Boor [25], which starts with a uniform mesh and aims to construct a mesh
that solves the equidistribution problem (4.32).
In [69] it is shown that (4.32) does not need to be enforced strictly. The de Boor
algorithm, which we are going to describe now, can be stopped when the relaxed
equidistribution principle

γ 
N
Qi hi ≤ hj Qj for i = 1, . . . , N,
N j=1

with a user-chosen constant γ > 1 is satisfied.

Algorithm:
1. Initialisation: Fix N and choose the constant C0 > 1. The initial mesh ω [0] is
uniform with mesh size 1/N .
2. For k = 0, 1, . . . , given the mesh ω [k] , compute the discrete solution uN,[k] on
[k] [k] [k]
this mesh. Set hi = xi − xi−1 for each i. Let the piecewise-constant monitor
function M̃ [k] be defined by
   
[k]
M̃ [k] (x) := Qi := Qi uN,[k] , ω [k] for x ∈ xki−1 , xki .

Then the total integral of the monitor function M̃ [k] is

 1 
N
[k] [k]
I [k] := M̃ [k] (t)dt = hj Qj .
0 j=1

3. Test mesh: If

[k] [k] I [k]

max hj Qj ≤ γ ,
j=1,...,N N

then go to Step 5. Otherwise, continue to Step 4.

4. Generate a new mesh by equidistributing the monitor function M̃ [k] of the current
computed solution: Choose the new mesh ω [k+1] such that

 [k+1]
xi
I [k]
M̃ [k] (t)dt = , i = 0, . . . , N.
[k+1]
xi−1 N

x
(Since 0 M [k] (t)dt is increasing in x, the above relation clearly determines the
mesh ω [k+1] uniquely.) Return to Step 2.
5. Set ω ∗ = ωN and uN,∗ = uN,[k] then stop.
[k]
4.2 A Simple Upwind Difference Scheme 97

In [69] it is shown that the stopping criterion is met after O (| ln ε|) iterations
and the error of the numerical solution obtained satisfies (4.31) with a constant
C = C(γ).
Remark 4.19. Beckett [19] notes that when γ is chosen close to 1 the algorithm
becomes numerically unstable. The mesh starts to oscillate: Mesh points moved into
the layer region in one iteration are moved back out of it in the next iteration. Thus,
the parameter γ must not be chosen too small. Values used in various publications
are 2 and 1.2, but may be problem dependent. ♣
To avoid these oscillations Linß [86] rewrites (4.31) as

(xi − xi−1 )2 + (uN

i − ui−1 )
N 2

= (xi+1 − xi )2 + (uN
i+1 − ui )
N 2
for i = 1, . . . , N − 1.

Then he treats the system of (4.3) and (4.31) as a map

 
(0, 1] → IR2(N +1) : ε → ω̄ε , uN
ε

and applies a continuation method combining an explicit Euler method (predictor)

with a Newton method (corrector). The iteration matrices in each Newton step are
seven diagonal and in an example the numerical costs are approximately of order
N |ln(N ε)|. However, convergence of this method is not proved in [86].

4.2.5 An Alternative Convergence Proof

In this section we shall demonstrate how the (∞ , 1 ) stability (4.8b) can be
exploited to study convergence of the scheme (4.3) on S-type meshes. The results
are less general than those of Sect. 4.2.2, but can be generalised to two dimensions;
cf. Sect. 9.3.2. In our presentation we follow [106].
By (4.8b), we have
 
u − uN  ≤ β −1 Lu − f 1,ω . (4.33)
∞,ω̄

Thus, the maximal nodal error is bounded by a discrete 1 norm of the truncation
error ζ := Lu − f :

N −1
ζ1,ω = hj+1 |ζj |.
j=0

Using the solution decomposition u = v + w of Theorem 3.48 and a triangle in-

equality, we can bound the truncation error pointwise:
' (  ' ( 
|ζi | ≤  Lv i − fi  +  Lw i  .
98 4 Finite Difference Schemes for Convection-Diffusion Problems

Separate Taylor expansions for the two solution components and the derivative
bounds of Theorem 3.48 yield
 
hi+1 |ζi | ≤ C hi+1 + hi + e−βxi−1 /ε (4.34a)

and
  
hi+1 |ζi | ≤ C |hi+1 − hi | 1 + ε−1 e−βxi−1 /ε
 2   (4.34b)
2 −2 −βxi−1 /ε
+ hi + hi+1 1 + ε e .

In the analysis, assume the mesh generating function ϕ̃ of the S-type mesh sat-
isfies (2.8) and that σ ≥ 2. For the sake of simplicity suppose ϕ̃ is nondecreasing.
This leads to a mesh that does not condense on [0, τ ] as we move away from the
layer, i.e., hi ≤ hi+1 for i = 1, . . . , qN − 1, which is reasonable for the given
problem.
Now let us bound the 1 norm of the truncation error. Apply (4.34a) to bound
hi+1 |ζi | for i = qN, qN + 1 and (4.34b) otherwise. We get


qN −1  
ζ1,ω ≤ C (hi+1 − hi ) 1 + ε−1 e−βxi /ε
i=1

  
+ h2i + h2i+1 1 + ε−2 e−βxi−1 /ε
  (4.35)
+ C h + e−βxqN −1 /ε + e−βxqN /ε


N −1  
+C N −2 1 + ε−2 e−βxi−1 /ε .
i=qN +2

We bound the terms on the right-hand side separately in reverse order.

Let H denote the (constant) mesh size on [τ, 1]. Then for i = qN + 2, . . . , N − 1

 2
ε−2 e−βxi−1 /ε ≤ ε−2 e−βH/ε e−βτ /ε ≤ C H/ε e−βH/ε ≤ C,

since xi−1 ≥ xN/2 + H = τ + H and σ ≥ 2. Thus,


N −1  
N −2 1 + ε−2 e−βxi−1 /ε ≤ CN −1 . (4.36)
i=qN +2
4.2 A Simple Upwind Difference Scheme 99

Furthermore,

h + e−βxqN −1 /ε + e−βxqN /ε
  (4.37)
≤ h + 1 + eβhqN /ε e−βxqN /ε ≤ h + CN −σ ,

by (2.11).
Next we bound the first sum in (4.35). We have


qN −1
 
hi+1 − hi + h2i + h2i+1 ≤ 3h (4.38)
i=1

and


qN −1
(hi+1 − hi ) e−βxi /ε
i=1


qN −1  
= −h1 e−βx1 /ε + hi e−βxi−1 /ε − e−βxi /ε + hqN e−βxqN −1 /ε .
i=2

The mean value theorem, (2.11) and (2.12) imply

 
 −βxi−1 /ε  β
e − e−βxi /ε  ≤ hi e−βxi−1 /ε ≤ CεN −1 max |ψ  |e−βxi−1 /(2ε) .
ε
Therefore, it follows that


qN −1 
qN
hi
ε−1 (hi+1 − hi ) e−βxi /ε ≤ CN −1 max |ψ  | e−βxi−1 /(2ε) .
i=1 i=1
ε

Ineq. (2.11) also gives


qN  τ
hi
e−βxi−1 /(2ε) ≤ C ε−1 e−βx/(2ε) dx ≤ C.
i=1
ε 0

Hence,


qN −1
ε−1 (hi+1 − hi ) e−βxi /ε ≤ CN −1 max |ψ  |. (4.39)
i=1

Similar calculations yield

qN −1 
   −βxi−1 /ε 
−2 
ε  2 2
hi + hi+1 e  ≤ CN −1 max |ψ  |. (4.40)
 
i=1
100 4 Finite Difference Schemes for Convection-Diffusion Problems

Substituting (4.36)–(4.40) into (4.35) and applying (4.33), we get the uniform
error bound
   
u − uN  ≤ C h + N −1 max |ψ  | .
∞,ω̄

In [106] the authors proceed—using more detailed bounds on the discrete

Green’s function—to prove the sharper bound
 
u − uN  ≤ CN −1
∞,ω̄∩[τ,1]

for the error outside of the layer region provided that (2.14) is satisfied by the mesh
generating function.

4.2.6 The Truncation Error and Barrier Function Technique

We now consider the convection-diffusion problem

−εu − bu + cu = f in (0, 1), u(0) = u(1) = 0 (4.1)

discretised by
' N(
L̂u i := −εuN
x̄x̂;i − bi ux;i + ci ui = fi
N N
for i = 1, . . . , N − 1,
(4.41)
uN N
0 = γ0 , uN = γ1 .

In contrast to the scheme (4.3) this method is first-order consistent in the mesh nodes
on arbitrary meshes.
The analysis of this section uses the truncation error and barrier function tech-
nique developed by Kellogg and Tsan [61]. This was adapted to the analysis
of Shishkin meshes by Stynes and Roos [153] and later used for other meshes
also [137]. This technique can be used for problems in two dimensions too; see
Sect. 9.1 or [81, 107]. We demonstrate this technique by sketching the convergence
analysis for S-type meshes. For more details the reader is referred to [137].
The matrix associated with L̂ is an M -matrix. Similar to (4.5), we have the fol-
lowing comparison principle for two mesh functions v, w ∈ IRN +1 :

⎫
L̂v ≤ L̂w on ω, ⎬
v0 ≤ w0 , =⇒ v ≤ w on ω̄. (4.42)
⎭
vN ≤ wN
4.2 A Simple Upwind Difference Scheme 101

Theorem 4.20. Let ω be a S-type mesh with σ ≥ 2; see Sect. 2.1.3. Assume that the
function ϕ̃ is piecewise differentiable and satisfies (2.8) and (2.14). Then the error
of the simple upwind scheme satisfies
  
  C h + N −1
max |ψ 
| for i = 0, . . . , qN − 1,
u i − u i  ≤
N
 
C h + N −1 for i = qN, . . . , N.

Proof. The numerical solution uN is split analogously to the splitting of u = v + w

of Theorem 3.48: uN = v N + wN with

[L̂v N ]i = fi for i = 1, . . . , N − 1, v0N = v(0), vN

N
= v(1) = γ1

and

[L̂wN ]i = 0 for i = 1, . . . , N − 1, w0N = w(0), wN

N
= w(1) = 0.

Then the error is u − uN = (v − v N ) + (w − wN ) and we can estimate the er-

ror in v and w separately. For the regular solution component v Taylor expansions
and (3.34a) give
   
L̂(v − v N )i  = [L̂v]i − (Lv)i  ≤ Ch for i = 1, . . . , N − 1.

Furthermore, (v − v N )0 = (v − v N )N = 0. Then the comparison principle (4.42)

yields
 
vi − viN  ≤ C(1 − xi )h

with some constant C, which is independent of ε. Thus,

 
v − v N  ≤ Ch. (4.43)
∞,ω

Using (4.42), one can show that

i 
3 −1
 N βhk
wi  ≤ w̄iN := C 1+ for i = 0, . . . , N. (4.44)
2ε
k=1

For ξ ≥ 0 we have ln(1 + ξ) ≥ ξ − ξ 2 /2 which implies

  2 
1
qN
−σ/2 βhk
w̄iN ≤ N
w̄qN ≤N exp ≤ CN −1 for i = qN, . . . , N ;
2 2ε
k=1
102 4 Finite Difference Schemes for Convection-Diffusion Problems

see Remark 2.4. Hence,

   
wi − wiN  ≤ |wi | + wiN  ≤ CN −1 for i = qN, . . . , N, (4.45)

where we have used (3.34b).

For the truncation error with respect to the layer part w, Taylor expansions
and (3.34b) give
   
   
[L̂(w − wN )]i  = [L̂w]i  ≤ Cε−2 (hi + hi+1 ) e−βxi−1 /ε

≤ Cε−1 e−βxi /(2ε) N −1 max |ψ  |

≤ Cε−1 w̄iN N −1 max |ψ  | for i = 1, . . . , qN − 1,

 
by (2.11) and (2.12). Note that w0 − w0N = 0 and wqN − wqN
N 
≤ CN −1 . There-
fore, (4.42) yields
  !
(w − wN )i  ≤ C N −1 + w̄iN N −1 max |ψ  | , for i = 0, . . . , qN − 1,

for C chosen sufficiently large. Thus,

 
wi − wiN  ≤ CN −1 max |ψ  | for i = 0, . . . , qN − 1.

Combine (4.43) and (4.45) with the last inequality to complete the proof. 


Remark 4.21. We are not aware of any results for B-type meshes that make use of
this truncation error and barrier function technique. Also note that this technique
needs σ ≥ 2, while in Sect. 4.2.2 only σ ≥ 1 was assumed. ♣

Remark 4.22. The technique of Sect. 4.2.2 also provides error estimates for the ap-
proximation of the first-order differences:
' 
ε  (u − uN )x ∞,ω ≤ Cϑcd (ω̄).
[1]

In [33] the authors use the barrier function technique to establish that the upwind
scheme (4.41) on standard Shishkin meshes satisfies

   CN −1 ln N for i = 0, . . . , qN − 1,
ε uN − u x;i  ≤
−1
CN for i = qN, . . . , N − 1.

However, the technique in [33] makes strong use of the piecewise uniformity of the
mesh. ♣
4.2 A Simple Upwind Difference Scheme 103

4.2.7 Discontinuous Coefficients and Point Sources

Consider the convection-diffusion problem in conservative form with a point source

or a discontinuity of the convection coefficient at d ∈ (0, 1):

Lc u := −εu − (bu) + cu = f + αδd , in (0, 1),

(4.46)
u(0) = γ0 , u(1) = γ1 ,

where δd is the shifted Dirac-delta function δd (x) = δ(x − d) with d ∈ (0, 1). As-
sume that b ≥ β1 > 0 on (0, d) and b ≥ β2 > 0 on (d, 1) and set β = min {β1 , β2 }.
For the sake of simplicity, also assume that c ≥ 0 and c − b ≥ 0 on [0, 1]. Properties
of the exact solution are studied in Sect. 3.4.1.3.
Following [88], we consider the upwinded finite difference method: Find uN ∈
N +1
IR with
' c N( − N
L u i := −εuN x̄x;i − (b u )x;i + ci ui = fi + Δd,i for i = 1, . . . , N − 1,
N

uN N
0 = γ0 , uN = γ1 ,
(4.47)

where v − (x) := lim v(s), vi− := v − (xi ) and

s→x−0

h−1
i+1 if d ∈ [xi , xi+1 ),
Δd;i :=
0 otherwise

is an approximation of the shifted Dirac-delta function.

The discrete operator Lc enjoys the stabilityproperty (4.8c).
 Therefore, it is suf-
ficient to derive bounds for the truncation error Lc (u − uN )−1,∞,ω . Adapting the
notation from Sect. 4.2.2, we set
 1
     
Ac v (x) := εv  (x) + b− v (x) + cv (s)ds,
x
 1 
α if xi ≤ d,
F(x) := f (s)ds +
x 0 otherwise,
 −1
c
 −
 N
[A v]i := εvx̄;i + b v i
+ hk+1 (cv)k
k=i

and
−1


N
α if xi ≤ d,
Fi := hk+1 fk +
k=i 0 otherwise.
104 4 Finite Difference Schemes for Convection-Diffusion Problems

Inspecting (4.46) and (4.47), we see

Ac u − F ≡ const on (0, 1) and Ac uN − F ≡ const on ω.

Then, analogously to (4.13), we obtain

(Ac u − Ac u − F + F)i

N −1  xN

= ε (ux̄ − u )i + hk+1 (ck uk − fk ) − (cu − f ) (x)dx
k=i xi

since the contributions from the δ functions and its discretisation cancel.
Proceeding along the lines of Sect. 4.2.2, we get.
Theorem 4.23. Let u be the solution of (4.46). Then the error of the simple upwind
scheme (4.47) satisfies

   
u − uN  ≤ C max 1 + |u (x)| dx.
ε,∞,ω i=1,...,N Ii

Corollary 4.24. Theorem 4.23 and the derivative bounds (3.36) yield
 
u − uN  [1]
≤ Cϑcdi (ω̄),
ε,∞,ω

where
  
1 + ε−1 e−β1 s/pε + Hd (s)ε−1 e−β2 (s−d)/pε ds,
[p]
ϑcdi (ω̄) := max
k=1,...,N Ik

and Hd is the shifted Heaviside function.

Remark 4.25. Layer-adapted meshes for (4.46) have been introduced in Sect. 2.1.5.
We have the uniform error bounds

  CN −1 ln N for the Shishkin mesh and
u − uN  ≤
ε,∞,ω −1
CN for the Bakhvalov mesh

[p]
if σ1 ≥ 1 and σ2 ≥ 1; see Sect. 2.1.5 for the bounds on ϑcdi (ω̄). ♣

Numerical results
Let us verify experimentally the theoretical result of Theorem 4.23. Our test prob-
lem is

−εu − u = x + δ1/2 in (0, 1), u(0) = u(1) = 0. (4.48)

4.2 A Simple Upwind Difference Scheme 105

Table 4.2 The upwind difference scheme

for (4.48); errors in the discrete maximum
norm
Bakhvalov mesh Shishkin mesh
N error rate error rate
27 2.822e-2 0.95 3.898e-2 0.78
28 1.458e-2 0.97 2.277e-2 0.81
29 7.447e-3 0.98 1.299e-2 0.84
210 3.779e-3 0.99 7.280e-3 0.85
211 1.909e-3 0.99 4.027e-3 0.87
212 9.610e-4 0.99 2.204e-3 0.88
213 4.828e-4 1.00 1.197e-3 0.89
214 2.422e-4 1.00 6.454e-4 0.90
215 1.214e-4 1.00 3.462e-4 0.91
216 6.080e-5 — 1.848e-4 —

The results presented in Table 4.2 are in fair agreement with Theorem 4.23. Again
the Bakhvalov mesh gives more accurate results than the Shishkin mesh.

Further remarks
The traditional truncation error and barrier function technique of Sect. 4.2.6 can also
be applied to problems with interior layers. Farrell et al. [35] consider the problem
of finding u ∈ C 2 ((0, d) ∩ (d, 1)) ∪ C 1 [0, 1] such that

−εu − bu = f in (0, d) ∪ (d, 1), u(0) = u(1) = 0,

where at the point d ∈ (0, 1) the convection coefficient changes sign:

b(x) > 0 for x ∈ (0, d), b(x) < 0 for x ∈ (d, 1) and |b(x)| ≥ β > 0.

The solution u and its derivatives satisfy

 (k) 
u (x) ≤ C 1 + ε−k e−β|x−d|/ε for k = 0, 1, . . . , q and x ∈ [0, 1],

where the maximal order q depends on the smoothness of the data. Using the barrier
function technique of Sect. 4.2.6, in [35] the authors establish the error bound
 
u − uN  ≤ CN −1 ln N
∞,ω

for the simple upwind scheme (4.41) on a Shishkin mesh.

106 4 Finite Difference Schemes for Convection-Diffusion Problems

4.2.8 Quasilinear Problems

We now extend the results of Sect. 3.4.1 and 4.2.1 to the class of quasilinear prob-
lems described by

T c u := −εu − b(·, u) + c(·, u) = 0 in (0, 1),

(4.49)
u(0) = γ0 , u(1) = γ1

with 0 < ε  1, ∂u b ≥ β > 0 and ∂u c ≥ 0 and its simple upwind discretisation

' c N(
T u i := −εuN x̄x;i − b(·, u )x;i + c(·, u )i = 0 for i = 1, . . . , N − 1,
N N

uN N
0 = γ0 , uN = γ1 .

First, for the solution u of (4.49) and its derivatives, the bounds (3.30) hold true
too; see [166]:
 (k) 
u (x) ≤ C 1 + ε−k e−βx/ε for k = 0, 1, . . . , q and x ∈ [0, 1],

where the maximal order q depends on the smoothness of the data.

Next we use a standard linearisation technique to study stability properties of T c .
For any two functions v, w ∈ W 1,∞ (0, 1) define the linear operator

Lc y = Lc [v, w]y := −εy  − (py) + qy,

with
 1  
p(x) = ∂u b x, w(x) + s(v − w)(x) ds ≥ β
0

and
 1  
q(x) = ∂u c x, w(x) + s(v − w)(x) ds ≥ 0.
0

The linearised operator Lc is constructed such that Lc (v − w) = T c v − T c w on

(0, 1). The analysis of Sect. 3.4.1 can be applied to Lc . We get

|||v − w|||ε,∞ ≤ T c v − T c w−1,∞

for all v, w ∈ W 1,∞ with v − w ∈ W10,∞ .

Similarly, we linearise T c . For arbitrary mesh functions v, w ∈ IRN +1 set

' (
[Lc y]i = Lc [v, w]y i := −εyx̄x;i − (py)x;i + qi yi
4.2 A Simple Upwind Difference Scheme 107

with p and q defined above. Again the linearised operator Lc is constructed such
that Lc (v − w) = T c v − T c w on ω. The technique from Sect. 4.2.1 can be used to
obtain

|||v − w|||ε,∞,ω ≤ T v − T w−1,∞,ω

for all v, w ∈ IRN +1 with v − w ∈ IR0N +1 .

In order to conduct an error analysis, take v = u and w = uN and proceed as in

Sect. 4.2.2 and 4.2.4 to get a priori and a posteriori error bounds.

Remark 4.26. Discretisations of quasilinear problems can also be analysed using the
truncation error and barrier function technique of Sect. 4.2.6 or using the (∞ , 1 )
stability (4.8b); see [36, 149] and [106], respectively. ♣

4.2.9 Derivative Approximation

In a number of applications the user is more interested in the approximation of the

gradient or of the flow than in the solution itself. In Sect. 4.2.2 the following error
bound for the weighted derivative was established:
  
  [1]
ε  u − uN  ≤ Cϑcd (ω̄).
∞

Note that u (0) ≈ ε−1 by (3.30). Therefore, multiplying by ε in this estimate is the
correct weighting. However, looking at the bounds (3.30) for the derivative of u,
we see that the derivative is bounded uniformly away from the layer, where we
therefore expect that a similar bound holds without the weighting by ε.

Theorem 4.27. Let u be the solution of (4.1) and uN that of (4.3). Then
    2
u i − u N
x̄;i
 ≤ Ch −1
ϑ
[2]
(ω̄) for i = 1, . . . , N.
i cd

Proof. We work from the error expansion of Sect. 4.2.3:

 
  u i − ψi − u N
i − ui−1 − ψi−1 − ui−1
N
ψi − ψi−1
u−u N
x̄;i
= +
hi hi

Then
    2
 u − uN  ≤ Ch−1 ϑ[2] (ω̄) , (4.50)
x̄;i i cd
108 4 Finite Difference Schemes for Convection-Diffusion Problems

by Lemma 4.13 and Theorem 4.14. Furthermore,

 
1  
|ui − ux̄;i | =  (s − xi−1 )u (s)ds
hi  
I
 i    2
C
(s − xi−1 ) 1 + ε−2 e−βs/ε ds ≤ Ch−1
[2]
≤ i ϑcd (ω̄) ,
hi Ii

by (3.30) and Lemma 4.16. Finally, a triangle inequality yields the assertion. 


Layer-adapted meshes.
Let us illustrate Theorem 4.27 by applying it to two standard layer-adapted meshes.
Bakhvalov meshes (Sect. 2.1.1) can be generated by equidistributing
  
Kβ βξ
MBa (ξ) = max 1, exp − for ξ ∈ [0, 1].
ε σε

Clearly MBa is continuous and monotonically decreasing. Therefore,

 1 
1
MBa (s)ds = MBa (s)ds ≤ hi MBa (xi−1 )
N 0 Ii

and
  
1 Kβ βxi−1
≤ CN MBa (xi−1 ) = CN max 1, exp − .
hi ε σε

Now, (2.6) and Theorem 4.27 yield

  
   Kβ βxi−1
u i − u N
x̄;i
 ≤ CN −1
max 1, exp − if σ ≥ 2.
ε σε

A very similar result was established by Kopteva and Stynes [72] through a different
technique.
Shishkin meshes (Sect. 2.1.3). For these meshes the local step sizes satisfy
σε ln N
hi = for i = 1, . . . , qN and hi ≥ N −1 for i = qN + 1, . . . , N.
qβ N

Hence, Theorem 4.27 gives


   −1 −1
for i = 1, . . . , qN − 1,
ui − ux̄;i  ≤ Cε N ln N
N
CN −1 ln2 N for i = qN, . . . , N.
4.3 Second-Order Difference Schemes 109

Outside the layer region this result is slightly suboptimal. Both in [43] and in [72]
it was shown by means of barrier function techniques that the approximation is a
factor of ln N better, i.e.,
  
ui+1 − uN 
x;i ≤ CN
−1
ln N for i = qN + 1, . . . , N.

4.3 Second-Order Difference Schemes

As simple upwinding yields only low accuracy, it is natural to look for higher-order
alternatives. For one-dimensional problems inverse-monotone schemes exist that
are second-order accurate. One will be studied in Sect. 4.3.1. However, the con-
struction of inverse-monotone difference schemes in two or more dimensions is an
open problem.
Sect. 4.3.2 summarises stability and convergence results for an unstabilised cen-
tral difference scheme.
Possible other approaches to higher-order schemes include:
• the combination of two (or more) approximations by a first-order upwind scheme
on nested meshes by means of the Richardson extrapolation technique.
• their combination of simple upwinding with higher-order unstabilised schemes
using defect correction.
Both approaches have the advantage that linear problems involving only stabilised
operators have to be solved. Sect. 4.3.3 is devoted to these techniques.
Finally, we like to mention the HODIE technique which was used by Clavero
et al. [28] to construct and analyse second- and third-order compact schemes on
Shishkin meshes.

4.3.1 Second-Order Upwind Schemes

Because of their stability properties, they can be analysed with the techniques simi-
lar to those of Sect. 4.2. Consider the convection-diffusion problem in conservative
form:

Lc u := −εu − (bu) + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1 . (4.9)

Let ρi , i = 1, . . . , N be arbitrary with ρi ∈ [1/2, 1]. Define the weighted step

sizes

χi = ρi+1 hi+1 + (1 − ρi )hi for i = 1, . . . , N − 1, χ0 = χN = 0.

110 4 Finite Difference Schemes for Convection-Diffusion Problems

Then following Andreev and Kopteva [11], our discretisation is: Find uN ∈ IRN +1
such that
' ρ N(
L u i = fρ;i for i = 1, . . . , N − 1, uN N
0 = γ0 , uN = γ1 , (4.51)

where

[Lρ v]i := −εvx̄x́;i − (ρbv + (1 − ρ)(bv)− )x́;i + (cv)ρ;i ,

vi+1 − vi
vx́;i = , v−;i = vi−1
χi
and

ρi+1 vi+1 + (1 − ρi+1 + ρi )vi + (1 − ρi )vi−1

vρ;i = .
2
The approximation of the first-order derivative is a weighted combination of up-
winded and downwinded operators. At first glance the approximation of the lowest-
order term and of the right-hand side seems to be very non-standard. It is chosen
such that
 xρ;i+1/2
χi gρ;i is a second-order approximation of g(x)dx
xρ;i−1/2

with xρ;i−1/2 = xi−1 + ρi hi . For ρ ≡ 1/2 we obtain a central difference scheme,

while for ρ ≡ 1 the mid-point upwind scheme is recovered.
This second-order upwind scheme is very similar to the streamline-diffusion
FEM, which is studied in Sect. 5.3.2.

4.3.1.1 Stability of the Discrete Operator

The stability analysis of the operator Lρ is complicated by the positive contribution

of the discretisation (cuN )ρ;i of the lowest order term to the offdiagonal entries of
the system matrix. It is difficult to ensure the correct sign pattern for the application
of the M -matrix criterion (Lemma 3.14). Instead we follow [85] which adapts the
technique from [11].
Set

i−1
[Aρ v]i := εvx̄;i + ρ(bv)i + (1 − ρi )(bv)i−1 − χj (cv)ρ;j , i = 1, . . . , N.
j=1

This operator is related to Lρ by (Aρ v)x́ = −Lρ v. Then any function v ∈ IR0n+1
can be represented as
WN
vi = V i − Wi ,
VN
4.3 Second-Order Difference Schemes 111

where V and W are the solution of the difference equations

[Aρ V ]i = 1, i = 1, 2, . . . , N, V0 = 0

and

[Aρ W ]i = [Aρ v]i + c, i = 1, 2, . . . , N, W0 = 0

for any constant c ∈ IR.

Proposition 4.28. Assume that

 
1 ε
1 ≥ ρi ≥ max ,1 − for i = 1, . . . , N, (4.52a)
2 bi−1 hi
and
c∞ h ≤ β/4. (4.52b)

Then the matrix associated with Aρ is an M -matrix.

Proof. First (4.52a) ensures that the offdiagonal entries of Aρ are nonpositive,
while (4.52b) implies that the diagonal entries are positive.
For any monotonically increasing mesh function zi ≥ 0 we have

c∞  
i−1 i−2
β
[A z]i > ρi bi zi −
ρ
χj (zj+1 + zj ) ≥ zi − c∞ χj zj+1 ,
2 j=1 4 j=1

by (4.52).
Now let
i 
3 
4c∞
z0 = z1 = z2 = 1, and zi = 1+ χk−2 for i = 3, . . . , N.
β
k=3
(4.53)

Clearly zi ≤ e4 c ∞ /β
and
β β
zi − c∞ χi−2 zi−1 , ≥ zi−1 , by (4.52b).
4 4
Then induction for i yields
β
[Aρ z]i > for i = 1, . . . , N.
4
Finally, application of Lemma 3.14 with the test function ei = zi completes the
proof.
112 4 Finite Difference Schemes for Convection-Diffusion Problems

The M -matrix property of Aρ and the function z from (4.53) can now be used to
establish bounds on V and W :
4 4
0 < Vi ≤ zi ≤ e 4 c /β
and |Wi | ≤ Vi Aρ v + c∞,ω , i = 1, . . . , N.
β β

We get our main stability result.

Theorem 4.29. Let ρ and h satisfy (4.52). Then the operator Lρ is (∞ , w−1,∞ )-
stable with
8 4
v∞,ω ≤ e c ∞ /β
min Aρ v + c∞,ω for all v ∈ IR0N +1 .
β c∈IR

Remark 4.30. The (∞ , 1 ) stability

 −1
N
 
v∞,ω ≤ C χk [Lρ v]k 
k=1

is an immediate consequence of the negative-norm stability.

Analyses of second-order upwind schemes based on this type of stability inequal-
ity were given by Andreev and Savin [12] for a modification of Samarskii’s scheme
on a Shishkin mesh [12], and on Bakhvalov meshes [4] and by Linß [87] for quasi-
linear problems discretised on S-type meshes. ♣

4.3.1.2 Error Analysis

We now study the approximation error of the scheme (4.51). Following [11, 85], we
base our analysis on the (l∞ , w−1,∞ ) stability of Theorem 4.29.
Choose

1/2 if hi ≤ 2ε/bi−1 ,
ρi = (4.54)
1 otherwise.

This choice satisfies the assumptions of Theorem 4.29. Therefore,

   
u − uN  ≤ C min Aρ (u − uN ) + c∞,ω . (4.55)
∞,ω c∈IR

Set
 
  xρ;1/2   xρ;1/2
(Ac v) (x) := εv  (x) + bv + cv (s)ds, F := f (s)ds
x x

and

i−1
Fiρ := − χk fρ;k
k=1
4.3 Second-Order Difference Schemes 113

Inspecting (4.9) and (4.51), we see that

Ac u − F ≡ α on (0, 1) and Aρ uN − F ρ ≡ a on ω

with constants α and a because Lc v = −(Ac v) and f = −F  on (0, 1), and
Lρ v = (Aρ v)x́ and f = Fx́ρ on ω. Take c = a − α in (4.55) in order to get
   
u − uN  ≤ C max [Aρ u]i − (Ac u) (xρ;i ) + F(xρ;i ) − Fiρ . (4.56)
∞,ω i=1,...,N

Set g := cu − f ,

[B ρ u]i := εux̄;i + ρi bi ui + (1 − ρi )bi−1 ui−1

and
 
B(x) := εu (x) + bu (x).

Then

[Aρ u]i − (Ac u) (xρ;i ) + F(xρ;i ) − Fiρ

 xρ;i−1/2 
i−1
(4.57)
= [B ρ u]i − (B c u) (xρ;i ) + g(s)ds − χj gρ;j .
xρ;1/2 j=1

When bounding the first term on the right-hand side of (4.57), we have to distinguish
two cases: σi = 1 and σi = 1/2.
For σi = 1 we have
  
' ρ (   ui − ui−1 ε
B u i − Bu (xρ;i ) = ε − ui = u (t)(t − xi−1 )dt.
hi hi Ii

Thus,
  
' h (   
 B u − Bu (xρ,i ) ≤ C 1 + ε−2 e−βt/ε (t − xi−1 )dt, (4.58)
i
Ii

by (3.30) and because ε/hi < b∞ /2 for ρi = 1.

Next, consider σi = 1/2. Then
' ρ (  
B u i − Bu (xρ,i )
 
ui − ui−1 bi ui + bi−1 ui−1
=ε − ui−1/2 + − bi−1/2 ui−1/2 ,
hi 2

where ui−1/2 = u(xi−1/2 ). Taylor expansions for u and u about xi give

  
 ui − ui−1  3ε   

ε  
− ui−1/2  ≤ u (t)(t − xi−1 )dt
hi 2 Ii
114 4 Finite Difference Schemes for Convection-Diffusion Problems

and
  
 bi ui + bi−1 ui−1  3  
 − b u  ≤ (bu) (t)(t − xi−1 )dt.
 2
i−1/2 i−1/2 
2 Ii

From this and (3.30) we see that (4.58) holds for σi = 1/2 too.
Finally, we bound the second term of the right-hand side of (4.57):
 xρ,j+1/2  xρ,j+1/2
g(s)ds − χρ,j gρ,j = (g(s) − gρ,j ) ds.
xρ,j−1/2 xρ,j−1/2

The representation
 xj+1

g(s) = gj+1 − gj+1 (xj+1 − s) + g  (t)(t − s)dt
s

yields

  xj+1  
gρ,j − g(s) − (xρ,j − s)gj+1
 ≤2 g  (t)(t − xj−1 )dt.
xj−1

Next,
   xj+1
 xρ,j+1/2    
  g (t)(t − xj−1 )dt
 g(s)ds − χρ,j gρ,j  ≤ 2(hj + hj+1 )
 xρ,j−1/2  xj−1
 xj+1  
≤ C(hj + hj+1 ) 1 + ε−2 e−βt/ε (t − xj−1 )dt,
xj−1

by (3.30) and because g = cu − f .

Combining this estimate with (4.56), (4.55) and (4.58), we get
  
  xi+1
u − uN  ≤C max 1 + ε−2 e−βt/ε (t − xi−1 )dt.
∞,ω i=1,...,N −1 xi−1

Finally, Lemma 4.16 gives the following convergence result.

Theorem 4.31. Let uN be the approximate solution to (4.9) obtained by the dif-
ference scheme (4.51) with ρ chosen according to (4.54). Assume c∞ h ≤ β/4.
Then
   2
u − uN  ≤ C ϑ
[2]
(ω̄) .
∞,ω cd
4.3 Second-Order Difference Schemes 115

Quasilinear problems
The conclusion of the Theorem also holds when (4.51) is adapted to discretise the
quasilinear problem

−εu − b(·, u) + c(·, u) = 0 in (0, 1), u(0) = γ0 , u(1) = γ1

with 0 < ε  1, ∂u b ≥ β > 0 and ∂u c ≥ 0. The scheme reads: Find uN ∈ IRn+1

such that
 
x̄x́;i − ρb(·, u ) + (1 − ρ)b(·, u )− x́;i + c(xρ;i , uρ;i ) = 0
−εuN N N N
on ω,
uN N
0 = γ0 , uN = γ1

with the stabilisation parameter chosen to satisfy, e. g,


1/2 if hi ≤ 2ε/b∞ ,
ρi =
1 otherwise.

Discontinuous coefficients and point sources

Consider the convection-diffusion problem (4.46) with a point source:

Lc u := −εu − (bu) + cu = f + αδd , in (0, 1), u(0) = γ0 , u(1) = γ1 ,

with the shifted Dirac-delta function δd (x) = δ(x − d). The coefficient b may also
have a discontinuity at x = d. Assume that b ≥ β1 > 0 on (0, d) and b ≥ β2 > 0
on (d, 1).
Using (4.51) we seek an approximation uN ∈ IRn+1 with
' (
Lρ uN i
= fρ;i + Δd,ρ;i for i = 1, . . . , N − 1, uN N
0 = γ0 , uN = γ1

with

χ−1
i if d ∈ [xρ;i−1/2 , xρ;i+1/2 ),
Δd,ρ;i :=
0 otherwise.

Then the above technique and the a priori bounds (3.36) for the derivatives of u
yield the error estimate [88]
   2
u − uN  ≤ C ϑ
[2]
i (ω̄) ,
∞,ω cd

[2]
where ϑcdi (ω̄) has been defined in Sect. 2.1.5.
116 4 Finite Difference Schemes for Convection-Diffusion Problems

Remark 4.32. Roos and Zarin [143] study the difference scheme generated by the
streamline diffusion FEM on Shishkin and on Bakhvalov-Shishkin meshes for the
discretisation of a problem with a point source. They prove (almost) second-order
convergence in the discrete maximum norm too. ♣

A posteriori error estimates

in the maximum norm for (4.9) discretised by (4.51) can be derived using the
(L∞ , W −1,∞ )-stability (3.29c) of the continuous operator Lc . However, compared
to Sect. 4.2.4 the analysis becomes more technical. Therefore, we refer the reader
to the article by Kopteva [64]. A flavour of the technique is given in Sect. 4.3.3.1
where a defect-correction method is analysed.

4.3.1.3 The Barrier Function Technique

Stynes and Roos [153] study a hybrid difference scheme on a Shishkin mesh (with
q = 1/2 and σ > 4). Their scheme uses central differencing on the fine part of the
mesh and the mid-point upwind scheme on the coarse part.
Let us consider the convection-diffusion problem

Lu := −εu − bu + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1 , (4.1)

with b ≥ β > 0 and c ≥ 0 on [0, 1]. This is discretised on a Shishkin mesh—see

Sect. 2.1.3—using the difference scheme
' N(
Lu i = f˜i for i = 1, . . . , N − 1, uN N
0 = γ0 , uN = γ1 (4.59)

with

−εvx̄x̂;i − bi vx̊;i + ci vi if bi hi ≤ 2ε,
[Lv]i :=
−εvx̄x̂;i − bi+1/2 vx;i + (ci vi + ci+1 vi+1 )/2 otherwise,

and

fi if bi hi ≤ 2ε,
f˜i :=
fi+1/2 otherwise.

For N larger than a certain threshold value N0 , the matrix associated with L is an
M -matrix and central differencing is used exclusively on the fine part of the mesh.
4.3 Second-Order Difference Schemes 117

Theorem 4.33. Let ω be a Shishkin mesh with σ ≥ 2; see Sect. 2.1.3. Then the
error of the upwinded scheme (4.59) applied to (4.1) satisfies

  −2
ui − ui  ≤ CN ln N for i = 0, . . . , qN − 1,
2
N
CN −2 for i = qN, . . . , N,

if N is larger than a certain threshold value.

Remark 4.34. A similar scheme generated by streamline-diffusion stabilisation was

analysed by Stynes and Tobiska [154] with special emphasis on the choice of the
mesh parameter σ. There it was first established that the mesh parameter σ should
be chosen equal (or greater than) to the formal order of the scheme. ♣

Proof (of Theorem 4.33). Start with the truncation error. When 2ε < bi hi we have
the bound
 
  xi+1
 
[Lg]i − (Lg)i+1/2  ≤ C ε |g  (s)| ds
xi−1
 (4.60a)
 xi+1
+ hi+1 [|g  (s)| + |g  (s)|] ds ,
xi

otherwise we use

  xi+1 ' (
[Lg] − (Lg)  ≤ C ε|g  (s)| + |g  (s)| ds (4.60b)
i i
xi−1

and, if hi = hi+1 ,

  xi+1 ' (
[Lg] − (Lg)  ≤ Chi ε|g (4) (s)| + |g  (s)| ds. (4.60c)
i i
xi−1

For the analysis we split the numerical solution uN analogously to the splitting
u = v + w of Theorem 3.48 and Remark 3.50: uN = v N + wN with
' N(
Lv i = f˜i for i = 1, . . . , N − 1, v0N = v(0), vN
N
= v(1)

and
' N(
Lw i = 0 for i = 1, . . . , N − 1, w0N = w(0), wN
N
= w(1).

Then the error is u − uN = (v − v N ) + (w − wN ) and we estimate the error in v

and w separately.
118 4 Finite Difference Schemes for Convection-Diffusion Problems

For the regular solution component Theorem 3.48, Remark 3.50 and (4.60) give

' (   −1
 L(v − v )  = [Lv] − f˜i  ≤ CN
N for i = qN,
i
i
CN −2 otherwise.

Note that (v − v N )0 = (v − v N )N = 0. Now set

⎧
⎪
⎪
⎨1  −1
for i = 0, . . . , qN,
ϕi = 3i
βhk
⎪
⎪ 1+ for i = qN + 1, . . . , N.
⎩ ε
k=qN +1

Clearly ϕ0 ≥ 0 and ϕN ≥ 0. Furthermore,

⎧
⎨0 for i = qN,
[Lϕ]i ≥ β β(1 − q)
⎩ ≥ N for i = qN.
2hqN +1 2

Application of a comparison principle with the barrier function CN −2 (1 − xi + ϕi )

yields
 
v − v N  ≤ CN −2 , (4.61)
∞,ω

since the matrix associated with L is inverse monotone as mentioned before.

Next, consider the layer component w. Let
⎧ i  −1 3 qN  −1
⎪
⎪ 3 βhk βhk
⎪
⎪ 1 + + 1 + for i = 1, . . . , qN,
⎨ ε ε
k=1 k=1
ψi :=
⎪ 3 i  −1
⎪
⎪ βhk
⎪
⎩2 1+ for i = qN, . . . , N.
ε
k=1

The inverse monotonicity of the discrete operator L yields

 N
wi  ≤ |v(0) − γ0 | ψi for i = 0, . . . , N,

because Lψ ≥ 0. Furthermore, |wi | ≤ Ce−βxi /ε ≤ Cψi . Thus,

 
wi − wiN  ≤ Cψi for i = 0, . . . , N.

Now the argument that lead to (4.45) for the first-order scheme is imitated to
establish
 
wi − wiN  ≤ CN −2 for i = qN, . . . , N, (4.62)

if σ ≥ 2 in the construction of the Shishkin mesh (Sect. 2.1.3).

4.3 Second-Order Difference Schemes 119

For i = 1, . . . , qN − 1 the truncation error with respect to w satisfies

' (
 L(w − wN )  ≤ CN −2 ln2 N ε−1 e−βxi−1 /ε ≤ CN −2 ln2 N ε−1 ψ̃i ,
i

by (4.60c), Theorem 3.48 and Remark 3.50, where

i 
3 −1
βhk
ψ̃i := 1+ .
2ε
k=1

The inverse monotonicity of L gives

 
(w − wN )i  ≤ CN −2 ln2 N ψ̃i for i = 1, . . . , qN − 1, (4.63)

because
' (
Lψ̃ i ≥ Cε−1 ψ̃i for i = 1, . . . , qN − 1
   
and because both w0 − w0N  ≤ CN −2 and wqN − wqN N 
≤ CN −2 .
Combining (4.61), (4.62) and (4.63), we are finished. 


4.3.2 Central Differencing

In numerical experiments [34, 50, 120] it was observed that central differencing on
Shishkin meshes yields almost second-order accuracy.
A drawback of central difference approximations is their lack of stability. The
discretisations are not maximum-norm stable. It will be seen in Sect. 4.3.2.1 that
the use of layer-adapted meshes induce some additional stability. However, the dis-
crete systems remain difficult to solve efficiently by means of iterative solvers. The
system matrices have eigenvalues with large imaginary parts. This becomes a par-
ticularly important issue when solving higher-dimensional problems.
We shall consider the discretisation
' N(
Lu i = fi for i = 1, . . . , N − 1, uN N
0 = γ0 , uN = γ1 (4.64)

of (4.9), where
[Lv]i := −εvx̄x̂;i − (bv)x̊;i + ci vi .

Similar to (4.4) this scheme is equivalent to a FEM with piecewise linear trial
and test functions, but with the trapezium rule

gi−1 + gi
g(s)ds ≈ hi
Ii 2

used to approximate the integrals on each subinterval Ii of the partition.

120 4 Finite Difference Schemes for Convection-Diffusion Problems

4.3.2.1 Stability

A first analysis was conducted by Andreev and Kopteva [10] who prove that central
differencing on a Shishkin mesh is (l∞ , l1 ) stable. This result was later generalised
by Kopteva [70].
Theorem 4.35. Assume that
 
3 N  5
 ε 1 ε 1  1
 − + ≤ (4.65)
 hi bi−1 2 h i bi 2  4
i=1

and that hi ≤ μhj for i ≤ j with some constant μ. Then the central difference
operator L is (l∞ , l1 ) stable with

N −1
81 
v∞,ω ≤ h̄i |[Lv]i | . (4.66)
4β i=1

Furthermore, let m be such that hi ≤ 2ε/bi−1 for i = 1, . . . , m and

hm+1 > 2ε/bm . Then the operator L is (l∞ , w−1,∞ )-stable with
 
N −1 
11  
v∞,ω ≤ max  h̄k [Lv]k  .
2β j=1,...,N −1  
k=j

for any mesh function v with [Lv]i = 0 for i > m.

Proof. The argument is very technical and therefore not presented here. Instead the
reader is referred to the original work by Kopteva [70]. 


4.3.2.2 A Priori Error Bounds

Based on Theorem 4.35 Kopteva [70] established convergence results for central
differencing on two types of layer-adapted meshes:

  CN −2 for Bakhvalov meshes with σ > 2,
u − uN  ≤ (4.67)
∞,ω −2 2
CN ln N for Shishkin meshes with σ > 2;

The (l∞ , l1 ) stability (4.66) was used by Roos and Linß [138] to prove
   2
u − uN  ≤ C h + max |ψ  | N −1 (4.68)
∞,ω

on S-type meshes with σ ≥ 3. A similar result was given by Kopteva and Linß [71]
for certain quasilinear problems of type (4.49).
4.3 Second-Order Difference Schemes 121

Another approach to study central differencing on Shishkin meshes is that of

Lenferink [76, 77]. He eliminates every other unknown to get a scheme whose sys-
tem matrix is an M -matrix.

4.3.2.3 Derivative Approximation

For the central-difference scheme (4.64) on S-type meshes with σ ≥ 3 we have the
second order bound
   2
   −1
ε  uN
x̄;i − u i−1/2  ≤ C h + N max |ψ  | .

The proof in [138] uses the bound (4.68) for the discretisation error, then interprets
the scheme as a finite element method with inexact integration and finally applies a
finite element technique [173] to get the bound for the derivative approximation.

4.3.3 Convergence Acceleration Techniques

In the early 1980s Hemker [51] proposed the use of defect-correction methods when
solving singularly perturbed problems. However, the first rigorous proof of uniform
convergence of a defect-correction scheme was not published before 2001 (Fröhner
et al. [43]). Various analyses by Nikolova and Axelsson [16, 126] are at least not
rigorous with regard to the robustness, i. e. the ε-independence of the error con-
stants, while the analysis by Fröhner and Roos [44] turned out to be technically
unsound [42].

4.3.3.1 Defect Correction

Let us consider the defect correction method from [43] for our model convection-
diffusion problem in conservative form:

Lc u := −εu − (bu) + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1 . (4.9)

It is based on the upwind scheme

' c N(  N
L u i := −εuN
x̄x;i − bu x;i
+ ci uN
i = fi (4.10)

combined with the unstabilised second-order central difference scheme

' c N(  N
6 u := −εuN
L i x̄x̂;i − bu x̊;i
+ ci uN
i = fi .
122 4 Finite Difference Schemes for Convection-Diffusion Problems

With this notation we can formulate the defect correction method. This two-stage
method is the following:
1. Compute an initial first-order approximation ũN using simple upwinding:
' (
Lc ũN i
= fi for i = 1, . . . , N − 1, ũN N
0 = γ0 , ũN = γ1 . (4.69a)

2. Estimate the defect ζ in the differential equation by means of the central differ-
ence scheme:

6 c ũN ]i − fi
ζi = [L for i = 1, . . . , N − 1. (4.69b)

3. Compute the defect correction Δ by solving

[Lc Δ]i = κi ζi , for i = 1, . . . , N − 1, Δ0 = ΔN = 0. (4.69c)

%
with κi = h̄i hi+1 .
4. Then the final computed solution is

uN = ũN − Δ. (4.69d)

Remark 4.36. At first glance both the upwind discretisation and the particular
weighting of the residual in (4.69c) appear a bit non-standard. No justification for
these choices is provided by [43, 93]. An argument that suggests this particular
choice is presented in [101].
Furthermore, the weighting becomes the standard κi = 1 on uniform meshes;
however, when used on non-uniform meshes, κi = 1 might reduce the order of
convergence which is illustrated by numerical experiments in [101]. ♣

In the analysis of the method we use the following notation:

 
  1  1
(Ac v) (x) := εv  (x) + bv (x) + cv (s)ds, F(x) := f (s)ds,
x x
 −1  −1
' c ( N N
A v i := εvx̄;i + (bv)i + hk+1 (cv)k , Fi := hk+1 fk
k=i k=i

and

' ( (bv)i + (bv)i−1 N 

N
6c v
A := εvx̄;i + + h̄k (cv)k , F6i := h̄k fk .
i 2
k=i k=i

The differential equation (4.9) yields

Ac u − F ≡ α = const, (4.70)
4.3 Second-Order Difference Schemes 123

while (4.69b) and (4.69c) imply

6c ũN − F6) ≡ a = const .

Ac Δ − (A (4.71)

A priori analysis
The negative norm stability (4.8c) of the operator Lc yields for the error of the
defect-correction method
   
u − uN  ≤ 2(Ac − A6c )(u − ũN )
ε,∞,ω ∞,ω
 c  (4.72)
 6 6 
+ 2 A u − F + α ∞,ω ,

by (4.70) and (4.71), where α is the constant from (4.70).

The second term in (4.72) is the truncation error of the central difference scheme.
It is formally of second order. The first term is the so called relative consistency
error. While the error u − ũN of the simple upwind scheme is only of first order,
the hope is that Ac and A 6c are sufficiently close to gain second order in this term
too.
Consider the relative consistency error first. Let η := u − ũN denote the error of
the simple upwind scheme. A straight-forward calculation and summation by parts
give

 −1
' c ( N
6c )η = (bη)i − (bη)i−1 +
(A − A hk+1
(cη)k−1 − (cη)k hi
− (cη)i ,
i 2 2 2
k=i+1

which can be bounded by

'  
 (  c∞
c 6
 (A − A )η i  ≤ b∞ +
c
max |ηi − ηi−1 |
2 i=1,...,N
 
 c ∞ + c∞
+ h b ∞ + η∞,ω .
2

Thus,
   2
 c 
6c )η 
(A − A ≤C max |ηi − ηi−1 | + hη∞,ω
[2]
≤ C ϑcd (ω̄) ,
∞,ω i=1,...,N
(4.73)

[1] [2]
by (4.50) and because h ≤ ϑcd (ω̄) ≤ ϑcd (ω̄).

Remark 4.37. The first term, the maximum difference of the error of the upwind
scheme in two adjacent mesh points, constituted the main difficulty in [43]. With
the error expansion of Sect. 4.2.3 this has become a simple task. ♣
124 4 Finite Difference Schemes for Convection-Diffusion Problems

Next, let us bound the truncation error of the central difference scheme. By (4.70)
we have (A6c u − F6)i − α = (A6c u − F6)i − (Au − F)
i−1/2 . Hence,

     (bu) + (bu) 

 6c 6      i i−1 
(A u − F )i − α  ≤ ε ux̄;i − u i−1/2  +  − (bu)i−1/2 
2
  1  (4.74)
 N 
 
+ h̄k gk − g(s)ds
 xi−1/2 
k=i

with g = f − cu. Using Taylor expansions for u, u and (bu) about x = xi , we

obtain
  3ε     2
 
ε ux̄;i − ui−1/2  ≤ (s − xi−1 )u (s)ds ≤ C ϑcd (ω̄)
[2]
(4.75)
2 Ii

and
    2
 (bu)i + (bu)i−1  3  
 − (bu)  ≤ (s − x ) (bu)
(s) ds ≤ C ϑ
[2]
(ω̄) ,
 2
i−1/2 
2 Ii
i−1 cd

by (4.9), (3.30) and Lemma 4.16.

For the last term in (4.74) a Taylor expansion gives
  xk 
h h2k   h3
 k 
 gk − gk−1/2 − g(s)ds ≤ k g  ∞,(xk−1/2 ,xk )
 2 8 xk−1/2  8
(4.76a)
   2
≤ Ch3k 1 + ε−2 e−βxk−1/2 /ε ≤ Chk ϑcd (ω̄) ,
[2]

where we have used (3.30) and Proposition 4.15 with x = xk−1/2 and σ = 2.
Furthermore, we have
  xk+1/2 
 hk+1 h2k+1  
 g(s)ds
 2 gk + 8 gk+1/2 −
xk
 xk+1/2  2 (4.76b)
(σ − xk ) |g  (σ)| dσ ≤ Chk+1 ϑcd (ω̄) ,
[2]
≤ hk+1
xk

by (3.30) and Lemma 4.16. Combine these two estimates:

N  1 
 
 
 h̄k gk − g(s)ds
 xi−1/2

k=i
   2
≤ C ϑcd (ω̄)2 + h2i 1 + ε−2 e−βxi−1/2 /ε
[2] [2]
≤ C ϑcd (ω̄) ,

by Proposition 4.15.
4.3 Second-Order Difference Schemes 125

Therefore,
 c   2
6 u − F6 − α
A ≤ C ϑ
[2]
(ω̄) ,
∞,ω cd

by (4.9), (3.30) and Lemma 4.16.

Collect (4.72), (4.73) and the last inequality to get the main result of this section.

Theorem 4.38. Let u be the solution of (4.9) and uN that of the defect correction
method (4.69). Then
   2
u − uN  [2]
≤ C ϑcd (ω̄) .
ε,∞,ω

Derivative approximation
For x ∈ Ii a triangle inequality gives
   I 
ε u (x) − uN   
x̄;i ≤ ε |u (x) − ux̄;i | + ε (u − u )x̄;i
N 
 2
≤ ε |u (x) − ux̄;i | + C ϑcd (ω̄) ,
[2]

by Theorem 4.38. The representation

  x
 1
u (x) − ux̄;i = u (t)dtds. (4.77)
hi Ii s

yields

ε |u (x) − ux̄;i | ≤ Cϑcd (ω̄),

[1]
by (3.30).

Thus, in general we only have a first-order approximation for the ε-weighted deriva-
tive:

ε |u (x) − ux̄;i | ≤ Cϑcd (ω̄) for x ∈ Ii .

[2]

This result is sharp.

For the midpoint xi−1/2 of the mesh interval Ii we expand (4.77) to give
  x t
1
u (xi−1/2 ) − ux̄;i = u (ξ)dξdtds.
hi Ii s xi−1/2

The right-hand side can be bounded using (3.30) and Lemma 4.16. We get
   2
ε u (xi−1/2 ) − ux̄;i  ≤ C ϑcd (ω̄) ,
[2]
126 4 Finite Difference Schemes for Convection-Diffusion Problems

and finally
   2
ε u (xi−1/2 ) − uN  [2]
x̄;i ≤ C ϑ cd (ω̄) .

This means that the midpoints of the mesh intervals are superconvergence points
for the derivative and we can define a recovery operator R for the derivative. For a
given mesh function v ∈ IRN +1 , let Rv be that function! that is piecewise linear on
the mesh ω̂ = 0, x1+1/2 , x2+1/2 , . . . , xN −1−1/2 , 1 and satisfies

(Rv)i−1/2 = vx̄;i for i = 1, . . . , N.

Then one can prove

   2
ε u − RuN ∞ ≤ C ϑcd (ω̄) .
[2]

A posteriori analysis
The next result is an extension of Lemma 2.2 in [64] which gave bounds in the
mesh points only. It is an essential ingredient for the analysis of second-order app-
roximations.
Theorem 4.39. Let the hypothesis of Theorem 3.45 be satisfied. Let ψ be the solu-
tion of the boundary value problem

Lc ψ = −Ψ  in (0, 1), u(0) = u(1) = 0

with

Ψ (x) = Ai−1/2 (x − xi−1/2 ) for x ∈ (xi−1 , xi ).

Then
 * +
  hi h2i
ψ∞ ∗
≤ C max  
Ai−1/2 min , ,
i=1,...,N b∞ 4ε
where

2b∞ + c∞ + β
C∗ = .
2β

Proof. Let x ∈ (0, 1) be arbitrary, but fixed. The Green’s function representation
gives
 1 
N
u(x) = ∂ξ G(x, ξ)F (ξ)dξ = Ai−1/2 Ji
0 i=1
4.3 Second-Order Difference Schemes 127

with

Ji := ∂ξ G(x, ξ)(ξ − xi−1/2 )dξ.
Ii

A first bound for the Ji ’s is

  
  hi
 
|Ji | ≤  ∂ξ G(x, ξ)(ξ − xi−1/2 )dξ  ≤ |∂ξ G(x, ξ)| dξ, (4.78)
Ii 2 Ii

while integration by parts yields

 * 2 +
h (ξ − xi−1/2 )2
Ji = ∂ξ2 G(x, ξ) i − dξ.
Ii 8 2

Hence

h2  2 
|Ji | ≤ i ∂ξ G(x, ξ) dξ
8 Ii

h2i ' (
≤ δx (ξ) + b∞ |∂ξ G(x, ξ)| + c(ξ)G(x, ξ) dξ.
8ε Ii

This estimate is combined with (4.78) to give

* +
h2i hi ' (
|Ji | ≤ min , δx (ξ) + b∞ |∂ξ G(x, ξ)| + c∞ G(x, ξ) dξ.
8ε 2b∞ Ii

 
Multiply by Ai−1/2 , take sums for i = 1, . . . , N , use a discrete Hölder inequality
and note that
 1' ( 2b∞ + c∞
δx (ξ) + b∞ |∂ξ G(x, ξ)| + c∞ G(x, ξ) dξ ≤ 1 + ,
0 β

by Theorem 3.20. This completes the proof. 



With these stability results at hand we can now derive our a posteriori error
bounds. We shall identify any mesh function v with its piecewise linear nodal inter-
polant.
Theorem 4.40. Let the hypothesis of Theorem 3.45 be satisfied. Set g := f − cuN .
Then the error of the defect-correction method satisfies
 
u − uN  ≤ η := η1 + η2 + η3 + η4 + η5
∞
128 4 Finite Difference Schemes for Convection-Diffusion Problems

with
 
hi h2i    
η1 := C ∗ max min , gi−1/2 + buN x̄,i  ,
i=1,...,N b∞ ε
 
  N
−1
hk+1 − hk 
1   1  
η2 := max hi (bΔ)x̄,i  , η3 := max  ck Δk  ,
β i=1,...,N β i=1,...,N −1  2 
k=i

1  3 
N
η4 := h g ∞,Ii
6β i=1 i

and

3  
η5 := max h2i 2 g  ∞,Ii + (buN ) ∞,I .
4β i=1,...,N i

Proof. By (4.70) and (4.71) we have, for x ∈ (xi−1 , xi ),

 
Ac u − uN (x)
' c N(   '  (
= F(x) − F6i + A
6u
i
6c Δ + α − a.
− Ac uN (x) − Ac − A i

Recalling the definitions of F, F6, Ac , Ac and A

6c , we obtain the representation
 
Ac u − uN (x)
 1  N  

N
bu i + buN i−1  N 
= g(s)ds − h̄k gk + − bu (x)
x 2 (4.79)
k=i

N −1
hi hk+1 − hk
− (bΔ)x̄,i − ck Δk + α − a.
2 2
k=i

Taylor expansions yield

 1 
N  1
g(s)ds − h̄k gk = (g − g I )(s)ds + (xi−1/2 − x)gi−1/2 + μi (x),
x k=i xi

where g I is the piecewise linear interpolant of g, and

   
buN i
+ buN i−1    
− buN (x) = (xi−1/2 − x) buN x̄,i + μ̃i (x)
2
4.3 Second-Order Difference Schemes 129

with

3h2i  3h2i  
(buN ) 
μi ∞,Ii ≤ g ∞,Ii and μ̃i ∞,Ii ≤ ∞,Ii
4 8
Substitute the above two equations into (4.79). We get
 
A u − uN (x)
 1
    
= (g − g I )(s)ds + xi−1/2 − x gi−1/2 + buN x̄,i
xi

N −1
hi hk+1 − hk
− (bΔ)x̄,i − ck δk + (μi + μ̃i ) (x) + α − a.
2 2
k=i

Furthermore,
 1 
  N
 (g − g I )(s)ds ≤ 1 h3i g  ∞,Ii .
  12
xi i=1

Finally, note that (Av) = −Lv. Use Theorems 3.45 and 4.39 to complete the
proof. 


Remark 4.41. The error estimate of Theorem 4.40 contains terms, namely η4 and
η5 , that in general have to be approximated, for example

gi − gi−1 gi − 2gi−1/2 + gi−1

g ≈ , g  ≈ 4
hi h2i

and

(buN )i − 2(buN )i−1/2 + (buN )i−1

(buN ) ≈ 4 .
h2i

The additional errors introduced this way are of third order and therefore decay
rapidly when the mesh is refined. ♣

An adaptive mesh algorithm

Based on Theorem 4.40, the de Boor algorithm described in Sect. 4.2.4.2 can be
adapted for the defect-correction method by choosing
 1/2

5
N
Qi = Qi (u , Δ, ω) := ρ0 + ρk ηk;i (4.80)
k=1
130 4 Finite Difference Schemes for Convection-Diffusion Problems

with
 
hi h2i     



η1;i := min , gi−1/2 + buN x̄,i  , η2;i := hi (bΔ)x̄,i  ,
b∞ 4ε
 
N−1
hk+1 − hk  2  
 
η3;i :=  ck Δk  , η4;i := gi − 2gi−1/2 + gi−1  ,
 2  3β
k=i
1  
η5;i := hi |gi − gi−1 | + (buN )i − 2(buN )i−1/2 + (buN )i−1 
2β

and non-negative weights ρ .

Remark 4.42. The square root in (4.80) is necessary because the underlying method
is formally of second order. ♣

Remark 4.43. The numerical experiments in [101] indicate that η1 contains suffi-
cient information to steer the mesh adaptation. Therefore ηk , k = 2, . . . , 5, can
be set to zero, however ρ0 must not in order to avoid mesh starvation in regions
where the solution does not vary much. This reduces the computational costs in the
remeshing phase of the de Boor algorithm. ♣

4.3.3.2 Richardson Extrapolation

Richardson extrapolation on layer-adapted meshes was first analysed by Natividad

and Stynes [124]. They study a simple first-order upwind scheme on a Shishkin
mesh and prove that Richardson extrapolation improves the accuracy to almost
second order, although the underlying scheme is only of first order. The analysis
in [124] is based on comparison principles and barrier function techniques.
Here we shall pursue an alternative approach similar to the one in [93] that is
based on the (l∞ , w−1,∞ ) stability and on the error expansion of Sect. 4.2.3. Again
consider the conservative form of our model problem:

Lc u := −εu − (bu) + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1 . (4.9)

Given an arbitrary mesh ω̄, let ω̄  : 0 = x1/2 < x1 < x1+1/2 < · · · < xN = 1 be
the mesh obtained by uniformly bisecting ω̄. Let ũN be the solution of the upwind
scheme (4.10) on ω̄ and
 
u2N = u2N 2N 2N 2N 2N
0 , u1/2 , u1 , . . . , uN −1/2 , uN

that of the difference scheme on ω̃. Since (4.10) is a first-order scheme we combine
ũN and ũ2N by

uN 2N
i := 2ũi − ũN
i for i = 0, . . . , N,

in order to get a second-order approximation defined on the coarser mesh ω̄.

4.3 Second-Order Difference Schemes 131

In addition to the notation introduced on p. 122 set

 −1
' c ( vi − vi−1
N
ck vk + ck+1/2 vk+1/2
à v i := 2ε + bi vi + hk+1
hi 2
k=i

and

N −1
fk + fk+1/2
F̃i := hk+1 .
2
k=i

The differential equation (4.9) and the difference equation (4.10) yield

Ac u − F ≡ α, Ac ũN − F ≡ a and Ãc ũ2N − F̃ ≡ ã.

A direct calculation gives

Ac (2ũ2N − uN − u)i + (Ac u − F)i−1/2

 
ui − ui−1 
= −ε − ui−1/2
hi
+ bi (ũ2N
i − ui ) − bi−1/2 (ũ2N
i−1/2 − ui−1/2 )

hi
− (cũ2N − cu)i−1/2
2
 −1 
N
' (
+ hk+1 (cũ2N − cu)k+1/2 − (cũ2N − cu)k
k=i
 1 
N −1 
hi
+ g(s)ds − gi−1/2 − hk+1 gk+1/2
i−1/2 2
k=i

[2]
with g = cu − f . The first term on the right-hand side is bounded by Cϑcd (ω̄)2 ,
[2]
see (4.75) The second and third term can be bounded by Cϑcd (ω̃)2 using the tech-
[2]
nique that gave (4.73). The last term is also bounded by Cϑcd (ω̄)2 , since similar
to (4.76) we have
  xk 
h h2k    2
 k  [2]
 gk−1/2 − gk−1/2 − g(s)ds ≤ Chk ϑcd (ω̄) ,
 2 8 xk−1/2 

and
  xk+1/2   2
 hk+1 h2k+1  
 g + g − g(s)ds ≤ Chk+1 ϑ[2] (ω̄) .
 2 k+1/2
8 k+1/2  cd
xk

Finally, using the stability inequality (4.8c) we obtain the following convergence
result.
132 4 Finite Difference Schemes for Convection-Diffusion Problems

Theorem 4.44. Let uN be the approximate solution to (4.9) obtained by the

Richardson extrapolation technique applied to the simple upwind scheme (4.10).
Then
   2
u − uN  [2]
≤ C ϑcd (ω̄) .
ε,∞,ω

Corollary 4.45. Theorem 4.44 and interpolation error bounds (see Sect. 5.1) give
     2
u − uN  + ε u − RuN  ≤ C ϑ[2] (ω̄) ,
∞ ∞ cd

where R is the recovery operator from Sect. 4.3.3.1.

4.3.4 A Numerical Example

The following tables contain the results of test computations for test problem (4.14):

−εu − u + 2u = ex−1 , u(0) = u(1) = 0.

We test the performance of:

• second-order upwinding
• central differencing
• defect correction and
• Richardson extrapolation.
In the experiments we have chosen ε = 10−8 . The meshes have been constructed
with parameters σ = 2, β = 1 and q = K = 1/2.
The numerical results in Tables 4.3 and 4.4 are clear illustrations of the conver-
gence estimates of Theorems 4.31, 4.38 and 4.44 and of (4.67). Furthermore, as the
theory predicts, all four methods give higher accuracy on Bakhvalov meshes than
on Shishkin meshes.
Finally, we consider a modified Shishkin mesh which is constructed as fol-
lows. Pick the transition point τ = 2εβ −1 ln N as usual. Set h = 2τ /N and
H = 2(1 − τ )/N . Then the mesh is defined by
⎧
⎪
⎪ if i ≤ N/2,
⎨h
hi = 4H/3 if i is odd and i > N/2,
⎪
⎪
⎩2H/3 if i is even and i > N/2.

Thus, instead of a uniform mesh on each of the two subdomains [0, τ ] and [τ, 1] we
use non-uniform, though very regular sub meshes. A similar mesh was considered
in [26].
4.3 Second-Order Difference Schemes 133

Table 4.3 Second-order schemes on Shishkin meshes

second order central defect Richardson
N upwinding differencing correction extrapolation
27 3.29e-04 1.61 1.37e-04 1.55 2.32e-04 1.64 2.33e-04 1.57
28 1.07e-04 1.66 4.66e-05 1.62 7.47e-05 1.68 7.86e-05 1.64
29 3.40e-05 1.70 1.52e-05 1.67 2.33e-05 1.72 2.53e-05 1.69
210 1.05e-05 1.73 4.79e-06 1.70 7.06e-06 1.75 7.86e-06 1.72
211 3.17e-06 1.75 1.47e-06 1.73 2.10e-06 1.78 2.38e-06 1.75
212 9.43e-07 1.77 4.43e-07 1.76 6.12e-07 1.79 7.07e-07 1.77
213 2.77e-07 1.79 1.31e-07 1.78 1.76e-07 1.81 2.07e-07 1.79
214 8.02e-08 1.80 3.83e-08 1.79 5.03e-08 1.82 5.97e-08 1.81
215 2.30e-08 1.81 1.11e-08 1.81 1.42e-08 1.83 1.71e-08 1.82
216 6.54e-09 1.83 3.16e-09 1.82 4.00e-09 1.84 4.84e-09 1.83
217 1.85e-09 1.84 8.95e-10 1.83 1.12e-09 1.85 1.36e-09 1.84
218 5.17e-10 — 2.52e-10 — 3.10e-10 — 3.81e-10 —

Table 4.4 Second-order schemes on Bakhvalov meshes

second order central defect Richardson
N upwinding differencing correction extrapolation
27 5.81e-04 2.36 9.67e-05 1.99 2.74e-04 1.94 1.16e-04 1.94
28 1.13e-04 2.11 2.43e-05 2.02 7.14e-05 1.97 3.01e-05 1.97
29 2.62e-05 2.05 5.98e-06 1.99 1.83e-05 1.98 7.67e-06 1.99
210 6.34e-06 2.27 1.51e-06 2.00 4.63e-06 1.99 1.93e-06 1.99
211 1.31e-06 2.36 3.77e-07 2.00 1.16e-06 2.00 4.86e-07 2.00
212 2.55e-07 2.11 9.45e-08 2.01 2.92e-07 2.00 1.22e-07 2.00
213 5.90e-08 2.05 2.34e-08 1.99 7.30e-08 2.00 3.05e-08 2.00
214 1.43e-08 2.05 5.89e-09 2.00 1.83e-08 2.00 7.63e-09 2.00
215 3.46e-09 2.00 1.47e-09 2.00 4.57e-09 2.00 1.91e-09 2.00
216 8.64e-10 2.00 3.68e-10 2.01 1.14e-09 2.00 4.77e-10 2.00
217 2.16e-10 2.00 9.16e-11 2.00 2.86e-10 2.00 1.19e-10 2.00
218 5.40e-11 — 2.30e-11 — 7.14e-11 — 2.98e-11 —

For this mesh ϑcd (ω̄) ≤ CN −1 ln N and almost second order convergence is
[2]

guaranteed for the upwind scheme, defect correction and Richardson extrapola-
tion. This order of convergence is observed in our computational experiments; see
Table 4.5. However, for central differencing the observed rate is only one. Thus,
on this mesh the assumption (4.65) must be violated. In [26] it is shown that for
b = const and c ≡ 0 the stability constant in (4.66) blows up for N → ∞.

Remark 4.46. This means that for central differencing a general result like
   2
u − uN  ≤ C ϑ
[2]
(ω̄)
∞,ω cd

cannot hold. ♣
134 4 Finite Difference Schemes for Convection-Diffusion Problems

Table 4.5 Second-order schemes on modified Shishkin meshes

second order central defect Richardson
N upwinding differencing correction extrapolation
27 1.30e-03 1.61 6.73e-03 0.98 5.94e-04 1.46 7.66e-04 1.46
28 4.27e-04 1.66 3.41e-03 0.99 2.16e-04 1.57 2.77e-04 1.58
29 1.35e-04 1.70 1.72e-03 0.99 7.25e-05 1.65 9.25e-05 1.65
210 4.17e-05 1.72 8.63e-04 1.00 2.31e-05 1.70 2.94e-05 1.70
211 1.26e-05 1.75 4.32e-04 1.00 7.11e-06 1.74 9.04e-06 1.74
212 3.76e-06 1.77 2.16e-04 1.00 2.13e-06 1.77 2.71e-06 1.76
213 1.10e-06 1.79 1.08e-04 1.00 6.24e-07 1.79 7.97e-07 1.79
214 3.20e-07 1.80 5.41e-05 1.00 1.81e-07 1.81 2.31e-07 1.80
215 9.18e-08 1.81 2.70e-05 1.00 5.16e-08 1.82 6.63e-08 1.82
216 2.61e-08 1.82 1.35e-05 1.00 1.46e-08 1.83 1.88e-08 1.83
217 7.37e-09 1.84 6.74e-06 1.02 4.11e-09 1.84 5.31e-09 1.84
218 2.07e-09 — 3.33e-06 — 1.15e-09 — 1.49e-09 —

This observation has far reaching consequences, in particular in higher dimen-

sions, where it is difficult, if not impossible, to construct uniform or nearly uniform
meshes. Therefore, stabilisation in regions where the mesh is coarse becomes
essential.

4.4 Systems

We now leave the scalar convection-diffusion equation and move on to systems of

equations of this type.

4.4.1 Weakly Coupled Systems in One Dimension

Consider the weakly coupled problem from Sect. 3.4.2:

Lu := − diag(ε)u − diag(b)u + Au = f on (0, 1),

(4.81)
u(0) = u(1) = 0,

where ε = (ε1 , . . . , ε )T and the small parameter εk is in (0, 1] for k = 1, . . . , .

Assume that for each k one has akk ≥ 0 and either bk ≥ βk or bk ≤ −βk on [0, 1]
with positive constants bk .
We follow [96] and discretise (4.81) by means
 ofthe simple upwind scheme that
was studied in detail in Sect. 4.2: Find u ∈ IR0N +1 such that
' N(
Lu i = f i for k = 1, . . . , N − 1, (4.82)
4.4 Systems 135

T  T
where f i = f (xi ) = (f1;i , f2;i , . . . , f ,i ) , Lv := (Lv)1 , (Lv)2 , . . . , (Lv) ,


(Lv)k := Λk vk + akm vm
m=1
m=k

and

−εk vx̄x;i − bk;i vx;i + akk;i vi if bk > 0,
[Λk v]i :=
−εk vxx̄;i − bk;i vx̄;i + akk;i vi if bk < 0.

4.4.1.1 Stability

The stability analysis for the discrete operator is conducted along the lines of the
continuous analysis. By the  of L and the Λk ’s we have, for any vector-
 definition
valued mesh function v ∈ IR0N +1 ,


Λ k vk = − akm vm + (Lv)k on ω, k = 1, . . . , . (4.83)
m=1
m=k

Then Theorem 4.3 yields

    
  (Lv)k   (Lv)k 
vk ∞,ω + γ̃km um ∞,ω 
≤ min   ,  
m=1
akk ∞,ω  bk 
∞,ω
m=k

for, k = 1, . . . , , where the  ×  constant matrix Γ̃ = Γ̃ (A, b) = (γ̃km ) is—as in

Sect. 3.4.2—defined by
    
 akm   akm 
γ̃kk = 1 and γ̃km = − min  
 akk  , bk 
 for k = m.
∞,ω ∞,ω

We reach the following stability result.

Theorem 4.47. Assume that the matrix A has non-negative diagonal entries. As-
sume that Γ̃ (A) is inverse-monotone. Then for i = 1, . . . ,  one has
    
 −1   (Lv)k   (Lv)k 
vi ∞,ω̄ ≤ Γ̃ ik
min  akk 
 ,  bk 

k=1 ∞,ω ∞,ω

 
for any mesh function v ∈ IR0N +1 .
136 4 Finite Difference Schemes for Convection-Diffusion Problems

Remark 4.48. Theorem 4.47 implies

 
v∞,ω̄ := max vk ∞,ω̄ ≤ C Lv∞,ω for all v ∈ IR0N +1 ,
k=1,...,

i.e., the operator L is (∞ , ∞ )-stable although it does not obey a comparison
principle. ♣

Corollary 4.49. Under the hypotheses of Theorem 4.47 the discrete problem (4.81)
has a unique solution uN , and
 N
u  ≤ C f ∞,ω
∞,ω̄

for some constant C.

Remark 4.50. One can also use (4.8c) when bounding (Lv)k in (4.83) to establish
that
 
v∞,ω̄ ≤ C max (Lv)k −1,∞ for all v ∈ IR0N +1 .
k=1,...,

This allows to analyse the difference scheme (4.82) when applied to problems whose
source terms consist of generalised functions like the δ-distribution. ♣

4.4.1.2 A Priori Error Analysis

Following [96], we split the error η := u − uN into two parts ψ, ϕ ∈ (IR0N +1 )

as η = ψ + ϕ with


Λk ψk = (Lη)k and Λ k ϕk = − akm ηm on ω, k = 1, . . . , .
m=1
m=k

A triangle inequality and Theorem 4.3 yield


ηk ∞,ω ≤ ψk ∞,ω − γ̃km ηm ∞,ω , k = 1, . . . , .
m=1
m=k

Assuming that Γ̃ is an M -matrix, we obtain

 
u − uN  ≤ C ψ∞,ω ,
∞,ω

and we are left with bounding ψ.

The components of ψ are the solutions of scalar problems to which the technique
of Sect. 4.2.2 can be applied. The following general error bound is obtained.
4.4 Systems 137

Theorem 4.51. Assume that the matrix A has non-negative diagonal entries.
Assume that Γ̃ (A) is inverse-monotone. Let u and uN be the solutions of (4.81)
and (4.82). Then
 *  +
 
u − uN  ≤ C max 1+ |um (s)| ds.
∞,ω̄ i=1,...,N Ii m=1

Corollary 4.52. The a priori bounds on the um in Sect. 3.4.2.2 can be used to derive
more explicit error bounds:
 
u − uN  [1]
≤ Cϑcd, (ω̄).
∞,ω̄

[p]
Note, the quantity ϑcd, (ω̄), p > 0, has been defined in Sect. 2.1.6.

Remark 4.53. We immediately obtain, for example,


  CN −1 for Bakhvalov meshes,
u − uN  ≤
∞,ω −1
CN ln N for Shishkin meshes,

when the mesh parameters satisfy σm ≥ 1. ♣

4.4.1.3 A Posteriori Error Analysis

Alternatively, one can appeal to the strong stability (3.29c) of the scalar continuous
operators and combine the arguments of Sect. 4.2.4 and 4.4.1.1, in order to get the
a posteriori bound
*  +
  
u − uN  ≤ C max hk+1 1 + u N 
m,x;k .
∞ k=0,...,N −1
m=1

The constant(s) involved in this error bound can be specified explicitly; cf.
Sect. 4.2.4.

4.4.2 Strongly Coupled Systems

We now consider strongly coupled systems of convection-diffusion type, i.e., for

each k one has bkm ≡ 0 for some k = m. Strong coupling causes interactions
between boundary layers that are not fully understood at present. The main papers
on this problem are [1, 100, 127, 128].
138 4 Finite Difference Schemes for Convection-Diffusion Problems

The general strongly coupled two-point boundary-value problem in conservative

 
form is: Find u ∈ C 2 (0, 1) ∩ C[0, 1]) such that

Lc u := − diag(ε)u − (Bu) + Au = f in Ω := (0, 1),

(4.84)
u(0) = u(1) = 0,

where as before f = (f1 , . . . , f )T , while A = (akm ) and B = (bkm ) are  × 

matrices, and the  ×  matrix diag(ε) is diagonal with k th entry εk for all k. Fur-
thermore, let bkk ≥ βk or bkk ≤ −βk on [0, 1] with positive constants βk .
We follow [100] and discretise (4.84) using the upwind scheme of Sect. (4.2) for
each equation of the system: Find uN ∈ (IR0N +1 ) such that
' c N(
L u i = fi for i = 1, . . . , N − 1, (4.85)

 T
where Lc v := (Lc v)1 , (Lc v)2 , . . . , (Lc v) ,

 
(Lc v)k := Λck vk − (bkm vm )x + akm vm if bkk > 0,
m=1 m=1
m=k m=k

 
(Lc v)k := Λck vk − (bkm vm )x̄ + akm vm if bkk < 0,
m=1 m=1
m=k m=k

and

−εk vx̄x;i − (bkk v)x;i + akk;i vi if bkk > 0,
[Λck v]i :=
−εk vxx̄;i − (bkk v)x̄;i + akk;i vi if bkk < 0.

4.4.2.1 Stability

The stability analysis for the difference operator L is analogous to that for the con-
tinuous operator L in Sect. 3.4.3.
Define the  ×  matrix Υ ω = Υ ω (A, B) = (γkm ) by

2 bkm ∞,ω + akm 1,ω

γkk = 1 and γkm = − for k = m.
βk

Introduce the discrete maximum norms

|||v|||εk ,∞,ω := εk |[vx ||∞,ω + βk v∞,ω̄ for v ∈ IRN +1

4.4 Systems 139

and
 
|||v|||ε,∞,ω := max |||vk |||εk ,∞,ω for v ∈ IRN +1 .
k=1,...,

Theorem 4.54. Assume that for each k = 1, . . . , 

bkk ≤ −βk or bkk ≥ βk on [0, 1]

with positive constants βk and that

akk ≥ 0 and bkk ≥ 0 on [0, 1].

Suppose Υ ω (A, B) is inverse-monotone. Then the operator Lc is (∞ , w−1,∞ )

stable with

 
|||vk |||εk ,∞,ω ≤ Υ −1
ω km
(Lc v)m −1,∞,ω for k = 1, . . . , ,
m=1

and for all v ∈ (IR0N +1 ) .

Proof. For the sake of simplicity in the presentation, we restrict ourselves to the
case when bkk > 0 for all k.
Let v ∈ (IR0N +1 ) be arbitrary. Then the definition of Lc and the Λck yields

 
Λck vk = (Lc v)k + (bkm vm )x − akm vm
m=1 m=1
m=k m=k

Apply Theorem 4.3 and Remark 4.4 to get

 
|||vk |||εk ,∞,ω ≤ 2 bkm ∞,ω + akm 1,ω vm ∞,ω + (Lc v)k −1,∞,ω .
m=1
m=k

Recall the definition of Υ ω and rearrange the last inequality


γkm |||vm |||εm ,∞,ω ≤ (Lc v)k −1,∞,ω for k = 1, . . . , .
m=1

Using the inverse monotonicity of Υ ω , we are finished. 


140 4 Finite Difference Schemes for Convection-Diffusion Problems

Corollary 4.55. Suppose the hypotheses of Theorem 4.54 are satisfied. Then the
difference equation (4.85) possesses a unique solution uN , with
 N 
u  ≤ C max fm −1,∞,ω
ε,∞,ω m=1,...,

for some constant C that is independent of both ε and the mesh.

Remark 4.56. In general, the operator Lc does not obey a comparison principle.
Nonetheless it is (∞ , ∞ )-stable, i.e.,
 
v∞,ω̄ ≤ C Lc v∞,ω for all v ∈ IR0N +1 ,

by Theorem 4.54. ♣

4.4.2.2 A Priori Error Analysis

Adapt the argument of Sect. 4.4.1.2 as in [100] and split the error η := uN − u into
two parts ψ, ϕ ∈ (IR0N +1 ) as η = ψ + ϕ with


Λck ψk = (Lη)k and Λck ϕk = ((bkm ηm )x − akm ηm ) on ω.
m=1
m=k

Recalling the definition of Υ ω = (γkm ), we use a triangle inequality and Theo-

rem 4.3 to obtain


γkm |||ηm |||εm ∞,ω ≤ |||ψk |||εk ,∞,ω for k = 1, . . . , .
m=1

Next, if Υ is inverse monotone then

 
u − uN  ≤ C |||ψ|||ε,∞,ω ,
ε,∞,ω

and we are left with bounding |||ψk |||εk ,∞,ω , k = 1, . . . , .

The components of ψ are the solutions of scalar problems. In [100] the technique
of Sect. 4.2.2 is used to obtain the following general error bound.
Theorem 4.57. Let the hypothesis of Theorem 4.54 be satisfied. Then the error of
the upwind scheme (4.85) applied to (4.84) satisfies
 *  +
 
u − uN  ≤ C max 1+ 
|um (s)| ds.
ε,∞,ω i=1,...,N Ii m=1
4.4 Systems 141

Corollary 4.58. By (3.56) we have uk 1 ≤ C for k = 1, . . . , . Therefore, there

exists a mesh ω ∗ such that
       
1 1
1+ |um (x)| dx = 1+ |um (x)| dx ≤ CN −1
Ii m=1
N 0 m=1

and on this mesh one consequently has

 
u − uN  ≤ CN −1 .
ε,∞,ω ∗

Remark 4.59. As satisfactory pointwise bounds on |uk | are unavailable, this result
does not give an immediate explicit convergence result on, e.g., a Bakhvalov or
Shishkin mesh. ♣
Remark 4.60. When εk = ε for k = 1, . . . , , Theorem 3.57 yields
  
 
u − uN  ≤ C max 1 + e −βx/ε [1]
dx = Cϑcd (ω̄).
ε,∞,ω i=1,...,N Ii

The system behaves like the scalar equation of Section 4.2 and appropriately adapted
meshes can be constructed as for scalar problems.
In [127] one also finds an error analysis for a system with a single parameter,
but the analysis is limited to Shishkin meshes and uses a more traditional truncation
error and barrier function argument. Furthermore, higher regularity of the solution
is required. On the other hand, in certain situations the analysis of [127] is valid
under less restrictive hypotheses on the entries of the matrices A and B than the
requirement that Υ ω be inverse-monotone. ♣

4.4.2.3 A Posteriori Error Bounds

Using the strong stability results of Theorem 3.54, we can follow [100] to obtain the
a posteriori error bound
  
 
u − uN  ≤ C max hi 1 + |uN
m;x̄;i | .
ε,∞ i=1,...,N
m=1

Remark 4.61. The de Boor algorithm (see Sect. 4.2.4.2) can be used to adaptively
generate meshes for (4.84) by choosing
  
N
2 1/2
Qi = hi 1+ um;x̄;i .
m=1

Numerical examples are presented in [100], but a complete analysis of the adaptive
algorithm is not given. ♣
142 4 Finite Difference Schemes for Convection-Diffusion Problems

4.4.2.4 Numerical Results

We now present the results of two numerical experiments in order to illustrate the
conclusions of Theorem 4.57.
First, we consider a test problem with two equations.
     1−2x 
4 + x −1 0 −x2 e
− diag (ε) u − u + u= in (0, 1), (4.86)
1 − 2x −3 1 1 cos 2x

subject to homogeneous Dirichlet boundary conditions u(0) = u(1) = 0.

For this problem our convergence theory applies since
   
1 12 −7 −1 1 12 7
Γ = and Γ = ≥ 0.
12 −12 12 5 12 12

We have b1 ≥ 4 and −b2 ≥ 3. Therefore, one expects that the solution exhibits
two layers: one at x = 0 behaving like e−4x/ε1 and the other at x = 1 that behaves
like e−3(1−x)/ε2 . Also the first-order derivative of u is expected to satisfy

|ui (x)| ≤ C 1 + ε−1

1 e
−4x/ε1
+ ε−1
2 e
−3(1−x)/ε2
, i = 1, 2.

Note, we do not have proper proof for these derivative bounds. The difficulties in
proving them have been explained in Section 3.4.3.1.
The exact solution of (4.86) is not available. Therefore, we compare the
numerical solution with that obtained by Richardson extrapolation as before.
We consider Bakhvalov and Shishkin meshes and the adaptive de Boor algorithm;
see Remark 4.61. The construction of layer-adapted meshes for overlapping layers
is explained in Section 2.1.6.
The results of our test computations are contained in Table 4.6. For both a priori
adapted meshes the expected (almost) first order of uniform convergence is con-
firmed. For the adaptive algorithm first order is also observed although the numerical
rates are “less stable”.

Table 4.6 Simple upwinding for a system of two convection-diffusion equations

N Shishkin mesh Bakhvalov mesh adaptive algorithm
ηN ρN ηN rN ηN rN
3·2 7
3.131e-02 0.83 8.194e-03 0.99 7.600e-03 0.96
3 · 28 1.843e-02 0.88 4.119e-03 0.99 3.903e-03 1.05
3 · 29 1.051e-02 0.91 2.073e-03 0.99 1.880e-03 0.99
3 · 210 5.870e-03 0.94 1.044e-03 0.99 9.477e-04 0.98
3 · 211 3.229e-03 0.95 5.240e-04 1.00 4.821e-04 0.94
3 · 212 1.758e-03 0.97 2.620e-04 1.00 2.505e-04 1.00
3 · 213 9.491e-04 0.98 1.308e-04 1.00 1.254e-04 1.15
3 · 214 5.092e-04 0.98 6.540e-05 1.00 5.643e-05 0.91
3 · 215 2.717e-04 0.99 3.269e-05 1.00 3.006e-05 0.99
3 · 216 1.444e-04 — 1.634e-05 — 1.516e-05 —
4.5 Problems with Turning Point Layers 143

Table 4.7 Simple upwinding for a convection-diffusion problem with three equations
N Shishkin mesh Bakhvalov mesh adaptive algorithm
ηN ρN ηN rN ηN rN
4·2 3
2.453e-01 0.49 1.780e-01 0.86 1.850e-01 0.64
4 · 24 1.809e-01 0.63 9.777e-02 0.93 1.187e-01 0.75
4 · 25 1.218e-01 0.75 5.124e-02 0.97 7.062e-02 0.87
4 · 26 7.612e-02 0.84 2.623e-02 0.98 3.863e-02 0.96
4 · 27 4.505e-02 0.90 1.327e-02 0.99 1.991e-02 0.92
4 · 28 2.569e-02 0.94 6.673e-03 1.00 1.051e-02 1.02
4 · 29 1.430e-02 0.97 3.346e-03 1.00 5.167e-03 0.97
4 · 210 7.824e-03 0.98 1.676e-03 1.00 2.642e-03 1.06
4 · 211 4.233e-03 1.00 8.384e-04 1.00 1.270e-03 1.05
4 · 212 2.272e-03 — 4.193e-04 — 6.115e-04 —

The second test problem consists of three convection-diffusion equations.

⎛ ⎞ ⎛ x ⎞
3 1 0 e
− diag (ε) u − −x 5 + x −1 u = cos x ⎠ in (0, 1),
 ⎝ 2 ⎠  ⎝
1 − x 0 −5 1 + x2

with boundary conditions u(0) = u(1) = 0. This time:

⎛ ⎞ ⎛ ⎞
15 10 0 225 150 60
1 ⎝ 1 ⎝
Γ = 7 15 6 ⎠ and Γ −1 = 159 225 90 ⎠ ≥ 0.
15 119
9 0 15 135 90 155

We expect layers e−3x/ε1 , e−5x/ε2 and e−5(1−x)/ε3 to form and adapt the mesh
accordingly.
Table 4.7 gives the numerical results for our second example. A comparison
with Table 4.6 reveals that the behaviour of the method is similar to that for
Example (4.86).

4.5 Problems with Turning Point Layers

This section considers linear convection-diffusion problems with a boundary turning

point: Find u ∈ C 2 (0, 1) ∩ C[0, 1] such that

Lu := −εu − pbu + pcu = f in (0, 1), u(0) = γ0 , u(1) = γ1 , (4.87)

where p(x) = xκ , κ > 0, b ≥ β > 0 and c ≥ 0 in (0, 1).

We are aware of four publications analysing numerical methods for this prob-
lem with κ = 1. Liseikin [113] constructs a special transformation and solves the
144 4 Finite Difference Schemes for Convection-Diffusion Problems

transformed problem on a uniform mesh. The method obtained is proved to be

first-order uniformly convergent in the discrete maximum norm. Vulanović [167]
studies an upwind-difference scheme on a layer-adapted Bakhvalov-type mesh and
proves convergence in a discrete 1 norm. This result is generalised in [170] for
quasilinear problems. However, this norm fails to capture the layers present in
the solution. Therefore, the problem is not singularly perturbed in the sense of
Def. 1.1. In [112] the authors establish almost first-order convergence for an upwind-
difference scheme on a Shishkin mesh. Here we follow [92] and study (4.87) with
arbitrary κ > 0.

4.5.1 A First-Order Upwind Scheme

The boundary-value problem (4.87) is discretised using simple upwinding: Find

u ∈ IRN +1 such that
' N(
Lu i = pi fi for i = 1, . . . , N − 1, uN N
0 = γ0 , uN = γ1 (4.88)

with

[Lv]i := −εvx̄x;i − pi bi vx;i + pi ci vi .

4.5.1.1 Stability of the Discretisation

The matrix associated with it is an L0 -matrix. Lemma 3.14 with the test function
ei = 1 − xi verifies that it is an M -matrix. Therefore, the operator L satisfies a
discrete comparison principle. That is, for any mesh functions v, w ∈ IRN +1
⎫
Lv ≤ Lw on ω, ⎬
v0 ≤ w0 , =⇒ v≤w on ω̄.
⎭
vN ≤ wN

Lemma 3.17 with ei = 1 − xi yields

v∞,ω̄ ≤ Lv/pb∞,ω for all v ∈ IR0N +1 .

Alternatively, if c > 0 on [0, 1], then Lemma 3.17 with e ≡ 1 gives

v∞,ω̄ ≤ Lv/pc∞,ω for all v ∈ IR0N +1 .

Thus, the operator L is (∞ , ∞,1/p ) stable in the 1/p-weighted maximum norm.
4.5 Problems with Turning Point Layers 145

For the solution uN of (4.88) this implies

    
 N ! f  f 
ui  
≤ max |γ0 |, |γ1 | + min   ,   .
∞,ω̄ b ∞,ω  c ∞,ω

Green’s function.
Lemma 4.62. Assume that

p > 0, p ≥ 0, b ≥ β > 0 and c ≥ 0 on (0, 1). (4.89)

Then
⎧
⎪
⎪1 for i = 0, . . . , j,
1 ⎨ 3 i  −1
0 ≤ Gi,j ≤ Ĝi,j := βpν hν+1
pj β ⎪
⎪ 1+ for i = j + 1, . . . , N.
⎩ ε
ν=j+1

Proof. Let j be arbitrary, but fixed. G·,j solves

[LG·,j ]i = δi,j , i = 1, . . . , N − 1, G0,j = GN,j = 0

with

h−1
i+1 for i = j,
δi,j =
0 otherwise.

We shall show that Ĝ·,j is a barrier function for G·,j . Clearly Ĝ0,j ≥ 0 and
ĜN,j ≥ 0.
Next, verify that
⎧
⎨0 for i = 0, . . . , j − 1,
Ĝx;i,j = Ĝi,j βpi
⎩− for i = j, . . . , N − 1,
ε + βpi hi+1

and
⎧
⎨0 for i = 1, . . . , j,
Ĝx̄;i,j = Ĝi,j
⎩− βpi−1 for i = j + 1, . . . , N.
ε
Hence,
' (
LĜ·,j i
≥ pi ci Ĝi,j ≥ 0 for i = 1, . . . , j − 1,
146 4 Finite Difference Schemes for Convection-Diffusion Problems
 
' ( βpj 1
LĜ·,j j ≥ + pj cj Ĝj,j ≥ ,
hj+1 hi+1

and
 
' ( pi − pi−1
LĜ·,j i ≥ β + pi ci Ĝi,j ≥ 0 for i = j + 1, . . . , N − 1,
hi+1

because p ≥ 0 on [0, 1]. Thus, Ĝ is a barrier function for G. 



Remark 4.63. The proof of Lemma 4.62 simplifies the argument in [112] where a
barrier function was constructed for the adjoint problem. ♣

Remark 4.64. Numerical results indicate that when p(x) = xκ , κ ≥ 0, one has the
sharper bound
 −κ
Gi,j ≤ C ε1/(κ+1) + ξj for all i, j = 1, . . . , N − 1,

but, we do not have a rigorous proof for this.

Theorem 4.65. Assume the data satisfies (4.89). Then the discrete operator L is
(∞ , 1,1/p ) stable

v∞,ω ≤ β −1 Lv/p1,ω

for all v ∈ IR0N +1 with a 1/p-weighted 1 norm.

Proof. For any function v ∈ IR0N +1 we have


N
vi = hj+1 Gi,j [Lv]j , i = 1, . . . , N − 1.
j=1

Then Lemma 4.62 yields the assertion of the theorem. 



Remark 4.66. An immediate consequence of Theorem 4.65 for the simple upwind
scheme is
 
u − uN  ≤ β −1 (Lu − pf ) /p1,ω .
∞,ω

Thus, the error of the numerical solution in the maximum norm is bounded by an
1 -type norm of the truncation error weighted with the inverse of the coefficient of
the convection term. This was used in [112] to establish uniform almost first-order
convergence on Shishkin meshes for κ = 1. ♣
4.5 Problems with Turning Point Layers 147

4.5.2 Convergence on Shishkin Meshes

In [112] convergence of the upwind scheme (4.88) applied to (4.87) with κ = 1

was studied. Starting from the observation that for any fixed m > 0 there exists a
constant C = C(m) such that
   
β̃x2 x
exp − ≤ C exp −m √
2ε ε

a piecewise uniform mesh is constructed as follows: Fix the mesh transition point
 √ 
2 ε
τ = min q, ln N .
m

Then divide [0, τ ] uniformly into qN subintervals and [τ, 1] into (1 − q)N subinter-
vals.
Using the stability inequality of Theorem 4.65, it is then shown that
 
u − uN  ≤ CN −1 (ln N ) .
2
∞,ω

The details of the analysis are similar to the argument in Sect. 4.2.5.
The general case of an arbitrary κ > 0 has been considered in [92]. We will give
a brief summary of that paper now. This time the transition point τ is chosen as
follows:
  1/(κ+1) 
ε(κ + 1)
τ = min q, σ ln N .
β̃

Here we shall consider τ < q which is the interesting case.

First Lemma 4.62 is sharpened to

Cτ ε−1 for j = 1, . . . , qN − 1,
Gi,j ≤
β −1 ξj−κ for j = qN, . . . , N − 1.

Note that τ ε−1 = Cμ−κ (ln N )

1/(κ+1)
.

Remark 4.67. It is argued in [92] based on numerical evidence that the logarithmic
factor is superfluous and one has
−κ
Gi,j ≤ C (μ + ξj ) for i, j = 1, . . . , N − 1,

but no rigorous analysis is provided. ♣

148 4 Finite Difference Schemes for Convection-Diffusion Problems

The analysis in [92] proceeds along the lines of Sect. 4.2.5 using the solution
decomposition in Theorem 3.63 to establish
 
u − uN  ≤ CN −1 (ln N )
2/(κ+1)
(4.90)
∞,ω

if σ ≥ 2.
−κ
Remark 4.68. If one had Gi,j ≤ C (μ + ξj ) , then (4.90) could be sharpened to
 
u − uN  ≤ CN −1 (ln N )
1/(κ+1)
.
∞,ω

Also we do not have any theory for arbitrary meshes. This is due to a lack of stronger
negative-norm stability inequalities for both the continuous and the discrete opera-
tors. More work in this direction is required. ♣

4.5.3 A Numerical Example

We verify experimentally the convergence result of (4.90). Our test problem is the
semilinear differential equation

−εu (x) − xκ (2 − x)u (x) + xκ eu(x) = 0 for x ∈ (0, 1),

u(0) = u(1) = 0.

The exact solution of this problem is not available. We therefore estimate the ac-
curacy of the numerical solution by comparing it with the numerical solution on
a higher order method: Richardson extrapolation. For our tests we take β̃ = 1,
q = 1/2 and ε = 10−12 .
Table 4.8 displays the results of the numerical test. For comparison rea-
sons Table 4.9 contains the rates which can be expected if the error bound

Table 4.8 Simple upwinding on Shishkin meshes for turning point problems
κ = 1/2 κ=1 κ=2 κ=3
N error rate error rate error rate error rate
27 6.171e-3 0.88 5.335e-3 0.92 4.675e-3 0.95 4.411e-3 0.96
28 3.358e-3 0.90 2.829e-3 0.93 2.426e-3 0.96 2.270e-3 0.97
29 1.803e-3 0.91 1.484e-3 0.94 1.249e-3 0.96 1.160e-3 0.97
210 9.592e-4 0.92 7.737e-4 0.95 6.401e-4 0.97 5.899e-4 0.98
211 5.069e-4 0.93 4.014e-4 0.95 3.269e-4 0.97 2.993e-4 0.98
212 2.666e-4 0.93 2.075e-4 0.96 1.666e-4 0.98 1.516e-4 0.98
213 1.396e-4 0.94 1.070e-4 0.96 8.473e-5 0.98 7.669e-5 0.98
214 7.292e-5 0.94 5.506e-5 0.96 4.305e-5 0.98 3.876e-5 0.99
215 3.798e-5 0.94 2.828e-5 0.96 2.185e-5 0.98 1.958e-5 0.99
216 1.973e-5 — 1.451e-5 — 1.108e-5 — 9.881e-6 —
4.5 Problems with Turning Point Layers 149

Table 4.9 Expected “convergence rates” for (ln N )p/(κ+1) N −1

κ = 1/2 κ=1 κ=2 κ=3
N p=2 p=1 p=2 p=1 p=2 p=1 p=2 p=1
27 0.74 0.87 0.81 0.90 0.87 0.94 0.90 0.95
28 0.77 0.89 0.83 0.92 0.89 0.94 0.92 0.96
29 0.80 0.90 0.85 0.92 0.90 0.95 0.92 0.96
210 0.82 0.91 0.86 0.93 0.91 0.95 0.93 0.97
211 0.83 0.92 0.87 0.94 0.92 0.96 0.94 0.97
212 0.85 0.92 0.88 0.94 0.92 0.96 0.94 0.97
213 0.86 0.93 0.89 0.95 0.93 0.96 0.95 0.97
214 0.87 0.93 0.90 0.95 0.93 0.97 0.95 0.98
215 0.88 0.94 0.91 0.95 0.94 0.97 0.95 0.98

N −1 for p = 2 and p = 1. The rates observed are closer to

p/(κ+1)
is (ln N )
those expected for p = 1. This supports the hypothesis of Remark 4.68.
Chapter 5
Finite Element and Finite Volume Methods

In this chapter we consider finite element and finite volume discretisations of

Lu := −εu − bu + cu = f in (0, 1), u(0) = u(1) = 0, (5.1)

with b ≥ β > 0. Its associated variational formulation is: Find u ∈ H01 (0, 1) such
that

a(u, v) = f (v) for all v ∈ H01 (0, 1), (5.2)

where

a(u, v) := ε (u , v  ) − (bu , v) + (cu, v)

and
 1 
f (v) := (f, v) := f v (x)dx. (5.3)
0

Throughout assume that

c + b /2 ≥ γ > 0. (5.4)

This condition guaranties the coercivity of the bilinear form in (5.2):

|||v|||ε := ε v  0 + γ v0 ≤ a(v, v)

2 2 2
for all v ∈ H01 (0, 1).

This is verified using standard arguments, see e.g. [141]. If b ≥ β > 0 then (5.4) can
always be ensured by a transformation ū(x) = u(x)eδx with δ chosen appropriately.
We assume this transformation has been carried out.
We start our investigations with interpolation-error estimates and a Galerkin
discretisations of (5.1)—including aspects of convergence, superconvergence, and
postprocessing of the derivatives. Then stabilised finite element methods are con-
sidered. We finish with an upwinded finite volume method.

T. Linß, Layer-Adapted Meshes for Reaction-Convection-Diffusion Problems, Lecture 151

Notes in Mathematics 1985, DOI 10.1007/978-3-642-05134-0 5,
c Springer-Verlag Berlin Heidelberg 2010
152 5 Finite Element and Finite Volume Methods

5.1 The Interpolation Error

In this section we study the error in linear interpolation. The argument follows [84].
Let ω̄ be an arbitrary mesh. Let wI denote the piecewise-linear function that inter-
polates to w at the nodes of ω̄.
In this section let us assume the function ψ ∈ C 2 [0, 1] admits the derivative
bounds

|ψ  (x)| ≤ C 1 + ε−2 e−βx/ε . (5.5)

For example, the solution u of the boundary-value problem (5.1) belongs to this
class of functions, see Sect. 3.4.1.2.
Proposition 5.1. Suppose ψ satisfies (5.5). Then
* +2
 
ψ − ψ I  ≤C 1+ε e −1 −βx/2ε
dx
∞,I i
Ii

for any mesh interval Ii = [xi−1 , xi ].

Proof. For the interpolation error on Ii we have the representation
  x  s
 I  1
ψ − ψ (x) = ψ  (t)dtdξds.
hi Ii xi−1 ξ

The right-hand side can be estimated to give


 I  
 ψ − ψ (x) ≤ (ξ − xi−1 ) |ψ  (ξ)| dξ.
Ii

Using Lemma 4.16 and (5.5) to bound the right-hand side, we are finished. 

Theorem 5.2. Suppose ψ satisfies (5.5). Then
 I     2
ψ − ψ  ≤ ψ I − ψ  ≤ C ϑ[2] (ω̄)
0 ∞ cd

and
 I 
ψ − ψ  ≤ Cϑ[2] (ω̄).
ε cd

Remark 5.3. The quantity

  
1 + ε−1 e−βs/pε ds,
[p]
ϑcd (ω̄) := max
k=0,...,N −1 Ik

[p]
was introduced in Sect. 2.1, where bounds on ϑcd (ω̄) for various layer-adapted
meshes are given too. ♣
5.1 The Interpolation Error 153

Proof (of Theorem 5.2). The bound on the L∞ error is an immediate consequence
[2]
of Prop. 5.1 and the definition of ϑcd .
1
For the error in the H norm, use integration by parts to get
  
 I  
2
1   2 1  
 ψ −ψ  = ψ − ψI (x) dx = − ψ  (x) ψ − ψ I (x)dx.
0 0 0

Thus,
 
 I  
2   1  
 ψ − ψ  ≤ ψ I − ψ ∞ |ψ  (x)| dx ≤ Cε−1 ψ I − ψ ∞
0 0

by a Hölder inequality and (5.5). Finally, combine this with the bound for the L2
norm of the interpolation error to obtain the energy-norm estimate. 


Remark 5.4. Proposition 5.1 can be used to give local estimates for the interpolation
error too. For example on S-type meshes (see Sect. 2.1.3) one has
 I   
ψ − ψ  ≤ ψ I − ψ ∞,[τ,1] ≤ CN −2 , if σ ≥ 2,
0,[τ,1]

for the interpolation error outside the layer region. This is in general a sharper bound
than that implicitly given by Theorem 5.2. ♣

Remark 5.5. The maximum-norm interpolation error bound can be generalised to

Lagrange interpolation with polynomial of arbitrary degree p ≥ 0.
Fix 0 ≤ ξ0 < ξ1 < · · · < ξp ≤ 1. Define an interpolant Ip ψ of ψ by

Ip ψ I ∈ Πp
i

and

(Ip ψ) (xi−1 + ξk hi ) = ψ (xi−1 + ξk hi ) for i = 1, . . . , N, k = 0, . . . , p.

If
 
 (p+1) 
ψ (x) ≤ C 1 + ε−(p+1) e−βx/ε

then
 p+1
[p+1]
Ip ψ − ψ∞ ≤ C ϑcd (ω̄) .

This result applies, for example, to the solution u of (5.1). ♣

154 5 Finite Element and Finite Volume Methods

5.2 Linear Galerkin FEM

We start from the weak formulation (5.2). Let ω̄ be an arbitrary mesh and let V ω
denote the space of continuous, piecewise linear functions on ω that vanish for x = 0
and x = 1. Then our discretisation is: Find uN ∈ V ω such that

a(uN , v) = f (v) for all v ∈ V ω . (5.6)

The coercivity of a(·, ·) guarantees the existence of unique solutions of both (5.2)
and of (5.6).

Notation. Throughout this section we use  · 1 to denote the L1 norm. This cannot
be confused with the H1 norm because we only use the weighted H1 norm |||·|||ε .

5.2.1 Convergence

Based on the interpolation error bounds of Sect. 5.1 we conduct an error analysis
for the Galerkin FEM on S-type meshes (see Sect. 2.1.3). The technique we shall
use was developed by Stynes and O’Riordan [152] for standard Shishkin meshes
and later generalised for S-type meshes by Linß and Roos [82, 137]. The technique
can be extended to discretisations of two-dimensional problems using triangular or
rectangular elements on tensor-product S-type meshes; see Sect. 9.2.2.1.
Theorem 5.6. Let ω̄ be an S-type mesh with σ ≥ 2 whose mesh generating function
ϕ̃ satisfies (2.8) and

max |ψ  | ln1/2 N ≤ CN. (5.7)

Then
   
u − uN  ≤ C h + N −1 max |ψ  |
ε

for the error of the Galerkin FEM.

Remark 5.7. The additional assumption (5.7) does not constitute a major restriction.
For example both the standard Shishkin mesh and the Bakhvalov-Shishkin mesh
satisfy this condition. ♣

Proof (Proof of Theorem 5.6). Let η = uI − u and χ = uI − uN . For the inter-

polation error η, we get from Sect. 5.1, the derivative bounds (3.30) and from (2.9)
that
 
|||η|||ε ≤ C h + N −1 max |ψ  | . (5.8)
5.2 Linear Galerkin FEM 155

To bound χ we start from the coercivity of a(·, ·) and the orthogonality of the
Galerkin method:

|||χ|||ε ≤ a(χ, χ) = a(η, χ) = ε(η  , χ ) + (bη, χ ) + ((c + b )η, χ) .

2

Apply the Cauchy-Schwarz inequality to the diffusion and reaction terms and the
Hölder inequality to the convection term to estimate
 
|||χ|||ε ≤ C |||η|||ε |||χ|||ε + C η∞,[0,τ ] χ 1,[0,τ ] + η∞,[τ,1] χ 1,[τ,1] ,
2

where τ is the mesh transition point in the S-type mesh. On [0, τ ] we use

χ 1,[0,τ ] ≤ Cτ 1/2 χ 0,[0,τ ] ≤ C ln1/2 N |||χ|||ε ,

while on [τ, 1] we have by an inverse inequality

χ 1,[τ,1] ≤ CN χ1,[τ,1] ≤ CN χ0,[τ,1] ≤ CN |||χ|||ε .

These two bounds and the interpolation results of Sect. 5.1 yield

|||χ|||ε ≤ C h + N −1 max |ψ  |
 2
+ h + N −1 max |ψ  | ln1/2 N + N −1 .

Thus,
 
|||χ|||ε ≤ C h + N −1 max |ψ  | ,

where we have used (5.7). Applying a triangle inequality and the bounds for |||η|||ε
and |||χ|||ε , we complete the proof. 


Remark 5.8. Sun and Stynes [157] use a similar technique to study the Galerkin-
FEM on standard Shishkin meshes for higher-order problems. ♣

Remark 5.9. We are not aware of a general convergence theory for the Galerkin
FEM on arbitrary layer-adapted meshes.
Roos [136] proves the optimal uniform error estimate
 
u − uN  ≤ CN −1
ε

for the Galerkin FEM on a special B-type mesh under the assumption that ε ≤ N −1 .
The key ingredient in his analysis is the use of a special quasi-interpolant with an
improved stability property. However, he points out that this technique cannot be
extended to higher dimensions. ♣
156 5 Finite Element and Finite Volume Methods

5.2.2 Supercloseness

In the preceding section we have seen that the Galerkin FEM is (almost)
 first-order

convergent in the ε-weighted energy norm. Now we prove that uI − uN  con-
  ε
verges faster than u − uN ε . This means the numerical approximation is closer
to the interpolant of the exact solution than to the solution itself. This phenomenon
is called supercloseness. Our analysis follows [83, 176] where two-dimensional
problems are studied.
Theorem 5.10. Let ω̄ be an S-type mesh with σ ≥ 5/2 whose mesh generating
function ϕ̃ satisfies (2.8). Then
 I   
u − uN  ≤ C h2 ln1/2 N + N −2 max |ψ  |2 (5.9)
ε

for the solution of the Galerkin FEM.

Proof. For the sake of simplicity we assume that b is constant. Let again η = uI − u
and χ = uI − uN . Then

a(η, χ) = ε(η  , χ ) − (bη  , χ) + (cη, χ)

For the diffusion term, integration by parts gives

 
η  χ = ηχ |xii−1 − ηχ = 0,
x

Ii Ii

because η(xi−1 ) = η(xi ) = 0 and because χ is linear. Thus, (η  , χ ) = 0. The

reaction term is easily bounded using the Cauchy-Schwarz inequality:
 2
|(cη, χ)| ≤ C η0 χ0 ≤ C h + N −1 max |ψ  | χ0 ,

by Theorem 5.2.
We are left with the convection term. Recalling the decomposition (3.34), we
split as follows:
 τ  τ
(η  , χ) = − (wI − w)χ − (v I − v)χ
0 0
 1  1 (5.10)
− (w − w)χ +
I
(v I − v) χ.
τ τ

The four terms on the right-hand side are bounded separately.

(i) By a standard interpolation error result

 I 
w − w ≤ Ch2i w 0,Ii .
0,I i
5.2 Linear Galerkin FEM 157

This and the bounds for the derivatives of w give

  qN 
 4
 I  qN
hi
w − w2 ≤ C h4i ε −4 −2βx/ε
e dx ≤ C hi e−2βxi−1 /ε
0,(0,τ ) ε
i=1 Ii i=1

 4 
qN
≤ C N −1 max |ψ  | hi e(4/σ−2)βxi /ε ,
i=1

by (2.12). Next, from (2.11)


hi e (4/σ−2)βxi /ε
≤ Chi e (4/σ−2)βxi−1 /ε
≤C e(4/σ−2)βx/ε dx.
Ii

Therefore,

 I   4 τ
w − w2 ≤ C N −1 max |ψ  | e(4/σ−2)βx/ε dx
0,(0,τ )
0
 4
≤ Cε N −1 max |ψ  | ,

where we have used σ > 2. This result and the Cauchy-Schwarz inequality
yield
 τ 
   
 (wI − w)χ  ≤ C N −1 max |ψ  | 2 |||χ||| . (5.11)
  ε
0

(ii) To bound the second term we proceed as follows:

 I  
qN
v − v 2 ≤ C h4i v  0,Ii ≤ Ch4 v  0,(0,τ ) ≤ Ch4 ε ln N,
2 2
0,(0,τ )
i=1

since |v  | ≤ C on [0, 1]. Hence,

 τ 
 
 (v I − v)ηχ  ≤ Ch2 ln1/2 N |||χ||| . (5.12)
  ε
0

1
(iii) Now we consider τ (w − wI )χ . The argument splits the integral once more,
but first let us recall that the mesh on (τ, 1) is uniform with mesh diameter
H: N −1 ≤ H ≤ N −1 /(1 − q). We have
 I 
w − w2 ≤ CN −1 e−2βτ /ε ≤ CN −6
0,I qN

since σ ≥ 5/2. Thus,

 
 xqN +1 
 (w − wI )χ  ≤ CN −2 χ0 ,

τ
158 5 Finite Element and Finite Volume Methods

by an inverse inequality. Next we have

 −1
 I  N
w − w2 ≤2 H w∞,Ii
0,(x qN +1 ,1)
i=qN +1


N −1  xN −1
≤C He−2βxi /ε ≤ C e−2βx/ε dx ≤ CεN −5 .
i=qN +1 τ

Thus,
 1 
 
 (w − wI )χ  ≤ CN −2 |||χ||| . (5.13)
  ε
τ

(iv) To bound the last term in (5.10) we use the integral identity
 
  1  
v − vI χ= v  Ei2 χ
Ii 6 Ii
 2   2 xi (5.14)
1 hi 1
 hi 
− v χ+ v χ
3 2 Ii 3 2 xi−1

with
  2 
1 hi 1
Ei (x) = (x − xi−1/2 ) − 2
= (x − xi−1 ) (x − xi ) .
2 2 2

This expansion formula holds true for arbitrary functions v ∈ W 3,∞ (Ii ) and
linear functions χ; cf. [78]. We get
  
1  
I  1 1   H 2  H2 1
v−v χ= v E 2 χ −

(v χ) (τ ) − v  χ.
τ 6 τ 12 12 τ

Assuming more regularity of the data, the decomposition (3.34) can be sharp-
ened to give |v  | ≤ C. This yields
 
 1  
I 

 v−v χ ≤ CH 3 χ 1,(τ,1) + CH 2 |χ(τ )| + CH 2 χ1,(τ,1)

τ
≤ CH 2 (|||χ|||ε + |χ(τ )|) ,

by an inverse inequality. Finally, we estimate

 τ 
 

|χ(τ )| =  χ (s)ds ≤ τ 1/2 χ 0,(0,τ ) ≤ C ln1/2 N |||χ|||ε .

0
5.2 Linear Galerkin FEM 159

Thus,
 
 1  
I 

 v−v χ ≤ CH 2 ln1/2 N |||χ|||ε . (5.15)

τ

Combine (5.10)-(5.15) to get for the convection term

 
|(η  , χ)| ≤ C h2 ln1/2 N + N −2 max |ψ  |2 |||χ|||ε .

This inequality, the bounds for the diffusion and reaction terms and the coercivity
of a(·, ·) yield the proposition of the theorem. 


Remark 5.11. Surprisingly, the major difficulty in the proof does not arise from the
layer term, but from the regular solution component. To cope with this, the special
integral expansion formula (5.14) by Lin had to be used. ♣

Remark 5.12. Another attempt at a supercloseness result for convection-diffusion

problems is [175]. In that paper, finite elements with piecewise polynomials of de-
gree p ≥ 1 are used on a piecewise uniform mesh with transition point
 
1 ε(p + 3/2)
τ = min , ln(N + 1) .
2 β

It was established that when the regular solution component v lies in the finite ele-
ment space then
 p+1
  ln(N + 1)
Qp u − uN  ≤ C ,
ε N

where Qp u is the (p + 1)-point Gauss-Lobatto interpolant of u. This is a superclose-

ness result, because in general, one has for the interpolation error
 p
ln(N + 1)
|||Qp u − u|||ε ≤ C .
N

However, the assumption that the regular solution component lies in the finite
element space is not very reasonable. If it were to hold for two different values of
the mesh parameter N then v ∈ Πp , because Shishkin meshes for different N are
not nested.
This too illustrates the technical difficulties with the regular solution compo-
nent just mentioned. Unfortunately—unlike (5.14) for linear elements—no expan-
sion formulae for the convection term are available for quadratic or higher-order
elements. ♣
160 5 Finite Element and Finite Volume Methods

5.2.3 Gradient Recovery and a Posteriori Error Estimation

Supercloseness results like Theorem 5.10 are basic ingredients for the superconver-
gent recovery of gradients, see for instance [2]. Furthermore, if a superconvergent
recovery operator is available, then it is possible to define an a posteriori error es-
timator that is asymptotically exact. The presentation follows [138], where further
details can be found.
First, we define for a given v ∈ V ω a recovery operator for the derivative. Set
xi − x x − xi−1
(Rv) (x) := αi−1 + αi for x ∈ Ii , i = 2, . . . , N − 1,
hi hi

where αi denotes the weighted average of the constant values of v  on the subinter-
vals adjacent to xi :
 
hi+1 v  I + hi v  I
i i+1
αi := .
hi + hi+1

For the boundary intervals we simply extrapolate the well-defined linear function of
the adjacent interval.  
Our aim is to prove a superconvergence estimate for ε1/2 u − RuN 0 that is
 
superior to that of Theorem 5.6 for ε1/2 u − (uN ) 0 . The key ingredients are
the supercloseness property of the Galerkin solution, i.e. Theorem 5.10, and the
consistency and stability of the recovery operator R.
Consistency: Let v be a quadratic function on I˜i , the union of Ii and its adjacent
mesh intervals. Then
 
R vI = v on Ii . (5.16a)

Stability:

Rv0,Ii ≤ C v  0,I˜i for all v ∈ V ω . (5.16b)

We start our analysis from a triangle inequality:

        
u − RuN  ≤ u − R uI  + R uI − uN  .
0 0 0

The second term in this inequality can be bounded  using Theorem

  5.10 and (5.16b).
Thus, we are left with the problem of estimating u − R uI 0 . To take advantage
of the consistency property (5.16a) we introduce a quadratic approximation of u on
I˜i : Qi u. Using a triangle inequality, we obtain
     
u − R uI  ≤ u − (Qi u) 0,I
0,Ii
 
i
   
+ (Qi u) − R (Qi u)I 0,Ii + R (Qi u − u)I 0,Ii .

5.2 Linear Galerkin FEM 161

The second term vanishes because of (5.16a). The last term can be bounded using
(5.16b) and the stability of the linear interpolation in H 1 , i. e., |v I |1 ≤ C|v|1 .
We get

    
N
  
u − R uI 2 ≤ C u − (Qi u) 2 ˜ .
0 0,I i
i=1

Note this H 1 stability of the interpolation operator holds true only in the one-
dimensional case. In two dimensions the L∞ stability of the interpolation operator
has to be used instead, see Sect. 9.2.2.5.
Choosing Qi u to be, e. g., that bilinear function that coincides with u at the
midpoint and both endpoints of I˜i and estimating the interpolation error carefully,
see [138], we obtain


N
    
ε u − (Qi u) 2 ˜ ≤ C h + N −1 max |ψ  | 4 if σ ≥ 2.
0,I i
i=1

Combining these estimates, we get the following result.

Theorem 5.13. Let ω̄ be a S-type mesh with σ ≥ 5/2 whose mesh generating func-
tion ϕ̃ satisfies (2.8). Then the error of the recovered gradient of the Galerkin FEM
satisfies
   
ε1/2 u − RuN 0 ≤ C h2 ln1/2 N + N −2 max |ψ  |2 .

Remark 5.14. Using RuN instead of u , we get an asymptotically exact  erroresti-

mator for the weighted H 1 -seminorm of the finite element error ε1/2 u − U  0 on
S-type meshes:
   
ε1/2 u − (uN ) 0 = ε1/2 RuN − (uN ) 0
 
+ O h2 ln1/2 N + N −2 max |ψ  |2 .

In the generic case one has

   
ε1/2 u − (uN ) 0 = O h + N −1 max |ψ  | .

Thus, the above error estimator is asymptotically exact for N → ∞. ♣

Remark 5.15. There are various other means of postprocessing to obtain supercon-
vergent approximations for the derivatives. ♣
162 5 Finite Element and Finite Volume Methods

5.2.4 A Numerical Example

Let us briefly illustrate our theoretical results for the linear Galerkin FEM on S-type
meshes when applied to the test problem
−εu − u + 2u = ex−1 in (0, 1), u(0) = u(1) = 0.

For our tests we take ε = 10−8 which is a sufficiently small choice to bring out the
singularly perturbed nature of the problem.
We consider three different S-type meshes: the original Shishkin mesh, the
Bakhvalov-Shishkin mesh and a mesh with a rational mesh characterising func-
tion ψ. The results of our test computations are presented in Tables 5.1, 5.2 and 5.3.
They are clear illustrations of the a priori error bounds given in Theorems 5.6, 5.10
and 5.13.

Table 5.1 Galerkin FEM on a S-type mesh with a rational ψ (m = 2)

   I   
u − uN  u − uN  ε1/2  u − RuN  0
ε ε
N error rate error rate error rate
28 1.158e-2 0.50 5.889e-4 0.99 3.289e-3 0.98
29 8.213e-3 0.50 2.959e-4 1.00 1.673e-3 0.99
210 5.817e-3 0.50 1.483e-4 1.00 8.435e-4 0.99
211 4.117e-3 0.50 7.424e-5 1.00 4.236e-4 1.00
212 2.912e-3 0.50 3.714e-5 1.00 2.123e-4 1.00
213 2.060e-3 0.50 1.858e-5 1.00 1.062e-4 1.00
214 1.456e-3 — 9.290e-6 — 5.315e-5 —

Table 5.2 Galerkin FEM on a standard Shishkin mesh

   I   
u − uN  u − uN  ε1/2  u − RuN  0
ε ε
N error rate error rate error rate
28 6.166e-3 0.83 1.624e-4 1.66 8.045e-4 1.64
29 3.470e-3 0.85 5.151e-5 1.69 2.585e-4 1.69
210 1.928e-3 0.86 1.592e-5 1.72 8.025e-5 1.72
211 1.060e-3 0.87 4.818e-6 1.75 2.432e-5 1.75
212 5.784e-4 0.88 1.434e-6 1.77 7.242e-6 1.77
213 3.133e-4 0.89 4.211e-7 1.79 2.126e-6 1.79
214 1.687e-4 — 1.221e-7 — 6.164e-7 —

Table 5.3 Galerkin FEM on a Bakhvalov-Shishkin mesh

   I   
u − uN  u − uN  ε1/2  u − RuN  0
ε ε
N error rate error rate error rate
28 1.357e-3 1.00 5.382e-6 1.99 4.173e-5 2.00
29 6.800e-4 1.00 1.353e-6 2.00 1.043e-5 2.00
210 3.403e-4 1.00 3.393e-7 2.00 2.610e-6 2.00
211 1.702e-4 1.00 8.497e-8 2.00 6.528e-7 2.00
212 8.514e-5 1.00 2.126e-8 2.00 1.632e-7 2.00
213 4.258e-5 1.00 5.317e-9 2.01 4.082e-8 2.00
214 2.129e-5 — 1.321e-9 — 1.020e-8 —
5.3 Stabilised FEM 163

5.3 Stabilised FEM

We have seen that the Galerkin FEM on S-type meshes has good approximation
properties. Unfortunately the linear systems generated are difficult to solve iter-
atively. Therefore, stabilisation is essential. We shall restrict ourselves to artificial
viscosity stabilisation and to the streamline-diffusion FEM. Other stabilisation tech-
niques, including:
• discontinuous Galerkin FEM (dGFEM),
• continuous interior penalties (CIP) and
• local projection stabilisation (LPFEM),

have been considered in the literature. However, with regard to the classification of
layer-adapted meshes these contributions are negligible. Nonetheless, some of the
results for these methods will be mentioned in Sect. 9.2.

5.3.1 Artificial Viscosity Stabilisation

The simplest way to stabilise discretisation methods for convection-diffusion prob-

lems consists of altering the diffusion coefficient a priori, the extra diffusion added
being called artificial viscosity. Typically artificial viscosity proportional to the
stepsize is used.
Let κ > 0 be an arbitrary constant. Then our stabilised FEM is: Find uN ∈ V ω
such that

aκ (uN , v) = f (v) for all v ∈ V ω ,

where
 
aκ (u, v) := (ε + κh̄)u , v  − (bu − cu, v)

and

h̄(x) ≡ hi for x ∈ Ii .

The bilinear form aκ (·, ·) is coercive with respect to the norm

  1/2
|||v|||κ := (ε + κh̄)v  , v  + γv20 ,

which is stronger than the ε-weighted energy norm. This is the reason for the im-
proved stability of the method.
164 5 Finite Element and Finite Volume Methods

Because of the added artificial viscosity, the method does not satisfy the orthog-
onality property which slightly complicates the convergence analysis. Assume a
S-type mesh is used. Let η = uI − u and χ = uI − uN again. Then
 
|||χ|||κ ≤ aκ (χ, χ) = a(η, χ) + κh̄(uI ) , χ
2

= a(η, χ) + κ (h̄η  , χ ) + κ (h̄u , χ ) .

Bounds for the first term have been derived in Sect. 5.2. The second term (h̄η  , χ )
vanishes, while the last term, which is the inconsistency of the method, satisfies
 
κ |(h̄u , χ )| ≤ Cκ h ln1/2 N + N −1 max |ψ  | |||χ|||ε .

The proof recycles some ideas from Sect. 5.2.1 and 5.2.2 and is therefore omitted.
For more details see also [148]. We get
 I 
u − uN  ≤ C h(h + κ) ln1/2 N
κ
  (5.17)
+ κ + N −1 max |ψ  | N −1 max |ψ  | .

If we choose κ = O (1), i. e., we add artificial viscosity proportional to the local

mesh size, we get
 I     
u − uN  + u − uN  ≤ C h ln1/2 N + N −1 max |ψ  | ,
κ ε

by the interpolation error estimate (5.8).

Comparing (5.9) and (5.17), we see that  the
 order of accuracy of the Galerkin
FEM is not affected if we take κ = O N −1 . This results in improved stability
compared to the Galerkin method and the discrete systems—in particular for higher-
dimensional problems—are slightly easier to solve by means of standard iterative
methods. On the other hand, the method is not as stable as if κ = O (1) were chosen.

5.3.2 Streamline-Diffusion Stabilisation

The most popular and most frequently studied stabilised FEM is the streamline-
diffusion finite element method (SDFEM) which is also referred to as the
streamline-upwind Petrov-Galerkin method (SUPG). This kind of stabilisation
was introduced by Hughes and Brooks [54]. Given a mesh ω and a finite element
space V ω , this method can be written as: Find uN ∈ V ω such that

 
N
 
a(uN , v) + δi f − LuN bv  = (f, v) for all v ∈ V ω , (5.18)
i=1 Ii
5.3 Stabilised FEM 165

where the stabilisation parameters δi are chosen according to the local mesh Peclét
number:

κ0 hi if P ei > 1, b∞,Ii hi
δi = with P ei =
2 −1 2ε
κ1 hi ε if P ei ≤ 1,

with user chosen positive constants κ0 and κ1 . In contrast to the artificial-viscosity

stabilisation, this method is consistent with (5.1) since u satisfies (5.18) for
all v ∈ H01 (0, 1). Another advantage—though it becomes relevant only in higher
dimensions—is the reduction of crosswind smear because artificial viscosity is
added only in the streamline direction.
The second-order upwind schemes of Sect. 4.3.1 may be regarded as versions
of the SDFEM with linear test and trial functions and inexact integration. While in
the one-dimensional case it is always possible to choose the stabilisation parameters
δi such that the resulting scheme is inverse monotone, this is in general impossi-
ble in higher dimensions. Therefore, alternative techniques have to be developed to
study the SDFEM. Here we shall consider convergence in the streamline-diffusion
norm |||·|||SD naturally associated with the bilinear form of the method. This tech-
nique can be extended to two-dimensional problems; see Sect. 9.2.4 or [155].

5.3.2.1 Energy-Norm Error Estimates

We study the SDFEM on S-type meshes ω̄. For the sake of simplicity we con-
sider (5.1) with constant b. Let V0ω ⊂ H01 (0, 1) be the linear space of piecewise-
affine functions on ω̄ that vanish at the boundary. We rewrite (5.18) as: Find
uN ∈ V0ω such that

aSD (uN , v) := a(uN , v) + astab (uN , v) = f (v) + fstab (v) for all v ∈ V0ω

where a(·, ·) is the bilinear form of the Galerkin FEM,


N 
astab (w, v) := −δ (−εw − bw + cw)bv  ,
i=qN +1 Ii


N 
fstab (v) := −δ f bv 
i=qN +1 Ii

and

κ0 H if bH/2ε > 1,
δ=
2
κ1 H /ε otherwise.

Here H denotes again the mesh size on the coarse part of the mesh.
166 5 Finite Element and Finite Volume Methods

The streamline-diffusion norm naturally associated with aSD (·, ·) is

 2
|||v|||SD := ε v  0 + γ v0 + δ 1/2 bv  0,(τ,1) .
2 2 2

Provided the maximum step size h is smaller than some threshold value, the bilinear
form aSD (·, ·) is coercive with respect to the streamline-diffusion norm:
2
aSD (v, v) ≥ 1
2 |||v|||SD for all v ∈ V0ω ;

see [141]. The bilinear form also satisfies the Galerkin-orthogonality property

aSD (u − uN , v) = 0 for all v ∈ V0ω .

This is the starting point of our error analysis. Let η = uI − u and χ = uI − uN .

Then
2
1
2 |||χ|||SD ≤ a(η, χ) + astab (η, χ). (5.19)

For the first term we have from the proof of Theorem 5.10
 
|a(η, χ)| ≤ C h2 ln1/2 +N −2 max |ψ  |2 |||χ|||ε .

It remains to bound astab (η, χ). We have

 1
astab (η, χ) = δ (εu + bη  + cη) bχ .
τ

1
Element-wise integration by parts yields τ
η  χ = 0. Furthermore,
  
 1 
δ cηbχ  ≤ Cδ 1/2 η0,(τ,1) δ 1/2 bχ 0

λ
≤ Cδ 1/2 N −2 δ 1/2 bχ 0 ≤ CN −2 |||χ|||SD ,

by the bounds for the interpolation error from Sect. 5.1.

1
To bound the remaining term τ u χ we use the decomposition of u into a
regular and a layer component; see Theorem 3.48. For the regular component v we
have
 1  1  τ
v  χ = − v  χ − v  χ .
τ τ 0

Hence,
 
 1 
 v  χ  ≤ Cχ0 + C (ε ln N ) χ 0 ,
1/2

τ
5.3 Stabilised FEM 167

by the bounds for the derivatives of v. Thus,

  1 
 
εδ   
v χ  ≤ CN −2 ln1/2 N |||χ|||SD ,

τ

since the choice of δ implies εδ ≤ CH 2 ≤ CN −2 .

For the layer component w we estimate as follows:
 
 1 
εδ  
w bχ  ≤ εδ 1/2 w 1,(τ,1) δ 1/2 bχ ∞,(τ,1)

τ
≤ Cδ 1/2 N −5/2 H −1/2 δ 1/2 bχ 0,(τ,1) ,

by an inverse inequality and because σ ≥ 5/2. We get

 1 
 

εδ  
w bχ  ≤ CN −2 |||χ|||SD .

τ

Collecting these results, the second term in (5.19) is bounded by

|astab (η, χ)| ≤ CN −2 ln1/2 N |||χ|||ε .

We summarise our results.

Theorem 5.16. Let ω̄ be an S-type mesh with σ ≥ 5/2 whose mesh generating
function ϕ̃ satisfies (2.8). Then the error of the SDFEM satisfies
 I   
u − uN  ≤ C h2 ln1/2 N + N −2 max |ψ  |2 .
SD

Remark 5.17. This is a superconvergence result like Theorem 5.10. Similar to

Sect. 5.2.3 it is possible to construct recovery operators to obtain a second-order
approximations of the gradient of the exact solution. ♣

5.3.2.2 Maximum-Norm Error Estimates

Chen and Xu [26] consider a modification of (5.18). Find ũN ∈ V0ω

N 
  
N
a(ũ , v) + δ̃i f − LũN bv  = (f, v) for all v ∈ V0ω ,
i=1 Ii

where
 
h2i hi
δ̃i (x) = min , (xi − x)(x − xi−1 ) for x ∈ Ii ,
2ε b

i.e., δ̃i is a bubble function on Ii instead of a constant.

168 5 Finite Element and Finite Volume Methods

For this modified SDFEM it is shown in [26] that if b = const. and c ≡ 0 then
 
u − ũN  ≤ C min u − v .
∞ ω v∈V0 ∞

Thus, the method is quasi-optimal in the maximum norm.

Suppose b = const. Then
 
δ̃i bw bv  = δi bw bv  for all w, v ∈ V0ω
Ii Ii

with
 
1 h2i hi
δi = min , ,
6 2ε b

i. e., δi is the mean value of δ̃i on Ii . Therefore, the modified and the original
SDFEM generate the same difference stencil, but different discretisations of the
right-hand side. It follows that

  N
  
aSD uN − ũN , v = δ̃i − δi f − f I bv  .
i=1 Ii

Next, the stability of the discretisation, which was established in [26], implies
 N   
u − ũN  ≤ C f − f I  .
∞ ∞

Finally, the triangle inequality yields for the solution of the original SDFEM
 
   
u − uN  ≤ C min u − v + f − f I  .
∞ v∈V0ω ∞ ∞

5.4 An Upwind Finite Volume Method

Let us finish this chapter by considering finite volume discretisations of (5.1).

Although the construction of finite volume methods differs from finite difference
and finite element methods, they are typically analysed as special finite differ-
ence methods or—more often—as nonconforming finite element methods. Here we
like to highlight both approaches. In particular, this section is intended to prepare
our later investigation of the FVM in two dimensions in Sect. 9.3. There a detailed
construction of the method can be found too.
We shall assume

c ≥ γ > 0, c + b ≥ γ > 0 when studying the FVM as a FDM (5.20)

5.4 An Upwind Finite Volume Method 169

and

c + b /2 ≥ γ > 0 in the FEM context. (5.21)

In the latter case the variational formulation (5.2) will be used.

Given an arbitrary mesh ω̄ our FVM reads: Find uN ∈ IR0N +1 such that
' (
Lρ uN i = fi for i = 1, . . . , N − 1, (5.22)

where
   
[Lρ v]i := −εvx̄x̂;i − ρ −μi+1/2 bi+1/2 vx̂;i − ρ μi−1/2 bi−1/2 vx̌;i + ci uN
i

 
with μi+1/2 = bi+1/2 hi+1 /ε and bi+1/2 = b xi+1/2 .
The method can also be written in variational form: Find uN ∈ IR0N +1 such that
 
aρ (uN , v) := Lρ uN , v ω = (f, v)ω =: fρ (v) for all v ∈ IR0N +1 ,

where


N −1
(w, v)ω := h̄i wi vi .
i=1

The crucial point is the choice of the controlling function ρ : IR → [0, 1]. It has to
provide the correct weighting between the two one-sided difference approximations
for the first-order derivative. Possible choices for ρ include:
⎧  
⎪
⎪
1
−
t
⎨ 1 for t = 0,
t exp t − 1
ρI (t) =
⎪
⎪
⎩1 for t = 0,
2
⎧
⎨1/(2 + t) for t ≥ 0,
ρS (t) =
⎩(1 − t)/(2 − t) for t < 0,

and, with m ≥ 0,
⎧
⎪
⎪0 for t > m,
⎪
⎨
ρU,m (t) = 1
for t ∈ [−m, m],
⎪
⎪
2
⎪
⎩
1 for t < −m;
170 5 Finite Element and Finite Volume Methods

ρI
1 ρs
ρU,m

ρ 0.8

0.6

0.4

0.2

−40 −30 −20 −10 0 10 20 30 40

t

Fig. 5.1 The stabilising functions ρI , ρS and ρU,m

see Fig. 5.1. The full upwind stabilisation ρU,0 is due to Baba and Tabata [17], while
ρU,m with m > 0 was introduced by Angermann [13]. For ρI and ρS we get slight
modifications of the schemes of Il’in [55] and of Samarski [144]. Further choices of
ρ are mentioned in [13] and [62] where also a detailed derivation of the method in
two dimensions can be found.
The constant choice ρ ≡ 12 generates a central difference scheme, while the
choice ρU,0 gives a scheme with upwinded one-sided difference approximation of
the first-order derivative which is very similar to the upwind scheme analysed in
Sect. 4.2. If a different ρ is used—in particular when ρ is Lipschitz continuous in a
neighbourhood of 0—then the first-order derivatives are approximated by weighted
combinations of upwinded and downwinded operators. This weighting provides an
adaptive transition from an upwinded to a central difference approximation when
the local mesh size is small enough. In this case higher accuracy is achieved while
retaining the good stability of the scheme.
Important properties of ρ are

(ρ0 ) t → tρ(t) is Lipschitz continuous,

' (
(ρ1 ) ρ(t) + ρ(−t) − 1 t = 0 for all t ∈ IR,
' (
(ρ2 ) 1/2 − ρ(t) t ≥ 0 for all t ∈ IR,
(ρ3 ) 1 − tρ(t) ≥ 0 for all t ∈ IR.

Condition (ρ1 ) ensures both the consistency of the scheme and the local conserva-
tion of the fluxes, while (ρ2 ) guarantees the coercivity of the bilinear form aρ (·, ·)
and (ρ3 ) the inverse monotonicity of the scheme.
5.4 An Upwind Finite Volume Method 171

5.4.1 Stability of the FVM

Coercivity of the bilinear form aρ (·, ·)

The consistency condition (ρ1 ) and summation by parts yield


N
(vi − vi−1 )2
aρ (v, v) = ε
i=1
hi
N *
 +
1  
+ − ρ μi−1/2 bi−1/2 (vi − vi−1 )2 (5.23)
i=1
2

 −1 * +
1 
N
+ h̄i ci + bi+1/2 − bi−1/2 vi2 .
i=1
2

Assume b is Hölder continuous with coefficient α ∈ (0, 1]. Then

 
bi+1/2 − bi−1/2 − h̄i bi  ≤ h̄i hα b . (5.24)
C 1,α [0,1]

Thus, if (5.21) is satisfied and if the maximum mesh size h is smaller than some
threshold value h∗ then

 −1 * + N −1
1  γ 
N
h̄i ci + bi+1/2 − bi−1/2 vi2 ≥ h̄i vi2 . (5.25)
i=1
2 2 i=1

Let
2 2 2 γ 2
|||v|||ρ := ε |v|1,ω + |v|ρ,ω + v0,ω
2
with


N 
N −1
2 (vi − vi−1 )2
|v|1,ω := , v0,ω := h̄i vi2
i=1
hi i=1

and

N *
 +
2 1  
|v|ρ,ω := − ρ μi−1/2 bi−1/2 (vi − vi−1 )2
i=1
2

Note, |||·|||ρ is a well-defined norm when (ρ2 ) is satisfied. The coercivity of the
discrete bilinear form aρ (·, ·) follows from (5.23) and (5.25).
172 5 Finite Element and Finite Volume Methods

Theorem 5.18. Assume conditions (ρ1 ), (ρ2 ) and (5.21) are satisfied. Let b ∈
C 1,α [0, 1] with Hölder exponent α ∈ (0, 1]. Then the bilinear form aρ (·, ·) is co-
ercive with respect to the FV-norm |||·|||ρ , i.e.,
2
aρ (v, v) ≥ |||v|||ρ for all v ∈ IR0N +1 ,

provided the maximum mesh size h is smaller than some threshold value which is
independent of the perturbation parameter ε.

Remark 5.19. Note when ρ ≡ 12 the stabilisation is switched off. Nonetheless Theo-
rem 5.18 states coercivity of the bilinear form with respect to the discrete ε-weighted
2
energy norm |||v|||ε,ω := ε|v|21,ω + γ2 v20,ω . However, in the case ρ ≡ 12 the scheme
is coercive with respect to a stronger norm which results in enhanced stability of the
method. ♣

Inverse monotonicity
Let r+ , r− , q ≥ γ > 0 and χ > 0 be arbitrary mesh functions with

βhi+1
ri+ ≥ and 1 ≥ ri− ≥ 0 for i = 1, . . . , N − 1 (5.26)
αε
with a constant α > 0. Consider the difference operator
ε ' ( ε ' (
[Lχ v]i := − 1 + ri+ vx;i + 1 − ri− vx̄;i + qi vi . (5.27)
χi χi

We study this more general situation because it will also serve as an auxiliary result
in Sect. 9.3.2 when two dimensional problems will be investigated. The FVM (5.22)
belongs to this class of schemes provided that (ρ3 ) holds.
Clearly, 1+ri+ ≥ 1 and 1−ri− ≥ 0. Hence, the system matrix associated with Lχ
possesses nonnegative offdiagonal entries and is therefore a L0 -matrix. Application
of Lemma 3.14 with the test vector e = 1 yields the inverse monotonicity of Lχ . In
particular, we get the stability inequality

v∞,ω̄ ≤ Lχ v/q∞,ω for all v ∈ IR0N +1 .

The inverse monotonicity can be used to study the Green’s function associated
with Lχ and derive stability inequalities similar to those of Sect. 4.2.1.

The (∞ , 1 ) stability

The Green’s function G·,j associated with the mesh node xj satisfies

[Lχ G·,j ]i = δi,j for i = 1, . . . , N − 1, G0,j = GN,j = 0,

5.4 An Upwind Finite Volume Method 173

where

χ−1
i for i = j,
δi,j :=
0 otherwise.

Given G, we can represent any function v ∈ IR0N +1 as

vi = (Gi,· , Lχ v)χ , (5.28)

with the scalar product


N −1
(w, v)χ := χj wj vj .
j=1

Let
⎧α
⎪
⎪ for i = 0, . . . , j,
⎨β
Ĝi,j = α 3
i   −1
⎪ βhk
⎪
⎩β 1+ for i = j + 1, . . . , N.
αε
k=j+1

Lemma 5.20. Suppose (5.26) holds true. Then the Green’s function G associated
with Lχ satisfies

0 ≤ Gi,j ≤ Ĝi,j ≤ α/β for i, j = 0, . . . , N.

Proof. If (5.26) holds then the operator Lχ is inverse monotone and therefore satis-
fies a discrete comparison principle. The lower bound on G is easily verified using
the barrier function v ≡ 0.
In order to establish the upper bound, it is sufficient to show that for fixed j we
have G0,j ≤ Ĝ0,j , GN,j ≤ ĜN,j and
' ( ' (
Lχ G·,j i ≤ Lχ Ĝ·,j i for i = 1, . . . , N − 1,

(i) First we check the boundary conditions. Clearly Ĝi,j > 0 for all i, j. Thus,

0 = G0,j ≤ Ĝ0,j and 0 = GN,j ≤ ĜN,j for j = 1, . . . , N − 1.

' (
(ii) Next, for i < j we have Lχ Ĝ·,j i = qi Ĝi,j ≥ 0 since both q and Ĝ are
positive.
(iii) For i > j we have
' ( β ' ( β
Ĝ·,j x;i
= Ĝi,j and Ĝ·,j x̄;i
=− Ĝi,j .
αε + βhi+1 αε
174 5 Finite Element and Finite Volume Methods

Thus,
    
' ( ε 1 + ri+ β 1 − ri− β
Lχ Ĝ·,j i ≥ − Ĝi,j ≥ 0,
χi αε + βhi+1 αε

by (5.26).
(iv) For i = j a combination of the arguments from (ii) and (iii) yields
 
' ( ε 1 + ri+ β 1 ' (
Lχ Ĝ·,j j ≥ Ĝj,j ≥ = Lχ G·,j j ,
χj αε + βhj+1 χj

by (5.26).

This Lemma and (5.28) give the (∞ , 1 ) stability of the method:
Theorem 5.21. Suppose (5.26) holds. Then the operator Lχ defined in (5.27) satis-
fies the stability inequality
N −1
α   
v∞,ω ≤ χi [Lχ v]i  for all v ∈ IR0N +1 .
β i=1

Remark 5.22. An error analysis of the upwind FVM using this (∞ , 1 ) stability
can be conducted along the lines of Sect. 4.2.5; see also [90]. ♣

The (∞ , w−1,∞ ) stability

Now let us restrict our attention to difference operators of the type
ε ' ( ε ' (
[Lκ v]i := − 1 + ρ+
i vx;i + 1 − ρ−
i vx̄,i + ci vi
χi χi
with
   
bi+κ hi+1 bi+κ hi+1 bi−1+κ hi bi−1+κ hi
ρ+
i := ρ − , ρ−
i := ρ
ε ε ε ε
and
κ ∈ [0, 1], χi = κhi+1 + (1 − κ)hi and bi+κ = b(xi + κhi+1 ).

The FVM (5.22) is recovered for κ = 1/2, while the finite difference scheme of
Sect. 4.2 is obtained when κ = 1 and ρ = ρU,0 .

Remark 5.23. Condition (5.26) with α = supt<0 1/ρ(t) follows from (ρ3 ). ♣

Assuming that (ρ1 ) holds, the Green’s function G solves for fixed i
' ∗ (
Lκ Gi,· j = δi,j for j = 1, . . . , N − 1, Gi,0 = Gi,N = 0
5.4 An Upwind Finite Volume Method 175

where L∗κ is the adjoint operator to Lκ with respect to the scalar product (·, ·)χ . Note
− −1
(ρ1 ) implies ρ+
i + ρi+1 = ε bi+κ hi+1 . Then it is verified that

' ∗ ( ε ' ( ' (

Lκ v j = − 1 − ρ−
j+1 vx;j − 1 + ρj−1 vx̂;j
+
χj
 
bj+κ − bj−1+κ
+ cj + vj .
χj

Assume c + b ≥ γ > 0 and let b be Hölder continuous with coefficient λ ∈ (0, 1].
Then
bj+κ − bj−1+κ
cj + ≥0
χj

if the maximum step size h is sufficiently small, independent of ε; cf. (5.24). Pro-
ceeding as in Sect. 4.2.1, one can show

Gi,j ≥ Gi,j−1 for j = 1, . . . , i and Gi,j ≤ Gi,j−1 for j = i + 1, . . . , N,

i.e., Gi,· is piecewise monotone.

Theorem 5.24. Suppose (ρ1 ), (ρ3 ) and (5.20) hold true. Assume b ∈ C 1,λ [0, 1]
with Hölder exponent λ ∈ (0, 1]. Then the operator Lκ satisfies the stability
inequality
N −1 
2α  
v∞,ω ≤ 
min  χj [Lκ v]j + C  for all v ∈ IR0N +1 ,
β C∈IR j=· 
∞,ω

with α = 1/ inf t<0 ρ(t) ≤ 2, provided the maximum step size h is smaller than
some threshold value that is independent of the perturbation parameter ε.

5.4.2 Convergence in the Energy Norm

In this section we study the convergence in the energy norm |||·|||ρ of the finite vol-
ume method on S-type meshes (see Sect. 2.1.3) with σ ≥ 2. The controlling function
ρ is assumed to satisfy conditions (ρ0 ), (ρ1 ) and (ρ2 ).
Our analysis starts from coercivity of aρ (·, ·) (see Theorem 5.18) and follows the
standard approach of the Strang Lemma. Let η = uI − u and χ = uI − uN , where
we use uN for both the pointwise defined solution of (5.22) and its piecewise-linear
extension on the mesh ω̄.
From (5.2) and (5.22) we get

aρ (χ, χ) = a(η, χ) + aρ (uI , χ) − a(uI , χ) + f (χ) − fρ (χ).

176 5 Finite Element and Finite Volume Methods

Then Theorem 5.18 yields

2
|||χ|||ρ ≤ |a(η, χ)| + |fρ (χ) − f (χ)|
    (5.29)
+ r(uI , χ) − rρ (uI , χ) + c(uI , χ) − cρ (uI , χ)

with
 1 
N −1  1
I
r(u , χ) = I
cu χ, I
rρ (u , χ) = h̄i ci ui χi , c(u , χ) = −
I
b(uI ) χ
0 i=1 0

and

 −1
N
 
cρ (uI , χ) = − ρ μi+1/2 bi+1/2 (ui+1 − ui )
i=1
 
+ ρ −μi−1/2 bi−1/2 (ui − ui−1 ) χi .

The four terms on the right-hand side of (5.29) will be bounded separately.

(i) The first term has been analysed in Sect. 5.2. We have under the assumptions
of Theorem 5.6
 
|a(η, χ)| ≤ C h + N −1 max |ψ  | |||χ|||ε,ω , (5.30)

because the discrete and continuous energy norms are equivalent for functions
from V ω .
(ii) Next we bound the error arising from the discretisation of the right-hand side f .
Denoting by ϕi the usual basis functions for linear finite elements, we have
 
   hi 
 f ϕi (x)dx − fi 
 2
Ii
   x  
 hi  h2i  
=  fi + 
f (s)ds ϕi (x)dx − fi  ≤ f   .
2 2 ∞
Ii xi

Thus,
  
N
−1 xi+1   
 
|f (χ) − fh (χ)| =  χi f ϕi (x)dx − h̄i fi 
 xi−1 
i=1
(5.31)

N −1
 
≤ f ∞ h h̄i |χi | ≤ f ∞ hχ0,ω .
i=1
5.4 An Upwind Finite Volume Method 177

(iii) The next term in line is r(uI , χ) − rρ (uI , χ). By the definition of rρ (·, ·)
and r(·, ·), we have


N −1
rρ (uI , χ) − r(uI , χ) = si χi ,
i=1

where
 
xi+1   xi+1 ' I  (
si := cuI ϕi (x)dx − h̄i ci ui = cu (x) − ci ui ϕi (x)dx.
xi−1 xi−1

We have
 I  
 cu (x) − ci ui 
 x  
  I   xi+1
≤  
cu ds ≤ C 1 + ε−1 e−βs/ε ds ≤ Cϑcd (ω̄)
[1]

xi xi−1

[p]
for x ∈ [xi−1 , xi+1 ]. (The quantity ϑcd (ω̄) has been introduced in Sect. 2.1.)
[1]
Hence, |si | ≤ Cϑcd (ω̄)h̄i and we obtain
 
rρ (uI , χ)i − r(uI , χ) ≤ Cϑ[1] (ω̄)χ0,ω . (5.32)
cd

(iv) Finally consider the convection term. We have

cρ (uI , χ) − c(uI , χ)
N 
 I  
= b(u ) χ (x)dx
i=1 Ii

'     (  
− ρ −μi−1/2 χi−1 + ρ μi−1/2 χi bi−1/2 ui − ui−1

and

    χi + χi−1
b(uI ) χ (x)dx = bi−1/2 ui − ui−1
Ii 2
  x 
u − u
b (s)ds
i i−1
+ χ(x) dx.
Ii xi−1/2 hi
178 5 Finite Element and Finite Volume Methods

Combine these two equations and use (ρ1 ) to get

cρ (uI , χ) − c(uI , χ)
N * +
1     
= − ρ μi−1/2 χi − χi−1 ui − ui−1 bi−1/2
i=1
2 (5.33)
N 
 
 x
u − u
b (s)ds
i i−1
+ χ(x) dx.
i=1 Ii xi−1/2
hi

[1]
Note that |ui − ui−1 | ≤ Cϑcd (ω̄). Thus, for the second term in (5.33) one has
   
N  

x
 ui − ui−1  [1]
 b (s)ds χ(x) dx ≤ Cϑcd (ω̄) χ0,ω . (5.34)
 Ii xi−1/2 h i 
i=1

[1]
Next we bound the first sum in (5.33). For i ≤ qN use |ui − ui−1 | ≤ Cϑcd (ω̄)
again to obtain
 qN * + 
 1   
 
 − ρ μi−1/2 (χi − χi−1 ) (ui − ui−1 ) bi−1/2 
 2 
i=1
(5.35)

qN
≤ Cϑ1 (ω) |χi − χi−1 | ≤ Cϑ1 (ω)ε1/2 ln1/2 N |χ|1,ω ,
i=1

by a discrete Cauchy-Schwarz inequality.

For i > qN we use the splitting u = v + w of the exact solution according to
Theorem 3.48. Starting with the layer term w, we have wi ≤ CN −2 for i ≥ qN .
Hence,
 
 N * + 
  1   
 − ρ μ (χ − χ ) (w − w ) b 
 i−1/2 i i−1 i i−1 i−1/2 
i=qN +1 2 
(5.36)
N
≤ CN −2 (|χi | + |χi−1 |) ≤ CN −1 χ0,ω ,
i=qN +1

by a discrete Cauchy-Schwarz inequality and an inverse inequality that can be used

because N −1 ≤ hi for i > qN . Finally, consider the regular solution component v.
To simplify the notation let
*  +
1 bi−1/2 hi
γi−1/2 := bi−1/2 −ρ .
2 ε
5.4 An Upwind Finite Volume Method 179

Summation by parts yields


N
  
γi−1/2 vi − vi−1 χi − χi−1
i=qN +1

= γqN +1/2 (vqN − vqN +1 ) χqN


N −1
− γi−1/2 (vi+1 − 2vi + vi−1 ) χi
i=qN +1

 −1
N
 
+ γi−1/2 − γi+1/2 (vi+1 − vi ) χi .
i=qN +1

Taylor expansions give |vi+1 − 2vi + ≤ CN −2 and |vi − vi−1 | ≤ CN −1 ,

 vi−1 | −1
 
while (ρ0 ) implies γi−1/2 − γi+1/2 ≤ CN . Thus,
 
 N 
  
 γ (v − v ) (χ − χ ) 
 i−1/2 i i−1 i i−1 
i=qN +1 
 
≤ CN −1 χ0,ω + |χqN | ≤ CN −1 ln1/2 N |||χ|||ρ ,

because


qN
|χqN | ≤ |χi − χi−1 | ≤ C ln1/2 N ε1/2 |χ|1,ω .
i=1

Collecting (5.33)–(5.37), we get

 
cρ (uI , χ) − c(uI , χ) ≤ Cϑ1 (ω) ln1/2 N |||χ||| . (5.37)
ρ

Now all terms on the right-hand side of (5.29) have been bounded; see (5.30),
(5.32), (5.31), (5.34) and (5.37). Divide by |||χ|||ρ . Then recall the interpolation error
bounds of Sect. 5.1 and note that ϑcd (ω̄) ≤ C (h + max |ψ  |) for S-type meshes
[1]

with σ ≥ 1. We get the main result of this section.

Theorem 5.25. Let ω̄ be an S-type mesh with σ ≥ 2 whose mesh generating func-
tion ϕ̃ satisfies (2.8) and max |ψ  | ln1/2 N ≤ CN . Assume (ρ0 ), (ρ1 ) and (ρ2 ) hold.
Then
     
u − uN  + uI − uN  ≤ C h + N −1 max |ψ  | ln1/2 N
ε ρ

for the error of the upwind FVM (5.22).

180 5 Finite Element and Finite Volume Methods

5.4.3 Convergence in the Maximum Norm

With the results of Sect. 5.4.1 at hand, the simplest maximum-norm analysis is based
on the (∞ , w−1,∞ ) stability. Set

'    ( vi − vi−1
[Aρ v]i := ε 1 + μi−1/2 ρ −μi−1/2 − ρ μi−1/2
hi
−1
vi + vi−1
N

+ bi−1/2 + h̄j cj + bj+1/2 − bj−1/2 vj .
2 j=i

If condition (ρ1 ) is satisfied then Lρ v = −(Aρ v)x̂ and Theorem 5.24 yields
4
v∞,ω ≤ min Aρ v + a∞,ω for all v ∈ IR0N +1 .
β a∈IR

Integrate (5.1) to see that

 xN −1/2
εui−1/2 + (bu)i−1/2 + ((c + b ) u − f ) (s)ds ≡ α
xi−1/2

for all i = 1, . . . , N.

Thus,
  4
u − uN  ≤ max |Mi | , (5.38a)
∞,ω β i=1,...,N

where
   
ui − ui−1 ui + ui−1
Mi := ε − ui−1/2 + bi−1/2 − ui−1/2
hi 2
−1 
N
 xN −1/2
+ h̄j cj + bj+1/2 − bj−1/2 uj − (c + b ) (s)u(s)ds
j=i xi−1/2


N −1  xN −1/2
− h̄j fj + f (s)ds
j=i xi−1/2
*    +
bi−1/2 hi bi−1/2 hi
+ bi−1/2 ρ − −ρ (ui − ui−1 ) .
ε ε
(5.38b)

[1]
All terms except for the last one can be bounded by ϑcd (ω̄) using the technique
from Sect. 4.2.2. When bounding the last term note that ρ(t) ∈ [0, 1] for all t ∈ IR
and that 
|u (s)| ds ≤ Cϑcd (ω̄).
[1]
|ui − ui−1 | ≤
Ii
5.4 An Upwind Finite Volume Method 181

Theorem 5.26. Suppose (ρ1 ) and (ρ3 ) hold. Then the error of the upwind finite
volume method (5.22) satisfies
 
u − uN  [1]
≤ Cϑcd (ω̄).
∞,ω

It was mentioned earlier that the accuracy of the scheme is improved when the
function ρ is Lipschitz continuous in a neighbourhood of t = 0, say on an inter-
val [−m, m]. We will briefly illustrate this using a standard Shishkin mesh with
mesh parameter σ ≥ 2.
We work from (5.38). The arguments from Sect. 4.3.3 are used to bound the first
four terms by ϑcd (ω̄)2 ≤ CN −2 ln2 N .
[2]

We are left with the last term in (5.38). On a Shishkin mesh with σ ≥ 1:

CN −1 ln N for i = 1, . . . , qN,
|ui − ui−1 | ≤
CN −1 for i = qN + 1, . . . , N.

This can be verified using the decomposition of Theorem 3.48.

Furthermore, if hi ≤ mεβ −1 then
    
 
ρ − bi−1/2 hi − ρ bi−1/2 hi  ≤ C hi ,
 ε ε  ε

because ρ is Lipschitz continuous on [−m, m]. Hence, in the layer region of the
Shishkin mesh we have
    
 
ρ − bi−1/2 hi − ρ bi−1/2 hi  ≤ CN −1 ln N for i = 1, . . . , qN,
 ε ε 

while on the coarse mesh region ρ(t) ∈ [0, 1] for all t ∈ IR is used. We obtain
    
 
ρ − bi−1/2 hi − ρ bi−1/2 hi  |ui − ui−1 | ≤ CN −1 .
 ε ε 

All terms in (5.38) have been bounded by either CN −1 or by CN −2 ln2 N .

Theorem 5.27. Assume ρ is Lipschitz continuous on [−m, m]. Let (ρ1 ) and (ρ3 )
hold. Then the error of the upwind FVM (5.22) on a standard Shishkin mesh satisfies
 
u − uN  ≤ CN −1
∞,ω

if N is larger than some threshold value which is independent of the perturbation

parameter ε.
Remark 5.28. On a standard Shishkinmesh the use
 of aLipschitz
 continuous func-
tion ρ improves the accuracy from O N −1 ln N to O N −1 . ♣
182 5 Finite Element and Finite Volume Methods

Table 5.4 The upwind FVM on a standard Shishkin mesh

ρU,0 ρU,10 ρS ρI
N error rate error rate error rate error rate
27 4.236e-3 0.84 3.855e-3 0.99 3.855e-3 0.99 3.855e-3 0.99
28 2.364e-3 0.86 1.942e-3 0.99 1.942e-3 0.99 1.942e-3 0.99
29 1.303e-3 0.87 9.745e-4 1.00 9.745e-4 1.00 9.745e-4 1.00
210 7.111e-4 0.89 4.882e-4 1.00 4.882e-4 1.00 4.882e-4 1.00
211 3.850e-4 0.89 2.443e-4 1.00 2.443e-4 1.00 2.443e-4 1.00
212 2.071e-4 0.90 1.222e-4 1.00 1.222e-4 1.00 1.222e-4 1.00
213 1.108e-4 0.91 6.115e-5 1.00 6.115e-5 1.00 6.115e-5 1.00
214 5.904e-5 0.91 3.059e-5 1.00 3.059e-5 1.00 3.059e-5 1.00
215 3.133e-5 0.92 1.531e-5 1.00 1.531e-5 1.00 1.531e-5 1.00
216 1.658e-5 — 7.672e-6 — 7.672e-6 — 7.672e-6 —

5.4.4 A Numerical Example

Table 5.4 displays the results of test computations using the upwind FVM with
various stabilising functions ρ, when applied to the test problem (4.14) and contains
the maximum nodal errors. For our tests we have chosen a standard Shishkin mesh
with σ = 1 and q = 1/2. The results of the numerical tests are in agreement with
Theorem 5.26 and 5.27. Comparing the numbers for ρU,0 with those for other ρs,
we clearly see an improvement in the accuracy when ρ is Lipschitz continuous in a
neighbourhood of t = 0. Also notice there is no (visible) difference in using either
of those Lipschitz continuous ρs.
Chapter 6
Discretisations of Reaction-Convection-Diffusion
Problems

This chapter is concerned with discretisations of the stationary linear reaction-

convection-diffusion problem

−εd u − εc bu + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1 ,

with b ≥ 1 and c ≥ 1 on [0, 1].

In particular, we shall study the special case of scalar reaction-diffusion problems

−ε2 u + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1 ,

and its vector-valued counterpart

−E 2 u + Au = f in (0, 1), u(0) = γ 0 , u(1) = γ 1 .

6.1 Reaction-Diffusion

This section is concerned with scalar reaction-diffusion problems

Lu := −ε2 u + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1 , (6.1)

where c ≥ ρ2 on [0, 1] and ρ > 0 is a constant.

Analytical properties of (6.1) were studied in Sect. 3.3, while layer-adapted
meshes for it have been introduced in Sect. 2.2. The crucial quantity for reaction-
diffusion problems is
  
1 + ε−1 e−ρx/pε + e−ρ(1−x)/pε dx.
[p]
ϑrd (ω̄) := max
k=1,...,N Ik

Using (6.1) as a model problem, a convergence analysis is conducted for variants

of the linear FEM with convergence established in the energy norm.
Next a general convergence theory in the maximum norm is derived for central
differencing on arbitrary meshes. The close relationship between the differential

T. Linß, Layer-Adapted Meshes for Reaction-Convection-Diffusion Problems, Lecture 183

Notes in Mathematics 1985, DOI 10.1007/978-3-642-05134-0 6,
c Springer-Verlag Berlin Heidelberg 2010
184 6 Discretisations of Reaction-Convection-Diffusion Problems

operator and its discretisation is highlighted. We then move on to a maximum-norm

error analysis for the linear FEM and a special 4th -order scheme. Finally, central
differencing for systems of reaction-diffusion equations is studied.

6.1.1 Linear Finite Elements

Consider (6.1) with homogeneous boundary conditions. Its weak formulation is:
Find u ∈ H01 (0, 1) such that

a(u, v) := ε2 (u , v  ) + r (u, v) = f (v) for all v ∈ H01 (0, 1) (6.2)

1 
with r(u, v) := (cu, v) and f (v) := (f, v) := 0 f v (s)ds.
Given a mesh ω̄, let V0ω be the space of continuous functions that are piecewise
linear on the mesh ω̄. Clearly V0ω ⊂ H01 (0, 1). The standard Galerkin-FEM approx-
imation is: Find uN ∈ V0ω such that

a(uN , v) = f (v) for all v ∈ V0ω .

Typically, the integrals in the bilinear form r(·, ·) and in the linear functional f (·)
cannot be evaluated exactly. Therefore, they have to be approximated:

r(w, v) ≈ r̂(w, v) and f (v) ≈ fˆ(v).

Different approximations yield different variants of the FEM. Later we shall con-
sider four possible choices for r(·, ·) and f (·).
With this notation our FEM is: Find uN ∈ V0ω such that
   
â(uN , v) := ε2 (uN ) , v  + r̂ uN , v = fˆ(v) for all v ∈ V0ω . (6.3)

The norm naturally associated with the weak formulation is the energy norm
1/2
|||v|||ε2 := ε2 v  0 + ρ2 v0
2 2
.

It is typically used in the convergence analysis of FEMs. Clearly a(·, ·) is coercive

in the energy norm:
2
|||v|||ε2 ≤ a(v, v) for all v ∈ H01 (0, 1).

However, the coercivity of â(·, ·) depends on the approximation used for the reaction
term and has to be investigated separately. It is one ingredient in the error analysis
for (6.3). The other ingredient is bounds for the interpolation error.
6.1 Reaction-Diffusion 185

6.1.1.1 The Interpolation Error

Let again wI denote the piecewise linear interpolant of w. Throughout this section
let us assume the function ψ ∈ C 2 [0, 1] admits the derivative bounds
 
|ψ  (x)| ≤ C 1 + ε−2 e−ρx/ε + e−ρ(1−x)/ε . (6.4)

For example, the solution u of the boundary-value problem (6.1) belongs to this
class of functions, see Sect. 3.3.1.2. And so does cu − f = ε2 u whose interpolation
error will appear in the later analysis too.

Proposition 6.1. Suppose ψ satisfies (6.4). Then

*   +2
 
ψ − ψ I  ≤C 1 + ε−1 e−ρx/2ε + e−ρ(1−x)/2ε dx
∞,I i
Ii

for all mesh intervals Ii = [xi−1 , xi ].

Proof. For x ∈ Ii and an arbitrary integrable functions χ set

  x  s
1
(Ji χ) (x) := χ(t)dtdξds.
hi Ii xi−1 ξ

For triple integrals of this structure we have the two bounds


|(Ji χ) (x)| ≤ (ξ − xi−1 ) |χ(ξ)| dξ (6.5a)
Ii

and

|(Ji χ) (x)| ≤ (xi − s) |χ(s)| ds. (6.5b)
Ii

These integrals can be further bounded using Lemma 4.16. Let χ : Ij → IR be any
function with χ ≥ 0 on Ii . Then
 2
1
(Ji χ) (x) ≤ χ(t)1/2 dt if χ is decreasing, (6.6a)
2 Ii

and
 2
1
(Ji χ) (x) ≤ χ(t)1/2 dt if χ is increasing. (6.6b)
2 Ii
186 6 Discretisations of Reaction-Convection-Diffusion Problems

For the interpolation error on Ii we have the representation

   
ψ − ψ I (x) = Ji (ψ  ) (x) for x ∈ Ii . (6.7)

Next we would like to apply (6.5) and (6.6). Therefore, we split ψ  into two parts
that can be bounded by monotone functions—one decreasing and the other increas-
ing. Set

 ψ  (x) for x ≤ 1/2,
ψ̄I := ψ − ψ̄D and ψ̄D (x) :=
0 otherwise.

Clearly,
   
ψ̄D (x) ≤ C 1 + ε−2 e−ρx/ε and ψ̄I (x) ≤ C 1 + ε−2 e−ρ(1−x)/ε ,

by (6.4). Hence, using (6.5) and (6.6), we obtain

  

Ji (ψ )∞,Ii ≤ C 1 + ε−1 e−ρx/2ε + ε−1 e−ρ(1−x)/2ε dx.
Ii

Recalling (6.7), we are finished. 



Theorem 6.2. Suppose ψ satisfies (6.4). Then

     2
ψ − ψ I  ≤ ψ − ψ I  ≤ C ϑ[2] (ω̄)
0 ∞ rd

and
   
ψ − ψ I  [2] [2]
≤ C ε1/2 + ϑrd (ω̄) ϑrd (ω̄).
ε2

Proof. The bound on the L∞ error is an immediate consequence of Prop. 6.1 and
[2]
the definition of ϑrd .
For the error in the H 1 norm we proceed as follows using integration by parts
  
  
2
1   2 1  
 ψ − ψI  = ψ − ψI (x) dx = − ψ  (x) ψ − ψ I (x)dx.
0 0 0

Next, a Hölder inequality gives

 
  
2   1  2
 ψ − ψ I  ≤ ψ − ψ I ∞ |ψ  (x)| dx ≤ Cε−1 ϑrd (ω̄) ,
[2]
0 0

by (6.4). Finally, combine this with the bound for the L2 norm of the interpolation
error to obtain the energy-norm estimate. 

6.1 Reaction-Diffusion 187

Remark 6.3. The interpolation error in the energy norm is an order of magnitude
better than for convection-diffusion problems. This is because this norm fails to
capture the wider boundary layers in reaction-diffusion problems:
 −ρx/ε   
e  = O ε1/2 for ε → 0.
ε2

Therefore, the reaction-diffusion problem is not singularly perturbed in the energy

norm in the sense of Def. 1.1. ♣

Remark 6.4. For Lagrange interpolation with polynomials of arbitrary degree p ≥ 0,

cf. Remark 5.5, we have
 p+1
[p+1]
Ip u − u∞ ≤ C ϑrd (ω̄)

for the solution u of (6.1). ♣

6.1.1.2 Convergence in the Energy Norm

With interpolation error bounds at hand, we can now return to the convergence anal-
ysis for the FEM (6.3).
Assume the bilinear form â(·, ·) is V0ω -coercive with respect to the energy norm.
That is, there exists a positive constant γ such that
2
γ |||v|||ε2 ≤ â(v, v) for all v ∈ V0ω . (6.8)

Set η := uI − u and χ := uI − uN . Then by a triangle inequality

 
u − uN  ≤ |||η|||ε2 + |||χ|||ε2 .
ε2

Theorem 6.2 yields

 
[2] [2]
|||η|||ε2 ≤ C ε1/2 + ϑrd (ω̄) ϑrd (ω̄),

and we are left with bounding |||χ|||ε2 .

Starting from (6.8), we get
2  
γ |||χ|||ε2 ≤ â(χ, χ) = r̂ uI , χ − r(u, χ) + f (χ) − fˆ(χ), (6.9)

where we have used (6.2) and (6.3).

We shall consider four variants of the FEM characterised by different approxi-
mations of the reaction term and of the source term:
FEM-0: r̂(w, v) = r(w, v) and fˆ(v) = f (v), i.e., no quadrature is used.
FEM-1: r̂(w, v) = (cI w, v) and fˆ(v) = (f I , v),
188 6 Discretisations of Reaction-Convection-Diffusion Problems

FEM-2: r̂(w, v) = ((cw)I , v) and fˆ(v) = (f I , v),

FEM-3: r̂(w, v) = (cw, v)ω and fˆ(v) = (f, v)ω with the discrete 2 -product


N −1
(w, v)ω := h̄i wi vi .
i=1

This method is generated by applying the trapezium rule

 
hi+1  
g(s)ds ≈ gi + gi+1 = g I (s)ds
Ii 2 Ii

to (cw, v) and (f, v). It is equivalent to the standard central difference scheme
which will be subject of Sect. 6.1.2.

Remark 6.5. A direct calculation shows that (w, v)ω = (w, v) for all w, v ∈ V0ω . ♣

Theorem 6.6. Let u be the solution of (6.1) and uN its approximation by FEM-0,
FEM-1 or FEM-2. Then
 I   2
u − uN  2 ≤ C ϑ[2] (ω̄) , (6.10)
ε rd

and
   
u − uN  [2] [2]
≤ C ε1/2 + ϑrd (ω̄) ϑrd (ω̄).
ε2

Corollary 6.7. For FEM-0, FEM-1 and FEM-2 we have the uniform first-order
convergence result
 
u − uN  [2]
≤ Cϑrd (ω̄).
ε2

However, the worst case is not when ε is small, but when ε = 1. This is observed in
numerical experiments [74].

We give short proofs of Theorem 6.6 for the various FEMs now.

FEM-0: r̂(w, v) = r(w, v), fˆ(v) = f (v).

Clearly â(·, ·) = a(·, ·). Therefore, (6.8) holds with γ = 1. Inequality (6.9) yields
2    
|||χ|||ε2 ≤ r uI − u, χ = c(uI − u), χ

and
   2
|||χ|||ε2 ≤ c∞ uI − u0 ≤ C ϑrd (ω̄) ,
[2]

by Theorem 6.2.
6.1 Reaction-Diffusion 189

FEM-1: r̂(w, v) = (cI w, v), fˆ(v) = (f I , v).

The coercivity of â(·, ·) is verified upon noting that
    2
cI w, w ≥ ρ2 w, w = ρ2 w0 .

Thus, (6.8) is satisfied with γ = 1. Starting from (6.9) again, we get

2      
|||χ|||ε2 ≤ (cI − c)uI , χ + c(uI − u), χ + f − f I , χ .

Appealing to Theorem 6.2 again, we get (6.10) for FEM-1.

FEM-2: r̂(w, v) = ((cw)I , v), fˆ(v) = (f I , v).

This time establishing coercivity is slightly more involved. Let w ∈ V0ω be arbitrary.
A direct calculation gives

hi+1   N

N −1 −1
hi+1
r̂(w, w) = ci wi2 + ci+1 wi+1
2
+ (ci + ci+1 ) wi wi+1 .
i=0
3 i=0
6

We bound the second term from below:

ci + ci+1  2 
(ci + ci+1 ) wi wi+1 ≥ − 2
wi + wi+1
2
   
hi+1  hi+1 
≥ − ci + c ∞ wi2 − ci+1 + c ∞ wi+1
2
.
2 2

Thus, if the maximum step size h is sufficiently small, dependent on κ, but indepen-
dent of ε, then

hi+1  

N −1
r̂(w, w) ≥ ci wi2 + ci+1 wi+1
2

i=0
8

 −1
hi+1  2  ρ2
N
≥ ρ2 2
wi + wi+1 = w20 .
i=0
8 4

Hence, â(·, ·) is coercive and (6.8) holds true for γ = 1/4.

Next, (6.9) and the Cauchy-Schwarz inequality yield
1  
|||χ|||ε2 ≤ q I − q 0 , with q := cu − f.
4
Theorem 6.2 applies to q and we obtain (6.10) for FEM-2.
190 6 Discretisations of Reaction-Convection-Diffusion Problems

FEM-3: r̂(w, v) = (cw, v)ω , fˆ(v) = (f, v)ω .

In view of Remark 6.5, ineq. (6.8) holds true with γ = 1. Then by (6.9)

2
1  
|||χ|||ε2 ≤ (q, χ)ω − (q, χ) = (qχ)I − qχ (x)dx
0
−1 2
  N hi+1
= q I − q, χ + (qi+1 − qi ) χx;i ,
i=0
6

where again q = cu − f = ε2 u . The first term on the right-hand has just been
bounded when analysing FEM-2. Unfortunately, in view of the last term—in par-
ticular the presence of the discrete derivative χx —it seems impossible to obtain a
convergence result as general as Theorem 6.6. On a S-type mesh, one might reason
as in Sect. 5.2.1 by using an inverse inequality on the coarse-mesh region, but rely
on the small mesh sizes inside the layer to gain the necessary powers of ε.

6.1.2 Central Differencing

Given an arbitrary mesh ω̄ the most frequently considered finite-difference approx-

imation of (6.1) is: Find uN ∈ IRN +1 such that
' N(
Lu i = fi for i = 1, . . . , N − 1, uN N
0 = γ0 , uN = γ1 (6.11)

with

[Lv]i := −ε2 vx̄x̂;i + ci vi for v ∈ IRN +1 .

The difference operators were introduced in Sect. 4.1. As mentioned before this
difference scheme is equivalent to FEM-3 in the preceding section: Find uN ∈ V0ω
such that
   
ac uN , v = fc v := (f, v)ω for all v ∈ V0ω , (6.12)

where

ac (w, v) := ε2 (w , v  ) + (cw, v)ω = ε2 [wx , vx )ω + (cw, v)ω ,

and


N −1 
N −1
[w, v)ω := hi+1 wi vi , (w, v)ω := h̄i wi vi .
i=0 i=1
6.1 Reaction-Diffusion 191

Taking as test functions v the standard hat-function basis in V0ω , we see

that (6.11) and (6.12) are equivalent. In particular, using summation by parts
(cf. [146]) it is verified that

ac (w, v) = (Lw, v)ω = (w, Lv)ω for all w, v ∈ V0ω .

Apparently the operator L is self-adjoint as is its continuous counterpart L.

6.1.2.1 Stability

The matrix associated with the difference operator L is an L0 -matrix because all
off-diagonal entries are non-positive. Application of the M -criterion (Lemma 3.14)
with a test vector with components ei = 1 establishes the inverse monotonicity of L.
Thus, L satisfies a comparison principle: For any mesh functions v, wIRN +1
⎫
Lv ≤ Lw on ω, ⎬
v0 ≤ w0 , =⇒ v ≤ w on ω̄. (6.13)
⎭
vN ≤ wN

This comparison principle and Lemma 3.17 provide a priori bounds for the solu-
tion of (6.11):
 N !
u  ≤ max |γ0 |, |γ1 | + f /c for i = 0, . . . , N.
ω̄ ∞,ω

It also gives the stability inequality

v∞,ω̄ ≤ Lv/c∞,ω for all v ∈ IR0N +1 . (6.14)

Green’s function estimates

Using the discrete Green’s function G : ω̄ 2 → IR : (xi , ξj ) → Gi,j = G(xi , ξj )
associated with L and Dirichlet boundary conditions, any mesh function v ∈ IR0N +1
can be represented as

vi = ac (v, Gi,· ) = (Lv, Gi,· )ω = (v, LGi,· )ω for i = 1, . . . , N − 1, (6.15)

because the operator L is self-adjoint, i.e., L = L∗ . This also implies Gi,j = Gj,i
for all i, j = 0, . . . , N .
Taking for v the standard basis in V0ω , we see that for fixed i = 1, . . . , N − 1

[LGi,· ]j = δi,j for j = 1, . . . , N − 1, Gi,0 = Gi,N = 0, (6.16)

192 6 Discretisations of Reaction-Convection-Diffusion Problems

where

h̄−1
i if i = j,
δi,j :=
0 otherwise

is the discrete equivalent of the Dirac-δ distribution.

Lemma 6.8. For any fixed μ > 0

 
v∞ ≤ μ [vx , vx )ω + μ−1 + 1 (v, v)ω
2
for all v ∈ V0ω .

Proof. This is Theorem 2 in chap. V, §4 of [145]. 



Theorem 6.9. The Green’s function G associated with the discrete operator L and
Dirichlet boundary conditions satisfies

1 + ερ−1
0 ≤ Gi,j ≤ for 0 ≤ i, j ≤ N,
ερ
Gξ;i,j ≥ 0 for 0 ≤ j < i < N,
Gξ;i,j ≤ 0 for 0 ≤ i ≤ j < N.

Proof. The positivity of G follows from the inverse monotonicity of L.

The upper bound on G is proved using an argument from [147]. Let i be arbitrary,
but fixed. Set Γ := Gi,· . Then Lemma 6.8 with μ = ε/ρ yields

1 + ερ−1  
ε [Γξ , Γξ )ω + ε−1 γ 2 (Γ, Γ )ω
2
Γ ∞,ω ≤
ρ
1 + ερ−1  
≤ ε [Γx , Γx )ω + ε−1 (cΓ, Γ )ω
ρ
1 + ερ−1
= Γi ,
ρε

by (6.15). The upper bound on G follows.

When proving the monotonicity, note that

ε2 Gξ;i,j = ε2 Gξ;i,j−1 + h̄j cj Gi,j for 0 < j < i

and

ε2 Gξ;i,j = ε2 Gξ;i,j+1 − h̄j+1 cj+1 Gi,j+1 for i < j < N − 1.

The non-negativity of G implies Gξ;i,0 ≥ 0 and Gξ;i,N −1 ≤ 0. Then the piecewise

monotonicity of G follows by induction. 

6.1 Reaction-Diffusion 193

Theorem 6.10. The Green’s function G associated with the discrete operator L
satisfies
 2
2 1 + ερ−1 2
cGi,· 1,ω ≤ 1, Gξ;i,· 1,∗ ω ≤ , Gξξ;i,· 1,ω ≤
ερ ε2

for all i = 1, . . . , N − 1.

Proof. First multiply (6.16) by h̄j and sum for j = 1, . . . , N − 1.

 −1
N
 
h̄j cj Gi,j = 1 + ε2 Gξ̄;i,N − Gξ;i,0 ≤ 1,
j=1

because Gξ;i,0 ≥ 0 and Gξ;i,N −1 ≤ 0. This is the bound on the c-weighted 1 norm.
Next
−1  2

i−1 
N
2 1 + ερ−1
Gξ;i,· 1,∗ ω = hj+1 Gξ;i,j − hj+1 Gξ;i,j = 2Gi,i ≤ ,
j=0 j=i
ερ
(6.17)

by Theorem 6.9.
Finally, a triangle inequality, (6.16) and (6.17) give the bound on Gξξ . 


6.1.2.2 A Priori Error Bounds

The analysis follows [95, 103]. By (6.15) we have for the error u − uN in the mesh
node xi
   
u − uN i
= ac u − uN , Gi,· = ac (u, Gi,· ) − fc (Gi,· ) .

For simplicity we set Γ := Gi . We identify the mesh function Γ with that function
from V0ω that coincides with Γ at the mesh nodes. Using the weak form of the
differential equation, we get
 
u − uN i
= ac (u, Γ ) − a (u, Γ ) + f (Γ ) − fc (Γ )
= (cu − f, Γ )ω − (cu − f, Γ ) .

Note, if v0 = vN = 0 then
 1 I
(w, v)ω = wv (x)dx, (6.18)
0

where wI denotes again the piecewise linear interpolant of w.

194 6 Discretisations of Reaction-Convection-Diffusion Problems

Setting q := cu − f = ε2 u , we obtain the error representation

 1 
   I
u−u N
i
= qΓ − qΓ (x)dx. (6.19)
0

We are left with bounding the interpolation error for qΓ . To this end we shall avail
of the derivative bounds derived in Sect. 3.3.1.2 and repeat some of the arguments
from Sect. 6.1.1.1.
By (6.7), for x ∈ [xj , xj+1 ] we have
   I    
qΓ (x) − qΓ (x) = 2Γx;j Jj (q  ) (x) + Jj (q  Γ ) (x). (6.20)

Next, we wish to apply (6.5) and (6.6) to the right-hand side of (6.20). There-
fore, we split q  into two parts that can be bounded by monotone functions—one
decreasing and the other increasing. Set

 q  (x) for x ≤ 1/2,
qI := q − qD and qD (x) :=
0 otherwise.

From Sect. 3.3.1.2 we have

ε−1 |qD (x)| ≤ C 1 + ε−2 e−ρx/ε

and

ε−1 |qI (x)| ≤ C 1 + ε−2 e−ρ(1−x)/ε .

Hence, using (6.5) and (6.6), we obtain

 2
Jj (q  )∞,Ij ≤ Cε ϑrd (ω̄) .
[2]
(6.21)

The second integral in (6.20) is bounded in a similar manner. Set


 q  (x) for x ≤ 1/2,
q̄I := q − q̄D and q̄D (x) :=
0 otherwise.

From Sect. 3.3.1.2:

|(Γ q̄D ) (x)| ≤ C (Γj + Γj+1 ) 1 + ε−2 e−ρx/ε

and

|(Γ q̄I (x))| ≤ C (Γj + Γj+1 ) 1 + ε−2 e−ρ(1−x)/ε ,

6.1 Reaction-Diffusion 195

since Γ is piecewise linear and positive. Thus,

 2
Jj (q  Γ )∞,Ij ≤ C (Γj + Γj+1 ) ϑrd (ω̄) .
[2]
(6.22)

Applying (6.21) and (6.22) to (6.20), we obtain

    2
 I  [2]
 Γ q − (Γ q) (x) ≤ C ϑrd (ω̄) (εΓx̄;j + Γj+1 + Γj ) for x ∈ (xj , xj+1 ).

Integrate over [0, 1] to get

 1   2  
   
 (Γ q) − (Γ q)I
(x)dx ≤ C ϑ[2] (ω̄) ε Γ  ∗ ω + Γ 1,ω .
  rd x 1,
0

Finally, recall (6.19) and Theorem 6.10. We arrive at the main convergence result
of this section.
Theorem 6.11. Let u be the solution of the reaction-diffusion problem (6.1) and uN
its central difference approximation (6.11). Suppose c, f ∈ C 2 [0, 1]. Then
   2
u − uN  ≤ C ϑ
[2]
(ω̄) .
∞,ω rd

Corollary 6.12. Theorems 6.2 and 6.11 yield

   2
u − uN  ≤ C ϑ[2] (ω̄) ,
∞ rd

   
by a triangle inequality and because uI − uN ∞ = uI − uN ∞,ω .

6.1.2.3 A Posteriori Error Analysis

The first a posteriori analysis for central differencing was conducted by Kopteva;
see [66]. We slightly modify her argument. It is based on the bounds for the Green’s
function G associated with the continuous operator L; see Theorem 3.31.
Let x ∈ (0, 1) be arbitrary, but fixed. Set Γ = G(x, ·). Then
     
u − uN (x) = a u − uN , Γ = f (Γ ) − a uN , Γ
 
= f (Γ ) − fc (Γ ) + ac (uN , Γ ) − a uN , Γ
   
= cuN − f, Γ ω − cuN − f, Γ ,

by (6.12). Setting q̂ := cuN − f , we have

 1 
   I
u − uN (x) = q̂Γ − q̂Γ (s)ds,
0
196 6 Discretisations of Reaction-Convection-Diffusion Problems

where (6.18) was used. Expand the integrand

 I     I
q̂Γ − q̂Γ = q̂ I − q̂ Γ + q̂ I Γ I − Γ + (q̂Γ ) − q̂ I Γ I

in order to obtain the error representation

     
u − uN (x) = q̂ I − q̂, Γ + q̂ I , Γ I − Γ
 1  (6.23)
I
+ (q̂Γ ) − q̂ I Γ I (s)ds.
0

The terms on the right-hand side are estimated separately.

A Hölder inequality and Theorem 3.31 give
   
 q̂ − q̂ I , Γ  ≤ ρ−2 q̂ − q̂ I  .
∞

For the second term in (6.23), integration by parts yields

 
 
q̂ I (s) Γ I − Γ (s)ds = (s − xk+1 ) (s − xk ) Qk (s)Γ  (s)ds,
Ik Ik

where

q̂k+1 + q̂k q̂k+1 − q̂k   xk + xk+1

Qk (s) = + s − xk+1/2 , xk+1/2 = .
4 6hk+1 2

Thus
  
   h2k+1  !
 
q̂ (x) Γ − Γ (s)ds ≤
I I
max |q̂k | , |q̂k+1 | |Γ  (s)| ds.
 8
Ik Ik

By Theorem 3.31, we have Γ  1 ≤ 2ε−2 . Hence,

 I I  h2k+1 !
 q̂ , Γ − Γ  ≤ max max |q̂k | , |q̂k+1 | .
k=0,...,N −1 4ε 2

For the last term in (6.23) a direct calculation yields

  
I hk+1
(q̂Γ ) − q̂ I Γ I (s)ds = (q̂k+1 − q̂k ) (Γk+1 − Γk )
Ik 6

hk+1
= (q̂k+1 − q̂k ) Γ  (s)ds.
6 Ik
6.1 Reaction-Diffusion 197

Thus
   
 1  hk+1
 I
(q̂Γ ) − q̂ Γ
I I
(s)ds ≤ max |q̂k+1 − q̂k | ,
 k=0,...,N −1 6ερ
0

because Γ  1 ≤ 1/(ερ), by Theorem 3.31.

All terms in (6.23) have been bounded. We get
Theorem 6.13. Let u be the solution of the reaction-diffusion problem (6.1) and uN
its central difference approximation (6.11). Let q̂ := cuN − f . Then
 
u − uN  ≤ η1 + η2 + η3 ,
∞

where
1  
q̂ − q̂ I  ,
h2k+1 !
η1 := 2 ∞
η2 := max 2
max |q̂k | , |q̂k+1 |
ρ k=0,...,N −1 4ε

and

hk+1
η3 := max |q̂k+1 − q̂k | .
k=0,...,N −1 6ερ
 
Remark 6.14. The term q̂ − q̂ I ∞ in the a posteriori bound can be further ex-
panded as follows
   I
q̂ − q̂ I = f − f I − c − cI uN − cI uN + cuN .

Thus
       
q̂ − q̂ I  ≤ f − f I  + c − cI  uN 
∞ ∞ ∞ ∞
1  N 
+ max  u N
− uk · |ck+1 − ck | .
4 k=0,...,N −1 k+1

The first two terms involve (continuous) norms of the data. These have to be approx-
imated numerically with sufficient accuracy. At least O(h2 ) is required. However,
higher order is desirable to ensure all non explicitly computable terms in the error
estimator are of higher order and decay rapidly as the mesh is refined. ♣
Remark 6.15. Invoking the difference equation (6.11), we see that η3 implicitly con-
tains third-order discrete derivatives of uN .

hk+1 εh2  
|q̂k+1 − q̂k | = k+1 uN 
x̄x̂x;k .
6ερ 6ρ

The a posteriori error bound in [66, Theorem 3.3] is given using these higher-order
operators.
198 6 Discretisations of Reaction-Convection-Diffusion Problems

The argument in [66] proceeds to show that

    2
u − uN  ≤ C max
1/2
hk+1 1 + |ux̄x̂;k | + |ux̄x̂;k+1 |
1/2
.
∞ i=0,...,N −1

The constant C is independent of ε, but not specified in [66]. Therefore, this latest
inequality cannot be used for reliable a posteriori error estimation. Nonetheless, it
is useful for steering adaptive mesh generation. ♣

Adaptive mesh generation

Based on Theorem 6.13, the de Boor algorithm in Sect. 4.2.4.2 can be adapted for
the problem and method under consideration. One only needs to redefine the Qi :
1/2
Qi = Qi (uN , Δ, ω) := κ0 + κ1 η1;i κ2 η2;i + κ3 η3;i

with

  h2i !
η1;i := q̂i−1 − 2q̂i−1/2 + q̂i  , η2;i := max |q̂i−1 | , |q̂i | ,
ε2
hi
η3;i := |q̂i − q̂i−1 |
ε
and non-negative weights κk . Note that maxi η1;i is a second order approximation
of η1 in Theorem 6.13.
In view of Remark 6.15 one can also use the de Boor algorithm with
1/2 1/2
Qi = 1 + |ux̄x̂;i−1 | + |ux̄x̂;i | .

Numerical experiment for this variant of the algorithm are documented in [68].

6.1.2.4 An Alternative Convergence Proof

Traditional finite difference analysis aims at directly exploiting the maximum-norm

stability or using barrier function techniques. In higher dimensions they are often
the only tool available, because of a lack of stronger stability results.
We now present an error analysis for central differencing on a Bakhvalov mesh
that solely uses (6.14). In the layer regions this mesh is not approximately equidis-
tant. Consequently, the truncation error of the difference scheme is apparently only
first order at points in the layers, but a more delicate analysis given in [18] shows
that the truncation error at every mesh point is in fact O(N −2 ) uniformly in ε. We
use the description of the mesh by a mesh generating functions, see Sect. 2.1.1.
For any ψ ∈ C 4 [0, 1], Taylor expansions show that
   
[Lψ − Lψ]  ≤ Cε2 ψ   (6.24a)
i [x i−1 ,xi+1 ]
6.1 Reaction-Diffusion 199

and
     
[Lψ − Lψ]  ≤ Cε2 |hi − hi+1 | ψi  + (hi + hi+1 )2 ψ (4)  (6.24b)
i [x i−1 ,xi+1 ].

When σε ≥ ρq the mesh is uniform with mesh size N −1 . Furthermore, ε−1 ≤ C.

Thus,

Lu − Luω ≤ CN −2 ,

by Theorem 3.35 and (6.24).

Now consider the case σε < ρq. For simplicity we will consider only the layer
at x = 0 and assume that xi = ϕ(ti ) ≤ 1/2.
From the construction of ϕ one must have τ < q. It follows that 1 < χ (τ ) < q̄,
where we set q̄ = 1/(1 − 2q). Define the auxiliary points τ1 and τ2 in (0, q) by
χ (τ1 ) = q̄ and χ (τ2 ) = 1. Then τ2 = q − σε/ρ < τ < τ1 = q − σε(1 − 2q)/ρ
because χ > 0 on [0, q).
(i) ϕ (t) ≤ χ (τ ) ≤ q̃ for t ∈ [0, 1]. Thus,
 ti
hi = ϕ (t)dt ≤ q̃N −1 for i = 1, . . . , N. (6.25)
ti−1

(ii) For t ≤ ti < q we have ϕ (t) ≤ χ (t) = σε/ρ(q − t) ≤ σε/ρ(q − ti ). Hence,

for ti ≤ q − N −1 ,
 ti
σε 2σε
hi = ϕ (t)dt ≤ N −1 ϕ (ti ) ≤ ≤ . (6.26)
ti−1 ρN (q − ti ) ρN (q − ti−1 )

(iii) hi+1 − hi = xi+1 − 2xi + xi−1 = ϕ (t∗i )N −2 for some t∗i ∈ [ti−1 , ti+1 ].
Now
σε 1 1 ρq̄
ϕ (t) ≤ χ (τ ) = and ≤ = ,
ρ(q − τ )2 q−τ q − τ1 σε

which gives

ρq̄ 2 −2
|hi+1 − hi | ≤ N . (6.27)
σε
Furthermore, we have the bound
σε 4σε 2
ϕ (t∗i ) ≤ ≤ for ti ≤ q − ,
ρ(q − ti+1 )2 ρ(q − ti )2 N

which yields
4σε 2
|hi+1 − hi | ≤ for ti ≤ q − . (6.28)
N 2 ρ(q − ti )2 N
200 6 Discretisations of Reaction-Convection-Diffusion Problems

(iv)
 σ
−ρxi /ε q − ti
e = for ti ≤ τ (6.29)
q

and
 σ
−ρxi /ε σε
e ≤ for ti ≥ τ2 . (6.30)
ρq

Henceforth let σ ≥ 2. Using (6.24), Theorem 3.35, (6.25) and (6.30), we get

|[Lu − Lu]i | ≤ CN −2 for τ2 ≤ ti−1 ,

which is the region outside the layer. For ti ≤ q − 2N −1 (the layer region),
from (6.24) and Theorem 3.35 one arrives at

|[Lu − Lu]i | ≤ C |hi − hi+1 | ε2 + C |hi − hi+1 | ε−1 e−ρxi /ε

+ C (hi + hi+1 ) ε2 + C (hi + hi+1 ) ε−2 e−ρxi−1 /ε .
2 2

To bound the first term on the right-hand side use (6.27); for the second term, use
(6.28) and (6.29); for the third term, use (6.25); and for the fourth, use (6.26), (6.29)
and q − ti−1 ≤ 3(q − ti )/2. This yields

|[Lu − Lu]i | ≤ CN −2 for ti ≤ q − 2N −1 .

We are left with the transition region where τ2 > ti−1 and ti > q − 2N −1 . Thus,
2 1 σε 1 1
q− < ti < τ2 + =q− + <q+ .
N N ρ N N

Notice that the first two inequalities here imply that ε < 3ρ/(σN ). Use (6.24):
 
|[Lu − Lu]i | ≤ C ε2 + e−ρxi−1 /ε ≤ CN −2 ,

by (6.30) and ε ≤ CN −1 .
Thus, on a Bakhvalov
  mesh the truncation error in the maximum norm is bounded
uniformly by O N −2 . Application of the stability inequality (6.14) gives the uni-
form error bound
 
u − uN  ≤ CN −2 , if σ ≥ 2.
∞,ω

We have recovered Theorem 6.11 for a Bakhvalov mesh by means of a different

kind of analysis.
6.1 Reaction-Diffusion 201

6.1.2.5 Discontinuous Data

Assume the reaction coefficient c or the source term f possess

 a discontinuity
 in a
point d ∈ (0, 1). Then (6.1) reads: Find u ∈ C 1 [0, 1] ∩ C 2 (0, 1) \ {d} such that

−ε2 u + cu = f in (0, 1) \ {d}, u(0) = γ0 , u(1) = γ1 .

How should the central differencing scheme (6.11) be generalised to deal with the
discontinuous data?
Assuming d = xκ ∈ ω, i.e., the discontinuity is in a mesh point, a naive finite
difference approach would seek a mesh function uN with
' N(
Lu i = fi for xi ∈ ω \ {xκ }, uN N N N
x;κ = ux̄;κ , u0 = γ0 , uN = γ1 .

This method on a Shishkin mesh was analysed in [37]. Only first order (up to a
logarithmic factor) was established. The numerical experiments in [37] show that
when ε is moderate at best first order can be achieved.
The continuity of the derivative in xκ = d is discretised by imposing

uN N
x;κ = ux̄;κ .

However, the one-sided difference operators are first-order approximation only. This
might explain the drop in accuracy.
Instead, we start from the variational formulation (6.12): Find uN ∈ V0ω such
that
     
ε2 uN , v  − cuN , v ω = (f, v)ω for all v ∈ V0ω ,

where


N −1  1
I
(w, v)ω := h̄i wi vi = (wv) (s)ds.
i=1 0

For discontinuous functions the nodal interpolant does not exist. However, since the
point d of discontinuity is a mesh point, we can define wI element wise:

xk+1 − x x − xk
wI (x) = w(xk + 0) + w(xk+1 − 0) for x ∈ Ik .
hk+1 hk+1

Clearly, when w ∈ C[0, 1], we recover the standard linear nodal interpolant.
202 6 Discretisations of Reaction-Convection-Diffusion Problems
1 I
Evaluating 0
(wv) (s)ds, we get


N −1
hi w(xi − 0) + hi+1 w(xi + 0)
(w, v)ω = h̄i w̃i vi , with w̃i = .
i=1
2h̄i

Note that w̃i = w(xi ) for i = κ.

The resulting difference scheme is: Find uN ∈ IRN +1 such that

LuN := −ε2 uN N ˜ on ω, uN = γ0 , uN = γ1 .
x̄x̂ + c̃u = f 0 N

This scheme was analysed by Boglaev and Pack [22]. They establish uniform con-
vergence of first order. Roos and Zarin [142] consider the scheme with w̃i =
(w(xi − 0) + w(xi + 0)) /2, but in the critical point xκ they have hκ = hκ+1 .
In that article uniform second order convergence is proved for Shishkin meshes and
for Bakhvalov-Shishkin meshes.
Using the derivative bounds derived in Sect. 3.3.1.3, the analysis of Sect. 6.1.2.2
needs only minor modifications to get the pointwise error bound
   2
u − uN  ≤ C ϑ[2]i (ω̄) ,
∞ rd

[p]
where ϑrdi (ω̄) has been defined in Sect. 2.2.1.

6.1.3 A Non-Monotone Scheme

In this section we shall present maximum-norm error bounds for a FEM applied
to (6.1). We consider FEM-2. It generates the difference scheme
' N( 7
Lu i := −ε2 uN N 6 for i = 1, . . . , N − 1,
x̄x̂;i + cu i = fi
(6.31)
uN
0 = γ0 , uN
N = γ1 ,

where
hi wi−1 2wi hi+1 wi+1
6i :=
w + + .
h̄i 6 3 h̄i 6

We see that the discretisation of the reaction term cu generates positive off-diagonal
entries in the system matrix. This results in a scheme that is—unlike the central dif-
ference scheme studied before—not inverse monotone. Nonetheless, a maximum-
norm error analysis can be conducted. We follow [97].
6.1 Reaction-Diffusion 203

6.1.3.1 Stability

Although L is not inverse-monotone, it possesses a core that is:

2ci
[Λv]i := −ε2 vx̄x̂;i + vi .
3

This is the standard central finite difference approximation of −ε2 u + 23 cu and
can be generated by means of the bilinear form
 1
2 I
ε2 (w , v  ) + cwv (s)ds.
3 0

By (6.14) we have
 
3 Λv 
v∞,ω̄ ≤   for all v ∈ IR0N +1 . (6.32)
2  c ∞,ω

Theorem 6.16. Suppose c ∈ C 0,α [0, 1]. Let κ ∈ (0, 1) be arbitrary, but fixed. Then
 
3   Lv 

v∞,ω̄ ≤ for all v ∈ IR0N +1 ,
1 − κ  c ∞,ω

provided that the maximum step size h is smaller than some threshold value that is
independent of ε.
Remark 6.17. Theorem 6.16 means the non-monotone scheme (6.31) is (∞ , ∞ )-
stable although the underlying operator L is not inverse monotone and does not
satisfy a maximum principle. ♣
Proof. Let v ∈ IR0N +1 be an arbitrary mesh function. Then
hi ci−1 hi+1 ci+1
[Λv]i = − vi−1 − vi+1 + [Lv]i .
h̄i 6 h̄i 6

Thus,
hi ci−1 + hi+1 ci+1
|[Λv]i | ≤ v∞,ω̄ + |[Lv]i | for i = 1, . . . , N − 1,
6h̄i

and the stability inequality (6.32) yields

 
3 hi ci−1 + hi+1 ci+1 3  Lv 
v∞,ω̄ ≤ max v∞,ω̄ +   . (6.33)
2 i=1,...,N −1 6ci h̄i 2  c ∞,ω

Since c ∈ C 0,α [0, 1], there exists a constant M with

α
|c(x) − c(ξ)| ≤ M |x − ξ| for all x, ξ ∈ [0, 1]. (6.34)
204 6 Discretisations of Reaction-Convection-Diffusion Problems

Therefore,
hi ci−1 + hi+1 ci+1 1 M hα 1+κ
≤ + 2
≤ ,
6ci h̄i 3 3ρ 3

provided h is smaller than some threshold value that is independent of ε. Now, the
proposition follows from (6.33). 


For our convergence analysis, we shall also require bounds on the discrete
Green’s function Gi associated with L and the mesh node xi . With Theorem 6.10
we have
 1 
N −1  1
I 3 3
(cGi ) (s)ds = h̄j cj Gi,j ≤ and Gi (s)ds ≤ . (6.35)
0 j=1
2 0 2ρ2

6.1.3.2 A Priori Error Analysis

Theorem 6.18. Let u be the solution of (6.1) and uN that of (6.31). Then
   2
u − uN  ≤ C ϑ[2] (ω̄) ,
∞ rd

provided that h is smaller than some threshold value that is independent of ε.

Proof. By a triangle inequality

     
u − uN  ≤ u − uI  + u − uN  .
∞ ∞ ∞,ω̄

The interpolation error was studied in Sect. 6.1.1.1.

Let η = uN − u and q := f − cu. Then the Green’s-function representation and
Eqs. (6.1) and (6.31) yield—after some calculations—

   I
 2 1
ηi = q I − q, Gi − (cη) , Gi + (ceGi )I (6.36)
3 0
 
where we used (uI ) , Gi = (u , Gi ), because Gi is piecewise constant. The first
term on the right-hand side of (6.36) is bounded using a Hölder inequality, Theorem
6.2 and (6.35):
 I   2
 q − q, Gi  ≤ C ϑ[2] (ω̄) .
rd

For the remaining terms in (6.36) a straight-forward calculation gives

  
2 1
I 1  hk
(cηGi )I − (cη) , Gi = − (ck ηk Gi,k−1 + ck−1 ηk−1 Gi,k ) .
3 0 3 2
6.1 Reaction-Diffusion 205

Let κ ∈ (0, 1) be arbitrary, but fixed. Then, using (6.34) and (6.35), we can estimate
as follows:
  1    1 M hα 
2 1+κ
 (cηG )I
− (cη)
I
, G 
i ≤ + η∞,ω̄ ≤ η∞,ω̄ .
3 i
2 2ρ2 2
0

Thus,
 2 1 + κ
[2]
|ηi | ≤ C ϑrd (ω̄) + η∞,ω̄ .
2

Taking the maximum over i = 1, . . . , N − 1, we get the general error bound of the
theorem. 


Remark 6.19. In contrast to the analysis for central differencing, only bounds for the
L1 norm of the Green’s function have been used, but no bounds on its derivative.
Also no third-order derivative of u is required. Only the second-order derivative is
used when Theorem 6.2 is invoked. ♣

Remark 6.20. The proof is easily adapted to deal with discontinuities in the right-
hand side or in the reaction coefficient. ♣

6.1.3.3 A Posteriori Error Analysis

The analysis in [97] is along the lines of the analysis for central differencing in
Sect. 6.1.2.3.
Set Γ = G(x, ·) and q̂ := cuN − f . Then
     
u − uN (x) = q̂ I − q̂, Γ + q̂ I , Γ I − Γ . (6.37)

Note that compared with (6.23) the term

 1  
I
(q̂Γ ) − q̂ I Γ I (s)ds
0

does not appear in (6.37).

Both terms on the right-hand side of (6.37) have been bounded in Sect. 6.1.2.3.
Theorem 6.21. Let u be the solution of the reaction-diffusion problem (6.1) and uN
its approximation by (6.31). Let q̂ := cuN − f . Then

    h2k+1 !
u − uN  ≤ 1 q̂ − q̂ I  + max max |q̂k | , |q̂k+1 | .
∞ ρ2 ∞ k=0,...,N −1 4ε 2
206 6 Discretisations of Reaction-Convection-Diffusion Problems

Remark 6.22. Compared with central differencing, we see the term

hk+1
max |q̂k+1 − q̂k |
k=0,...,N −1 6ερ

does not feature in the a posteriori estimate for the non-monotone scheme.
By Remark 6.15 this term corresponds to a discrete third-order derivative of uN .
Also note that in the analysis, no bounds on the derivative of the Green’s function
are required. ♣

6.1.4 A Compact Fourth-Order Scheme

In this section we consider a compact finite difference scheme of order four. Given
an arbitrary mesh ω̄ we seek a mesh function uN ∈ IRN +1 satisfying
' N(
Lu i := cli uN c N r N
i−1 + ci ui + ci ui+1

= qil fi−1 + qic fi + qir fi+1 =: [Qf ]i , i = 1, . . . , N − 1.

The coefficients c and q are determined so that the scheme is exact for polynomials
up to degree four. That is [Lp]i = [Q(Lp)]i for all p ∈ Π4 . This construction yields
a difference scheme whose system matrix is not inverse monotone.
A similar approach was used in [27] where in order to ensure inverse monotonic-
ity, the high-order approximation is used only when the local mesh size is small
enough to give non-positive off-diagonal entries. In all other mesh points central
differencing is used. This hybrid method is shown to be third-order convergent uni-
formly in ε on a Shishkin mesh; see [27, § 2]. An alternative approach to obtain
a higher-order difference approximation while maintaining the M -matrix property
can be found in [46]. However, the construction of that scheme requires explicit
knowledge of the derivatives of the data (c and f ) and subtle modifications in those
points where the mesh is non-uniform.
We shall follow the analysis in [98] and see that inverse monotonicity is not a
prerequisite for the maximum-norm error analysis of higher-order schemes.

6.1.4.1 Discretisation

The exactness of the scheme for polynomials up to degree four and the normalisation
condition ql;i + qc;i + qr;i = 1, i = 1, . . . , N − 1, yield the difference scheme: Find
uN ∈ IRN +1 such that
' N( ' (
Lu i := −ε2 uN N
x̄x̂,i + Q(cu ) i = [Qf ]i for i = 1, . . . , N − 1,
N N
(6.38)
u0 = γ0 , uN = γ1 ,
6.1 Reaction-Diffusion 207

where

1 − μ−
i 4 + μ− +
i + μi 1 − μ+
i
[Qv]i := vi−1 + vi + vi+1
6 6 6

with
h2i+1 h2i
μ−
i := and μ+
i := .
2hi h̄i 2hi+1 h̄i

6.1.4.2 Stability

For the stability analysis we consider an arbitrary mesh ω̄ with maximal step size h.
Although L is not inverse-monotone, it possesses a core that is:

' ( α− 4 + μ− +
i + μi α+
[Λv]i := −ε2 δx2 v i + i ri vi−1 + ri vi + i ri vi+1
6 6 6

where
! !
αi− = min 0, 1 − μ−
i , αi+ = min 0, 1 − μ+
i .

The matrix associated with Λ is an L0 matrix with row sums βi /6, where

βi := 4 + αi− + αi+ + μ− +
i + μi .

Therefore, it is an M -matrix and we can conclude that

 
 6 [Λv]i 
vω̄ ≤ max    for all v ∈ IRN +1 , (6.39)
i=1,...,N −1 ci βi  0

by Lemma 3.17.
Theorem 6.23. Suppose c ∈ C 1 [0, 1]. Let κ ∈ (0, 1) be arbitrary, but fixed. Then
 
3   Lv 

vω̄ ≤ for all v ∈ IR0N +1 ,
2 − κ c ω


provided that h is smaller than some threshold value that depends on κ and c only.
Proof. First, note that

h3i+1 + h3i
μ− +
i + μi = ≥ 1, (6.40)
hi hi+1 (hi + hi+1 )

by the relation between cubic, arithmetic and geometric means.

208 6 Discretisations of Reaction-Convection-Diffusion Problems

Furthermore,

hi+1 ≥ hi =⇒ αi+ = 0, and hi+1 ≤ hi =⇒ αi− = 0.

Therefore, at least one of αi− and αi+ is zero. Without loss of generality we assume
αi− = 0. This implies 0 ≤ 1 − μ− i ≤ 1.
We distinguish two cases αi+ = 0 and αi+ = 1 − μ+ i .

(i) Suppose 0 = αi+ . Then 0 ≤ 1 − μ+

i ≤ 1 and

1 − μ− 1 − μ+
[Λv]i = [Lv]i − i
ci vi−1 − i
ci vi+1
6 6
1 − μ− 1 − μ+
− i
(ci−1 − ci ) vi−1 − i
(ci+1 − ci ) vi+1 .
6 6
We estimate as follows:
  
 1 − μ− (ci−1 − ci ) ≤ hi c ∞ ≤ κci .
i

Similarly,
  
 1 − μ+ (ci+1 − ci ) ≤ hi+1 c ∞ ≤ κci ,
i

provided h is sufficiently small—depending on κ and c only. We also have

   
1 − μ−  + 1 − μ+  = 2 − μ− − μ+ ≤ 1, by (6.40).
i i i i

Therefore,

1 + 2κ
|[Λv]i | ≤ |[Lv]i | + vω̄ . (6.41)
6

Note that βi = 4 + μ−
i + μi ≥ 5, by (6.40). Hence,
+

   
 6 [Λv]i  6  [Lε v]i  1 + 2κ
 ≤  
 ci βi  5  ci  + 5 vω̄ . (6.42)

(ii) If αi+ = 1 − μ+
i ≤ 0 then

1 − μ− 1 − μ−
[Λv]i = [Lv]i − i
ci vi−1 − i
(ci−1 − ci ) vi−1
6 6
1 − μ+
− i
(ci+1 − ci ) vi+1 .
6
6.1 Reaction-Diffusion 209

The second and third term on the right-hand

 side
 are bounded as in case (i). For the
last term note that 1 − μ+ ≤ 0 yields 1 − μ+  ≤ μ+ and therefore,
i i i
  
 1 − μ+ (ci+1 − ci ) ≤ μ+ hi+1 c ∞ ≤ hi c ∞ ≤ κci ,
i i

for sufficiently small h. Thus, (6.41) holds for this case too. Furthermore, for βi we
have βi = 5 + μ− i ≥ 5. Consequently, (6.42) holds for all i = 1, . . . , N − 1.
Combining (6.39) with (6.42), we are finished. 


Remark 6.24. The discretisation (6.38) is (∞ , ∞ )-stable although the underlying
operator is in general not inverse monotone and therefore does not satisfy a maxi-
mum principle. ♣

Remark 6.25. The argument presented here sharpens Theorem 1 in [98] by giving a
smaller stability constant. ♣

6.1.4.3 Nodal Error Analysis

We shall apply the difference scheme (6.38) on a Shishkin mesh with σ ≥ 4. For
this we have h ≤ N −1 /(1 − 2q).
Let η = u − uN denote the error of the scheme. We start our analysis by decom-
posing the error of the scheme as η = ψ + ϕ, where the two parts ψ, ϕ ∈ IR0N +1
solve

[Λψ]i = [Lη]i = ε2 [Q(u ) − ux̄x̂ ]i on ω,

and
ci−1 ci+1
[Λϕ]i = − ηi−1 − ηi+1
6 6
−
μ μ+
+ i (ci−1 − ci ) ηi−1 + i (ci+1 − ci ) ηi+1 on ω.
6 6

Let κ ∈ (0, 1) be arbitrary, but fixed. Then using arguments from our stability anal-
ysis, we get
 
 6 [Λϕ]i  1 + 2κ
 
 βi ci  ≤ 5
ηω̄ for i = 1, . . . , N − 1,

if N is greater than some threshold value that is independent of ε. The stability

inequality (6.39) yields
1 + 2κ
ϕω ≤ ηω̄ .
5
210 6 Discretisations of Reaction-Convection-Diffusion Problems

Application of a triangle inequality gives

1 + 2κ
ηω̄ ≤ ψω̄ + ϕω̄ ≤ ψω̄ + ηω̄ .
5
Hence, the error can be bounded in terms of ψ:
  5
u − uN  ≤ ψω̄ . (6.43)
ω̄ 4 − 2κ

We are left with estimating ψ which will be done using a truncation error and
barrier function technique.
Let g ∈ C 6 [xi−1 , xi+1 ]. Then Taylor expansions give
⎧
⎪
⎪C g  [xi−1 ,xi+1 ] if hi = hi+1 ,
⎪
⎪
  ⎪ ⎨C g   − 
  [xi ,xi+1 ] + Cμi hi g [xi−1 ,xi ] if hi ≤ hi+1 ,
[Q(g  ) − gx̄x̂ ]i  ≤  
⎪C (hi + hi+1 ) g (5) [x ,x ]
3
⎪
⎪ and
⎪
⎪   i−1 i+1
⎩Ch4 g (6)  if hi = hi+1 .
i [x
i−1 ,x ]
i+1

(6.44)

We consider the two distinct cases for the mesh transition point: τ = q and τ < q.
In the first case, the mesh
 is  uniform with hi = 1/N for i = 1, . . . , N . Moreover,
ε−1 ≤ C ln N . Thus, u(6)  ≤ Cε−2 ln4 N , by Theorem 3.35. Now the fourth
bound of (6.44) yields
 
 
|[Λψ]i | = ε2 [Q(u ) − ux̄x̂ ]i  ≤ CN −4 ln4 N, i = 1, . . . , N − 1.

Hence,

ψω̄ ≤ CN −4 ln4 N if τ = q, (6.45)

by (6.39).
The case when τ < q requires a more detailed argument employing the decom-
position of u.
 
(i) For xi ∈ (0, τ ) ∪ (1 − τ, 1), use the fourth bound of (6.44), u(6)  ≤ Cε−6
and hi = hi+1 ≤ CεN −1 ln N to get
 
 
|[Λψ]i | = ε2 [Q(u ) − ux̄x̂ ]i  ≤ CN −4 ln4 N for xi ∈ (0, τ ) ∪ (1 − τ, 1).

(ii) For xi ∈ (τ, 1 − τ ), split the truncation error according to the decomposition
of u:

ε2 |[Q(u ) − ux̄x̂ ]i | = ε2 |[Q(v  ) − vx̄x̂ ]i | + ε2 |[Q(w ) − wx̄x̂ ]i | .

6.1 Reaction-Diffusion 211
 
The first term is bounded using the fourth estimate of (6.44), ε2 v (6)  ≤ C
and hi = hi+1 ≤ CN −1 . We obtain ε2 |[Q(v  ) − vx̄x̂ ]i | ≤ CN −4 . In order to
bound the truncation error with respect to the layer part w, use the first estimate
of (6.44) to get ε2 |[Q(w ) − wx̄x̂ ]i | ≤ CN −σ . Hence,

|[Λψ]i | ≤ CN −4 for xi ∈ (τ, 1 − τ ).

(iii) For xi ∈ {τ, 1 − τ }, split the truncation error again. For the regular solution
component v the third estimate of (6.44) gives

ε2 |[Q(v  ) − vx̄x̂ ]i | ≤ CεN −3 ,

while for w we have by the second bound of (6.44)

   
 
ε2 [Q(w ) − wx̄x̂ ]qN  ≤ C e−ρxqN /ε + μ−
qN N
−1
ln N e−ρ(1−x(1−q)N +1 )/ε
≤ CN −σ + CN σ/qN N −σ−1 μ−
qN ln N
≤ CN −σ + CN −σ−1 μ−
qN ln N,

with an analogous estimate for i = (1 − q)N . Collecting the various bounds

for the truncation error, we get (for τ < q)

ε2 |[Q(u ) − ux̄x̂ ]i |
⎧
⎪
⎪CεN −3 + CN −5 μ−
⎨ i ln N if i = qN,
≤ CN −4 ln4 N + CεN −3 + CN −5 μ+
i ln N if i = (1 − q)N,
⎪
⎪
⎩0 otherwise.
(6.46)

Finally, we use an idea from [122]. Define the mesh function z ∈ IR0N +1 by
⎧
⎪
⎪xi /τ for i = 0, . . . , qN,
⎨
zi := 1 for i = qN, . . . , (1 − q)N, and
⎪
⎪
⎩(1 − x )/τ for i = (1 − q)N, . . . , N.
i

A direct calculation verifies

⎧ 2
⎪ ε μ−
i ri
⎪
⎨ hi h̄i + 6 if i = qN,
2ci 1 ε2 μ+i ri
[Λz]i = zi + hi+1 h̄i + 6 if i = (1 − q)N and
3 qN ⎪
⎪
⎩
0 otherwise.
212 6 Discretisations of Reaction-Convection-Diffusion Problems

Note that

ε2 ε2 ερ(1 − 2q)N
, ≥ .
hqN h̄qN h(1−q)N +1 h̄(1−q)N 2σ ln N

Therefore, recalling the truncation error bound (6.46), we can

 use the discrete
comparison principle (6.13) with the barrier function CN −4 ln4 N + zi ln N
to see that (6.45) holds for τ < q too.
Finally, (6.43) yields the main convergence result of this section.
Theorem 6.26. Let u be the solution of the boundary value problem (6.1) and uN
that of the finite difference scheme (6.38) on a Shishkin mesh with σ ≥ 4. Then
 
u − uN  ≤ CN −4 ln4 N,
∞,ω̄

provided that N is larger than some threshold value that depends on c only.

Corollary 6.27. Clustering three adjacent and equidistant mesh intervals and fit-
ting a cubic function through the numerical approximation on the four associated
mesh points, we obtain a cubic C 0 -spline SuN that approximates u on the whole
domain:
 
u − SuN  ≤ CN −4 ln4 N.
∞

Remark 6.28. Approximations of the derivatives can be obtained by differentiat-

ing SuN :
   (4−k)
εk (u − SuN )(k)  ≤ C N −1 ln N for k = 1, 2, 3.

Note that in those mesh points where two different cubic functions are concatenated
to give SuN , we have different left- and right-sided derivatives, however for both,
the above bound holds.
Better approximations of the second-order derivative are obtained by appealing
to (6.1):
 
u (x) ≈ m(x) := ε−2 cSuN − f (x).

We have the error bound

ε2 u − m ≤ CN −4 ln4 N.

This readily follows from Theorem 6.26. ♣

6.1 Reaction-Diffusion 213

6.1.4.4 Numerical Results

To illustrate the theoretical results, we consider the problem

−ε2 u + (1 + x2 + cos x)u = x9/2 + sin x, in (0, 1), (6.47)

with homogeneous Dirichlet boundary conditions.

The exact solution is not available. Therefore, we estimate the error for uN and
SuN by comparing them to the numerical solution ũN obtained on a mesh ω̃ that is
seven times as fine. More precisely ω̃ is constructed from ω by uniformly dividing
each of its mesh intervals into seven equidistant subintervals. We take the estimates
   
u − uN  ≈ η N := ũN − uN ∞,ω
∞,ω

and
   
u − SuN  ≈ η̃ N := ũN − SuN  .
∞ ∞,ω̃

 p
Since we have an error bound of the form C N −1 ln N , we also compute the
“Shishkin” rates of convergence:

ln η N − ln η 2N
pN = .
ln(2 ln N ) − ln(ln(2N ))

We choose N divisible by three and define SuN on macro intervals [x3k , x3(k+1) ],
k = 0, . . . , N/3 − 1.
The left half of Table 6.1 displays the results of our test computations for the
Shishkin mesh. They are in good agreement with Theorem 6.26 and Corollary 6.27.
The right half of the table contains results for a modified Bakhvalov mesh: First

Table 6.1 Compact fourth order scheme for (6.47), ε = 10−4 (identical numbers for ε =
10−4k , k = 2, . . . , 6)
Shishkin mesh Bakhvalov mesh
N ηN pN η̃ N p̃N ηN πN η̃ N π̃ N
3·2 7
1.151e-05 3.99 3.979e-04 3.66 2.644e-08 4.01 4.915e-07 3.99
3 · 28 1.123e-06 4.00 4.701e-05 3.81 1.641e-09 4.00 3.090e-08 4.00
3 · 29 1.045e-07 4.00 4.890e-06 3.89 1.024e-10 4.00 1.936e-09 4.00
3 · 210 9.375e-09 4.00 4.672e-07 3.94 6.394e-12 4.00 1.212e-10 4.00
3 · 211 8.160e-10 4.00 4.212e-08 3.97 3.996e-13 4.00 7.580e-12 4.00
3 · 212 6.925e-11 4.00 3.644e-09 3.98 2.497e-14 4.00 4.739e-13 4.00
3 · 213 5.750e-12 4.00 3.058e-10 3.99 1.561e-15 4.00 2.963e-14 4.00
3 · 214 4.686e-13 4.00 2.506e-11 4.00 9.756e-17 4.00 1.852e-15 4.00
3 · 215 3.756e-14 4.00 2.014e-12 4.00 6.097e-18 4.00 1.157e-16 4.00
3 · 216 2.967e-15 — 1.594e-13 — 3.811e-19 — 7.234e-18 —
214 6 Discretisations of Reaction-Convection-Diffusion Problems

we construct a standard Bakhvalov mesh with N/3 mesh intervals, which are then
subdivided into three subintervals of equal length. This gives our computational
mesh ω. This modification is necessary because the stability constant of S depends
on the local ratio of the mesh sizes, which on a Bakhvalov mesh depends on ε. For
this mesh, we expect uniform convergence of order N −4 . This is clearly observed
in the numerical experiments.

6.2 Systems of Reaction-Diffusion Type

We now leave the scalar problems and move on to systems of reaction-diffusion

 
equations: Find u ∈ C 2 (0, 1) ∩ C[0, 1] such that

Lu := −E 2 u + Au = f in (0, 1), u(0) = u(1) = 0, (6.48)

where E = diag(ε1 , . . . , ε ) and the small parameters εk are in (0, 1].

The analytical behaviour of the solution to (6.48) has been studied in Sect. 3.3.2.
Again we assume the coupling matrix A has positive diagonal entries and is diago-
nally dominant with

 
 akm 
β := max  
k=1,...,  akk  < 1. (6.49)
m=1 ∞
m=k

Define κ > 0 by

κ2 := (1 − β) min min akk (x).

k=1,..., x∈[0,1]

For simplicity in our presentation we assume that

κ
ε1 ≥ ε2 ≥ · · · ≥ ε and ε1 ≤ .
4
The first chain of inequalities can always be achieved by renumbering the equations,
while the last inequality provides a threshold value for the validity of our analysis.

6.2.1 The Interpolation Error

We consider piecewise linear interpolation. Using the derivative bounds of Theorem

3.43 for the solution of (6.48), we can apply the technique in Section 6.1.1.1 to
establish interpolation error bounds for the components of u.
6.2 Systems of Reaction-Diffusion Type 215

Theorem 6.29. Suppose the assumptions of Theorem 3.43 hold true. Then
     2
uk − uIk  ≤ uk − uIk  ≤ C ϑ[2] (ω̄) , k = 1, . . . , ,
0 ∞ rd,k

and
 
εk uk − uIk 1 ≤ Cεk ϑrd,k (ω̄),
1/2 [2]
k = 1, . . . , ,

where

  
k  

[p] −1 −κs/pεm −κ(1−s)/pεm
ϑrd,k (ω̄) := max 1+ εm e +e ds
i=0,...,N −1 Ii m=1

for k = 1, . . . , .

6.2.2 Linear Finite Elements

In view of Lemma 3.40 we may assume without loss of generality that A is coercive,
i.e., there exists a positive constant α such that

v T A(x)v ≥ αv T v for all x ∈ [0, 1], and v ∈ IR . (6.50)

As usual with finite element discretisations, we consider the weak formulation

of (6.48): Find u ∈ H01 (0, 1) such that

B(u, v) = F (v) for all v ∈ H01 (0, 1) ,

with
 
B(u, v) := ε2m (um , vm

)+ (amk uk , vm )
m=1 m=1 k=1

and
  1 
F (v) := (fm , vm ), (w, v) = wv (s)ds.
m=1 0

The natural norm on H01 (0, 1) associated with the bilinear form B(·, ·) is the
energy norm |||·|||ε2 defined by

2
 2 2 2
 2
|||v|||ε2 := ε2m |vm |1 + α v0 , v0 := vm 0 .
m=1 m=1
216 6 Discretisations of Reaction-Convection-Diffusion Problems

Because of (6.50) the bilinear form B(·, ·) is coercive with respect to the energy
norm, i.e.,
2
|||v|||ε2 ≤ B(v, v) for all v ∈ H01 (0, 1) .

If f ∈ L2 (0, 1) then F is a linear continuous functional on H01 (0, 1) . Therefore,

the Lax-Milgram Lemma ensures the existence of a unique solution u ∈ H01 (0, 1)
of the variational formulation of (6.48).
Given a mesh ω̄, we seek an approximation uN ∈ (V0ω ) such that

B̂(uN , v) = F̂ (v) for all v ∈ (V0ω ) , (6.51)

where
 
B̂(u, v) := ε2m (um , vm

)+ (aImk uk , vm )
m=1 m=1 k=1

and

I
F̂ (v) := (fm , vm ).
m=1

Thus, our FEM incorporates quadrature to approximate the integrals.

The coercivity of A, i.e. (6.50), implies the coercivity of its piecewise linear
interpolant AI :

v T AI (x)v ≥ αv T v for all x ∈ [0, 1], and v ∈ IR .

Consequently, the bilinear form B̂(·, ·) is also coercive with

2
|||v|||ε2 ≤ B̂(v, v) for all v ∈ H01 (0, 1) . (6.52)

Set η := uI − u and χ := uI − uN . Then by a triangle inequality

 
u − uN  ≤ |||η|||ε2 + |||χ|||ε2 .
ε2

Theorem 6.29 yields

 
1/2 [2] [2]
|||η|||ε2 ≤ C ε1 + ϑrd, (ω) ϑrd, (ω),

and we are left with bounding |||χ|||ε2 .

Starting from (6.52), we get
 
2
  
|||χ|||ε2 ≤ B̂(χ, χ) = amk uk − amk uk + fm − fm , χm .
I I I

m=1 k=1
6.2 Systems of Reaction-Diffusion Type 217

Use
   
aImk uIk − amk uk = aImk − amk uIk + amk uIk − uk ,

the Cauchy-Schwarz inequality and Theorem 6.29 to get

 
2 1/2 [2] [2]
|||χ|||ε2 ≤ C ε1 + ϑrd, (ω) ϑrd, (ω) χ0 .

We obtain the following convergence results in the energy norm.

Theorem 6.30. Let u be the solution of (6.48) and uN its approximation by (6.51).
Then
 I   2
u − uN  2 ≤ C ϑ[2] (ω̄) ,
ε rd,

and
   
u − uN  1/2 [2] [2]
≤ C ε1 + ϑrd, (ω̄) ϑrd, (ω̄).
ε2

Remark 6.31. A similar result is given in [99], but there the effect of numerical
integration is not taken into account. ♣
Remark 6.32. Like in the scalar case, the energy norm |||·|||ε2 fails to capture the
layers present in the solution. ♣

6.2.3 Central Differencing

We consider the discretisation of (6.48) by standard central differencing on meshes

 
ω̄ that for the moment are arbitrary. That is, we seek uN ∈ IR0N +1 such that
' (
LuN i
:= − diag(E)2 uN N
x̄x̂;i + Ai ui = f i for i = 1, . . . , N − 1,
(6.53)
uN
0 = uN
N = 0.

6.2.3.1 Stability

Our analysis follows that of [104] and is based on the stability properties of
Section 6.1.2.1 for scalar operators.
 
Let v ∈ IR0N +1 be arbitrary. Then

−ε2k vk;x̄x̂ + akk vk = (Lv)k − akm vm for k = 1, . . . , .
m=1
m=k
218 6 Discretisations of Reaction-Convection-Diffusion Problems

The stability inequality (6.14) and a triangle inequality then yield

   
  akm   (Lv)k 
vk ∞,ω̄ −   um  ≤   .
 akk  ∞,ω  akk 
m=1 ∞ ∞,ω
m=k

Define the  ×  constant matrix Γ = Γ (A) = (γkm ) by

 
 akm 
γkk = 1 and γkm = −    for k = m.
akk ∞

Suppose that Γ is inverse-monotone, i.e., that Γ is invertible and Γ −1 ≥ 0. Then

we obtain the following stability result for the difference operator L.
Theorem 6.33. Assume the matrix A has positive diagonal entries. Assume that
Γ (A) is inverse-monotone. Then for k = 1, . . . ,  one has
 
   (Lv)m 
vk ∞,ω̄ ≤ Γ −1  
km  a 
m=1 mm ∞,ω

 
for any function mesh function v ∈ IR0N +1 .
Corollary 6.34. Under the hypotheses
  of Theorem 6.33 the discrete problem (6.53)
has a unique solution uN , and uN ∞,ω̄ ≤ C f ∞,ω for some constant C.
Thus, the operator L is (∞ , ∞ ) stable, or maximum-norm stable although in
general it is not inverse-monotone.

6.2.3.2 A Priori Error Bounds

Let η := u − uN denote the error of the discrete solution. We decompose the solu-
tion error as η = ϕ + ψ, where the components ϕk and ψk of ϕ and ψ respectively
are the solutions of
 
−ε2k ϕk;x̄x̂ + akk ϕk = −ε2k uk,x̄x̂ − uk on ω, ϕk;0 = ϕk;N = 0

and

−ε2i ψk;x̄x̂ + akk ψk = − akm ηm on ω, ψk;0 = ψk;N = 0.
m=1
m=k

Assume that the matrix Γ (A) is inverse-monotone. Then for each k one has
 
  akm 
ηi ∞,ω̄ ≤ ϕi ∞,ω̄ + ψi ∞,ω̄ ≤ ϕi ∞,ω̄ +  
 akk  ηm ∞,ω̄ ,
m=1 ∞
m=k
6.2 Systems of Reaction-Diffusion Type 219

by (6.14). Gathering together the η terms and invoking the inverse-monotonicity of

Γ (A), we get
 
u − uN  ≤ C ϕ∞,ω̄ .
∞,ω̄

Each component ϕi of ϕ is the solution of a scalar problem and can be anal-

ysed using the technique in Sect. 6.1.2.2 combined with the derivative bounds of
Theorem 3.43. This was done in [104, § 3.2] to deduce the following result:
Theorem 6.35. Let the matrix A and the source term f be twice continuously dif-
ferentiable. Assume A possesses positive diagonal entries and satisfies (6.49). Then
the error in the solution of (6.53) satisfies
   2
u − uN  ≤ C ϑ
[2]
(ω̄) .
∞,ω̄ rd,

Corollary 6.36. Identifying uN with a piecewise linear function on ω̄, we have

   2
u − uN  ≤ C ϑ[2] (ω̄) ,
∞ rd,

by Theorems 6.29 and 6.35.

Remark 6.37. An alternative analysis based on comparison principles with special

barrier functions was used in [102, 118]. This technique has the more restrictive
condition that A be an M -matrix and up to now has been applied successfully only
to Shishkin meshes. ♣

6.2.3.3 Numerical Results

We now present the results of some numerical experiments in order to illustrate

the conclusions of the preceding section, and to check if the error estimates of
Theorem 6.35 are sharp.
The test problem is

−ε21 u1 + 3u1 + (1 − x)(u2 − u3 ) = ex , u1 (0) = u1 (1) = 0,

−ε22 u2 + 2u1 + (4 + x)u2 − u3 = cos x, u2 (0) = u2 (1) = 0,
−ε23 u3 + 2u1 + 3u3 = 1 + x2 , u3 (0) = u3 (1) = 0.

In the construction of the Bakhvalov and the Shishkin mesh (see Sect. 2.2.2) we take
κ/p = 0.8.
The exact solutions to the test problems is not available, so we estimate the ac-
curacy of the numerical solution by comparing it to the numerical solution of the
Richardson extrapolation method, which is of higher order: Let uN ε be the solution
220 6 Discretisations of Reaction-Convection-Diffusion Problems

of the difference scheme on the original mesh and ũ2Nε that on the mesh obtained
by uniformly bisecting the original mesh. This yields the estimated error

4
uN − ũ2N


ηεN := ∞,ω
.
3 ε ε

The uniform errors for a fixed N are estimated by taking the maximum error over a
wide range of ε, namely

η N := max ηεN .
ε1 ,...,ε =10−0 ,...,10−24

 p
For the Shishkin mesh we have an error bound of the form C N −1 ln N . There-
fore, we compute the “Shishkin” rates of convergence using the formula

ln η N − ln η 2N
pN = ,
ln(2 ln N ) − ln(ln(2N ))

while for the Bakhvalov mesh, the standard formula

 
ρN = ln η N − ln η 2N / ln 2

is used.
The results of our test computations displayed in Table 6.2 are in agreement with
Theorem 6.35. The Bakhvalov mesh gives second order accuracy, while the results
for the Shishkin mesh are spoiled by the typical logarithmic factor.

Table 6.2 Central differencing on layer-adapted

meshes for systems of reaction-diffusion type
Shishkin mesh Bakhvalov mesh
N ηN pN ηN ρN
8·2 3
4.895e-02 1.22 4.910e-03 2.08
8 · 24 2.276e-02 1.50 1.157e-03 2.05
8 · 25 8.854e-03 1.67 2.790e-04 2.04
8 · 26 3.091e-03 1.77 6.771e-05 2.01
8 · 27 1.014e-03 1.83 1.676e-05 2.00
8 · 28 3.201e-04 1.88 4.179e-06 2.00
8 · 29 9.831e-05 1.91 1.044e-06 2.00
8 · 210 2.955e-05 1.94 2.610e-07 2.00
8 · 211 8.725e-06 1.96 6.527e-08 2.00
8 · 212 2.537e-06 — 1.632e-08 —
6.3 Reaction-Convection-Diffusion 221

6.2.3.4 A Posteriori Error Bounds

By interpreting the components of uN as piecewise linear functions, one can con-

duct an a posteriori error analysis that combines the technique in Sect. 6.1.2.3 with
Theorem 3.36 to give

   
uk − uN 
k ∞ ≤ Γ −1 km (η1;m + η2;m + η3;m ) , k = 1, . . . , ,
m=1

where
 
I 
q̂m − q̂m h2i+1
∞
η1;m := , η2;m := max max {|qk;i | , |qk;i+1 |}
ρ2m i=0,...,N −1 4ε2 k
hi+1
η3;m := max |q̂m;i+1 − q̂m;i |
i=0,...,N −1 6εm ρm


q̂m = fm − amν uN ν and ρm = min amm (x)1/2 .
x∈[0,1]
ν=1

Remark 6.38. Since the a posteriori analysis uses the stability of the differential
operator, a posteriori bounds can also be derived for non-monotone discretisations,
like schemes generated by FEMs. If for example, FEM-2 (see p. 187) is used, then
the η3 terms in the above estimate will disappear; cf. Sect. 6.1.3.3. ♣

6.3 Reaction-Convection-Diffusion

Consider the scalar reaction-convection-diffusion problem

Lu := −εd u − εc bu + cu = f in (0, 1), u(0) = γ0 , u(1) = γ1 , (6.54)

where b ≥ 1 and c ≥ 1 on [0, 1], εd ∈ (0, 1] and εc ∈ [0, 1].

Analytical properties of (6.54) were studied in Sect. 3.2, while layer-adapted
meshes for it have been introduced in Sect. 2.2.
We briefly recall some of the notation. Let λ0 < 0 and λ1 > 0 be the solution of

−εd λ(x)2 − εc b(x)λ(x) + c(x) = 0. (6.55)

Set

μ0 := max λ0 (x) < 0 and μ1 := min λ1 (x) > 0.

x∈[0,1] x∈[0,1]
222 6 Discretisations of Reaction-Convection-Diffusion Problems

The quantity used for presenting a priori error bounds is


1 + |μ0 | eμ0 s/pε + μ1 e−μ1 (1−s)/pε ds.
[p]
ϑrd (ω̄) := max
i=1,...,N Ii

6.3.1 The Interpolation Error

Let wI denote the piecewise linear interpolant of w again. Adapting the techniques
from Sects. 5.1 and 6.1.1.1, and using the derivative bounds of Theorem 3.29 we
obtain the following result.
Proposition 6.39. Let u be the solution of (6.54). Let p > 2 be arbitrary, but fixed.
Then
   2
 
u − uI  ≤C 1 + |μ0 | eμ0 s/pε + μ1 e−μ1 (1−s)/pε ds
∞,Ii
Ii

for all mesh intervals Ii = [xi−1 , xi ].

Define the εd -weighted energy norm by
2 2 2
|||v|||εd := εd |v|1 + v0 .

Theorem 6.40. Let u be the solution of (6.54). Let p > 2 be arbitrary, but fixed.
Then
   2
u − uI  ≤ C ϑ[p] (ω̄)
∞ rcd

and
   
u − uI  −1/2
≤ C μ1
[p] [p]
+ ϑrcd (ω̄) ϑrcd (ω̄).
εd

for any p > 2.

Proof. The bound on the L∞ error is an immediate consequence of Prop. 6.39 and
[p]
the definition of ϑrcd .
For the error in the H 1 semi-norm use integration by parts to get
 1
   
u − u I 2 = − u (x) u − uI (x)dx.
1
0

Note, that
 1
εd |u (x)| dx ≤ Cεd |μ0 | ≤ Cμ−1
1 .
0

The assertion of the theorem follows. 


6.3 Reaction-Convection-Diffusion 223

Remark 6.41. The results of Theorem 6.40 hold for arbitrary p > 2, but not for
p = 2. This is because of the derivative bounds provided by Theorem 3.29.

6.3.2 Simple Upwinding

This section studies a simple first-order upwind difference scheme for (6.54). The
analysis essentially follows [95].
Our scheme is: Find uN ∈ IRN +1 such that
' N(
Lu i := −εd uN
x̄x;i − εc bi ux;i + ci ui = fi
N N
for i = 1, . . . , N − 1,
(6.56)
uN
0 = γ0 , uN
N = γ1 .

The difference operators were introduced in Sect. 4.1.

The variational formulation of (6.56) is: Find uN ∈ V ω with uN
0 = γ0 and
uN
N = γ 1 such that
   
au uN , v = fu v := (f, v)ω for all v ∈ V0ω , (6.57)

where

au (w, v) := εd [wx , vx )ω − (εc bux − cw, v)ω

and


N −1 
N −1
[w, v)ω := hi+1 wi vi , (w, v)ω := hi+1 wi vi .
i=0 i=1

Taking as test functions v the standard hat-function basis in V0ω , we see

that (6.56) and (6.57) are equivalent. In particular, using summation by parts it
is verified that

au (w, v) = (Lw, v)ω = (w, L∗ v)ω for all w, v ∈ V0ω ,

where
' ∗ (
L v i := −εd vx̄x;i + εc (bv)x̆;i + ci vi .

is the adjoint operator to L with respect to the scalar product (·, ·)ω .
224 6 Discretisations of Reaction-Convection-Diffusion Problems

6.3.2.1 Stability

The matrix associated with the difference operator L is an L0 -matrix. Application

of Lemma 3.14 with a constant positive test function establishes the inverse mono-
tonicity of L. Thus, L satisfies a comparison principle. This comparison principle
gives the (∞ , ∞ )-stability inequality

v∞,ω̄ ≤ Lv/c∞,ω for all v ∈ IR0N +1

and provides a priori bounds for the solution of (6.56):

 N !
u  ≤ max |γ0 |, |γ1 | + f /c for i = 0, . . . , N.
ω̄ ∞,ω

Green’s function estimates

Using the discrete Green’s function G : ω̄ 2 → IR : (xi , ξj ) → Gi,j = G(xi , ξj )
associated with L and Dirichlet boundary conditions, any mesh function v ∈ IR0N +1
can be represented as

vi = au (v, Gi,· ) = (Lv, Gi,· )ω = (v, L∗ Gi,· )ω for i = 1, . . . , N − 1.

Taking for v the standard hat basis in V0ω , we see that for fixed i = 1, . . . , N − 1

[L∗ Gi,· ]j = δi,j for j = 1, . . . , N − 1, Gi,0 = Gi,N = 0, (6.58)

where

h−1
i+1 if i = j,
δi,j :=
0 otherwise

is the discrete equivalent of the Dirac-δ distribution.

Theorem 6.42. Assume that b ≥ 1 and c ≥ 1 on [0, 1], then there exists a positive
constant C such that

0 ≤ Gi,j ≤ Kμ1 for i, j = 0, . . . , N,

where
 !
b∞ c∞ max b∞ , c∞ if εc > 0,
K := 3/2
c∞ if εc = 0.

Furthermore, if b ≥ 0 on [0, 1] then

Gξ;i,j ≥ 0 for 0 ≤ j < i < N,

Gξ;i,j ≤ 0 for 0 ≤ i ≤ j < N.
6.3 Reaction-Convection-Diffusion 225

Proof. Let j be fixed. Set

⎧ j
⎪
⎪ 3
⎪
⎪ (1 + μ1 hk )
−1
for 0 ≤ i < j ≤ N,
⎪
⎪
⎪
⎨k=i+1
Γi := 1 for 0 ≤ i = j ≤ N,
⎪
⎪
⎪
⎪ 3i
⎪
⎪ (1 − μ0 hk )
−1
for 0 ≤ j < i ≤ N.
⎪
⎩
k=j+1

We shall use the discrete comparison principle to show that Kμ1 Γ is an upper bound
for G·,j .
Clearly Γ0 ≥ 0 and ΓN ≥ 0.
Next, a direct calculation gives
⎧
⎨μ1 for 0 ≤ i < j ≤ N,
Γx;i = Γi μ0
⎩ for 0 ≤ j ≤ i < N
1 − μ0 hi+1

and
⎧ μ1
⎨ for 0 < i ≤ j ≤ N,
Γx̄;i = Γi 1 + μ1 hi+1
⎩μ for 0 ≤ j < i ≤ N.
0

For i < j we have

 
' ( −εd μ21  
LΓ i = Γi − εc bi μ1 + ci ≥ Γi −εd μ21 − εc bi μ1 + ci ≥ 0,
1 + μ1 hi+1

by Proposition 3.19.
For i > j
 
' ( −εd μ20 − εc bi μ0 −εd μ20 − εc bi μ0 + ci
LΓ i = Γi + ci ≥ Γi ≥ 0,
1 − μ0 hi+1 1 − μ0 hi+1

where Proposition 3.19 was used again.

For i = j
 
' ( εd μ0 μ1 μ0
LΓ j = − − − ε c bj + cj
hj+1 1 − μ0 hj+1 1 + μ1 hj 1 − μ0 hj+1
−εd μ0 − (εc bj μ0 − cj ) hj+1
≥ .
hj+1 (1 − μ0 hj+1 )
226 6 Discretisations of Reaction-Convection-Diffusion Problems

The function x → λ0 (x) is continuous. Therefore there exists a ξ ∈ [0, 1] such that
λ0 (ξ) = μ0 . Furthermore, recall b ≥ 1, c ≥ 1 and μ0 < 0. Hence,
 
' ( 1 −εd μ0 − εc b(ξ)μ0 − c(ξ) hj+1
LΓ j ≥
max {bj , cj } hj+1 (1 − μ0 hj+1 )
−εd μ0 c(ξ)
= = , by (6.55)
hj+1 max {bj , cj } hj+1 max {bj , cj } λ1 (ξ)
1
≥ , by (3.6).
hj+1 Kμ1

Thus, Kμ1 Γ is a barrier function for G·,j , for all j.

In order to verify the monotonicity of Gi,· use the argument from the proof of
Theorem 4.1. 


Theorem 6.43. Assume that b ≥ 1 and c ≥ 1 on [0, 1], then there exists a positive
constant C such that

cGi,· 1,ω ≤ 1 for i = 0, . . . , N,

Furthermore, if b ≥ 0 on [0, 1] then

Gξ;i,· 1,ω ≤ Cμ1 for i = 0, . . . , N.

Proof. To verify the bound on the 1 -norm of G multiply (6.58) by hj+1 and sum
for j = 1, . . . , N − 1. Then use the positivity of G.
For the bound on Gξ use the piecewise monotonicity of Gξ to get

Gξ;i,· 1,ω ≤ 2Gi,i ≤ Cμ1 ,

by Theorem 6.42. 


6.3.2.2 A Priori Error Analysis

Theorem 6.44. Let the assumptions of Theorems 3.29 and 6.43 be satisfied. Let u
be the solution of (6.54) and uN its approximation by (6.56). Then
 N 
u − u  ≤ Cϑ[p] (ω̄)
∞ rcd

for any p > 1.

6.3 Reaction-Convection-Diffusion 227

Proof. Let Γ := Gi,· ∈ V0ω be the Green’s function associated with L and the mesh
node xi . Then
 N 
i − ui = au u − u, Γ = fu (Γ ) − au (u, Γ )
uN
= fu (Γ ) − f (Γ ) + a(u, Γ ) − au (u, Γ ), by (6.54)

= (εc bux − cu + f, Γ )ω − (εc bu − cu + f, Γ ) .

We obtain the representation for the nodal error:

 −1 
N
' (
i − u i = εc
uN u (s) (bΓ )(s) − bi Γi ds
i=0 Ii+1


N −1    (6.59)
+ (cu − f )i Γi − (cu − f )(s)Γ (s) ds
i=0 Ii+1

=: T1 + T0 .

To bound the first term we proceed as follows:

|(bΓ )(s) − bi Γi | ≤ b∞ |Γi+1 − Γi | + b ∞ hi+1 Γi ,

because Γ is piecewise linear. Thus,

|T1 | ≤ Cεc ϑrcd (ω̄) b∞ Γx 1,ω + b ∞ Γ 1,ω

[p]
(6.60)

for any p > 1, by Theorem 3.29.

Next we estimate T0 .

 −1 
N
' (
T0 = (cu − f )i − (cu − f )(s) Γi ds
i=0 Ii+1

 −1 
N
' (
+ (cu − f )(s) Γi − Γ (s) ds.
i=0 Ii+1

Hence,

 
|T0 | ≤ Γ 1,ω max (f − cu) (s)ds
i=1,...,N Ii

 
+ Γx 1,ω max εd u (s) + εc b(s)u (s)ds ,
i=1,...,N Ii
228 6 Discretisations of Reaction-Convection-Diffusion Problems

where (6.54) was used. To bound u and u apply Theorem 3.29. Then, for any
p > 1,
 
|T0 | ≤ Cϑrcd (ω̄) Γ 1,ω + μ−1
[p]
1 Γx 1,ω ,

where we have used that εd |μ0 | + εd |μ1 | ≤ Cμ−1

1 and εc ≤ Cμ1 .
−1

Finally, combining the last inequality with (6.59) and (6.60), we get
 N   
ui − ui  ≤ Cϑ[p] (ω̄) Γ  + μ−1 Γx 
rcd 1,ω 1 1,ω .

Application of Theorem 6.43 completes the analysis. 



Remark 6.45. The truncation error and barrier function technique (see Sect. 4.2.6)
was used in [80, 105] when studying the difference scheme

−εd uN
x̄x̂;i − εc bi ux;i + ci ui = fi
N N
(6.61)

on a Shishkin mesh. The authors establish the typical rate of N −1 ln N if the critical
mesh parameter satisfies σ > 2. In [80, 105] hybrid difference schemes were also
studied that raise the order of convergence to almost second order. ♣

Remark 6.46. Gracia et al. [47] combine (6.61) with the mid-point upwind scheme
 N 
−εd uN
x̄x̂;i − εc bi+1/2 ux;i + ci+1/2 ui + ui+1 /2 = fi+1/2
N N

and central differencing

x̄x̂;i − εc bi ux̊;i + ci ui = fi
−εd uN N N

in order to obtain an inverse monotone method that is second-order consistent in the

maximum norm for all values of εd , εc and N .
In [47] that scheme is shown to be uniformly convergent on a Shishkin mesh with
 N 
u − u  ω ≤ CN −2 ln3 N.

The analysis in [47], using solution decompositions and distinguishing different pa-
rameter regimes, is very tedious. Results for general meshes are not available.
Surla et al. [158] design an inverse monotone spline difference scheme. They
prove uniform convergence on a Shishkin mesh with
 N 
u − u  ω ≤ CN −2 ln2 N,

if εd , εc ≤ CN −1 . ♣
6.3 Reaction-Convection-Diffusion 229

6.3.2.3 A Numerical Example

In this section a brief illustration for the a priori error bounds of Theorem 6.44 is
shown. The test problem is:

−εd u + εc u + u = cos πx in (0, 1), u(0) = u(1) = 0.

Its solution is easily computed using, e. g., MAPLE.

Indicating by uN εd ,εc that the numerical approximation depends on N , εd and εc ,
and by uεd ,εc that the exact solution depends on εd and εc , we estimate the uniform
error by
 N 
η N := max uε ,ε − uIε ,ε  .
d c d c ∞
εd =1,10 ,...,10−12
−1

εc =1,10−1 ,...,10−12 ,0

The rates% of convergence are computed using the standard formula rN =

log2 η N η 2N . We also compute the constant in the error estimate, i. e., for a
given theoretical error bound η N ≤ Cϑ(N ), we approximate the constant in the
error estimate by C N = η N /ϑ(N ).
Table 6.3 displays the results of our test computations. The numbers confirm the
results of Theorem 6.44 for both the Shishkin and the Bakhvalov mesh.

6.3.2.4 A Posteriori Error Analysis

The analysis for all schemes starts from the Green’s-function representation (3.1)
and utilises Theorem 3.23, when bounds on the norms of the Green’s function G are
required.

Table 6.3 Uniform nodal errors of the upwind difference scheme ap-
plied to a model reaction-convection-diffusion problem
Shishkin mesh Bakhvalov mesh
N ηN rN CN ηN rN CN
29 1.227e-2 0.83 1.007 3.404e-3 0.99 1.743
210 6.905e-3 0.85 1.020 1.713e-3 1.00 1.754
211 3.824e-3 0.87 1.027 8.592e-4 1.00 1.760
212 2.093e-3 0.88 1.031 4.302e-4 1.00 1.762
213 1.136e-3 0.89 1.032 2.153e-4 1.00 1.763
214 6.119e-4 0.90 1.033 1.077e-4 1.00 1.764
215 3.279e-4 0.91 1.033 5.384e-5 1.00 1.764
216 1.749e-4 0.91 1.033 2.692e-5 1.00 1.764
217 9.290e-5 0.92 1.033 1.346e-5 1.00 1.764
218 4.922e-5 — 1.034 6.731e-6 — 1.764
230 6 Discretisations of Reaction-Convection-Diffusion Problems

Let
 
q := f − cuN + εc b uN .

Clearly q may have discontinuities at the mesh points because uN ∈ V ω . Therefore,

set

qi+ = lim q(x) and qi− = lim q(x)

x→xi +0 x→xi −0

for all mesh nodes xi .

Fix x ∈ (0, 1) and set Γ := G(x, ·). Then (3.1) and (6.54) yield
     
u − uN (x) = a u − uN , Γ = f (Γ ) − a uN , Γ
 
= (f − fu ) (Γ ) + (au − a) uN , Γ
N *  +
 
= +
qi−1 Γi−1 hi − qΓ (s)ds .
i=1 Ii

We see the approximation error is bounded by the error of the one-sided rectangle
rule used for integrating qΓ . A Taylor expansion gives
  
 +   
hi q Γi−1 − bΓ (ξ)dξ 
 i−1
Ii
  
 
≤ hi |q (s)|Γ (s)ds + |q(s)| |Γ (s)|ds
I Ii
 i   
 
≤ hi q /c∞,Ii c(s)Γ (s)ds + q∞,Ii |Γ (s)|ds .
Ii Ii

Using the bounds on the Green’s function from Theorem 3.23, we arrive at the main
result of this section.
Theorem 6.47. Suppose (3.8) is satisfied, then the error of the first-order upwind
scheme (6.56) satisfies
 
u − uN  ≤ η1u + η2u ,
∞

where ηku := max ηk,i

u
and
i=1,...,N

2hi
u
η1,i := hi q  /c∞,Ii , u
η2,i := q∞,Ii .
εd (μ1 − μ0 )
6.3 Reaction-Convection-Diffusion 231

Remark 6.48. The error has been bounded in terms of the numerical solution uN
and of the data of the problem. The two parts of the error bound can be further
expanded. For example:
        
q   f  − c uN   εc b   N 
hi 
c
 ≤ hi 


 + 
 c − 1
ui+1 − uN
i
 . (6.62)
∞,Ii c ∞,Ii ∞,Ii

Apparently, sampling of the data is inevitable. However, instead of sampling (6.62)

it seems advisable to sample the ηiu :
 
 q− − q+   
 i i−1  2hi qi−1/2 
u
η1,i ≈ u
η̃1,i :=   and u
η2,i ≈ η̃2,i
u
:=
 ci−1/2  εd (μ1 − μ0 )

This avoids the use of a triangle inequality and therefore gives in general sharper
upper bounds for the error. ♣

Remark 6.49. The leading term in the estimate of Theorem 6.47 is a discrete first-
order derivative. Therefore, the de Boor algorithm of Sect. 4.2.4.2 equidistributing
the arc-length of the numerical solution can be used for adaptive mesh gener-
ation, when the simple upwind scheme (6.56) is applied to the boundary-value
problem (6.54). ♣
 Part II
Two Dimensional Problems
Chapter 7
The Analytical Behaviour of Solutions

In this chapter we gather a number of analytical properties for singularly perturbed

elliptic boundary-value problems of the general type

−εΔu − b∇u + cu = f in Ω ⊂ IR2 , u|∂Ω = g,

with a small positive parameter ε, the convection field b = (b1 , b2 )T and functions
b1 , b2 , c, f : Ω̄ → IR, and g : ∂Ω → IR.

7.1 Preliminaries

The function spaces required here are the spaces of Hölder continuous functions.
Let D ∈ IR2 be a convex domain and let α ∈ (0, 1]. For nonnegative integer k,
we use C k (D) to denote the space of functions whose derivatives up to order k are
continuous on D; when k = 0 we write simply C(D). We put the usual supremum
(semi-)norms on C k (D):
 
|v|k,D := max sup ∂xi ∂yj v(x) and vk,D := max |v|l,D
i+j=k x∈D 0≤l≤k

For convenience we drop D from the notation when D = Ω. We denote by C 0,α (D)
the space of all functions that are Hölder continuous of degree α on D. Let

|v(x) − v(x )|
v0,α,D := sup ,
x=x x − x α
x,x ∈D

where  ·  is the Euclidean norm in IR2 . Then the norm in C 0,α (D) is

v0,α,D := |v|0,D + v0,α,D .

T. Linß, Layer-Adapted Meshes for Reaction-Convection-Diffusion Problems, Lecture 235

Notes in Mathematics 1985, DOI 10.1007/978-3-642-05134-0 7,
c Springer-Verlag Berlin Heidelberg 2010
236 7 The Analytical Behaviour of Solutions

For each positive integer k, the space C k,α (D) consists of all Hölder continuous
functions whose derivatives up to order k are also Hölder continuous. For
v ∈ C k,α (D) we define the seminorms
8 9
vl,α,D := max ∂xi ∂yj v 0,α,D for l = 1, . . . , k
i+j=l

and

|v|l,α,D := |v|l,D + vl,α,D for l = 1, . . . , k.

The norm in C k,α (D) is

vk,α,D := max |v|l,α,D .

l=0,...,k

7.1.1 Stability

Consider the general elliptic second-order differential operator

Lu := −∇ (a∇u) + b∇u + cu

with a, b1 , b2 , c : Ω → IR and a > 0 on Ω̄.

Definition 7.1. The differential operator L is called inverse monotone or of pos-

itive type if it obeys a comparison principle. That is, for any two functions
v, w ∈ C 2 (0, 1) ∩ C[0, 1]

Lv ≤ Lw in Ω,
=⇒ v ≤ w on Ω̄.
v≤w on Ω̄

Lemma 7.2. Let there exist a function ψ ∈ C 2 (Ω) ∩ C(Ω̄) with ψ > 0 on Ω̄ and
Lψ > 0 in Ω. Then the operator L is inverse monotone.

Proof. The proof is by contradiction; see, e.g. [131]. 



Similar to Lemma 3.5 we have:

Lemma 7.3. Suppose there exists a function ψ ∈ C 2 (Ω) ∩ C(Ω̄) with ψ > 0 on Ω̄
and Lψ > 0 in Ω. Then for any function v ∈ C 2 (Ω) ∩ C(Ω̄) with v|∂Ω = 0
 
 Lv 
|v| ≤ ψ   on Ω̄.
Lψ ∞
7.1 Preliminaries 237

Remark 7.4. Under the assumptions of Lemma 7.3 the operator L is (L∞ , L∞ )
stable, or maximum-norm stable. That is

v∞ ≤ K Lv∞
&
with the stability constant K = ψ∞ min |(Lψ) (x)|. ♣
x∈Ω̄

7.1.2 Regularity of Solutions

In this Section we state compatibility conditions from [48, 162] that guarantee
smoothness of the solution of the boundary value problem

−εΔu − bT ∇u + cu = f in Ω = (0, 1)2 ,

(7.1)
u(·, 0) = gs , u(·, 1) = gn , u(0, ·) = gw , u(1, ·) = ge on [0, 1].

Let

F := f + bT ∇u − cu.

We define linear functionals Λk at the corner (0, 0) of the domain by

Λk (f, b, c, g) := εgs(2k) (0) + ε(−1)k+1 gw

(2k)
(0)

k
(7.2)
+ (−1)i−1 ∂x2(k−i) ∂y2(i−1) F (0, 0).
i=1

We also write Λ1k := Λk . In a similar way linear functionals Λlk , for l = 2, 3, 4, are
defined at the other vertices of Ω. For Λlk to be well defined, the data must possess
sufficient smoothness. We set

X k,α := (f, b, c, g) : f, b1 , b2 , c ∈ C k−2,α (Ω̄)

and gs , gw , gn , ge ∈ C k,α [0, 1] .

The following result is adapted from [48] which in turn is based on earlier work
by Volkov [162].
Theorem 7.5. Let α ∈ (0, 1] be arbitrary, but fixed. Suppose that (f, b, c, g) ∈
X 2,α and that Λl0 (f, b, c, g) = 0 for l = 1, 2, 3, 4. Then the solution u of (7.1) lies
in C 1,α (Ω̄).
Suppose (f, b, c, g) ∈ X 2k,α and u ∈ C 2k−1,α (Ω̄), then the linear function-
als Λlj (f, b, c, g) are well defined for j = 0, . . . , k and Λlj (f, b, c, g) = 0 for
238 7 The Analytical Behaviour of Solutions

j = 0, . . . , k − 1. We have Λlk (f, b, c, g) = 0 for l = 1, 2, 3, 4 if and only if

u ∈ C 2k,α (Ω̄). If in addition (f, b, c, g) ∈ X 2k+1,α , then u ∈ C 2k+1,α (Ω̄).
This theorem is used to derive compatibility conditions in terms of the data of
the problem (7.1), i. e., in terms of the functions b1 , b2 , c, f and the gs. For k = 0
the linear functional defined in (7.2) becomes

Λ0 (f, b, c, g) = εgs (0) − εgw (0) = 0.

Similarly Λl0 (f, b, c, g) = 0 for l = 2, 3, 4, so u ∈ C 1,α (Ω̄). Consequently, we

have u(0, 0) = gs (0), ∂x u(0, 0) = gs (0) and ∂y u(0, 0) = gw 
(0). Thus, the linear
functional Λ1 (f, b, c, g) can be reformulated as a local functional of the data:

Λ1 (f, b, c, g) = εgs (0) + εgw


(0) + f (0, 0)
+ b1 (0, 0)gs (0) + b2 (0, 0)gw

(0) − c(0, 0)gs (0).

Remark 7.6. Han and Kellogg [48] point out that in general the functionals Λk can-
not be reformulated as local functionals for k ≥ 2. For example, Λ2 (f, b, c, g)
contains the term ∂y b1 (0, 0)∂x ∂y u(0, 0) that cannot be expressed in terms of the
data of the problem by using (7.1) and continuity arguments. However, for reaction-
diffusion equations, i.e. vanishing convective field b, we have
 
Λ2 (f, 0, c, g) = εgs(4) (0) + f (·, 0) − c(·, 0)gs (0)
 
− εgw(4)
(0) − f (0, ·) − c(0, ·)gw (0) .

7.2 Reaction-Diffusion

We consider a singularly perturbed system of  ≥ 2 of reaction-diffusion equations

on the square. We seek the solution u = (u1 , u2 , . . . , u )T such that

Lu := −ε2 Δu + Au = f in Ω = (0, 1)2 , u|∂Ω = g, (7.3)

where the parameter ε satisfies 0 < ε  1.

Assume that a diagonal constant matrix D = diag(d11 , . . . , d ) exists with
dkk > 0 for all k and a constant μ > 0 such that

v T DAv ≥ μ2 v T v in Ω for all v ∈ IR . (7.4)

Under this hypothesis it is not true in general that the system (7.3) obeys a compar-
ison principle.
7.2 Reaction-Diffusion 239

Remark 7.7. In [57], which generalises the one-dimensional analysis of [18], the
authors analyse the case D = I. In [58], the coupling matrix A is assumed to
be a strictly diagonally dominant M -matrix, so Lemma 3.40 ensures that (7.4) is
satisfied.
Here we present an idea from [111] which extends the analysis of [18, 57] by
showing that it remains valid under the more general hypothesis (7.4). ♣

For vector-valued functions v ∈ C(G) , we introduce the norm

1/2
vD,G := sup v(x)T Dv(x) .
x∈G

Furthermore, set δ 2 := max dkk .

k=1,...,

7.2.1 Stability

The following stability result is a slight modification of an argument from [18, 57].
Note how neatly it side-steps the absence of a comparison principle for L.

Theorem 7.8. Let w ∈ C 2 (Ω) ∩ C(Ω̄) . Then

wD,Ω̄ ≤ max δ 2 μ−2 LwD,Ω , wD,∂Ω .

Proof. Set ϕ = wT Dw/2. Observe that 2ϕ ≤ δ 2 wT w on Ω̄ and

−wT DΔw = −Δϕ + ∂x wT D∂x w + ∂y wT D∂y w ≥ −Δϕ.

Taking the scalar product of Dw with −ε2 Δw + Aw = Lw and using the coer-
civity of DA, we get

2μ2
−ε2 Δϕ + ϕ ≤ wT DLw in Ω.
δ2
Hence, by the comparison principle for scalar problems,
 2 
δ  

 1 2
ϕ∞,Ω̄ ≤ max w DLw ∞,Ω , wD,∂Ω .
T
2μ2 2

It follows from the Cauchy-Schwarz inequality that

wD,Ω̄ = 2 ϕ∞,Ω̄ ≤ wD,Ω̄ max δ 2 μ−2 LwD,Ω , wD,∂Ω ,

2

and we are done. 


240 7 The Analytical Behaviour of Solutions

Remark 7.9. Theorem 7.8 implies that L is (L∞ , L∞ ) stable. ♣

Corollary 7.10. The solution u of (7.3) satisfies u∞,Ω̄ ≤ C, where the con-
stant C depends only on A, f and g.

7.2.2 Derivative Bounds

Let q ∈ {0, 1, 2} and α ∈ (0, 1]. Assume that we have A ∈ C 2(q+1),α (Ω̄) × and
f ∈ C 2(q+1),α (Ω̄) , and that the restriction of g to each edge of Ω̄ yields a function
lying in C 2(q+1),α [0, 1] , and that g is continuous at each of the corners; this implies
that g ∈ C 0,α (∂Ω) . Finally, suppose that at the south-west corner of the domain
we have the compatibility conditions

g s (0) − g w (0) = 0. (7.5a)

If q ≥ 1 also assume

−ε2 g s (0) − ε2 g w (0) + A(0, 0)g s (0) = f (0, 0) (7.5b)

and, if q = 2,
 
s (0) + f (·, 0) − A(·, 0)g s
ε2 g (4) (0)
  (7.5c)
− ε2 g (4)
w (0) − f (0, ·) − A(0, ·)g w (0) = 0

with corresponding conditions at the other three corners of the domain Ω̄.
These assumptions imply that the boundary-value problem (7.3) has a unique
solution u in ∈ C 2q+2,α (Ω) ∩ C 2q+1,α (Ω̄) ; see [75].
Corollary 7.10 enables the invocation of Schauder-type estimates; see [75]. Con-
sequently, we have the following result:
Lemma 7.11. Let q ∈ {1, 2, 3} be fixed. Suppose the compatibility conditions (7.5)
are satisfied at all corners of the domain Ω̄. Then the solution u of (7.3) satisfies
the bounds
 m k−m 
∂x ∂y u ≤ Cε−k for m = 0, . . . , k, k = 0, . . . , 2q − 1; (7.6a)
∞,Ω̄

furthermore,
 2m 2(q−m) 
∂x ∂y u∞,Ω̄ ≤ Cε−2q for m = 0, . . . , q. (7.6b)
7.2 Reaction-Diffusion 241

Remark 7.12. The inequality (7.6b), which stems from a result of Volkov [162],
is somewhat surprising since in general, one does not have u ∈ C 2q,α (Ω̄) ; note
that (7.6b) is valid only for certain derivatives of order 2q. ♣

We now derive sharper bounds on the pure derivatives of u, that show that the
large values seen in Lemma 7.11 do in fact decay rapidly as one moves away
from ∂Ω.
Theorem 7.13. Let u be the solution of (7.3). Let ρ ∈ (0, μ/δ) be arbitrary but
fixed. Suppose (7.5) hold true for some q ∈ {0, 1, 2}, then there exists a constant C,
which is independent of ε, such that
 k    
∂x u(x, y) ≤ C 1 + ε−k e−ρx/ε + e−ρ(1−x)/ε (7.7a)

and
 k    
∂y u(x, y) ≤ C 1 + ε−k e−ρy/ε + e−ρ(1−y)/ε , (7.7b)

for all (x, y) ∈ Ω̄ and k = 1, . . . , 2(q + 2).

Proof. Fix ρ ∈ (0, μ/δ) and set

 
Bk (x) := 1 + ε−k e−ρx/ε + e−ρ(1−x)/ε .

(i) First let us consider k ≤ 2q + 1. We use induction on k. For k = 0 the theorem

follows from Corollary 7.10. For k = 1, . . . , 2q+1, differentiating (7.3) k times
with respect to x gives

  
k  k−m  m
k−1
−ε2
Δ∂xk u + A∂xk u = ∂xk f − ∂x A ∂x u =: ϕk
m=0
m

with

|ϕk (x, y)| ≤ CBk−1 (x),

where the bound on ϕk is a consequence of the inductive hypothesis. Define ũ

by ∂xk u = Bk ũ. Then
 
B B  ϕk
−ε2 Δũ − 2ε2 k ∂x ũ + A − ε2 k ũ = .
Bk Bk Bk
242 7 The Analytical Behaviour of Solutions

Taking the inner product with D ũ and setting vk = 12 ũT D ũ we obtain, as in

the proof of Theorem 7.8,

Bk
−ε2 Δvk − 2ε2 ∂x vk + 2(μ2 δ −2 − ρ2 )vk ≤ Cũ,
Bk

by (7.4) and because Bk (x) ≤ ε−2 ρ2 Bk (x) and ϕk (x, y) ≤ CBk (x).
Boundary conditions for vk follow from Lemma 7.11: |vk | ≤ Cũ on ∂Ω.
Application of a comparison principle for scalar equations yields
2
ũ∞,Ω̄ = 2 vk ∞,Ω̄ ≤ C ũ∞,Ω̄ .

Dividing by ũ∞,Ω̄ gives ũ∞,Ω̄ ≤ C; then, recalling the definition of ũ,

we obtain the first inequality of the theorem. The second inequality is proved in
the same way.
(ii) Now consider the pure derivatives of order (2q + 2). The proof is different
since these derivatives are discontinuous at the corners of the domain and con-
sequently do not lie in C(Ω̄).
For each (x, y) ∈ Ω, let Γ (x, y) denote the distance from (x, y) to the near-
est of the four corners of Ω. Set Ωε = {(x, y) ∈ Ω : Γ (x, y) < ε}. From (7.6b)
we see immediately that (7.7a) holds true for k = 2q + 2 on Ω̄ε , because the
sum of exponentials in (7.7a) is then bounded below by a positive constant.
On the other hand, u ∈ C 2q+2,α (Ω \ Ωε ) since the corners of Ω̄ are ex-
cluded from Ω \ Ωε ; as regards boundary data for ∂x2q+2 u on ∂(Ω \ Ωε ), first
(7.6b) implies that (7.7a) is valid on ∂Ωε and on {(x, y) ∈ ∂(Ω \ Ωε ) : x =
 (2q+2) 
0 or 1}, then (7.3) gives ∂x u ≤ C on {(x, y) ∈ ∂(Ω \ Ωε ) : y =
0 or 1}. The maximum principle argument of (i) can now be invoked to prove
(7.7a) on Ω \ Ωε for k = 2q + 2.

The inequality (7.7b) follows similarly. 



Remark 7.14. The bounds of Theorem 7.13 were obtained without constructing any
decomposition of u. ♣

Remark 7.15. Consider the case of a single equation, i.e.,  = 1. Then one can
apply a maximum principle argument directly to (7.3), i.e., most of the argument
of Theorem 7.8 can be discarded, and following our subsequent analysis, one again
obtains the bounds of Theorems 7.13. These bounds are sharper than the bounds
obtained in [29] via a lengthy decomposition of u. ♣

Remark 7.16. When the above arguments are applied to a reaction-diffusion system
posed on a one-dimensional domain, this yields a slight improvement of the a priori
bounds of [18]. ♣
7.3 Convection-Diffusion 243

7.3 Convection-Diffusion

We now consider the two-dimensional convection-diffusion problem

−εΔu − b · ∇u + cu = f in Ω, u = g on Γ = ∂Ω. (7.8)

Its solution may typically exhibit three different types of layers: interior layers,
parabolic boundary layers and regular boundary layers. Let us assume that Ω is
a domain with a regular boundary that has a uniquely defined outward normal n
almost everywhere. Then the boundary can be divided into three parts:
!
Γ − := x ∈ Γ : bT n < 0! inflow boundary,
Γ 0 := x ∈ Γ : bT n = 0! characteristic boundary and
Γ + := x ∈ Γ : bT n > 0 outflow boundary.

With this notation the layers can be classified as follows:

Regular Boundary Layers occur at the outflow boundary Γ + and have a width
of O (ε ln(1/ε)). They are often also called exponential boundary layers.
Parabolic Boundary Layers occur at characteristic boundaries Γ 0 where the
boundary is parallel to the characteristics of the vector field b. They are there-
fore also called
√ characteristic boundary layers. In the non-degenerate case, their
width is O ( ε ln(1/ε)).
Interior Layers arise, e. g., from discontinuities in the boundary data at the in-
flow boundary Γ − and are propagated across the domain along the characteristics
of the vector field b. They are similar in nature to parabolic boundary layers and
therefore
√ also called characteristic or parabolic interior layers. Their thickness
is O ( ε ln(1/ε)).

7.3.1 Regular Layers

Consider the convection-diffusion problem

Lu := −εΔu − b · ∇u + cu = f in Ω = (0, 1)2 , u = 0 on Γ = ∂Ω, (7.9)

i. e., (7.8) on the unit square with homogeneous Dirichlet boundary conditions. The
regularity of its solution was studied in Sect. 7.1.2.
We assume that (b1 , b2 ) > (β1 , β2 ) > 0 on Ω̄ with constants β1 and β2 . These
assumptions on b imply that the solution has exponential or regular layers along the
sides x = 0 and y = 0.
244 7 The Analytical Behaviour of Solutions

A standard method to gain insight into the layer structure of the solution is the
method of matched asymptotic expansions. In [108] this approach is complemented
with a careful analysis of the remainder term of the expansion to obtain a solution
decomposition.
Theorem 7.17. Let f ∈ C 4,α (Ω̄) for some α ∈ (0, 1). Let n ≥ 2 be an integer.
Suppose that f satisfies the compatibility conditions

f (0, 0) = f (0, 1) = f (1, 1) = f (1, 0) = 0,

that
   
f f
(1, 1) = ∂x
∂y (1, 1),
b1 b2
      
f f f
∂y ∂x − D0 2
(1, 1) = ∂x (1, 1),
b1 b1 b2
         
f f f f
2
∂y ∂ x − D0 ∂x − D0 − 2D1 (1, 1)
b1 b1 b1 b1
 
3 f
= ∂x (1, 1)
b2

and that
     
f f
b2 ∂x2 2
(1, 1) = b1 ∂y (1, 1),
b2 b1

where D0 v := −∂y vb2 /b1 + vc/b1 and D1 v := ∂y v∂x (b2 /b1 ) − v∂x (c/b1 ). If
n ≥ 4, we assume in addition that

∂x b2 (0, 0) = ∂y b1 (0, 0).

Then the boundary value problem (7.9) has a solution u ∈ C 3,α (Ω̄), and this
solution can be decomposed as u = v + w1 + w2 + w12 , where

vC 2 (Ω̄) + εα |v|C 2,α (Ω̄) ≤ C ,

while for all x, y ∈ [0, 1] we have

 
 i 
∂x w1 (x, ·) ν,α ≤ Cε−i e−β1 x/ε ,
C ({x}×[0,1])
 
 j 
∂y w2 (·, y) μ,α ≤ Cε−j e−β2 y/ε
C ([0,1]×{y})
7.3 Convection-Diffusion 245

and
 
 i 
∂x ∂y jw12 (x, y) ≤ Cε−(i+j) e−(β1 x+β2 y)/ε

for 0 ≤ μ, ν ≤ 2 and 0 ≤ i, j ≤ n. Moreover, for all (x, y) ∈ Ω we have

|Lw1 (x, y)| ≤ Cεe−β1 x/ε , |Lw2 (x, y)| ≤ Cεe−β2 y/ε

and

|Lw12 (x, y)| ≤ Cεe−(β1 x+β2 y)/ε .

Proof. The argument is very technical an therefore omitted. The interested reader is
referred to the original publication [108]. 

The regular solution component is defined via solutions of hyperbolic problems.
Unlike elliptic operators, these first-order operators do not possess smoothing prop-
erties. Because of this we have to assume high regularity of f and a large number
of compatibility conditions in Theorem 7.17, but we expect such a decomposition
to exist under less restrictive assumptions. Similar ideas have been pursued in [30]
and [121], but compatibility issues are either not considered or dealt with incor-
rectly; see Remarks 5.1, 5.2 and 5.5 in [108].
Remark 7.18. If for the analysis of a scheme less regularity of the various compo-
nents of the decomposition is required, then some of the compatibility conditions
can be discarded, see [108, Remark 5.3]. ♣
Remark 7.19. A different approach is used by Roos [135] who defines the regular
solution component as the solution of an elliptic problem on an extended domain.
Therefore, the construction requires less regularity and compatibility of the data,
but only bounds for the first order derivatives of the components of u are obtained
in [135]. ♣

7.3.2 Characteristic Layers

Consider (7.8) with convective field b = (b, 0), i.e.,

−εΔu − bux + cu = f in Ω = (0, 1)2 , u = 0 on ∂Ω (7.10)

with b ≥ β > 0 and c ≥ 0 on Ω̄.

Near the outflow boundary at x = 0 a regular layer e−βx/ε will form, while along
the characteristic boundaries y√= 0 and y = 1√characteristic (or parabolic) layers
appear which behave like e−y/ ε and e−(1−y)/ ε , resp.
246 7 The Analytical Behaviour of Solutions

A detailed study of (7.10) has been conducted by Kellogg and Stynes [59,60] for
the case of constant b and c. Their analysis also takes into account the formation of
corner singularities when the data does not satisfy certain compatibility condition.

Theorem 7.20. Suppose b > 0 and c are constant. Let f ∈ C 8 (Ω̄) satisfy

f (0, 0) = f (0, 1) = f (1, 1) = f (1, 0) = 0.

Then the solution u of (7.10) can be decomposed as

u = v + w1 + w2 + w12 ,

where for all x, y ∈ [0, 1] and 0 ≤ i + j ≤ 2, we have the pointwise estimates

 i j   
∂x ∂y v(x, y) ≤ C, ∂xi ∂yj w1 (x, y) ≤ Cε−i e−βx/ε ,
 i j   √ √ 
∂x ∂y w2 (x, y) ≤ Cε−j/2 e−y/ ε + e−(1−y)/ ε ,
 i j   √ √ 
∂x ∂y w12 (x, y) ≤ Cε−(i+j/2) e−βx/ε e−y/ ε + e−(1−y)/ ε

and for 0 ≤ i + j ≤ 3 the L2 bounds

 i j   i j 
∂x ∂y v  ≤ C, ∂x ∂y w1  ≤ Cε−i+1/2 ,
0,Ω 0,Ω
 i j   i j 
∂x ∂y w2  ≤ Cε−j/2+1/4 , ∂x ∂y w12  ≤ Cε−i−j/2+3/4 ,
0,Ω 0,Ω

except for ∂x2 ∂y w2 for which

 2 
∂x ∂y w2  ≤ Cε−1/2 .
0,Ω

Proof. See [59, 60]. 


Chapter 8
Reaction-Diffusion Problems

In this chapter numerical methods for singularly perturbed reaction-diffusion

equations on the square are studied. Find u ∈ C 2 (Ω) ∪ C(Ω̄) such that

Lu := −ε2 Δu + cu = f in Ω = (0, 1)2 , u|∂Ω = g, (8.1)

where the parameter ε satisfies 0 < ε  1 and c ≥ γ > 0 on Ω.

8.1 Central Differencing

We consider central differencing for a system of  ≥ 2 singularly perturbed reaction-

diffusion equations on the square. Thus, instead of (8.1) we consider the more
general problem of finding a vector-valued function u = (u1 , u2 , . . . , u )T such
that

Lu := −ε2 Δu + Au = f in Ω = (0, 1)2 , u|∂Ω = g, (8.2)

where the parameter ε satisfies 0 < ε  1.

Assume that a diagonal constant matrix D = diag(d11 , . . . , d ) exists with
dkk > 0 for all k and a constant μ > 0 such that

v T DAv ≥ μ2 v T v in Ω for all v ∈ IR .

Let δ 2 = max dkk . The analytical behaviour of the solution to (8.2) was studied
k=1,...,
in Sect. 7.2.
Consider an arbitrary tensor-product mesh ω̄ = ω̄x × ω̄y on Ω̄, with

ω̄x : 0 = x0 < x1 < · · · < xN = 1

and

ω̄y : 0 = y0 < y1 < · · · < yN = 1.

T. Linß, Layer-Adapted Meshes for Reaction-Convection-Diffusion Problems, Lecture 247

Notes in Mathematics 1985, DOI 10.1007/978-3-642-05134-0 8,
c Springer-Verlag Berlin Heidelberg 2010
248 8 Reaction-Diffusion Problems

Let ω = ω̄ ∩ Ω and ∂ω = ω̄ ∩ ∂Ω. Set hi := xi − xi−1 and kj := yj − yj−1 . Given

N
a mesh function {vi,j }i,j=0 , introduce the standard difference operators
vij − vi−1,j vi+1,j − vij hi + hi+1
vx̄;ij = and vx̂;ij = with h̄i = .
hi h̄i 2

Analogously we define vȳ;ij , vŷ;ij and k̄ j .

Our discretisation is: Find uN such that
' N(  
Lu i,j := −ε2 uN N N
x̄x̂ + uȳ ŷ i,j + A(xi , yj )ui,j = f i,j in ω, (8.3a)
uN
i,j = g i,j on ∂ω, (8.3b)

T
where v x̄x̂ = (v1;x̄x̂ , v2;x̄x̂ , . . . , v ;x̄x̂ ) for vector-valued mesh functions.

8.1.1 Stability

For vector-valued mesh functions v, we set

1/2
vD,G := max v(x)T Dv(x) , for any G ⊂ ω̄.
x∈G

Theorem 8.1. The discrete operator L is (∞ , ∞ )-stable and satisfies the stability
inequality

wD,ω̄ ≤ δ 2 μ−2 max LwD,ω , wD,∂ω .

for arbitrary vector-valued functions w defined on ω̄.

Proof. This proof is a discrete analogue of the argument for Theorem 7.8. Let
ϕ = 12 wT Dw. Note that 2ϕ ≤ δ 2 wT w on ω̄ and

 dkk  
ϕx̄x̂;i,j = wTi,j Dwx̄x̂;i,j + 2
hi+1 wk;x;i,j 2
+ hi wk;x̄;i,j
2h̄i
k=1

with a similar identity for ϕȳŷ . Thus,

wTi,j D (wx̄x̂;i,j + wȳŷ;i,j ) ≤ ϕx̄x̂;i,j + ϕȳŷ;i,j .

Taking the scalar product of Dw with −ε2 (wx̄x̂ + wȳŷ ) + Aw = Lw and using
the coercivity of DA, we get

−ε2 (ϕx̄x̂ + ϕȳŷ ) + 2μ2 δ −2 ϕ ≤ wT DLw on ω .

8.1 Central Differencing 249

Furthermore, |ϕ| ≤ 12 w2D,∂ω on ∂Ω. A standard discrete maximum principle for

scalar problems, with a constant barrier function, yields
 2 
δ  wT DLw
 1 2
ϕ∞,ω̄ ≤ max ,
∞,ω 2
wD,∂ω .
2μ2

It follows from the Cauchy-Schwarz inequality that

wD,ω̄ = 2 ϕ∞,ω̄ ≤ wD,ω̄ max δ 2 μ−2 LwD,ω , wD,∂ω ,

2

and we are finished. 



Remark 8.2. Theorem 8.1 implies that the linear system (8.3) possesses a unique
solution uN . ♣

8.1.2 Convergence on Layer-Adapted Meshes

In the layer regions the Bakhvalov mesh is not approximately equidistant. Therefore,
the truncation error of the difference scheme is only first order at points in the layers.
Nonetheless, it is straightforward to modify the analysis from Sect. 6.1.2.4 for a two-
point boundary value problem to the two-dimensional problem (8.1), and to show
that the truncation error is of order N −2 uniformly in ε.
The analysis of central differencing for reaction-diffusion problems on Shishkin
meshes on the domain Ω has been carried out by Clavero et al. [29] in the case of
a single equation, and by Kellogg et al. [58] for a system of equations, but the error
analysis in both papers relies on the lengthy construction of a decomposition of the
solution. Here we shall present the much simpler analysis from [57].

Theorem 8.3. Suppose that the compatibility conditions (7.5a,b) are satisfied. Let
ρ ∈ (0, μ/δ) be arbitrary, but fixed. Assume that the mesh parameter σ satisfies
σ ≥ 2. Then

  CN −2 for Bakhvalov meshes,
u − u 
N
≤
∞,ω̄
CN −2 ln2 N for Shishkin meshes.

Proof. Theorem 8.1 yields

     
u − uN  ≤ C L(u − uN )∞,ω = C Lu − Lu∞,ω (8.4)
∞,ω̄

because u = uN on ∂ω. For the mth component of the truncation error we have
   
(Lu − Lu)m = ε2 ∂x2 um − um;x̄x̂ + ε2 ∂y2 um − um;ȳŷ
for m = 1, . . . , .
250 8 Reaction-Diffusion Problems

For any ψ ∈ C 4 (Ω) with ∂xk ψ bounded on Ω̄ for k = 0, . . . , 4, Taylor expansions

show that
' (   
 2
 ∂x ψ − ψx̄x̂ i,j  ≤ C ∂x2 ψ (·, yj )∞,[xi−1 ,xi+1 ] , (8.5a)
' (    
 2
 ∂x ψ − ψx̄x̂ i,j  ≤ C (hi + hi+1 ) ∂x3 ψ (·, yj )∞,[x ,x ]
i−1 i+1
(8.5b)

and
' (   
 2
 ∂x ψ − ψx̄x̂ i,j  ≤ C |hi − hi+1 | ∂x3 ψ(xi , yj ) (8.5c)
 
+ C(hi + hi+1 )2 ∂x4 ψ(·, yj )∞,[x .
i−1 ,xi+1 ]

There are analogous bounds for ∂y2 ψ − ψȳŷ .

Bakhvalov meshes
On choosing σ ≥ 2, an application of the technique in Sect. 6.1.2.4, combined with
the bound (7.7a) on ∂xk um for k = 2, 3, 4, yields
  

ε2  ∂x2 um − um;x̄x̂ i,j  ≤ CN −2 for i, j = 1, . . . , N − 1 and m = 1, . . . , .

There is a corresponding bound for ∂y2 um − um,ȳŷ . Thus, the truncation error is
uniformly bounded by CN −2 . Invoking (8.4), the proof of Theorem 8.3 for the
Bakhvalov mesh is complete.

Shishkin meshes
First consider the case where τ = q ≤ σερ−1 ln N . Then the mesh is uniform with
mesh
 size N−1 . Furthermore, ε−1 ≤ C ln N . Hence, Theorem 7.13 and (8.5c) give
Lu − Lu ≤ CN −2 ln2 N . Invoking (8.4), Theorem 8.3 follows in this case.
∞,ω
Now suppose that τ = σερ−1 ln N < q. Let x∗ = 2ερ−1 ln(1/ε). For each
m ∈ {1, . . . , } and (x, y) ∈ Ω̄, set
⎧ 4
⎪  (x − x∗ )ν
⎪
⎪ ∂xν um (x∗ , y) for 0 ≤ x ≤ x∗ ,
⎪
⎪
⎪
⎨ν=0 ν!
vm (x, y) = um (x, y) for x∗ ≤ x ≤ 1 − x∗ ,
⎪
⎪
⎪
⎪4
(x − x∗ )ν ν
⎪
⎪ ∂x um (1 − x∗ , y) for 1 − x∗ ≤ x ≤ 1,
⎩
ν=0
ν!

and wm (x, y) = um (x, y) − v(x, y). Then Theorem 7.13, and the choice of x∗ ,
imply that
 k   
∂x vm (x, y) ≤ C 1 + ε2−k
8.1 Central Differencing 251

and
 k   
∂x wm (x, y) ≤ Cε−k e−ρx/ε + e−ρ(1−x)/ε for 0 ≤ k ≤ 4.

That is, u = v+w, i.e., we have decomposed u into a sum of a regular component v
and a layer component w = (w1 , . . . , w )T . We remark that our decomposition does
not in general satisfy Lv = f and Lw = 0; these additional properties, which are
not needed here, have been obtained for various singular perturbation problems via
more complicated analyses.
Split the truncation error by writing
' 2 ( ' ( ' (
∂x um − um;x̄x̂ i,j = ∂x2 vm − vm;x̄x̂ i,j + ∂x2 wm − wm;x̄x̂ i,j .

When bounding the truncation error in v, use (8.5b) for i = qN or i = (1 − q)N ,

i.e., at the mesh transition points, and at other points use (8.5c). For the layer term w,
invoke (8.5a) for i = qN, . . . , (1 − q)N and (8.5c) otherwise. This yields

' (  CεN −1 for i ∈ {qN, (1 − q)N },

ε  ∂x2 um − δx2 um i,j  ≤ CN −2 ln2 N +
2
0 otherwise.

We cannot invoke (8.4) to get the desired error bound. Instead we proceed as
follows. Write u − uN = ψ + ϕ, where ψ and ϕ are the solutions of
 
Lψ = ε2 ∂x2 u − ux̄x̂ in ω, ψ=0 on ∂ω,
 
Lϕ = ε2 ∂y2 u − uȳŷ in ω, ϕ=0 on ∂ω.

In order to bound ψ we use the technique of Theorem 8.1. Set Ψ = 12 ψ T ψ. Then

 
−ε2 (Ψx̄x̂ + Ψȳŷ ) + 2μ2 δ −2 Ψ ≤ C ψ∞,ω Lψ  in ω, Ψ = 0 on ∂ω.

One can apply a discrete comparison principle for scalar equations using the barrier
function from [122]:
 
Ψ̄i,j = C1 ψ∞,ω N −2 ln2 N + τ ε−1 χi ,

where
⎧
⎪
⎪ −1
⎨xi τ for i = 0, . . . , qN,
χi := 1 for i = qN, . . . , (1 − q)N,
⎪
⎪
⎩(1 − x )τ −1 for i = (1 − q)N, . . . , N.
i
252 8 Reaction-Diffusion Problems

Then with C1 chosen sufficiently large, independently of ε, one obtains Ψ ∞,ω̄ ≤

CΨ̄ ∞,ω̄ . Hence,

ψ∞,ω̄ ≤ CN −2 ln2 N.

An identical bound is obtained for ϕ. The assertion of the theorem follows for the
Shishkin mesh. 


Remark 8.4. Our convergence analysis is based on derivative bounds whose deriva-
tion requires that A ∈ C 4,α (Ω̄) × and f ∈ C 4,α (Ω̄) , and that the compatibility
conditions (7.5a) and (7.5b) are satisfied. Therefore, u ∈ C 3,α (Ω̄) ∩ C 6 (Ω).
Andreev [8] analyses central differencing for a scalar reaction-diffusion equation
assuming that only the lowest order compatibility condition (7.5a) is satisfied. In
this case one has u ∈ C 1,α (Ω̄) ∩ C 6,α (Ω) only. Nonetheless, it is shown in [8] that
on a Shishkin mesh
 
u − uN  ≤ CN −2 ln4 N.
∞,ω̄

The key ingredient is a careful analysis of the corner singularity arising from the
violation of the compatibility condition (7.5b). ♣

Remark 8.5. The situation becomes more involved when the reaction-diffusion
equation is considered on an L-shaped domain:

−ε2 Δu + cu = f in Ω = (−1, 1) \ [0, 1]2 , u|∂Ω = g,

ε ∈ (0, 1], c ≥ ρ2 on Ω̄ with a positive constant ρ.

In addition to boundary layers along all six sides of the domain, there is also a
singularity at the re-entrant corner (0, 0). Because of this, the solution only lies in
the Hölder space C 2/3 (Ω̄). Andreev and Kopteva [9] consider central differencing
on a Shishkin mesh with additional refinement at the corner (0, ! 0).
Fix the transition parameter τ = min 1/3, 2ερ−1 ln N, 13 . The mesh ω̄x in
x-direction is then constructed by dividing each of the four intervals [−1, τ − 1],
[τ − 1, −τ ], [τ, 1 − τ ] and [1 − τ, 1] into M = N/6 equidistant subintervals. On
[−τ, τ ] a polynomial Shishkin mesh with m = 3 is used; see (2.13), i.e., the mesh
3
points are fixed by xi = τ (i/M ) , i = −M, . . . , M .
The final mesh on Ω̄ is ω̄ = ω̄x2 \ (0, 1)2 ; see Fig. 8.1. For central differencing
on this mesh the pointwise error bound
 
u − uN  ≤ CN −2 ln2 N
∞,ω̄

is established in [9]. ♣
8.1 Central Differencing 253

Fig. 8.1 Shishkin mesh and geometric refinement for a reaction-diffusion problem with boundary
layers and a corner singularity. The right plot zooms into the region [−τ, τ ]2 to better visualise the
polynomial refinement

8.1.3 Numerical Results

We now present the results of some numerical experiments in order to illustrate the
conclusions of Theorem 4.8, and to check if they are sharp.

Example 1.
 
−ε2 Δu1 + 2u1 + xu2 = sin π(x + y) in Ω, u1 |∂Ω = 0
−ε2 Δu2 + (1 + y 2 )u1 + (3 + x)u2 = 3x(1 − x) in Ω, u1 |∂Ω = 0.

For this problem D = I and μ2 ≈ 1.12.

Example 2.

−ε2 Δu1 + (3 + x)u1 + yu2 + x2 u3 = 1 in Ω, u1 |∂Ω = 0

−ε Δu2 + yu1 + (4 − y)u2 + xyu3 = 0
2
in Ω, u2 |∂Ω = 0
−ε Δu3 + x u1 + xyu2 + (3 − x )u3 = 1
2 2 2
in Ω, u3 |∂Ω = 0

with γ 2 ≈ 1.32.
For both problems we take ρ = 1 and σ = 2 in the construction of the meshes.
The perturbation parameter ε is chosen to be 10−8 .
The exact solutions to both test problems is not available, so we estimate the
accuracy of the numerical solution by comparing it with the numerical solution ob-
tained from Richardson extrapolation over two meshes, which has a higher order
254 8 Reaction-Diffusion Problems

of accuracy. Thus, let uN

ε be the solution of the difference scheme on the original
Shishkin mesh and ũ2N
ε that on the mesh obtained by uniformly bisecting the orig-
inal mesh. Then the extrapolated solution is

ε − uε
4ũ2N N
uR,N
ε := .
3
We then compute the error
  4 
ηεN := uN R,N 
ε − uε ∞,ω
= uN 2N 
ε − ũε ∞,ω
.
3
The results of our test computations are displayed in the following table.

Example 1 Example 2
Shishkin mesh Bakhvalov mesh Shishkin mesh Bakhvalov mesh
N ηN rN CN ηN rN CN ηN rN CN ηN rN CN
16 1.45e-2 0.93 0.48 5.17e-3 1.81 1.32 1.18e-2 0.80 0.39 4.53e-3 1.82 1.16
32 7.57e-3 1.41 0.65 1.37e-3 1.94 1.40 6.47e-3 1.20 0.55 1.19e-3 1.95 1.22
64 2.84e-3 1.47 0.67 3.52e-4 1.99 1.44 2.82e-3 1.46 0.67 3.03e-4 1.98 1.24
128 1.02e-3 1.58 0.71 8.86e-5 2.00 1.45 1.02e-3 1.58 0.71 7.61e-5 2.00 1.25
256 3.43e-4 1.65 0.73 2.22e-5 2.00 1.45 3.41e-4 1.65 0.73 1.90e-5 2.00 1.25
512 1.09e-4 — 0.73 5.55e-6 — 1.45 1.09e-4 — 0.73 4.76e-6 — 1.25

As well as the uniform errors η N , we also give experimental rates of convergence

rN which are computed by means of the formula
 % 
rN = log2 η N η 2N .

We also estimate the constants in the error estimate, i. e., if we have the theoretical
error bound η N ≤ Cϑ(N ), then we compute the quantity C N = η N /ϑ(N ).
The numerical results are in accordance with Theorem 8.3 and illustrate its sharp-
ness. The 2nd example deliberately does not satisfy the compatibility condition
f = 0 in the corners. Nonetheless, we observe (almost) second order convergence.
This illustrates Andreev’s findings.

8.2 Arbitrary Bounded Domains

So far we have considered only rectangular domains or domains that can be assem-
bled from two rectangles. Kopteva [67] considers the semilinear problem

−ε2 Δu + b(·, u) = 0, in Ω, u=g on ∂Ω, (8.6)

8.2 Arbitrary Bounded Domains 255

with an arbitrary bounded domain Ω ⊂ IR2 . Depending on the precise assumptions

on the data b and g, the problem may possess multiple solutions with a boundary
layer of width O (ε ln(1/ε)) along all of ∂Ω; see Fig. 8.2.
The domain is discretised as in Figure 8.3, using layer-adapted tensor-product
meshes of Bakhvalov and Shishkin types along the boundary with the typical re-
finement perpendicular to it. On this part of the mesh Kopteva’s method uses a
central finite difference approximation adapted to the curvilinearity of the mesh. In
the interior region, lumped mass linear finite elements on a quasiuniform
 Delaunay
triangulation is used. The number of mesh nodes is O h−2 , where h is the maxi-
mum side length of mesh elements. The key feature of the method is that it uses an
M -matrix discretisation of the Laplace operator.

−1
0.5
1
0 0.5
−0.5 0
−0.5
−1 −1

Fig. 8.2 Typical solution of (8.6)

0.6

0.4

0.2

−0.2

−0.4

−0.6

−0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8

Fig. 8.3 A Bakhvalov-type mesh for (8.6)

256 8 Reaction-Diffusion Problems

The resulting method is shown in [67] to be uniformly convergent if ε ≤ Ch in

the following sense. For any solution u of (8.6), there exists a discrete solution uN
such that the nodal error satisfies

  Ch2 | ln h|2 for a Shishkin mesh,
u − u 
N
≤
∞,ω̄
Ch2 for a Bakhvalov mesh.

Figs. 8.2 and 8.3 were kindly provided by N. Kopteva.

Chapter 9
Convection-Diffusion Problems

This chapter is devoted to numerical methods for the convection-diffusion problem

−εΔu − b∇u + cu = f in Ω = (0, 1)2 , u|∂Ω = 0, (9.1)

with b1 ≥ β1 > 0, b2 ≥ β2 > 0 on [0, 1]2 , i.e., problems with regular boundary
layers at the outflow boundary x = 0 and y = 0. The analytical behaviour of the
solution of (9.1) was studied in Sect. 7.3.1.
Results for problems with characteristic layers will only be mentioned briefly.

9.1 Upwind Difference Schemes

We shall consider discretisations of (9.1) on a tensor product mesh ω̄ = ω̄x × ω̄y

with N mesh intervals in both coordinate directions.
(N +1)2
The simple upwind scheme for (9.1) is: Find uN ∈ IR0 such that
' N(
Lu ij = fij for i, j = 1, . . . , N − 1 (9.2)

with
' N(  
x̄x̂;ij + uȳ ŷ;ij − b1;ij ux;ij − b2;ij uy;ij + cij uij ,
Lu ij := −ε uN N N N N

and

vi+1,j − vi vij − vi−1,j vi+1,j − vij

vx;ij = , vx̄;ij = and vx̂;ij = ,
hi+1 hi h̄i

h̄i = (hi + hi+1 ) /2 and analogous definitions for vy;ij , vȳ;ij , vŷ;ij and k̄ j .
This scheme on layer-adapted meshes was first studied by Shishkin who estab-
lished the maximum-norm error estimate

u − uN ∞,ω ≤ CN −1 ln2 N

T. Linß, Layer-Adapted Meshes for Reaction-Convection-Diffusion Problems, Lecture 257

Notes in Mathematics 1985, DOI 10.1007/978-3-642-05134-0 9,
c Springer-Verlag Berlin Heidelberg 2010
258 9 Convection-Diffusion Problems

on Shishkin meshes; see [121]. He also proved [151, §3, Theorem 2.3]
 p
u − uN ∞,ω ≤ C N −1 ln2 N

with p = 1/4 and p = 1/8 (depending on the precise assumptions on the data) if
the solution is less smooth.
Here we shall present the technique from [107] which gives a sharper error es-
timate. This technique is an extension of the truncation error and barrier function
technique from Sect. 4.2.6 to two dimensions.

9.1.1 Stability

The matrix associated with L is an L0 -matrix. Application of the M -criterion

(Lemma 3.14) with the test function vij = (1 − xi )/β1 establishes the inverse
monotonicity of L. As a consequence, we have the (∞ , ∞ )-stability inequality
      
 Lv   Lv   Lv   2
v∞,ω̄ ≤ min  
 b1  , 
 b2  ,
 c 
 for all v ∈ IR0N +1 ,
∞,ω ∞,ω ∞,ω

by Lemma 3.17.
Because of the inverse monotonicity we also have the following discrete compar-
 2
ison principle. For any two mesh functions v, w ∈ IRN +1

Lv ≤ Lw in ω and
=⇒ v ≤ w on ω̄.
v≤w on ∂ω

It will be used repeatedly in the convergence analysis.

9.1.2 Pointwise Error Bounds

Consider the upwind scheme (9.2) on a tensor-product Shishkin-type mesh ω̄ =

ω̄x ×ω̄y where the two one-dimensional meshes in x- and y-direction are constructed
as described in Sect. 2.1.3. Give mesh parameters σ > 0 and q > 0 the mesh uses
transition points in the S-type mesh are
   
σε σε
τx := min q, ln N and τy := min q, ln N .
β1 β2

Fig. 2.19 displays a plot of the resulting mesh.

9.1 Upwind Difference Schemes 259

Theorem 9.1. Assume the solution u of (9.1) can be decomposed as in Theorem

7.17 with α = 1 and n = 3. Let the mesh be a tensor-product S-type mesh with
σ ≥ 2. Suppose the mesh generating function ϕ̃ satisfies (2.8) and (2.14). Then the
error of the simple upwind scheme (9.2) satisfies
⎧  
  ⎨C h + N −1 for i, j = qN, . . . , N,
uij − uN 
ij ≤
⎩C h + N −1 max |ψ  | otherwise.

Proof. We adapt the truncation error and barrier function technique of Sect. 4.2.6 to
two space dimensions.
Recalling the decomposition of Theorem 7.17, we split the numerical solution in
a similar manner:

uN = v N + w1N + w2N + w12

N
,

where

Lv N = Lv, Lw1N = Lw1 , Lw2N = Lw2 , N

Lw12 = Lw12 on ω,

and

v N = v, w1N = w1 , w2N = w2 , N
w12 = w12 on ∂ω.

For the regular solution component a Taylor expansion, the derivative bounds of
Theorem 7.17 and the inverse monotonicity of L give
 
v − v N  ≤ Ch.
∞,ω

For the term representing the layer at x = 0, we have, similarly to (4.44),

i 
3 −1
β1 hk
0 ≤ w1;ij
N
≤ w̄1;i
N
:= C 1+ for i, j = 0, . . . , N.
2ε
k=1

Thus
 
N 
w1;ij − w1;ij ≤ CN −1 for i = qN, . . . , N, j = 0, . . . , N ;

see the argument that led to (4.45). Now let i < qN . Taylor expansions and
Theorem 7.17 give
   
L(w1 − w1N )ij  ≤ C h + ε−2 (hi + hi+1 ) e−β1 xi−1 /ε
 
≤ C h + ε−1 w̄1;i
N
N −1 max |ψ  | .
260 9 Convection-Diffusion Problems

Application of the comparison principle with the barrier function

 
C N −1 + h + w̄1;i
N
N −1 max |ψ  |

with C sufficiently large yields

   
N 
w1;ij − w1;ij ≤ C h + N −1 max |ψ  |
for i = 0, . . . , qN − 1, j = 0, . . . , N.

For the boundary layer at y = 0, the same type of argument is used in order to
obtain
 
N 
w2;ij − w2;ij ≤ CN −1 for i = 0, . . . , N, j = qN, . . . , N

and
   
w2;ij − w2;ij
N 
≤ C h + N −1 max |ψ  |
for i = 0, . . . , N, j = 0, . . . , qN − 1.

Finally, for the corner layer term one first shows

i 
3 j 
−1 3 −1
  β1 hk β2 kl
w12;ij − w12;ij
N  ≤ w̄12;ij
N
:= C 1+ 1+
2ε 2ε
k=1 l=1
for i, j = 0, . . . , N,

which implies
 
w12;ij − w12;ij
N  ≤ CN −1 if i ≥ qN or j ≥ qN.

In a second step the truncation error is estimated using Taylor expansions:

 
L(w12 − w12
N
)ij  ≤ Cε−1 w̄12;ij
N
N −1 max |ψ  |.

And the discrete comparison principle yields

 
w12;ij − w12;ij
N  ≤ CN −1 max |ψ  | for i, j = 0, . . . , qN − 1.

Collecting the bounds for the various components, we are finished. 



Remark 9.2. We are not aware of any a priori error estimates for arbitrary meshes
similar to those of Sect. 4.2.2 for one-dimensional problems. This seems to be due
to a lack of strong negative-norms stability inequalities. ♣
9.1 Upwind Difference Schemes 261

Remark 9.3. In [107] a modified, hybrid scheme on a standard Shishkin mesh is

considered. It is based on simple upwinding, but employs central differencing when-
ever the mesh allows one to do this without losing stability. For this scheme the
above technique gives the maximum-norm error bound
 
u − uN  ≤ CN −1 .
∞,ω

The improved bound is because central differencing improves the error terms of
order N −1 max |ψ  | in the above proof to order N −2 max |ψ  |2 . ♣

A numerical example
We briefly illustrate our theoretical findings for the simple upwind difference
scheme on S-type meshes and for the hybrid scheme when applied to the test
problem

−εΔu − (2 + x)ux − (3 + y 2 )uy + u = f in Ω = (0, 1)2 , (9.3a)

u = 0 on Γ = ∂Ω, (9.3b)

where the right-hand side is chosen such that

πx    
u(x, y) = cos 1 − e−2x/ε (1 − y)3 1 − e−3y/ε (9.3c)
2
is the exact solution. This function exhibits typical boundary layer behaviour. For
our tests we take ε = 10−8 , which is a sufficiently small choice to bring out the
singularly perturbed nature of the problem. Table 9.1 displays the results of our text
computations. They are in agreement with the theoretical findings.

Table 9.1 Upwind and hybrid difference scheme on S-type meshes

simple upwinding hybrid scheme
standard Shishkin mesh with Bakhvalov- standard
Shishkin mesh 2 transition points Shishkin mesh Shishkin mesh
N error rate error rate error rate error rate
16 9.6379e-2 0.50 9.0430e-2 0.73 9.3261e-2 0.74 1.1072e-1 0.88
32 6.8194e-2 0.59 5.4533e-2 0.76 5.5803e-2 0.90 5.9962e-2 0.94
64 4.5364e-2 0.66 3.2138e-2 0.79 2.9916e-2 0.93 3.1328e-2 0.97
128 2.8636e-2 0.72 1.8606e-2 0.84 1.5665e-2 0.97 1.6031e-2 0.98
256 1.7360e-2 0.77 1.0416e-2 0.87 8.0140e-3 0.98 8.1081e-3 0.99
512 1.0182e-2 0.80 5.6941e-3 0.90 4.0529e-3 0.99 4.0768e-3 1.00
1024 5.8286e-3 0.83 3.0602e-3 0.91 2.0379e-3 1.00 2.0440e-3 1.00
2048 3.2776e-3 — 1.6247e-3 — 1.0219e-3 — 1.0234e-3 —
262 9 Convection-Diffusion Problems

9.1.3 Error Expansion

Kopteva [65] derives an error expansion for the simple upwind scheme (9.2) on
standard Shishkin meshes. Let hc and hf denote the coarse and fine mesh sizes
in the Shishkin mesh. Provided that ε ≤ CN −1 , she proves that the error can be
expanded as
hf
ij − uij = hc Φij +
uN Ψ + Rij
ε ij
with
 
b1 (0, 0)x + b2 (0, 0)y
Φ(x, y) = ϕ(x, y) + ϕ(0, 0) exp −
ε
   
b1 (0, y)x b2 (x, 0)y
− ϕ(0, y) exp − − ϕ(x, 0) exp −
ε ε

and

x b21 (0, y)w̃1 + b21 (0, 0)w̃12 y b22 (x, 0)w̃2 + b22 (0, 0)w̃12
Ψ (x, y) = + ,
ε 2 ε 2

where the w̃’s satisfy bounds similar to those of Theorem 7.17 and ϕC 1,1 ≤ C,
while for the remainder we have

CN −2 for i, j = qN, . . . , N,
Rij ≤
−2 2
CN ln N otherwise.

This expansion is used in [65] to derive error bounds for Richardson extrapolation
and for the approximation of derivatives.

Richardson extrapolation
Let ũN be the upwind difference solution on the mesh obtained by uniformly bi-
secting the original mesh ω̄ and let Π ũN be the obvious restriction of ũN to ω̄.
Then

' (   N −2 for i, j = qN, . . . , N − 1,

 2Π ũ − u − u ij  ≤ C
N N
−2 2
N ln N otherwise [65].

These results are neatly illustrated by the numbers in Table 9.2 which display the
results of Richardson extrapolation applied to our test problem (9.3).
9.2 Finite Element Methods 263

Table 9.2 Richardson extrapolation on a Shishkin mesh

fine mesh region coarse mesh region
N error rate error rate
16 1.3869e-2 1.08 3.7171e-3 1.44
32 6.5448e-3 1.23 1.3733e-3 1.74
64 2.7918e-3 1.38 4.1086e-4 1.87
128 1.0703e-3 1.49 1.1271e-4 1.93
256 3.8049e-4 1.58 2.9616e-5 1.96
512 1.2701e-4 1.64 7.5975e-6 1.98
1024 4.0623e-5 — 1.9234e-6 —

Derivative approximation
In [65] the bounds
 N 
(u − u)x;ij 
⎧
⎪
⎪ −1
for i, j = qN, . . . , N − 1,
⎨N
≤ C N −1 ln2 N for i = qN, . . . , N − 1, j = 0, . . . , qN − 1,
⎪
⎪
⎩ε−1 N −1 ln N otherwise

are given with analogous results for (uN − u)y .

9.2 Finite Element Methods

This section is concerned with finite element discretisations of (9.1).

The variational formulation of (9.1) is as follows. Find u ∈ H01 (Ω) such that

a(u, v) = f (v) for all v ∈ H01 (Ω), (9.4)

where

a(u, v) = ε(∇u, ∇v) − (b · ∇u, v) + (cu, v) and f (v) = (f, v)

with

(u, v) := u(x, y)v(x, y)dxdy.
Ω

If

c+ 1
2 div b ≥ γ > 0 (9.5)
264 9 Convection-Diffusion Problems

Fig. 9.1 Triangulations into rectangles and triangles on tensor-product layer-adapted meshes

then the bilinear form a(·, ·) is coercive, i.e.,

2  
a(v, v) ≥ |||v|||ε := ε ∂x v20 + ∂y v20 + γv20 for all v ∈ H01 (Ω),

and (9.4) possesses a unique solution u ∈ H01 (Ω).

We shall restrict ourselves to tensor-product meshes ω̄ := ω̄x × ω̄y as in Sect. 9.1.
Set Ii := [xi−1 , xi ], Jj := [yj−1 , yj ] and Tij := Ii × Jj . We shall consider both bi-
linear elements on rectangles and linear elements on triangles with the triangulation
obtained by drawing either diagonal in each of the mesh rectangles; see Fig. 9.1.

9.2.1 The Interpolation Error

The first important results are bounds for the interpolation error. We denote by ψ I
the piecewise bilinear function that interpolates to ψ at the nodes of the mesh ω̄. The
meshes we consider are characterised by high aspect ratios of the mesh elements.
Because of this anisotropy, standard interpolation theory cannot be applied. There
have been a number of contributions to extend the theory to anisotropic elements,
e.g., [15, 171, 172]. The first uniform interpolation error estimates for layer-adapted
meshes, namely Shishkin meshes, were derived by Stynes and O’Riordan [152] and
Dobrowolski and Roos [30]. Here we shall give the more general results from [84].
Set
     
1 + ε−1 e−β1 x/(pε) dx + 1 + ε−1 e−β2 y/(2ε) dy
[p]
Θcd (Tij ) :=
Ii Jj

and
[p] [p]
ϑcd (ω̄) := max Θcd (Tij ).
i,j=1,...,N
9.2 Finite Element Methods 265

Theorem 9.4. Suppose the assumptions of Theorem 7.17 are satisfied. Then the
maximum-norm error of bilinear interpolation on a tensor-product mesh satisfies

 I   2
u − u ≤ C Θ
[2]
(T ij ) , (9.6)
∞,Tij cd
 
ε ∇(uI − u)∞,T ≤ CΘcd (Tij )
[1]
(9.7)
ij

and for the ε-weighted energy norm

 I 
u − u ≤ Cϑ[2] (ω̄) .
ε cd

Proof. First Theorem 7.17 implies

 i j     
∂x ∂y u(x, y) ≤ C 1 + ε−i e−β1 x/ε × 1 + ε−j e−β2 y/ε (9.8)

for i + j ≤ 2.
(i) Let (x, y) ∈ Tij . Then
  
  1 x ξ
u − uI (x, y) = ∂x2 u(τ, yj−1 )dτ dξdσ
hi Ii xi−1 σ
  y  ξ
1
+ ∂ 2 u(xi−1 , τ )dτ dξdσ
kj Jj yj−1 σ y
 x  y
+ ∂x ∂y u(ξ, τ )dτ dξ
xi−1 yj−1
 
x − xi−1 y − yj−1
− ∂x ∂y u(ξ, τ )dτ dξ.
hi kj Ii Jj

Applying the technique from Prop. 5.1, we immediately see that the first two
[2]
terms are bounded by Θcd (Tij )2 . The third and fourth term are clearly bounded
[1]
by Θcd (Tij )2 . Ineq. (9.6) follows.
(ii) Next, we have
  x  y
  1
∂x u − u (x, y) =
I 2
∂ u(τ, yj−1 )dτ dσ + ∂x ∂y u(x, τ )dτ
hi Ii σ x yj−1
 
1 y − yj−1
− ∂x ∂y u(ξ, τ )dτ dξ.
hi kj Ii Jj

Thus,
 
     2 
∂x u − uI (x, y) ≤ ∂x u(τ, yj−1 ) dτ + 2 |∂x ∂y u(x, τ )| dτ
Ii Jj

and (9.7) follows from (9.8).

266 9 Convection-Diffusion Problems

(iii) In order to bound the interpolation error in the H 1 seminorm, use integration
by parts:

   
∂x (uI − u)2 = ∂x2 u(x, y) uI − u (x, y)dxdy
0
Ω
 −1  1 (9.9)
N
 
+ uI − u (xi , y)Ki (y)dy ,
i=1 0

where

Ki (y) := ∂x uI (xi − 0, y) − ∂x uI (xi + 0, y) .

For y ∈ Jj we have
y − yj−1 yj − y
Ki (y) = (ux̄;ij − ux;i+1,j ) + (ux̄;i,j−1 − ux;i+1,j−1 ) .
kj kj

By the mean-value theorem there exists a ξi,j ∈ Ii , such that ux̄;i,j = ∂x u(ξi,j , yj ).
Therefore,
 xi+1
 2 
|ux̄;ij − ux;i+1,j | = |∂x u(ξi,j , yj ) − ∂x u(ξi+1,j , yj )| ≤ ∂x u(ξ, yj ) dξ .
xi−1

We get
 xi+1  2 
|Ki (y)| ≤ max ∂x u(ξ, y) dξ .
y∈[0,1] xi−1

This and a Hölder inequality applied to (9.9) yield

 I 
(u − u)x 2
0
  1 
   2   2 
≤ uI − u∞ ∂x u(x, y) dxdy + 2 max ∂x u(x, y) dx
y∈[0,1] 0

−1  I
Ω
≤ Cε 
u −u ∞,

by (9.8). The interpolation error in the L2 norm is bounded by its L∞ norm. We get
the second bound of the theorem. 


Remark 9.5. The second part of the proof when the H 1 seminorm is considered
works for bilinear elements, but not for linear ones. Nonetheless, for S-type meshes
and linear elements, the conclusions of the theorem hold too; see [82, 137]. ♣

Remark 9.6. Error bounds for particular layer-adapted meshes can be immediately
concluded using the results from Sections 2.1.1 and 2.1.3. ♣
9.2 Finite Element Methods 267

9.2.2 Galerkin FEM

Let V0ω ⊂ H01 (Ω) be the space of piecewise linear/bilinear functions on the given
triangulation that vanish on the boundary of Ω. Then our discretisation is as follows:
Find uN ∈ V0ω such that
 
a uN , v = f (v) for all v ∈ V0ω .

The coercivity of a(·, ·) guarantees the existence of a unique solution uN ∈ V0ω .

9.2.2.1 Convergence

Convergence of the (bi)linear Galerkin FEM on standard Shishkin meshes was first
studied by Stynes and O’Riordan [152]. Their technique was later adapted by Linß
and Roos to the analysis of more general S-type meshes [82, 137].
The mesh transition points in the S-type mesh are
   
σε σε
τx := min q, ln N and τy := min q, ln N
β1 β2

with mesh parameters σ > 0 and q ∈ (0, 1) arbitrary, but fixed and with qN ∈ IN .
Divide the domain Ω as in Fig. 9.2: Ω̄ = Ω11 ∪ Ω21 ∪ Ω12 ∪ Ω22 .
Corollary 9.7. Let ω̄ = ω̄x × ω̄y be a S-type mesh with σ ≥ 2. Then Theorem 9.4
implies
   2  
u − uI  ≤ C h + N −1 max |ψ  | , u − uI  ≤ CN −2 ,
∞,Ω\Ω 22 ∞,Ω 22

Ω12 Ω22

Ω11 = [0, τx ] × [0, τy ],

Ω21 = [τx , 1] × [0, τy ],
Ω12 = [0, τx ] × [τy , 1],
τy
Ω11 Ω21 Ω22 = [τx , 1] × [τy , 1].

τx
Fig. 9.2 Dissection of Ω for tensor-product S-type meshes
268 9 Convection-Diffusion Problems

and
   
u − uI  ≤ C h + N −1 max |ψ  | .
ε

Furthermore, by the Cauchy-Schwarz inequality,

   2
u − uI  ≤ Cε1/2 ln1/2 N h + N −1 max |ψ  |
0,Ω\Ω 22

and
 
u − uI  ≤ CN −2 .
0,Ω 22

Where we have used the results of Sect. 2.1.3 too.

Theorem 9.8. Let ω̄ = ω̄x × ω̄y be a tensor-product S-type mesh with σ ≥ 2, whose
mesh generating function ϕ̃ satisfies (2.8) and
 
h + N −1 max |ψ  | ln1/2 N ≤ C. (9.10)

Then
   
u − uN  ≤ C h + N −1 max |ψ  |
ε

for both linear elements on triangles and bilinear elements on rectangles.

Proof. The proof is along the lines of Sect. 5.2.1 exploiting the tensor-product struc-
ture of the mesh and the solution decomposition of Theorem 7.17; see also [82]. Let
η = uI − u and χ = uI − uN . Bounds for the interpolation error η are provided by
Corollary 9.7.
Bounding χ, we start from the coercivity of a(·, ·) and the orthogonality of the
Galerkin method, i. e.,
2  
|||χ|||ε ≤ a(χ, χ) = a(η, χ) = ε(∇η, ∇χ) + (η, bT ∇χ) + (div b + c)η, χ

≤ C |||η|||ε |||χ|||ε + C η0,Ω22 ∇χ0,Ω22

+ ηL∞ (Ω\Ω22 ) ∇χL1 (Ω\Ω22 ) .

On Ω \ Ω22 the Cauchy-Schwarz inequality yields

,
∇χL1 (Ω\Ω22 ) ≤ C τx + τy ∇χ0 ≤ C ln1/2 N |||χ|||ε ,

while on Ω22 an inverse inequality yields

∇χ0,Ω22 ≤ CN χ0,Ω22 ≤ CN |||χ|||ε .

9.2 Finite Element Methods 269

because hi ≥ N −1 and kj ≥ N −1 for Tij ∈ Ω22 . These two bounds and Corollary
9.7 give
 2
|||χ|||ε ≤ C h + N −1 max |ψ  | + h + N −1 max |ψ  | ln1/2 N + N −1 .

Thus,
 
|||χ|||ε ≤ C h + N −1 max |ψ  | ,

where we have used (9.10). A triangle inequality and the bounds for |||η|||ε and |||χ|||ε
complete the proof. 


9.2.2.2 Supercloseness

Similar to the one dimensional case, the Galerkin FEM using bilinear elements on
rectangular S-type meshes enjoys a superconvergence property; see [83, 176]. Note
that this superconvergence result generally does not hold for linear elements on
triangles as numerical experiments confirm [109].
In contrast to the one-dimensional case where we have ((uI − u) , χ ) = 0 for
arbitrary χ ∈ V ω , we do not have (∇(uI − uN ), ∇χ) = 0 here because uI − u
vanishes in the mesh points only, but not at the inter-element boundaries. This com-
plicates the analysis and requires higher regularity of the solution. In particular, we
shall assume that the solution u can be decomposed as u = v + w1 + w2 + w12 ,
where
 i j 
∂x ∂y v(x, y) ≤ C, (9.11a)
 i j 
∂x ∂y w1 (x, y) ≤ Cε−i e−β1 x/ε , (9.11b)
 i j 
∂x ∂y w2 (x, y) ≤ Cε−j e−β2 y/ε (9.11c)

and
 i j 
∂x ∂y w12 (x, y) ≤ Cε−(i+j) e−(β1 x+β2 y)/ε (9.11d)

for i + j ≤ 3 and x, y ∈ [0, 1].

Theorem 9.9. Let ω̄ = ω̄x × ω̄y be a tensor-product S-type mesh with σ ≥ 5/2 that
satisfies (2.8). Then the error of the Galerkin FEM uN satisfies
 I   
u − uN  ≤ C h2 ln1/2 N + N −2 max |ψ  |2 .
ε
270 9 Convection-Diffusion Problems

Proof. The coercivity and Galerkin orthogonality of a(·, ·) give

 I    
u − uN 2 ≤ a u − uI , uI − uN 
ε
 
≤ ε  ∇(u − uI ), ∇(uI − uN ) 
 
 
+  bT ∇(u − uI ) − c(u − uI ), uI − uN  .

In the Sect. 9.2.2.3 we shall show that for all χ ∈ V ω

   
ε  ∇(u − uI ), ∇χ  ≤ C h2 + N −2 max |ψ  |2 |||χ|||ε (9.12)

and
 
 T 
 b ∇(u − uI ) − c(u − uI ), χ 
  (9.13)
≤ C h2 ln1/2 N + N −2 max |ψ  |2 |||χ|||ε .

Thus,
 I     
u − uN 2 ≤ C h2 ln1/2 N + N −2 max |ψ  |2 uI − uN  .
ε ε

 
Divide by uI − uN ε to complete the proof.

Theorem 9.9 yields


 I  −2 2
u − uN  ≤ CN ln N for the standard Shishkin mesh and
ε
C(ε2 + N −2 ) ln1/2 N for the Bakhvalov-Shishkin mesh;

see Sect. 2.1.3 for the bounds on h and max |ψ  |.

Remark 9.10. Another superconvergence result was established by Zhang [176],

who studied convergence of the Galerkin FEM on Shishkin meshes in a discrete
version of the energy norm, where ∇(u − uN ) is replaced by a piecewise-constant
approximation based on the midpoints of the rectangles of the triangulation. ♣

9.2.2.3 Detailed Analysis, Proofs of (9.12) and (9.13)

In the analysis we require error estimates for interpolation on anisotropic elements,

which were derived by Apel and Dobrowolski [15]. Furthermore, a sharp bound for
the L2 -norm error of the interpolation error for the layer terms is needed. We shall
also use special error expansion formulae derived by Lin [78].
9.2 Finite Element Methods 271

Preliminaries
Let Tij = Ii × Jj be an element of the triangulation. Set

(x − xi−1/2 )2 h2 (y − yj−1/2 )2 kj2

Fi (x) = − i and Gj (y) = − ,
2 8 2 8

where (xi−1/2 , yj−1/2 ) is the midpoint of the mesh rectangle Tij . Denote the east,
north, west and south edges of Tij by lk;ij for k = 1, . . . , 4, respectively.
Lemma 9.11 (Lin Identities [78]). For any function g ∈ C 3 (T ij ) and any χ ∈ V ω
we have the identities
   
1  
∂x (g − g I )χx = Gj ∂x χ − G2j ∂x ∂y χ ∂x2 ∂y g , (9.14a)
Tij Tij 3
   
∂x (g − g )χy =
I
− Fj ∂x2 gχx
Tij l4;ij l2;ij
   (9.14b)
 
+ Gj ∂x ∂y2 g ∂y χ − Fi ∂x ∂y χ + Fj ∂x2 ∂y g∂x χ
Tij

and

∂x (g − g I )χ
Tij
 * +
1  2   
= Gj (χ − Fi ∂x χ) − Gj ∂y χ − Fi ∂x ∂y χ ∂x ∂y2 g (9.14c)
Tij 3
   2  * +
hi 2 1  2  h2i
+ − χ∂x g + Fi ∂x χ − χ ∂x3 g .
l1;ij l2;ij 12 Tij 6 12

An immediate consequence of (9.14a) is

   
   kj2   kj3   
∂x (g − g ), ∂x χ Tij  ≤
I  2 
|∂x χ| ∂x ∂y g + ∂x ∂y χ ∂x ∂y2 g  .
8 Tij 24 Tij

with the Cauchy-Schwarz and an inverse inequality giving

     

∂x (g − g I ), ∂x χ Tij  ≤ Ckj2 ∂x χ0,Tij ∂x ∂y2 g 0,Tij . (9.15)

Lemma 9.12 ([15, Theorem 3]). Let Tij ∈ Ω N and p ∈ [1, ∞]. Assume that g lies
in Wp2 (Tij ). Let g I denote the bilinear function that interpolates to g at the vertices
of Tij . Then
 
∂x (g − g I ) ≤ C hi ∂x2 gLp (Tij ) + kj ∂x ∂y gLp (Tij ) .
L p (Tij )
272 9 Convection-Diffusion Problems

Proposition 9.13. Let ω̄ = ω̄x × ω̄y be a tensor-product S-type mesh that satis-
fies (2.8). Then for w = w1 + w2 + w12
   
w − wI  ≤ C ε 1/2 −σ
N + N −σ−1/2
(9.16a)
0,Ω22

and
   2
w − wI  ≤ Cε1/2 h + N −1 max |ψ  | if σ > 2. (9.16b)
0,Ω\Ω 22

 
Proof. (i) When proving (9.16a), we bound w0,Ω22 and wI 0,Ω separately and
22
apply a triangle inequality. Clearly,

w0,Ω22 ≤ Cε1/2 N −σ , (9.17)

by (9.11) and a direct calculation.

In order to bound the L2 norm of wI we split Ω22 into two subdomains

S := [xqN +1 , 1] × [yqN +1 , 1] and Ω22 \ S.

Note that Ω22 \ S consists of only one ply of mesh rectangles along the interface
between the coarse and the fine mesh regions. We have
 I 2    2
w  ≤ 2(1 − q)N − 1 hqN +1 kqN +1 wI ∞,Ω .
0,Ω 22 \S 22

Hence,
 I
w  ≤ CN −σ−1/2 . (9.18)
0,Ω 22 \S

For Tij ⊂ S we estimate as follows:

 I 2  2
w  ≤ hi kj wI ∞,T
0,Tij
  
ij

≤C e−2β1 x/ε + e−2β2 y/ε + e−2(β1 x+β2 y)/ε ,
Ii−1 Jj−1

by (9.11) and since the mesh on Ω22 is uniform. We get

   
 I 2 1 1
w  ≤ C e−2β1 x/ε + e−2β2 y/ε + e−2(β1 x+β2 y)/ε
0,S
τx τy

and
 I
w  ≤ Cε1/2 N −σ . (9.19)
0,S

Collecting (9.17), (9.18) and (9.19), we obtain (9.16a).

9.2 Finite Element Methods 273

(ii) Before starting the proof of (9.16b) note that by (2.11) and (2.12)

βhi
≤ N −1 max |ψ  |eβ1 xi /σε ≤ CN −1 max |ψ  | eβ1 x/σε for x ∈ Ii .
σε

(α) First let us study w1 − w1I . For Tij ⊂ Ω12 ∪ Ω11 we have by Lemma 9.12
and (9.11)
 
 
w1 − w1I 2
≤ C hi kj
4
ε−4 e−2β1 x/ε dx
0,T ij
Ii

  
−2 −2β1 x/ε −2β1 x/ε
+ h2i kj3 ε e dx + kj5 e dx
Ii Ii

 4
≤ C N −1 max |ψ  | + h kj e−(2−4/σ)β1 x/ε dx.
Ii

Summing over all elements in Ω11 ∪ Ω12 , we get


   4 τx
w1 − w1I 2 ≤C N −1
max |ψ | + h
e−(2−4/σ)β1 x/ε dx.
0,Ω 11 ∪Ω12
0

Thus,
   2
w1 − w1I  ≤ Cε1/2 h + N −1 max |ψ  |
0,Ω 11 ∪Ω12

because σ > 2 is assumed.

On Ω21 we estimate as follows
  ,   ,
w1 − w1I  ≤ meas Ω21 w1 − w1I ∞,Ω ≤ meas Ω21 w1 ∞,Ω21
0,Ω 21 21

≤ Cε1/2 ln1/2 N e−β1 τx /ε ≤ Cε1/2 N −2

because σ > 2.
Therefore,
   2
w1 − w1I  ≤ Cε1/2 h + N −1 max |ψ  | (9.20)
0,Ω\Ω 22

since max |ψ  | ≥ 1.
(β) Clearly a similar argument yields
   2
w2 − w2I  ≤ Cε1/2 h + N −1 max |ψ  | . (9.21)
0,Ω\Ω 22
274 9 Convection-Diffusion Problems

(γ) Finally, let us bound w12 − w12

I
. For Tij ⊂ Ω11 we have by Lemma 9.12
and (9.11)
 
   2 
w12 − w12 I 2
≤ C hi + kj 2 2
ε−4 e−2(β1 x+β2 y)/ε dydx
0,T ij
Ii Jj
 
 4
≤ C h + N −1 max |ψ  | e−(2−4/σ)(β1 x+β2 y)/ε dydx .
Ii Jj

Summing over all elements in Ω11 , we get

 
I 2
w12 − w12
0,Ω 11
 
 −1 
4 τx τy
≤C N max |ψ | + h e−(2−4/σ)(β1 x+β2 y)/ε dydx .
0 0

Hence,
   2
I 
w12 − w12 ≤ Cε N −1 max |ψ  | + h . (9.22)
0,Ω 11

On Ω12 ∪ Ω21 we estimate as follows:

  ,  
I 
w12 − w12 ≤ meas Ω12 ∪ Ω21 w12 − w12
I 
0,Ω12 ∪Ω21 ∞,Ω12 ∪Ω21

≤ ε1/2 N −σ ln1/2 N w12 ∞,Ω12 ∪Ω21

≤ Cε1/2 N −σ ln1/2 N ≤ Cε1/2 N −2 , (9.23)

because σ > 2.
Collect (9.20)–(9.23) to complete the proof. 


Proof of (9.12)
(i) Using (9.15), we obtain for Tij ⊂ Ω11 ∪ Ω21
  

 ∂x (u − uI ), ∂x χ Tij 
  
 
≤ Ckj2  1 + ε−1 e−β1 x/ε 1 + ε−2 e−β2 y/ε  ∂x χ0,Tij
0,Tij
  
 
≤ Ckj2 1 + ε−2 e−β2 yj−1 /ε 1 + ε−1 e−β1 x/ε  ∂x χ0,Tij
0,Tij
 2 



≤ C h + N −1 max |ψ  | 1 + ε−1 e−β1 x/ε  ∂x χ0,Tij ,
0,Tij
9.2 Finite Element Methods 275

by (2.12) and since eβ2 kj /ε ≤ C because of (2.8). Application of the Cauchy-

Schwarz inequality yields
  
 
ε  ∂x (u − uI ), ∂x χ Ω ∪Ω 
2  
11 21
  
≤ Cε h + N −1 max |ψ  | 1 + ε−1 e−β1 x/ε  ∂x χ0,Ω11 ∪Ω21 .
0,Ω11 ∪Ω21

Hence,
    2
 
ε  ∂x (u − uI ), ∂x χ Ω
11 ∪Ω21
 ≤ C h + N −1 max |ψ  | |||χ|||ε . (9.24)

(ii) An argument similar to (i) gives

  
 
ε  ∂x ((v + w1 ) − (v + w1 )I ), ∂x χ Ω ∪Ω 
 
12 22

 
≤ Cεh2 1 + ε−1 e−β1 x/ε  ∂x χ0,Ω12 ∪Ω22
0,Ω12 ∪Ω22

≤ Ch |||χ|||ε .
2
(9.25)

(iii) Next we consider w := w2 + w12 for Tij ⊂ Ω12 . The stability of the interpo-
lation operator and our bounds on the derivatives of w2 and w12 yield
 
∂x (w − wI )
∞,T
 
ij

≤ ∂x w∞,Tij + ∂x wI ∞,T ≤ C ∇w∞,Tij ≤ Cε−1 N −σ .

ij

Thus,
  

ε  ∂x (w − wI ), ∂x χ Ω12  ≤ CN −σ ∂x χ1,Ω12 ≤ CN −σ ε1/2 ln1/2 N |||χ|||ε ,

since meas Ω12 = O (ε ln N ). Therefore,

   

ε ∂x (w2 + w12 ) − (w2 + w12 )I ), ∂x χ Ω  ≤ CN −2 |||χ|||ε , (9.26)
12

because σ > 2.
(iv) Finally, let us bound the terms involving w2 and w12 on Ω22 . Using
Lemma 9.12 and (9.11) we get
 
∂x (w2 − w2I ) ≤ Cε−1/2 N −σ
0,Ω 22

and
 
∂x (w12 − w12
I 
) 0,Ω ≤ Cε−1 N −2σ .
22
276 9 Convection-Diffusion Problems

Thus,
  

ε  ∂x (w2 − w2I ), ∂x χ Ω  ≤ CN −2 |||χ|||ε (9.27)
22

and
  

ε  ∂x (w12 − w12
I
), ∂x χ Ω  ≤ CN −2σ ∂x χ0,Ω22 ≤ CN −2σ+1 χ0,Ω22 ,
22

(9.28)

by an inverse inequality.
Collect (9.24)-(9.28) to obtain
   2
ε  ∂x (u − uI ), ∂x χ  ≤ C h + N −1 max |ψ  | |||χ|||ε for all χ ∈ V ω .
 
Obviously, we have an identical bound for  ∂y (u − uI ), ∂y χ  which completes
the proof of (9.12).

Proof of (9.13)
Recalling the decomposition (9.11), we set w = w1 + w2 + w12 . Then integration
by parts yields
    

− bT ∇(u − uI ), χ + c(u − uI ), χ 
   

≤  bT ∇(v − v I ), χ  +  w − wI , bT ∇χ  (9.29)
   

+  c(v − v I ), χ + (c + div b)(w − wI ), χ 

The terms on the right-hand side are analysed separately.

First
   

 c(v − v I ), χ + (c + div b)(w − wI ), χ 
 
≤ C v − v I 0 + w − wI 0 χ0 .

Adapting the technique from Sect. 9.2.1 it is shown that

 2
v − v I 0 + w − wI 0 ≤ C h + N −1 max |ψ  | ,

since v and w satisfy derivative bounds similar to those of u. Thus,

     2

 c(v − v I ), χ + (c + div b)(w − wI ), χ  ≤ C h + N −1 max |ψ  | |||χ|||ε .
(9.30)
9.2 Finite Element Methods 277

Next let us bound the second and third term in (9.29). The Cauchy-Schwarz
inequality and Proposition 9.13 yield
 

 w − wI , bT ∇χ 
≤ Cw − wI 0,Ω22 ∇χ0,Ω22 + Cw − wI 0,Ω\Ω22 ∇χ0,Ω\Ω22
 
≤ C ε1/2 N −5/2 + N −3 ∇χ0,Ω22
 2
+ Cε1/2 h + N −1 max |ψ  | ∇χ0,Ω\Ω22
 2
≤ C h + N −1 max |ψ  | |||χ|||ε , (9.31)

 used an inverse inequality and that on Ω22 the mesh is uniform with
where we have
mesh size O N −1 .
 
Finally, we study the term bT ∇(v − v I ), χ . Let b1;ij = b1 (xi , yj ) for all i, j.
Using the second identity of Lemma 9.11, we get
 
b1 ∂x (v − v I ), χ
    
= b1;ij ∂x (v − v I ), χ T + (b1 − b1;ij )∂x (v − v I ), χ T
ij ij
Tij ∈Ω N
   
1  1
= b1;ij Fi2 ∂x χ − h2i χ ∂x3 v
Tij 6 12
Tij ∈Ω N
  

 1  2   

+ Gj χ − Fi ∂x χ − Gj ∂y χ − Fij ∂x ∂y χ ∂x ∂y v 2
3
−1 
1     2 
N N
+ b1;i+1,j hi+1 − b1;ij hi
2 2
χ∂x v (xi , y)dy
12 i=1 j=1 Jj
  
+ (b1 − b1;ij )∂x (v − v I ), χ T
ij
Tij ∈Ω N

=: I1 + I2 + I3 . (9.32)

Use (9.11) to obtain

   
|I1 | ≤ C h2i + kj2 χ1,Tij + hi ∂x χ1,Tij
Tij ∈Ω N
 
+ kj2 kj ∂y χ1,Tij + hi kj ∂x ∂y χ1,Tij .

Thus,
  
|I1 | ≤ C h2i + kj2 χ1,Tij ≤ Ch2 χL1 (Ω) ≤ Ch2 χ0 . (9.33)
Tij ∈Ω N
278 9 Convection-Diffusion Problems

For I2 we proceed as follows: First

 i 
 
   
χvxx (xi , y)dy = ∂x χ∂x2 v + χ∂x3 v (x, y)dxdy
Jj k=1 Jj Ik

yields


qN 
   2 
b1;i+1,j h2i+1 − b1;ij h2i χ∂x v (xi , y)dy
i=1 Jj


qN  
   
= b1;qN +1,j h2qN +1 − b1;ij h2i ∂x χ∂x2 v + χ∂x3 v (x, y)dxdy .
i=1 Jj Ii

Thus,
qN   

N
   2  
 b1;i+1,j hi+1 − b1;ij hi
2 2
χ∂x v (xi , y)dy 
i=1 j=1 Jj
 
≤ Ch ∂x χ∂x2 v + ∂x3 χv 1,Ω ∪Ω
2


11 12

≤ Ch2 ε1/2 ln1/2 N ∂x χ0 + χ0
≤ Ch2 ln1/2 N |||χ|||ε , (9.34)

by (9.11). Furthermore, for i = qN + 1, . . . , N , we have

  
   2  
b h 2
 1;i+1,j i+1 − b h
1;ij i
2
χ∂x v (xi , y)dy  ≤ Ch2i χ1,Tij ,
Jj

 
because b1;i+1,j − b1;ij  ≤ Chi+1 , hi = hi+1 ≤ h and

 2 
χ∂x v (xi , y)dy ≤ Ch−1
i χ1,Tij ,
Jj

by an inverse inequality. We get

  
 N
−1 
N
   2  
 
 b1;i+1,j hi+1 − b1;ij hi
2 2
χ∂x v (xi , y)dy  ≤ Ch2 χ0 . (9.35)
 Jj 
i=qN +1 j=1

For I3 we have the following bound:

         
|I3 | ≤ b1 − b1;ij  h i
∂x2 v  + kj
∂x ∂y vxy  χ
∞,T ∞,Tij ∞,T 1,T ij ij ij
Tij ∈ΩN

≤ Ch2 χ0 , (9.36)

by Lemma 9.12 and (9.11).

9.2 Finite Element Methods 279

Collect (9.32)–(9.36) to obtain

 
 b1 ∂x (v − v I ), χ  ≤ Ch2 ln1/2 N |||χ||| (9.37)
ε

with the analogously bound

 
 b2 ∂y (v − v I ), χ  ≤ Ch2 ln1/2 N |||χ||| . (9.38)
ε

Substituting (9.30), (9.31), (9.37) and (9.38) into (9.29), we are finished.

9.2.2.4 Maximum-Norm Error Bounds

In this section we use Theorem 9.9 and the interpolation error bounds from
Sect. 9.2.1 to obtain bounds for the error of the Galerkin method in the maxi-
mum norm.
Start with the region Ω22 , where the mesh is quasi-uniform with mesh size
O N −1 :
 I   
u − uN  ≤ CN uI − uN 0,Ω
∞,Ω 22
 22

≤ C N h ln N + N −1 max |ψ  |2 .
2 1/2

 
Thus, on a standard Shishkin mesh, where h = O N −1 , one gets
 
u − uN  ≤ CN −1 ln2 N,
∞,Ω 22

while for the Bakhvalov-Shishkin mesh we have

 
u − uN  ≤ CN −1 ln1/2 if ε ≤ CN −1 ,
∞,Ω 22

 
because for this mesh h = O max{N −1 , ε} .
Now let (xi , yj ) be any mesh node in Ω21 . Then following [152, pp. 11,12] we
obtain
 I 
(u − uN )(xi , yj )
 xi   
  τx  
=  (u − u )(x, yj )dx ≤ CN
I N ∂x (uI − uN )
0 0 Jj
 1/2  
≤ CN εN −1 ln N ∇(uI − uN )
0,[0,τx ]×Jj
 
≤ CN 1/2 ln1/2 N uI − uN ε .
280 9 Convection-Diffusion Problems

Thus,
 
u − uN ∞,Ω21 ≤ CN 1/2 ln1/2 N h2 ln1/2 N + N −2 max |ψ  |2 ,

by Theorems 9.4 and Theorem 9.9. Clearly identical bounds hold on Ω12 .
Apply this result to get bounds for particular S-type meshes:
⎧
⎪
⎨CN
−3/2
ln5/2 N for standard Shishkin meshes,
 
u − uN  ≤ CN −3/2 ln N for Bakhvalov-Shishkin meshes
∞,Ω12 ∪Ω21 ⎪
⎩
with ε ≤ CN −1 .

9.2.2.5 Gradient Recovery

Similar to Sect. 5.2.3 a gradient recovery operator can be defined for the bilinear
Galerkin FEM, that gives approximations of the gradient which are superior to those
of Theorem 9.8. We follow [139].

Notation. In this section let  · 1,D denote the standard norm in H 1 (D).

Let T be a rectangle of Ω N and let T: be the patch associated with T , consisting

of all rectangles that have a common corner with T (see Fig. 9.3). We define for
v ∈ V ω the recovered gradient Rv as follows:
 First we compute the gradient of v
at the midpoints of the mesh rectangles γ ij := ∇v(xi−1/2 , yj−1/2 ) . Then these
values are bilinearly interpolated to give the values of Rv at the mesh points of the
triangulation, viz.,

  
1
hi+1−m kj+1−n
Rv ij = αij := γ . (9.39)
h
m,n=0 i
+ hi+1 kj + kj+1 i+m,j+n

yj+1
b b
T: r yj

b b

T
yj−1

yj−2
xi−2 xi−1 xi xi+1

Fig. 9.3 Mesh rectangle T and associated patch T:

9.2 Finite Element Methods 281

Bilinear interpolation is again used to extend the recovered gradient from the mesh
nodes to the whole of Ω:
  xi − x yj − y x − xi−1 yj − y
RwN (x, y) := αi−1,j−1 + αi,j−1
hi kj hi kj
xi − x y − yj−1 x − xi−1 y − yj−1
+ αi−1,j + αi,j
hi kj hi kj
for (x, y) ∈ Tij , i, j = 2, . . . , N − 1.

For the boundary rectangles, we simply extrapolate the well-defined bilinear func-
tion of the adjacent rectangles.
Lemma 9.14. R : V ω → V ω × V ω is a linear operator with the following
properties:

(locality) Rv on T depends only on values of v on the patch T:,

   
(stability) Rv  ≤ C v 1,∞,T: for all v ∈ V ω , (9.40a)
∞,T
   
Rv  ≤ C v 1,T: for all v ∈ V ω , (9.40b)
0,T

(consistency) Rv I = ∇v on T for all v that are quadratic on T:. (9.40c)

Proof. The first three properties are immediate consequences of the definition of R,
while (9.40c) is verified by a Taylor expansion of v.
Now, given any continuous function v on T:, we denote by Qv that quadratic
function on T: with
 
Qv (Pk ) = v(Pk ) for k = 1, . . . , 6 (see Fig. 9.4).

This set of degrees of freedom is unisolvent and thus our Lagrange interpolant Qv
is well defined.
The decomposition (9.11) and a careful analysis yield the following bounds for
quadratic interpolation.

P1 r Pr2 rP3

r P6

r r
P4 P5

Fig. 9.4 Definition of the quadratic interpolant on the patch T:

282 9 Convection-Diffusion Problems

Lemma 9.15. Let ω̄ = ω̄x × ω̄y be a tensor-product S-type mesh with σ ≥ 3 that
satisfies (2.8). Assume that the solution u of (9.1) can be decomposed as in (9.11).
Then
  
  C h + N −1
max |ψ  2
| for T ⊂ Ω \ Ω22 ,
εu − Qu1,∞,T: ≤
CN −2 for T ⊂ Ω22 ,
  3
  C h + N −1 max |ψ  | for T ⊂ Ω \ Ω22 ,
u − Qu : ≤
∞,T −3
CN for T ⊂ Ω22

and
  2

1/2 
 C(meas T̃ )1/2 h + N −1 max |ψ  | for T ⊂ Ω \ Ω22 ,
ε u − Qu1,T: ≤
−3
CN for T ⊂ Ω22 .

We would like to estimate the difference between the gradient and the recovered
gradient in the ε-weighted H 1 seminorm. We start from
     
ε1/2 ∇u − RuN 0 ≤ ε1/2 ∇u − RuI 0 + ε1/2 R(uI − uN )0 , (9.41)

by a triangle inequality. For the second term in (9.41), the stability property (9.40b)
of the recovery operator and the superconvergence result of Theorem 9.9 yield
   2
ε1/2 R(uI − uN )0 ≤ C h + N −1 max |ψ  | ln1/2 N. (9.42)

In the next result we estimate the first term in (9.41).

Proposition 9.16. Let ω̄ = ω̄x ×ω̄y be a tensor-product S-type mesh with σ ≥ 3 that
satisfies (2.8). Assume that the solution u of (9.1) can be decomposed as in (9.11)
and that

min {hqN , kqN } ≥ CεN −1 . (9.43)

Then
   2
ε1/2 ∇u − RuI 0 ≤ C h + N −1 max |ψ  | ln1/2 N.

Proof. For any T ∈ Ω N , the consistency property (9.40c) of the recovery operator
yields
     
∇u − RuI  ≤ ∇(u − Qu) +R(u − Qu)I  , (9.44)
0,T 0,T 0,T
9.2 Finite Element Methods 283

since R(Qu)I = ∇Qu. For the interpolation operator we can use the stability
estimates
 I  I  
v  ≤ Cv∞,T and v  ≤ C v 1,∞,T .
∞,T 1,∞,T

To estimate the second term in (9.44), we bound the L2 norm by the L∞ norm and
apply the stability property (9.40a) of the recovery operator:
   
R(u − Qu)I  ≤ (meas T )1/2 R(u − Qu)I ∞,T
0,T
 
≤ C(meas T )1/2 (u − Qu)I  :.
1,∞,T
(9.45)

Thus, for T ∈
/ Ω22 we have
   2
R(u − Qu)I  ≤ Cε−1 (meas T )1/2 h + N −1 max |ψ  | , (9.46)
0,T

by Lemma 9.15.
Next we consider T ∈ Ω22 . We apply an inverse inequality and the L∞ stability
of bilinear interpolation to (9.45) to get
   −1  
R(u − Qu)I  ≤ CN −1 minT: h u − u∗  : .
0,T ∞,T

 
If minT: h = O N −1 then
   
R(u − Qu)I  ≤ C u − Qu∞,T: ≤ CN −3 , (9.47)
0,T

by Lemma 9.15. Otherwise — for the elements T along the transition from the fine
to the coarse mesh — we have to estimate more carefully:

     (meas T )1/2  
R(u − Qu)I  ≤ (u − Qu)I  : ≤ u − Qu .
0,T 1,T minT h ∞,T
T ∈T:

From (9.43), we have

   
ε1/2 R(u − Qu)I 0,T ≤ C u − Qu∞,T: ≤ CN −3 , (9.48)

by Lemma 9.15. Combining (9.46), (9.47) and (9.48), we have

 
ε1/2 R(u − Qu)I 0,T
  2
Cε−1/2 (meas T )1/2 h + N −1 max |ψ  | for T ⊂ Ω \ Ω22 ,
≤
CN −3 for T ⊂ Ω22 .
284 9 Convection-Diffusion Problems

We use the last estimate of Lemma 9.15 and (9.44) to obtain

 
ε1/2 ∇u − RuI 0,T
  2
Cε−1/2 (meas T )1/2 h + N −1 max |ψ  | for T ⊂ Ω \ Ω22 ,
≤
CN −3 for T ⊂ Ω22 .

Recalling that
    
∇u − RuI 2 = ∇u − RuI 2
0 0,T
T ∈Ω N

 
and meas Ω \ Ω22 = O (ε ln N ), the proof is complete. 


Remark 9.17. The condition (9.43) is satisfied if for example, ϕ̃ in Sect. 2.1.3 is
bounded from below by a positive constant independently of ε and N . Both the
original Shishkin mesh and the Bakhvalov-Shishkin mesh satisfy this condition. ♣

As a consequence of (9.41), (9.42) and Proposition 9.16 we have the following

result:
Theorem 9.18. Let ω̄ = ω̄x ×ω̄y be a tensor-product S-type mesh with σ ≥ 5/2 that
satisfies (2.8) and (9.43). Assume that the solution u of (9.1) can be decomposed as
in (9.11). Then
   2
ε1/2 ∇u − RuN 0 ≤ C h + N −1 max |ψ  | ln1/2 N.

Remark 9.19. Similar to the one-dimensional case, using RuN instead of ∇uN , we
get an asymptotically exact
 error estimator
 for the weighted H 1 -seminorm of the
1/2  N 
finite element error ε ∇(u − u ) 0 on S-type meshes. ♣

9.2.2.6 Numerical Tests

Let us verify our theoretical results for the Galerkin FEM using bilinear trial and test
functions on S-type meshes when applied to the test problem (9.3). In our computa-
tions we have chosen ε = 10−8 and σ = 3 for themeshes. In  the tables we compare
both the error in the ε-weighted energy norm u − uN  with the error in the
  ε
discrete energy norm uI − uN ε , and the accuracy of the gradient approxima-
tion ∇uN with that of the recovered gradient approximation RuN . The errors are
estimated using a 4th -order Gauß-Legendre formula on each mesh rectangle. The
rates of convergence are computed in the usual way. Tables 9.3 and 9.4 are clear
illustrations of Theorems 9.8, 9.9 and 9.18.
9.2 Finite Element Methods 285

Table 9.3 Shishkin mesh

   I     
u − uN  u − uN  ε1/2  ∇u − ∇uN  0 ε1/2  ∇u − RuN  0
ε ε
N error rate error rate error rate error rate
16 2.6900e-1 0.63 5.2110e-2 1.25 2.6898e-1 0.63 9.5425e-1 2.86
32 1.7359e-1 0.72 2.1896e-2 1.43 1.7359e-1 0.72 1.3141e-1 1.81
64 1.0556e-1 0.77 8.1467e-3 1.53 1.0556e-1 0.77 3.7507e-2 1.48
128 6.1881e-2 0.80 2.8137e-3 1.60 6.1881e-2 0.80 1.3479e-2 1.56
256 3.5421e-2 0.83 9.2543e-4 1.65 3.5421e-2 0.83 4.5685e-3 1.64
512 1.9936e-2 0.85 2.9398e-4 1.69 1.9936e-2 0.85 1.4687e-3 1.69
1024 1.1078e-2 — 9.0961e-5 — 1.1078e-2 — 4.5612e-4 —

Table 9.4 Bakhvalov-Shishkin mesh

   I     
u − uN  u − uN  ε1/2  ∇u − ∇uN  0 ε1/2  ∇u − RuN  0
ε ε
N error rate error rate error rate error rate
16 1.2475e-1 1.00 7.9084e-3 2.00 1.2471e-1 1.00 5.0012e-1 3.43
32 6.2574e-2 1.00 1.9800e-3 2.00 6.2569e-2 1.00 4.6315e-2 3.09
64 3.1312e-2 1.00 4.9620e-4 2.00 3.1311e-2 1.00 5.4227e-3 2.43
128 1.5659e-2 1.00 1.2425e-4 2.00 1.5659e-2 1.00 1.0044e-3 2.08
256 7.8298e-3 1.00 3.1096e-5 2.00 7.8298e-3 1.00 2.3690e-4 2.01
512 3.9149e-3 1.00 7.7789e-6 2.00 3.9149e-3 1.00 5.8638e-5 2.00
1024 1.9575e-3 — 1.9460e-6 — 1.9575e-3 — 1.4624e-5 —

9.2.3 Artificial Viscosity Stabilisation

In Sect. 5.3 we studied a FEM with artificial viscosity stabilisation in one dimension.
It can be generalised to two dimensions as follows: Set

h̄ := diag(h̄, k̄) with h̄ := hi in Ii × (0, 1) and k̄ := kj in (0, 1) × Jj

and let κ ≥ 0 be an arbitrary constant. Then we add artificial viscosity of order

κh̄ in x-direction and of order κk̄ in y-direction, i. e., we consider the following
discretisation. Find uN ∈ V0ω such that
 
aκ (uN , v) := a(uN , v) + κ h̄∇uN , ∇v = (f, v) for all v ∈ V0ω .

The norm naturally associated with aκ (·, ·) is

 1/2
2
|||v|||κ := |||v|||ε + κ (h̄∇v, ∇v) ≥ |||v|||ε , for all v ∈ H 1 (Ω).

The bilinear form aκ (·, ·) is coercive with respect to this norm, because
2
aκ (v, v) ≥ |||v|||κ for all v ∈ H01 (Ω). (9.49)

In our analysis we follow Schneider et al. [148], but refine it by explicitly moni-
toring the dependence on κ. Let again η = uI − u denote the interpolation error and
286 9 Convection-Diffusion Problems

χ = uI − uN the difference between interpolated and exact solution. Because of

the added artificial viscosity the discretisation does not satisfy the Galerkin orthog-
onality condition, but we have

aκ (χ, χ) = a(η, χ) + κ(h̄∇η, ∇χ) + κ(h̄∇u, ∇χ) . (9.50)

(i) For the first term we have two bounds from Sections 9.2.2.1 and 9.2.2.2:

|a(η, χ)|
⎧
⎪ −1 
⎨h + N max |ψ | for general linear and,
≤ C |||χ|||ε bilinear elements, (9.51)
⎪
⎩ 2 1/2
h ln N + N −2 max |ψ  |2 for bilinear elements.
 
(ii) Next we bound κ h̄∇η, ∇χ . Let Tij be arbitrary. Then
    
 
h̄ηx , χx T = hi ∂x η∂x χ = ∂x η ∂x χ.
ij
Jj Ii Jj Ii Ii

Thus,
      

 h̄∂x η, ∂x χ Tij  ≤ 2η ∞,Tij ∂x χ1,Tij .

Consequently, we have
 
 h̄∂x η, ∂x χ 
   2  1/2  
≤ C N −2 ∂x χ0,Ω + N −1 max |ψ  | ε ln N ∂x χ
Ω\Ω
22 22

−1 
≤ CN max |ψ | ln 1/2
N |||χ|||ε ,

by
 an inverse  inequality and (9.5). An analogous estimate holds for
 k̄∂y η, ∂y χ . Hence,

 
κ  h̄∇η, ∇χ  ≤ CκN −1 max |ψ  | ln1/2 N |||χ|||ε . (9.52)
 
 ∇χ has to be considered.
(iii) Finally, h̄∇u,  We restrict ourselves to bounding
h̄∂x u, ∂x χ since the term k̄∂y u, ∂y χ can be treated analogously. Using the
decomposition of Theorem 7.17, we get
 
h̄∂x u, ∂x χx
   
= h̄∂x (v + w2 ), ∂x χ Ω ∪Ω + h̄∂x (v + w2 ), ∂x χ Ω ∪Ω
 
11 12
 21
 22
+ h̄∂x (w1 + w12 ), ∂x χ Ω ∪Ω + h̄∂x (w1 + w12 ), ∂x χ Ω ∪Ω .
11 12 21 22
(9.53)
9.2 Finite Element Methods 287

The Cauchy-Schwarz inequality and Theorem 7.17 yield

  
 
 h̄∂x (v + w2 ), ∂x χ Ω11 ∪Ω12 
 1/2   (9.54)
≤ Ch ε ln N ∂x χ ≤ Ch ln1/2 N |||χ|||ε .
Ω 11 ∪Ω12

On Ω21 ∪ Ω22 we have

 
h̄∂x (v + w2 ), ∂x χ Ω ∪Ω
21 22
 1 1
=H ∂x (v + w2 )∂x χdxdy
0 τx
 1  
 1
= −H ∂x (v + w2 )χ (τx , y) + ∂x2 (v + w2 )χdx dy.
0 τx

Thus,
    
  
1
 h̄∂x (v + w2 ), ∂x χ Ω21 ∪Ω22  ≤ CN −1 χ0 + |χ(τx , y)| dy . (9.55)
0

Note that
   
1 1  τx 
|χ(τx , y)| dy =  ∂x χdx dy ≤ ∂x χ1,Ω11 ∪Ω12 ≤ C ln1/2 N |||χ|||ε .

0 0 0

We apply this inequality to (9.55) to obtain

 
(h̄∂x (v + w2 ), ∂x χ)  ≤ CN −1 ln1/2 N |||χ||| . (9.56)
Ω21 ∪Ω22 ε

Now we bound the last two terms in (9.53). Using Theorem 7.17 we get, for any
Tij ∈ Ω N ,
  

 h̄∂x (w1 + w12 ), ∂x χ Tij 
   
 
≤C −1 −β1 x/ε
ε e dx  
∂x χ dx dy.
Jj Ii Ii

This implies that

  

 h̄∂x (w1 + w12 ), ∂x χ Tij 
⎧  
⎨CN −1 max |ψ  |∂x χ for Tij ⊂ Ω11
N
∪ Ω12
N
,
≤   1,Tij
⎩CN −2 ∂x χ for Tij ⊂ Ω21
N
∪ Ω22
N
.
1,Tij
288 9 Convection-Diffusion Problems

Therefore,
  
 
 h̄∂x (w1 + w12 ), ∂x χ Ω11 ∪Ω12 
 
≤ CN −1 max |ψ  |χx 1,Ω ≤ CN −1 max |ψ  | ln1/2 N |||χ|||ε
11 ∪Ω12

and
    
 
 h̄∂x (w1 + w12 ), ∂x χ Ω21 ∪Ω22  ≤ CN −1 χ0 ,

by an inverse inequality.
Combine the last two bounds with (9.53), (9.54) and (9.56) to get
 
 h̄∂x u, ∂x χ  ≤ CN −1 max |ψ  | ln1/2 N |||χ||| .
ε

 
With an analogous estimate for k̄∂y u, ∂y χ we have

κ |(h̄∇u, ∇χ)| ≤ CκN −1 max |ψ  | ln1/2 N |||χ|||ε . (9.57)

Finally, combine (9.49)–(9.52) and (9.57) in order to obtain the main result of
this section.
Theorem 9.20. Let ω̄ := ω̄x × ω̄y be a tensor-product S-type mesh with σ ≥ 2 that
satisfies (2.8). Then the upwind-FEM solution uN satisfies
 I   
u − uN  ≤ C 1 + κ ln1/2 N N −1 max |ψ  |
κ

and, for bilinear elements and σ ≥ 5/2,

 I 
u − uN  ≤ C κN −1 max |ψ  | ln1/2 N + h2 ln1/2 N + N −2 max |ψ  |2 .
κ

A consequence of Theorem 9.20 and Sect. 9.2.1 is the following bound of the
error in the ε-weighted energy norm:
   
u − uN  ≤ C h + N −1 max |ψ  | ln1/2 N .
ε

Remark 9.21. The supercloseness property of the Galerkin FEM with bilinear ele-
ments is not affected if we take κ = O N −1 . However, for the efficient treatment
of the discrete systems, the choice κ = O (1) is more appropriate which then results
in a loss of the supercloseness property. ♣
Remark 9.22. The |||·|||κ bounds imply that the method gives uniform convergent
approximations of the gradient on the coarse mesh region Ω22 . For example, for a
Shishkin mesh, where max |ψ  | ≤ C ln N and h ≤ 2N −1 , we have
  
κ1/2 N −1/2 ∇ uI − uN 0,Ω ≤ C κN −1 ln3/2 +N −2 ln2 N .
22
9.2 Finite Element Methods 289

Thus,

  I  CN −1/2 ln3/2 if κ = O (1) ,
∇ u − uN  ≤  
0,Ω 22
CN −1 ln2 if κ = O N −1 .

Note that in contrast to the streamline-diffusion FEM, we have full control of the
gradient, while for SDFEM one has uniform bounds for the streamline derivative
b · ∇(uI − uN )0,Ω22 only; see Sect. 9.2.4. ♣

Remark 9.23. Suboptimal maximum-norm error bounds on Ω22 can be obtained by

application of the discrete Sobolev inequality

χ∞,Ω22 ≤ C ln1/2 N ∇χ0,Ω22 , (9.58)

that holds true for piecewise-polynomial functions χ that vanish on a part of the
boundary of finite length, see [160, Lemma 5.4] or [63]. We get

  CN −1/2 ln2 N if κ = O (1) ,
u − u 
N
≤  
∞,Ω22
CN −1 ln5/2 N if κ = O N −1 .

Bounds for the maximum-norm error on Ω21 ∪ Ω12 can be obtained using the tech-
nique from Sect. 9.2.2.4. ♣

9.2.4 Streamline-Diffusion FEM

Introduced by Hughes and Brooks [54], this method is the most commonly used
stabilised FEM for the discretisation of convection-diffusion and related problems.
Starting from the weak formulation (9.4), we add weighted residuals in order to
stabilise the method. Then the SDFEM reads: Find uN ∈ V0ω such that

aSD (uN , v) = a(uN , v) + astab (uN , v) = fSD (v) for all v ∈ V0ω

with

astab (uN , v) := δT (LuN , −b · ∇v)T
T ∈Ω N

and

fSD (v) := f (v) + δT (f, −b · ∇v)T
T ∈Ω N
290 9 Convection-Diffusion Problems

and user chosen stabilisation parameters δT ≥ 0. We clearly have the Galerkin

orthogonality property

aSD (u − uN , v) = 0 for all v ∈ V0ω . (9.59)

Let V0ω be our finite element space consisting of piecewise (bi)linear functions
that vanish on ∂Ω. It is shown in, e.g., [141, §III.3.2.1], that if

0 ≤ δT ≤ γc−2
∞,T for all T ∈ Ω N , (9.60)

then
1 2
aSD (v, v) ≥ |||v|||SD for all v ∈ V0ω , (9.61)
2
with the streamline-diffusion norm

2 2

|||v|||SD := |||v|||ε + δT (b · ∇v, b · ∇v)T .
T ∈Ω N

9.2.4.1 Convergence in the Streamline-Diffusion Norm

Stynes and Tobiska [155] analyse the SDFEM using piecewise bilinear finite ele-
ments on standard Shishkin meshes for problems with regular layers. Here we shall
extend the technique from [155] to our more general class of S-type meshes, but
still consider piecewise bilinear test and trial functions.
Partition the domain Ω̄ as in Fig. 9.2. We follow standard recommendations [141,
p. 307] and set

δ if T ⊂ Ω22 ,
δT := (9.62a)
0 otherwise,
and

δ0 N −1 if ε ≤ N −1 ,
δ := (9.62b)
δ1 ε−1 N −2 otherwise.

with positive constants δ0 and δ1 . Clearly δ ≤ max {δ0 , δ1 } N −1 and there-

fore (9.60) is satisfied for N sufficiently large, independent of ε.
Note that in the layer regions Ω\Ω22 , the stabilisation is switched off because
there the streamline-diffusion stabilisation would be negligible compared to the nat-
ural stability induced by the discretisation of the diffusion term.
Our error analysis again starts from the coercivity (9.61) and the Galerkin or-
thogonality (9.59). Let again η = uI − u and χ = uI − uN . Then
1 2
|||χ|||SD ≤ a(η, χ) + astab (η, χ).
2
9.2 Finite Element Methods 291

For the first term we have

 
|a(η, χ)| ≤ C h2 ln1/2 N + N −2 max |ψ  |2 |||χ|||ε ,

see Sect. 9.2.2.3, while the stabilisation term


astab (η, χ) = δ (εΔu + b · ∇η − cη, b · ∇χ)T
T ⊂Ω22

still has to be analysed. This was done in [155]. Using (9.14b) as a crucial ingredient,
Stynes and Tobiska derive the bound

|astab (η, χ)| ≤ CN −2 ln1/2 N |||χ|||SD .

Eventually we get the following convergence results.

Theorem 9.24. Let ωx × ωy be a tensor-product S-type mesh with σ ≥ 5/2 that
satisfies (2.8). Then the SDFEM solution uN satisfies
 I   
u − uN  ≤ C h 2 1/2
ln N + N −2
max |ψ  2
|
SD

Remark 9.25. Theorems 9.4 and 9.24 give

   
u − uN  ≤ C h + N −1 max |ψ  | .
ε

Thus, Theorem 9.24 is a supercloseness result for the SDFEM.

Furthermore,
   
ε ∇u − RuN 0 ≤ C h2 + N −2 max |ψ  |2 ln1/2 N,

where R is the recovery operator from Sect. 9.2.2.5. ♣

9.2.4.2 Maximum-Norm Error Bounds

Clearly the technique for the Galerkin FEM from Sect. 9.2.2.4 can be applied to
give pointwise error bounds for the SDFEM with bilinear test and trial functions
within the layer regions Ω12 and Ω21 , while on the coarse mesh region Ω22 , we can
employ (9.58). We get
⎧ −3/2
⎪
⎨ CN ln5/2 N for standard Shishkin meshes
u − uN ∞,Ω\Ω11 ≤ CN −3/2 ln N for Bakhvalov-Shishkin meshes
⎪
⎩
with ε ≤ CN −1 .
(9.63)
292 9 Convection-Diffusion Problems

Adapting Niijima’s technique [125], Linß & Stynes [110] study the SDFEM with
piecewise linear test and trial functions on Shishkin meshes. For technical reasons
a modified version of the SDFEM with artificial crosswind diffusion added on Ω22
is studied. Furthermore, it is assumed that the convective field b is constant. The
method reads as follows. Find uN ∈ V0ω such that

aSD (uN , v) + (ε∗ b⊥ · ∇uN , b⊥ · ∇v) = fSD (v) for all v ∈ V0ω

with

   !
1 −b2 max 0, N −3/2 − ε on Ω22 ,
b⊥ := and ε∗ :=
b b1 0 otherwise.

If ε ≤ N −3/2 , then for any point (x, y) ∈ Ω the analysis in [110] yields
⎧ −1/2
⎪
⎪ CN ln3/2 N if (x, y) ∈ Ω22 ,
  ⎨
(u − uN )(x, y) ≤ CN −3/4 ln3/2 N if (x, y) ∈ Ω \ Ω22 ,
⎪
⎪
⎩
CN −11/8 ln1/2 N if (x, y) ∈ (λ∗ , 1)2 ,

 
where λ∗ = O N −3/4 ln N . The analysis in [110] includes more detailed results
and also deals with the case ε ≥ N −3/2 . Numerical experiments in [109] show
convergence of almost second order on the coarse part of the mesh, while inside
the boundary layers, the rates are smaller than one. For bilinear elements, almost
second-order convergence in the maximum norm is observed globally, but no rigor-
ous analysis is yet available.

9.2.4.3 A Numerical Example

Let us verify the theoretical results when the SDFEM is applied to our test prob-
lem (9.3). In the computations we have chosen ε = 10−8 and σ=  3. 
The tables display the error in the ε-weighted energy norm u − uN ε , in the
 
discrete SD-norm uI − uN SD and in the maximum-norm. Tables 9.5 and 9.6
clearly illustrate Theorem 9.24, while for the maximum-norm errors (9.63) appears
to be suboptimal: instead of convergence of order (almost) 3/2 we observe (almost)
2nd order.

9.2.4.4 Higher-Order Elements

In [156] Stynes and Tobiska study the SDFEM with Qp elements, p > 1 on tensor-
product Shishkin meshes.
9.2 Finite Element Methods 293

Table 9.5 The SDFEM on a Shishkin mesh

   I   
u − uN  u − uN   u − uN 
ε SD ∞
N error rate error rate error rate
16 3.3542e-1 0.75 2.0654e-1 1.04 1.7673e-1 1.14
32 1.9932e-1 0.82 1.0021e-1 1.33 8.0261e-2 1.41
64 1.1259e-1 0.83 3.9957e-2 1.50 3.0251e-2 1.51
128 6.3418e-2 0.83 1.4151e-2 1.59 1.0635e-2 1.61
256 3.5718e-2 0.84 4.6849e-3 1.65 3.4956e-3 1.66
512 1.9989e-2 0.85 1.4886e-3 1.69 1.1063e-3 1.70
1024 1.1087e-2 — 4.5993e-4 — 3.4131e-4 —

Table 9.6 The SDFEM on a Bakhvalov-Shishkin mesh

   I   
u − uN  u − uN   u − uN 
ε SD ∞
N error rate error rate error rate
16 1.3415e-1 1.07 4.9909e-2 1.92 5.1204e-2 1.89
32 6.3934e-2 1.02 1.3161e-2 1.98 1.3793e-2 1.96
64 3.1488e-2 1.01 3.3354e-3 2.00 3.5346e-3 1.99
128 1.5681e-2 1.00 8.3621e-4 2.00 8.8983e-4 2.00
256 7.8326e-3 1.00 2.0910e-4 2.00 2.2291e-4 2.00
512 3.9153e-3 1.00 5.2263e-5 2.00 5.5756e-5 2.00
1024 1.9575e-3 — 1.3063e-5 — 1.3940e-5 —

The transition points in the Shishkin mesh are

   
(p + 1)ε (p + 1)ε
τx := min q, ln N and τy := min q, ln N ,
β1 β2

otherwise the construction of the mesh is unchanged. The stabilisation parameters

δT are chosen as in (9.62).
We introduce a special vertices-edges-element interpolant [79] as follows. Let v
be a given function. On each element T ∈ Ω N the interpolant Iv is defined by

(Iv)(xi ) = v(xi ) for i = 1, . . . , 4,

 
(Iv)ϕ = vϕ for all ϕ ∈ Pp−2 (i ) and for i = 1, . . . , 4,
i i

and
 
(Iv)ψ = vψ for all ψ ∈ Qp−2 (T ),
T T

where xi are the four vertices of T and i are the four edges of T . Pk (i ) is the space
of polynomials of degree at most k in the single variable whose axis is parallel to
the edge i .
294 9 Convection-Diffusion Problems

In [156] the authors proceed by deriving interpolation error bounds for Iu on

anisotropic meshes. For example, the pointwise interpolation error is shown to
satisfy

CN −(p+1) if x ∈ Ω̄22 ,
|(u − Iu)(x)| ≤  p+1
C N −1 ln N otherwise.

However, the main result in [156] is

 N 
u − Iu ≤ CN −(p+1/2) .
SD

This is a supercloseness result, because in general for Qp elements, one can expect
at best
 N   
u − u ≤ C N −1 ln N p .
ε

Postprocessing can be used to obtain an approximation of u of order p + 1/2.

Remark 9.26. Matthies [119] considers a different approach to stabilised FEM: lo-
cal projection stabilisation on Shishkin-type meshes. Using Qp elements inside the
layers and enriched Qp elements in the region where the mesh is coarse, he proves
convergence and supercloseness results that resemble those by Stynes and Tobiska
for SDFEM. ♣

9.2.5 Characteristic Layers

We now consider (9.1) with parabolic layers. This is, we assume the convective field
is b = (b, 0) and seek a solution u to

−εΔu − bux + cu = f in Ω = (0, 1)2 , u = 0 on ∂Ω (9.64)

with b ≥ β > 0 and c ≥ 0 on Ω̄. The main contributions here are by Franz et al.
[38–41].
Analytical properties of this problem have been studied in Sect. 7.3.2. There will
be an exponential layer at x = 0 and parabolic layers along the boundaries y = 0
and y = 1.
When discretising (9.64), we use tensor-product Shishkin-type meshes with N
mesh intervals in each coordinate direction. The mesh in x-direction is a mesh for
one-dimensional convection-diffusion equations (Sect. 2.1.3), while the mesh in
y-direction is a mesh for a reaction-diffusion problem (Sect. 2.2). The transition
parameters for these meshes are
 
σε q √
λx := min q, ln N and λy := min , σ ε ln N
β 2
9.2 Finite Element Methods 295

with the mesh parameters σ > 0 and q ∈ (0, 1/2). A plot of the resulting mesh
is displayed in Fig. 2.21 on p. 29. The mesh characterising function is again de-
noted by ψ.

Interpolation error
Adapting the technique from Sect. 9.2.1, we obtain for all mesh rectangles Tij

   
u − uI  ≤C 1 + ε−1 e−βx/2ε dx
∞,T ij
Ii
 2
 √ 
−1/2 −y/2 ε
+ 1+ε e dy
Jj

and
  
  
u − uI  ≤ C max 1 + ε−1 e−βx/2ε dx
∞ i=1,...,N Ii
 2
 √ 
−1/2 −y/2 ε
+ max 1+ε e dy .
j=1,...,N Jj

Bounds for particular meshes (Bakhvalov or Shishkin meshes) can immediately be

concluded as has been done before.
Energy-norm bounds. The L2 part of the norm is easily bounded by the maximum
norm. Therefore, we consider the H1 part only. Reasoning as in Sect. 9.2.1 and using
Theorem 7.20, we get
    1/2
ε1/2 ∂x u − uI 0 ≤ C u − uI ∞

and
    1/2
ε1/2 ∂y u − uI 0 ≤ Cε1/4 u − uI ∞ .

The last estimate highlights a problem of the energy norm applied to problems with
characteristic layers: It fails to capture these layers. Nonetheless, it is the natural
norm associated with the weak formulation of (9.64).

Galerkin FEM
In [38, 39] for finite elements with piecewise bilinear test and trial functions the
following bounds in the ε-weighted energy norm are given.
296 9 Convection-Diffusion Problems

Theorem 9.27. Let uN be the piecewise bilinear Galerkin approximation on a

S-type mesh with σ ≥ 5/2 then
 I    2
u − uN  ≤ C h + N −1 ln1/4 N + k + N −1 max |ψ  |
ε

and
   
u − uN  ≤ C ε1/4 h + k + N −1 max |ψ  | ,
ε

where h and k are the maximum step sizes in x- and y-direction, resp.

Remark 9.28. The first bound is a supercloseness result and allows for postprocess-
ing that gives higher-order accurate approximations of the gradient. ♣

Streamline-diffusion FEM
The choice of S-type meshes naturally divides the domain Ω into four (six) subre-
gions, see Fig. 9.5. Ω12 covers the exponential layer, Ω21 the parabolic layer and
Ω22 the corner layer and Ω11 the remaining region which does not have layers. On
each of the four subdomains we allow the streamline-diffusion parameter δ to take
different values: δij on Ωij .

Theorem 9.29. Let uN be the piecewise bilinear streamline-diffusion approxima-

tion on a S-type mesh with σ ≥ 5/2. Suppose the stabilisation parameters satisfy
 2
δ12 ≤ δ0 ε h + N −1 max |ψ  | , δ21 ≤ δ0 ε−1/4 N −2 , δ22 ≤ δ0 ε3/4 N −2

and
!
δ11 ≤ δ0 min N −1 , ε−1 N −2

Ω22 Ω21

Ω11 := [λx , 1] × [λy , 1 − λy ]

Ω12 Ω11 Ω12 := [0, λx ] × [λy , 1 − λy ]
Ω21 := [λx , 1] × ([0, λy ] ∪ [1 − λy , 1])
Ω22 := [0, λx ] × ([0, λy ] ∪ [1 − λy , 1])

Ω22 Ω21

Fig. 9.5 Dissection of Ω

9.3 Finite Volume Methods 297

with a constant δ0 that is independent of ε. Then

 I    1/4 2

u − uN  ≤ C h + N −1
ln N + k + N −1
max |ψ 
|
SD

and
   
u − uN  ≤ C ε1/4 h + k + N −1 max |ψ  | ,
ε

where h and k are the maximum step sizes in x- and y-direction, resp.

Proof. See [38, 40]. 



Remark 9.30. The amount of stabilisation inside the exponential layer Ω12 and in
the corner Ω22 is negligible and can be switched off. On the coarse-mesh region Ω11
the stabilisation is standard. However, inside the characteristic layer, i.e. on Ω21 the
negative power of ε in the upper bound for δ21 is surprising.
This is essentially due to the aforementioned weakness of the SD-norm and the
ε-weighted energy norm, which fail to capture the parabolic layer. For the  w2
 term
in the decomposition of Theorem 7.20 one has |||w2 |||ε = |||w2 |||SD = O ε1/4 .
An alternative, though heuristic
  approach, in [94] using residual free bubbles
suggests the choice δ22 = O N −2 . ♣

Local projection stabilisation

LPFEM for problems with characteristic layers is studied by Franz and Matthies
[41]. The results are similar to those for streamline-diffusion stabilisation.

9.3 Finite Volume Methods

In this section we consider an inverse-monotone finite volume discretisation for

problem (7.8). This scheme was introduced by Baba and Tabata [17] and later gen-
eralised by Angermann [13, 14]. For a detailed derivation of the method, the reader
is referred to [13, 62].
When working on arbitrary partitions we follow Angermann [13]. Further sta-
bility results for tensor-product meshes and uniform convergence of the method in
both a discrete energy norm and in the maximum norm are due to other authors.
References will be given when appropriate.
For the moment let Ω ⊂ IR2 be an arbitrary domain with polygonal boundary.
Consider the problem

−εΔu − b∇u + cu = f in Ω, u = 0 on Γ = ∂Ω (9.65)

298 9 Convection-Diffusion Problems

with 0 < ε  1 and

1
c+ div b ≥ γ > 0 on Ω. (9.66)
2

Let ω̄ = {xi } ⊂ Ω̄ be a set of mesh points. Let Λ and ∂Λ be the sets of indices
of interior and boundary mesh points, i.e.,

Λ := {i : xi ∈ Ω} and ∂Λ := {i : xi ∈ ∂Ω}.

Set Λ̄ := Λ ∪ ∂Λ. Partition the domain Ω into subdomains

!
Ωi := x ∈ Ω : x − xi  < x − xj  for all j ∈ Λ̄ with i = j for i ∈ Λ̄,

where · is the Euclidean norm in IR2 . We define Γij = ∂Ωi ∩∂Ωj and we say that
two mesh nodes xi = xj are adjacent if and only if mij := meas1D Γij = 0. By Λi
we denote the set of indices of all mesh nodes that are adjacent to xi . Furthermore,
set dij := xi − xj  and mi = meas2D Ωi . We denote by nij the outward normal
on the boundary part Γij of Ωi . Let h, the mesh size, be the maximal distance
between two adjacent mesh nodes. Set Nij := −nij · b ((xi + xj )/2); see Fig. 9.6.
For;a reasonable discretisation of the boundary conditions, we shall assume that
Γ ⊂ i∈∂Λ Ω̄i .

Ωj
xj

nij

Γij

xi

Ωi

Fig. 9.6 Mesh cell of the FVM

9.3 Finite Volume Methods 299

Then our discretisation of (9.65) is as follows. Find uN : Λ̄ → IR such that

[Lρ uN ]i = fi mi for i ∈ Λ, uN
i =0 for i ∈ ∂Λ, (9.67a)

with uN
i being the numerical approximation of u(xi ),

  
ε
[Lρ v]i := mij − Nij ρij (vi − vj ) + ci mi vi , (9.67b)
dij
j∈Λi

ρij = ρ(Nij dij /ε) and a function ρ : IR → [0, 1]. Possible choices for ρ are given
in Sect. 5.4 which studies the one-dimensional version of the FVM. Again we shall
assume that ρ satisfies

(ρ0 ) t → tρ(t) is Lipschitz continuous,

' (
(ρ1 ) ρ(t) + ρ(−t) − 1 t = 0 for all t ∈ IR,
' (
(ρ2 ) 1/2 − ρ(t) t ≥ 0 for all t ∈ IR,
(ρ3 ) 1 − tρ(t) ≥ 0 for all t ∈ IR.

Note that the constant choice ρ ≡ 12 , which generates a generalised central dif-
ference scheme, satisfies conditions (ρ1 ) and (ρ2 ), but not (ρ3 ). Conditions (ρ1 )
and (ρ2 ) guarantee the coercivity of the weak formulation associated with (9.67),
while (ρ3 ) ensures the inverse monotonicity of the scheme when the coefficient c is
strictly positive.

9.3.1 Coercivity of the Method

The FVM can be written in variational form: Find

!
uN ∈ V0ω := v ∈ IRcard Λ̄ : vk = 0 for k ∈ ∂Λ

such that

aρ (uN , v) = fρ (v) for all v ∈ V0ω ,

with
 
aρ (w, v) := [Lρ w]i vi and fρ (v) := fi mi vi .
i∈Λ i∈Λ̄
300 9 Convection-Diffusion Problems

When studying the coercivity of the scheme we split the bilinear form into three
parts representing the diffusion, convection and reaction terms:

aρ (w, v) = εdρ (w, v) + cρ (w, v) + rρ (w, v)

with
  mij
dρ (w, v) = (wi − wj )vi ,
dij
i∈Λ̄ j∈Λi

cρ (w, v) = − mij Nij ρij (wi − wj ) vi
i∈Λ̄ j∈Λi

and

rρ (w, v) = ci mi wi vi .
i∈Λ̄

These three terms will be studied separately.

Changing the order of summation and renaming the indices yields
  mij   mij
(vi − vj )vi = − (vi − vj )vj
dij dij
i∈Λ̄ j∈Λi i∈Λ̄ j∈Λi

Therefore,

1   mij
dρ (v, v) = (vi − vj )2 =: |v|21,ω (9.68)
2 dij
i∈Λ̄ j∈Λi

which is a positive definite term.

Remark 9.31. Given a mesh function v ∈ V0ω define a function ṽ ∈ H01 (Ω) that
coincides with v in the mesh points, and that is piecewise linear on a Delaunay
triangulation associated with the set of mesh points ω̄. Then

|v|21,ω = (∇ṽ, ∇ṽ) = |ṽ|21 ;

see [13]. ♣
Next consider the convection term. By definition we have mij = mji , dij = dji
and Nij = −Nji . Furthermore, (ρ1 ) implies Nji ρji = Nij (ρij − 1). Hence,
 
mij Nij ρij (vi − vj )vi = − mij Nij (ρij − 1) (vi − vj )vj
i∈Λ̄ j∈Λi i∈Λ̄ j∈Λi
9.3 Finite Volume Methods 301

and
 
1 1 2
cρ (v, v) = mij Nij − ρij (vi − vj )
2 2
i∈Λ̄ j∈Λi

1  
− mij Nij vi2 − vj2 .
4
i∈Λ̄ j∈Λi

Introducing
 
1 1  2
|v|2ρ,ω := mij Nij − ρij vi − vj ,
2 2
i∈Λ̄ j∈Λi

which is a well-defined semi-norm when (ρ2 ) is satisfied, we have

1 2 
cρ (v, v) = |v|2ρ,ω − vi mij Nij .
2
i∈Λ̄ j∈Λi

This and (9.68) yield

⎛ ⎞
 1 
aρ (v, v) = ε|v|21,ω + |v|2ρ,ω + mi vi2 ⎝ci − mij Nij ⎠ .
2
i∈Λ̄ j∈Λi

Note that

mi div bi + mij Nij = O (h) .
j∈Λi

This implies
γ 2

aρ (v, v) ≥ ε|v|21,ω + |v|2ρ,ω + v20,ω =: |||v|||ρ with v20,ω := mi vi2 ,
2
i∈Λ̄

provided h is sufficiently small, independent of the perturbation parameter ε.

We summarise the result of our stability analysis.
Theorem 9.32. Assume the discretisation (9.67) satisfies conditions (ρ1 ) and (ρ2 ).
Suppose (9.66) holds true. Then the bilinear form aρ (·, ·) is coercive with respect to
the norm |||·|||ρ , i.e.,
2
aρ (v, v) ≥ |||v|||ρ for all v ∈ V0ω

if h is sufficiently small, independent of the perturbation parameter ε.

302 9 Convection-Diffusion Problems

Remark 9.33. When ρ ≡ 12 , i.e. when the stabilisation is switched off, the bilinear
form is coercive with respect to the discrete ε-weighted energy norm
2 γ
|||v|||ε,ω := ε|v|21,ω + v20,ω .
2

However, when ρ ≡ 12 then we have coercivity of the scheme in a stronger norm,

which results in enhanced stability properties of the FVM. ♣

9.3.2 Inverse Monotonicity

Let the function ρ, which describes the FVM method, satisfy (ρ1 ) and (ρ3 ). Fur-
thermore, assume that c > 0 on Ω̄. Then recalling the definition (9.67), we have
ε
− Nij ρij ≥ 0.
dij

Hence, the diagonal entries of the matrix associated with Lρ are positive while
the off-diagonal ones are non-positive. Thus, the system matrix is an L0 matrix.
Next note that for v ≡ 1 we have [Lρ v]i = ci mi > 0. Therefore, application
of the M -criterion (Lemma 3.14) verifies the inverse monotonicity of Lρ , while
Lemma 3.17 gives the (∞ , ∞ )-stability inequality

v∞,ω ≤ f /c∞,ω .

Since Lρ is inverse monotone, it enjoys a comparison principle. That is if two

mesh functions v and w satisfy

[Lρ v]i ≤ [Lρ w]i for all i ∈ Λ,
=⇒ vi ≤ wi for all i ∈ Λ̄.
vi ≤ wi for all i ∈ ∂Λ

Remark 9.34. These results hold true with no restrictions imposed on the convective
field b and with (ρ2 ) possibly violated. ♣

The Green’s function on a tensor-product mesh

Using the inverse monotonicity of Lρ , we now study the Green’s functions asso-
ciated with Lρ and derive an anisotropic stability inequality on a general tensor-
product mesh ω̄ := ω̄x × ω̄y , with N mesh intervals in each coordinate direction.
A stability result of this kind was first established by Andreev for a simple upwind
difference scheme; see [7].
9.3 Finite Volume Methods 303

Setting h̄i := (hi+1 + hi )/2,

   
b1;i+1/2,j hi+1 b1;i−1/2,j hi
ρ+
1;ij := ρ − b1;i+1/2,j , ρ−
1;ij := ρ b1;i−1/2,j ,
ε ε
vij − vi−1,j vi+1,j − vij vi,j − vi−1,j
vx̄;ij := , vx̂;ij = and vx̌;ij =
hi h̄i h̄i
−
with analogous definitions for ρ+2 , ρ2 , vȳ , vŷ , vy̌ and k̄, we can rewrite (9.67) as:
  2
Find uN ∈ IR0 N +1
such that
' (   −
x̄x̂;ij + uȳ ŷ;ij − ρ1,ij ux̂,ij − ρ1,ij ux̌,ij
Lρ uN ij := −ε uN N + N N

−
− ρ+
2,ij uŷ,ij − ρ2,ij uy̌,ij + cij uij = fij
N N N

for i, j = 1, . . . , N − 1.
Any mesh function v that vanishes on the boundary can be represented using the
Green’s function:


N −1
vij = (v, Gij,·· )ρ := h̄k k̄ l Gij,kl [Lv]kl , (9.69)
k,l=1

where Gij,kl = G(xi , yj , ξk , ηl ) solves for fixed k and l

[Lρ G··,kl ]ij = δx;ik δy;jl on ω, Gij,kl = 0 on ∂ω

with
 
h̄−1
i if i = k, k̄ −1
j if j = l,
δx;ik = and δy;jl =
0 otherwise, 0 otherwise.

The adjoint operator to Lρ is

' ∗ (      − 
Lρ v kl = −ε vξ̄ξ̂;kl + vη̄η̂;kl + ρ+
1 v ξ̌;kl + ρ1 v ξ̂;kl
   − 
+ ρ+
2 v η̌;kl + ρ2 v η̂;kl + ckl vkl

and the Green’s function solves, for fixed i and j,

[L∗ Gij,·· ]kl = δx;ik δy;jl on ω, Gij,kl = 0 on ∂ω. (9.70)

In our subsequent analysis the following mean value theorem is used.

304 9 Convection-Diffusion Problems

Lemma 9.35. Let ϕ, g ∈ IRN +1 with gj ≥ 0 and m ≤ ϕj ≤ M for j = 1, . . . ,

N − 1. Then there exists a constant ϕ̃ ∈ [m, M ] with


N −1 
N −1
k̄ j ϕj gj = ϕ̃ k̄ j gj .
j=1 j=1

Let i and j be fixed. First, the inverse monotonicity of L yields Gij,kl ≥ 0.

Next, multiplying (9.70) by k̄ l and summing for l = 1, . . . , N − 1, we obtain the
one-dimensional equation
N −1  N −1 
 
−ε k̄ l Gij,·l + k̄ l ρ+
1,·l Gij,·l
l=1 ξ̄ ξ̌,k l=1 ξ̌,k

N −1  −1
 
N
+ k̄ l ρ−
1,·l Gij,·l + k̄ l ckl Gij,kl = δx;ik − Fk ,
l=1 ξ̂,k l=1

where
< = < =
ρ+
2;k,N −1 hN ρ+
2;k,0 h1
Fk = −ε 1 + Gη̄;ij,kN + ε 1 − Gη̄;ij,k1 ≥ 0,
ε ε

by (ρ3 ) and since G ≥ 0.

Defining


N −1
G̃k := k̄ l Gij,kl = Gij,·l 1,ω , for k = 0, . . . , N,
l=1

we see that according to Lemma 9.35 there exist mesh functions ρ̃+ , ρ̃− , c̃ with
ρ̃+ ≥ β1 , ρ̃− ≥ β1 and c̃ ≥ γ such that
   
−εG̃ξ̄ξ̌,k + ρ̃+ G̃ ξ̌,k + ρ̃− G̃ ξ̂,k + c̃k G̃k = δx;ik − Fk .

Let Γ = Γm,k be the Green’s function of the operator

−
[Lv]k = −εvξ̄ξ̌,k − ρ̃+
k vξ̌,k − ρ̃k vξ̂,k + c̃k vk .

Then G̃ can be written as


N −1
G̃k = Γi,k − h̄m Γm,k Fm ,
m=1
9.3 Finite Volume Methods 305

The nonnegativity of Γ and F gives

1
G̃k ≤ Γi,k ≤ ,
β1 inf t<0 ρ(t)

by Lemma 5.20 and Remark 5.23. We get the first inequality of the following theo-
rem. The other one is proved analogously.

Theorem 9.36. Suppose the control function ρ enjoys properties (ρ0 ) and (ρ3 ).
Then the Green’s function associated with L satisfies


N −1 
N −1
α α
max k̄ l Gij,kl ≤ and max h̄k Gij,kl ≤
i,j,k=1,...,N −1 β1 i,j,l=1,...,N −1 β2
l=1 k=1

with α = 1/ inf t<0 ρ(t) ≤ 2.

Finally, we use these bounds on the Green’s function to derive stability estimates
for the operator Lρ . For any mesh function v : ω̄ → IR that vanish on ∂ω, introduce
the norm


N −1
vA := h̄k max |vkl | .
l=1,...,N −1
k=1

Its dual norm with respect to the discrete scalar product (·, ·)ρ is


N −1
vA∗ = max k̄ k |vkl | ,
k=1,...,N −1
l=1

cf. [21, Theorem 2]. The representation (9.69) gives

|vij | ≤ Gij,·· A∗ vA .

Application of Theorem 9.36 yields our final stability result which is an extension
of the (∞ , 1 ) stability of Sect. 4.2.5 and a generalisation of [7].

Theorem 9.37. Suppose the control function ρ enjoys properties (ρ0 ) and (ρ3 ).
Then the operator Lρ is (∞ , 1 ⊗ ∞ ) stable with

α α
v∞,ω ≤ Lρ v 1⊗ ∞
and v∞,ω ≤ Lρ v ∞⊗ 1
β1 β2

with α = 1/ inf t<0 ρ(t) ≤ 2,

306 9 Convection-Diffusion Problems

 −1
N
 
w := h̄k max wkl 
1⊗ ∞
l=1,...,N −1
k=1

and

 −1
N
 
w := k̄ l max wkl .
∞⊗ 1
k=1,...,N −1
l=1

9.3.3 Convergence

Energy norm
Starting from the coercivity of the bilinear form aρ (·, ·), see Theorem 9.32, the
analysis proceeds along the lines of Sect. 5.4.2 resembling many of the details also
used for the Galerkin FEM in two dimensions, see Sect. 9.2.2. Eventually one gets
   
u − uN  + u − uN  ≤ CN −1 max |ψ  | ln1/2 N
ρ ε

for tensor-product meshes of Shishkin-type with σ ≥ 2; see also [174].

Maximum norm
The pointwise errors can be bounded using the hybrid stability inequalities from
Theorem 9.37, see [90]. We give a very brief outline of the argument.
The truncation error is split according to the decomposition of Theorem 7.17.
Then either of the two bounds from Theorem 9.37 is applied. Section 4.2.5 gives a
flavour of the technical details. For a S-type mesh with σ ≥ 2 we obtain
 
u − uN  ≤ CN −1 max |ψ  |.
∞,ω

If ρ is Lipschitz continuous in (−m, m) with m > 0, then there exists an Nm > 0

independent of the perturbation parameter ε such that on a standard Shishkin mesh
with σ ≥ 2
 
u − uN  ≤ CN −1 for N ≥ Nm .
∞,ω

In the latter case the stabilisation is reduced when the local mesh size is small
enough, thus giving higher accuracy inside the layers.
9.3 Finite Volume Methods 307

9.3.3.1 Numerical Tests

We verify our theoretical results for the upwind FEM on Shishkin meshes when ap-
plied to the test problem (9.3). For our tests we take ε = 10−8 , which is a sufficiently
small choice to bring out the singularly perturbed nature of the problem.
We test the method for three different choices of the controlling function ρ. The
errors are measured in the discrete energy and maximum norm and in the FVM-
norm.
For ρU,0 (see Table 9.7) we observe convergence of almost first order, namely
N −1 ln N , in all three norms, while for both ρU,2 and ρI —which are Lipschitz
continuous—the errors behave like O(N −1 ); see Tables 9.8 and 9.9. Note, this is
covered by our analysis for the maximum norm only.

Table 9.7 FVM on Shishkin meshes, ρ = ρU,0

     
u − uN   u − uN   u − uN 
ρ ε,ω ∞,ω
N error rate error rate error rate
16 2.7575e-1 0.68 2.0623e-1 0.55 1.8112e-1 0.62
32 1.7198e-1 0.75 1.4052e-1 0.66 1.1770e-1 0.71
64 1.0230e-1 0.79 8.9046e-2 0.73 7.1880e-2 0.76
128 5.8999e-2 0.83 5.3575e-2 0.79 4.2537e-2 0.80
256 3.3292e-2 0.85 3.1081e-2 0.82 2.4483e-2 0.83
512 1.8493e-2 0.87 1.7579e-2 0.85 1.3786e-2 0.85
1024 1.0153e-2 0.88 9.7672e-3 0.87 7.6456e-3 0.87
2048 5.5247e-3 — 5.3576e-3 — 4.1908e-3 —

Table 9.8 FVM on Shishkin meshes, ρ = ρI

     
u − uN   u − uN   u − uN 
ρ ε,ω ∞,ω
N error rate error rate error rate
16 1.5894e-1 0.83 8.9598e-2 0.80 7.5370e-2 0.70
32 8.9627e-2 0.92 5.1417e-2 0.90 4.6297e-2 0.84
64 4.7445e-2 0.96 2.7514e-2 0.95 2.5790e-2 0.92
128 2.4388e-2 0.98 1.4222e-2 0.98 1.3610e-2 0.96
256 1.2360e-2 0.99 7.2279e-3 0.99 6.9899e-3 0.98
512 6.2219e-3 1.00 3.6430e-3 0.99 3.5418e-3 0.99
1024 3.1214e-3 1.00 1.8288e-3 1.00 1.7827e-3 1.00
2048 1.5633e-3 — 9.1618e-4 — 8.9431e-4 —

Table 9.9 FVM on Shishkin meshes, ρ = ρU,2

     
u − uN   u − uN   u − uN 
ρ ε,ω ∞,ω
N error rate error rate error rate
16 1.5359e-1 0.81 8.2430e-2 0.77 7.6384e-2 0.72
32 8.7574e-2 0.91 4.8263e-2 0.88 4.6337e-2 0.85
64 4.6686e-2 0.95 2.6272e-2 0.93 2.5790e-2 0.92
128 2.4120e-2 0.98 1.3773e-2 0.96 1.3610e-2 0.96
256 1.2270e-2 0.99 7.0752e-3 0.98 6.9899e-3 0.98
512 6.1928e-3 0.99 3.5935e-3 0.99 3.5418e-3 0.99
1024 3.1122e-3 1.00 1.8132e-3 0.99 1.7827e-3 1.00
2048 1.5605e-3 — 9.1144e-4 — 8.9431e-4 —
Conclusions and Outlook

This monograph has attempted to provide a general framework for the convergence
analysis of a variety of numerical methods using layer-adapted meshes for the solu-
tion of certain classes of singularly perturbed problems. While for some problems
satisfactory answers have been presented, there is still a large number of open is-
sues. We would like to summarise the results here and to point out some directions
for future research.
Difference schemes in one dimension have been covered successfully to a large
extent, although there are a few minor open questions. For example, the analysis for
the fourth-order scheme for reaction-diffusion problems in Sect. 6.1.4 and the anal-
ysis for turning-point problems are restricted to Shishkin meshes so far. Similarly,
for the two-parameter problem in Sect. 6.3, no analysis for arbitrary meshes and a
second-order scheme is available.
For finite element methods—in both one and two dimensions—the situation is
different. Here results on arbitrary meshes are restricted to the interpolation error.
Convergence and superconvergence estimates are known for special meshes, namely
Shishkin-type meshes, only. A general framework for this very important class of
methods is still missing.
For systems of strongly coupled convection-diffusion problems, we only have a
limited grasp of the situation. Even for one-dimensional problems there are still ba-
sic difficulties: when different diffusion parameters are present, can sharp pointwise
bounds on derivatives be proved?
For two-dimensional convection-diffusion problems, further work on stability
bounds is needed to improve our understanding of these problems. In particular,
sharp estimates for the Green’s functions and negative-norm stability inequalities
are required. In one dimensions these turned out to be the key ingredient for the
convergence theory for arbitrary meshes, for the a posteriori error analysis and for
dealing with strongly coupled convection-diffusion equations.
Two parameter problems of reaction-convection-diffusion type in two dimen-
sions were considered in a small number of publications, but the presentations are
typically very technical, mainly because of complicated solution decompositions in-
volving a large number of different terms. Also the techniques used are very similar
to that known from convection-diffusion problems with a single parameter. Further
research is required to derive a general comprehensive theory.

309
310 Conclusions and Outlook

Time-dependent problems have not been included in this book because only a
few results are available that are insufficient for the development of a general the-
ory. The vast majority of results use first-order backward Euler for discretisation
in time. Higher-order time discretisations by A-stable Runge-Kutta methods have
been considered in the literature, but their analysis is incomplete because off a lack
of resolvent estimates for non-uniform meshes.
This account of open issues in the field is naturally incomplete. Studying some of
the material presented in the book, the reader will certainly discover further mathe-
matical problems worthy of investigating.
References

1. Abrahamsson, L.R., Keller, H.B., Kreiss, H.-O.: Difference approximations for singular per-
turbations of systems of ordinary differential equations. Numer. Math. 22, 367–391 (1974)
2. Ainsworth, M., Oden, J.: A posteriori error estimation in finite element analysis. Pure and
Applied Mathematics. Wiley, Chichester (2000)
3. Alhumaizi, K.: Flux-limiting solution techniques for simulation of reaction-diffusion-
convection system. Commun. Nonlinear Sci. Numer. Simul. 12(6), 953–965 (2007)
4. Andreev, V.B.: Convergence of a modified Samarskii’s monotonic scheme on a smoothly
condensing grid. Comput. Math. Math. Phys. 38(8), 1212–1224 (1998)
5. Andreev, V.B.: The Green function and a priori estimates of solutions of monotone three-point
singularly perturbed finite-difference schemes. Differ. Equ. 37(7), 923–933 (2001)
6. Andreev, V.B.: A priori estimates for solutions of singularly perturbed two-point boundary
value problems. Mat. Model. 14, 5–16 (2002)
7. Andreev, V.B.: Anisotropic estimates of the green function for a singularly perturbed two-
dimensional monotone convection-diffusion equation operator and its applications. Comput.
Math. Math. Phys. 43(4), 521–528 (2003)
8. Andreev, V.B.: On the accuracy of grid approximations of nonsmooth solutions of a singularly
perturbed reaction-diffusion equation in the square. Differ. Equ. 42(7), 954–966 (2006)
9. Andreev, V.B., Kopteva, N.: Pointwise approximation of corner singularities for a singularly
perturbed reaction-diffusion equation in an l-shaped domain. Math. Comp. 77(264), 2125–
2139 (2008)
10. Andreev, V.B., Kopteva, N.V.: A study of difference schemes with the first derivative approx-
imated by a central difference ratio. Comput. Math. Math. Phys. 36(8), 1065–1078 (1996)
11. Andreev, V.B., Kopteva, N.V.: On the convergence, uniform with respect to a small parameter,
of monotone three-point finite-difference approximations. Differ. Equations 34(7), 921–929
(1998)
12. Andreev, V.B., Savin, I.A.: The uniform convergence with respect to a small parameter of
A. A. Samarskij’s monotone scheme and its modification. Comput. Math. Math. Phys. 35(5),
581–591 (1995)
13. Angermann, L.: Numerical solution of second-order elliptic equations on plane domains.
Math. Model. Numer. Anal. 25, 169–191 (1991)
14. Angermann, L.: Error estimates for the finite-element solution of an elliptic singularly per-
turbed problem. IMA J. Numer. Anal. 15(2), 161–196 (1995)
15. Apel, T., Dobrowolski, M.: Anisotropic interpolation with applications to the finite element
method. Computing 47(3-4), 277–293 (1992)
16. Axelsson, O., Nikolova, M.: Adaptive refinement for convection-diffusion problems based on
a defect-correction technique and finite difference method. Computing 58(1), 1–30 (1997)
17. Baba, K., Tabata, M.: On a conservative upwind finite element scheme for convective diffu-
sion equations. RAIRO Anal. Numer. 15, 3–25 (1981)
18. Bakhvalov, N.S.: Towards optimization of methods for solving boundary value problems in
the presence of boundary layers. Zh. Vychisl. Mat. i Mat. Fiz. 9, 841–859 (1969). In Russian

311
312 References

19. Beckett, G.: The robust and efficient numerical solution of singularly perturbed boundary
value problems using grid adaptivity. Ph. D. thesis, University of Strathclyde (1999)
20. Beckett, G., Mackenzie, J.: Convergence analysis of finite difference approximations on
equidistributed grids to a singularly perturbed boundary value problem. Appl. Numer. Math.
35(2), 87–109 (2000)
21. Besov, O., Il’in, V., Nikol’skiı̆, S.: Integral representations of functions and embedding theo-
rems. Nauka, Moscow (1975)
22. Boglaev, I., Pack, S.: A uniformly convergent method for a singularly perturbed semilinear
reaction–diffusion problem with discontinuous data. Appl. Math. Comput. 182(1), 244–257
(2006)
23. Boglaev, I.P.: Approximate solution of a nonlinear boundary value problem with a small pa-
rameter at the highest-order derivative. Comput. Math. Math. Phys. 24(6), 30–35 (1984)
24. Boor, C.d.: Good approximation by splines with variable knots. In: A. Meir, A. Sharma (eds.)
Spline Functions and Approximation Theory, pp. 57–72. Birkhäuser Verlag, Basel (1973)
25. Boor, C.d.: Good Approximation by Splines with Variable Knots II. No. 363 in Lecture Notes
in Mathematics. Springer, Berlin (1973)
26. Chen, L., Xu, J.: Stability and accuracy of adapted finite element methods for singularly
perturbed problems. Numer. Math 109, 167–191 (2008)
27. Clavero, C., Gracia, J.L.: High order methods for elliptic and time dependent reaction-
diffusion singularly perturbed problems. Appl. Math. Comput. 168(2), 1109–1127 (2005)
28. Clavero, C., Gracia, J.L., Lisbona, F.: High order methods on shishkin meshes for singular
perturbation problems of convection-diffusion type. Numer. Algorithms 22(1), 73–97 (1999)
29. Clavero, C., Gracia, J.L., O’Riordan, E.: A parameter robust numerical method for a two
dimensional reaction-diffusion problem. Math. Comput. 74(252), 1743–1758 (2005)
30. Dobrowolski, M., Roos, H.-G.: A priori estimates for the solution of convection-diffusion
problems and interpolation on Shishkin meshes. Z. Anal. Anwend. 16(4), 1001–1012 (1997)
31. Dorfi, E.A., Drury, L.O’C.: Simple adaptive grids for 1-d initial value problems. J. Comput.
Phys. 69, 175–195 (1987)
32. Ewing, R.E. (ed.): The mathematics of reservoir simulation, Frontiers in Applied Mathemat-
ics, vol. 1. Society for Industrial and Applied Mathematics, Philadelphia (1983)
33. Farrell, P.A., Hegarty, A.F., Miller, J.J.H., O’Riordan, E., Shishkin, G.I.: The ε-uniform con-
vergence of the discrete derivatives for singularly perturbed problems. Preprint TCDM-9907,
Trinity College Dublin (1999)
34. Farrell, P.A., Hegarty, A.F., Miller, J.J.H., O’Riordan, E., Shishkin, G.I.: Robust computa-
tional techniques for boundary layers, Applied Mathematics and Mathematical Computation,
vol. 16. Chapman & Hall/CRC Press, Boca Raton, FL (2000)
35. Farrell, P.A., Hegarty, A.F., Miller, J.J.H., O’Riordan, E., Shishkin, G.I.: Singularly perturbed
convection–diffusion problems with boundary and weak interior layers. J. Comput. Appl.
Math. 166(1), 133–151 (2004)
36. Farrell, P.A., O’Riordan, E., Miller, J.J.H., Shishkin, G.I.: Parameter-uniform fitted mesh
method for quasilinear differential equations with boundary layers. Comput. Meth. Appl.
Math. 1(2), 154–172 (2001)
37. Farrell, P.A., O’Riordan, E., Shishkin, G.I.: A class of singularly perturbed semilinear differ-
ential equations with interior layers. Math. Comput. 74(252), 1759–1776 (2005)
38. Franz, S.: Singularly perturbed problems with characteristic layers: Supercloseness and post-
processing. Ph.D. thesis, TU Dresden, Dresden (2008)
39. Franz, S., Linß, T.: Superconvergence analysis of the galerkin fem for elliptic problems with
characteristic layers. Numer. Methods Partial Differential Equations 24(1), 144–164 (2008)
40. Franz, S., Linß, T., Roos, H.-G.: Superconvergence analysis of the SDFEM for elliptic prob-
lems with characteristic layers. Appl. Numer. Math. 58(12), 1818–1829 (2008)
41. Franz, S., Matthies, G.: Local projection stabilisation on S-type meshes for convection-
diffusion problems with characteristic layers. Preprint MATH-NM-07-2008, Institut für
Numerische Mathematik, TU Dresden (2008)
42. Fröhner, A.: Private communication (2001)
References 313

43. Fröhner, A., Linß, T., Roos, H.-G.: Defect correction on shishkin-type meshes. Numer.
Algorithms 26(3), 281–299 (2001)
44. Fröhner, A., Roos, H.-G.: The ε-uniform convergence of a defect correction method on a
Shiskin mesh. Appl. Numer. Math. 37(1-2), 79–94 (2001)
45. Gartland, E.C.: Graded-mesh difference schemes for singularly perturbed two-point boundary
value problems. Math. Comput. 51(184), 631–657 (1988)
46. Gracia, J.L., Lisbona, F., Clavero, C.: High order ε-uniform methods for singularly perturbed
reaction-diffusion problems. In: L. Vulkov, J. Waśniewski, P. Yalamov (eds.) Proceedings
NAA 2000, no. 1988 in Lecture Notes in Computer Science, pp. 350–358. Springer, Berlin
(2001)
47. Gracia, J.L., O’Riordan, E., Pickett, M.L.: A parameter robust second order numerical method
for a singularly perturbed two-parameter problem. Appl. Numer. Math. 56(7), 962–980
(2006)
48. Han, H., Kellogg, R.B.: Differentiability properties of solutions of the equation −ε2 Δu +
ru = f (x, y) in a square. SIAM J. Math. Anal. 21, 394–408 (1990)
49. Havik, E., Hemker, P., Hoffmann, W.: Application of the over-set grid technique to a model
singular perturbation problem. Computing 65(4), 339–356 (2000)
50. Hegarty, A.F., Miller, J.J.H., O’Riordan, E., Shishkin, G.I.: Use of central-difference opera-
tors for solution of singularly perturbed problems. Commun. Numer. Methods Eng. 10(4),
297–302 (1994)
51. Hemker, P.W.: The use of defect correction for the solution of a singularly perturbed o.d.e.
Seminarber., Humboldt-Univ. Berlin, Sekt. Math. 46, 91–103 (1982)
52. Hirsch, C.: Numerical computation of internal and external flows. Volume 1: Fundamentals
of numerical discretization. Wiley Series in Numerical Methods in Engineering. John Wiley
& Sons, Chichester (1988)
53. Hirsch, C.: Numerical computation of internal and external flows. Volume 2: Computational
methods for inviscid and viscous flows. Wiley Series in Numerical Methods in Engineering.
John Wiley & Sons, Chichester (1990)
54. Hughes, T.J.R., Brooks, A.: A multidimensional upwind scheme with no crosswind diffusion.
In: T.J.R. Hughes (ed.) Finite Element Methods for Convection Dominated Flows, pp. 19–35.
AMD. Vol. 34, ASME, New York (1979)
55. Il’in, A.M.: A difference scheme for a differential equation with a small parameter affecting
the highest derivative. Mat. Zametki 6, 237–248 (1969)
56. Jakob, M.: Heat transfer. John Wiley & Sons, Chichester (1959)
57. Kellogg, R.B., Linß, T., Stynes, M.: A finite difference method on layer-adapted meshes for
an elliptic reaction-diffusion system in two dimensions. Math. Comp. 77(264), 2085–2096
(2008)
58. Kellogg, R.B., Madden, N., Stynes, M.: A parameter-robust numerical method for a system
of reaction-diffusion equations in two dimensions. Numer. Methods Partial Differ. Equations
24(1), 312–334 (2008)
59. Kellogg, R.B., Stynes, M.: Corner singularities and boundary layers in a simple convection-
diffusion problem. J. Differential Equations 213, 81–120 (2005)
60. Kellogg, R.B., Stynes, M.: Sharpened bounds for corner singularities and boundary layers in
a simple convection-diffusion problem. Appl. Math. Lett. 20, 539–544 (2007)
61. Kellogg, R.B., Tsan, A.: Analysis of some difference approximations for a singular perturba-
tion problem without turning points. Math. Comput. 32, 1025–1039 (1978)
62. Knabner, P., Angermann, L.: Numerical methods for elliptic and parabolic partial differential
equations, Texts in Applied Mathematics, vol. 44. Springer, New York (2003)
63. Kopteva, N.: The two-dimensional sobolev inequality in the case of an arbitrary grid. Comput.
Math. Math. Phys. 38(4), 574–577 (1998)
64. Kopteva, N.: Maximum norm a posteriori error estimates for a one-dimensional convection-
diffusion problem. SIAM J. Numer. Anal. 39(2), 423–441 (2001)
65. Kopteva, N.: Error expansion for an upwind scheme applied to a two-dimensional convec-
tion-diffusion problem. SIAM J. Numer. Anal. 41(5), 1851–1869 (2003)
314 References

66. Kopteva, N.: Maximum norm a posteriori error estimates for a 1d singularly perturbed semi-
linear reaction-diffusion problem. IMA J. Numer. Anal. 27(3), 576–592 (2007)
67. Kopteva, N.: Maximum norm error analysis of a 2d singularly perturbed semilinear reaction-
diffusion problem. Math. Comput. 76(258), 631–646 (2007)
68. Kopteva, N., Madden, N., Stynes, M.: Grid equidistribution for reaction-diffusion problems
in one dimension. Numer. Algorithms 40(3), 305–322 (2005)
69. Kopteva, N., Stynes, M.: A robust adaptive method for a quasi-linear one-dimensional
convection-diffusion problem. SIAM J. Numer. Anal. 39(4), 1446–1467 (2001)
70. Kopteva, N.V.: Uniform convergence with respect to a small parameter of a scheme with
central difference on refining grids. Comput. Math. Math. Phys. 39(10), 1594–1610 (1999)
71. Kopteva, N.V., Linß, T.: Uniform second order pointwise convergence of a central differ-
ence approximation for a quasilinear convection-diffusion problem. J. Comput. Appl. Math.
137(2), 257–267 (2001)
72. Kopteva, N.V., Stynes, M.: Approximation of derivatives in a convection-diffusion two-point
boundary value problem. Appl. Numer. Math. 39(1), 47–60 (2001)
73. Kreiss, H.-O., Lorenz, J.: Initial-boundary value problems and the Navier-Stokes equations,
Pure and Applied Mathematics, vol. 136. Academic Press, Inc., Boston (1989)
74. Kunert, G.: A note on the energy norm for a singularly perturbed model problem. Computing
69(3), 265–272 (2002)
75. Ladyzhenskaya, O.A., Ural’tseva, N.N.: Linear and Quasilinear Elliptic Equations. Academic
Press, New York and London (1968)
76. Lenferink, W.: Pointwise convergence of approximations to a convection-diffusion equation
on a Shishkin mesh. Appl. Numer. Math. 32(1), 69–86 (2000)
77. Lenferink, W.: Some superconvergence results for finite element discretizations on a Shishkin
mesh of a convection-diffusion problem. In: J.J.H.M. L. G. Vulkov, G.I. Shishkin (eds.)
Analytical and Numerical Methods for Convection-Dominated and Singularly Perturbed
Problems. Nova Science Publishers, Inc., Huntington, NY (2000)
78. Lin, Q.: A rectangle test for finite element analysis. In: Proceedings of Systems Science &
Systems Engineering, pp. 213–216. Great Wall (H. K.) Culture Publish Co. (1991)
79. Lin, Q., Yan, N., Zhou, A.: A rectangle test for interpolated finite elements. In: Proc. Syst.
Sci. Eng., pp. 217–229. Great Wall (H.K.) Culture Publish Co. (1991)
80. Linß, T.: Untersuchungen zur numerischen Behandlung singulär gestörter Randwert- und An-
fangsrandwertaufgaben auf Shishkin-Gittern. Diplomarbeit, Technische Universität Dresden
(1996)
81. Linß, T.: An upwind difference scheme on a novel Shishkin-type mesh for a linear convection-
diffusion problem. J. Comput. Appl. Math. 110(1), 93–104 (1999)
82. Linß, T.: Analysis of a Galerkin finite element method on a Bakhvalov-Shishkin mesh for a
linear convection-diffusion problem. IMA J. Numer. Anal. 20(4), 621–632 (2000)
83. Linß, T.: Uniform superconvergence of a Galerkin finite element method on Shishkin-type
meshes. Numer. Methods Partial Differ. Equations 16(5), 426–440 (2000)
84. Linß, T.: The necessity of Shishkin-decompositions. Appl. Math. Lett. 14, 891–896 (2001)
85. Linß, T.: Sufficient conditions for uniform convergence on layer-adapted grids. Appl. Numer.
Math. 37(1-2), 241–255 (2001)
86. Linß, T.: Uniform pointwise convergence of finite difference schemes using grid equidistri-
bution. Computing 66(1), 27–39 (2001)
87. Linß, T.: Uniform second order pointwise convergence of a finite difference approximation
for a quasilinear convection-diffusion problem. Comput. Math. Math. Phys. 41(6), 898–909
(2001)
88. Linß, T.: Finite difference schemes for convection-diffusion problems with a concentrated
source and a discontinuous convection field. Comput. Methods Appl. Math. 2(1), 41–49
(2002)
89. Linß, T.: Solution decompositions for linear convection-diffusion problems. Z. Anal.
Anwend. 21(1), 209–214 (2002)
90. Linß, T.: Uniform pointwise convergence of an upwind finite volume method on layer-adapted
meshes. ZAMM, Z. Angew. Math. Mech. 82(4), 247–254 (2002)
References 315

91. Linß, T.: Layer-adapted meshes for convection-diffusion problems. Comput. Methods Appl.
Mech. Eng. 192(9-10), 1061–1105 (2003)
92. Linß, T.: Robustness of an upwind finite difference scheme for semilinear convection-
diffusion problems with boundary turning points. J. Comput. Math. 21(4), 401–410 (2003)
93. Linß, T.: Error expansion for a first-order upwind difference scheme applied to a model
convection-diffusion problem. IMA J. Numer. Anal. 24(2), 239–253 (2004)
94. Linß, T.: Anisotropic meshes and streamline-diffusion stabilization for convection-diffusion
problems. Commun. Numer. Methods Eng. 21(10), 515–525 (2005)
95. Linß, T.: Sufficient conditions for uniform convergence on layer-adapted meshes for one-
dimensional reaction-diffusion problems. Numer. Algorithms 40(1), 23–32 (2005)
96. Linß, T.: Analysis of an upwind finite-difference scheme for a system of coupled singularly
perturbed convection-diffusion equations. Computing 79, 23–32 (2007)
97. Linß, T.: Maximum-norm error analysis of a non-monotone FEM for a singularly perturbed
reaction-diffusion problem. BIT 47(2), 379–391 (2007)
98. Linß, T.: Robust convergence of a compact fourth-order finite difference scheme for reaction-
diffusion problems. Numer. Math. 111(2), 239–249 (2008)
99. Linß, T.: Analysis of a fem for a coupled system of singularly perturbed reaction-diffusion
equations. Numer. Algorithms 50(3), 283–291 (2009)
100. Linß, T.: Analysis of a system of singularly perturbed convection-diffusion equations with
strong coupling. SIAM J. Numer. Anal. 47(3), 1847–1862 (2009)
101. Linß, T., Kopteva, N.: A posteriori error estimation for a defect-correction method applied to
convection-diffusion problems. Preprint, University of Limerick (2008)
102. Linß, T., Madden, N.: Accurate solution of a system of coupled singularly perturbed reaction-
diffusion equations. Computing 73(2), 121–133 (2004)
103. Linß, T., Madden, N.: Parameter uniform approximations for time-dependent reaction-
diffusion problems. Numer. Methods Partial Differ. Equations 23(6), 1290–1300 (2007)
104. Linß, T., Madden, N.: Layer-adapted meshes for a system of coupled singularly perturbed
reaction-diffusion problem. IMA J. Numer. Anal. 29(1), 109–125 (2009)
105. Linß, T., Roos, H.-G.: Analysis of a finite-difference scheme for a singularly perturbed prob-
lem with two small parameters. J. Math. Anal. Appl. 289(2), 355–366 (2004)
106. Linß, T., Roos, H.-G., Vulanović, R.: Uniform pointwise convergence on Shishkin-type
meshes for quasilinear convection-diffusion problems. SIAM J. Numer. Anal. 38(3), 897–
912 (2000)
107. Linß, T., Stynes, M.: A hybrid difference scheme on a Shishkin mesh for linear convection-
diffusion problems. Appl. Numer. Math. 31(3), 255–270 (1999)
108. Linß, T., Stynes, M.: Asymptotic analysis and Shishkin-type decomposition for an elliptic
convection-diffusion problem. J. Math. Anal. Appl. 261, 604–632 (2001)
109. Linß, T., Stynes, M.: Numerical methods on Shishkin meshes for linear convection-diffusion
problems. Comput. Methods Appl. Mech. Engrg. 190(28), 3527–3542 (2001)
110. Linß, T., Stynes, M.: The SDFEM on Shishkin meshes for linear convection-diffusion prob-
lems. Numer. Math. 87(3), 457–484 (2001)
111. Linß, T., Stynes, M.: Systems of singularly perturbed differential equations. Comput. Methods
Appl. Math. (in press)
112. Linß, T., Vulanović, R.: Uniform methods for semilinear problems with an attractive boundary
turning point. Novi Sad J. Math. 31(2), 99–114 (2001)
113. Liseikin, V.: The use of special transformations in the numerical solution of boundary layer
problems. Comput. Math. Math. Phys. 30(1), 43–53 (1990)
114. Liseikin, V.D.: Layer resolving grids and transformations for singular perturbation problems.
VSP BV, Utrecht (2001)
115. Liseikin, V.D.: On the numerical solution of singularly perturbed problems with turning
points. Comput. Math. Math. Phys. 41(1), 55–83 (2001)
116. Liseikin, V.D., Yanenko, N.N.: On the numerical solution of equations with interior and exte-
rior boundary layers on a nonuniform mesh. In: BAIL III, Proc. 3rd Int. Conf. Boundary and
interior layers, pp. 68–80. Dublin, Ireland (1984)
316 References

117. Mackenzie, J.: Uniform convergence analysis of an upwind finite-difference approximation

of a convection-diffusion boundary value problem on an adaptive grid. IMA J. Numer. Anal.
19(2), 233–249 (1999)
118. Madden, N., Stynes, M.: A uniformly convergent numerical method for a coupled system
of two singularly perturbed linear reaction-diffusion problems. IMA J. Numer. Anal. 23(4),
627–644 (2003)
119. Matthies, G.: Local projection methods on layer-adapted meshes for higher order discretisa-
tions of convection diffusion problems. Appl. Numer. Math. 59(10), 2515–2533 (2009)
120. Miller, J.J.H., O’Riordan, E., Shishkin, G.I.: On piecewise-uniform meshes for upwind- and
central-difference operators for solving singularly perturbed problems. IMA J. Numer. Anal.
15(1), 89–99 (1995)
121. Miller, J.J.H., O’Riordan, E., Shishkin, G.I.: Fitted numerical methods for singular pertur-
bation problems. Error estimates in the maximum norm for linear problems in one and two
dimensions. World Scientific, Singapore (1996)
122. Miller, J.J.H., O’Riordan, E., Shishkin, G.I., Shishkina, L.P.: Fitted mesh methods for prob-
lems with parabolic boundary layers. Math. Proc. R. Ir. Acad. 98A(2), 173–190 (1998)
123. Morton, K.W.: Numerical solution of convection-diffusion problems, Applied Mathematics
and Mathematical Computation, vol. 12. Chapman & Hall, London (1996)
124. Natividad, M.C., Stynes, M.: Richardson extrapolation for a convection–diffusion problem
using a Shishkin mesh. Appl. Numer. Math. 45(2-3), 315–329 (2003)
125. Niijima, K.: Pointwise error estimates for a streamline diffusion finite element scheme.
Numer. Math 56, 707–719 (1990)
126. Nikolova, M., Axelsson, O.: Uniform in ε convergence of defect-correction method for
convection-diffusion problems. University of Nijmegen, Deptartment of Mathematics, Re-
port No. 9905 (1999)
127. O’Riordan, E., Stynes, J., Stynes, M.: A parameter-uniform finite difference method for a
coupled system of convection-diffusion two-point boundary value problems. Numer. Math.
Theor. Meth. Appl. 1, 176–197 (2008)
128. O’Riordan, E., Stynes, M.: Numerical analysis of a strongly coupled system of two singularly
perturbed convection-diffusion problems. Adv. Comput. Math. (to appear)
129. Ostrowski, A.: Über die Determinanten mit überwiegender Hauptdiagonale. Comment. Math.
Helv. 10, 69–96 (1937)
130. Polak, S., den Heijer, C., Schilders, W.: Semiconductor device modelling from the numerical
point of view. Int. J. Numer. Methods Eng. 24, 763–838 (1987)
131. Protter, M.H., Weinberger, H.F.: Maximum principles in differential equations, 2nd edn.
Springer, Berlin, Heidelberg (1984)
132. Qiu, Y., Sloan, D.M.: Analysis of difference approximations to a singularly perturbed two-
point boundary value problem on an adaptively generated grid. J. Comput. Appl. Math.
101(1-2), 1–25 (1999)
133. Qiu, Y., Sloan, D.M., Tang, T.: Numerical solution of a singularly pertured two-point bound-
ary value problem using equidistribution: Analysis of convergence. J. Comput. Appl. Math.
116(1), 121–143 (2000)
134. Roos, H.-G.: Layer-adapted grids for singular perturbation problems. ZAMM, Z. Angew.
Math. Mech. 78(5), 291–309 (1998)
135. Roos, H.-G.: Optimal convergence of basic schemes for elliptic boundary value problems
with strong parabolic layers. J. Math. Anal. Appl. 267(1), 194–208 (2002)
136. Roos, H.-G.: Error estimates for linear finite elements on Bakhvalov-type meshes. Appl.
Math., Praha 51(1), 63–72 (2006)
137. Roos, H.-G., Linß, T.: Sufficient conditions for uniform convergence on layer-adapted grids.
Computing 63(1), 27–45 (1999)
138. Roos, H.-G., Linß, T.: Gradient recovery for singularly perturbed boundary value problems.
I: One-dimensional convection-diffusion. Computing 66(2), 163–178 (2001)
139. Roos, H.-G., Linß, T.: Gradient recovery for singularly perturbed boundary value problems.
II: Two-dimensional convection-diffusion. Math. Models Methods Appl. Sci. 11(7), 1169–
1179 (2001)
References 317

140. Roos, H.-G., Skalický, T.: A comparison of the finite element method on Shishkin and
Gartland-type meshes for convection-diffusion problems. CWI Q. 10(3-4), 277–300 (1997)
141. Roos, H.-G., Stynes, M., Tobiska, L.: Robust Numerical Methods for Singularly Perturbed
Differential Equations, Springer Series in Computational Mathematics, vol. 24, 2nd edn.
Springer, Berlin (2008)
142. Roos, H.-G., Zarin, H.: A second-order scheme for singularly perturbed differential equations
with discontinuous source term. J. Numer. Math. 10(4), 275–289 (2002)
143. Roos, H.-G., Zarin, H.: The streamline-diffusion method for a convection-diffusion problem
with a point source. J. Comput. Appl. Math. 150(1), 109–128 (2003)
144. Samarskii, A.A.: Monotonic difference schemes for elliptic and parabolic equations in the
case of a non-selfadjoint elliptic operator. Comput. Math. Math. Phys. 5(3), 212–217 (1965)
145. Samarskii, A.A., Andreev, V.B.: Difference methods for elliptic equations. Nauka, Moscow
(1976)
146. Samarskij, A.A.: Theory of difference schemes. Nauka, Moscow (1989)
147. Savin, I.A.: On the rate of convergence, uniform with respect to a small parameter, of a dif-
ference scheme for an ordinary differential equation. Comput. Math. Math. Phys. 35(11),
1417–1422 (1995)
148. Schneider, D., Roos, H.-G., Linß, T.: Uniform convergence of an upwind finite element
method on layer-adapted grids. Comput. Methods Appl. Mech. Eng. 190(35-36), 4519–4530
(2001)
149. Shishkin, G.: A difference approximation of a singularly perturbed boundary-value problem
for quasilinear elliptic equations degenerating into first-order equations. Comput. Math. Math.
Phys. 32(4), 467–480 (1992)
150. Shishkin, G.I.: Grid approximation of singularly perturbed elliptic and parabolic equations.
Second doctorial thesis, Keldysh Institute, Moscow (1990). In Russian
151. Shishkin, G.I.: Discrete approximation of singularly perturbed elliptic and parabolic equa-
tions. Russian Academy of Sciences, Ural Section, Ekaterinburg (1992)
152. Stynes, M., O’Riordan, E.: A uniformly convergent Galerkin method on a Shishkin mesh for
a convection-diffusion problem. J. Math. Anal. Appl. 214(1), 36–54 (1997)
153. Stynes, M., Roos, H.-G.: The midpoint upwind scheme. Appl. Numer. Math. 23(3), 361–374
(1997)
154. Stynes, M., Tobiska, L.: A finite difference analysis of a streamline diffusion method on a
Shishkin mesh. Numer. Algorithms 18(3-4), 337–360 (1998)
155. Stynes, M., Tobiska, L.: The SDFEM for a convection-diffusion problem with a boundary
layer: Optimal error analysis and enhancement of accuracy. SIAM J. Numer. Anal. 41(5),
1620–1642 (2003)
156. Stynes, M., Tobiska, L.: Using rectangular qp elements in the sdfem for a convection-diffusion
problem with a boundary layer. Appl. Numer. Math. 58, 1789–1802 (2008)
157. Sun, G., Stynes, M.: Finite-element methods for singularly perturbed high-order elliptic two-
point boundary value problems. ii: Convection-diffusion-type problems. IMA J. Numer. Anal.
15(2), 197–219 (1995)
158. Surla, K., Uzelac, Z., Teofanov, L.: The discrete minimum principle for quadratic spline dis-
cretization of a singularly perturbed problem. Math. Comp. Sim. 79, 2490–2505 (2009)
159. Tartar, L.: Une nouvelle caractérisation des M matrices. Rev. Française Informat. Recherche
Opérationelle 5(Ser. R-3), 127–128 (1971)
160. Thomée, V.: Galerkin finite element methods for parabolic problems, Springer Series in Com-
putational Mathematics, vol. 25. Springer, Berlin, Heidelberg (1997)
161. van Veldhuizen, M.: Higher order methods for a singularly perturbed problem. Numer. Math.
30, 267–279 (1978)
162. Volkov, E.A.: Differentiability properties of solutions of boundary value problems for the
Laplace and Poisson equations on a rectangle. Proc. Steklov Inst. Math. 77, 101–126 (1965)
163. Vulanovic, R.: On a numerical solution of a type of singularly perturbed boundary value
problem by using a special discretization mesh. Zb. Rad., Prir.-Mat. Fak., Univ. Novom Sadu,
Ser. Mat. 13, 187–201 (1983)
318 References

164. Vulanović, R.: Mesh construction for discretization of singularly perturbed boundary value
problems. Ph. D. thesis, University of Novi Sad (1986)
165. Vulanović, R.: Non-equidistant generalizations of the Gushchin-Shchennikov scheme.
Z. Angew. Math. Mech. 67, 625–632 (1987)
166. Vulanović, R.: A uniform numerical method for quasilinear singular perturbation problems
without turning points. Computing 41(1-2), 97–106 (1989)
167. Vulanović, R.: On numerical solution of a mildly nonlinear turning point problem. Math.
Model. Numer. Anal. 24(6), 765–783 (1990)
168. Vulanović, R.: A higher-order scheme for quasilinear boundary value problems with two
small parameters. Computing 67(4), 287–303 (2001)
169. Vulanović, R.: A priori meshes for singularly perturbed quasilinear two-point boundary value
problems. IMA J. Numer. Anal. 21(1), 349–366 (2001)
170. Vulanović, R., Lin, P.: Numerical solution of quasilinear attractive turning point problems.
Comput. Math. Appl. 23(12), 75–82 (1992)
171. Ženı́šek, A., Vanmaele, M.: Applicability of the bramble-hilbert lemma in interpolation prob-
lems of narrow quadrilateral isoparametric finite elements. J. Comput. Appl. Math. 63(1-3),
109–122 (1995)
172. Ženı́šek, A., Vanmaele, M.: The interpolation theorem for narrow quadrilateral isoparametric
finite elements. Numer. Math. 72(1), 123–141 (1995)
173. Wahlbin, L.B.: Superconvergence in Galerkin finite element methods, Lecture Notes in Math-
ematics, vol. 1605. Springer, Berlin, Heidelberg (1995)
174. Wollstein, D., Linß, T., Roos, H.-G.: A uniformly accurate finite volume discretization for a
convection-diffusion problem. ETNA, Electron. Trans. Numer. Anal. 13, 1–11 (2002)
175. Zhang, Z.: Finite element superconvergence approximation for one-dimensional singularly
perturbed problems. Numer. Methods Partial Differ. Equations 18(3), 374–395 (2002)
176. Zhang, Z.: Finite element superconvergence on Shishkin mesh for 2-d convection-diffusion
problems. Math. Comp. 72(243), 1147–1177 (2003)
Index

adaptive mesh movement, 96, 129, 198 195, 204, 212, 219, 226, 249, 258,
268, 288, 291
error bounds
barrier function, 34 a priori, 306
discrete, 100, 116, 251, 259 error expansion, 87, 262
extrapolation, 130, 262

comparison principle, 34, 236

discrete, 80, 100, 191, 258, 302 FEM
for discontinuous functions, 36
artificial viscosity, 163, 285
for Robin boundary conditions, 60
Galerkin, 154, 184, 215, 267, 295
compatibility conditions, 237, 244
higher order, 292
consistency error
streamline-diffusion, 164, 289, 296
relative, 123
convergence
robust/uniform, 3
Green’s function, 36, 39, 58
discrete, 80, 145, 191, 224, 302
de Boor algorithm, 96
defect correction, 121
derivative approximation, 107, 125, 263, 288 interpolation, 152, 185, 214, 222, 264, 281,
gradient recovery, 160, 280 293, 295
difference operators, 78, 248, 257 inverse monotonicity, 34, 37, 172, 236
difference scheme
1st order, 79, 144, 223, 257
2nd order, 110, 190, 202, 217, 247
layer, 2
4th order, 206
characteristic, 28, 245, 294
central, 119, 190, 217, 247
exponential, 6, 19, 25, 27, 243
compact, 206
interior, 16, 21
monotone, 79, 110, 144, 190, 223, 257
overlapping layers, 17, 23
non-monotone, 119, 202, 206
discontinuous data, 16, 21, 51, 63, 103, 115, turning point, 143
201 turning-point, 15, 69
layer-damping transformations, 5

error bound
a posteriori, 94, 127, 137, 141, 161, 197, matrix
205, 221, 230 H-matrix, 54
a priori, 85, 100, 104, 114, 117, 125, 131, L0 -matrix, 37
137, 140, 148, 154, 179, 181, 188, M -criterion, 37

319
320 Index

M -matrix, 37, 111, 144, 258 artificial viscosity, 163

coercive, 54 streamline-diffusion, 164
inverse monotone, 37 stability
mesh (A, B), 35
— characterising function, 10 (L∞ , L1 ), 45, 59, 67
— equidistribution, 5 (L∞ , L∞ ), 35, 45, 53, 59, 65, 67, 237,
— generating function, 5 240
Bakhvalov, 7 (L∞ , L1,1/p ), 71
Bakhvalov-type, 9 (L∞ , L∞,1/p ), 70
Shishkin, 9 (L∞ , W −1,∞ ), 45, 65, 67
Shishkin-type, 10 (∞ , 1 ), 83, 97, 112, 174
monitor function, 5 (∞ , 1 ⊗ ∞ ), 305
(∞ , ∞ ), 38, 80, 136, 140, 209, 218, 224,
248, 258, 302
norms, 42 (∞ , 1,1/p ), 146
discrete, 78 (∞ , ∞,1/p ), 144
(∞ , w−1,∞ ), 83, 112, 136, 139, 175
stretching function, 5
point source, 16, 63, 103, 115 supercloseness, 156, 167, 269, 291, 294
systems
convection-diffusion
quadrature, 187 strongly coupled, 66, 137
quasilinear problem, 106, 115 weakly coupled, 64, 134
reaction-diffusion, 52, 214, 238, 247

singular perturbation, 2
stabilisation truncation error, 100, 251, 259
Lecture Notes in Mathematics
For information about earlier volumes
please contact your bookseller or Springer
LNM Online archive: springerlink.com

Vol. 1801: J. Azéma, M. Émery, M. Ledoux, M. Yor Approximation. Cetraro, Italy 2000. Editors: F. Amoroso,
(Eds.), Séminaire de Probabilités XXXVI (2003) U. Zannier (2003)
Vol. 1802: V. Capasso, E. Merzbach, B. G. Ivanoff, Vol. 1820: F. Hiai, H. Kosaki, Means of Hilbert Space
M. Dozzi, R. Dalang, T. Mountford, Topics in Spatial Operators (2003)
Stochastic Processes. Martina Franca, Italy 2001. Editor: Vol. 1821: S. Teufel, Adiabatic Perturbation Theory in
E. Merzbach (2003) Quantum Dynamics (2003)
Vol. 1803: G. Dolzmann, Variational Methods for Crys- Vol. 1822: S.-N. Chow, R. Conti, R. Johnson, J. Mallet-
talline Microstructure – Analysis and Computation Paret, R. Nussbaum, Dynamical Systems. Cetraro, Italy
(2003) 2000. Editors: J. W. Macki, P. Zecca (2003)
Vol. 1804: I. Cherednik, Ya. Markov, R. Howe, G. Vol. 1823: A. M. Anile, W. Allegretto, C. Ringhofer,
Lusztig, Iwahori-Hecke Algebras and their Representa- Mathematical Problems in Semiconductor Physics.
tion Theory. Martina Franca, Italy 1999. Editors: V. Bal- Cetraro, Italy 1998. Editor: A. M. Anile (2003)
doni, D. Barbasch (2003) Vol. 1824: J. A. Navarro González, J. B. Sancho de Salas,
Vol. 1805: F. Cao, Geometric Curve Evolution and Image C ∞ – Differentiable Spaces (2003)
Processing (2003) Vol. 1825: J. H. Bramble, A. Cohen, W. Dahmen, Mul-
Vol. 1806: H. Broer, I. Hoveijn. G. Lunther, G. Vegter, tiscale Problems and Methods in Numerical Simulations,
Bifurcations in Hamiltonian Systems. Computing Singu- Martina Franca, Italy 2001. Editor: C. Canuto (2003)
larities by Gröbner Bases (2003) Vol. 1826: K. Dohmen, Improved Bonferroni Inequal-
Vol. 1807: V. D. Milman, G. Schechtman (Eds.), Geomet- ities via Abstract Tubes. Inequalities and Identities of
ric Aspects of Functional Analysis. Israel Seminar 2000- Inclusion-Exclusion Type. VIII, 113 p, 2003.
2002 (2003) Vol. 1827: K. M. Pilgrim, Combinations of Complex
Vol. 1808: W. Schindler, Measures with Symmetry Prop- Dynamical Systems. IX, 118 p, 2003.
erties (2003) Vol. 1828: D. J. Green, Gröbner Bases and the Computa-
Vol. 1809: O. Steinbach, Stability Estimates for Hybrid tion of Group Cohomology. XII, 138 p, 2003.
Coupled Domain Decomposition Methods (2003) Vol. 1829: E. Altman, B. Gaujal, A. Hordijk, Discrete-
Vol. 1810: J. Wengenroth, Derived Functors in Functional Event Control of Stochastic Networks: Multimodularity
Analysis (2003) and Regularity. XIV, 313 p, 2003.
Vol. 1811: J. Stevens, Deformations of Singularities Vol. 1830: M. I. Gil’, Operator Functions and Localiza-
(2003) tion of Spectra. XIV, 256 p, 2003.
Vol. 1812: L. Ambrosio, K. Deckelnick, G. Dziuk, Vol. 1831: A. Connes, J. Cuntz, E. Guentner, N. Hig-
M. Mimura, V. A. Solonnikov, H. M. Soner, Mathemat- son, J. E. Kaminker, Noncommutative Geometry, Mar-
ical Aspects of Evolving Interfaces. Madeira, Funchal, tina Franca, Italy 2002. Editors: S. Doplicher, L. Longo
Portugal 2000. Editors: P. Colli, J. F. Rodrigues (2003) (2004)
Vol. 1813: L. Ambrosio, L. A. Caffarelli, Y. Brenier, Vol. 1832: J. Azéma, M. Émery, M. Ledoux, M. Yor
G. Buttazzo, C. Villani, Optimal Transportation and its (Eds.), Séminaire de Probabilités XXXVII (2003)
Applications. Martina Franca, Italy 2001. Editors: L. A. Vol. 1833: D.-Q. Jiang, M. Qian, M.-P. Qian, Mathemati-
Caffarelli, S. Salsa (2003) cal Theory of Nonequilibrium Steady States. On the Fron-
Vol. 1814: P. Bank, F. Baudoin, H. Föllmer, L.C.G. tier of Probability and Dynamical Systems. IX, 280 p,
Rogers, M. Soner, N. Touzi, Paris-Princeton Lectures on 2004.
Mathematical Finance 2002 (2003) Vol. 1834: Yo. Yomdin, G. Comte, Tame Geometry with
Vol. 1815: A. M. Vershik (Ed.), Asymptotic Combi- Application in Smooth Analysis. VIII, 186 p, 2004.
natorics with Applications to Mathematical Physics. Vol. 1835: O.T. Izhboldin, B. Kahn, N.A. Karpenko,
St. Petersburg, Russia 2001 (2003) A. Vishik, Geometric Methods in the Algebraic Theory
Vol. 1816: S. Albeverio, W. Schachermayer, M. Tala- of Quadratic Forms. Summer School, Lens, 2000. Editor:
grand, Lectures on Probability Theory and Statistics. J.-P. Tignol (2004)
Ecole d’Eté de Probabilités de Saint-Flour XXX-2000. Vol. 1836: C. Nǎstǎsescu, F. Van Oystaeyen, Methods of
Editor: P. Bernard (2003) Graded Rings. XIII, 304 p, 2004.
Vol. 1817: E. Koelink, W. Van Assche (Eds.), Orthogonal Vol. 1837: S. Tavaré, O. Zeitouni, Lectures on Probabil-
Polynomials and Special Functions. Leuven 2002 (2003) ity Theory and Statistics. Ecole d’Eté de Probabilités de
Vol. 1818: M. Bildhauer, Convex Variational Problems Saint-Flour XXXI-2001. Editor: J. Picard (2004)
with Linear, nearly Linear and/or Anisotropic Growth Vol. 1838: A.J. Ganesh, N.W. O’Connell, D.J. Wischik,
Conditions (2003) Big Queues. XII, 254 p, 2004.
Vol. 1819: D. Masser, Yu. V. Nesterenko, H. P. Schlick- Vol. 1839: R. Gohm, Noncommutative Stationary Pro-
ewei, W. M. Schmidt, M. Waldschmidt, Diophantine cesses. VIII, 170 p, 2004.
Vol. 1840: B. Tsirelson, W. Werner, Lectures on Probabil- Vol. 1866: O.E. Barndorff-Nielsen, U. Franz, R. Gohm,
ity Theory and Statistics. Ecole d’Eté de Probabilités de B. Kümmerer, S. Thorbjønsen, Quantum Independent
Saint-Flour XXXII-2002. Editor: J. Picard (2004) Increment Processes II. Structure of Quantum Lévy
Vol. 1841: W. Reichel, Uniqueness Theorems for Vari- Processes, Classical Probability, and Physics. Editors: M.
ational Problems by the Method of Transformation Schürmann, U. Franz, (2005)
Groups (2004) Vol. 1867: J. Sneyd (Ed.), Tutorials in Mathematical Bio-
Vol. 1842: T. Johnsen, A. L. Knutsen, K3 Projective Mod- sciences II. Mathematical Modeling of Calcium Dynam-
els in Scrolls (2004) ics and Signal Transduction. (2005)
Vol. 1843: B. Jefferies, Spectral Properties of Noncom- Vol. 1868: J. Jorgenson, S. Lang, Posn (R) and Eisenstein
muting Operators (2004) Series. (2005)
Vol. 1844: K.F. Siburg, The Principle of Least Action in Vol. 1869: A. Dembo, T. Funaki, Lectures on Probabil-
Geometry and Dynamics (2004) ity Theory and Statistics. Ecole d’Eté de Probabilités de
Vol. 1845: Min Ho Lee, Mixed Automorphic Forms, Saint-Flour XXXIII-2003. Editor: J. Picard (2005)
Torus Bundles, and Jacobi Forms (2004) Vol. 1870: V.I. Gurariy, W. Lusky, Geometry of Müntz
Vol. 1846: H. Ammari, H. Kang, Reconstruction of Small Spaces and Related Questions. (2005)
Inhomogeneities from Boundary Measurements (2004) Vol. 1871: P. Constantin, G. Gallavotti, A.V. Kazhikhov,
Vol. 1847: T.R. Bielecki, T. Björk, M. Jeanblanc, M. Y. Meyer, S. Ukai, Mathematical Foundation of Turbu-
Rutkowski, J.A. Scheinkman, W. Xiong, Paris-Princeton lent Viscous Flows, Martina Franca, Italy, 2003. Editors:
Lectures on Mathematical Finance 2003 (2004) M. Cannone, T. Miyakawa (2006)
Vol. 1848: M. Abate, J. E. Fornaess, X. Huang, J. P. Vol. 1872: A. Friedman (Ed.), Tutorials in Mathemati-
Rosay, A. Tumanov, Real Methods in Complex and CR cal Biosciences III. Cell Cycle, Proliferation, and Cancer
Geometry, Martina Franca, Italy 2002. Editors: D. Zait- (2006)
sev, G. Zampieri (2004) Vol. 1873: R. Mansuy, M. Yor, Random Times and En-
Vol. 1849: Martin L. Brown, Heegner Modules and Ellip- largements of Filtrations in a Brownian Setting (2006)
tic Curves (2004) Vol. 1874: M. Yor, M. Émery (Eds.), In Memoriam Paul-
Vol. 1850: V. D. Milman, G. Schechtman (Eds.), Geomet- André Meyer - Séminaire de Probabilités XXXIX (2006)
ric Aspects of Functional Analysis. Israel Seminar 2002- Vol. 1875: J. Pitman, Combinatorial Stochastic Processes.
2003 (2004) Ecole d’Eté de Probabilités de Saint-Flour XXXII-2002.
Vol. 1851: O. Catoni, Statistical Learning Theory and Editor: J. Picard (2006)
Stochastic Optimization (2004) Vol. 1876: H. Herrlich, Axiom of Choice (2006)
Vol. 1852: A.S. Kechris, B.D. Miller, Topics in Orbit Vol. 1877: J. Steuding, Value Distributions of L-Functions
Equivalence (2004) (2007)
Vol. 1853: Ch. Favre, M. Jonsson, The Valuative Tree Vol. 1878: R. Cerf, The Wulff Crystal in Ising and Percol-
(2004) ation Models, Ecole d’Eté de Probabilités de Saint-Flour
Vol. 1854: O. Saeki, Topology of Singular Fibers of Dif- XXXIV-2004. Editor: Jean Picard (2006)
ferential Maps (2004) Vol. 1879: G. Slade, The Lace Expansion and its Applica-
Vol. 1855: G. Da Prato, P.C. Kunstmann, I. Lasiecka, tions, Ecole d’Eté de Probabilités de Saint-Flour XXXIV-
A. Lunardi, R. Schnaubelt, L. Weis, Functional Analytic 2004. Editor: Jean Picard (2006)
Methods for Evolution Equations. Editors: M. Iannelli, Vol. 1880: S. Attal, A. Joye, C.-A. Pillet, Open Quantum
R. Nagel, S. Piazzera (2004) Systems I, The Hamiltonian Approach (2006)
Vol. 1856: K. Back, T.R. Bielecki, C. Hipp, S. Peng, Vol. 1881: S. Attal, A. Joye, C.-A. Pillet, Open Quantum
W. Schachermayer, Stochastic Methods in Finance, Bres- Systems II, The Markovian Approach (2006)
sanone/Brixen, Italy, 2003. Editors: M. Fritelli, W. Rung- Vol. 1882: S. Attal, A. Joye, C.-A. Pillet, Open Quantum
galdier (2004) Systems III, Recent Developments (2006)
Vol. 1857: M. Émery, M. Ledoux, M. Yor (Eds.), Sémi- Vol. 1883: W. Van Assche, F. Marcellàn (Eds.), Orthogo-
naire de Probabilités XXXVIII (2005) nal Polynomials and Special Functions, Computation and
Vol. 1858: A.S. Cherny, H.-J. Engelbert, Singular Application (2006)
Stochastic Differential Equations (2005) Vol. 1884: N. Hayashi, E.I. Kaikina, P.I. Naumkin,
Vol. 1859: E. Letellier, Fourier Transforms of Invariant I.A. Shishmarev, Asymptotics for Dissipative Nonlinear
Functions on Finite Reductive Lie Algebras (2005) Equations (2006)
Vol. 1860: A. Borisyuk, G.B. Ermentrout, A. Friedman, Vol. 1885: A. Telcs, The Art of Random Walks (2006)
D. Terman, Tutorials in Mathematical Biosciences I. Vol. 1886: S. Takamura, Splitting Deformations of Dege-
Mathematical Neurosciences (2005) nerations of Complex Curves (2006)
Vol. 1861: G. Benettin, J. Henrard, S. Kuksin, Hamilto- Vol. 1887: K. Habermann, L. Habermann, Introduction to
nian Dynamics – Theory and Applications, Cetraro, Italy, Symplectic Dirac Operators (2006)
1999. Editor: A. Giorgilli (2005) Vol. 1888: J. van der Hoeven, Transseries and Real Dif-
Vol. 1862: B. Helffer, F. Nier, Hypoelliptic Estimates and ferential Algebra (2006)
Spectral Theory for Fokker-Planck Operators and Witten Vol. 1889: G. Osipenko, Dynamical Systems, Graphs, and
Laplacians (2005) Algorithms (2006)
Vol. 1863: H. Führ, Abstract Harmonic Analysis of Con- Vol. 1890: M. Bunge, J. Funk, Singular Coverings of
tinuous Wavelet Transforms (2005) Toposes (2006)
Vol. 1864: K. Efstathiou, Metamorphoses of Hamiltonian Vol. 1891: J.B. Friedlander, D.R. Heath-Brown,
Systems with Symmetries (2005) H. Iwaniec, J. Kaczorowski, Analytic Number Theory,
Vol. 1865: D. Applebaum, B.V. R. Bhat, J. Kustermans, Cetraro, Italy, 2002. Editors: A. Perelli, C. Viola (2006)
J. M. Lindsay, Quantum Independent Increment Pro- Vol. 1892: A. Baddeley, I. Bárány, R. Schneider, W. Weil,
cesses I. From Classical Probability to Quantum Stochas- Stochastic Geometry, Martina Franca, Italy, 2004. Editor:
tic Calculus. Editors: M. Schürmann, U. Franz (2005) W. Weil (2007)
Vol. 1893: H. Hanßmann, Local and Semi-Local Bifur- Vol. 1918: B. Roberts, R. Schmidt, Local Newforms for
cations in Hamiltonian Dynamical Systems, Results and GSp(4) (2007)
Examples (2007) Vol. 1919: R.A. Carmona, I. Ekeland, A. Kohatsu-
Vol. 1894: C.W. Groetsch, Stable Approximate Evalua- Higa, J.-M. Lasry, P.-L. Lions, H. Pham, E. Taflin,
tion of Unbounded Operators (2007) Paris-Princeton Lectures on Mathematical Finance 2004.
Vol. 1895: L. Molnár, Selected Preserver Problems on Editors: R.A. Carmona, E. Çinlar, I. Ekeland, E. Jouini,
Algebraic Structures of Linear Operators and on Function J.A. Scheinkman, N. Touzi (2007)
Spaces (2007) Vol. 1920: S.N. Evans, Probability and Real Trees. Ecole
Vol. 1896: P. Massart, Concentration Inequalities and d’Été de Probabilités de Saint-Flour XXXV-2005 (2008)
Model Selection, Ecole d’Été de Probabilités de Saint- Vol. 1921: J.P. Tian, Evolution Algebras and their Appli-
Flour XXXIII-2003. Editor: J. Picard (2007) cations (2008)
Vol. 1897: R. Doney, Fluctuation Theory for Lévy Vol. 1922: A. Friedman (Ed.), Tutorials in Mathematical
Processes, Ecole d’Été de Probabilités de Saint-Flour BioSciences IV. Evolution and Ecology (2008)
XXXV-2005. Editor: J. Picard (2007) Vol. 1923: J.P.N. Bishwal, Parameter Estimation in
Vol. 1898: H.R. Beyer, Beyond Partial Differential Equa- Stochastic Differential Equations (2008)
tions, On linear and Quasi-Linear Abstract Hyperbolic Vol. 1924: M. Wilson, Littlewood-Paley Theory and
Evolution Equations (2007) Exponential-Square Integrability (2008)
Vol. 1899: Séminaire de Probabilités XL. Editors: Vol. 1925: M. du Sautoy, L. Woodward, Zeta Functions
C. Donati-Martin, M. Émery, A. Rouault, C. Stricker of Groups and Rings (2008)
(2007) Vol. 1926: L. Barreira, V. Claudia, Stability of Nonauto-
Vol. 1900: E. Bolthausen, A. Bovier (Eds.), Spin Glasses nomous Differential Equations (2008)
(2007) Vol. 1927: L. Ambrosio, L. Caffarelli, M.G. Crandall,
Vol. 1901: O. Wittenberg, Intersections de deux L.C. Evans, N. Fusco, Calculus of Variations and Non-
quadriques et pinceaux de courbes de genre 1, Intersec- Linear Partial Differential Equations. Cetraro, Italy 2005.
tions of Two Quadrics and Pencils of Curves of Genus 1 Editors: B. Dacorogna, P. Marcellini (2008)
(2007) Vol. 1928: J. Jonsson, Simplicial Complexes of Graphs
Vol. 1902: A. Isaev, Lectures on the Automorphism (2008)
Groups of Kobayashi-Hyperbolic Manifolds (2007) Vol. 1929: Y. Mishura, Stochastic Calculus for Fractional
Vol. 1903: G. Kresin, V. Maz’ya, Sharp Real-Part Theo- Brownian Motion and Related Processes (2008)
rems (2007) Vol. 1930: J.M. Urbano, The Method of Intrinsic Scaling.
Vol. 1904: P. Giesl, Construction of Global Lyapunov A Systematic Approach to Regularity for Degenerate and
Singular PDEs (2008)
Functions Using Radial Basis Functions (2007)
Vol. 1931: M. Cowling, E. Frenkel, M. Kashiwara,
Vol. 1905: C. Prévôt, M. Röckner, A Concise Course on
A. Valette, D.A. Vogan, Jr., N.R. Wallach, Representation
Stochastic Partial Differential Equations (2007)
Theory and Complex Analysis. Venice, Italy 2004.
Vol. 1906: T. Schuster, The Method of Approximate
Editors: E.C. Tarabusi, A. D’Agnolo, M. Picardello
Inverse: Theory and Applications (2007)
(2008)
Vol. 1907: M. Rasmussen, Attractivity and Bifurcation Vol. 1932: A.A. Agrachev, A.S. Morse, E.D. Sontag,
for Nonautonomous Dynamical Systems (2007) H.J. Sussmann, V.I. Utkin, Nonlinear and Optimal
Vol. 1908: T.J. Lyons, M. Caruana, T. Lévy, Differential Control Theory. Cetraro, Italy 2004. Editors: P. Nistri,
Equations Driven by Rough Paths, Ecole d’Été de Proba- G. Stefani (2008)
bilités de Saint-Flour XXXIV-2004 (2007) Vol. 1933: M. Petkovic, Point Estimation of Root Finding
Vol. 1909: H. Akiyoshi, M. Sakuma, M. Wada, Methods (2008)
Y. Yamashita, Punctured Torus Groups and 2-Bridge Vol. 1934: C. Donati-Martin, M. Émery, A. Rouault,
Knot Groups (I) (2007) C. Stricker (Eds.), Séminaire de Probabilités XLI (2008)
Vol. 1910: V.D. Milman, G. Schechtman (Eds.), Geo- Vol. 1935: A. Unterberger, Alternative Pseudodifferential
metric Aspects of Functional Analysis. Israel Seminar Analysis (2008)
2004-2005 (2007) Vol. 1936: P. Magal, S. Ruan (Eds.), Structured Popula-
Vol. 1911: A. Bressan, D. Serre, M. Williams, tion Models in Biology and Epidemiology (2008)
K. Zumbrun, Hyperbolic Systems of Balance Laws. Vol. 1937: G. Capriz, P. Giovine, P.M. Mariano (Eds.),
Cetraro, Italy 2003. Editor: P. Marcati (2007) Mathematical Models of Granular Matter (2008)
Vol. 1912: V. Berinde, Iterative Approximation of Fixed Vol. 1938: D. Auroux, F. Catanese, M. Manetti, P. Seidel,
Points (2007) B. Siebert, I. Smith, G. Tian, Symplectic 4-Manifolds
Vol. 1913: J.E. Marsden, G. Misiołek, J.-P. Ortega, and Algebraic Surfaces. Cetraro, Italy 2003. Editors:
M. Perlmutter, T.S. Ratiu, Hamiltonian Reduction by F. Catanese, G. Tian (2008)
Stages (2007) Vol. 1939: D. Boffi, F. Brezzi, L. Demkowicz, R.G.
Vol. 1914: G. Kutyniok, Affine Density in Wavelet Durán, R.S. Falk, M. Fortin, Mixed Finite Elements,
Analysis (2007) Compatibility Conditions, and Applications. Cetraro,
Vol. 1915: T. Bıyıkoǧlu, J. Leydold, P.F. Stadler, Italy 2006. Editors: D. Boffi, L. Gastaldi (2008)
Laplacian Eigenvectors of Graphs. Perron-Frobenius and Vol. 1940: J. Banasiak, V. Capasso, M.A.J. Chap-
Faber-Krahn Type Theorems (2007) lain, M. Lachowicz, J. Miȩkisz, Multiscale Problems in
Vol. 1916: C. Villani, F. Rezakhanlou, Entropy Methods the Life Sciences. From Microscopic to Macroscopic.
for the Boltzmann Equation. Editors: F. Golse, S. Olla Bȩdlewo, Poland 2006. Editors: V. Capasso, M. Lachow-
(2008) icz (2008)
Vol. 1917: I. Veselić, Existence and Regularity Prop- Vol. 1941: S.M.J. Haran, Arithmetical Investigations.
erties of the Integrated Density of States of Random Representation Theory, Orthogonal Polynomials, and
Schrödinger (2008) Quantum Interpolations (2008)
Vol. 1942: S. Albeverio, F. Flandoli, Y.G. Sinai, SPDE in Vol. 1968: R. Weissauer, Endoscopy for GSP(4) and the
Hydrodynamic. Recent Progress and Prospects. Cetraro, Cohomology of Siegel Modular Threefolds (2009)
Italy 2005. Editors: G. Da Prato, M. Röckner (2008) Vol. 1969: B. Roynette, M. Yor, Penalising Brownian
Vol. 1943: L.L. Bonilla (Ed.), Inverse Problems and Imag- Paths (2009)
ing. Martina Franca, Italy 2002 (2008) Vol. 1970: M. Biskup, A. Bovier, F. den Hollander, D.
Vol. 1944: A. Di Bartolo, G. Falcone, P. Plaumann, Ioffe, F. Martinelli, K. Netočný, F. Toninelli, Methods of
K. Strambach, Algebraic Groups and Lie Groups with Contemporary Mathematical Statistical Physics. Editor:
Few Factors (2008) R. Kotecký (2009)
Vol. 1945: F. Brauer, P. van den Driessche, J. Wu (Eds.), Vol. 1971: L. Saint-Raymond, Hydrodynamic Limits of
Mathematical Epidemiology (2008) the Boltzmann Equation (2009)
Vol. 1946: G. Allaire, A. Arnold, P. Degond, T.Y. Hou, Vol. 1972: T. Mochizuki, Donaldson Type Invariants for
Quantum Transport. Modelling, Analysis and Asymp- Algebraic Surfaces (2009)
totics. Cetraro, Italy 2006. Editors: N.B. Abdallah, Vol. 1973: M.A. Berger, L.H. Kauffmann, B. Khesin,
G. Frosali (2008) H.K. Moffatt, R.L. Ricca, De W. Sumners, Lectures on
Vol. 1947: D. Abramovich, M. Mariño, M. Thaddeus, Topological Fluid Mechanics. Cetraro, Italy 2001. Editor:
R. Vakil, Enumerative Invariants in Algebraic Geo- R.L. Ricca (2009)
metry and String Theory. Cetraro, Italy 2005. Editors: Vol. 1974: F. den Hollander, Random Polymers: École
K. Behrend, M. Manetti (2008) d’Été de Probabilités de Saint-Flour XXXVII – 2007
Vol. 1948: F. Cao, J-L. Lisani, J-M. Morel, P. Musé, (2009)
F. Sur, A Theory of Shape Identification (2008) Vol. 1975: J.C. Rohde, Cyclic Coverings, Calabi-Yau
Vol. 1949: H.G. Feichtinger, B. Helffer, M.P. Lamoureux, Manifolds and Complex Multiplication (2009)
N. Lerner, J. Toft, Pseudo-Differential Operators. Quan- Vol. 1976: N. Ginoux, The Dirac Spectrum (2009)
tization and Signals. Cetraro, Italy 2006. Editors: L. Vol. 1977: M.J. Gursky, E. Lanconelli, A. Malchiodi,
Rodino, M.W. Wong (2008) G. Tarantello, X.-J. Wang, P.C. Yang, Geometric Analysis
Vol. 1950: M. Bramson, Stability of Queueing Networks, and PDEs. Cetraro, Italy 2001. Editors: A. Ambrosetti,
Ecole d’Eté de Probabilités de Saint-Flour XXXVI-2006 S.-Y.A. Chang, A. Malchiodi (2009)
(2008) Vol. 1978: M. Qian, J.-S. Xie, S. Zhu, Smooth Ergodic
Vol. 1951: A. Moltó, J. Orihuela, S. Troyanski, Theory for Endomorphisms (2009)
M. Valdivia, A Non Linear Transfer Technique for Vol. 1979: C. Donati-Martin, M. Émery, A. Rouault,
Renorming (2009) C. Stricker (Eds.), Śeminaire de Probablitiés XLII (2009)
Vol. 1952: R. Mikhailov, I.B.S. Passi, Lower Central and Vol. 1980: P. Graczyk, A. Stos (Eds.), Potential Analysis
Dimension Series of Groups (2009) of Stable Processes and its Extensions (2009)
Vol. 1953: K. Arwini, C.T.J. Dodson, Information Geo- Vol. 1981: M. Chlouveraki, Blocks and Families for
metry (2008) Cyclotomic Hecke Algebras (2009)
Vol. 1954: P. Biane, L. Bouten, F. Cipriani, N. Konno, Vol. 1982: N. Privault, Stochastic Analysis in Discrete
N. Privault, Q. Xu, Quantum Potential Theory. Editors: and Continuous Settings. With Normal Martingales
U. Franz, M. Schuermann (2008) (2009)
Vol. 1955: M. Bernot, V. Caselles, J.-M. Morel, Optimal Vol. 1983: H. Ammari (Ed.), Mathematical Modeling in
Transportation Networks (2008) Biomedical Imaging I. Electrical and Ultrasound Tomo-
Vol. 1956: C.H. Chu, Matrix Convolution Operators on graphies, Anomaly Detection, and Brain Imaging (2009)
Groups (2008) Vol. 1984: V. Caselles, P. Monasse, Geometric Descrip-
Vol. 1957: A. Guionnet, On Random Matrices: Macro- tion of Images as Topographic Maps (2010)
scopic Asymptotics, Ecole d’Eté de Probabilités de Saint- Vol. 1985: T. Linß, Layer-Adapted Meshes for Reaction-
Flour XXXVI-2006 (2009) Convection-Diffusion Problems (2010)
Vol. 1958: M.C. Olsson, Compactifying Moduli Spaces
for Abelian Varieties (2008) Recent Reprints and New Editions
Vol. 1959: Y. Nakkajima, A. Shiho, Weight Filtrations
on Log Crystalline Cohomologies of Families of Open Vol. 1702: J. Ma, J. Yong, Forward-Backward Stochas-
Smooth Varieties (2008) tic Differential Equations and their Applications. 1999 –
Vol. 1960: J. Lipman, M. Hashimoto, Foundations of Corr. 3rd printing (2007)
Grothendieck Duality for Diagrams of Schemes (2009) Vol. 830: J.A. Green, Polynomial Representations of
Vol. 1961: G. Buttazzo, A. Pratelli, S. Solimini, GLn , with an Appendix on Schensted Correspondence
E. Stepanov, Optimal Urban Networks via Mass Trans- and Littelmann Paths by K. Erdmann, J.A. Green and
portation (2009) M. Schoker 1980 – 2nd corr. and augmented edition
Vol. 1962: R. Dalang, D. Khoshnevisan, C. Mueller, (2007)
D. Nualart, Y. Xiao, A Minicourse on Stochastic Partial Vol. 1693: S. Simons, From Hahn-Banach to Monotonic-
Differential Equations (2009) ity (Minimax and Monotonicity 1998) – 2nd exp. edition
Vol. 1963: W. Siegert, Local Lyapunov Exponents (2009) (2008)
Vol. 1964: W. Roth, Operator-valued Measures and Inte- Vol. 470: R.E. Bowen, Equilibrium States and the Ergodic
grals for Cone-valued Functions and Integrals for Cone- Theory of Anosov Diffeomorphisms. With a preface by
valued Functions (2009) D. Ruelle. Edited by J.-R. Chazottes. 1975 – 2nd rev.
Vol. 1965: C. Chidume, Geometric Properties of Banach edition (2008)
Spaces and Nonlinear Iterations (2009) Vol. 523: S.A. Albeverio, R.J. Høegh-Krohn, S. Maz-
Vol. 1966: D. Deng, Y. Han, Harmonic Analysis on zucchi, Mathematical Theory of Feynman Path Integral.
Spaces of Homogeneous Type (2009) 1976 – 2nd corr. and enlarged edition (2008)
Vol. 1967: B. Fresse, Modules over Operads and Functors Vol. 1764: A. Cannas da Silva, Lectures on Symplectic
(2009) Geometry 2001 – Corr. 2nd printing (2008)
LECTURE NOTES IN MATHEMATICS 123
Edited by J.-M. Morel, F. Takens, B. Teissier, P.K. Maini

Editorial Policy (for the publication of monographs)

1. Lecture Notes aim to report new developments in all areas of mathematics and their
applications - quickly, informally and at a high level. Mathematical texts analysing new
developments in modelling and numerical simulation are welcome.
Monograph manuscripts should be reasonably self-contained and rounded off. Thus
they may, and often will, present not only results of the author but also related work
by other people. They may be based on specialised lecture courses. Furthermore, the
manuscripts should provide sufficient motivation, examples and applications. This clearly
distinguishes Lecture Notes from journal articles or technical reports which normally are
very concise. Articles intended for a journal but too long to be accepted by most journals,
usually do not have this “lecture notes” character. For similar reasons it is unusual for
doctoral theses to be accepted for the Lecture Notes series, though habilitation theses may
be appropriate.
2. Manuscripts should be submitted either online at www.editorialmanager.com/lnm to
Springer’s mathematics editorial in Heidelberg, or to one of the series editors. In general,
manuscripts will be sent out to 2 external referees for evaluation. If a decision cannot yet
be reached on the basis of the first 2 reports, further referees may be contacted: The author
will be informed of this. A final decision to publish can be made only on the basis of the
complete manuscript, however a refereeing process leading to a preliminary decision can
be based on a pre-final or incomplete manuscript. The strict minimum amount of material
that will be considered should include a detailed outline describing the planned contents
of each chapter, a bibliography and several sample chapters.
Authors should be aware that incomplete or insufficiently close to final manuscripts
almost always result in longer refereeing times and nevertheless unclear referees’ recom-
mendations, making further refereeing of a final draft necessary.
Authors should also be aware that parallel submission of their manuscript to another
publisher while under consideration for LNM will in general lead to immediate rejection.
3. Manuscripts should in general be submitted in English. Final manuscripts should contain
at least 100 pages of mathematical text and should always include
– a table of contents;
– an informative introduction, with adequate motivation and perhaps some historical re-
marks: it should be accessible to a reader not intimately familiar with the topic treated;
– a subject index: as a rule this is genuinely helpful for the reader.
For evaluation purposes, manuscripts may be submitted in print or electronic form (print
form is still preferred by most referees), in the latter case preferably as pdf- or zipped
ps-files. Lecture Notes volumes are, as a rule, printed digitally from the authors’ files.
To ensure best results, authors are asked to use the LaTeX2e style files available from
Springer’s web-server at:
ftp://ftp.springer.de/pub/tex/latex/svmonot1/ (for monographs) and
ftp://ftp.springer.de/pub/tex/latex/svmultt1/ (for summer schools/tutorials).
 Additional technical instructions, if necessary, are available on request from:
[email protected].
4. Careful preparation of the manuscripts will help keep production time short besides en-
suring satisfactory appearance of the finished book in print and online. After acceptance
of the manuscript authors will be asked to prepare the final LaTeX source files and also
the corresponding dvi-, pdf- or zipped ps-file. The LaTeX source files are essential for
producing the full-text online version of the book (see
http://www.springerlink.com/openurl.asp?genre=journal&issn=0075-8434 for the exist-
ing online volumes of LNM).
The actual production of a Lecture Notes volume takes approximately 12 weeks.
5. Authors receive a total of 50 free copies of their volume, but no royalties. They are entitled
to a discount of 33.3% on the price of Springer books purchased for their personal use, if
ordering directly from Springer.
6. Commitment to publish is made by letter of intent rather than by signing a formal contract.
Springer-Verlag secures the copyright for each volume. Authors are free to reuse material
contained in their LNM volumes in later publications: a brief written (or e-mail) request
for formal permission is sufficient.

Addresses:
Professor J.-M. Morel, CMLA,
École Normale Supérieure de Cachan,
61 Avenue du Président Wilson, 94235 Cachan Cedex, France
E-mail: [email protected]
Professor F. Takens, Mathematisch Instituut,
Rijksuniversiteit Groningen, Postbus 800,
9700 AV Groningen, The Netherlands
E-mail: [email protected]
Professor B. Teissier, Institut Mathématique de Jussieu,
UMR 7586 du CNRS, Équipe “Géométrie et Dynamique”,
175 rue du Chevaleret,
75013 Paris, France
E-mail: [email protected]

For the “Mathematical Biosciences Subseries” of LNM:

Professor P.K. Maini, Center for Mathematical Biology,
Mathematical Institute, 24-29 St Giles,
Oxford OX1 3LP, UK
E-mail: [email protected]
Springer, Mathematics Editorial, Tiergartenstr. 17,
69121 Heidelberg, Germany,
Tel.: +49 (6221) 487-259
Fax: +49 (6221) 4876-8259
E-mail: [email protected]