...

Kuzmin-Transport2010.pdf

by user

on
Category: Documents
9

views

Report

Comments

Transcript

Kuzmin-Transport2010.pdf
A Guide to Numerical Methods
for Transport Equations
Dmitri Kuzmin
2010
Contents
1
Getting Started . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1 Introduction to Flow Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Mathematics of Transport Phenomena . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.1 Conservation Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.2 Convective and Diffusive Fluxes . . . . . . . . . . . . . . . . . . . . . . .
1.2.3 The Generic Transport Equation . . . . . . . . . . . . . . . . . . . . . . . .
1.2.4 Initial and Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . .
1.2.5 Weighted Residual Formulation . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Taxonomy of Reduced Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.1 Elliptic Transport Equations . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.2 Hyperbolic Transport Equations . . . . . . . . . . . . . . . . . . . . . . . .
1.3.3 Parabolic Transport Equations . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.4 Summary of Model Problems . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4 Space Discretization Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.1 Computational Meshes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.2 Semi-Discrete Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.3 Finite Difference Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.4 Finite Volume Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.5 Finite Element Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5 Systems of Algebraic Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5.1 Time-Stepping Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5.2 Direct vs. Iterative Solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5.3 Explicit vs. Implicit Schemes . . . . . . . . . . . . . . . . . . . . . . . . . .
1.6 Fundamental Design Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.6.1 Numerical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.6.2 Physical Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.6.3 The Basic Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.7 Scope of This Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
1
3
4
5
7
9
9
11
11
12
13
14
15
15
17
19
20
23
25
26
27
29
30
31
33
34
38
v
vi
Contents
2
Finite Element Approximations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1 Discretization on Unstructured Meshes . . . . . . . . . . . . . . . . . . . . . . . .
2.1.1 Group Finite Element Formulation . . . . . . . . . . . . . . . . . . . . . .
2.1.2 Properties of Discrete Operators . . . . . . . . . . . . . . . . . . . . . . . .
2.1.3 Conservation and Mass Lumping . . . . . . . . . . . . . . . . . . . . . . .
2.1.4 Variational Gradient Recovery . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.5 Treatment of Nonlinear Fluxes . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.6 Conservative Flux Decomposition . . . . . . . . . . . . . . . . . . . . . .
2.1.7 Relationship to Finite Volumes . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.8 Edge-Based Data Structures . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.9 Compressed Row Storage . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Stabilization of Convective Terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.1 First-Order Upwinding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.2 Artificial Diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.3 Streamline Upwinding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.4 Petrov-Galerkin Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.5 Taylor-Galerkin Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.6 Discontinuity Capturing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.7 Interior Penalty Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.8 Modulated Dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Discontinuous Galerkin Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.1 Upwind DG Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.2 Taylor Basis Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.3 The Barth-Jespersen Limiter . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.4 The Vertex-Based Limiter . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.5 Limiting Higher-Order Terms . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
41
41
43
46
48
51
53
54
59
61
64
67
67
68
70
71
72
77
79
80
84
84
85
87
88
89
90
3
Maximum Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
3.1 Properties of Linear Transport Models . . . . . . . . . . . . . . . . . . . . . . . . . 91
3.1.1 The Laplace Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.1.2 Equations of Elliptic Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
3.1.3 Equations of Hyperbolic Type . . . . . . . . . . . . . . . . . . . . . . . . . . 98
3.1.4 Equations of Parabolic Type . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.1.5 Singularly Perturbed Problems . . . . . . . . . . . . . . . . . . . . . . . . . 105
3.2 Matrix Analysis for Steady Problems . . . . . . . . . . . . . . . . . . . . . . . . . . 106
3.2.1 The Discrete Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
3.2.2 M-Matrices and Monotonicity . . . . . . . . . . . . . . . . . . . . . . . . . 108
3.2.3 Discrete Maximum Principles . . . . . . . . . . . . . . . . . . . . . . . . . . 110
3.2.4 Desirable Mesh Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
3.3 Matrix Analysis for Unsteady Problems . . . . . . . . . . . . . . . . . . . . . . . . 115
3.3.1 Semi-Discrete DMP Constraints . . . . . . . . . . . . . . . . . . . . . . . . 115
3.3.2 Fully Discrete DMP Constraints . . . . . . . . . . . . . . . . . . . . . . . . 118
3.3.3 Positive Time-Stepping Methods . . . . . . . . . . . . . . . . . . . . . . . 120
3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
Contents
vii
4
Algebraic Flux Correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.1 Nonlinear High-Resolution Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.1.1 Design Philosophy and Tools . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.1.2 Artificial Diffusion Operators . . . . . . . . . . . . . . . . . . . . . . . . . . 129
4.1.3 Conservative Flux Decomposition . . . . . . . . . . . . . . . . . . . . . . 133
4.1.4 Limited Antidiffusive Correction . . . . . . . . . . . . . . . . . . . . . . . 134
4.1.5 The Generic Limiting Strategy . . . . . . . . . . . . . . . . . . . . . . . . . 136
4.1.6 Summary of Algorithmic Steps . . . . . . . . . . . . . . . . . . . . . . . . . 138
4.2 Solution of Nonlinear Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
4.2.1 Successive Approximations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
4.2.2 Defect Correction Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
4.2.3 Underrelaxation and Smoothing . . . . . . . . . . . . . . . . . . . . . . . . 142
4.2.4 Positivity-Preserving Solvers . . . . . . . . . . . . . . . . . . . . . . . . . . 144
4.2.5 Accuracy vs. Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
4.3 Steady Transport Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
4.3.1 Upwind-Biased Flux Correction . . . . . . . . . . . . . . . . . . . . . . . . 148
4.3.2 Relationship to TVD Limiters . . . . . . . . . . . . . . . . . . . . . . . . . . 151
4.3.3 Gradient-Based Slope Limiting . . . . . . . . . . . . . . . . . . . . . . . . 152
4.3.4 Reconstruction of Local Stencils . . . . . . . . . . . . . . . . . . . . . . . 154
4.3.5 Background Dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
4.3.6 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
4.4 Unsteady Transport Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
4.4.1 Nonlinear FEM-FCT Schemes . . . . . . . . . . . . . . . . . . . . . . . . . 165
4.4.2 Zalesak’s Limiter Revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
4.4.3 Flux Linearization Techniques . . . . . . . . . . . . . . . . . . . . . . . . . 170
4.4.4 Predictor-Corrector Algorithms . . . . . . . . . . . . . . . . . . . . . . . . 171
4.4.5 Positive Time Integrators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
4.4.6 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
4.5 Limiting for Diffusion Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
4.5.1 The Galerkin Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
4.5.2 Positive-Negative Splitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
4.5.3 Symmetric Slope Limiter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
4.5.4 Treatment of Nonlinearities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
4.5.5 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
5
Error Estimates and Adaptivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
5.2 Galerkin Weak Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
5.3 Global Error Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
5.4 Local Error Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
5.5 Numerical Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
Chapter 1
Getting Started
In this chapter, we start with a brief introduction to numerical simulation of transport
phenomena. We consider mathematical models that express certain conservation
principles and consist of convection-diffusion-reaction equations written in integral,
differential, or weak form. In particular, we discuss the qualitative properties of
exact solutions to model problems of elliptic, hyperbolic, and parabolic type. Next,
we review the basic steps involved in the design of numerical approximations and
the main criteria that a reliable algorithm should satisfy. The chapter concludes with
an outline of the rationale behind the scope and structure of the present book.
1.1 Introduction to Flow Simulation
Fluid dynamics and transport phenomena, such as heat and mass transfer, play a
vitally important role in human life. Gases and liquids surround us, flow inside our
bodies, and have a profound influence on the environment in which we live. Fluid
flows produce winds, rains, floods, and hurricanes. Convection and diffusion are responsible for temperature fluctuations and transport of pollutants in air, water or soil.
The ability to understand, predict, and control transport phenomena is essential for
many industrial applications, such as aerodynamic shape design, oil recovery from
an underground reservoir, or multiphase/multicomponent flows in furnaces, heat exchangers, and chemical reactors. This ability offers substantial economic benefits
and contributes to human well-being. Heating, air conditioning, and weather forecast have become an integral part of our everyday life. We take such things for
granted and hardly ever think about the physics and mathematics behind them.
The traditional approach to investigation of a physical process is based on observations, experiments, and measurements. The amount of information that can
be obtained in this way is usually very limited and subject to measurement errors.
Moreover, experiments are only possible when a small-scale model or the actual
equipment has already been built. An experimental investigation may be very timeconsuming, dangerous, prohibitively expensive, or impossible for another reason.
1
2
1 Getting Started
Alternatively, an analytical or computational study can be performed on the basis
of a suitable mathematical model. As a rule, such a model consists of several differential and/or algebraic equations which make it possible to predict how the quantities of interest evolve and interact with one another. A drawback to this approach is
the fact that complex physical phenomena give rise to complex mathematical equations that cannot be solved analytically, i.e., using paper and pencil.
The most detailed models of fluid flow are based on ‘first principles’, such as the
conservation of mass, momentum, and energy. Mathematical equations that embody
these fundamental principles have been known for a very long time but used to be
practically worthless until numerical methods and digital computers were invented.
The second half of the twentieth century has witnessed the advent of Computational
Fluid Dynamics (CFD), a new branch of applied mathematics that deals with numerical simulation of fluid flows. Nowadays, computer codes based on CFD models are
used routinely to predict a variety of increasingly complex flow phenomena.
The quality of simulation results depends on the choice of the model and on the
accuracy of the numerical method. In spite of the inevitable numerical and modeling
errors, approximate solutions may provide a lot of valuable information at a fraction
of the cost that a full-scale experimental investigation would require. Moreover, the
sampling of relevant data is free of errors due to a flow disturbance caused by probes.
A further advantage of the computational approach is the fact that it can be applied to
flows in domains with arbitrarily large or small dimensions under realistic operating
conditions. High pressures, toxic chemicals or hot temperatures pose no hazard to
a CFD practitioner. Last but not least, simultaneous computation of instantaneous
density, velocity, pressure, temperature, and concentration fields is feasible. Clearly,
no experimental technique can capture the evolution of all flow variables throughout
the domain. However, experiments are still required to determine the values of input
parameters for a mathematical model and to validate the computational results.
The choice of a CFD model is dictated by the nature of the physical process to be
simulated, by the objectives of the numerical study, and by the available resources.
As a rule of thumb, the mathematical model should be as detailed as possible without making the computations too expensive. The use of a universally applicable
model makes it difficult to develop and implement an efficient numerical algorithm.
In many cases, the desired information can be obtained using a simplified version
that exploits some a priori knowledge of the flow pattern or incorporates empirical correlations supported by theoretical or experimental studies. Thus, a hierarchy
of fundamental, phenomenological, and empirical models is usually available for
particularly difficult problems, such as the numerical simulation of turbulence.
Over the past three decades, the market for CFD software has expanded rapidly,
and remarkable progress has been made in the development of numerical algorithms. An astonishing variety of finite difference, finite element, finite volume, and
spectral schemes were developed for the equations of fluid mechanics and applied
to virtually every flow problem of practical importance. Modern CFD codes are
equipped with automatic mesh generation/adaptation tools, reliable error control
mechanisms, and efficient iterative solvers for sparse linear systems. Unstructured
mesh methods are available for flows in complex geometries. Problems with moving
1.2 Mathematics of Transport Phenomena
3
boundaries and free interfaces can be solved in a fixed or moving reference frame.
Parallelization and vectorization make it possible to perform large-scale computations with more than a billion of degrees of freedom. The rapid growth of computing power has stimulated implementation of sophisticated models and extended the
range of possible applications to problems as complex as turbulent multiphase flows
and fluid-structure interaction. Nowadays, 3D simulations of unsteady transport processes can be performed on a laptop or desktop computer, whereas supercomputers
were required to simulate steady 2D problems a couple of decades ago.
If something sounds too good to be true, it probably is. In spite of the abovementioned recent advances, there is still a lot of room of improvement when it
comes to reliable simulation of transport phenomena. The user of a commercial
CFD code might be unaware of the numerous subtleties, trade-offs, compromises,
and ad hoc tricks involved in the computation of beautiful colorful pictures. Usually, there is no guarantee that these pictures are quantitatively correct. If the same
problem is solved using another mesh, another time step, and/or another numerical
scheme, then a qualitatively different solution may be obtained. Hence, the results of
a CFD simulation should not be taken at their face value even if they look ‘nice’ and
plausible. In other cases, the approximate solution may exhibit spurious oscillations
and/or assume nonphysical negative values. This behavior is typical of problems
with discontinuities and steep fronts that cannot be resolved properly on a given
mesh. Therefore, it might be necessary to refine the mesh and/or adjust the coefficients of the numerical scheme if nonphysical solution behavior is detected. Ideally,
the numerical algorithm should do it automatically by adapting itself to the nature
of the problem at hand so as to compute accurate solutions in an efficient way. The
goal of this book is to present a general approach to the design of such algorithms.
1.2 Mathematics of Transport Phenomena
In Part I, we dwell on the numerical treatment of differential equations that govern
the evolution of scalar fluid properties. The derivation of these equations is usually
based on certain conservation principles, as applied to an arbitrary control volume
V ⊂ Rd , where d = 1, 2, or 3 is the number of space dimensions. If the fluid is in
motion, it may flow in and out across the control surface S which forms the boundary
of V , see Fig. 1. Individual molecules may travel across the interface even if the fluid
is at rest. Therefore, the physical and chemical properties of the fluid inside V are
influenced by those of the surrounding medium. Moreover, some quantities, such
as mass, momentum, and energy, are conserved. That is, they may move from one
place to another but cannot emerge out of nothing or disappear spontaneously. The
physical forces that transport, produce or destroy these quantities are well-known,
and reliable mathematical models are available. Thus, conservation principles can
be expressed in terms of differential equations that describe all relevant transport
mechanisms, such as convection (also called advection), diffusion, and dispersion.
4
1 Getting Started
1.2.1 Conservation Principles
Let c(x,t) ∈ R denote the concentration (amount per unit mass) of a scalar conserved
quantity at point x ∈ V and time t ≥ 0. The corresponding concentration per unit
volume is given by u = ρ c, where ρ is the density of the carrier fluid. The total
amount of the conserved variable inside V is given by the volume integral
Z
u(x,t) dx =
V
Z
V
ρ (x,t)c(x,t) dx.
(1.1)
We will call this integral the mass and speak of mass conservation even if c represents the energy, a single velocity component, or another dimensional quantity.
Obviously, the variation of (1.1) depends on the rate at which c enters or leaves
V through the boundary S. This rate is called the flux and denoted by
f(x,t) = ( f 1 , . . . , f d ),
where f k corresponds to the rate of transport in the k−th coordinate direction, per
unit area and time. If ds = n ds is an infinitesimally small patch of S with the unit
outward normal n, then the mass crossing this patch per unit time is f · n ds.
In the simplest case, the flux vector f is a linear function of u and/or ρ ∇c, where
∇=
∂
∂
,...,
∂ x1
∂ xd
T
is the vector of partial derivatives that defines the gradient and divergence operators.
Chemical reactions, heating, cooling, and similar processes give rise to interior
sources or sinks that generate s(x,t) units of mass per unit volume and time. Thus,
the temporal variation of (1.1) satisfies an integral conservation law of the form
∂
∂t
Z
V
u(x,t) dx +
Z
S
f · n ds =
Z
s(x,t) dx.
V
Fig. 1.1 A fixed control volume V bounded by the control surface S.
(1.2)
1.2 Mathematics of Transport Phenomena
5
The surface integral is the mass that leaves V per unit area and time, whereas the
right-hand side of (1.2) corresponds to the mass produced inside V per unit time.
If the functions u(x,t) and f(x,t) are differentiable, then the divergence theorem,
as applied to the surface integral in (1.2), yields the identity
Z ∂ u(x,t)
+ ∇ · f(x,t) − s(x,t) dx = 0.
∂t
V
Since the choice of V is arbitrary, the expression in the square brackets must vanish,
so the evolution of u(x,t) is governed by the partial differential equation (PDE)
∂ u(x,t)
+ ∇ · f(x,t) = s(x,t).
∂t
(1.3)
If the divergence theorem is applicable, this differential form of the conservation
law is equivalent to the underlying integral form (1.2). However, the latter is more
fundamental since it does not contain any space derivatives. If the flux f(x,t) does
not depend on the gradients of c, then the generalized solution may exhibit very
steep gradients or even discontinuities. Such solutions satisfy (1.2) but not (1.3)
since discontinuous functions are not differentiable in the classical sense.
1.2.2 Convective and Diffusive Fluxes
The modeling of the flux function f should reflect the nature of the involved transport
processes. Convective effects arise when fluids flow and transport the quantities
of interest downstream. For example, consider a horizontal pipe filled with water
which flows from left to right at constant speed. In experimental studies, the flow
pattern is commonly visualized by tracking a set of small tracer particles which are
convected with the flow as time goes on. Suppose that some particles are white,
while others are black. The distribution of particles at time t = 0 is displayed in
Fig. 1.2a. Since the water is in motion, it carries the suspension of tracer particles
towards the outlet. If we keep injecting black particles at the left end of the pipe,
(a)
(b)
Fig. 1.2 Transport of tracer particles in a pipe filled with moving water.
6
1 Getting Started
a snapshot of the particle distribution at a later time t > 0 might look as depicted
in Fig. 1.2b. Similarly, if we vary the temperature of the water at the inlet, this will
affect the temperature distribution inside the pipe and, eventually, at the outlet. This
effect is utilized in many heating and cooling devices that we use in everyday life.
Mass and heat may also be transported from one place to another by diffusion or
heat conduction, respectively. Random molecular motion induces diffusive fluxes
even if the fluid is at rest. To illustrate this process, consider a tank filled with liquid
in which two distinct chemical species are dissolved. In Fig. 1.3, the black and
white circles represent the molecules of species A and B, respectively. Initially, these
species are separated by a diaphragm that divides the tank into two parts (Fig. 1.3a).
When the diaphragm is removed, some white molecules may cross the interface
and end up in the left half of the tank. Conversely, black ones may travel in the
opposite direction and end up in the right half (Fig. 1.3b). After a certain time, the
mixture will become homogeneous, and each half of the tank will contain the same
number of black and white molecules (Fig. 1.3c). In the context of central heating,
convection transports hot water into the radiator but heat transfer inside the room is
of a diffusive nature since it is driven primarily by the temperature gradients.
(a) initial state
(b) intermediate state
(c) equilibrium state
Fig. 1.3 Random motion of molecules across an interface in a stationary liquid.
In general, the transport of conserved quantities from regions of high concentration into regions of low concentration may be caused by random molecular motion or turbulence. Molecular diffusion represents the natural tendency of a physical
system towards an equilibrium, whereas turbulent dispersion is due to unresolved
eddies that enhance the macroscopic mixing rate. The corresponding mathematical
models look the same but the coefficients differ by orders of magnitude. In what
follows, both molecular and turbulent mixing will be referred to as ‘diffusion.’
Let us now describe the above transport processes in terms of formulas rather
than words. Assume that the velocity field v(x,t) is known. The volume of fluid that
crosses an infinitesimally small patch ds = n ds during a short time interval dt is
dV = (v · n ds)dt.
1.2 Mathematics of Transport Phenomena
7
Since u = ρ c is the mass of a conserved quantity per unit volume, the amount of
mass transported in the normal direction n per unit area and time is given by
u dV
= (v · n)u.
dsdt
If n is taken to be the unit vector along the coordinate direction xd , the above expression yields the d−th component of the convective flux
fC = v(x,t)u.
(1.4)
Since diffusion-like processes are driven by the gradients of the concentration
field, a typical model for the corresponding flux vector is as follows
fD = −D(x,t)ρ ∇c,
(1.5)
where D = {di j } is a symmetric positive definite matrix of diffusion coefficients. If
D = dI, where I is the D × D identity matrix, then the scalar diffusivity d(x,t) > 0
is the same for all coordinate directions, and the diffusive flux reduces to
fD (x,t) = −d(x,t)ρ ∇c.
(1.6)
In the realm of mass and heat transfer, this definition of fD follows from Fick’s law
of mass diffusion and Fourier’s law of heat conduction, respectively.
In general, both convective and diffusive effects must be taken into account, so
f(x,t) = v(x,t)u − D(x,t)ρ ∇c.
(1.7)
However, the rates of convective and diffusive transport may be quite different. For
example, the transport of pollutants in a river is dominated by convection, whereas
the spreading of pollutants in a lake is dominated by diffusion (dispersion).
The relative strength of fC and fD can be expressed in terms of the Peclet number
Pe =
v0 L0
,
d0
(1.8)
where v0 is a reference velocity, L0 is a geometric length scale, and d0 is a diffusion
coefficient. The dimensionless Peclet number is infinite in the limit of pure convection (f = fC , D = 0) and vanishes in the limit of pure diffusion (f = fD , v = 0).
1.2.3 The Generic Transport Equation
Substitution of u = ρ c and (1.7) into (1.3) yields the generic transport equation
∂ ρc
+ ∇ · (vρ c) − ∇ · (D ρ ∇c) = s.
∂t
(1.9)
8
1 Getting Started
The terms that appear in this equation admit the following physical interpretation
•
•
•
•
the rate-of-change term ∂∂ρtc is the net gain/loss of mass per unit volume and time;
the convective term ∇ · (vρ c) is due to the downstream transport with velocity v;
the diffusive term −∇ · (D ρ ∇c) is due to a nonuniform spatial distribution of c;
the source or sink term s combines all other effects that create or destroy ρ c.
For the time being, we assume that the parameters ρ , v, D, and s are known. In reallife applications, they may depend on the concentration c and/or other variables.
In particular, conservation laws of the form (1.9) constitute the Navier-Stokes
equations, in which the conserved variables are the mass, momentum, and total
energy. The simplest component of this PDE system is the continuity equation
∂ρ
+ ∇ · (ρ v) = 0
∂t
(1.10)
which is responsible for mass conservation and corresponds to (1.9) with c ≡ 1 and
s = 0. Note that the diffusive term vanishes since the gradient of c is zero. If viscosity
and heat conduction are neglected, then the Navier-Stokes equations reduce to the
Euler equations that describe inviscid gas flows at high speeds. Advanced CFD
models based on the Euler and Navier-Stokes equations will be treated in Part II.
The common structure of mathematical models which are based on (systems of)
scalar conservation laws of the form (1.9) suggests a systematic approach to analysis, discretization, and coding. This strategy facilitates the development, implementation, and testing of numerical methods for advanced CFD applications. In addition
to the conceptual and algorithmic simplicity, it offers a simple way to investigate the
solution behavior in important limiting cases (steady state, pure convection, pure
diffusion etc.) and design simple test problems that can be solved analytically.
The generic transport equation (1.9) can also be written in terms of u = ρ c. If the
density ρ is constant or the velocity is redefined as v := v + (D∇ρ )/ρ , then (1.9) is
a convection-diffusion-reaction (CDR) equation for the mass variable u
∂u
+ ∇ · (vu) − ∇ · (D∇u) = s.
∂t
(1.11)
This equation and some simplifications thereof will serve as basic models in Part I.
If the velocity field v is incompressible, that is, ∇ · v = 0, then the vector identity
∇ · (vu) = v · ∇u + (∇ · v)u
(1.12)
makes it possible to write the left-hand side of (1.11) in the nondivergent form
∂u
+ v · ∇u − ∇ · (D∇u) = s.
∂t
(1.13)
Another possibility is to take the average of (1.11) and (1.13), which gives a skewsymmetric form of the convective term. All three formulations are equivalent for
divergence-free velocity fields but only (1.11) is conservative for ∇ · v 6= 0.
1.2 Mathematics of Transport Phenomena
9
1.2.4 Initial and Boundary Conditions
The same differential equation may describe an amazing variety of flow patterns, so
some additional information is required to complete the problem statement. In practical applications, the processes to be investigated take place in a concrete geometry
(e.g., in turbines, chemical reactors, heat exchangers, car engines etc.) during a finite
interval of time. The choice of the domain and of the time interval to be considered
is dictated by the nature of the problem at hand, by the objectives of the analytical
or numerical study, and by the available resources. Another important aspect is the
choice of initial and/or boundary conditions that lead to a well-posed problem.
Let Ω ⊂ Rd be a bounded domain and (0, T ) be the time interval of interest. In
general, the boundary Γ of Ω may consist of an inflow part Γ− = {x ∈ Γ | v · n < 0},
an outflow part Γ+ = {x ∈ Γ | v · n > 0}, and a solid wall Γ0 = {x ∈ Γ | v · n = 0},
where n denotes the unit outward normal to the boundary at the point x ∈ Γ .
Since the CDR equation contains a time derivative, it must be supplemented by
an initial condition that defines the distribution of mass at t = 0
u(x, 0) = u0 (x),
∀x ∈ Ω .
(1.14)
Furthermore, the fluid inside Ω interacts with the surrounding medium, so it is also
necessary to prescribe suitable boundary conditions on Γ . If the values of u are
known on ΓD ⊂ Γ , they can be imposed as Dirichlet boundary conditions
u(x,t) = uD (x,t),
∀x ∈ ΓD , ∀t ∈ (0, T ).
(1.15)
As a rule, this boundary condition is used at the inlet Γ− and/or on the solid wall Γ0 .
Alternatively, a given normal flux may be prescribed on the complementary
boundary part ΓN = Γ \ΓD . The so-defined Neumann boundary condition reads
f · n = g(x,t),
∀x ∈ ΓN , ∀t ∈ (0, T ).
(1.16)
The involved flux f may consist of a convective and/or a diffusive part, depending
on the information available. If f = fC or the diffusive flux fD is required to vanish,
then the right-hand side of (1.16) is given by g = (v · n)u on Γ± and g = 0 on Γ0 .
1.2.5 Weighted Residual Formulation
The classical solution u of the CDR equation must belong to the space of functions
which are continuous with continuous partial derivatives of first and second order.
In other words, it must be very smooth. In order to broaden the class of admissible
functions, it is worthwhile to consider an integral or weak form of the conservation
law. The corresponding generalized solution is supposed to satisfy the strong form
of (1.11) for sufficiently smooth data but exist even if the divergence theorem is not
applicable and the underlying conservation law holds only in an integral sense.
10
1 Getting Started
A very general approach to the derivation of weak forms for a given PDE is called
the method of weighted residuals. The residual of equation (1.11) is defined as
∂ ū
+ ∇ · (vū) − ∇ · (D∇ū) − s
∂t
R(ū) =
(1.17)
so that R(u) = 0 if u is the exact solution of the CDR equation. Thus, the magnitude
of the residual R(ū) measures the accuracy of an approximate solution ū ≈ u.
Obviously, a zero residual remains unchanged if we multiply it by a suitable
weighting (or test) function and integrate over the domain of interest. Hence,
Z
Ω
wR(u) dx = 0,
∀w ∈ W ,
(1.18)
where W is a space of weighting functions vanishing on ΓD . Mathematically speaking, the residual R(u) must be orthogonal to all w ∈ W . The weak solution u resides
in a space V of functions satisfying the Dirichlet boundary conditions (1.15).
Since the residual of (1.11) is given by (1.17), the associated weak form reads
Z
∂u
+ ∇ · (vu) − ∇ · (D∇u) − s dx = 0,
∀w ∈ W .
(1.19)
w
∂t
Ω
If the number of test functions is infinite, formulations (1.19) and (1.11) are equivalent. Otherwise, the residual R(u) may be nonzero even if u satisfies (1.19).
The rationale for the use of a weighted residual formulation is the possibility
to shift some derivatives onto the test function w using integration by parts. The
Green’s formula, as applied to the diffusive term in (1.19), yields
Z
Z ∂u
w + w∇ · (vu) + ∇w · (D∇u) − ws dx −
w(D∇u) · n ds = 0. (1.20)
∂t
Ω
ΓN
Since w = 0 on ΓD , the surface integral is taken over the boundary part ΓN = Γ \ΓD .
The functions u ∈ V and w ∈ W are required to possess generalized derivatives
of first order. If the convective term is also integrated by parts, one obtains
Z Z
∂u
w − ∇w · (vu − D∇u) − ws dx +
w(vu − D∇u) · n ds = 0. (1.21)
∂t
Ω
ΓN
The surface integrals that pop up in (1.20) and (1.21) contain the normal components
of the diffusive and total flux, respectively. Since Neumann boundary conditions of
the form (1.16) are prescribed on ΓN , the corresponding integral is given by
Z
ΓN
w f · n ds =
Z
ΓN
wg ds.
(1.22)
Thus, flux boundary conditions fit naturally into the weak form of the transport
equation. Dirichlet boundary conditions (1.15) are imposed in a strong sense, i.e.,
they are built into the definition of the spaces V and W in which u and w reside.
1.3 Taxonomy of Reduced Models
11
The elegance and generality of the weighted residual method lies in the choice
of the test functions w. For example, if we substitute w ≡ 1 and ΓN = Γ into (1.21),
then the integral form (1.2) of the CDR equation is recovered
∂
∂t
Z
Ω
u dx +
Z
Γ
(vu − D∇u) · n ds =
Z
Ω
s dx.
(1.23)
On the other hand, it is possible to enforce (1.11) in a strong sense by setting the
residual to zero at a collocation point x0 ∈ Ω . To this end, we substitute the Dirac
delta function w(x) = δ (x − x0 ) into (1.19) and obtain the pointwise identity
∂u
+ ∇ · (vu) − ∇ · (D∇u) = s
∂t
at
x = x0 .
Hence, the weighted residual formulation unites the integral, differential, and weak
forms of the conservation law. The existence of several equivalent representations,
as presented above, makes it possible to choose the one which is easier to handle for
a given choice of functional spaces and boundary conditions. This flexibility turns
out to be very useful when it comes to the design of numerical approximations.
1.3 Taxonomy of Reduced Models
The behavior of analytical and numerical solutions to the strong or weak form
of (1.11) depends on the interplay of the four terms that appear in this equation.
The time derivative may be large for transient transient processes but vanish in the
steady-state limit. The relative importance of convective and diffusive effects depends on the Peclet number (1.8). The reactive and diffusive terms are zero in the
continuity equation (1.10) but may be dominant in other transport models.
Many useful model problems can be constructed on the basis of the CDR equation by omitting some terms and/or making additional assumptions. These simplifications may affect the type of the partial differential equation, the choice of initial and boundary conditions, the qualitative properties of exact solutions, and the
performance of numerical schemes. A good algorithm must be sufficiently robust,
accurate, and efficient for all possible manifestations of the transport equation.
1.3.1 Elliptic Transport Equations
If the convective and diffusive fluxes are in equilibrium with the source term, then
the time derivative of the transported quantity vanishes, and (1.11) reduces to
∇ · (vu − D∇u) = s
in Ω .
(1.24)
12
1 Getting Started
This second-order PDE is of elliptic type provided that the matrix of diffusion coefficients D(x) is symmetric positive definite for all x ∈ Ω . In elliptic problems, information propagates in all directions at infinite speed. The variation of u at any point
x1 ∈ Ω may influence the solution at any other point x2 ∈ Ω and vice versa. Boundary conditions of Dirichlet or Neumann type are to be prescribed on Γ = ΓD ∪ ΓN ,
whereas no initial conditions are required for stationary problems like (1.24).
Elliptic CDR equations describe equilibrium transport phenomena and may represent the steady-state limit of a transient process. Indeed, if the velocity field, the
diffusion coefficients, and the boundary conditions do not depend on time, then the
solution of (1.11) will eventually become stationary and satisfy equation (1.24).
Steady diffusion-reaction processes are described by equation (1.24) with v = 0
in Ω .
−∇ · (D∇u) = s
(1.25)
Elliptic PDEs of this form play an important role, e.g., in mathematical modeling
of flows in porous media. In this context, the relationship v = −D∇u is called the
Darcy law, in which v and u represent the velocity and pressure fields, respectively.
If the diffusion tensor is defined as D = dI, where d is a constant diffusion coefficient, then (1.25) divided by d yields the following Poisson equation
−∆ u = f ,
in Ω ,
(1.26)
where ∆ = ∇2 denotes the Laplacian operator and f = s/d. The Laplace equation
∆ u = 0,
in Ω
(1.27)
is recovered for f = 0. In particular, it can be used to compute the potential of the
velocity field v = −∇u for incompressible irrotational flows (∇ · v = 0, ∇ × v = 0).
1.3.2 Hyperbolic Transport Equations
The next model problem to be considered is that of purely convective transport. For
D = 0, equation (1.24) degenerates into a hyperbolic PDE of first order
∇ · (vu) = s
in Ω .
(1.28)
In this case, information is transported at finite speeds along the streamlines of
the stationary velocity field v(x). The nature of hyperbolic problems requires that
boundary conditions be specified only on the inflow part Γ− , where v·n < 0. It would
be inappropriate and incorrect to prescribe any boundary conditions elsewhere.
The unsteady version of the convection-reaction equation (1.28) is given by
∂u
+ ∇ · (vu) = s
∂t
in Ω × (0, T ).
(1.29)
1.3 Taxonomy of Reduced Models
13
The most prominent example is the continuity equation (1.10) with u = ρ and s = 0.
Like any other PDE of first order, equation (1.29) is of hyperbolic type. The
direction and speed of convective transport depend on the velocity field v(x,t). As
before, boundary conditions are to be prescribed only at the inlet Γ− . The initial
condition is given by (1.14), and information travels forward in time. That is, the
distribution of u at any time instant t¯ depends on the previous evolution history but
only the solution at a later time may be influenced by what happens at t = t¯.
Analytical solutions to (1.28) and (1.29) can be constructed by the method of
characteristics to be presented in subsequent chapters. Due to the lack of diffusive
effects, hyperbolic conservation laws admit discontinuous and, possibly, nonunique
weak solutions. Such problems are particularly difficult to solve numerically, although a lot of information about the properties of exact solutions is available.
1.3.3 Parabolic Transport Equations
If the fluid is at rest, then the contribution of the convective term to the CDR equation
(1.11) vanishes, so the evolution of u is driven by diffusion and reaction
∂u
− ∇ · (D∇u) = s
∂t
in Ω × (0, T ).
(1.30)
If diffusion is isotropic then D∇u = d∇u and the above equation assumes the form
∂u
− d∆ u = s
∂t
in Ω × (0, T ).
(1.31)
This model describes unsteady transport processes like mass diffusion or heat conduction. The redistribution of u continues until the time derivative vanishes, and the
solution of the elliptic Poisson equation (1.26) is obtained in the steady state limit.
Unsteady partial differential equations of second order are of parabolic type if
their stationary counterparts are elliptic. Both initial and boundary conditions of the
form (1.14)–(1.16) are required for time-dependent transport models based on the
parabolic equations (1.11), (1.30), and (1.31). As in the case of hyperbolic PDEs,
information propagates forward in time, and there is no backward influence.
Steady CDR equations with v 6= 0 can also be parabolic if there is a predominant
flow direction, and the diffusive transport in this direction is neglected. For example,
let v = (1, 0, 0) and D = diag{0, d, d}. Then equation (1.24) can be written as
2
∂u
∂ u ∂ 2u
+
= s,
(1.32)
−d
∂ x1
∂ x22 ∂ x32
where x1 represents a time-like coordinate such that information is convected downstream, and no recirculation takes place. This problem is parabolic and exhibits the
same structure as the unsteady diffusion-reaction equation (1.30)–(1.27) in 2D.
14
1 Getting Started
1.3.4 Summary of Model Problems
As we have seen, the CDR equation (1.11) represents a rich variety of model problems. If the time derivative, convection, or diffusion are neglected, the resulting
equation looks simpler than (1.11). Ironically, it may be more difficult to treat numerically. In particular, computation of stationary solutions is hard, unless a good
initial guess is available. Therefore, it is common practice to march solutions to
the steady state by solving the corresponding unsteady PDE subject to arbitrary
initial conditions. Similarly, the lack of diffusive terms in hyperbolic models may
adversely affect the performance of a numerical scheme designed to solve (1.11).
Partial differential equations and numerical methods are easier to analyze in one
space dimension. In the 1D case, the operator ∇ reduces to the partial derivative
with respect to x, the matrix D degenerates into a scalar diffusion coefficient d, and
the unsteady convection-reaction-diffusion equation (1.11) assumes the form
∂u
∂u
∂ 2u
+ v − d 2 = s.
∂t
∂x
∂x
Again, it can be used to generate model problems for the whole range of possible
PDE types. The taxonomy of reduced transport models is summarized in Table 1.1.
If reaction is not important, we set s = 0. This does not change the PDE type. The
presented models will help us to develop, evaluate, and compare numerical solution
techniques. A detailed analysis of each model will be performed in Chapter 3.
Table 1.1 Summary of models for convection, diffusion, and reaction processes.
PDE type
multidimensional
one-dimensional
v ∂∂ ux − d ∂∂ x2u = s
2
∇ · (vu − D∇u) = s
elliptic
−∇ · (D∇u) = s
−d ∂∂ x2u = s
∇ · (vu) = s
v ∂∂ ux = s
hyperbolic
∂u
∂t
∂u
∂t
parabolic
∂u
∂t
2
∂u
∂t
+ ∇ · (vu) = s
− ∇ · (D∇u) = s
+ ∇ · (vu − D∇u) = s
∂u
∂t
∂u
∂t
+ v ∂∂ ux = s
− d ∂∂ x2u = s
2
+ v ∂∂ ux − d ∂∂ x2u = s
2
1.4 Space Discretization Techniques
15
1.4 Space Discretization Techniques
Computers are of little help in obtaining closed-form analytical solutions to a PDE
model. However, they can be programmed to solve algebraic equations very fast.
Replacing calculus by algebra, it is possible to compute approximate solutions to
the CDR equation and more advanced mathematical models. To this end, the computational domain, the unknown solution, and its partial derivatives need to be discretized, so as to obtain a set of algebraic equations for the function values at a
finite number of discrete locations. We will begin with the discretization in space
and discuss time-stepping techniques for unsteady PDEs in Section 1.5.1.
1.4.1 Computational Meshes
Recall that the integral conservation law (1.2) which has led us to (1.9) and (1.11)
was formulated for a fixed control volume (CV) of finite size. Instead of looking at
the whole flow field at once, we have focused our attention on what is happening in a
small subdomain. A similar approach is used to discretize differential equations that
embody physical conservation principles. The unknowns of the discrete problem are
associated with a computational mesh or grid which represents a subdivision of the
1D, triangles
domain Ω ⊂ Rd into many small control volumes Ωk (e.g., intervals in S
or quadrilaterals in 2D, tetrahedra or hexahedra in 3D) such that Ω̄ ≈ k Ω̄k .
Many excellent texts are devoted to automatic generation and adaptation of computational meshes, see [55, 115, 118, 226, 318]. Mesh generation is easy for domains of rectangular shape but difficult in the case of curvilinear boundaries, internal obstacles, and small-scale features. Depending on the geometric complexity of
Ω , the mesh may be structured, block-structured, or unstructured (see Fig. 1.4).
In the one-dimensional case, the computational domain Ω = (a, b) is an interval.
A subdivision of this interval into N subintervals Ωk = (xk−1 , xk ) of equal size
∆x =
b−a
N
yields the simplest representative of structured meshes. The N + 1 grid points
xi = i∆ x,
∀i = 0, 1, . . . , N
(1.33)
are numbered from left to right. Each interior grid point xi has two nearest neighbors
whose indices i ± 1 and coordinates xi±1 are known. The spacing ∆ xk = xk − xk−1
can also be nonuniform if higher mesh resolution is desired in some regions.
In multidimensions, a structured mesh is a net of grid lines (Fig.1.4, a–c) which
can be numbered consecutively. In the simplest case, formula (1.33) is used to discretize each coordinate axis. Again, the search for nearest neighbors is easy, and
their number is the same for each interior point. The involved data structures and
numerical algorithms are almost as simple as in the one-dimensional case. A nonuni-
16
(a) structured, uniform
1 Getting Started
(b) structured, deformed
(c) structured, perturbed
(d) block-structured, 2 subdomains
(e) unstructured, triangular
(f) unstructured, quadrilateral
Fig. 1.4 Examples of computational meshes for two-dimensional domains.
form body-fitted mesh can be generated and mapped onto a uniform Cartesian grid
[4, 318]. However, since all grid lines must begin and end on the boundary, an
attempt to obtain higher resolution in zones of particular interest may entail unintended and, sometimes, harmful mesh refinement in other parts of the domain [104].
The generation of a block-structured mesh is based on a two-level subdivision,
whereby a number of overlapping or nonoverlapping subdomains (blocks) are discretized using structured meshes [55, 104]. Figure 1.4 (d) displays such a mesh
that consists of two components. The use of multiple blocks makes it easier to deal
with nonrectangular domains and moving objects. Domain decomposition methods
[55, 143, 209, 210, 277] can be employed to distribute the work between multiple
processors in a parallel computing environment. Local mesh refinement and solution algorithms tailored to structured meshes can be applied blockwise but special
1.4 Space Discretization Techniques
17
care is required to transfer information between the blocks in a conservative manner. It is also possible to use structured grids in some subdomains (e.g., to generate a
Cartesian core [234] or resolve a boundary layer) and unstructured ones elsewhere.
Domains of particularly complex geometric shape call for the use of a fully unstructured mesh. Recent advances in computational geometry make it possible to
generate such meshes automatically for 2D and 3D problems as complex as flow
around a car, a submarine, or a space shuttle [226]. Unstructured mesh methods are
very flexible and well suited for mesh adaptivity. An arbitrary number of elements
are allowed to meet at a single vertex (see Fig.1.4, e–f), so it is easy to insert extra
grid points in regions of insufficient mesh resolution. Conversely, it is possible to
remove points in regions where a coarser mesh would suffice. Since transport processes move information from one place into another, it is worthwhile to adjust the
mesh in the course of simulation, so as to achieve high accuracy at a low cost.
Of course, the flexibility offered by unstructured meshes is not a free lunch since
sophisticated data structures are required to handle the irregular connectivity pattern, and data access is rather slow. Moreover, efficient solution methods are more
difficult to develop and program than in the case of (block-)structured grids. Nevertheless, most successful general-purpose CFD codes are based on unstructured
meshes. We refer to the recent monograph by Löhner [226] for a comprehensive
introduction to this approach and a unique collection of state-of-the-art algorithms.
The nodes of the mesh may be fixed or move in a prescribed fashion. The corresponding numerical algorithms can be classified into Eulerian, Lagrangian, and
Arbitrary Lagrangian Eulerian (ALE) ones. An Eulerian method is based on a fixed
mesh, i.e., the positions of mesh points do not change as time goes on. The nodes of
a Lagrangian mesh move with the flow velocity, so that the convective transport is
built into the mesh motion and only diffusive fluxes need to be discretized. A major
drawback to this approach is the fact that the shape and size of mesh cells cannot
be controlled. As a consequence, mesh tangling is possible, unless global or local
remeshing is performed on a regular basis. Within the ALE framework, some nodes
may remain fixed, while others may move with arbitrary velocities. This is the most
general formulation which is often used for simulation of flows with free interfaces
and fluid-structure interaction (FSI). Moving meshes are of value for many applications but, for simplicity, only Eulerian methods will be discussed in this book.
1.4.2 Semi-Discrete Problem
Given a suitably designed computational mesh, the continuous function u(x,t) is
approximated by a finite number of nodal values {ui } which may be associated with
vertices, edges, faces, cells, or control volumes. Depending on the type of approximation, these degrees of freedom may represent, e.g., pointwise function values,
cell averages, or coefficients of piecewise-polynomial basis functions. If the governing equation models an unsteady process, then the degrees of freedom are timedependent and should be updated step-by-step, as explained in Section 1.5.1.
18
1 Getting Started
The discretization in space is required to obtain a system of equations for the
nodal values of the approximate solution. Of course, the number of equations should
be the same as the number of unknowns. The only derivative that may still appear
in each equation is the one with respect to time. For example, the semi-discretized
unsteady CDR equation (1.11) for the nodal value ui can be written as
du j
(1.34)
∑ mi j dt + ∑(ci j + di j )u j = ri ,
j
j
where the coefficients ci j , di j , and ri are due to convection, diffusion, and reaction,
respectively. The use of the total derivative notation in the left-hand side of equation
(1.34) is appropriate since no space derivatives are present anymore. The weights
mi j distribute the gain or loss of mass, if any, between node i and its neighbors.
If mi j = 0, ∀ j 6= i, (1.34) reduces to the ordinary differential equation (ODE)
mii
dui
+ (ci j + di j )u j = ri .
dt ∑
j
(1.35)
Since the right-hand sides of (1.34) and (1.35) depend on the solution values at several nodes, the semi-discrete equations must be integrated in time simultaneously.
The matrix form of a space discretization given by (1.34) or (1.35) reads
M
du
+ (C + D)u = r,
dt
(1.36)
where u = {ui } denotes the vector of time-dependent nodal values, M = {mi j } is the
mass matrix, C = {ci j } is the discrete transport operator, D = {di j } is the discrete
diffusion operator, and r = {ri } is the vector of discretized source or sink terms.
As a rule, the matrices M, C, and D are sparse. That is, most of their entries are
equal to zero and do not need to be stored. The sparsity pattern of the discrete operators depends on the type of the underlying mesh (structured or unstructured) and
on the numbering of nodes. The mass matrix M is diagonal or symmetric positive
definite. Ideally, the discrete diffusion operator D should also be symmetric, as required by the properties of its continuous counterpart [289]. The discrete convection
operator C is nonsymmetric since the flow direction must be taken into account. For
example, this matrix can be skew-symmetric (C = −CT ) or upper/lower triangular.
In the case of a steady governing equation, such as (1.24), the time derivative
vanishes, so the semi-discrete problem (1.36) reduces to the algebraic system
(C + D)u = r.
(1.37)
The space discretization of pure convection, pure diffusion, and zero reaction models from Table 1.1 corresponds to D = 0, C = 0, and r = 0 respectively.
Polynomials play an important role in the discretization process since they are
easy to differentiate and integrate. The most popular discretization techniques based
on polynomial approximations are the finite difference, finite volume, and finite el-
1.4 Space Discretization Techniques
19
ement method. Spectral and boundary element methods are also worth mentioning
but their range of applicability is rather limited, so they will not be discussed in this
book. For a general introduction to numerical methods for differential equations,
we refer to [76, 104, 149, 276]. In this section, we will introduce some basic discretization concepts and present a concise summary of the approximations involved.
An in-depth presentation of numerical schemes tailored to the convection-diffusionreaction equation and other transport models will follow in subsequent chapters.
1.4.3 Finite Difference Methods
The finite difference method (FDM) is the oldest among the discretization techniques for partial differential equations. Many modern numerical schemes for transport phenomena trace their origins to finite difference approximations developed
in the late 1950s through early 1980s. The derivation and implementation of FDM
are particularly simple on structured meshes which are topologically equivalent to a
uniform Cartesian grid. The nodal value of the approximate solution at node i
ui (t) ≈ u(xi ,t)
(1.38)
is a pointwise approximation to the true solution of the partial differential equation.
Taylor series expansions or polynomial fitting techniques are used to approximate
all space derivatives in terms of ui and/or solution values at a number of neighboring
nodes. For example, if we consider the uniform 1D mesh given by (1.33), then
ui+1 − ui−1
∂u
≈
∂x i
2∆ x
is a second-order approximation to the first derivative of u at node i, whereas
2 ∂ u
ui+1 − 2ui + ui−1
≈
∂ x2 i
(∆ x)2
is a second-order approximation to the second derivative. On a nonuniform mesh,
the coefficients are different and must be derived individually for each grid point.
Alternatively, a mapping onto a regular Cartesian grid may be employed [4].
After all space derivatives have been approximated by finite differences, the
semi-discrete counterpart of the unsteady CDR equation (1.11) can be written in
the generic form (1.35) which determines the relationship between ui and the solution values at a certain number of neighboring nodes. The set of all grid points that
make a nonzero contribution to the equation for ui is called the stencil.
The resulting finite difference discretization is of the form (1.36), where M is a
diagonal matrix. The coefficients of the matrices C and D depend on the parameters
of the model, on the choice of finite difference approximations, and on the mesh
size. Global mesh refinement results in higher accuracy but increases the size of the
20
1 Getting Started
algebraic system and, consequently, the computational cost. If the mesh is structured, then the stencil of each interior point has the same size, and nearest neighbors
are easy to identify provided that the grid lines are numbered consecutively. In the
case of an unstructured mesh, the number of neighbors may vary, and the distribution of grid points is nonuniform. Fitting a polynomial to scattered data is feasible
but computationally expensive and difficult to implement. For this reason, the use of
unstructured meshes is rather uncommon in the realm of finite difference methods.
1.4.4 Finite Volume Methods
Due to the growing demand for numerical simulation of transport processes in 2D
and 3D domains of complex shape, the finite difference method has eventually lost
its leadership position. Nowadays, general-purpose CFD codes are typically based
on the finite volume method (FVM) which yields a finite-difference like approximation on a uniform Cartesian grid but is readily applicable to unstructured meshes.
Finite volume methods for the CDR equation (1.11) are based on the underlying
integral conservation law. Inserting the flux f = vu − D∇u into (1.2), one obtains
∂
∂t
Z
u(x,t) dx +
Vi
Z
Si
(vu − D∇u) · n ds =
Z
s(x,t) dx,
(1.39)
Vi
where Vi is a control volume (CV) bounded by the control surface Si , and n is the
unit outward normal. In cell-centered finite volume methods, Vi = Ωi is a single cell
of the computational mesh, see Fig. 1.5. Alternatively, a dual tessellation can be used
to define Vi for a vertex-centered finite volume method. In the two-dimensional case,
the dual cell Vi around the vertex xi can be constructed by joining the midpoints of
mesh edges and the centroids of the neighboring cells, as shown in Fig. 1.6.
The outward normals to the interface Si j = Si ∩ S j between any pair of adjacent
control volumes Vi and V j have opposite signs, so the integrals over internal boundaries cancel out if equations (1.39) are summed over i. Hence, the integral conservation law (1.2) holds not only for all Vi but also for the whole domain V = Ω
∂
∂t
Z
Ω
u(x,t) dx +
Z
Γ
(vu − D∇u) · n ds =
Z
Ω
s(x,t) dx.
(1.40)
Equations (1.39) and (1.40) express the local and global conservation principles,
respectively. The former implies the latter but the reverse is generally not true.
The degrees of freedom for cell-centered or vertex-centered FVM can be defined
as mean values over the CVs associated with cells and vertices, respectively. Let
ui (t) =
1
|Vi |
Z
Vi
u(x,t) dx,
(1.41)
1.4 Space Discretization Techniques
Vi
21
n
Vj
Vi
n
Vj
Fig. 1.5 Control volumes for a cell-centered FVM in two dimensions.
Vi
Vj
n
Vj
Vi
n
xi
n
xj
xi
xj
n
Fig. 1.6 Control volumes for a vertex-centered FVM in two dimensions.
where |Vi | denotes the volume of Vi . According to (1.39), the evolution of the sodefined mean value ui is governed by the integro-differential equation
|Vi |
dui
+
dt
Z
Si
(vu − D∇u) · n ds = |Vi |si
(1.42)
in which si denotes the average rate of production inside Vi , that is,
si (t) =
1
|Vi |
Z
s(x,t) dx.
(1.43)
Vi
By definition, the boundary Si of the control volume Vi consists of several patches
Si =
[
Si j .
(1.44)
j
For any index j 6= i, the patch Si j = Si ∩ S j is the interface between Vi and one of its
neighbors V j . The index j = i is reserved for boundary patches Sii = Si ∩ Γ .
Splitting the surface integral into a sum over patches, one can write (1.42) as
dui
|Vi |
+
dt ∑
j
Z
Si j
(vu − D∇u) · n ds = |Vi |si .
(1.45)
If the control volumes for a cell-centered or vertex-centered FVM are defined as
shown in Figs. 1.5–1.6, then the normal to Si j is constant or piecewise-constant.
22
1 Getting Started
The total flux across the patch Si j of the control surface Si is given by
Z
fi j =
Si j
(vu − D∇u) · n ds.
(1.46)
If the patch Si j is the common boundary of Vi and V j , then j 6= i and f ji = − fi j , so
that the mass leaving Vi is equal to the mass entering V j and vice versa.
Equation (1.45) can be expressed in terms of the integrated fluxes fi j as follows
|Vi |
dui
+ fi j = |Vi |si
dt ∑
j
(1.47)
and transformed into an equation of the form (1.35) with mii = |Vi | and ri = |Vi |si
|Vi |
dui
+ (ci j + di j )u j = |Vi |si .
dt ∑
j
(1.48)
To this end, it is necessary to approximate the integrated fluxes fi j in terms of the
mean values ui . The derivation of (1.48) involves two levels of approximation [104]
• Surface and volume integrals are approximated using numerical quadrature (cubature) which requires evaluation of the integrand at one or more locations.
• Interpolation techniques are employed to approximate the function values and
derivatives at the quadrature points in terms of the primary unknowns ui .
By the midpoint rule, the mean values ui (t) and si (t) represent second-order accurate
approximations to u(x̄i ,t) and s(x̄i ,t) evaluated at the center of mass
x̄i =
1
|Vi |
Z
x dx.
Vi
The surface integral in the right-hand side of (1.46) can also be evaluated by the
midpoint rule. If the normal to the interface Si j is not constant, numerical integration is performed patchwise. Interpolation is required to approximate the flux function at the quadrature points, because only ui (t) ≈ u(x̄i ,t) are available. If the same
interpolation formula is used on both sides of the interface, then the finite volume
discretization (1.48) is conservative, both locally and globally. Summation over i
yields a discrete counterpart of the integral conservation law (1.40) for V = Ω .
The finite volume method is promoted in the majority of introductory courses
and textbooks on numerical methods for CFD. It is relatively easy to understand
and implement, especially in the case of first- and second-order approximations on
structured meshes. Higher-order schemes are difficult to derive within the framework of the above-mentioned two-level approximation strategy that involves interpolation and integration. Finite volumes lend themselves to the discretization of
hyperbolic equations but the approximation of diffusive fluxes requires numerical
differentiation, which makes it rather difficult to achieve high accuracy and preserve
the symmetry of the continuous diffusion operator at the discrete level.
1.4 Space Discretization Techniques
23
1.4.5 Finite Element Methods
The finite element method (FEM) is a relative newcomer to CFD and a very promising alternative to finite differences and finite volumes. It is usually used in conjunction with unstructured meshes and provides the best approximation property when
applied to elliptic and parabolic problems at relatively low Peclet numbers. The development of high-resolution finite element schemes for hyperbolic and convectiondominated transport equations is a topic of active research. A summary of the author’s contributions to this field will be the main highlight of the present book.
Finite element approximations to the CDR equation are based on the weighted
residual method. The weak formulations (1.20) and (1.21) can be written as
∂u
+ c(w, u) + d(w, u) = r(w),
∀w ∈ W .
(1.49)
w,
∂t
The operators c(·, ·) and d(·, ·) are associated with the weak form of the convective
and diffusive terms, respectively. The reactive part r(·) combines the contributions
of the source/sink term s and of the surface integral (1.22), if any. The scalar product
(·, ·) is defined in the space L2 (Ω ) of functions that are square integrable in Ω
(w, v) =
Z
Ω
∀v, w ∈ L2 (Ω ).
wv dx,
The approximate solution uh ∈ Vh to problem (1.49) is defined as follows
uh (x,t) = ∑ u j (t)ϕ j (x),
(1.50)
j
where {ϕi } is a set of basis functions spanning the finite-dimensional space Vh . As
a rule, these basis functions are required to possess the following properties [76]
• there exists a set of nodes xi ∈ Ω such that ϕi (xi ) = 1 and ϕi (x j ) = 0, ∀ j 6= i;
• the restriction of ϕi to each cell is a polynomial function of local coordinates.
Due to the first property and (1.50), the nodal values of the approximate solution
are given by uh (xi ,t) = ui (t). The points xi are usually located at the vertices of the
mesh. Other possible locations are the midpoints/barycenters of edges, faces, and
cells [76]. A typical piecewise-linear basis function ϕi is depicted in Fig. 1.7.
Let Wh be the space of test functions spanned by {ψi }. Using u = uh and w = ψi
in (1.49), one obtains a semi-discrete equation of the form (1.34), where
mi j = (ψi , ϕ j ),
di j = d(ψi , ϕ j ),
ci j = c(ψi , ϕ j ),
ri = r(ψi ).
The conventional Galerkin method takes the trial functions ϕ j and test functions ψi
from the same space Wh = Vh . That is, ψi = ϕi , so the number of equations equals
the number of unknowns. Sometimes it is worthwhile to consider test functions
from Wh 6= Vh spanned by the same number of basis functions ψi 6= ϕi . Such finite
24
1 Getting Started
ϕi
xi
Fig. 1.7 A piecewise-linear finite element basis function on a triangular mesh.
element approximations are known as Petrov-Galerkin methods and offer certain
advantages, e.g., in the case of convection-dominated transport problems.
The finite element method is supported by a large body of mathematical theory
that makes it possible to obtain rigorous error estimates and proofs of convergence.
Moreover, it can be combined with h− p adaptivity, whereby the local mesh size and
the order of polynomials are chosen so as to obtain the best possible resolution. The
finite element mesh (also called triangulation) is usually unstructured, and the shape
of mesh cells can be fitted to the shape of a curvilinear boundary. Matrix assembly is
performed element-by-element in a fully automatic way. The remarkable generality
and flexibility of the FEM makes it very powerful. Almost all codes for structural
mechanics problems are based on finite element approximations, and a lot of current
research is aimed at the development of adaptive FEM for fluid dynamics.
Finite elements and finite volumes have a lot in common and are largely equivalent in the case of low-order polynomials [8, 166, 226, 300]. The traditional
strengths of FVM and FEM, as applied to the CDR equation, are complementary.
Convective terms call for the use of an upwind-biased discretization which is easier
to construct within the finite volume framework. On the other hand, the finite element approach takes the lead when it comes to the discretization of diffusive terms.
Therefore, many hybrid FVM-FEM schemes have been proposed. For example, finite element shape functions are used to interpolate the fluxes for a vertex-centered
finite volume method [104]. Conversely, FVM-like approximations of convective
terms are frequently employed to achieve the upwinding effect in finite element
codes [11, 322]. Fluctuation splitting (alias residual distribution) methods [54, 79]
represent another attempt to bridge the gap between the FVM and FEM worlds.
A current trend in CFD is towards the use of discontinuous Galerkin (DG) methods [66] which represent a generalization of FVM and incorporate some of their
most attractive features, such as local conservation. At the same time, the treatment
of diffusive fluxes is not straightforward and requires special care, as in the case of
classical FVM. We welcome the advent of DG methods but feel that the potential of
continuous finite elements has not yet been exploited to the full extent in the context
of transport equations. This is why the high-resolution finite element schemes to be
presented in this book will be based on conventional Galerkin discretizations.
1.5 Systems of Algebraic Equations
25
1.5 Systems of Algebraic Equations
Space discretization of a stationary problem leads to a set of coupled algebraic equations which do not contain any derivatives and, therefore, do not require any further
discretization. An algebraic system like (1.37) needs to be solved just once since the
nodal values of the approximate solution are assumed to be independent of time.
If the problem at hand is nonstationary, then semi-discrete equations of the form
(1.34) must be integrated in time using a suitable numerical method. To this end, the
time interval (0, T ) is discretized in much the same way as the spatial domain for
one-dimensional problems. Consider a sequence of discrete time levels
0 = t 0 < t 1 < . . . < t K = T.
The time step ∆ t n = t n+1 − t n may be constant or variable. In the former case
t n = n∆ t,
∀n = 0, 1, . . . , K.
The value of the approximate solution at node i and time level t n is denoted by
uni ≈ ui (t n ).
Due to the initial condition (1.14), the value of u0i = ui (0) is assumed to be known.
In principle, the time can be treated just like an extra space dimension. Finite difference, finite volume, and finite element methods are readily applicable to functions
of x = (x1 , . . . , xd ,t) ∈ Rd+1 defined in the space-time domain Ω × (0, T ). However,
simultaneous computation of uni for all nodes and time levels is often too expensive.
Since information propagates forward in time, it is worthwhile to take advantage of
this fact and advance the numerical solution in time step-by-step.
Time-stepping (or marching) methods initialize the vector of discrete nodal
values by u0 = {u0i } and use un = {uni } as initial data for the computation of
un+1 = {uin+1 }. This solution strategy is faster and requires less memory than a coupled space-time discretization. First, the size of the algebraic systems to be solved at
each time step depends only on the number of spatial degrees of freedom and not on
the number of time levels. In other words, a huge system is replaced by a sequence
of smaller ones, which results in considerable savings of computer time. Second,
intermediate data are overwritten as soon as they are not needed anymore. Last but
not least, a nonphysical dependence of un on un+1 is ruled out by construction.
Regardless of the methods chosen to discretize the continuous problem in space
and time, the result is an algebraic system that can be written in the generic form
Au = b,
(1.51)
where A = {ai j } is a sparse matrix, u = un+1 is the vector of unknowns, and b is
evaluated using the previously computed data from one or more time levels. The
discretization is said to be explicit if A is a diagonal matrix and implicit otherwise.
26
1 Getting Started
1.5.1 Time-Stepping Techniques
A semi-discrete system like (1.36) can be discretized in time using a wealth of methods developed for numerical integration of ODEs and differential algebraic equations (DAEs). This approach is called the method of lines (MOL) since the problem at hand consists of many one-dimensional subproblems for the nodal values
ui (t) to be integrated over the time interval (t n ,t n+1 ) subject to the initial condition
ui (t n ) = uni . The decoupling of space and time coordinates makes it possible to use
any discretization technique applicable to an initial value problem of the form
du
= F(u,t),
dt
t n < t ≤ t n+1 ,
u(t n ) = un .
(1.52)
Within the MOL framework, the solution u is the vector of time-dependent nodal
values, whereas the vector F(u,t) contains the discretized space derivatives, sources,
sinks, and boundary conditions. In the case of our DAE system (1.36)
MF(u,t) = r(t) − (C + D)u(t).
(1.53)
If the underlying velocity field and/or the diffusion tensor depend on t, then so do
the coefficients of the discrete operators C = {ci j } and D = {di j }, respectively.
The simplest time-stepping methods are based on a finite difference discretization of the time derivative that appears in (1.36) and (1.52). Let t n+1 = t n + ∆ t and
un+1 − un
= θ F(un+1 ) + (1 − θ )F(un ),
∆t
0 ≤ θ ≤ 1.
(1.54)
This generic formula unites the first-order accurate forward Euler method (θ = 0)
un+1 = un + ∆ tF(un ),
the second-order accurate Crank-Nicolson scheme which corresponds to θ =
un+1 = un +
∆t
[F(un+1 ) + F(un )],
2
(1.55)
1
2
(1.56)
and the first-order accurate backward Euler method (θ = 1) which yields
un+1 = un + ∆ tF(un+1 ).
(1.57)
In general, the right-hand side of (1.54) is a weighted average of F(u,t) evaluated
at the old and new time level. Depending on the value of θ , the resulting time discretization can be explicit or implicit. In an update like (1.55), each nodal value uin+1
can be calculated explicitly using the data from the previous time level. In the case
of implicit methods (θ > 0), each algebraic equation contains several unknowns,
whence a simultaneous update of all nodal values is required. Explicit methods are
easier to implement but implicit methods are more stable, as explained below.
1.5 Systems of Algebraic Equations
27
To obtain a fully discrete counterpart of system (1.36), all we have to do is to
multiply (1.54) by the mass matrix M and invoke (1.53). The result is an algebraic
system of the form (1.51), where the left-hand side matrix A is given by
A = M + θ ∆ t(C + D),
and the right-hand side b is the sum of all terms that do not depend on un+1
b = [M − (1 − θ )∆ t(C + D)]un + ∆ t[θ rn+1 + (1 − θ )rn ].
This discretization is fully explicit if θ = 0 and the mass matrix M is diagonal. The
latter condition holds for any finite difference or finite volume scheme. However,
many finite element approximations produce nondiagonal mass matrices, so that a
linear system needs to be solved even if the forward Euler method is employed.
The generic θ −scheme (1.54) belongs to the family of two-level methods since
only un is involved in the computation of un+1 . Such time-stepping methods can be
at most second-order accurate. Higher-order approximations must use information
from additional time levels. In Runge-Kutta methods, all time levels t n+α , where
0 ≤ α ≤ 1, belong to the interval [t n ,t n+1 ], and a predictor-corrector strategy is
employed to compute un+1 . Another possibility is to integrate a polynomial fitted
to the values of F(u,t) at t n+α , . . . ,t n−β , where α and β are nonnegative integers.
Multipoint methods of Adams-Bashforth (explicit) and Adams-Moulton (implicit)
type correspond to α = 0 and α = 1, respectively. Their pros and cons, as compared
to Runge-Kutta time-stepping schemes of the same order, are explained in [104].
1.5.2 Direct vs. Iterative Solvers
The last step in the development of a numerical algorithm is the solution of the
algebraic system (1.51) that results from the discretization of a continuous problem.
In the case of an explicit scheme, the computation of u = A−1 b is trivial
ui =
bi
,
aii
∀i.
Otherwise, the tools of numerical linear algebra are required to solve (1.51). The
choice of the solution method is largely independent of the underlying discretization
techniques but the size and structure of the matrix A need to be taken into account.
1.5.2.1 Direct Methods
Direct methods for solving linear systems of the form (1.51) accomplish this task in
one step. The input parameters are the matrix A and the right-hand side b. The result
is the solution vector u. In the absence of roundoff errors, this solution is exact.
28
1 Getting Started
Direct methods perform well for linear systems of moderate size but the memory
requirements and CPU time increase nonlinearly with the number of unknowns. As
a rule of thumb, direct solvers are not to be recommended for very large systems.
If the sparsity pattern of the matrix A exhibits some regular structure, then this
structure can sometimes be exploited to design a fast direct solver, such as the
Thomas algorithm for tridiagonal matrices. This algorithm can also be embedded in
Alternating-Direction-Implicit (ADI) solvers for banded matrices that result from
2D and 3D discretizations on structured meshes. The multifrontal method implemented in the open-source software package UMFPACK [330] is one of the fastest
direct solvers for nonsymmetric sparse linear systems. The rapid increase in computer memory and the possibility of efficient implementation on parallel supercomputers have revived the interest in direct solvers for large-scale applications [226].
1.5.2.2 Iterative Methods
Iterative methods solve linear systems of the form (1.51) using a sequence of explicit updates starting with an initial guess which must be supplied as another input
parameter. Such an algorithm is said to be convergent if each update brings the solution closer to that of (1.51) which is recovered after sufficiently many iterations.
In practice, it is neither necessary nor wise to iterate until convergence. The iterative process is terminated when certain stopping criteria are satisfied. As a rule,
these criteria amount to monitoring the differences between two successive iterates
and/or the residuals measured in a suitably chosen norm. Stopping too early gives
rise to large iteration errors; stopping too late results in a waste of CPU time.
Iterative solvers use a rather small amount of computer memory. Many of them
do not even require that the matrix A be available. All they need is a subroutine that
evaluates the residual of the linear system for a given tentative solution. The most
efficient iterative algorithms are based on multigrid methods [131, 345]. If properly
configured, they can solve a system of N equations using as few as O(N) arithmetic
operations, as compared to O(N 3 ) for direct solvers based on Gaussian elimination
and O(N 2 ) for forward/backward substitution, given a precomputed LU factorization. Clearly, this makes a big difference, especially if N is as large as 106 and more.
Thus, an iterative solution strategy pays off for large sparse linear systems.
If the system to be solved corresponds to the discretization of a stationary problem, a good initial guess is rarely available. The default is zero, and many iterations
are usually required to achieve the prescribed tolerance. The computational effort
associated with advancing the solution of an unsteady problem from one time level
to the next is much smaller. Since un provides a relatively good guess for un+1 , just
a few iterations per time step are normally required to obtain a sufficiently accurate
solution. Therefore, each update is typically much cheaper than that performed with
a direct solver. If the time step is very small, then a single iteration may suffice, and
the cost of solving the linear system is comparable to that of a fully explicit update.
As the time step increases, so does the number of iterations for the linear solver.
1.5 Systems of Algebraic Equations
29
In light of the above, an iterative solver for the discretization of a stationary
problem may be more expensive and/or difficult to implement than a time-stepping
method for the corresponding unsteady problem. It might be easier to march the
solution to the steady state. If we discretize system (1.36) in time by an explicit or
implicit method, prescribe arbitrary initial conditions, and run the code for a sufficiently long time, then we will end up with the solution of (1.37). This popular
approach to steady-state computations is called pseudo time-stepping since it represents an iterative solver in which the time step serves as a relaxation parameter. In
this case, the accuracy of the time discretization is not important, and the artificial
time step should be chosen so as to reach the steady state limit as fast as possible.
Note that the use of excessively large time steps or inappropriate parameter settings
may cause divergence, which makes iterative methods less robust than direct ones.
1.5.2.3 Nonlinear Systems
If the coefficients of the discrete problem depend on the unknown solution, then the
algebraic system (1.51) is nonlinear. In this case, an iterative solution strategy is a
must. Starting with a suitably chosen initial guess, the currently available solution
values are used to update the matrix coefficients and obtain the next guess by solving
a nominally linear system. This process is repeated until the changes and/or the
residual of the nonlinear system become small enough. In principle, both direct and
iterative methods can be used to solve the involved linear systems. However, the use
of direct methods is usually impractical since they spend an excessive amount of
CPU time on solving linear systems with only tentative coefficients.
Within a fully iterative approach, the coefficients are calculated using the solution
values from the previous outer iteration, and a few inner iterations are performed
to obtain an improved solution. Since the coefficients are tentative, the amount of
work spent on the solution of each linear system should not be inordinately large
[268]. As soon as the residuals have decreased by a factor of 10 or so, one may stop
the inner iteration process and proceed to the calculation of the coefficients for the
next outer iteration. Well-balanced stopping criteria make it possible to minimize the
computational cost that depends on the total number of inner and outer iterations.
Of course, an iterative method is of little value if it does not converge. If the
nonlinearity is too strong, the solution may exhibit oscillatory behavior that inhibits
convergence. A possible remedy to this problem is the use of underrelaxation techniques. The basic idea is to take a weighted average of the old and new data, so as
to slow down the changes of solution values and matrix coefficients [268].
1.5.3 Explicit vs. Implicit Schemes
In the case of an fully explicit scheme, no linear or nonlinear systems need to be
solved. Explicit algorithms are easy to code/parallelize and require a modest amount
30
1 Getting Started
of computer memory. However, the time step may not exceed a certain threshold that
depends on the Peclet number and on the smallest mesh size. Otherwise, the scheme
may become unstable and produce meaningless numbers going to infinity. The lack
of stability is the price to be paid for algorithmic simplicity of explicit schemes. The
cost of a single solution update is minimal but an inordinately large number of time
steps may be required to perform simulation over a given interval of time.
Implicit methods produce nondiagonal matrices, whence each algebraic equation
contains several unknowns and cannot be solved in a stand-alone fashion. The design of an efficient implicit algorithm is particularly difficult if the underlying PDE
and/or the discretization procedure are nonlinear. The cost per time step is large as
compared to that of an explicit solution update. Also, the programming of iterative
solvers for (1.51) is time-consuming, and their efficiency depends on the parameter
settings, stopping criteria etc. On the other hand, most implicit schemes are unconditionally stable, and the use of large time steps makes it possible to reach the final
time faster than with an explicit scheme subject to a restrictive stability limit. Of
course, it should be borne in mind that the accuracy of the time discretization and
the convergence behavior of iterative solvers also depend on the time step.
It is essential to distinguish between truly transient problems and the ones in
which the solution varies slowly and/or becomes stationary in the long run. The
optimal choice of the time-stepping scheme depends on whether the goal is
• to perform a transient computation in which evolution details are important, or
• to predict the long-term flow behavior or to compute a steady-state solution.
Depending on the objectives of the simulation to be performed, an explicit or implicit solution strategy may be preferable. Explicit schemes lend themselves to the
treatment of problems in which the use of small time steps is dictated by accuracy
considerations. If only the steady-state solution is of interest, then it is worthwhile
to use local time-stepping (different time steps for different nodes) and/or an unconditionally stable implicit scheme, such as the backward Euler method (1.57). If it
is not known in advance, whether the transport process to be simulated is steady or
unsteady, it is possible to start with an explicit scheme and switch to an implicit one
if the ratio (un+1 − un )/∆ t becomes small as compared to other terms. Therefore,
both explicit and implicit solvers belong into a general-purpose CFD toolbox.
1.6 Fundamental Design Principles
No numerical method is perfect, and many compromises are involved in the design
process. The quality of numerical approximations to convection-diffusion equations
depends on the underlying mesh, on the properties of the employed discretization
techniques, and on the Peclet number. The mesh size and time step should also be
chosen carefully, especially in the case of conditionally stable explicit schemes. All
of the above-mentioned factors influence the coefficients of the algebraic system
(1.51) for the nodal values of the approximate solution. The properties of the matrix
1.6 Fundamental Design Principles
31
A and of the discrete operators involved in the assembly of the right-hand side b may
make or break the entire numerical model. The type of the governing equation and
the smoothness of the solution may also play an important role. Some algorithms are
tailored to equations of a certain type and/or perform poorly if the solution exhibits
steep gradients. Other methods do not work at all or produce nice-looking results
which have little in common with the true solution of the mathematical model.
A good numerical method must fulfil a number of prerequisites dictated by the
physics, mathematical theory, and numerical analysis of the problem at hand. These
criteria lead to a set of rules that guarantee a certain level of accuracy and robustness for a sufficiently broad range of applications. Some of the fundamental design
principles are summarized in this section, and their implications are explained.
1.6.1 Numerical Analysis
The difference between the exact solution u of the continuous problem and an approximate solution uh∆ t produced by a computer code is the sum of numerical errors.
Aside from programming bugs, we can distinguish between discretization errors,
roundoff errors, and iteration errors. The discretization error εh∆ t depends on the
mesh size h and time step ∆ t. It can be estimated using Taylor series expansions or,
in the case of finite element methods, sophisticated tools of functional analysis.
Roundoff errors due to the finite precision of computer arithmetics are usually
much smaller than εh∆ t , whereas iteration errors depend on the prescribed tolerances
and stopping criteria for linear solvers. A properly designed numerical scheme must
be sufficiently accurate and converge to the exact solution of the differential equation as the mesh size h and time step ∆ t are refined. Therefore, it must contain
inherent mechanisms to control the magnitude of the total error in the course of simulation. A rigorous analysis of consistency, stability, and convergence is required to
evaluate new discretization techniques and identify the range of their applicability.
1.6.1.1 Consistency
A numerical method is said to be consistent if the discretization error εh∆ t goes to
zero as h → 0 and ∆ t → 0. Consistency refers to the relationship between the exact
solutions of the continuous and discrete problems. In essence, it guarantees that the
discretization is asymptotically correct. Of course, finite values of h and ∆ t are used
in practice to keep simulations affordable. Since the computational cost increases
rapidly with the number of unknowns, it is natural to require that the discretization
error εh∆ t become smaller if we take a finer mesh and/or time step. Moreover, it is
desirable to have some idea of how much accuracy we can gain by doing so. This
information can be inferred from an a priori error estimate of the form
εh∆ t = O(h p , ∆ t q ).
(1.58)
32
1 Getting Started
The corresponding discretization is consistent if p > 0 and q > 0. Its spatial and temporal accuracy is of order p and q, respectively. This gives an asymptotic estimate
of the rate at which εh∆ t shrinks as the mesh and/or time step are refined.
It is worth mentioning that the formal order of approximation is not the sole indicator of accuracy and, in many cases, not even a particularly good one. The absolute
values of the errors produced by two schemes of the same order may differ significantly, and a low-order scheme might perform better than a high-order one on a
coarse mesh. Strictly speaking, a priori estimates based on Taylor series expansions
are not applicable to discontinuous solutions, Also, the contribution of higher-order
terms may become nonnegligible if the corresponding derivatives are too large. As a
consequence, the error εh∆ t might decrease much slower than expected. Last but not
least, even a consistent scheme may fail to converge if it turns out to be unstable.
1.6.1.2 Stability
A numerical method is said to be stable if numerical errors, e.g., due to roundoff, are
not amplified, and the approximate solution remains bounded. This criterion applies
to time-stepping schemes and iterative solvers alike. Stability refers to the relationship between the exact solution of the discrete problem and the actually computed
solution that includes roundoff and iteration errors. Mathematical tools of stability analysis are available for linear problems with constant coefficients. The most
popular technique is the von Neumann method. Nonlinear problems are more difficult to analyze and may require a stronger form of stability, see Chapter 3. Some
approximations enjoy unconditional stability, others are stable under certain conditions for the choice of input parameters. Unstable schemes may be fixed by adding
extra terms provided that the discretization remains consistent. This idea leads to a
rich variety of stabilized schemes to be presented in what follows.
1.6.1.3 Convergence
A numerical method is said to be convergent if the numerical solution of the discrete
problem approaches the exact solution of the differential equation as the mesh size
and time step go to zero. High-order methods converge to smooth solutions faster
than low-order ones. Consistency and stability are the necessary and sufficient conditions of convergence for finite difference approximations to well-posed linear initial value problems. This statement is known as the Lax equivalence theorem. In the
nonlinear case, compactness is the main ingredient of convergence proofs [216].
In practical computations, convergence must be verified numerically by running
the same simulation on a series of successively refined meshes and varying the time
step size. The results of this grid convergence study can also be used to estimate the
genuine order of accuracy [104, 218, 283]. If the discrepancy between the solutions
obtained with several different meshes and time steps is insignificant, this indicates
that the numerical errors are small, and the results are close enough to the exact
1.6 Fundamental Design Principles
33
solution of the differential equation. Otherwise, the refinement process should be
continued to make sure that the method converges to a grid-independent solution.
1.6.2 Physical Constraints
Consistency, stability, and convergence are the three cornerstones of numerical approximation. A discretization that meets all of these requirements is guaranteed to
produce an accurate solution provided that h and ∆ t are sufficiently small. However,
the definition of “sufficiently small” is highly problem-dependent. If a numerical
scheme fails to resolve a small-scale feature properly on a given mesh, it typically
reacts by generating large numerical errors and/or nonphysical side effects, such as
a spontaneous loss/gain of mass or spurious oscillations, also known as ‘wiggles’.
The strongest violation of physical realism is likely to occur in the vicinity of discontinuities, moving fronts, interior/boundary layers, and other regions in which the
solution gradients are steep. In many cases, small imperfections may be tolerated but
there are situations in which nonphysical solution behavior is totally unacceptable.
Discretization techniques that may give rise to artificial sources/sinks or negative
concentrations should be avoided. Therefore, certain physical constraints, such as
conservation and boundedness, may need to be enforced at the discrete level [104].
1.6.2.1 Conservation
Since mathematical models of transport phenomena are based on conservation principles, similar principles should apply to the approximate solution. If the quantity
of interest is conserved, then numerical errors can only distribute it improperly. Discrete conservation is a constraint that forces the numerical algorithm to reproduce
an important qualitative property of the physical system correctly. The Lax-Wendroff
theorem states that if a consistent and conservative scheme converges, then it converges to a weak solution of the conservation law. Convergence to nonphysical weak
solutions, such as shocks moving at wrong speeds, is ruled out by this theorem.
In the asymptotic limit, even a nonconservative scheme will produce correct results if it is consistent and stable. Unfortunately, it is difficult to tell how fine the
mesh and time step must be chosen to keep conservation errors small enough. Thus,
even solutions to rather simple problems may behave in an unpredictable manner.
A finite difference scheme proves conservative if it can be written in terms of
numerical fluxes from one grid point into another. Finite volume and discontinuous Galerkin methods are conservative by construction, both globally and locally.
The continuous Galerkin FEM provides global conservation [134] and is claimed
to be locally conservative by some authors [158, 162]. Conservation may be lost if
inaccurate quadrature rules or nonstandard approximations are employed. Any modification of a conservative scheme is dangerous and must be examined critically.
34
1 Getting Started
1.6.2.2 Boundedness
In many applications, the transported quantity must stay within certain bounds for
physical reasons. For example, densities and temperatures must be nonnegative; volume and mass fractions must be bounded by 0 and 1. Under certain assumptions,
analytical solutions to scalar transport equations are known to attain their maxima
and/or minima on the boundary of the domain. In unsteady problems, the local extrema of initial data may also serve as upper or lower bounds. Maximum and minimum principles for PDEs of different types are presented in Chapter 3. Similar
constraints can be formulated for the nodal values of the discrete solution.
If a numerical approximation fails to satisfy an a priori bound based on the
known properties of the exact solution, it can easily be repaired by clipping all
undershoots and overshoots. However, pointwise correction of nodal values is a
dangerous practice since the conservation property may be lost. Some low-order
approximations are conservative and bounded but their accuracy leaves a lot to be
desired. High-order schemes perform well for smooth data but may produce unbounded solutions in the neighborhood of steep fronts. A conservative and bounded
convergent scheme of high order is rather difficult to design and expensive to run
but the results are usually rewarding. Indeed, if the total amount of the conserved
quantity is correct, its distribution is accurate, and all relevant upper/lower bounds
are satisfied, then the numerical solution must be very close to the exact one. Boundedness implies strong stability, so a consistent and bounded scheme is convergent.
1.6.2.3 Causality
In some situations, it is important to make sure that information travels in the right
direction (downstream and forward in time) and at the right speed. This principle
is known as casuality [265]. Some algortihms transmit information too far or fast;
others fail to reflect the one-way pattern of wave propagation. The principle of casuality requires that large differences between the analytical and numerical domains
of dependence be avoided, as far as possible. A good numerical method should be
faithful to the nature of the physical and mathematical problem to be solved.
1.6.3 The Basic Rules
Since the end product of the discretization process is an algebraic system, the above
design principles impose certain restrictions on the coefficients of the numerical
scheme. Four basic rules that ensure conservation, boundedness, and physical realism were formulated by Patankar [268] three decades ago. In this section, we restate
these guidelines in a form suitable for our purposes. Later, we will put them on a
firm mathematical basis, explain their far-reaching implications, and use them as a
1.6 Fundamental Design Principles
35
design tool. As we will see, Patankar was far ahead of his time in recognizing the
importance of algebraic constraints for the development of numerical methods.
1.6.3.1 Mass Balance
The first basic rule is: no mass should be created or destroyed inside the domain by
the discretized convective and diffusive terms. The global mass balance may only
change due to sources, sinks, and nonzero fluxes across the boundary.
In our systems (1.36) and (1.37), the discrete mass that belongs to node i is given
by m|i = ∑ j mi j u j , where mi j is a coefficient of the mass matrix M. The sum of its
coefficients should be equal to the area/volume of the computational domain.
The evolution of masses m|i is governed by (1.34) or, in the case of a diagonal
mass matrix, (1.35). Summing over i, one obtains the global mass balance
"
#
dm
= ∑ ri − ∑(ci j + di j )u j ,
(1.59)
dt
i
j
where m = ∑i m|i is the total mass that must satisfy a discrete conservation principle.
Depending on the choice of discretization techniques, the correctness of the mass
balance (1.59) can be checked and maintained in (at least) two different ways:
• Finite volume methods and some of their finite difference counterparts can be
expressed in terms of numerical fluxes fi j defined as suitable approximations to
(1.46). For the scheme to be conservative, each pair of fluxes must satisfy
f ji = − fi j ,
∀ j 6= i.
(1.60)
The flux fi j corresponds to the amount of mass transported by convection and/or
diffusion from node i into node j. Since the flux f ji is the negative of fi j , what is
subtracted from one node is added to another. Hence, mass is conserved.
• Finite element methods can also be written in terms of numerical fluxes that
satisfy (1.60), see Chapter 3. Another useful criterion is based on the properties
of the discrete convection and diffusion operators C = {ci j } and D = {di j }. Note
that the contribution of u j to the right-hand side of (1.59) vanishes if
∑ ci j = 0,
i
∑ di j = 0.
(1.61)
i
Therefore, the scheme is conservative if the matrices C and D have zero column
sums, except for a small set of nodes located on the boundary or next to it. This
property was mentioned in [43] in the context of finite difference discretizations.
At the stage of testing and debugging, it is useful to check if the total mass evolves
in the right way. If this is not the case, the code is likely to contain a pernicious bug.
36
1 Getting Started
1.6.3.2 Zero Row Sums
The second basic rule is: if a continuous operator produces zero when applied to a
constant, so should its discrete counterpart. If we consider (1.13) with s ≡ 0, this
equation remains valid if the solution and its boundary values are increased by an
arbitrary constant. For the solution of the discrete problem to possess the same property, it is sufficient to require that the matrices C and D have zero row sums
∑ ci j = 0,
j
∑ di j = 0.
(1.62)
j
Then the discrete solution is defined up to an additive constant. Uniqueness follows
from the Dirichlet boundary conditions to be prescribed for at least one node.
By definition, the diffusive flux is proportional to the gradient of u, so the rows
of D should always satisfy the zero-sum rule. However, the row sums of C may be
nonzero if its continuous counterpart is given by (1.12) and the velocity field is not
divergence-free. In this case, the governing equation contains a zeroth-order term of
the form (∇ · v)u, so the rule is not applicable to the discrete convection operator. If
we force the numerical solution to behave in a certain way, then our intention is to
mimic some qualitative properties of the exact solution. The basic rules should not
be used blindly in situations when the underlying assumptions do not hold.
1.6.3.3 Positive Coefficients
The third basic rule is: if convection and diffusion are the only processes to be
simulated, the nodal value uin+1 should not decrease as result of increasing any other
nodal value that appears in the discretized equation for node i. Conversely, it should
not increase if another nodal value is decreased, all other things being fixed.
In the absence of a reactive term, the fully discrete problem can be written as
Aun+1 = Bun ,
(1.63)
where A = {bi j } and B = {bi j } contain the coefficients of the implicit and explicit
part, respectively. The nodal value uin+1 satisfies the following algebraic equation
aii uin+1 = ∑ bi j unj − ∑ ai j un+1
j .
(1.64)
j6=i
j
Obviously, the requirements of the third rule are satisfied if the coefficients of all
nodal values that contribute to this equation have the same sign. Following Patankar
[268], we choose them to be positive. Strictly speaking, we require that
aii > 0,
ai j ≤ 0,
bii ≥ 0,
bi j ≥ 0,
∀i,
∀ j 6= i.
(1.65)
(1.66)
1.6 Fundamental Design Principles
37
In the context of explicit finite difference schemes, positivity constraints of this form
were formulated by Book et al. [43] as early as in 1975. If ai j = 0 for all j 6= i, then
it is obvious that a scheme of the form (1.64) is positivity preserving, i.e.,
un ≥ 0
⇒
un+1 ≥ 0,
∀n.
(1.67)
In the implicit case, a proof of this property is based on the M-matrix property of A
which ensures that all coefficients of A−1 are nonnegative, see Chapter 3.
As we will see, conditions (1.65)–(1.66) and similar algebraic constraints can be
used to achieve many favorable properties, such as positivity, monotonicity, nonincreasing total variation, and the discrete maximum principle, to name just a few.
1.6.3.4 Negative Slopes
The fourth basic rule is: if the discretization of convective and diffusive terms is
positivity-preserving, inclusion of a reactive part should not destroy this property.
Sources and sinks may reverse the sign of analytical and numerical solutions
alike. It is not unusual that they trigger numerical instabilities, cause divergence of
iterative solvers, or give rise to nonphysical artifacts. Hence, the treatment of zerothorder terms requires special care. Assume that a nonlinear reactive term ri can be
split into a (positive) source and a (negative) sink proportional to ui as follows
ri = βi − αi ui .
(1.68)
The linearization parameters αi and βi may depend on the unknown solution values.
It is instructive to consider the extreme situation in which ri is very large as
compared to other terms and, therefore, equation (1.34) reduces to ri = 0 or
αi ui = βi .
(1.69)
To secure convergence and keep the numerical solution positive, the splitting of the
reactive term should be designed so that αi ≥ 0 and βi ≥ 0 for all i, see [268].
Even if the contribution of other terms to the discretized equation cannot be neglected, it is worthwhile to express the reactive term ri into the form (1.68) with
nonnegative coefficients αi and βi . This negative-slope linearization technique was
invented by Patankar [268] to preserve the sign of inherently positive variables.
Guidelines for constructing splittings of the form (1.68) can be found in his book.
The aspects of positivity preservation for linearized reactive terms of the form
(1.68) with βi = ci ui ≥ 0 were also analyzed by MacKinnon and Carey [246] who
mentioned that such a splitting is possible, e.g., for zeroth-order kinetics (radiative
decay: r = −cu ⇒ α = c ≥ 0, β = 0), a combination of first- and second-order
kinetics (population growth: r = c1 u − c2 u2 ⇒ α = c2 u ≥ 0, β = c1 u ≥ 0), and
reactions of fractional order (for instance: r = −cu3/2 ⇒ α = cu1/2 ≥ 0, β = 0).
38
1 Getting Started
1.7 Scope of This Book
The models, techniques, and concepts presented so far were chosen so as to get
started with numerical simulation of transport phenomena without going too much
into detail. The range of topics we have covered is certainly eclectic and incomplete. Other texts should be consulted for an in-depth introduction to the field of
Computational Fluid Dynamics.
This book is concerned with the numerical treatment of transport equations that
cannot be handled using conventional discretization techniques, or the results are
unsatisfactory. If the solution of the continuous problem varies on a length scale
shorter than the mesh size, small-scale features cannot be captured accurately by
any numerical scheme. Thus, insufficient mesh resolution gives rise to large errors
that may result in an incorrect qualitative behavior of approximate solutions.
Each scheme reacts in its own way when it encounters an unresolvable subgridscale feature. Typical side effects are strong numerical diffusion and/or spurious
oscillations (wiggles, ripples). Arguably, an overly diffusive scheme is the lesser of
the two evils if it is guaranteed to be bounded. Another arguable viewpoint is [128]:
“don’t suppress the wiggles — they’re telling you something!” We take the liberty
to formulate another rule: a good algorithm must contain just as much numerical
dissipation as is necessary to avoid nonphysical artifacts. Moreover, the mesh should
be refined adaptively in troublesome regions where small-scale effects are present.
In most cases, the bizarre behavior of numerical solutions is due the fact that
• the model is dominated by convection, anisotropic diffusion, or stiff reaction;
• discontinuities, steep gradients, and/or interior/boundary layers are involved;
• the employed numerical scheme violates at least one of the four basic rules.
A possible remedy is to take a suitable high-order discretization and constrain it
at the algebraic level so as to enforce desirable properties without losing too much
accuracy. This is the approach that we will pursue and promote in this book.
Most of the material presented in this text is not really new. Many excellent books
and review articles have been written about numerical methods for convectiondiffusion equations and hyperbolic conservation laws. The main reasons that have
led the author to retell the story are as follows:
•
•
•
•
useful techniques are scattered over a vast body of literature and difficult to find;
many algorithms are inherently explicit or require the use of structured meshes;
texts overloaded with complex mathematical theory are unreadable to engineers;
rigorous convergence proofs may disguise the fact that the method does not really
work when applied to problems in which subgrid-scale effects are important.
In the introductory part, we select a number of particularly good and relatively simple methods. We analyze the properties of these methods and put them in a unified framework so as to highlight existing similarities. Also, we present some wellknown concepts in a new light and interpret them from the algebraic viewpoint. This
preliminary study is intended to pave the way for various generalizations.
1.7 Scope of This Book
39
The algebraic flux correction paradigm [200] to be introduced in Chapter 4 is
the methodology developed by the author and his coworkers to enforce boundedness in a conservative manner. It will be explained in the context of scalar transport
equations discretized by explicit and implicit finite element methods on arbitrary
meshes. Special-purpose algorithms will be presented for the numerical treatment
of stationary and time-dependent problems alike. However, our main goal is to show
how the underlying design philosophy works and to equip the reader with tools that
make it possible to fix a given discretization building on the four basic rules and
similar algebraic constraints.
A large share of CFD research is performed by people living in the parallel
worlds of ‘viscous incompressible’ or ‘inviscid compressible’ flows. The former
is populated by practitioners who are interested primarily in elliptic/parabolic problems. People from this group typically favor implicit algorithms, finite element discretizations, and unstructured meshes. Inviscid flows are governed by hyperbolic
conservation laws, and the traditional solution strategy relies on explicit finite difference or finite volume discretizations. Of course, there are many notable exceptions
to this rule, and some diffusion of ideas takes place at the interface between the two
worlds. In our experience, a lot can be gained by looking at what is going on both
sides of this interface.
Space and time do not permit a comprehensive review of all promising numerical
schemes and discretization concepts. We will not discuss the recent developments
in the realm of finite volume and discontinuous Galerkin methods. These two mainstream trends are covered in many texts and research papers. Instead, we focus on
continuous Galerkin methods and make them fit for the numerical treatment of transport equations at arbitrary Peclet numbers. Many methods presented in this book
were chosen with this objective in mind. Of course, this choice has been influenced
by the author’s personal preferences, views, and research objectives. However, the
results of recent comparative studies [174, 175] indicate that it is a fairly good one.
Chapter 2
Finite Element Approximations
This chapter presents a self-contained review of some promising discretization and
stabilization techniques for multidimensional transport equations in two and three
space dimensions. The methods to be discussed do not require directional splitting and are readily applicable to unstructured meshes. The group finite element
interpolation of the flux function provides a handy link between finite element and
finite volume approximations. The existence of a conservative flux decomposition
paves the way to various extensions of one-dimensional algorithms. Also, it leads
to efficient edge-based data structures that offer a number of significant advantages
as compared to the traditional element-based implementation. In this chapter, we
consider unstructured grid methods for convection-diffusion equations and discuss
relevant algorithmic details, such as matrix assembly. Furthermore, we analyze the
properties of discrete operators and describe some popular approaches to the design
of stabilized finite element methods for stationary and time-dependent problems.
2.1 Discretization on Unstructured Meshes
Unstructured grid methods are commonly employed if the domain of interest has
a complex geometrical shape. The finite element approach provides a particularly
convenient framework for the development of general-purpose software packages.
Its advantages include but are not limited to the flexible choice of basis functions,
simple treatment of natural boundary conditions, fully automatic matrix assembly,
and the possibility of mesh adaptation backed by rigorous mathematical theory.
The roots and the traditional strength of finite element methods lie in the field of
elliptic problems and applications to structural engineering, whereas finite volumes
lend themselves to numerical simulation of transport phenomena. One of the reasons
for this state of affairs is that many one-dimensional concepts and geometric design
criteria introduced in the context of finite difference schemes are relatively easy to
extend to finite volume discretizations on unstructured meshes, while an extension
to finite elements is often nontrivial or even impossible. In the case of low-order
41
42
2 Finite Element Approximations
approximations, it turns out that finite element and finite volume schemes are largely
equivalent, although their derivation is based on entirely different premises.
The existing similarities between the two approaches to discretization of transport equations on unstructured meshes have been recognized, documented, and exploited by many authors [8, 56, 166, 300, 301]. In particular, edge-based finite element methods have been formulated in terms of numerical fluxes that move the mass
from one node into another without changing the global balance [226, 239, 259].
This formulation is particularly well suited for compressible flows. In spite of its
enormous potential, as demonstrated by spectacular simulation results for aerodynamic applications [22, 23, 226], many finite element practitioners are unaware of
its existence or reluctant to use an unconventional methodology. It is hoped that the
present book will provide additional evidence in favor of edge-based algorithms.
The most general model problem to be dealt with in the present chapter is a scalar
transport equation to be solved in a domain Ω with boundary Γ = ΓD ∪ ΓN
∂u
+∇·f = s
∂t
in Ω ,
(2.1)
where u is a conserved scalar quantity, s is a source term, and f is a generic flux
function that consists of a diffusive and/or convective part. We will also consider the
steady-state counterpart of equation (2.1) in which the time derivative is omitted.
A typical set of initial and boundary conditions for problem (2.1) is as follows
u = u0 ,
u = g,
f · n = h,
at t = 0,
on ΓD ,
on ΓN ,
(2.2)
where n denotes the unit outward normal. The appropriate choice of initial/boundary
conditions depends on the problem at hand and on the available information.
In what follows, we consider continuous (linear or multilinear) finite element discretizations and illustrate their relationship to finite volumes. Starting with a weak
form of (2.1), we derive the coefficient matrices and examine their properties. We
revisit the aspects of conservation, upwinding, and artificial diffusion in the FEM
context. Also, we split the contributions of discrete convection and diffusion operators into skew-symmetric numerical fluxes that represent the rate of bilateral mass
transfer between two neighboring nodes. The use of edge-based data structures is
feasible but not mandatory. Furthermore, it is possible to assemble matrices elementby-element but define extra stabilization terms in terms of numerical fluxes associated with edges of the sparsity graph. This approach simplifies the implementation
of edge-based artificial diffusion methods in an existing finite element code.
The presentation is focused on simple stabilization mechanisms suitable for multidimensional transport problems at moderately large Peclet numbers. Moreover, it
is assumed that the computational mesh is sufficiently regular and does not contain
inordinately stretched/deformed elements. Advanced mathematical theory and nonlinear high-resolution schemes for computing nonoscillatory solutions to problems
with steep fronts on arbitrary meshes will be developed in the next two chapters.
2.1 Discretization on Unstructured Meshes
43
2.1.1 Group Finite Element Formulation
The starting point for a finite element discretization of the scalar conservation law
(2.1) is its weak form that corresponds to the weighted residual formulation
Z
Z
∂u
w
ws dx.
(2.3)
+ ∇ · f dx =
∂t
Ω
Ω
This relation must hold for any test function w vanishing on the boundary part ΓD ,
see Section 1.2.5. To get started, consider a linear flux function of the form
f = vu − ε ∇u,
(2.4)
where v is the velocity vector and ε is a constant diffusion coefficient. After substitution of the so-defined flux f and integration by parts, equation (2.3) becomes
Z
Z
Z ∂u
ws dx +
w + w∇ · (vu) + ∇w · (ε ∇u) dx =
w(ε ∇u) · n ds. (2.5)
∂t
Ω
Ω
ΓN
The integral over ΓN is evaluated using the diffusive flux given by the Neumann
boundary conditions. If the total flux is prescribed on ΓN , then integration by parts
should also be applied to the convective term. The corresponding weak form is
Z
Z
Z ∂u
w − ∇w · (vu − ε ∇u) dx =
w(vu − ε ∇u) · n ds. (2.6)
ws dx −
∂t
Ω
Ω
ΓN
The next step is to represent the approximate solution uh as a linear combination
of piecewise-polynomial basis functions {ϕi } spanning a finite-dimensional space
uh = ∑ u j ϕ j .
(2.7)
j
This definition yields a convenient separation of variables, such that the spatial and
temporal variations of uh are associated with u j (t) and ϕ j (x), respectively.
In the course of differentiation, time derivatives are applied to the nodal values
u j , whereas gradient and divergence operators are applied to the basis functions ϕ j .
For example, the approximate diffusive flux at a given point x is proportional to
∇uh = ∑ u j ∇ϕ j .
(2.8)
j
The conventional approximation of the convective flux involves the multiplication
of (2.7) by the instantaneous velocity that can be interpolated in the same way
vh = ∑ v j ϕ j .
j
(2.9)
44
2 Finite Element Approximations
For the time being, the nodal values v j are assumed to be known. In real-life applications, they are computed numerically from a momentum balance equation.
The straightforward product rule is not the only and often not the best way to approximate the convective flux that appears in (2.5) and (2.6). Instead of introducing
separate trial solutions for each variable, it is possible to interpolate the nodal values
of a function that depends on a group of variables. This approximation technique is
known as the group finite element formulation [109, 110, 111]. It eliminates the need
for dealing with products of basis functions and leads to simpler discretized equations, which results in significant savings especially in the case of highly nonlinear
problems. In particular, this kind of approximation provides a natural treatment of
inviscid fluxes in conservation laws. In many situations, it turns out to be more accurate than the independent approximation of the involved variables [109, 110].
The group finite element approximation of the convective flux is given by
(vu)h = ∑(v j u j )ϕ j .
(2.10)
j
Obviously, this function is easier to differentiate than the product vh uh . In particular,
the divergence of the convective flux in equation (2.5) is approximated by
∇ · (vu)h = ∑ u j (v j · ∇ϕ j ).
(2.11)
j
That is, the conservative semi-discrete form of the convective term is the sum of
nodal values multiplied by the convective derivatives of the basis functions.
If formulation (2.5) is adopted, the resultant semi-discrete problem is as follows
Z
Ω
∂ uh
dx +
(wh ∇ · (vu)h + ∇wh · (ε ∇uh )) dx
∂t
Z
ZΩ
Z
wh
=
Ω
wh s dx +
ΓN
wh (ε ∇uh ) · n ds,
(2.12)
where the weighting function wh is taken from a suitable finite-dimensional space.
The finite element discretization based on alternative weak form (2.6) reads
Z
Ω
wh
∂ uh
∇wh · ((vu)h − ε ∇uh ) dx
dx −
∂t
Z
ZΩ
Z
=
wh s dx −
Ω
ΓN
wh ((vu)h − ε ∇uh ) · n ds.
(2.13)
Remark 2.1. In the case wh ≡ 1 and ΓN = Γ , the divergence theorem can be used to
show that both discretizations reduce to the integral form of the conservation law
∂
∂t
Z
Ω
uh dx +
Z
Γ
((vu)h − ε ∇uh ) · n ds =
Z
Ω
s dx,
which proves that the group finite element formulation is globally conservative.
2.1 Discretization on Unstructured Meshes
45
In the Galerkin method, the weighting functions wh are taken from the same finitedimensional space as the basis functions. Substitution of (2.7), (2.8), and (2.11) into
(2.12) with wh = ϕi yields the following equation for nodal value ui
Z
Z
du j
∑ Ω ϕi ϕ j dx dt + ∑ Ω (ϕi (v j · ∇ϕ j ) + ∇ϕi · (ε ∇ϕ j )) dx u j
j
j
=
Z
Ω
ϕi s dx −
Z
ΓN
ϕi h ds.
(2.14)
The value of the flux h = −(ε ∇u) · n is furnished by the Neumann boundary conditions prescribed on ΓN . If the nodal value ui is associated with a point lying on
ΓD = Γ \ΓN , then equation (2.14) should be replaced by the algebraic relation ui = gi
that follows from the Dirichlet boundary conditions. In a practical implementation,
it is convenient to do so after the matrix assembly process has been fully completed.
Therefore, we will first consider the discretized equations before the imposition of
Dirichlet boundary conditions but bear in mind that some of these equations are actually redundant and need to be eliminated to obtain a well-posed discrete problem.
The system of equations (2.14) for all nodes can be written in matrix form as
MC
du
= Ku + r,
dt
(2.15)
where MC = {mi j } is the consistent mass matrix, K = {ki j } is the negative of the
discrete transport operator, and r = {ri } is given by the right-hand side of (2.14)
ri =
Z
Ω
ϕi s dx −
Z
ΓN
ϕi h ds.
(2.16)
The coefficients of the matrices MC and K can also be inferred from (2.14)
mi j =
Z
Ω
ϕi ϕ j dx,
ki j = −ci j · v j − ε di j ,
(2.17)
where the nodal velocities v j may vary with time, the diffusion coefficient ε is constant by assumption, and the value of ki j depends on the discretized derivatives
ci j =
Z
Ω
ϕi ∇ϕ j dx,
di j =
Z
Ω
∇ϕi · ∇ϕ j dx.
(2.18)
For the discretization based on (2.13), the values of mi j and di j are the same but
ki j = c ji · v j − ε di j
(2.19)
and the contribution of the surface integral to (2.16) should be assembled using the
total flux h = (vu − ε ∇u) · n prescribed on the Neumann boundary part ΓN .
Remark 2.2. If the boundary value u = g rather than the flux is available, it can be
treated as Neumann boundary condition with h = (v · n)g. The advantages of weakly
imposed Dirichlet boundary conditions have been explored in [25, 176, 212, 293].
46
2 Finite Element Approximations
In the notation of definitions (2.16)–(2.19) the i−th equation assumes the form
∑ mi j
j
du j
= ∑ ki j u j + ri ,
dt
j
(2.20)
where the actual values of ki j and ri depend on whether (2.12) or (2.13) is employed.
At the initialization stage, the coefficients mi j , ci j , and di j need to be evaluated
using numerical integration. In finite element codes, the matrix assembly process is
fully automatic and involves transformations to a reference element on which the
local basis functions are defined [76, 176]. In the case of a time-dependent velocity
field, the convective part of the discrete operator K must be updated at each time
step. On a fixed Eulerian mesh, this can be accomplished in a very efficient way
since the coefficients ci j and di j do not change as time evolves. If they are stored, the
computation of ki j from the above formulas can be performed at a fraction of the cost
that the repeated use of element-by-element matrix assembly would require. This is
one of the remarkable features peculiar to the group finite element formulation.
The above derivation of the semi-discrete equations is valid for Lagrangian elements of arbitrary shape and order. Traditionally, linear and bilinear approximations
have been the workhorse of finite element methods for the equations of fluid dynamics. In many cases, the lack of coercivity or the presence of unresolvable small-scale
effects makes the use of higher-order basis functions impractical. Also, the intricate coupling between the degrees of freedom makes it more difficult to control the
behavior of numerical solutions than in the case of linear or bilinear elements.
2.1.2 Properties of Discrete Operators
Consistency requires that approximations (2.7) and (2.8) be exact at least for constant functions. To this end, the sum of basis functions must equal 1 everywhere
∑ ϕ j ≡ 1.
(2.21)
j
Differentiating (2.21), we deduce that the gradients of ϕ j must sum to zero
∑ ∇ϕ j ≡ 0.
(2.22)
j
This property ensures that the gradient of uh is zero if uh is a constant function
u j = c,
∀j
⇒
∇uh ≡ 0.
uh ≡ c,
Due to (2.21), the i−th row sum of the consistent mass matrix MC = {mi j } is
∑ mi j =
j
Z
Ω
ϕi ∑ ϕ j dx =
j
Z
Ω
ϕi dx.
2.1 Discretization on Unstructured Meshes
47
Furthermore, summation over all indices yields the volume (area) of the domain Ω
∑ ∑ mi j =
i
j
Z
∑ ϕi dx = |Ω |.
(2.23)
Ω i
The discrete Laplacian D = {di j } is symmetric with zero row and column sums
∑ di j =
Z
∑ di j =
Z
j
i
Ω
∇ϕi · ∑ ∇ϕ j dx = 0,
(2.24)
∑ ∇ϕi · ∇ϕ j dx = 0.
(2.25)
j
Ω i
The discrete gradient/divergence operator C = {ci j } is nonsymmetric and
∑ ci j =
Z
ϕi ∑ ∇ϕ j dx = 0,
∑ ci j =
Z
∑ ϕi ∇ϕ j dx =
j
i
Ω
(2.26)
j
Ω i
Z
Ω
∇ϕ j dx.
(2.27)
The zero row sum property of C and D is dictated by the second basic rule from
Section 1.6.3. It ensures that the result of numerical differentiation equals zero if all
nodal values of the numerical solution coincide and define a constant field.
By definition, finite element basis functions have compact support. That is, ϕi
and all of its derivatives are zero outside the set of elements which contain node
xi . Likewise, products of basis functions and/or their derivatives are nonvanishing
only on a small patch Ωi j of elements, as shown in Fig. 2.1 for a triangular mesh.
Hence, all volume and surface integrals shrink to those over Ωi j and its boundary
Γi j , respectively. This property guarantees that the resulting matrices are sparse.
For example, the assembly of the mass matrix MC involves the computation of
mi j =
Z
Ω
ϕi ϕ j dx =
Z
Ωi j
ϕi ϕ j dx
which can be performed element-by-element using numerical or exact integration.
0000000000000000000
1111111111111111111
1111111111111111111
0000000000000000000
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
Γii
0000000000000000000
1111111111111111111
Ωii
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
xi
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
0000000000000000000
1111111111111111111
xj
00000000000
11111111111
11111111111
00000000000
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
Γij
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
Ωij
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
xi11111111111
00000000000
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
Fig. 2.1 Integration regions for linear basis functions on a triangular mesh.
xj
48
2 Finite Element Approximations
Assume that at least the mean values of ϕi ϕ j are continuous across interelement
boundaries. Then integration by parts within each cell Ωk ∈ Ωi j reveals that
ci j + c ji =
Z
Ωi j
(ϕi ∇ϕ j + ϕ j ∇ϕi ) dx =
Z
Γi j ∩Γ
ϕi ϕ j n ds.
(2.28)
The integrals of ϕi ϕ j over an interface between any pair of cells that belong to
Ωi j cancel out due to continuity. The integral over the outer boundary of Ωi j is
nonvanishing only on Γi j ∩ Γ since the product of the two basis functions is zero
elsewhere. In the case of linear and bilinear elements, formula (2.28) implies that
c ji = −ci j unless both nodes lie on the boundary Γ and belong to the same element.
In particular, cii = 0 for any interior node i due to the fact that ϕi vanishes on the
whole outer boundary Γii of the integration region Ωii , as depicted in Fig. 2.1.
Due to the local support property, relation (2.28) can be reformulated as follows
ci j + c ji = bi j ,
bi j =
Z
Γ
ϕi ϕ j n ds.
(2.29)
The vector-valued weights bi j distribute the surface area between pairs of nodes in
much the same way as the coefficients mi j distribute the volume of the domain Ω .
By virtue of (2.29) and (2.26), the column sums (2.27) of C are equal to
∑ ci j = ∑ bi j = b j ,
i
bj =
i
Z
Γ
ϕ j n ds.
(2.30)
If the basis function ϕ j is associated with an internal node, then it vanishes on the
boundary Γ and the j−th column sum b j equals zero. Otherwise the result is
b j = n js j,
nj =
1
sj
Z
Γ
ϕ j n ds,
sj =
Z
Γ
ϕ j ds,
(2.31)
where s j is the surface area associated with a boundary node j and n j is the corresponding unit outward normal vector. Since the basis functions sum to unity
∑sj =
j
Z
∑ ϕ j ds = |Γ |.
Γ j
The above properties of the coefficient matrices MC , C and D are not only of purely
theoretical interest. As we will see shortly, they are of fundamental importance for
the analysis and design of finite element approximations on unstructured meshes.
2.1.3 Conservation and Mass Lumping
The fact that the finite element discretization (2.20) of the transport equation is
globally conservative can be easily inferred from the above properties of coefficient
2.1 Discretization on Unstructured Meshes
49
matrices. The total ‘mass’ of the numerical solution uh is defined as the integral
m(t) =
Z
uh (x,t) dx = ∑ mi ui ,
Ω
(2.32)
i
where the row sum mi represents the share of the domain Ω associated with node i
mi =
Z
Ω
ϕi dx = ∑ mi j .
(2.33)
j
Summing equations (2.20) over i, one obtains the global mass balance equation
dm
= ∑ ∑ ki j u j + ∑ ri .
dt
j i
i
If the coefficients ki j are defined by (2.19), then the properties of C and D imply
∑ ∑ ki j u j = ∑(v j u j ) · ∑ c ji − ∑ u j ∑ ε di j = 0.
j
i
j
i
j
i
Thus, only source terms and fluxes across the boundary may affect the total mass.
The semi-discrete scheme with ki j defined by (2.17) is also conservative since
∑ ∑ ki j u j = − ∑(v j u j ) · ∑(bi j − c ji ) − ∑ ui ∑ ε di j
j
i
j
i
j
i
= − ∑(v j u j ) · ∑ bi j = − ∑(v j u j ) · b j .
j
i
(2.34)
j
By definition of the convective flux (2.11) and of the integrated normal vector b j
∑(v j u j ) · b j =
j
Z
Γ
(vu)h · n ds.
Again, the total mass may change only due to internal sources and boundary fluxes.
Remark 2.3. The imposition of Dirichlet boundary conditions in the strong sense
may apparently invalidate the proof of global conservation. Indeed, the elimination
of the corresponding equations from the algebraic system means that wh = ∑i ϕi ≡ 1
is no longer an admissible test function, and the column sums of the resultant coefficient matrices are different. This problem can be rectified by including a set of
equations for the boundary fluxes weighted by the omitted basis functions [158].
Remark 2.4. Clearly, the global conservation statement remains valid for Dirichlet
boundary conditions imposed in a weak sense, e.g., as proposed in Remark 2.2.
It is common practice to approximate the consistent mass matrix MC by a diagonal matrix ML , so as to update the solution in a fully explicit way or make an implicit
algorithm more robust and efficient. This trick is particularly useful in steady-state
50
2 Finite Element Approximations
computations since the approximation of the time derivative has no influence on
the final solution. In the case of time-dependent problems, mass lumping is usually
undesirable since it may adversely affect the phase accuracy of the finite element
scheme. Moreover, it may result in a loss of the intrinsic conservation property.
In general, a diagonal mass matrix ML is a conservative approximation to MC if
∑(ML u)i = ∑(MC u)i
i
(2.35)
i
for an arbitrary vector u. In particular, the sum of its elements should be equal to
the volume/area of the computational domain, as required by (2.23) and (2.35) with
u ≡ 1. These requirements are clearly satisfied by the row-sum mass lumping technique, whereby ML is defined as a diagonal matrix of weights given by (2.33)
mi = ∑ mi j .
ML = diag{mi },
(2.36)
j
By definition, the sum of all nodal values multiplied by the diagonal coefficients of
the so-defined lumped mass matrix ML is equal to the total mass (2.32).
Mass lumping can also be performed using diagonal scaling or inexact evaluation
of the coefficients mi j by a nodal quadrature rule that produces a diagonal matrix. In
fact, the integrals given by (2.33) can be interpreted as weights of a Newton-Cotes
numerical integration scheme. This observation makes it possible to estimate the
quadrature error and quantify the loss of accuracy due to mass lumping [134].
The quadrature-based approach to mass lumping is attractive from the viewpoint
of analysis but a word of caution is in order. It turns out that only the row-sum
formula (2.33) conserves mass in the above sense [134]. Indeed, if we suppose that
there is another diagonal mass matrix M̃L = diag{m̃i } satisfying (2.35) then
∑(ML u)i − ∑(M̃L u)i = ∑(mi − m̃i )ui = 0.
i
i
i
Since u is arbitrary, this means that mi = m̃i for all i, which proves the uniqueness
of the lumped mass matrix ML that enjoys the global conservation property.
Linear and bilinear finite element approximations give rise to mass matrices with
mi j ≥ 0,
mi > 0,
∀i, j
since the basis functions are nonnegative everywhere. In higher-order approximations, the off-diagonal entries of MC are of variable sign, so row-sum lumping may
produce ML with zero or negative diagonal entries. For example, the lumped mass
matrix for the six-node quadratic finite element approximation on a triangular mesh
assigns zero masses to all vertex-based degrees of freedom [73]. Such a lumpedmass Galerkin discretization can be interpreted as a finite volume method with edgecentered degrees of freedom [271]. A possible remedy is to enrich the finite element
basis by adding a bubble function associated with the center of gravity [73].
2.1 Discretization on Unstructured Meshes
51
2.1.4 Variational Gradient Recovery
In many situations, numerical differentiation of the approximate solution is required
to calculate certain derived quantities such as fluxes or curvatures. By definition,
the first derivatives of uh are given by (2.8) but the so-defined gradient ∇uh is
a piecewise-constant function which is discontinuous at interelement boundaries.
Therefore, a direct evaluation of the derivatives at nodes is not feasible, and some
kind of postprocessing is needed to extract information from the available data.
A continuous approximation to the gradient of the solution uh can be defined as
gh = ∑ g j ϕ j ,
(2.37)
j
where g j is a suitable approximation to the vector of first derivatives at node j. It can
be determined, for example, by fitting a polynomial to the values of the piecewiseconstant consistent gradient ∇uh evaluated at a set of points surrounding node j.
These points can be placed so as to exploit the superconvergence phenomenon [190]
and increase the accuracy of approximation by orders of magnitude [362, 363, 364].
Superconvergent postprocessing techniques are frequently employed for purposes
of a posteriori error estimation and adaptive mesh refinement, see Chapter 7.
Another way to determine the nodal values of gh is to perform the L2 -projection
Z
Ω
ϕi gh dx =
Z
Ω
ϕi ∇uh dx,
∀i.
(2.38)
This approach is a classical example of variational gradient recovery which can be
used repeatedly to calculate the nodal values of higher-order space derivatives.
Remark 2.5. The L2 -projection can also be carried out using a different set of basis/test functions {ψi } for the representation of the averaged gradient gh . However,
the choice ψi = ϕi is usually the most natural and economical one [3, 261].
Substitution of (2.8) and (2.37) into (2.38) yields a linear system of the form
MC g = Cu,
(2.39)
where MC = {mi j } is the consistent mass matrix and C = {ci j } is the discrete gradient operator analyzed in the previous sections. As usual, the boldface notation means
that the number on unknowns per mesh node equals that of space dimensions.
The linear system (2.39) can be solved, e.g., using Richardson’s iteration preconditioned by the lumped mass matrix ML which is a usable approximation to MC
ĝ(m+1) = ĝ(m) + ML−1 [Cu − MC ĝ(m) ],
m = 0, 1, 2, . . . .
(2.40)
Since ML is a diagonal matrix, its inversion is trivial. Due to the diagonal dominance
and positive-definiteness of MC , convergence is typically fast. In practical calculations, a fixed number (up to 3) of iterations are usually performed. A single iteration
52
2 Finite Element Approximations
with initial zero guess corresponds to the lumped-mass version of system (2.39)
g = ML−1 Cu.
(2.41)
This kind of gradient reconstruction is very cheap and its accuracy is sufficient for
most practical purposes. Also, the lumped-mass L2 -projection (2.41) is less likely
to generate undershoots and overshoots than gradient recovery based on (2.39).
Due to the zero row sum property, the diagonal coefficients of C are given by
cii = − ∑ ci j .
j6=i
Hence, the equation for the nodal gradient gi admits the following representation
gi =
1
1
ci j u j =
mi ∑
m
i
j
∑ ci j (u j − ui ).
j6=i
Variational gradient recovery, such as the above lumped-mass L2 -projection, can
also be used to approximate the divergence of a vector field. For example
(∇ · g)i =
1
1
ci j · g j =
∑
mi j
mi
∑ ci j · (g j − gi )
(2.42)
j6=i
yields an approximate value of the Laplacian ∆ u at node i. As an alternative, secondorder derivatives can be recovered directly using integration by parts [226]
Z
Ω
ϕi (∆ u)h dx =
Z
Γ
ϕi (n · ∇uh ) ds −
Z
Ω
∇ϕi · ∇uh dx,
∀i.
(2.43)
After row-sum mass lumping, the i−th equation in this system can be written as
(∆ u)i =
1
1
hi j u j −
∑
mi j
mi
∑ di j (u j − ui ),
j6=i
where di j and hi j represent the contributions of the volume and surface integrals to
the right-hand side of (2.43), respectively. Neumann boundary conditions (if any)
can be built into the integral over ΓN . This approach to recovery of (∆ u)i is faster
and typically more accurate than repeated application of first derivatives.
The divergence of the velocity field vh can be approximated in the same way
as that of the averaged gradient in equation (2.42). Interestingly enough, the result
is proportional to the row sums of the discrete transport operator K = {ki j } with
coefficients ki j = −ci j · v j − ε di j . Indeed, using property (2.24) one obtains
∑ ki j = − ∑ ci j · v j ≈ −mi (∇ · v)i .
j
j
Loosely speaking, the velocity field is discretely divergence-free if the group finite
element approximation (2.10) of the convective flux leads to K with zero row sums.
2.1 Discretization on Unstructured Meshes
53
It is instructive to decompose the i−th component of the vector Ku as follows
∑ ki j u j = ∑ ki j (u j − ui ) + ui ∑ ki j .
j6=i
j
(2.44)
j
In light of the above, the following approximations are associated with each term
∑ ki j u j ≈ −mi (∇ · (vu))i − mi (ε∆ u)i ,
(2.45)
∑ ki j (u j − ui ) ≈ −mi (v · ∇u)i − mi (ε∆ u)i ,
(2.46)
ui ∑ ki j = −ui ∑ ci j · v j ≈ −mi ui (∇ · v)i .
(2.47)
j
j6=i
j
j
In conclusion, the sum over j 6= i is the ‘incompressible’ part of the discrete transport
operator K. The ‘compressible’ part (2.47) vanishes if K has zero row sums.
2.1.5 Treatment of Nonlinear Fluxes
In many practical applications, the diffusive terms can be neglected, while the flux
function f(u) depends on the unknown solution u in a nonlinear way. In this case,
equation (2.1) represents a hyperbolic conservation law which can be discretized in
much the same way as the convective part of a linear transport equation. Let
fh = ∑ f j ϕ j ,
f j = f(u j )
(2.48)
j
be the group finite element approximation of the inviscid flux to be inserted into
Z
Ω
wh
∂ uh
dx +
∂t
Z
Ω
Z
wh ∇ · fh dx =
Ω
wh s dx.
(2.49)
Remark 2.6. As before, the global balance equation can be inferred from (2.49) with
wh = ∑i ϕi ≡ 1 by applying the divergence theorem element-by-element
∂
∂t
Z
Ω
uh dx +
Z
Γ
fh · n ds =
Z
Ω
s dx.
The semi-discrete Galerkin equation that corresponds to wh = ϕi is given by
∑ mi j
j
du j
= − ∑ ci j · f j + ri .
dt
j
(2.50)
Note that the right-hand side of this equation contains a linear combination of nodal
fluxes f j = f(u j ) that depend on the solution values u j at the corresponding nodes.
If an explicit time-stepping strategy is adopted, the fluxes can be evaluated using
54
2 Finite Element Approximations
the available solution from the previous time level. The matrix assembly process for
implicit algorithms involves a suitable linearization of the discretized fluxes.
Due to the zero row sum property (2.26) of the matrix C = {ci j }, we have
∑ mi j
j
du j
= − ∑ ci j · (f j − fi ) + ri .
dt
j6=i
(2.51)
Furthermore, it is possible to transform this equation into the quasi-linear form
∑ mi j
j
du j
= − ∑ ci j · vi j (u j − ui ) + ri ,
dt
j
(2.52)
where vi j is a characteristic velocity such that (2.51) and (2.52) are equivalent
(
f(u j )−f(ui )
if u j 6= ui ,
u j −ui ,
vi j =
(2.53)
′
f (ui ),
if u j = ui .
As in the one-dimensional case, this definition guarantees that shocks, if any, move
at correct speed satisfying the Rankine-Hugoniot condition (see Chapter 3).
In fact, the analytical derivative f′ (ui ) with respect to u can be replaced by zero.
The following equivalence relation holds no matter how vi j is defined for u j = ui
ci j · vi j (u j − ui ) = ci j · (f j − fi ).
To convert (2.52) into the usual form (2.20), the entries of K = {ki j } are defined as
kii = − ∑ ki j ,
j6=i
ki j = −ci j · vi j ,
∀ j 6= i.
(2.54)
Remark 2.7. Since vi j = v ji , the relationship between the off-diagonal entries of K
follows from (2.29). For any pair of internal nodes, c ji = −ci j implies k ji = −ki j .
Likewise, the symmetric boundary part bi j makes a symmetric contribution to K.
Remark 2.8. For a linear convective flux f = vu, the coefficients ki j = −ci j · vi j may
differ from those given by (2.17) with ε = 0. On the one hand, the above-mentioned
preservation of (skew-)symmetry speaks in favor of definition (2.54). On the other
hand, the formula ki j = −ci j · v j is certainly more economical for linear problems.
2.1.6 Conservative Flux Decomposition
In finite difference and finite volume methods for conservation laws, the distribution of mass inside the domain satisfies local balance equations formulated in terms
of numerical fluxes that transport the mass from one node into another. The mass
associated with a given node depends on the sum of incoming and outgoing fluxes.
2.1 Discretization on Unstructured Meshes
55
The advantages of a flux-based formulation are twofold. First, it is guaranteed to be
conservative and reflects the physical nature of convection and diffusion processes.
Second, the rules of the game are remarkably simple: (i) each pair of neighboring
nodes may trade mass, (ii) the mass added to one node is subtracted from another,
(iii) mass can be imported or exported across the boundaries of the domain.
The integral nature of the finite element method makes it difficult to ascertain
how the mass exchange between individual nodes takes place. Due to global conservation, a decomposition of the residual into numerical fluxes is possible. However,
their definition is not as obvious as in the case of finite differences or finite volumes.
The discretized convective and/or diffusive terms represent the net amount of mass
received by node i from its neighbors. Unfortunately, it is difficult to identify the
contribution of each neighbor, and multiple solutions to this problem may exist.
In the early 1990s, a proper definition of numerical fluxes was found for continuous piecewise-linear Galerkin approximations on triangular and tetrahedral meshes
[17, 18, 22, 269, 300]. This was a major breakthrough in the development of edgebased finite element methods for compressible flows [225, 226, 239, 259]. Most of
such algorithms are fully explicit and rely heavily on certain properties of linear
basis functions. As a consequence, they are not suitable for bilinear approximations
and other finite elements. This lack of generality rules out the use of quadrilateral,
hexahedral, and hybrid meshes which might be desirable for various reasons.
A very useful flux decomposition technique was introduced by Selmin et al.
[300, 301, 302] who developed a unified framework for the implementation of unstructured grid methods based on finite element and finite volume discretizations. It
provides a simple way to generate numerical fluxes and artificial diffusion operators
for Galerkin approximations based on Lagrangian finite elements of various shape
and order. Furthermore, element-based data structures are only needed to assemble
the matrices of coefficients involved in the computation of numerical fluxes. The
transition to a flux-based representation leads to efficient algorithms that operate
with node pairs, as in the case of finite difference and finite volume schemes.
Following Selmin et al. [300, 301, 302] and Löhner [226], we consider a finite element approximation to equation (2.1) and define the corresponding fluxes in
terms of the off-diagonal coefficients ci j and di j associated with the discretization
of first-order and second-order derivatives, respectively. The flux decomposition to
be presented can serve as a vehicle for extending classical high-resolution schemes
to unstructured meshes. Many examples of such generalizations can be found in
[239, 243], where the edge-based data structure of Peraire et al. [269] was used to
perform upwinding and flux limiting on a triangular mesh of linear finite elements.
2.1.6.1 Inviscid Fluxes
For any linear or nonlinear flux function f(u), the group finite element discretization
of the conservation law (2.1) yields an equation of the form (2.50). The task is
to express the right-hand side of this semi-discrete equation in terms of internodal
fluxes that take the mass from one node and give it to another node without changing
56
2 Finite Element Approximations
the global balance. The zero row sum property (2.26) leads to the representation
∑ ci j · f j = fi · ∑ ci j + ∑ ci j · f j = ∑ ci j · (fi + f j ).
j
j
j
(2.55)
j
Furthermore, the coefficient ci j can be decomposed into the symmetric internal part
ai j and the skew-symmetric boundary part bi j given by formula (2.29)
ci j =
ai j + bi j
,
2
ai j = ci j − c ji ,
bi j = ci j + c ji .
(2.56)
The sum of terms proportional to ai j admits the following flux decomposition
∑ ai j ·
j
fi + f j
= ∑ fi j ,
2
j6=i
fi j = ai j ·
fi + f j
,
2
∀ j 6= i.
(2.57)
The average of f j and fi indicates that this formula is a finite element counterpart of
the central difference approximation. Since a ji = −ai j , the flux f ji = − fi j has the
same magnitude and opposite sign. This property guarantees mass conservation.
According to (2.29) and (2.48), the contribution of bi j to (2.55) is represented by
∑ bi j ·
j
fi + f j
= fii ,
2
fii =
Z
Γ
ϕi
fi + fh
· n ds,
2
∀i.
(2.58)
Here and below, the index j = i is reserved for boundary fluxes. On the Neumann
boundary part ΓN , integration should be performed using h = f · n in place of fh · n.
On the Dirichlet boundary part ΓD , the i−th equation is overwritten by the constraint
ui = gi which should be used to calculate fi = f(gi ) in (2.57) and elsewhere.
Remark 2.9. The boundary flux fii can also be evaluated using inexact nodal quadrature or approximation fh ≈ fi which has the same effect as mass lumping [301]
bi = ∑ bi j = ni si ,
fii ≈ bi · fi ,
(2.59)
j
where the unit outward normal vector ni and the surface area si are defined by (2.31).
In the worst-case scenario, approximation (2.59) generates an error of order h. On
the other hand, the unique definition of the normal at boundary nodes makes the
numerical treatment of boundary conditions remarkably simple and natural [301].
Summarizing the results, the sum of internal and boundary fluxes is given by
∑ ci j · f j = ∑ fi j .
j
(2.60)
j
The same flux decomposition is obtained if integration by parts is performed to shift
the derivatives onto the test function. Indeed, property (2.29) confirms that
∑ bi j · f j − ∑ c ji · f j = ∑ fi j .
j
j
j
(2.61)
2.1 Discretization on Unstructured Meshes
57
Remark 2.10. Another possibility to define the flux fi j is to take [200, 201, 203]
fi j = ai j ·
fi + f j
f j − fi
− bi j ·
= ci j · fi − c ji · f j ,
2
2
∀ j 6= i.
(2.62)
The difference between the fluxes (2.57) and (2.62) for j 6= i implies that
fii = ∑ bi j ·
j
f j − fi
fi + f j
+ ∑ bi j ·
= ∑ bi j · f j ,
2
2
j
j
∀i.
The so-defined boundary flux fii represents the first term in the left-hand side of
(2.61) and approximates the surface integral that results from integration by parts
∑ bi j · f j =
j
Z
Γ
ϕi fh · n ds.
In the interior of the domain, the two definitions of fi j are equivalent since bi j = 0.
Remark 2.11. Flux decompositions (2.60) and (2.61) are valid not only for linear and
bilinear basis functions but also for higher-order (quadratic, cubic) finite elements.
However, the straightforward two-point flux approximation may cease to reflect the
physical nature of transport processes. As the polynomial order increases, so does
the number of nonzero matrix entries and possible flux decompositions. Ideally,
the definition of numerical fluxes for higher-order FEM should (i) involve solution
values at more than two nodes and (ii) allow mass exchange only between nearest
neighbors, as in the case of finite difference and finite volume schemes [193].
2.1.6.2 Viscous Fluxes
Several alternative approaches to the numerical treatment of diffusive fluxes such as
g = −ε ∇u
have been explored by finite element practitioners [226, 239, 300, 301]. The first
possibility is to determine the nodal values gi ≈ −(ε ∇u)i using the L2 -projection
u j + ui
u j + ui
ε
ε
+ bi j
gi = − ∑ ci j u j = − ∑ ai j
mi j
mi j6=i
2
2
(cf. Section 2.1.4) or another gradient recovery technique. The corresponding fluxes
gi j can be derived in the same way as their inviscid counterparts, that is,
gi + g j
gii ≈ bi · gi ,
gi j = ai j ·
,
∀ j 6= i.
(2.63)
2
The use of gradient recovery makes such a formulation akin to finite difference and
finite volume schemes. The combined treatment of convective and diffusive terms
58
2 Finite Element Approximations
leads to conceptually simple and computationally efficient algorithms. A potential
drawback to this approach is that the evaluation of gi j involves information from
two layers of points [239, 301]. The consistent Galerkin approximation achieves superior accuracy with a compact stencil, i.e., using data from nearest neighbors only.
The result is a sparse matrix D which contains the coefficients of diffusive fluxes.
A symmetric matrix D = {di j } is called a discrete diffusion operator if it has zero
row and column sums [205, 200]. That is, its entries should satisfy the relations
∑ di j = ∑ di j = 0,
j
di j = d ji ,
∀i, j.
i
(2.64)
For example, consider the standard Galerkin discretization of the diffusive term
−
Z
Ω
w∇ · (ε ∇u) dx =
Z
∇w · (ε ∇u) dx −
Ω
Z
ΓN
wn · (ε ∇u) ds.
If the diffusion coefficient ε is constant, this weak statement leads to the formula
Z
Ω
∇ϕi · (ε ∇uh ) dx = ε ∑ di j u j ,
j
di j =
Z
Ω
∇ϕi · ∇ϕ j dx.
Due to (2.24) and (2.25) the resultant matrix is a typical representative of discrete
diffusion operators. The contribution of the surface integral vanishes if homogeneous Neumann boundary conditions are imposed. In general, it can be evaluated
using the known value of n · (ε ∇u) or the compact approximation [301]
gii = −
Z
ΓN
ϕi n · (ε ∇u)h ds ≈ gi · ni ,
where ni denotes the outward normal vector defined in the same way as in (2.59).
Remark 2.12. Interestingly enough, the difference between the consistent mass matrix MC = {mi j } and its lumped counterpart ML = diag{mi } is also a discrete diffusion operator in the above sense. By definition, mi j = m ji and mi = ∑ j mi j so that
∑(mi j − mi ) = ∑(mi j − mi ) = 0.
j
i
The diffusive nature of MC − ML is well known [134] and has been exploited to
design various stabilization terms for finite element schemes [26, 87, 232, 301, 302].
Any discrete diffusion operator D defines the conservative flux decomposition
(Du)i = ∑ di j u j = ∑ gi j ,
j
j6=i
gi j = di j (u j − ui ).
(2.65)
For any pair of nodes i and j, the skew-symmetric fluxes gi j and g ji = −gi j are proportional to the difference between the two nodal values, which results in smoothing
or steepening of solution profiles depending on the sign of the coefficient di j . As we
2.1 Discretization on Unstructured Meshes
59
will see later, these properties make it possible to adjust the magnitude of the flux
gi j so as to enforce monotonicity or remove excessive numerical diffusion.
2.1.7 Relationship to Finite Volumes
To illustrate the relationship between finite element and finite volume discretizations of a conservation law like (2.1), consider an arbitrary control volume Vi with
S
boundary Si = j Si j which consists of several patches Si j . For all j 6= i, the interface between the control volumes Vi and V j is denoted by Si j = Si ∩ S j . The notation
Sii = Si ∩ Γ refers to the union of boundary patches that belong to Γ , if any.
The local mass balance for an arbitrary control volume Vi can be written as
∂
∂t
Z
u dx +
Vi
Z
f · n ds =
Si
Z
s dx.
(2.66)
Vi
The volume of the integration region Vi and the mean value of u are given by
Z
mi =
Vi
dx = |Vi |,
ui =
1
mi
Z
u dx.
Vi
Let ri denote the right-hand side of (2.66). Splitting the surface integral, one obtains
mi
dui
+
dt ∑
j
Z
Si j
f · n ds = ri .
In finite volume methods, numerical integration and interpolation are employed to
approximate the integral over each portion Si j of the control surface Si by
fi j ≈
Z
Si j
f · n ds.
A second-order approximation of central difference type is based on the definition
fii = bi · fi ,
fi j = ai j ·
f j + fi
,
2
∀ j 6= i.
The so-defined numerical fluxes fi j are of the same form as those for the group finite
element formulation but the metric vectors ni and ai j are defined as
bi =
Z
n ds,
Sii
ai j =
Z
n ds,
Si j
∀ j 6= i.
(2.67)
It is easy to verify that these quantities possess the following properties [301]
∑ ai j + bi = 0,
j6=i
a ji = −ai j ,
∀ j 6= i.
(2.68)
60
2 Finite Element Approximations
The first property reflects the fact that a constant flux cannot change the value of ui
at an interior node. It can be readily inferred from (2.67) and the integral relation
Z
n ds =
Z
Vi
Si
∇1 dx = 0,
Si =
[
Si j .
j
The second property in (2.68) is a consequence of the fact that the ‘outward’ normal
for Vi is the negative of that for V j . The skew-symmetry of the coefficient vectors ai j
and a ji ensures that f ji = − fi j , as required by local and global conservation.
Remark 2.13. So far no restrictions have been imposed on the shape of Vi . The degrees of freedom ui are typically associated with cells or vertices of the underlying
mesh. In the latter case, a dual mesh of control volumes covers the whole domain.
Remark 2.14. The original and/or dual mesh are only needed to identify the connections between nodes and generate coefficients that depend on the geometric properties of control volumes. These coefficients are assigned to nodes or node pairs which
participate in the mass exchange. A typical implementation involves visiting each
pair of nodes and sending the fluxes fi j to the corresponding equations [300].
On an unstructured triangular mesh, the control volume Vi can be built around
each node i as sketched in Fig. 2.2. First, each triangle is subdivided into six subelements delimited by the medians whose intersection lies at the center of gravity. Then
Vi is defined as the union of subelements having the point xi as a vertex. This yields
a median dual mesh which represents a subdivision of the computational domain
into nonoverlapping polygonal cells. The area of each cell Vi equals the diagonal
entry mi of the lumped mass matrix ML = diag{mi } for a piecewise linear Galerkin
discretization on a triangulation with the same vertices [8, 300]. Moreover, it turns
out that the corresponding metric quantities ai j and ni are identical [226, 301]. The
same relationship between linear finite elements and vertex-centered finite volume
approximations holds for tetrahedral meshes in the three-dimensional case [301].
Vi
Vj
xj
xi
Fig. 2.2 Control volume Vi for a vertex-centered FVM on a triangular mesh.
2.1 Discretization on Unstructured Meshes
61
The bridge between the finite element and finite volume approaches makes it
possible to combine their advantages and develop a unified framework for the discretization of transport equations on arbitrary meshes [301]. The flux for a finite element scheme can be manipulated in the same way as that in finite volume methods.
Conversely, the consistent mass matrix and the Galerkin approximation of diffusive
terms become portable to a finite volume code. Furthermore, the metric coefficients
ai j , bi , and mi can be calculated using element-by-element matrix assembly or the
alternative geometric approach. In the case of bilinear finite elements, there is no
equivalent finite volume representation since it is impossible to construct a dual
mesh with the same connectivity pattern. However, a conservative flux decomposition and the use of data structures that operate with node pairs are still feasible.
2.1.8 Edge-Based Data Structures
One of the basic tasks involved in the practical implementation of an unstructured
grid method is the assembly of sparse matrices and/or matrix-vector products that
constitute the algebraic system to be solved. A lion’s share of computer time is spent
on retrieving information from large arrays and locating array elements which need
to be updated. The search process depends on the links between individual elements,
nodes, fluxes, and other data items. Hence, the overall performance of the code is
strongly influenced by the choice of data structures and storage techniques.
In most finite element codes, global matrices like MC , C, and D are assembled by
visiting all elements sequentially and collecting their contributions to the integrals
that define the coefficients of discrete operators. The three basic operations are:
gathering nodal data pertaining to each element, processing this data, and scattering
the results back to nodes ([226], p. 187). To perform numerical integration, one
needs to know the numbers and locations of nodes that belong to a given element.
Thus, a typical element-based data structure includes arrays in which the Cartesian
coordinates of each node and the connectivity list of each element are stored.
After the matrix assembly, all information about the geometrical and topological
properties of the underlying mesh can be discarded. Furthermore, the evaluation
of right-hand sides and matrix-vector products can be performed using an edgebased data structure which operates with pairs of nodes, as in the case of finite
volume schemes. Such data structures offer algorithmic simplicity, low memory
requirements, and a major reduction in indirect addressing [22, 229, 225, 269, 301].
Moreover, they are amenable to large-scale parallel computing [56, 226, 245].
Throughout this book, the term ‘edge’ refers to a pair of nodes associated with
→
−
a pair of nonzero off-diagonal matrix entries, i.e., with an edge i j of the graph
that represents the sparsity pattern (stencil) of the numerical scheme. In the finite
element context, any pair of basis functions with overlapping supports defines such
an edge. On a simplex mesh, the number of internodal links for linear finite elements
equals the number of physical mesh edges. Bilinear approximations on rectangular
62
2 Finite Element Approximations
Fig. 2.3 Internodal links for linear and bilinear finite elements.
or mixed meshes give rise to extra diagonal links, see Fig. 2.3. Hence, there is no
one-to-one correspondence between the graph of the matrix and that of the mesh.
To get started with edge-based data structures, consider a sparse matrix A = {ai j }
with neq rows/columns and nnz entries that may assume nonzero values. If storage
is allocated for ai j and/or a ji with j > i, then the node pair {i, j} defines an edge in
the above algebraic sense. The list of edges is stored in a two-dimensional connectivity array kedge(1 : 2, 1 : nedge) with two rows and nedge columns such that the
two node numbers associated with the edge number iedge = 1, . . . , nedge are
i = kedge(1, iedge),
j = kedge(2, iedge).
Remark 2.15. The columnwise storage of edge data is adopted since matrices are
stored by columns in M ATLAB, F ORTRAN, and other programming environments.
Another array aedge(1 : 2, 1 : nedge) is used to store the off-diagonal matrix entries
ai j = aedge(1, iedge),
a ji = aedge(2, iedge).
The first row of aedge is associated with the upper triangular part ( j > i), while
the second row contains the nonzero entries of the lower triangular part ( j < i).
For symmetric matrices, a one-dimensional array aedge is enough. A matrix with
zero row/column sums is completely defined by the contents of kedge and aedge.
Otherwise, an extra array adiag(1 : neq) is used to store the diagonal entries
aii = adiag(i).
This one-dimensional array is sufficient to represent a diagonal matrix such as ML .
2.1 Discretization on Unstructured Meshes
63
In summary, an elementary edge-based data structure includes the edge connectivity list and arrays that store the diagonal and/or off-diagonal part of a matrix with
up to nnz = neq + 2 ∗ nedge nonzero entries. In what follows, this representation of
sparse matrices will be referred to as the Compressed Edge Storage (CES) format.
The following examples illustrate some possibilities to evaluate the matrix-vector
product Au and add the result to the residual or right-hand side vector b. Here and
below, pseudo-code segments are written in the F ORTRAN 90 terminology which we
believe is self-explanatory and readily portable to other programming languages.
Example 2.1. Given a nonsymmetric matrix A = {adiag, aedge, kedge} with nonvanishing row sums, its diagonal entries are processed in a loop over nodes
do i = 1, neq
b(i) = b(i) + adiag(i) ∗ u(i)
end do
while the contribution of off-diagonal entries is included in a loop over edges
do iedge = 1, nedge
i = kedge(1, iedge)
j = kedge(2, iedge)
b(i) = b(i) + aedge(1, iedge) ∗ u(j)
b(j) = b(j) + aedge(2, iedge) ∗ u(i)
end do
Example 2.2. If the matrix A has zero row sums, then a loop over edges is enough
do iedge = 1, nedge
i = kedge(1, iedge)
j = kedge(2, iedge)
diff = u(j) − u(i)
b(i) = b(i) + aedge(1, iedge) ∗ diff
b(j) = b(j) − aedge(2, iedge) ∗ diff
end do
Example 2.3. If the matrix A is symmetric with zero row sums, then it represents a
discrete diffusion operator. Hence, a conservative flux decomposition is feasible
do iedge = 1, nedge
i = kedge(1, iedge)
j = kedge(2, iedge)
flux = aedge(iedge) ∗ (u(j) − u(i))
b(i) = b(i) + flux
b(j) = b(j) − flux
end do
Other kinds of internodal fluxes can be inserted into the vector b in the same way.
64
2 Finite Element Approximations
The edge-based formulation offers an efficient way to perform sparse matrixvector multiplications in explicit algorithms and iterative solvers. In the latter case,
preconditioners should also be applied edge-by-edge if the CES format is adopted.
The design of such preconditioners is addressed in [57, 58]. Reportedly, they are
more efficient than their element-by-element counterparts. A further reduction in
indirect addressing can be achieved using advanced concepts like stars, superedges,
and chains which represent extensions of the single-edge data structure [56, 225].
2.1.9 Compressed Row Storage
A complete transition to an edge-based data structure might be impractical in an existing finite element code that employs a different format for storage of sparse matrices in assembly routines and/or linear solvers. However, the edge-by-edge evaluation of internodal fluxes is nontrivial if the coefficients ci j and di j are not available
in the CES format. Moreover, the entries of global matrices may depend on edge
data, such as the mean velocity vi j in definition (2.54). Therefore, some parts of the
code call for an edge-based implementation and it is essential to have fast access to
all matrix entries associated with a given edge. Alternatively, element matrices can
be disassembled into edge contributions and combined to form edge matrices [56].
The Compressed Row Storage (CRS) and Compressed Column Storage (CCS)
formats are often used in unstructured mesh codes. In either case, all nonzero entries
of a matrix A are packed into a one-dimensional array aval(nnz). This array is
filled row-by-row or column-by-column in the CRS and CCS versions, respectively.
In other words, the CRS form of a matrix A is equivalent to the CCS form of AT .
Hence, the decision as to which format is better suited for a given application is
a matter of whether the matrix itself or its transposed is needed more frequently.
Without loss of generality, we assume the former and discuss the CRS format as
implemented in the open-source finite element software library F EAT 2D [40].
Given the array aval(nnz) of nonzero entries stored row-by-row, two auxiliary
integer arrays, kcol(1 : nnz) and kptr(1 : neq + 1), make it possible to find an
entry with given row and column numbers. The value of j = kcol(ij) is the column number of ai j = aval(ij) for any index from the range ij = 1, . . . , nnz. The
beginning of the i-th row is indicated by ii = kptr(i). That is, if i = 1, . . . , neq is
the row number of ai j = aval(ij), then kptr(i) ≤ ij ≤ kptr(i + 1) − 1. The last
element kptr(neq + 1) = nnz + 1 of the row pointer array kptr corresponds to the
position where the next row would begin. This entry is introduced for programming
convenience in order to avoid treating the last row differently from the others.
For our purposes, it is worthwhile to store the diagonal entry of each row first
[40] so that kcol(ii) = i if ii = kptr(i). This convention provides fast access
to the diagonal part. The remaining column indices are stored in ascending order
so that kptr(i) + 1 ≤ ij < ik ≤ kptr(i + 1) − 1 implies kcol(ij) < kcol(ik).
While this restriction is not a part of the standard CRS format, it plays a pivotal role
in the development of the edge-based assembly algorithms to be presented below.
2.1 Discretization on Unstructured Meshes
65
Example 2.4. Consider the following square matrix with nnz = 12 nonzero entries


1 2 0 7
 2 4 3 0 

A=
 0 3 6 5 .
7 0 5 8
In the sorted CSR format, this 4 × 4 matrix is represented by the three arrays
aval = (1, 2, 7, 4, 2, 3, 6, 3, 5, 8, 7, 5),
kcol = (1, 2, 4, 2, 1, 3, 3, 2, 4, 4, 1, 3),
kptr = (1, 4, 7, 10, 13).
(2.69)
Note that the underlined diagonal entries of A are stored in positions indicated by
kptr and followed by other elements of the same row traversed from left to right.
Example 2.5. Given a matrix A = {aval, kcol, kptr} in the CSR format, the following algorithm returns the sum of the matrix-vector product Au and a vector b
do i = 1, neq
do ij = kptr(i), kptr(i + 1) − 1
j = kcol(ij)
b(i) = b(i) + aval(ij) ∗ u(j)
end do
end do
(2.70)
Example 2.6. If A has zero row sums, an alternative way to evaluate and add Au is
do i = 1, neq
do ij = kptr(i) + 1, kptr(i + 1) − 1
j = kcol(ij)
b(i) = b(i) + aval(ij) ∗ (u(j) − u(i))
end do
end do
(2.71)
The edge-by-edge assembly/modification of the matrix A = {aval, kcol, kptr}
or its conversion into the CES format A = {adiag, aedge, kentry} involves visiting
all node pairs {i, j} with j > i and searching for elements of aval that correspond to
matrix entries with row/column indices i and j. The positions of the diagonal entries
aii = aval(ii) and a j j = aval(jj) are given by ii = kptr(i) and jj = kptr(j),
respectively. Let all node pairs {i, j} be processed in the same order in which they
are encountered in (2.70) and (2.71). Then the off-diagonal entry ai j = aval(ij)
is readily available but the location of a ji = aval(ji) is more difficult to find. In
principle, it can be determined by jumping to the beginning of row j and checking
if kcol(ji) = i for ji = kptr(j) + 1, . . . , kptr(j + 1) − 1. However, the fact that
the elements of kcol are sorted makes it possible to get the value of ji for free.
66
2 Finite Element Approximations
Let ksep(1 : neq) be an integer array that will serve as a pointer to the position
of a ji for the current node pair. The edge-by-edge processing of a matrix and its
storage in the CES format can be performed by the following algorithm [254]
iedge = 0
ksep(1 : neq) = kptr(1 : neq)
do i = 1, neq
ii = kptr(i); adiag(i) = aval(ii)
do ij = ksep(i) + 1, kptr(i + 1) − 1
j = kcol(ij); ksep(j) = ksep(j) + 1
ji = ksep(j); jj = kptr(j)
...
iedge = iedge + 1
kedge(1, iedge) = i
kedge(2, iedge) = j
aedge(1, iedge) = aval(ij)
aedge(2, iedge) = aval(ji)
...
end do
end do
(2.72)
Initially, the separator array ksep points to the beginning of each row, where the
diagonal entry is stored. For i = 1, the next row element (if any) belongs to the upper
triangular matrix since there is no j < 1. All rows are visited sequentially in the
outer loop. First, the value of aii = aval(ii) is retrieved and packed into adiag(i).
Then the lists of edges and off-diagonal entries with row or column index i are
extruded in the inner loop. Due to the above sorting convention, the column indices
j = kcol(ij) increase with ij. When a new entry ai j = aval(ij) with j > i is
encountered, the edge counter iedge and the entry ksep(j) are incremented by 1.
The new value of ji = ksep(j) marks the position of a ji = aval(ji). When row j
is reached in the outer loop, all rows with i < j have already been processed, so that
ksep(j) + 1 is the position of the first row element from the upper triangular part.
Remark 2.16. As an exercise that illustrates how ksep evolves, it is instructive to
consider the sample matrix (2.69) and unroll the loops using paper and pencil.
Remark 2.17. Algorithm (2.72) differs from the original version proposed by the
author and cited in [254] in that it operates with the upper triangular rather than
lower triangular part. This eliminates the need for an extra auxiliary array.
The body of the inner loop may include further statements that involve manipulations with matrix entries indexed by i and j. In particular, the edge-by-edge assembly of aval for a discrete transport operator of the form (2.54) and/or the evaluation
of numerical fluxes like (2.57) and (2.65) can be performed in this way. Since the
values ai j and a ji are not stored contiguously in aval, the permanent ‘jumping’ may
incur a high overhead cost due to slow memory access. Therefore, it is advisable to
use the CES format for storage and edge-by-edge processing of sparse matrices.
2.2 Stabilization of Convective Terms
67
2.2 Stabilization of Convective Terms
The standard Galerkin method produces accurate solutions to elliptic and parabolic
transport equations as long as the Peclet number is relatively small (some notable
exceptions to this rule are considered in Section 4.5). However, the presence of
convective terms deprives the Galerkin FEM of the best approximation property
which it is known to possess in the case of self-adjoint (symmetric) operators.
At high Peclet numbers, the discrete transport operator K is dominated by the
nonsymmetric convective part and exhibits very unfavorable properties. Since the
Galerkin discretization of convective terms is akin to a central difference approximation, it tends to produce spurious oscillations, also known as wiggles. Moreover,
an iterative algorithm or an explicit time integration scheme may become unstable.
The lack of robustness can be rectified by adding some artificial diffusion or
using modified test functions to construct an upwind-biased finite element scheme.
In either case, the Galerkin operator K is replaced by its stabilized counterpart K̄
which may or may not be associated with a continuous bilinear form. The difference
between the two matrices represents a stabilization operator D = {di j }. That is,
K̄ = K + D,
k̄i j = ki j + di j .
(2.73)
Of course, it is essential to ensure that the modified scheme remains consistent and
conservative. Furthermore, the application of D should not make it too diffusive.
In this section, we review some traditional stabilization tools developed in an attempt to suppress the wiggles or, at least, to keep them small (bounded). The presentation of linear stabilized FEM is followed by a brief introduction to residual-based
shock-capturing techniques for problems with discontinuities and steep fronts. Last
but not least, a prototype of nonoscillatory finite element schemes to be presented
in Chapter 6 is constructed building on the concept of modulated dissipation.
2.2.1 First-Order Upwinding
In one space dimension, the first-order upwind difference scheme (UDS) constitutes a nonoscillatory, albeit inaccurate, alternative to the second-order central difference (CDS) and Galerkin finite element methods (GFEM) which are plagued by
numerical instabilities at mesh Peclet numbers greater than 2. The 1D examples in
Chapter 2 illustrate the relationship between these basic discretization techniques.
In an unstructured mesh environment, the design of UDS-like finite element
methods is typically based on a vertex-centered finite volume approach to the discretization of convective terms [8, 11, 167, 288, 322]. Instead, the inviscid flux
(2.57) associated with the standard Galerkin approximation can be replaced by [300]
fi j = ai j ·
fi + f j |ai j · vi j |
−
(u j − ui ),
2
2
∀ j 6= i,
(2.74)
68
2 Finite Element Approximations
where vi j is the averaged velocity defined by (2.53). Substitution into (2.74) yields
ai j · fi , if ai j · vi j > 0,
∀ j 6= i.
fi j =
ai j · f j , if ai j · vi j < 0,
The so-defined multidimensional generalization of UDS proves nonoscillatory for
arbitrary Peclet numbers [300]. Due to the equivalence between linear finite elements and vertex-centered finite volume CDS, it leads to the same set of discrete
equations as the geometric approach based on the construction of a dual mesh.
It is worth mentioning that the practical implementation of first-order upwinding
in a finite element code does not require a reformulation of the discrete problem in
terms of numerical fluxes. In fact, the replacement of (2.57) by (2.74) is equivalent
to adding a discrete diffusion operator D = {di j } with coefficients given by
dii = − ∑ di j ,
di j =
j6=i
|ai j · vi j |
,
2
∀ j 6= i.
Note that the matrix D is symmetric with zero row and column sums, as required by
(2.64). A general approach to the design of such artificial diffusion operators and to
‘discrete upwinding’ for FEM on unstructured meshes is presented in Chapter 6.
2.2.2 Artificial Diffusion
First-order upwinding is rarely used in the realm of finite elements because it results
in strong smearing and does not fit into the usual variational framework. Instead, the
Galerkin transport operator K is typically stabilized by manipulating the bilinear
form a(·, ·) that defines ki j = −a(ϕi , ϕ j ) for any pair of nodes i and j. Consider
ā(w, u) := a(w, u) + b(w, u),
(2.75)
where b(u, v) represents a linear or nonlinear stabilization term. Ideally, this part
should vanish if u is the exact solution of the continuous problem. For practical
purposes, it is sufficient to make sure that b(u, v) → 0 as the mesh h goes to zero.
The matrix K̄ with coefficients k̄i j = −ā(ϕi , ϕ j ) assumes the form (2.73), where
di j = −b(ϕi , ϕ j ).
In a classical artificial diffusion method, the stabilization operator is defined by
b(w, u) = ∑
k
Z
Ωk
∇w · (D∇u) dx,
(2.76)
where D denotes a tensor diffusivity. Typically, the amount of artificial diffusion
depends on the local mesh size h and on the magnitude of the velocity vector v.
2.2 Stabilization of Convective Terms
69
The simplest way to offset the intrinsic negative diffusion of the Galerkin scheme
is to apply isotropic balancing dissipation of the form D = δ I , where
δ =α
|v|h
,
2
(2.77)
is a scalar diffusion coefficient and I is the identity tensor. The free parameter α
determines the magnitude of δ and that of the additional discretization error.
Using a constant value α ∈ [0, 1] is feasible but the resulting scheme is at most
first-order accurate, no matter how high the degree of basis functions {ϕi } is. This
dramatic loss of accuracy is clearly undesirable. Therefore, the parameter α is typically defined as a monotonically increasing function of the mesh Peclet number
Pe h =
|v|h
ε
and evaluated separately for each element Ωk . The local mesh size h may refer, e.g.,
to the longest edge or to the diameter of the area/volume equivalent circle (sphere).
As a rule of thumb, the solution is well-resolved and no stabilization is required
for Pe h ≤ 2. A larger value of Pe h indicates that the flow is too fast or the mesh is too
coarse. Hence, it is necessary to refine the mesh or apply some artificial diffusion.
Example 2.7. Consider the steady one-dimensional convection-diffusion equation
v
d2 u
du
+ε 2 = 0
dx
dx
(2.78)
discretized by linear finite elements on a uniform mesh of size h = ∆ x. The artificial
diffusion method with α ≡ 1 corresponds to the first-order upwind approximation of
the convective term. On the other hand, the solution to equation (2.78) with constant
coefficients is nodally exact if α is defined by [86]
Pe h
2
α = coth
−
.
(2.79)
2
Pe h
To reduce the computational cost associated with repeated evaluation of α , it is
common practice to replace (2.79) by the ‘doubly asymptotic’ approximation
Pe h
α = min 1,
.
(2.80)
6
Neither (2.79) nor (2.80) is guaranteed to be a perfect choice in the case of multidimensional transport equations, variable coefficients, and unstructured meshes.
Still, these definitions of the parameter α are frequently used by default in FEM
codes [86]. Many other formulas have been proposed but the optimal value of α is
highly problem-dependent and difficult to determine from a priori considerations.
In advanced artificial diffusion methods, the definition of α may depend not only
on the local Peclet number Pe h but also on the derivatives of the velocity or pressure
70
2 Finite Element Approximations
fields, on the residuals of the continuous problem, and on other quantities that measure the smoothness of a given solution [170, 172, 226]. The design of such schemes
can be based on empirical principles or backed by solid mathematical theory. Both
approaches have been employed in CFD computations with considerable success.
2.2.3 Streamline Upwinding
In multidimensional problems, both convection and diffusion transport information
in certain directions. The velocity field v determines the direction and speed of convective transport, whereas the net diffusive flux depends on the definition of D. In
the above example, the stabilization term b(w, u) consists of element contributions
associated with a weak form of δ ∆ u. The results are frequently polluted by numerical crosswind diffusion that could be removed without making the scheme unstable.
This has led the developers of stabilized FEM to introduce anisotropic balancing
dissipation that acts along the streamlines of v but not transversely [157, 181].
A family of widespread streamline upwind (SU) finite element methods is based
on (2.76) with an anisotropic tensor diffusivity of the form [47, 86, 176, 322]
D = τ v ⊗ v = {Dξ η },
(2.81)
where τ is the so-called intrinsic time. The componentwise form of (2.81) reads
Dξ η = τ vξ vη ,
∀ξ , η ∈ {x, y, z}.
After some rearrangements, the corresponding stabilization term (2.76) becomes
b(w, u) = ∑
k
Z
Ωk
τ (v · ∇w)(v · ∇u) dx.
(2.82)
Since the integrand represents a weak form of the second convective derivative
(v · ∇)2 u = v · ∇(v · ∇u),
the term b(w, u) incorporates streamline diffusion (SD) into the Galerkin scheme.
The stabilization parameter τ for the streamline upwind method is defined as
τ=
δ
αh
=
,
2
|v|
2|v|
where δ is given by (2.77) and depends on the choice of α . The default is (2.79) or
(2.80), while α ≡ 1 corresponds to first-order upwinding along the streamlines.
The lack of crosswind diffusion results in a smaller deviation from the original Galerkin scheme than the use of D = δ I . However, the SD stabilization fails
to remove undershoots and overshoots in the vicinity of steep layers [172]. Some
remedies to this deficiency of streamline upwinding are discussed in Section 2.2.6.
2.2 Stabilization of Convective Terms
71
2.2.4 Petrov-Galerkin Methods
Instead of modifying the Galerkin bilinear form a(·, ·), streamline diffusion can be
introduced within the framework of a consistent Petrov-Galerkin method. Let
L u = v · ∇u − ∇ · (ε ∇u) + σ u = s
in Ω
(2.83)
be the conservative (σ = ∇ · v) or nonconservative (σ = 0) form of a stationary
convection-diffusion equation. Using the usual notation for the L2 scalar product
(w, u) =
Z
Ω
wu dx,
the Galerkin weak form of the above model problem can be written as follows
a(w, u) = (w, s).
For a pure Dirichlet problem, w = 0 on Γ and the bilinear form a(·, ·) reads
a(w, u) = ∑
k
Z
Ωk
(wv · ∇u + ∇w · (ε ∇u) + wσ u) dx.
(2.84)
The inclusion of a SU stabilization term (2.82) transforms this integral form into
ā(w, u) = a(w, u) + ∑
k
=∑
k
Z
Ωk
Z
Ωk
τ (v · ∇w)(v · ∇u) dx
(w̄v · ∇u + ∇w · (ε ∇u) + wσ u) dx,
where the convective term is multiplied by the nonconforming test function
w̄ = w + τ v · ∇w.
(2.85)
In the streamline upwind / Petrov-Galerkin (SUPG) method [47], this test function is applied to all components of L u and the bilinear form is redefined as
ā(w, u) = a(u, w̄).
(2.86)
Unlike classical SU methods, this approach to stabilization of convective terms ensures that the residual of the associated weak form vanishes if u is the exact solution
of (2.83). This desirable property is called strong consistency and can be exploited
to maintain optimal accuracy in a given finite-dimensional space ([276], p. 269).
The advent of SUPG was followed by the development of other Petrov-Galerkin
methods that differ in the definition of the operator P for the stabilization term
b(w, u) = ∑
k
Z
Ωk
τ (Pw)(L u − s) dx.
(2.87)
72
2 Finite Element Approximations
In unsteady problems, the transport operator L includes the contribution of a (discretized) time derivative. The original SUPG formulation corresponds to
Pw = v · ∇w.
The resulting stabilization term b(w, u) is nonsymmetric, unless L = P. The symmetry can be restored by using the Galerkin / least squares (GLS) method [159]
Pw = L w.
Reportedly, a better weighting of the reactive term σ u is offered by the formula
Pw = −L ∗ w,
where L ∗ stands for the adjoint of L . In this basic subgrid scale (SGS) method
[86] and in its predecessors [92, 113] the sign of the symmetric terms is reversed.
Remark 2.18. All of the above Petrov-Galerkin methods reduce to the classical
SUPG stabilization if σ = 0 and the test function w is linear inside each element.
Many other definitions of the operator P are possible. In most cases, the differences
between the resulting solutions are marginal. For a comprehensive review and a
detailed comparative study of stabilized FEM based on various generalizations or
modifications of SUPG, we refer to [70, 86, 172, 276] and references therein.
Remark 2.19. The use of edge-based data structures in the context of SUPG is addressed in [56]. It is shown that element matrices and residuals associated with the
stabilized bilinear form can be decomposed into edge contributions. Hence, global
matrix assembly and the computation of matrix-vector products can be performed in
a loop over edges rather than elements. This approach to implementation of SUPG
offers significant savings in terms of the CPU time and memory requirements [56].
2.2.5 Taylor-Galerkin Methods
A major drawback of SU methods and their Petrov-Galerkin counterparts is the
uncertainty regarding the choice of the stabilization parameter. The default setting
given by formula (2.79) is designed to produce a nodally exact solution to (2.78) but
might be inappropriate for more involved convection-diffusion-reaction equations.
Guessing the right value of α for time-dependent problems is particularly difficult
since both amplitude and phase errors must be taken into account. Fortunately, there
is a conceptually simple and parameter-free alternative to streamline upwinding.
An accurate and stable finite element approximation to the transport equation
∂u
+L u = s
∂t
in Ω
(2.88)
2.2 Stabilization of Convective Terms
73
can be constructed within the framework of Taylor-Galerkin (TG) methods [85, 86,
90, 275]. The main advantages of these finite element schemes are their inherent
stability, improved phase accuracy, and the absence of free parameters.
2.2.5.1 Second-Order TG Approximation
In Taylor-Galerkin methods, the time discretization plays a pivotal role. Instead of
manipulating the Galerkin space discretization, the time-stepping method is chosen
so as to stabilize it in natural way, perhaps, under a certain time step restriction. TG
algorithms do not contain free parameters and are directly applicable to multidimensional transport problems. To give a simple example that illustrates their relationship
to SU/SD methods, consider the second-order accurate approximation [70, 74]
un+1 − un
=
∆t
∂u
∂t
n
∆t
+
2
∂ 2u
∂ t2
n+θ
(2.89)
which leads to the explicit (θ = 0) or semi-implicit (θ = 1) Lax-Wendroff / TG2
method. The stability limit (if any) for the time step ∆ t depends on the structure of
the spatial differential operator L and on the number of space dimensions [89].
Invoking the governing equation (2.88) and assuming that all of its coefficients
are independent of t, the involved time derivatives are expressed as follows
∂u
= s − L u,
∂t
∂ 2u
∂u
= −L
= L (L u − s)
∂ t2
∂t
(2.90)
and plugged into equation (2.89). Next, the residual of the semi-discrete scheme
un+1 − un
∆t
+ L un + L (s − L un+θ ) = s
∆t
2
(2.91)
is multiplied by the test function w and integrated over the domain Ω . This yields
un+1 − un
w,
(2.92)
+ a(un , w) + b(un+θ , w) = (w, s),
∆t
where the contributions of the time derivatives (2.90) are represented by the terms
a(w, u) = (w, L u − s),
b(w, u) =
∆t
(w, L (s − L u)).
2
The latter represents an explicit or implicit correction to the forward Euler / Galerkin
discretization which makes it more stable and second-order accurate in time.
After integration by parts, the following representation of b(w, u) is obtained
b(w, u) =
∆t
∆t
(L ∗ w, s − L u) =
2
2
Z
Ω
(L ∗ w)(s − L u) dx
(2.93)
74
2 Finite Element Approximations
for a pure Dirichlet problem such that w = 0 on the boundary Γ . This term is identical to (2.87) with τ = ∆ t/2 and P = −L ∗ . Like many other stable FEM approximations, (2.92) turns out to be a disguised Petrov-Galerkin method [70, 290].
Remark 2.20. The equivalence between Taylor-Galerkin and SUPG-like discretizations can be exploited to define the stabilization parameter τ for transient computations, or the local time step ∆ t for marching the solution to a steady state [38, 74].
2.2.5.2 Linear Hyperbolic Equations
A standard model problem is the linear hyperbolic equation given by (2.88) with
L u = v · ∇u.
Substitution of the streamline diffusion operator L 2 for
∆t
b(w, u) =
2
Z
Ω
∇ · (vw)(v · ∇u) dx −
Z
Γ
∂2
∂ t2
in (2.89) leads to
w v · n (v · ∇u) ds .
In incompressible flow problems, the velocity field is divergence-free (∇ · v = 0) and
the volume integral reduces to the SU stabilization term (2.82) with τ = ∆ t/2
Z
Z
∆t
(v · ∇w)(v · ∇u) dx − w v · n (v · ∇u) ds .
b(w, u) =
2
Ω
Γ
When a Dirichlet boundary condition is imposed at the inlet Γ− = {x ∈ Γ | v · n < 0},
as required by the nature of hyperbolic problems, the surface integral reduces to that
over Γ+ = {x ∈ Γ | v · n > 0}. Such boundary terms are missing in the original SU
formulation (2.82) but have been incorporated into the GLS method [95]. The failure
to include them might cause spurious reflections at outflow boundaries [38, 89].
2.2.5.3 Nonlinear Hyperbolic Equations
Next, consider a generic conservation law of the form (2.1) which corresponds to
L u = ∇ · f(u)
with a nonlinear inviscid flux f(u). Invoking approximation (2.89), one obtains
un+1 = un + ∆ t(s − ∇ · fn ) +
(∆ t)2
∇ · ftn+θ ,
2
(2.94)
where fn = f(un ) and ftn is the corresponding time derivative. By the chain rule
df
df ∂ u
.
(2.95)
= −a∇ · f,
a=
ft =
du ∂ t
du
2.2 Stabilization of Convective Terms
75
Again, the Galerkin weak form of the semi-discrete scheme (2.94)–(2.95) admits
representation (2.92). Integration by parts in the bilinear form a(·, ·) yields
Z
a(w, u) =
Γ
w f(u) · n ds −
Z
Ω
∇w · f(u) dx.
(2.96)
By virtue of (2.95), the bilinear form b(·, ·) for the second-order term is given by
Z
Z
∆t
b(w, u) =
(a · ∇w)∇ · f(u) dx − w (a · n)∇ · f(u) ds .
(2.97)
2
Ω
Γ
In an implicit TG algorithm, the term b(un+θ , w) may be linearized using [226]
fn+θ ≈ fn + θ an (un+1 − un ).
For θ = 0, the contribution of a(un , w) and b(un , w) to (2.92) can be written as [87]
ā(un , w) =
Z
Γ
w fn+1/2 · n ds −
Z
Ω
∇w · fn+1/2 dx,
(2.98)
where the flux fn+1/2 represents a second-order accurate approximation to f(un+1/2 )
fn+1/2 = fn +
∆t n
∆t
f = fn − an ∇ · fn .
2 t
2
(2.99)
Note that there is no need to differentiate the (possibly discontinuous) flux function
in (2.96) and (2.98). However, the evaluation of (2.97) and (2.99) might be cumbersome and expensive, especially in the case of nonlinear hyperbolic systems in which
the scalar characteristic speed a is replaced by a Jacobian matrix [230]. This is why
a fractional-step approach to the treatment of such problems is frequently adopted.
2.2.5.4 Two-Step Implementation
An alternative implementation [6, 231] of the explicit second-order Taylor-Galerkin
method for (2.1) is based on the two-step Runge-Kutta time-stepping scheme
∆t ∂ u n
∆t
= un + (s − ∇ · fn ),
(2.100)
un+1/2 = un +
2 ∂t
2
n+1/2
∂u
= un + ∆ t(s − ∇ · fn+1/2 ).
un+1 = un + ∆ t
(2.101)
∂t
In contrast to (2.99), the flux fn+1/2 is evaluated using the intermediate solution
fn+1/2 = f(un+1/2 ).
(2.102)
76
2 Finite Element Approximations
The finite difference discretization of (2.100)–(2.101) is known as the Richtmyer
scheme. In the 1980s, a finite element version of this popular predictor-corrector
algorithm was developed [6, 7, 231] and combined with the flux-corrected transport
(FCT) methodology to enforce monotonicity on unstructured meshes [232, 233].
The weak form of the midpoint rule corrector (2.101) can be written as follows
un+1 − un
+ ā(un , w) = (w, s),
(2.103)
w,
∆t
where ā(·, ·) is the bilinear form defined by (2.98). The only difference as compared
to the one-step TG2 method (2.92) is that the intermediate flux fn+1/2 is defined
by (2.102) rather than (2.99). Donea and Huerta ([86], p. 158) perform matrix assembly for the discrete counterpart of (2.103) using the strong form of (2.100) to
calculate the divergence of fn+1/2 at the numerical integration points. Furthermore,
they illustrate the influence of various flux representations by numerical examples.
Alternatively, the Richtmyer-TG scheme can be implemented using a usual finite
element discretization for the first step. The corresponding weak form reads
!
un+1/2 − un
w,
+ a(un , w) = (w, s),
(2.104)
∆ t/2
where a(·, ·) is the Galerkin bilinear form (2.96). The second step is given by
un+1 − un
w,
+ a(un+1/2 , w) = (w, s).
(2.105)
∆t
The approximate solutions un+1/2 and un+1 may belong to different finite dimensional spaces. The use of equal-order interpolations widens the stencil of the discrete scheme and is not suitable for steady-state computations since the standard
Galerkin approximation a(w, u) = (w, s) is recovered for u = un = un+1/2 = un+1 .
Therefore, it is common practice to use piecewise-constant basis/test functions for
un+1/2 and (multi-)linear elements for un+1 [87, 231]. This strategy eliminates the
need for extra stabilization at steady state [226]. For the linear convection equation
in 1D, it leads to the same algebraic system as the TG scheme based on (2.99).
The reader is referred to Löhner et al. [226, 231, 232] for a detailed presentation
of the two-step Taylor-Galerkin method (2.104)–(2.105) including practical implementation details and examples that demonstrate the advantages of this approach.
2.2.5.5 Edge-Based Formulation
In an edge-based finite element code, the Richtmyer-TG scheme can be formulated
in terms of numerical fluxes. The discrete counterpart of (2.102) is [226, 228]
n+1/2
fi j
n+1/2
= ai j · f(ui j
),
(2.106)
2.2 Stabilization of Convective Terms
77
n+1/2
where ai j denotes the vector of weights given by (2.56) and ui j
n+1/2
ui j
=
uni + unj
2
∆t
+
2
∂u
∂t
n
is defined by
.
ij
The value of the time derivative at the edge midpoint can be approximated, e.g., by
si + s j (∇ · f)i + (∇ · f) j
∂u
−
,
=
∂t ij
2
2
where (∇ · f)i and (∇ · f) j are obtained using (2.42) with f in place of g. Note that
the resulting flux (2.106) depends on the solution values at more than two nodes.
As a cheap alternative, the following simplification can be envisaged [226, 228]
si + s j (xi − x j ) · (fi − f j )
∂u
=
−
.
(2.107)
∂t ij
2
|xi − x j |2
This definition involves a straightforward extension of the 1D approximation
f j − fi
∂f
,
j = i + 1.
≈
∂x ij
∆x
While the use of formula (2.107) in two or three dimensions seems to lack a theoretical justification, it makes the edge-based TG algorithm very efficient [226].
2.2.6 Discontinuity Capturing
Linear stability of a numerical scheme is insufficient to guarantee that discontinuities and fronts are resolved in a nonoscillatory fashion. All finite element methods
that rely on streamline diffusion (SD) stabilization of convective terms are known
to produce undershoots and overshoots in regions where the solution gradients are
steep and not aligned with the flow direction. In some situations, imperfections are
small in magnitude and can be tolerated. In other cases, it is essential to ensure that
the numerical solution remains nonnegative and/or devoid of spurious oscillations.
Nonoscillatory finite element approximations can be constructed by increasing
the amount of numerical dissipation in regions where linear stabilization is insufficient. This process is commonly referred to as discontinuity-capturing or shockcapturing [69, 161, 290] even if the solution is differentiable but exhibits abrupt
changes across thin layers. Since the location of layers is generally unknown, it must
be determined using suitable smoothness sensors such as gradients or residuals.
Discontinuity capturing (DC) involves the construction of a nonlinear artificial
diffusion operator and adding its contribution to the stabilization term
b̄(w, u) = b(w, u) + c(w, u).
78
2 Finite Element Approximations
The extra dissipation c(w, u) should be designed so as to control the solution gradient
∇u not only along the streamlines but in all relevant directions. Consider [86]
c(w, u) = ∑
k
Z
Ωk
τ̂ (v̂ · ∇w)R(u) dx,
(2.108)
where τ̂ is another free parameter, v̂ is a solution-dependent vector function, and
R(u) = L u − s
is the residual of the governing equation. Obviously, an additional term of the form
(2.108) does not destroy the strong consistency of a Petrov-Galerkin formulation.
The definition of v̂ proposed by Hughes and Mallet [161] for |∇u| =
6 0 reads
v · ∇u
v̂ =
∇u,
(2.109)
|∇u|2
which corresponds to a projection of the velocity v onto the gradient of u. The
resultant vector field v̂ is called the effective transport velocity [290] since
v̂ · ∇u = v · ∇u.
(2.110)
Of course, there is no need for any discontinuity capturing if the solution u is constant. Therefore, v̂ = 0 is the natural setting for the degenerate case |∇u| = 0.
The combination of (2.108) and (2.109) gives rise to isotropic artificial dissipation, as shown by the following representation of the nonlinear DC term
c(w, u) = ∑
k
Z
Ωk
ν (u)∇w · ∇u dx,
(2.111)
where the coefficient ν (u) depends on the gradient and residual of u as follows
( v·∇u
if |∇u| =
6 0,
τ̂ |∇u|
2 R(u),
ν (u) =
0,
if |∇u| = 0.
An obvious drawback to this kind of artificial viscosity is the fact that it becomes
negative and may destabilize the finite element discretization if (v · ∇u)R(u) < 0.
To ensure that the sign of ν (u) is correct, the corresponding vector v̂ can be
redefined using the residual instead of the convective derivative [117, 136, 177]
R(u)
v̂ =
∇u.
|∇u|2
Substitution into (2.108) yields a dissipative term of the form (2.108), where
( R(u) 2
, if |∇u| =
6 0,
τ̂
|∇u|
ν (u) =
0,
if |∇u| = 0.
2.2 Stabilization of Convective Terms
79
As before, the nonlinearity of ν (u) stems from the definition of v̂ in terms of ∇u.
In the context of SUPG methods, a reasonable value of τ̂ is even more difficult
to find than that of the linear stabilization parameter τ . A typical setting is [160]
τ̂ = max{0, τ (v̂) − τ (v)},
where the dependence of τ (v) on v is the same as in the linear term b(w, u). Thus,
no extra numerical diffusion is added in the case when v̂ = v is parallel to ∇u.
Alternatively, an anisotropic DC operator can be designed to act in the crosswind
direction only. The removal of streamline diffusion from (2.111) leads to [69]
Z
(v · ∇w)(v · ∇u)
c(w, u) = ∑
dx.
ν (u) ∇w · ∇u −
|v|2
k Ωk
This nonlinear crosswind diffusion term is equivalent to (2.76) with D given by
v⊗v
D = ν (u) I −
,
|v|2
where I denotes the unit tensor and it is tacitly assumed that |v| =
6 0. The use of
anisotropic artificial diffusion makes it possible to avoid excessive damping.
De Sampaio and Coutinho [290] combine the streamline diffusion and discontinuity capturing operators within the framework of Petrov-Galerkin and TaylorGalerkin methods for unsteady convection-diffusion equations. To this end, they
blend the flow velocity v and effective transport velocity v̂ as follows
v̄ = γ v + (1 − γ )v̂,
0 ≤ γ ≤ 1.
To ensure that v̂ is well defined, the setting γ = 1 is used whenever |∇u| = 0. The
dissipative term b̄(w, u) differs from its linear counterpart b(w, u) in that the convective derivatives are evaluated using the above combined velocity v̄ rather than v.
This replacement is consistent since v̄ · ∇u = v · ∇u at the continuum level [290].
Many alternative definitions of b̄(w, u) and of the involved parameters can be
found in the literature. We refer to John et al. [172, 174] for a comprehensive review
of the state of the art and a detailed comparative study of available techniques.
2.2.7 Interior Penalty Methods
A promising new approach to linear stabilization and discontinuity capturing involves the addition of interior penalty terms that control the jumps of the solution
gradient. This far-reaching idea was originally proposed by Douglas and Dupont
[91] and recently revived by Burman and Hansbo [49, 50] who called it edge stabilization and put it on a firm theoretical basis. In recent years, the finite element community has come to recognize and explore the potential of such schemes [293, 327].
80
2 Finite Element Approximations
Edge stabilization methods penalize the jumps of the normal derivative via [50]
b(w, u) = ∑
Z
k6=l Γkl
γ h2 [n · ∇w] [n · ∇u] ds,
(2.112)
where γ is a free parameter and [q] denotes the jump of a given quantity q across the
common boundary Γkl = Ω̄k ∪ Ω̄l of two adjacent mesh elements Ωk and Ωl .
In 3D problems, the interface Γkl is not an edge but a face of the computational
mesh. Hence, continuous interior penalty (CIP) is a more appropriate name for stabilization via jump terms of the form (2.112). It is also worth mentioning that the
(geometric) definition of an edge differs from that given earlier in this section.
In contrast to Petrov-Galerkin and Taylor-Galerkin methods, the symmetric CIP
term (2.112) is independent of the underlying equation and vanishes if [n · ∇u] = 0.
The amount of stabilization is determined solely by the smoothness of the numerical
solution and the results are typically rather insensitive to the choice of γ . The price
to be paid for these benefits is a denser matrix with a nonstandard sparsity pattern.
The discontinuity capturing term for a CIP method can be defined as follows [52]
c(w, u) = ∑ δ (u)
k6=l
∑
Z
E∈Γkl E
(e · ∇w)(e · ∇u)
de,
|e · ∇u|
(2.113)
where e is a vector parallel to an edge E of Γkl . In two space dimensions, the second
sum consists of a single term that represents a line integral over Γkl = E.
Burman and Ern [52] employ the following definition of the smoothness sensor
δ (u) = γ̂ h
Z
Γkl
|[n · ∇u]| ds
(2.114)
and prove that no spurious maxima/minima are generated if γ̂ is sufficiently large.
Remark 2.21. The edge/face integrals that appear in (2.112)–(2.114) are commonly
evaluated by the midpoint rule which is exact in the case of linear finite elements.
2.2.8 Modulated Dissipation
In the late 1980s, the diffusive nature of row-sum mass lumping for (multi-)linear
finite elements was exploited to construct explicit Taylor-Galerkin schemes that introduce modulated dissipation in the vicinity of discontinuities and steep fronts.
In these methods, the added mass diffusion admits a conservative flux decomposition which facilitates an extension of flux-corrected transport (FCT) algorithms
and total variation diminishing (TVD) schemes to finite elements [87, 298, 299].
As explained in Chapter 4, the basic idea behind such high-resolution schemes is
to suppress dispersive ripples using a combination of (linear) first-order numerical
diffusion and (nonlinear) balancing antidiffusion. The latter must be controlled by a
suitable sensor/limiter capable of detecting discontinuities and steep gradients.
2.2 Stabilization of Convective Terms
81
2.2.8.1 High-Order Scheme
Consider a time-dependent conservation law (2.1) and its weak form associated with
a linearly stable (Taylor-Galerkin or Petrov-Galerkin) finite element scheme
un+1 − un
+ ā(w, un ) = (w, s).
w,
∆t
Inserting the finite-dimensional (linear or multilinear) counterparts of u and w, one
ends up with a linear system that relates the vectors of new and old nodal values
MC un+1 = MC un + ∆ trn ,
(2.115)
where MC is the consistent mass matrix and r is the vector of nodal increments
mi j = (ϕi , ϕ j ),
ri = ∑ k̄i j u j ,
k̄i j = ā(ϕi , ϕ j ).
j
The replacement of MC by its lumped counterpart ML , as defined in (2.36), is feasible but degrades the phase accuracy of FEM for transient problems. Note that the use
of MC in the left-hand side of (2.115) results in an implicit coupling of the degrees
of freedom, although the underlying time discretization is fully explicit. However,
the symmetry and diagonal dominance of MC make it possible to solve (2.115) efficiently by the following iterative algorithm proposed by Donea et al. [85]
ML u(m+1) = MC un + ∆ trn + (ML − MC )u(m) ,
m = 0, . . . , L − 1.
(2.116)
The first and last iterate are given by u(0) = un and un+1 = u(L) , respectively. The
derivation of this preconditioned Richardson’s scheme is based on an approximate
factorization of the consistent mass matrix [85, 275]. In the context of explicit
Taylor-Galerkin schemes, the three-pass solver (L = 3) was found to be optimal.
2.2.8.2 Low-Order Scheme
Next, consider (2.116) with L = 1 and u(0) = 0. The resulting explicit approximation
ML un+1 = MC un + ∆ trn
(2.117)
corresponds to the original finite element scheme (2.115) with mass lumping in the
left-hand side only. For our purposes, it is worthwhile to write (2.117) in the form
ML un+1 = ML un + ∆ trn + (MC − ML )un .
(2.118)
As already mentioned in Section 2.1.6.2, the matrix MC − ML represents a discrete
diffusion operator. Its contribution renders the solution to problem (2.118) nonoscillatory (for sufficiently small time steps) but results in a dramatic loss of accuracy.
82
2 Finite Element Approximations
Remark 2.22. For 1D hyperbolic problems, the low-order counterpart (2.118) of the
TG2 scheme corresponds to the Lax-Wendroff method with added dissipation.
p It is
stable and monotone [140, 298] for Courant numbers ν in the range |ν | ≤ 2/3.
2.2.8.3 Selective Mass Lumping
As explained in Section 2.1.6.2, the explicit mass diffusion built into the right-hand
side of (2.118) can be expressed as a conservative sum of internodal fluxes
(MC u − ML u)i = ∑ mi j u j − mi ui = ∑ mi j (u j − ui ).
(2.119)
j6=i
j
The amount of numerical diffusion can be reduced using modulation coefficients
αi j ∈ [0, 1] to replace MC − ML by another symmetric matrix D = {di j } such that
dii = − ∑ di j ,
di j = (1 − αi j )mi j ,
j6=i
∀ j 6= i.
The setting αi j ≡ 0 corresponds to D = MC − ML , while the use of 0 < αi j ≤ 1 leads
to a less diffusive approximation. The matrix form of the resulting scheme reads
ML un+1 = ML un + ∆ trn + Dun .
(2.120)
The last term vanishes for αi j ≡ 1, which corresponds to standard mass lumping
ML un+1 = ML un + ∆ trn .
(2.121)
Varying the modulation coefficients αi j between 0 and 1, it is possible to switch between (2.117) and (2.121) in a conservative fashion. This strategy can be interpreted
as selective mass lumping in the right-hand side of the low-order scheme (2.117).
2.2.8.4 Two-Stage Implementation
If the problem at hand is truly time-dependent, it is advantageous to split a finite
element scheme with modulated dissipation into two stages, so as to separate convective transport from added mass diffusion as follows [298, 299, 87]
MC uH = MC un + ∆ trn ,
n+1
ML u
H
= (ML + D)u .
(2.122)
(2.123)
The first stage corresponds to (2.115) and inherits its superb phase accuracy. The
superscript H stands for “high-order.” At the second stage, the amplitudes are corrected by adding a certain amount of mass diffusion in regions where steep gradients
are detected by the smoothness sensor built into the definition of the coefficients αi j .
2.2 Stabilization of Convective Terms
83
The high-order solution un+1 = uH is obtained for D = 0. The associated loworder scheme corresponds to D = MC − ML and un+1 = ML−1 MC uH . Importantly,
neither first-order nor modulated dissipation introduces any phase errors. That is,
the predictor uH is smeared but not displaced. This is what makes (2.122)–(2.123)
more attractive than the one-step implementation (2.120) of selective mass lumping.
2.2.8.5 Modulation Coefficients
In essence, the use of modulated mass diffusion is a discontinuity capturing technique that operates at the fully discrete level. The overall complexity and performance of the resulting scheme depend on the philosophy behind the computation
of the modulation coefficients αi j . First or second derivatives of flow variables can
be used to construct empirical smoothness sensors but the presence of a free parameter undermines the practical utility of such schemes. Alternatively, the values
of αi j can be determined using Zalesak’s fully multidimensional FCT algorithm
[226, 232, 299] or symmetric TVD limiters [87, 299]. Below we outline the former
approach since it is easier to implement and more accurate for unsteady problems.
Explicit FCT algorithms add limited antidiffusion to a nonoscillatory low-order
solution uL . In the present context, this “transported and diffused” solution is computed from (2.120) or (2.123) with D = MC − ML . It follows that un+1 is given by
mi uin+1 = mi uLi + ∑ αi j mi j (ui − u j ).
(2.124)
j6=i
The last term consists of limited antidiffusive fluxes which are evaluated using the
solution u = un for (2.120) and u = uH for (2.123). The correction factors αi j are
chosen so that uin+1 is bounded by the local extrema umax
and umin
defined by
i
i
umax
= max uLj ,
i
j∈Si
umin
= min uLj ,
i
j∈Si
where Si = { j | mi j 6= 0} is the stencil of node i. The definition of αi j for (2.124) is
based on the following algorithm [226, 355] which is known as Zalesak’s limiter
1. Compute the sums of positive/negative raw antidiffusive fluxes fi j = mi j (ui − u j )
Pi+ = ∑ max{0, fi j },
j6=i
Pi− = ∑ min{0, fi j }.
j6=i
2. Define the upper/lower bounds such that no spurious maxima/minima can emerge
max
Q+
− uLi ),
i = mi (ui
min
Q−
− uLi ).
i = mi (ui
3. Evaluate the correction factors αi j and α ji for each pair of antidiffusive fluxes
R±
i
Q±
i
= min 1, ± ,
Pi
αi j =
−
min{R+
i , R j }, if f i j ≥ 0,
− +
min{Ri , R j }, if fi j < 0.
84
2 Finite Element Approximations
This parameter-free definition of αi j gurantees that un+1 given by (2.124) satisfies
umin
≤ uin+1 ≤ umax
,
i
i
∀i.
We refrain from going into detail at this point because an in-depth presentation of
finite element FCT schemes based on Zalesak’s limiter will follow in Section 4.4.
2.3 Discontinuous Galerkin Methods
Discontinuous Galerkin (DG) methods [66, 67, 108, 147] represent one of the most
promising current trends in computational fluid dynamics. The frequently mentioned advantages of this approach include local conservation and the ease of constructing high-order approximations on unstructured meshes. Moreover, DG methods are well suited for hp-adaptivity and parallelization.
One of the major bottlenecks in the design of high-order DG methods for
convection-dominated transport problems is the lack of reliable mechanisms that
ensure nonlinear stability and effectively suppress spurious oscillations. A number
of successful discontinuity capturing and slope limiting techniques are available for
DG finite element methods [37, 48, 68, 154, 186, 188, 320] and their finite difference/volume counterparts [19, 251, 350, 342]. However, no universally applicable methodology has been developed to date. Since the accuracy of monotonicitypreserving schemes degenerates to first order at local extrema, free parameters or
heuristic indicators are frequently employed to distinguish between troubled cells
and regions where the solution varies smoothly. In some cases, the results leave a
lot to be desired. Also, the use of limiters may cause severe convergence problems
in steady state computations [342].
In this section, we present a parameter-free, non-clipping slope limiter [195] for
high-resolution DG-FEM on arbitrary meshes. A hierarchical approach to adaptive
p-coarsening is pursued. The Taylor series form [237, 251, 350] of a polynomial
shape function is considered, and the involved derivatives are limited so as to control the variations of lower-order terms. The corresponding upper and lower bounds
are defined using the data from elements sharing a vertex. This strategy yields a remarkable gain of accuracy, as compared to traditional compact limiters that search
the von Neumann (common face) neighbors of a given element [19, 68, 188].
2.3.1 Upwind DG Formulation
A simple model problem that will serve as a vehicle for our presentation of slopelimited DG approximations is the linear convection equation
∂u
+ ∇ · (vu) = 0
∂t
in Ω ,
(2.125)
2.3 Discontinuous Galerkin Methods
85
where u(x,t) is a scalar quantity transported by a continuous velocity field v(x,t).
Let n denote the unit outward normal to the boundary Γ of the domain Ω . The initial
and boundary conditions are given by
u|t=0 = u0 ,
u|Γin = g,
Γin = {x ∈ Γ | v · n < 0}.
Multiplying (2.125) by a sufficiently smooth test function w, integrating over Ω ,
and using Green’s formula, one obtains the following weak formulation
Z Z
∂u
w − ∇w · vu ∆ x + wuv · n ds = 0,
∀w.
(2.126)
∂t
Ω
Γ
In the discontinuous Galerkin method, the domain Ω is decomposed into a finite
number of cells Ωe , and a local polynomial basis {ϕ j } is employed to define the
restriction of the approximate solution uh ≈ u to Ωe via
uh (x,t)|Ωe = ∑ u j (t)ϕ j (x),
j
∀x ∈ Ωe .
(2.127)
The globally defined uh is piecewise-polynomial and may have jumps at interelement boundaries. The meaning of the coefficients u j depends on the choice of the
basis functions. A local version of (2.126) can be formulated as
Z Z
∂ uh
− ∇wh · vuh ∆ x + wh ûh v · n ds = 0,
wh
∀wh ,
(2.128)
∂t
Ωe
Γe
where wh is an arbitrary test function from the DG space spanned by ϕi . Since uh is
multiply defined on Γe , the surface integral is calculated using the solution value ûh
from the upwind side of the interface, that is,

lim uh (x + δ n,t),


 δ →+0
g(x,t),
ûh (x,t)|Γe =


 lim uh (x − δ n,t),
δ →+0
v · n < 0, x ∈ Ω̄ \Γin ,
v · n < 0, x ∈ Γin ,
(2.129)
v · n ≥ 0, x ∈ Ω̄ .
In the case of a piecewise-constant approximation, the result is equivalent to the
first-order accurate upwind finite volume scheme. The DG formulation for general
conservation laws and systems thereof is described, e.g., in [67, 68].
2.3.2 Taylor Basis Functions
In a discontinuous Galerkin method of degree p ≥ 0, the shape function uh |Ωe is
given by (2.127), where the number of basis functions depends on p. Clearly, many
alternative representations are possible, and some choices are better than others. For
accuracy and efficiency reasons, it is worthwhile to consider an orthogonal basis
86
2 Finite Element Approximations
such that M is a diagonal matrix and its inversion is trivial. For example, tensor
products of Legendre polynomials are commonly employed on quadrilaterals and
hexahedra [37]. The Gram-Schmidt orthonormalization procedure [108, 365], Dubiner’s basis functions [48, 147], and Bernstein-Bézier [102] polynomials are suitable for the construction of hierarchical approximations on triangular meshes. In
general, one set of basis functions may be used for matrix assembly and another for
limiting or visualization purposes. Due to the local nature of DG methods, conversion between a pair of alternative bases is straightforward and relatively efficient.
Following Luo et al. [237], we restrict our discussion to quadratic polynomials
uh |Ωe ∈ P2 (Ωe ) and consider the 2D Taylor series expansion
2
2 uh (x, y) = uc + ∂∂ ux (x − xc ) + ∂∂ uy (y − yc ) + ∂∂ xu2 (x−x2 c )
c
c
c
(2.130)
2
2 2 + ∂∂ yu2 (y−y2 c ) + ∂∂x∂uy (x − xc )(y − yc )
c
c
about the centroid (xc , yc ) of a cell Ωe . Introducing the volume averages
1
|Ω e |
ūh =
Z
Ωe
uh ∆ x,
xn ym =
1
|Ω e |
Z
Ωe
xn ym ∆ x,
the quadratic function uh can be expressed in the equivalent form [237, 251, 350]
uh (x, y) = ūh + ∂∂ ux (x − xc ) + ∂∂ uy (y − yc )
c
c
(x−xc )2
(x−xc )2
∂ 2u +
− 2
+ ∂ x2 2
c
+
h
∂ 2u ∂ x∂ y c
∂ 2u ∂ y2 c
(y−yc )2
2
−
(y−yc )2
2
(2.131)
i
(x − xc )(y − yc ) − (x − xc )(y − yc ) .
This representation has led Luo et al. [237] to consider the local Taylor basis
ϕ1 = 1,
ϕ5 =
ϕ2 =
(y−yc )2
2 ∆ y2
−
x−xc
∆x ,
(y−yc )2
,
2 ∆ y2
ϕ3 =
y−yc
∆y ,
ϕ6 =
ϕ4 =
(x−xc )2
2∆ x 2
2
c)
− (x−x
,
2∆ x 2
(2.132)
(x−xc )(y−yc )−(x−xc )(y−yc )
.
∆ x∆ y
The scaling by ∆ x = (xmax −xmin )/2 and ∆ y = (ymax −ymin )/2 is required to obtain a
well-conditioned system [237]. The normalized degrees of freedom are proportional
to the cell mean value ūh and derivatives of uh at (xc , yc )
2 uh (x, y) = ūh ϕ1 + ∂∂ ux ∆ x ϕ2 + ∂∂ uy ∆ y ϕ3 + ∂∂ xu2 ∆ x2 ϕ4
c
c
c
(2.133)
2 2 ∂ u
∂ u 2
+ ∂ y2 ∆ y ϕ5 + ∂ x∂ y ∆ x∆ y ϕ6 .
c
c
Note that the cell averages are decoupled from other degrees of freedom since
2.3 Discontinuous Galerkin Methods
Z
Ωe
ϕ12 ∆ x = |Ωe |,
87
Z
Ωe
ϕ1 ϕ j ∆ x = 0,
2 ≤ j ≤ 6.
On a uniform mesh of rectangular elements, the whole Taylor basis (2.132) is orthogonal, as shown by Cockburn and Shu [68]. On a triangular mesh, this is not
the case even for the linear part {ϕ1 , ϕ2 , ϕ3 } since the L2 inner product of ϕ2 and
ϕ3 is nonvanishing. However, the consistent mass matrix M may be ‘lumped’ by
setting all off-diagonal entries equal to zero. In contrast to the case of a typical
Lagrange basis, this modification is conservative because it does not affect the decoupled equation for the mean value of uh in Ωe .
2.3.3 The Barth-Jespersen Limiter
The above Taylor series representation is amenable to p-adaptation and limiting.
In the context of finite volume and DG finite element methods, a slope limiter is
a postprocessing filter that constrains a polynomial shape function to stay within
certain bounds. Many unstructured grid codes employ the algorithm developed by
Barth and Jespersen [19] for piecewise-linear data. Given a cell average ūh = uc and
the gradient (∇u)c , the goal is to determine the maximum admissible slope for a
constrained reconstruction of the form
uh (x) = uc + αe (∇u)c · (x − xc ),
0 ≤ αe ≤ 1,
x ∈ Ωe .
(2.134)
Barth and Jespersen [19] define the correction factor αe so that the final solution
values at a number of control points xi ∈ Γe are bounded by the maximum and minimum centroid values found in Ωe or in one of its neighbors Ωa having a common
boundary (edge in 2D, face in 3D) with Ωe . That is,
max
umin
e ≤ u(xi ) ≤ ue ,
∀i.
(2.135)
Due to linearity, the solution uh attains its extrema at the vertices xi of the cell Ωe .
To enforce condition (2.135), the correction factor αe is defined as [19]

n max o
−uc

, if ui − uc > 0,
min
1, ueui −u


c
(2.136)
αe = min 1, n
o if ui − uc = 0,
i 
min

 min 1, ue −uc , if ui − uc < 0,
ui −uc
where ui = uc + (∇u)c · (xi − xc ) is the unconstrained solution value at xi .
The above algorithm belongs to the most popular and successful limiting techniques, although its intrinsic non-differentiability tends to cause severe convergence
problems at steady state [251, 342]. Another potential drawback is the elementwise
which implies that
and umin
definition of umax
e
e
88
2 Finite Element Approximations
Ωe
xi
Ωa
Fig. 2.4 Vertices and neighbors of Ωe on a triangular mesh.
• the bounds for u(xi ) satisfying (2.135) at a vertex xi depend on the element number e and may be taken from neighbors that do not contain xi ,
• no constraints are imposed on the difference between the solution values in elements meeting at a vertex but having no common edge/face,
• the results are rather sensitive to the geometric properties of the mesh.
In particular, problems are to be expected if Ωe has sharp angles, as in Fig. 2.4.
2.3.4 The Vertex-Based Limiter
In light of the above, the accuracy of limited reconstructions can be significantly improved if the bounds for variations ui − uc at the vertices of Ωe are constructed using
the maximum and minimum values in the elements containing the vertex xi . The sodefined umax
and umin
may be initialized by a small/large constant and updated in a
i
i
loop over elements Ωe as follows:
umax
:= max{uc , umax
},
i
i
umin
:= min{uc , umin
i
i }.
(2.137)
The elementwise correction factors αe for (2.134) should guarantee that
umin
≤ u(xi ) ≤ umax
,
i
i
∀i.
This vertex-based condition can be enforced in the same way as (2.135)
n max o

u −u

min
1, iui −uc c , if ui − uc > 0,


αe = min 1, n
o if ui − uc = 0,
i 
min

 min 1, ui −uc , if ui − uc < 0.
(2.138)
(2.139)
ui −uc
Obviously, the only difference as compared to the classical Barth-Jespersen (BJ)
max
max
limiter is the use of uimin in place of uemin . This subtle difference turns out to be the
key to achieving high accuracy with p-adaptive DG methods.
2.3 Discontinuous Galerkin Methods
89
In fact, the revised limiting strategy resembles the elementwise version of the
finite element flux-corrected transport (FEM-FCT) algorithm developed by Löhner
et al. [232]. In explicit FCT schemes, umax
and umin
represent the local extrema of
i
i
a low-order solution. In accordance with the local discrete maximum principle for
unsteady problems, data from the previous time level can also be involved in the
estimation of admissible upper/lower bounds.
2.3.5 Limiting Higher-Order Terms
The quality of the limiting procedure is particularly important in the case of a highorder DG method [186]. Poor accuracy and/or lack of robustness restrict the practical utility of many parameter-dependent algorithms and heuristic generalizations of
limiters tailored for piecewise-linear functions.
Following Yang and Wang [350], we multiply all derivatives of order p by a
(p)
common correction factor αe . The limited counterpart of (2.131) becomes
o
n (1) ∂ u ∂u
uh (x, y) = ūh + αe
∂ x (x − xc ) + ∂ y (y − yc )
c
(2)
+ αe
+
c
(x−xc )2
(x−xc )2
∂ 2u +
− 2
2
∂ x2 c
∂ 2u ∂ y2 c
(y−yc )2
2
−
(y−yc )2
2
(2.140)
h
io
(x − xc )(y − yc ) − (x − xc )(y − yc ) .
∂ 2u ∂ x ∂ y c
(1)
(2)
In our method, the values of αe and αe are determined using the vertex-based or
standard BJ limiter, as applied to the linear reconstructions
2 ∂ u ∂ u ∂ 2 u (2)
(2)
α
ux (x, y) =
+
(y
−
y
)
,
(2.141)
(x
−
x
)
+
x
c
c
∂ x c
∂ x 2 c
∂ x ∂ y c
2 ∂ u ∂ u ∂ 2 u (2)
(2)
α
+
(x
−
x
)
+
(y
−
y
)
,
(2.142)
uy (x, y) =
y
c
c
∂ y c
∂ x ∂ y c
∂ y2 c
∂ u ∂ u (1)
(x
−
x
)
+
(y
−
y
)
.
(2.143)
u(1) (x, y) = ūh + αe
c
c
∂ x c
∂ y c
The last step is identical to (2.134). In the first and second step, first-order derivatives
with respect to x and y are treated in the same way as cell averages, while secondorder derivatives represent the gradients to be limited.
Since the mixed second derivative appears in (2.141) and (2.142), the correction
(2)
factor αe for the limited quadratic reconstruction (2.140) is defined as
(2)
(2)
(2)
αe = min{αx , αy }.
The first derivatives are typically smoother and should be limited using
(2.144)
90
2 Finite Element Approximations
(1)
(1)
(2)
αe := max{αe , αe }
(2.145)
to avoid the loss of accuracy at smooth extrema. It is important to implement the
limiter as a hierarchical p-coarsening algorithm, as opposed to making the assumption [68] that no oscillations are present in uh if they are not detected in the linear
part. In general, we begin with the highest-order derivatives (cf. [186, 350]) and
calculate a nondecreasing sequence of correction factors
(p)
αe
(q)
:= max αe ,
p≤q
p ≥ 1.
(2.146)
(q)
As soon as αe = 1 is encountered, no further limiting is required since definition
(p)
(2.146) implies that αe = 1 for all p ≤ q. Remarkably, there is no penalty for
using the maximum correction factor. At least for scalar equations, discontinuities
are resolved in a sharp and nonoscillatory manner.
For a numerical study of the above slope limiter, we refer to [195].
2.4 Summary
In this chapter, we dealt with the design of unstructured grid methods for scalar
transport equations. The first part was concerned with the standard Galerkin finite
element approximation of conserved variables, fluxes, and derived quantities. The
analysis of discrete operators has shown that they possess some interesting and useful properties. The aspects of mass conservation were discussed in some detail. The
semi-discrete scheme was expressed in terms of numerical fluxes, and a relationship
to finite volumes was established. Last but not least, edge-based algorithms and data
structures were introduced as an alternative to the traditional element-by-element
programming strategy. The above concepts and tools lend themselves to numerical
simulation of compressible flows and convection-dominated transport problems.
The second part was devoted to finite element approximations of convective
terms. A survey of representative Petrov-Galerkin and Taylor-Galerkin schemes was
included to introduce the key ideas but the reader may want to consult, e.g., the recent book by Donea and Huerta [86] for further details and numerical examples.
Also, we have presented a hierarchical slope limiter for discontinuous Galerkin
methods. The aspects of flux/slope limiting for continuous finite elements are addressed in Chapter 4, where we pursue an algebraic approach to the design of artificial diffusion operators on the basis of generalized FCT and TVD algorithms.
Chapter 3
Maximum Principles
In this chapter, we elaborate on the qualitative behavior of solutions to multidimensional equations of elliptic, hyperbolic, and parabolic type. We analyze the properties of differential operators and derive a priori bounds that depend on the initial
and/or boundary conditions. Maximum and minimum principles are formulated for
each PDE model. If we include a proof, we try to keep it rigorous but simple. The
obtained estimates lead to a set of algebraic and geometric constraints on the coefficients of the numerical scheme and the shape of mesh elements, respectively.
In particular, we consider a handy generalization of Harten’s TVD theorem to
multidimensional discretizations on unstructured meshes. We show that the nonnegativity of off-diagonal coefficients is sufficient for the space discretization of an unsteady transport equation to be local extremum diminishing (LED) and/or positivitypreserving. Furthermore, we address the implications of the time-stepping method
and the properties of discrete operators. The third basic rule, as postulated in Section 1.6.3.3, is reinforced by criteria based on the concept of monotone matrices.
The material to be covered provides the theoretical background and useful design
criteria for the derivation of algebraic flux correction schemes in the next chapter.
3.1 Properties of Linear Transport Models
The theory of partial differential equations makes it possible to perform a detailed
analysis and validation of the mathematical models we are interested in. A particularly useful and important analytical tool is the maximum principle which also implies positivity preservation. In the absence of zeroth-order terms, solutions to some
elliptic PDEs of second order are known to attain their maxima and minima on the
boundary of the domain. If a positive source is included, the solution cannot assume
a negative value at any interior point if nonnegative boundary data are prescribed. In
unsteady problems of hyperbolic and parabolic type, the initial time level represents
another inflow boundary of the space-time domain. Therefore, the upper and lower
bounds for the exact solution may also be influenced by the choice of initial data.
91
92
3 Maximum Principles
There are several reasons for the importance of maximum principles. On the one
hand, they usually represent certain physical constraints that should apply to a given
mathematical model. On the other hand, useful information about the solutions of
differential solutions becomes available, even if the solutions themselves are unknown. Upper/lower bounds, uniqueness proofs, and comparison principles can be
obtained using elementary calculus. Last but not least, discrete maximum principles
play an important role in the development of numerical methods, so we feel that a
structured and self-contained review of their continuous counterparts is in order.
In this section, we restrict ourselves to a study of transport problems from Section 1.3. For simplicity, we assume that all coefficients are known and do not depend
on the solution. However, the maximum principles to be established are applicable
more generally, and the assumption of linearity may be waived in many cases.
3.1.1 The Laplace Operator
The maximum principle for harmonic functions, i.e., functions that satisfy the
Laplace equation, was known to Gauss already in 1839. A far-reaching generalization is due to Hopf [152] who proved that if a function satisfies a partial differential
inequality of second order and attains a maximum in the interior of the domain,
then this function is constant. Building on this result, strong and weak maximum
principles have been established for PDEs of different types [120, 211, 273].
For simplicity, let us start with the maximum principle for the Laplace operator
∆=
∂2
∂2
+
.
.
.
+
= ∇2
∂ x12
∂ xd2
that appears in the left-hand side of the Poisson equation −∆ u = f to be solved in a
bounded domain Ω ⊂ Rd , where d is the number of space dimensions.
Consider a twice continuously differentiable function u ∈ C2 (Ω ) ∩ C0 (Ω̄ ). If u
has a local maximum at an interior point x ∈ Ω , then the partial derivatives of first
and second order must satisfy the following conditions at this point ([273], p. 51)
∂u
= 0,
∂ xk
∂ 2u
≤ 0,
∂ xk2
∀k = 1, . . . , d.
(3.1)
Obviously, this cannot be the case if the Laplacian of u is strictly positive
∆ u > 0 in Ω .
It turns out that maxima are attained on the boundary even if this inequality is not
strict. This result is known as the weak maximum principle (cf. [120], p. 15).
Theorem 3.1. If the Laplacian of u ∈ C2 (Ω ) ∩C0 (Ω̄ ) satisfies the inequality
∆ u ≥ 0 in Ω ,
(3.2)
3.1 Properties of Linear Transport Models
93
then the maximum of u(x) over all x ∈ Ω̄ is attained on the boundary Γ , that is,
max u = max u.
Ω̄
Γ
(3.3)
Remark 3.1. The corresponding strong maximum principle states that µ = maxΩ̄ u
cannot be attained at any interior point x ∈ Ω unless u ≡ µ is constant.
Different proofs of Theorem 3.1 can be found in the literature [120, 273]. The following one can be readily extended to steady convection-diffusion equations.
Proof. Following Gilbarg and Trudinger ([120], p. 45), we construct a proof by
contradiction. Let µ = maxΓ u be the maximum over x ∈ Γ and consider the function
w = max{0, u − µ }.
(3.4)
By construction, w ≥ 0 in Ω̄ and w = 0 on Γ . The theorem requires that w ≡ 0 in
Ω . Suppose that w(x) > 0 at an interior point x ∈ Ω . Due to continuity, there is a
neighborhood Ω∗ ⊂ Ω such that w = u − µ > 0 in Ω∗ and w = 0 on its boundary Γ∗ .
Since the derivatives of u and w are equal in Ω∗ , (3.2) and (3.4) imply that
w∆ w = w∆ u ≥ 0
in Ω∗ .
(3.5)
Integrating this product over Ω∗ , invoking Green’s formula for integration by parts,
and using the assumption that w is zero on the boundary Γ∗ , we obtain
Z
Ω∗
w∆ w dx = −
Z
Ω∗
|∇w|2 dx.
Obviously, the right-hand side of this relation cannot be positive, while the lefthand side is nonnegative due to (3.5). This can only be the case if w is constant
in Ω̄∗ . However, a constant w cannot satisfy w > 0 in Ω∗ and w = 0 on Γ∗ . This
contradiction proves the weak maximum principle formulated in Theorem 3.1. Theorem 3.2. If the Laplacian of u ∈ C2 (Ω ) ∩C0 (Ω̄ ) satisfies the inequality
∆ u ≤ 0 in Ω ,
then the minimum of u(x) over all x ∈ Ω̄ is attained on the boundary Γ , that is,
min u = min u.
Ω̄
Γ
(3.6)
Proof. This estimate follows from Theorem 3.1 applied to −u. Due to the equivalence of maximum and minimum principles, it is enough to prove the former. Corollary 3.1. Let u ∈ C2 (Ω ) ∩C0 (Ω̄ ) be the solution of the Poisson equation
−∆ u = f
in Ω .
(3.7)
Then maxΩ̄ u = maxΓ u and/or minΩ̄ u = minΓ u if f ≤ 0 and/or f ≥ 0, respectively.
94
3 Maximum Principles
Definition 3.1. A function u ∈ C2 (Ω ) ∩ C0 (Ω̄ ) is called subharmonic if ∆ u ≥ 0 in
Ω , superharmonic if ∆ u ≤ 0 in Ω , and harmonic if ∆ u = 0 in Ω [273].
If the right-hand side of (3.7) is zero, then both estimates are applicable. Hence,
the weak maximum principle for harmonic functions can be formulated as follows.
Corollary 3.2. Let u ∈ C2 (Ω ) ∩C0 (Ω̄ ) be the solution of the Laplace equation
∆ u = 0 in Ω .
Then this harmonic function attains its maxima and minima on the boundary Γ
min u ≤ u(x) ≤ max u,
Γ
Γ
∀x ∈ Ω̄ .
(3.8)
This double inequality gives an a priori estimate of u(x) in Ω in terms of its values
on Γ which are known if boundary conditions of Dirichlet type are prescribed.
Corollary 3.3. Let u be subharmonic and v be harmonic in Ω . If u = v on Γ , then
u≤v
in Ω .
Proof. Consider the function w = u − v which is subharmonic in Ω and vanishes on
the boundary Γ . By the maximum principle, this function is nonpositive in Ω . Similarly, a superharmonic function u provides an upper bound for its harmonic
counterpart v if u = v on Γ . This fact explains the names given in Definition 3.1.
3.1.2 Equations of Elliptic Type
The maximum principle established for the Poisson and Laplace equations can be
extended to many other problems including steady transport equations of the form
Lu=s
in Ω ,
(3.9)
where the divergence of convective and/or diffusive fluxes is represented by
L u = ∇ · (vu − D∇u).
(3.10)
Definition 3.2. A second-order operator L of the form (3.10) is called elliptic at a
point x if the matrix D(x) is symmetric positive definite at this point. It is called
uniformly elliptic in a domain Ω if it is elliptic at each point x ∈ Ω .
Ellipticity has some interesting implications. It is known from linear algebra that
any symmetric positive definite matrix D admits the factorization ([273], p. 59)
D = R−1Λ R = CT C,
(3.11)
3.1 Properties of Linear Transport Models
95
where Λ is the diagonal matrix of positive
eigenvalues, R is the orthogonal matrix
√
(R−1 = RT ) of eigenvectors, and C = Λ R defines the linear transformation
CT x̃ = x.
By the chain rule, the first derivatives of u with respect to x̃ and x are related by
˜ = C∇u,
∇u
(3.12)
and the sum of second derivatives with respect to the coordinates x̃ corresponds to
˜ · (∇u)
˜ = ∇ · (CT C∇u) = ∇ · (D∇u).
∆˜ u = ∇
(3.13)
Therefore, the diffusive term can be expressed in terms of the Laplacian ∆˜ associated with the x̃ coordinates under the above linear transformation. This remarkable
property indicates that the theory developed for the Poisson and Laplace equations
should be applicable to other models based on elliptic PDEs of second order.
If a convective flux is included, its contribution to the transport equation can be
decomposed into the streamline derivative and a ‘reactive’ term as follows
∇ · (vu) = v · ∇u + (∇ · v)u.
(3.14)
The physical meaning of the divergence ∇ · v is the rate at which the volume of a
moving fluid parcel changes as it travels through the flow field (see [4], p. 48). If
∇ · v ≡ 0, then the flow is incompressible and the solution of the transport equation
(3.9) is bounded by its boundary values as in the case of the Laplace operator.
Theorem 3.3. Let the function u ∈ C2 (Ω ) ∩C0 (Ω̄ ) satisfy the differential inequality
L u = ∇ · (vu − D∇u) ≤ 0 in Ω .
If the diffusion tensor D is symmetric positive definite and ∇ · v ≡ 0, then
max u = max u.
Ω̄
Γ
(3.15)
Proof. The proof is similar to that of Theorem 3.1. Let µ = maxΓ u and consider
w = max{0, u − µ }
such that w ≥ 0 in Ω and w = 0 on Γ . Again, the goal is to prove that w ≡ 0 in Ω .
Suppose that there is a subdomain Ω∗ ⊂ Ω such that w = u − µ > 0 in Ω∗ and
w = 0 on its boundary Γ∗ . Using (3.14) and the fact that ∇ · v ≡ 0, we obtain
L w = v · ∇w − ∇ · (D∇w).
Since the partial derivatives of u and w are equal in Ω∗ , this gives the estimate
wL w = wL u ≤ 0.
(3.16)
96
3 Maximum Principles
Next, we integrate the convective term by parts and invoke ∇ · v ≡ 0 again to get
Z
Ω∗
wv · ∇w dx = −
Z
Ω∗
w∇ · (vw) dx = −
Z
Ω∗
wv · ∇w.
Since the left- and right-hand sides are the same up to the sign, this integral is zero.
Therefore, only the diffusive term might make a nonvanishing contribution to
Z
Ω∗
wL w dx =
Z
Ω∗
∇w · (D∇w) dx.
(3.17)
The left-hand side of this relation is nonpositive due to (3.16). Since the diffusion
tensor D was assumed to be symmetric positive definite, ∇w · (D∇w) > 0. Thus, the
right-hand side of (3.17) is strictly positive, which yields a contradiction. In the case of an arbitrary velocity field, convective effects may create internal
maxima/minima in regions where the term (∇ · v)u is nonvanishing. Therefore, we
can only prove a weaker result known as nonnegativity or positivity (preservation).
Theorem 3.4. Let the function u ∈ C2 (Ω ) ∩C0 (Ω̄ ) satisfy the differential inequality
L u = ∇ · (vu − D∇u) ≥ 0 in Ω .
(3.18)
If the diffusion tensor D is symmetric positive definite in Ω and u ≥ 0 on Γ , then
in Ω .
u≥0
Proof. We prove this weak minimum principle using a generalization of the idea
presented in [289] for a simpler convection-diffusion equation in divergence form.
Suppose, contrary to the theorem, that u ≥ 0 on Γ and u(x) < 0 at an interior
point x ∈ Ω . Then there is a subdomain Ω∗ ⊂ Ω such that u < 0 in Ω∗ and u = 0 on
its boundary Γ∗ . By the divergence theorem, we have the integral identity
Z
Ω∗
L u dx = −
Z
Γ∗
n · (D∇u) ds,
(3.19)
where n denotes the unit outward normal to Γ∗ . The contribution of the convective
flux n · (vu) to the surface integral vanishes since u = 0 on Γ∗ . The remainder
n · (D∇u) = ñ · ∇u
(3.20)
is the rate at which u changes as we approach the boundary Γ∗ moving in the direction ñ := n · D. Since we assume that u < 0 in Ω∗ and u = 0 on Γ∗ , this rate of
change is strictly positive if we go in the outward direction, that is, if ñ · n > 0.
Due to the assumption that the matrix D is symmetric positive definite, we have
ñ · n = n · D · n > 0.
3.1 Properties of Linear Transport Models
97
It follows that the directional derivative (3.20) is positive, whereas L u ≥ 0 by assumption (3.18). Therefore, the left- and right-hand sides of equation (3.19) have
nonmatching signs, which leads to a contradiction and concludes the proof. Remark 3.2. Reversing the sign of u in Theorems 3.3 and 3.4, one can prove the
corresponding minimum principle and sign preservation for negative functions, respectively. Following a common convention, we restrict ourselves to the analysis of
maximum principles and nonnegativity (positivity) constraints in what follows.
Remark 3.3. All of the above upper and lower bounds reflect the qualitative properties of differential operators and not of any particular boundary value problem. This
is why no restrictions have been imposed on the choice of boundary conditions.
Elliptic partial differential equations of second order are usually endowed with
boundary conditions of Dirichlet or mixed (Dirichlet-Neumann) type. In the former
case, the boundary value problem for the transport equation (3.9)–(3.10) reads
∇ · (vu − D∇u) = s, in Ω ,
(3.21)
u = g on Γ .
The weak maximum principle established in Theorems 3.3 and 3.4 yields the following a priori estimates of the solution u ∈ C2 (Ω ) ∩C0 (Ω̄ ) in terms of g ∈ C0 (Γ ).
Theorem 3.5. Let the diffusion tensor D be symmetric positive definite and ∇ · v ≡ 0
in Ω . Then a solution of problem (3.21) satisfies the maximum principle
s≤0
⇒
max u = max g.
Γ
Ω̄
Theorem 3.6. Let the diffusion tensor D be symmetric positive definite in Ω . Then
a solution of problem (3.21) with arbitrary v satisfies the positivity constraint
s ≥ 0,
g≥0
⇒
u ≥ 0.
Here and below, inequalities are meant to hold in the whole range of function values.
Corollary 3.4. Let the diffusion tensor D be symmetric positive definite in Ω . Then
there is at most one solution to a linear problem of the form (3.21).
Proof. If we suppose that there exist two different solutions u and v, then the function w = u−v satisfies (3.21) with s = 0 and g = 0. Due to Theorem 3.6, as applied to
w and −w, this implies that 0 ≤ w ≤ 0 in Ω , which can only be the case if w ≡ 0. Remark 3.4. The existence of a unique solution is not guaranteed by this Corollary.
Corollary 3.5. Let the linear operator L be given by (3.10), where D is symmetric
positive definite in Ω . If L u ≥ L v in Ω and u ≥ v on Γ , then
u≥v
in Ω̄ .
98
3 Maximum Principles
Proof. Since the function w = u − v satisfies the Dirichlet problem (3.21) with s ≥ 0
and g ≥ 0, the fact that w ≥ 0 in Ω follows from Theorem 3.6. This relationship between the solutions of the same partial differential equation
with different boundary data is called the comparison principle ([120], p. 263).
Remark 3.5. If ∇ · v ≡ 0, then L u = L (u + c) for any constant c. Hence, if u is the
unique solution of the Dirichlet problem (3.21) with s = 0, then u + c is the unique
solution of the same PDE with the Dirichlet boundary data given by g + c.
Theorems 3.3 and 3.4 are also applicable to (3.9) with boundary conditions of
Dirichlet-Neumann type. However, the corresponding estimates are of little practical value since the solution u is not known on the whole boundary. For further information on maximum principles for elliptic problems we refer to [120, 180, 273].
3.1.3 Equations of Hyperbolic Type
Convection-reaction models are based on first-order PDEs, and maximum principles
are obtained in an entirely different way. Convective transport of information by a
prescribed velocity field v takes place along parametric curves that represent the
characteristics of the hyperbolic equation. In the Lagrangian reference frame which
corresponds to the viewpoint of an observer moving with the flow velocity, the problem reduces to a set of ODEs to be integrated along the characteristics subject to the
prescribed initial and/or boundary conditions. This knowledge makes it possible to
predict how the solution evolves as the fluid moves through the flow field.
In experimental fluid dynamics, it is common practice to visualize the flow motion by releasing and tracking markers, such as small particles or colored dye. The
flow lines traced by these markers are the physical prototypes of characteristics.
Definition 3.3. A streamline is a curve tangent to the velocity vector at each point.
As long as a fluid particle cannot have two different velocities at the same point,
streamlines of an instantaneously defined velocity field v(x) cannot intersect. If the
function v is Lipschitz-continuous, i.e., there is a constant C such that
|v(x1 ) − v(x2 )| ≤ C|x1 − x2 |,
∀x1 , x2 ∈ Ω ,
then there is exactly one streamline through a given point [176]. In steady flow,
markers released at successive time instants are exposed to the same flow field and
follow the same path. Thus, streamlines coincide with the trajectories of tracers.
Definition 3.4. A pathline is the trajectory followed by an individual fluid particle.
If the velocity field is known, a pathline is described by a system of ordinary
differential equations for the Cartesian coordinates of a moving marker. Hence, it is
possible to track the markers and monitor their properties mathematically rather than
experimentally. This is the idea behind the method of characteristics that we will use
to analyze the properties of steady and unsteady convective transport models.
3.1 Properties of Linear Transport Models
99
3.1.3.1 Steady Convective Transport
Steady convection-reaction processes can be described by the hyperbolic equation
∇ · (vu) = s in Ω
(3.22)
which can also be written in the generic form (3.9) with L u = ∇ · (vu) and D ≡ 0.
In the absence of diffusion, boundary conditions can only be prescribed on the
inflow boundary, as required by the one-way nature of convective transport. Let
u = g on Γ− = {x ∈ Γ | v · n < 0}.
(3.23)
No boundary conditions are prescribed on the complementary part Γ0 ∪ Γ+ , where
Γ0 = {x ∈ Γ | v · n = 0},
Γ+ = {x ∈ Γ | v · n > 0}.
Although the velocity field v(x) is assumed to be steady, the fluid is in motion. If a
marker is launched at a point x0 ∈ Γ− and time instant t0 , then it will move along the
streamline/pathline through x0 until it leaves the domain Ω at the outlet Γ+ . Let
x̂(t) = x(t, x0 ,t0 )
(3.24)
denote the instantaneous position x of the marker at time t ≥ t0 . The ‘hat’ notation
will also refer to functions of the position vector x̂(t), such as the velocity
v̂(t) = v(x̂(t)).
Since the marker is moving with the prescribed velocity v̂(t), its pathline is given
by the set of points x̂(t) ∈ Ω which satisfy the following Cauchy problem
dx̂(t)
= v̂(t),
dt
x̂(t0 ) = x0 .
(3.25)
The so-defined parametric curves x̂(t), as depicted in Fig. 3.1, are the streamlines of
the velocity field and the characteristics of the linear hyperbolic equation (3.22).
By the chain rule, the total derivative of the function û(t) = u(x̂(t)) is given by
d
∂ u dx̂k
dû
ˆ
=∑
= v̂ · ∇u.
dt k=1 ∂ xk dt
(3.26)
ˆ stands for ∇ applied at x̂(t). Invoking (3.22) and (3.14), we obtain
The notation ∇
ˆ = ŝ − (∇
ˆ · v)û.
v̂ · ∇u
Therefore, the evolution of û(t) along the characteristic through x0 is governed by
dû
+ r̂û = ŝ,
dt
û(t0 ) = g0 ,
(3.27)
100
3 Maximum Principles
Γ+
Ω
Γ−
Γ+
x̂(t)
x0
Γ−
Fig. 3.1 Characteristics of the steady two-dimensional convection equation.
ˆ · v represents the rate of volumetric compressibility, and the value of
where r̂ = ∇
g0 = g(x0 )
is available from the Dirichlet boundary conditions (3.23) prescribed at the inlet Γ− .
The analytical solution of the Cauchy problem (3.27) is as follows [86, 88]
Z t
1
û(t) =
g0 + γ (τ )ŝ(τ ) dτ ,
(3.28)
γ (t)
0
where the contribution of r̂û is taken into account by the auxiliary function
Z t
γ (t) = exp
r̂(τ ) dτ .
0
Remark 3.6. In the case r̂ ≡ 0 and ŝ ≡ 0, solution (3.28) reduces to the identity
û(t) = g0 ,
∀t ≥ t0 .
(3.29)
This means that the solution of (3.22) is constant along the streamlines/characteristics.
Remark 3.7. By using the chain rule, we tacitly assumed that u is differentiable. This
assumption can be relaxed. If the prescribed boundary condition (3.23) has a jump at
some point x0 ∈ Γ− , then the weak solution of the linear hyperbolic equation (3.22)
will remain discontinuous along the entire characteristic x̂(t) through x0 = x̂(t0 ).
To obtain the solution u(x) of problem (3.22)–(3.23) at a point x ∈ Ω , one needs
to solve (3.27) along the characteristic that passes through x and satisfies (3.25).
Moreover, the following upper/lower bounds can be readily inferred from (3.28).
Theorem 3.7. A solution of problem (3.22)–(3.23) with ∇ · v ≡ 0 satisfies
s≤0
⇒
max u = max g.
Ω̄
Γ−
3.1 Properties of Linear Transport Models
101
In elliptic problems, maxima and minima could be attained anywhere on the boundary. The above theorem takes advantage of the fact that convection is a one-way
process, which restricts the possible location of maxima/minima to the inlet Γ− .
If the velocity field is not divergence-free, local extrema can emerge in the interior of the domain Ω but equation (3.28) still implies positivity preservation.
Theorem 3.8. A solution of problem (3.22)–(3.23) with arbitrary v satisfies
s ≥ 0,
g≥0
⇒
u ≥ 0.
Corollary 3.6. If equation (3.22) is linear, then there is at most one solution.
Corollary 3.7. If w = u − v satisfies (3.22)–(3.23) with s ≥ 0 and g ≥ 0, then
u≥v
in Ω̄ .
Linearity is essential for the proof of uniqueness and for the comparison principle.
3.1.3.2 Unsteady Convective Transport
In unsteady problems, the solution u(x,t) is defined in a bounded space-time domain Ω × (0, T ), and information is ‘convected’ forward in time with unit velocity.
Adding a time derivative to equation (3.22), one obtains its unsteady counterpart
∂u
+ ∇ · (vu) = s in Ω × (0, T ).
∂t
(3.30)
It is also hyperbolic, so boundary conditions are required only at the inlet Γ− . Let
u(x,t) = g(x,t),
∀x ∈ Γ− .
(3.31)
Since the time level t = 0 represents another ‘inflow boundary’ of the space-time
domain Ω × (0, T ), it is also necessary to prescribe a suitable initial condition
u(x, 0) = u0 (x),
∀x ∈ Ω .
(3.32)
If the velocity field v(x,t) is time-dependent, then the trajectory of a marker
depends not only on the position x0 but also on the time t0 at which it is launched.
The characteristics of equation (3.30) are defined as pathlines given by (3.25) with
v̂(t) = v(x̂(t),t). The origin of each characteristic x̂(t) is located either at the inlet
(x0 ∈ Γ− for t0 > 0) or at the initial time level (x0 ∈ Ω for t0 = 0), see Fig. 3.2.
Definition 3.5. The substantial derivative is the time rate of change along a pathline
∂u
Du
:=
+ v · ∇u.
Dt
∂t
(3.33)
102
3 Maximum Principles
Ω
x̂(t)
Ω
x0
Fig. 3.2 Characteristics of the unsteady two-dimensional convection equation.
Differentiating the function û(t) = u(x̂(t),t) along the characteristics, we obtain
d
∂ u dx̂k
dû(t) ∂ u
Dû
=
+∑
=
,
dt
∂ t k=1 ∂ xk dt
Dt
where
Dû
Dt
(3.34)
is evaluated at point x̂(t) and time t. Due to equation (3.30), it satisfies
Dû
ˆ · v)û.
= ŝ − (∇
Dt
Substitution into the right-hand side of (3.34) leads to a Cauchy problem of the form
(3.27), where the value of g0 is known from the initial/boundary data
g(x0 ,t0 ), if x0 ∈ Γ− , t0 > 0,
(3.35)
g0 =
u0 (x0 ),
if x0 ∈ Ω , t0 = 0.
Again, the instantaneous value of û(t) is given by (3.28), and the solution may be
discontinuous along the characteristic x̂(t) if there is a jump at point x0 and time t0 .
Theorem 3.9. A solution of problem (3.30)–(3.32) with ∇ · v ≡ 0 satisfies
s≤0
⇒
max u = max g or
Ω̄
Γ− ×[0,T ]
max u = max u0 .
Ω̄
Ω
In contrast to steady convection, the possible locus of maxima and minima includes
not only Γ− × [0, T ] but also the ‘time inlet’ Ω × {0} of the space-time domain.
Theorem 3.10. A solution of problem (3.30)–(3.31) with arbitrary v satisfies
s ≥ 0,
g ≥ 0,
u0 ≥ 0
⇒
u ≥ 0.
Corollary 3.8. If equation (3.30) is linear, then there is at most one solution.
3.1 Properties of Linear Transport Models
103
Corollary 3.9. Let w = u − v be a solution of (3.30)–(3.32) with s ≥ 0 and g ≥ 0. If
the corresponding initial data satisfy u0 ≥ v0 , then
u≥v
in Ω̄ × [0, T ].
Linearity is essential for the proof of uniqueness and for the comparison principle.
3.1.4 Equations of Parabolic Type
Unsteady transport equations in which both convection and diffusion are taken into
account are of parabolic type. The most general formulation of such a model is
∂u
+ L u = s in Ω × (0, T ),
∂t
(3.36)
where the linear operator L is the same as that for steady convection-diffusion
L u = ∇ · (vu − D∇u).
(3.37)
This model shares some features of its elliptic and hyperbolic counterparts which
are obtained by neglecting the time derivative and/or diffusive terms, respectively.
Definition 3.6. If L is an elliptic operator of second order, then
∂
∂t
+L is parabolic.
Loosely speaking, parabolic equations are elliptic in space and hyperbolic in
time. On the one hand, information may travel in arbitrary space directions, so the
space variables are two-way coordinates [268]. On the other hand, the time is always
a one-way coordinate since changes that occur at a given instant can only influence
the solution at the same or later time. Current happenings depend on the evolution
history and affect future events but have no influence on what happened in the past.
Due to the presence of a time derivative and second-order space derivatives, both
initial data and boundary conditions are to be prescribed. In contrast to unsteady hyperbolic problems, the distinction between inlets and outlets is irrelevant. A boundary condition is required at each point x ∈ Γ , no matter if v · n is positive or negative.
Consider an initial boundary value problem that consists of the parabolic PDE
∂u
+ ∇ · (vu − D∇u) = s in Ω × (0, T )
∂t
(3.38)
supplemented by an initial condition and boundary conditions of Dirichlet type
u(x, 0) = u0 (x),
u(x,t) = g(x,t),
∀x ∈ Ω ,
∀x ∈ Γ , ∀t ∈ (0, T ].
(3.39)
(3.40)
Our experience with transport equations of elliptic and hyperbolic type enables us
to prove the following maximum principle for the above parabolic problem.
104
3 Maximum Principles
Theorem 3.11. Let the diffusion tensor D be symmetric positive definite in Ω . Then
a solution of problem (3.38)–(3.40) with ∇ · v ≡ 0 satisfies
s≤0
⇒
max u = max g or
Ω̄
Γ ×[0,T ]
max u = max u0 .
Ω
Ω̄
(3.41)
Proof. Obviously, any values prescribed on Γ and at t = 0 satisfy the requirement
stated in (3.41). To prove the maximum principle (3.41), it is sufficient to show that
no new maxima can be generated in the interior of Ω at any time t ∈ (0, T ].
Consider g0 = maxΓ− ×[0,T ] g or g0 = maxΩ u0 , whichever is greater. By definition,
g0 = u(x0 ,t0 ), where x0 ∈ Ω and t0 = 0 or x0 ∈ Γ− and t0 > 0. The question is
whether the so-defined initial peak will grow or decay as it convected through the
flow field subject to diffusion and reaction. Along the corresponding pathline x̂(t),
equation (3.38) can be written in terms of the substantial derivative (3.34) thus:
dû
ˆ · (D ∇u)
ˆ + ŝ − (∇
ˆ · v)û,
=∇
dt
where the last term is zero since ∇ · v ≡ 0. Hence, the evolution of û(t) along the
pathline through x0 is governed by the following initial value problem
dû
ˆ · (D ∇u)
ˆ + ŝ,
=∇
dt
û(t0 ) = g0 .
In incompressible flow problems, convection alone cannot change the amplitude
of the peak g0 but diffusion and reaction surely can. Since the diffusion tensor D is
symmetric positive definite, it admits a factorization of the form (3.11) which leads
ˆ · (D ∇u)
ˆ = ∆˜ u and
to (3.13). Under the linear transformation (3.12), we have ∇
dû
= ∆˜ u + ŝ.
dt
(3.42)
If û(t) is a local maximum, then conditions (3.1) hold in the transformed coordinate
system, so the right-hand side of (3.42) is nonpositive for s ≡ 0. Therefore, the
initial peak û(t0 ) = g0 cannot increase along the pathline x̂(t), which proves that the
maximum µ = maxΩ̄ u must be attained either in Ω at t = 0 or on Γ × [0, T ]. Theorem 3.12. Let the diffusion tensor D be symmetric positive definite in Ω . Then
a solution of problem (3.38)–(3.40) with arbitrary v satisfies
s ≥ 0,
g ≥ 0,
u0 ≥ 0
⇒
u ≥ 0.
(3.43)
Proof. If the velocity field v(x,t) is not divergence-free, then the nonvanishing ‘reˆ · v)û must be added to the right-hand side of equation (3.42). As a
active’ term (∇
consequence, a positive minimum û(t0 ) = g0 may decrease along x̂(t). However, as
ˆ · v)û vanishes, so û(t) cannot
soon as its value reaches the zero level, the term (∇
decrease any further for the reasons explained in the proof of Theorem 3.11. Corollary 3.10. If equation (3.38) is linear, then there is at most one solution.
3.1 Properties of Linear Transport Models
105
Corollary 3.11. Let w = u − v be a solution of (3.38)–(3.40) with s ≥ 0 and g ≥ 0.
If the corresponding initial data satisfy u0 ≥ v0 in Ω , then
u≥v
in Ω̄ × [0, T ].
Linearity is essential for the proof of uniqueness and for the comparison principle.
3.1.5 Singularly Perturbed Problems
In hyperbolic problems, boundary conditions are required only at the inlet. If we add
a small diffusive term, this hardly makes any difference as far as the partial differential equation itself is concerned. However, extra boundary conditions are required
for the so-defined problem which is nominally elliptic or parabolic. Even if the diffusion coefficient is very small, the solution of the perturbed problem may turn out
to be a poor approximation to that of the original one and vice versa. Moreover, the
right formulation of a maximum/minimum principle changes with the PDE type.
In perturbation theory, an approximation to a regularly perturbed problem can
be obtained by simply setting the small parameter to zero. Problems that cannot be
approximated properly in this way are referred to as singularly perturbed.
In fluid dynamics, the small parameter is usually the diffusion coefficient. A classical example is the following singularly perturbed elliptic problem [176, 288]
∇ · (vu − ε ∇u) = s in Ω ,
(3.44)
u = g on Γ ,
where 0 < ε ≪ 1 is very small. The solution u of this convection-dominated problem
is smooth (differentiable) but its derivatives may become very large as ε → 0.
Definition 3.7. A zone in which u or its derivatives change abruptly is called a layer.
In steady transport problems, the location of layers is fixed and sometimes known
a priori. Their thickness decreases as the ratio |v|/ε increases. To identify a possible
cause of layers, consider the hyperbolic counterpart of problem (3.44)
∇ · (vu) = s in Ω ,
(3.45)
u = g on Γ− .
Note that the boundary data are prescribed at the inlet Γ− since the solution of a
hyperbolic equation cannot be forced to satisfy any boundary condition elsewhere.
If the solution of problem (3.45) has a jump along a characteristic through a
point x0 ∈ Γ− , then diffusion will smear it over a zone of finite thickness. As
√ a
consequence, the solution of (3.44) will exhibit an internal layer of width O( ε )
around the characteristic [176]. Discontinuous diffusion coefficients and singular
sources may also give rise to internal layers. Moreover, a boundary layer of width
106
3 Maximum Principles
Γ+
O(ε)
Γ−
√
O( ε)
Ω
Γ+
x̂(t)
x0
Γ−
O(ε)
Fig. 3.3 Interior and boundary layers for a singularly perturbed transport equation.
O(ε ) will form if the boundary conditions prescribed on Γ \Γ− are incompatible
with the boundary values of the solution to the reduced problem (3.45), see Fig. 3.3.
In the unsteady case, the solution of the singularly perturbed parabolic problem
 ∂u
in Ω × (0, T ),
 ∂ t + ∇ · (vu − ε ∇u) = s
(3.46)
u=g
on Γ × (0, T ),

u = u0 in Ω at t = 0
may also have internal and boundary layers. The former may be caused by discontinuities in initial data, inflow boundary conditions, or coefficients. As time evolves,
internal layers are convected downstream and smeared by diffusion. Their thickness
and the rate of smearing depend on the value of the perturbation parameter ε .
3.2 Matrix Analysis for Steady Problems
If the solution of a given boundary value problem satisfies a maximum principle,
then a properly designed approximation should behave in the same way. A numerical scheme that does not generate spurious global extrema in the interior of the
computational domain is said to satisfy a discrete maximum principle (DMP). As in
the continuous case, the precise formulation of this criterion is problem-dependent.
In particular, the zero row sum property (second rule from Section 1.6.3) has the
same implications as the constraint ∇ · v ≡ 0 in continuous maximum principles.
In the context of finite difference approximations to linear elliptic problems, sufficient conditions of DMP were formulated and proven by Varga [340] as early as
in 1966. These conditions are related to the concept of monotone operators and,
in particular, M-matrices which play an important role in numerical linear algebra
[339, 354]. A general approach to DMP analysis for finite difference operators was
developed by Ciarlet [63]. Its extension to finite elements in [64] features a proof of
uniform convergence, as well as simple geometric conditions that ensure the validity
3.2 Matrix Analysis for Steady Problems
107
of DMP for a piecewise-linear Galerkin discretization of the (linear) model problem
−∆ u + ru = s, in Ω ,
(3.47)
u = g on Γ ,
on a triangular mesh under the assumption that r ≥ 0 in Ω . The results obtained in
[63, 64, 339] have illustrated the significance of DMPs for the analysis and design of
numerical approximations. Various extensions and generalizations were published
during the past three decades, see [61, 100, 151, 179, 180, 341, 311] and references
therein. The frequently cited monograph by Ikeda [167] is devoted entirely to DMP
for finite element models of convection-diffusion phenomena.
Some low-order approximations of transport equations are known to satisfy a
DMP unconditionally or under rather mild restrictions on the angles or aspect ratios
of mesh cells. However, most a priori proofs are based on a set of sufficient conditions which become overly restrictive in the case of higher-order discretizations,
singularly perturbed convection-diffusion equations, and anisotropic diffusion problems. A possible remedy to this problem is proposed in the next chapter.
In this section, we review the algebraic constraints that ensure DMP and/or positivity preservation for steady transport problems of elliptic and hyperbolic type. A
brief summary of the corresponding geometric conditions will also be presented.
3.2.1 The Discrete Problem
Consider the steady transport-reaction equation (3.9) discretized by a finite difference, finite volume, or finite element scheme. Let the approximate solution uh ,
where the subscript h refers to the mesh size, be determined by a finite number N̄
of degrees of freedom u1 , . . . , uN̄ that represent pointwise nodal values, control volume averages, or coefficients of piecewise-polynomial basis functions, respectively.
Hence, all information about the solution uh can be packed into a vector u ∈ RN̄ .
Furthermore, the differential operator L acting on functions defined at infinitely
many locations is replaced by a discrete operator Lh acting on vectors of length N̄
Lh : RN̄ → RN̄ .
Regardless of the underlying approximation technique, we define this mapping as
Lh u = Au,
where A = {ai j } is a sparse N̄ × N̄ matrix and u = {ui } is the vector of nodal values.
The sparsity pattern of A depends on the mesh, on the type of discretization, and
on the numbering of nodes. Since some nodal values are known from the Dirichlet
boundary conditions, the size of the algebraic system reduces accordingly.
Let the first N nodes be associated with the unknown degrees of freedom, and the
rest with the Dirichlet boundary values. This numbering convention implies that the
108
3 Maximum Principles
discrete operator A and the vector of nodal values u can be partitioned as follows
uΩ
AΩ Ω AΩΓ
, u=
.
(3.48)
A=
AΓ Ω AΓ Γ
uΓ
The subscripts Ω and Γ refer to row/column numbers from NΩ = {1, . . . , N} and
NΓ = {N + 1, . . . , N̄}, respectively. Thus, uΩ = {u1 , . . . , uN } is the vector of unknowns, whereas uΓ = {uN+1 , . . . , uN̄ } is given by the prescribed boundary data
uΓ = g.
(3.49)
In this notation, the system of algebraic equations for the components of uΩ reads
AΩ Ω uΩ = bΩ − AΩΓ g,
(3.50)
where bΩ is the contribution of sources and Neumann boundary conditions, if any.
In a practical implementation, it is convenient to incorporate the Dirichlet boundary conditions into the N̄ × N̄ matrix A and solve the extended linear system [63]
Āu = b,
(3.51)
where the matrix Ā and right-hand side b are defined so as to enforce (3.49)
bΩ
AΩ Ω AΩΓ
, b=
.
(3.52)
Ā =
0
I
g
In other words, the N̄ × N̄ matrix Ā is constructed from A by setting AΓ Ω := 0 and
AΓ Γ := I, where I denotes the identity matrix with N̄ − N rows and columns.
If the solution of the continuous problem satisfies Theorem 3.5 or 3.7, it is natural
to require that the maxima of uΩ be bounded by those of uΓ = g. Likewise, all
nodal values should be nonnegative if Theorem 3.6 or 3.8 is applicable. To verify
the validity of DMP, one needs to analyze the properties of the discrete operator Ā.
3.2.2 M-Matrices and Monotonicity
A key ingredient of the mathematical theory behind the discrete maximum principles and positivity preservation is the following monotonicity concept [63, 339].
Definition 3.8. A regular matrix A = {ai j } is called monotone if A−1 ≥ 0.
This kind of monotonicity is equivalent to the requirement that, for any vector u,
Au ≥ 0
⇒
u ≥ 0.
3.2 Matrix Analysis for Steady Problems
109
Of course, it is impractical to compute the inverse of A and check the sign of its
entries. Instead, we will deal with a special class of matrices which are known to be
monotone under certain constraints on the sign and magnitude of their coefficients.
Definition 3.9. A regular matrix A = {ai j } is called an M-matrix if A−1 ≥ 0 and
ai j ≤ 0,
∀ j 6= i.
In other words, an M-matrix is a monotone matrix with nonpositive off-diagonal
entries. These properties ensure positivity and convergence of iterative solvers.
Definition 3.10. A matrix A = {ai j } is called diagonally dominant (by rows) if
|aii | ≥ ∑ |ai j |,
j6=i
∀i.
(3.53)
Such a matrix is called strictly diagonally dominant if all inequalities are strict
|aii | > ∑ |ai j |,
j6=i
∀i.
(3.54)
Definition 3.11. A matrix A = {ai j } of size N × N is called irreducible if there is no
N × N permutation matrix P such that the following transformation is possible
A11 A12
T
,
PAP =
0 A22
where the size of A11 is M ×M, the size of A22 is (N −M)×(N −M), and 1 ≤ M < N.
It turns out that a matrix A is irreducible if and only if its directed graph is
strongly connected ([339], p. 20) or, equivalently, if and only if for any i and j 6= i
there is a sequence of distinct indices i = n0 , n1 , . . . , nl = j such that [132]
ank−1 nk 6= 0,
1 ≤ k ≤ l.
Remark 3.8. In the context of linear systems, irreducibility ensures that it is impossible to extract a subsystem that can be solved independently. Matrices that result
from discretization of partial differential equations are irreducible in most cases.
Definition 3.12. A matrix A = {ai j } is irreducibly diagonally dominant if it is irreducible and diagonally dominant, with strict dominance for at least one row i.
The following theorem yields a set of sufficient conditions which are commonly
employed in DMP analysis based on the M-matrix property of discrete operators.
Theorem 3.13. If A = {ai j } is a strictly or irreducibly diagonally dominant N × N
matrix with aii > 0, ∀i = 1, . . . , N and ai j ≤ 0, ∀ j 6= i, then A−1 ≥ 0.
110
3 Maximum Principles
Proof. A common approach to the proof of this theorem is based on the splitting
A = D −C,
where the diagonal part D = diag(A) > 0 is nonsingular and C ≥ 0. The diagonal
dominance makes it possible to prove that the spectral radius ρ of B = D−1C ≥ 0
satisfies ρ (B) < 1. This condition holds if and only if the series
(I − B)−1 = I + B + B2 + B3 + . . .
converges, see [131, 339] for technical details. Hence, A−1 = (I − B)−1 D−1 ≥ 0. If all diagonal entries of A are strictly positive and there are no positive offdiagonal ones, then diagonal dominance (3.53) requires that all row sums be nonnegative. The following definition summarizes the corresponding sign conditions.
Definition 3.13. A matrix A = {ai j } is said to be of nonnegative type [64, 69] if
aii > 0,
∀i,
ai j ≤ 0,
∀ j 6= i,
∑ ai j ≥ 0, ∀i.
(3.55)
(3.56)
(3.57)
j
Corollary 3.12. By Theorem 3.13, a nonnegative-type matrix A is an M-matrix if
inequality (3.57) is strict or A is irreducible and (3.57) is strict for at least one row.
Note that conditions (3.55)–(3.56) impose the same constraints as the third basic
rule from Section 1.6.3. The second basic rule is satisfied if (3.57) holds as equality.
Under the assumptions of Corollary 3.12, the nonnegativity conditions are sufficient (but not necessary) for the matrix A to be monotone. Some other useful criteria
related to M-matrices and monotonicity can be found in [46, 132, 153, 311, 346].
3.2.3 Discrete Maximum Principles
Given a discretization of the form (3.51), the monotonicity of the matrix Ā makes it
possible to prove discrete counterparts of all maximum, minimum, and comparison
principles established in Section 3.1. The uniqueness of the solution vector u follows
from the regularity of Ā. The usual approach to the proof of monotonicity is based on
Theorem 3.13 and Corollary 3.12 since the nonnegativity conditions (3.55)–(3.57)
are easy to verify for an arbitrary space discretization of the transport equation.
To prove that the solution of problem (3.51) attains its maximum on the set NΓ of
Dirichlet boundary nodes, we need a discrete counterpart of the incompressibility
constraint ∇ · v ≡ 0. At the continuous level, it implies that L u = L (u + c) for
an arbitrary constant c. According to the second basic rule from Section 1.6.3, the
discrete operator A should have zero row sums to inherit this property. Thus, the
global discrete maximum principle for nodal values can be formulated as follows.
3.2 Matrix Analysis for Steady Problems
111
Theorem 3.14. If the matrix Ā is given by (3.52), AΩ Ω is monotone, AΩΓ ≤ 0, and
N̄
∑ ai j = 0,
∀i ∈ NΩ ,
j=1
(3.58)
then the solution of (3.51) satisfies the global discrete maximum principle
bΩ ≤ 0
⇒
max ui = max g j .
i
(3.59)
j
Proof. Let bΩ ≤ 0 and consider w = u − µ , where µ = max j g j is the largest Dirichlet boundary value. Due to (3.51) and the zero-sum property (3.58), we have
N̄
∑ ai j w j =
j=1
N̄
∑ ai j u j − µ
j=1
N̄
∑ ai j =
j=1
N̄
∑ ai j u j ,
j=1
∀i ∈ NΩ .
(3.60)
It follows that AΩ Ω wΩ +AΩΓ wΓ = bΩ ≤ 0, where AΩΓ wΓ ≥ 0. Since AΩ Ω is monotone, wΩ = A−1
Ω Ω [bΩ − AΩΓ wΓ ] ≤ 0 so that ui ≤ max j g j for all i ∈ NΩ .
Since the matrix Ā is sparse, only nearest neighbors make a nonzero contribution
to the right-hand side of the algebraic equation for an interior node i ∈ NΩ . The
following theorem states that ui is bounded by the solution values at neighbor nodes.
Theorem 3.15. If the matrix Ā is of nonnegative type and condition (3.58) holds,
then the solution of (3.51) satisfies the local discrete maximum principle [20]
bi ≤ 0
⇒
ui ≤ max u j ,
j∈Ni
∀i ∈ NΩ ,
(3.61)
where Ni := { j 6= i | ai j 6= 0} is the set of neighbors that form the stencil of node i.
Proof. Let i ∈ NΩ be any interior node. The equation for the nodal value ui reads
aii ui = bi −
∑
ai j u j .
(3.62)
j∈Ni
The zero sum property (3.58) implies that the involved coefficients satisfy
ai j
= −1.
j∈Ni aii
∑
Due to the assumptions that Ā is of nonnegative type and bi ≤ 0, this yields
!
ai j
ai j
bi
− ∑
u j ≤ − max u j ∑
= max u j ,
(3.63)
ui =
aii j∈N
aii
j∈Ni
j∈Ni
j∈Ni aii
i
which proves that ui is bounded by the maximum over the stencil of i ∈ NΩ .
Corollary 3.13. If bi = 0 and u j = ū, ∀ j ∈ Ni , then ui = ū by the local DMP.
112
3 Maximum Principles
That is, if the source term is absent and the solution has a constant value µ at all
neighboring nodes, then ui must also assume this value. This property is guaranteed
by the zero row sum condition (3.58). Only a poor discretization of the transport
equation with ∇ · v ≡ 0 would produce ui 6= ū in this situation ([268], p. 39).
Remark 3.9. Since estimate (3.63) holds for any interior node, successive application of the local DMP can be used to prove (3.59) if AΩ Ω is irreducible, cf. [69, 344].
If the row sums of Ā are nonvanishing for some i ∈ NΩ , then the DMP may cease
to hold but positivity preservation can be inferred from the fact that Ā is monotone.
Theorem 3.16. If the matrix Ā is given by (3.52), where AΩ Ω is monotone and
AΩΓ ≤ 0, then discretization (3.51) is positivity-preserving, that is,
b≥0
⇒
u ≥ 0.
(3.64)
Proof. The matrix Ā given by (3.52) is regular if and only if the block AΩ Ω is
−1
regular. Furthermore, the inverse matrices Ā−1 and AΩ
Ω are related by the formula
−1
−1
AΩΓ
AΩ Ω −AΩ
−1
Ω
Ā =
.
(3.65)
0
I
−1
−1
If AΩ
Ω ≥ 0 and AΩΓ ≤ 0, then Ā is monotone and u = Ā b ≥ 0 for any b ≥ 0.
Remark 3.10. In the case of linear or bilinear finite elements, the interpolant uh satisfies a local maximum principle inside each cell. That is, the value of the approximate
solution at any interior point x ∈ Ω is bounded by the nodal values at the vertices of
the cell containing x, so that Theorems 3.14–3.16 can be used to estimate uh (x).
Remark 3.11. Many other definitions and proofs of DMP can be found in the literature. The following result [311] is one of the most general and elegant formulations.
Theorem 3.17. If Ā is an M-matrix that satisfies condition (3.57), then
max ui ≤ max u j ,
1≤i≤N̄
j∈N+
N+ = {1 ≤ j ≤ N̄ | b j > 0}.
In the case N+ = 0,
/ the right-hand side of the above inequality is taken to be zero.
This theorem states that maxima occur on a set of nodes where the right-hand side of
system (3.51) is positive. Remarkably, the degrees of freedom associated with NΩ
and NΓ are treated equally. For a proof, the interested reader is referred to [311].
Theorem 3.18. If Ā is a linear monotone operator and Āu ≥ Āv, then u ≥ v.
Proof. Due to linearity and monotonicity, Ā(u − v) ≥ 0 ⇒ u − v ≥ 0.
Unfortunately, only some low-order approximations can be both linear and monotone. To achieve higher accuracy, one of these requirements has to be sacrificed.
3.2 Matrix Analysis for Steady Problems
113
Theorem 3.19. Any linear monotone operator that results from a discretization of
first-order space derivatives can be at most first-order accurate.
This order barrier is known as the Godunov theorem [123]. It is responsible for the
tradeoff between devastating numerical diffusion and undershoots/overshoots. Since
first-order accuracy is usually insufficient, the only way to avoid both side effects is
to adjust the coefficients of the discrete transport operator in an adaptive way so as
to enforce relevant DMP conditions a posteriori, that is, for a given data set.
Theorem 3.20. Any linear monotone operator that results from a discretization of
second-order space derivatives can be at most second-order accurate.
This result is also disappointing but second-order accuracy if often sufficient for
practical purposes. A simple proof of Theorems 3.19 and 3.20 for finite difference
schemes is available in [164] on pp. 118-120. The lack of monotonicity for quadratic
finite element discretizations of the Laplace operator was first reported in [151].
3.2.4 Desirable Mesh Properties
Some finite element approximations to elliptic problems like (3.47) are known to
satisfy the DMP conditions (3.55)–(3.57) on a suitably designed mesh. The derivation of geometric constraints that ensure monotonicity has been one of the primary
research directions in the DMP analysis for finite element schemes [64, 100, 179,
180]. Below, we present some useful geometric criteria in the context of linear and
bilinear Galerkin discretizations of the Laplace operator in two space dimensions.
Definition 3.14. A triangular mesh is called strongly acute if all angles are smaller
than π /2 and weakly acute (or nonobtuse) if all angles are not greater than π /2.
Theorem 3.21. The discrete Laplace operator Ā for the linear finite element approximation on a triangular mesh of weakly acute type is monotone [18, 64].
This classical result dates back to the paper by Ciarlet and Raviart [64]. In the 3D
case, a tetrahedral mesh is said to be of acute type if all internal angles between the
faces of tetrahedra are not greater than π /2. Again, this condition ensures that the
discrete Laplace operator is monotone if linear finite elements are employed [189].
Definition 3.15. A triangular mesh is a Delaunay triangulation if no vertex of this
mesh is inside the circumcircle of any triangle to which it does not belong.
Theorem 3.22. The discrete Laplace operator Ā for the linear finite element approximation on a Delaunay triangulation is monotone [18].
Delaunay triangulations maximize the minimum angle as to avoid excessively
stretched triangles. It is known that there exists a unique Delaunay triangulation for
114
3 Maximum Principles
any set of points that do not lie on the same line. Moreover, fast algorithms are available for creating such triangulations [55, 115, 226], which makes them very popular
with finite element practitioners. Figure 3.4 displays a simple 2D Delaunay triangulation generated using the MATLAB function delaunay. In three dimensions, no
vertex of a tetrahedron is inside the circumsphere of any other tetrahedron. However,
the discrete Laplacian operator for a linear FEM approximation on an arbitrary 3D
Delaunay triangulation may fail to be a matrix of nonnegative type [18]. This does
not necessarily cause a violation of the DMP but it cannot be ruled out anymore.
Definition 3.16. A rectangular mesh is√called nonnarrow if the ratio of longest and
shortest mesh edge is not greater than 2 for any rectangle [100].
Theorem 3.23. The discrete Laplace operator Ā for the bilinear finite element approximation on a rectangular mesh of nonnarrow type is monotone [61, 100].
This theorem explains why iterative solution techniques that rely on the M-matrix
property may experience convergence problems when applied to discretizations of
second-order PDEs on quadrilateral/hexahedral meshes with high aspect ratios.
Geometric DMP conditions for various elliptic and parabolic problems have been
formulated building on the above results [100, 116, 179, 180]. Even if convective
effects are present, it is desirable to use a sufficiently regular mesh that satisfies the
above conditions, so that at least the discrete diffusion operator poses no hazard to
DMP. Moreover, it may offset a nonmonotone convective part if the Peclet number
is small or a sufficiently large amount of artificial diffusion is added [51, 69]. Alternatively, the upwind triangle method or other techniques can be used to construct a
monotone approximation of convective terms [11, 167, 200, 288, 348].
1
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
0.8
Fig. 3.4 A two-dimensional Delaunay triangulation with 15 points.
1
3.3 Matrix Analysis for Unsteady Problems
115
Many finite element approximations do not produce a monotone matrix, or more
sophisticated tools than Theorem 3.13 are required to prove monotonicity. It is not
unusual that spurious maxima or minima of significant amplitude are generated in
regions where small-scale features are present and the mesh is too coarse. On the
other hand, a well-resolved numerical solution satisfies the DMP even if it violates
the sufficient monotonicity conditions that impose unrealistically severe restrictions
on the properties of the mesh and on the choice of polynomial approximations.
3.3 Matrix Analysis for Unsteady Problems
Unsteady transport processes are governed by equations of parabolic and hyperbolic
type in which convective terms may be dominant. The time derivative can also be
interpreted as ‘convection’ in the positive t−direction. In inviscid flow problems, the
discrete maximum principle and positivity preservation can be enforced within the
framework of monotone, monotonicity-preserving, and total variation diminishing
(TVD) methods for 1D hyperbolic conservation laws. Many representatives of these
schemes are explicit and/or essentially one-dimensional. Their extensions to 2D/3D
rely on dimensional splitting, which rules out the use of unstructured meshes.
It turns out that any multidimensional TVD scheme is at most first-order accurate, except in certain trivial cases [124, 216]. The concept of local extremum diminishing (LED) schemes [170, 171] makes it possible to design approximations that
enjoy the TVD property in the 1D case and provide a weaker form of monotonicity in multidimensions. Interestingly enough, all of the above DMP criteria impose
essentially the same constraints on the coefficients of the space discretization.
In transient computations, the time-stepping method should be chosen so as to
keep the solution free of undershoots and overshoots. Since the Godunov order barrier applies to time discretizations as well, only the first-order accurate backward
Euler method may be used with arbitrary time steps. In all other cases, monotonicity conditions impose a certain upper bound on the time step. This constraint may
be the same or more stringent than the usual stability condition, if any. The use of
the consistent mass matrix in finite element discretizations of unsteady hyperbolic
and parabolic problems may also cause some complications [33, 100, 116].
3.3.1 Semi-Discrete DMP Constraints
In this section, we start with the DMP analysis for semi-discrete schemes that can
be written as a system of differential algebraic equations (DAEs) of the form
M
du
= Cu + r,
dt
(3.66)
116
3 Maximum Principles
where u is the vector of nodal values, M = {mi j } is the mass matrix, C = {ci j } is
the negative of the discrete transport operator, and r is the vector of nodal sources.
A semi-discrete maximum principle for a space discretization of the form (3.66)
is particularly easy to establish if M is a diagonal matrix. Finite difference and finite
volume discretizations satisfy this requirement from the outset. In the realm of finite
elements, it is commonly enforced using row-sum mass lumping, see Section 2.1.2.
Consider the i−th equation in system (3.66) with the mass matrix M = diag{mi }
mi
dui
= ∑ ci j u j + ri .
dt
j
(3.67)
Since the coefficient matrix C = {ci j } is sparse, only nearest neighbors from the set
Ni = { j 6= i | ci j 6= 0}
can contribute to the right-hand side of (3.67). If the coefficients ci j sum to zero
∑ ci j = 0,
(3.68)
j
then it follows that cii = − ∑ j6=i ci j , whence equation (3.67) can be written as
mi
dui
= ∑ ci j (u j − ui ) + ri .
dt
j6=i
(3.69)
Theorem 3.24. The following (local) semi-discrete maximum principle holds for the
solution of equation (3.69) with ri ≤ 0 if mi > 0 and ci j ≥ 0 for all j 6= i
ui ≥ u j ,
∀ j ∈ Ni
⇒
dui
≤ 0.
dt
(3.70)
Proof. The theorem states that a maximum ui = max j u j cannot increase. Indeed,
dui
1
=
dt
mi
ri
∑ ci j (u j − ui ) + mi
(3.71)
j6=i
is nonpositive due to the fact that mi > 0, ri ≤ 0, and ci j (u j − ui ) ≤ 0, ∀ j 6= i.
In a similar vein, a minimum ui = min j u j cannot decrease if ri is nonnegative
ui ≤ u j ,
∀ j ∈ Ni
⇒
dui
≥ 0.
dt
(3.72)
In the case ri = 0, the above semi-DMP states that neither maxima nor minima are
enhanced, which has led Jameson [170, 171] to introduce the following definition.
Definition 3.17. A semi-discrete scheme of the form (3.69) with ri = 0 is called
local extremum diminishing (LED) if estimates (3.70) and (3.72) hold for all i.
3.3 Matrix Analysis for Unsteady Problems
117
Corollary 3.14. The following conditions are sufficient for (3.69) to be LED
mi > 0,
ri = 0,
ci j ≥ 0,
∀i, ∀ j 6= i.
(3.73)
Remark 3.12. Flux-corrected transport (FCT) algorithms [42, 355] impose essentially the same constraints (no new maxima or minima, no growth of existing ones).
Remark 3.13. The term local implies that maxima and minima are taken over the
stencil of individual nodes. A global semi-discrete maximum principle and, hence,
L∞ -stability can be readily inferred from the LED criterion and Theorem 3.24.
The concept of LED schemes provides [171] “a convenient basis for the construction
of nonoscillatory schemes on both structured and unstructured meshes.” Indeed,
conditions like (3.73) are easy to check and enforce for arbitrary discretizations.
In the context of three-point finite difference schemes for 1D problems, the nonnegativity of both off-diagonal coefficients is also required by Harten’s TVD conditions [137]. It turns out that there is a close relationship between TVD and LED
space discretizations. In one space dimension, the total variation
TV (u,t) = ∑ |ui+1 (t) − ui (t)|,
∀t ≥ 0
i
must be a nonincreasing function of time for a semi-discrete scheme to be TVD. If
zero boundary values are prescribed at both endpoints of the 1D domain, then [171]
!
TV (u,t) = 2
∑
j∈Nmax
u j (t) −
∑
uk (t) ,
(3.74)
k∈Nmin
where Nmax and Nmin contain the indices of local maxima and minima, respectively
Nmax = { j | u j ≥ u j±1 },
Nmin = {k | uk ≤ uk±1 }.
If the LED constraint holds, then these maxima and minima cannot grow, whence
TV (u,t1 ) ≥ TV (u,t2 ),
∀t1 ≤ t2 .
Therefore, a semi-discrete LED scheme enjoys the TVD property in one dimension.
Remark 3.14. Due to Theorems 3.19 and 3.20, a linear LED/TVD discretization of
convective/diffusive terms is at most first/second-order accurate, respectively.
It is certainly incorrect to demand that the numerical scheme be local extremum
diminishing if the exact solution does not satisfy the LED constraint. However, it is
still possible to prove the following property implied by Theorems 3.10 and 3.12.
Definition 3.18. A semi-discrete scheme of the form (3.67) is called positive if
ui (0) ≥ 0,
∀i
⇒
ui (t) ≥ 0,
∀i, ∀t > 0.
(3.75)
118
3 Maximum Principles
To avoid a common misunderstanding, we emphasize that the numerical solution is
not forced to be positive if there is j 6= i such that u j (0) < 0. Positivity preservation
means that the numerical scheme cannot produce nonphysical negative values, i.e.,
undershoots. Likewise, an initially nonpositive solution should preserve its sign,
so that no overshoots are generated. Sink terms may destroy positivity and require
special treatment in accordance with the fourth basic rule from Section 1.6.3.
Theorem 3.25. The following conditions are sufficient for (3.67) to be positive
mi > 0,
ri ≥ 0,
ci j ≥ 0,
∀i, ∀ j 6= i.
(3.76)
Proof. Suppose that ui (t) = 0 and u j (t) ≥ 0 for all j ∈ Ni . Then the time derivative
of ui satisfies (3.71) and is nonnegative under the above sufficient conditions. Corollary 3.15. Let u and v be solutions computed by a linear positive scheme using
the initial data ui (0) ≥ vi (0), ∀i, all other settings being fixed. Then
ui (t) ≥ vi (t),
∀i,
∀t ≥ 0.
Remark 3.15. In the 1D case, finite difference schemes that satisfy such a comparison principle for initial data are monotone by definition [216].
3.3.2 Fully Discrete DMP Constraints
After the discretization in time and implementation of the Dirichlet boundary conditions, the fully discrete counterpart of (3.66) can be written in the form
uΩ
b
AΩ Ω AΩΓ
= Ω ,
(3.77)
0
I
uΓ
bΓ
where uΩ is the vector of unknowns for the current time step, bΓ is the vector of
prescribed boundary values, and bΩ depends on the previously computed data
bΩ = BΩ Ω gΩ + BΩΓ gΓ + sΩ .
The matrix blocks BΩ Ω and BΩΓ contain the coefficients of the explicit part that
depends on the vector of old nodal values g = gΩ ∪ gΓ . The remainder sΩ represents
the contribution of source terms and Neumann boundary conditions, if any.
For a two-level time-stepping method, u = un+1 and g = un , where the superscripts refer to the time levels t n+1 and t n = t n+1 − ∆ t, respectively. In fractional-step
algorithms, a pair of intermediate solutions may also be denoted by u and g.
Definition 3.19. The global discrete maximum principle holds for (3.77) if
sΩ ≤ 0
⇒
ui ≤ max g j ,
j
∀i ∈ NΩ .
(3.78)
3.3 Matrix Analysis for Unsteady Problems
119
Theorem 3.26. Let the first NΩ rows sums of A and B be equal (cf. [101]), i.e.,
N̄
N̄
j=1
j=1
∑ ai j = ∑ bi j ,
∀i ∈ NΩ
(3.79)
and the block AΩ Ω be regular. Then the solution uΩ of problem (3.77) satisfies the
global discrete maximum principle under the following sign conditions
A−1
Ω Ω ≥ 0,
AΩΓ ≤ 0,
BΩ Ω ≥ 0,
BΩΓ ≥ 0.
(3.80)
Proof. Consider w = u − µ and v = g − µ ≤ 0, where µ = max j g j . Due to (3.79)
N̄
∑ (bi j v j − ai j w j ) + si =
j=1
N̄
∑ (bi j g j − ai j u j ) + si = 0,
j=1
∀i ∈ NΩ .
(3.81)
It follows that AΩ Ω wΩ + AΩΓ wΓ = BΩ Ω vΩ + BΩΓ vΓ + sΩ , where sΩ ≤ 0 by assumption and wΓ ≤ 0, v ≤ 0 by definition. Invoking (3.80), we infer that
−1
wΩ = AΩ
Ω [BΩ Ω vΩ + BΩΓ vΓ − AΩΓ wΓ + sΩ ] ≤ 0.
This estimate proves the global DMP which requires that ui ≤ µ for all i ∈ NΩ .
As in the case of steady transport equations, it is also possible to estimate the
unknown nodal value ui in terms of the data defined at a few neighboring nodes.
Definition 3.20. The local discrete maximum principle holds for (3.77) if
si ≤ 0
⇒
ui ≤ µi ,
∀i ∈ NΩ ,
(3.82)
where µi denotes the maximum taken over {u j | ai j 6= 0, j 6= i} and {g j | bi j 6= 0}.
Theorem 3.27. The solution of (3.77) satisfies (3.82) subject to the row-sum constraint (3.79) and conditions of the third basic rule from Section 1.6.3, i.e.,
aii > 0,
ai j ≤ 0,
bii ≥ 0,
bi j ≥ 0,
∀i,
∀ j 6= i.
(3.83)
(3.84)
Proof. Let i ∈ NΩ be any interior node. Following the proof of Theorem 3.26, we
introduce the auxiliary functions w = u − µi and v = g − µi . By definition of µi for
the local DMP, w j ≤ 0 for all j 6= i such that ai j 6= 0, and vk ≤ 0 for 1 ≤ k ≤ N̄ such
that bik 6= 0. The row sum property leads to a relation of the form (3.81), whence
aii wi =
N̄
N̄
j=1
j6=i
∑ bi j v j − ∑ ai j w j + si .
(3.85)
Conditions (3.83)–(3.84) imply that aii > 0 and the right-hand side of equation
(3.85) is nonpositive for si ≤ 0. This proves that wi ≤ 0, that is, ui ≤ µi . 120
3 Maximum Principles
Remark 3.16. Note that the steady-state counterpart (3.51) of the local discrete maximum principle (3.77) is recovered for u = g. Thus, pseudo-time stepping can be
used not only to march the solution to a steady state but also to prove DMP for
stationary problems. Of course, a proof of convergence should also be provided.
It remains to formulate sufficient conditions of positivity preservation for discretizations of transport equations in which sources, sinks, or compressibility effects
may destroy the ‘equal row sum’ property (3.79) of the two coefficient matrices.
Due to the presence of a strictly diagonally dominant mass matrix, the usual way to
establish positivity preservation for time-dependent problems is as follows.
Theorem 3.28. If the coefficients of (3.77) satisfy conditions (3.83)–(3.84) and
∑ ai j > 0,
j
∀i ∈ NΩ ,
(3.86)
then such a discretization is guaranteed to be positivity-preserving, that is,
sΩ ≥ 0,
g≥0
⇒
uΩ ≥ 0.
Proof. Due to (3.83)–(3.84) and (3.86), the block AΩ Ω is monotone by Corollary
(3.12). Thus, uΩ = A−1
Ω Ω [BΩ Ω gΩ + BΩΓ gΓ − AΩΓ uΓ + sΩ ] ≥ 0.
Corollary 3.16. If solutions u and v are computed by the same linear positivitypreserving scheme using the data g ≥ h, all other settings being fixed, then u ≥ v.
This comparison principle represents a fully discrete counterpart of Corollary 3.15.
As usual, the proof is based on Theorem 3.28 applied to the vector w = u − v.
3.3.3 Positive Time-Stepping Methods
The fully discrete maximum principles may turn out to be more restrictive than their
semi-discrete counterparts. Even if the space discretization is designed to be monotone, the final algebraic system may fail to satisfy the DMP conditions, especially
in the case of a finite element approximation with a nondiagonal mass matrix.
According to Theorem 3.28, the following set of sufficient conditions guarantees
positivity preservation for a discretization that can be cast into the form (3.77)
• the diagonal block AΩ Ω has (i) positive diagonal entries, (ii) nonpositive offdiagonal entries, and (iii) positive row sums (strict diagonal dominance).
• there are no positive coefficients in AΩΓ and negative ones in BΩ Ω or BΩΓ .
Furthermore, the global and local DMP hold under the additional condition (3.79).
Some time-stepping schemes preserve positivity, at least if the time step is chosen so as to satisfy the above algebraic constraints on the coefficients of discrete
operators. For example, let (3.66) be discretized in time by the θ −scheme
3.3 Matrix Analysis for Unsteady Problems
M
121
un+1 − un
= θ Cun+1 + (1 − θ )Cun + r,
∆t
(3.87)
where 0 ≤ θ ≤ 1 is an implicitness parameter. This formulation combines the forward Euler (θ = 0), Crank-Nicolson (θ = 12 ), and backward Euler (θ = 1) methods.
Collecting all terms that depend on u = un+1 and g = un , we can multiply (3.87)
by the time step ∆ t and write this algebraic system in the equivalent form
Aun+1 = Bun + s,
where s = ∆ tr and the involved coefficient matices are defined as follows
A = M − θ ∆ tC,
B = M + (1 − θ )∆ tC.
Therefore, the objective is to check the sign and magnitude of the coefficients
ai j = mi j − θ ∆ tci j ,
bi j = mi j + (1 − θ )∆ tci j .
(3.88)
Suppose that the underlying space discretization is of positive type, that is,
ci j ≥ 0,
∀i, ∀ j 6= i.
(3.89)
This condition holds for some upwind-type discretizations of convective terms, as
well as for centered schemes if there is enough (physical or artificial) diffusion.
Moreover, the geometric DMP conditions for finite element approximations of diffusive terms may require that the mesh be of acute/nonnarrow type, see Section
3.2.4. Alternatively, flux/slope limiters can be used to enforce condition (3.89).
If the mass matrix M is diagonal, all off-diagonal coefficients have the right sign
mi j = 0,
ci j ≥ 0
⇒
ai j ≤ 0,
bi j ≥ 0,
∀i, ∀ j 6= i.
In the case of (linear and bilinear) finite elements, the consistent mass matrix has
some positive off-diagonal entries. Thus, the corresponding coefficients bi j remain
nonnegative but ai j ≤ 0 only if the time step ∆ t satisfies the following lower bound
∆t ≥
mi j
= ∆ tmin ,
θ ci j
∀i, ∀ j 6= i.
(3.90)
Obviously, this requirement is impossible to fulfil with θ = 0 or ci j = 0. Thus,
the time-stepping scheme must be implicit (θ > 0) and the diffusive term must be
nonvanishing for a consistent-mass finite element method to satisfy the sufficient
conditions of positivity preservation [100, 116, 246]. To circumvent this problem,
some modifications of the mass matrix were proposed by Berzins [33, 34, 35].
Conditions (3.83) and (3.86) for the diagonal coefficient aii require that
∑[mi j − θ ∆ tci j ] > 0,
j
∀i.
(3.91)
122
3 Maximum Principles
The diagonal coefficient bii is nonnegative under the following condition
mii + (1 − θ )∆ tcii ≥ 0,
∀i
(3.92)
which is responsible for stability and boundedness of the explicit part. The backward Euler method satisfies this condition automatically but the Crank-Nicolson
scheme is positivity-preserving only for sufficiently small time steps, although it is
unconditionally stable. In the case of the forward Euler method, the time step must
be small anyway for stability reasons, so positivity can be achieved at no extra cost.
Remark 3.17. Condition (3.92) with θ < 1 is particularly restrictive if the coefficient
cnii assumes a large negative value. This situation occurs, e.g., if diffusion plays a
dominant role or the divergence of the velocity field is large and positive.
Due to (3.91) and (3.92) the upper bound for largest admissible time step is
∑ j mi j
mii
∆ t ≤ min
∀i.
(3.93)
= ∆ tmax ,
,
max{0, θ ∑ j ci j } (θ − 1)cii
In the case of a 1D hyperbolic conservation law discretized in space by the explicit
first-order upwind scheme, this bound reduces to the standard CFL condition.
Theorem 3.29. If the space discretization satisfies (3.89), while the time step satisfies (3.90) and (3.93), then the θ −scheme (3.87) is positivity-preserving.
The restrictions on the choice of θ and ∆ t are very stringent in the case of consistentmass FEM since the time step ∆ t cannot be greater than ∆ tmax and smaller than
∆ tmin . The use of diagonal mass matrices and/or implicit algorithms is preferable
from the standpoint of positivity preservation, see also [100, 116, 164, 246].
The analysis of the θ −scheme can be extended to other time discretizations. For
example, let M = diag{mi } and consider an explicit L-stage Runge-Kutta method
Mu(l) =
l−1
(k)
(k) (k)
Mu
+
,
γ
θ
∆
tC
u
kl
kl
∑
k=0
n
u(0) = u ,
un+1 = u(L) ,
l = 1, . . . , L.
(3.94)
(3.95)
This time discretization is called a TVD Runge-Kutta method if it preserves the
TVD property of the underlying space discretization for scalar conservation laws in
1D. Such time-stepping schemes were introduced by Shu and Osher [304, 306] and
analyzed by Gottlieb and Shu [125]. Other (non-TVD but linearly stable) RungeKutta methods can generate spurious oscillations even if the semi-discrete scheme
is local extremum diminishing and/or positive, see [125] for a numerical example.
Theorem 3.30. A Runge-Kutta method of the form (3.94)–(3.95) with
0 ≤ γkl ≤ 1,
l−1
∑ γkl = 1,
k=0
0 ≤ θkl ≤ 1
(3.96)
3.3 Matrix Analysis for Unsteady Problems
123
is positivity-preserving if the time step ∆ t satisfies condition (3.92) for θ = 0 [125].
Proof. Conditions (3.96) imply that the right-hand side of (3.94) is a convex combination of forward Euler predictors with ∆ t replaced by θkl ∆ t, where θkl ∈ [0, 1].
Therefore, positivity is preserved under condition (3.92) with θ = 0. Remark 3.18. The above proof of positivity is only valid for diagonal (lumped) mass
matrices. Otherwise, positive off-diagonal coefficients ai j = mi j violate (3.84).
In the review paper by Gottlieb, Shu, and Tadmor [126], explicit high-order timestepping schemes that comply with the requirements of Theorem 3.30 were renamed
into strong stability-preserving (SSP) time discretizations. This more suitable term
refers to the ability of SSP methods to maintain boundedness not only in the total
variation norm but also in other norms. If the forward Euler method is SSP, so are its
high-order counterparts, perhaps under a different restriction on the time step [126].
The optimal (in terms of the time step restriction and computational cost) SSP
Runge-Kutta scheme of second order is the well-known Heun method [125]
Mu(1) = Mun + ∆ tCn un ,
1
Mun+1 =
M(un + u(1) ) + ∆ tC(1) u(1) .
2
(3.97)
(3.98)
The final solution un+1 represents the average of the forward Euler predictor u(1)
and a backward Euler corrector evaluated using u(1) in place of un+1 .
The optimal SSP Runge-Kutta time discretization of third order is given by [125]
Mu(1) = Mun + ∆ tCn un ,
1
M(3un + u(1) ) + ∆ tC(1) u(1) ,
Mu(2) =
4
1
n+1
Mu
=
M(un + 2u(2) ) + 2∆ tC(2) u(2) .
3
(3.99)
(3.100)
(3.101)
For a comprehensive review and systematic study of SSP Runge-Kutta / multistep
methods that provide high accuracy and low storage, the reader is referred to [126,
305]. Aspects of positivity preservation and monotonicity concepts for numerical
integration of initial value problems are also addressed in [164], pp. 185–196.
Since SSP time-stepping methods are usually of the same form and incur approximately the same computational cost per time step as traditional ODE solvers, it is
worthwhile to use them whenever possible. Even if the corresponding theoretical
restriction on the time step is smaller than the linear stability bound, an SSP method
tends to be more stable in the range of time steps that lie in-between [305]. In many
cases, there is no penalty for taking time steps far beyond the SSP bound because
its derivation is usually based on sufficient (rather than necessary) conditions.
124
3 Maximum Principles
3.4 Summary
The abundance of theorems and proofs in the present chapter makes it very different from the rest of the book. The mastery of this material is not important for
programmers and users of CFD codes. However, the presented theory illustrates the
close relationship between the physical nature of transport processes and qualitative
properties of solutions to partial differential equations of different types. Moreover,
numerical approximations were shown to inherit these properties under certain restrictions on the coefficients of the algebraic systems to be solved. This knowledge
can contribute to the development of numerical methods for transport equations.
Since most analytical and numerical studies are restricted to a certain class of
problems (steady/unsteady, viscous/inviscid etc.), no unified theory of continuous
and discrete maximum principles seems to be available to date. This chapter was
written in an attempt to fill this gap and illustrate some striking similarities between
positivity and monotonicity constraints that have been known under different names
and developed independently by different groups of researchers. Another goal was
to retrace the basic steps involved in the discretization process and discuss the implications of each step in terms of maximum principles and positivity preservation. As
we have seen, a simple analysis of matrix properties provides a link to the governing
equation and valuable information about the properties of numerical solutions.
The theoretical framework presented in this chapter rests on a set of sufficient
conditions that rule out the onset of spurious undershoots or overshoots. To this
end, the left-hand side matrix is required to be (irreducibly or strictly) diagonally
dominant with nonpositive off-diagonal coefficients. Its diagonal entries and all coefficients of the explicit part are required to be nonnegative. These conditions are
known from classical texts on numerical linear algebra [339, 354] and coincide with
the requirements of Patankar’s basic rules [268] frequently referred to in this book.
As a useful byproduct, one obtains computable bounds for admissible time steps.
Some discretizations of transport equations are guaranteed to be monotone but
their accuracy is restricted by the order barriers that apply to linear approximations
of convective and diffusive terms. The only way to achieve higher accuracy while
maintaining monotonicity is to devise a smart feedback mechanism that extracts
information from the approximate solution and constrains the coefficients of the
numerical scheme on the basis of this information. This design principle leads to
the algebraic flux correction paradigm to be introduced in the next chapter.
Chapter 4
Algebraic Flux Correction
Multidimensional transport problems with interior/boundary layers or discontinuities represent a formidable challenge for numerical techniques, especially in the
case of unstructured meshes and implicit time-stepping schemes. As usual, the reason is the tradeoff between spurious oscillations and excessive numerical diffusion.
In the realm of finite elements, it is common practice to combat the latter evil by
adding some anisotropic artificial diffusion acting in the streamline direction only.
Ironically, the Galerkin discretization of diffusive terms may violate the discrete
maximum principle (DMP) instead of helping the convective part to satisfy it. This
multidimensional side effect is rarely taken into account. Moreover, conventional
stabilization techniques operate at the continuous level and involve free parameters
which are highly problem-dependent. It is difficult to achieve the M-matrix property
and maintain monotonicity in this fashion. This is why even stabilized finite element
methods with favorable theoretical properties tend to produce oscillatory results.
In this chapter, we revisit the algebraic constraints that guarantee the validity of
a DMP and enforce them in a mass-conserving way using a set of diffusive and
antidiffusive fluxes. After a brief presentation of the basic ideas, we introduce the
algebraic flux correction (AFC) paradigm which will serve as a general framework
for the design of multidimensional flux/slope limiters in an unstructured grid environment. We address the iterative treatment of nonlinear algebraic systems and the
optimal choice of the limiting strategy. Finally, we apply the developed tools to finite
element discretizations of elliptic, hyperbolic, and parabolic transport problems.
4.1 Nonlinear High-Resolution Schemes
The trend towards the use of unstructured mesh methodologies in general-purpose
CFD codes has stimulated a lot of research on fully multidimensional generalizations of classical high-resolution schemes for transport equations. While Godunovtype methods can be readily integrated into finite volume codes, there is no natural
extension to continuous (linear or bilinear) Galerkin discretizations. Similarly, the
125
126
4 Algebraic Flux Correction
development of finite element schemes based on algebraic flux limiting techniques
requires a major revision, or at least a new interpretation, of their finite difference
prototypes which are typically explicit and/or tailored for Cartesian meshes.
In the late 1980s and early 1990s, flux-corrected transport (FCT) and total variation diminishing (TVD) schemes were extended to explicit algorithms based on linear and bilinear Galerkin finite element discretizations [8, 232, 244, 266, 298, 299].
Conservative flux decompositions, edge-based data structures, and the equivalence
to finite volumes have made it possible to generalize many one-dimensional concepts, such as ‘upwind difference’ or ‘slope ratio’, in a rather straightforward way
[239, 243]. These remarkable advances have formed the basis for the development
of high-resolution finite element schemes for compressible CFD and aerodynamics.
However, they were met with little enthusiasm by the theoretically oriented fraction of the FEM community, perhaps, due to the lack of mathematical rigor and a
possible loss of the Galerkin orthogonality as a result of such ‘variational crimes.’
As of this writing, most finite element schemes for convection-dominated transport equations still rely on linear stabilization. For decades, the mainstream approach has been represented by the Streamline Upwind Petrov-Galerkin (SUPG)
and Galerkin Least Squares (GLS) methods [172]. A variety of ‘improvements’ and
‘optimal’ values of free parameters have been proposed. Furthermore, front capturing techniques have been devised for problems with interior and boundary layers
in an attempt to suppress spurious oscillations. In most cases, the involved stabilization mechanisms are designed at the continuous level using heuristic arguments
and ad hoc parameter fitting rather than a rigorous mathematical theory which ensures the validity of the discrete maximum principle. Therefore, undershoots and/or
overshoots are to be expected whenever the solution develops steep gradients [172].
Even if the ripples are relatively small, they may cause irrecoverable damage in situations when positivity preservation is a must for physical and numerical reasons.
In recent years, interior penalty (edge stabilization) techniques [51, 50, 293, 327]
have become increasingly popular with finite element practitioners. In this approach,
the amount of stabilization is proportional to the jumps of the gradient across interelement boundaries. The resulting solutions are not as sensitive to the choice of
free parameters as in the case of SUPG-like methods. The inclusion of a nonlinear
shock capturing term with a sufficiently large coefficient makes it possible to prove
the weak DMP property [52]. Edge stabilization appears to be a very promising
methodology but it is not a free lunch since the addition of jump terms leads to a
finite element discretization with a wider stencil and a different sparsity pattern.
The latest comparative studies of finite element methods for stationary and timedependent transport problems [174, 175] speak in favor of high-resolution schemes
based on the algebraic flux correction (AFC) paradigm [200, 203, 205, 206]. The
basic idea is very simple: if a given discretization fails to satisfy the sufficient conditions of the discrete maximum principle, they can be enforced by adding a discrete
diffusion operator that adjusts itself adaptively to the local solution behavior. This
design principle represents a ‘black-box’ approach to the construction of constrained
high-order discretizations, whereby all the necessary information is inferred from
the entries of a given matrix. Algebraic flux correction schemes can be equipped
4.1 Nonlinear High-Resolution Schemes
127
with symmetric or upwind-biased flux limiters which differ in the definition of local
extremum diminishing (LED) upper and lower bounds. It is also possible to constrain the local slopes edge-by-edge so as to limit the jumps of the gradient.
The marriage of implicit FEM-FCT schemes [191, 203, 205] and their multidimensional FEM-TVD counterparts [206] within the framework of algebraic flux
correction [192, 200] was followed by the investigation of many related concepts
and limiting techniques [194, 196, 204, 258]. As the methodology has evolved and
matured, the growing number of publications has made it difficult for the readers to
keep track of recent developments and choose the right algorithms. Therefore, the
time has come to review the state of the art, summarize the most important results,
and give some practical recommendations. This is the goal of the present chapter.
4.1.1 Design Philosophy and Tools
As a standard model problem, consider an unsteady conservation law of the form
∂u
+ ∇ · f = 0 in Ω .
∂t
(4.1)
For the linear convection-diffusion equation, the flux function f(u) is given by
f = vu − ε ∇u,
where v is a known velocity field and ε is a constant diffusion coefficient. The case
of a nonuniform diffusion tensor D(x,t) will be considered in Section 4.5.
The above problem is endowed with appropriate boundary conditions imposed
in a strong or weak sense. The initial condition is given by the formula
u(x, 0) = u0 (x),
∀x ∈ Ω .
Modern front-capturing methods for (4.1) constrain the coefficients of a highorder scheme so as to keep it as accurate as possible without generating undershoots
or overshoots. The basic ingredients of such nonlinear discretizations are [357]
1.
2.
3.
4.
physical or mathematical constraints that guarantee certain qualitative properties;
a stable high-order scheme that satisfies 1 only for sufficiently smooth solutions;
a monotone low-order scheme which is guaranteed to satisfy 1 for arbitrary data;
a simple mechanism for blending 2 and 3 so as to enforce 1 in an adaptive fashion.
Of course, it is also essential to maintain mass conservation. To this end, it is sufficient to express the differences between 2, 3, and 4 in terms of internodal fluxes.
In algebraic flux correction schemes, the constraints to be imposed (item 1) are
based on the theory of discrete maximum principles (DMP) and positivity preservation, as presented in Chapter 3. A discrete diffusion operator is employed to convert
an accurate high-order discretization (item 2) into a nonoscillatory low-order one
128
4 Algebraic Flux Correction
(item 3), after which a limited amount of compensating antidiffusion is applied in
order to prevent a global loss of accuracy (item 4). All of these manipulations are
performed at the discrete level using the set of sufficient conditions (3.83)–(3.84) to
enforce and maintain the M-matrix property which guarantees monotonicity. This
methodology represents a generalization of classical FCT and TVD schemes.
Consider equation (4.1) discretized in space by a centered finite difference, finite
volume, or finite element method (item 1) on a structured or unstructured mesh
∑ mi j
j
du j
= ∑ ki j u j ,
dt
j
(4.2)
where u j is a time-dependent nodal value, mi j is an entry of the mass matrix, and ki j
is an entry of the discrete transport operator (see Chapter 2). This approximation is
supposed to be conservative, linear, and more than first-order accurate.
Also, consider a low-order discretization (item 2) with a diagonal mass matrix
mi
dui
= ∑ li j u j .
dt
j
(4.3)
The job of (4.2) is to approximate smooth data with high precision, whereas (4.3)
must produce nonoscillatory solutions even in the presence of steep fronts.
By definition, a space discretization of equation (4.1) is positivity-preserving if
ui (0) ≥ 0,
∀i
⇒
ui (t) ≥ 0,
∀i, ∀t > 0.
A sufficient condition for (4.3) to possess this property is given by Theorem 3.25
mi > 0,
li j ≥ 0,
∀i, ∀ j 6= i.
(4.4)
If the coefficient matrix L = {li j } has zero row sums, then (4.3) is equivalent to
mi
dui
= ∑ li j (u j − ui ).
dt
j6=i
(4.5)
Such a semi-discrete scheme proves not only positivity-preserving but also local
extremum diminishing (LED) under conditions (4.4). Theorem 3.24 states that
ui ≥ u j ,
∀ j 6= i
⇒
dui
≤ 0,
dt
whence a maximum cannot increase. Likewise, a minimum cannot decrease since
ui ≤ u j ,
∀ j 6= i
⇒
dui
≥ 0.
dt
After the discretization in time by the standard θ −scheme or a suitable RungeKutta method, conditions (4.4) ensure the validity of a discrete maximum principle,
perhaps under additional restrictions on the time step size, see Section 3.3.3.
4.1 Nonlinear High-Resolution Schemes
129
At large Peclet numbers, both (4.2) and (4.3) may fail to resolve the solution
properly. Due to the Godunov theorem [123], a linear high-order discretization of
the form (4.2) cannot be positivity-preserving for arbitrary data. In fact, it tends to
produce spurious oscillations in the neighborhood of steep fronts. This problem can
be cured by adding some artificial diffusion. On the other hand, the accuracy of a
linear low-order discretization like (4.3) can be enhanced by removing excessive
numerical diffusion. As long as the diffusive and antidiffusive terms admit a conservative flux decomposition, they do not affect the global mass balance but make it
possible to improve the distribution of mass and satisfy the discrete maximum principle. The algebraic flux correction methodology to be presented below provides a
general approach to finding the right amount of artificial diffusion and antidiffusion.
4.1.2 Artificial Diffusion Operators
For the time being, we assume that the mesh is sufficiently regular for the discretization of the diffusive term to satisfy the LED criterion. Under this assumption, undershoots and overshoots are caused by the contributions of the convective term and/or
of a nondiagonal mass matrix. In the case of finite difference and finite volume
approximations, the use of first-order upwinding leads to the least diffusive linear
positivity-preserving scheme. For linear and bilinear finite element discretizations,
the same effect can be achieved by adding a discrete diffusion operator [191, 205].
The system of equations (4.2) for a FEM discretization of (4.1) can be written as
MC
du
= Ku,
dt
(4.6)
where u is the vector of time-dependent nodal values, MC = {mi j } is the consistent
mass matrix, and K = {ki j } is (the negative of) the discrete transport operator. The
coefficients of MC and K are defined and calculated as explained in Chapter 2.
Remark 4.1. The skew-symmetric part 12 (K − K T ) is associated with a centered discretization of −v · ∇, whereas 21 (K + K T ) − diag{K} is a discrete (anti-)diffusion
operator. The latter may include streamline diffusion used for stabilization purposes
and/or to achieve better phase accuracy, as in the case of Taylor-Galerkin methods.
A scheme of the form (4.6) is neither LED nor positivity-preserving as long as
∃ mi j 6= 0,
∃ ki j < 0,
j 6= i.
In order to enforce the discrete maximum principle for (4.6) it is sufficient to
• perform row-sum mass lumping and replace the consistent mass matrix MC by
ML = diag{mi },
mi = ∑ mi j ,
j
(4.7)
130
4 Algebraic Flux Correction
• approximate the discrete transport operator K by its low-order counterpart
L = K + D,
li j ≥ 0,
∀ j 6= i,
(4.8)
where D = {di j } stands for an artificial diffusion operator such that
∑ di j = ∑ di j = 0,
j
di j = d ji ,
i
∀i, j.
(4.9)
These manipulations lead to a linear positivity-preserving scheme of the form
ML
du
= Lu.
dt
(4.10)
For every pair of nonzero off-diagonal entries ki j and k ji , the artificial diffusion
coefficient di j = d ji should ensure that li j = ki j + di j ≥ 0 and l ji = k ji + di j ≥ 0, as
required by (4.8). Therefore, the lower bound for di j is [170, 191, 205, 301]
di j = max{−ki j , 0, −k ji } = d ji ,
∀ j 6= i.
(4.11)
This is just enough to eliminate all negative off-diagonal entries of the high-order
operator K. Artificial diffusion coefficients that enforce positivity in this way were
used to construct low-order schemes for FCT as early as in the mid-1970s [43].
For the row sums of D = {di j } to be zero, its diagonal entries are defined as
dii := − ∑ di j ,
j6=i
∀i.
(4.12)
By construction, di j = d ji for all i and j. Hence, the resulting matrix satisfies (4.9)
and is a representative of discrete diffusion operators defined in Section 2.1.6.2.
Remark 4.2. If the matrix K is skew-symmetric, as in the case of ki j = −k ji given
by (2.54), then the diffusion coefficient (4.11) reduces to di j = |ki j | for all j 6= i.
Remark 4.3. In nonlinear inviscid flow problems, it might happen that ki j < 0 and,
consequently, li j = 0 for some i and all j 6= i. If the matrix K has zero row sums, then
lii = 0 as well, which implies that L is reducible and actually singular. The contribution of the mass matrix or ‘relaxation by inertia’ (see next section) renders the fully
discrete problem well-posed but the numerical solution may exhibit spurious kinks.
These artifacts typically occur at stagnation points, where the velocity reverses its
sign creating an internal ‘inlet’ with unspecified ‘inflow’ values. A common remedy
is to replace the smallest artificial diffusion coefficient (4.11) by [139, 235]
(
di j ,
if di j ≥ δ ,
di j := di2j +δ 2
∀ j 6= i,
2δ , if di j ≤ δ ,
where δ > 0 is a small parameter that does not allow di j to vanish and produce a row
of zero entries. In gas dynamics, this trick is known as the entropy fix. The threshold
δ may be taken constant or designed to be a function of the local flow conditions.
4.1 Nonlinear High-Resolution Schemes
131
Remark 4.4. Physical diffusion can be built into the matrices K and L before or after
the computation of D. In the former case, the value of the artificial diffusion coefficient di j given by (4.11) is reduced accordingly. This may or may not be desirable.
For example, the negative numerical diffusion inherent to the standard Galerkin discretization of convective terms would offset some physical diffusion and result in
artificial steepening of solution profiles. On the other hand, if the high-order scheme
contains some background dissipation and its leading truncation error is of a diffusive nature, then it is worthwhile to minimize the amount of numerical diffusion.
In practice, the elimination of negative off-diagonal entries is performed step-bystep without assembling the global matrix D. Instead, artificial diffusion can be built
into the operator K in a loop over edges of the sparsity graph, see Section 2.1.8. By
definition, each edge is a pair of nodes {i, j} that corresponds to a pair of nonzero
off-diagonal coefficients ki j and k ji . The required solution update is as follows
kii := kii − di j ,
k ji := k ji + di j ,
ki j := ki j + di j ,
k j j := k j j − di j .
(4.13)
Without loss of generality, the edges of the sparsity graph are oriented so that
ki j ≤ k ji .
(4.14)
This orientation convention implies that node i is located ‘upwind’ and corresponds
to the row number of the negative off-diagonal entry to be eliminated (if any).
Equation (4.12) implies that the diagonal-entries of L = K + D are given by
lii := kii − ∑ di j .
(4.15)
j6=i
Furthermore, it can readily be seen that the row sums of K and L = K + D are equal
∑ li j = ∑(ki j + di j ) = ∑ ki j .
j
j
j
The ordinary differential equation for the nodal value ui can be represented as
mi
dui
= ∑ li j (u j − ui ) + ui ∑ ki j .
dt
j
j6=i
(4.16)
If K has zero row sums, then the low-order scheme is local extremum diminishing
mi
dui
= ∑ li j (u j − ui ),
dt
j6=i
li j ≥ 0,
∀ j 6= i.
Otherwise, the second term in the right-hand side of (4.16) represents a nonvanishing discrete counterpart of −u∇ · v which is responsible for compressibility effects.
As in the continuous case, convective transport by a nonuniform velocity field may
132
4 Algebraic Flux Correction
concentrate the mass in certain regions or create zones of low concentration. Then
the low-order discretization (4.16) is no longer LED but still positivity-preserving.
Example 4.1. To clarify the implications of (4.13), consider the 1D model problem
∂u
∂u
+v
= 0,
∂t
∂x
(4.17)
where the velocity v is constant and strictly positive. The computational domain is
Ω = (0, 1) and a Dirichlet boundary condition u(0) = g is prescribed at the inlet.
On a uniform mesh of linear finite elements, the standard Galerkin method yields
K=

...
v
1


2
0 −v
v 0 −v
v 0 −v

...


.

Note that this skew-symmetric tridiagonal matrix has zero row sums. The diagonal
entries and column sums, except for the first and last one, are also equal to zero.
For any interior node, mi = ∆ x, where ∆ x is is the constant mesh size. Hence, the
lumped-mass version of (4.2) is equivalent to the central difference scheme
dui
ui+1 − ui−1
+v
= 0.
dt
2∆x
Since ki j = − 2v for j = i + 1, the artificial diffusion coefficient (4.11) is di j =



L=

...
v −v 0
v −v 0
v −v
v
2
and

0
...


,

which corresponds to the first-order accurate upwind difference approximation
dui
ui − ui−1
+v
= 0.
dt
∆x
Thus, the elimination of negative off-diagonal entries from a skew-symmetric operator K can be interpreted as discrete upwinding [200]. For any pair of nodes i and
j = i + 1 numbered in accordance with (4.14), the grid point xi lies upstream of x j .
After the discretization in time by the standard θ −scheme, the upwind method
is stable and positive under the CFL-like condition (3.92) which reduces to
v
1
∆t
≤
,
∆x 1−θ
0 ≤ θ < 1.
(4.18)
4.1 Nonlinear High-Resolution Schemes
133
Of course, there is no time step restriction for the backward Euler method (θ = 1)
which corresponds to ‘upwinding in time’ and is only first-order accurate.
4.1.3 Conservative Flux Decomposition
The replacement of the high-order discretization (4.6) by the low-order one (4.10)
ensures positivity preservation but introduces a first-order perturbation error which
manifests itself in strong smearing effects. The next step towards the construction
of an algebraic flux correction scheme involves a decomposition of this error into
internodal fluxes which can be used to restore high accuracy in regions where the
solution is well resolved and no modifications of the original scheme are required.
By construction, the difference between the residuals of(4.6) and (4.10) is
f = (ML − MC )
du
− Du.
dt
(4.19)
The zero row sum property of the artificial diffusion operator D implies that its
contribution to the i−th component of the vector f can be written in the form
(Du)i = ∑ di j u j = ∑ di j (u j − ui )
j
(4.20)
j6=i
and looks similar to the right-hand side of a local extremum diminishing scheme.
The error due to row-sum mass lumping can be decomposed in a similar way
(MC u − ML u)i = ∑ mi j u j − mi ui = ∑ mi j (u j − ui ).
j
(4.21)
j6=i
Due to (4.20)–(4.21) and symmetry, the total error (4.19) induced by mass lumping
and artificial diffusion admits a conservative decomposition into internodal fluxes
fi = ∑ fi j ,
j6=i
f ji = − fi j .
(4.22)
The amount of mass transported by the raw antidiffusive flux fi j is given by
d
fi j = mi j + di j (ui − u j ),
∀ j 6= i.
(4.23)
dt
Every pair of fluxes fi j and f ji can be associated with an edge of the sparsity graph
which represents a pair of nonzero off-diagonal entries with indices i and j.
Remark 4.5. After the discretization in time, the derivative with respect to t is replaced by a finite difference. In steady-state problems, its contribution is zero.
Giving the flux fi j to node i and f ji = − fi j to its neighbor j does not create or destroy
mass. The addition of raw antidiffusive fluxes (4.23) to the right-hand side of (4.16)
134
4 Algebraic Flux Correction
removes the error induced by the row-sum mass lumping and artificial diffusion
mi
dui
=
dt
∑ li j (u j − ui ) + ui ∑ ki j
j6=i
j
dui du j
−
− ∑ di j (u j − ui ) − ∑ mi j
dt
dt
j6=i
j6=i
dui du j
.
−
= ∑ ki j u j − ∑ mi j
dt
dt
j
j6=i
Moving all time derivatives into the left-hand side, one obtains an equation of the
form (4.2) which corresponds to the original high-order discretization (4.6).
4.1.4 Limited Antidiffusive Correction
Some of the raw antidiffusive fluxes fi j are harmless but others are responsible for
the violation of the positivity constraint by the high-order scheme. Such fluxes need
to be canceled or limited so as to keep the scheme positivity-preserving. In the
process of flux correction, every antidiffusive flux fi j is multiplied by a solutiondependent correction factor αi j ∈ [0, 1] before it is inserted into the equation
mi
dui
= ∑ li j u j + f¯i ,
dt
j
f¯i = ∑ αi j fi j .
(4.24)
j6=i
Of course, the fluxes fi j and f ji = − fi j must be limited using the same correction
factor αi j = α ji to maintain skew-symmetry and, hence, conservation of mass.
By construction, the high-order discretization (4.6) and its low-order counterpart
(4.10) are recovered for αi j = 1 and αi j = 0, respectively. The former setting is
usually acceptable in regions where the solution is smooth and well-resolved. However, the magnitude of antidiffusive fluxes may need to be reduced elsewhere, so
as to prevent the formation of undershoots or overshoots. As a rule of thumb, the
solution-dependent correction factors αi j should be chosen as close to 1 as possible
without violating the positivity constraint. Since discretization (4.24) is nonlinear in
the choice of αi j , it has the potential of being more than first-order accurate.
The flux-corrected semi-discrete scheme (4.24) can be written in the matrix form
M̄C
du
= K̄u.
dt
(4.25)
The coefficients of the partially lumped mass matrix M̄C = {m̄i j } are given by
m̄ii := mi − ∑ m̄i j ,
j6=i
m̄i j = αi j mi j ,
∀ j 6= i,
(4.26)
4.1 Nonlinear High-Resolution Schemes
135
while the structure of the nonlinear transport operator K̄ = {k̄i j } is as follows
k̄ii := lii + ∑ αi j di j ,
k̄i j = li j − αi j di j ,
j6=i
∀ j 6= i.
(4.27)
Suppose that the solution to (4.25) satisfies an equivalent nonlinear ODE system
ML
du
= L̄u,
dt
(4.28)
where L̄ = {l¯i j } has no negative off-diagonal coefficients and is defined so that
L̄u = Lu + f¯.
(4.29)
Such a space discretization is positivity-preserving and so is (4.25) because it was
assumed to have the same solution. After the discretization in time, additional restrictions may apply to the time step for an explicit or semi-implicit algorithm.
To ensure the existence of representation (4.28) with l¯i j ≥ 0 for all j 6= i, it is
sufficient to find αi j such that the sum of antidiffusive fluxes is constrained by
+
Q−
i ≤ ∑ αi j f i j ≤ Qi ,
(4.30)
j6=i
where Q±
i are local extremum diminishing upper and lower bounds of the form
Q+
i = ∑ qi j max{0, u j − ui },
(4.31)
Q−
i = ∑ qi j min{0, u j − ui }.
(4.32)
j6=i
j6=i
The coefficients qi j must be nonnegative for all j 6= i. The best choice of these
parameters depends on the problem at hand and on the limiting strategy (see below).
When appropriate values of qi j have been fixed, it remains to define the correction
factors αi j = α ji . It is always possible to satisfy (4.30) by setting αi j = 0 but a
properly designed flux limiter returns αi j ≈ 1 if the fluxes fi j and f ji are harmless.
Due to (4.30)–(4.32), there exists a matrix Q̄ = {q̄i j } of nonlinear coefficients
q̄ii := − ∑ q̄i j ,
j6=i
0 ≤ q̄i j ≤ qi j ,
∀ j 6= i
(4.33)
such that the sum of limited antidiffusive fluxes can be expressed in the LED form
f¯i = ∑ q̄i j (u j − ui ),
j6=i
q̄i j ≥ 0,
∀ j 6= i.
(4.34)
In essence, this representation guarantees that the term f¯i is equivalent to a sum
of ‘diffusive’ and, therefore, acceptable edge contributions. Substitution into (4.29)
yields L̄ = L + Q̄, which proves positivity preservation at the semi-discrete level.
136
4 Algebraic Flux Correction
The beauty and generality of the above approach to flux limiting lie in the flexible choice of the parameters qi j that define the upper and lower bounds. Any set
of nonnegative bounded values (0 ≤ qi j < ∞, ∀ j 6= i) is acceptable from the viewpoint of positivity preservation at the semi-discrete level. Hence, the definition of
these parameters is dictated by accuracy and efficiency considerations. To get close
enough to the high-order solution, the magnitude of qi j should be sufficiently large.
On the other hand, it cannot be chosen arbitrarily large for the following reasons:
• In explicit algorithms, the CFL-like positivity condition (3.92) for the largest
admissible time step depends on the sum of qi j and may become too restrictive.
• In implicit schemes and steady-state solvers, the nonlinear antidiffusive term is
updated in an iterative way (see Section 4.2). Inordinately large values of qi j may
cause severe convergence problems and should be avoided, especially if intermediate solutions may fail to conserve mass and/or to stay positivity-preserving.
A number of ways to define qi j will be discussed in Sections 4.3–4.5. As we will
see, (quasi)-stationary and time-dependent transport problems may require different
treatments. For the time being, let us keep the values of qi j unspecified and present
a general approach to the practical computation of the correction factors αi j .
4.1.5 The Generic Limiting Strategy
In the process of flux correction, each node i may receive both positive and negative
antidiffusive fluxes from its neighbors. Although some fluxes may cancel out, the
formula for αi j should be failsafe even in the worst-case scenario. Given a vector of
nodal values u and a set of nonnegative coefficients qi j , the sum of positive fluxes is
required to be smaller than the upper bound (4.31), while the sum of negative ones
may not fall below the lower bound (4.32). Hence, the admissible portion of a raw
antidiffusive flux depends on its sign [355]. In Sections 4.3–4.4, we will consider
algebraic flux correction schemes based on the following generic algorithm [192]
1. Compute the sums of positive and negative antidiffusive fluxes to be limited
Pi+ = ∑ max{0, fi j },
j6=i
Pi− = ∑ min{0, fi j }.
(4.35)
j6=i
2. Generate local extremum diminishing upper and lower bounds of the form
Q+
i = ∑ qi j max{0, u j − ui },
j6=i
Q−
i = ∑ qi j min{0, u j − ui }.
(4.36)
j6=i
3. Evaluate the nodal correction factors for the positive and negative part
Q−
Q+
−
+
i
i
Ri = min 1, − .
Ri = min 1, + ,
Pi
Pi
(4.37)
4.1 Nonlinear High-Resolution Schemes
137
4. Perform flux limiting using edge-based correction factors αi j such that
+
Ri , if fi j > 0,
αi j ≤
α ji = αi j .
R−
i , if f i j < 0,
(4.38)
Definition (4.38) guarantees that an upper bound of the form (4.30) holds for the
sum of limited antidiffusive fluxes into node i. Indeed, it is easy to verify that
− −
+
+ +
Q−
i ≤ Ri Pi ≤ ∑ αi j f i j ≤ Ri Pi ≤ Qi .
j6=i
Remark 4.6. This multidimensional limiting strategy traces its origins to Zalesak’s
algorithm [355, 357]. Although classical (two-step) FCT methods do not fit into this
framework, we adopt the same notation to highlight the existing similarities.
Remark 4.7. The upper/lower bounds Q±
i may consist of a single term associated
with a local maximum umax
or
minimum
umin
taken over the stencil of node i
i
i
+ max
Q+
− ui ),
i = qi (ui
− min
Q−
− ui ).
i = qi (ui
(4.39)
The nonnegative coefficients q±
i can be defined as the sums of qi j ≥ 0 for all j 6= i.
It remains to give a formal definition of αi j ≤ R±
i and j 6= i. This definition
depends on whether an upwind-biased or a symmetric limiting strategy is capable of
producing larger values of αi j . Here and below, the terms ‘upwind’ and ‘downwind’
refer to the order of nodes i and j under convention (4.14) on edge orientation.
• Symmetric flux limiters do not distinguish between upwind and downwind
nodes, treating them equally. For each pair of nodes {i, j}, the raw antidiffusive
flux fi j is added to the sum Pi± and subtracted from the sum Pj∓ . Flux correction
is performed using the minimum of nodal correction factors for i and j
∓
αi j = min{R±
i , R j },
α ji = αi j .
• Upwind-biased flux limiters take advantage of the fact that (a large portion of)
the raw antidiffusive flux f ji into a downwind node j is compensated by the
diffusive edge contribution l ji (ui − u j ), where l ji > 0. In this approach, only the
upwind sum Pi± is incremented, and the correction factors are defined as
αi j = R±
i ,
α ji := αi j .
• General-purpose flux limiters compare the magnitude of the raw antidiffusive
flux f ji to that of the diffusive edge contribution l ji (ui − u j ). On the basis of this
comparison, an upwind-biased or a symmetric limiter is invoked, cf. [192].
Remark 4.8. Instead of constraining the sums of antidiffusive fluxes, it is possible to
adjust the local slopes ui − u j in a stand-alone fashion, as in the 1D case. Then the
±
quantities Pi+ , bounds Q±
i , and correction factors Ri are defined edge-by-edge.
138
4 Algebraic Flux Correction
4.1.6 Summary of Algorithmic Steps
Let us summarize the basic steps involved in the derivation of an algebraic flux
correction scheme. The starting point was a linear high-order discretization
MC
du
= Ku,
dt
∃ j 6= i :
mi j 6= 0,
ki j < 0.
(4.40)
To achieve the desired matrix properties, we performed row-sum mass lumping and
applied an artificial diffusion operator D designed so as to eliminate all negative offdiagonal coefficients of K. These manipulations have led us to a low-order counterpart of (4.40) which is the least diffusive among linear positivity-preserving ones
ML
du
= Lu,
dt
L = K + D,
li j ≥ 0,
∀ j 6= i.
(4.41)
Finally, we removed excessive artificial diffusion using a set of internodal fluxes fi j
multiplied by solution-dependent correction factors αi j ∈ [0, 1]. The end product is
a nonlinear blend of (4.40) and (4.41) with solution-dependent matrix entries
M̄C
du
= K̄u,
dt
∃ j 6= i :
m̄i j 6= 0,
k̄i j < 0.
(4.42)
Note that the limited antidiffusive correction reintroduces some off-diagonal entries
of wrong sign. However, there exists a matrix Q̄ of coefficients given by (4.33) such
that problem (4.42) has the same solution as the equivalent nonlinear system
ML
du
= L̄u,
dt
L̄ = L + Q̄,
l¯i j ≥ 0,
∀ j 6= i.
(4.43)
In Sections 4.3–4.5, we will present a number of multidimensional flux limiters
which guarantee the existence of the matrix L̄ without constructing it explicitly.
The link between representations (4.42) and (4.43) is given by the formula
ML
du
= Lu + f¯,
dt
f¯i = ∑ αi j fi j = ∑ q̄i j (u j − ui ).
j6=i
(4.44)
j6=i
Conservation and positivity of the flux-corrected scheme follow from the fact that
α ji = αi j ,
f ji = − fi j ,
q̄i j ≥ 0,
∀ j 6= i.
(4.45)
Up to now, we have explored the principles of algebraic flux correction in a rather
abstract setting, so as develop a general framework in which to work. An in-depth
presentation of some upwind-biased and symmetric flux limiters will follow in Sections 4.3–4.5. The main objective of this chapter is not to promote any particular
algorithm but to retrace the steps involved in the design process and explain their
ramifications. It is hoped that this background information will enable the interested
reader to develop new high-resolution schemes building on similar concepts.
4.2 Solution of Nonlinear Systems
139
4.2 Solution of Nonlinear Systems
The implementation of flux correction in an unstructured mesh code requires further considerations regarding the choice of time-stepping schemes and/or iterative
solution methods. In explicit algorithms, the limited antidiffusive term resides in the
right-hand side and can be readily evaluated using the solution from the previous
time step. The implementation of such a scheme is rather straightforward and will
not be discussed here. An implicit time discretization makes it possible to operate
with larger time steps but the correction factors αi j must be calculated in an iterative
way, even in the case of a linear transport equation. The repeated solution of linear
subproblems followed by an update of αi j is likely to pay off only if robust and
efficient solution algorithms are available. In this section, we address the numerical
treatment of nonlinear algebraic systems in implicit flux correction schemes.
4.2.1 Successive Approximations
After the time discretization by an implicit θ −scheme, equation (4.44) becomes
An+1 un+1 = Bn un + ∆ t f¯(un+1 , un ).
(4.46)
The matrices An+1 and Bn represent the contribution of the low-order part. The entries of these matrices depend on the time step ∆ t and on the parameter θ ∈ (0, 1]
An+1 = ML − θ ∆ tLn+1 ,
(4.47)
B = ML + (1 − θ )∆ tL .
n
n
(4.48)
The settings θ = 21 and θ = 1 correspond to the Crank-Nicolson and backward Euler methods, respectively. The latter is first-order accurate but offers unconditional
stability and positivity preservation. The former is second-order accurate and unconditionally stable but positive only under a CFL-like condition of the form (3.92).
In the case of a nonlinear governing equation and/or a nonstationary velocity
field, the entries of An+1 and Bn need to be updated as the solution and time evolve.
Furthermore, the antidiffusive term f¯(un+1 , un ) depends on un+1 in a nonlinear way,
so an iterative approach to the solution of the algebraic system (4.46) is required.
Consider a sequence of successive approximations {u(m) } to u = un+1 . At the
beginning of the first outer iteration, the value of u(0) is guessed making use of
the previously computed data and initial/boundary conditions. A reasonable initial
guess for an unsteady transport problem is u(0) = un or u(0) = 2un − un−1 . These
settings correspond to the constant and linear extrapolation in time, respectively.
Given an approximation u(m) , its successor u(m+1) can be calculated as follows
A(m) u(m+1) = Bn un + ∆ t f¯(u(m) , un ),
m = 0, 1, . . .
(4.49)
140
4 Algebraic Flux Correction
The sum of limited antidiffusive fluxes remains in the right-hand side and is evaluated using the current iterate u(m) in place of un+1 . If the coefficients of the matrix
A(m) = ML − θ ∆ tL(m) depend on the solution, they also needs to be recalculated.
This straightforward solution procedure is referred to as fixed-point iteration, also
known as the Picard iteration and successive approximation (substitution).
A convenient way to enforce Dirichlet boundary conditions for a given node i is
to replace the corresponding row of A(m) by that of the identity matrix
aii := 1,
∀ j 6= i
ai j := 0,
(4.50)
and substitute the boundary value gi for the i-th element of the right-hand side [322].
In most cases, linear systems (4.49) are also solved iteratively. Since the loworder operator A(m) was designed to meet the requirements of Corollary 3.12, it
proves to be an M-matrix. This favorable property ensures that a small number of
inner iterations are typically sufficient for practical purposes. The number of outer
iteration cycles (4.49) depends on the evolution of the residual (alias defect)
r(m) = Bn un − A(m) u(m) + ∆ t f¯(u(m) , un ).
(4.51)
All elements of r(m) associated with Dirichlet boundary nodes should be set to zero.
The Euclidean (or maximum) norm ||r(m) || of the residual is a good indicator of
how close the current approximation u(m) is to the final solution un+1 . A typical set
of stopping criteria for an iterative solution procedure like (4.49) is as follows
||r(m+1) || ≤ min{ε1 , ε2 ||r(0) ||},
||u(m+1) − u(m) ||
< ε3 ,
||u(m+1) ||
(4.52)
where ε1 , ε2 , and ε3 are the tolerances prescribed for defects and relative changes.
Suppose that M outer iterations are required to meet the above stopping criteria
and obtain un+1 := u(M) . Then the effective rate of convergence can be defined as
ρM =
||r(M) ||
||r(0) ||
!1
M
.
In essence, this is the average factor by which the magnitude of the defect shrinks
during a single outer iteration. Note that ρM < 1 is required for convergence.
The above methodology is also suitable for computing stationary solutions with
ML = 0,
θ = 1,
∆ t = 1.
Alternatively, the solution can be marched to the steady state by the unconditionally positive backward Euler method with large and, possibly, variable pseudo-time
steps. The relative solution changes and/or convergence rates can be used to determine the optimal time step size adaptively. When the solution begins to approach
the steady state, the removal of the mass matrix can greatly speed up convergence,
while removing it too soon can have the opposite effect [307].
4.2 Solution of Nonlinear Systems
141
In steady-state computations, it is advisable to switch off the nonlinear antidiffusive term f¯ at the startup stage since the low-order solution is usually inexpensive
to compute and closer to the flux-corrected one than an arbitrary initial guess u(0) .
4.2.2 Defect Correction Schemes
Fixed-point iteration is one of the simplest techniques for solving nonlinear algebraic equations. In many cases, the computational cost can be drastically reduced
by using another iterative solver tailored to the properties of the continuous problem and/or of the employed discretization techniques. In addition, special measures
may need to be taken to secure convergence to a steady state or the ability to operate
with large time steps. Hence, it is worthwhile to consider a general class of iterative
solution techniques to which the basic fixed-point iteration (4.49) belongs.
Many iterative solution methods for system (4.46) can be formally written in the
following generic form which will be referred to as the defect correction scheme
u(m+1) = u(m) + [Ā(m) ]−1 r(m) ,
m = 0, 1, 2 . . . ,
(4.53)
where Ā(m) is a suitable ‘preconditioner’ and r(m) is the residual given by (4.51). As
before, the Dirichlet boundary conditions for node i are implemented by setting
āii := 1,
(m)
ui
:= gi ,
āi j := 0,
(m)
ri
∀ j 6= i,
:= 0.
In practice, the ‘inversion’ of Ā(m) is performed by solving the linear subproblem
Ā(m) ∆ u(m+1) = r(m) ,
m = 0, 1, 2, . . .
(4.54)
After a few inner iterations, the increment ∆ u(m+1) is applied to the last iterate
u(m+1) = u(m) + ∆ u(m+1) .
(4.55)
The iteration process is terminated when the defect r(m) and the relative solution
changes ∆ u(m+1) become sufficiently small in the sense of criteria (4.52).
The implementation of a defect correction cycle involves the following tasks:
•
•
•
•
Assembly of Ā(m) and r(m) , imposition of Dirichlet boundary conditions.
Monitoring the residual norms, relative changes, and convergence rates.
Implicit computation of the solution increments ∆ u(m+1) from (4.54).
Explicit computation of the new approximation u(m+1) from (4.55).
Ideally, the preconditioner Ā(m) should be designed so that (i) matrix assembly is
relatively fast, (ii) linear subproblems (4.54) can be solved efficiently, and (iii) convergence is achieved with a small number of outer iterations. These requirements
are often in conflict with one another, so some compromises need to be made.
142
4 Algebraic Flux Correction
If the time step ∆ t is very small, the approximate solution can be updated in a
fully explicit fashion using (4.53) with a Jacobi-like diagonal preconditioner
Ā = diag{mi − θ ∆ t lii }.
(4.56)
The number of outer iterations for (4.53) preconditioned in this way can be as small
as 1 if a good initial guess is available. Thus, the computational cost per time step
might be comparable to that for a conditionally stable explicit algorithm.
At intermediate and large time steps, the default preconditioner for (4.53) is
Ā = ML − θ ∆ tL.
(4.57)
Substitution of (4.51) and (4.57) into (4.53) reveals that the resulting defect correction scheme is equivalent to the standard fixed-point iteration given by (4.49).
Furthermore, particularly severe nonlinearities can be handled using Newton-like
methods in which Ā is a suitable approximation to the Jacobian matrix
∂ ri
.
(4.58)
Ā ≈ −
∂uj
This definition requires (numerical) differentiation of the residual r(u) with respect
to each element of the solution vector u. In the context of algebraic flux correction,
the matrix J(u) can be approximated by divided differences and assembled edgeby-edge, as proposed by Möller [254]. If properly configured, the resulting discrete
Newton method converges much faster than the defect correction scheme (4.53)
preconditioned by (4.57). For implementation details we refer to [254, 255, 258].
The optimal choice of iterative solvers for sparse linear systems (4.54) with a
nondiagonal and nonsymmetric matrix Ā(m) also depends on the time step size and
on the matrix properties. As long as ∆ t is relatively small, the preconditioner is dominated by the lumped mass matrix and a basic iteration of Jacobi or Gauß-Seidel
(SOR) type may suffice. In this case, each inner iteration has approximately the
same cost as one step of an explicit algorithm. As the time step and the condition
number increase, BiCGSTAB, GMRES, and multigrid methods are to be preferred.
The Gauß-Seidel iteration and ILU with Cuthill-McKee renumbering can serve as
smoothers/preconditioners for intermediate and large time steps, respectively. Due
to the M-matrix property of the low-order operator (4.57), its ILU factorization exists and is unique [249]. It can be used as a preconditioner for inner iterations even if
the actual matrix to be ‘inverted’ is the Jacobian (4.58) for Newton’s method [255].
4.2.3 Underrelaxation and Smoothing
In many cases, it is desirable to reduce the changes between two successive iterates so as to stabilize the solution and make it converge ‘slowly but surely.’ This
can be accomplished by limiting the increments ∆ u(m+1) computed in (4.54) before
4.2 Solution of Nonlinear Systems
143
applying them to the old iterate u(m) . To this end, formula (4.55) is replaced by
(m+1)
ui
(m)
= ui
(m)
+ ωi
(m+1)
∆ ui
,
(4.59)
(m)
where 0 < ωi ≤ 1 for all i. This is a classical example of explicit underrelaxation
which amounts to combining the latest and previous nodal values [104, 106, 268].
The relaxation is said to be ‘heavy’ if a larger weight is given to the latter, that is, if
(m)
(m)
ωi < 0.5. In other cases, ‘light’ underrelaxation with ωi ≥ 0.5 is appropriate.
Unfortunately, there are no general rules for the choice of relaxation factors. They
can be assigned a fixed value (e.g., ω = 0.8) or chosen adaptively so as to control
the evolution of residuals [173] or minimize the error in an appropriate norm [322].
Typically, a range of admissible values is specified for a variable relaxation factor
(m)
0 < ω min ≤ ωi
≤ ω max ≤ 1.
Alternatively, implicit underrelaxation can be performed to make a given precondi(m)
tioner Ā(m) more diagonally dominant [104, 268]. For example, let γi ≥ 1 and
(m)
(m) (m)
āii .
āii := γi
(4.60)
The same effect can be achieved by means of ‘relaxation through inertia’ [268],
(m)
whereby a nonnegative number σi ≥ 0 is added to each diagonal entry
(m)
(m)
(m)
āii := āii + σi
.
(4.61)
Obviously, the additive version is related to the multiplicative one by the formula
(m)
σi
(m)
= ( γi
(m)
− 1)aii .
In steady-state computations, implicit underrelaxation is equivalent to the use of
variable pseudo-time steps [104, 268]. Conversely, local time-stepping can be in(m)
(m)
terpreted as underrelaxation. The optimal values of γi and σi are problemdependent and may change in the course of simulation. It is advisable to perform
stronger underrelaxation at the startup stage and gradually adjust relaxation factors
so as to control the residuals, relative solution changes, and convergence rates.
Another way to accelerate the defect correction scheme is known as residual
smoothing. If the residual r(m) exhibits an oscillatory behavior that hampers convergence, it is worthwhile to replace it by a smooth approximation r̄(m) . For instance,
the latter can be constructed by introducing some implicit mass diffusion [301]
(m)
mi r̄i
+ ∑ ωi j mi j (r̄i
(m)
j6=i
(m)
(m)
− r̄ j ) = mi ri ,
(4.62)
where ωi j is a positive weight and mi j ≥ 0 is an entry of the consistent mass matrix.
144
4 Algebraic Flux Correction
In practice, r̄(m) is obtained with a few sweeps of the Jacobi iteration [301]
!
mi + ∑ ωi j mi j r̄i
(m,l+1)
(m)
= mi ri
+ ∑ ωi j mi j r̄ j
(m,l)
.
j6=i
j6=i
Remark 4.9. A converged solution does not depend on how it was computed. However, underrelaxation and residual smoothing may render intermediate results nonconservative. To avoid stopping outer iterations too soon, it is worthwhile to check
the total mass if the tolerances for residuals and relative changes are rather slack.
4.2.4 Positivity-Preserving Solvers
The use of limiters in unstructured grid methods for steady-state problems is frequently associated with severe convergence problems [342]. Sometimes the nonlinearity is so strong that even heavy underrelaxation and/or residual smoothing are
of little help. The residuals decrease steadily until the initial error is reduced by
several orders of magnitude, after which convergence stalls. Although the discrete
maximum principle would hold for the fully converged solution, the linearization inherent to an iterative solver may give rise to undershoots/overshoots. Conversely, the
oscillatory solution behavior may be responsible for the lack of convergence. Therefore, it is worthwhile to start with a monotone low-order solution and configure the
defect correction scheme so as to ensure that each step is positivity-preserving.
At steady state, the residual of a scalar transport equation discretized by an algebraic flux correction scheme reduces to (4.29) which corresponds to
r(m) = L(m) u(m) + Q̄(m) u(m) = L(m) u(m) + f¯(m) ,
(4.63)
where Q̄ = {q̄i j } is a matrix with zero row sums and nonnegative off-diagonal entries given by (4.33). By definition, the antidiffusive term can be written as
f¯i = ∑ αi j di j (ui − u j ) = ∑ q̄i j (u j − ui ).
j6=i
(4.64)
j6=i
After the imposition of Dirichlet boundary conditions, the nodes can be numbered
so as to cast the defect correction scheme (4.53) into the partitioned form
"
# "
# (m) (m+1)
(m)
(m) −1
(m)
uΩ
uΩ
rΩ .
Ā
Ā
ΩΩ
ΩΓ
(4.65)
(m+1) =
(m) +
0
I
0
uΓ
uΓ
As in Chapter 3, the subscripts Ω and Γ refer to row numbers associated with the
vectors of unknowns uΩ and Dirichlet boundary values uΓ = g, respectively.
Again, the iterative solution procedure involves the evaluation of residuals, solution of linear systems, and updating the solution step-by-step. The choice of the
preconditioner Ā(m) defines a splitting of the residual (4.63) into a part that varies
4.2 Solution of Nonlinear Systems
145
linearly with u(m+1) and the remainder b(m) which depends on u(m) only. That is,
each defect correction cycle (4.65) represents a linearized problem of the form
# "
"
#"
#
(m)
(m)
(m)
(m+1)
bΩ
ĀΩ Ω ĀΩΓ
uΩ
.
(4.66)
(m+1) =
uΓ
0
I
g
Theorem 3.16 states that the resulting solution is positivity-preserving if the lefthand side matrix is monotone and the corresponding bΩ is nonnegative for g ≥ 0.
Mild nonlinearities incurred by rather diffusive flux limiters can be handled with
the stationary counterpart of (4.57). This linearization leads to (4.66) with
b = f¯.
Ā = −L,
(4.67)
By definition, the matrix in the left-hand side of the linearized system (4.66) is
monotone but there is no guarantee that b(u) ≥ 0 if u ≥ 0. Hence, convergence is
required for positivity preservation and, sometimes, vice versa. To give an example
of a defect correction scheme that maintains positivity in each step, consider
Ā = −(L + Q̄),
b = 0.
(4.68)
In this case, the monotonicity of Ā is sufficient to keep all solutions to (4.66) nonnegative given g ≥ 0. However, only the fully converged solution is guaranteed to
conserve mass. This is in contrast to the usual situation in which every update is
conservative but intermediate solutions may assume nonphysical negative values.
Remark 4.10. In the context of pseudo-time-stepping with implicit TVD schemes,
splitting (4.68) corresponds to the linearized nonconservative implicit (LNI) form
[353]. The preconditioner Ā is defined in terms of the coefficients involved in the
proof of Harten’s theorem [137]. Aspects of strict positivity preservation in linearized TVD approximations are also investigated in [178].
If a generic limiter of the form (4.35)–(4.38) is employed, the following strategy
can be used to determine the coefficients (4.33) of the nonlinear operator Q̄
1. Compute the limited sums P̄i± of positive and negative antidiffusive fluxes
P̄i+ = ∑ max{0, αi j fi j },
j6=i
P̄i− = ∑ min{0, αi j fi j }.
(4.69)
j6=i
2. Retrieve the local extremum diminishing upper and lower bounds (4.36)
Q+
i = ∑ qi j max{0, u j − ui },
j6=i
Q−
i = ∑ qi j min{0, u j − ui }.
(4.70)
j6=i
3. Assemble Q̄ = {q̄i j } using (4.33) with off-diagonal coefficients q̄i j given by
R̄+
i =
P̄i+
,
Q+
i
R̄−
i =
P̄i−
,
Q−
i
q̄i j =
R̄+
i qi j , if u j > ui ,
R̄−
i qi j , if u j < ui .
(4.71)
146
4 Algebraic Flux Correction
Remark 4.11. The correction factors applied to qi j are q ji are generally not equal.
Preconditioners of LNI type are to be recommended for steady state computations in which the use of (4.57) would inhibit convergence or require taking impractically small time steps. For linear transport equations, the need to recompute
the coefficients q̄i j and update Ā(m) after each flux/defect correction step makes matrix assembly more expensive than that for (4.57). However, the differences are not
so pronounced in the case of nonlinear real-life problems since even the discrete
transport operator of low order needs to be reassembled after each outer iteration.
In fact, the defect correction scheme (4.65) with (4.68) is not the only and not the
cheapest way to keep intermediate solutions nonoscillatory. Another representative
of such schemes can be constructed using negative slope linearization (cf. the fourth
basic rule in Section 1.6.3) in the LED form (4.64) of the antidiffusive term
(m) (m+1)
(m) (m)
(m+1)
.
f¯i
= ∑ q̄i j u j − ∑ q̄i j ui
(4.72)
j6=i
j6=i
The defect correction scheme (4.66) proves positivity-preserving if we take [204]
Ā = −L + diag{σ̄i },
σ̄i = ∑ q̄i j
(4.73)
j6=i
such that the i-th component of the right-hand side vector bΩ is given by
bi = ∑ q̄i j u j .
j6=i
Remark 4.12. This definition of the preconditioner Ā can be interpreted as adaptive
relaxation through inertia (4.61) as applied to the low-order operator Ā = −L.
The values of q̄i j depend on the correction factors R̄±
i which need to be recalculated at each outer iteration using (4.69)–(4.71). Alternatively, the inertia term σi
can be defined in terms of the uncorrected coefficients qi j . The preconditioner
Ā = −L + diag{σi },
σi = ∑ qi j
(4.74)
j6=i
differs from (4.73) in the value of the inertia terms σi . The right-hand side becomes
bi = ∑ q̄i j u j + ∑ (qi j − q̄i j )ui = f¯ + σi ui .
j6=i
j6=i
Positivity preservation follows from the fact that 0 ≤ q̄i j ≤ qi j for all j 6= i. In a
practical implementation, the right-hand side vector bΩ is assembled as follows
bi = ∑ αi j fi j + σi ui ,
j6=i
fi j = di j (ui − u j ),
∀ j 6= i.
In contrast to solvers based on (4.68) and (4.73), there is no need to compute q̄i j .
4.3 Steady Transport Problems
147
Remarkably, positivity is maintained if negative off-diagonal entries of the preconditioner Ā are set to zero. For example, if an M-matrix Ā = {āi j } is replaced by
Ā := diag{āii }, the corresponding modification of the right-hand side is
bi := bi − ∑ āi j u j ,
j6=i
āi j ≤ 0,
j 6= i.
Hence, the replacement of Ā, as defined in (4.68), (4.73), or (4.73), by its diagonal
or triangular part does not destroy positivity but convergence will slow down.
Remark 4.13. A slowly converging defect correction scheme can be accelerated
within the framework of a nonlinear full approximation storage / full multigrid
(FAS-FMG) solution strategy. In this case, a basic iteration of the form (4.65) with
a diagonal or upper/lower triangular preconditioner Ā can serve as a smoother.
4.2.5 Accuracy vs. Convergence
In our experience, there is a tradeoff between the accuracy of the flux limiting procedure and convergence of an iterative defect correction scheme. Any enhancement
of the flux limiter that makes it possible to accept more antidiffusion is likely to
have an adverse effect on the nonlinear convergence rates. Conversely, more diffusive schemes converge better but the results are less accurate. In many cases, it is
worthwhile to sacrifice some accuracy if this would make computations much faster.
In other situations, the use of the least diffusive flux limiters is desirable and more
work needs to be invested in the development of a robust iterative solver. The right
constellation of parameter settings for inner and outer iterations is almost certain to
exist. Therefore, there is often no need to give up or use ad hoc tricks (like ‘freezing’
the correction factors [239, 342]) when convergence problems are encountered.
4.3 Steady Transport Problems
The term algebraic flux correction was introduced in [200] as a common name for
high-resolution schemes based on node-oriented flux limiters of FCT and TVD type.
The striking similarities between the two flux limiting techniques were exploited in
[192], where algorithm (4.35)–(4.38) was first presented. The differences between
algebraic FCT and TVD schemes were also analyzed and explained. The former approach is readily applicable to finite element discretizations with a consistent mass
matrix. The amount of admissible antidiffusion is inversely proportional to the time
step, which makes the imposed constraints less restrictive as the time step is refined.
This is a desirable feature if the problem at hand is unsteady, and a small time step
is required to capture the evolution details. However, the use of large time steps results in a loss of accuracy, which compromises the advantages of unconditionally
148
4 Algebraic Flux Correction
stable implicit algorithms. Moreover, severe convergence problems are observed in
the steady-state limit and stationary solutions depend on the pseudo-time step.
On the other hand, algebraic flux correction of TVD type is independent of
the time step and lends itself to the treatment of stationary transport problems. Of
course, it can also be employed in transient computations but the use of small time
steps has no direct influence on the correction factors and, therefore, does not lead to
a marked improvement. Moreover, the need for mass lumping makes the solutions
less accurate than those produced by a consistent-mass FEM-FCT scheme.
Upwind-biased TVD limiters can be integrated into unstructured grid codes and
applied edge-by-edge [8, 239] or node-by-node [206], so as to control the slope ratio
for a local 3-point stencil or the net antidiffusive flux, respectively. In either case, the
resulting scheme proves local extremum diminishing (LED) but, strictly speaking,
the use of standard limiter functions like minmod or superbee does not guarantee
that a second-order accurate approximation is recovered in regions of smoothness.
Indeed, the raw antidiffusive flux for an algebraic flux correction scheme is uniquely
defined by (4.23) rather than by an arbitrary combination of Lax-Wendroff (central
difference) and Beam-Warming (second-order upwind) fluxes. Thus, straightforward extensions of classical TVD schemes may exhibit unexpected behavior when
applied to multidimensional transport problems on strongly nonuniform meshes.
The above considerations have stimulated the search for alternatives to algebraic
FEM-TVD schemes proposed in [206]. Several algorithms [192, 194, 204] were
designed to constrain a given high-order discretization and revert to it in regions
where no flux limiting is required. In the present section, we review the current
state of the art and apply algebraic flux correction to steady convection-dominated
transport equations. The design of FCT algorithms for transient problems and the
numerical treatment of anisotropic diffusion are addressed in subsequent sections.
4.3.1 Upwind-Biased Flux Correction
An upwind-biased limiting strategy is to be recommended for simulation of stationary and weakly time-dependent transport processes at high Peclet numbers. Before
embarking on the development of flux correction schemes for this class of problems,
some implications of the edge orientation convention (4.14) need to be discussed.
If the off-diagonal entry ki j is dominated by the skew-symmetric convective part
→
−
ki′ j = (ki j − k ji )/2 and the edge i j is oriented in accordance with (4.14) then
ki j < 0 < k ji ,
di j = −ki j > 0,
0 = li j < l ji .
(4.75)
The correction factor αi j should ensure that bounds of the form (4.30) hold for a
given set of parameters qi j ≥ 0. Instead of checking these bounds for both nodes,
the antidiffusive flux received by the downwind node j can be absorbed into
k̄ ji (ui − u j ) = l ji (ui − u j ) − αi j fi j ,
(4.76)
4.3 Steady Transport Problems
149
where l ji is positive and 0 ≤ αi j ≤ 1. This edge contribution is of LED type if k̄ ji ≥ 0.
If the problem at hand is stationary or the time-dependent part of the raw antidiffusive flux fi j can be neglected, then definition (4.23) reduces to
fi j = di j (ui − u j ),
f ji = − fi j .
(4.77)
Substitution into (4.76) reveals that the flux-corrected coefficient k̄ ji is given by
k̄ ji = l ji − αi j di j = k ji + (1 − αi j )di j
and proves nonnegative, except at critical points where the velocity changes its sign
so that both off-diagonal entries of K are negative (a rather unusual situation).
To make sure that k̄ ji ≥ 0, an arbitrary antidiffusive flux fi j can be replaced by
fi j := minmod{ fi j , l ji (ui − u j )},
f ji := − fi j .
(4.78)
The minmod function returns zero if its arguments do not have the same sign. Otherwise, the argument with the smallest magnitude is returned. That is,

 min{a, b, . . .}, if a > 0, b > 0, . . .
(4.79)
minmod{a, b, . . .} = max{a, b, . . .}, if a < 0, b < 0, . . .

0,
otherwise.
In many situations, minmod prelimiting (4.78) does not change the magnitude
of the original antidiffusive flux fi j or reduces it by a small amount. The remaining
part can be handled using the upwind-biased version of algorithm (4.35)–(4.38).
The sums of positive and negative antidiffusive fluxes fi j received by node i can
be decomposed into the contributions of upwind (ki j > k ji ) and downwind (ki j ≤ k ji )
neighbors. In light of the above, only the latter part still needs to be limited. Let
Pi+ =
∑
ki j ≤k ji
max{0, fi j },
Pi− =
∑
ki j ≤k ji
min{0, fi j }.
(4.80)
To enforce the positivity constraint, a set of nonnegative coefficients qi j is employed
to define local extremum diminishing upper and lower bounds of the form
Q+
i = ∑ qi j (u j − ui ),
j6=i
Q−
i = ∑ qi j (u j − ui ).
(4.81)
j6=i
The flux fi j is limited using the nodal correction factor for the upwind node, i.e.,
Q±
i
R±
=
min
1,
,
i
Pi±
αi j =
R+
i , if f i j ≥ 0,
R−
i , if f i j < 0,
(4.82)
assuming that (4.14) holds. The flux f ji is multiplied by the same correction factor
α ji := αi j ,
∀ j 6= i,
ki j ≤ k ji .
150
4 Algebraic Flux Correction
Obviously, the amount or raw antidiffusion that can be accepted by the above
upwind-biased flux limiter depends on the definition of qi j in (4.81). Choosing a
value that is too large is as bad as choosing one that is too small. It is worthwhile
to define the upper/lower bounds so that a classical TVD scheme is recovered for
linear convection in 1D. As shown in [192], this requirement is satisfied for
qi j := li j ≥ 0,
∀ j 6= i.
(4.83)
This definition guarantees that (i) limited antidiffusive fluxes have the same effect
as the local extremum diminishing low-order part and (ii) their contribution to the
residual of the flux-corrected scheme is of the same or smaller magnitude.
In the nontrivial case, di j > 0 and 0 = li j < l ji due to (4.14). Hence, the upper
and lower bounds Q±
i consist of upstream edge contributions (ki j > k ji ) only. The
→
−
contribution of the edge i j can be inserted into the sums Pi± and Q±j as follows
Pi+ := Pi+ + max{0, fi j },
Pi− := Pi− + min{0, fi j },
Q+j := Q+j + max{0, l ji (ui − u j )},
Q−j := Q−j + min{0, l ji (ui − u j )}.
After the computation of the correction factors αi j = R±
i from (4.82), the limited
antidiffusive fluxes are inserted into the global vector f¯ that appears in (4.44)
f¯i := f¯i + αi j fi j ,
f¯j := f¯j − αi j fi j .
Thus, a typical implementation of algorithm (4.80)–(4.82) involves two loops over
edges (one to assemble Pi± and Q±
i , the other to perform flux limiting) and one loop
over nodes. The latter is required to evaluate the nodal correction factors R±
i .
Remark 4.14. If Dirichlet boundary conditions are imposed at node i, then the nodal
value ui is fixed. Therefore, there is no need to limit fi j and we can set R±
i := 1.
As an alternative to (4.83), the bounds Q±
i can be defined in terms of [194]
qi j := di j ≥ 0,
∀ j 6= i.
(4.84)
This version is designed to ensure that (i) limited antidiffusive fluxes have the same
effect as artificial diffusion built into the low-order part and (ii) their contribution to
the residual of the flux-corrected scheme is of the same or smaller magnitude.
In a loop over edges, a raw antidiffusive flux of the form fi j = di j (ui − u j ) is
added to the upwind sum Pi± and to the upper/lower bounds for both nodes [194]
+
Q+
i := Qi + max{0, − f i j },
−
Q−
i := Qi + min{0, − f i j },
Q+j := Q+j + max{0, fi j },
Q−j := Q−j + min{0, fi j }.
The 1D version of this algorithm, as applied to the pure convection equation discretized by linear finite elements, corresponds to the minmod limiter which is more
diffusive than the 1D counterpart of (4.83). In multidimensions, the differences are
not so large, while the flux correction scheme based on (4.84) converges better.
4.3 Steady Transport Problems
151
In either case, the imposed constraints (4.81) can be made less restrictive by
taking the largest/smallest possible value of the difference u j − ui , that is,
δ umax
:= max(u j − ui ),
i
δ umin
:= min(u j − ui ).
i
j
j
(4.85)
can be initialized by zero and updated
and δ umin
The so-defined increments δ umax
i
i
edge-by-edge in the same loop as the sums of antidiffusive fluxes to be limited
δ umax
:= max{δ umax
, u j − ui },
i
i
δ umin
:= min{δ umin
i
i , u j − ui },
δ umax
:= max{δ umax
j
j , ui − u j },
min
δ umin
j := min{δ u j , ui − u j }.
(4.86)
The resulting ‘lumped’ upper/lower bounds Q±
i are of the form (4.39), where
,
umax
= ui + δ umax
i
i
umin
= ui + δ umin
i
i
(4.87)
and the coefficients q±
i = ∑ j6=i qi j are nonnegative since qi j ≥ 0 for all j 6= i. This
enhancement makes it possible to resurrect a larger portion of the raw antidiffusive
flux but may inhibit convergence to the steady state and/or give rise to ‘terracing.’
4.3.2 Relationship to TVD Limiters
In fact, algebraic flux correction schemes based on standard TVD limiters [200, 206]
can also be written in the form (4.80)–(4.82). The corresponding coefficients qi j are
given by (4.83) but the raw antidiffusive flux (4.77) is replaced by
+
ri , if ui > u j ,
fi j = Φ̂ (ri )di j (ui − u j ),
ri =
ri− , if ui < u j ,
where Φ̂ (r) is a function associated with a linear second-order scheme in Sweby’s
diagram [216, 315]. The generalized smoothness indicator ri± is defined as the ratio
of edge contributions with positive and negative coefficients [192, 200, 206]
max
ri± =
∑ j6=i max{0, ki j − k ji } min {0, u j − ui }
min
∑ j6=i min{0, ki j − k ji } max {0, u j − ui }
.
For the 1D model problem (4.17) discretized by linear finite elements on a uniform
mesh, the so-defined ri reduces to the ratio of consecutive gradients and [192]
αi j = Φ (ri ),
Φ (r) = max{0, min{2, Φ̂ (r), 2r}}.
The Galerkin flux (4.77) corresponds to Φ̂ (r) ≡ 1 and αi j = max{0, min{1, 2ri }},
which leads to the limited central difference scheme. The minmod and superbee
152
4 Algebraic Flux Correction
limiters are associated with Φ̂ (r) = min{1, ri } and Φ̂ (r) = max{1, ri }, respectively.
The graphs of other limiter functions lie somewhere in-between.
Even though the multidimensional version of the FEM-TVD algorithm with
Φ̂ (r) 6= 1 was found to produce good results even on nonuniform triangular meshes
[201], it does not revert to the original Galerkin scheme and may fail to stay secondorder accurate for smooth data. Instead of manipulating the raw antidiffusive flux in
an uncontrollable manner, it is preferable to add some background diffusion to the
underlying high-order scheme or adjust the parameters qi j as explained above.
4.3.3 Gradient-Based Slope Limiting
Ideally, the correction factors αi j should approach unity in regions where the solution is well-resolved. In particular, no diffusion or antidiffusion should be applied,
even on a nonuniform grid, if the solution varies linearly in the vicinity of node i.
This desirable property is known as linearity preservation [54, 250]. It was found to
maintain consistency and, normally, second-order accuracy on arbitrary meshes.
Classical TVD schemes are linearity-preserving (LP) if Φ (1) = 1, which is the
case for all standard limiter functions. Unfortunately, it is difficult to prove the LP
property for a multidimensional algebraic flux correction scheme based on algorithm (4.80)–(4.82). It is easier to do so if the fluxes are constrained individually so
as to limit the jump of the directional derivative along the edge. To this end, let the
correction factor αi j be defined in terms of the limited slope s̄i j such that
s̄i j = αi j (ui − u j ).
To find the right value of s̄i j , one needs a LED-type estimate of the solution gradient at node i. As explained in Section 2.1.4, a continuous approximation to nodal
gradients can be obtained, for example, using the lumped-mass L2 -projection
gi =
1
cik uk ,
mi ∑
k
(4.88)
where mi is a diagonal entry of the lumped mass matrix ML . The coefficient vectors
ci j =
Z
Ω
ϕi ∇ϕ j dx
constitute the discrete gradient operator C = {ci j } which has zero row sums so that
gi =
1
cik (uk − ui ).
mi k6∑
=i
For any pair of nodes i and j, a usable approximation to the difference ui − u j is
si j = gi · (xi − x j ).
(4.89)
4.3 Steady Transport Problems
153
The slope si j can be estimated in terms of the maxima and minima (4.87) thus:
− ui ) ≤ si j ≤ γi j (umax
− ui ),
γi j (umin
i
i
(4.90)
where the LED upper and lower bounds depend on the nonnegative coefficients
γi j =
1
|cik · (xi − x j )|,
mi k6∑
=i
∀ j 6= i.
(4.91)
Finally, the constrained slope s̄i j is taken to be ui − u j or twice the upper/lower
bound (4.90) for the extrapolated value si j , whichever is smaller in magnitude
− ui ), ui − u j }, if ui > u j ,
min{2γi j (umax
i
s̄i j =
(4.92)
max{2γi j (umin
− ui ), ui − u j }, if ui < u j .
i
Remark 4.15. A symmetric version of this formula was proposed in [204] in the
context of algebraic flux correction schemes for anisotropic diffusion problems.
As the mesh is refined, the difference between the local slopes shrinks and s̄i j approaches ui −u j . This guarantees consistency and linearity preservation for the highresolution scheme (4.44) in which the limited antidiffusive fluxes are given by
f¯i j = αi j di j (ui − u j ) = di j s̄i j ,
∀ j 6= i.
(4.93)
Furthermore, the sum of limited antidiffusive fluxes can be written in the LED form
− min
max
Q−
− ui ) ≤ ∑ αi j fi j ≤ q+
− ui ) = Q+
i ,
i (ui
i = qi (ui
(4.94)
j6=i
where the parameters q±
i combine the coefficients of all positive/negative slopes
q−
i =
∑
ui <u j
q̄i j ,
q+
i =
∑
q̄i j ,
(4.95)
∀ j 6= i.
(4.96)
ui >u j
0 ≤ q̄i j ≤ qi j = 2γi j di j ,
This representation proves that the slope-limited scheme is positivity-preserving.
Remark 4.16. The performance of the slope limiter (4.92) depends on the smoothness of the solution and on the quality of underlying gradient recovery method. If
the gradient is discontinuous, the standard lumped-mass L2 -projection (4.88) may
converge too slowly. To avoid sampling data from both sides of an internal interface,
adaptive gradient reconstruction techniques of ENO type [20] can be employed.
Example 4.2. In one dimension, the lumped-mass L2 -projection (4.88) with mi = ∆ x
and ci±1/2 = ±1/2 reduces to the second-order accurate central difference
ui+1 − ui−1
1 ui − ui−1 ui+1 − ui
=
+
.
gi =
2
∆x
∆x
2∆ x
154
4 Algebraic Flux Correction
For any interior node, the local maxima and minima of the 1D grid function u are
umax
= max{ui−1 , ui , ui+1 },
i
umin
= min{ui−1 , ui , ui+1 }.
i
Estimate (4.90) with γi j = 1 and j = i + 1 yields the upper and lower bounds
umin
− ui ≤ ∆ xgi ≤ umax
− ui .
i
i
Finally, the one-dimensional version of formula (4.92) can be written as follows
s̄i j = minmod{2(ui−1 − ui ), ui − ui+1 }.
(4.97)
The minmod function, as defined in (4.79), compares the signs and magnitudes of
the two slopes. If a local maximum or minimum is attained at node i, then the slope
ratio is negative and, therefore, s̄i j = 0. Otherwise, the result is s̄i j = ui − ui+1 or a
slope of the same sign and smaller magnitude. Limiting is performed only if the two
slopes have opposite signs or their magnitudes differ by a factor of two and more.
Remark 4.17. In the case of the linear convection equation (4.17), the limited antidiffusive flux (4.93) is the same as that obtained with algorithm (4.80)–(4.83).
4.3.4 Reconstruction of Local Stencils
The traditional approach to implementation of high-resolution schemes on unstructured meshes is based on reconstruction of one-dimensional stencils associated with
mesh edges [8, 170, 244]. The key idea is to extend the edge connecting the vertices
xi and x j in both directions so as to define a pair of dummy nodes. The interpolated
or extrapolated solution values at these points are used to define the local slopes that
make it possible to design the diffusive and antidiffusive fluxes as in the 1D case.
In upwind-biased algorithms, just one dummy node is involved. Given a pair of
neighboring nodes i and j such that convention (4.14) holds, the position of the third
node for an equidistant local stencil is xk = 2xi − x j . Alternatively, the dummy node
can be placed at the intersection xl of the straight line through xi and x j with the
boundary of the adjacent element E located upstream of the point xi , see Fig. 4.1.
E
xk
xl
xj
xi
Fig. 4.1 Upwind stencil reconstruction on an unstructured triangular mesh.
4.3 Steady Transport Problems
155
4.3.4.1 Construction of SLIP Schemes
The flux correction process involves edge-by-edge reconstruction and limiting of
slopes for the local 1D stencil. The slope estimated using the data at xi and xk is
si j = uk − ui .
(4.98)
If the dummy node is located at the point xl , then the formula for si j becomes
si j =
|xi − xl |
(ul − ui ).
|x j − xi |
(4.99)
In either case, the limited antidiffusive flux (4.93) can be defined in terms of
s̄i j = minmod{2si j , ui − u j }.
(4.100)
This formula is similar to (4.92) and leads to an algorithm that belongs to the class
of upstream slope-limited positive (SLIP) schemes introduced by Jameson [170].
A wealth of other high-resolution schemes, including straightforward generalizations of classical TVD and MUSCL methods, can be derived using the reconstructed
slope si j to design the modified numerical flux for a finite element or finite volume
discretization on an unstructured mesh [8, 170, 226, 239, 243]. Although such algorithms are sensitive to the orientation of mesh edges and may lack a rigorous theoretical justification, they belong to the most successful unstructured mesh methods
for convection-dominated transport problems and hyperbolic conservation laws.
4.3.4.2 Recovery via FEM Interpolation
The quality of a SLIP-like scheme depends on the structure of the slopes si j given by
(4.98) or (4.99). Let the unknown solution value at a dummy node x̄ be interpolated
using linear or multilinear finite element basis functions {ϕi } such that
u(x̄) = ∑ u j ϕ j (x̄)
(4.101)
j
is a positivity-preserving convex average of the nodal values u j at the vertices of the
element that contains the point x̄. For the upwind triangle E depicted in Fig. 4.1,
linear interpolation is performed using local basis functions {ϕ̂1 , ϕ̂2 , ϕ̂3 }. Without
loss of generality, assume that ϕ̂3 = ϕi |E is the one associated with û3 = ui . Since
this basis function vanishes on the edge where the point x̄ = xl resides, we have
ul = û1 ϕ̂1 (xl ) + û2 ϕ̂2 (xl ) = ξ û1 + (1 − ξ )û2 ,
0 ≤ ξ ≤ 1.
(4.102)
Likewise, the value of uk can be determined using local basis functions defined
on the actual element to which x̄ = xk belongs. This reconstruction technique is
recommended in [239] but it may require a costly search for the host element and
156
4 Algebraic Flux Correction
sampling data from points that are not nearest neighbors of node i. In any case, if
the limited slope s̄i j is proportional to u(x̄) − ui and has the same sign, then (4.101)
implies that the flux f¯i j = di j s̄i j is of LED type and there is no threat to positivity.
4.3.4.3 Gradient-Based Reconstruction
Another popular approach to the reconstruction of si j is gradient-based extrapolation
uk = ui + si j ,
si j = (∇u)i · (xi − x j ).
(4.103)
The nodal gradient (∇u)i can be approximated using the upwind-sided derivatives
or gradient averaging. In the latter case, the above formula reduces to (4.89) which
may fail to possess the LED property if (∇u)i = gi is recovered via the lumped-mass
L2 -projection. In the upwind-biased version, the gradient is obtained by differentiating the solution uh |E restricted to the first mesh cell crossed by the line xi xk
(∇u)i = ∑ ûk ∇ϕ̂k (xi ).
k
As before, ϕ̂k denotes a local basis function associated with the upwind element E
and ûk is the corresponding nodal value. For linear triangles, the result is [170, 239]
si j = ϕ̂1 (û1 − û3 ) + ϕ̂2 (û2 − û3 ),
where û3 = ui under the node numbering convention adopted in (4.102). Since ϕ̂1
and ϕ̂2 are nonnegative, the slope si j is LED and so is its limited counterpart s̄i j .
For a detailed comparative study of various techniques for reconstruction of local 1D stencils in edge-based finite element codes, we refer to Lyra [239, 243, 240].
Slope limiting based on (4.99)–(4.101) is probably the most attractive among the
considered alternatives since it is relatively simple, positivity-preserving, and applicable not only to simplex meshes but also to quadrilateral/hexahedral/hybrid ones.
4.3.5 Background Dissipation
The nondissipative nature of the central difference scheme and Galerkin finite element methods is known to have an adverse effect on the convergence of their
flux-limited counterparts. Therefore, it is generally preferable to use another base
scheme which incorporates some streamline diffusion or dampens oscillatory modes
by means of higher-order dissipation [170, 239, 357]. Alternatively, the stabilization
effect can be achieved in the process of flux correction. Consider a pair of slopes
si j = ei j · (∇u)i ,
s ji = e ji · (∇u) j ,
ei j = xi − x j
(4.104)
4.3 Steady Transport Problems
157
obtained with any of the above gradient reconstruction techniques. To introduce
background dissipation, let the raw antidiffusive flux (4.77) be redefined as [170]
si j − s ji
(∇u)i + (∇u) j
fi j = di j
= di j ei j ·
.
(4.105)
2
2
In upwind-biased flux correction schemes, this replacement should be carried out
before the minmod prelimiting (4.78) is applied. Furthermore, the reconstructed
slope (si j − s ji )/2 should be used instead of ui − u j in formulas like (4.92) and
(4.100), whereas the upper and lower bounds for the limiter remain unchanged.
Remark 4.18. The one-dimensional counterpart of the flux (4.105) is as follows
ui−1 − ui + ui+1 − ui+2
,
j = i + 1.
(4.106)
fi j = di j
2
The Taylor series expansion reveals that (4.105) is related to (4.77) via [226]
ei j ·
(∇u)i + (∇u) j
|ei j |2 ′′
≈ ui − u j −
(u j − u′′i ),
2
4
where u′′ = (ei j · ∇)2 u denotes the second directional derivative of u along the edge.
Hence, a simple way to introduce fourth-order damping into a centered scheme
is to augment the corresponding flux (4.23) by a dissipative term proportional to the
difference between the approximate nodal values of second derivatives. Such terms
extend the stencil of the numerical scheme and are not of LED type but positivity
preservation is enforced by the flux/slope limiter applied to fi j . As a result, background dissipation is added in smooth regions and low-order diffusion elsewhere.
The use of high-order fluxes with a dissipative component makes the constrained
scheme more robust and less susceptible to ‘terracing’ or similar side effects.
In multidimensions, the scalar-valued Laplacian is cheaper to calculate than the
gradient. This is the rationale behind the following definition [226, 250, 301, 302]
fi j = di j (ui − u j ) − di′′j |x j − xi |2 ((∆ u) j − (∆ u)i ),
(4.107)
where di′′j is a nonnegative coefficient associated with fourth-order dissipation. The
nodal Laplacians (∆ u)i and (∆ u) j can be recovered as explained in Section 2.1.4 or
approximated using the off-diagonal entries of the consistent mass matrix [301].
d
Remark 4.19. In the 1D case, definition (4.107) with di′′j = 2i j reduces to (4.106) if
the second derivatives at nodes i and j are approximated by central differences
(∆ u)i =
ui−1 − 2ui + ui+1
,
(∆ x)2
xi = i∆ x,
∀i.
Remark 4.20. Fourth-order dissipation is linearity-preserving since all second derivatives of a linear function are zero. This property turns out to be a valuable tool for
the theoretical analysis of accuracy and consistency on general triangulations [250].
158
4 Algebraic Flux Correction
4.3.6 Numerical Examples
To assess the accuracy of flux-limited Galerkin approximations to stationary transport equations, a comparative study is performed for high-resolution schemes of
Upwind-LED type. The four methods under investigation are abbreviated by
ULED-0
ULED-1
ULED-2
ULED-3
the low-order scheme which corresponds to αi j ≡ 0,
algebraic flux correction of TVD type (4.80)–(4.83),
slope limiting based on gradient recovery and (4.92),
the upstream SLIP scheme given by (4.99)–(4.101).
In one space dimension, the last three algorithms are equivalent to one another and
produce antidiffusive fluxes proportional to the limited slope (4.97). The use of upper/lower bounds (4.84) instead of (4.83) has little influence on the accuracy and
qualitative behavior of ULED-1, at least for the test problems considered here.
Numerical solutions are marched to the steady state using pseudo-time stepping of backward Euler type. Nonlinear systems are solved by the defect correction
scheme (4.53) preconditioned by the monotone low-order operator (4.57). Implicit
underrelaxation with ω ≡ 0.8 is performed to secure convergence to a steady state.
4.3.6.1 Circular Convection
The first test problem is taken from [156]. Consider the hyperbolic conservation law
∇ · (vu) = 0
in Ω = (−1, 1) × (0, 1).
(4.108)
This equation describes steady circular convection if the velocity field is defined as
v(x, y) = (y, −x).
The exact solution and inflow boundary conditions for this test case are given by
p
G(r), if 0.35 ≤ r = x2 + y2 ≤ 0.65,
u(x, y) =
0,
otherwise,
where G(r) is a function that defines the shape of the solution profile along the
inflow (−1 ≤ x < 0) and outflow (0 < x ≤ 1) part of the boundary Γ at y = 0.
Since equation (4.108) is linear, both smooth and discontinuous data propagate
along the characteristics that coincide with the streamlines of the stationary velocity
field (see Chapter 3). The ability of a numerical scheme to maintain smooth peaks
and discontinuities is tested by imposing inflow boundary conditions defined by
2r + 1
G1 (r) = cos2 5π
,
G2 (r) ≡ 1.
3
4.3 Steady Transport Problems
159
The test problems that deal with circular convection of G = G1 and G = G2 will be
referred to as CC1 and CC2, respectively. All numerical solutions are computed on
a uniform mesh of bilinear finite elements which is successively refined to perform
a grid convergence study. The errors are measured in the discrete norms
Emax = max |u(xi ) − ui | ≈ ||u − uh ||∞ ,
i
r
E2 = ∑ mi |u(xi ) − ui |2 ≈ ||u − uh ||2 ,
(4.109)
(4.110)
i
where mi = Ω ϕi dx denotes a diagonal coefficient of the lumped mass matrix or,
equivalently, the area of the control volume associated with the mesh point xi .
Figure 4.2a displays the steady-state solution to CC1 computed by ULED-2 on a
mesh with spacing h = 1/64. The outflow profiles produced by the four schemes on
this mesh are compared in Fig. 4.2b. The uncorrected low-order solution (ULED-0)
R
(a)
(b)
(c)
1.2
Exact
ULED−0
ULED−1
ULED−2
ULED−3
1
−1
10
0.8
E
2
0.6
−2
0.4
10
ULED−0
ULED−1
0.2
ULED−2
ULED−3
0
1/256
−0.2
1/128
1/64
h
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Fig. 4.2 Circular convection, ULED results for smooth data.
1/32
160
4 Algebraic Flux Correction
is not only strongly smeared by numerical diffusion but also asymmetric. The other
three curves are much closer to the dashed line that depicts the exact solution. The
highest accuracy is achieved with ULED-2, followed by ULED-1 and ULED-3.
The log-log plot in Fig. 4.2c shows the variation of the E2 error with the mesh
size h. The order of accuracy p = log2 (E2 (h)/E2 (h/2)) estimated on the finest mesh
level (h = 1/128) equals {0.68, 1.59, 1.89, 1.84} for ULED-0 through ULED-3, respectively. The values of E2 and Emax for all meshes are presented in Table 4.1.
The inability of LED methods to distinguish a smooth peak from a spurious maximum/minimum is the reason why even flux-corrected versions fail to attain secondorder accuracy in this example. Nevertheless, a large portion of artificial diffusion
can be removed in the process of flux correction as long as the solution is smooth.
In the test case CC2, the exact solution is discontinuous at the inlet and remains
so along the streamlines of the incompressible velocity field. The numerical solution
produced by ULED-2 with h = 1/64 is shown in Fig. 4.3a. It is devoid of under(a)
(b)
(c)
1.2
Exact
ULED−0
ULED−1
ULED−2
ULED−3
1
0.8
E
2
0.6
0.4
−1
10
ULED−0
0.2
ULED−1
ULED−2
ULED−3
0
1/256
−0.2
1/128
1/64
h
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Fig. 4.3 Circular convection, ULED results for discontinuous data.
1/32
4.3 Steady Transport Problems
161
Table 4.1 Circular convection, smooth data, convergence history.
ULED-0
E2
Emax
h
1/32
1/64
1/128
1/256
0.209e0
0.157e0
0.107e0
0.666e-1
0.637e0
0.512e0
0.375e0
0.244e0
ULED-1
E2
Emax
0.616e-1
0.235e-1
0.731e-2
0.242e-2
ULED-2
E2
Emax
0.258e0
0.998e-1
0.375e-1
0.132e-1
0.551e-1
0.204e-1
0.595e-2
0.160e-2
0.235e0
0.917e-1
0.340e-1
0.118e-1
ULED-3
E2
Emax
0.734e-1
0.282e-1
0.845e-2
0.236e-2
0.296e0
0.118e0
0.452e-1
0.162e-1
Table 4.2 Circular convection, discontinuous data, convergence history.
h
1/32
1/64
1/128
1/256
ULED-0
E2
Emax
0.292e0
0.237e0
0.198e0
0.166e0
0.600e0
0.561e0
0.573e0
0.540e0
ULED-1
E2
Emax
0.154e0
0.110e0
0.873e-1
0.613e-1
ULED-2
E2
Emax
0.605e0
0.562e0
0.667e0
0.550e0
0.152e0
0.108e0
0.860e-1
0.601e-1
0.597e0
0.566e0
0.683e0
0.557e0
ULED-3
E2
Emax
0.163e0
0.119e0
0.944e-1
0.678e-1
0.584e0
0.570e0
0.660e0
0.569e0
shoots/overshoots and exhibits a fairly high resolution. Figure 4.3b reveals that the
qualitative behavior of the four methods is the same as that for CC1. Again, the
low-order solution (ULED-0) is strongly smeared and asymmetric, whereas the differences between the results produced by ULED-1 through ULED-3 are marginal.
The definition of upper/lower bounds in ULED-2 enables the slope limiter to accept
more antidiffusion than in the case of ULED-1 or ULED-3. The latter proves to be
the most diffusive among the three flux correction schemes under investigation.
The convergence history presented in Fig. 4.3c and Table 4.2 confirms that the
presence of a discontinuous profile has an adverse effect on the overall performance
of numerical schemes. In this example, mesh refinement does not necessarily improve the value of Emax , while the effective order of accuracy with respect to E2
deteriorates to {0.25, 0.51, 0.52, 0.48} for ULED-0 through ULED-3, respectively.
Still, the flux-corrected versions converge twice as fast as the low-order scheme.
In conclusion, no discretization technique can resolve smooth and discontinuous
profiles equally well. However, flux correction still pays off as long as it increases
the effective order of accuracy by a factor of 2 as compared to the underlying loworder scheme. It is not the absolute value of the error but the ratio of errors and
convergence rates that determines which scheme is the best for a given problem.
4.3.6.2 Convection-Diffusion
Another popular test case is the singularly perturbed convection-diffusion equation
∇ · (vu − ε ∇u) = 0
in Ω = (0, 1) × (0, 1)
(4.111)
162
4 Algebraic Flux Correction
which is nominally elliptic but may give rise to arbitrarily sharp internal and boundary layers (see Section 3.1.5). Following John and Knobloch [172], we consider
π
π
ε = 10−8
v ≡ cos , − sin
,
3
3
and impose a Dirichlet boundary condition which is discontinuous at x0 = (0, 0.7)
0 if x = 1 or y ≤ 0.7,
u(x, y) =
1 otherwise.
The exact solution to the above boundary value problem has an internal layer along
the streamline through the point x0 and a boundary layer next to the line y = 0.
To our knowledge, John and Knobloch [172] were the first to perform a detailed and systematic comparative study of conventional stabilized FEM for transport equations with small diffusion and sharp layers. It turns out that even the use
of a nonlinear shock-capturing viscosity may fail to prevent a violation of the discrete maximum principle. In this section, we contribute the results computed with
the four ULED schemes that guarantee the validity of the DMP by construction.
The discretization of (4.111) is performed using linear and bilinear finite elements on three uniform meshes. The first (Grid 1) and second (Grid 2) one are triangular and have the same vertices as a Cartesian mesh (Grid 3) with equal spacing
in both coordinate directions. The orientation of mesh edges is shown in Fig. 4.4.
The total number of nodes (vertices) is N = (1 + 1/h)2 , where h is the mesh size.
The numerical solutions in Fig. 4.5 were produced by ULED-0 and ULED-2 on
Grid 1 with 4,225 degrees of freedom, which corresponds to h = 1/64. As expected,
the low-order solution (left diagram) exhibits a stronger smearing of the interior and
boundary layers. To assess the rate of smearing and enable a quantitative comparison
of different methods, the following benchmark quantities are introduced in [172]
r
smearint = x2 − x1 ,
smearexp =
∑ (min{0, ui − 1})2 ,
xi ∈Ω 2
where Ω2 = {(x, y) ∈ Ω | x ≥ 0.7}. The points x1 and x2 are chosen so that [172]
0.1 = uh (x1 , 0.25) ≤ uh (x, 0.25) ≤ uh (x2 , 0.25) = 0.9,
∀x ∈ (x1 , x2 ).
To determine the values of x1 and x2 for the cutline y = 0.25, the approximate solution uh is evaluated on a one-dimensional subgrid with spacing 10−5 [172].
The magnitudes of smearint and smearexp measure the thickness of the internal
and (exponential) boundary layer, respectively. Two additional benchmark quantities (oscint and oscexp ) are defined in [172] to quantify the pollution by undershoots
and overshoots. In the present study, both of them are identically equal to zero.
Table 4.3 displays the results obtained using algebraic flux correction on the three
meshes with h = 1/64. The reader is invited to compare these results to those produced by SUPG-like finite element methods in [172]. While the low-order scheme
(ULED-0) is characterized by inordinately large values of smearint and smearexp ,
4.3 Steady Transport Problems
163
Grid 2
Grid 1
Grid 3
Fig. 4.4 John-Knobloch benchmark: computational meshes for h = 1/8.
ULED-0 solution
ULED-2 solution
Fig. 4.5 John-Knobloch benchmark: numerical solutions on Grid 1, h = 1/64.
the ULED antidiffusive correction is seen to reduce the amount of smearing, while
keeping the numerical solution free of spurious oscillations in the vicinity of internal and boundary layers. As in the previous example, the most accurate solutions
are obtained with ULED-2. This flux correction scheme receives the highest possible score 10 as defined in [172] for Grid 1. The different pattern of linear triangles in
Grid 2 and the bilinear approximation on Grid 3 give rise to stronger smearing but
the results are still quite good as compared to other finite element methods [172].
Table 4.3 John-Knobloch benchmark: results for all grids with h = 1/64.
ULED-0
ULED-1
ULED-2
ULED-3
Grid smearint smearexp smearint smearexp smearint smearexp smearint smearexp
1
2
3
0.1176
0.2457
0.1929
9.1e-6
1.5494
0.8525
0.0388
0.0665
0.0570
7.8e-6
0.4140
0.2222
0.0379
0.0615
0.0566
7.9e-6
0.3544
0.1961
0.0416
0.0751
0.0675
7.9e-6
0.4180
0.2240
164
4 Algebraic Flux Correction
4.4 Unsteady Transport Problems
In unsteady transport problems, the Galerkin discretization of space derivatives can
also be fixed using any of the flux correction schemes presented in the previous
section. Within the framework of the method of lines (MOL), integration in time
can be performed by an arbitrary explicit or implicit algorithm. However, the use of
small time steps in transient computations makes the upper/lower bounds considered so far more restrictive than necessary to keep the scheme positivity-preserving.
Moreover, the time-dependent part of the raw antidiffusive flux (4.23) for a finite
element discretization may become dominant. Neglecting it would make the mass
lumping error irrecoverable and significantly degrade the phase accuracy. The use
of an upwind-biased algorithm with minmod prelimiting (4.78) may also result in a
serious loss of accuracy. In such situations, algebraic flux correction schemes based
on the flux-corrected transport (FCT) methodology typically perform much better.
FCT was the first nonlinear high-resolution scheme to produce sharp and monotone solutions even in the limit of pure convection [41]. The early FCT algorithms
of Boris, Book, and Hain [42, 43, 44] involve two basic steps:
1. Advance the solution in time by an explicit low-order scheme that incorporates
enough numerical diffusion to suppress undershoots and overshoots.
2. Correct the solution using antidiffusive fluxes limited in such a way that no new
maxima or minima can form and existing extrema cannot grow.
This predictor-corrector strategy is typical of diffusion-antidiffusion (DAD) methods [82]. The job of the numerical diffusion built into the low-order scheme is to
maintain positivity and provide good phase accuracy. The antidiffusive correction is
intended to reduce the amplitude errors in a local extremum diminishing manner.
Zalesak’s fully multidimensional FCT algorithm [355] is based on blending explicit high- and low-order approximations so as to constrain the maximum and minimum increments to each nodal value. A detailed presentation of the underlying
design philosophy can be found in [357]. Zalesak’s limiter has had a significant impact on the development of the algebraic flux correction paradigm and served as a
prototype for the generic limiting strategy presented in Section 4.1.5. In contrast to
TVD schemes and extensions thereof, flux limiters of FCT type operate at the fully
discrete level and are designed to accept as much antidiffusion as possible.
The combination of FCT with finite elements and unstructured meshes dates back
to the explicit algorithms of Parrott and Christie [266] and Löhner et al. [232, 233].
Several implicit FEM-FCT schemes were published by the author and his coworkers
[191, 203, 205, 258]. The rationale for the use of an implicit time discretization
stems from the fact that the CFL stability condition becomes prohibitively restrictive
in the case of strongly nonuniform velocity fields and/or locally refined meshes.
Woodward and Colella ([347], p. 119) conclude that “adaptive grid schemes have a
major drawback – they demand an implicit treatment of the flow equations.” This
statement reflects a widespread prejudice that implicit schemes are computationally
expensive. As a matter of fact, the cost of an implicit algorithm depends on the
choice of iterative methods, parameter settings, and stopping criteria. If the time
4.4 Unsteady Transport Problems
165
step is very small, then a good initial guess is available and the sparse linear system
can be solved with 1-2 iterations of the Jacobi or Gauß-Seidel method. Thus, the cost
per time step approaches that of an explicit finite difference or finite volume scheme.
As the time step increases, so does the number of iterations, and more sophisticated
linear algebra tools (smoothers, preconditioners) may need to be employed.
When antidiffusive fluxes depend on the unknown solution, the nonlinear algebraic system must be replaced by a sequence of linearized ones in which the antidiffusive term is evaluated using the previously computed data. Sometimes, too
many flux/defect correction cycles are required to obtain a fully converged solution,
especially if the Courant number is large and the contribution of the consistent mass
matrix cannot be neglected. The use of a discrete Newton method [258] makes it
possible to accelerate convergence but the computational cost per time step is still
rather high as compared to that of a fully explicit algorithm. This is unacceptable
since the time step for FCT must be chosen relatively small for accuracy reasons.
In this section, we consider both the nonlinear FEM-FCT procedure and simple
predictor-corrector algorithms of diffusion-antidiffusion type. In order to reduce the
cost of flux correction, we linearize the antidiffusive fluxes about a nonoscillatory
end-of-step solution computed by an explicit or implicit low-order scheme [196].
Going back to the roots of FCT, we correct this “transported and diffused” solution directly instead of modifying the algebraic system and solving it again. This
fractional-step approach seems to provide the best cost-accuracy ratio [175, 196].
4.4.1 Nonlinear FEM-FCT Schemes
A family of implicit FEM-FCT algorithms was developed in [191, 203, 205] using
the algebraic approach to flux correction. Consider system (4.44) discretized in time
by the two-level θ −scheme which yields a nonlinear system of the form (4.46)
[ML − θ ∆ tLn+1 ]un+1 = [ML + (1 − θ )∆ tLn ]un + ∆ t f¯(un+1 , un ),
(4.112)
where 0 ≤ θ ≤ 1 is the degree of implicitness. The i−th element of the vector f¯ is
f¯i = ∑ f¯i j ,
j6=i
f¯i j = αi j fi j ,
0 ≤ αi j ≤ 1.
(4.113)
The raw antidiffusive flux fi j is the fully discrete counterpart of (4.23) defined as
n
n
fi j = [mi j (uin+1 − un+1
j ) − mi j (ui − u j )]/∆ t
n+1
n n
n
+ θ din+1
− un+1
j (ui
j ) + (1 − θ )di j (ui − u j ).
(4.114)
Interestingly enough, the contribution of the consistent mass matrix to fi j combines
a truly antidiffusive implicit part and a diffusive explicit part. Mass diffusion of
the form D = MC − ML offers a cheap way to construct the nonoscillatory loworder scheme within the framework of explicit FEM-FCT algorithms [205, 232].
166
4 Algebraic Flux Correction
However, the associated time step restriction (3.92) is more severe than that for the
artificial diffusion operator D with variable coefficients given by (4.11) and (4.12).
The forward Euler version (θ = 0) of (4.112) is not to be recommended. If the
underlying high-order discretization (αi j ≡ 1) is linearly unstable, then aggressive
flux limiting may result in a significant distortion of solution profiles. Also, firstorder time accuracy is insufficient for simulation of unsteady phenomena. If a fully
explicit solution strategy is preferred, then a Lax-Wendroff/Taylor-Galerkin method
or a TVD Runge-Kutta scheme, such as (3.97)–(3.98), should be employed.
The fixed-point iteration method (4.49) transforms (4.112) into a sequence of
linear systems for approximations {u(m) } to the end-of-step solution un+1
[ML − θ ∆ tL(m) ]u(m+1) = [ML + (1 − θ )∆ tLn ]un + ∆ t f¯(u(m) , un ).
(4.115)
(0)
A natural initial guess is u(0) = un such that fi j = dinj (uni − unj ). In our experience,
convergence is faster if the contribution of the time derivative is approximated by
(0)
n n
n
fi j = [mi j (uni − unj ) − mi j (uin−1 − un−1
j )]/∆ t + di j (ui − u j ).
Each solution update of the form (4.115) can be split into three steps [191, 196]
1. Compute an explicit low-order approximation to un+1−θ by solving
ML ũ = [ML + (1 − θ )∆ tLn ]un .
(4.116)
2. Apply limited antidiffusive fluxes to the intermediate solution ũ
ML ū = ML ũ + ∆ t f¯(u(m) , un ).
(4.117)
3. Solve the linear system for the new approximation to un+1
[ML − θ ∆ tL(m) ]u(m+1) = ML ū.
(4.118)
The auxiliary solution ũ depends only on un and needs to be computed just once at
the first outer iteration (m = 0). For the explicit update in Step 1 to be positivitypreserving, the time step ∆ t must satisfy (3.92) with mii = mi and cii = liin
mi + (1 − θ )∆ tliin ≥ 0,
0 ≤ θ < 1.
(4.119)
The correction factors αi j for Step 2 are determined using Zalesak’s multidimensional FCT limiter to be presented below. This algorithm guarantees that ū ≥ 0 for
ũ ≥ 0. Step 3 is positivity-preserving under condition (3.91). In summary,
un ≥ 0
⇒
ũ ≥ 0
⇒
ū ≥ 0
⇒
u(m+1) ≥ 0
provided that the time step ∆ t satisfies (3.91) and (3.92) for the given θ ∈ (0, 1].
4.4 Unsteady Transport Problems
167
4.4.2 Zalesak’s Limiter Revisited
The computation of the correction factors αi j involved in the assembly of (4.113) is
based on the fully multidimensional flux-corrected transport algorithm [355]. This
limiting strategy represents a symmetric version of (4.35)–(4.38). Again, the objective is to make sure that neither positive nor negative antidiffusive fluxes can
conspire to create/enhance a local extremum [355, 357].
4.4.2.1 Prelimiting
Flux correction might be beneficial even in the unlikely case when fi j has the same
sign as ũ j − ũi and poses no threat to positivity. Such an outlier flattens the solution
profile instead of steepening it. As a consequence, numerical ripples may develop
within the bounds imposed on the flux-corrected solution. In the Boris-Book limiter
[42] and some FEM-FCT algorithms [228], the sign of a defective antidiffusive flux
is reversed and the amplitude is limited in the usual way. This trick results in a sharp
resolution of discontinuities but may produce excessive antidiffusion elsewhere.
A safer remedy is to cancel fi j if it is directed down the gradient of ũ. That is,
fi j := 0,
if
fi j (ũ j − ũi ) > 0.
(4.120)
Similarly to (4.78), this manipulation should be done before the evaluation of αi j .
Zalesak [355] argued that the effect of (4.120) is marginal and cosmetic in nature
since the vast majority of antidiffusive fluxes have the right sign and steepen the solution gradient. This remark might have led many readers to disregard equations (14)
and (14′ ) in [355]. Two decades later, the need for ‘prelimiting’ of the form (4.120)
was emphasized by DeVore [80] who explained its ramifications and demonstrated
that it may lead to a marked improvement of accuracy. In our experience, this optional correction step is certainly worth including in FEM-FCT algorithms [205].
In the case of a finite element discretization, the contribution of the consistent
mass matrix provides better phase accuracy but may reverse the sign of the antidiffusive spatial part (4.77) or significantly increase its magnitude. For particularly
sensitive problems, the minmod function (4.79) can be used to redefine fi j as
fi j := minmod{ fi j , di j (ũi − ũ j )},
(4.121)
where di j is a nonnegative coefficient. The default value (4.11) prevents the flux fi j
from becoming diffusive or more antidiffusive than its lumped-mass counterpart.
4.4.2.2 Flux Correction
In the context of multidimensional FCT algorithms, the solution-dependent correction factors αi j are chosen so as to ensure that antidiffusive fluxes acting in concert
168
4 Algebraic Flux Correction
shall not cause the solution value ūi to exceed some maximum value ũmax
or fall
i
below some minimum value ũmin
.
Assuming
the
worst-case
scenario,
positive
and
i
negative fluxes should be limited separately, as proposed by Zalesak [355]
1. Compute the sums of positive/negative antidiffusive fluxes into node i
Pi+ = ∑ max{0, fi j },
j6=i
Pi− = ∑ min{0, fi j }.
(4.122)
j6=i
2. Determine the distance to a local maximum/minimum and the bounds
Q+
i =
mi max
(ũ − ũi ),
∆t i
Q−
i =
mi min
(ũ − ũi ).
∆t i
3. Evaluate the nodal correction factors for the net increment to node i
Q−
Q+
−
+
i
i
Ri = min 1, − .
Ri = min 1, + ,
Pi
Pi
4. Check the sign of each raw antidiffusive flux and adjust its magnitude
−
min{R+
if fi j > 0,
i , R j },
¯fi j = αi j fi j ,
αi j =
− +
min{Ri , R j }, if fi j < 0.
(4.123)
(4.124)
(4.125)
The local maximum ũmax
and minimum ũmin
are defined as in (4.87) and the correi
i
sponding solution increments are assembled in a loop over edges using (4.86).
The symmetric limiting strategy (4.122)–(4.125) guarantees that the second step
(4.117) of the generic FEM-FCT algorithm is positivity-preserving since
+
max
ũmin
= ũi + Q−
.
i
i ≤ ūi ≤ ũi + Qi = ũi
Remark 4.21. It is worthwhile to set R±
i := 1 if Dirichlet boundary conditions are
imposed at node i and, therefore, the value of ui is invariant to the choice of αi j .
Remark 4.22. The need for a repeated evaluation of the nodal correction factors R±
i
can be avoided if the limited antidiffusive flux f¯i j is redefined as follows [258]
f¯i j = minmod{ fi j , ᾱinj dinj (uni − unj )},
where the correction factors ᾱinj may exceed 1. They are calculated at the beginning
±
±
of each time step using Zalesak’s limiter with R±
i = Qi /Pi instead of (4.124).
mi
±
The upper and lower bounds Q±
i are of the form (4.39), where qi = ∆ t . The
appearance of the time step in the denominator turns out to be a blessing or a curse,
depending on the purpose of simulation. On the one hand, the constraints become
less restrictive and, consequently, a larger portion of the raw antidiffusive flux fi j
is retained as the time step is refined. This makes FCT the method of choice for
transient computations. On the other hand, the use of large ∆ t results in a loss of
accuracy, and severe convergence problems may occur in the steady state limit.
4.4 Unsteady Transport Problems
169
A well-known problem associated with flux correction of FCT type is clipping
[41, 355]. Since the sum of limited antidiffusive fluxes is forced to become local
extremum diminishing, existing peaks lose a little bit of amplitude during each time
step. TVD-like methods also fail to recognize smooth extrema, while geometric
schemes based on ENO/WENO reconstruction can handle them fairly well [20].
Another infamous byproduct of flux limiting is known as terracing. It manifests
itself in a distortion of smooth profiles and represents ‘an integrated, nonlinear effect
of residual phase errors’ [265] or, loosely speaking, ‘the ghosts of departed ripples’
[41].
Remark 4.23. As shown in [200, 203], it is possible to incorporate accepted antidiffusion into the auxiliary solution ũ so that only the rejected portion of the flux fi j
needs to be constrained in the next cycle. This simplifies the task of the flux limiter
and enables it to remove more artificial diffusion. The resulting iterative FEM-FCT
algorithm yields a crisp resolution of steep gradients and alleviates peak clipping
but might aggravate terracing and slow down the convergence of outer iterations.
4.4.2.3 Slack Bounds
Zalesak’s limiter is designed to accept as much antidiffusion as the reference solution ũ can accommodate. Instead, it is possible to relax the upper/lower bounds
Q±
i and adjust the time step if this is necessary to satisfy a CFL-like condition. For
example, the local extrema of un can be used to define Q±
i as follows [192]
Q+
i =
mi − mii max
(ui − uni ),
∆t
Q−
i =
mi − mii min
(ui − uni ),
∆t
(4.126)
where mi − mii = ∑ j6=i mi j is the difference between the diagonal entries of the consistent and lumped mass matrices. The resulting correction factors αi j will typically
be smaller than those obtained with (4.123). However, the difference between the
solutions produced by the two versions will shrink and eventually vanish as the time
step is refined. As soon as ∆ t becomes sufficiently small, the accuracy of the results
depends solely on the resolving power of the underlying high-order scheme.
The use of slack bounds (4.126) eliminates the need for evaluating an intermediate solution of low order. Combining (4.116) and (4.117), one obtains
ML ū = [ML + (1 − θ )∆ tLn ]un + ∆ t f¯(u(m) , un ).
Since the sum of limited antidiffusive fluxes satisfies estimate (4.94) with u = un
and q±
i = (mi − mii )/∆ t, the fully discrete scheme is positivity-preserving if [258]
mii + (1 − θ )∆ tliin ≥ 0,
0 ≤ θ < 1.
This CFL-like condition is more restrictive than (4.119). To rectify this, the upper
and lower bounds (4.126) can be defined implicitly using un+1 rather than un .
170
4 Algebraic Flux Correction
4.4.3 Flux Linearization Techniques
A major drawback of the nonlinear FEM-FCT algorithm is the need to recompute
the raw antidiffusive fluxes (4.114) and the correction factors (4.125) after every
solution update. The fixed point iteration method (4.115) keeps intermediate solutions positivity-preserving and the convergence of inner iterations is fast owing to
the M-matrix property of the low-order preconditioner. Unfortunately, the lagged
treatment of limited antidiffusion results in slow convergence of outer iterations. At
large time steps, as many as 50 sweeps may be required to obtain a fully converged
solution (see the numerical examples below). The lumped-mass version, which corresponds to (4.114) with mi j = 0, converges faster but is not to be recommended
for strongly time-dependent problems. The convergence of outer iterations can be
accelerated using a discrete Newton method [258]. However, the costly assembly of
the (approximate) Jacobian operator is unlikely to pay off in transient computations
that call for the use of small time steps and, in many cases, explicit algorithms.
A suitable linearization technique can simplify a FEM-FCT algorithm and make
it much more efficient. For instance, the implicit part of (4.114) can be evaluated
using the solution uH ≈ un+1 of the high-order system (αi j ≡ 1) given by
[MC − θ ∆ tK n+1 ]uH = [MC + (1 − θ )∆ tK n ]un .
(4.127)
In this case, the right-hand side of system (4.118) needs to be assembled just once
per time step. If the underlying continuous problem is linear, then the left-hand side
matrix does not change either, and a single iteration of the form (4.118) is required to
obtain the flux-corrected end-of-step solution un+1 . Thus, the computational effort
reduces to one call of Zalesak’s limiter and two linear systems per time step [205,
258]. If the governing equation is nonlinear, so are the two systems to be solved.
In practice, it is usually much more expensive to solve (4.127) than (4.118). The
main reason is the lack of the M-matrix property which has an adverse effect on the
behavior of linear solvers. As the time step increases, inner iterations may fail to
converge, even if advanced linear algebra tools are employed. A robust alternative
to the brute-force approach is to compute the high-order predictor uH using fixedpoint iteration (4.115) with αi j ≡ 1. However, this linearization strategy is rather
inefficient since the flux-limited version of (4.115) tends to converge faster [258].
Another possibility is to linearize fi j about the solution of the low-order system
[ML − θ ∆ tLn+1 ]uL = [ML + (1 − θ )∆ tLn ]un .
(4.128)
Unlike (4.127), this linear system can be solved efficiently but the flux-corrected solution un+1 computed with (4.116)–(4.118) turns out too diffusive (see [337], Section 5.2). Moreover, it differs from uH even if Zalesak’s limiter returns αi j ≡ 1.
Instead, the flux fi j can be linearized about the solution of a nonlinear system
[ML − θ ∆ t K̄ n+1 ]ūn+1 = [ML + (1 − θ )∆ t K̄ n ]un
(4.129)
4.4 Unsteady Transport Problems
171
associated with a flux correction scheme of the form (4.80)–(4.82) which is supposed to be less accurate but more efficient than the nonlinear FEM-FCT algorithm.
Since the raw antidiffusive flux is replaced by an accurate and smooth approximation, no ripples are typically generated if the optional prelimiting step is omitted.
4.4.4 Predictor-Corrector Algorithms
After years of research aimed at making implicit FEM-FCT schemes more efficient,
the author has come to favor predictor-corrector methods in which the computation
of a tentative solution uL is followed by an explicit flux correction step [196]
ML un+1 = ML uL + ∆ t f¯(uL , un ),
(4.130)
as in the case of classical diffusion-antidiffusion methods based on FCT [42, 82].
The low-order predictor uL can be calculated using (4.128) or any other time
discretization of (4.10), e.g., the explicit TVD Runge-Kutta method (3.97)–(3.98).
The raw antidiffusive fluxes for the flux correction step (4.130) are defined as
L
L
fi j = mi j (u̇Li − u̇Lj ) + din+1
j (ui − u j ),
(4.131)
where u̇L denotes a finite difference approximation of the time derivative. This quantity can be computed, e.g., using the leapfrog method as applied to (4.6)
u̇L = MC−1 [K n+1 uL ].
(4.132)
Since the inverse of MC is full, let successive approximations to u̇L be computed
using Richardson’s iteration preconditioned by the lumped mass matrix [85]
u̇(m+1) = u̇(m) + ML−1 [K n+1 uL − MC u̇(m) ],
m = 0, 1, . . .
(4.133)
starting with u̇(0) = 0 or u̇(0) = (uL − un )/∆ t. Convergence is typically very fast (1-5
iterations are enough) since the consistent mass matrix MC is well-conditioned.
As an alternative to iterating until (4.133) converges, the first lumped-mass approximation u̇(1) = ML−1 [K n+1 uL ] or the corresponding low-order solution
u̇L = ML−1 [Ln+1 uL ]
(4.134)
can be used to predict the raw antidiffusive flux fi j . This technique yields a smooth
but diffusive approximation to the time derivative. Its accuracy can be enhanced
using an upwind-biased flux limiter applied to (4.131) with u̇L = 0. The result is
u̇L = ML−1 [K̄ n+1 uL ],
(4.135)
where K̄ is the nonlinear operator with built-in antidiffusion. Since the computation
of u̇L is performed just once per time step, the associated overhead cost is acceptable.
172
4 Algebraic Flux Correction
The ‘high-order’ solution uH produced by (4.130)–(4.131) with αi j ≡ 1 satisfies
ML uH = ML uL + ∆ t[(ML − MC )u̇L − Dn+1 uL ],
where the vector of approximate time derivatives u̇L is given by (4.132), (4.134), or
(4.135). This uH is typically less oscillatory than the solution of (4.127).
The above linearization strategy offers a number of significant advantages. First,
the low-order predictor uL can be calculated by an arbitrary (explicit or implicit)
time-stepping method. In the case of an implicit algorithm, iterative solvers are fast
due to the M-matrix property of the low-order operator. Second, the leapfrog time
discretization of the antidiffusive flux is second-order accurate with respect to the
time level t n+1 on which uL and fi j are defined. Third, instead of three different
solutions (un , un+1 , and ũ) only the smooth predictor uL is involved in the computation of fi j and αi j for the correction step (4.130). No prelimiting is required for the
L
L
lumped-mass version (u̇L = 0) since the flux fiLj = din+1
j (ui − u j ) is truly antidiffusive. For u̇L 6= 0, it can serve as the upper bound for minmod prelimiting (4.121).
4.4.5 Positive Time Integrators
The efficiency of implicit FEM-FCT schemes for unsteady transport problems is
hampered not only by their nonlinearity but also by the severe time step restriction
(4.119). The second-order accurate Crank-Nicolson version (θ = 21 ) is linearly stable for arbitrary time steps but may produce oscillatory solutions if (4.119) does not
hold. The backward Euler method (θ = 1) is unconditionally positivity-preserving
(UPP) but only first-order accurate in time. An analog of the Godunov theorem [123]
states that no linear time integration scheme of higher order can be UPP [126]. The
unsatisfactory state of affairs can be rectified by selecting the optimal degree of
implicitness θi j ∈ [0, 1] for each node pair so as to localize the CFL-like condition
[337, 338] or to enforce monotonicity constraints by means of time limiters [96].
For a variable-order θ −scheme to stay conservative, not only the diffusive and
antidiffusive terms but also the underlying high-order discretization must be expressed in terms of internodal fluxes. The derivation of a conservative flux decomposition for the Galerkin finite element scheme was addressed in Section 2.1.6. After
the discretization in time, the corresponding low-order scheme can be written as
mi uL = mi un − ∆ t ∑ fˆi j ,
(4.136)
j
where the numerical flux fˆi j consists of a centered convective part augmented by
physical and/or numerical diffusion. Following the notation of previous chapters,
the subscript j = i is reserved for fluxes across the boundary of the domain.
Let the discretization in time be performed by the local θ -scheme [337, 338]
ˆn
fˆi j = θi j fˆin+1
j + (1 − θi j ) f i j ,
0 ≤ θi j ≤ 1.
(4.137)
4.4 Unsteady Transport Problems
173
Since fˆi j is a convex average of fˆin+1
and fˆinj , a numerical scheme of this form is
j
conservative even if the parameters θi j vary in space and time. Instead of adjusting
the time step size so as to preserve positivity with a constant value of θ , the degree of
implicitness can be tailored to the local Courant number and/or to a given solution.
The CFL-like condition (4.119) suggests the following linear combination of the
Crank-Nicolson and backward Euler time integration methods, cf. [337, 338]
mi
1
θi j = min{θi , θ j },
θi = max
.
(4.138)
, 1 + min 0,
2
∆ tliin
This choice is to be recommended for problems in which local mesh refinement
and/or a nonuniform velocity field result in a strong variation of the Courant number.
Furthermore, a tentative solution uL calculated by the backward Euler method
(θi j ≡ 1) or using the local θ -scheme (4.137)–(4.138) can be improved by including
ˆfi j − fˆn+1/2 = θi j − 1 ( fˆn+1 − fˆinj )
ij
ij
2
into the raw antidiffusive flux fi j for a nonlinear or linearized FEM-FCT algorithm.
A more sophisticated approach to the design of time limiters for implicit highresolution schemes is considered in [96]. The nonlinear L-TRAP scheme proposed
therein is based on the following definition of the implicitness parameters
θi j =
θi + θ j
,
2
1
θi = 1 − r̄i ,
2
where r̄i is a correction factor that controls the jump of the temporal slope ratio at a
given node i. A linear and nonlinear stability analysis is performed for the first-order
upwind discretization of the linear convection equation in 1D. The numerical results
for scalar conservation laws and hyperbolic systems confirm that a linear time integration scheme of high order can be forced to satisfy monotonicity constraints using
an extension of tools and concepts originally developed for space discretizations.
4.4.6 Numerical Examples
A properly designed high-resolution scheme should be (i) at least second-order accurate for smooth data and (ii) capable of resolving arbitrary small-scale features
without excessive smearing or steepening of the transported profiles. To evaluate
the accuracy and efficiency of FEM-FCT, we consider a suite of time-dependent
benchmark problems which are representative and challenging enough to indicate
how the methods under investigation would behave in real-life applications [196].
In what follows, a comparative study is performed for both explicit and implicit
FEM-FCT algorithms. The ones based on the TVD Runge-Kutta (3.97)–(3.98),
Crank-Nicolson (θ = 12 ), and backward Euler (θ = 1) time-stepping schemes will be
174
4 Algebraic Flux Correction
referred to as RK-FCT-L, CN-FCT-L, and BE-FCT-L, respectively. The last digit in
these abbreviations refers to the type of flux linearization, if any. The fully nonlinear
version from Section 4.4.1 is assigned the number L = 1, whereas L = 2 stands for
the linearization about the solution uH of system (4.127). Both of these algorithms
require the use of an implicit θ −scheme, so RK-FCT-1 and RK-FCT-2 are not available. The predictor-corrector strategy (4.131) corresponds to L = 3 if u̇L is given by
(4.132), while L = 4 if approximation (4.134) is adopted. In the former case, the
consistent mass matrix MC is ‘inverted’ using 5 iterations of the form (4.133).
By default, systems (4.118), (4.127), and (4.128) are solved by the Gauss-Seidel
method. BiCGSTAB with ILU preconditioning and Cuthill-McKee renumbering is
invoked if convergence fails because of too large a time step. All computations are
performed on a laptop computer using the Intel Fortran Compiler for Linux.
In this section, the following quantities serve as the measure of the difference
between the analytical solution u and a given numerical approximation uh
E1 = ∑ mi |u(xi ) − ui | ≈ ||u − uh ||1 ,
(4.139)
i
E2 =
r
∑ mi |u(xi ) − ui |2 ≈ ||u − uh ||2 .
(4.140)
i
The goal of the below numerical study is to investigate how the above errors and the
CPU times depend on the mesh size h, time step ∆ t, and linearization technique.
4.4.6.1 Solid Body Rotation
Solid body rotation illustrates the ability of a numerical scheme to transport initial
data without distortion [218, 355]. Consider the linear convection equation
∂u
+ ∇ · (vu) = 0
∂t
in Ω = (0, 1) × (0, 1)
(4.141)
which is hyperbolic and of the form (4.1). The incompressible velocity field
v(x, y) = (0.5 − y, x − 0.5)
(4.142)
corresponds to a counterclockwise rotation about the center of the square domain
Ω . Homogeneous Dirichlet boundary conditions are prescribed at the inlets.
The exact solution to (4.141)–(4.142) depends solely on the initial state u0 and
reproduces it exactly after each full revolution. Hence, the challenge of this test is
to preserve the shape of u0 as accurately as possible. Following LeVeque [218], we
consider a slotted cylinder, a sharp cone, and a smooth hump (Fig. 4.6). Initially, the
geometry of each body is given by a function G(x, y) defined within the circle
q
1
(x − x0 )2 + (y − y0 )2 ≤ 1
r(x, y) =
r0
4.4 Unsteady Transport Problems
175
Fig. 4.6 Initial data / exact solution at the final time t = 2π .
of radius r0 = 0.15 centered at a certain point with Cartesian coordinates (x0 , y0 ).
For the slotted cylinder, the reference point is (x0 , y0 ) = (0.5, 0.75) and [218]
(
1 if |x − x0 | ≥ 0.025 or y ≥ 0.85,
G(x, y) =
0 otherwise.
The conical body is centered at (x0 , y0 ) = (0.5, 0.25) and its geometry is defined by
G(x, y) = 1 − r(x, y).
The peak of the hump is located at (x0 , y0 ) = (0.25, 0.5) and the shape function is
G(x, y) =
1 + cos(π r(x, y))
.
4
In the rest of the domain Ω , the solution of equation (4.141) is initialized by zero.
Figure 4.7 displays the results produced by the four Crank-Nicolson FEM-FCT
schemes after one full revolution (t = 2π ). These numerical solutions were computed on a uniform mesh of bilinear elements with h = 1/128 and ∆ t = 10−3 . Prelimiting of the form (4.120) was performed for CN-FCT-1 through CN-FCT-3, while
CN-FCT-4 was found to produce ripple-free solutions even without prelimiting.
The diagrams in Fig. 4.8 depict the convergence history for (4.139) and the total
CPU time as a function of the mesh size h. It can be seen that the linearized schemes
CN-FCT-2 and CN-FCT-3 are almost as accurate as CN-FCT-1. The norms of the
error for CN-FCT-4 are larger on all meshes but decrease at a faster rate than those
for CN-FCT-3. The effective order of accuracy p = log2 (E1 (h)/E1 (h/2)) estimated
176
4 Algebraic Flux Correction
(a) CN-FCT-1
(b) CN-FCT-2
(c) CN-FCT-3
(d) CN-FCT-4
Fig. 4.7 Solid body rotation, CN-FCT schemes, Q1 elements, ∆ t = 10−3 , t = 2π .
CN−FCT−1
CN−FCT−2
CN−FCT−3
CN−FCT−4
CN−FCT−1
CN−FCT−2
CN−FCT−3
CN−FCT−4
3
E
1
CPU
10
2
10
−2
10
1
10
1/256
1/128
1/64
h
1/32
1/256
1/128
1/64
1/32
h
Fig. 4.8 Solid body rotation, convergence history and CPU times for CN-FCT.
using h = 1/128 equals {0.82, 0.81, 0.70, 0.81} for CN-FCT-1 through CN-FCT-4,
respectively. As in the case of steady convection, the presence of a discontinuous
profile is the reason why the errors decrease so slowly with mesh refinement.
A comparison of the CPU times illustrates the gain of efficiency offered by the
predictor-corrector algorithms CN-FCT-3 and CN-FCT-4. The cost of CN-FCT-2 is
significantly higher and even exceeds that for CN-FCT-1 on the coarsest mesh. This
4.4 Unsteady Transport Problems
177
is explained by the slow convergence of the linear solver for the ill-conditioned highorder system (4.127). In the case of CN-FCT-1, defect correction was performed as
long as required to make the normalized residual smaller than the tolerance 10−5 .
The results in Tables 4.4–4.6 illustrate the performance of different time-stepping
methods and flux linearization techniques. The local Courant number ν = |v| ∆ht
varies between zero and νmax = ∆2ht in each test. The errors and CPU times are measured for the numerical solutions computed on a Cartesian mesh with h = 1/128.
The entry in the last column is the average number of outer iterations required to
reach the above tolerance for CN-FCT-1. In the case of linearized FCT schemes,
there is no need for iterative defect correction anymore. This is why NIT equals 1.
For time steps as small as ∆ t = 10−3 , the second-order accurate RK-FCT and
CN-FCT schemes produce essentially the same results (see Table 4.4). The firstorder temporal accuracy of BE-FCT is noticeable but spatial errors are clearly dominant. Furthermore, it can be seen that the predictor-corrector approach to flux linearization reduces the difference between the cost (per time step) of explicit and
implicit FEM-FCT algorithms. A further gain of efficiency can be achieved using a
Jacobi-like iterative method to update the discrete solution in a fully explicit way.
Table 4.5 demonstrates that the errors for BE-FCT become disproportionately
large as compared to those for RK-FCT and CN-FCT as ∆ t is increased by a factor
of 10. Since the backward Euler method is equivalent to the first-order backward
difference approximation of the time derivative, it turns out overly diffusive at large
time steps. The main advantage of BE-FCT is its unconditional stability and positivity preservation for arbitrary time steps. The poor accuracy of the results in Table 4.6
indicates that no time-accurate solutions can be obtained with time steps that violate
the CFL stability condition in the whole domain. However, if the Courant number ν
exceeds unity only in small subdomains, where the velocity is unusually large and/or
adaptive mesh refinement is performed, then a local loss of accuracy is acceptable
if the use of large time steps would make the computation much more efficient.
No results for RK-FCT and linearized CN-FCT are presented in Table 4.6 since
these schemes turn out to be unstable for the time step ∆ t = 10−1 that exceeds
Table 4.4 Solid body rotation: results for h = 1/128, ∆ t = 10−3 , νmax = 0.064.
RK-FCT-3
RK-FCT-4
CN-FCT-1
CN-FCT-2
CN-FCT-3
CN-FCT-4
BE-FCT-1
BE-FCT-2
BE-FCT-3
BE-FCT-4
E1
E2
CPU
NIT
1.1754e-2
2.1913e-2
1.0622e-2
1.0980e-2
1.1729e-2
2.1902e-2
1.9818e-2
2.0069e-2
2.1131e-2
2.7443e-2
5.9882e-2
8.3066e-2
5.6411e-2
5.7370e-2
5.9818e-2
8.3045e-2
7.5392e-2
7.5862e-2
7.9686e-2
9.2886e-2
127
84
343
263
156
116
280
255
155
110
1.0
1.0
3.5
1.0
1.0
1.0
2.5
1.0
1.0
1.0
178
4 Algebraic Flux Correction
Table 4.5 Solid body rotation: results for h = 1/128, ∆ t = 10−2 , νmax = 0.64.
RK-FCT-3
RK-FCT-4
CN-FCT-1
CN-FCT-2
CN-FCT-3
CN-FCT-4
BE-FCT-1
BE-FCT-2
BE-FCT-3
BE-FCT-4
E1
E2
CPU
NIT
1.8289e-2
2.4417e-2
1.2867e-2
1.3552e-2
1.7018e-2
2.3676e-2
5.5943e-2
5.6119e-2
5.7247e-2
5.8198e-2
7.5075e-2
8.8419e-2
6.2870e-2
6.5033e-2
7.3535e-2
8.7242e-2
1.3651e-1
1.3675e-1
1.3966e-1
1.4102e-1
13
8
173
27
17
13
155
36
17
13
1.0
1.0
19.7
1.0
1.0
1.0
15.9
1.0
1.0
1.0
Table 4.6 Solid body rotation: results for h = 1/128, ∆ t = 10−1 , νmax = 6.4.
CN-FCT-1
BE-FCT-1
BE-FCT-3
BE-FCT-4
E1
E2
7.3711e-2
1.0519e-1
1.0504e-1
1.0506e-1
1.6587e-1
2.0244e-1
2.0250e-1
2.0251e-1
CPU
NIT
54
92
4
3
35.0
51.3
1.0
1.0
the upper bound imposed by the CFL-like condition (4.119). CN-FCT-1 remains
stable and more accurate than BE-FCT but the solution is no longer guaranteed to be
positivity-preserving. The results for BE-FCT-2 are missing due to the failure of the
BiCGSTAB solver for the high-order system (4.127). At large time steps, the cost
of a nonlinear FEM-FCT algorithm becomes very high due to slow convergence
of inner and outer iterations. In the case of BE-FCT-1, more than 50 flux/defect
correction steps are required to advance the solution from one time level to the next
in Table 4.6. BE-FCT-3 or BE-FCT-4 produce the same results 30 times faster.
4.4.6.2 Swirling Flow
In the next example, we consider the same equation, the same domain, and the same
initial data as before but the velocity field is given by the formula [218]
v(x, y,t) = (sin2 (π x) sin(2π y)g(t), − sin2 (π y) sin(2π x)g(t)),
(4.143)
where g(t) = cos(π t/T ) on the time interval 0 ≤ t ≤ T . This incompressible velocity field describes a swirling deformation flow that provides a more severe test for
numerical schemes than solid body rotation with a constant angular velocity.
Since the velocity v vanishes on the boundaries of the domain Ω = (0, 1)× (0, 1),
no boundary conditions need to be prescribed in the case of pure convection. The
function g(t) is designed so that the flow slows down and eventually reverses its
4.4 Unsteady Transport Problems
179
direction as time evolves. The exact solution at t = T reproduces the initial profile
depicted in Fig. 4.6 although the flow field has a fairly complicated structure.
The numerical solutions in Fig. 4.9–4.10 were computed by CN-FCT using linear finite elements and ∆ t = 10−3 . The underlying triangular mesh has the same
vertices and twice as many cells as its quadrilateral counterpart with h = 1/128.
The snapshots in Fig. 4.9 correspond to the time of maximum deformation t = T /2
and those in Fig. 4.10 to the final time T = 1.5. Although the solution undergoes
significant deformations in the course of simulation, the shape of the initial data is
recovered fairly well. As in the case of solid body rotation, erosion of the slotted
cylinder is stronger for CN-FCT-4 than for the other three schemes. On the other
hand, the artificial steepening of smooth profiles is alleviated since the linearized
antidiffusive flux is smooth and does not need to be prelimited in this particular test.
The error norms and CPU times for all FEM-FCT algorithms as applied to the
swirling flow problem are presented in Tables 4.7–4.9. Since the velocity field v
is nonstationary, the coefficients of the discrete operators K = {ki j } and D = {di j }
need to be updated after each time step. As explained in Chapter 2, the group FEM
approximation offers a simple and efficient way to do so. Since matrix assembly
claims a larger share of the CPU time, the difference between the cost of explicit
(a) CN-FCT-1
(b) CN-FCT-2
(c) CN-FCT-3
(d) CN-FCT-4
Fig. 4.9 Swirling deformation, CN-FCT schemes, P1 elements, ∆ t = 10−3 , t = 0.75.
180
4 Algebraic Flux Correction
(a) CN-FCT-1
(b) CN-FCT-2
(c) CN-FCT-3
(d) CN-FCT-4
Fig. 4.10 Swirling deformation, CN-FCT schemes, P1 elements, ∆ t = 10−3 , t = 1.5.
and implicit schemes is smaller than in the case of a stationary velocity field. In
Table 4.7, the differences between the CPU times for RK-FCT-4, CN-FCT-4, and
BE-FCT-4 are marginal since the convergence of the Gauss-Seidel solver is fast.
At intermediate and large time steps, the convergence of implicit solvers slows
down. This is the price to be paid for robustness. Tables 4.8–4.9 summarize the results for ∆ t = 10−2 and ∆ t = 10−1 . Both explicit RK-FCT algorithms turned out
unstable, while the linear solver for BE-FCT-2 failed in the latter test. Again, linearization about the low-order predictor uL proved more efficient than the nonlinear
FEM-FCT scheme and the one linearized about the high-order solution uH . The
associated loss of accuracy is acceptable, especially in the case of backward Euler.
4.4.6.3 Convection-Diffusion
To investigate the interplay between physical and numerical diffusion, we apply the
four Crank-Nicolson FEM-FCT algorithms to the parabolic equation
∂u
+ ∇ · (vu − ε ∇u) = 0
∂t
in Ω = (−1, 1) × (−1, 1),
(4.144)
4.4 Unsteady Transport Problems
181
Table 4.7 Swirling deformation: results for h = 1/128, ∆ t = 10−3 , t = 1.5.
RK-FCT-3
RK-FCT-4
CN-FCT-1
CN-FCT-2
CN-FCT-3
CN-FCT-4
BE-FCT-1
BE-FCT-2
BE-FCT-3
BE-FCT-4
E1
E2
CPU
NIT
1.4440e-2
2.4558e-2
1.2043e-2
1.3049e-2
1.4300e-2
2.4493e-2
2.4606e-2
2.5185e-2
2.5334e-2
3.1814e-2
6.6023e-2
8.9130e-2
5.8858e-2
6.1370e-2
6.5626e-2
8.8983e-2
8.4485e-2
8.5713e-2
8.5644e-2
1.0039e-1
45
36
122
60
50
42
112
60
49
41
1.0
1.0
5.6
1.0
1.0
1.0
4.8
1.0
1.0
1.0
Table 4.8 Swirling deformation: results for h = 1/128, ∆ t = 10−2 , t = 1.5.
CN-FCT-1
CN-FCT-2
CN-FCT-3
CN-FCT-4
BE-FCT-1
BE-FCT-2
BE-FCT-3
BE-FCT-4
E1
E2
2.2380e-2
2.3670e-2
2.4119e-2
2.8809e-2
6.4479e-2
6.4621e-2
6.3877e-2
6.4827e-2
8.1277e-2
8.4051e-2
8.6538e-2
9.6268e-2
1.4867e-1
1.4885e-1
1.4760e-1
1.4907e-1
CPU
NIT
38
10
6
5
53
11
6
5
17.9
1.0
1.0
1.0
21.1
1.0
1.0
1.0
Table 4.9 Swirling deformation: results for h = 1/128, ∆ t = 10−1 , t = 1.5.
CN-FCT-1
CN-FCT-3
CN-FCT-4
BE-FCT-1
BE-FCT-3
BE-FCT-4
E1
E2
6.3013e-2
6.4958e-2
6.3189e-2
1.1173e-1
1.1155e-1
1.1155e-1
1.3422e-1
1.3829e-1
1.3556e-1
2.0886e-1
2.0870e-1
2.0870e-1
CPU
NIT
12
1
1
11
1
1
24.1
1.0
1.0
17.7
1.0
1.0
where v(x, y) = (−y, x) is the velocity field and ε = 10−3 is the diffusion coefficient.
The initial and boundary conditions are defined using an analytical solution that
describes convection and diffusion of a rotating Gaussian hill [214]
u(x, y,t) =
1 − r2
e 4ε t ,
4πε t
r2 = (x − x̂)2 + (y − ŷ)2 ,
where x̂ and ŷ are the time-dependent coordinates of the moving peak
x̂(t) = x0 cost − y0 sint,
ŷ(t) = −x0 sint + y0 cost.
(4.145)
182
4 Algebraic Flux Correction
Since u(x, y, 0) = δ (x0 , y0 ), where δ is the Dirac delta function, it is worthwhile
to start the numerical simulation at t0 > 0. As time goes on, the moving peak is
gradually smeared by diffusion and flux limiting becomes redundant. The purpose
of this test is to investigate how FEM-FCT algorithms can handle such situations.
Because of phase errors, the peak of an approximate solution uh may be displaced. Its Cartesian coordinates x̂h = (x̂h , ŷh ) can be estimated as follows
x̂h (t) =
Z
Ω
xuh (x, y,t) dx,
ŷh (t) =
Z
Ω
yuh (x, y,t) dx.
The smearing caused by physical and numerical diffusion is quantified via
σh2 (t) =
Z
Ω
rh2 uh (x, y,t) dx,
rh2 = (x − x̂h )2 + (y − ŷh )2 .
The difference between σh2 and the variance σ 2 = 4ε t of the exact solution (4.145)
to the convection-diffusion equation defines the relative dispersion error
Edisp (t) =
σh2 (t) − σ 2 (t) σh2 (t)
− 1.
=
σ 2 (t)
4ε t
Positive values of Edisp imply that a given approximate solution is overly diffusive,
while negative dispersion errors indicate that some physical diffusion is offset by
numerical antidiffusion inherent to a nondissipative space discretization.
As a direct measure of peak clipping, we introduce the relative amplitude error
Epeak (t) =
max (t)
umax
umax
h (t)
h (t) − u
=
− 1,
max
u (t)
umax (t)
denote the global maxima of the analytical and numerical
where umax and umax
h
solutions, respectively. Hence, the value of Epeak is positive if the top of the rotating
Gaussian hill is too high and negative in the presence of clipping effects.
The numerical experiment begins at t0 = π /2 with a peak located at the point
(x0 , y0 ) = (0.0, 0.5). The initial shape of the Gaussian hill and the solution produced
by CN-FCT-4 after one full revolution (t = 5π /2) are displayed in Fig. 4.11. This
simulation was performed on a uniform mesh of bilinear elements using h = 1/128
and ∆ t = 10−3 . Flux correction was applied to the convective part of the discrete
transport operator, whereas the diffusive part was left unchanged. While the latter
is of nonnegative type (on such a regular mesh), the Galerkin discretization of the
convective term is too antidiffusive. Therefore, it is not desirable to minimize the
amount of artificial diffusion. On the other hand, physical diffusion may be taken
into account if some background dissipation is included in the high-order scheme.
The convergence history and CPU times for t = 5π /2 and ∆ t = 10−3 are presented in Fig. 4.12. Surprisingly enough, the convergence of the nonlinear version
slows down as the mesh is refined. On the finest mesh, the solution obtained with
CN-FCT-1 is even less accurate than that produced by CN-FCT-4. The difference
between the corresponding CPU times is about 50 percent, which is not as much
4.4 Unsteady Transport Problems
183
as in the case of solid body rotation. The convergence behavior of all linearized
FEM-FCT schemes is satisfactory. In this test, the estimated order of accuracy
p = log2 (E1 (h)/E1 (h/2)) equals {1.3, 2.5, 2.4, 2.7} for the solutions computed with
CN-FCT-1 through CN-FCT-4 on the two finest meshes. Remarkably, the cheapest
algorithm converges at the fastest rate which is higher than second order. It is well
known that the presence of the consistent mass matrix makes the 1D Galerkin discretization of pure convection problems fourth-order (!) accurate on a uniform mesh
of linear finite elements (see [86], p. 96). This leads to a significant gain of accuracy
as compared to centered finite difference or finite volume discretizations.
While fourth-order accuracy is no longer guaranteed for multidimensional problems, nonsmooth data, and nonuniform meshes or velocity fields, mass lumping
tends to degrade the accuracy of transient solutions. Likewise, the definition of
u̇L 6= 0 has a strong influence on the final solution. The use of a low-order approximation in CN-FCT-4 makes it more efficient but less accurate than CN-FCT-3.
Figure 4.13 shows how the relative dispersion and amplitude errors vary with the
mesh size. On the coarsest mesh, all four schemes produce rather diffusive results.
Although the peak height predicted by CN-FCT-3 turns out to be very accurate,
further mesh refinement reveals that this is just a coincidence. On finer meshes, the
relative errors Edisp and Epeak approach zero in a monotone fashion. The slightly
antidiffusive behavior of CN-FCT-1 through CN-FCT-3 is due to the nondissipative
nature of the underlying Galerkin discretization. While peak clipping is pronounced
on coarse meshes, this hardly justifies the use of ad hoc adjustments that increase
the complexity of the algorithm and/or may cause a loss of positivity preservation.
(a) initial peak at t0 =
π
2
(b) CN-FCT solution at t =
Fig. 4.11 Snapshots of the Gaussian hill: Q1 elements, h = 1/128, ∆ t = 10−3 .
5π
2
184
4 Algebraic Flux Correction
CN−FCT−1
CN−FCT−2
CN−FCT−3
CN−FCT−4
3
CN−FCT−1
CN−FCT−2
CN−FCT−3
CN−FCT−4
10
−1
10
2
E
1
CPU
10
−2
10
1
10
−3
10
1/256
1/128
1/64
1/32
1/256
h
1/128
1/64
1/32
h
Fig. 4.12 Gaussian hill: convergence history and CPU times for CN-FCT.
0.7
0.6
0.1
CN−FCT−1
CN−FCT−2
CN−FCT−3
CN−FCT−4
0
0.5
−0.1
0.4
Epeak
Edisp
−0.2
0.3
−0.3
0.2
CN−FCT−1
−0.4
CN−FCT−2
CN−FCT−3
0.1
CN−FCT−4
−0.5
0
−0.1
1/256
1/128
1/64
1/32
−0.6
1/256
h
1/128
1/64
1/32
h
Fig. 4.13 Gaussian hill: dispersion (left) and amplitude (right) errors for CN-FCT.
4.5 Limiting for Diffusion Operators
As a matter of fact, not only convective but also diffusive terms may be the cause of
undershoots and overshoots if the computational mesh and/or the diffusion tensor
are anisotropic. This is why discrete diffusion operators may also require some kind
of flux correction [29, 163, 349]. Monotone finite volume schemes were recently
developed for anisotropic diffusion problems on unstructured meshes [215, 221].
However, they are merely positivity-preserving and, generally, do not satisfy the
discrete maximum principle. Moreover, their derivation is based on a design philosophy which is not applicable to finite element approximations. The methodology
proposed in [222] is based on constrained optimization and requires a priori knowledge of the solution bounds. Also, the solution of the constrained optimization problem can become prohibitively expensive as the number of unknowns increases.
In this section, we extend the algebraic flux correction paradigm to anisotropic
diffusion problems and enforce the DMP using a symmetric version [204] of the
slope limiter based on gradient reconstruction. The resulting discretization is akin
to symmetric limited positive (SLIP) schemes [170] but the upper and lower bounds
4.5 Limiting for Diffusion Operators
185
are given in terms of local maxima and minima, as in the case of FCT algorithms.
The numerical study to be presented demonstrates that this kind of slope limiting
renders the constrained Galerkin approximation local extremum diminishing and
keeps it sufficiently accurate when applied to problems with smooth solutions.
4.5.1 The Galerkin Discretization
Consider an elliptic boundary value problem that describes steady diffusive transport of a scalar quantity u in a multidimensional domain Ω with boundary Γ
−∇ · (D∇u) = s, in Ω ,
(4.146)
u = g on Γ ,
where D(x) is a (possibly anisotropic) diffusion tensor and s(x) is a source or sink.
The finite element approach to solution of (4.146) is based on the weak form
a(w, u) = (s, w),
(4.147)
where a(·, ·) is a bilinear form, u is the weak solution, w is any admissible test
function, and (·, ·) is the usual shorthand notation for the scalar product in L2 (Ω )
(w, u) =
Z
Ω
wu dx.
The bilinear form associated with the weak form of the model problem (4.146) reads
a(w, u) =
Z
Ω
∇w · (D∇u) dx.
(4.148)
Given a set of finite element basis functions {ϕi }, substitution of u ≈ ∑ j u j ϕ j and
w = ϕi into (4.147) yields the standard Galerkin discretization
∑ a(ϕi , ϕ j )u j = (ϕi , s),
j
∀i.
This is a linear system of the form Au = b with A = {ai j } and b = {bi } given by
ai j = a(ϕi , ϕ j ),
bi = (ϕi , s).
(4.149)
Due to (4.148) the extended matrix A is symmetric with zero row and column sums
ai j = a ji ,
∑ ai j = ∑ ai j = 0.
i
j
(4.150)
186
4 Algebraic Flux Correction
However, this discrete diffusion operator may fail to be of nonnegative type in the
sense of Definition 3.13 if the given mesh and/or the tensor D are anisotropic. To
prevent a violation of the DMP, some coefficients of A may need to be adjusted.
4.5.2 Positive-Negative Splitting
As before, the process of algebraic flux correction starts with conservative elimination of entries that have wrong sign. Now the stiffness matrix A resides in the
left-hand side of the linear system, so the ‘bad’ part is A+ = {a+
i j } with [311]
a+
i j := max{0, ai j },
∀ j 6= i.
(4.151)
The diagonal coefficients of A+ are defined so that it has zero row and column sums
+
a+
ii := − ∑ ai j ,
j6=i
∀i.
(4.152)
The complement A− := A − A+ represents the ‘good’ part of the stiffness matrix
A = A+ + A− .
(4.153)
By virtue of (4.150)–(4.152), the i−th component of the vector A± u is given by
±
(A± u)i = ∑ a±
i j u j = ∑ ai j (u j − ui )
j6=i
j
and can be expressed in terms of numerical fluxes from one node into another
(A± u)i = − ∑ fi±j ,
fi±j = a±
i j (ui − u j ).
j6=i
(4.154)
The fluxes fi−j and fi+j represent the diffusive and antidiffusive edge contributions to
row i, respectively. The i−th element of the residual vector r = b − Au is
ri = bi + ∑ ( fi+j + fi−j ).
j6=i
To enforce monotonicity constraints, the raw antidiffusive flux fi+j is replaced by
+
f¯i+j = a+
i j s̄i j = αi j f i j ,
(4.155)
where s̄i j denotes the limited slope and αi j ∈ [0, 1] is the correction factor such that
s̄i j = αi j (ui − u j ).
4.5 Limiting for Diffusion Operators
187
Assuming that the nonnegative-type matrix A− is irreducible, Corollary 3.12 states
that the discrete maximum principle holds at least for αi j ≡ 0 and s̄i j ≡ 0. However,
the so-defined ‘low-order’ scheme may turn out to be inconsistent since the elimination of positive off-diagonal entries a+
i j from a discrete diffusion operator introduces
a perturbation error of one order lower than that for a convective flux [193]. If the
antidiffusive part A+ of the stiffness matrix is omitted, the modified scheme may
converge to a wrong solution as the mesh size h is refined. Thus, an antidiffusive
correction is a must and it is essential to guarantee that f¯i+j → fi+j as h → 0.
4.5.3 Symmetric Slope Limiter
Since it is rather difficult to maintain/prove consistency within the framework of a
purely algebraic approach, flux correction is performed using a linearity-preserving
slope limiter based on gradient reconstruction. Due to the symmetry of the stiffness
matrix, nodes i and j should be treated equally. For example, a modification of
formula (4.92) leads to the following definition of the limited slope [204]
− ui ), ui − u j , 2γ ji (u j − umin
if ui > u j ,
min{2γi j (umax
j )},
i
(4.156)
s̄i j =
max )},
−
u
),
u
−
u
,
2
(u
−
u
if ui < u j .
γ
max{2γi j (umin
i
i
j
ji j
i
i
The nonnegative coefficients γi j and γ ji are obtained from a LED estimate of the
form (4.90). The corresponding unlimited slopes si j and s ji are given by (4.104),
where the nodal gradients are approximated using the lumped-mass L2 -projection.
Remark 4.24. A symmetric counterpart of the upstream SLIP limiter (4.100) based
on the reconstruction of local 1D stencils can be formulated in the same way. The
one-dimensional version of formula (4.156) depends on three consecutive slopes
and amounts to a double application of the one-sided slope limiter (4.97)
s̄i j = minmod{2(ui−1 − ui ), ui − ui+1 , 2(ui+1 − ui+2 )}.
If the sign of ui − ui+1 differs from that of a neighboring slope, then s̄i j = 0. Otherwise, the result is at most twice as large in magnitude as ui−1 − ui and ui+1 − ui+2 .
Remark 4.25. If i is an internal node, while j is a node on the boundary, then the positive coefficients a+ji and a+j j pose no hazard to monotonicity of the discrete problem.
Indeed, the corresponding algebraic equation is replaced by the Dirichlet boundary
conditions before the linear solver is invoked. Hence, the one-sided slope limiting
strategy (4.92) may be employed to constrain the antidiffusive flux fi+j into node i.
As the mesh is refined, the difference between the local slopes shrinks and s̄i j approaches ui − u j . The validity of the discrete maximum principle follows from estimates (4.94)–(4.96) with di j = a+
i j and limited antidiffusive fluxes satisfying
q̄i j (uk − ui ) = f¯i+j = − f¯ji+ = −q̄ ji (ul − u j ),
188
4 Algebraic Flux Correction
where uk = umax
or uk = umin
and ul = umin
or ul = umax
depending on the sign of
i
j
i
j
ui − u j . The edge contributions received by nodes i and j are of LED type since
0 ≤ q̄i j ≤ qi j = 2γi j a+
ij,
0 ≤ q̄ ji ≤ q ji = 2γ ji a+ji .
(4.157)
Of course, the constrained Galerkin scheme stays conservative since f¯i+j + f¯ji+ = 0.
4.5.4 Treatment of Nonlinearities
After slope limiting, the corrected fluxes f¯i+j are inserted into the residual vector
r̄ = b − A− u + f¯+ ,
f¯i+ = ∑ f¯i+j
(4.158)
j6=i
and a defect correction scheme of the form (4.65) is employed to solve the associated nonlinear system. The first guess u(0) can be obtained by solving the linear
system A− uL = b or AuH = b. The ‘low-order’ solution u(0) = uL is guaranteed to
be monotone but its accuracy might be very poor due to the lack of consistency.
On the other hand, the unconstrained Galerkin solution u(0) = uH enjoys the ‘best
approximation property’ but may violate the discrete maximum principle. Before
proceeding to the next step, it is worthwhile to trim the undershoots and overshoots,
if any. This is acceptable since u(0) is just an arbitrary guess which has no influence
on the converged solution and, therefore, is not required to be conservative.
The simplest preconditioner for subsequent cycles is, again, the linear operator
Ā = A− . It does not need to be reassembled and linear solvers for (4.66) converge
rapidly owing to the M-matrix property. However, the convergence of outer iterations tends to be very slow, or even fail, if the anisotropy effects are too strong.
Residual smoothing (4.62) makes it possible to achieve monotone convergence but
only the final solution is guaranteed to satisfy the discrete maximum principle. The
alternative is to trade conservation for positivity and use one of the preconditioners
presented in Section 4.2.4. Then intermediate results are devoid of spurious oscillations and converge to the mass-conserving final solution in a monotone fashion.
The relaxation factors for preconditioners (4.73) and (4.74) with L = −A− depend on the nonnegative coefficients q̄i j and qi j related by (4.157). The former version converges faster but the nonlinear diagonal part of Ā must be updated along
with the solution. Anyhow, thousands of defect correction steps may be required to
make the Euclidean norm of the residual as small as 10−10 on a fine mesh. Then the
nonlinearity of the slope-limited Galerkin scheme results in a high overhead cost.
Setting the off-diagonal entries of Ā to zero leads to a fully explicit solution
strategy but the number of outer iterations increases further. As in the case of linear
systems that result from the discretization of elliptic problems, convergence rates
deteriorate as the mesh is refined. Multigrid acceleration of the basic defect correction scheme seems to be a promising way to make computations more efficient.
4.5 Limiting for Diffusion Operators
189
4.5.5 Numerical Examples
In the below numerical study, we test the ability of the symmetric slope limiter
(4.156) to enforce the DMP for (4.146) with an anisotropic diffusion tensor. Also,
we present a grid convergence study for test problems with smooth and discontinuous data. The difference between the numerical solution uh and the exact solution u
is measured in the discrete norms (4.109) and (4.110). The defect correction scheme
preconditioned by (4.74) is employed as the iterative solver for nonlinear systems.
4.5.5.1 Nonsmooth solutions
To illustrate the benefits of slope limiting, we present two examples in which the
existence of steep gradients represents a challenge to the conventional Galerkin
discretization. The computational domain Ω = (0, 1)2 \[4/9, 5/9]2 for the first test
problem (TP1) is depicted in Fig. 4.14a. The outer and inner boundary of Ω are
denoted by Γ0 and Γ1 , respectively. Consider the Dirichlet boundary conditions
u = −1 on Γ0 ,
u = 1 on Γ1 .
(4.159)
Let the diffusion tensor D be the symmetric positive definite 2 × 2 matrix given by
k1 0
cos θ sin θ
R(θ ) =
R(θ ),
,
(4.160)
D = R(−θ )
− sin θ cos θ
0 k2
where k1 = 100 and k2 = 1 are the diffusion coefficients associated with the axes
of the Cartesian coordinate system rotated by the angle θ = −π /6. The source/sink
term s is taken to be zero. By the continuous maximum principle, the exact solution
to the elliptic problem (4.146) is bounded by the Dirichlet boundary data g = ±1.
However, the diffusion tensor (4.160) is highly anisotropic, which may result in a
violation of the DMP even if a regular mesh of acute/nonnarrow type is employed.
(a)
(b)
(c)
Fig. 4.14 TP1: (a) computational domain Ω , (b) triangular mesh, (c) quadrilateral mesh.
190
4 Algebraic Flux Correction
(a) umin = −1.055, umax = 1.0
(b)
umin = −1.054, umax = 1.0
Fig. 4.15 TP1: unconstrained solutions, (a) triangular mesh, (b) quadrilateral mesh.
(a)
umin = −1.0, umax = 1.0
(b)
umin = −1.0, umax = 1.0
Fig. 4.16 TP1: constrained solutions, (a) triangular mesh, (b) quadrilateral mesh.
The verification of the DMP property is performed for linear and bilinear finite
elements on two uniform meshes (see Fig. 4.14b-c). In both cases, the total number of nodes is 1,360. The number of mesh elements equals 2,560 for the triangular
mesh and 1,280 for the quadrilateral one. The numerical solutions computed on
these meshes by the standard Galerkin method are displayed in Fig. 4.15. Both of
them attain correct maximum values but exhibit spurious minima that fall below
the theoretical lower bound umin = −1 by about 5%. Although the undershoots are
relatively small, they might be totally unacceptable in some situations. For example, if the scalar variable u is responsible for phase transitions, such undershoots
can trigger a nonphysical process. Since it is rather difficult to ‘repair’ a DMPviolating solution [222], it is worthwhile to use a scheme that does not produce
undershoots/overshoots in the first place. The constrained Galerkin solutions com-
4.5 Limiting for Diffusion Operators
191
puted on the same meshes using the slope limiter (4.156) are shown in Fig. 4.16.
Both of them satisfy the DMP perfectly, and no other side effects are observed.
The second test problem (TP2) stems from a benchmark suite for anisotropic
diffusion problems on general grids ([146], see Test 9: anisotropy and wells). The
diffusion tensor is given by (4.160) with k1 = 1, k2 = 10−3 , and θ = 67.5◦ . As before, the source term is zero. The domain Ω = (0, 1)2 \(Ω̄4,6 ∪ Ω̄8,6 ) has two square
holes that correspond to cells (4, 6) and (8, 6) of a uniform grid with 11 × 11 cells.
The Dirichlet boundary conditions prescribed on Γ1 = ∂ Ω̄4,6 and Γ2 = ∂ Ω̄8,6 are
u = 0 on Γ1 ,
u = 1 on Γ2 .
(4.161)
Homogeneous Neumann boundary conditions are applied at the outer boundary Γ0
of Ω . For a detailed description of this benchmark problem we refer to [146].
The numerical solutions obtained with 11 × 11 bilinear finite elements are shown
in Fig. 4.17. On such a coarse mesh, the unconstrained Galerkin method produces
undershoots and overshoots of about 14%. Other discretization methods compared
(a) umin = −0.139 umax = 1.139
(b)
umin = 0.0 umax = 1.0
Fig. 4.17 TP2: bilinear elements, (a) unconstrained solution, (b) constrained solution.
192
4 Algebraic Flux Correction
in [146] behave in the same way, whereas the slope-limited solution is uniformly
bounded by the Dirichlet boundary values, as required by the maximum principle.
4.5.5.2 Smooth solutions
Next, we study the approximation properties of the proposed technique as applied
to problems with smooth solutions. Usually, even the conventional Galerkin scheme
does not violate the discrete maximum principle for this type of problems. Thus, no
slope limiting is actually required for smooth data. The goal of the numerical experiments to be performed is to compare the accuracy and convergence behavior of the
constrained nonlinear scheme to those of the underlying Galerkin discretization.
The diffusion tensor and source for the third test problem (TP3) are given by
100 0
(4.162)
D=
,
s(x, y) = 50.5 sin(π x) sin(π y).
0 1
With these parameter settings, the exact solution to the Dirichlet problem (3.21) is
u(x, y) =
1
sin(π x) sin(π y).
2π 2
(4.163)
In accordance with this formula, homogeneous Dirichlet boundary conditions are
imposed. The problem is solved on a sequence of distorted triangular and quadrilateral meshes. Given a uniform grid with spacing h, its distorted counterpart is
generated by applying random perturbations to the coordinates of internal nodes
x := x + αξx h
y := y + αξy h,
where ξx and ξy are random numbers with values in the range from −0.5 to 0.5. The
parameter α ∈ [0, 1] quantifies the degree of distortion. The default value α = 0.4
was adopted to introduce sufficiently strong grid deformations without tangling.
(a)
umin = 0.0, umax = 0.05
(b)
umin = 0.0, umax = 0.05
Fig. 4.18 TP3: numerical solutions, (a) triangular mesh, (b) quadrilateral mesh.
4.5 Limiting for Diffusion Operators
193
Table 4.10 TP3: grid convergence study for the unconstrained Galerkin scheme.
h
1/16
1/32
1/64
1/128
1/256
triangular meshes
E2
Emax
0.158e-3
0.445e-4
0.112e-4
0.320e-5
0.820e-6
quadrilateral meshes
E2
Emax
0.576e-3
0.154e-3
0.473e-4
0.140e-4
0.467e-5
0.113e-3
0.270e-4
0.693e-5
0.176e-5
0.441e-6
0.396e-3
0.113e-3
0.351e-4
0.789e-5
0.231e-5
Table 4.11 TP3: grid convergence study for the constrained Galerkin scheme.
h
1/16
1/32
1/64
1/128
1/256
triangular meshes
E2
Emax
0.293e-3
0.656e-4
0.146e-4
0.321e-5
0.826e-6
quadrilateral meshes
E2
Emax
0.136e-2
0.407e-3
0.121e-3
0.140e-4
0.467e-5
0.265e-3
0.616e-4
0.104e-4
0.204e-5
0.468e-6
0.103e-2
0.337e-3
0.847e-4
0.211e-4
0.642e-5
In this test, the results produced by the Galerkin scheme and by its slope-limited
counterpart are optically indistinguishable. The diagrams in Fig. 4.18 show the
solutions computed using linear and bilinear finite elements with h = 1/16. The
corresponding grid convergence study is presented in Tables 4.10–4.11. On coarse
meshes, the slope limiter tends to ‘clip’ smooth peaks, as in the case of FEM-FCT
methods. To alleviate the undesirable decay of admissible local extrema, the sufficient conditions of DMP should be replaced by a weaker monotonicity constraint.
As the mesh is refined and resolution improves, the slope limiter is gradually
deactivated, and the error norms approach those for the Galerkin method. The results
presented in Tables 4.10–4.11 indicate that slope limiting does not degrade the order
of convergence, and peak clipping becomes less pronounced on finer meshes.
4.5.5.3 Heterogeneous diffusion
The last example (TP4) is designed to test the ability of a discretization technique
to handle problems with discontinuous coefficients. Let the diffusion tensor D be a
piecewise-constant function defined in the unit square Ω = (0, 1)2 as follows
D1 ,
if x < 0.5,
(4.164)
D(x, y) =
D2 ,
if x > 0.5,
where
D1 =
10
,
01
D2 =
10 3
.
3 1
194
4 Algebraic Flux Correction
This heterogeneous diffusion tensor has a jump in value and direction of anisotropy
across the line x = 0.5. The source term s is also discontinuous along this line
4.0,
if x < 0.5,
s(x, y) =
(4.165)
−5.6,
if x > 0.5.
For D and s defined as above, an analytical solution to problem (4.146) is given by
1 − 2y2 + 4xy + 2y + 6x,
if x ≤ 0.5,
u(x, y) =
(4.166)
b2 y2 + c2 xy + d2 x + e2 y + f2 , if x > 0.5.
Substitution into (4.146) yields the following values of the involved coefficients
6 − 4D2 (1, 2)
4(D2 (1, 2) + 1)
, d2 =
,
D2 (1, 1)
D2 (1, 1)
4D2 (1, 1) − 2D2 (1, 2) − 2
4D2 (1, 1) + 2D2 (1, 2) − 3
e2 =
, f2 =
.
D2 (1, 1)
D2 (1, 1)
b2 = −2,
c2 =
Again, the discretization is performed using linear and bilinear finite elements on
distorted meshes. These meshes are constructed as explained in the previous subsection but nodes that lie on the line x = 0.5 are shifted in the y−direction only.
The unconstrained Galerkin solutions for h = 1/16 are presented in Fig. 4.19.
Their limited counterparts look the same but a comparison of the error norms presented in Tables 4.12–4.13 reveals significant differences between the convergence
histories of the slope-limited version on triangular and quadrilateral meshes. Although the solution consists of two smooth patches, its gradient is discontinuous
across the internal interface x = 0.5. Moreover, the kink in the solution profile makes
(a)
umin = 1.0, umax = 6.5
(b)
umin = 1.0, umax = 6.5
Fig. 4.19 TP4: numerical solutions, (a) triangular mesh, (b) quadrilateral mesh.
4.6 Summary
195
Table 4.12 TP4: grid convergence study for the unconstrained Galerkin scheme.
h
1/16
1/32
1/64
1/128
1/256
triangular meshes
E2
Emax
0.101e-2
0.328e-3
0.841e-4
0.211e-4
0.551e-5
0.337e-2
0.133e-2
0.374e-3
0.107e-3
0.351e-4
quadrilateral meshes
E2
Emax
0.473e-3
0.154e-3
0.426e-4
0.109e-4
0.286e-5
0.167e-2
0.636e-3
0.242e-3
0.514e-4
0.166e-4
Table 4.13 TP4: grid convergence study for the constrained Galerkin scheme.
h
1/16
1/32
1/64
1/128
1/256
triangular meshes
E2
Emax
0.244e-2
0.161e-2
0.101e-2
0.281e-3
0.140e-3
0.155e-1
0.187e-1
0.109e-1
0.438e-2
0.214e-2
quadrilateral meshes
E2
Emax
0.473e-3
0.154e-3
0.426e-4
0.109e-4
0.286e-5
0.167e-2
0.636e-3
0.242e-3
0.514e-4
0.166e-4
the outcome of the slope limiting procedure highly mesh-dependent. Note that the
solution is smooth along the y-axis and piecewise-smooth along the x−axis. This is
why the constrained and unconstrained solutions coincide on quadrilateral meshes.
On the other hand, some edges of the triangular mesh are directed skew to the
kink so that the corresponding solution differences are large, whereas the distance to
the nearest local maximum or minimum, as defined in (4.87), is small. This places a
heavy burden on the slope limiter which is forced to reject a large percentage of the
antidiffusive flux in accordance with (4.156). The approximation of discontinuous
gradients by means of the standard L2 -projection (4.88) can also be responsible for
the relatively slow convergence on distorted triangular meshes. In summary, this test
problem turns out to be very easy or rather difficult, depending on the orientation
of mesh edges. It was included to identify the limitations of the proposed limiting
strategy, discuss their ramifications, and illustrate the need for further research.
4.6 Summary
The algebraic flux correction paradigm considered in this chapter provides a set of
general rules, concepts, and tools that make it possible to enforce positivity constraints and the discrete maximum principle in an adaptive way. The presented algorithms are based on a generalization of classical high-resolution schemes. All of
them exhibit a common structure but the actual flux/slope limiting procedure can be
tailored to the specific properties of the transport problem under investigation.
196
4 Algebraic Flux Correction
In the context of finite element methods, an upwind-biased limiting strategy of
TVD type is appropriate if the Peclet number is large and the mass lumping error
is negligible/acceptable. The inclusion of the consistent mass matrix and the use
of symmetric FCT limiters are to be recommended for transient computations with
small time steps. Similarly, the best approach to the solution of linear and nonlinear
systems depends on the properties of the continuous and discrete problems.
The presentation of algebraic flux correction schemes in this chapter was aimed
primarily at finite element practitioners who share our viewpoint (or can be convinced) that positivity preservation is more important than the Galerkin orthogonality when it comes to numerical solution of problems with steep fronts. Instead
of manipulating the variational formulation and fitting a free parameter, we have
shown that artificial diffusion operators can be constructed at the discrete level so as
to control the contribution of matrix entries associated with antidiffusive fluxes.
Remarkably, the same limiter routine can be employed to enforce the positivity
constraint for linear and multilinear elements in 2D and 3D, on structured and unstructured meshes. The origin of discrete operators makes no difference as far as the
M-matrix property is concerned. However, the flux limiter must be designed to keep
the perturbation of the algebraic system as small as possible. The demand for high
resolution is particularly difficult to meet in the case of higher-order finite elements
because the fluxes may depend on solution values at more than two nodes, and even
the construction of an optimal low-order scheme becomes a nontrivial task.
In essence, flux correction is intended to increase the local order of accuracy in
smooth regions and decrease it when a front is detected. Hence, a further extension can be envisaged within the framework of a variable-order (p-adaptive) finite
element scheme with hierarchical basis functions. The basic idea is as follows:
1. Compute a first approximation ū using (multi-)linear elements and flux limiting.
2. Increase the polynomial order in smooth regions where no limiting is performed.
3. Recalculate the solution using the nonoscillatory predictor ū as the initial guess.
On the one hand, it is difficult to prove that the discrete maximum principle will
continue to hold in regions where the lowest-order Galerkin approximation produces
monotone results. On the other hand, it is intuitively clear that the risk of a DMP
violation is minimal and worth taking, given the lack of cost-effective alternatives.
Also, it is always possible to decrease the order of basis functions if the solution
develops local maxima or minima that are not present in ū computed at stage 1.
The unavoidable errors due to low-order artificial diffusion that cannot be removed in the vicinity of internal or boundary layers can be compensated by means
of local mesh refinement. The use of mesh adaptation for improving the accuracy of
numerical approximations to transport equations is addressed in Chapter 5.
Chapter 5
Error Estimates and Adaptivity
In this chapter, we discuss some aspects of mesh adaptation for steady transport
equations. The goal-oriented error estimator developed in [197, 199] is used as a
refinement criterion. The error in the value of a linear target functional is measured
in terms of weighted residuals that depend on the solutions to the primal and dual
problems. The Galerkin orthogonality error is taken into account and turns out to
be dominant whenever flux or slope limiters are activated to enforce monotonicity
constraints. The localization of global errors is performed using a natural decomposition of the involved weights into nodal contributions. The developed simulation
tools are applied to a linear convection problem in two space dimensions.
5.1 Introduction
The goal-oriented approach to error estimation [14, 27, 185, 295, 309] is applicable
not only to elliptic PDEs but also to hyperbolic conservation laws [141, 142, 310].
In most cases, the error in the quantity of interest is estimated using the duality argument, Galerkin orthogonality, and a direct decomposition of the weighted residual
into element contributions. The most prominent representative of such error estimators is the Dual Weighted Residual (DWR) method of Becker and Rannacher
[27, 28]. The recent paper by Meidner et al. [248] is a rare example of a DWR estimate that does not require Galerkin orthogonality or information about the cause of
its possible violation.
Kuzmin and Korotov [197] applied the DWR method to steady convectiondiffusion equations and obtained a simple estimate of local Galerkin orthogonality errors due to flux limiting or other ‘variational crimes.’ In contrast to the usual
approach, the weighted residuals are decomposed into nodal (rather than element)
contributions. In regions of insufficient mesh resolution, the computable Galerkin
orthogonality error comes into prominence. The mesh adaptation strategy to be presented below takes advantage of this fact.
197
198
5 Error Estimates and Adaptivity
5.2 Galerkin Weak Form
Steady convective transport of a conserved scalar quantity u in a domain Ω with
boundary Γ can be described by the linear hyperbolic equation
in Ω .
∇ · (vu) = s
(5.1)
Here v is a stationary velocity field and s is a volumetric source/sink. Due to hyperbolicity, a Dirichlet boundary condition is imposed at the inlet
on Γin = {x ∈ Γ | v · n < 0},
u = uD
(5.2)
where n is the unit outward normal and uD is the prescribed boundary data.
The weak form of the above boundary value problem can be written as
a(w, u) = b(w),
∀w.
(5.3)
For brevity, we refrain from an explicit definition of functional spaces. The bilinear
form a(·, ·) and the linear functional b(·) are defined by
a(w, u) =
Z
w ∇ · (vu)∆ x −
Ω
b(w) =
Z
Ω
ws∆ x −
Z
Z
Γin
wuv · n ds,
wuD v · n ds.
Γin
(5.4)
(5.5)
The inflow boundary conditions are imposed weakly via the surface integrals.
The differentiation of vu in (5.4) can be avoided using integration by parts
a(w, u) =
Z
Γ
wuv · n ds −
Z
Ω
∇w · (vu)∆ x.
(5.6)
This representation implies that a discontinuous weak solution u is admissible. In
linear hyperbolic problems of the form (5.1), singularities travel along the streamlines of v. They may be caused by a jump in the value of s or uD .
5.3 Global Error Estimates
Let uh be a continuous function that may represent an approximate solution to (5.1)–
(5.2) or a finite element interpolant of discrete nodal values. The numerical error
e = u − uh can be measured using the residual of (5.3)
ρ (w, uh ) = b(w) − a(w, uh ).
(5.7)
Obviously, the value of ρ (w, uh ) depends not only on the quality of uh but also on
the choice of w. In goal-oriented estimates, this weight carries information about the
5.3 Global Error Estimates
199
quantities of interest. The objectives of a numerical study are commonly defined in
terms of a linear output functional, such as [310]
j(u) =
Z
Ω
gu∆ x +
Z
Γout
huv · n ds,
g, h ∈ {0, 1}.
(5.8)
The piecewise-constant function g picks out a subdomain, for example, an interior
or boundary layer, where a particularly accurate approximation to u is desired. The
selector h picks out a portion of the outflow boundary Γout = {x ∈ Γ | v · n > 0},
where the convective flux is to be controlled.
In order to estimate the error j(e) in the numerical value of the output functional,
consider the dual or adjoint problem [27, 28] associated with (5.3)
a(z, e) = j(e),
∀e.
(5.9)
The surface integral in (5.8) implies the weakly imposed Dirichlet boundary condition z = h on Γout [310]. The error j(e) and residual (5.7) are related by
j(u − uh ) = a(z, u − uh ) = ρ (z, uh ).
(5.10)
An arbitrary numerical approximation zh to the exact solution z of the dual problem
(5.9) can be used to decompose the so-defined error as follows
j(u − uh ) = ρ (z − zh , uh ) + ρ (zh , uh ).
(5.11)
If Galerkin orthogonality holds for the numerical approximation uh , then ρ (zh , uh ) =
0. Thus, the computable term ρ (zh , uh ) is omitted in most goal-oriented error estimates for finite element discretizations. However, the orthogonality condition is
frequently violated due to numerical integration, round-off errors, slack tolerances
for iterative solvers, and flux limiting.
Since the exact dual solution z is usually unknown, the derivation of a computable
error estimate involves another approximation ẑ ≈ z such that
j(u − uh ) ≈ ρ (ẑ − zh , uh ) + ρ (zh , uh ).
(5.12)
The magnitudes of the two residuals can be estimated separately as follows:
|ρ (ẑ − zh , uh )| ≤ Φ ,
|ρ (zh , uh )| ≤ Ψ ,
(5.13)
where the globally defined bounds Φ and Ψ are assembled from contributions of
individual nodes or elements, as explained in the next section.
The reference solution ẑ is commonly obtained from zh using some sort of postprocessing. If ρ (zh , uh ) = 0, then the estimate j(u − uh ) ≈ 0 that follows from (5.12)
with ẑ = zh is worthless, hence the need to compute ẑ on another mesh or interpolate
it using higher-order polynomials [197, 295]. On the other hand, the setting ẑ = zh is
not only acceptable but also optimal for nonlinear flux-limited discretizations such
that j(u − uh ) ≈ ρ (zh , uh ) 6= 0. In situations when the term ρ (z − zh , uh ) is non-
200
5 Error Estimates and Adaptivity
negligible, extra work needs to be invested into the recovery of a superconvergent
approximation ẑ 6= zh .
5.4 Local Error Estimates
The global upper bounds Φ and Ψ make it possible to verify the accuracy of the
approximate solution uh but the estimated errors in the quantity of interest must be
localized to find the regions where a given mesh is too coarse or too fine. A straightforward decomposition of weighted residuals into element contributions results in
an oscillatory distribution and a strong overestimation of local errors. In particular,
the restriction of the term ρ (zh , uh ) to a single element Ωk can be large in magnitude
even if Galerkin orthogonality is satisfied globally (positive and negative contributions cancel out).
Following Schmich and Vexler [295], we decompose Φ and Ψ into local bounds
associated with the nodes of the mesh on which zh is defined. Let
zh = ∑ zi ϕi ,
(5.14)
i
where {ϕi } is a set of Lagrange basis functions such that ∑i ϕi ≡ 1 and
ẑ − zh = ∑ wi ,
i
wi = ϕi (ẑ − zh ).
(5.15)
The contribution of node i to the bounds Φ and Ψ is defined as in [197]
Φi = |ρ (wi , uh )|,
Ψi = |ρ (zi ϕi , uh )|.
(5.16)
If the residual is orthogonal to the test function ϕi , then Ψi = 0. A nonvanishing
value of Ψi implies a local violation of Galerkin orthogonality.
The magnitude of j(u − uh ) is estimated by the sum of local errors, i.e.,
Φ = ∑ Φi ,
i
Ψ = ∑ Ψi .
(5.17)
i
Finally, an optional conversion into element contributions is performed for mesh
adaptation purposes. Introducing the continuous error function
ξ = ∑ ξi ϕi ,
i
Φi + Ψi
ξi = R
,
Ω ϕi ∆ x
(5.18)
the following representation of the total error η = Φ + Ψ is obtained [197]
η = ∑ ηk ,
k
ηk =
Z
Ωk
ξ ∆ x.
(5.19)
5.5 Numerical Experiments
201
In a practical implementation, the midpoint rule is employed to calculate ηk .
5.5 Numerical Experiments
In this section, the presented high-resolution finite element scheme, goal-oriented
error estimator, and hierarchical mesh adaptation algorithm are applied to a test
problem from [156]. Consider equation (5.1) with s ≡ 0 and
v(x, y) = (y, −x)
in Ω = (−1, 1) × (0, 1).
This incompressible velocity field corresponds to steady rotation about (0, 0).
The exact solution and inflow boundary conditions are given by [156]
p
1, if 0.35 ≤ x2 + y2 ≤ 0.65,
u(x, y) =
0, otherwise.
The so-defined discontinuous inflow profile (−1 ≤ x < 0, y = 0) undergoes circular
convection and propagates along the streamlines of v(x, y) all the way to the outlet
(0 < x ≤ 1, y = 0), while its shape remains the same.
Let j(u) be defined by (5.8) with g = 1 in ω = (−0.1, 0.1) × (0, 1) and g = 0
elsewhere. The function h is defined as the trace of g on Γout . The exact value of j(u)
is 6.04497e−02. The solution shown in Fig. 5.1 (a) was computed by the FEM-LED
scheme described in Chapter 4 on a uniform mesh of bilinear elements with spacing
h = 1/80. Owing to algebraic flux correction, the resolution of the discontinuous
front is remarkably sharp, and no undershoots or overshoots are observed. However,
it is obvious that there is actually no need for such a high resolution beyond x > 0.1
if it is enough to have an accurate approximation in the small subdomain ω . Indeed,
whatever is happening downstream of ω has no influence on the solution in this
subdomain. This is illustrated by Fig. 5.1 (b) which shows the solution to the dual
problem computed by the FEM-LED scheme on the same mesh.
Goal-oriented error analysis is performed using estimate (5.12) with ẑ = zh . This
setting implies that Φ = 0 and η = Ψ is the Galerkin orthogonality error caused by
flux limiting. Remarkably, the resulting global estimates are in a good agreement
with the exact error which is illustrated in Table 5.1 for different grid spacings. The
sharpness of the obtained error estimates is measured using the absolute and relative
effectivity indices [197]
|( j(u − uh )| − η η
.
Ieff =
,
Irel = |( j(u − uh )|
|( j(u)|
We remark that the value of Ieff is unstable and misleading when the denominator
is very small or zero, and the evaluation of integrals is subject to rounding errors.
The relative effectivity index Ieff is free of this drawback and exhibits monotone
convergence as the mesh is refined (see Table 5.1).
202
5 Error Estimates and Adaptivity
h
1/10
1/20
1/40
1/80
1/160
| j(u − uh )|
2.009555e-03
4.401534e-04
1.312391e-04
4.283158e-05
1.254089e-05
η (zh , uh )
2.115012e-03
3.640322e-04
1.025215e-04
3.535738e-05
1.072697e-05
Ieff
1.05
0.82
0.78
0.82
0.85
Irel
1.744541e-03
1.259248e-03
4.750662e-04
1.236433e-04
3.000709e-05
Table 5.1 Circular convection: exact vs. estimated global error.
The adaptive hybrid mesh presented in Fig. 5.2 is refined along the discontinuity
lines of u but only until they cross the outflow boundary of ω . Using a finer mesh
beyond the line x = 0.1 would not improve the accuracy of the solution uh inside
ω . The smallest mesh width is h = 1/320, which corresponds to more than 200,000
cells in the case of global mesh refinement.
Since the dual weight zh contains built-in information regarding the transport
of errors and goals of simulation, such error estimators furnish a better refinement
criterion than, for example, error indicators based on gradient recovery [362]. In the
latter case, unnecessary mesh refinement would take place along the discontinuities
located downstream of the subdomain ω .
5.6 Summary
A goal-oriented error estimate was derived for LED discretizations of a steady transport equation. The loss of Galerkin orthogonality in the process of flux limiting was
shown to provide valuable feedback for mesh adaptation. The local orthogonality
error was employed to generate an adaptive mesh for circular convection in a 2D
domain. Diffusive terms can be included using gradient recovery to stabilize the
residuals and infer a proper distribution of local errors [197]. Further work will concentrate on goal-oriented error estimation for unsteady flow problems.
5.6 Summary
203
(a) primal solution
(b) dual solution
Fig. 5.1 Circular convection: FEM-LED discretization, h = 1/80.
204
5 Error Estimates and Adaptivity
(a) primal solution
(b) computational mesh
Fig. 5.2 Circular convection: FEM-LED discretization, 5,980 cells.
References
1. M. Aftosmis and N. Kroll, A quadrilateral based second-order TVD method for unstructured
adaptive methods. AIAA Paper, 91-0124, 1991.
2. M. Ainsworth and J.T. Oden, A Posteriori Error Estimation in Finite Element Analysis. John
Wiley & Sons, New York, 2000.
3. M. Ainsworth, J.Z. Zhu, A.W. Craig, O.C. Zienkiewicz, Analysis of the Zienkiewicz-Zhu aposteriori error estimator in the finite element method. Int. J. Numer. Methods Engrg. 28:9
(1989) 2161–2174.
4. J.D. Anderson, Jr., Computational Fluid Dynamics. The Basics with Applications. McGrawHill, 1995.
5. J.D. Anderson, Jr., Modern Compressible Flow: With Historical Perspective, McGraw-Hill,
1990.
6. F. Angrand, A. Derieux, L. Loth, G. Vijayasundaram, Simulation of Euler transonic flows by
means of explicit finite element type schemes. INRIA Research Report 250, 1983.
7. F. Angrand and A. Derieux, Some explicit triangular finite element schemes for the Euler
equations. Int. J. Numer. Methods Fluids 4 (1984) 749–764.
8. P. Arminjon and A. Dervieux, Construction of TVD-like artificial viscosities on 2dimensional arbitrary FEM grids. INRIA Research Report 1111, 1989.
9. M. Arora and P.L. Roe, A well-behaved TVD limiter for high-resolution calculations of unsteady flow. J. Comput. Phys. 132 (1997) 3–11.
10. Athena test suite, http://www.astro.virginia.edu/VITA/ATHENA/dmr.html.
11. K. Baba and M. Tabata, On a conservative upwind finite element scheme for convective diffusion equations. RAIRO Numerical Analysis 15 (1981) 3–25.
12. I. Babuška and W.C. Rheinboldt, Error estimates for adaptive finite element computations.
SIAM J. Numer. Anal. 15:4 (1978) 736–354.
13. I. Babuška and W.C. Rheinboldt, A posteriori error estimates for the finite element method.
Int. J. Numer. Methods Engrg. 12:10 (1978) 1597–1615.
14. W. Bangerth and R. Rannacher, Adaptive finite element methods for differential equations.
Lectures in Mathematics, ETH Zürich, Birkhäuser, 2003.
15. R.E. Bank, A.H. Sherman, A. Weiser, Some refinement algorithms and data structures for
regular local mesh refinement. In: R. Stepleman (ed.), Scientific Computing, Applications of
Mathematics and Computing to the Physical Sciences, IMACS Transactions on Scientific
Computation, Vol. I. North-Holland, Amsterdam, 1983, 3–17.
16. R.E. Bank and R.K. Smith, Mesh smoothing using a posteriori error estimates. SIAM J. Numer. Anal. 34 (1997) 979–997.
17. T.J. Barth, Numerical aspects of computing viscous high Reynolds number flows on unstructured meshes. AIAA Paper, 91-0721, 1991.
18. T.J. Barth, Aspects of unstructured grids and finite volume solvers for the Euler and NavierStokes equations. In: Lecture Series 1994-05, von Karman Institute for Fluid Dynamics, Brussels, 1994.
205
206
References
19. T. Barth and D.C. Jespersen, The design and application of upwind schemes on unstructured
meshes. AIAA Paper, 89-0366, 1989.
20. T. Barth and M. Ohlberger, Finite volume methods: foundation and analysis. In: E. Stein,
R. de Borst, T.J.R. Hughes (eds), Encyclopedia of Computational Mechanics, Volume 1: Fundamentals. John Wiley & Sons, 2004, 439–474.
21. G.K. Batchelor, An Introduction to Fluid Dynamics. Cambridge University Press, 2000.
22. J.D. Baum and R. Löhner, Numerical simulation of pilot/seat ejection from an F-16. AIAA
Paper, 93-0783, 1993.
23. J.D. Baum, H. Luo, R. Löhner, Validation of a new ALE adaptive unstructured moving body
methodology for multi-store ejection simulations. AIAA Paper, 95-1792, 1995.
24. C.E. Baumann and T.J. Oden, A discontinuous hp-finite element method for the solution of
the Euler and Navier-Stokes equations. Int. J. Numer. Methods Fluids 31 (1999) 79–95.
25. Y. Bazilevs and T.J.R. Hughes, Weak imposition of Dirichlet boundary conditions in fluid
mechanics. Comput. Fluids 36:1 (2007) 12–26.
26. R. Becker and P. Hansbo, A simple pressure stabilization method for the Stokes equation.
Commun. Numer. Methods Engrg., in press.
27. R. Becker and R. Rannacher, A feed-back approach to error control in finite element methods:
Basic analysis and examples. East-West J. Numer. Math. 4 (1996) 237–264.
28. R. Becker and R. Rannacher, An optimal control approach to a posteriori error estimation in
finite element methods. Acta Numerica 10 (2001) 1–102.
29. J.-M. Beckers, H. Burchard, E. Deleersnijder, P.P. Mathieu, Numerical discretization of rotated diffusion operators in ocean models. Monthly Weather Review 28:8 (2000) 2711–2733.
30. M.J. Berger and P. Colella, Local adaptive mesh refinement for shock hydrodynamics. J.
Comput. Phys. 82:1 (1989) 64–84.
31. M.J. Berger and R.J. LeVeque, Adaptive mesh refinement using wave-propagation algorithms
for hyperbolic systems. SIAM J. Numer. Anal. 35:6 (1998) 1439–1461.
32. M.J. Berger and J. Oliger, Adaptive mesh refinement for hyperbolic partial differential equations. J. Comput. Phys. 53:3 (1984) 484–512.
33. M. Berzins, Modified mass matrices and positivity preservation for hyperbolic and parabolic
PDEs. Commun. Numer. Methods Engrg. 17:9 (2001) 659–666.
34. M. Berzins, Variable-order finite elements and positivity preservation for hyperbolic PDEs.
Appl. Numer. Math. 48:3-4 (2004) 271–292.
35. M. Berzins, Preserving positivity for hyperbolic PDEs using variable-order finite elements
with bounded polynomials. Appl. Numer. Math. 52 (2005) 197–217.
36. K.S. Bey and J.T. Oden, hp-version discontinuous Galerkin methods for hyperbolic conservation laws. Comput. Methods Appl. Mech. Engrg. 133 (1996) 259–286.
37. R. Biswas, K. Devine, and J. E. Flaherty, Parallel adaptive finite element methods for conservation laws. Appl. Numer. Math. 14 (1994) 255–284.
38. H. Blank, M. Rudgyard, A. Wathen, Stabilized finite element methods for steady incompressible flow. Comput. Methods Appl. Mech. Engrg. 174 (1999) 91–105.
39. J. Blasco, R. Codina, A. Huerta, Analysis of Fractional Step Finite Element Methods for the
Incompressible Navier-Stokes Equations. Monograph CIMNE 38, 1997.
40. H. Blum, J. Harig, S. Müller, P. Schreiber, S. Turek, FEAT2D/3D: Finite Element Analysis
Tools, User Manual, Release 1.3, University of Heidelberg, 1995.
41. D.L. Book, The conception, gestation, birth, and infancy of FCT. In: D. Kuzmin, R. Löhner,
S. Turek (eds), Flux-Corrected Transport: Principles, Algorithms, and Applications. Springer,
Berlin, 2005, 5–28.
42. J.P. Boris and D.L. Book, Flux-Corrected Transport: I. SHASTA, a fluid transport algorithm
that works. J. Comput. Phys. 11 (1973) 38–69.
43. D.L. Book, J.P. Boris, K. Hain, Flux-Corrected Transport: II. Generalizations of the Method.
J. Comput. Phys. 18 (1975) 248–283.
44. J.P. Boris and D.L. Book, Flux-Corrected Transport: III. Minimal-error FCT algorithms. J.
Comput. Phys. 20 (1976) 397–431.
45. J.P. Boris, F.F. Grinstein, E.S. Oran, R.J. Kolbe, New insights into Large Eddy Simulation.
Fluid Dynamics Research 10:4–6 (1992) 199–227.
References
207
46. J.H. Bramble and B.E. Hubbard, New monotone type approximations for elliptic problems.
Math. Comp 18 (1964) 349–367.
47. A.N. Brooks and T.J.R. Hughes, Streamline upwind Petrov-Galerkin formulations for convection dominated flows with particular emphasis on the incompressible Navier-Stokes equations. Comput. Methods Appl. Mech. Engrg. 32 (1982) 199–259.
48. A. Burbeau, P. Sagaut, and C.-H. Bruneau, A problem-independent limiter for high-order
Runge-Kutta discontinuous Galerkin methods. J. Comput. Phys. 169 (2001) 111–150.
49. E. Burman, A unified analysis for conforming and nonconforming stabilized finite element
methods using interior penalty. SIAM J. Numer. Anal. 43 (2005) 2012–2033.
50. E. Burman and P. Hansbo, Edge stabilization for Galerkin approximations of convectiondiffusion-reaction problems. Comput. Methods Appl. Mech. Engrg. 193 (2004) 1437–1453.
51. E. Burman and A. Ern, Nonlinear diffusion and discrete maximum principle for stabilized
Galerkin approximations of the convection-diffusion-reaction equation. Comput. Methods
Appl. Mech. Engrg. 191 (2002) 3833–3855.
52. E. Burman and A. Ern, Stabilized Galerkin approximation of convection-diffusion-reaction
problems: discrete maximum principle and convergence. Math. Comput. 74 (2005) 1637–
1652.
53. G. Cantin, C. Loubignac, C. Touzot, An iterative scheme to build continuous stress and displacement solutions. Int. J. Numer. Methods Engrg. 12 (1978) 1493–1506.
54. J.-C. Carette, H. Deconinck, H. Paillère, P.L. Roe, Multidimensional upwinding: Its relation
to finite elements. Int. J. Numer. Methods Fluids 20:8-9 (1995) 935–955.
55. G.F. Carey, Computational Grids: Generation, Adaptation, and Solution Strategies. Taylor &
Francis, 1997.
56. L. Catabriga and A.L.G.A. Coutinho, Implicit SUPG solution of Euler equations using edgebased data structures. Comput. Methods Appl. Mech. Engrg. 191 (2002) 3477–3490.
57. L. Catabriga, M.A.D. Martins, A.L.G.A. Coutinho, J.L.D. Alves, An Edge-based Preconditioner for Non-symmetric Finite Element Equations, In: proceedings of the 18th Ibero-Latin
American Congress on Computational Methods in Engineering. Brazil, October 29-31, 1997,
Vol. 3, 1249–1255.
58. L. Catabriga, M.A.D. Martins, A.L.G.A. Coutinho, J.L.D. Alves, Clustered edge-by-edge
preconditioners for non-symmetric finite element equations. In: CD-ROM proceedings of the
4th World Congress on Computational Mechanics. Buenos Aires, Argentina, 29 June to 2
July, 1998.
59. K.-Y. Chien, Predictions of channel and boundary-layer flows with a low-Reynolds-number
turbulence model. AIAA Journal 20 (1982) 33–38.
60. A.J. Chorin, Numerical solution of the Navier–Stokes equations. Math. Comp. 22 (1968)
745–762.
61. I. Christie and C. Hall, The maximum principle for bilinear elements. Int. J. Numer. Methods
Engrg. 20 (1984) 549–553.
62. M.A. Christon, P.M. Gresho, S.B. Sutton, Computational predictability of natural convection
flows in enclosures. In: K. J. Bathe (ed.), Computational Fluid and Solid Mechanics, Elsevier,
2001, 1465–1468.
63. P.G. Ciarlet, Discrete maximum principle for finite-difference operators. Aequationes Math.
4 (1970) 338–352.
64. P.G. Ciarlet and P.-A. Raviart, Maximum principle and convergence for the finite element
method. Comput. Methods Appl. Mech. Engrg. 2 (1973) 17–31.
65. R. Clift, J.R. Grace, M.E. Weber, Bubbles, Drops and Particles. Academic Press, 1978.
66. B. Cockburn, G.E. Karniadakis, C.-W. Shu, The development of discontinuous Galerkin
methods. In: Discontinuous Galerkin Methods. Theory, Computation and Applications,
LNCSE 11, Springer, 2000, 3–50.
67. B. Cockburn and C.-W. Shu, Runge-Kutta discontinuous Galerkin methods for convectiondominated problems. J. Sci. Comput. 16 (2001) 173–261.
68. B. Cockburn and C.-W. Shu, The Runge-Kutta discontinuous Galerkin method for conservation laws V. Multidimensional Systems. J. Comput. Phys. 141 (1998) 199–224.
208
References
69. R. Codina, A discontinuity-capturing crosswind-dissipation for the finite element solution of
the convection-diffusion equation. Comput. Methods Appl. Mech. Engrg. 110 (1993) 325–
342.
70. R. Codina, Comparison of some finite element methods for solving the diffusion-convectionreaction equation. Comput. Methods Appl. Mech. Engrg. 156:1–4 (1998) 185–210.
71. R. Codina and M. Cervera, Block-iterative algorithms for nonlinear coupled problems. In:
M. Papadrakakis and G. Bugeda (eds), Advanced Computational Methods in Structural
Mechanics, Chapter 7. Theory and Engineering Applications of Computational Methods,
CIMNE, Barcelona, 1996.
72. R. Codina and O. Soto, Finite element implementation of two-equation and algebraic stress
turbulence models for steady incompressible flows. Int. J. Numer. Methods Fluids 30:3 (1999)
309–333.
73. G. Cohen, P. Joly, J.E. Roberts, N. Tordjman, Higher order triangular finite elements with
mass lumping for the wave equation. SIAM J. Numer. Anal. 38:6 (2001) 2047–2078.
74. G. Comini, M. Manzan, C. Nonino, Analysis of finite element schemes for convection-type
problems. Int. J. Numer. Methods Fluids 20 (1995) 443–458.
75. M. Crouzeix and P.A. Raviart, Conforming and non-conforming finite elements for solving
the stationary Stokes equations. R.A.I.R.O. Anal. Numér. 7 (1973) 33–76.
76. C. Cuvelier, A. Segal, A.A. van Steenhoven, Finite Element Methods and Navier-Stokes
Equations. Kluwer, 1986.
77. J.F. Dannenhoffer and J.R. Baron, Robust grid adaptation for complex transonic flows. AIAA
Paper, 86-0495, 1986.
78. D.L. Darmofal and B. van Leer, Local preconditioning: Manipulating mother nature to fool
father time. In: D. A. Caughey et al. (eds), Frontiers of Computational Fluid Dynamics, Singapore: World Scientific Publishing, 1998, 211–239.
79. H. Deconinck, H. Paillère, R. Struijs, P.L. Roe, Multidimensional upwind schemes based on
fluctuation-splitting for systems of conservation laws. Comput. Mech. 11:5-6 (1993) 323–340.
80. C.R. DeVore, An improved limiter for multidimensional flux-corrected transport. NASA Report, AD-A360122, 1998.
81. M.J. Dı́az, F. Hecht, B. Mohammadi, New progress in anisotropic grid adaptation for inviscid and viscous flows simulations. In: Proceedings of the 4th Annual International Meshing
Roundtable. Sandia National Laboratories, 1995.
82. G.S. Dietachmayer, A comparison and evaluation of some positive definite advection
schemes. In: J. Noyle and R. May (eds), Computational Techniques and Applications, Elsevier, 1986, 217–232.
83. J. Donea, A Taylor-Galerkin method for convective transport problems. Int. J. Numer. Meth.
Engrg. 20 (1984) 101–120.
84. J. Donea, S. Giuliani, H. Laval, L. Quartapelle, Finite element solution of the unsteady NavierStokes equations by a fractional step method. Comput. Methods Appl. Mech. Engrg. 30 (1982)
53–73.
85. J. Donea, S. Giuliani, H. Laval, L. Quartapelle, Time-accurate solution of advection-diffusion
equations by finite elements. Comput. Methods Appl. Mech. Engrg. 193 (1984) 123–145.
86. J. Donea and A. Huerta, Finite Element Methods for Flow Problems. John Wiley & Sons,
Chichester, 2003.
87. J. Donea, V. Selmin, L. Quartapelle, Recent developments of the Taylor-Galerkin method for
the numerical solution of hyperbolic problems. Numerical methods for fluid dynamics III,
Oxford, 1988, 171–185.
88. J. Donea and L. Quartapelle, An introduction to finite element methods for transient advection
problems. Comput. Methods Appl. Mech. Engrg. 95 (1992) 169–203.
89. J. Donea, L. Quartapelle, V. Selmin, An analysis of time discretization in the finite element
solution of hyperbolic problems. J. Comput. Phys. 70 (1987) 463–499.
90. J. Donea, B. Roig, A. Huerta, High-Order Accurate Time-Stepping Schemes for ConvectionDiffusion Problems. Monograph CIMNE 42, Barcelona, 1998.
References
209
91. J. Douglas Jr. and T. Dupont, Interior penalty procedures for elliptic and parabolic Galerkin
methods. In: R. Glowinski and J.-L. Lions (eds) Computing Methods in Applied Sciences,
Lecture Notes in Physics, vol. 58, Springer, Berlin, 1976, 207–216.
92. J. Douglas, Jr. and J. Wang, An absolutely stabilized finite element method for the Stokes
problem. Math. Comp. 52 (1989) 495–508.
93. D.A. Drew and S.L. Passman, Theory of Multicomponent Fluids. Springer, 1999.
94. D. Drikakis and W. Rider, High-Resolution Methods for Incompressible and Low-Speed
Flows: Fundamentals and Applications. Springer, 2004.
95. J.-J. Droux and T.J.R. Hughes, A boundary integral modification of the Galerkin least squares
formulation for the Stokes problem. Comput. Methods Appl. Mech. Engrg. 113 (1994) 173–
182.
96. K. Duraisamy, K.D. Baeder, J.-G. Liu, Concepts and application of time-limiters to high
resolution schemes. J. of Sci. Comput. 19:1-3 (2003) 139–162.
97. M.S. Engelman, V. Haroutunian, I. Hasbani, Segregated finite element algorithms for the numerical solution of large–scale incompressible flow problems. Int. J. Numer. Methods Fluids
17 (1993) 323–348.
98. M.S. Engelman, R.L. Sani, P.M. Gresho, The implementation of normal and/or tangential
boundary conditions in finite element codes for incompressible fluid flow. Int. J. Numer. Methods Fluids 2 (1982) 225–238.
99. K. Eriksson, D. Estep, P. Hansbo, C. Johnson, Introduction to adaptive methods for differential equations. Acta Numerica 4 (1995) 105–158.
100. I. Faragó, R. Horváth, S. Korotov, Discrete maximum principle for linear parabolic problems
solved on hybrid meshes. Appl. Numer. Math. 53 (2005) 249–264.
101. I. Faragó and R. Horváth, Continuous and discrete parabolic operators and their qualitative
properties. IMA J. Numer. Anal. In press, doi:10.1093/imanum/drn032.
102. R.T. Farouki, T.N.T. Goodman, and T. Sauer, Construction of orthogonal bases for polynomials in Bernstein form on triangular and simplex domains. Computer Aided Geometric Design
20 (2003) 209–230.
103. M. Feistauer, J. Felcman, I. Straškraba, Mathematical and Computational Methods for Compressible Flow. Clarendon Press, Oxford, 2003.
104. J.H. Ferziger and M. Perić, Computational Methods for Fluid Dynamics. Springer, 1996.
105. J.H. Ferziger and M. Perić, Further discussion of numerical errors in CFD. Int. J. Numer.
Methods Fluids 23 (1996) 1–12.
106. FIDAP 8 Theory Manual, December 1998.
107. B. A. Finlayson. Numerical Methods for Problems with Moving Fronts. Ravenna Park Publishing, Seattle, 1992.
108. J.E. Flaherty, L. Krivodonova, J.-F. Remacle, and M.S. Shephard, Aspects of discontinuous
Galerkin methods for hyperbolic conservation laws. Finite Elements in Analysis and Design
38 (2002) 889–908.
109. C.A.J. Fletcher, The group finite element formulation, Comput. Methods Appl. Mech. Engrg.
37 (1983) 225–243.
110. C.A.J. Fletcher, A comparison of finite element and finite difference solutions of the oneand two-dimensional Burgers’ equations. J. Comput. Phys. 51 (1983) 159–188.
111. C.A.J. Fletcher, Computational Techniques for Fluid Dynamics. Springer, 1988.
112. P.A. Forsyth. A control volume finite element approach to NAPL groundwater contamination. SIAM J. Sci. Stat. Comput. 12 (1991) 1029–1057.
113. L.P. Franca, S.L. Frey, T.J.R. Hughes, Stabilized finite element methods: I. Application to
the advective-diffusive model. Comput. Methods Appl. Mech. Engrg. 95 (1992) 253–276.
114. L.A. Freitag, On combining Laplacian and optimization-based mesh smoothing techniques.
Trends in Unstructured Mesh Generation ASME Applied Mechanics Division, 220 (1997)
37–44.
115. P.J. Frey and P.L. George, Automatic Mesh Generation, Applications to Finite Methods. Hermes Science Publishing, Oxford, Paris, 2000.
210
References
116. H. Fujii, Some remarks on finite element analysis of time-dependent field problems. In:
Y. Yamada and R.H. Gallagher (eds), Theory and Practice in Finite Element Structural Analysis, Univ. Tokyo Press, Tokyo, 1973, 91–106.
117. A.C. Galeão and E.G.D. do Carmo, A consistent approximate upwind Petrov-Galerkin
Method for convection-dominated problems. Comput. Methods Appl. Mech. Engrg. 68:1
(1988) 83–95.
118. P.L. George, Automatic Mesh Generation, Applications to Finite Methods. John Wiley &
Sons, New York, 1991.
119. J. Geppert, Adaptive Gitterverfeinerung bei Flux-Limiter-Verfahren. PhD thesis, University
of Hamburg, 1996.
120. D. Gilbarg and N.S. Trudinger, Elliptic Partial Differential Equations of Second Order (2nd
edition). Grudlehren der Mathematischen Wissenschaften 224, Springer, 1983.
121. V. Girault and P.A. Raviart, Finite Element Methods for Navier–Stokes Equations, Springer,
Berlin-Heidelberg, 1986.
122. R. Glowinski, Finite Element Methods for Incompressible Viscous Flow. In: P.G. Ciarlet and
J.L. Lions (eds), Handbook of Numerical Analysis, Vol. IX: Numerical Methods for Fluids
(Part 3), North-Holland, Amsterdam, 2003, 3–1176.
123. S.K. Godunov, Finite difference method for numerical computation of discontinuous solutions of the equations of fluid dynamics. Mat. Sb. 47 (1959) 271–306.
124. J.B. Goodman and R.J. LeVeque, On the accuracy of stable schemes for 2D scalar conservation laws. Math. Comp. 45 (1985) 15–21.
125. S. Gottlieb and C.-W. Shu, Total Variation Diminishing Runge-Kutta schemes. Math. Comp.
67 (1998) 73–85.
126. S. Gottlieb, C.-W. Shu, and E. Tadmor, Strong stability-preserving high-order time discretization methods. SIAM Review 43 (2001) 89–112.
127. P.M. Gresho, On the theory of semi–implicit projection methods for viscous incompressible
flow and its implementation via a finite element method that also introduces a nearly consistent mass matrix, Part 1: Theory, Part 2: Implementation. Int. J. Numer. Methods Fluids 11
(1990) 587–659.
128. P.M. Gresho and R.L. Lee, Don’t suppress the wiggles — they’re telling you something!
In: T.J.R. Hughes (ed.), Finite Element Methods for Convection Dominated Flows, American
Society of Mechanical Engineers, New York, 1979, 37-61.
129. P.M. Gresho and R.L. Sani, Incompressible Flow and the Finite Element Method. Vol. 1:
Advection-diffusion, Vol. 2: Isothermal laminar flow. John Wiley & Sons, 2000.
130. F.F. Grinstein and C. Fureby, On monotonically integrated Large Eddy Simulation of turbulent flows based on FCT algorithms. In: D. Kuzmin, R. Löhner, S. Turek (eds) Flux-Corrected
Transport: Principles, Algorithms, and Applications. Springer, 2005, 79–104.
131. W. Hackbusch, Multigrid Methods and Applications. Springer, Berlin, 1985.
132. W. Hackbusch, Elliptic Differential Equations: Theory and Numerical Treatment. Springer,
Berlin, 1992
133. K. Hain, The partial donor cell method. J. Comput. Phys. 73 (1987) 131–147.
134. P. Hansbo, Aspects of conservation in finite element flow computations. Comput. Methods
Appl. Mech. Engrg. 117 (1994) 423–437.
135. P. Hansbo, A free-Lagrange finite element method using space-time elements. Comput. Methods Appl. Mech. Engrg. 188:1–3 (2000) 347–361.
136. P. Hansbo and C. Johnson, Adaptive streamline diffusion methods for compressible flow
using conservation variables. Comput. Methods Appl. Mech. Engrg. 87 (1991) 267–280.
137. A. Harten, High resolution schemes for hyperbolic conservation laws. J. Comput. Phys. 49
(1983) 357–393.
138. A. Harten, On a class of high resolution total-variation-stable finite-difference-schemes.
SIAM J. Numer. Anal. 21 (1984) 1-23.
139. A. Harten and J. Hyman, Self adjusting grid methods for one-dimensional hyperbolic conservation laws. J. Comput. Phys. 50 (1983) 235–269.
140. A. Harten, J. Hyman, P. Lax, On finite-difference approximations and entropy conditions for
shocks. Comm. Pure Appl. Math. 29 (1976) 297–322.
References
211
141. R. Hartmann, Adaptive FE methods for conservation equations. In: H. Freistühler and
G. Warnecke (eds.), Hyperbolic Problems: Theory, Numerics, Applications, ISNM 141.
Birkhäuser, Basel, 2001, 495-503.
142. R. Hartmann and P. Houston, Adaptive discontinuous Galerkin finite element methods for
the compressible Euler equations. J. Comput. Phys. 183 (2002) 508–532.
143. F.K. Hebeker and Yu. A. Kuznetsov, Unsteady convection and convection-diffusion problems
via direct overlapping domain decomposition methods. Numer. Meth. PDEs 14:3 (1998) 387–
406.
144. P.W. Hemker and B. Koren, Defect correction and nonlinear multigrid for steady Euler equations. In: W.G. Habashi and M.M. Hafez (eds), Computational Fluid Dynamics Techniques.
London: Gordon and Breach Publishers, 1995, 699–718.
145. D. Hempel, Rekonstruktionsverfahren auf unstrukturierten Gittern zur numerischen Simulation von Erhaltungsprinzipien. PhD thesis, University of Hamburg, 1999.
146. R. Herbin and F. Hubert, Benchmark on discretization methods for anisotropic diffusion
problems on general grids. In: R. Eymard and J.-M. Herard (eds), Finite Volumes for Complex
Applications V, 2008, 659–692.
147. J.S. Hesthaven and T. Warburton, Nodal Discontinuous Galerkin Methods: Algorithms, Analysis, and Applications. Springer Texts in Applied Mathematics 54, Springer, New York, 2008.
148. E. Hinton and J. Campbell, Local and global smoothing of discontinuous finite element functions using least squares method. Int. J. Numer. Methods Engrg. 8:3 (1974) 461–480.
149. C. Hirsch, Numerical Computation of Internal and External Flows. Vol. I: Fundamentals of
Numerical Discretization. John Wiley & Sons, Chichester, 1990.
150. C. Hirsch, Numerical Computation of Internal and External Flows. Vol. II: Computational
Methods for Inviscid and Viscous Flows. John Wiley & Sons, Chichester, 1990.
151. W. Höhn and H.-D. Mittelmann, Some remarks on the discrete maximum-principle for finite
elements of higher order. Computing 27:2 (1981) 145–154.
152. E. Hopf, Elementare Bemerkungen über die Lösungen partieller Differentialgleichungen
zweiter Ordnung vom elliptischen Typus, Sitzungsber. Preuss. Akad. Wissensch. 19 (1927)
147–152.
153. R. A. Horn and C.R. Johnson, Topics in Matrix Analysis. Cambridge University Press, Cambridge, 1991.
154. H. Hoteit, Ph. Ackerer, R. Mosé, J. Erhel, and B. Philippe, New two-dimensional slope limiters for discontinuous Galerkin methods on arbitrary meshes. Int. J. Numer. Meth. Engrg. 61
(2004) 2566–2593.
155. P. Houston, E. Süli, J.A. Mackenzie, G. Warnecke, A posteriori error analysis for numerical
approximations of Friedrichs systems. Numer. Math. 82:3 (1999) 433–470.
156. M.E. Hubbard, Non-oscillatory third order fluctuation splitting schemes for steady scalar
conservation laws. J. Comput. Phys. 222 (2007) 740–768.
157. T.J.R. Hughes and A. Brooks, A multidimensional upwind scheme with no crosswind diffusion. In: T.J.R. Hughes (ed.), Finite Element Methods for Convection Dominated Flows,
AMD 34, AMSE, New York, 1979, 19–35.
158. T.J.R. Hughes, G. Engel, L. Mazzei, M.G. Larson, The continuous Galerkin method is locally
conservative. J. Comput. Phys. 163:2 (2000) 467–488.
159. T.J.R. Hughes, L. Franca, G. Hulbert, A new finite element formulation for fluid dynamics:
VIII. The Galerkin least-squares method for advective-diffusive equations. Comput. Methods
Appl. Mech. Engrg. 73 (1989) 173–189.
160. T.J.R. Hughes, M. Mallet and A. Mizukami, A new finite element formulation for computational fluid dynamics. II. Beyond SUPG. Comput. Methods Appl. Mech. Engrg. 54:3 (1986)
341–355.
161. T.J.R. Hughes and M. Mallet, A new finite element formulation for computational fluid dynamics. IV. A discontinuity capturing operator for multidimensional advective-diffusive systems. Comput. Methods Appl. Mech. Engrg. 58 (1986) 329–336.
162. T.J.R. Hughes and G.N. Wells, Conservation properties for the Galerkin and stabilised forms
of the advection-diffusion and incompressible Navier-Stokes equations. Comput. Methods
Appl. Mech. Engrg. 194:9-11 (2005) 1141–1159. Erratum: Comput. Methods Appl. Mech.
Engrg. 195:9–12 (2006) 1277–1278.
212
References
163. W. Hundsdorfer and C. Montijn, A note on flux limiting for diffusion discretizations. IMA J.
of Numer. Anal. 24:4 (2004) 635–642.
164. W. Hundsdorfer and J.G. Verwer, Numerical Solution of Time-Dependent AdvectionDiffusion-Reaction Equations. Springer, 2003.
165. S. Hysing, S. Turek, D. Kuzmin, N. Parolini, E. Burman, S. Ganesan and L. Tobiska, Quantitative benchmark computations of two-dimensional bubble dynamics. Int. J. Numer. Meth.
Fluids. In press, doi: 10.1002/fld.1934
166. S. Idelsohn and E. Oñate, Finite elements and finite volumes. Two good friends. Int. J. Numer.
Methods Engrg. 37 (1994) 3323–3341.
167. T. Ikeda, Maximum Principle in Finite Element Models for Convection-Ciffusion Phenomena. North-Holland: Mathematics Studies 4, Kinokuniya, Tokyo, 1983.
168. F. Ilinca, J.-F. Hétu, D. Pelletier, A unified finite element algorithm for two-equation models
of turbulence. Comp. & Fluids 27:3 (1998) 291–310.
169. C. Ilinca, X.D. Zhang, J.Y. Trépanier, R. Camarero, A comparison of three error estimation
techniques for finite-volume solutions of compressible flows. Comput. Methods Appl. Engrg.
189:4 (2000) 1277–1294.
170. A. Jameson, Computational algorithms for aerodynamic analysis and design. Appl. Numer.
Math. 13 (1993) 383–422.
171. A. Jameson, Analysis and design of numerical schemes for gas dynamics 1. Artificial diffusion, upwind biasing, limiters and their effect on accuracy and multigrid convergence. Int.
Journal of CFD 4 (1995) 171–218.
172. V. John and P. Knobloch, On spurious oscillations at layers diminishing (SOLD) methods
for convection-diffusion equations: Part I - A review. Comput. Methods Appl. Mech. Engrg.
196:17–20 (2007) 2197–2215.
173. V. John and P. Knobloch, On spurious oscillations at layers diminishing (SOLD) methods
for convection-diffusion equations: Part II - Analysis for P1 and Q1 finite elements. Comput.
Methods Appl. Mech. Engrg. 197 (2008) 1997–2014.
174. V. John and E. Schmeyer, Finite element methods for time-dependent convection-diffusionreaction equations with small diffusion. Preprint, Saarland University, 2008.
http://www.math.uni-sb.de/ag/john/tcd.pdf
175. V. John and E. Schmeyer, On finite element methods for 3D time-dependent convectiondiffusion-reaction equations with small diffusion. Preprint, Saarland University, 2008.
http://www.math.uni-sb.de/ag/john/john schmeyer.bail2008.pdf
176. C. Johnson, Numerical Solution of Partial Differential Equations by the Finite Element
Method. Studentlitteratur, Lund, 1987
177. C. Johnson, A. Szepessy, P. Hansbo, On the convergence of shock-capturing streamline diffusion finite element methods for hyperbolic conservation laws. Math. Comput. 54 (1990)
107–129.
178. T. Jongen and Y.P. Marx, Design of an unconditionally stable, positive scheme for the K − ε
and two-layer turbulence models. Comput. Fluids 26 (1997) no. 5, 469-487.
179. J. Karátson, S. Korotov, M. Křı́žek, On discrete maximum principles for nonlinear elliptic
problems. Math. Comp. Simulation 76 (2007) 99–108.
180. J. Karátson and S. Korotov, Discrete maximum principles for finite element solutions of
nonlinear elliptic problems with mixed boundary conditions. Numer. Math. 99 (2005) 669–
698.
181. D.W. Kelly, S. Nakazawa, O.C. Zienkiewicz, J.C. Heinrich, A note on anisotropic balancing
dissipation in finite element approximation to convection diffusion problems, Int. J. Numer.
Methods Engrg. 15 (1980) 1705–1711.
182. J. Kim, Investigation of Separation and Reattachment of Turbulent Shear Layer: Flow over
a Backward Facing Step. PhD thesis, Stanford University, 1978.
183. J. Kim, P. Moin, R.D. Moser, Turbulence statistics in fully developed channel flow at low
Reynolds number. J. Fluid Mech., 177 (1987) 133–166.
184. B. Koren, A robust upwind discretization method for advection, diffusion and source terms.
In: C.B. Vreugdenhil et al. (eds), Numerical methods for advection - diffusion problems.
Braunschweig: Vieweg. Notes Numer. Fluid Mech. 45 (1993) 117–138.
References
213
185. S. Korotov, P. Neittaanmäki, S. Repin, A posteriori error estimation of goal-oriented quantities by the superconvergence patch recovery. J. Numer. Math. 11 (2003) 33-59.
186. L. Krivodonova, Limiters for high-order discontinuous Galerkin methods. J. Comput. Phys.
226 (2007) 879–896.
187. L. Krivodonova and M. Berger, High-order accurate implementation of solid wall boundary
conditions in curved geometries. J. Comput. Phys. 211 (2006) 492–512.
188. L. Krivodonova, J. Xin, J.-F. Remacle, N. Chevaugeon, and J.E. Flaherty, Shock detection and limiting with discontinuous Galerkin methods for hyperbolic conservation laws.
Appl. Numer. Math. 48 (2004) 323–338.
189. M. Křı́žek and Q. Lin, On diagonal dominance of stiffness matrices in 3D. East-West J.
Numer. Math. 3 (1995) 59–69.
190. M. Křı́žek and P. Neittaanmäki, On superconvergence techniques. Acta Appl. Math. 9 (1987)
175–198.
191. D. Kuzmin, Positive finite element schemes based on the flux-corrected transport procedure.
In: K. J. Bathe (ed.), Computational Fluid and Solid Mechanics, Elsevier, 2001, 887–888.
192. D. Kuzmin, On the design of general-purpose flux limiters for implicit FEM with a consistent
mass matrix. I. Scalar convection. J. Comput. Phys. 219 (2006) 513–531.
193. D. Kuzmin, On the design of algebraic flux correction schemes for quadratic finite elements.
J. Comput. Appl. Math. 218:1 (2008) 79–87.
194. D. Kuzmin, Algebraic flux correction for finite element discretizations of coupled systems.
In: E. Oñate, M. Papadrakakis, B. Schrefler (eds) Computational Methods for Coupled Problems in Science and Engineering II, CIMNE, Barcelona, 2007, 653–656.
195. D. Kuzmin, A vertex-based hierarchical slope limiter for p-adaptive discontinuous Galerkin
methods. J. Comput. Appl. Math. 233 (2010) 3077–3085.
196. D. Kuzmin, Explicit and implicit FEM-FCT algorithms with flux linearization. J. Comput.
Phys. In press, doi:10.1016/j.jcp.2008.12.011
197. D. Kuzmin and S. Korotov, Goal-oriented a posteriori error estimates for transport problems.
Math. Comput. Simul. 80 (2010) 2674-1683.
198. D. Kuzmin, O. Mierka, S. Turek, On the implementation of the k − ε turbulence model in
incompressible flow solvers based on a finite element discretization. Int. J. Comp. Sci. Math.
1:2–4 (2007) 193-206.
199. D. Kuzmin and M. Möller, Goal-oriented mesh adaptation for flux-limited approximations
to steady hyperbolic problems. J. Comput. Appl. Math. 233 (2010) 3113–3120.
200. D. Kuzmin and M. Möller, Algebraic flux correction I. Scalar conservation laws. In:
D. Kuzmin, R. Löhner, S. Turek (eds) Flux-Corrected Transport: Principles, Algorithms, and
Applications. Springer, 2005, 155–206.
201. D. Kuzmin and M. Möller, Algebraic flux correction II. Compressible Euler equations. In:
D. Kuzmin, R. Löhner, S. Turek (eds) Flux-Corrected Transport: Principles, Algorithms, and
Applications. Springer, 2005, 207–250.
202. D. Kuzmin, M. Möller and S. Turek, Multidimensional FEM-FCT schemes for arbitrary
time-stepping. Int. J. Numer. Methods Fluids 42 (2003) 265–295.
203. D. Kuzmin, M. Möller, S. Turek, High-resolution FEM-FCT schemes for multidimensional
conservation laws. Computer Methods Appl. Mech. Engrg. 193 (2004) 4915–4946.
204. D. Kuzmin, M. Shashkov, D. Svyatskiy, A constrained finite element method satisfying the
discrete maximum principle for anisotropic diffusion problems on arbitrary meshes. Appl.
Math. Report 373, University of Dortmund, 2008. Submitted to J. Comput. Phys.
205. D. Kuzmin and S. Turek, Flux correction tools for finite elements. J. Comput. Phys. 175
(2002) 525–558.
206. D. Kuzmin and S. Turek, High-resolution FEM-TVD schemes based on a fully multidimensional flux limiter. J. Comput. Phys. 198 (2004) 131–158.
207. D. Kuzmin and S. Turek, Multidimensional FEM-TVD paradigm for convection-dominated
flows. In: P. Neittaanmäki, T. Rossi, K. Majava, O. Pironneau (eds) CD-ROM Proceedings of
the IV European Congress on Computational Methods in Applied Sciences and Engineering
(ECCOMAS 2004), Jyväskylä, Finland, 24-28 July 2004. Vol. II, ISBN 951-39-1869-6.
214
References
208. D. Kuzmin and S. Turek, Numerical simulation of turbulent bubbly flows. In: G.P. Celata,
P. Di Marco, A. Mariani, R.K. Shah (eds) Two-Phase Flow Modeling and Experimentation.
Edizioni ETS, Pisa, 2004, Vol. I, 179-188.
209. Yu. A. Kuznetsov, New algorithms for approximate realization of implicit difference schemes. Soviet. J. Numer. Anal. Math. Modelling 3 (1988) 99–114.
210. Yu. A. Kuznetsov, Domain decomposition methods for unsteady convection diffusion problems. Comput. Methods Appl. Sci. Engin. (Proceedings of the Ninth International Conference,
Paris 1990) SIAM, Philadelphia (1990) 211–227.
211. O.A. Ladyzhenskaya and N.N. Ural’tseva, Linear and Quasilinear Elliptic Equations. Academic Press, New York, 1968.
212. R.J. Labeur and G.N. Wells, A Galerkin interface stabilisation method for the advectiondiffusion and incompressible Navier-Stokes equations. Computer Meth. Appl. Mech. Engrg.
196 (2007) 4985–5000.
213. M. Laforest, M.A. Christon, T.E. Voth, A survey of error indicators and error estimators for
hyperbolic problems. Available online at
http://www.mgi.polymtl.ca/marc.laforest/pages/sand rep.ps.
214. A. Lapin, University of Stuttgart. Private communication, 2001.
215. C. Le Potier, Schema volumes finis monotone pour des operateurs de diffusion fortement
anisotropes sur des maillages de triangle non structures. C.C. Acad. Sci. Paris, Ser. I, 341
(2005) 787–792.
216. R.J. LeVeque, Numerical Methods for Conservation Laws. Birkhäuser, 1992.
217. R.J. LeVeque, CLAWPACK – a software package for solving multi-dimensional conservation laws. In: proceedings of the 5th International Conference on Hyperbolic Problems,
http://www.amath.washington.edu/∼claw/, 1994.
218. R.J. LeVeque, High-resolution conservative algorithms for advection in incompressible flow.
SIAM J. Numer. Anal. 33 (1996) 627–665.
219. R.J. LeVeque, Finite Volume Methods for Hyperbolic Problems, Cambridge University Press,
2002.
220. A.J. Lew, G.C. Buscaglia, P.M. Carrica, A note on the numerical treatment of the k − ε
turbulence model. Int. J. of Comp. Fluid Dyn. 14 (2001) 201–209.
221. K. Lipnikov, M. Shashkov, D. Svyatskiy, Yu. Vassilevski, Monotone finite volume schemes
for diffusion equations on unstructured triangular and shape-regular polygonal meshes. J.
Comput. Phys. 227 (2007) 492–512.
222. R. Liska and M. Shashkov, Enforcing the discrete maximum principle for linear finite element solutions of second-order elliptic problems. Commun. Comput. Phys. 3:4 (2007) 852–
877.
223. R. Löhner, An adaptive finite element scheme for transient problems in CFD. Comput. Methods Appl. Mech. Engrg. 61 (1987) 323–338.
224. R. Löhner, Adaptive remeshing for transient problems. Comput. Methods Appl. Mech. Engrg.
75 (1989) 195–214.
225. R. Löhner, Edges, stars, superedges and chains. Comput. Methods Appl. Mech. Engrg. 111
(1994) 255–263.
226. R. Löhner, Applied CFD Techniques: An Introduction Based on Finite Element Methods (2nd
edition). John Wiley & Sons, Chichester, 2008.
227. R. Löhner and J.D. Baum, Adaptive h-refinement on 3-D unstructured grids for transient
problems. Int. J. Numer. Methods Fluids 14 (1992) 1407–1419.
228. R. Löhner and J.D. Baum, 30 Years of FCT: Status and Directions. In: D. Kuzmin, R. Löhner,
S. Turek (eds) Flux-Corrected Transport: Principles, Algorithms, and Applications. Springer,
2005, 131–154.
229. R. Löhner and M. Galle, Minimization of indirect addressing for edge-based field solvers.
Commun. Numer. Methods Eng. 18:5 (2002) 335–343.
230. R. Löhner, K. Morgan, O.C. Zienkiewicz, The solution of non-linear hyperbolic equation
systems by the finite element method. Int. J. Numer. Methods Fluids 4 (1984) 1043–1063.
231. R. Löhner, K. Morgan, O.C. Zienkiewicz, An adaptive finite element procedure for compressible high speed flows. Comput. Methods Appl. Mech. Engrg. 51 (1985) 441–465.
References
215
232. R. Löhner, K. Morgan, J. Peraire, M. Vahdati, Finite element flux-corrected transport (FEMFCT) for the Euler and Navier-Stokes equations. Int. J. Numer. Meth. Fluids 7 (1987) 1093–
1109.
233. R. Löhner, K. Morgan, M. Vahdati, J.P. Boris, D.L. Book, FEM-FCT: combining unstructured grids with high resolution. Commun. Appl. Numer. Methods 4 (1988) 717–729.
234. R. Löhner, H. Luo, J.D. Baum, Selective edge removal for unstructured grids with Cartesian
cores. J. Comput. Phys. 206:1 (2005) 208–226.
235. H. Luo, J.D. Baum, R. Löhner, Edge-based finite element scheme for the Euler equations.
AIAA Journal 32 (1994) 1183–1190.
236. H. Luo, J.D. Baum, and R. Löhner, Fast p-multigrid discontinuous Galerkin method for
compressible flows at all speeds. AIAA Journal 46 (2008) 635–652.
237. H. Luo, J.D. Baum, and R. Löhner, A discontinuous Galerkin method based on a Taylor basis
for the compressible flows on arbitrary grids. J. Comput. Phys. 227 (2008) 8875–8893.
238. J.F. Lynn, Multigrid Solution of the Euler Equations with Local Preconditioning. PhD thesis,
University of Michigan, 1995.
239. P.R.M. Lyra, Unstructured Grid Adaptive Algorithms for Fluid Dynamics and Heat Conduction. PhD thesis, University of Wales, Swansea, 1994.
240. P.R.M. Lyra, O. Hassan, K. Morgan, Adaptive unstructured grid solutions of hypersonic viscous flows. In: K.W. Morton and M.J. Baines (Eds) Numerical Methods for Fluid Dynamics
V, Oxford University Press, Oxford, 1996, 465–472.
241. P.R.M. Lyra and K. Morgan, A review and comparative study of upwind biased schemes for
compressible flow computation. I: 1-D first-order schemes. Arch. Comput. Methods Eng. 7:1
(2000) 19–55.
242. P.R.M. Lyra and K. Morgan, A review and comparative study of upwind biased schemes for
compressible flow computation. II: 1-D higher-order schemes. Arch. Comput. Methods Eng.
7:3 (2000) 333–377.
243. P.R.M. Lyra and K. Morgan, A review and comparative study of upwind biased schemes for
compressible flow computation. III: Multidimensional extension on unstructured grids. Arch.
Comput. Methods Eng. 9:3 (2002) 207–256.
244. P.R.M. Lyra, K. Morgan, J. Peraire, J. Peiro, TVD algorithms for the solution of the compressible Euler equations on unstructured meshes. Int. J. Numer. Methods Fluids 19 (1994)
827–847.
245. P.R.M. Lyra, R.B. Willmersdorf, M.A.D. Martins, A.L.G.A. Coutinho, Parallel implementation of edge-based finite element schemes for compressible flow on unstructured grids, Proceedings of the 3rd International Meeting on Vector and Parallel Processing, Faculdade de
Engenharia da Universidade do Porto, Porto, Portugal, 21.-23. Juni 1998.
246. R.J. MacKinnon and G.F. Carey, Positivity preserving flux-limited finite-difference and
finite-element methods for reactive transport. Int. J. Numer. Methods Fluids 41 (2003) 151–
183.
247. G. Medić and B. Mohammadi, NSIKE - an incompressible Navier-Stokes solver for unstructured meshes. INRIA Research Report 3644, 1999.
248. D. Meidner, R. Rannacher and J. Vihharev, Goal-oriented error control of the iterative solution of finite element equations. J. Numer. Math. 17 (2009) 143–172.
249. J.A. Meijerink and H.A. van der Vorst, Iterative solution method for linear systems of which
the coefficient matrix is a symmetric M-matrix. Math. Comput. 31 (1977) 148–162.
250. K. Mer, Variational analysis of a mixed element/volume scheme with fourth-order viscosity
on general triangulations. Comput. Methods Appl. Mech. Engrg. 153 (1998) 45–62.
251. K. Michalak and C. Ollivier-Gooch, Limiters for unstructured higher-order accurate solutions
of the Euler equations. In: Proceedings of the AIAA Forty-Sixth Aerospace Sciences Meeting,
2008.
252. B. Mohammadi and O. Pironneau, Analysis of the k-epsilon Turbulence Model. John Wiley
& Sons, 1994.
253. B. Mohammadi and O. Pironneau, Applied Shape Optimization for Fluids. Numerical Mathematics and Scientific Computation. Oxford University Press, Oxford, 2001.
216
References
254. M. Möller, Efficient solution techniques for implicit finite element schemes with flux limiters.
Int. J. Numer. Methods Fluids 55:7 (2007) 611–635.
255. M. Möller, Adaptive High-Resolution Finite Element Schemes. PhD thesis, Dortmund University of Technology, 2008.
256. M. Möller and D. Kuzmin, Adaptive mesh refinement for high-resolution finite element
schemes. Int. J. Numer. Methods Fluids 52 (2006) 545–569.
257. M. Möller and D. Kuzmin, On the use of slope limiters for the design of recovery based error
indicators. In: A. Bermudez de Castro, D. Gomez, P. Quintela, P. Salgado (eds), Numerical
Mathematics and Advanced Applications, Springer, 2006, 233–240.
258. M. Möller, D. Kuzmin, D. Kourounis, Implicit FEM-FCT algorithms and discrete Newton
methods for transient convection problems. Int. J. Numer. Methods Fluids 57:6 (2008) 761–
792.
259. K. Morgan and J. Peraire, Unstructured grid finite element methods for fluid mechanics.
Reports on Progress in Physics. 61:6 (1998) 569–638.
260. K.W. Morton, Numerical Solution of Convection-Diffusion Problems. Chapman & Hall,
1996.
261. J.T. Oden and H.J. Brauchli, On the calculation of consistent stress distributions in finite
element approximations. Int. J. Numer. Methods Engrg. 3 (1971) 317–325.
262. J.T. Oden and S. Prudhomme, Goal-oriented error estimation and adaptivity for the finite
element methods. Comput. Math. Appl. 41 (2000) 735–756.
263. J.T. Oden and J.N. Reddy, Note on an approximate method for computing consistent conjugate stresses in elastic finite elements. Int. J. Numer. Methods Engrg. 6 (1973) 55–61.
264. J.T. Oden, T. Strouboulis, P. Devloo, Adaptive finite element methods for the analysis of
inviscid compressible flow: Part I. Fast refinement/unrefinement and moving mesh methods
for unstructured meshes. Comput. Methods Appl. Mech. Engrg. 59 (1986) 327–362.
265. E.S. Oran and J.P. Boris, Numerical Simulation of Reactive Flow (2nd edition). Cambridge
University Press, 2001.
266. A.K. Parrott, M.A. Christie, FCT applied to the 2-D finite element solution of tracer transport
by single phase flow in a porous medium. In: Numerical Methods for Fluid Dynamics, Oxford
Univ. Press, London, 1986, 609–619.
267. G. Patnaik, J.P. Boris, F.F. Grinstein, J.P. Iselin, Large scale urban simulations with FCT.
In: D. Kuzmin, R. Löhner, S. Turek (eds), Flux-Corrected Transport: Principles, Algorithms,
and Applications. Springer, 2005, 105–130.
268. S.V. Patankar, Numerical Heat Transfer and Fluid Flow. McGraw-Hill, New York, 1980.
269. J. Peraire, M. Vahdati, J. Peiro, K. Morgan, The construction and behaviour of some unstructured grid algorithms for compressible flows. Numerical Methods for Fluid Dynamics IV,
Oxford University Press, 1993, 221-239.
270. O. Pironneau, Finite Element Methods for Fluids. John Wiley & Sons, Chichester; Masson,
Paris, 1989.
271. L. Postma and J.-M. Hervouet, Compatibility between finite volumes and finite elements
using solutions of shallow water equations for substance transport. Int. J. Numer. Methods
Fluids 53 (2007) 1495–1507.
272. A. Prohl, Projection and Quasi-Compressibility Methods for Solving the Incompressible
Navier-Stokes Equations. Advances in Numerical Mathematics. B.G. Teubner, Stuttgart,
1997.
273. M.H. Protter and H.F. Weinberger, Maximum Principles in Differential Equations. PrenticeHall, 1967.
274. S. Prudhomme and J.T. Oden, Computable error estimators. In: T.J. Barth and H. Deconinck (eds), Error Estimation and Adaptive Discretization Methods in Computational Fluid
Dynamics, LNCSE 25. Springer, Berlin, 2002, 207–268.
275. L. Quartapelle, Numerical Solution of the Incompressible Navier-Stokes Equations. ISNM
113, Birkäuser, Basel, 1993.
276. A. Quarteroni and A. Valli, Numerical Approximation of Partial Differential Equations.
Springer, Berlin, 1994.
References
217
277. A. Quarteroni and A. Valli, Domain Decomposition Methods for Partial Differential Equations. Oxford University Press, Oxford, 1999.
278. R. Rannacher and S. Turek, A simple nonconforming quadrilateral Stokes element. Numer.
Meth. PDEs 8:2 (1992) 97–111.
279. W. J. Rider, Methods for extending high-resolution schemes to non-linear systems of hyperbolic conservation laws. Int. J. Numer. Methods Fluids 17:10 (1993) 861–885.
280. M.C. Rivara, Design and data structure of fully adaptive multigrid finite element software.
ACM Trans. Math. Software 10 (1984) 242–264.
281. M.C. Rivara, New mathematical tools and techniques for the refinement and/or improvement
of unstructured triangulations. Proc. 5th Int. Meshing Roundtable 96, Pittsburgh, 1996: 77-86.
282. M.C. Rivara, Using longest-side bisection techniques for the automatic refinement of delaunay triangulations. Int. J. Numer. Methods Engrg. 40 (1997) 581–597.
283. P.J. Roache, A method for uniform reporting of grid refinement studies. ASME J. Fluids
Engrg. 116 (1994) 405–413.
284. P.J. Roache, Verification and Validation in Computational Science and Engineering. Hermosa Publishers, New Mexico, 1998.
285. P.L. Roe, Approximate Riemann solvers, parameter vectors and difference schemes. J. Comput. Phys. 43 (1981) 357–372.
286. A. Rohde, Eigenvalues and eigenvectors of the Euler equations in general geometries. AIAA
Paper, 2001-2609, 2001.
287. C.J. Roy, Grid convergence error analysis for mixed-order numerical schemes. AIAA Journal
41:4 (2003) 595–604.
288. H.G. Roos, M. Stynes, L. Tobiska, Numerical Methods for Singularly Perturbed Differential
Equations. Springer, 1996.
289. A. Samarskii and P. Vabishchevich, Numerical Methods for Solving Convection-Diffusion
Problems (in Russian). Editorial URSS, Moscow, 1999.
290. P.A.B. de Sampaio and A.L.G.A. Coutinho, A natural derivation of discontinuity capturing
operator for convection-diffusion problems. Comput. Methods Appl. Mech. Engrg. 190 (2001)
6291–6308.
291. E. Schall, D. Leservoisier, A. Dervieux, B. Koobus, Mesh adaptation as a tool for certified
computational aerodynamics. Int. J. Numer. Methods Fluids 45 (2004) 179–196.
292. M. Schäfer and S. Turek (with support of F. Durst, E. Krause, R. Rannacher), Benchmark
computations of laminar flow around cylinder, In E.H. Hirschel (ed.), Flow Simulation with
High-Performance Computers II, Vol. 52 von Notes on Numerical Fluid Mechanics, Vieweg,
1996, 547–566.
293. F. Schieweck, On the role of boundary conditions for CIP stabilization of higher order finite
elements. Electronic Transactions on Numerical Analysis 32 (2008) 1–16.
294. R. Schmachtel, Robuste lineare und nichtlineare Lösungsverfahren für die inkompressiblen
Navier-Stokes-Gleichungen. PhD thesis, University of Dortmund, 2003.
295. M. Schmich and B. Vexler, Adaptivity with dynamic meshes for space-time finite element
discretizations of parabolic equations. SIAM J. Sci. Comput. 30:1 (2008) 369–393.
296. P. Schreiber, A new finite element solver for the nonstationary incompressible Navier–Stokes
equations in three dimensions, PhD thesis, University of Heidelberg, 1996.
297. P. Schreiber and S. Turek, An efficient finite element solver for the nonstationary incompressible Navier–Stokes equations in two and three dimensions. Proc. Workshop Numerical
Methods for the Navier–Stokes Equations, Heidelberg, Oct. 25–28, 1993, Vieweg.
298. V. Selmin, Finite element solution of hyperbolic equations. I. One-dimensional case. INRIA
Research Report 655, 1987.
299. V. Selmin, Finite element solution of hyperbolic equations. I I. Two-dimensional case. INRIA
Research Report 708, 1987.
300. V. Selmin, The node-centred finite volume approach: bridge between finite differences and
finite elements. Comput. Methods Appl. Mech. Engrg. 102 (1993) 107–138.
301. V. Selmin and L. Formaggia, Unified construction of finite element and finite volume discretizations for compressible flows. Int. J. Numer. Methods Engrg. 39 (1996) 1–32.
218
References
302. V. Selmin and L. Quartapelle, A unified approach to build artificial dissipation operators for
finite element and finite volume discretisations. In: K. Morgan et al. (eds), Finite Elements in
Fluids, CIMNE / Pineridge Press, 1993, 1329–1341.
303. R.A. Shapiro, Adaptive Finite Element Solution Algorithm for the Euler Equations. Notes on
Numerical Fluid Mechanics, Vol. 32. Vieweg, Braunschweig-Wiesbaden, 1991.
304. C.-W. Shu, Total-variation-diminishing time discretizations. SIAM J. Sci. Stat. Comput. 9
(1988) 1073–1084.
305. C.-W. Shu, A Survey of Strong Stability Preserving High Order Time Discretizations. Scientific Computing Report Series 2001-18, Brown University, 2001.
306. C.-W. Shu and S. Osher, Efficient implementation of essentially non-oscillatory shockcapturing schemes. J. Comput. Phys. 77 (1988) 439–471.
307. T.M. Smith, R.W. Hooper, C.C. Ober, A.A. Lorber, Intelligent Nonlinear Solvers for Computational Fluid Dynamics. Conference Paper, Presentation at the 44th AIAA Aerospace Sciences
Meeting and Exhibit, Reno NV, January 2006.
308. G. Sod, A survey of several finite difference methods for systems of nonlinear hyperbolic
conservation laws. J. Comput. Phys. 27 (1978) 1–31.
309. P. Šolin and L. Demkowicz, Goal-oriented hp-adaptivity for elliptic problems. Comput.
Methods Appl. Mech. Engrg. 193 (2004) 449–468.
310. C. Steiner and S. Noelle, On adaptive timestepping for weakly instationary solutions of hyperbolic conservation laws via adjoint error control. Commun. Numer. Meth. Engrg. (2009),
in press, doi:10.1002/cnm.1183.
311. G. Stoyan, On maximum principles for monotone matrices. Linear Algebra Appl. 78 (1986)
147–161.
312. E. Süli, A-posteriori error analysis and adaptivity for finite element approximations of hyperbolic problems. In: M. Ohlberger, D. Kröner, C. Rohde (eds), An Introduction to Recent
Developments in Theory and Numerics for Conservation Laws, LNCSE 5, Springer, 1998,
123–194.
313. E. Süli, P. Houston, B. Senior, hp-discontinuous Galerkin finite element methods for nonlinear hyperbolic problems. Int. J. Numer. Methods Fluids 40:1–2 (2001) 153–169.
314. E. Süli and P. Houston, Adaptive finite element approximation of hyperbolic problems. In:
T.J. Barth and H. Deconinck (eds) Error Estimation and Adaptive Discretization Methods in
Computational Fluid Dynamics. LNCSE 25. Springer, Berlin, 2002, 269–344.
315. P.K. Sweby, High resolution schemes using flux limiters for hyperbolic conservation laws.
SIAM J. Numer. Anal. 21 (1984) 995–1011.
316. S. Thangam and C.G. Speziale, Turbulent flow past a backward-facing step: A critical evaluation of two-equation models. AIAA Journal 30:5 (1992) 1314–1320.
317. J. F. Thompson, B. Soni, and N. Weatherill (Eds) Handbook of Grid Generation. CRC Press
LLC , Boca Raton, FL, 1998,
318. J.F. Thompson, Z.U.A. Warsi, C.W. Mastin, Numerical Grid Generation: Foundations and
Applications. North-Holland, Amsterdam, 1985.
319. E.F. Toro, Riemann Solvers and Numerical Methods for Fluid Dynamics. Springer, 1999.
320. S. Tu and S. Aliabadi, A slope limiting procedure in discontinuous Galerkin finite element
method for gasdynamics applications. Int. J. Numer. Anal. Model. 2 (2005) 163–178.
321. S. Turek, On discrete projection methods for the incompressible Navier-Stokes equations:
An algorithmical approach. Comput. Methods Appl. Mech. Engrg. 143 (1997) 271–288.
322. S. Turek, Efficient Solvers for Incompressible Flow Problems: An Algorithmic and Computational Approach, LNCSE 6, Springer, 1999.
323. S. Turek et al., FEATFLOW: finite element software for the incompressible Navier-Stokes
equations. User manual, University of Dortmund, 2000. Available at http://www.featflow.de.
324. S. Turek, C. Becker, S. Kilian, Hardware-oriented numerics and concepts for PDE software.
Special Journal Issue for PDE Software, Elsevier, International Conference on Computational
Science ICCS’2002, Amsterdam, 2002, FUTURE 1095 (2003), 1-23.
325. S. Turek, C. Becker, S. Kilian, Some concepts of the software package FEAST. In:
J.M. Palma, J. Dongarra, V. Hernandes (eds), VECPAR’98 - Third International Conference
for Vector and Parallel Processing, Lecture Notes in Computer Science, Springer, Berlin,
1999.
References
219
326. S. Turek and S. Kilian, An example for parallel ScaRC and its application to the incompressible Navier-Stokes equations, Proc. ENUMATH’97, World Science Publ., 1998.
327. S. Turek and A. Ouazzi, Unified edge-oriented stabilization of nonconforming FEM for incompressible flow problems: Numerical investigation. J. Numer. Math. 15:4 (2007) 299–322.
328. S. Turek and R. Schmachtel, Fully coupled and operator-splitting approaches for natural
convection. Int. Numer. Meth. Fluids 40 (2002) 1109–1119.
329. S. Turek and D. Kuzmin, Algebraic flux correction III. Incompressible flow problems. In:
D. Kuzmin, R. Löhner, S. Turek (eds), Flux-Corrected Transport: Principles, Algorithms,
and Applications. Springer, 2005, 251–296.
330. UMFPACK web site, http://www.cise.ufl.edu/research/sparse/umfpack/.
331. P. Vabishchevich, Monotone difference schemes for convection/diffusion problems. Differential Equations 30:3 (1994) 466–474.
332. P. Vabishchevich, Iterative methods for solving convection-diffusion problem. Comput. Meth.
Appl. Math. 2:4 (2002) 410–444.
333. A.M.P. Valli, G.F. Carey, A.L.G.A. Coutinho, Finite element simulation and control of nonlinear flow and reactive transport. In: Proceedings of the 10th International Conference on
Numerical Methods in Fluids. Tucson, Arizona, 1998: 450–455.
334. A.M.P. Valli, G.F. Carey, A.L.G.A. Coutinho, Control strategies for timestep selection in
simulation of coupled viscous flow and heat transfer. Commun. Numer. Methods Eng. 18:2
(2002) 131–139.
335. S.P. Vanka, Implicit multigrid solutions of Navier–Stokes equations in primitive variables. J.
Comput. Phys. 65 (1985) 138–158.
336. J. van Kan, A second–order accurate pressure–correction scheme for viscous incompressible
flow. SIAM J. Sci. Stat. Comp. 7 (1986) 870–891.
337. P. van Slingerland, An Accurate and Robust Finite Volume Method for the Advection Diffusion Equation. M.Sc. thesis, Delft University of Technology, June 2007.
338. P. van Slingerland, M. Borsboom, C. Vuik, A local theta scheme for advection problems
with strongly varying meshes and velocity profiles. Report 08-17, Department of Applied
Mathematical Analysis, Delft University of Technology, June 2008.
339. R.S. Varga, Matrix Iterative Analysis. Prentice-Hall, Englewood Cliffs, 1962.
340. R.S. Varga, On discrete maximum principle. SIAM J. Numer. Anal. 3 (1966) 355–359.
341. T. Vejchodsky and P. Solin, Discrete maximum principle for higher-order finite elements in
1D. Math. Comp. 76 (2007) 1833–1846.
342. V. Venkatakrishnan, Convergence to steady state solutions of the Euler equations on unstructured grids with limiters. J. Comput. Phys. 118 (1995) 120–130.
343. R. Verfürth, A Review of a Posteriori Error Estimation and Adaptive Mesh-Refinement Techniques. Wiley-Teubner, 1996.
344. P. Wesseling, Principles of Computational Fluid Dynamics. Springer, Berlin-Heidelberg,
2001.
345. P. Wesseling, An Introduction to Multigrid Methods. John Wiley & Sons, Chichester, 1992.
Corrected Reprint. Philadelpha: R.T. Edwards, Inc., 2004.
346. G. Windisch, M-matrices in Numerical Analysis. Teubner, Leipzig, 1989.
347. P.R. Woodward and P. Colella, The numerical simulation of two-dimensional fluid flow with
strong shocks. J. Comput. Phys. 54 (1984) 115–173.
348. J. Xu and L. Zikatanov, A monotone finite element scheme for convection-diffusion equations. Math. Comput. 68 (1999) 1429–1446.
349. M. Xue, High-order monotonic numerical diffusion and smoothing. Monthly Weather Review
128:8 (2000) 2853–2864.
350. M. Yang and Z.J. Wang, A parameter-free generalized moment limiter for high-order methods on unstructured grids, AIAA-2009-605.
351. H.C. Yee, Numerical approximations of boundary conditions with applications to inviscid
gas dynamics. NASA Report TM-81265, 1981.
352. H.C. Yee, Construction of explicit and implicit symmetric TVD schemes and their applications. J. Comput. Phys. 43 (1987) 151–179.
220
References
353. H.C. Yee, R.F. Warming, A. Harten, Implicit Total Variation Diminishing (TVD) schemes
for steady-state calculations. J. Comput. Phys. 57 (1985) 327–360.
354. D. Young, Iterative Solution of Large Linear Systems, Academic Press, New York, 1971.
355. S.T. Zalesak, Fully multidimensional flux-corrected transport algorithms for fluids. J. Comput. Phys. 31 (1979) 335–362.
356. S.T. Zalesak, A preliminary comparison of modern shock-capturing schemes: linear advection. In: R. Vichnevetsky and R. Stepleman (eds), Advances in Computer Methods for PDEs.
Publ. IMACS, 1987, 15–22.
357. S.T. Zalesak, The design of Flux-Corrected Transport (FCT) algorithms for structured grids.
In: D. Kuzmin, R. Löhner, S. Turek (eds), Flux-Corrected Transport: Principles, Algorithms,
and Applications. Springer, 2005, 29–78.
358. X.D. Zhang, D. Pelletier, J.-Y. Tr’epanier, R. Camarero, Numerical assessment of error estimators for Euler equations. AIAA Journal 39:9 (2001) 1706–1715.
359. Z. Zhang and A. Naga, A posteriori error estimates based on polynomial preserving recovery.
SIAM J. Numer. Anal. 42:4 (2004) 1780–1800.
360. Z. Zhang and A. Naga, A new finite element gradient recovery method: Superconvergence
property. SIAM J. Sci. Comput. 26:4 (2005) 1192–1213.
361. O.C. Zienkiewicz and R.L. Taylor, The Finite Element Method. Butterworth-Heinemann,
Oxford, 2000.
362. O.C. Zienkiewicz and J.Z. Zhu, A simple error estimator and adaptive procedure for practical
engineering analysis. Int. J. Numer. Methods Engrg. 24:2 (1987) 337–357.
363. O.C. Zienkiewicz and J.Z. Zhu, The superconvergent patch recovery and a posteriori error
estimates. Part 1: The recovery techniques. Int. J. Numer. Methods Engrg. 33:7 (1992) 1331–
1364.
364. O.C. Zienkiewicz and J.Z. Zhu, The superconvergent patch recovery and a posteriori error
estimates. Part 2: Error estimates and adaptivity. Int. J. Numer. Methods Engrg. 33:7 (1992)
1365–1382.
365. M. Zitka, P. Solin, T. Vejchodsky, and F. Avila, Imposing orthogonality to hierarchic higherorder finite elements. Math. Comput. Simul. 76 (2007) 211–217.
Fly UP