Finite Volume Methods

Robert Eymard1 , Thierry Gallouët2 and Raphaèle Herbin3

January 2003. This manuscript is an update of the preprint

n0 97-19 du LATP, UMR 6632, Marseille, September 1997
whichappeared in Handbook of Numerical Analysis,
P.G. Ciarlet, J.L. Lions eds, vol 7, pp 713-1020

1
Ecole Nationale des Ponts et Chaussées, Marne-la-Vallée, et Université de Paris XIII
2
Ecole Normale Supérieure de Lyon
3
Université de Provence, Marseille
Contents

1 Introduction 4
1.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 The finite volume principles for general conservation laws . . . . . . . . . . . . . . . . . . 6
1.2.1 Time discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.2 Space discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Comparison with other discretization techniques . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 General guideline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 A one-dimensional elliptic problem 12

2.1 A finite volume method for the Dirichlet problem . . . . . . . . . . . . . . . . . . . . . . . 12
2.1.1 Formulation of a finite volume scheme . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.1.2 Comparison with a finite difference scheme . . . . . . . . . . . . . . . . . . . . . . 14
2.1.3 Comparison with a mixed finite element method . . . . . . . . . . . . . . . . . . . 15
2.2 Convergence theorems and error estimates for the Dirichlet problem . . . . . . . . . . . . 16
2.2.1 A finite volume error estimate in a simple case . . . . . . . . . . . . . . . . . . . . 16
2.2.2 An error estimate using finite difference techniques . . . . . . . . . . . . . . . . . . 19
2.3 General 1D elliptic equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.1 Formulation of the finite volume scheme . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.2 Error estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.3 The case of a point source term . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4 A semilinear elliptic problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.4.1 Problem and Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.4.2 Compactness results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4.3 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3 Elliptic problems in two or three dimensions 32

3.1 Dirichlet boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1.1 Structured meshes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.1.2 General meshes and schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.1.3 Existence and estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.1.4 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.1.5 C 2 error estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.1.6 H 2 error estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.2 Neumann boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.2.1 Meshes and schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.2.2 Discrete Poincaré inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.2.3 Error estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.2.4 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.3 General elliptic operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.3.1 Discontinuous matrix diffusion coefficients . . . . . . . . . . . . . . . . . . . . . . . 78

1
 Version Juillet 2003 2

3.3.2 Other boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

3.4 Dual meshes and unknowns located at vertices . . . . . . . . . . . . . . . . . . . . . . . . 84
3.4.1 The piecewise linear finite element method viewed as a finite volume method . . . 84
3.4.2 Classical finite volumes on a dual mesh . . . . . . . . . . . . . . . . . . . . . . . . 85
3.4.3 “Finite Volume Finite Element” methods . . . . . . . . . . . . . . . . . . . . . . . 88
3.4.4 Generalization to the three dimensional case . . . . . . . . . . . . . . . . . . . . . 89
3.5 Mesh refinement and singularities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.5.1 Singular source terms and finite volumes . . . . . . . . . . . . . . . . . . . . . . . . 90
3.5.2 Mesh refinement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.6 Compactness results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

4 Parabolic equations 95
4.1 Meshes and schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.2 Error estimate for the linear case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.3 Convergence in the nonlinear case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.3.1 Solutions to the continuous problem . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.3.2 Definition of the finite volume approximate solutions . . . . . . . . . . . . . . . . . 103
4.3.3 Estimates on the approximate solution . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.3.4 Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.3.5 Weak convergence and nonlinearities . . . . . . . . . . . . . . . . . . . . . . . . . . 114
4.3.6 A uniqueness result for nonlinear diffusion equations . . . . . . . . . . . . . . . . . 115

5 Hyperbolic equations in the one dimensional case 119

5.1 The continuous problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.2 Numerical schemes in the linear case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.2.1 The centered finite difference scheme . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.2.2 The upstream finite difference scheme . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.2.3 The upwind finite volume scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.3 The nonlinear case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.3.1 Meshes and schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.3.2 L∞ -stability for monotone flux schemes . . . . . . . . . . . . . . . . . . . . . . . . 132
5.3.3 Discrete entropy inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
5.3.4 Convergence of the upstream scheme in the general case . . . . . . . . . . . . . . . 134
5.3.5 Convergence proof using BV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
5.4 Higher order schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

6 Multidimensional nonlinear hyperbolic equations 145

6.1 The continuous problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
6.2 Meshes and schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
6.2.1 Explicit schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
6.2.2 Implicit schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.2.3 Passing to the limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.3 Stability results for the explicit scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
6.3.1 L∞ stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
6.3.2 A “weak BV ” estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
6.4 Existence of the solution and stability results for the implicit scheme . . . . . . . . . . . . 156
6.4.1 Existence, uniqueness and L∞ stability . . . . . . . . . . . . . . . . . . . . . . . . 156
6.4.2 “Weak space BV ” inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
6.4.3 “Time BV ” estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
6.5 Entropy inequalities for the approximate solution . . . . . . . . . . . . . . . . . . . . . . . 164
6.5.1 Discrete entropy inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
6.5.2 Continuous entropy estimates for the approximate solution . . . . . . . . . . . . . 166
 Version Juillet 2003 3

6.6 Convergence of the scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

6.6.1 Convergence towards an entropy process solution . . . . . . . . . . . . . . . . . . 174
6.6.2 Uniqueness of the entropy process solution . . . . . . . . . . . . . . . . . . . . . . 175
6.6.3 Convergence towards the entropy weak solution . . . . . . . . . . . . . . . . . . . . 179
6.7 Error estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
6.7.1 Statement of the results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
6.7.2 Preliminary lemmata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
6.7.3 Proof of the error estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
6.7.4 Remarks and open problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
6.8 Nonlinear weak-? convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
6.9 A stabilized finite element method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
6.10 Moving meshes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

7 Systems 197
7.1 Hyperbolic systems of equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
7.1.1 Classical schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
7.1.2 Rough schemes for complex hyperbolic systems . . . . . . . . . . . . . . . . . . . . 200
7.1.3 Partial implicitation of explicit scheme . . . . . . . . . . . . . . . . . . . . . . . . . 203
7.1.4 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
7.1.5 Staggered grids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
7.2 Incompressible Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
7.2.1 The continuous equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
7.2.2 Structured staggered grids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
7.2.3 A finite volume scheme on unstructured staggered grids . . . . . . . . . . . . . . . 209
7.3 Flows in porous media . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
7.3.1 Two phase flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
7.3.2 Compositional multiphase flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
7.3.3 A simplified case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
7.3.4 The scheme for the simplified case . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
7.3.5 Estimates on the approximate solution . . . . . . . . . . . . . . . . . . . . . . . . . 219
7.3.6 Theorem of convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
Bibliography
 Chapter 1

Introduction

The finite volume method is a discretization method which is well suited for the numerical simulation of
various types (elliptic, parabolic or hyperbolic, for instance) of conservation laws; it has been extensively
used in several engineering fields, such as fluid mechanics, heat and mass transfer or petroleum engineer-
ing. Some of oden
the important features of the finite volume method are similar to those of the finite element
method, see Oden [1991]: it may be used on arbitrary geometries, using structured or unstructured
meshes, and it leads to robust schemes. An additional feature is the local conservativity of the numerical
fluxes, that is the numerical flux is conserved from one discretization cell to its neighbour. This last
feature makes the finite volume method quite attractive when modelling problems for which the flux is of
importance, such as in fluid mechanics, semi-conductor device simulation, heat and mass transfer. . . The
finite volume method is locally conservative because it is based on a “ balance” approach: a local balance
is written on each discretization cell which is often called “control volume”; by the divergence formula,
an integral formulation of the fluxes over the boundary of the control volume is then obtained. The fluxes
on the boundary are discretized with respect to the discrete unknowns.
Let us introduce the method more precisely on simple examples, and then give a description of the
discretization of general conservation laws.

1.1 Examples
Two basic examples can be used to introduce the finite volume method. They will be developed in details
in the following chapters.

introte Example 1.1 (Transport equation) Consider first the linear transport equation

ut (x, t) + div(vu)(x, t) = 0, x ∈ IR 2 , t ∈ IR + ,
(1.1) introtransp
u(x, 0) = u0 (x), x ∈ IR 2

where ut denotes the time derivative of u, v ∈ C 1 (IR 2 , IR 2 ), and u0 ∈ L∞ (IR 2 ). Let T be a mesh of
IR 2 consisting of polygonal bounded convex subsets of IR 2 andintrotransp
let K ∈ T be a “control volume”, that
is an element of the mesh T . Integrating the first equation of (1.1) over K yields the following “balance
equation” over K:
Z Z
ut (x, t)dx + v(x, t) · nK (x)u(x, t)dγ(x) = 0, ∀t ∈ IR + , (1.2) introbaltr
K ∂K

K. Let k ∈ IR ∗+ be a constant time discretization

where nK denotes the normal vector to ∂K, outward tointrobaltr
step and let tn = nk, for n ∈ IN. Writing equation (1.2) at time tn , n ∈ IN and discretizing the time

4
 5

partial derivative by the Euler explicit scheme suggests to find an approximation u (n) (x) of the solution
introtransp
of (1.1) at time tn which satisfies the following semi-discretized equation:
Z Z
1
(u(n+1) (x) − u(n) (x))dx + v(x, tn ) · nK (x)u(n) (x)dγ(x) = 0, ∀n ∈ IN, ∀K ∈ T , (1.3) introsemidis
k K ∂K

where dγ denotes the one-dimensional Lebesgue measure on ∂K and u(0) (x) = u(x, 0) = u0 (x). We need
to define the discrete unknowns for the (finite volume) space discretization. We shall be concerned here
principally with the so-called “cell-centered” finite volume method in which each discrete unkwown is
(n)
associated with a control volume. Let (uK )K∈T ,n∈IN denote the discrete unknowns. For K ∈ T , let EK
in ∂K, and for σ ⊂ ∂K, let nK,σ denote the unit normal to σ
be the set of edges which are included introsemidisc
outward to K. The second integral in (1.3) may then be split as:
Z X Z
v(x, tn ) · nK (x)u(n) (x)dγ(x) = v(x, tn ) · nK,σ u(n) (x)dγ(x); (1.4) introflux1
∂K σ∈EK σ

for σ ⊂ ∂K, let Z

(n)
vK,σ = v(x, tn )nK,σ (x)dγ(x).
σ
introflux1
Each term of the sum in the right-hand-side of (1.4) is then discretized as
(
(n) (n) (n)
(n) vK,σ uK if vK,σ ≥ 0,
FK,σ = (n) (n) (n) (1.5) introflux2
vK,σ uL if vK,σ < 0,

where L denotes the neighbouring control volume to K with common edge σ. This “upstream” or
“upwind” choice is classical for transport equations; it may be seen, from the mechanical point of view,
as the choice of the “upstream information” with respect to the location of σ. This choice is crucial inhyper1d
the mathematical
hypmd
analysis; it ensures the stability properties of the finite volume scheme (see chapters 5
introtransp
and 6). We have therefore derived the following finite volume scheme for the discretization of ( 1.1):
 X (n)
 m(K) (n+1) (n)

 (uK − uK ) + FK,σ = 0, ∀K ∈ T , ∀n ∈ IN,
k
Z σ∈E K (1.6)


 u(0)
K = u 0 (x)dx,
K

(n) introflux2
where m(K) denotes the measure of the control volume K and FK,σ is defined in (1.5). This scheme
is locally conservative in the sense that if σ is a common edge to the control volumes K and L, then
FK,σ = −FL,σ . This property is important in several application fields; it will later be shown to be a key
ingredient in the mathematical proof of convergence. Similar schemes for the discretization of linear or
hyper1dhypmd
nonlinear hyperbolic equations will be studied in chapters 5 and 6.

introde Example 1.2 (Stationary diffusion equation) Consider the basic diffusion equation

−∆u = f on Ω =]0, 1[×]0, 1[,
(1.7) introdiffeq
u = 0 on ∂Ω.
introdiffeq
Let T be a rectangular mesh. Let us integrate the first equation of (1.7) over a control volume K of the
mesh; with the same notations as in the previous example, this yields:
X Z Z
−∇u(x) · nK,σ dγ(x) = f (x)dx. (1.8)
σ∈EK σ K
 6

For each control volume K ∈ T , let x

R K be the center of K. Let σ be the common edge between the
control volumes K and L. The flux − σ ∇u(x) · nK,σ dγ(x), may be approximated by the following finite
difference approximation:
m(σ)
FK,σ = − (uL − uK ), (1.9) introfluxint
dσ
where (uK )K∈T are the discrete unknowns and dσ is the distance between xK and xL . This finite
difference approximation of the first order derivative ∇u · n on the edges of the mesh (where n denotes
the unit normal vector) is consistent: the truncation error on the flux is of order h, where h is the
maximum length of the edges of the mesh. It is necessary for this to be true that the points x K be the
intersections of the orthogonal bisectors of the edges of K. Indeed, this is the case here since the control
volumes are rectangular. This property is satisfied by other meshes which will be studied hereafter. It is
crucial for the discretization of diffusion operators.
In the case where the edge σ is part of the boundary, then dσ denotes the distance R between the center
xK of the control volume K to which σ belongs and the boundary. The flux − σ ∇u(x) · nK,σ dγ(x), is
then approximated by
m(σ)
FK,σ = uK , (1.10) introfluxext
dσ
introdiffeq
Hence the finite volume scheme for the discretization of (1.7) is:
X
FK,σ = m(K)fK , ∀K ∈ T , (1.11)
σ∈EK
introfluxint
introfluxext
where FK,σ is defined by (1.9) and (1.10), and f parabolic
ellmd K
ellund
denotes (an approximation of) the mean value of f
on K. We shall see later (see chapters 2, 3 and 4) that the finite volume scheme is easy to generalize
to a triangular mesh, whereas the finite difference method is not. As in the previous example, the finite
volume scheme is locally conservative, since for any edge σ separating K from L, one has F K,σ = −FL,σ .

1.2 The finite volume principles for general conservation laws

The finite volume method is used for the discretization of conservation laws. We gave in the above section
two examples of such conservation laws. Let us now present the discretization of general conservation
laws by finite volume schemes. As suggested by its name, a conservation law expresses the conservation
of a quantity q(x, t). For instance, the conserved quantities may be the energy, the mass, or the number
of moles of some chemical species. Let us first assume that the local form of the conservation equation
may be written as

qt (x, t) + divF(x, t) = f (x, t), (1.12) conserv.loc

conserv.loc
at each point x and each time t where the conservation of q is to be written. In equation ( 1.12), (·)t
denotes the time partial derivative of the entity within the parentheses, div represents the space divergence
operator: divF = ∂F1 /∂x1 + . . . + ∂Fd /∂xd , where F = (F1 , . . . , Fd )t denotes a vector function depending
on the space variable x and on the time t, xi is the i-th space coordinate, for i = 1, . . . , d, and d is the
space dimension, i.e. d = 1, 2 or 3; the quantity F is a flux which expresses a transport mechanism of
q; the “source term” f expresses a possible volumetric exchange, due for instance to chemical reactions
between the conserved quantities.

Thanks to the physicist’s work, the problem can be closed by introducing constitutive laws which relate
q, F, f with some scalar or vector unknown u(x, t), function of the space variable x and of the time t. For
example, the components of u can be pressures, concentrations, molar fractions of the various chemical
species by unit volume. . . The quantity q is often given by means of a known function q̄ of u(x, t), of the
space variable x and of the time t, that is q(x, t) = q̄(x, t, u(x, t)). The quantity F may also be given
by means of a function of the space variable x, the time variable t and of the unknown u(x, t) and (or)
 7

introte
by meansconserv.loc
of the gradient of u at point (x, t). . . . The transport equation of Example 1.1 is a particular
case of (1.12) with q(x, t) = u(x, t), F(x, t) = vu(x, t) and f (x, t) = f (x); so is the stationary diffusion
introde
equation of Example 1.2 with q(x, t) = u(x), F(x, t) = −∇u(x), and f (x, t) = f (x). The source term f
may also be given by means of a function of x, t and u(x, t).

Example 1.3 (The one-dimensional Euler equations) Let us consider as an example of a system
of conservation laws the 1D Euler equations for equilibrium real gases; these equations may be written
conserv.loc
under the form (1.12), with
 ρ   ρu 
q = ρu and F = ρu2 + p ,
E u(E + p)
where ρ, u, E and p are functions of the space variable x and the time t, and refer respectively to the
density, the velocity, the total energy and the pressure of the particular gas under consideration. The
system of equations is closed by introducing the constitutive laws which relate p and E to the specific
volume τ , with τ = ρ1 and the entropy s, through the constitutive laws:

∂ε u2
p= (τ, s) and E = ρ(ε(τ, s) + ),
∂τ 2
where ε is the internal energy per unit mass, which is a given function of τ and s.
conserv.loc
Equation (1.12) may be seen as the expression of the conservation of q in an infinitesimal domain; it is
formally equivalent to the equation
Z Z Z t2 Z
q(x, t2 )dx − q(x, t1 )dx + F(x, t) · nK (x)dγ(x)dt
K K t1 ∂K Z t2 Z (1.13) conserv.vol
= f (x, t)dxdt,
t1 K

for any subdomain K and for all times t1 and t2 , where nK (x) is the unit normal vector to the boundary
conserv.vol
∂K, at point x, outward to K. Equation (1.13) expresses the conservation law in subdomain K between
times t1 and t2 . Here and in the sequel, unless otherwise mentionned, dx is the integration symbol for
the d-dimensional Lebesgue measure in IR d and dγ is the integration symbol for the (d − 1)-dimensional
Hausdorff measure on the considered boundary.

1.2.1 Time discretization

conserv.loc
The time discretization of Equation (1.12) is performed by introducing an increasing sequence (tn )n∈IN
with t0 = 0. For the sake of simplicity, only constant time steps will be considered here, keeping in mind
that the generalization to variable time steps is straightforward. Let k ∈ IR ?+ denote the time step, and
conserv.loc
let tn = nk, for n ∈ IN. It can be noted that Equation (1.12) could be written with the use of a space-
conserv.loc
time divergence. Hence, Equation (1.12) could be either discretized using a space-time finite volume
discretization or a space finite volume discretization with a time finite difference scheme (the explicit
Euler scheme, for instance). In the first case, theconserv.loc
conservation law is integrated over a time interval and
a space “control volume” as in the formulation (1.12). In the latter case, it is only integrated space wise,
and the time derivative is approximated by a finite difference scheme; with the explicit Euler scheme, the
term (q)t is therefore approximated by the differential quotient (q (n+1) − q (n) )/k, and q (n) is computed
with an approximate value of u at time tn , denoted by u(n) . Implicit and higher order schemes may also
be used.

1.2.2 Space discretization

conserv.loc
In order to perform a space finite volume discretization of equation (1.12), a mesh T of the domain Ω of
IR d , over which the conservation law is to be studied, is introduced. The mesh is such that Ω = ∪ K∈T K,
 8

where an element of T , denoted by K, is an open subset of Ω and is called a control volume. Assumptions
on the meshes will be needed for the definition of the schemes; they also depend on the type of equation
to be discretized.
(n)
For the finite volume schemes considered here, the discrete unknowns at time tn are denoted by uK ,
(n)
K ∈ T . The value uK is expected to be some approximation of u on conserv.loc
the cell K at time tn . The basic
principle of the classical finite volume method is to integrate equation (1.12) over each cell K of the mesh
conserv.vol
T . One obtains a conservation law under a nonlocal form (related to equation ( 1.13)) written for the
volume K. Using the Euler time discretization, this yields
Z Z Z
q (n+1) (x) − q (n) (x)
dx + F(x, tn ) · nK (x)dγ(x) = f (x, tn )dx, (1.14) eVF3
K k ∂K K

where nK (x) is the unit normal vector to ∂K at point x, outward to K.

The remaining step in order to define the finite volume scheme is therefore the approximation of the “flux”,
(n)
F(x, tn ) · nK (x), across the boundary ∂K of each control volume, in terms of {uL , L ∈ T } (this flux
n+1
approximation has to be done in terms of {uL , L ∈ T } if one chooses the implicit Euler scheme instead of
the explicit Euler scheme for the time discretization). More precisely, omittingRthe terms on the boundary
of Ω, let K|L = K ∩ L, with K, L ∈ T , the exchange term (from K to L), K|L F(x, tn ) · nK (x)dγ(x),
between the control volumes K and L during the time interval [tn , tn+1 ) is approximated by some quantity,
(n) (n)
FK,L , which is a function of {uM , M ∈ T } (or a function of {un+1
M , M ∈ T } for the implicit Euler scheme,
(n)
or more generally a function of {uM , M ∈ T } and {un+1
M , M ∈ T } if the time discretization is a one-step
(n)
method). Note that FK,L = 0 if the Hausdorff dimension of K ∩ L is less than d − 1 (e.g. K ∩ L is a
point in the case d = 2 or a line segment in the case d = 3).
Let us point out that two important features of the classical finite volume method are
(n) (n)
1. the conservativity, that is FK,L = −FL,K , for all K and L ∈ T and for all n ∈ IN.
2. the “consistency” of the approximation of F(x, tn ) · nK (x), which has to be defined for each relation
type between F and the unknowns.
These properties, together with adequate stability properties which are obtained by estimates on the
approximate solution, will give some convergence properties of the finite volume scheme.

1.3 Comparison with other discretization techniques

The finite volume method is quite different from (but sometimes related DB
to) the finite difference method
or the finite element
finitedifference ciarlet
method. On these classical methods
finitelement
see e.g. Dahlquist
mixedfinitelement
and Björck [1974],
Thomée [1991], Ciarlet, P.G. [1978], Ciarlet [1991], Roberts and Thomas [1991].
Roughly speaking, the principle of the finite difference method is, given a number of discretization points
which may be defined by a mesh, to assign one discrete unknown per discretization point, and to write
one equation per discretization point. At each discretization point, the derivatives of the unknown are
replaced by finite differences through the use of Taylor expansions. The finite difference method becomes
difficult to use when the coefficients involved in the equation are discontinuous (e.g. in the case of
heterogeneous media). With the finite volume method, discontinuities of the coefficients will not be any
problem if the mesh is chosen suchell1Dthat the discontinuities of the coefficients occur on the boundaries of
ell2Dgen
the control volumes (see sections 2.3 and 3.3, for elliptic problems). Note that the finite volume scheme
is often called “finite difference scheme” or “cell centered difference scheme”. Indeed, in the finite volume
method, the finite difference approach can be used for the approximation of the fluxes on the boundary
of the control volumes. Thus, the finite volume scheme differs from the finite difference scheme in that
the finite difference approximation is used for the flux rather than for the operator itself.
 9

ciarlet
The finite element method (see e.g. Ciarlet, P.G. [1978]) is based on a variational formulation, which is
written for both the continuous and the discrete problems, at least in the case of conformal finite element
methods which are considered here. The variational formulation is obtained by multiplying the original
equation by a “test function”. The continuous unknown is then approximated by a linear combination
of “shape” functions; these shape functions are the test functions for the discrete variational formulation
(this is the so called “Galerkin expansion”); the resulting equation is integrated over the domain. The
finite volume method is sometimes called a “discontinuous finite element method” since the original
equation is multiplied by the characteristic function of each grid cell which is defined by 1K (x) = 1, if
x ∈ K, 1K (x) = 0, if x ∈ / K, and the discrete unknown may be considered as a linear combination of
shape functions. However, the techniques used to prove the convergence of finite element methods do not
generally apply for this choice of test functions. In the following chapters, the finite volume method will
be compared in more detail with the classical and the mixed finite element methods.

From the industrial point of view, the finite volume method is known as a robust and cheap method
for the discretization of conservation laws (by robust, we mean a scheme which behaves well even for
particularly difficult equations, such as nonlinear systems of hyperbolic equations and which can easily be
extended to more realistic and physical contexts than the classical academic problems). The finite volume
method is cheap thanks to short and reliable computational coding for complex problems. It may be more
adequate than the finite difference method (which in particular requires a simple geometry). However,
in some cases, it is difficult to design schemes which give enough precision. Indeed, the finite element
method can be much more precise than the finite volume method when using higher order polynomials,
but it requires an adequate functional framework which is not always available in industrial problems.
Other more precise methods are, for instance, particle methods or spectral methods but these methods
can be more expensive and less robust than the finite volume method.

1.4 General guideline

The mathematical theory of finite volume schemes has recently been undertaken. Even though we choose
here to refer to the class of scheme which is the object of our study as the ”finite volume” method, we
must point out that there are several methods with different names (box method, control volume finite
element methods, balance method to cite only a few) which may be viewed as finite volume methods.
The name ”finite difference” has also often been used referring to the finite volume method. We shall
mainly quote here the works regarding the mathematical analysis of the finite volume method, keeping
in mind that there exist numerous works on applications of the finite volume methods in the applied
sciences, some references to which may be found in the books which are cited below.
Finite volume methods for convection-diffusionsamarski1
samarskitich
equations seem to havesamarski2
been first introduced in the early
sixties by Tichonov and Samarskii [1962], Samarskii [1965] and Samarskii [1971].
The convergence theory of such schemes in several space dimensions has only recently been undertaken. In
samarskilaz
the case of vertex-centered finite volume schemes, Heinrich
studies were carried BR
out by Samarskii, Lazarov
Cai
and
Makarov
CMMC
[1987] in the case of Cartesian meshes, Heinrich
Vanselow
[1986], Bank and Rose [1986], Cai [1991],
Cai, Mandel and Mc Cormick suli
mortonsuli
[1991] and Vanselow
mackenzie
[1996] in the case of unstructured
mortonss
meshes;
see also Morton andshashkov
Süli [1991], Süli [1989], Mackenzie, and Morton [1992], Morton, Stynes
and Süli [1997] and Shashkov
WM
[1987] in the case of quadrilateral
FS
meshes. Cell-centered finite
weiser
volume
schemes are addressed in Manteuffel
lazarovsiam
and White [1986], Forsyth and Sammon [1988], Weiser and
Wheeler
vass
[1988] and Lazarov, Mishev and Vassilevski
VF4
[1996] in
rouenrh
the case of Cartesian
rouenlazarov
meshes and
in Vassileski,
mishev
Petrova and Lazarov [1992], Herbin [1995], Herbin [1996], Lazarov and Mishev
rouencoudiere
[1996], Mishev [1998] in the coudiere
case of triangular or Voronoı̈ meshes; let us also mention Coudière, Vila
and Villedieu [1996] and Coudière, Vila and Villedieu [1999] where more general meshes are
treated, with, however, a somewhat technical geometrical condition. In the pureABMO diffusion case, the cell
centered finite volume method has also
angermann
been analyzed
BMO
with finite element tools: Agouzal, Baranger,
arbogast
Maitre and Oudin [1995], Angermann [1996], Baranger, Maitre and Oudin [1996], Arbogast,
 10

angermann
Wheeler
FFL2
and Yotov [1997], Angermann [1996]. Semilinear convection-diffusion are studied in
Feistauer,cvnl
Felcman and Lukacova-Medvidova [1997] with a combined finite element-finite volume
method, Eymard, Gallouët and Herbin [1999] with a pure finite volume scheme.
Concerning nonlinear hyperbolic conservation laws, the one-dimensional case is now classical; let us
godlewski-ellipses
mentionleveque
the following books on numerical methods
godlewski-springer
for hyperbolic problems:
kronerbook
Godlewski and Raviart
[1991], LeVeque [1990], Godlewski and Raviart [1996], Kröner [1997], and referencescgh therein. In
the multidimensional case, let us mention
kroner1
the convergence results which where
Coquelconv
obtained in Champier,
Gallouët and Herbin [1993], Kr öner and Rokyta [1994], Cockburn, Coquel
Coquel Vi
and LeFloch
[1995] and the error estimateseggh
of Cockburn, Coquel and LeFloch [1994] and Vila [1994] in the
case of an explicit scheme and Eymard, Gallouët, Ghilani and Herbin [1998] in the case of explicit
and implicit schemes.

The purpose of the following chapters is to lay out a mathematical framework for the convergence and
error analysis of the finite volume method for the discretization of elliptic, parabolic or hyperbolic
cgh
partial
differential equations under conservative
VF4
form, following
eggh
the philosophy of the works of Champier,
Gallou
cvnl
ët and Herbin [1993], Herbin [1995], Eymard, Gallouët, Ghilani and Herbin [1998]
and Eymard, Gallouët and Herbin [1999]. In order to do so, we shall describe the implementation of
the finite volume method on some simple (linear or non-linear) academic problems, and develop the tools
which are needed for the mathematical analysis. This approach will help to determine the properties of
finite volume schemes which lead to “good” schemes for complex applications.
ellund
Chapter 2 introduces the finite volume discretization of an elliptic operator in one space dimension.
The resulting numerical scheme is compared to finite difference, finite element and mixed finite element
methods in this particular case. An error estimate is given; this estimate is in fact contained in results
shown later in the multidimensional case; however, with the one-dimensional case, one can already un-
derstand
WM
the basic principles of the convergence
FS
proof, and understand the difference with the proof of
Manteuffel and White [1986] or Forsyth and Sammon [1988], which does not seem to generalize
to the unstructured meshes. In particular, it is made clear that, although the finite volume scheme is not
consistent in the finite difference sense since the truncation error does not tend to 0, the conservativity of
the scheme, together with a consistent approximation of the fluxes and some “stability” allow the proof of
convergence. The scheme and the error estimate are then generalized to the case of a more general elliptic
operator allowing discontinuities in the diffusion coefficients. Finally, a semilinear problem is studied, for
which a convergence result is proved. The principle of the proof ofellmd
this result may be used for nonlinear
problems in several space dimensions. It will be used in Chapter 3 in order to prove convergence results
for linear problems when no regularity on the exact solution is known.
ellmd
In Chapter 3, the discretization of elliptic problems in several space dimensions by the finite volume
method is presented. Structured meshes are shown to be an easy generalization of the one-dimensional
case; unstructured meshes are then considered, for Dirichlet and Neumann conditionscvnl
on the boundary
of the domain. In both cases, admissible meshes are defined, and, following Eymard, Gallouët and
Herbin [1999], convergence results (with no regularity on the data) and error estimates assuming a
C 2 or H 2 regular solution to the continuous problems are proved. As in the one-dimensional case, the
conservativity of the scheme, together with a consistent approximation of the fluxes and some “stability”
are used for the proof of convergence. In addition to the properties already used in the one-dimensional
case, the multidimensional estimates require the use of a “discrete Poincaré” inequality which is proved
in both Dirichlet and Neumann cases, along with some compactness properties which are also used and
are given in the last section. It is then shown how to deal with matrix diffusion coefficients and more
general boundary conditions. Singular sources and mesh refinement are also studied.
parabolic ellmd
Chapter 4 deals with the discretization of parabolic problems. Using the same concepts as in Chapter 3,
an error estimate is given in the linear case. A nonlinear degenerate parabolic problem is then studied,
for which a convergence result is proved, thanks to a uniqueness result which is proved at the end of the
chapter.
hyper1d
Chapter 5 introduces the finite volume discretization of a hyperbolic operator in one space dimension.
 11

Some basics on entropy weak solutions to nonlinear hyperbolic equations are recalled. Then the concept
of stability of a scheme is explained on a simple linear advection problem, for which both finite difference
and finite volume schemes are considered. Some well known schemes are presented with a finite volume
formulation in the nonlinear case. A proof of convergencehypmd
using a “weak BV inequality” which was found
to be crucial in the multidimensional case (Chapter 6) is given in the one-dimensional case for the sake
of clarity. For the sake of completeness, the proof of convergence based on “strong BV estimates” and
the Lax-Wendroff theorem is also recalled, although it does not seem to extend to the multidimensional
case with general meshes.
hypmd
In Chapter 6, finite volume schemes for the discretization of multidimensional nonlinear hyperbolic con-
servation equations are studied. Under suitable assumptions, which are satisfied by several well known
schemes, it is shown that the considered schemes are L∞ stable (this is classical) but also satisfy some
“weak BV inequality”. This eggh
“weak BV ” inequality is the key estimate to the proof of convergence of
the schemes. Following Eymard, Gallouët, Ghilani and Herbin [1998], both time implicit and
explicit discretizations are considered. The existence of the solution to the implicit scheme is proved.
The approximate solutions are shown to satisfy some discrete entropy inequalities. Using the weak BV
estimate, the approximate solution is also shown to satisfy some continuous entropy inequalities. Intro-
ducing the concept of “entropy process solution”DP
to the nonlinear hyperbolic equations (which is similar
to the notion of measure valued solutions of DiPerna [1985]), the approximate solutions are proved to
converge towards an entropy process solution as the mesh size tends to 0. The entropy process solution is
shown to be unique, and is therefore equal to the entropy weak solution, which concludes the convergence
of the approximate solution towards the entropy weak solution. Finally error estimates are proved for
both the explicit and implicit schemes.
The last chapter is concerned with systems of equations. In the case of hyperbolic systems which are
considered in the first part, little is known concerning the continuous problem, so that the schemes which
are introduced are only shown to be efficient by numerical experimentation. These “rough” schemes
seem to be efficient for complex cases such as the Euler equations for real gases. The incompressible
Navier-Stokes equations Pat
are then considered; after recalling the classical staggered grid finite volume
formulation (see e.g. Patankar [1980]), a finite volume scheme defined on a triangular ellmd
mesh for the
Stokes equation is studied. In the case of equilateral triangles, the tools of Chapter 3 allow to show that
the approximate velocities converge to the exact velocities. Systems arising from modelling multiphase
flow in porous media are then considered. The convergence of the approximate
hypmd
finite volume solution for
a simplified case is then proved with the tools introduced in Chapter 6.

More precise references to recent works on the convergence of finite volume methods will be made in the
following chapters. However, we shall not quote here the numerous works on applications of the finite
volume methods in the applied sciences.
 Chapter 2

A one-dimensional elliptic problem

ellund introde
The purpose of this chapter is to give some developments of the example 1.2 of the introduction in the
one-dimensional case. The formalism needed to define admissible finite volume meshes is first given
and applied to the Dirichlet problem. After some comparisons with other relevant schemes, convergence
theorems and error estimates are provided. Then, the case of general linear elliptic equations is handled
and finally, a first approach of a nonlinear problem is studied and introduces some compactness theorems
in a quite simple framework; these compactenss theorems will be useful in further chapters.

2.1 A finite volume method for the Dirichlet problem

Diric1D
2.1.1 Formulation of a finite volume scheme
fvse1d
The principle of the finite volume method will be shown here on the academic Dirichlet problem, namely a
second order differential operator without time dependent terms and with homogeneous Dirichlet bound-
ary conditions. Let f be a given function from (0, 1) to IR, consider the following differential equation:

−uxx(x) = f (x), x ∈ (0, 1),

u(0) = 0, (2.1) elliptic1D
u(1) = 0.
elliptic1D
If f ∈ C([0, 1], IR), there exists a unique solution u ∈ C 2 ([0, 1], IR) to Problem (2.1). In the sequel, this
exact solution will be denoted by u. Note that the equation −uxx = f can be written in the conservative
form div(F) = f with F = −ux .
In order to compute a numerical approximation to the solution of this equation, let us define a mesh,
denoted by T , of the interval (0, 1) consisting of N cells (or control volumes), denoted by K i , i = 1, . . . , N ,
and N points of (0, 1), denoted by xi , i = 1, . . . , N , satisfying the following assumptions:
meshund Definition 2.1 (Admissible one-dimensional mesh) An admissible mesh of (0, 1), denoted by T , is
given by a family (Ki )i=1,···,N , N ∈ IN ? , such that Ki = (xi− 12 , xi+ 12 ), and a family (xi )i=0,···,N +1 such
that

x0 = x 12 = 0 < x1 < x 23 < · · · < xi− 21 < xi < xi+ 21 < · · · < xN < xN + 21 = xN +1 = 1.
One sets
N
X
hi = m(Ki ) = xi+ 12 − xi− 12 , i = 1, . . . , N, and therefore hi = 1,
i=1
h− +
i = xi − xi− 12 , hi = xi+ 21 − xi , i = 1, . . . , N,
hi+ 21 = xi+1 − xi , i = 0, . . . , N,
size(T ) = h = max{hi , i = 1, . . . , N }.

12
 13

The discrete unknowns are denoted by ui , i = 1, . . . , N , and are expected to be some approximation of
u in the cell Ki (the discrete unknown ui can be viewed as an approximation of the mean value of u
overelliptic1D
Ki , or of the value of u(xi ), or of other values
eVF3
of u in the control volume Ki . . . ). The first equation
of (2.1) is integrated over each cell Ki , as in (1.14) and yields
Z
−ux (xi+ 21 ) + ux (xi− 21 ) = f (x)dx, i = 1, . . . , N.
Ki

A reasonable choice for the approximation of −ux (xi+ 21 ) (at least, for i = 1, . . . , N − 1) seems to be the
differential quotient
ui+1 − ui
Fi+ 21 = − .
hi+ 12
This approximation is consistent in the sense that, if u ∈ C 2 ([0, 1], IR), then

? u(xi+1 ) − u(xi )
Fi+ 1 = − = −ux (xi+ 21 ) + 0(h), (2.2) consistflux
2 hi+ 12

where |0(h)| ≤ Ch, C ∈ IR + only depending on u.

Remark 2.1 Assume that xi is the center of Ki . Let ũi denote the mean value over Ki of the exact
elliptic1D
solution u to Problem (2.1). One may then remark that |ũi − u(xi )| ≤ Ch2i , with some C only depending
on u; it follows easily that (ũi+1 − ũi )/hi+ 21 = ux (xi+ 12 ) + 0(h) also holds, for i = 1, . . . , N − 1 (recall
that h = max{hi , i = 1, . . . , N }). Hence the approximation of the flux is also consistent if the discrete
unknowns ui , i = 1, · · · , N , are viewed as approximations of the mean value of u in the control volumes.

The Dirichlet boundary conditions are taken into account by using the values imposed at the boundaries to
computeRthe fluxes on these boundaries. Taking these boundary conditions into consideration and setting
fi = h1i Ki f (x)dx for i = 1, . . . , N (in an actual computation, an approximation of fi by numerical
elliptic1D
integration can be used), the finite volume scheme for problem (2.1) writes

Fi+ 12 − Fi− 12 = hi fi , i = 1, . . . , N (2.3) VF1

ui+1 − ui
Fi+ 21 = − , i = 1, . . . , N − 1, (2.4) VF2
hi+ 21
u1
F 12 = − , (2.5) VF3
h 21
uN
FN + 21 = . (2.6) VF4
hN + 21
VF2 VF3 VF4
Note that (2.4), (2.5), (2.6) may also be written
ui+1 − ui
Fi+ 21 = − , i = 0, . . . , N, (2.7) VF2m
hi+ 12
setting

u0 = uN +1 = 0. (2.8) VF3m
VF1 VF4
The numerical scheme (2.3)-(2.6) may be written under the following matrix form:

AU = b, (2.9) AUb
VF3m
where U = (u1 , . . . , uN )t , b = (b1 , . . . , bN )t , with (2.8) and with A and b defined by
 14

1  ui+1 − ui ui − ui−1 
(AU )i = − + , i = 1, . . . , N, (2.10) matrix
hi hi+ 12 hi− 21
Z
1
bi = f (x)dx, i = 1, . . . , N, (2.11) rhsdf
h i Ki

elliptic1D
angot1 Remark 2.2 There are other finite volume schemes for problem (2.1).
meshund
1. For instance, it isVF1possible,
VF4
in Definition
VF1
2.1, to take x1 ≥ 0, xN ≤ 1 and, for the VF3 definitionVF4of the
scheme (that is (2.3)-(2.6)),VF2
to write ( 2.3) only for i = 2, . . . , N − 1 and to replace ( 2.5) and ( 2.6) by
u1 = uN = 0 (note that (2.4) does not change). For this so-calledVF1 “modified
VF4
finite volume” scheme,
h2conv
it is also possible to obtain an error estimate as for the scheme (2.3)-(2.6) (see Remark 2.5). Note
that, with this scheme, the union of all control volumes for which the “conservation law” is written
is slightly different from [0, 1] (namely [x3/2 , xN −1/2 ] 6= [0, 1]) .
2. Another possibility is to take (primary) unknowns associated
keller
to the croisille
boundaries of the control
volumes. We shall not consider this case here (cf. Keller [1971], Courbet and Croisille
[1998]).

2.1.2 Comparison with a finite difference scheme

fdvf fvse1d
With the same notations as in Section 2.1.1, consider that ui is now an approximation of u(xi ). It is
interesting to notice that the expression
1  ui+1 − ui ui − ui−1 
− +
hi hi+ 12 hi− 21
is not a consistent approximation of −uxx(xi ) in the finite difference DB sense, that is the error made by
replacing the derivative by a difference quotient (the truncation error Dahlquist and Björck [1974])

t AUb rhsdf
does not tend to 0 as h tends to 0. Indeed, let U = u(x1 ), . . . , u(xN ) ; with the notations of (2.9)-(2.11),
the truncation error may be defined as

r = AU − b,
with r = (r1 , . . . , rN )t . Note that for f regular enough, which is assumed in the sequel, bi = f (xi ) + 0(h).
An estimate of r is obtained by using Taylor’s expansion:
1 1
u(xi+1 ) = u(xi ) + hi+ 21 ux (xi ) + h2i+ 1 uxx (xi ) + h3i+ 1 uxxx(ξi ),
2 2 6 2

for some ξi ∈ (xi , xi+1 ), which yields

1 hi+ 21 + hi− 12
ri = − uxx (xi ) + uxx (xi ) + 0(h), i = 1, . . . , N,
hi 2
which does not, in general tend to 0 as h tends to 0 (except in particular cases) as may be seen on the
simple following example:
meshund
Example 2.1 Let f ≡ 1 and consider a mesh of (0, 1), in the sense of Definition 2.1, satisfying hi = h
for even i, hi = h/2 for odd i and xi = (xi+1/2 + xi−1/2 )/2, for i = 1, . . . , N . An easy computation shows
that the truncation error r is such that

ri = − 41 , for even i
ri = + 12 , for odd i.
Hence sup{|ri |, i = 1, . . . , N } 6→ 0 as h → 0.
 15

VF1 VF4
Therefore, the scheme obtained from (2.3)-(2.6) is not consistent in the finite difference sense, even
though it is consistent in the finite volume sense, that is, the numerical approximation of the fluxes is
conservative and the truncation error on the fluxes tends to 0 as h tends to 0. elliptic1D
If, for instance, xi is the center of Ki , for i = 1, . . . , N , it is well known that for problem (2.1), the
consistent finite difference scheme would be, omitting boundary conditions,
h u ui − ui−1 i
4 i+1 − ui
− + = f (xi ), i = 2, . . . , N − 1, (2.12) FD
2hi + hi−1 + hi+1 hi+ 21 hi− 12

Remark 2.3 Assume that xi is, for i = 1, . . . , N , the center of Ki and that the discrete unknown ui of
the finite volume scheme is considered as an approximation of the mean value ũi of u over Ki (note that
ũi = u(xi ) + (h2i /24)uxx(xi ) + 0(h3 ), if u ∈ C 3 ([0, 1], IR)) instead of u(xi ), then again, the finite volume
scheme, considered once more as a finite difference scheme, is not consistent in the finite difference sense.
Indeed, let R̃ = AŨ − b, with Ũ = (ũ1 , . . . , ũN )t , and R̃ = (R̃1 , . . . , R̃N )t , then, in general, R̃i does not
go to 0 as h goes to 0. In fact, it will be shown later that the finite volume scheme, when seen as a finite
difference scheme, is consistent in the finite difference sense if ui is considered as an approximation of
u(xi ) − (h2i /8)uxx(xi ).FSThis is the idea upon which the first proof offdconvergence
convergence by Forsyth and Sammon
in 1988 is based, see Forsyth and Sammon [1988] and Section 2.2.2.
elliptic1D
In the case of Problem (2.1),
FD
both the finite volume and finite difference schemes are convergent. The
finite difference scheme (2.12) is convergent since it is stable, in the sense that kXk∞ ≤ CkAXk∞ ,
for all X ∈ IR N , where C is a constant and kXk∞ = sup(|X1 |, . . . , |XN |), X = (X1 , . . . , XN )t , and
consistent in the usual finite difference sense. Since A(U − U ) = R, the stability property implies that
kU − U k∞ ≤ CkRk∞ which goes to 0, as h goes to 0, by definition VF1 VF4
of the consistency in the finite
difference sense. Thefvconvergence1D
convergence of the finite volume scheme (2.3)-(2.6) needs some more work and is
described in Section 2.2.1.

2.1.3 Comparison with a mixed finite element method

The finite volume method has often be thought of as a elliptic1D
kind of mixed finite element method. Nevertheless,
we show here that, on the
elliptic1D
simple Dirichlet problem ( 2.1), the two methods yield two different schemes.
For Problem (2.1), the discrete unknowns of the finite volume method are the values ui , i = 1, . . . , N .
However, the finite volume method also introduces one discrete unknown at each of the control volume
extremities, namely the numerical flux between the corresponding control volumes. Hence, the finite
volume method for elliptic problems may appear closely related to the mixedelliptic1D
finite element method.
Recall that the mixed finite element method consists in introducing in Problem (2.1) the auxiliary variable
q = −ux , which yields the following system:

q + ux = 0,
qx = f ;
assuming f ∈ L2 ((0, 1)), a variational formulation of this system is:

q ∈ H 1 ((0, 1)), u ∈ L2 ((0, 1)), (2.13) efm0

Z 1 Z 1
q(x)p(x)dx = u(x)px (x)dx, ∀ p ∈ H 1 ((0, 1)), (2.14) efm1
0 0
Z 1 Z 1
qx (x)v(x)dx = f (x)v(x)dx, ∀ v ∈ L2 ((0, 1)). (2.15) efm2
0 0
meshund
Considering an admissible mesh of (0, 1) (see Definition 2.1), the usual discretization of this variational
formulation consists in taking the classical piecewise linear finite element functions for the approximation
H of H 1 ((0, 1)) and the piecewise constant finite element for the approximation L of L2 ((0, 1)). Then,
the discrete unknowns are {ui , i = 1, . . . , N } and {qi+1/2 , i = 0, . . . , N } (ui is an approximation of u in
 16

Ki and qi+1/2 is an approximation of −ux (xi+1/2 )). The discrete equations are obtained by performing
a Galerkin expansion of u and q with respect to the natural basis functions ψl , l = 1, . . . , N (spanning
efm1
L), and ϕj+1/2 , j = 0, . efm2. . , N (spanning H) and by taking p = ϕi+1/2 , i = 0, . . . , N in (2.14) and
v = ψk , k = 1, . . . , N in (2.15). Let h0 = hN +1 = 0, u0 = uN +1 = 0 and q−1/2 = qN +3/2 = 0; then the
discrete system obtained by the mixed finite element method has 2N + 1 unknowns. It writes
hi + hi+1 hi hi+1
qi+ 21 ( ) + qi− 12 ( ) + qi+ 32 ( ) = ui − ui+1 , i = 0, . . . , N,
3 6 6
Z
qi+ 21 − qi− 12 = f (x)dx, i = 1, . . . , N.
Ki

Note that the unknowns qi+1/2 cannot be eliminated from the system. The resolution of this system of
equationsVF1 not give the same values {ui , i = 1, . . . , N } than those obtained by using the finite volume
doesVF4
scheme (2.3)-(2.6). In fact it is easily seen that, in this case, the finite volume scheme can be obtained
fromefm1
the mixed finite element scheme by using the following numerical integration for the left handside
of (2.14): Z
g(xi+1 ) + g(xi )
g(x)dx = hi .
Ki 2
This is also true for some two-dimensional elliptic problems and therefore the finiteABMO
volume error estimates
for these problems may be obtained
BMO
via the mixed finite element theory, see Agouzal, Baranger,
Maitre and Oudin [1995], Baranger, Maitre and Oudin [1996].

2.2 Convergence theorems and error estimates for the Dirichlet

problem
2.2.1 A finite volume error estimate in a simple case
convergence1D
We shall now prove the following error estimate, which will be generalized to more general elliptic problems
and in higher space dimensions.

eelapl1D Theorem 2.1 elliptic1D

Let f ∈ C([0, 1], IR) and let u ∈ C 2 ([0, 1], IR) be the (unique)meshundsolution of Problem (2.1). Let T =
(Ki )i=1,...,N be an admissible mesh in the sense of Definition 2.1. Then, there exists a unique vector
VF1 VF4
U = (u1 , . . . , uN )t ∈ IR N solution to (2.3) -(2.6) and there exists C ≥ 0, only depending on u, such that
N
X (ei+1 − ei )2
≤ C 2 h2 , (2.16) eel1
i=0
hi+ 12
and

|ei | ≤ Ch, ∀i ∈ {1, . . . , N }, (2.17) eel2

with e0 = eN +1 = 0 and ei = u(xi ) − ui , for all i ∈ {1, . . . , N }.

ee1dell
This theorem is in fact a consequence of Theorem 2.3, which gives an error estimate for the finite volume
discretization of a more general operator. However, we now give the proof of the error estimate in this
first simple case.
eelapl1D
Proof of Theorem 2.1
VF1 VF4
that there exists a unique vector U = (u1 , . . . , uN )t ∈ IR N solution to (2.3)-(2.6). Indeed,
First remark VF1
multiplying (2.3) by ui and summing for i = 1, . . . , N gives
 17

N
X −1 XN
u21 (ui+1 − ui )2 u2N
+ + = ui hi f i .
h 12 i=1
hi+ 12 hN + 12 i=1
VF1
Therefore, if fi = 0 for any i ∈ {1, . . . , N }, then the unique solution to (2.3) is obtained by taking ui = 0,
VF1
t N
i ∈ {1,
for anyVF2 VF4
. . . , N }. This gives existence and uniqueness of U = (u 1 , . . . , u N ) ∈ IR solution to ( 2.3)
(with (2.4)-(2.6)).
eel1
One now proves (2.16). Let

F i+ 12 = −ux (xi+ 21 ), i = 0, . . . , N,
Integrating the equation −uxx = f over Ki yields

F i+ 12 − F i− 21 = hi fi , i = 1, . . . , N.
VF1
By (2.3), the numerical fluxes Fi+ 12 satisfy

Fi+ 12 − Fi− 12 = hi fi , i = 1, . . . , N.

Therefore, with Gi+ 21 = F i+ 21 − Fi+ 21 ,

Gi+ 12 − Gi− 21 = 0, i = 1, . . . , N.
consistflux
Using the consistency of the fluxes (2.2), there exists C > 0, only depending on u, such that
?
Fi+ 1 = F i+ 1 + Ri+ 1 and |Ri+ 1 |. ≤ Ch, (2.18) errconst
2 2 2 2

Hence with ei = u(xi ) − ui , for i = 1, . . . , N , and e0 = eN +1 = 0, one has

ei+1 − ei
Gi+ 21 = − − Ri+ 21 , i = 0, . . . , N,
hi+ 21
so that (ei )i=0,...,N +1 satisfies
ei+1 − ei ei − ei−1
− − Ri+ 12 + + Ri− 12 = 0, ∀i ∈ {1, . . . , N }. (2.19) eqerror
hi+ 2
1 hi− 21
eqerror
Multiplying (2.19) by ei and summing over i = 1, . . . , N yields
N
X N
X N
X N
X
(ei+1 − ei )ei (ei − ei−1 )ei
− + =− Ri− 12 ei + Ri+ 12 ei .
i=1
hi+ 12 i=1
hi− 12 i=1 i=1
errconst
Noting that e0 = 0, eN +1 = 0 and reordering by parts, this yields (with (2.18))
N
X N
X
(ei+1 − ei )2
≤ Ch |ei+1 − ei |. (2.20) caspie1
i=0
hi+ 12 i=0

The Cauchy-Schwarz inequality applied to the right hand side gives

N
X X
N
(e − ei )2  12 X
N  12
i+1
|ei+1 − ei | ≤ hi+ 12 . (2.21) caspie2
i=0 i=0
hi+ 21 i=0
N
X caspie2 caspie1 eel1
Since hi+ 12 = 1 in (2.21) and from (2.20), one deduces (2.16).
i=0
 18

i
X caspie2 eel1 eel2
Since, for all i ∈ {1, . . . , N }, ei = (ej − ej−1 ), one can deduce, from (2.21) and (2.16) that (2.17)
j=1
holds.

Remark 2.4 The error estimate given in this section does not use the discrete maximum principle (that
is the fact that fi ≥ 0, for all i = 1, . . . , N , implies ui ≥ 0, for all i = 1, . . . , N ), which is used in the
proof of error estimates by the finite difference techniques, but the coerciveness of the elliptic operator,
as in the proof of error estimates by the finite element techniques.

h2conv Remark 2.5

1. The above proof of convergence gives an error estimate of order h. It is sometimes possible to obtain
an error estimate of order h2 . Indeed, this is the case, at least if u ∈ C 4 ([0, 1], IR), if xi is the center
| ≤ Ch2 , for all i ∈ {1, . . . , N }, where C only
. . , N . One obtains, in this case, |eiBMO
of Ki for all i = 1, .FS
depends on u (see Forsyth and Sammon [1988] or Baranger, Maitre and Oudin [1996]).
2. It is also possible to obtain an error
angot1 angot1
estimate for the modified finite volume scheme described in
the first item of Remark 2.2 page 14. It is even possible to obtain an error estimate of order h2 in
the case x1 = 0, xN = 1 and assuming that xi+1/2 = (1/2)(xi + xi+1 ), for all i = 1, . . . , N − 1. In
fact, case, one obtains |Ri+1/2 | ≤ C1 h2 , for all i = 1, . . . , N − 1. Then, the proof of Theorem
in thiseel1
eelapl1D
2.1 gives (2.16) with h4 instead of h2 which yields |ei | ≤ C2 h2 , for all i ∈ {1, . . . , N } (where C1
and C2 are only depending on u). Note that this modified finite volume scheme is also consistent
in the finite difference sense. Then, the finite difference techniques yield also an error estimate on
|ei |, but only of order h.
3. It could be tempting to try and find error estimates with respect to the mean value of the exact
solution on the control volumes rather than with respect to its value at some point of the control
volumes. This is not such a good idea: indeed, if xi is not the center of Ki (this will be the general
case in several space dimensions), then one does not have (in general) |ẽi | ≤ C3 h2 (for some C3
only depending on u) with ẽi = ũi − ui where ũi denotes the mean value of u over Ki .

Remark 2.6
eelapl1D
1. If the assumption f ∈ C([0, 1], IR) is replaced by the assumption f ∈ L2 ((0,eelapl1D
1)) in Theorem 2.1,
then u ∈ H 2 ((0, 1)) instead of C 2 ([0, 1], IR), but the estimates of Theorem 2.1 still hold. Then,
the consistency of the fluxes must be obtained with a Taylor expansion with an integral remainder.
This is feasible for C 2 functions, and since the remainder
testh2
only depends on the H 2 norm, a density
testh2 cvnl
argument allows to conclude; see also Theorem 3.4 page 56 and Eymard, Gallouët and Herbin
[1999].
eelapl1D
2. If the assumption f ∈ C([0, 1], IR) is replaced by the assumption f ∈ L1 ((0, 1)) in Theorem 2.1,
then u ∈ C 2 ([0, 1], IR) no longer holds (neither does u ∈ H 2 ((0, 1))), but the convergence still holds;
indeed there exists C(u, h), only depending on u and h, such that C(u, h) → 0, as h → 0, and
|eerrconst
i | ≤ C(u, h), for all i = 1, . . . , N . The proof is similar to the one above, except that the estimate
(2.18) is replaced by |Ri+1/2 | ≤ C1 (u, h), for all i = 0, . . . , N , with some C1 (u, h), only depending
on u and h, such that C(u, h) → 0, as h → 0.
eel1
Remark 2.7 Estimate (2.16) can be interpreted
eel2
as a “discrete H01 ” estimate on the error. A theoretical
∞
result which underlies the L estimate (2.17) is the fact that if Ω is an open bounded subset of IR, then
H01 (Ω) is imbedded in L∞ (Ω). This is no longer true in higher dimension. In two space dimensions,
thesejmf
for instance, a discrete version of the imbedding of H01 in Lp allows to obtain (see e.g. Fiard [1994]) norminf
kekp ≤ Ch, for all finite p, which in turn yields kek∞ ≤ Ch ln h for convenient meshes (see Corollary 3.1
norminf
page 62).
 19

The important features needed for the above proof seem to be the consistency of the approximation of
the fluxes and the conservativity of the scheme; this conservativity is natural the fact that the scheme is
obtained by integrating the equation over each cell, and the approximation of the flux on any interface
is obtained by taking into account the flux balance (continuity of the flux in the case of no source term
on the interface).
The above proof generalizes to other elliptic problems, such as a convection-diffusion equation of the form
−uxx + aux + bu = f , and to equations of the form −(λux )x = f where λ ∈ L∞ may be discontinuous,
and is such that there exist α and β in IR ?+ such that α ≤ λ ≤ β. These generalizations are studied
in the next section. Other generalizations include similar problems in 2 (or 3) space dimensions, with
meshes consisting of rectangles (parallepipeds), triangles (tetrahedra), or general meshes of Voronoı̈ type,
and the corresponding evolutive (parabolic) problems. These generalizations will be addressed in further
chapters. eel2
Let us now give a proof of Estimate (2.17), under slightly different conditions, which uses finite difference
techniques.

2.2.2 An error estimate using finite difference techniques

fdconvergence
Convergence can be obtainedFS
via a method similar to that WMof the finite difference proof of convergence
Fa1
(following, for instance, Forsyth and Sammon [1988], Manteuffel and White [1986], Faille elliptic1D
[1992a]). Most of these methods, are, however, limited to the finite volume method for Problem ( 2.1).
fdvf
Using the notations of Section 2.1.2 (recall that U = (u(x1 ), . . . , u(xN ))t , and r = AU − b = 0(1)), the
idea is to find U “close” to U , such that

AU = b + r, with r = 0(h).
FS
This value of U was found in Forsyth and Sammon [1988] and is such that U = U − V , where

h2i uxx (xi )

V = (v1 , . . . , vN )t and vi = , i = 1, . . . , N.
8
Then, one may decompose the truncation error as

r = A(U − U ) = AV + r with kV k∞ = 0(h2 ) and r = 0(h).

V
The existence of such a V is given in Lemmastab
2.1. In order to prove the convergence of the scheme, a
stability property is established in Lemma 2.2.
meshund meshund
V Lemma 2.1 Let T = (Ki )i=1,···,N be an admissible mesh of (0, 1), in the sense of Definition 2.1 page 12,
such that xi is the center of Ki for all i = 1, . . . , N . Let αT > 0 be such that hi > αT h for all i = 1, . .elliptic1D
.,N
(recall that h = max{h1 , . . . , hN }). Let U = (u(x1 ), . . . , u(xN ))t ∈ IR N , where u is the solutionmatrix
to (2.1),
and assume u ∈ C 3 ([0, 1], IR). Let A be the matrix defining the
matrix VF1
numerical
VF4
scheme, given in (2.10) page
14. Then there exists a unique U = (u1 , . . . , uN ) solution of (2.3)-(2.6) and there exists r and V ∈ IR N
such that
r = A(U − U ) = AV + r, with kV k∞ ≤ Ch2 and krk∞ ≤ Ch,
where C only depends on u and αT .
V
Proof of Lemma 2.1
eelapl1D
The existence and uniqueness of U is classical (it is also proved in Theorem 2.1).
For i = 0, . . . N , define

u(xi+1 ) − u(xi )
Ri+ 21 = − + ux (xi+ 12 ).
hi+ 12
Remark that
 20

1
ri = (R 1 − Ri− 21 ), for i = 0, . . . , N, (2.22) aieaie
hi i+ 2
where ri is the i−th component of r = A(U − U ).
The computation of Ri+ 21 yields

Ri+ 12 = − 41 (hi+1 − hi )uxx (xi+ 12 ) + 0(h2 ), i = 1, . . . , N − 1,

R 12 = − 41 h1 uxx (0) + 0(h2 ), RN + 12 = 14 hN uxx (1) + 0(h2 ).
h2i uxx (xi )
Define V = (v1 , . . . , vN )t with vi = 8 , i = 1, . . . , N . Then,
vi+1 − vi
− = Ri+ 21 + 0(h2 ), i = 1, . . . , N − 1,
hi+ 21
2v1
− = R 21 + 0(h2 ),
h1
2vN
= RN + 21 + 0(h2 ).
hN
aieaie
Since hi ≥ αT h, for i = 1, . . . , N , replacing Ri+ 12 in (2.22) gives that ri = (AV )i + 0(h), for i = 1, . . . , N ,
and kV k∞ = 0(h2 ). Hence the lemma is proved.

stab Lemma
meshund
2.2 (Stability) Let T = (Ki )i=1,···,N be an admissible meshmatrix
of [0, 1] in the sense of Definition
2.1. Let A be the matrix defining the finite volume scheme given in (2.10). Then A is invertible and
1
kAk−1
∞ ≤ . (2.23) eqstab
4
stab
Proof of Lemma 2.2
First we prove a discrete maximum principle; indeed if bi ≥ 0, for all i = 1, . . . , N , and if U is solution of
AU = b then we prove that ui ≥ 0 for all i = 1, . . . , N .
Let a = min{ui , i = 0, . . . , N + 1} (recall that u0 = uN +1 = 0) and i0 = min{i ∈ {0, . . . , N + 1}; ui = a}.
If i0 6= 0 and i0 6= N + 1, then
1  ui0 − ui0 −1 ui +1 − ui0 
− 0 = bi0 ≥ 0,
hi 0 hi0 − 21 hi0 + 21
this is impossible since ui0 +1 − ui0 ≥ 0 and ui0 − ui0 −1 < 0, by definition of i0 . Therefore, i0 = 0 or
N + 1. Then, a = 0 and ui ≥ 0 for all i = 1, . . . , N .
Note that, by linearity, this implies that A is invertible.
Next, we shall prove that there exists M > 0 such that kA−1 k∞ ≤ M (indeed, M = 1/4 is convenient).
Let φ be defined on [0, 1] by φ(x) = 12 x(1 − x). Then −φxx (x) = 1 for all x ∈ [0, 1]. Let Φ = (φ1 , . . . , φN )
with φi = φ(xi ); if A represented the usual finite difference approximation of the second order derivative,
then we would have AΦ = 1, since the difference quotient approximation of the second order derivative
VF1 VF4
of a second order polynomial is exact (φxxx = 0). Here, with the finite volume scheme (2.3)-(2.6), we
have AΦ − 1 = AW (where 1 denotes the vector of IR N the components of which are all equal to 1), with
h2 V
W = (w1 , . . . , wN ) ∈ IR N such that Wi = − 8i (see proof of Lemma 2.1). Let b ∈ IR N and AU = b, since
A(Φ − W ) = 1, we have
A(U − kbk∞ (Φ − W )) ≤ 0,
this last inequality being meant componentwise. Therefore, by the above maximum principle, assuming,
without loss of generality, that h ≤ 1, one has
kbk∞
ui ≤ kbk∞ (φi − wi ), so that ui ≤ .
4
 21

(note that φ(x) ≤ 18 ). But we also have

A(U + kbk∞ (Φ − W )) ≥ 0,

and again by the maximum principle, we obtain

kbk∞
ui ≥ − .
4
Hence kU k∞ ≤ 14 kbk∞ . This shows that kA−1 k∞ ≤ 14 .
V
This stability result, together with the existence of V given by Lemma 2.1, yields the convergence of the
finite volume scheme, formulated in the next theorem.
meshund
Theorem
meshund
2.2 Let T = (Ki )i=1,···,N be an admissible mesh of [0, 1] in the sense of Definition 2.1 page
12. Let αT ∈ IR ?+ be such that hi ≥ αT h, for all i = 1, . . . , N (recall that h = max{h1 , . . . , hN }). Let
elliptic1D
U = (u(x1 ), . . . , u(xN ))t ∈ IR N , and assume u ∈ C 3 ([0, 1], IR) (recall
VF1
thatVF4
u is the solution to (2.1)). Let
U = (u1 , . . . , uN ) be the solution given by the numerical scheme (2.3)-(2.6). Then there exists C > 0,
only depending on αT and u, such that kU − U k∞ ≤ Ch.
stab
rubarbare Remark 2.8 In the proof of Lemma 2.2, it was shown that A(U − V ) = b + 0(h); therefore, if, once
again, the finite volume scheme is considered as a finite difference scheme, it is consistent, in the finite
difference sense, when ui is considered to be an approximation of u(xi ) − (1/8)h2i uxx (xi ).
V
wc Remark 2.9 With the notations of Lemma 2.1, let r be the function defined by

r(x) = ri , if x ∈ Ki , i = 1, . . . , N,

the function r does not necessarily go to 0 (as h goes to 0) in the L∞ norm (and even in the L1 norm),
but, thanks to the conservativity of the scheme, it goes to 0 in L∞ ((0, 1)) for the weak-? topology, that
is Z 1
r(x)ϕ(x)dx → 0, as h → 0, ∀ϕ ∈ L1 ((0, 1)).
0
This property will be called “weak consistency”
Fa1
in the sequel and may also be used to prove the conver-
gence of the finite volume scheme (see Faille [1992a]).

The proof of convergence described above may be easily generalized to the two-dimensional Laplace
equation −∆u = f in two and three space dimensions if a rectangular or a parallepipedic mesh is used,
provided that the solution u is of class C 3 . However, it does not seem to be easily generalized to other
types of meshes.

2.3 General 1D elliptic equations

ell1D
2.3.1 Formulation of the finite volume scheme
This section is devoted to the formulation and to the proof of convergence of a finite volume scheme for
a one-dimensional linear convection-diffusion equation, with a discontinuous diffusion coefficient. The
scheme can be generalized in the two-dimensional and three-dimensional cases (for a spacevfquatre
discretization
vfquatre
which uses, for instance, simplices or parallelepipedes or a “Voronoı̈ mesh”, see Section 3.1.2 page 37)
and to other boundary conditions.

Let λ ∈ L∞ ((0, 1)) such that there exist λ and λ ∈ IR ?+ with λ ≤ λ ≤ λ a.e. and let a, b, c, d ∈ IR, with
b ≥ 0, and f ∈ L2 ((0, 1)). The aim, here, is to find an approximation to the solution, u, of the following
problem:
−(λux )x (x) + aux (x) + bu(x) = f (x), x ∈ [0, 1], (2.24) eq1
 22

u(0) = c, u(1) = d. (2.25) cl

The discontinuity of the coefficient λ may arise for instance for the permeability of a porous medium,
the ratio between the permeability of sand and the permeability of clay being of an order of 10 3 ; heat
conduction in a heterogeneous medium can also yield such discontinuities, since the conductivities of the
different components of the medium may be quite different. Note that the assumption b ≥ 0 ensures the
existence of the solution to the problem.
eq1 cl
Remark 2.10 Problem (2.24)-(2.25) has a unique solution u in the Sobolev space H 1 ((0, 1)). This
2
R x but is not, in general, of class C (even if λ(x) = 1, for all x 1∈ [0, 1]).
solution is continuous (on [0, 1])
Note that one has −λux (x) = 0 g(t)dt + C, where C is some constant and g = f − aux − bu ∈ L ((0, 1)),
so that λux is a continuous function and ux ∈ L∞ ((0, 1)).
meshund meshund
Let T = (Ki )i=1,···,N be an admissible mesh, in the sense of Definition 2.1 page 12, such that the
discontinuities of λ coincide with the interfaces of the mesh.
Diric1D eq1
The notations being the same as in section 2.1, integrating Equation (2.24) over Ki yields
Z Z
−(λux )(xi+ 12 ) + (λux )(xi− 21 ) + au(xi+ 12 ) − au(xi− 12 ) + bu(x)dx = f (x)dx, i = 1, . . . , N.
Ki Ki

Let (ui )i=1,···,N be the discrete unknowns. In the case a ≥ 0, which will be considered in the sequel,
the convective term au(xi+1/2 ) is approximated by aui (“upstream”) because of stability considerations.
Indeed, this choice always yields a stability result whereas the approximation of au(x i+1/2 ) by (a/2)(ui +
ui+1 ) (with the approximation of the other terms as it is done below) yields a stable scheme if ah ≤ 2λ,
R λ. The case a ≤ 0 is easily handled in the
for a uniform mesh of size h and a constant diffusion coefficient
same way by approximating au(xi+1/2 ) by aui+1 . The term Ki bu(x)dx is approximated by bhi ui . Let
R
us now turn to the approximation Hi+1/2 of −λux (xi+1/2 ). Let λi = h1i Ki λ(x)dx; since λ|Ki ∈ C 1 (K̄i ),
there exists cλ ∈ IR + , only depending on λ, such that |λi − λ(x)| ≤ cλ h, ∀ x ∈ Ki . In order that the
scheme be conservative, the discretization of the flux at xi+1/2 should have the same value on Ki and
Ki+1 . To this purpose, we introduce the auxiliary unknown ui+1/2 (approximation of u at xi+1/2 ). Since
on Ki and Ki+1 , λ is continuous, the approximation of −λux may be performed on each side of xi+1/2
by using the finite difference principle:
ui+ 21 − ui
Hi+ 21 = −λi on Ki , i = 1, . . . , N,
h+i
ui+1 − ui+ 12
Hi+ 21 = −λi+1 on Ki+1 , i = 0, . . . , N − 1,
h−
i+1

with u1/2 = c, and uN +1/2 = d, for the boundary conditions. (Recall that h+ i = xi+1/2 − xi and
h−
i = x i − x i−1/2 ). Requiring the two above approximations of λu x (x i+1/2 ) to be equal (conservativity
of the flux) yields the value of ui+1/2 (for i = 1, . . . , N − 1):
λi+1 λi
ui+1 + ui +
h−
i+1 h i
ui+ 21 = (2.26) uipud
λi+1 λi
+ +
h−
i+1 hi
which, in turn, allows to give the expression of the approximation Hi+ 21 of λux (xi+ 21 ):

Hi+ 12 = −τi+ 21 (ui+1 − ui ), i = 1, . . . , N − 1,

λ1
H 12 = − − (u1 − c),
h1 (2.27) vfsig
λN
HN + 12 = − + (d − uN )
hN
 23

with
λi λi+1
τi+ 12 = , i = 1, . . . , N − 1. (2.28) tau1D
hi λi+1 + h−
+
i+1 λi

Example 2.2 If hi = h, for all i ∈ {1, . . . , N }, and xi is assumed to be the center of Ki , then h+
i =
h− h
i = 2 , so that
2λi λi+1 ui+1 − ui
Hi+ 12 = − ,
λi + λi+1 h
and therefore the mean harmonic value of λ is involved.
eq1 cl
The numerical scheme for the approximation of Problem (2.24)-(2.25) is therefore,

Fi+ 21 − Fi− 12 + bhi ui = hi fi , ∀i ∈ {1, . . . , N }, (2.29) eqd1

1
R xi+ 1
with fi = hi xi− 1
2 f (x)dx, for i = 1, . . . , N , and where (Fi+ 12 )i∈{0,...,N } is defined by the following
2
expressions

Fi+ 12 = −τi+ 21 (ui+1 − ui ) + aui , ∀i ∈ {1, . . . , N − 1}, (2.30) eqd2

λ1 λN
F 12 = − (u1 − c) + ac, FN + 21 = − + (d − uN ) + auN . (2.31) cld
h−
1 h N

Remark 2.11 In the case a ≥ 0, the choice of the approximation of au(xi+1/2 ) by aui+1 would yield an
unstable scheme, except for h small enough (when a ≤ 0, the unstable scheme is aui ).
tau1D eqd2 cld eqd1
Taking (2.28), (2.30) and (2.31) into account, the numerical scheme (2.29) yields a system of N equations
with N unknowns u1 , . . . , uN .

2.3.2 Error estimate

ce1r
ee1dell Theorem 2.3
Let a, b ≥ 0, c, d ∈ IR, λ ∈ L∞ ((0,eq11)) such
cl
that λ ≤ λ ≤ λ a.e. with some λ, λ ∈ IR ?+ and f ∈ L1 ((0, 1)).
(unique) solution of (2.24)-(2.25). Let T = (Ki )i=1,···,N be an admissible mesh, in the sense of
Let u be themeshund
Definition 2.1, such that λ ∈ C 1 (K i ) and f ∈ C(K i ), for all i = 1, · · · , N . Let γ = max{kuxxkL∞(Ki ) , i =
1, · · · , N } and δ = max{kλkL∞ (Ki ) , i = 1, · · · , N }. Then,
tau1D cld
1. there exists a unique vector U = (u1 , . . . , uN )t ∈ IR N solution to (2.28)-(2.31),
2. there exists C, only depending on λ, λ, γ and δ, such that
N
X
τi+ 21 (ei+1 − ei )2 ≤ Ch2 , (2.32) ee1
i=0
tau1D
where τi+ 12 is defined in (2.28), and

|ei | ≤ Ch, ∀i ∈ {1, . . . , N }, (2.33) ee2

with e0 = eN +1 = 0 and ei = u(xi ) − ui , for all i ∈ {1, . . . , N }.

 24

ee1dell
Proof of Theorem 2.3
tau1D cld
Step 1. Existence eqd1
and uniqueness of the solution to (2.28)-(2.31).
Multiplying (2.29) by ui and summing for i = 1, . . . , N yields that if c = d = 0 and fi = 0 for any i ∈
tau1D cld
{1, . . . , N }, then the unique solution to (2.28)-(2.31) is obtained by taking ui = 0, for any i ∈ {1, . . . , N }.
tau1D cld
This yields existence and uniqueness of the solution to (2.28)-(2.31).

Step 2. Consistency of the fluxes.

Recall that h = max{h1 , . . . , hN }. Let us first show the consistency of the fluxes.
?
Let H i+1/2 = −(λux )(xi+1/2 ) and Hi+1/2 = −τi+1/2 (u(xi+1 )−u(xi )), for i = 0, . . . , N , with τ1/2 = λ1 /h−
1
and τN +1/2 = λN /h+ ?
N . Let us first show that there exists C1 ∈ IR + , only depending on λ, λ, γ and δ,
such that
?
Hi+ 1 = H i+ 1 + Ti+ 1 ,
2 2 2
(2.34) lambdaconsis
|Ti+ 12 | ≤ C1 h, i = 0, . . . , N.
In order to show this, let us introduce

?,−
u(xi+ 21 ) − u(xi ) ?,+
u(xi+1 ) − u(xi+ 21 )
Hi+ 1 = −λi and Hi+ 1 = −λi+1 ; (2.35) Gast
2 h+
i
2 h−
i+1

since λ ∈ C 1 (K̄i ), one has u ∈ C 2 (K̄i ); hence, there exists C ∈ IR ?+ , only depending on γ and δ, such
that
?,− − −
Hi+ 1 = H i+ 1 + R
2 i+ 1
, where |Ri+ 1 | ≤ Ch, i = 1, . . . , N, (2.36) eqRun
2 2 2

and
?,+ + +
Hi+ 1 = H i+ 1 + R
2 i+ 1
, where |Ri+ 1 | ≤ Ch, i = 0, . . . , N − 1. (2.37) eqRde
2 2 2
lambdaconsist
This yields (2.34) for i = 0 and i = N .
The following equality:
?,− − ?,+ +
H i+ 12 = Hi+ 1 − R = Hi+ 1 − R , i = 1, . . . , N − 1, (2.38) Gbarast
2 i+ 1 2 i+ 1 2 2

yields that
λi+1 λi
u(xi+1 ) + + u(xi )
h− h
u(xi+ 21 ) = i+1 i
+ Si+ 12 , i = 1, . . . , N − 1, (2.39) lambdauipud
λi λi+1
+
h+i h−
i+1

where
+ −
Ri+ 1 − R
i+ 1
S i+ 12 = λi
2
λi+1
2

h+
+ h−
i i+1

so that
+ −
1 hi hi+1 + −
|Si+ 21 | ≤ − |Ri+ 21 − Ri+ 12 |.
λ h+
i + h i+1
lambdauipud ?,− Gast
Let us replace the expression (2.39) of u(xi+1/2 ) in Hi+1/2 defined by (2.35) (note that the computation
uipud vfsig
is similar to that performed in (2.26)-(2.27)); this yields

?,− λi
Hi+ 1 = −τi+ 1 (u(xi+1 ) − u(xi )) − Si+ 21 , i = 1, . . . , N − 1. (2.40) computGast
2 2
h+
i
 25

Gbarast ?
Using (2.38), this implies that Hi+ 1 = H i+ 1 + Ti+ 1 where
2 2
2

− + − λ
|Ti+ 12 | ≤ |Ri+ 1 | + |R
i+ 1
− Ri+ 1| .
2 2 2 2λ
eqRun eqRde
Using (2.36) and (2.37), this last inequality yields that there exists C1 , only depending on λ, λ, γ, δ, such
that
?
|Hi+ 1 − H i+ 1 | = |Ti+ 1 | ≤ C1 h, i = 1, . . . , N − 1.
2 2
2
lambdaconsist
Therefore (2.34) is proved.
Define now the total exact fluxes;

F i+ 12 = −(λux )(xi+ 21 ) + au(xi+ 12 ), ∀i ∈ {0, . . . , N },

and define
?
Fi+ 1 = −τi+ 1 (u(xi+1 ) − u(xi )) + au(xi ), ∀i ∈ {1, . . . , N − 1},
2 2

λ1 λN
F 1? = − ?
− (u(x1 ) − c) + ac, FN + 21 = − + (d − u(xN )) + auN .
2 h1 hN
lambdaconsist
Then, from (2.34) and the regularity of u, there exists C2 , only depending on λ, λ, γ and δ, such that
?
Fi+ 1 = F i+ 1 + Ri+ 1 , with |Ri+ 1 | ≤ C2 h, i = 0, . . . , N. (2.41) consistflux1
2 2 2 2

Hence the numerical approximation of the flux is consistent.

Step 3. Error estimate.

eq1
Integrating Equation (2.24) over each control volume yields that

F i+ 12 − F i− 21 + bhi (u(xi ) + Si ) = hi fi , ∀i ∈ {1, . . . , N }, (2.42) eqdc1

i ∈ IR is such that there exists C3 only depending on u such that |Si | ≤ C3 h, for i = 1, . . . , N .
where Sconsistflux1D
Using (2.41) yields that
? ?
1 − F
Fi+ i− 1 + bhi (u(xi ) + Si ) = hi fi + Ri+ 2 − Ri− 2 , ∀i ∈ {1, . . . , N }. (2.43) eqfast1D
1 1
2 2
eqd1 eqfast1D
Let ei = u(xi ) − ui , for i = 1, . . . , N , and e0 = eN +1 = 0. Substracting (2.29) from (2.43) yields

−τi+ 12 (ei+1 − ei ) + τi− 12 (ei − ei−1 ) + a(ei − ei−1 ) + bhi ei = −bhi Si + Ri+ 21 − Ri− 12 , ∀i ∈ {1, . . . , N }.

Let us multiply this equation by ei , sum for i = 1, . . . , N , reorder the summations. Remark that
N
X N +1
1 X
ei (ei − ei−1 ) = (ei − ei−1 )2
i=1
2 i=1
and therefore
N
X N +1 N N N
a X X X X
τi+ 12 (ei+1 − ei )2 + (ei − ei−1 )2 + bhi e2i = − bhi Si ei − Ri+ 12 (ei+1 − ei ).
i=0
2 i=1 i=1 i=1 i=0
consistflux1D
Since |Si | ≤ C3 h and thanks to (2.41), one has
 26

N
X N
X N
X
τi+ 12 (ei+1 − ei )2 ≤ bC3 hi h|ei | + C2 h|ei+1 − ei |.
i=0 i=1 i=1

PN P  12 P  21
N 2 N 1
Remark that |ei | ≤ j=1 |ej − ej−1 |. Denote by A = i=0 τi+ 12 (ei+1 − ei ) and B = i=0 τi+ 1 .
2
The Cauchy-Schwarz inequality yields
N
X
A2 ≤ bC3 hi hAB + C2 hAB.
i=1

Now, since

XN XN
1 λ − + − + + −
≤ (h i+1 + h i ), (h i+1 + h i ) = 1, with h 0 = h N +1 = 0, and hi = 1,
τi+ 12 λ2 i=0 i=1
ee1
one obtains that A ≤ C4 h, with C4 only depending on λ, λ, γ and δ, which
ee2
yields Estimate (2.32).
Applying once again the Cauchy-Schwarz inequality yields Estimate (2.33).

2.3.3 The case of a point source term

In many physical problems, some discontinuous or point source terms appear. In the case where a
source term exists at the interface xi+1/2 , the fluxes relative to Ki and Ki+1 will differ because of this
source term. The computation of the fluxes is carried out in a similar way, writing that the sum of the
approximations of the fluxes must be equalclto the source term at the interface. Consider again the one-
eq1
dimensional conservation problem (2.24), (2.25) (with, for the sake of simplification, a = b = c = d = 0,
we use below the notations of the previous section), but assume now that at x ∈ (0, 1), a point source of
intensity α exists. In this case, the problem may be written in the following way:

−(λux (x))x = f (x), x ∈ (0, x) ∪ (x, 1), (2.44) eqdisc

u(0) = 0, (2.45) cldisc1

u(1) = 0, (2.46) cldisc2

(λux )+ (x) − (λux )− (x) = −α, (2.47) fluxdisc

where
(λux )+ (x) = lim (λux )(x) and (λux )− (x) = lim (λux )(x).
x→x,x>x x→x,x<x
fluxdisc
Equation (2.47) states that the flux is discontinuous at point x. Another formulation of the problem is
the following:

−(λux )x = g in D0 ((0, 1)), (2.48) eqmatdisc

u(0) = 0, (2.49) cldisc21

u(1) = 0, (2.50) cldisc22

where g = f + αδx , where δx denotes the Dirac measure, which is defined by < δx , ϕ >D0 ,D = ϕ(x), for
any ϕ ∈ D((0, 1)) = Cc∞ ((0, 1), IR), and D 0 ((0, 1)) denotes the set of distributions on (0,1), i.e. the set of
continuous linear forms on D((0, 1)).
Assuming the mesh to be such that x = xi+1/2 for some i ∈ 1, . . . , N − 1, the equation corresponding to
−
R
the unknown ui is Fi+1/2 − Fi−1/2 = Ki f (x)dx, while the equation corresponding to the unknown ui+1
+
R ±
is Fi+3/2 − Fi+1/2 = Ki+1 f (x)dx. In order to compute the values of the numerical fluxes Fi+1/2 , one
 27

must take the source term into account while writing the conservativity of the flux; hence at x i+1/2 , the
+ − fluxdisc
two numerical fluxes at x = x, namely Fi+ 1 and F
i+ 1
, must satisfy, following Equation (2.47),
2 2

+ −
Fi+ 1 − Fi+ 1 = α. (2.51) ddf
2 2

+ −
Next, the fluxes Fi+1/2 and Fi+1/2 must be expressed in terms of the discrete variables uk , k = 1, . . . , N ;
in order to do so, introduce the auxiliary variable ui+1/2 (which will be eliminated later), and write

+
ui+1 − ui+ 12
Fi+ 1 = −λi+1
2 h−
i+1

−
ui+ 21 − ui
Fi+ 1 = −λi .
2 h+
i
ddf
Replacing these expressions in (2.51) yields
−
h+
i hi+1 λi+1 λi
ui+ 12 = [ − ui+1 + ui + α].
(h−
i+1 λi + +
hi λi+1 ) hi+1 h+
i

and therefore

+ h+
i λi+1 λi λi+1
Fi+ 1 = − α − (ui+1 − ui )
2 hi+1 λi + h+
i λi+1 h− +
i+1 i + hi λi+1
λ

− −h−
i+1 λi λi λi+1
Fi+ 1 = α − (ui+1 − ui ).
2 h−
i+1 λi + h+
i λi+1 h−
i+1 λi+ h+i λi+1

Note that the source term α is distributed on either side of the interface proportionally to the coefficient
λ, and that, when α = 0, the above expressions lead to

+ − λi λi+1
Fi+ 1 = F
i+ 1
=− − (ui+1 − ui ).
2 2 hi+1 λi + h+
i λi+1
ee1dell
Note that the error estimate given in Theorem 2.3 still holds in this case (under adequate assumptions).

2.4 A semilinear elliptic problem

2.4.1 Problem and Scheme
This section is concerned with the proof of convergence for some nonlinear problems. We are interested,
as an example, by the following problem:

−uxx (x) = f (x, u(x)), x ∈ (0, 1), (2.52) enl1d

u(0) = u(1) = 0, (2.53) enl1dcl

with a function f : (0, 1) × IR → IR such that

f (x, s) is measurable with respect to x ∈ (0, 1) for all s ∈ IR

(2.54) carat
and continuous with respect to s ∈ IR for a.e. x ∈ (0, 1),

f ∈ L∞ ((0, 1) × IR). (2.55) bornee

enl1d enl1dcl
It is possible to prove that there exists at least one weak solution to (2.52), (2.53), that is a function u
such that
 28

Z 1 Z 1
u ∈ H01 ((0, 1)), ux (x)vx (x)dx = f (x, u(x))v(x)dx, ∀v ∈ H01 ((0, 1)). (2.56) enl1dw
0 0
enl1dw
Note that (2.56) is equivalent to “u ∈ H01 ((0, 1)) and −uxx = f (·, u) in the distribution sense in (0, 1)”.
The proof of the existence
deimling
of such a solution is possible by using, for instance,enl1dcr
the Schauder’s fixed point
theorem (see e.g. Deimling [1980]) or by using the convergence theorem 2.4 which is proved in the
sequel.
meshund enl1d enl1dcl
Let T be an admissible mesh of [0, 1] in the sense of Definition 2.1. In order to discretize (2.52), (2.53),
let us consider the following (finite volume) scheme

Fi+ 12 − Fi− 12 = hi fi (ui ), i = 1, . . . , N, (2.57) enl1dd

ui+1 − ui
Fi+ 12 = − , i = 0, . . . , N, (2.58) enl1df
hi+ 12

u0 = uN +1 = 0, (2.59) enl1dcld
1
R
with fi (ui ) = hi Ki f (x, ui )dx, i = 1, . . . , N . The discrete unknowns are therefore u1 , . . . , uN .
enl1dcr
In order to give a convergence result for this scheme (Theorem 2.4), one first proves
enl1dd enl1dcld
the existence of a
enl1dd enl1dcld
solution
enl1dl1
to ( 2.57)-( 2.59), a stability result, that
enl1dl2
is, an estimate on the solution of ( 2.57)-( 2.59) (Lemma
2.3) and a compactness lemma (Lemma 2.4).
carat bornee
enl1dl1 Lemma 2.3 (Existence and stability result) Let f : (0,meshund 1) × IR → IR satisfying ( 2.54), (2.55) and
T be an admissible mesh of
enl1dd enl1dcld
(0, 1) in the sense of Definition 2.1. Then, there exists (u1 , . . . , uN )t ∈ IR N
solution of (2.57)-(2.59) and which satisfies:
N
X (ui+1 − ui )2
≤ C, (2.60) h1dest
i=0
hi+ 21
for some C ≥ 0 only depending on f .
enl1dl1
Proof of Lemma 2.3
h1dest
kf kL∞enl1dcld
Define M = enl1dd ((0,1)×IR) . The proof of estimate (2.60) is given in a first step, and the existence of a
solution to (2.57)-(2.59) in a second step.

Step 1 (Estimate) enl1dd enl1dcld

Let V = (v1 , . . . , vN )t ∈ IR N , there exists a unique U = (u1 , . . . ,eelapl1D
uN )t ∈ IR N
solution of (2.57)-(2.59)
eelapl1D
with fi (vi ) instead of fi (ui ) in the right hand-side (see Theorem 2.1 page 16). One sets U = F (V ), so
enl1dd enl1dcld
that F is a continuous application from IR N to IR N , and (u1 , . . . , uN ) is a solution to (2.57)-(2.59) if and
t
only if U = (uenl1dd
1 , . . . , uN ) is a fixed point to F .
Multiplying (2.57) by ui and summing over i yields
N
X N
X
(ui+1 − ui )2
≤M hi |ui |, (2.61) preest
i=0
hi+ 21 i=1

and from the Cauchy-Schwarz inequality, one has

N
X (uj+1 − uj )2
1
|ui | ≤ 2
, i = 1, . . . , N,
j=0
hj+ 12
preest
then (2.61) yields, with C = M 2 ,
 29

N
X (ui+1 − ui )2
≤ C. (2.62) h1destbis
i=0
hi+ 21
h1dest enl1dd enl1dcld
This gives, in particular, Estimate (2.60) if (u1 , . . . , uN )t ∈ IR N is a solution of (2.57)-(2.59) (that is
ui = vi for all i).

Step 2 (Existence)
The application F : IR N → IR N defined above is continuous and, taking in IR N the norm
N
X (vi+1 − vi )2
1
kV k = 2
, for V = (v1 , . . . , vN )t , with v0 = vN +1 = 0,
i=0
hi+ 12

one has F (BM ) ⊂ BM , where BM is the closed ball of radius M and center 0 in IR N . Then, F has a
deimling
fixed point in BM thanks to the Brouwer fixed point theorem (see e.g. Deimling [1980]). This fixed
enl1dd enl1dcld
point is a solution to (2.57)-(2.59).

2.4.2 Compactness results

.

enl1dl2 Lemma 2.4 (Compactness) meshund h1dest

For an admissible mesh T of (0, 1) (see definition 2.1), let (u1 , . . . , uN )t ∈ IR N satisfy (2.60) for some
C ∈ IR (independent of T ) and let uT : (0, 1) → IR be defined by uT (x) = ui if x ∈ Ki , i = 1, . . . , N .
Then, the set {uT , T admissible mesh of (0, 1)} is relatively compact in L2 ((0, 1)). Furthermore, if
uTn → u in L2 ((0, 1)) and size(Tn ) → 0, as n → ∞, then, u ∈ H01 ((0, 1)).
enl1dl2
Proof of Lemma 2.4
A possible proof is to use “classical” compactness results, replacing uT by a continuous function, say
uT , piecewise affine, such that uT (xi ) = ui for i = 1, . . . , N , and uT (0) = u (1) = 0. The set {uT , T
Liebig T Liebig
admissible mesh of (0, 1)} is then bounded in H01 ((0, 1)), see Remark 3.9 page 49.
Another proof is given here, the interest of which is its simple generalization tovfquatre
multidimensional
vfquatre
cases
(such as the case
rellichd rellichd
of one unknown per triangle in 2 space dimensions, see Section 3.1.2 page 37 and Section
3.6 page 92) when the construction of such a function, uT , “close” to uT and bounded in H01 ((0, 1))
(independently of T ), is not so easy.

In order to have uT defined on IR, one sets uT (x) = 0 for x ∈

/ [0, 1]. The proof may be decomposed into
four steps.

the set {uT , T an admissible mesh of (0, 1)} is bounded in L2 (IR). Indeed, this
Step 1. First remark thath1dest
an easy consequence of (2.60), since one has, for all x ∈ [0, 1] (since u0 = 0 and by the Cauchy-Schwarz
inequality),
N
X N
X (ui+1 − ui )2 1
|uT (x)| ≤ |ui+1 − ui | ≤ ( ) 2 ≤ C.
i=0 i=0
hi+ 12

Step 2. Let 0 < η < 1. One proves, in this step, that

kuT (· + η) − uT k2L2 (IR) ≤ Cη(η + 2h). (2.63) cleest

(Recall that h = size(T ).)
 30

Indeed, for i ∈ {0, . . . , N } define χi+1/2 : IR → IR, by χi+1/2 (x) = 1, if xi+1/2 ∈ [x, x+η] and χi+1/2 (x) =
0, if xi+1/2 ∈
/ [x, x + η]. Then, one has, for all x ∈ IR,
N
X
2
(uT (x + η) − uT (x))2 ≤ |ui+1 − ui |χi+ 21 (x)
i=0

X
N
(ui+1 − ui )2 X N 
≤ χi+ 12 (x) χi+ 12 (x)hi+ 21 . (2.64) clest
i=0
hi+ 21 i=0
P R
Since N i=0 χ i+1/2 (x)hi+1/2 ≤ η +
clest
2h, for all x ∈ IR, and IR χi+1/2 (x)dx = η, for all i ∈ {0, . . . , N },
cleest
integrating (2.64) over IR yields (2.63).
cleest
Step 3. For 0 < η < 1, Estimate (2.63) implies that

kuT (· + η) − uT k2L2 (IR) ≤ 3Cη.

rellichd
This
Kolm
givesKolm
(with Step 1), by the Kolmogorov compactness theorem (recalled in Section 3.6, see Theorem
3.9 page 93), the relative compactness of the set {uT , T an admissible mesh of (0, 1)} in L2 ((0, 1)) and
also in L2 (IR) (since uT = 0 on IR \ [0, 1]).
enl1dl2 kolmh10 kolmh10
Step 4. In order to conclude the proof of Lemma 2.4, one may use Theorem 3.10 page 93, which we prove
here in the one-dimensional case for the sake of clarity. Let (Tn )n∈IN be a sequence of admissible meshes
of (0, 1) such that size(Tn ) → 0 and uTn → u, in L2 ((0, 1)), as n → ∞. Note that uTn → u, in L2 (IR),
cleest
with u = 0 on IR \ [0, 1]. For a given η ∈ (0, 1), let n → ∞ in (2.63), with uTn instead of uT (and size(Tn )
instead of h). One obtains

u(· + η) − u 2
k kL2 (IR) ≤ C. (2.65) preh1
η
Since (u(· + η) − u)/η tends to Du (the distribution derivative of u) in the distribution sense, as η → 0,
preh1
Estimate (2.65) yields that Du ∈ L2 (IR). Furthermore,
enl1dl2
since u = 0 on IR \ [0, 1], the restriction of u to
(0, 1) belongs to H01 ((0, 1)). The proof of Lemma 2.4 is complete.

2.4.3 Convergence
enl1dl1 enl1dl2
The following convergence result follows from lemmata 2.3 and 2.4.
carat bornee
enl1dcr Let f : (0, 1) × IR → IR satisfying (2.54), (2.55).
Theorem 2.4 meshund For an admissible mesh, T , of (0, 1)
enl1dd enl1dcld
t N
(see Definition 2.1),
enl1dl1
let (u 1 , . . . , u N ) ∈ IR be a solution to ( 2.57)-( 2.59) (the existence of which is given
by Lemma 2.3), and let uT : (0, 1) → IR by uT (x) = ui , if x ∈ Ki , i = 1, . . . , N .
Then, for any sequence (Tn )n∈IN of admissible meshes such that size(Tn ) → 0, as n → ∞, there exists a
subsequence, still denoted n )n∈IN , such that uTn → u, in
by (Tenl1dcl
enl1d
L2 ((0, 1)), as n → ∞, where u ∈ H01 ((0, 1))
enl1dw
is a weak solution to (2.52), (2.53) (that is, a solution to (2.56)).
enl1dcr
Proof of Theorem 2.4
Let N be a sequence of admissible meshes of (0, 1) such that size(T n ) → 0, as n → ∞. By lemmata
(T ) enl1dl2
enl1dl1n n∈I
2.3 and 2.4, there exists a subsequence, still denoted by (Tn )n∈IN , such that uTn → u, in L2 ((0, 1)), as
n → ∞, where u ∈ H01 ((0, 1)). In order to conclude, it only remains to prove that −uxx = f (·, u) in the
distribution sense in (0, 1).
To prove this, let ϕ ∈ Cc∞ ((0, 1)). Let T be an admissible mesh of (0, 1), and ϕ
enl1dd enl1dcld i = ϕ(xi ), i = 1, . . . , N ,
enl1dd
and ϕ0 = ϕN +1 = 0. If (u1 , . . . , uN ) is a solution to (2.57)-(2.59), multiplying (2.57) by ϕi and summing
over i = 1, . . . , N yields
 31

Z 1 Z 1
uT (x)ψT (x)dx = fT (x)ϕT (x)dx, (2.66) yenamar
0 0
where
1 ϕi − ϕi−1 ϕi+1 − ϕi
ψT (x) = ( − ), fT (x) = f (x, ui ) and ϕT (x) = ϕi , if x ∈ Ki .
hi hi− 21 hi+ 12
Note that, thanks to the regularity of the function ϕ,
ϕi+1 − ϕi
= ϕx (xi+ 21 ) + Ri+ 12 , |Ri+ 12 | ≤ C1 h,
hi+ 12
with some C1 only depending on ϕ, and therefore
Z N Z
ui  
1 X XN
uT (x)ψT (x)dx = ϕx (xi− 12 ) − ϕx (xi+ 12 ) dx + ui (Ri− 12 − Ri+ 21 )
0 h
i=1 Ki i i=1
Z 1 N
X
= −uT (x)θT (x)dx + Ri+ 21 (ui+1 − ui ),
0 i=0

with u0 = uN +1 = 0, where the piecewise constant function

X ϕx (xi+ 1 ) − ϕx (xi− 1 )
θT = 2 2
1 Ki
hi
i=1,N

tends to ϕxx as hyenamar

tends to 0.
Let us consider (2.66) with Tn instead
h1dest
of T ; thanks to the Cauchy-Schwarz inequality, a passage to the
limit as n → ∞ gives, thanks to (2.60),
Z 1 Z 1
− u(x)ϕxx (x)dx = f (x, u(x))ϕ(x)dx,
0 0
enl1dcr
and therefore −uxx = f (·, u) in the distribution sense in (0, 1). This concludes the proof of Theorem 2.4.
Note that the crucial idea of this proof is to use the property of consistency of the fluxes on the regular
test function ϕ.

sublinear Remark
enl1dcr
2.12 It is possible to give some extensions of the results of this section. For instance, Theorem
bornee
2.4 is true with an assumption of “sublinearity” enl1dw
on f instead of (2.55). Furthermore, in order to have
both existence and uniqueness of the solution
enl1dcr
to (2.56) bornee
carat
and a rate of convergence (of order h) in Theorem
2.4, it is sufficient to assume, instead of (2.54) and (2.55), that f ∈ C 1 ([0, 1] × IR, IR) and that there
exists γ < 1, such that (f (x, s) − f (x, t))(s − t) ≤ γ(s − t)2 , for all (x, s) ∈ [0, 1] × IR.
 Chapter 3

Elliptic problems in two or three

dimensions
ellmd
The topic of this chapter is the discretization of elliptic problems in several space dimensions by the
ellund
finite volume method. The one-dimensional case which was studied in Chapter 2 is easily generalized
to nonuniform rectangular or parallelipedic meshes. However, for general shapes of control volumes,
the definition of the scheme (and the proof of convergence) requires some assumptions which define an
“admissible mesh”. Dirichlet and Neumann boundary conditions are both considered. In both cases, a
discrete Poincaré inequality is used, and the stability of the scheme is proved by establishing estimates
on the approximate solutions. The convergence of the scheme without any assumption on the regularity
of the exact solution is proved; this result may be generalized, under adequate assumptions, to nonlinear
equations. Then, again in both the Dirichlet and Neumann cases, an error estimate between the finite
volume approximate solution and the C 2 or H 2 regular exact solution to the continuous problems are
proved. The results scvfe
are generalized to the case of matrix diffusion coefficients and more general boundary
conditions. Section 3.4 is devoted to finite volume schemes written with unknowns located at the vertices.
Some links between the finite element method, F1
the “classical” finite volume method and the “control
meshref
volume finite element” method introduced by Forsyth [1989] are given. Section 3.5 is devoted to the
treatment of singular sources and to mesh refinement; under suitable assumption,
rellichd
it can be shown that
error estimates still hold for “atypical” refined meshes. Finally, Section 3.6 is devoted to the proof of
compactness results which are used in the proofs of convergence of the schemes.

3.1 Dirichlet boundary conditions

sdbc
Let us consider here the following elliptic equation

−∆u(x) + div(vu)(x) + bu(x) = f (x), x ∈ Ω, (3.1) elldifstaf

with Dirichlet boundary condition:

u(x) = g(x), x ∈ ∂Ω, (3.2) elldifstab

where

ellH Assumption 3.1

1. Ω is an open bounded polygonal subset of IR d , d = 2 or 3,

2. b ≥ 0,
3. f ∈ L2 (Ω),

32
 33

4. v ∈ C 1 (Ω, IR d ); divv ≥ 0,
5. g ∈ C(∂Ω, IR) is such that there exists g̃ ∈ H 1 (Ω) such that γ(g̃) = g a.e. on ∂Ω.

Here, and in the sequel, “polygonal” is used for both d = 2 and d = 3 (meaning polyhedral in the latter
case) and γ denotes the trace operator from H 1 (Ω) into L2 (∂Ω). Note also that “a.e. on ∂Ω” is a.e. for
the d − 1-dimensional Lebesgue measure on ∂Ω.
ellH
Under Assumption 3.1,
elldifstaf
by the Lax-Milgram theorem, there exists a unique variational solution u ∈ H 1 (Ω)
elldifstab
of Problem (3.1)-(ciarlet
3.2). (For the study of elliptic problems and their discretization by finite element
methods, see e.g. Ciarlet, P.G. [1978] and references therein). This solution satisfies u = w + g̃, where
g̃ ∈ H 1 (Ω) is such that γ(g̃) = g, a.e. on ∂Ω, and w is the unique function of H01 (Ω) satisfying
Z  
∇w(x) · ∇ψ(x) + div(vw)(x)ψ(x) + bw(x)ψ(x) dx =
Z  Ω  (3.3) ellsolvar
−∇g̃(x) · ∇ψ(x) − div(vg̃)(x)ψ(x) − bg̃(x)ψ(x) + f (x)ψ(x) dx, ∀ψ ∈ H01 (Ω).
Ω

3.1.1 Structured meshes

smesh
If Ω is a rectangle (d = 2) or a parallelepiped (d = 3), it may then be meshed with rectangular or
parallelepipedic control volumes. In this case, the one-dimensional scheme may easily be generalized.

Rectangular meshes for the Laplace operator

let Ω = (0, 1)×(0, 1), and f ∈ C 2 (Ω, IR) (the three dimensional
Let us for instance consider the case delldifstaf
= 2,elldifstab
case is similar). Consider Problem (3.1)-(3.2) and assume here that b = 0, v = 0 and g = 0 (the general
case is considered later, on general unstructured meshes). The problem reduces to the pure diffusion
equation:

−∆u(x, y) = f (x, y), (x, y) ∈ Ω,

(3.4) lapl2d
u(x, y) = 0, (x, y) ∈ ∂Ω.
In this section, it is convenient to denote by (x, y) the current point of IR 2 (elsewhere, the notation x is
used for a point or a vector of IR d ).
Let T = (Ki,j )i=1,···,N1 ;j=1,···,N2 be an admissible
meshund
mesh of (0, 1) × (0, 1), that is, satisfying the following
assumptions (which generalize Definition 2.1)

hyprect Assumption 3.2 Let N1 ∈ IN? , N2 ∈ IN ? , h1 , . . . , hN1 > 0, k1 , . . . , kN2 > 0 such that
N1
X N2
X
hi = 1, ki = 1,
i=1 i=1

and let h0 = 0, hN1 +1 = 0, k0 = 0, kN2 +1 = 0. For i = 1, . . . , N1 , let x 21 = 0, xi+ 21 = xi− 21 + hi , (so that
xN1 + 21 = 1), and for j = 1, . . . , N2 , y 21 = 0, yj+ 21 = yj− 21 + kj , (so that yN2 + 21 = 1) and

Ki,j = [xi− 12 , xi+ 12 ] × [yj− 21 , yj+ 12 ].

Let (xi )i=0,N1 +1 , and (yj )j=0,N2 +1 , such that

xi− 12 < xi < xi+ 21 , for i = 1, . . . , N1 , x0 = 0, xN1 +1 = 1,

yj− 21 < yj < yj+ 21 , for j = 1, . . . , N2 , y0 = 0, yN2 +1 = 1,

and let xi,j = (xi , yj ), for i = 1, . . . , N1 ,, j = 1, . . . , N2 ; set

hi − = xi − xi− 21 , hi + = xi+ 21 − xi , for i = 1, . . . , N1 , hi+ 21 = xi+1 − xi , for i = 0, . . . , N1 ,

 34

kj − = yj − yj− 21 , kj + = yj+ 21 − yj , for j = 1, . . . , N2 , kj+ 21 = yj+1 − yj , for j = 0, . . . , N2 .

Let h = max{(hi , i = 1, · · · , N1 ), (kj , j = 1, · · · , N2 )}.
lapl2d
As in the 1D case, the finite volume scheme is found by integrating the first equation of ( 3.4) over each
control volume Ki,j , which yields
 Z yj+ 1 Z y 1
i+

 −
2
u (x 1 , y)dy +
2
ux (xi− 12 , y)dy

 x i+ 2
yj− 1 yi− 1
Z x 21 Z x 21 Z


i+
2
1 )dx −
i+
2

 + u y (x, y j− 2 u y (x, y 1 )dx =
j+ 2 f (x, y)dx dy.
xi− 1 xi− 1 Kij
2 2

The fluxes are then approximated by differential quotients with respect to the discrete unknowns (u i,j , i =
1, · · · , N1 , j = 1, · · · , N2 ) in a similar manner to the 1D case; hence the numerical scheme writes

Fi+ 12 ,j − Fi− 12 ,j + Fi,j+ 21 − Fi,j− 12 = hi,j fi,j , ∀ (i, j) ∈ {1, . . . , N1 } × {1, . . . , N2 }, (3.5) vfrect

where hi,j = hi × kj , fi,j is the mean value of f over Ki,j , and

kj
Fi+ 12 ,j = − (ui+1,j − ui,j ), for i = 0, · · · , N1 , j = 1, · · · , N2 ,
hi+ 12
hi (3.6) fluxrect
Fi,j+ 12 =− (ui,j+1 − ui,j ), for i = 1, · · · , N1 , j = 0, · · · , N2 ,
kj+ 21

u0,j = uN1 +1,j = ui,0 = ui,N2 +1 = 0, for i = 1, . . . , N1 , j = 1, . . . , N2 . (3.7) clrect

vfrectclrect
The numerical scheme (3.5)-(3.7) is therefore clearly conservative and the numerical approximations of
the fluxes can easily be shown to be consistent.

Proposition 3.1 (Error

lapl2d
(0, 1) × (0, 1) and f ∈ L2 (Ω). Let u be the unique varia-
estimate) Let Ω =hyprect
tional solution to (3.4). Under Assumptions 3.2, let ζ > 0 be such that hi ≥ ζh for i = 1, . .vfrect . , N1clrect
and
kj ≥ ζh for j = 1, . . . , N2 . Then, there exists a unique solution (ui,j )i=1,···,N1 ,j=1,···,N2 to (3.5)-(3.7).
Moreover, there exists C > 0 only depending on u, Ω and ζ such that
X (ei+1,j − ei,j )2 X (ei,j+1 − ei,j )2
kj + hi ≤ Ch2 (3.8) hunzrect
i,j
h i+ 1
2 i,j
k j+ 1
2

and
X
(ei,j )2 hi kj ≤ Ch2 , (3.9) lderect
i,j

where ei,j = u(xi,j ) − ui,j , for i = 1, · · · , N1 , j = 1, · · · , N2 .

above proposition, since f ∈ L2 (Ω) and Ω is convex, it is well known that the variational solution
In thelapl2d
2
u to (3.4) belongs
testh2
to Htesth2
(Ω). We do not give here the proof of this proposition since it is in fact included
lazarovsiam
in Theorem 3.4 page 56 (see also Lazarov, Mishev and Vassilevski [1996] where the case u ∈ H s ,
s ≥ 32 is also studied).
hunzrect lderect
In the case u ∈ C 2 (Ω), the estimates (3.8) and (3.9) can be shown with the same technique as inellesterr
thesejmf
the 1D
2
case (see
ellesterr
e.g. Fiard [1994]). If u ∈ C then the
hunzrect lderect
above estimates are a consequence of Theorem 3.3 page
52; in this case, the value C in (3.8) and (3.9) independent of ζ, and therefore the assumption hi ≥ ζh
for i = 1, . . . , N1 and kj ≥ ζh for j = 1, . . . , N2 is no longer needed.
hunzrect lderect
Relation (3.8) can be seen as an estimate of a “discrete H01 norm” of the error, while relation (3.9) gives
an estimate of the L2 norm of the error.
 35

vfrectclrect
h2conv2d Remarkangot1
3.1 Some slight modifications of the scheme (3.5)-(3.7) are possible, as in the first item of
angot1
Remark 2.2 page 14. It is also possible to obtain,
lderect
(or L∞ ) h2conv
sometimes, an “h2 ” estimate on the L2 h2conv norm
4 2
of the error (that is “h ” instead of “h ” in (3.9)),h2conv
exactly as in the 1D case, see Remark 2.5 page 18. In
the case equivalent to the second case of Remark 2.5, the point xi,j is not necessarily the center of Ki,j .
vfrectfluxrect
When the mesh is no longer rectangular, the scheme (3.5)-(3.6) is not easy to generalize if keepingFa1
to a 5
points scheme. In particular, the consistency of the fluxes or the conservativity can be lost, see Faille
[1992a], which yields a bad numerical behaviour of the scheme. One way to keep both properties is to
introduce a 9-points scheme.

Quadrangular meshes: a nine-point scheme

bounded polygonal subset of IR 2 , and f be a regular function from Ω to IR. We still
Let Ω be an open lapl2d
consider Problem 3.4, turning back to the usual notation x for the current point of IR 2 ,

−∆u(x) = f (x), x ∈ Ω,
(3.10) Lapl-2D
u(x) = 0, x ∈ ∂Ω.
Lapl-2D
Let T be a mesh defined over Ω; then, integrating the first equation of (3.10) over any cell K of the mesh
yields Z Z
− gradu · nK = f,
∂K K
where nK is the normal to the boundary ∂K, outward to K. Let uK denote the discrete unknown
associated to the control volume K ∈ T . In order to obtain a numerical scheme, if σ is a common edge
to K ∈ T and L ∈ T (denoted by K|L) or if σ is an edge of K ∈ T belonging to ∂Ω, the expression
gradu · nK must be approximated on σ by using the discrete unknowns. The study of the finite volume
scheme in dimension 1 and the above straightforward generalization to the rectangular case showed that
the fundamental properties of the method seem to be
1. conservativity: in the absence of any source term on K|L, the approximation of gradu · n K on K|L
which is used in the equation associated with cell K is equal to the approximation of −gradu · n L
which is used in the equation associated with cell L. This property is naturally obtained when
using a finite volume scheme.
2. consistency of the fluxes: taking for uK the value of u in a fixed point of K (for instance, the center
of gravity of K), where u is a regular function, the difference between gradu · nK and the chosen
approximation of gradu · nK is of an order less or equalcafe to that cafe
of the mesh
creme
size.creme
This need of
consistency will be discussed in more detail: see remarks 3.2 page 37 and 3.8 page 49
fluxrect
Several computer codes use the following “natural” extension of (3.6) for the approximation of gradu·nK
on K ∩ L:
uL − u K
gradu · nK = ,
dK|L
where dK|L is the distance between the center of the cells K and L. This choice, however simple, is far
from optimal, at least in the case of a general (non rectangular) mesh, because the fluxes thus obtained
are not consistent; this yields Fa1
important errors,
Fa2
especially in the case where the mesh cells are all oriented
in the same direction, see Faille [1992a], Faille [1992b]. This problem may be avoided by modifying
the approximation of gradu · nK so as to make it consistent. However, one must be careful, in doing so,
to maintain the conservativity of the scheme. To this purpose, a 9-points scheme was developped, which
is denoted by FV9.
Let us describe now how the flux gradu · nK is approximated by the FV9 scheme. Assume here, for
the sake of clarity, that the mesh T is structured; indeed, vf9fig
it consists in a set of quadrangular cells
{Ki,j , i = 1, . . . , N ; j = 1, . . . , M }. As shown in Figure 3.1, let Ci,j denote the center of gravity of the cell
 36

Ki,j , σi,j−1/2 , σi+1/2,j , σi,j+1/2 , σi−1/2,j the four edges to Ki,j and ηi,j−1/2 , ηi+1/2,j , ηi,j+1/2 , ηi−1/2,j
their respective orthogonal bisectors. Let ζi,j−1/2 , (resp. ζi+1/2,j , ζi,j+1/2 , ζi−1/2,j ) be the lines joining
points Ci,j and Ci,j−1 (resp. Ci,j and Ci+1,j , Ci,j and Ci,j+1 , Ci−1,j and Ci,j ).

ηi,j+1/2
q
Ci,j+1
Ci+1,j+1
 ζ 1
i+1/2,j+1

Ui,j+1/2
σi,j+1/2 * Ki,j+1
Ki,j
Ci−1,j
 YCi,j
ζi−1/2,j
Di,j+1/2

Figure 3.1: FV9 scheme vf9fig

vf9fig
Consider for instance the edge σi,j+1/2 which lies between the cells Ki,j and Ki,j+1 (see Figure 3.1). In
order to find an approximation of gradu · nK , for K = Ki,j , at the center of this edge, we shall first
derive an approximation of u at the two points Ui,j+1/2 and Di,j+1/2 which are located on the orthogonal
bisector ηi,j+1/2 of the edge σi,j+1/2 , on each side of the edge. Let φi,j+1/2 be the approximation of
−gradu · nK at the center of the edge σi,j+1/2 . A natural choice for φi,j+1/2 consists in taking

uU D
i,j+1/2 − ui,j+1/2
φi,j+1/2 = − , (3.11) flux
d(Ui,j+1/2 , Di,j+1/2 )
where uU D
i,j+1/2 and ui,j+1/2 are approximations of u at Ui,j+1/2 and Di,j+1/2 , and d(Ui,j+1/2 , Di,j+1/2 )
is the distance between points Ui,j+1/2 and Di,j+1/2 .
The points Ui,j+1/2 and Di,j+1/2 are chosen so that they are located on the lines ζ which join the centers
of the neighbouring cells. The points Ui,j+1/2 and Di,j+1/2 are therefore located at the intersection of
the orthogonal bisector ηi,j+1/2 with the adequate ζ lines, which are chosen according to the geometry
of the mesh. More precisely,

Ui,j+1/2 = ηi,j+1/2 ∩ ζi−1/2,j+1 if ηi,j+1/2 is to the left of Ci,j+1

= ηi,j+1/2 ∩ ζi+1/2,j+1 otherwise
Di,j+1/2 = ηi,j+1/2 ∩ ζi−1/2,j if ηi,j+1/2 is to the left of Ci,j
= ηi,j+1/2 ∩ ζi+1/2,j otherwise
In order to satisfy the property of consistency of the fluxes, a second order approximation of u atvf9fig
points
Ui,j+1/2 and Di,j+1/2 is required. In the case of the geometry which flux is described in Figure 3.1, the
following linear approximations of uU D
i,j+1/2 and ui,j+1/2 can be used in (3.11);
 37

d(Ci,j+1 , Ui,j+1/2 )
uU
i,j+1/2 = αui+1,j+1 + (1 − α)ui,j+1 where α =
d(Ci,j+1 , Ci+1,j+1 )
d(Ci,j , Di,j+1/2 )
uD
i,j+1/2 = βui−1,j + (1 − β)ui,j where β =
d(Ci−1,j , Ci,j )
The approximation of gradu · nK at the center of a “vertical” edge σi+1/2,j is performed in a similar way,
by introducing the points Ri+1/2,j intersection of the orthogonal bisector ηi+1/2,j and, according to the
geometry, of the line ζi,j−1/2 or ζi,j+1/2 , and Li+1/2,j intersection of ηFa1
i+1/2,j and ζi+1,j−1/2 or ζi+1,j+1/2 .
Note that the outmost grid cells require a particular treatment (see Faille [1992a]).
The scheme which is described above is stable under a geometrical condition on the family of meshes
which is considered. Since the fluxes are consistent and the scheme is conservative, wc
it also
wc
satisfies a
property
ell1D
of “weak consistency”, that
Lapl-2D
is, as in the one dimensional case (see remark 2.9 page 21 of Section
2.3), the exact solution of (3.10) satisfies the numerical scheme with an error which tends to 0 in L∞ (Ω)
for the weak-?Fa1topology. Under adequate restrictive assumptions, the convergence of the scheme can be
deduced, see Faille [1992a].
Numerical tests were performed for the Laplace operator Fa1
and for operators of the type −div( Λ grad.),
where Λ is a variable and discontinuous matrix (see Faille
ell1D
[1992a]); the discontinuities of Λ are treated
in a similar way as in the 1D case (see Section 2.3). Comparisons with solutions which were obtained
by the bilinear finite element method, and with known analytical solutions, were performed. The results
given by the VF9 scheme and by the finite element scheme were very similar.
The two drawbacks of this method are the fact that it is a 9-points scheme, and therefore computationally
expensive, and that it yields a nonsymmetric matrix even if the original continuous operator is symmetric.
Also, its generalization to three dimensions is somewhat complex.

cafe Remark 3.2 The proof of convergence of this scheme is hindered by the lack of consistency for the
ellcvf4
discrete adjoint operator (see Section 3.1.4). An error estimate is also difficult to obtain because the
numerical flux at an interface K|L cannot be written under the form τK|L (u − uL ) with τK|Lcoudiere
Fa1 K
> 0. Note,
however, that under some geometrical assumptions on the mesh, see Faille [1992a] and Coudière,
Vila and Villedieu [1999], error estimates may be obtained.

3.1.2 General meshes and schemes

vfquatre
Let us now turn to the discretization of convection-diffusion problems on general structured or non
structured grids, consisting of any polygonal (recall that we shall call “polygonal” any polygonal domain
of IR 2 or polyhedral domain or IR 3 ) control volumes (satisfying adequate geometrical conditions which
are stated in the sequel) and not necessarily ordered in a Cartesian grid. The advantage of finite volume
schemes using non structured meshes is clear for convection-diffusion equations. On one hand, the stability
and convergence properties of the finite volume scheme (with an upstream choice for the
meshdirichlet
convective
meshdirichlet
flux)
meshneumanmeshneuman
ensure a robust scheme for any admissible mesh as defined in Definitions 3.1 page 37 and 3.5 page 63
below, without any need for refinement in the areas of a large convection flux. On the other hand, the
use of a non structured mesh allows the computation of a solution for any shape of the physical domain.

We saw in the previous section that a consistent discretization of the normal flux −∇u·n over the interface
of two control volumes K and L may be performed with a differential quotient involving values of the
unknown located on the orthogonal line to the interface between K and L, on either side of this interface.
This remark suggests the following definition of admissible finite volume meshes for the discretization of
diffusion problems. We shall only consider here, for the sake of simplicity, the case of polygonal domains.
The case of domains with a regular boundary does not introduce any supplementary difficulty other than
complex notations. The definition of admissible meshes and notations introduced in this definition are
CCfig1
illustrated in Figure 3.2

meshdirichlet Definition 3.1 (Admissible meshes) Let Ω be an open bounded polygonal subset of IR d , d = 2, or 3.
An admissible finite volume mesh of Ω, denoted by T , is given by a family of “control volumes”, which
 38

are open polygonal convex subsets of Ω , a family of subsets of Ω contained in hyperplanes of IR d , denoted
by E (these are the edges (two-dimensional) or sides (three-dimensional) of the control volumes), with
strictly positive (d − 1)-dimensional measure, and a family of points of Ω denoted by P satisfying the
following properties (in fact, we shall denote, somewhat incorrectly, by T the family of control volumes):

(i) The closure of the union of all the control volumes is Ω;

(ii) For any K ∈ T , there exists a subset EK of E such that ∂K = K \ K = ∪σ∈EK σ. Furthermore,
E = ∪K∈T EK .
(iii) For any (K, L) ∈ T 2 with K 6= L, either the (d − 1)-dimensional Lebesgue measure of K ∩ L is 0
or K ∩ L = σ for some σ ∈ E, which will then be denoted by K|L.
(iv) The family P = (xK )K∈T is such that xK ∈ K (for all K ∈ T ) and, if σ = K|L, it is assumed that
xK 6= xL , and that the straight line DK,L going through xK and xL is orthogonal to K|L.
(v) For any σ ∈ E such that σ ⊂ ∂Ω, let K be the control volume such that σ ∈ EK . If xK ∈ / σ, let
DK,σ be the straight line going through xK and orthogonal to σ, then the condition DK,σ ∩ σ 6= ∅
is assumed; let yσ = DK,σ ∩ σ.

In the sequel, the following notations are used.

The mesh size is defined by: size(T ) = sup{diam(K), K ∈ T }.
For any K ∈ T and σ ∈ E, m(K) is the d-dimensional Lebesgue measure of K (it is the area of K in the
two-dimensional case and the volume in the three-dimensional case) and m(σ) the (d − 1)-dimensional
measure of σ.
The set of interior (resp. boundary) edges is denoted by Eint (resp. Eext ), that is Eint = {σ ∈ E; σ 6⊂ ∂Ω}
(resp. Eext = {σ ∈ E; σ ⊂ ∂Ω}).
The set of neighbours of K is denoted by N (K), that is N (K) = {L ∈ T ; ∃σ ∈ EK , σ = K ∩ L}.
If σ = K|L, we denote by dσ or dK|L the Euclidean distance between xK and xL (which is positive) and
by dK,σ the distance from xK to σ.
If σ ∈ EK ∩ Eext , let dσ denote the Euclidean distance between xK and yσ (then, dσ = dK,σ ).
For any σ ∈ E; the “transmissibility” through σ is defined by τσ = m(σ)/dσ if dσ 6= 0.
In some results and proofs given below, there are summations over σ ∈ E0 , with E0 = {σ ∈ E; dσ 6= 0}.
For simplicity, (in these results and proofs) E = E0 is assumed.

m(σ) ∂Ω
σ
xK yσ xL dσ
xK yσ
dσ
K
K L

Figure 3.2: Admissible meshes CCfig1

Remark 3.3 (i) The definition of yσ for σ ∈ Eext requires that yσ ∈ σ. However, In many cases, this
may be relaxed. The condition xK ∈ K may also be relaxed as described, for instance, in
condition tm
Example 3.1 below.
 39

(ii) The condition xK 6= xL if σ = K|L, is in fact quite easy to satisfy: two neighbouring control volumes
K, L which do not satisfy it just have to be collapsed into a new control volume M with x M = xK = xL ,
and the edge K|L removed from the set of edges. The new mesh thus obtained is admissible.

tm Example 3.1 (Triangular meshes) Let Ω be an open bounded polygonal subset of IR 2 . Let T be a
family of open triangular disjoint subsets of Ω such that two triangles having a common edge have also two
common vertices. Assume that all angles of the triangles are less than π/2. This last condition is sufficient
for the orthogonal bisectors to intersect inside each triangle, thus naturally defining the points x K ∈ K.
One obtains an admissible mesh. In the case of an elliptic operator, the finite volume scheme defined on
such
VF4
a grid using differential quotients for the approximation of the normal flux yields a 4-point scheme
Herbin [1995]. This scheme does not lead to a finite difference scheme consistent with the continuous
diffusion operator (using a Taylor expansion). The consistency is only verified for the approximation of
the fluxes, but this, together with the conservativity of the scheme yields the convergence of the scheme,
as it is proved below.
Note that the condition that all angles of the triangles are less than π/2 (which yields x K ∈ K) may
be relaxed (at least for the triangles the closure of which are in Ω) to the so called “strict Delaunay
condition” which is that the closure of the circumscribed circle to each triangle of the mesh does not
contain any other triangle of the mesh. For such a mesh, the point xK (which is the intersection of the
easyvf1easyvf3
orthogonal
easyvf2
bisectors of the edges of K) is not always in K, but the scheme ( 3.17)-( 3.19) is convenient since
(3.18) yields a consistent approximation of the diffusion fluxes and since the transmissibilities (denoted
by τK|L ) are positive.

evoronoi Example 3.2 (Voronoı̈ meshes) Let Ω be an open bounded polygonal subset of IR d . An admissible
finite volume mesh can be built by using the so called “Voronoı̈” technique. Let P be a family of points
of Ω. For example, this family may be chosen as P = {(k1 h, . . . , kd h), k1 , . . . kd ∈ ZZ } ∩ Ω, for a given
h > 0. The control volumes of the Voronoı̈ mesh are defined with respect to each point x of P by

Kx = {y ∈ Ω, |x − y| < |z − y|, ∀z ∈ P, z 6= x}.

Recall that |x − y| denotes the euclidean distance between x and y.
meshdirichlet
Voronoı̈ meshes are admissible in the
meshdirichlet
sense of Definition 3.1 if the assumption “on the boundary”, namely
part (v) of Definition 3.1, is satisfied. Indeed, this is true, in particular, if the number of points x∈P
meshdirichlet
which are located on ∂Ω is “large enough”. Otherwise, the assumption (v) of Definition 3.1 may be
replaced by the weaker assumption “d(yσ , σ) ≤ size(T ) for any σ ∈ Eext ” which is much easier to satisfy.
Note alsoellschema
that ellschemad
a slight modification
ellschemad
of the treatment of the boundary conditions in the finite volume
scheme ( 3.20)-( 3.23) page 42 allows
ellcvgce ellcvgceellesterr ellesterr
us to obtain convergence and error estimates results (as in theorems
3.1 page 46 and 3.3 page 52) for all Voronoı̈ meshes. Thisangot1
angot1
modification is the obvious generalization of
the scheme described in the first item of Remarkellschema
2.2 page 14 for the 1D case. It consists in replacing, for
K ∈ T such that EK ∩ Eext 6= ∅, the equation (3.20), associated to this control volume, by the equation
uK = g(zK ), where zK is some point on ∂Ω ∩ ∂K. In fact, Voronoı̈ meshes often satisfy the following
property:
EK ∩ Eext 6= ∅ ⇒ xK ∈ ∂Ω
meshdirichlet ellschema
ellschemad
and the mesh is therefore admissible in the sense of Definition 3.1 (then, the scheme (3.20)-(3.23) page
ellschemad
42 yields uK = g(xK ) if K ∈ T is such that EK ∩ Eext 6= ∅).
An advantage of the Voronoı̈ method is that it easily leads to meshes on non polygonal domains Ω.

Let us now introduce the space of piecewise constant functions associated to an admissible mesh and
some “discrete H01 ” norm for this space. This discrete norm will be used to obtain stability properties
which are given by some estimates on the approximate solution of a finite volume scheme.

Xmesh Definition 3.2 Let Ω be an open bounded polygonal subset of IR d , d = 2 or 3, and T an admissible
mesh. Define X(T ) as the set of functions from Ω to IR which are constant over each control volume of
the mesh.
 40

defh10d Definition 3.3 (Discrete H01 norm) Let Ω be an open boundedmeshdirichlet

polygonal subset of IR d , d = 2 or 3,
meshdirichlet
and T an admissible finite volume mesh in the sense of Definition 3.1 page 37.
For u ∈ X(T ), define the discrete H01 norm by
X  21
kuk1,T = τσ (Dσ u)2 , (3.12) normh1d
σ∈E

where τσ = m(σ)/dσ and

Dσ u = |uK − uL | if σ ∈ Eint , σ = K|L,
Dσ u = |uK | if σ ∈ Eext ∩ EK ,
where uK denotes themeshdirichlet
value taken by u on the control volume K and the sets E, Eint , Eext and EK are
meshdirichlet
defined in Definition 3.1 page 37.

The discrete H01 norm is used in the following sections to prove the congergence of finite volume schemes
and, under some regularity conditions, to give error estimates. It is related to the H 01 norm, see the
ellcvgce
convergence of the norms in Theorem 3.1. One of the tools used below is the following “discrete Poincaré
temam
inequality” which may also be found in Temam [1977]:

thinpoin Lemma 3.1 (Discrete Poincaré inequality) Let Ω be an open bounded polygonal subset of IR d , d = 2
meshdirichlet
or 3, T an admissible finite volume mesh in the sense of Definition 3.1 and u ∈ X(T ) (see Definition
Xmesh
3.2), then
kukL2 (Ω) ≤ diam(Ω)kuk1,T , (3.13) ellinpoin
defh10d defh10d
where k · k1,T is the discrete H01 norm defined in Definition 3.3 page 40.

dirich-part Remark 3.4 (Dirichlet condition on part of the boundary) This lemma gives a discrete Poincaré
inequality for Dirichlet boundary conditions on the boundary ∂Ω. In the case of a Dirichlet condition on
part of the boundary only, it is still possible to prove a Discretethinpoin
boundarythinpoin
condition provided that the
polygonal bounded open set Ω is also connex, thanks to Lemma ( 3.1) page 40 proven in the sequel.
thinpoin
Proof of Lemma 3.1
For σ ∈ E, define χσ from IR d × IR d to {0, 1} by χσ (x, y) = 1 if σ ∩ [x, y] 6= ∅ and χσ (x, y) = 0 otherwise.

Let u ∈ X(T ). Let d be a given unit vector. For all x ∈ Ω, let Dx be the semi-line defined by its origin, x,
and the vector d. Let y(x) such that y(x) ∈ Dx ∩∂Ω and [x, y(x)] ⊂ Ω, where [x, y(x)] = {tx+(1−t)y(x),
t ∈ [0, 1]} (i.e. y(x) is the first point where Dx meets ∂Ω).

Let K ∈ T . For a.e. x ∈ K, one has

X
|uK | ≤ Dσ u χσ (x, y(x)),
σ∈E
defh10d defh10d
where the notations Dσ u and uK are defined in Definition 3.3 page 40. We write the above inequality
for a.e x ∈ Ω and not for all x ∈ Ω in order to account for the cases where an edge or a vertex of the
mesh is included in the semi-line [x, y(x)]; in both cases one may not write the above inequality, but there
are only a finite number of edges and vertices, and since d is fixed, the above inequality may be written
almost everywhere.
Let cσ = |d · nσ | (recall that ξ · η denotes the usual scalar product of ξ and η in IR d ). By the Cauchy-
Schwarz inequality, the above inequality yields:
X (Dσ u)2 X
|uK |2 ≤ χσ (x, y(x)) dσ cσ χσ (x, y(x)), for a.e. x ∈ K. (3.14) bon
d σ cσ
σ∈E σ∈E

Let us show that, for a.e. x ∈ Ω,

 41

X
dσ cσ χσ (x, y(x)) ≤ diam(Ω). (3.15) prelim
σ∈E

Let x ∈ K, K ∈ T , such that σ ∩ [x, y(x)] for all σ ∈ E, and [x, y(x)] does
contains at most one point, prelim
prelim
not contain any vertex of T (proving (3.15) for such points x leads to (3.15) a.e. on Ω, since d is fixed).
There exists σx ∈ Eext such that y(x) ∈ σx . Then, using the fact that the control volumes are convex,
one has:
X
χσ (x, y(x))dσ cσ = |(xK − xσx ) · d|.
σ∈E
meshdirichlet prelim
Since xK and xσx ∈ Ω (see Definition 3.1), this gives (3.15).
bon prelim
Let us integrate (3.14) over Ω; (3.15) gives
XZ X (Dσ u)2 Z
|uK |2 dx ≤ diam(Ω) χσ (x, y(x))dx.
d σ cσ Ω
K∈T K σ∈E
R
Since Ω χσ (x, y(x))dx ≤ diam(Ω)m(σ)cσ , this last inequality yields
XZ X m(σ)
|uK |2 dx ≤ (diam(Ω))2 |Dσ u|2 dx.
K dσ
K∈T σ∈E

Hence the result.

elldifstaf
elldifstab elldifstab
Let T be an admissible mesh. Let us now define a finite volume scheme to discretize (3.1), (3.2) page 32.
Let
Z
1
fK = f (x)dx, ∀K ∈ T . (3.16) frhs
m(K) K
Let (uK )K∈T denote the discrete unknowns. In order to describe the scheme
ell1D
in the most general way, one
introduces some auxiliary unknowns (as in the 1D case, see Section 2.3), namely the fluxes FK,σ , for all
K ∈ T and σ ∈ EK , and some (expected) approximation of u in σ, denoted by uσ , for all R σ ∈ E. For K ∈ T
and σ ∈ EK , let nK,σ denote the normal unit vector to σ outward to K and vK,σ = σ v(x) · nK,σ dγ(x).
Note that dγ is the integration symbol for the (d − 1)-dimensional Lebesgue measure on the considered ell1D
hyperplane.
ell1D
In order to discretize the convection term div(v(x)u(x)) in a stable way (see Section 2.3
page 21), let us define the upstream choice uσ,+ of u on an edge σ with respect to v in the following way.
If σ = K|L, then uσ,+ = uK if vK,σ ≥ 0, and uσ,+ = uL otherwise; if σ ⊂ K ∩ ∂Ω, then uσ,+ = uK if
vK,σ ≥ 0 and uσ,+ = g(yσ ) otherwise.
Let us first assume that the points xK are located in the interior of each control volume, and are therefore
not located on the edges, hence dK,σ > 0 for any σ ∈ EK , where dK,σ is the distance from xK to σ. A
finite volume scheme can be defined by the following set of equations:
X X
FK,σ + vK,σ uσ,+ + bm(K)uK = m(K)fK , ∀K ∈ T , (3.17) easyvf1
σ∈EK σ∈EK

FK,σ = −τK|L (uL − uK ), ∀σ ∈ Eint , if σ = K|L, (3.18) easyvf2

FK,σ = −τσ (g(yσ ) − uK ), ∀σ ∈ Eext such that σ ∈ EK . (3.19) easyvf3

In the general case, the center of the cell may be located on an edge. This is the case for instance when
constructing Voronoı̈ meshes with some of the original points located on the boundary ∂Ω. In this case,
the following formulation of the finite volume scheme is valid, and is equivalent to the above scheme if
no cell center is located on an edge:
 42

X X
FK,σ + vK,σ uσ,+ + bm(K)uK = m(K)fK , ∀K ∈ T , (3.20) ellschema
σ∈EK σ∈EK

FK,σ = −FL,σ , ∀σ ∈ Eint , if σ = K|L, (3.21) ellschemab

FK,σ dK,σ = −m(σ)(uσ − uK ), ∀σ ∈ EK , ∀K ∈ T , (3.22) ellschemac

uσ = g(yσ ), ∀σ ∈ Eext . (3.23) ellschemad

ellschema
ellschemad
Note that (3.20)-(3.23) always lead, after an easy elimination of the auxiliary unknowns, to a linear
system of N equations with N unknowns, namely the (uK )K∈T , with N = card(T ).

Remark 3.5
ellschemac
1. Note that one may have, for some σ ∈ EK , xK ∈ σ, and therefore, thanks to (3.22), uσ = uK .
ellschemad
2. The choice uσ = g(yσ ) in (3.23) needs some discussion. Indeed, this choice is possible since g is
assumed to belong to C(∂Ω, IR) and then is everywhere defined on ∂Ω. In the case where the
elldifstaf
elldifstab elldifstab
solution to (3.1), (3.2) page 32 belongs to H 2 (Ω) (which yields g ∈ C(∂Ω, IR)), it is clearly the
“good choice” since it yields the consistency of fluxes (even though an error estimate also holds
with other choices for uσ , the choice given below is, for instance, possible). If g ∈ H 1/2 (and not
continuous), the value g(yσ ) is not necessarily defined. Then, another choice for uσ is possible, for
instance,
Z
1
uσ = g(x)dγ(x).
m(σ) σ
ellcvgcediraf
With this latter choice for uσ , a convergence result also holds, see Theorem 3.2.
meshdirichlet
For the sake of simplicity, it is assumed in Definition 3.1 that xK 6= xL , for all K, L ∈ T . This condition
may be relaxed; it simply allows an easy expression of the numerical flux FK,σ = −τK|L (uL − uK ) if
σ = K|L.

3.1.3 Existence and estimates

Let us first prove the existence of the approximate solution and an estimate on this solution. This estimate
ellinpoin
ensures the stability of the scheme and will be obtained by using the discrete rellichd
Poincaré rellichd
inequality ( 3.13)
and will yield convergence thanks to a compactness theorem given in Section 3.6 page 92.
ellH
ellexistu Lemma 3.2 (Existence and estimate) Under Assumptions 3.1, let T be an admissible
meshdirichlet
meshdirichlet
mesh
ellschema
in the
ellschemad
sense of Definition 3.1 page 37; there exists a unique solution (uK )K∈T to Xmesh
equations (3.20)-(3.23).
Furthermore, assuming g = 0 and defining uT ∈ X(T ) (see Definition 3.2) by uT (x) = uK for a.e.
x ∈ K, and for any K ∈ T , the following estimate holds:

kuT k1,T ≤ diam(Ω)kf kL2 (Ω) , (3.24) ellestimx

defh10d
where k · k1,T is the discrete H01 norm defined in Definition 3.3.
ellexistu
Proof of Lemma 3.2
ellschema
ellschemad
Equations (3.20)-(3.23) lead, after an easy elimination of the auxiliary unknowns, to a linear system of
N equations with N unknowns, namely the (uK )K∈T , with N = card(T ).

Step 1 (existence and uniqueness)

 43

Assume that (uK )K∈T satisfies this linear system with g(yσ ) ellschemab
ellschema
σ ∈ Eext , and fK = 0 for all
= 0 for anyellschemac
K ∈ T . Let us multiply (3.20) by uK and sum over K; from (3.21) and (3.22) one deduces
X X X X X
b m(K)u2K + FK,σ uK + vK,σ uσ,+ uK = 0, (3.25) ell1e
K∈T K∈T σ∈EK K∈T σ∈EK

which gives, reordering the summation over the set of edges

X X X  
b m(K)u2K + τσ (Dσ u)2 + vσ uσ,+ − uσ,− uσ,+ = 0, (3.26) ell7e
K∈T σ∈E σ∈E

where
|Dσ u| =
R |uK − uL |, if σ = K|L and |Dσ u| = |uK |, if σ ∈ EK ∩ Eext ;
vσ = | σ v(x) · ndγ(x)|, n being a unit normal vector to σ;
uσ,− is the downstream value to σ with respect to v, i.e. if σ = K|L, then uσ,− = uK if vK,σ ≤ 0, and
uσ,− = uL otherwise; if σ ∈ EK ∩ Eext , then uσ,− = uK if vK,σ ≤ 0 and uσ,− = uσ if vK,σ > 0.
Note that uσ = 0 if σ ∈ Eext .

Now, remark that

X 1X  
vσ uσ,+ (uσ,+ − uσ,− ) = vσ (uσ,+ − uσ,− )2 + (u2σ,+ − u2σ,− ) (3.27) vfqconv1
2
σ∈E σ∈E

and, thanks to the assumption divv ≥ 0,

X X Z  Z
vσ (u2σ,+ − u2σ,− ) = v(x) · nK dγ(x) u2K = (divv(x))u2T (x)dx ≥ 0. (3.28) vfqconv11
σ∈E K∈T ∂K Ω

Hence,
X X
bkuT k2L2 (Ω) + kuT k21,T = b m(K)u2K + τσ (Dσ u)2 ≤ 0, (3.29) ell7f
K∈T σ∈E
ell7f
One deduces, from (3.29), that uK = 0 for all K ∈ T .
This proves
ellschema
the existence and the uniqueness of the solution (uK )K∈T , of the linear system given by
ellschemad
(3.20)-(3.23), for any {g(yσ ), σ ∈ Eext } and {fK , K ∈ T }.

Step 2 (estimate) ellschema ellschemab

ellschemac
vfqconv1 vfqconv11
Assume g = 0. Multiply (3.20) by uK , sum over K; then, thanks to (3.21), (3.22), (3.27) and (3.28) one
has
X
bkuT k2L2 (Ω) + kuT k21,T ≤ m(K)fK uK .
K∈T

By the Cauchy-Schwarz inequality, this inequality yields

X 1
X
2 21
kuT k21,T ≤ ( m(K)u2K ) 2 ( m(K)fK ) ≤ kf kL2(Ω) kuT kL2 (Ω) .
K∈T K∈T
ellinpoin
Thanks to the discrete Poincaré inequality (3.13), this yields kuT k1,T ≤ kf kL2 (Ω) diam(Ω), which con-
cludes the proof of the lemma.
ellschema
ellschemad
Let us now state a discrete maximum principle which is satisfied by the scheme (3.20)-(3.23); this is an
interesting stability property, even though it will not be used in the proofs of the convergence and error
estimate.
 44

ellH ellH
maxppe Proposition
meshdirichlet
3.2 Under Assumption 3.1frhs
meshdirichlet
page 32, let T be an admissible mesh in the sense of Definition
3.1 page 37, let (fK )K∈T be defined by (
ellschema
3.16).
ellschemad
If fK ≥ 0 for all K ∈ T , and g(yσ ) ≥ 0, for all σ ∈ Eext ,
then the solution (uK )K∈T of (3.20)-(3.23) satisfies uK ≥ 0 for all K ∈ T .
maxppe
Proof of Proposition 3.2
Assume that fK ≥ 0 for all K ∈ T and g(yσ ) ≥ 0 for all σ ∈ Eext . Let a = min{uK , K ∈ T }. Let K0 be
a control volume such that uK0 = a. Assume first that K is an “interior” control volume, in the sense
ellschema 0
that EK ⊂ Eint , and that uK0 ≤ 0. Then, from (3.20),
X X
FK0 ,σ + vK0 ,σ uσ,+ ≥ 0; (3.30)
σ∈EK0 σ∈EK0

since for any neighbour L of K0 one has uL ≥ uK0 , then, noting that divv ≥ 0, one must have uL = uK0
for any neighbour L of K. Hence, setting B = {K ∈ T , uK = a}, there exists K ∈ B such that EK 6⊂ Eint ,
that is K is a control volume “neighbouring the boundary”.
Assume then that K0 is a control volume neighbouring the boundary and that uK0 = a < 0. Then, for
ellschemac ellschemad
an edge σ ∈ Eext ∩ EK , relationsmaxppe
(3.22) and (3.23) yield g(yσ ) < 0, which is in contradiction with the
assumption. Hence Proposition 3.2 is proved.

Remark 3.6 The maximum principle

ellschema
immediately yields the existence andellexistu
ellschemad
uniqueness of the solution
of the numerical scheme (3.20)-(3.23), which was proved directly in Lemma 3.2.

3.1.4 Convergence
ellcvf4
Let us now show the convergence of approximate solutions obtained by the above finite volume scheme
ellexistu kolmh10
when the size of the mesh tends to 0. One uses Lemma 3.2 together with the compactness theorem 3.10
kolmh10
given at the end of this chapter to prove the convergence result. In order to use Theorem 3.10, one needs
the following lemma.

h1dtot Lemma 3.3 Let Ω be an

meshdirichlet
open bounded set of IR d , d = 2 or
meshdirichlet
3. Let T be an admissible mesh in the sense
Xmesh
of Definition 3.1 page 37 and u ∈ X(T ) (see Definition 3.2). One defines ũ by ũ = u a.e. on Ω, and
ũ = 0 a.e. on IR d \ Ω. Then there exists C > 0, only depending on Ω, such that

kũ(· + η) − ũk2L2 (IR d ) ≤ kuk21,T |η|(|η| + C size(T )), ∀η ∈ IR d . (3.31) itransh10

h1dtot
Proof of Lemma 3.3
For σ ∈ E, define χσ from IR d × IR d to {0, 1} by χσ (x, y) = 1 if [x, y] ∩ σ 6= ∅ and χσ (x, y) = 0 if
[x, y] ∩ σ = ∅.
Let η ∈ IR d , η 6= 0. One has
X
|ũ(x + η) − ũ(x)| ≤ χσ (x, x + η)|Dσ u|, for a.e. x ∈ Ω
σ∈E
defh10d defh10d
(see Definition 3.3 page 40 for the definition of Dσ u).
This gives, using the Cauchy-Schwarz inequality,
X |Dσ u|2 X
|ũ(x + η) − ũ(x)|2 ≤ χσ (x, x + η) χσ (x, x + η)dσ cσ , for a.e. x ∈ IR d , (3.32) astucecos
d σ cσ
σ∈E σ∈E
η
where cσ = |nσ · |η| |, and nσ denotes a unit normal vector to σ.

Let us now prove that there exists C > 0, only depending on Ω, such that
 45

X
χσ (x, x + η)dσ cσ ≤ |η| + C size(T ), (3.33) aie
σ∈E

for a.e. x ∈ IR d .

Let x ∈ IR d such that σ ∩ [x, x + η]aie

contains at most one point,aie for all σ ∈ E, and [x, x + η] does not
contain any vertex of T (proving (3.33) for such points x gives (3.33) for a.e. x ∈ IR d , since η is fixed).
Since Ω is not assumed to be convex, it may happen that the line segment [x, x + η] is not included in Ω.
In order to deal with this, let y, z ∈ [x, x + η] such that y 6= z and [y, z] ⊂ Ω; there exist K, L ∈ T such
that y ∈ K and z ∈ L. Hence,
X η
χσ (y, z)dσ cσ = |(y1 − z1 ) · |,
|η|
σ∈E

where y1 = xK or yσ with σ ∈ Eext ∩ EK and z1 = xL or yσ̃ with σ̃ ∈ Eext ∩ EL , depending on the position
of y and z in K or L respectively.
Since y1 = y + y2 , with |y2 | ≤ size(T ), and z1 = z + z2 , with |z2 | ≤ size(T ), one has
η
|(y1 − z1 ) · | ≤ |y − z| + |y2 | + |z2 | ≤ |y − z| + 2 size(T )
|η|
and
X
χσ (y, z)dσ cσ ≤ |y − z| + 2 size(T ). (3.34) aier
σ∈E
aie
Note that this yields (3.33) with C = 2 if [x, x + η] ⊂ Ω.
Since Ω has a finite number of sides, the line segment [x, x + η] intersects ∂Ω a finite number of times;
hence there exist t1 , . . . , tn such that 0 ≤ t1 < t2 < . . . < tn ≤ 1, n ≤ N , where N only depends on Ω
(indeed, it is possible to take N = 2 if Ω is convex and N equal to the number of sides of Ω for a general
Ω) and such that
X X X
χσ (x, x + η)dσ cσ = χσ (xi , xi+1 )dσ cσ ,
σ∈E i=1,n−1 σ∈E
oddi

for i = 1, . . . , n, xi ∈ ∂Ω if ti ∈
with xi = x + ti η,aier / {0, 1} and [xi , xi+1 ] ⊂ Ω if i isaieodd.
Then, thanks to (3.34) with y = xi and z = xi+1 , for aie i = 1, . . . , n − 1, one hasitransh10
(3.33) with C = 2(N − 1)
(in particular, if Ω is convex, C = 2 is convenient for (3.33) and therefore for (3.31) as we shall see below).
h1dtot
In order to conclude the proof of Lemma 3.3, remark that, for all σ ∈ E,
Z
χσ (x, x + η)dx ≤ m(σ)cσ |η|.
IR d
astucecos d aie
Therefore, integrating (3.32) over IR yields, with (3.33),
X m(σ)

kũ(· + η) − ũk2L2 (IR d ) ≤ |Dσ u|2 |η|(|η| + C size(T )).
dσ
σ∈E

We are now able to state the convergence theorem. We shall first prove the convergence result in the case
of homogeneous Dirichlet boundary conditions, i.e.
ellcvgcediraf
g = 0; the nonhomogenous case is then considered in
ellcvgcediraf cvnl
the two-dimensional case (see Theorem 3.2 page 51), following Eymard, Gallouët and Herbin [1999].
 46

ellcvgce Theorem
ellH
3.1 (Convergence, homogeneous Dirichlet boundary conditions)
ellH
Under
meshdirichlet
Assumption
meshdirichlet
3.1 page 32 with g = 0, let T be an admissible mesh (in the
ellschema
sense of
ellschemad
Definition 3.1 page 37). Let (uK )K∈T
the system given by equations (3.20)-(3.23) (existence and uniqueness of (uK )K∈T are
be the solution ofellexistu
given in Lemma 3.2). Define uT ∈ X(T ) by uT (x) = uK for a.e. x ∈ K, andelldifstaf
for any K ∈ T . Then uT
elldifstab
converges in L2 (Ω) to the unique variational solution u ∈ H01 (Ω) of Problem (3.1), (3.2) as size(T ) → 0.
Furthermore kuT k1,T converges to kukH01 (Ω) as size(T ) → 0.

convplus Remark 3.7

ellcvgce
1. In Theorem 3.1, the hypothesis f ∈ L2 (Ω) is not necessary. It is used essentially to obtain a bound
on kuT k1,T . In order to pass to the limit, the hypothesis “f ∈ L1 (Ω)” is sufficient. Then, in
ellcvgce
Theorem 3.1, the hypothesis f ∈ L2 (Ω) can be replaced by f ∈ Lp (Ω) for some p > 1, if d = 2,
and for p ≥ 56 , if d = 3, provided that the meshes satisfy, for some fixed ζ > 0, dK,σ ≥ ζdσ , for all
σ ∈ EK and for all control volumes K. Indeed, one obtains, lq
in this
lq
case, a bound on kuT k1,T by
using a “discrete Sobolev inequality” (proved in Lemma 3.5 page 60).
elldifstaf
is also possible to obtain convergence results, towards a “very weak solution” of Problem ( 3.1),
Itelldifstab
(3.2), with only f ∈ L1 (Ω), by working with some discrete equivalent of the W01,q -norm, with
d
q < d−1 . This is not detailed here.
ellcvgce
2. In Theorem 3.1, it is also possible to prove enl1dcr
convergence results when f (x) is replaced by some
enl1dcr
nonlinear function f (x, u(x))
enl1dcr enl1dcr
as in Theorem 2.4 page 30. The proof is an easy adaptation of that
of Theorem 2.4 page 30.
ellcvgce
Proof of Theorem 3.1
Let Y be the set ofmeshdirichlet
approximate solutions, that is the set of uT where T is an admissible mesh in the
meshdirichlet
sense of Definition
ellsolvar
3.1 page 37. First, we want to prove that uT tends to the unique solution (in H01 (Ω))
to (3.3) as size(T ) → 0.
ellexistu ellinpoin
Thanks to Lemma 3.2 and to the discrete Poincaré inequality (3.13), there exists C1 ∈ IR, only depending h1dtot
on Ω and f , such that kuT k1,T ≤ C1 and kuT kL2kolmh10
(Ω) ≤ C1 for all u ∈ Y . Then, thanks to Lemma 3.3
kolmh10 T
and to the compactness result given in Theorem 3.10 page 93, the set Y is relatively compact in L2 (Ω)
and any possible limit (in L2 (Ω)) of a sequence (uTn )n∈IN ⊂ Y (such that size(T ) → 0) belongs to H01 (Ω).
ellsolvar n
Therefore, thanks to the uniqueness of the solution (in H01 (Ω)) of (3.3), it is sufficient to prove that if
(uTn )n∈IN ⊂ Y converges
ellsolvar
towards some u ∈ H01 (Ω), in L2 (Ω), and size(Tn ) → 0 (as n → ∞), then u is
the solution to (3.3). We prove this result below, omiting the index n, that is assuming uT → u in L2 (Ω)
as size(T ) → 0.

Let ψ ∈ Cc∞ (Ω) and let size(T ) be small enough so that ψ(x) = 0 if x ∈ K and K ∈ T is such that
ellschema
∂K ∩ ∂Ω 6= ∅. Multiplying (3.20) by ψ(xK ), and summing the result over K ∈ T yields

T1 + T 2 + T 3 = T 4 , (3.35) ell1g
with
X
T1 = b m(K)uK ψ(xK ),
K∈T
X X
T2 = − τK|L (uL − uK )ψ(xK ),
K∈T L∈N (K)
X X
T3 = vK,σ uσ,+ ψ(xK ),
K∈T σ∈EK
X
T4 = m(K)ψ(xK )fK .
K∈T
 47

First remark that, since uT tends to u in L2 (Ω),

Z
T1 → b u(x)ψ(x)dx as size(T ) → 0.
Ω

Similarly, Z
T4 → f (x)ψ(x)dx as size(T ) → 0.
Ω

Let us now turn to the study of T2 ;

X
T2 = − τK|L (uL − uK )(ψ(xK ) − ψ(xL )).
K|L∈Eint

Consider the following auxiliary expression:

Z
T20 = uT (x)∆ψ(x)dx
X Ω Z
= uK ∆ψ(x)dx
K
K∈T
X Z
= (uK − uL ) ∇ψ(x) · nK,L dγ(x).
K|L∈Eint K|L
Z
Since uT converges to u in L2 (Ω), it is clear that T20 tends to u(x)∆ψ(x) dx as size(T ) tends to 0.
Ω
Define
Z
1 ψ(xL ) − ψ(xK )
RK,L = ∇ψ(x) · nK,L dγ(x) − ,
m(K|L) K|L dK|L
where nK,L denotes the unit normal vector to K|L, outward to K, then
X
|T2 + T20 | = | m(K|L)(uK − uL )RK,L |
K|L∈Eint
h X (uK − uL )2 X i1/2
≤ m(K|L) m(K|L)dK|L (RK,L )2 ,
dK|L
K|L∈Eint K|L∈Eint

Regularity properties of the function ψ give the existence of C2 ∈ IR, only depending on ψ, such that
|RK,L | ≤ C2 size(T ). Therefore, since
X
m(K|L)dK|L ≤ dm(Ω),
K|L∈Eint
ellestimx
from Estimate (3.24), we conclude that
R T2 + T20 → 0 as size(T ) → 0.
Let us now show that T3 tends to − Ω v(x)u(x)∇ψ(x)dx as size(T ) → 0. Let us decompose T3 = T30 +T300
where
X X
T30 = vK,σ (uσ,+ − uK )ψ(xK )
K∈T σ∈EK

and Z
X X
T300 = vK,σ uK ψ(xK ) = divv(x)uT (x)ψT (x)dx,
K∈T σ∈EK Ω
 48

where ψT is defined by ψT (x) = ψ(xK ) if x ∈ K, K ∈ T . Since uT → u and ψT → ψ in L2 (Ω) as

size(T ) → 0 (indeed, ψT → ψ uniformly on Ω as size(T ) → 0) and since divv ∈ L∞ (Ω), one has
Z
T300 → divv(x)u(x)ψ(x)dx as size(T ) → 0.
Ω

Let us now rewrite T30 as T30 = T3000 + r3 with

X X Z
T3000 = (uσ,+ − uK ) v(x) · nK,σ ψ(x)dγ(x)
K∈T σ∈EK σ

and Z
X X
r3 = (uσ,+ − uK ) v(x) · nK,σ (ψ(xK ) − ψ(x))dγ(x).
K∈T σ∈EK σ

Thanks to the regularity of v and ψ, there exists C3 only depending on v and ψ such that
X
|r3 | ≤ C3 size(T ) |uK − uL |m(K|L),
K|L∈Eint

which yields, with the Cauchy-Schwarz inequality,

X 1
X 1
|r3 | ≤ C3 size(T )( τK|L |uK − uL |2 ) 2 ( m(K|L)dK|L ) 2 ,
K|L∈Eint K|L∈Eint
ellestimx
from which one deduces, with Estimate (3.24), that r3 → 0 as size(T ) → 0.
Next, remark that
X X Z X Z
T3000 = − uK v(x) · nK,σ ψ(x)dγ(x) = − uK div(v(x)ψ(x))dx.
K∈T σ∈EK σ K∈T K
R
This implies (sinceR uT → u in L2 (Ω)) that T3000 → − Ω
div(v(x)ψ(x))u(x)dx, so that T30 has the same
limit and T3 → − Ω v(x) · ∇ψ(x)u(x)dx.
ell1g
Hence, letting size(T ) → 0 in (3.35) yields that the function u ∈ H01 (Ω) satisfies
Z  
bu(x)ψ(x) − u(x)∆ψ(x) − v(x)u(x)∇ψ(x) − f (x)ψ(x) dx = 0, ∀ψ ∈ Cc∞ (Ω),
Ω
ellsolvar
which, in turn, yields (3.3) thanks to the fact that u ∈ H01 (Ω), and to the density of Cc∞ (Ω) in H01 (Ω).
concludes the proof of uT → u in L2 (Ω) as size(T ) → 0, where u is the unique solution (in H01 (Ω))
Thisellsolvar
to (3.3).
S Let us now prove that kuT k1,T tends to kukH01 (Ω) in the pure diffusion case, i.e. assuming b = 0 and
v = 0. Since
Z Z
kuT k21,T = fT (x)uT (x)dx → f (x)u(x)dx as size(T ) → 0,
Ω Ω

where fT is defined from Ω to IR by fT (x) = fK a.e. on K for all K ∈ T , it is easily seen that
Z
kuT k21,T → f (x)u(x)dx = kuk2H 1 (Ω) as size(T ) → 0.
0
Ω
ellcvgce
This concludes the proof of Theorem 3.1.
 49

ellcvgce
creme Remark 3.8 (Consistency for the adjoint operator) The proof of Theorem 3.1 uses the property
of consistency of the (diffusion) fluxes on the test functions. This property consists in writing the
consistency of the fluxes for the adjoint operator to the discretized Dirichlet operator. This consistency is
achieved thanks to that of fluxes for the discretized Dirichlet operator and to the fact that this operator
is self adjoint. In fact, any discretization of the Dirichlet operator giving
cafe
“L2 -stability”
cafe
and consistency
of fluxes on its adjoint, yields a convergence result
errordirichlet
sh2ee
(see also Remark 3.2 page 37). On the contrary, the
error estimates proved in sections 3.1.5 and 3.1.6 directly use the consistency for the discretized Dirichlet
operator itself.

Liebig Remark 3.9 (Finite volume schemes and H 1 approximate solutions)

In the above proof, we showed that a sequence of approximate solutions (which are piecewise constantkolmh10
functions) converges in L2 (Ω) to a limit which is in H01 (Ω). An alternative to the use of Theorem 3.10 is
d
the construction of a bounded sequence in H 1 (IR ) from the sequence of approximate solutions. This can
be performed by convoluting the approximate solution with a mollifier “of size size(T )”. Using Rellich’s
compactness theorem and the weak sequential compactness of the bounded sets of H 1 , one obtains that
the limit of the sequence of approximate solutions is in H01 .

Let us now deal with the case of non homogeneous Dirichlet boundary conditions, in which case g ∈
H 1/2 (∂Ω) is no longer assumed to be 0. The proof uses the following preliminary result:

thuh1 Lemma 3.4 Let Ω be an open bounded polygonal subset of IR 2 , g̃ ∈ H 1 (Ω) and g = γ(g̃) (recall that γ is
the “trace” operator from H 1 (Ω) to H 1/2 (∂Ω)). Let T be an admissible mesh (in the sense of Definition
meshdirichlet
meshdirichlet
3.1 page 37) such that, for some ζ > 0, the inequality dK,σ ≥ ζdiam(K) holds for all control volumes
K ∈ T and for all σ ∈ EK , and let M ∈ IN be such that card(EK ) ≤ M for all K ∈ T . Let us define g̃K
for all K ∈ T by
Z
1
g̃K = g̃(x)dx
m(K) K
and g̃σ for all σ ∈ Eext by
Z
1
g̃σ = g(x)dγ(x).
m(σ) σ

Let us define
 X X  21
N (g̃, T ) = τK|L (g̃K − g̃L )2 + τσ (g̃K(σ) − g̃σ )2 , (3.36) Ngtilde
σ=K|L∈Eint σ∈Eext

where K(σ) = K if σ ∈ Eext ∩ EK . Then there exists C ∈ IR + , only depending on ζ and M , such that

N (g̃, T ) ≤ Ckg̃kH 1 (Ω) . (3.37) ellthuh1

thuh1
Proof of Lemma 3.4
thuh1
Lemma 3.4 is given in the two dimensional case, an analogousthuh1 result is possible in the three dimensional
case. Let Ω, g̃, T , ζ, M satisfying the hypotheses of Lemma 3.4. By a classical argument of density, one
may assume that g̃ ∈ C 1 (Ω, IR).
A first step consists in proving that there exists C1 ∈ IR + , only depending on ζ, such that
Z
diam(K)
(g̃K − g̃σ )2 ≤ C1 |∇g̃(x)|2 dx, ∀K ∈ T , ∀σ ∈ EK , (3.38) thuh11
m(σ) K

where g̃K (resp. g̃σ ) is the mean value of g̃ on K (resp. σ), for K ∈ T (resp. σ ∈ E). Indeed, without
loss of generality, one assumes that σ = {0} × J0 , with J0 is a closed interval of IR and K ⊂ IR + × IR.
Let α = max{x1 , x = (x1 , x2 )t ∈ K} and a = (α, β)t ∈ K. In the following, a is fixed. For all x1 ∈ (0, α),
let J(x1 ) = {x2 ∈ IR, such that (x1 , x2 )t ∈ K}, so that J0 = J(0).
 50

For a.e. x = (x1 , x2 )t ∈ K and a.e., for the 1-Lebesgue measure, y = (0, y)t ∈ σ (with y ∈ J0 ), one sets
z(x, y) = ta+(1−t)y with t = xα1 . Note that, since K is convex, z(x, y) ∈ K and z(x, y) = (x1 , z2 (x1 , y))t ,
with z2 (x1 , y) = xα1 β + (1 − xα1 )y.
One has, using the Cauchy-Schwarz inequality,
2
(g̃K − g̃σ )2 ≤ (A + B), (3.39) thuh111
m(K)m(σ)
where
Z Z

2
A= g̃(x) − g̃(z(x, y)) dγ(y)dx,
K σ
and
Z Z

2
B= g̃(z(x, y)) − g̃(y) dγ(y)dx.
K σ
Let us now obtain a bound of A. Let Di g̃, i = 1 or 2, denote the partial derivative of g̃ w.r.t. the
components of x = (x1 , x2 )t ∈ IR 2 . Then,
Z αZ Z Z x2

2
A= D2 g̃(x1 , s)ds dydx2 dx1 .
0 J(x1 ) J(0) z2 (x1 ,y)

The Cauchy-Schwarz inequality yields

Z αZ Z Z

2
A ≤ diam(K) D2 g̃(x1 , s) dsdydx2 dx1
0 J(x1 ) J(0) J(x1 )

and therefore
Z
3

2
A ≤ diam(K) D2 g̃(x) dx. (3.40) thuh12
K
One now turns to the study of B, which can be rewritten as
Z αZ Z Z x1
β−y
2
B= [D1 g̃(s, z2 (s, y)) + D2 g̃(s, z2 (s, y))]ds dydx2 dx1 .
0 J(x1 ) J(0) 0 α
The Cauchy-Schwarz inequality and the fact that α ≥ ζdiam(K) give that
1
B ≤ 2diam(K)(B1 + B2 ), (3.41) thuh112
ζ2
with
Z α Z Z Z x1
2
Bi = Di g̃(s, z2 (s, y)) dsdydx2 dx1 , i = 1, 2.
0 J(x1 ) J(0) 0

First, using Fubini’s theorem, one has

Z Z α Z α Z

2
Bi = Di g̃(s, z2 (s, y)) dx2 dx1 dsdy.
J(0) 0 s J(x1 )

Therefore
Z α Z

2
Bi ≤ diam(K) Di g̃(s, z2 (s, y)) (α − s)dyds.
0 J(0)

Then, with the change of variables z2 = z2 (s, y), one gets

 51

Z α Z

2 α − s
Bi ≤ diam(K) Di g̃(s, z2 ) dz2 ds.
0 J(s) 1 − αs
Hence Z
2

2
Bi ≤ diam(K) Di g̃(x) dx. (3.42) thuh13
K

2 thuh111 thuh12 thuh112 thuh13 thuh11
Using the fact that m(K) ≥ πζ 2 diam(K) , (3.39), (3.40), (3.41) and (3.42), one concludes (3.38).
ellthuh1
In order to conclude the proof of (3.37), one remarks that
 2 X X
N (g̃, T ) ≤2 τσ (g̃K − g̃σ )2 .
K∈T σ∈EK
thuh11
Because, for all K ∈ T and σ ∈ EK , dσ ≥ ζdiam(K), one gets thanks to (3.38), that
 2 X X C1 Z
N (g̃, T ) ≤ 2 |∇g̃(x)|2 dx.
ζ K
K∈T σ∈EK

The above inequality shows that

 2 Z
C1
N (g̃, T ) ≤ 2M |∇g̃(x)|2 dx,
ζ Ω
ellthuh1
which implies (3.37).

ellcvgcediraf Theorem 3.2 (Convergence, non homogeneous ellH

Dirichlet
ellH
boundary condition)
Assume items 1, 2, 3 and 4 of Assumption 3.1 page 32 and g ∈ H 1/2 (∂Ω). ∈ IR + and M ∈ IN
Let ζmeshdirichlet
meshdirichlet
be given values. Let T be an admissible mesh (in the sense of Definition 3.1 page 37) such that dK,σ ≥
ζdiam(K) for all control volumes K ∈ T and for all σ ∈ Eellschema
K , and card(EK ) ≤ M for all K ∈ T . Let
ellschemac
(uK )K∈T be the solution of the system given by equations (3.20)-(3.22) and
Z
1
uσ = g(x)dγ(x), ∀σ ∈ Eext . (3.43) ellschemadi
m(σ) σ
ellexistu
(note that the proofs of existence and uniqueness of (uK )K∈T which were given in Lemma 3.2 page
ellexistu
42 remain valid). Define uT ∈ X(T ) by uT (x) = uK for a.e. x ∈ K and for K ∈ T . Then, uT
any elldifstab
elldifstaf
converges, in L2 (Ω), to the unique variational solution u ∈ H 1 (Ω) of Problem (3.1), (3.2) as size(T ) → 0.
ellcvgcediraf
Proof of Theorem 3.2
The proof is only detailed for the case b = 0 and v = 0 (the extension of the proof to the general case
ellcvgce ellcvgce
is straightforward using the proof of Theorem 3.1 page 46). Let g̃ ∈ H 1 (Ω) be such that the trace of
g̃ on ∂Ω is equal
R to g. One defines ũT ∈ X(T ) by ũT = uT − g̃T where g̃T ∈ X(T ) is defined by
1
g̃(x) = m(K) K
g̃(y)dy for all x ∈ K and all K ∈ T . Then (ũK )K∈T satisfies
X X
F̃K,σ = m(K)fK − GK,σ , ∀K ∈ T , (3.44) ellschemada
σ∈EK σ∈EK

F̃K,σ = −τK|L (ũL − ũK ), ∀σ ∈ Eint , if σ = K|L, (3.45) ellschembda

F̃K,σ = τσ (ũK ), ∀σ ∈ Eext such that σ ∈ EK . (3.46) ellschemcda

GK,σ = −τK|L (g̃L − g̃L ), ∀σ ∈ Eint , if σ = K|L, (3.47) ellschemdda

 52

GK,σ = −τσ (g̃σ − g̃L ), ∀σ ∈ Eext such that σ ∈ EK , (3.48) ellschemeda

1
R ellschemada
where g̃σ = m(σ) σ
g(x)dγ(x) Multiplying (3.44) by ũK , summing over K ∈ T , gathering by edges in the
right hand side and using the Cauchy-Schwarz inequality yields
X
kũT k21,T ≤ m(K)fK ũK + N (g̃, T )kũT k1,T ,
K∈T
Ngtilde Ngtilde defh10d defh10d
from the definition
thuh1
3.36) page 49 of N (g̃, T ) and Definitionellinpoin
(thuh1 3.3 page 40 of k · k1,T . Therefore, thanks
to Lemma 3.4 page 49 and the discrete Poincaré inequality (3.13), there exists C1 ∈ IR, only depending
on Ω, kg̃kH 1 (Ω) , ζ, M and f , such that kũT k1,T ≤ C1 and kũT kL2 (Ω) ≤ C1 . Letellsolvar
us now prove that ũT
in L2 (Ω),ellcvgce
converges ellcvgce towards the unique solution in H01 (Ω) to (3.3).
as size(T ) → 0, h1dtot We proceed
kolmh10
as in
kolmh10
Theorem 3.1 page 46. Using Lemma 3.3, the compactness
ellsolvar
result given in Theorem 3.10 page 93 and the
uniqueness of the solution (in H01 (Ω)) of (3.3), it is sufficient to prove that
ellsolvar
if ũ T converges towards some
ũ ∈ H01 (Ω), in L2 (Ω) as size(T ) → 0, then ũ is the solution to (3.3). In order to prove this result, let us
introduce the function g̃T defined by
Z
1
g̃T (x) = g̃(y)dy, ∀x ∈ K, ∀K ∈ T ,
m(K) K
which converges to g̃ in L2 (Ω), as size(T ) → 0. Then the function uTellsolvar
converges in L2 (Ω), as size(T ) → 0
1
to u = ũ + g̃ ∈ H (Ω) and the ellcvgce ũ is the unique solution of (3.3) is identical
proof thatellcvgce to the corresponding
ellcvgcediraf
part in the proof of Theorem 3.1 page 46. This completes the proof of Theorem 3.2.

Remark 3.10 A more simple proof of convergence for the finite volume scheme with non homogeneous
Dirichlet boundary condition can be made if g is the trace of a Lipschitz-continuous function g̃. In that
thuh1
case, ζ and M do not have to be introduced and Lemma 3.4 is not used. The scheme is defined with
uσ = g(yσ ) instead of the average value of g on σ, and the proof uses g̃(xK ) instead of the average value
of g̃ on K.

3.1.5 C 2 error estimate

rrordirichlet elldifstaf
elldifstab
Under adequate regularity assumptions on the solution of Problem (3.1)-(3.2), one may prove that the
ellschema
error between
ellschemad
the exact solution and the approximate solution given by the finite volume scheme ( 3.20)-
(3.23) is of order size(T ) = supK∈T diam(K), in a certain sense which we give in the following theorem:
ellH ellH meshdirichlet
ellesterr Theorem 3.3 Under Assumption 3.1 page
meshdirichlet
T be an admissible mesh as defined in Definition 3.1
32, letXmesh
Xmesh
page 37 and uT ∈ X(T ) (see Definition 3.2 page 39) be defined
ellschema
a.e.in Ω by uT (x) = uK for a.e. x ∈ K,
ellschemad
for all K ∈ T , where (u K ) K∈T
elldifstaf
is the
elldifstab
solution to ( 3.20)-( 3.23). Assume that the unique variational
solution u of Problem (3.1)-(3.2) satisfies u ∈ C 2 (Ω). Let, for each K ∈ T , eK = u(xK ) − uK , and
eT ∈ X(T ) defined by eT (x) = eK for a.e. x ∈ K, for all K ∈ T .
Then, there exists C > 0 only depending on u, v and Ω such that

keT k1,T ≤ Csize(T ), (3.49) eqesterrh1

defh10d
where k · k1,T is the discrete H01 norm defined in Definition 3.3,

keT kL2 (Ω) ≤ Csize(T ) (3.50) eqesterr

and
 53

X u − u Z 2
L K 1
m(σ)dσ − ∇u(x) · nK,σ dγ(x) +
σ∈Eint
dσ m(σ) σ
X
σ=K|L
 Z 2 (3.51) eqesterr1
g(yσ ) − uK 1
m(σ)dσ − ∇u(x) · nK,σ dγ(x) ≤ Csize(T )2 .
σ∈E
dσ m(σ) σ
ext
σ∈K∩∂Ω

Remark 3.11
eqesterrh1 eqesterr
1. Inequality (3.49) (resp. (3.50)) yields an estimate of order 1 for the discrete H01 norm (resp.
eqesterr
L2
1
norm) of the error on the solution. Note also that, since u ∈ C (Ω), one deduces, from (3.50),eqesterr1
the
existence of C only depending on u and Ω such that ku − uT kL2 (Ω) ≤ Csize(T ). Inequality (3.51)
may be seen as an estimate of order 1 for the L2 norm of the flux.
BMO
2. In Baranger, Maitre and Oudin [1996], finite element tools are used to obtain error estimates
of order size(T )2 in the case d = 2, v = b = g = 0 and if the elements
scvfe
of T are triangles of a finite
Delaunay
element mesh satisfying the Delaunay condition
h2conv
(see section 3.4 page 85).
h2conv h2conv2d h2conv2d
Note that this result is
quite different of those of the remarks 2.5 page 18 and 3.1 page 35, which are obtained by using a
higher order approximation of the flux.
ellesterr
3. The proof of Theorem 3.3 given below is close to that of error estimates for finite element schemes
in the sense that it uses the coerciveness of the operator
maxppe
(themaxppe
discrete Poincaré inequality) instead
of the discrete maximum principle of Proposition 3.2 page 44 (which is used for error estimates
with finite difference schemes).
ellesterr
Proof of Theorem 3.3
Let uT ∈ X(T ) ellschema
be defined a.e. in Ω by uT (x) = uK for a.e. x ∈ K, for all K ∈ T , where (uK )K∈T is
ellschemad
the solution to (3.20)-(3.23). Let us write the flux balance for any K ∈ T ;
X  Z Z
F K,σ + V K,σ + b u(x)dx = f (x)dx, (3.52) eller1
σ∈EK K K
R R
where F K,σ = − σ ∇u(x) · nK,σ dγ(x), and V K,σ = σ u(x)v(x) · nK,σ dγ(x) are respectively the diffusion
and convection fluxes through σ outward to K.
? ?
Let FK,σ and VK,σ be defined by
?
FK,σ = −τK|L (u(xL ) − u(xK )), ∀σ = K|L ∈ EK ∩ Eint , ∀K ∈ T ,

?
FK,σ d(xK , σ) = −m(σ)(u(yσ ) − u(xK )), ∀σ ∈ EK ∩ Eext , ∀K ∈ T ,

?
VK,σ = vK,σ u(xσ,+ ), ∀σ ∈ EK , ∀K ∈ T ,
where xσ,+ = xK (resp. xL ) if σ ∈ Eint , σ = K|L and vK,σ ≥ 0 (resp. vK,σ ≤ 0) and xσ,+ = xK (resp.
yσ ) if σ = EK ∩ Eext and vK,σ ≥ 0 (resp. vK,σ ≤ 0). Then, the consistency error on the diffusion and
convection fluxes may be defined as
1 ?
RK,σ = (F K,σ − FK,σ ), (3.53) eller2
m(σ)
1 ?
rK,σ = (V K,σ − VK,σ ), (3.54) eller2b
m(σ)
 54

Thanks to the regularity of u and v, there exists C1 ∈ IR, only depending on u and v, such that
|RK,σ | + |rK,σ | ≤ C1 size(T ) for any K ∈ T and σ ∈ EK . For K ∈ T , let
Z
ρK = u(xK ) − (1/m(K)) u(x)dx,
K

so that |ρK | ≤ C2 size(T ) with some C2 ∈ IR + only depending on u.

ellschema eller1 eller2 eller2b
Substract (3.20) to (3.52); thanks to (3.53) and (3.54), one has
X  X
GK,σ + WK,σ + bm(K)eK = bm(K)ρK − m(σ)(RK,σ + rK,σ ), (3.55) eller3
σ∈EK σ∈EK

where
?
GK,σ = FK,σ − FK,σ is such that

GK,σ = −τK|L (eL − eK ), ∀K ∈ T , ∀σ ∈ EK ∩ Eint , σ = K|L,

GK,σ d(xK , σ) = m(σ)eK , ∀K ∈ T , ∀σ ∈ EK ∩ Eext ,

?
with eK = u(xK ) − uK , and WK,σ = VK,σ − VK,σ = vK,σ (u(xσ,+ ) − uσ,+ )
eller3
Multiply (3.55) by eK , sum for K ∈ T , and note that
X X X m(σ)
GK,σ eK = |Dσ e|2 = kek21,T .
dσ
K∈T σ∈EK σ∈E

Hence

X X X X X
keT k21,T + vK,σ eσ,+ eK +bkeT k2L2 (Ω) ≤ b m(K)ρK eK − m(σ)(RK,σ +rK,σ )eK , (3.56) ellraslebol1
K∈T σ∈EK K∈T K∈T σ∈EK

where
eT ∈ X(T ), eT (x) = eK for a.e. x ∈ K and for all K ∈ T ,
|Dσ e| = |eK − eL |, if σ ∈ Eint , σ = K|L, |Dσ e| = |eK |, if σ ∈ EK ∩ Eext ,
eσ,+ = u(xσ,+ ) − uσ,+ .
By Young’s inequality, the first term of the left hand side satisfies:
X 1 1
| m(K)ρK eK | ≤ keT k2L2 (Ω) + C22 (size(T ))2 m(Ω). (3.57) ellyoung
2 2
K∈T
vfqconv1
vfqconv11 vfqconv11
Thanks to the assumption divv ≥ 0, one obtains, through a computation similar to (3.27)-(3.28) page 43
that X X
vK,σ eσ,+ eK ≥ 0.
K∈T σ∈EK
ellraslebol1ellyoung
Hence, (3.56) and (3.57) yield that there exists C3 only depending on u, b and Ω such that
1 X X
keT k21,T + bkeT k2L2 (Ω) ≤ C3 (size(T ))2 − m(σ)(RK,σ + rK,σ )eK , (3.58) ellraslebol2
2
K∈T σ∈EK

Thanks to the property of conservativity, one has RK,σ = −RL,σ and rK,σ = −rL,σ for σ ∈ Eint such that
σ = K|L. Let Rσ = |RK,σ | and rσ = |rK,σ | if σ ∈ EK . Reordering the summation over the edges and
from the Cauchy-Schwarz inequality, one then obtains
 55

X X X
| m(σ)(RK,σ + rK,σ )eK | ≤ m(σ)(Dσ e)(Rσ + rσ ) ≤
K∈T σ∈EK σ∈E
 X m(σ)  12 X  21 (3.59) ellraslebol3
(Dσ e)2 m(σ)dσ (Rσ + rσ )2 .
dσ
σ∈E σ∈E
X ellraslebol2ellraslebol3
Now, since |Rσ + rσ | ≤ C1 size(T ) and since m(σ)dσ = d m(Ω), (3.58) and (3.59) yield the existence
σ∈E
of C4 ∈ IR + only depending on u, v and Ω such that
1
keT k21,T + bkeT k2L2 (Ω) ≤ C3 (size(T ))2 + C4 size(T )kek1,T .
2
Using again Young’s inequality, there exists C5 only depending on u, v, b and Ω such that

keT k21,T + bkeT k2L2 (Ω) ≤ C5 (size(T ))2 . (3.60) ellraslebol4

eqesterrh1 eqesterr
This inequality yields Estimate (3.49) and,ellinpoin
in the case b > 0, Estimate (3.50). In the case where b = 0,
ellraslebol4
one uses the discrete Poincaré inequality (3.13) and the inequality (3.60) to obtain

keT k2L2 (Ω) ≤ diam(Ω)2 C5 (size(T ))2 ,

eqesterr
which yields (3.50).
eqesterrh1
Remark now that (3.49) can be written
X u − u u(xL ) − u(xK ) 2
L K
m(σ)dσ − +
σ∈Eint
dσ dσ
σ=K|L
X  g(y ) − u u(yσ ) − u(xK ) 2 (3.61) caspie3
σ K
m(σ)dσ − ≤ (Csize(T ))2 .
σ∈Eext
dσ dσ
σ∈K∩∂Ω
eller2
From Definition (3.53) and the consistency of the fluxes, one has
X  u(x ) − u(x ) Z 2
L K 1
m(σ)dσ − ∇u(x) · nK,σ dγ(x) +
σ∈Eint
dσ m(σ) σ
σ=K|L
X  u(y ) − u(x ) Z 2
σ K 1
m(σ)dσ − ∇u(x) · nK,σ dγ(x) = (3.62) caspie4
σ∈Eext
dσ m(σ) σ
σ∈K∩∂Ω
X
m(σ)dσ Rσ2 ≤ dm(Ω)C12 (size(T ))2 .
σ∈E
caspie3 caspie4 eqesterr1
Then (3.61) and (3.62) give (3.51).

3.1.6 H 2 error estimate

sh2ee ellesterr testh2
In Theoremeqesterrh1
3.3, the hypothesis
eqesterr
u ∈ C 2 (Ω) was used. In the following theorem (Theorem 3.4), one obtains
Estimates (3.49)ramand (3.50), in the case b = v = 0 and assuming some additional assumption on the mesh
(see Definition 3.4 below), under the weaker assumption
h2plus
u ∈ H 2 (Ω). This additional assumption on the
ghv
mesh is not completely necessary (see Remark 3.13 and Gallouët, Herbin and Vignal [1999]). It is
eqesterrh1 eqesterr 2
also possible to
h2plus
obtain
ghv
Estimates ( 3.49) and (3.50) in the cases b 6= 0 or v 6= 0 assuming u ∈ Hlazarovsiam
(Ω) (see
Remark 3.13 and Gallouët, Herbin coudiere and Vignal [1999]). Some similar results are also in Lazarov,
Mishev and Vassilevski [1996] and Coudière, Vila and Villedieu [1999].
 56

ram Definition 3.4 (Restricted admissible meshes) Let Ω be an open bounded polygonal subset of IR d ,
d = 2 or 3. A restricted admissible finite volume mesh of Ω, denoted by T , is an admissible mesh in the
meshdirichlet
sense of Definition 3.1 such that, for some ζ > 0, one has dK,σ ≥ ζdiam(K) for all control volumes K
and for all σ ∈ EK .
ellH ellH
testh2 Theorem 3.4 (H 2 regularity) Under Assumption ram
3.1 page 32 with b = v = 0, let T be aXmesh
Xmesh
restricted
admissible mesh in the sense of Definition 3.4 and uT ∈ X(T ) (see Definition 3.2 page 39) be the
approximate solution defined Ω by uT (x) = uK for a.e. x ∈ K, for all K ∈ T , where (uK )K∈T
in ellschemad
ellschema
is
ellexistu
the (unique) solution to (3.20)-(3.23) (existence
ellsolvar
and uniqueness of (u )
K K∈T are given by Lemma 3.2).
Assume that the unique solution, u, of (3.3) (with b = v = 0) belongs to H 2 (Ω). For each control volume
K, let eK = u(xK ) − uK , and eT ∈ X(T ) defined by eT (x) = eKeqesterrh1
for a.e. x ∈ K, for
eqesterr
all K ∈ T .
eqesterr1
Then, there exists C, only depending on u, ζ and Ω, such that (3.49), (3.50) and (3.51) hold.

uish2 Remark 3.12

testh2
1. In Theorem 3.4, the function eT is still well defined, and so is the quantity “∇u · nσ ” on σ, for all
σ ∈ E. Indeed, since u ∈ H 2 (Ω) (and d ≤ 3), one has u ∈ C(Ω) (and then u(xK ) is well defined for
all control volumes K) and ∇u · nσ belongs to L2 (σ) (for the (d − 1)-dimensional Lebesgue measure
on σ) for all σ ∈ E.
ellH ellsolvar
2. Note that, under Assumption 3.1 with b = v = g = 0 the (unique) solution of (3.3) is necessarily
in H 2 (Ω) provided that Ω is convex.
testh2
Proof of Theorem 3.4
Let K be a control volume and σ ∈ EK . Define VK,σ = {txK + (1 − t)x, x ∈ σ, t ∈ [0, 1]}. For σ ∈ Eint ,
let Vσ = VK,σ ∪ VL,σ , if K and L are the control volumes such that σ = K|L. For σ ∈ Eext ∩ EK , let
Vσ = VK,σ .
The main part of the proof consists inram
proving the existence of some C, only depending on the space
dimension d and ζ (given in Definition 3.4), such that, for all control volumes K and for all σ ∈ EK ,
Z
(size(T ))2
|RK,σ |2 ≤ C |H(u)(z)|2 dz, (3.63) consh2
m(σ)dσ Vσ
whereH is the Hessian matrix of u and
d
X
|H(u)(z)|2 = |Di Dj u(z)|2 ,
i,j=1

and Di denotes the (weak) derivative with respect to the componenteller2zi of z = (z1 , · · · , zd )t ∈ IR d .
Recall that RK,σ is the consistency error on the diffusion flux (see (3.53)), that is:
Z
u(xL ) − u(xK ) 1
RK,σ = − ∇u(x) · nK,σ dγ(x), if σ ∈ Eint and σ = K|L,
dσ m(σ) σ
Z
u(yσ ) − u(xK ) 1
RK,σ = − ∇u(x) · nK,σ dγ(x), if σ ∈ Eext ∩ EK .
dσ m(σ) σ
uish2
Note that RK,σ is well defined, thanks to u ∈ H 2 (Ω), see Remark 3.12.
consh2 eqesterrh1 eqesterr
In Step 1, one proves (3.63), and, in Step 2, we conclude the proof of Estimates (3.49) and (3.50).
consh2
Step 1. Proof of (3.63).
Let σ ∈ E. Since u ∈ H 2 (Ω), the restriction of u to Vσ belongs to H 2 (Vσ ). The space C 2 (Vσ ) is dense in
necas
2
H (Vσ ) (see, for instance, Nečas [1967], this can be proved
consh2
quite easily be a regularization technique).
Then, by a density argument, one needs only to prove (3.63) for u ∈ C 2 (Vσ ). Therefore, in the remainder
of Step 1, it is assumed u ∈ C 2 (Vσ ).
 57

consh2
First, one proves (3.63) if σ ∈ Eint . Let K and L be the 2 control volumes such that σ = K|L.
It is possible to assume, for simplicity of notations and without loss of generality, that σ = 0 × σ̃, with
some σ̃ ⊂ IR d−1
ram
, and xram t t
K = (−α, 0) , xL = (β, 0) , with some α > ζdiam(K), β > ζdiam(L) (ζ is defined
in Definition 3.4 page 56).
Since u ∈ C 2 (Vσ ) a Taylor expansion gives for a.e. (for the (d − 1)-dimensional Lebesgue measure on σ)
x = (0, x̃)t ∈ σ,
Z 1
u(xL ) − u(x) = ∇u(x) · (xL − x) + H(u)(tx + (1 − t)xL )(xL − x) · (xL − x)tdt,
0
and
Z 1
u(xK ) − u(x) = ∇u(x) · (xK − x) + H(u)(tx + (1 − t)xK )(xK − x) · (xK − x)tdt,
0

where H(u)(z) denotes the Hessian matrix of u at point z.

Subtracting one equation to the other and integrating over σ yields (note that x L − xK = nK,σ dσ )
|RK,σ | ≤ BK,σ + BL,σ , with, for some C1 only depending on d,
Z Z 1
C1
BK,σ = |H(u)(tx + (1 − t)xK )||xK − x|2 tdtdγ(x). (3.64) donotpanic
m(σ)dσ σ 0
The quantity BL,σ is obtained withdonotpanic
BK,σ by changing K in L.
One uses a change of variables in (3.64). Indeed, one sets z = tx + (1 − t)xK . Since |xK − x| ≤ diam(K)
and dz = td−1 αdtdγ(x), one obtains, since z1 = (t − 1)α, z = (z1 , z)t ,
Z
C1 (diam(K))2 αd−2
BK,σ ≤ |H(u)(z)| dz.
m(σ)dσ VK,σ α(z1 + α)d−2
This gives, with the famous Cauchy-Schwarz inequality,
Z Z
C1 αd−3 (diam(K))2
1 1
1
BK,σ ≤ |H(u)(z)|2 dz 2 (d−2)2
dz 2 . (3.65) onrespirefor
m(σ)dσ VK,σ VK,σ (z1 + α)
onrespirefort
For d = 2, (3.65) gives
Z
C1 (diam(K))2 αm(σ)
12
21
BK,σ ≤ |H(u)(z)|2 dz ,
αm(σ)dσ 2 VK,σ

and therefore
Z
C1 (diam(K))2
12
BK,σ ≤ 1 1 1 |H(u)(z)|2 dz .
2 (m(σ)dσ ) (dσ α)
2 2 2 VK,σ

A similar estimate holds on BL,σ by changing K in L and αconsh2 in β. Since α, β ≥ ζdiam(K) and dσ =
α + β ≥ ζdiam(K), these estimates on BK,σ and BL,σ yield (3.63) for some C only depending on d and
ζ. R 1
For d = 3, the computation of the integral A = VK,σ (z1 +α) 2 dz by the following change of variable (see
fig-consist-h2
Figure (3.1.6)):
Z 0 Z
1 z1 + α
A= 2
( dz)dz1 , where t = .
−d (z 1 + α) z∈tσ̃ dK,σ
Now, Z Z
(z1 + α)2
dz = t2 dy = m(σ),
z∈tσ̃ y∈σ̃ α2
 58

m(σ) onrespirefort
and therefore A = α , and (3.65) yields that:
Z !1/2
C3 (diam(K))2 C3 size(T )
BK,σ ≤ |H(u)(z)|2 dz ≤√ kH(u)kL2 (VK,σ ) .
(m(σ) d2σ dK,σ )1/2 VK,σ 2 ζ (m(σ) dσ )1/2

xK = (−dK,σ , 0)t (0,0) z1

tσ̃

σ̃

ig-consist-h2

Figure 3.3: Consistency error, d = 3

onrespirefort
and therefore (3.65) gives:
Z 0 Z
C1 (diam(K))2 m(σ)
1
21
BK,σ ≤ dz1 2 |H(u)(z)|2 dz ,
m(σ)dσ −α α2 VK,σ

and then
Z
C1 (diam(K))2
12
BK,σ ≤ 1 1 |H(u)(z)|2 dz .
(m(σ)dσ ) (dσ α)
2 2 VK,σ
consh2
With a similar estimate on BL,σ , this yields (3.63) for some C only depending on d and ζ.
consh2
Now, one proves (3.63) if σ ∈ Eext . Let K be the control volume such that σ ∈ EK . One can assume,
= {2α} × σ̃ with σ̃ ⊂ IR d−1 and some α ≥ 12 ζdiam(K).
without loss of generality, that xK = meshdirichlet
0 and σ meshdirichlet
The above proof gives (see Definition 3.1 page 37 for the definition of yσ ), with some C2 only depending
on d,
Z Z
u(yσ ) − u(xK ) 1 2 (size(T ))2
| − ∇u(x) · nK,σ dγ(x)| ≤ C2 |H(u)(z)|2 dz, (3.66) consbord1
2α m(σ̂) σ̂ m(σ)dσ Vσ̂
with σ̂ = {(α x2 ), x ∈ σ̃}, and Vσ̂ = {tyσ + (1 − t)x, x ∈ σ̂, t ∈ [0, 1]} ∪ {txK + (1 − t)x, x ∈ σ̂, t ∈ [0, 1]}.
Note that m(σ̂) = m(σ)2d−1 and that Vσ̂ R⊂ Vσ . R
1 1
One has now to compare Iσ = m(σ) σ ∇u(x) · nK,σ dγ(x) with Iσ̂ = m(σ̂) σ̂ ∇u(x) · nK,σ dγ(x).
A Taylor expansion gives
Z Z 1
1
Iσ − Iσ̂ = H(u)(xK + t(x − xK ))(x − xK ) · nK,σ dtdγ(x).
m(σ) σ 1
2

The change of variables in this last integral z = xK + t(x − xK ), which gives dz = 2αtd−1 dtdγ(x), yields,
with Eσ = {tx + (1 − t)xK , x ∈ σ, t ∈ [ 12 , 1]} and some C3 only depending on d (note that t ≥ 12 ),
 59

Z
C3
|Iσ − Iσ̂ | ≤ |H(u)(z)||x − xK |dz.
m(σ)α Eσ

Then, from the Cauchy-Schwarz inequality and since |x − xK | ≤ diam(K),

Z
C4 (diam(K))2
|Iσ − Iσ̂ |2 ≤ |H(u)(z)|2 dz, (3.67) consbord2
m(σ)dσ Eσ

with some C4consbord1

only depending on d and
consbord2
ζ.
consh2
Inequalities (3.66) and (3.67) yield (3.63) for some C only depending on d and ζ.
consh2
One may therefore choose C ∈ IR + such that (3.63) holds for σ ∈ Eint or σ ∈ Eext . This concludes Step 1.
eqesterrh1
eqesterr eqesterr1
Step 2. Proof of Estimates (3.49), (3.50) and (3.51).eqesterr
eqesterrh1 ellinpoin
In order to obtain Estimate ellesterr
(3.49) (and therefore (3.50) from the discrete Poincaré inequality (3.13)),
ellraslebol1
one proceeds as in Theorem 3.3. Inequality (3.56) writes here, since RK,σ = −RL,σ , if σ = K|L,
X
keT k21,T ≤ Rσ |Dσ e|m(σ),
σ∈E

with Rσ = |RK,σ |, if σ ∈ EK . Recall also that |Dσ e| = |eK − eL | if σ ∈ Eint , σ = K|L and |Dσ e| = |eK |,
if σ ∈ Eext ∩ EK . Cauchy and Schwarz strike again:
X
12 X m(σ)
21
keT k21,T ≤ Rσ2 m(σ)dσ |Dσ e|2 .
dσ
σ∈E σ∈E
consh2
The main consequence of (3.63) is that
X XZ Z
m(σ)dσ Rσ2 ≤ C(size(T ))2 |H(u)(z)|2 dz = C(size(T ))2 |H(u)(z)|2 dz. (3.68) bouh
σ∈E σ∈E Vσ Ω

Then, one obtains

√ Z

21
keT k1,T ≤ Csize(T ) |H(u)(z)|2 dz .
Ω
eqesterrh1 R
This concludes the proof of (3.49)
eqesterr1 bouh
since u ∈ H 2 (Ω) implies Ω |H(u)(z)|2 dz < ∞. ellesterr
Estimate (3.51) followstesth2
from (3.68) in a similar manner as in the proof of Theorem 3.3. This concludes
the proof of Theorem 3.4.

h2plus Remark 3.13 (Generalizations)

1. By developping the method used to bound

testh2
of E ext ,
the consistency error on the flux on the elementsram
it
ram
is possible to replace, in Theorem 3.4, the hypothesis d K,σ ≥ ζdiam(K) in Definition 3.4 page
56 by the weaker hypothesis dσ ≥ ζdiam(σ) provided that Vσ is convex. Note also that, in this
case, the hypothesis xK ∈ K is not necessary, it suffices that xL − xK = dσ nK,σ , for all σ ∈ Eint ,
σ = K|L (for σ ∈ Eext , one always needs yσ − xK = dσ nK,σ ).
testh2
2. It is also possible to prove Theorem 3.4 if b 6= 0 or v 6= 0 (or, of course, b 6= 0 and v 6= 0). Indeed,
ellsolvar
if the solution, u, to (3.3) is not only in H 2 (Ω) but is also Lipschitz continuous on Ω (this is the
case if, for instance, there exists p > d such that u ∈ W 2,p (Ω)), the treatment of the consistency
ellesterr
error terms due to the terms involving b and v are exactly as in Theorem 3.3. If u is not Lipschitz
continuous on Ω, one has to dealcvnl
testh2
with the consistency error terms due to b andghv v similarly as in the
proof of Theorem 3.4 (see also Eymard, Gallouët and Herbin [1999] or Gallouët, Herbin
and Vignal [1999]).
 60

ellH ellH
It is also possible, essentially under Assumption 3.1 page 32, to obtain an Lq estimate of the error, for
sobolev
2 ≤ q < +∞ if d = 2, and for 1 ≤ q ≤ 6 if d = 3, see Coudière, Gallouët and Herbin [1998]. The
error estimate for the Lq norm is a consequence of the following lemma:

lq Lemma 3.5 (Discrete Sobolev Inequality) Let Ω be an open bounded polygonal subset of IR d and T
meshneumanmeshneuman
be a general finite volume mesh of Ω in the sense of definition 3.5 page 63, and let ζ > 0 be such that

∀K ∈ T , ∀σ ∈ EK , dK,σ ≥ ζdσ , and dK,σ ≥ ζdiam(K). (3.69) regularitebi

Xmesh Xmesh
Let be u ∈ X(T ) (see definition 3.2 page 39), then, there exists C > 0 only depending on Ω and ζ, such
that for all q ∈ [2, +∞), if d = 2, and q ∈ [2, 6], if d = 3,

kukLq (Ω) ≤ Cqkuk1,T , (3.70) ellinlq

defh10d defh10d
where k · k1,T is the discrete H01 norm defined in definition 3.3 page 40.
lq
Proof of Lemma 3.5
Let us first prove the two-dimensional case. Assume d = 2 and let q ∈ [2, +∞). Let d1 = (1, 0)t and
d2 = (0, 1)t ; for x ∈ Ω, let Dx1 and Dx2 be the straight lines going through x and defined by the vectors
d1 and d2 .
Let v ∈ X(T ). For all control volume K, one denotes by vK the value of v on K. For any control volume
K and a.e. x ∈ K, one has
X X
2
vK ≤ Dσ v χ(1)
σ (x) Dσ v χ(2)
σ (x), (3.71) lq1
σ∈E σ∈E
(1) (2)
where χσ and χσ are defined by

1 if σ ∩ Dxi =
6 ∅
χ(i)
σ (x) = for i = 1, 2.
0 if σ ∩ Dxi = ∅
lq1
Recall that Dσ v = |vK − vL |, if σ ∈ Eint , σ = K|L and Dσ v = |vK |, if σ ∈ Eext ∩ EK . Integrating (3.71)
over K and summing over K ∈ T yields
Z Z X X 
2
v (x)dx ≤ Dσ v χ(1)
σ (x) Dσ v χ(2)
σ (x) dx.
Ω Ω σ∈E σ∈E
(1) (2)
Note that χσ
(resp. χσ
) only depends on the second component x2 (resp. the first component x1 ) of
x and that both functions are non zero on a region the width of which is less than m(σ); hence
Z X 2
v 2 (x)dx ≤ m(σ)Dσ v . (3.72) poinbis
Ω σ∈E
poinbis
Applying the inequality (3.72) to v = |u|α sign(u), where u ∈ X(T ) and α > 1 yields
Z X 2
|u(x)|2α dx ≤ m(σ)Dσ v .
Ω σ∈E
α−1 α−1
Now, since |vK − vL | ≤ α(|uK | + |uL | )|uK − uL|, if σ ∈ Eint , σ = K|L and |vK | ≤ α(|uK |α−1 )|uK |,
if σ ∈ Eext ∩ EK ,
Z  12 X X
|u(x)|2α dx ≤α m(σ)|uK |α−1 Dσ u.
Ω K∈T σ∈EK
0 1 1
Using Hölder’s inequality with p, p ∈ IR + such that p + p0 = 1 yields that
 61

Z  12 X X
p1 X X |Dσ u|p
0

1
2α p(α−1)
|u(x)| dx ≤α |uK | m(σ)dK,σ p0
m(σ)dK,σ p0 .
Ω K∈T σ∈EK
d
K∈T σ∈EK K,σ
X
Since m(σ)dK,σ = 2m(K), this gives
σ∈EK
Z  21 1
Z

p1 X X |Dσ u|p
0

1
|u(x)|2α dx ≤ α2 p |u(x)|p(α−1) dx p0
m(σ)dK,σ p0 ,
Ω Ω d
K∈T σ∈EK K,σ
2α 2α
which yields, choosing p such that p(α − 1) = 2α, i.e. p = α−1 and p0 = α+1 ,

Z  2α
1
1 X X |Dσ u|p0
1
kukLq (Ω) = |u(x)|2α dx ≤ α2 p p0
m(σ)dK,σ p0 , (3.73) lq2
Ω d
K∈T σ∈EK K,σ
2 2
where q = 2α. Let r = p0 and r0 = 2−p0 , Hölder’s inequality yields
X X |Dσ u|p0 X X |Dσ u|2
p20 X X
1
p0
m(σ)dK,σ ≤ 2 m(σ)dK,σ m(σ)dK,σ r0 ,
d
K∈T σ∈EK K,σ
d
K∈T σ∈EK K,σ K∈T σ∈EK
lq2
replacing in (3.73) gives
1 2 1 1
kukLq (Ω) ≤ α2 p ( ) 2 (2m(Ω)) p0 r0 kuk1,T
ζ
ellinlq 1 1
and then (3.70) with, for instance, C = ( 2ζ ) 2 ((2m(Ω)) 2 + 1).
Let us now prove the three-dimensional case. Let d = 3. Using the same notations as in the two-
dimensional case, let d1 = (1, 0, 0)t , d2 = (0, 1, 0)t and d3 = (0, 0, 1)t ; for x ∈ Ω, let Dx1 , Dx2 and Dx3
be the straight lines going through x and defined by the vectors d1 , d2 and d3 . Let us again define the
(1) (2) (3)
functions χσ , χσ and χσ by

1 if σ ∩ Dxi 6= ∅
χ(i) (x) = for i = 1, 2, 3.
σ 0 if σ ∩ Dxi = ∅
Let v ∈ X(T ) and let A ∈ IR + such that Ω ⊂ [−A, A]3 ; we also denote by v the function defined on
[−A, A]3 which equals v on Ω and 0 on [−A, A]3 \ Ω. By the Cauchy-Schwarz inequality, one has:
Z A Z A
3
|v(x1 , x2 , x3 )| 2 dx1 dx2
−AZ −AZ
 A A  21 Z A Z A  12 (3.74) sob3d1
≤ |v(x1 , x2 , x3 )|dx1 dx2 |v(x1 , x2 , x3 )|2 dx1 dx2 .
−A −A −A −A

Now remark that

Z AZ A X Z A Z A X
|v(x1 , x2 , x3 )|dx1 dx2 ≤ Dσ v χ(3)
σ (x)dx1 dx2 ≤ m(σ)Dσ v.
−A −A σ∈E −A −A σ∈E

Moreover, computations which were already performed in the two-dimensional case give that
Z A Z A Z A Z A X X X 2
2
|v(x1 , x2 , x3 )| dx1 dx2 ≤ Dσ vχ(1)
σ (x) Dσ vχ(2)
σ (x)dx1 dx2 ≤ m(σx3 )Dσ v ,
−A −A −A −A σ∈E σ∈E σ∈E

where σx3 denotes the intersection of σ with the plane which contains the point (0, 0, x3 ) and is orthogonal
sob3d1
to d3 . Therefore, integrating (3.74) in the third direction yields:
 62

Z X  23
3
|v(x)| 2 dx ≤ m(σ)Dσ v . (3.75) sob3d2
Ω σ∈E
sob3d2
Now let v = |u|4 sign(u), since |vK − vL | ≤ 4(|uK |3 + |uL |3 )|uK − uL |, Inequality (3.75) yields:
Z h X X i 32
|u(x)|6 dx ≤ 4 |uK |3 Dσ um(σ) .
Ω K∈T σ∈EK
X
By Cauchy-Schwarz’ inequality and since m(σ)dK,σ = 3m(K), this yields
σ∈EK

√ X X m(σ)
kukL6 ≤ 4 3 (Dσ u)2 ,
dK,σ
K∈T σ∈EK

ellinlq √
and since dK,σ ≥ ζdσ , this yields (3.70) with, for instance, C = 4√ 3 .
ζ

poinbis
poin2 Remark 3.14 (Discrete Poincaré Inequality) In the above proof, Inequality ( 3.72) leads to another
thinpoin thinpoin
proof of some discrete Poincaré inequality (as in Lemma 3.1 page 40) in the two-dimensional case. Indeed,
let Ω be an open bounded polygonal
meshdirichlet
subset of IR 2 . Let T be an admissible finite volume mesh ofpoinbis
meshdirichlet
Ω in
the sense of Definition 3.1 page 37 (but more general
X meshes are possible). Let v ∈ X(T ). Then, ( 3.72),
the Cauchy-Schwarz inequality and the fact that m(σ)dσ = 2 m(Ω) yield
σ∈E

kvk2L2 (Ω) ≤ 2 m(Ω)kvk21,T .

A similar result holds in the three-dimensional case.
ellesterr ellesterr
norminf Corollary 3.1 testh2
Under the same assumptions and with the same notations as in Theorem 3.3 page 52, or
testh2
as in Theorem 3.4 page 56, and assuming that the mesh satisfies, for some ζ > 0, dK,σ ≥ ζdσ , for all
σ ∈ EK and for all control volume K, there exists C > 0 only depending on u, ζ and Ω such that

[1, 6] if d = 3,
keT kLq (Ω) ≤ Cqsize(T ); for any q ∈ (3.76) eqesterrq
[1, +∞) if d = 2,
m(K)
furthermore, there exists C ∈ IR + only depending on u, ζ, ζT = min{ size(T )2 , K ∈ T }, and Ω, such that

keT kL∞ (Ω) ≤ Csize(T )(| ln(size(T ))| + 1), if d = 2. (3.77)

keT kL∞ (Ω) ≤ Csize(T )2/3 , if d = 3. (3.78) eqesterrinf
norminf
Proof of Corollary 3.1
eqesterrh1 ellesterr testh2 ellinlq lq
Estimate (3.49) of Theorem 3.3 (or Theorem 3.4) andeqesterrinf
eqesterrq
Inequality (3.70) of Lemma 3.5 immediately yield
Estimate (3.76) in the case d = 2. Let us now prove (3.78). Remark that
 1  q1
keT kL∞ (Ω) = max{|eK |, K ∈ T } ≤ 2
keT kLq . (3.79) inf1
ζT size(T )

For d = 2, a study of the real function defined, for q ≥ 2, by q 7→ ln q + (1 − q2 ) ln h (with h = size(T ))

eqesterrq inf1
shows that its minimum is attained for q = −2 ln h, if ln h ≤ − 21 . And therefore (3.76) and (3.79) yield
eqesterrinf
(3.78). eqesterrq
The 3 dimensional case is an immediate consequence of (3.76) with q = 6.
 63

3.2 Neumann boundary conditions

nll
This section is devoted to the convergence proof of the finite volume scheme when Neumann boundary
conditions are imposed. The discretization of a general convection-diffusion
ell2Dgen
equation with Dirichlet,
Neumann and Fourier boundary conditions is considered in section 3.3 below, and the convection term is
largely studied in the previous section. Hence we shall limit here the presentation to the pure diffusion
operator. Consider the following elliptic problem:

−∆u(x) = f (x), x ∈ Ω, (3.80) nlldifstaf

with Neumann boundary conditions:

∇u(x) · n(x) = g(x), x ∈ ∂Ω, (3.81) nlldifstab

where ∂Ω denotes the boundary of Ω and n its unit normal vector outward to Ω.
The following assumptions are made on the data:

nllH Assumption 3.3

1. Ω is an open bounded polygonal connected subset of IR d , d = 2 or 3,

R R
2. g ∈ L2 (∂Ω), f ∈ L2 (Ω) and ∂Ω g(x)dγ(x) + Ω f (x)dx = 0.
nllH nlldifstaf
nlldifstab
Under Assumption
R 3.3, Problem (3.80), (3.81) has a unique (variational) solution, u, belonging to H 1 (Ω)
and such that Ω u(x)dx = 0. It is the unique solution of the following problem:
Z
1
u ∈ H (Ω), u(x)dx = 0, (3.82) nllspace
Ω
Z Z Z
∇u(x)∇ψ(x) = f (x)ψ(x)dx + g(x)γ(ψ)(x)dγ(x), ∀ψ ∈ H 1 (Ω). (3.83) nllsolvar
Ω Ω ∂Ω
1
Recall that γ is the “trace” operator from H 1 (Ω) to L2 (∂Ω) (or to H 2 (∂Ω)).

3.2.1 Meshes and schemes

msn
Admissible meshes
The definition of the scheme in the case of Neumann boundary conditions is easier, since the finite volume
scheme naturally introduces the fluxes on the boundaries in its formulation. Hence meshdirichlet
the class meshdirichlet
of admissible
meshes considered here is somewhat wider than the one considered in Definition 3.1 page 37, thanks to
the Neumann boundary conditions and the absence of convection term.

meshneuman Definition 3.5 (Admissible meshes) Let Ω be an open bounded polygonal connected subset of IR d ,
nlldifstaf
nlldifstab
d = 2, or 3. An admissible finite volume mesh of Ω for the discretization of Problem (3.80), (3.81), denoted
by T , is given by a family of “control volumes”, which are open disjoint polygonal convex subsets of Ω,
a family of subsets of Ω contained in hyperplanes of IR d , denoted by E (these are the “sides” of the
control volumes), with strictly positive (d − 1)-dimensional Lebesgue measure, andmeshdirichlet
a family of points of
meshdirichlet
Ω denoted by P satisfying properties (i), (ii), (iii) and (iv) of Definition 3.1 page 37.
meshdirichlet
meshdirichlet
The same notations as in Definition 3.1 page 37 are used in the sequel.

One defines the Xmesh

set X(T )Xmesh
of piecewise constant functions on the control volumes of an admissible mesh
as in Definition 3.2 page 39.
 64

defh1d Definition 3.6 (Discrete H 1 seminorm) Let Ω be an open boundedmeshneumanpolygonal subset of IR d , d = 2

or 3, and T an admissible finite volume mesh in the sense of Definition 3.5.
For u ∈ X(T ), the discrete H 1 seminorm of u is defined by
 X  21
|u|1,T = τσ (Dσ u)2 ,
σ∈Eint
meshdirichlet
meshdirichlet
where τσ = m(σ)
dσ and Eint are defined in Definition 3.1 page 37, uK is the value of u in the control volume
K and Dσ u = |uK − uL | if σ ∈ Eint , σ = K|L.

The finite volume scheme

meshneuman
Let T be an admissible mesh in the sense of Definition 3.5 . For K ∈ T , let us define:
Z
1
fK = f (x)dx, (3.84) nllcondsou
m(K) K
Z
1
gK = g(x)dγ(x) if m(∂K ∩ ∂Ω) 6= 0,
m(∂K ∩ ∂Ω) ∂K∩∂Ω (3.85) nllcondbor
gK = 0 if m(∂K ∩ ∂Ω) = 0.
nllcondsou nllcondbor
Recall that, in formula (3.84), m(K) denotes the d-dimensional Lebesgue measure of K, and, in (3.85),
m(∂K ∩ ∂Ω) denotes the (d − 1)-dimensional Lebesgue measure of ∂K ∩ ∂Ω. Note that g K = 0 if the
dimension of ∂K ∩ ∂Ωellschema than d −ellschemac
is lessellschemac 1. Let (uK )K∈T denote the discrete unknowns; the numerical
scheme is defined by (3.20)-(3.22) page 42, with b = 0 and v = 0. This yields:
X  
− τK|L uL − uK = m(K)fK + m(∂K ∩ ∂Ω)gK , ∀K ∈ T , (3.86) nllschema
L∈N (K)
meshdirichlet
meshdirichlet
meshneumanmeshneuman nllspace
(see the notations in Definitions 3.1 page 37 and 3.5 page 63). The condition (3.82) is discretized by:
X
m(K)uK = 0. (3.87) nllschemaO
K∈T
Xmesh Xmesh
Then, the approximate solution, uT , belongs to X(T ) (see Definition 3.2 page 39) and is defined by

uT (x) = uK , for a.e. x ∈ K, ∀K ∈ T .

nllschema nllschemaO
The following lemma gives existence and uniqueness of the solution of (3.86) and (3.87).
nllH meshneuman
nllexistu Lemma 3.6 Under Assumption 3.3.
nllcondsou
let T be an admissible mesh (see Definition 3.5) and {f
nllcondbor
∈ T },
K , K nllschemaO
nllschema
{gK , K ∈ T } defined by (3.84), (3.85). Then, there exists a unique solution (uK )K∈T to (3.86)-(3.87).
nllexistu
Proof of lemma 3.6
nllschema
Let N = card(T ). The equations (3.86) are a system of N equations with N unknowns, namely (uK )K∈T .
Ordering the unknowns (and the equations), this system can be written under a matrix form with a N ×N
matrix A. Using the connexity of Ω, the null space of this matrix is the set of “constant” vectors (that
isnllschema
uK = uL , for all K, L ∈ T ). Indeed, if fK = gK = 0 for all K ∈ T and {uK , K ∈ T } is solution of
nllschema
(3.86), multiplying (3.86) (for K ∈ T ) by uK and summing over K ∈ T yields
X
τσ (Dσ u)2 = 0,
σ∈Eint

where Dσ u = |uK − uL | if σ ∈ Eint , σ = K|L. This gives, thanks to the positivity of τσ and the connexity
of Ω, uK = uL , for all K, L ∈ T .
 65

nllschema
For general (fK )K∈T and (gK )K∈T , a necessary condition, in order that (3.86) has a solution, is that
X
(m(K)fK + m(∂K ∩ ∂Ω)gK ) = 0. (3.88) finetfin
K∈T

Since thenllschema
dimension of the null space of A is one, this condition is also a sufficient condition. Therefore,
finetfin
System (3.86) has a solution nllschemaO
if and only if (3.88) holds, and thisfinetfin
solution is unique up to an additive
constant. Adding
nllH
condition ( 3.87) yields uniqueness. Note that
nllexistu
( 3.88) holds thanks to the second item
of Assumption 3.3; this concludes the proof of Lemma 3.6.

3.2.2 Discrete Poincaré inequality

The proof of an error estimate, under a regularity assumption on the exact solution, and of a convergence
nllH
result, in the general case (under Assumption 3.3), requires a “discrete Poincaré” inequality as in the
case of the Dirichlet problem.

thinpoin1 Lemma 3.7 (Discrete mean Poincaré inequality) Let Ω be an open bounded polygonal connected
subset of IR d , d = 2 or 3. Then, there exists C ∈ IR + , only depending on Ω, such that for all
meshneumanmeshneuman
admissible
Xmesh Xmesh
meshes (in the sense of Definition 3.5 page 63), T , and for all u ∈ X(T ) (see Definition 3.2 page 39),
the following inequality holds:
Z
−1
kukL2(Ω) ≤ C|u|1,T + 2(m(Ω)) ( u(x)dx)2 ,
2 2
(3.89) ellinpoin1
Ω
defh1d
where | · |1,T is the discrete H 1 seminorm defined in Definition 3.6.
thinpoin1
Proof of Lemma 3.7
poincontr
The proof given here is a “direct proof”; another proof, by contradiction, is possible (see Remark 3.16).
Let T be an admissible mesh and u ∈ X(T ). Let mΩ (u) be the mean value of u over Ω, that is
Z
1
mΩ (u) = u(x)dx.
m(Ω) Ω

Since
kuk2L2(Ω) ≤ 2ku − mΩ (u)k2L2 (Ω) + 2(mΩ (u))2 m(Ω),
thinpoin1
proving Lemma 3.7 amounts to proving the existence of D ≥ 0, only depending on Ω, such that

ku − mΩ (u)k2L2 (Ω) ≤ D|u|21,T . (3.90) poineff

poineff
The proof of (3.90) may be decomposed into three steps (indeed, if Ω is convex, the first step is sufficient).
Step 1 (Estimate on a convex part of Ω)
Let ω be an open convex subset of Ω, ω 6= ∅ and mω (u) be the mean value of u on ω. In this step, one
proves that there exists C0 , depending only on Ω, such that
1
ku(x) − mω (u)k2L2 (ω) ≤ C0 |u|21,T . (3.91) pstep2
m(ω)
poineff thinpoin1
(Taking ω = Ω, this proves (3.90) and Lemma 3.7 in the case where Ω is convex.)

Noting that
Z Z Z
2 1

(u(x) − mω (u)) dx ≤ (u(x) − u(y))2 dy dx,
ω m(ω) ω ω
pstep2
(3.91) is proved provided that there exists C0 ∈ IR + , only depending on Ω, such that
 66

Z Z
(u(x) − u(y))2 dxdy ≤ C0 |u|21,T . (3.92) poinconv
ω ω

For σ ∈ Eint , let the function χσ from IR d × IR d to {0, 1} be defined by

χσ (x, y) = 1, if x, y ∈ Ω, [x, y] ∩ σ 6= ∅,
χσ (x, y) = 0, if x ∈
/ Ω or y ∈/ Ω or [x, y] ∩ σ = ∅.
(Recall that [x, y] = {tx + (1 − t)y, t ∈ [0, 1]}.) For a.e. x, y ∈ ω, one has, with Dσ u = |uK − uL | if
σ ∈ Eint , σ = K|L,
X
2
(u(x) − u(y))2 ≤ |Dσ u|χσ (x, y) ,
σ∈Eint

(note that the convexity of ω is used here) which yields, thanks to the Cauchy-Schwarz inequality,
X |Dσ u|2 X
(u(x) − u(y))2 ≤ χσ (x, y) dσ cσ,y−x χσ (x, y), (3.93) astcospoin
dσ cσ,y−x
σ∈Eint σ∈Eint

with
y−x
cσ,y−x = | · nσ |,
|y − x|
recall that nσ is a unit normal vector to σ, and that xK − xL = ±dσ nσ if σ ∈ Eint , σ = K|L. For a.e.
x, y ∈ ω, one has
X y−x
dσ cσ,y−x χσ (x, y) = |(xK − xL ) · |,
|y − x|
σ∈Eint

for some convenient control volumes K and L, depending on x, y and σ (the convexity of ω is used again
here). Therefore,
X
dσ cσ,y−x χσ (x, y) ≤ diam(Ω).
σ∈Eint
astcospoin
Thus, integrating (3.93) with respect to x and y in ω,
Z Z Z Z X |Dσ u|2
(u(x) − u(y))2 dxdy ≤ diam(Ω) χσ (x, y)dxdy,
ω ω ω ω σ∈Edσ cσ,y−x
int

which gives, by a change of variables,

Z Z Z X |Dσ u|2 Z

(u(x) − u(y))2 dxdy ≤ diam(Ω) χσ (x, x + z)dx dz. (3.94) astcospoin1
ω ω IR d dσ cσ,z ω
σ∈Eint

Noting that, if |z| > diam(Ω), χσ (x, x + z) = 0, for a.e. x ∈ Ω, and

Z
χσ (x, x + z)dx ≤ m(σ)|z · nσ | = m(σ)|z|cσ,z for a.e. z ∈ IR d ,
Ω
astcospoin1
therefore, with (3.94):
Z Z X m(σ)|Dσ u|2
(u(x) − u(y))2 dxdy ≤ (diam(Ω))2 m(BΩ ) ,
ω ω dσ
σ∈Eint
d
where BΩ denotes the ball of IR of center 0 and radius diam(Ω).
 67

poinconv pstep2
This inequality proves (3.92) and then (3.91) with C0 = (diam(Ω))2 m(BΩ ) (which only depends on Ω).
thinpoin1
Taking ω = Ω, it concludes the proof of Lemma 3.7 in the case where Ω is convex.

Step 2 (Estimate with respect to the mean value on pstep2

a part of the boundary)
In this step, one proves the same inequality than (3.91) but with the mean value of u on a (arbitrary)
part I of the boundary of ω instead of mω (u) and with a convenient C1 depending on I, Ω and ω instead
of C0 .
More precisely, let ω be a polygonal open convex subset of Ω and let I ⊂ ∂ω, with m(I) > 0 (m(I) is
the (d − 1)-Lebesgue measure of I). Assume that I is included in a hyperplane of IR d . Let γ(u) be the
“trace” of u on the boundary of ω, that is γ(u)(x) = uK if x ∈ ∂ω ∩ K, for K ∈ T . (If x ∈ K ∩ L, the
choice of γ(u)(x) between uK and uL does not matter). Let mI (u) be the mean value of γ(u) on I. This
step is devoted to the proof that there exists C1 , only depending on Ω, ω and I, such that

ku − mI (u)k2L2 (ω) ≤ C1 |u|21,T . (3.95) pstep3

For the sake of simplicity, only the case d = 2 is considered here. Since I is included in a hyperplane, it
may be assumed, without loss of generality, that I = {0} × J, with J ⊂ IR and ω ⊂ IR + × IR (one uses
here the convexity of ω).
Let α = max{x1 , x = (x1 , x2 )t ∈ ω} and a = (α, β)t ∈ ω. In the following, a is fixed. For a.e.
x = (x1 , x2 )t ∈ ω and for a.e. (for the 1-Lebesgue measure) y = (0, y)t ∈ I (with y ∈ J), one sets
z(x, y) = ta + (1 − t)y with t = x1 /α. Note that, thanks to the convexity of ω, z(x, y) = (z1 , z2 )t ∈ ω,
with z1 = x1 . The following inequality holds:

±(u(x) − γ(u)(y)) ≤ |u(x) − u(z(x, y))| + |u(z(x, y) − γ(u)(y))|.

In the following, the notation Ci , i ∈ IN? , will be used for quantities only depending on Ω, ω and I.
Let us integrate the above inequality over y ∈ I, take the power 2, from the Cauchy-Schwarz inequality,
an integration over x ∈ ω leads to
Z Z Z
2
(u(x) − mI (u))2 dx ≤ (u(x) − u(z(x, y)))2 dγ(y)dx
ω Z Z m(I) ω I
2
+ (u(z(x, y)) − u(y))2 dγ(y)dx.
m(I) ω I
Then,
Z
2
(u(x) − mI (u))2 dx ≤ (A + B),
ω m(I)
with, since ω is convex,
Z Z X
2
A= |Dσ u|χσ (x, z(x, y)) dγ(y)dx,
ω I σ∈Eint

and Z Z X
2
B= |Dσ u|χσ (z(x, y), y) dγ(y)dx.
ω I σ∈Eint

Recall that, for ξ, η ∈ Ω, χσ (ξ, η) = 1 if [ξ, η] ∩ σ 6= ∅ and χσ (ξ, η) = 0 if [ξ, η] ∩ σ = ∅. Let us now look
for some bounds of A and B of the form C|u|21,T .
The bound for A is easy. Using the Cauchy-Schwarz inequality and the fact that
X
cσ,x−z(x,y) dσ χσ (x, z(x, y)) ≤ diam(Ω)
σ∈Eint
η
(recall that cσ,η = | |η| · nσ | (for η ∈ IR 2 \ 0) gives
 68

Z Z X |Dσ u|2 χσ (x, z(x, y))

A ≤ C2 dxdγ(y).
ω I σ∈E cσ,x−z(x,y) dσ
int

Since z1 = x1 , one has cσ,x−z(x,y) = cσ,e , with e = (0, 1)t . Let us perform the integration of the right
hand side of the previous inequality, with respect to the first component of x, denoted by x 1 , first. The
result of the integration with respect to x1 is bounded by |u|21,T . Then, integrating with respect to x2
and y ∈ I gives A ≤ C3 |u|21,T .
In order to obtain a bound B, one remarks, as for A, that
Z Z X
|Dσ u|2 χσ (z(x, y), y)
B ≤ C4 dxdγ(y).
ω I cσ,y−z(x,y)dσ
σ∈Eint

In the right hand side of this inequality, the integration with respect to y ∈ I is transformed into an
integration with respect to ξ = (ξ1 , ξ2 )t ∈ σ, this yields (note that cσ,y−z(x,y) = cσ,a−y )
X |Dσ u|2 Z Z ψσ (x, ξ) |a − y(ξ)|
B ≤ C4 dxdγ(ξ),
dσ
σ∈Eint ω σ cI,a−y(ξ) |a − ξ|

where y(ξ) = sξ + (1 − s)a, with sξ1 + (1 − s)α = 0, and where ψσ is defined by

ψσ (x, ξ) = 1, if y(ξ) ∈ I and ξ1 ≤ x1

ψσ (x, ξ) = 0, if y(ξ) 6∈ I or ξ1 > x1 .
Noting that cI,a−y(ξ) ≥ C5 > 0, one deduces that
X |Dσ u|2 Z Z |a − y(ξ)|

B ≤ C6 ψσ (x, ξ) dx dγ(ξ) ≤ C7 |u|21,T ,
dσ σ ω |a − ξ|
σ∈Eint
pstep3
with, for instance, C7 = C6 (diam(ω))2 . The bounds on A and B yield (3.95).
poineff
Step 3 (proof of (3.90)) poineff
Let us now prove that there exists D ∈ IR + , only depending on Ω such that (3.90) hold. Since Ω is a
polygonal set (d = 2 or 3), there exists a finite number of disjoint convex polygonal sets, denoted by
{Ω1 , . . . , Ωn }, such that Ω = ∪ni=1 Ωi . Let Ii,j = Ωi ∩ Ωj , and B be the set of couples (i, j) ∈ {1, . . . , n}2
such that i 6= j and the (d − 1)-dimensional Lebesgue measure of Ii,j , denoted by m(Ii,j ), is positive.
Let mi denote the mean value of u on Ωi , i ∈ {1, . . . , n}, and mi,j denote the mean value of u on Ii,j ,
(i, j) ∈ B. (For σ ∈ Eint , in order that u be defined on σ, a.e. for the (d − 1)-dimensional Lebesgue
measure, let K ∈ T be a control volume such that σ ∈ EK , one sets u = uK on σ.) Note that mi,j = mj,i
for all (i, j) ∈ B.

Step 1 gives the existence of Ci , i ∈ {1, . . . , n}, only depending on Ω (since the Ωi only depend on Ω),
such that

ku − mi k2L2 (Ωi ) ≤ Ci |u|21,T , ∀i ∈ {1, . . . , n}, (3.96) morc

Step 2 gives the existence of Ci,j , i, j ∈ B, only depending on Ω, such that

ku − mi,j k2L2 (Ωi ) ≤ Ci,j |u|21,T , ∀(i, j) ∈ B.

Then, one has (mi − mi,j )2 m(Ωi ) ≤ 2(Ci + Ci,j )|u|21,T , for all (i, j) ∈ B. Since Ω is connected, the
above inequality yields the existence of M , only depending on Ω, such that |mi − mj | ≤ M |u|
morc 1,T
for
all (i, j) ∈ {1, . . . , n}2 , and therefore |mΩ (u) − mi | ≤ M |u| 1,T
poineff
for all i ∈ {1, . . . , n}. Then, ( 3.96) yields
the existence of D, only depending on Ω, such that (3.90) holds. This completes the proof of Lemma
thinpoin1
3.7.
 69

thinpoin1
An easy consequence of the proof of Lemma 3.7 is the following lemma. Although this lemma is not used
in the sequel, it is interesting in its own sake.

thinpoin2 Lemma 3.8 (Mean boundary Poincaré inequality) Let Ω be an open bounded polygonal connected
subset of IR d , d = 2 or 3. Let I ⊂ ∂Ω such that the (d − 1)- Lebesgue measure of I is positive. Then, there
exists C ∈ IR + , only depending on Ω and I, such thatXmesh
meshneumanmeshneuman
for all admissible
Xmesh
mesh (in the sense of Definition
3.5 page 63) T and for all u ∈ X(T ) (see Definition 3.2 page 39), the following inequality holds:

ku − mI (u)k2L2 (Ω) ≤ C|u|21,T

defh1d
where | · |1,T is the discrete H 1 seminorm defined in Definition 3.6 and mI (u) is the mean value of γ(u)
on I with γ(u) defined a.e. on ∂Ω by γ(u)(x) = uK if x ∈ σ, σ ∈ Eext ∩ EK , K ∈ T .

Note that this last lemma also gives as a by-product a discrete poincaré inequality in the case of a Dirichlet
dirich-part
boundary condition on a part of the boundary if the domain is assumed to be connex, see Remark 3.4.
thinpoin1thinpoin2
Finally, let us point out that a continuous version of lemmata 3.7 and 3.8 holds and that the proof is
similar and rather easier. Let us statethuh1
this continuous
thuh1
version which can be proved by contradiction or
with a technique similar to Lemma 3.4 page 49. The advantage of the latter is that it gives a more
explicit bound.

thinpoincont Lemma 3.9 Let Ω be an open bounded polygonal connected subset of IR d , d = 2 or 3. Let I ⊂ ∂Ω such
that the (d − 1)- Lebesgue measure of I is positive.
Then, there exists C ∈ IR + , only depending on Ω, and C̃ ∈ IR + , only depending on Ω and I, such that,
for all u ∈ H 1 (Ω), the following inequalities hold:
Z
kuk2L2(Ω) ≤ C|u|2H 1 (Ω) + 2(m(Ω))−1 ( u(x)dx)2
Ω
and

ku − mI (u)k2L2 (Ω) ≤ C̃|u|2H 1 (Ω) ,

R
where |·|H 1 (Ω) is the H 1 seminorm defined by |v|2H 1 (Ω) = k∇uk2(L2(Ω))d = Ω |∇v(x)|2 dx for all v ∈ H 1 (Ω),
and mI (u) is the mean value of γ(u) on I. Recall that γ is the trace operator from H 1 (Ω) to H 1/2 (∂Ω).

3.2.3 Error estimate

nllH meshneuman
Under Assumption 3.3, let T be an admissible
nllcondsou
nllcondbor nllexistu
mesh (see Definition 3.5) and {fK , K ∈nllschema
T }, {g K, K ∈ T }
nllschemaO
defined by (3.84), (3.85). By Lemma 3.6, there exists a unique solution (uK )K∈T to (3.86)-(3.87). Under
an additional regularity assumption on the exact solution, the following error estimate holds:
nllH nllH meshneumanmeshneuman
nllesterr Theorem 3.5 Under Assumption 3.3 page 63, let T be annllschema admissiblenllschemaO
mesh (see Definition 3.5 page 63)
nllcondsou nllcondbor
and h = size(T ). Let (uK )K∈T be the unique solution to (3.86) and
nllexistu
( 3.87) (thanks to ( 3.84) andXmesh
(3.85),
existence
Xmesh
and uniqueness of (u K ) K∈T is given in Lemma 3.6). Let u T ∈ X(T ) (see Definition 3.2 page
39) be defined by u T (x)
nllspacenllsolvar
= u K for a.e. x ∈ K, for all K ∈ T . Assume that the unique solution, u, to
Problem (3.82), (3.83) satisfies u ∈ C 2 (Ω).
Then there exists C ∈ IR + which only depends on u and Ω such that

kuT − ukL2 (Ω) ≤ Ch, (3.97) nqesterr

X Z
uL − u K 1
m(σ)dσ ( − ∇u(x) · nK,σ dγ(x))2 ≤ Ch2 . (3.98) nqesterr1
dσ m(σ) σ
σ=K|L∈Eint
 70

Recall that, in the above theorem, K|L denotes the element σ of Eint such that σ = ∂K ∩ ∂L, with K,
L∈T.
nllesterr
Proof of Theorem 3.5
Let CT ∈ IR be such that
X
u(xK )m(K) = 0,
K∈T

where u = u + CT .
Let, for each K ∈ T , eK = u(xK ) − uK , and eT ∈ X(T ) defined by eT (x) = eK for a.e. x ∈ K, for all
K ∈ T . Let us first prove the existence of C only depending on u and Ω such that

|eT |1,T ≤ Ch and keT kL2 (Ω) ≤ Ch. (3.99) nllestpre

nlldifstaf nlldifstaf nlldifstab nlldifstab
Integrating (3.80) page 63 over K ∈ T , and taking (3.81) page 63 into account yields:
X Z Z Z
∇u(x) · nK,σ dγ(x) = f (x)dx + g(x)dγ(x). (3.100) nller1
σ∈EK σ K ∂K∩∂Ω

For σ ∈ Eint such that σ = K|L, let us define the consistency error on the flux from K through σ by:
Z
1 u(xL ) − u(xK )
RK,σ = ∇u(x) · nK,σ dγ(x) − . (3.101) nller2
m(σ) σ dσ
nller2
Note that the definition of RK,σ remains with u instead of u in (3.101).
Thanks to the regularity of the solution u,
nller1 nller2
there exists C1 ∈ IR + , only depending on u, such that
nllschema
|RK,L | ≤ C1 h. Using (3.100), (3.101) and (3.86) yields
X
τK|L (eL − eK )2 ≤ dm(Ω)(C1 h)2 ,
K|L∈Eint
nllestpre
which gives the first part of (3.99). ellinpoin1
Thanks to the discrete Poincaré inequality (3.89) applied to the function eT , and since
X
m(K)eK = 0
K∈T
nllestpre
(which is the reason why eT was defined with u instead of u) one obtains the second part of (3.99), that
is the existence of C2 only depending on u and Ω such that
X
m(K)(eK )2 ≤ C2 h2 .
K∈T
nllestpre nqesterr
From (3.99), one deduces (3.97) from the fact that u ∈ C 1 (Ω). Indeed, let C2 be the Rmaximum value of
|∇u| in Ω. One has |u(x) − u(y)| ≤ C2 h, for all x, y ∈ K, for all K ∈ T . Then, from Ω u(x)dx = 0, one
deduces CT ≤ C2 h. Furthermore, one has
XZ X
(u(xK ) − u(x))2 dx ≤ m(K)(C2 h)2 = m(Ω)(C2 h)2 .
K∈T K K∈T

Then, noting that

XZ
kuT − (uK − u(x))2 dx
uk2L2 (Ω) =
K
XZ
K∈T
X
2 2
≤3 m(K)(eK ) + 3(CT ) m(Ω) + 3 (u(xK ) − u(x))2 dx
K∈T K∈T K
 71

nqesterr
yields (3.97).
nqesterr1
The proof of Estimate (3.98) is exactly the same as in the Dirichlet case. This property will be useful
in the study of the convergence of finite volume methods
casimplec
in the case of a system consisting of an elliptic
equation and a hyperbolic equation (see Section 7.3.6).

As for the Dirichlet problem, the hypothesis u ∈ C 2 (Ω) is notnramnecessary to obtain error estimates.
nllestpre nqesterr1
Assuming an additional assumption on the mesh (see Definition
ntesth2
3.7), Estimates ( 3.99) and (3.98) hold
2
under the
nqesterr
weaker assumption u ∈ H (Ω) (see Theorem 3.6 below). It is therefore also possible to obtain
(3.97) under the additional assumption that u is Lipschitz continuous.

nram Definition 3.7 (Neumann restricted admissible meshes) Let Ω be an open bounded polygonal
connected subset of IR d , d = 2 or 3. A restricted admissiblemeshneuman
mesh for the Neumann problem, de-
noted by T , is an admissible mesh in the sense of Definition 3.5 such that, for some ζ > 0, one has
dK,σ ≥ ζdiam(K) for all control volume K and for all σ ∈ EK ∩ Eint .
nllH nllH
ntesth2 Theorem 3.6 (H 2 regularity, Neumann nram
problem) Under Assumption 3.3 page 63, let XmeshT be an ad-
Xmesh
missible mesh in the sense of Definition 3.7 and h = size(T ). Let uT ∈ X(T ) (see Definition 3.2 page 39)
be the approximated solution nllschema
defined in ΩnllschemaO
by uT (x) = uK for a.e. x ∈ K, for all K ∈ T , where (uK )K∈T
nllcondsou nllcondbor
is the (unique) solution to (3.86) and
nllexistu
( 3.87) (thanks to ( 3.84) and (3.85), existence and uniqueness of
nllspacenllsolvar
(uK )K∈T is given in Lemma 3.6). Assume that the unique solution, u, of (3.82), (3.83) belongs to H 2 (Ω).
Let CT ∈ IR be such that X
u(xK )m(K) = 0 where u = u + CT .
K∈T

Let, for each control volume K ∈ T , eK = u(xK ) − uK , and eT ∈ X(T ) defined by eT (x) = eK for a.e.
x ∈ K, for all K ∈ T . nllestpre nqesterr1
Then there exists C, only depending on u, ζ and Ω, such that (3.99) and (3.98) hold.
ntesth2
Note that, in Theorem 3.6, the function eT is well defined, and the quantity “∇u · nσ ” is well defined on
uish2
σ, for all σ ∈ E (see Remark 3.12).
ntesth2
Proof of Theorem 3.6
testh2 testh2
The proof is very similar to that of Theorem 3.4 page 56, from which the same notations are used.
nram
There exists some C, depending only on the space dimension (d) and ζ (given in Definition 3.7), such
that, for all σ ∈ Eint ,
Z
h2
|Rσ |2 ≤ C |(H(u)(z)|2 dz, (3.102) nconsh2
m(σ)dσ Vσ
and therefore
X Z
m(σ)dσ Rσ2 ≤ Ch 2
|H(u)(z)|2 dz. (3.103) nbouh
σ∈Eint Ω
nconsh2 nbouh
The proof of (3.102) (from which (3.103) is an easy consequence) was already done in the proof of Theorem
testh2 nllestpre
3.4 (note that, here, there is no need to consider the case of σ ∈ Eext ). In order to obtain Estimate (3.99),
testh2
one proceeds as in Theorem 3.4. Recall
X
|eT |21,T ≤ Rσ |Dσ e|m(σ),
σ∈Eint

where |Dσ e| = |eK − eL | if σ ∈ Eint is such that σ = K|L; hence, from the Cauchy-Schwarz inequality,
one obtains that
 72

X
21 X m(σ)
12
|eT |21,T ≤ Rσ2 m(σ)dσ |Dσ e|2 .
dσ
σ∈Eint σ∈Eint
nbouh
Then, one obtains, with (3.103),
√ Z

21
|eT |1,T ≤ Ch |H(u)(z)|2 dz .
Ω
nllestpre nllestpre
This concludes the proof of the first part of
ellinpoin1
(3.99). The second part
nbouh
of (3.99) is a consequence of the
nqesterr1
discrete Poincaré inequality (3.89). Using (3.103) also easily leads
nqesterr
( 3.98). nllestpre
Note also that, if u is Lipschitz continuous,
nllesterr
Inequality (3.97) follows from the second part of (3.99) and
the definition of u as in Theorem 3.5. ntesth2
This concludes the proof of Theorem 3.6.
ntesth2 h2plus h2plus
Some generalizations of Theorem 3.6 are possible, as for the Dirichlet case, see Remark 3.13 page 59.

3.2.4 Convergence
nllH
A convergence result, under Assumption 3.3, may be proved without any regularity assumption on the
exact solution.
The proof of convergence uses the following preliminary inequality on the “trace” of an element of X(T )
on the boundary:

nlltrace Lemma 3.10 (Trace inequality) Let Ω be an open bounded polygonal connected subset of IR d , d = 2
or 3 (indeed, the connexity of Ω is not used in this lemma).
meshneumanmeshneuman
Let T Xmesh
Xmesh
be an admissible mesh, in the sense
of Definition 3.5 page 63, and u ∈ X(T ) (see Definition 3.2 page 39). Let uK be the value of u in the
control volume K. Let γ(u) be defined by γ(u) = uK a.e. (for the (d − 1)-dimensional Lebesgue measure)
on σ, if σ ∈ Eext and σ ∈ EK . Then, there exists C, only depending on Ω, such that

kγ(u)kL2 (∂Ω) ≤ C(|u|1,T + kukL2 (Ω) ). (3.104) nlltraceineq

Remark 3.15 The result stated in this lemma still holds if Ω is not assumed connected. Indeed, one
meshneumanmeshneuman
needs only modify (in an obvious way) the definition of admissible meshes (Definition 3.5 page 63) so as
to take into account non connected subsets.
nlltrace
Proof of Lemma 3.10
By compactness of the boundary of ∂Ω, there exists a finite number of open hyper-rectangles (d = 2 or
3), {Ri , i = 1, . . . , N }, and normalized vectors of IR d , {ηi , i = 1, . . . , N }, such that

 ∂Ω ⊂ ∪N i=1 Ri ,
ηi · n(x) ≥ α > 0 for all x ∈ Ri ∩ ∂Ω, i ∈ {1, . . . , N },

{x + tηi , x ∈ Ri ∩ ∂Ω, t ∈ IR + } ∩ Ri ⊂ Ω,

where α is some positive number and n(x) is the normal vector to ∂Ω at x, inward to Ω. Let {α i , i =
PN
1, . . . , N } be a family of functions such that i=1 αi (x) = 1, for all x ∈ ∂Ω, αi ∈ Cc∞ (IR d , IR + ) and
αi = 0 outside of Ri , for all i = 1, . . . , N . Let Γi = Ri ∩ ∂Ω; let us prove that there exists Ci only
depending on α and αi such that


kαi γ(u)kL2 (Γi ) ≤ Ci |u|1,T + kukL2 (Ω) . (3.105) trace1
nlltraceineq PN
The existence of C, only depending on Ω, such that (3.104) holds, follows easily (taking C = i=1 Ci ,
PN trace1
and using i=1 αi (x) = 1, note that α and αi depend only on Ω). It remains to prove (3.105).
 73

Let us introduce some notations. For σ ∈ E and K ∈ T , define χσ and χK from IR d × IR d to {0, 1}
by χσ (x, y) = 1, if [x, y] ∩ σ 6= ∅, χσ (x, y) = 0, if [x, y] ∩ σ = ∅, and χK (x, y) = 1, if [x, y] ∩ K 6= ∅,
χK (x, y) = 0, if [x, y] ∩ K = ∅.

Let i ∈ {1, . . . , N } and let x ∈ Γi . There exists a unique t > 0 such that x + tηi ∈ ∂Ri , let y(x) = x + tηi .
For σ ∈ E, let zσ (x) = [x, y(x)] ∩ σ if [x, y(x)] ∩ σ 6= ∅ and is reduced to one point. For K ∈ T , let
ξK (x), ηK (x) be such that [x, y(x)] ∩ K = [ξK (x), ηK (x)] if [x, y(x)] ∩ K 6= ∅.
One has, for a.e. (for the (d − 1)-dimensional Lebesgue measure) x ∈ Γi ,

X X
|αi γ(u)(x)| ≤ |αi (zσ (x))(uK − uL )|χσ (x, y(x)) + |(αi (ξK (x) − αi (ηK (x))uK |χK (x, y(x)),
σ=K|L∈Eint K∈T

that is,

|αi γ(u)(x)|2 ≤ A(x) + B(x) (3.106) trace2

with
X
A(x) = 2( |αi (zσ (x))(uK − uL )|χσ (x, y(x)))2 ,
σ=K|L∈Eint
X
B(x) = 2( |(αi (ξK (x)) − αi (ηK (x)))uK |χK (x, y(x)))2 .
K∈T

A bound on A(x) is obtained for a.e. x ∈ Γi , by remarking that, from the Cauchy-Schwarz inequality:
X |Dσ u|2 X
A(x) ≤ D1 χσ (x, y(x)) dσ cσ χσ (x, y(x)),
d σ cσ
σ∈Eint σ∈Eint

where D1 only depends on αi and cσ = |ηi · nσ |. (Recall that Dσ u = |uK − uL |.) Since
X
dσ cσ χσ (x, y(x)) ≤ diam(Ω),
σ∈Eint

this yields:
X |Dσ u|2
A(x) ≤ diam(Ω)D1 χσ (x, y(x)).
d σ cσ
σ∈Eint

Then, since Z
1
χσ (x, y(x))dγ(x) ≤ cσ m(σ),
Γi α
there exists D2 , only depending on Ω, such that
Z
A= A(x)dγ(x) ≤ D2 |u|21,T .
Γi

A bound B(x) for a.e. x ∈ Γi is obtained with the Cauchy-Schwarz inequality:

X X
B(x) ≤ D3 u2K χK (x, y(x))|ξK (x) − ηK (x)| |ξK (x) − ηK (x)|χK (x, y(x)),
K∈T K∈T

where D3 only depends on αi . Since

X Z
1
|ξK (x) − ηK (x)|χK (x, y(x)) ≤ diam(Ω) and χK (x, y(x))|ξK (x) − ηK (x)|dγ(x) ≤ m(K),
Γi α
K∈T
 74

there exists D4 , only depending on Ω, such that

Z
B= B(x)dγ(x) ≤ D4 kuk2L2 (Ω) .
Γi
trace2 trace1
Integrating (3.106) over Γi , the bounds on A and B lead (3.105) for some convenient Ci and it concludes
nlltrace
the proof of Lemma 3.10.

nlltraceineq Kolm
poincontr Remark Kolm
3.16 Using this “trace inequality” (3.104) and the Kolmogorov theorem (see Theorem 3.9
thinpoin1 thinpoin1
page 93, it is possible to prove Lemma 3.7 page 65 (Discrete Poincaré inequality) by way of contradic-
Rtion. Indeed, assume that there exists a sequence (un )n∈IN such that, for all n ∈ IN, kun kL2meshneuman
(Ω) = 1,
u (x)dx = 0, un ∈ X(Tn ) (where Tn is an admissible mesh in the sense of Definition 3.5) and
Ω n
1
|un |1,Tn ≤nllconv
n . Using the trace inequality, one proves that (un )n∈IN is relatively compact in L2 (Ω), as in
nllconv 2
Theorem 3.7 page 74. Then, one can assume R that un → u, in L (Ω),
R as n → ∞. The function u satisfies 1
kukL2(Ω) = 1, since kun kL2 (Ω) = 1, kolmh1
and Ω u(x)dx
kolmh1
= 0, since Ω un (x)dx = 0. Using |un |1,Tn ≤ n ,
a proof similar to that of Theorem 3.11 page 94, yields that Di u = 0, for all i ∈ {1, . . . , n} (even if
size(Tn ) 6→ 0, as n → ∞), where Di u is the derivative in the distribution sense with respect to xi of u.
RSince Ω is connected, one deduces that u is constant on Ω, but this is impossible since kukL2(Ω) = 1 and
Ω
u(x)dx = 0.

nllschema nllschemaO nllcondsou

now prove that the scheme (3.86) and (3.87), where (fK )K∈T and (gK )K∈T are given by (3.84)
Let usnllcondbor
and (3.85) is stable: the approximate solution given by the scheme is bounded independently of the mesh,
as we proceed to show.
nllH nllH
nllest Lemma 3.11 (Estimate for the Neumann problem) Under Assumption 3.3 page 63, let T be
meshneumanmeshneuman
an
nllschema
admissible mesh
nllschemaO
(in the sense of Definition 3.5 page 63). Let (u )
K K∈T
nllcondsou
be the
nllcondbor
unique solution to ( 3.86)
and (3.87), where (fK )K∈T and nllexistu
(gK )K∈T are given by (3.84) and (3.85);
Xmesh
the existence and uniqueness
of (uK )K∈T is given in Lemma 3.6. Let uT ∈ X(T ) (see Definition 3.2) be defined by uT (x) = uK for
a.e. x ∈ K, for all K ∈ T . Then, there exists C ∈ IR + , only depending on Ω, g and f , such that

|uT |1,T ≤ C, (3.107) nllestimate

defh1d defh1d
where | · |1,T is defined in Definition 3.6 page 64.
nllest
Proof of Lemma 3.11
nllschema
Multiplying (3.86) by uK and summing over K ∈ T yields
X X X
τK|L (uL − uK )2 = m(K)fK uK + uKσ gKσ m(σ), (3.108) nqesterr2a
K|L∈Eint K∈T σ∈Eext

where, for σ ∈ Eext , Knqesterr2a

nllestimate σ ∈ T is such nlltraceineq
that σ ∈ellinpoin1
E Kσ .
We get (3.107) from (3.108) using (3.104), (3.89) and the Cauchy-Schwarz inequality.
nllestimate
Using the estimate (3.107) on the approximate solution, a convergence result is given in the following
theorem.

nllconv Theorem 3.7 (Convergence

nllH nllH
in the case of the Neumannnllspace problem)
nllsolvar
Under Assumption 3.3 page 63, let u be the
meshneumanmeshneuman
unique solution to ( 3.82),( 3.83). Fornllschema
an admissible mesh (in
nllschemaO
the sense of Definition 3.5 page 63) Tnllcondsou
, let (uK )K∈T be the
nllcondbor
unique solution to ( 3.86) and ( 3.87) (where
(fK )K∈T and (g K )
nllexistu K∈T are given by ( 3.84) and ( 3.85), the
Xmesh
existence and uniqueness of (u K ) K∈T is given
in Lemma 3.6) and define uT ∈ X(T ) (see Definition 3.2) by uT (x) = uK for a.e. x ∈ K, for all K ∈ T .
Then,
 75

uT → u in L2 (Ω) as size(T ) → 0,
Z
|uT |21,T → |∇u(x)|2 dx as size(T ) → 0
Ω
and
γ(uT ) → γ(u) in L2 (Ω) for the weak topology as size(T ) → 0,
nlltrace
where the function γ(u) stands for the trace of u on ∂Ω in the sense given in Lemma 3.10 when u ∈ X(T )
1
and in the sense of the classical trace operator from H 1 (Ω) to L2 (∂Ω) (or H 2 (∂Ω)) when u ∈ H 1 (Ω).
nllconv
Proof of Theorem 3.7
Step 1 (Compactness) nllest
Denote by Y the set of approximate solutions uT for all admisible meshes T . Thanks to Lemma 3.11
ellinpoin1
and to the discrete Poincaré inequality (3.89), the set Y is bounded in L2 (Ω). Let us prove that Y is
relatively compact in L2 (Ω), and that, if (Tn )n∈IN is a sequence of admissible meshes such that size(Tn )
tends to 0 and uTn tends to u, in L2 (Ω),
Kolm
as n tends
kolmh1 kolmh1
to infinity, then u belongs to H 1 (Ω). Indeed, these
results follow from theorems 3.9 and 3.11 page 94, provided that there exists a real positive number C
only depending on Ω, f and g such that

kũT (· + η) − ũT k2L2 (IR d ) ≤ C|η|, for any admissible mesh T and for any η ∈ IR d , |η| ≤ 1, (3.109) cvneum1

and that, for any compact subset ω̄ of Ω,

kuT (· + η) − uT k2L2 (ω̄) ≤ C|η|(|η| + 2 size(T )), for any admissible mesh T
(3.110) cvneum2
and for any η ∈ IR d such that |η| < d(ω̄, Ωc ).
cvneum1
that ũT is defined by ũT (x) = uT (x) if x ∈ Ω and ũT (x) = 0 otherwise. In order to prove (3.109)
Recallcvneum2
and (3.110), define χσ from IR d × IR d to {0, 1} by χσ (x, y) = 1 if [x, y] ∩ σ 6= ∅ and χσ (x, y) = 0 if
[x, y] ∩ σ = ∅. Let η ∈ IR d \ {0}. Then:
X X
|ũ(x + η) − ũ(x)| ≤ χσ (x, x + η)|Dσ u| + χσ (x, x + η)|uσ |, for a.e. x ∈ Ω, (3.111) cvneum3
σ∈Eint σ∈Eext

where, for σ ∈ Eext , uσ = uK , and K is the control volume such that σ ∈ EK . Recall also that
cvneum2
Dσ u = |uK − uL |, if σ = K|L. Let us first prove Inequality (3.110). Let ω̄ be a compact subset of Ω. If
cvneum3
x ∈ ω̄ and h1dtot Ωc ), the second term of the right hand side of (3.111) is 0, aie
|η| < d(ω̄,h1dtot and the same proof as
in Lemma 3.3 page 44 gives, from an integration over ω̄ instead of Ω and from (3.33) with C = 2 since
[x, x + η] ⊂ Ω for x ∈ ω̄,

kuT (· + η) − uT k2L2 (ω̄) ≤ |u|21,T |η|(|η| + 2 size(T )). (3.112) preppar

cvneum1
In ordercvneum3
to prove (3.109), remark that the number of non zero terms in the second term of the right hand
side of (3.111) is, for a.e. x ∈ Ω, bounded by some real positive number, which only depends on Ω, which
can be taken, for instance, as the number of sides of Ω, denoted by N . Hence, with C1 = (N + 1)2 (which
only depends on Ω. Indeed, if Ω is convex, N = 2 is also convenient), one has

X X
|ũ(x + η) − ũ(x)|2 ≤ C1 ( χσ (x, x + η)|Dσ u|)2 + C1 χσ (x, x + η)u2σ , for a.e. x ∈ Ω. (3.113) cvneum4
σ∈Eint σ∈Eext

h1dtot h1dtot
Let us integrate this inequality over IR d . As seen in the proof of Lemma 3.3 page 44,
 76

Z X
2
χσ (x, x + η)|Dσ u| dx ≤ |u|21,T |η|(|η| + 2(N − 1)size(T ));
IR d σ∈E
int
nllest nllest
hence, by Lemma 3.11 page 74, there exists a real positive number C2 , only depending on Ω, f and g,
such that (if |η| ≤ 1)
Z X
2
χσ (x, x + η)|Dσ u| dx ≤ C2 |η|.
IR d σ∈E
int
cvneum4
Let us now turn to the second term of the right hand side of (3.113) integrated over IR d ;
Z X X


χσ (x, x + η)u2σ dx ≤ m(σ)|η|u2σ
IR d σ∈E σ∈Eext
ext
≤ kγ(uT )k2L2 (∂Ω) |η|;
nlltrace nllest ellinpoin1
therefore, thanks to Lemma 3.10, Lemma 3.11 and to the discrete Poincaré inequality (3.89), there exists
a real positive number C3 , only depending on Ω, f and g, such that
Z X

χσ (x, x + η)u2σ dx ≤ C3 |η|.
IR d σ∈E
ext
cvneum1
Hence (3.109) is proved for some real positive number C only depending on Ω, f and g.

Step 2 (Passage to the limit) nllspacenllsolvar

In this step, the convergence of uT to the solution of (3.82), (3.83) (in L2 (Ω) as size(T ) → 0) is first
proved. nllspacenllsolvar
Since the solution to (3.82), (3.83) is unique, and thanks to the compactness of the set Y described in
Stepnllspace sufficient to prove that, if uTn → u in L2 (Ω) and size(Tn ) → 0 as n → 0, then u is a solution
1, it isnllsolvar
to (3.82)-(3.83). nllschema
Let (Tn )n∈INnllschemaO
nllschemaO
be a sequence of admissible meshes and (uTn )n∈IN be the corresponding solutions to (3.86)-
(3.87) page 64 with T = Tn . Assume uTn → u in L2 (Ω) and size(T R n ) → 0 as n → 0. By Step 1, one has
H 1 (Ω) and since the mean value of uTn is zero,
u ∈ nllspace one
nllsolvar
also has Ω
u(x)dx = 0. Therefore, u is a solution
of (3.82). It remains to show that u satisfies (3.83). Since (γ(uTn ))n∈IN is bounded in L2 (∂Ω), one may
assume (up to a subsequence) that it converges to some v weakly in L2 (∂Ω). Let us first prove that
Z Z Z
− u(x)∆ϕ(x)dx + ∇ϕ(x) · n(x)v(x)dγ(x) = f (x)ϕ(x)dx
Ω ∂Ω Z Ω (3.114) cvneum5
+ g(x)ϕ(x)dγ(x), ∀ϕ ∈ C 2 (Ω),
∂Ω
nllsolvar
and then that u satisfies (3.83).

Let Tnllschema
be an admissible mesh, uT the corresponding approximate solution
nllschemaO
to the Neumann problem, given
nllcondsou nllcondbor
by (3.86) and (3.87), where (fK )K∈T and (gK )K∈T are given by (3.84) and (3.85) and let ϕ ∈ C 2 (Ω).
Let ϕK = ϕ(xK ), define ϕT by ϕT (x) = ϕK , for a.e. x ∈ K and for any control volume K, and
γ(ϕT )(x) = ϕK for a.e. x ∈ σ (for the (d − 1)-dimensional Lebegue measure), for any σ ∈ Eext and
control volume K such that σ ∈ EK .
nllschema
Multiplying (3.86) by ϕK , summing over K ∈ T and reordering the terms yields
X X Z Z
uK τK|L (ϕL − ϕK ) = f (x)ϕT (x)dx + γ(ϕT )(x)g(x)dγ(x). (3.115) cvneum6
K∈T L∈N (K) Ω ∂Ω

Using the consistency of the fluxes and the fact that ϕ ∈ C 2 (Ω), there exists C only depending on ϕ such
that
 77

X Z Z X
τK|L (ϕL − ϕK ) = ∆ϕ(x)dx − ∇ϕ(x) · n(x)dγ(x) + RK,L (ϕ),
L∈N (K) K ∂Ω∩∂K L∈N (K)

with RK,L = −RL,K , for all L ∈ N (K) and K ∈ T , and |RK,L | ≤ C4 m(K|L)size(T ), where C4 only
cvneum6
depends on ϕ. Hence (3.115) may be rewritten as
Z Z
− uT (x)∆ϕ(x)dx + ∇ϕ(x) · n(x)γ(uT )(x)dγ(x) + r(ϕ, T ) =
Ω Z∂Ω Z (3.116) cvneum7
f (x)ϕT (x)dx + γ(ϕT )(x)g(x)dγ(x),
Ω ∂Ω
where X
|r(ϕ, T )| = C4 |Dσ u|m(σ)size(T )
σ∈Eint
X m(σ)
12 X
21
≤ C4 |Dσ u|2 m(σ)dσ size(T )
dσ
σ∈Eint σ∈Eint
≤ C5 size(T ),
nllest
where C5 cvneum7
is a real positive number only depending on f , g, Ω and ϕ (thanks to Lemma 3.11).
cvneum5
Writing (3.116) with T = Tn and passing to the limit as n tends to infinity yields (3.114).
nllsolvar cvneum5
Let us now prove that u satifies (3.83). Since u ∈ H 1 (Ω), an integration by parts in (3.114) yields
Z Z
∇u(x) · ∇ϕ(x)dx + ∇ϕ(x) · n(x)(v(x) − γ(u)(x))dγ(x)
Ω Z ∂Ω Z (3.117) cvneum9
= f (x)ϕ(x)dx + g(x)ϕ(x)dγ(x), ∀ϕ ∈ C 2 (Ω),
Ω ∂Ω

where γ(u) denotes the trace of u on ∂Ω (which belongs

nllsolvar nllconv
to L2 (∂Ω)). In order to prove that u is solution
to (3.83) (this will conclude the proof of Theorem 3.7), it is sufficient, thanks to the density of C 2 (Ω) in
H 1 (Ω), to prove that v = γ(u) a.e. on ∂Ω (for the (d − 1) dimensional
cvneum9
Lebesgue measure on ∂Ω). Let us
now prove that v = γ(u) a.e. on ∂Ω by first remarking that (3.117) yields
Z Z
∇u(x) · ∇ϕ(x)dx = f (x)ϕ(x)dx, ∀ϕ ∈ Cc∞ (Ω),
Ω Ω

and therefore, by density of Cc∞ (Ω) in H01 (Ω),

Z Z
∇u(x) · ∇ϕ(x)dx = f (x)ϕ(x)dx, ∀ϕ ∈ H01 (Ω).
Ω Ω
cvneum9
With (3.117), this yields
Z
− ∇ϕ(x) · n(x)(v(x) − γ(u)(x))dγ(x) = 0, ∀ϕ ∈ C 2 (Ω) such that ϕ = 0 on ∂Ω. (3.118) cvneum10
∂Ω
cvneum10
There remains to show that the wide choice of ϕ in (3.118) allows to conclude v = γ(u) a.e. on ∂Ω (for
the (d − 1)-dimensional Lebesgue measure of ∂Ω). Indeed, let I be a part of the boundary ∂Ω, such that
I is included in a hyperplane of IR d . Assume that I = {0} × J, where J is an open ball of IR d−1 centered
|t|
on the origin. Let z = (a, z̃) ∈ IR d with a ∈ IR ?+ , z̃ ∈ IR d−1 and B = {(t, a−|t|
a y + a z̃); t ∈ (−a, a), y ∈ J};
assume that, for a convenient a, one has

a − |t| |t|
B ∩ Ω = {(t, y + z̃); t ∈ (0, a), y ∈ J}.
a a
 78

Let ψ ∈ Cc∞ (J), and for x = (x1 , y) ∈ IR × J, define ϕ1 (x) = −x1 ψ(y). Then,

∂ϕ1
ϕ1 ∈ C ∞ (IR d ) and = ψ on I.
∂n
(Recall that n is the normal unit vector to ∂Ω, outward to Ω.) Let ϕ2 ∈ Cc∞ (B) such that ϕ2 = 1 on
a neighborhood of {0} × {ψcvneum10
6= 0}, where {ψ 6= 0} = {x ∈ J; ψ(x) 6= 0}, and set ϕ = ϕ 1 ϕ2 ; ϕ is an
admissible test function in (3.118), and therefore
Z


ψ(y) γ(u)(0, y) − v(0, y) dy = 0,
J

which yields, since ψ is arbitrary in Cc∞ (J), v = γ(u) a.e. on I. Since J is arbitrary, this implies that
v = γ(u) a.e. on ∂Ω.
nllspace
nllsolvar
This conclude the proof of uT → u in L2 (Ω) as size(T ) → 0, where u is the solution to (3.82),(3.83).
Note also that the above proof gives (by way of contradiction) that γ(u T ) → γ(u) weakly in L2 (∂Ω), as
size(T ) → 0. nqesterr2a nllsolvar
Then, a passage to the limit in (3.108) together with (3.83) yields

|uT |21,T → k|∇u|k2L2 (Ω) , as size(T ) → 0.

nllconv
This concludes the proof of Theorem 3.7.
ellinlq
hypothesis “f ∈ L2 (Ω) g ∈
Note that, with some discrete Sobolev inequality (similar to (3.70)), the convplus
L2 (∂Ω)” may be relaxed in some way similar to that of Item 2 of Remark 3.7.

3.3 General elliptic operators

ell2Dgen
3.3.1 Discontinuous matrix diffusion coefficients
Meshes and schemes
Let Ω be an open bounded polygonal subset of IR d , d = 2 or 3. We are interested here in the discretiza-
tion of an elliptic operator with discontinuous matrix diffusion coefficients, which may appear in real
case problems such as electrical or thermal transfer problems or, more generally, diffusion problems in
heterogeneous media. In this case, the mesh is adaptedmeshdirichlet
to fit the discontinuities of the data. Hence
the definition of an admissible mesh given in Definition 3.1 must beell1D
adapted.ell1D
As an illustration, let us
consider here the following problem, which was studied in Section 2.3 page 21 in the one-dimensional
case:

−div(Λ∇u)(x) + div(vu)(x) + bu(x) = f (x), x ∈ Ω, (3.119) roueneq

u(x) = g(x), x ∈ ∂Ω, (3.120) rouencl
with the following assumptions on the data (one denotes by IR d×d the set of d × d matrices with real
coefficients):

hypdiscont Assumption 3.4

1. Λ is a bounded measurable function from Ω to IR d×d such that for any x ∈ Ω, Λ(x) is symmetric,
and that there exists λ and λ ∈ IR ?+ such that λξ · ξ ≤ Λ(x)ξ · ξ ≤ λξ · ξ for any x ∈ Ω and any
ξ ∈ IR d .
2. v ∈ C 1 (Ω, IR d ), divv ≥ 0 on Ω, b ∈ IR + .
3. f is a bounded piecewise continuous function from Ω to IR.
 79

4. g is such that there exists g̃ ∈ H 1 (Ω) such that γ(g̃) = g (a.e. on ∂Ω) and is a bounded piecewise
continuous function from ∂Ω to IR.
sdbc
(Recall that γ denotes the trace operator from H 1 (Ω) into L2 (∂Ω).) As in roueneq
hypdiscont
Section 3.1, under Assumption
rouencl
3.4, there exists a unique variational solution u ∈ H 1 (Ω) of Problem (3.119), (3.120). This solution
satisfies u = w + g̃, where g̃ ∈ H 1 (Ω) is such that γ(g̃) = g, a.e. on ∂Ω, and w is the unique function of
H01 (Ω) satisfying
Z  
Λ(x)∇w(x) · ∇ψ(x) + div(vw)(x)ψ(x) + bw(x)ψ(x) dx =
Z  Ω 
−Λ(x)∇g̃(x) · ∇ψ(x) − div(vg̃)(x)ψ(x) − bg̃(x)ψ(x) + f (x)ψ(x) dx, ∀ψ ∈ H01 (Ω).
Ω
roueneq rouencl
Let us now define an admissible mesh for the discretization of Problem (3.119)-(3.120).

rouenmesh Definition 3.8 (Admissible mesh for a general diffusion operator) Let Ω be an open bounded
polygonal subset of IR d , d = 2 or 3. An admissible finite volume mesh meshdirichlet
roueneq rouencl
for the discretization
meshdirichlet
of Problem
(3.119)-(3.120) is an admissible mesh T of Ω in the sense of Definition 3.1 page 37 where items (iv) and
(v) are replaced by the two following conditions:

(iv)’ The set T is such that

the restriction of g to each edge σ ∈ Eext is continuous.
For any K ∈ T , let ΛK denote the mean value of Λ on K, that is
Z
1
ΛK = Λ(x)dx.
m(K) K

There exists a family of points

P = (xK )K∈T such that xK = ∩σ∈EK DK,σ ∈ K,

where DK,σ is a straigth line perpendicular to σ with respect to the scalar product induced by Λ−1
K
such that DK,σ ∩ σ = DL,σ ∩ σ 6= ∅ if σ = K|L. Furthermore, if σ = K|L, let yσ = DK,σ ∩ σ(=
DL,σ ∩ σ) and assume that xK 6= xL .
(v)’ For any σ ∈ Eext , let K be the control volume such that σ ∈ EK and let DK,σ be the straight line
going through xK and orthogonal to σ with respect to the scalar product induced by Λ−1 K ; then,
there exists yσ ∈ σ ∩ DK,σ ; let gσ = g(yσ ).
meshdirichlet
meshdirichlet
The notations are are the same as those introduced in Definition 3.1 page 37.

We shall now define the discrete unknowns of the numerical scheme, with the same notations as in Section
vfquatre
3.1.2. As in the case of the Dirichlet problem, the primary unknowns (uK )K∈T will be used, which aim
to be approximations of the values u(xK ), and some auxiliary unknowns, namely the fluxes FK,σ , for
all K ∈ T and σ ∈ EK , and some (expected) approximation of u in σ, say uσ , for all σ ∈ E. Again,
these auxiliary unknowns are helpful to write the scheme, but they can be eliminated locally so that
the discrete equations will only be written with respect to the primary unknowns (uK )K∈T . For any
σ ∈ Eext , set uσ =rouencl
roueneq
g(yσ ). The finite volume scheme for the numerical
roueneq
approximation of the solution to
Problem (3.119)-(3.120) is obtained by integrating Equation (3.119) over each control volume K, and
approximating the fluxes over each edge σ of K. This yields
X X
FK,σ + vK,σ uσ,+ + m(K)buK = fK , ∀K ∈ T , (3.121) rouenvfn
σ∈EK σ∈EK

where
 80

R
vK,σ = σ v(x) · nK,σ dγ(x) (where nK,σ denotes the normal unit vector to σ outward to K); if σ =
Kσ,+ |Kσ,− , uσ,+ = uKσ,+ , where Kσ,+ is the upstream control volume, i.e. vK,σ ≥ 0, with K = Kσ,+ ;
if σ ∈ Eext , then uσ,+ = uK if vK,σ ≥ 0 (i.e. K is upstream to σ with respect to v), and uσ,+ = uσ
otherwise. R
FK,σ is an approximation of σ −ΛK ∇u(x) · nK,σ dγ(x); the approximation FK,σ is written with respect
to the discrete unknowns (uK )K∈T and (uσ )σ∈E . For K ∈ T and σ ∈ EK , let λK,σ = |ΛK nK,σ | (recall
that | · | denote the Euclidean norm).

• If xK 6∈ σ, a natural expression for FK,σ is then

uσ − u K
FK,σ = −m(σ)λK,σ .
dK,σ

Writing the conservativity of the scheme, i.e. FL,σ = −FK,σ if σ = K|L ⊂ Ω, yields the value of
uσ , if xL ∈
/ σ, with respect to (uK )K∈T ;

1 λK,σ λL,σ

uσ = λK,σ λL,σ
uK + uL .
+ dK,σ dL,σ
dK,σ dL,σ
uipud uipud
Note that this expression is similar to that of (2.26) page 22 in the 1D case.
• If xK ∈ σ, one sets uσ = uK .

Hence the value of FK,σ ;

• internal edges:
FK,σ = −τσ (uL − uK ), if σ ∈ Eint , σ = K|L, (3.122) rouenfluxni
where
λK,σ λL,σ
τσ = m(σ) if yσ 6= xK and yσ 6= xL
λK,σ dL,σ + λL,σ dK,σ
and
λK,σ
τσ = m(σ) if yσ 6= xK and yσ = xL ;
dK,σ

• boundary edges:

FK,σ = −τσ (gσ − uK ), if σ ∈ Eext and xK 6∈ σ, (3.123) rouenfluxnf

where

λK,σ
τσ = m(σ) ;
dK,σ
rouenvfn
if xK ∈ σ, then the equation associated to uK is uK = gσ (instead ofrouenvfn
that given by (3.121)) and
the numerical flux FK,σ is an unknown which may be deduced from (3.121).
rouenvfn rouenfluxnf
Remark 3.17 Note vfquatre
that if Λ = Id, then the scheme (3.121)-(3.123) is the same scheme than the one
described in Section 3.1.2.
 81

Error estimate
rouenesterr Theorem 3.8 hypdiscont
Let Ω be an open bounded polygonal subset of IR d , d = 2 or 3. Under Assumption 3.4, let u be the unique
roueneq rouencl
variationalroueneq
solution T be an admissible mesh for the discretization of
to Problem (3.119)-(3.120). Letrouenmesh
rouencl
Problem (3.119)-(3.120), in the sense of Definition 3.8. Let ζ1 and ζ2 ∈ IR + such that

ζ1 (size(T ))2 ≤ m(K) ≤ ζ2 (size(T ))2 ,

ζ1 size(T ) ≤ m(σ) ≤ ζ2 size(T ),
1 size(T ) ≤ dσ ≤ ζ2 size(T ).
ζvfquatreH2bis
Assuming moreover that
the restriction of f to K belongs to C(K), for any K ∈ T ;
the restriction of Λ to K belongs to C 1 (K, IR d×d ), for any Kroueneq
∈ T ; rouencl
the restriction of u (unique variational solution of Problem (3.119)-(3.120)) to K belongs to C 2 (K), for
any K ∈ T .
(Recall that C m (K, IR N ) = {v|K , v ∈ C m (IR d , IR N )} and C m (·) = C m (·, IR).)
rouenvfn rouenfluxnf
Then, there exists a unique family (uK )K∈T satisfying (3.121)-(3.123); furthermore, denoting by eK =
u(xK ) − uK , there exists C ∈ IR + only depending on ζ1 , ζ2 , γ = supK∈T (kD2 ukL∞(K) ) and δ = supK∈T
(kDΛkL∞(K) ) such that
X (Dσ e)2
m(σ) ≤ C(size(T ))2 (3.124) rouenesthuz
dσ
σ∈E

and
X
e2K m(K) ≤ C(size(T ))2 . (3.125) rouenestdeux
K∈T

Recall that Dσ e = |eL − eK | for σ ∈ Eint , σ = K|L and Dσ e = |eK | for σ ∈ Eext ∩ EK .
rouenesterr
Proof of Theorem 3.8
First, one ee1dell
may use Taylor expansions and the same technique
ell1D
as in the 1D case (see step 2 of the proof of
rouenfluxni rouenfluxnf
Theorem 2.3, Section 2.3) to R show that the expressions ( 3.122) and (3.123) are consistent approximations
of the exact diffusion flux σ −Λ(x)∇u(x) · nK,σ dγ(x), i.e. there exists C1 only depending on u and Λ
? ?
such that, for all σ ∈ E, with FK,σ = τσ (u(xL ) − u(xK )), if σ = K|L, and FK,σ = τσ (u(yσ ) − u(xK )), if
σ ∈ Eext ∩ EK ,
?
R
FK,σ − σ −Λ(x)∇u(x) · nK,σ dγ(x) = RK,σ ,
with |RK,σ | ≤ C1 size(T )m(σ).
There also exists C2 only depending on u and v such that, for all σ ∈ E,
R
vK,σ u(xKσ,+ ) − σ v · nK,σ u = rK,σ ,
with |rK,σ | ≤ C2 size(T )m(σ).
roueneq rouenvfn
Let us then integrate Equation (3.119) over each control volume, subtract to (3.121) and use the consis-
tency of the fluxes to obtain the following equation on the error:
 X X

 − G K,σ + vK,σ eσ,+ + m(K)beK =

σ∈E
X K σ∈E K


 (R K,σ + r K,σ ) + SK , ∀K ∈ T ,

σ∈EK

where GK,σ = τσ (eL − eK ), if σ = K|L, and GK,σ = τσ (−eK ), if σ ∈ Eext ∩ ERK , eσ,+ = eKσ,+ is the
error associated to the upstream control volume to σ and SK = b(m(K)u(xK ) − K u(x)dx) is such that
 82

ellesterr
|SK | ≤ m(K)C h, where C3 ∈ IR + only depends on u and b. Then, similarly to the proof of Theorem 3.3
ellesterr 3
page 52, let us multiply by eK , sum over K ∈ T , and use the conservativity of the scheme, which yields
K|L then RK,σ = −RL,σVF4
that if σ = rouenesthuz . A reordering of the summation over σ ∈ E yields the “discrete H01
estimate” (3.124). Then, following Herbin [1995], one shows the following discrete Poincaré inequality:

X X (Dσ e)2
e2K m(K) ≤ C4 m(σ), (3.126) ellinpoindis
dσ
K∈T σ∈E
rouenestdeux
where C4 only depends on Ω, ζ1 and ζ2 , which in turn yields the L2 estimate (3.125).

Remark 3.18 In the case where Λ is constant, or more generally, in the case where Λ(x) = λ(x)Id, where
thinpoin
λ(x) > 0, the proof of Lemma 3.1 is easily extended. However, for a general matrixrouenesthuz
Λ, the generalization
rouenestdeux
of this proof is not so clear; this is the ellinpoindisc
reason of theVF4
dependency of the estimates (3.124) and (3.125) on
ζ1 and ζ2 , which arises when proving (3.126) as in Herbin [1995].

3.3.2 Other boundary conditions

obt
The finite volume scheme may be used to discretize elliptic problems with Dirichlet or Neumann boundary
conditions, as we saw in the previous sections. It is also easily implemented in the case of Fourier (or
Robin) and periodic boundary conditions. The case of interface conditions between two geometrical
regions is also generally easy to implement; the purpose here is to present the treatment of some of
Angot
these
FH1
boundary and interface conditions. One may also refer to Angot [1989] and references therein, Fiard,
Herbin [1994] for the treatment of more complex boundary conditions and coupling terms in a system
of elliptic equations.
Let Ω be (for the sake of simplicity) the open rectangular subset of IR 2 defined by Ω = (0, 1) × (0, 2),
let Ω1 = (0, 1) × (0, 1), Ω2 = (0, 1) × (1, 2), Γ1 = [0, 1] × {0}, Γ2 = {1} × [0, 2], Γ3 = [0, 1] × {2},
Γ4 = {0} × [0, 2] and I = [0, 1] × {1}. Let λ1 and λ2 > 0, f ∈ C(Ω), α > 0, u ∈ IR, g ∈ C(Γ4 ), θ and
Φ ∈ C(I). Consider here the following problem (with some “natural” notations):

−div(λi ∇u)(x) = f (x), x ∈ Ωi , i = 1, 2, (3.127) pbeq

−λi ∇u(x) · n(x) = α(u(x) − u), x ∈ Γ1 ∪ Γ3 , (3.128) pbFourier

∇u(x) · n(x) = 0, x ∈ Γ2 , (3.129) pbNeumann

u(x) = g(x), x ∈ Γ4 , (3.130) pbDirichlet

(λ2 ∇u(x) · nI (x))|2 = (λ1 ∇u(x) · nI (x))|1 + θ(x), x ∈ I, (3.131) pbjumpflux

u|2 (x) − u|1 (x) = Φ(x), x ∈ I, (3.132) pbjumpun

t
where n denotes the unit normal vector to ∂Ω outward to Ω and nI = (0, 1) (it is a unit normal vector
to I). pbeq pbjumpun rouenmesh
Let T be an admissible mesh for the discretization of (3.127)-(3.132) in the sense of Definition 3.8. For the
pbeq
sake of simplicity, let us assume here that dK,σ > 0 for all K ∈ T , σ ∈ EK . Integrating Equation (3.127)
over each control volume K, and approximating the fluxes over each edge σ of K yields the following
finite volume scheme:
X
FK,σ = fK , ∀K ∈ T , (3.133) pbvfn
σ∈EK
R
where FK,σ is an approximation of σ −λi ∇u(x) · nK,σ dγ(x), with i such that K ⊂ Ωi .
Let NT = card(T ), NE = card(E), NE0 = card({σ ∈ E; σ 6⊂ ∂Ω ∪ I}), NEi = card({σ ∈ E; σ ⊂ Γi }), and
P4
NEI = card({σ ∈ E; σ ⊂ I}) (note that NE = NE0 + i=1 NEi + NEI ). Introduce the NT (primary) discrete
 83

unknowns (uK )K∈T ; note that the number of (auxiliary) unknowns of the type FK,σ is 2(NE0 + NEI ) +
P4 i
i=1 NE ; let us introduce the discrete unknowns (u σ )σ∈E , which aim to be approximations of u on σ.
pbjumpun
In order to take into account the jump condition (3.132), two unknowns of this type are necessary on
the edges σ ⊂ I, namely uσ,1 and uσ,2 . Hence the number of (auxiliary) unknowns of the type uσ is
P4
NE0 + i=1 NEi + 2NEI . Therefore, the total number of discrete unknowns is
4
X
Ntot = NT + 3NE0 + 4NEI +2 NEi .
i=1

Hence, it is convenient, in order to obtain a well-posed system, to write Ntot discrete equations. We
pbvfn
already have NT equations from (3.133). The expression of FK,σ with respect to the unknowns uK and
uσ is
uσ − u K
FK,σ = −m(σ)λi , ∀ K ∈ T ; K ⊂ Ωi (i = 1, 2), ∀ σ ∈ EK ; (3.134) pbvfnflux
dK,σ
P4 pbvfnflux
which yields 2(NE0 + NEI ) + i=1 NEi . (In (3.134), uσ stands for uσ,i if σ ⊂ I.)
Let us now take into account the various boundary and interface conditions:
pbFourier
• Fourier boundary conditions. Discretizing condition (3.128) yields

FK,σ = αm(σ)(uσ − u), ∀ K ∈ T , ∀ σ ∈ EK ; σ ⊂ Γ1 ∪ Γ3 , (3.135) pbvfnFourier

that is NE1 + NE3 equations.

pbNeumann
• Neumann boundary conditions. Discretizing condition (3.129) yields

FK,σ = 0, ∀ K ∈ T , ∀ σ ∈ EK ; σ ⊂ Γ2 , (3.136) pbvfnNeumann

that is NE2 equations.
pbDirichlet
• Dirichlet boundary conditions. Discretizing condition (3.130) yields

uσ = g(yσ ), ∀ σ ∈ E; σ ⊂ Γ4 , (3.137) pbvfnDirichl

that is NE4 equations.

• Conservativity of the flux. Except at interface I, the flux is continuous, and therefore

4
[
FK,σ = −FL,σ , ∀ σ ∈ E; σ 6⊂ ( Γi ∪ I) and σ = K|L, (3.138) pbvfnconserv
i=1

that is NE0 equations.

pbjumpflux
• Jump condition on the flux. At interface I, condition (3.131) is discretized into
Z
FK,σ + FL,σ = θ(x)ds, ∀ σ ∈ E; σ ⊂ I and σ = K|L; K ⊂ Ω2 , (3.139) pbvfnjumpflu
σ

that is NEI equations.

pbjumpun
• Jump condition on the unknown. At interface I, condition (3.132) is discretized into

uσ,2 = uσ,1 + Φ(yσ ), ∀ σ ∈ E; σ ⊂ I and σ = K|L. (3.140) pbvfnjumpun

that is another NEI equations.

 84

pbvfn pbvfnjumpun
Hence the total number of equations from (3.133) to (3.140) is Ntot , so that the numerical scheme can
be expected to be well posed.
pbeq pbjumpun
The pbvfn
finite volume scheme for the discretization of equations (3.127)-(3.132)
pbvfnjumpun
is therefore
ellschema
completely defined
ellschemad ellschemad
by (3.133)-(3.140). Particular cases nllschema
of thisnllschemaO
scheme are the schemes
nllschemaO
( 3.20)-( 3.23) page 42 (written for
Dirichlet boundary conditions) and (3.86)-(3.87) page 64 (written for Neumann boundary conditions and
no convection term) which were thoroughly studied in the two previous sections.

3.4 Dual meshes and unknowns located at vertices

scvfe
One of the principles of the classical finite volume method is to associate the discrete unknowns to the grid
cells. However, it is sometimes useful to associate the discrete unknowns with the vertices of the mesh;
for instance, the finite volume method may be used for the discretization of a hyperbolic equation coupled
systemes
with an elliptic equation (see Chapter 7). Suppose that an existing finite element code is implemented
for the elliptic equation and yields the discrete values of the unknown at the vertices of the mesh. One
might then want to implement a finite volume method for the hyperbolic equation with the values of the
unknowns at the vertices of the mesh. Note also that for some physical problems, e.g. the modelling of
two phase flow in porous media, the conservativity principle is easier to respect if the discrete unknowns
have the same location. For these various reasons, we introduce here some finite volume methods where
the discrete unknowns are located at the vertices of an existing mesh.
For the sake of simplicity, the treatment of the boundary conditions will be omitted here. Recall that
the construction of a finite volume method is carried out (in particular) along the following principles:

1. Divide the spatial domain in control volumes,

2. Associate to each control volume and, for time dependent problems, to each discrete time, one
discrete unknown,
3. Obtain the discrete equations (at each discrete time) by integration of the equation over the control
volume and the definition of one exchange term between two (adjacent) control volumes.

Recall, in particular, that the definition of one (and one only) exchange term between two control volumes
is important; this is called the property of conservativity of a finite volume method. The aim here is
to present finite volume methods for which the discrete unknowns are located at the vertices of the
mesh. Hence, to each vertex must correspond a control volume. Note that these control volumes may be
somehow “fictive” (see the next section); the important issue is to respect the principles given above in
the construction of the finite volume scheme. In the three following sections, we shall deal with the two
cvfe3D
dimensional case; the generalization to the three-dimensional case is the purpose of section 3.4.4.

3.4.1 The piecewise linear finite element method viewed as a finite volume
method
lfefv
We consider here the Dirichlet problem. Let Ω be a bounded open polygonal subset of IR 2 , f and g be
some “regular” functions (from Ω or ∂Ω to IR). Consider the following problem:

−∆u(x) = f (x), x ∈ Ω,
(3.141) rouene2D
u(x) = g(x), x ∈ ∂Ω.
rouene2D
Let us show that the “piecewise linear” finite element method for the discretization of ( 3.141) may be
a kind of finite volume method. Let M be a finite element mesh of Ω, consisting of triangles
viewed asciarlet
(see e.g. Ciarlet, P.G. [1978] for the conditions on the triangles), and let V ⊂ Ω be the set of vertices
of M. For K ∈ V (note that here K denotes a point of Ω), let ϕK be the shape function associated to
K in the piecewise linear finite element method for the mesh M. We remark that
 85

X
ϕK (x) = 1, ∀x ∈ Ω,
K∈V

and therefore
XZ
ϕK (x)dx = m(Ω) (3.142) consglo
K∈V Ω

and
X
∇ϕK (x) = 0, for a.e.x ∈ Ω. (3.143) rouencruu
K∈V

Using the latter equality, the discrete finite element equation associated to the unknown u K , if K ∈ Ω,
can therefore be written as
XZ Z
(uL − uK )∇ϕL (x) · ∇ϕK (x)dx = f (x)ϕK (x)dx.
L∈V Ω Ω

Then the finite element method may be written as

X Z
−τK|L (uL − uK ) = f (x)ϕK (x)dx, if K ∈ V ∩ Ω,
L∈V Ω

uK = g(K), if K ∈ V ∩ ∂Ω,
with
Z
τK|L = − ∇ϕL (x) · ∇ϕK (x)dx.
Ω
Under this form, the finite element method may be viewed as a finite volume method,consglo except that there
are no “real” control volumes associated to the vertices of M. Indeed, thanks to ( 3.142), the control
volume associated to K may be viewed as the support of ϕK “weighted” F1
by ϕK . This F2interpretation of
the finite
EG
element method as a finite volume method was also used in Forsyth [1989], Forsyth [1991]
and Eymard and Gallouët [1993] in order to design a numerical scheme for a transport equation for
which theHLvelocity field is the gradient of the pressure, which is itself the solution to an elliptic equation
(see also Herbin and Labergerie [1997] for numerical tests). This method is often referred to as the
”control volume finite element” method.
In this finite volume interpretation of the finite element scheme, the notion of “consistency of the fluxes”
does not appear. This notion of consistency, however, seems to be an interesting tool in the study of the
“classical” finite volume schemes.
Note that the (discrete) maximum principle is satisfied with this scheme if only if the transmissibilities
τK|L are nonnegative (for all K, L ∈ V with K ∈ Ω) ; this is the case under the classical Delaunay
condition; this condition states that the (interior of the) circumscribed circle (or sphere in the three
dimensional case) of any triangle (tetrahedron in the three dimensional case) of the mesh does not
contain any element of V. This is equivalent, in the case of two dimensional triangular meshes, to the
Delaunay fact that the sum of two opposite angles facing a common edge is less or equal π.

3.4.2 Classical finite volumes on a dual mesh

Let M be a mesh of Ω (M may consist of triangles, but it is not necessary) and V be the set of vertices
of M. In order to associate to each vertex (of M) a control volume (such that the whole spatial domain
is the “disjoint union” of the control volumes), a possibility is to construct a “dual mesh” which will be
meshdirichlet
meshdirichlet
denoted by T . In order for this mesh to be admissible in the sense of Definition 3.1 page 37, a simple way
 86

evoronoi evoronoi
is to use the Voronoı̈ mesh defined with V (see Example 3.2 page 39).nicolbook
For a description of the Delaunay-
Voronoı̈ discretization and its use for covolume methods, we refer to Nicolaides [1993] (and
ellschema
references
ellschemad ellschemad
therein). In order to write the “classical” finite volume scheme with this mesh (see (3.20)-( 3.23)
evoronoi
page 42),
a slight modification is necessary at the boundary for some particular M (see Example 3.2); this method
is denoted CFV/DM (classical finite volume on dual mesh); it is conservative, the numerical fluxes are
consistent, and the transmissibilities are nonnegative. Hence, the convergence results and error estimates
ellcvgce ellcvgceellesterr ellesterr
which were studied in previous sections hold (see, in particular, theorems 3.1 page 46 and 3.3 page 52).

A case of particular interest is found when the primal mesh (that is M) consists in triangles with acute
angles. One uses, as dual mesh, the
meshdirichlet
Voronoı̈ mesh defined with V. Then, the dual mesh is admissible in
meshdirichlet
the sense of Definition 3.1 page 37 and is constructed with the orthogonal bisectors of the edges of the
elements of M, parts of these orthogonal bisectors (and parts of ∂Ω) give the boundaries to the control
volumes of the dual mesh. In this case, the CFV/DM scheme is “close” to the piecewise linear finite
element scheme on the primal mesh. Let us elaborate on this point.
For K ∈ V, let K also denote the control volume (of the dual mesh) associated to K (in the sequel, the
sense of “K”, which denotes vertex or control volume, will not lead any confusion) and let ϕ K be the
shape function associated to the vertex K (in the piecewise linear finite element associated to M). The
term τK|L (ratio between the length of the edge K|L and the distance between vertices), which is used
in the finite volume scheme, verifies
Z
τK|L = − ∇ϕK (x) · ∇ϕL (x)dx.
Ω

The CFV/DM scheme (finite volume scheme on the dual mesh) writes
X Z
− τK|L (uL − uK ) = f (x)dx, if K ∈ V ∩ Ω,
L∈N (K) K

uK = g(K), if K ∈ V ∩ ∂Ω,
where K stands for an element of V or for the control volume (of the dual mesh) associated to this point.
The finite element scheme (on the primal mesh) writes
X Z
− τK|L (uL − uK ) = f (x)ϕK (x)dx, if K ∈ V ∩ Ω,
L∈N (K) Ω

uK = g(K), if K ∈ V ∩ ∂Ω.
Therefore, the only difference between the finite element and finite volume schemes is in the definition
of the right hand sides. Note that these right hand sides may be quite different. Consider for example a
node K whichcfvdm
is the vertex of four identical triangles featuring an angle of π2 at the vertex K, as depicted
in Figure 3.4, and denote by a the area of each of these triangles.
Then, for f ≡ 1, the right hand side computed for the discrete equation associated to the node K is equal
to a in the case of the finite element (piecewise linear finite element) scheme, and equal to 2a for the
dual mesh finite volume (CFV/DM) scheme. Both schemes may be shown to converge, by using finite
volume techniques for the CFV/DM scheme ciarlet
(see previous sections), and finite element techniques for the
piecewise linear finite element (see e.g.Ciarlet, P.G. [1978]).
Let us now weaken the hypothesis that all angles of the triangles of the primal mesh M are acute to the
so called Delaunay condition and the additional assumption
Vanselow
that an angle of an element of M is less or
equal π/2 if its opposite edge lies on ∂Ω (see e.g. Vanselow [1996]). Under this new assumption the
schemes (piecewise linear finite element finite element and CFV/DM with the Voronoı̈ mesh defined with
V) still lead to the same transmissibilities and still differ in the definition of the right hand sides.
Recall that the Delaunay condition states that no neighboring element (of M) is included in the circum-
scribed circle of an arbitrary element of M. This is equivalent to saying that the sum of two opposite
 87

Figure 3.4: An example of a triangular primal mesh (solid line) and a dual Voronoı̈ control volume
(dashed line) cfvdm

fig.delaunay
angles to an edge is less
meshdirichlet
or equal π. As shown in Figure 3.5, the dual mesh is still admissible in the sense
meshdirichlet
of Definition 3.1 page 37 and is still constructed with the orthogonal bisectors of the edges of the elements
of M, parts of these orthogonal bisectors (and parts of ∂Ω) give the boundaries to the control volumes
fig.delaunay
of the dual mesh (see Figure (3.5)) is not the case when M does not satisfy the Delaunay condition.

B
B

K L K L

A A

Delaunay case Non Delaunay case

Figure 3.5: Construction of the Voronoı̈ dual cells (dashed line) in the case of a triangular primal mesh
(solid line) with and without the Delaunay condition fig.delaunay

Consider now a primal mesh, M, consisting of triangles, but which does not satisfy the Delaunay condition
and let the dual mesh be the Voronoı̈ mesh defined with V. Then, the two schemes, piecewise linear finite
element and CFV/DM are quite different. If the Delaunay condition does not hold say between the
d and KBL
angles KAL d (the triplets (K, A, L) and (K, B, L) defining two elements of M), the sum of
R
these two angles is greater than π and the transmissibility τK|L = − Ω ∇ϕK (x) · ∇ϕL (x)dx between the
two control volumes associated respectively to K and L becomes negative with the piecewise linear finite
element scheme; there is no transmissibility between A and B (since A and B do not belong to a common
element of M). Hence the maximum principle is no longer respected for the finite element scheme, while
it remains valid for the CFV/DM finite volume scheme. This is due to the fact that the CFV/DM scheme
 88

allows an exchange term between A and B, with a positive transmissibility (and leads to no exchange
term between K and L), while the finite element scheme does not. Also note also that the common edge
to the control volumes (of the dual mesh) associated to A and B is not a part of an orthogonal bisector
of an edge of an element of M (it is a part of the orthogonal bisector of the segment [A, B]).

To conclude this section, note that an admissible mesh for the classical finite volume is generally not a
dual mesh of a primal triangular
VF4
mesh consisting of triangles (for instance, the general triangular meshes
which are considered in Herbin [1995] are not dual meshes to triangular meshes).

3.4.3 “Finite Volume Finite Element” methods

The “finite volume finite element” method for elliptic problems also uses a dual mesh T constructed from
a finite element primal mesh, such that each cell of T is associated with a vertex of the primal mesh M.
Let V again denote the set of vertices of M. As in the classical finite volume method, the conservation
law is integrated over each cell of the (dual) mesh. Indeed, this integration is performed only if the cell
is associated to a vertex (of the primal mesh) belonging to Ω.
rouene2D
Let us consider Problem (3.141). Integrating the conservation law over KP , where P ∈ V ∩ Ω and KP is
the control volume (of the dual mesh) associated to P yields
Z Z
− ∇u(x) · nP (x)dγ(x) = f (x)dx.
∂KP KP

(Recall that nP is the unit normal vector to ∂KP outward to KP .) Now, following
P the idea of finite
element methods, the function u is approximated by a Galerkin expansion M ∈V M ϕM , where the
u
functions ϕM are the shape functions of the piecewise linear finite element method. Hence, the discrete
unknowns are {uP , P ∈ V} and the scheme writes
X Z  Z
− ∇ϕM (x) · nP (x)dγ(x) uM = f (x)dx, ∀ P ∈ V ∩ Ω, (3.144) cvfe
M ∈V ∂KP KP

uP = g(P ), ∀ P ∈ V ∩ ∂Ω.
cvfe
Equations (3.144) may also be written under the conservative form
X Z
EP,Q = f (x)dx, ∀ P ∈ V ∩ Ω, (3.145) cvfe1
Q∈V KP

uP = g(P ), ∀ P ∈ V ∩ ∂Ω, (3.146) cvfe2

where
X Z
EP,Q = − ∇ϕM (x) · nP (x)dγ(x). (3.147) cvfe3
M ∈V ∂KP ∩∂KQ

Note that EQ,P = −EP,Q . Unfortunately, the exchange term EP,Q between P and Q is not, in general,
a function of the only unknowns uP and uQ (this property was used, in the previous sections,
cvfe
to obtain
rouencruu
convergence results of finite volume schemes). Another way to write (3.144) is, thanks to (3.143),
X Z  Z
− ∇ϕQ (x) · nP (x)dγ(x) (uQ − uP ) = f (x)dx, ∀ P ∈ V ∩ Ω.
Q∈V ∂KP KP
R 
Hence a new exchange term from P to Q might be ĒP,Q = − ∂KP ∇ϕQ (x) · nP (x)dγ(x) (uQ − uP ) and
the scheme is therefore conservative if ĒP,Q = −ĒQ,P . Unfortunately, this is not the case for a general
dual mesh.
 89

La
There are several ways of constructing a dual mesh from a primal mesh. A common way (see e.g. Fezoui,
Lanteri, Larrouturou and Olivier [1989]) is to take a primal mesh (M) consisting of triangles and
to construct the dual mesh with the medians (of the triangles of M), joining the centers of gravity of
the triangles to the midpoints of the edges of the primal mesh. The main interest of this way is that the
resulting scheme (called FVFE/M below, Finite Volume Finite Element with Medians) is very close to
the piecewise
cvfe1 cvfe3
linear finite element scheme associated to M. Indeed the FVFE/M scheme is defined by
(3.145)-(3.147) while the piecewise linear finite element scheme writes
X Z
EP,Q = f (x)ϕP (x)dx, ∀ P ∈ V ∩ Ω,
Q∈V Ω

uP = g(P ), ∀ P ∈ V ∩ ∂Ω,
cvfe3
where EP,Q is defined by (3.147).
These two schemes only differ by the right hand sides and, in fact, these right hand sides are “close” since
Z
m(KP ) = ϕP (x)dx, ∀ P ∈ V.
Ω
R
This is due to the fact that T ϕP (x)dx = m(T )/3 and m(KP ∩ T ) = m(T )/3, for all T ∈ M and all
vertex P of T .
Thus, convergence properties of the FVFE/M scheme can be proved by using the finite element techniques.
Recall however that the piecewise linear finite element scheme (and the FVFE/M scheme) does not satisfy
the (discrete) maximum principle if M does not satisfy the Delaunay condition.

There are other means to construct a dual mesh starting from a primal triangular mesh. One of them is
the Voronoı̈ mesh associated to the vertices of the primal mesh, another possibility is to join the centers
of gravity; in the latter case, the control volume associated to a vertex, say S, of the primal mesh is then
limited by the lines joining the centers of gravity of the neighboring triangles of which S is Barth
a vertex (with
some convenient modification for the vertices which are on the boundary of Ω). See also Barth [1994]
for descriptions of dual meshes.
Note that the proof of convergence which we designed for finite volume with admissible meshes does not
generalize to any “FVFE” (Finite Volume Finite Element) method for several reasons. In particular,
since the exchange term between P and Q (denoted by EP,Q ) is not, in general, a function of the only
unknowns uP and uQ (and even if it is the transmissibilities may become negative) and also since, as in
the case of the finite element method, the concept of consistency of the fluxes is not clear with the FVFE
schemes.

3.4.4 Generalization to the three dimensional case

cvfe3D
The methods described in the three above sections generalize to the three-dimensional case, in particular
when the primal mesh is a tetrahedral mesh. With such a mesh, the Delaunay condition no longer ensures
the non negativity of the transmissibilities in the case of the piecewise linear finite element method. It
is however possible to construct a dual mesh (the “three-dimensional Voronoı̈” mesh) to a Delaunay
triangulation such that the FVFE scheme leads to positive transmissibilities, and therefore such that the
Putti
maximum principle holds, see Cordes and Putti [1998].
Note that the theoretical results (convergence and vfquatre
error estimate) which were shown
vfquatrenll nll
for the classical finite
volume method on an admissible mesh (sections 3.1.2 page 37 and 3.2 page 63) still hold for CFV/DM
in three-dimensional, since the dual mesh is admissible.
 90

3.5 Mesh refinement and singularities

meshref
Some problems involve singular source terms. In the case of petroleum engineering for instance, one may
model (in two space dimensions) the well with a Dirac measure. Other problems may require a better
precision of some unknown in certain areas. This section is devoted to the treatment of this kind of
problem, either with an adequate treatment of the singularity or by mesh refinement.

3.5.1 Singular source terms and finite volumes

It is possible to take into account, in the discretization with the finite volume method, the singularities
of the solution of an elliptic problem. A common example is the study of wells in petroleum engineering.
As a model example we can consider the following problem, which appears, for instance, in the study of
a two phase flow in a porous medium. Let B be the ball of IR 2 of center 0 and radius rp (B represents a
well of radius rp ). Let Ω = (−R, R)2 be the whole domain of simulation; rp is of the order of 10 cm while
R can be of the order of 1 km for instance. An approximation to the solution of the following problem
is sought:

−div(∇u)(x) = 0, x ∈ Ω \ B,
u(x) = Pp , x ∈ ∂B, (3.148) sing
“BC”on ∂Ω,
where “BC” stands for some “smooth” boundary conditions on ∂Ω (for instance, Dirichlet or Neumann
condition). This system is a mathematical model (under convenient assumptions. . . ) of the two phase
flow problem, with u representing the pressure of the fluid and Pp an imposed pressure at the well. In
sing
order to discretize (3.148) with the finite volume method, a mesh T of Ω is introduced. For the sake of
simplicity, the elements of T are assumed to be squares of length h (the method is easily generalized to
other meshes). It is assumed that the well, represented by B, is located in the middle of one cell, denoted
by K0 , so that the origin 0 is the center of K0 . It is also assumed that the mesh size, h, is large with
respect to the radius of the well, rp (which is the case in real applications, where, for instance, h ranges
between 10 and 100 m). Following the principle of the finite volume method, one discrete unknown u K
per cell K (K ∈ T ) is introduced in order to discretize the following system:
Z
∇u(x) · nK (x)dγ(x) = 0, K ∈ T , K 6= K0 ,
Z∂K Z (3.149) singd
∇u(x) · nK0 (x)dγ(x) = ∇u(x) · nB (x)dγ(x),
∂K0 ∂B

where nP denotes the normal to ∂P , outward to P (with P = K, K0 or B).

Hence, we have to discretize ∇u · nK on ∂K (and ∇u · nB on ∂B) in terms of {uL, L ∈ T } (and “BC”
and Pp ).
The problems arise in the discretization of ∇u · nK0 and ∇u · nB . Indeed, if σ = K|L is the common edge
sing
to K and L (elements of T ), with K 6= K0 and L 6= K0 , since the solution of (3.148) is “smooth” enough
with respect to the mesh size, except “near” the well, ∇u · nK can be discretized by h1 (uL − uK ) on σ.
In order to discretize ∇u near the well, it is assumed that ∇u · nB is constant on ∂B. Let q(x) =
−2πrp ∇u · nB for x ∈ ∂B (recall that nB is the normal to ∂B, outward to B). Then q ∈ IR is a new
unknown, which satisfies
Z
−∇u · nB dγ(x) = q.
∂B
sing
Denoting by | · | the euclidian norm in IR 2 , and u the solution to (3.148), let v be defined by
q
v(x) = ln(|x|) + u(x), x ∈ Ω \ B, (3.150) reg1
2π
q
v(x) = ln(rp ) + Pp , x ∈ B. (3.151) reg2
2π
 91

Thanks to the boundary conditions satisfied by u on ∂B, the function v satisfies −div(∇v) = 0 on the
whole domain Ω, and therefore v is regular on the whole domain Ω. Note that, if we set
q
u(x) = − ln(|x|) + v(x), a.e. x ∈ Ω,
2π
then
−div(∇u) = qδ0 on Ω,
where δ0 is the Dirac mass at 0. A discretization of ∇u · nK0 is now obtained in the following way. Let σ
be the common edge to K1 ∈ T and K0 , since v is smooth, it is possible to approximate ∇v · nK0 on σ
by h1 (vK1 −reg2
vK0 ), where vKi is some approximation of v in Ki (e.g. the value of v at the center of Ki ).
Then, by (3.151), it is natural to set
q
v K0 = ln(rp ) + Pp ,
2π
reg1
and by (3.150),
q
v K1 = ln(h) + uK1 .
2π
reg1 q q
By (3.150) and from the fact R that the integral over σ of ∇( 2π ln(|x|)) · nK0 is equal to 4 , we find the
following approximation for σ ∇u · nK0 dγ:
q q h
− + ln( ) + uK1 − Pp .
4 2π rp
The discretization is now complete, there are as many equations as unknowns. The discrete unknowns
appearing in the discretized problem are {uK , K ∈ T , K 6= K0 } and q. Note that, up to now,
singd
the
singd
unknown uK0 has not been used. The discrete equations are given by (3.149) where each term of (3.149)
is replaced by its approximation in terms of {uK , K ∈ T , K 6= K0 } and q. In particular, the discrete
singd
equation “associated” to the unknown q is the discretization of the second equation of ( 3.149), which is

X4
q h
( ln( ) + uKi − Pp ) = 0, (3.152) do1
i=1
2π rp

where {Ki , i = 1, 2, 3, 4} are the four neighbouring cells to K0 .

It is possible to replace the unknown q by the unknown uK0 (as it is done in petroleum engineering) by
setting
q q h
uK 0 = − ln( ) + Pp , (3.153) uko
4 2π rp
the interest of which is that it yields the usual formula for the discretization of ∇u · nK0 on σ if σ is the
1
common edge to Kdo1 1 and K0 , namely h (uK1 − uK0 ); the discrete equation associated to the unknown
uK0 is then (from (3.152))
4
X
(uKi − uK0 ) = −q
i=1
uko
and (3.153) may be written as:
1
q = ip (Pp − uK0 ), with ip = .
− 14 + 1
2π ln( rhp )
This last equation defines ip , the so called “well-index” in petroleum engineering. With this formula for ip ,
the discrete unknowns are nowsingd{uK , K ∈ T }. The discretesingd
equations associated to {uK , K ∈ T , K 6= K0 }
are given by the first part of (3.149) where each terms of (3.149) is replaced by its approximation in terms
of {uK , K ∈ T } (using also “BC” on ∂Ω). The discrete equation associated to the unknown u K0 is
 92

4
X
(uKi − uK0 ) = −ip (Pp − uK0 ),
i=1

where {Ki , i = 1, 2, 3, 4} are the four neighbouring cells to K0 .

Note that the discrete unknown uK0 is somewhat artificial, it does not really represent the value of u in
q
K0 . In fact, if x ∈ K0 , the “approximate value” of u(x) is − 2π ln( |x| q q h
rp ) + Pp and uK0 = 4 − 2π ln( rp ) + Pp .

3.5.2 Mesh refinement

Mesh refinement consists in using, in certain areas of the domain, control volumes of smaller size than
elsewhere. In the case of triangular grids, a refinement may be performed for instance by dividing each
triangle in the refined area into four subtriangles, and those at the boundary of the refined area in two
triangles. Then, with some additional technique
meshdirichlet
(e.g. change ofrouenmesh
meshdirichlet
meshneumanmeshneuman
diagonal), one may obtain an admissible
rouenmesh
mesh in the sense
ellesterr ellesterr
of definitions
nllesterr nllesterr
3.1 page
rouenesterr
37, 3.5
rouenesterr
page 63 and 3.8 page 79; therefore the error estimates
3.3 page 52, 3.5 page 69 and 3.8 page 81 hold under the same assumptions.

In the case of rectangular grids, the same refining procedure leads to “atypical” nodes and edges, i.e. an
edge σ of a given control volume K may be common to two other control volumes, denoted by L and
M . This is also true in the triangular case if the triangles of the boundary of the refined area are left
untouched. smesh smesh
Let us consider for instance the same problem as in section 3.1.1 page 33, with the same assumptions
and notations, namely the discretization of

−∆u(x, y) = f (x, y), (x, y) ∈ Ω = (0, 1) × (0, 1),

u(x, y) = 0, (x, y) ∈ ∂Ω.
It is easily seen that,fluxrect
in this case, if the approximation of the fluxes is performed using differential
fluxrect
quotients such as in (3.6) page 34, the fluxes on the “atypical” edge σ cannot be consistent, since the
lines joining the centers of K and L and the centers of K and M are not orthogonal to σ. However, the
error which results from this lack of consistency can be controlled if the number of atypical edges is not
too large.
In the case of rectangular grids (with a refining procedure), denoting by Ea the set of “atypical” edges of a
given mesh T , i.e. edges with separate more than two control volumes, and Ta the set of “atypical” control
volumes, i.e. the control volumes containing an atypical edge in their boundaries; let e K denote the error
between u(xK ) and uK for each control volume K, and eT denote the piecewise constant function defined
by e(x) = eK for any x ∈ K, then one has
X 1
kekL2(Ω) ≤ C(size(T ) + ( m(K)) 2 ).
K∈Ta
ellesterr ellesterr belmouhoub
The proof is similar to that of Theorem 3.3 page 52. It is detailed in Belmouhoub [1996].

3.6 Compactness results

rellichd
This section is devoted to some functional analysis results which were used in the previous section. Let Ω
be a bounded open set of IR d , d ≥ 1. Two relative compactness results in L2 (Ω) for sequences “almost”
bounded in H 1 (Ω) which were used in the proof of convergence of the schemes are presented here. Indeed,
they are variations of the Rellich theorem (relative compactness in L2 (Ω) of a bounded sequence in H 1 (Ω)
or H01 (Ω)). The originality of these results is not the fact that the sequences are relatively compact in
L2 (Ω), which is an immediate consequence of the Kolmogorov theorem (see below), but the fact that the
eventual limit, in L2 (Ω), of the sequence (or of a subsequence) is necessarily in H 1 (Ω) (or in H01 (Ω) for
kolmh10
Theorem 3.10), a space which does not contain the elements of the sequence.
 93

theorem in L2 (Ω) which we now recall.

We shall make use in this section of the Kolmogorov compactnessbre
The essential part of the proof of this theorem may be found in Brezis [1983].

Kolm Theorem 3.9 Let ω be an open bounded set of IR N , N ≥ 1, 1 ≤ q < ∞ and A ⊂ Lq (ω). Then, A is
relatively compact in Lq (ω) if and only if there exists {p(u), u ∈ A} ⊂ Lq (IR N ) such that

1. p(u) = u a.e. on ω, for all u ∈ A,

2. {p(u), u ∈ A} is bounded in Lq (IR N ),
3. kp(u)(· + η) − p(u)kLq (IR N ) → 0, as η → 0, uniformly with respect to u ∈ A.

Let us now state the compactness results used in this chapter.

kolmh10 Theorem 3.10 Let Ω be an open bounded set of IR d with a Lipschitz continuous boundary, d ≥ 1, and
{un , n ∈ IN} a bounded sequence of L2 (Ω). For n ∈ IN, one defines ũn by ũn = un a.e. on Ω and ũn = 0
a.e. on IR d \ Ω. Assume that there exist C ∈ IR and {hn , n ∈ IN} ⊂ IR + such that hn → 0 as n → ∞
and

kũn (· + η) − ũn k2L2 (IR d ) ≤ C|η|(|η| + hn ), ∀n ∈ IN, ∀η ∈ IR d . (3.154) transh10

Then, {un , n ∈ IN} is relatively compact in L2 (Ω). Furthermore, if un → u in L2 (Ω) as n → ∞, then

u ∈ H01 (Ω).
kolmh10
Proof of Theorem 3.10
Since {hn , n ∈ IN} is bounded,
Kolm
the fact that {un , n ∈ IN} is relatively compact in L2 (Ω) is an immediate
consequence of Theorem 3.9, taking N = d, ω = Ω, q = 2 and p(un ) = ũn . Then, assuming that un → u
in L2 (Ω) as n → ∞, it is only necessary to prove that u ∈ H01 (Ω). Let us first remark that ũn → ũ in
L2 (IR d ), as n → ∞, with ũ = u a.e. on Ω and ũ = 0 a.e. on IR d \ Ω.
∞ d d
Then, for ϕ ∈ Ctransh10
c (IR ), one has, for all η ∈ IR , η 6= 0 and n ∈ IN, using the Cauchy-Schwarz inequality
and thanks to (3.154),
Z p
(ũn (x + η) − ũn (x)) C|η|(|η| + hn )
ϕ(x)dx ≤ kϕkL2 (IR d ) ,
IR d |η| |η|
which gives, letting n → ∞, since hn → 0,
Z √
(ũ(x + η) − ũ(x))
ϕ(x)dx ≤ CkϕkL2 (IR d ) ,
IR d |η|
and therefore, with a trivial change of variables in the integration,
Z √
(ϕ(x − η) − ϕ(x))
ũ(x)dx ≤ CkϕkL2 (IR d ) . (3.155) ueh1
IR d |η|
ueh1
Let {ei , i = 1, . . . , d} be the canonical basis of IR d . For i ∈ {1, . . . , d} fixed, taking η = hei in (3.155)
and letting h → 0 (with h > 0, for instance) leads to
Z √
∂ϕ(x)
− ũ(x)dx ≤ CkϕkL2 (IR d ) ,
IR d ∂x i

for all ϕ ∈ Cc∞ (IR d ).

This proves that Di ũ (the derivative of ũ with respect to xi in the sense of distributions) belongs to
L2 (IR d ), and therefore that ũ ∈ H 1 (IR d ). Since u is the restriction kolmh10
of ũ on Ω and since ũ = 0 a.e. on
IR d \ Ω, therefore u ∈ H01 (Ω). This completes the proof of Theorem 3.10.
 94

kolmh1 Theorem 3.11 Let Ω be an open bounded set of IR d , d ≥ 1, and {un , n ∈ IN} a bounded sequence of
L2 (Ω). For n ∈ IN, one defines ũn by ũn = un a.e. on Ω and ũn = 0 a.e. on IR d \ Ω. Assume that there
exist C ∈ IR and {hn , n ∈ IN} ⊂ IR + such that hn → 0 as n → ∞ and such that

kũn (· + η) − ũn k2L2 (IR d ) ≤ C|η|, ∀n ∈ IN, ∀η ∈ IR d , (3.156) transh11

and, for all compact ω̄ ⊂ Ω,

kun (· + η) − un k2L2 (ω̄) ≤ C|η|(|η| + hn ), ∀n ∈ IN, ∀η ∈ IR d , |η| < d(ω̄, Ωc ). (3.157) transh12

(The distance between ω̄ and IR d \ Ω is denoted by d(ω̄, Ωc ).)

Then {un , n ∈ IN} is relatively compact in L2 (Ω). Furthermore, if un → u in L2 (Ω) as n → ∞, then
u ∈ H 1 (Ω).
kolmh1
Proof of Theorem 3.11
kolmh10 transh11 Kolm
The proof is very similar to that of Theorem 3.10. Using assumption 3.156, Theorem 3.9 yields that {un ,
n ∈ IN} is relatively compact in L2 (Ω). Assuming now that un → u in L2 (Ω), as n → ∞, one has to
prove that u ∈ H 1 (Ω).
Let ϕ ∈ C ∞ (Ω) and ε > 0 such that ϕ(x) = 0 if the distance from x to IR d \ Ω is less than ε. Assumption
transh12 c
3.157 yields
Z p
(un (x + η) − un (x)) C|η|(|η| + hn )
ϕ(x)dx ≤ kϕkL2 (Ω) ,
Ω |η| |η|
for all η ∈ IR d such that 0 < |η| < ε. kolmh10
From this inequality, it may be proved, as in the proof of Theorem 3.10 (letting n → ∞ and using a
change of variables in the integration),
Z √
(ϕ(x − η) − ϕ(x))
u(x)dx ≤ CkϕkL2 (Ω) ,
Ω |η|
for all η ∈ IR d such that 0 < |η| < ε.
Then, taking η = hei and letting h → 0 (with h > 0, for instance) one obtains, for all i ∈ {1, . . . , d},
Z √
∂ϕ(x)
− u(x)dx ≤ CkϕkL2 (Ω) ,
Ω ∂xi

for all ϕ ∈ Cc∞ (Ω).

This proves that Di u (the derivative of u with respect to xi in the sense of distributions) belongs to
kolmh1
L2 (Ω), and therefore that u ∈ H 1 (Ω). This completes the proof of Theorem 3.11.
 Chapter 4

Parabolic equations
parabolic
The aim of this chapter is the study of finite volume schemes applied to a class of linear or nonlinear
parabolic problems. We consider the following transient diffusion-convection equation:

ut (x, t) − ∆ϕ(u)(x, t) + div(vu)(x, t) + bu(x, t) = f (x, t), x ∈ Ω, t ∈ (0, T ), (4.1) parabch

where Ω is an open polygonal bounded subset of IR d , with d = 2 or d = 3, T > 0, b ≥ 0, v ∈ IR d is,

for the sake of simplicity, a constant velocity field, f is a function defined on Ω × IR + which represents a
volumetric source term. The function ϕ is a nondecreasing Lipschitz continuous function, which arises in
the modelling of general diffusion processes. A simplified version of Stefan’s problem may be expressed
parabch
with the formulation (4.1) where ϕ is a continuous piecewise linear function,parabch
which is constant on an
interval. The porous medium equation is also included in equation (4.1), with ϕ(u) = umparabest , m > 1.
However, the linear case, i.e. ϕ(u) =
paranl
u, is of
paranl
full interest and the error estimate of section 4.2 will be
given in such a case. In section 4.3 page 102, we study the convergence of the explicit and of the implicit
Euler scheme for the nonlinear case with v = 0 and b = 0.

Remark 4.1 One could also consider a nonlinear convection term of the form div(vψ(u))(x, t) where
ψ ∈ C 1 (IR, IR). Such a nonlinear convection term will be largely studied in the framework of nonlinear
hyper1dhypmd
hyperbolic equations (chapters 5 and 6) and we restrain here to a linear convection term for the sake of
simplicity.

An initial condition is given by

u(x, 0) = u0 (x), x ∈ Ω. (4.2) parabcich

Let ∂Ω denote the boundary of Ω, and let ∂Ωd ⊂ ∂Ω and ∂Ωn ⊂ ∂Ω such that ∂Ωd ∪ ∂Ωn = ∂Ω and
∂Ωd ∩ ∂Ωn = ∅. A Dirichlet boundary condition is specified on ∂Ωd ⊂ ∂Ω. Let g be a real function
defined on ∂Ωd × IR + , the Dirichlet boundary condition states that

u(x, t) = g(x, t), x ∈ ∂Ωd , t ∈ (0, T ). (4.3) parabclch

A Neumann boundary condition is given with a function g̃ defined on ∂Ωn × IR + :

−∇ϕ(u)(x, t) · n(x) = g̃(x, t), x ∈ ∂Ωn , t ∈ (0, T ), (4.4) parabc2ch

where n is the unit normal vector to ∂Ω, outward to Ω.

Remark 4.2 Note that, formally, ∆ϕ(u) = div(ϕ0 (u)∇u). Then, parabch
if ϕ0 (u)(x, t) = 0 for some (x, t) ∈
Ω × (0, T ), the diffusion coefficient vanishes, so that Equation (4.1) is a “degenerate” parabolic equation.
In this case of degeneracy, the choice of the boundary conditions is important in order for the problem
to be well-posed. In the case where ϕ0 is positive, the problem is always parabolic.

95
 96

parabch
parabc2ch
In the next section, a finite volume scheme for the discretization of (4.1)-(4.4) is presented. An error
parabest
estimate in the linear caseparanl
(that is ϕ(u) = u) is given in section 4.2. Finally, a nonlinear (and degenerate)
case is studied in section 4.3; a convergence result is given for subsequences of sequences of approximate
solutions, and, when the weak solution is unique, for the whole set of approximate solutions. A uniqueness
result is therefore proved for the case of a smooth boundary.

4.1 Meshes and schemes

parasch
parabch
parabc2ch
In order to perform a finite volume discretization of system (4.1)-(4.4), admissible meshes are usedmeshdirichlet
in
a similar way to
meshdirichlet
the elliptic cases. Let T be an admissible mesh of Ω in the sense of Definition 3.1
page 37 with the additional assumption that any σ ∈ Eext is included in the closure of ∂Ωd or included
in the closure of ∂Ωn . The time discretization may be performed with a variable time step; in order
to simplify the notations, we shall choose a constant time step k ∈ (0, T ). Let Nk ∈ IN? such that
Nk = max{n ∈ IN, nk < T }, and we shall denote tn = nk, for n ∈ {0, . . . , Nk + 1}. Note that with a
variable time step, error estimates and convergence results similar to that which are given in the next
sections hold.

Denote by {unK , K ∈ T , n ∈ {0, . . . , Nk + 1}} the discrete unknowns; the value unK is an expected
approximation of u(xK , nk). parabch
In order to obtain the numerical scheme, let us integrate formally Equation (4.1) over each control volume
K of T , and time interval (nk, (n + 1)k), for n ∈ {0, . . . , Nk }:

Z Z (n+1)k Z
(u(x, tn+1 ) − u(x, tn ))dx − ∇ϕ(u)(x, t) · nK (x)dγ(x)dt+
ZK(n+1)k Z nk ∂K
Z (n+1)k Z Z (n+1)k Z (4.5) eqint
v · nK (x)u(x, t)dγ(x)dt + b u(x, t)dxdt = f (x, t)dxdt.
nk ∂K nk K nk K

where nK is the unit normal vector to ∂K, outward to K.

Recall that, as usual, the stability condition for an explicit discretization of a parabolic equation requires
the time step to be limited by a power two of the space step, which is generally too strong a condition eqint
in terms of computational cost. Hence the choice of an implicit formulation in the left hand side of ( 4.5)
which yields
Z Z
1
(u(x, tn+1 ) − u(x, tn ))dx − ∇ϕ(u)(x, tn+1 ) · nK (x)dγ(x)+
k
Z K Z∂K Z Z (4.6) eqinti
1 (n+1)k
v · nK (x)u(x, tn+1 )dγ(x) + b u(x, tn+1 )dxdt = f (x, t)dxdt,
∂K K k nk K
eqint
There now remains to replace in Equation (4.5) each term by its approximation with respect to the
eqinti
discrete unknowns (and the data). Before doing
parabch
so, let us remark that another way to obtain ( 4.6) is to
integrate (in space) formally Equation (4.1) over each control volume K of T , at time t ∈ (0, T ). This
gives
Z Z
ut (x, t)dx − ∇ϕ(u)(x, t) · nK (x)dγ(x)+
ZK ∂K Z Z (4.7) eqintrt
v · nK (x)u(x, t)dγ(x) + b u(x, t)dx = f (x, t)dx.
∂K K K
eqintrt
An implicit time discretization is then obtained by taking t = tn+1 in the left eqintrt
hand side of (4.7), and
replacing ut (x, tn+1 ) by (u(x, tn+1 ) − u(x, tn ))/k.
eqinti
For the right hand side of ( 4.7) a mean value of f
between eqintrt
tn and tn+1 may be used. This gives (4.6). It is also possible to take f (x, tn+1 ) in theparabest
right hand
side of (4.7). This latter choice is simpler for the proof of some error estimates (see Section 4.2).
 97

eqinti
Writing the approximation of the various terms in Equation (4.6) with respect to the discrete unknowns
(namely, {unK , K ∈ T , n ∈ {0, . . . , Nk + 1}}) and taking into account the initialparabch
and parabc2ch
boundary conditions
yields the following implicit finite volume scheme for the discretization ellmd
of ( 4.1)-( 4.4), usingellschemad
ellschema
the same
notations
ellschemad
and introducing some auxiliary unknowns as in Chapter 3 (see equations ( 3.20)-( 3.23) page
42):

un+1 − unK X X
n+1
m(K) K
+ FK,σ + vK,σ un+1 n+1
σ,+ + m(K)buK
n
= m(K)fK ,
k (4.8) parab
σ∈EK σ∈EK
∀K ∈ T , ∀n ∈ {0, . . . , Nk },
with
 
n
dK,σ FK,σ = −m(σ) ϕ(unσ ) − ϕ(unK ) for σ ∈ EK , for n ∈ {1, . . . , Nk + 1}, (4.9) parabflux

n n
FK,σ = −FL,σ for all σ ∈ Eint such that σ = K|L, for n ∈ {1, . . . , Nk + 1}, (4.10) parabcons

Z nk Z
n 1
FK,σ = g̃(x, t)dγ(x)dt for σ ∈ EK such that σ ⊂ ∂Ωn , for n ∈ {1, . . . , Nk + 1}, (4.11) parabc2
k (n−1)k σ

and

unσ = g(yσ , nk) for σ ⊂ ∂Ωd , for n ∈ {1, . . . , Nk + 1}, (4.12) parabcl
easyvf1
The upstream choice for the convection term is performed as in the elliptic case (see page 41, recall that
vK,σ = m(σ)v.nK,σ ),
 n
uK , if v · nK,σ ≥ 0,
unσ,+ = for all σ ∈ Eint such that σ = K|L, (4.13) parabconv
unL , if v · nK,σ < 0,
 n
n uK , if v · nK,σ ≥ 0,
uσ,+ = for all σ ∈ EK such that σ ⊂ ∂Ω. (4.14) parabconvb
unσ if v · nK,σ < 0,
parabflux
parabcl
Note that, in the same way as in the elliptic case, the unknowns un+1
σ may be eliminated using (4.9)-(4.12).
There remains to define the right hand side, which may be defined by:
Z (n+1)k Z
n 1
fK = f (x, t)dxdt, ∀K ∈ T , ∀n ∈ {0, . . . , Nk }, (4.15) parabvfnp
k m(K) nk K

or by:
Z
n 1
fK = f (x, tn+1 )dx, ∀K ∈ T , ∀n ∈ {0, . . . , Nk }. (4.16) parabvfnq
m(K) K

Initial conditions can be taken into account by different ways, depending on the regularity of the data
u0 . For example, it is possible to take
Z
1
u0K = u0 (x)dx, K ∈ T , (4.17) parnlcondini
m(K) K
or

u0K = u0 (xK ), K ∈ T . (4.18) parabvf3

 98

parab parabconvb parabvfnp

Remark 4.3 It is notparabvf3
parabvfnq parnlcondini
obvious to prove that the implicit finite volume scheme (4.8)-(4.14) (with (4.15) or
n+1
(4.16) and (4.17) or (4.18)) has a solution. Once the unknowns FK,σ are eliminated, a nonlinear system
of equations has to be solved. A proof of the existence and uniqueness parnleximp
of a solution to this system is
proved in the next section for the linear case, and is sketched in Remark 4.9 for the nonlinear case.

Remark 4.4 (Comparison with finite difference and finite element) Let us first consider the case
of the heat equation, that is the case where v = 0, b = 0, ϕ(s) = s for all s ∈ IR, with Dirichlet condi-
tion on the whole boundary (∂Ωd = ∂Ω). If the the mesh consists in rectangular control volumes with
constant space step in each direction, then the discretization obtained with the finite volume method
gives (as in the case of the Laplace operator), the same scheme than the one obtained with the finite
difference method (for which the discretization points are the centers of the elements ofATTT ) except at
the
KAM
boundary. In the general nonlinear case,
MEY
finite difference methods have been used in Attey [1974],
Kamenomostskaja, S.L. [1995] and Meyer [1973], for example. AG
Finite element methods haveCIA
also been classically used for this type of problem, see for instance Amiez
lfefv
and Gremaud [1991] or Ciavaldini [1975]. Following the notations of section 3.4.1, let M be a finite
ciarlet
element mesh of Ω, consisting of triangles (see e.g. Ciarlet, P.G. [1978] for the conditions on the
triangles), and let V ⊂ Ω be the set of vertices of M. For K ∈ V (note that here K denotes a point of
Ω), let ϕK be the shape function associated to K in the piecewise linear finite element method for the
parabch
mesh M. A finite element formulation for (4.1), with the implicit Euler scheme in time, yields for a node
/cv/in/Omega:
Z  Z Z
1
(un+1 (x) − un (x))ϕK (x)dx + ∇un+1 (x) · ∇ϕK (x)dx = f (x, tn+1 )ϕK (x)dx,
k Ω Ω Ω

Let us approximate un by the usual Galerkin expansion:

X X
un+1 = un+1
L ϕL and u =
n
unL ϕL
L∈V L∈V

where unLis expected to be an approximation of u at time tn and node L, for all L and n; replacing in
the above equation, this yields:

Z XZ Z
1 X
(un+1
j − u n
j )ϕ L (x)ϕ K (x)dx + u n+1
j ∇ϕ L (x) · ∇ϕ K (x)dx = f (x, tn+1 )ϕK (x)dx. (4.19) efp
k Ω Ω Ω
L∈V L∈V

Hence, the finite element formulation yields, at each time step, a linear system of the form CU n+1 +
AU n+1 = B (where U n+1 = (un+1 t
K )K∈V,K∈Ω , and A and C are N × N matrices); this scheme, however, is
generally used after a mass-lumping, i.e. by assigning to the diagonal term of C the sum of the coefficients
lfefv
of the corresponding line and transforming it into a diagonal matrix; we already saw in section 3.4.1 that
the part AU n+1 may be seen as a linearefpsystem derived from a finite volu;e formulation; hence the mass
lumping technique the left hand side of (4.19) to be seen as the result of a discretization by a finite volume
scheme.

4.2 Error estimate for the linear case

parabest
We consider, in this section, the linear case, ϕ(s) = s for all s ∈ IR, and assume ∂Ωdparabch
= ∂Ω, i.e. that a
parabc2ch
Dirichlet boundary condition is given on the whole boundary, in which case Problem (4.1)-(4.4) becomes

ut (x, t) − ∆u(x, t) + div(vu)(x, t) + bu(x, t) = f (x, t), x ∈ Ω, t ∈ (0, T ),

u(x, 0) = u0 (x), x ∈ Ω,
 99

u(x, t) = g(x, t), x ∈ ∂Ω, t ∈ (0, T );

parab parabconvb
the finite volume scheme (4.8)-(4.14) then becomes, assuming, for the sake of simplicity, that xK ∈ K
for all K ∈ T ,

un+1 − unK X X
n+1
m(K) K
+ FK,σ + vK,σ un+1 n+1
σ,+ + m(K)buK
n
= m(K)fK ,
k (4.20) parablin
σ∈EK σ∈EK
∀K ∈ T , ∀n ∈ {0, . . . , Nk },
with
n
FK,σ = −τK|L (unL − unK ) for all σ ∈ Eint such that σ = K|L, for n ∈ {1, . . . , Nk + 1}, (4.21) parablinflux

n
FK,σ = −τσ (g(yσ , nk) − unK ) for all σ ∈ EK such that σ ⊂ ∂Ω, for n ∈ {1, . . . , Nk + 1}, (4.22) parablincl
and

unσ,+ = unK , if v · nK,σ ≥ 0,
for all σ ∈ Eint such that σ = K|L, (4.23) parablinconv
unσ,+ = unL , if v · nK,σ < 0,

unσ,+ = unK , if v · nK,σ ≥ 0,
for all σ ∈ EK such that σ ⊂ ∂Ω. (4.24) parablinconv
unσ,+ = g(yσ , nk), if v · nK,σ < 0,
parabvfnq parabvf3
The source term and initial condition f and u0 , are discretized by (4.16) and (4.18).
guo
A convergence analysis of a one-dimensional vertex-centered scheme was performed in Guo and Stynes
[1997] by writing the scheme in a finite element framework. Here we shall use direct finite volume
techniques which also handle the multi-dimensional case.
The following theorem gives an L∞ estimate (on the approximate solution) and an error estimate. Some
easy generalizations are possible
parinfty
(for instance, the same theorem holds with b < 0, the only difference is
that in the L∞ estimate (4.25) the bound c also depends on b).

parabestim Theorem 4.1 Let Ω be an open polygonal bounded subset of IR d , T > 0, u ∈ C 2 (Ω × IR + , IR), b ≥ 0
and v ∈ IR d . Let u0 ∈ C 2 (Ω, IR) be defined by u0 = u(·, 0), let f ∈ C 0 (Ω × IR + , IR) be defined by
0
f = ut − div(∇u) + div(vu) + bu and g ∈ Cmeshdirichlet
(∂Ω × IR + , IR) defined by g = u on ∂Ω × IR + . Let T be an
meshdirichlet
admissible mesh in the sense of Definition
parablinparablinconvb
3.1 page 37
parab parabconvb
and k ∈ (0, T ). Then there exists a unique vector
parabvfnq parabvf3
(uK )K∈T satisfying (4.20)-(4.24) (or (4.8)-(4.14)) with (4.16) and (4.18). There exists c only depending
on u0 , T , f and g such that

sup{|unK |, K ∈ T , n ∈ {1, . . . , Nk + 1}} ≤ c. (4.25) parinfty

Furthermore, let enK = u(xK , tn ) − unK , for K ∈ T and n ∈ {1, . . . , Nk + 1}, and h = size(T ). Then there
exists C ∈ IR + only depending on b, u, v, Ω and T such that
X 1
( (enK )2 m(K)) 2 ≤ C(h + k), ∀ n ∈ {1, . . . , Nk + 1}. (4.26) estdeuxp
K∈T
parabestim
Proof of Theorem 4.1
For simplicity, let us assume that xK ∈ K for all K ∈ T . Generalization without this condition is
straightforward.
(i) Existence, uniqueness, and L∞ estimate parablin
n
For a given n ∈ {0, . . . , Nk }, parablin
set fK = 0 and unK = 0 in (4.20), and g(yσ , (n + 1)k) = 0 for all σ ∈ E such
ellexistu
⊂ ∂Ω. Multiplying (4.20) by un+1
that σellexistu K and using the same technique as in the proof of Lemma 3.2
page 42 yields that un+1
K = 0 for all K ∈ T . This yields the uniqueness of the solution {un+1
K , K ∈ T } to
 100

parablin
parablinconvb
(4.20)-(4.24) for given {unK ,parablin n
K ∈ Tparablinconvb
}, {fK , K ∈ T } and {g(yσ , (n + 1)k), σ ∈ E, σ ⊂ ∂Ωd }. The existence
follows immediately, since (4.20)-(4.24) is a finite dimensional linear system with respect to the unknown
{un+1
K , K ∈ T } (with as many unknowns as equations).
parinfty
Let us now prove the estimate (4.25).
Set mf = min{f (x, t), x ∈ Ω, t ∈ [0, 2T ]} and mg = min{g(x, t), x ∈ ∂Ω, t ∈ [0, 2T ]}.
Let n ∈ {0, . . . , Nk }. Then, we claim that

min{un+1 n
K , K ∈ T } ≥ min{min{uK , K ∈ T } + kmf , 0, mg }. (4.27) claim1
Indeed, if min{un+1
K , K ∈ T } < min{0, mg }, let K0 ∈ T such that un+1
K0 = min{un+1
K , K ∈ T }. Since
parablin
un+1
K0 < 0 and un+1
K0 < mg writing (4.20) with K = K0 and n leads to

un+1 n n n
K0 ≥ uK0 + kfK0 ≥ min{uK , K ∈ T } + kmf ,
claim1
this proves (4.27), which yields, by induction, that:

min{unK , K ∈ T } ≥ min{min{u0K , K ∈ T }, 0, mg } + nk min{mf , 0}, ∀ n ∈ {0, . . . , Nk + 1}.

Similarly,

max{unK , K ∈ T } ≤ max{max{u0K , K ∈ T }, 0, Mg } + nk max{Mf , 0}, ∀ n ∈ {0, . . . , Nk + 1},

with Mf = max{f (x, t), x ∈ Ω, t ∈ [0, 2T ]} and Mg = max{g(x, t), x ∈ ∂Ω, t ∈ [0, 2T ]}.
parinfty
This proves (4.25) with c = ku0 kL∞ (Ω) + kgkL∞(∂Ω×(0,2T )) + 2T kf kL∞(Ω×(0,2T )) .

(ii) Error estimate ellesterr ellesterr

As in the stationary case (see the proof of Theorem 3.3 page 52), one uses the regularity of the data
and the solution to write an equation for the error enK = u(xK , tn ) − unK , defined for K ∈ T and
∈ {0, . . . , Nk + 1}. Note that e0K = 0 for K ∈ T . Let n ∈ {0, . . . , Nk }. Integrating (in space) Equation
nparabch parabvfnq
n
(4.1) over each control volume K of T , at time t = tn+1 , gives, thanks to the choice of fK (see (4.16)),

Z Z   Z
n
ut (x, tn+1 )dx − ∇u(x, t) − vu(x, tn+1 ) · nK (x)dγ(x) + b u(x, tn+1 )dx = m(K)fK . (4.28) eqintrtl
K ∂K K

Note that, for all x ∈ K and all K ∈ T , a Taylor expansion yields, thanks to the regularity of u:

ut (x, tn+1 ) = (1/k)(u(xK , tn+1 ) − u(xK , tn )) + snK (x) with |snK (x)| ≤ C1 (h + k)
Z
with some C1 only depending on u and T . Therefore, defining SK n
= snK (x)dx, one has: |SK
n
| ≤
K
C1 m(K)(h + k). ellesterrellesterr
One follows now the lines of the proof of Theorem 3.3 page 52, adding the terms due to the time derivative
parablin eqintrtl
ut . Substracting (4.20) to (4.28) yields

en+1 − enK X 
m(K) K
+ Gn+1
K,σ + W n+1
K,σ + bm(K)en+1
K =
k
X σ∈E K (4.29) diffea
bm(K)ρnK − n
m(σ)(RK,σ n
+ rK,σ ) − SKn
, ∀K ∈ T ,
σ∈EK
meshdirichlet
meshdirichlet
where (with the notations of Definition 3.1 page 37),

Gn+1 n+1
K,σ = −τσ (eL − en+1
K ), ∀K ∈ T , ∀σ ∈ EK ∩ Eint , σ = K|L,
n+1 n+1
GK,σ = τσ eK , ∀K ∈ T , ∀σ ∈ EK ∩ Eext ,
n+1
WK,σ = m(σ)v · nK,σ (u(xσ,+ , tn+1 ) − un+1
σ,+ ),
 101

where xσ,+ = xK (resp. xL ) if σ ∈ Eint , σ = K|L and v · nK,σ ≥ 0 (resp. v · nK,σ < 0) and xσ,+ = xK
(resp. yσ ) if σ = EK ∩ Eext and v · nK,σ ≥ 0 (resp. v · nK,σ < 0),
Z
n n+1 1
ρK = u(xK , t )− u(x, tn+1 )dx,
m(K) K
Z
n
m(σ)RK,σ = τσ (u(xK , tn+1 ) − u(xL , tn+1 )) + ∇u(x, tn ).nK,σ dγ(x) if σ = K|L ∈ Eint ,
σ
Z
n
m(σ)RK,σ = τσ (u(xK , tn+1 ) − g(yσ , tn+1 ) + ∇u(x, tn ).nK,σ dγ(x) if σ ∈ EK ∩ Eint ,
σ
and Z
n
m(σ)rK,σ = v · nK,σ (m(σ)u(xσ,+ , tn+1 ) − (σ)u(x, tn+1 )dγ(x), for any σ ∈ E.
m
ellesterr
As in Theorem 3.3, thanks to the regularity of u, there exists C2 , only depending on u, v and T , such
n n
that |RK,σ | + |rK,σ | ≤ C2 h and |ρnK | ≤ C2 h, for any K ∈ T and σ ∈ EK .
diffea
(4.29) by en+1
Multiplying ellesterr , summing
K ellraslebol1 for K ∈ T , and performing the same computations as in the proof
ellraslebol4ellraslebol4
of Theorem 3.3 between (3.56) to (3.60) page 55 yields, with some C3 only depending on u, v, b, Ω and
T,
1X n+1 2 1 1
m(K)(eK ) + ken+1 k21,T + bken+1 k2L2 (Ω) ≤
k 2 T 2 T
K∈T
X 1X (4.30) finesterr
C3 h2 + C1 (h + k) m(K)|en+1
K |+ m(K)en+1 n
K eK ,
k
K∈T K∈T
n
where the second term of the right hand side is due to the bound on SK and where en+1
T is a piecewise
constant function defined by

en+1
T (x) = en+1
K , for x ∈ K, K ∈ T .
finesterr
Inequality (4.30) yields

ken+1
T k2L2 (Ω) ≤ 2kC3 h2 + 2kC1 m(Ω)(k + h)ken+1
T kL2 (Ω) + kenT k2L2 (Ω) ,
which gives


ken+1
T k2L2 (Ω) ≤ kenT k2L2 (Ω) + C4 kh2 + k(k + h)ken+1
T kL2 (Ω) , (4.31) An
where C4 ∈ IR + only depends on u, v, b, Ω and T . Remarking that for ε > 0, the following inequality
holds:

C4 k(k + h)ken+1
T kL2 (Ω) ≤ ε2 ken+1
T k2L2 (Ω) + (1/ε2 )C42 k 2 (k + h)2 ,
An
taking ε2 = k/(k + 1), (4.31) yields

ken+1
T k2L2 (Ω) ≤ (1 + k)kenT k2L2 (Ω) + C4 kh2 (1 + k) + (1 + k)2 C42 k(k + h)2 . (4.32) Anp
Anp
Then, if kenT k2L2 (Ω) 2
≤ cn (h + k) , with cn ∈ IR + , one deduces from (4.32), using h ≤ h + k and k < T ,
that
ken+1
T k2L2 (Ω) ≤ cn+1 (h + k)2 with cn+1 = (1 + k)cn + C5 k and C5 = C4 (1 + T ) + C42 (1 + T )2 .
(Note that C5 only depends on u, v, b, Ω and T ).
Choosing c0 = 0 (since ke0T kL2 (Ω) = 0), the relation between cn and cn+1 yields (by induction) cn ≤
estdeuxp
C5 e2kn . Estimate (4.26) follows with C 2 = C5 e4T .
 102

parabestim
Remark 4.5 The error estimate given in Theorem 4.1 may be generalized to the case of discontinuous
coefficients. The admissibility of the mesh is thenrouenmesh
redefined so that therouenrh
rouenmesh
data and the solution are piecewise
regular on the control volumes as in Definition 3.8 page 79, see also Herbin [1996].

4.3 Convergence in the nonlinear case

paranl
4.3.1 Solutions to the continuous problem
parnlpbcont parabch
parabc2ch
We consider Problem (4.1)-(4.4) with v = 0, b = 0, ∂Ωn = ∂Ω and g̃ = 0, that is a homogeneous
Neumann condition on the whole boundary, in which case the problem becomes

ut (x, t) − ∆ϕ(u)(x, t) = f (x, t), for (x, t) ∈ Ω × (0, T ), (4.33) paranlanf

with

∇ϕ(u)(x, t) · n(x) = 0, for (x, t) ∈ ∂Ω × (0, T ), (4.34) paranlanb

and the initial condition

u(x, 0) = u0 (x), for all x ∈ Ω. (4.35) paranlani

We suppose that the following hypotheses are satisfied:

parnlH Assumption 4.1

(i) Ω is an open bounded polygonal subset of IR d and T > 0.

(ii) The function ϕ ∈ C(IR, IR) is a nondecreasing locally Lipschitz continuous function.
(iii) The initial data u0 satisfies u0 ∈ L∞ (Ω).
(iv) The right hand side f satisfies f ∈ L∞ (Ω × IR ?+ ).
paranlanf
Equation (4.33) is a degenerate parabolic equation. Formally, ∆ϕ(u) = div(ϕ0 (u)∇u), so that, if ϕ0paranlanf
(u) =
0,paranlani
the diffusion coefficient vanishes. Let us give a definition of a weak
KAM
solution u to Problem ( 4.33)-
(LSU
4.35) (the proof of the existence of such a solution isMEIgiven in Kamenomostskaja,
Ol2
S.L. [1995],
Ladyženskaja, Solonnikov and Ural’ceva [1968], Meirmanov [1992], Oleinik [1960]).
parnlH paranlanf
paranlani
parnldefess Definition 4.1 Under Assumption 4.1, a measurable function u is a weak solution of (4.33)-(4.35) if

u ∈ L∞ (Ω × (0, T )),
Z TZ  
u(x, t)ψt (x, t) + ϕ(u(x, t))∆ψ(x, t) + f (x, t)ψ(x, t) dx dt + (4.36) paranlanw
Z0 Ω
u0 (x)ψ(x, 0)dx = 0, for all ψ ∈ AT ,
Ω

where AT = {ψ ∈ C 2,1 (Ω × [0, T ]), ∇ψ · n = 0 on ∂Ω × [0, T ], and ψ(·, T ) = 0}, and C 2,1 (Ω × [0, T ])
denotes the set of functions which are restrictions on Ω × [0, T ] of functions from IR d × IR into IR which
are twice (resp. once) continuously differentiable with respect to the first (resp. second) variable. (Recall
that, as usual, n is the unit normal vector to ∂Ω, outward to Ω.)

Remark 4.6 It is possible to use a solution in a stronger sense, using only one integration by parts for
the space term. It then leads to a larger test function space than AT .
 103

Remark 4.7 Note that the function u formally satisfies the conservation law
Z Z Z tZ
u(x, t)dx = u0 (x)dx + f (x, t)dxdt, (4.37) parnlcsvt
Ω Ω 0 Ω

for all t ∈ [0, T ]. This property is also satisfied by the finite volume approximation.

4.3.2 Definition of the finite volume approximate solutions

vfquatre vfquatremsn msn
As in sections 3.1.2 page 37 and 3.2.1 page 63, an admissible mesh of Ω is defined, with respect to which
a functional space is introduced: this space contains the approximate solutions obtained from the finite
volume discretization over the admissible mesh.

Xmeshk Definition 4.2 Letmeshneuman open bounded polygonal subset of IR d , T be an admissible mesh in the
Ω be anmeshneuman
sense of Definition 3.5 page 63, T > 0, k ∈ (0, T ) and Nk = max{n ∈ IN; nk < T }. Let X(T , k) be
the set of functions u from Ω × (0, (Nk + 1)k) to IR such that there exists a family of real values {unK ,
K ∈ T , n ∈ {0, . . . , Nk }}, with u(x, t) = unK for a.e. x ∈ K, K ∈ T and for a.e. t ∈ [nk, (n + 1)k),
n ∈ {0, . . . , Nk }.

Since we only consider, for the sake of simplicity, a Neumann

parab
boundary condition,parabflux
we can easily eliminate
parabcons parabc2
n
the unknowns FK,σ located at the edges in equation (4.8) using the equations (4.9), (4.10), and (4.11).
An explicit version of the scheme can then be written in the following way:

un+1 − unK X  
m(K) K
− τK|L ϕ(unL ) − ϕ(unK ) = m(K)fK
n
,
k (4.38) parnlschema
L∈N (K)
∀K ∈ T , ∀n ∈ {0, . . . , Nk }.
Z
1
u0K = u0 (x)dx, ∀K ∈ T , (4.39) parnlcondinj
m(K) K
Z (n+1)k Z
n 1
fK = f (x, t)dxdt, ∀K ∈ T , ∀n ∈ {0, . . . , Nk }. (4.40) parnlcondsou
k m(K) nk K

m(K|L) meshneumanmeshneuman
(Recall that τK|L = , see Definition 3.5 page 63.)
dK|L
parnlcondinj
Remark 4.8 The definition using the mean value in (4.39) is motivated by the lack of regularity assumed
on the data u0 .
parnlschema
parnlcondsou
The scheme (4.38)-(4.40) is then used to build an approximate solution, uT ,k ∈ X(T , k) by

uT ,k (x, t) = unK , ∀x ∈ K, ∀t ∈ [nk, (n + 1)k), ∀K ∈ T , ∀n ∈ {0, . . . , Nk }. (4.41) parnldefuapp

parnleximp Remark 4.9 The implicit finite volume scheme is defined by

un+1 − unK X  
m(K) K
− τK|L ϕ(un+1 n+1 n
L ) − ϕ(uK ) = m(K)fK ,
k (4.42) parnlschemai
L∈N (K)
∀K ∈ T , ∀n ∈ {0, . . . , Nk }.

The proof of the existence of un+1

K , for any n ∈ {0, . . . , Nk }, can be obtained using the following fixed
point method:
 104

n+1,0
uK = unK , for all K ∈ T , (4.43) parnlpfixei
and
n+1,m+1
uK − unK X  
n+1,m n+1,m+1 n
m(K) − τK|L ϕ(uL ) − ϕ(uK ) = m(K)fK ,
k (4.44) parnlpfixem
L∈N (K)
∀K ∈ T , ∀m ∈ IN.
parnlpfixem
Equation (4.44) gives a contraction property, which leads first to prove that for all K ∈ T , the sequence
n+1,m n+1,m
(ϕ(uK ))m∈IN converges. Then we deduce that (uK )m∈IN also converges.
We shall see further that all results obtained for the explicit scheme are also true, with convenient
adaptations, for the implicit scheme. The function uT ,k is then defined by uT ,k (x, t) = un+1
K , for all x ∈
K, for all t ∈ [nk, (n + 1)k).
parnlH
The mathematical problem is to study, under Assumption paranlanf
meshneuman
4.1 andparanlani
with a mesh in the sense of Definition
3.5, the convergence of uT ,k to a weak solution of Problem (4.33)-(4.35), when h = size(T ) → 0 and k → 0.
Exactly in the same manner as for the elliptic case, we shall use estimates on the approximate solutions
which are discrete versions of the estimates which hold on the solution of the continous problem and which
ensure the stability of the scheme. We present the proofs in the case of the explicit scheme and show in
several remarks how they can be extended to the case of the implicit scheme (which is significantly
CIA
easier
to study). The proof of convergence of the scheme uses a nlparaws
weak-? convergence
nlparaws
property, as in Ciavaldini
[1975], which is proved in a general setting in section 4.3.5 page
paranlanf
114. For the sake of completeness,
paranlani
the proof of uniqueness of the weak solution of Problem (4.33)-(4.35) is given for the case of a regular
boundary; this allowsparanlanf
to prove that the whole sequence of approximate solutions converges to the weak
paranlani
solution of problem (4.33)-( 4.35), in which case an admissible mesh for a smooth domain can easily be
parnlHT parnlHT
defined (see Definition 4.4 page 114).

4.3.3 Estimates on the approximate solution

Maximum principle
parnlH meshneumanmeshneuman
parnlestinf Lemma 4.1 Under Assumption 4.1, let T be an admissible mesh in the sense of Definition 3.5 page 63
ϕ(x) − ϕ(y)
and k ∈ (0, T ). Let U = ku0 kL∞(Ω) + T kf kL∞(Ω×(0,T )) , B = sup . Assume that the
−U ≤x<y≤U x−y
condition

m(K)
k≤ X , for all K ∈ T , (4.45) parnlcflt
B τK|L
L∈N (K)
parnlschema
parnldefuapp
is satisfied. Then the function uT ,k defined by (4.38)-(4.41) verifies

kuT ,k kL∞ (Ω×(0,T )) ≤ U. (4.46) parnlinfsta

parnlestinf
Proof of Lemma 4.1
and assume unK ∈ [−U, +U ] for all K ∈ T .
Let n ∈ {0, . . . , Nk − 1}parnlschema
Let K ∈ T , Equation (4.38) can be written as
 k X ϕ(unL ) − ϕ(unK )  n
un+1
K = 1− τK|L uK +
m(K) unL − unK
L∈N (K)
k X  ϕ(unL ) − ϕ(unK )  n n
τK|L uL + kfK ,
m(K) unL − unK
L∈N (K)
 105

ϕ(unL ) − ϕ(unK )
with the convention that n − un = 0 if unL − unK = 0.
u L
parnlcflt K
Thanks to the condition (4.45) and since ϕ is nondecreasing, the following inequality can be deduced:

|un+1 n
K | ≤ sup |uL | + kkf kL∞(Ω×(0,T )) .
L∈T

Then, since K is arbitrary in T ,

sup |un+1 n
K | ≤ sup |uL | + kkf kL∞ (Ω×(0,T )) . (4.47) parnlinfsta1
K∈T L∈T
parnlinfsta1
Using (4.47), an induction on n yields, for n ∈ {0, . . . , Nk }, supK∈T |unK | ≤ ku0 kL∞(Ω) +nkkf kL∞(Ω×(0,T )) ,
parnlinfsta
which leads to Inequality (4.46) since Nk k ≤ T .

Remark 4.10 Assume that there exist α, β, γ ∈ IR ?+ such that m(K) ≥ αhd , m(∂K) ≤ βhd−1 , for all
and dK|L ≥ γh, for all K|L ∈ Eint (recall that h = size(T )). Then, k ≤ Ch2 with C = (αγ)/(Bβ)
K ∈ T ,parnlcflt
yields (4.45).

steps, and (uTn ,kn )n∈IN the

Remark 4.11 Let (Tn , kn )n∈IN be a sequence of admissible meshes and timeparnlinfsta
associated sequence of approximate finite volume solutions; then , thanks to (4.46), there exists a function
u ∈ L∞ (Ω × (0, T )) and a subsequence of (uTn ,kn )n∈IN which converges to u for the weak-? topology of
L∞ (Ω × (0, T )).
parnlinfsta
= ku0 kL∞ (Ω) + 2T kf kL∞(Ω×(0,2T )) , for the implicit
Remark 4.12 Estimate (4.46) is also true, with Uparnlinfsta1
scheme, because the fixed point method guarantees (4.47) (with kf kL∞(Ω×(0,2T )) instead of kf kL∞(Ω×(0,T ))
and until n = Nk ), without any condition on k.

Space translates of approximate solutions

Let us now define a seminorm, which is the discrete version of the seminorm in the space L2 (0, T ; H 1 (Ω)).

defh1dk Definition 4.3 (Discrete L2 (0, T ; H 1 (Ω)) seminorm) Let Ω be anmeshneuman open bounded polygonal subset of
meshneuman
IR d , T an admissible finite volume mesh in the sense of Definition 3.5 page 63, T > 0, k ∈ (0, T ) and
Nk = max{n ∈ IN; nk < T }. For u ∈ X(T , k), let the following seminorms be defined by:
X
|u(·, t)|21,T = τK|L (unL − unK )2 , for a.e. t ∈ (0, T ) and n = max{n ∈ IN; nk ≤ t}, (4.48) normh1dsk
K|L∈Eint

and
Nk
X X
|u|21,T ,k = k τK|L (unL − unK )2 . (4.49) normh1dk
n=0 K|L∈Eint

Let us now state some preliminary lemmata to the use of Kolmogorov’s theorem (compactness properties
in L2 (Ω × (0, T ))) in the proof of convergence of the approximate solutions.

lsteftrx Lemma 4.2 Let Ω be an open bounded polygonal subset of IR d , T an admissible mesh in the sense of
meshneumanmeshneuman
Definition 3.5 page 63, T > 0, k ∈ (0, T ) and u ∈ X(T , k). For all η ∈ IR d , let Ωη be defined by
Ωη = {x ∈ Ω, [x, x + η] ⊂ Ω}. Then:

ku(· + η, ·) − u(·, ·)k2L2 (Ωη ×(0,T )) ≤ |u|21,T ,k |η|(|η| + 2 size(T )), ∀η ∈ IR d , (4.50) parnltrx
 106

lsteftrx
Proof of Lemma 4.2
h1dtot h1dtot cvneum2 cvneum2
Reproducing the proof of Lemma 3.3 page 44 (see also the proof of (3.110) page 75), we get, for a.e.
t ∈ (0, T ):

ku(· + η, t) − u(·, t)k2L2 (Ωη ) ≤ |u(·, t)|21,T |η|(|η| + 2 size(T )), ∀η ∈ IR d . (4.51) parnltrx1
parnltrx1 parnltrx
Integrating (4.51) on t ∈ (0, T ) gives (4.50).
lsteftrx
The set Ωη defined in Lemma 4.2 verifies Ω \ Ωη ⊂ ∪σ∈Eext ωη,σ
¯ , with ωη,σ = {y − tη, y ∈ σ, t lsteftrx
∈ [0, 1]}.
Then, m(Ω \ Ωη ) ≤ |η| m(∂Ω), since m(ω̄η ) ≤ ηm(σ). Then, an immediate corollary of Lemma 4.2 is the
following:

lsteftrxt Lemma 4.3 Let Ω be an open bounded polygonal subset of IR d , T an admissible mesh in the sense of
meshneumanmeshneuman
Definition 3.5 page 63, T > 0, k ∈ (0, T ) and u ∈ X(T , k). Let ũ be defined by ũ = u a.e. on Ω × (0, T ),
and ũ = 0 a.e. on IR d+1 \ Ω × (0, T ). Then:

(  
kũ(· + η, ·) − ũ(·, ·)k2L2 (IR d+1 ) ≤ |η| |u|21,T ,k (|η| + 2 size(T )) + 2m(∂Ω)kuk2L∞(Ω×(0,T )) ,
(4.52) parnltrxt
∀η ∈ IR d .
parnltrxt
Remark 4.13 Estimate (4.52) makes use of the L∞ (Ω×(0, T ))-norm of u ∈ X(T , k). A similar estimate
with the L2 (Ω × (0, T ))-norm of u (instead of the L∞ (Ω × (0, T ))-norm). Indeed, the right
may be provedparnltrxt
hand side of (4.52) may be replaced by Cη(|u|2
nllconv nllconv1,T ,k
+ kuk2L2(Ω×(0,T )) ), where C only depends on Ω. This
estimate is proved in Theorem 3.7 page 74 where it is used for the convergence of numerical schemes for
the Neumann problem (for which no L∞ estimate on the approximate solutions is available). The key to
nlltrace nlltrace
its proof is the “trace lemma” 3.10 page 72.

Let us now state the following lemma, which gives an estimate of the discrete L2 (0, T ; H 1 (Ω)) seminorm
of the nonlinearity.
parnlH meshneumanmeshneuman
parnlestimeg Lemma 4.4 Under Assumption 4.1, let T be an admissible mesh in the sense of Definition 3.5 page 63.
Let ξ ∈ (0, 1) and k ∈ (0, T ) such that

m(K)
k ≤ (1 − ξ) X , for all K ∈ T . (4.53) parnlcfl
B τK|L
L∈N (K)
parnlschema
parnldefuapp
Let uT ,k ∈ X(T , k) be given by (4.38)-(4.41).
Let U = ku0 kL∞ (Ω) + T kf kL∞(Ω×(0,T )) and B be the Lipschitz constant of ϕ on [−U, U ]. Then there
exists F1 ≥ 0, which only depends on Ω, T , ϕ, u0 , f and ξ such that

|ϕ(uT ,k )|21,T ,k ≤ F1 . (4.54) parnlestimx

parnlestimeg
Proof of lemma 4.4
parnlcfl parnlcflt parnlestinf
Let us first remark that the condition (4.53) is stronger than (4.45). Therefore, the result of lemma 4.1
parnlschema
holds, i.e. |unK | ≤ U , for all K ∈ T , n ∈ {0, . . . , Nk }. Multiplying equation (4.38) by kunK , and summing
the result over n ∈ {0, . . . , Nk } and K ∈ T yields:
Nk X
X
m(K)(un+1
K − unK )unK −
n=0K∈T
Nk X   Nk (4.55) parnl1e
X X X X
k τK|L ϕ(unL ) − ϕ(unK ) unK = k m(K)unK fK
n
.
n=0 K∈T L∈N (K) n=0 K∈T
 107

parnl1e
In order to obtain a lower bound on the first term on the left hand side of (4.55), let us first remark that:
1 n+1 2 1 n 2 1 n+1
(un+1
K − unK )unK =
(u ) − (uK ) − (uK − unK )2 . (4.56) parnl2e
2 K 2 2
parnlschema
Now, applying (4.38), using Young’s inequality, the following inequality is obtained:
h 1 X 2 n 2i
(fK )
(un+1
K − unK )2 ≤ k 2 (1 + ξ) τK|L (ϕ(unL ) − ϕ(unK )) + . (4.57) parnl3eb
m(K) ξ
L∈N (K)

which yields in turn, using the Cauchy-Schwarz inequality:

k2 h X ih X  2 i
(un+1
K − unK )2 ≤ (1 + ξ) τK|L τK|L ϕ(unL ) − ϕ(unK )
m(K)2
L∈N (K) L∈N (K) (4.58) parnl3e
n 2
(1 + ξ)(k fK )
+ .
ξ
parnlcfl
Taking condition (4.53) into account gives:

k h X  2 i (1 + ξ)(k f n )2
(un+1
K − unK )2 ≤ (1 − ξ 2 ) τK|L ϕ(unL ) − ϕ(unK ) + K
. (4.59) parnl4e
Bm(K) ξ
L∈N (K)
parnl2e parnl4e parnl1e
Using (4.56) and (4.59) leads to the following lower bound on the first term of the left hand side of (4.55):
Nk X
X 1X  
Nk +1 2
m(K)(un+1
K − unK )unK ≥ m(K) (uK ) − (u0K )2
n=0K∈T
2
K∈T
Nk
1 − ξ2 X Xh X  2 i
− k τK|L ϕ(unL ) − ϕ(unK ) (4.60) parnl5e
2B n=0
K∈T L∈N (K)
Nk
k(1 + ξ) X X
n 2
− k m(K)(fK ) .
2ξ n=0
K∈T
parnl1e
Z x the second term on the left hand side of (4.55). Let φ ∈ C(IR, IR) be defined by
Let us now handle
φ(x) = xϕ(x) − ϕ(y)dy, where x0 ∈ IR is an arbitrary given real value. Then the following equality
x0
holds:
Z un
L
φ(unL ) − φ(unK ) = unK (ϕ(unL ) − ϕ(unK )) − (ϕ(x) − ϕ(unL ))dx. (4.61) parnl6e
un
K

The following technical lemma is used here and several times in the sequel:

petitlemme Lemma 4.5 Let g : IR → IR be a monotone Lipschitz continuous function, with a Lipschitz constant
G > 0. Then:
Z d
1
| (g(x) − g(c))dx| ≥ (g(d) − g(c))2 , ∀c, d ∈ IR. (4.62) estplemme
c 2G
petitlemme
Proof of Lemma 4.5
petitlemme
In order to prove Lemma 4.5, we assume, for instance, that g is nondecreasing and c < d (the other
cases are similar). Then, one has g(s) ≥ h(s), for all s ∈ [c, d], where h(s) = g(c) for s ∈ [c, d − l] and
h(s) = g(c) + (s − d + l)G for s ∈ [d − l, d], with lG = g(d) − g(c), and therefore:
Z d Z d
l 1
(g(s) − g(c))ds ≥ (h(s) − g(c))ds = (g(d) − g(c)) = (g(d) − g(c))2 ,
c c 2 2G
 108

petitlemme
this completes the proof of Lemma 4.5.
petitlemme
parnl6e X X
Using Lemma 4.5, (4.61) and the equality τK|L (φ(unL ) − φ(unK )) = 0 yields:
K∈T L∈N (K)

Nk X
X X   Nk X
1 X X
− k τK|L ϕ(unL ) − ϕ(unK ) unK ≥ k τK|L (ϕ(unL ) − ϕ(unK ))2 . (4.63) parnl7e
n=0 K∈T L∈N (K)
2B n=0
K∈T L∈N (K)
parnlinfsta parnl1e
Since k < T we deduce from (4.46) that the right hand side of equation (4.55) satisfies
Nk
X X
| k m(K)unK fK
n
| ≤ 2T m(Ω)U kf kL∞(Ω×(0,2T )) . (4.64) parnl8e
n=0 K∈T
parnl1e parnl5e parnl7e parnl8e
Relations k < T , (4.55), (4.60), (4.63) and (4.64) lead to
Nk
ξ2 X X X
k τK|L (ϕ(unL ) − ϕ(unK ))2 ≤
2B n=0
K∈T L∈N (K) (4.65) parnl9e
 1+ξ  1
2T m(Ω)kf kL∞(Ω×(0,2T )) U + kf kL∞(Ω×(0,2T )) T + m(Ω)ku0 k2L∞ (Ω)
2ξ 2
which concludes the proof of the lemma.

parnlestimx
Remark 4.14 Estimate (4.54) also holds for the implicit scheme parnl2e
parnlschemaimp
, without any condition on k. One
multiplies (4.42) by un+1
K : the last term on the right hand side of ( 4.56) appears with the opposite sign,
which considerably simplifies the previous proof.

Time translates of approximate solutions

In order to fulfill the hypotheses of Kolmogorov’s theorem, the study of time translates must now be
performed. The following estimate holds:
parnlH parnlH meshneuman
lsteftrt Lemma 4.6 Under Assumption 4.1 page 102, let T be an
meshneuman
admissible
parnlschema
mesh in the sense of Definition 3.5
parnldefuapp
page 63 and k ∈ (0, T ). Let uT ,k ∈ X(T , k) be given by (4.38)-(4.41). Let U = kuT ,k kL∞(Ω×(0,T )) and B
be the Lipschitz constant of ϕ on [−U, U ]. Then:
(
kϕ(uT ,k (·, · + τ )) − ϕ(uT ,k (·, ·))k2L2 (Ω×(0,T −τ )) ≤
  (4.66) parnltrt
2Bτ |ϕ(uT ,k )|21,T ,k + BT m(Ω)U kf kL∞(Ω×(0,T )) , ∀τ ∈ (0, T ).
lsteftrt
Proof of Lemma 4.6
Let τ ∈ (0, T ). Since B is the Lipschitz constant of ϕ on [−U, U ], U = kuT ,k kL∞ (Ω×(0,T )) and ϕ is
nondecreasing, the following inequality holds:
Z  2 Z T −τ
ϕ(uT ,k (x, t + τ )) − ϕ(uT ,k (x, t)) dxdt ≤ B A(t)dt, (4.67) parnltrt1
Ω×(0,T −τ ) 0

where, for almost every t ∈ (0, T − τ ),

Z   
A(t) = ϕ(uT ,k (x, t + τ )) − ϕ(uT ,k (x, t)) uT ,k (x, t + τ ) − uT ,k (x, t) dx.
Ω
parnldefuapp
Let t ∈ (0, T − τ ). Using the definition of uT ,k (4.41), this may also be written:
X   
n (t) n (t) n (t) n (t)
A(t) = m(K) ϕ(uK1 ) − ϕ(uK0 ) uK1 − uK0 , (4.68) parnltrt3
K∈T
 109

n1 (t) ∈ {0, . . . , Nk } such that n0 (t)k ≤ t < (n0 (t) + 1)k and n1 (t)k ≤ t + τ < (n1 (t) + 1)k.
with n0 (t),parnltrt3
Equality (4.68) may be written as

X  n1 (t)
X 
n (t) n (t) n−1
A(t) = (ϕ(uK1 ) − ϕ(uK0 )) m(K)(unK − uK ) ,
K∈T n=n0 (t)+1

which also writes

X X
Nk 
n (t) n (t) n−1
A(t) = (ϕ(uK1 ) − ϕ(uK0 )) χn (t, t + τ )m(K)(unK − uK ) , (4.69) parnltrt4
K∈T n=1

χn (t, t + τ ) = 1 if nk ∈ (t, t + τ ] and χn (t, t + τ ) = 0 if nk ∈

withparnltrt4 / (t, t + τ ]. parnlschema
In (4.69), the order of summation between n and K is changed and the scheme (4.38) is used. Hence,
Nk
X hX
n (t) n (t)
A(t) = k χn (t, t + τ ) (ϕ(uK1 ) − ϕ(uK0 ))
 X n=1 K∈T i
n−1 n−1 n−1
τK|L (ϕ(uL ) − ϕ(uK )) + m(K)fK .
L∈N (K)

Gathering by edges, this yields:

Nk h X
X n (t) n (t) n (t) n (t)
A(t) = k τK|L (ϕ(uK1 ) − ϕ(uL1 ) − ϕ(uK0 ) + ϕ(uL0 ))
n=1 K|L∈Eint
X i
n−1 n−1 n (t) n (t) n−1
(ϕ(uL ) − ϕ(uK )) + (ϕ(uK1 ) − ϕ(uK0 ))m(K)fK χn (t, t + τ ).
K∈T

Using the inequality 2ab ≤ a2 + b2 , this yields:

1 1
A(t) ≤ A0 (t) + A1 (t) + A2 (t) + A3 (t), (4.70) parnltrt7
2 2
with
Nk
X X n (t) n (t)

A0 (t) = k χn (t, t + τ ) τK|L (ϕ(uL0 ) − ϕ(uK0 ))2 ,
n=1 K|L∈Eint

Nk
X X n (t) n (t)

A1 (t) = k χn (t, t + τ ) τK|L (ϕ(uL1 ) − ϕ(uK1 ))2 ,
n=1 K|L∈Eint

Nk
X X
n−1 n−1 2


A2 (t) = k χn (t, t + τ ) τK|L (ϕ(uL ) − ϕ(uK )) ,
n=1 K|L∈Eint

and
Nk
X X n (t) n (t) n−1


A3 (t) = k χn (t, t + τ ) (ϕ(uK1 ) − ϕ(uK0 ))m(K)fK .
n=1 K∈T

Note that, since t ∈ (0, T − τ ), n0 (t) ∈ {0, . . . , Nk }, and, for m ∈ {0, . . . , Nk }, n0 (t) = m if and only if
t ∈ [mk, (m + 1)k). Therefore,
Z T −τ Nk Z
X (m+1)k Nk
X X

A0 (t)dt ≤ k χn (t, t + τ ) τK|L (ϕ(um m 2
L ) − ϕ(uK )) dt,
0 m=0 mk n=1 K|L∈Eint
 110

which also writes

Z T −τ Nk
X Z (m+1)k  X
Nk  X
A0 (t)dt ≤ k χn (t, t + τ ) dt τK|L (ϕ(um m 2
L ) − ϕ(uK )) . (4.71) parnltrt12
0 m=0 mk n=1 K|L∈Eint

The change of variable t = s + (n − m)k yields

Z (m+1)k Z 2mk−nk+k Z 2mk−nk+k

χn (t, t + τ )dt = χn (s + (n − m)k, s + (n − m)k + τ )ds = χm (s, s + τ )ds,
mk 2mk−nk 2mk−nk

then, for all m ∈ {0, . . . , Nk },

Z (m+1)k X
Nk  Z
χn (t, t + τ ) dt ≤ χm (s, s + τ )ds = τ,
mk n=1 IR

s + τ ) = 1 if and only if mk ∈ (s, s + τ ] which is equivalent to s ∈ [mk − τ, mk).

since χm (s,parnltrt12
Therefore (4.71) yields
Z T −τ
A0 (t)dt ≤ τ |ϕ(uT ,k )|21,T ,k . (4.72) parnltrt14
0
Similarly:
Z T −τ
A1 (t)dt ≤ τ |ϕ(uT ,k )|21,T ,k . (4.73) parnltrt14b
0
R T −τ
Let us now study the term 0 A2 (t)dt:
Z T −τ Nk
X X Z T −τ
n−1 n−1 2
A2 (t)dt ≤ k τK|L (ϕ(uL ) − ϕ(uK )) χn (t, t + τ )dt. (4.74) parnltrt15
0 n=1 K|L∈Eint 0

R T −τ
Since 0 χn (t, t + τ ) ≤ τ (recall that χn (t, t + τ ) = 1 if and only if t ∈ [nk − τ, nk)), the following
inequality holds:
Z T −τ
A2 (t)dt ≤ τ |ϕ(uT ,k )|21,T ,k . (4.75) parnltrt16
0
In the same way:
Nk
X X Z T −τ
R T −τ

0
A3 (t)dt ≤ k m(K)2BU kf kL∞(Ω×(0,T )) χn (t, t + τ )dt
0 (4.76) parnltrt17
n=1 K∈T
≤ τ T m(Ω)2BU kf kL∞(Ω×(0,T )) .
parnltrt1
parnltrt7 parnltrt14
parnltrt17
parnltrt
Using inequalities (4.67), (4.70) and (4.72)-(4.76), (4.66) is proved.

parnltrt
Remark 4.15 Estimate (4.66) is again true for the implicit scheme , with kf kL∞(Ω×(0,2T )) instead of
kf kL∞(Ω×(0,T )) .
lsteftrt
An immediate corollary of Lemma 4.6 is the following.
 111

parnlH parnlH meshneuman

lsteftrtt Lemma 4.7 Under Assumption 4.1 page 102, let T be anparnlschema
meshneuman
admissible mesh in the sense of Definition 3.5
parnldefuapp
page 63 and k ∈ (0, T ). Let uT ,k ∈ X(T , k) be given by (4.38)-(4.41). Let U = kuT ,k kL∞ (Ω×(0,T ) and B
be the Lipschitz constant of ϕ on [−U, U ]. One defines ũ by ũ = uT ,k a.e. on Ω × (0, T ), and ũ = 0 a.e.
on IR d+1 \ Ω × (0, T ). Then:

kϕ(ũ(·, · + τ )) − ϕ(ũ(·, ·))k2L2 (IR d+1 ) ≤ 2|τ |B |ϕ(uT ,k )|21,T ,k +

BT m(Ω)U kf kL∞(Ω×(0,T )) + Bm(Ω)U 2 ,
∀τ ∈ IR.

4.3.4 Convergence
parnlH parnlH
parnlcvgce Theorem 4.2 Under Assumption 4.1 page 102, let U = ku0 kL∞(Ω) + T kf kL∞(Ω×(0,T )) and

ϕ(x) − ϕ(y)
B= sup .
−U ≤x<y≤U x−y
Let ξ ∈ (0, 1) be a given real value. For m ∈ IN, let Tm parnlcfl
meshneumanmeshneuman
be an admissible mesh in the sense of Definition
3.5 page 63 and k m
parnlschema
∈ (0, T
parnldefuapp
) satisfying the condition ( 4.53) with T = Tm and k = km . Let uTm ,km be
given by (4.38)-(4.41) with T = Tm and k = km . Assume that size(Tm ) → 0 as m → ∞.
Then, there exists a subsequence of the sequence ofparanlanf
approximate solutions, still denoted by (u Tm ,km )m∈IN ,
paranlani
which converges to a weak solution u of Problem (4.33)-(4.35), as m → ∞, in the following sense:
(i) uTm ,km converges to u in L∞ (Ω × (0, T )), for the weak-? topology as m tends to +∞,
(ii) (ϕ(uTm ,km )) converges to ϕ(u) in L1 (Ω × (0, T )) as m tends to +∞,
where uTm ,km and ϕ(uTm ,km ) also denote the restrictions of these functions to Ω × (0, T ).
parnlcvgce
Proof of Theorem 4.2
Let parnlcfl
us set um = uTm ,km and assume, without lossparnlestinf
of generality, that ϕ(0) = 0. First remark that,
parnlestinf
by (4.53), km → 0 as m → 0. Thanks to Lemma 4.1 page 104, the sequence (um )m∈IN is bounded in
L∞ (Ω × (0, T ). Then, there exists a subsequence, still denoted by (um )m∈IN , such that um converges, as
m → ∞, to u in L∞ (Ω × (0, T )), for the weak-? topology. Kolm Kolm
For the study of the sequence (ϕ(um ))m∈IN , we shall apply Theorem 3.9 page 93 with N = d + 1, q = 2,
ω = Ω×(0, T ) and p(v) = ṽ with ṽ defined, as usual, by ṽ = v on Ω×(0, T ) and ṽ = 0 on IR d+1 \Ω×(0, T ).
Kolm
The first and second itemsparnlestimeg
of Theorem 3.9 are clearly satisfied; let us prove hereafter that the third is
also satisfied. By Lemma 4.4, the sequence (|ϕ(um )|1,Tm ,km )m∈IN is bounded. Let η ∈ IR d and τ ∈ IR,
since

kϕ(ũm (· + η, · + τ )) − ϕ(ũm (·, ·))kL2 (IR d+1 ) ≤

kϕ(ũm (· + η, ·)) − ϕ(ũm (·, ·))kL2 (IR d+1 ) + kϕ(ũm (·, · + τ )) − ϕ(ũm (·, ·))kL2 (IR d+1 ) ,
lsteftrxtlsteftrtt Kolm
lemmata 4.3 and 4.7 give the third item of Theorem 3.9 and this yields the compactness of the sequence
(ϕ(um ))m∈IN in L2 (Ω × (0, T )).
Therefore, there exists a subsequence, still denoted by (ϕ(um ))m∈IN , and there exists χ ∈ L2 (Ω × (0, T ))
such that ϕ(uTm ,km ) converges, as m → ∞, to χ in L2 (Ω × (0, T )). Indeed, since (ϕ(um ))m∈IN is bounded
in L∞ (Ω × (0, T )), this convergence holds in Lq (Ω × (0, T )) for all 1 ≤ q < ∞. Furthermore, since ϕ is
parnlAF parnlAF
nondecreasing, Theorem 4.3 page 114 gives that χ = ϕ(u).

Up to now, the following properties have been shown to be satisfied by a convenient subsequence:

(i) (um )m∈IN converges to u, as m → ∞, in L∞ (Ω × (0, T )) for the weak-? topology,

(ii) (ϕ(um ))m∈IN converges to ϕ(u) in L1 (Ω × (0, T )) (and even in Lp (Ω × (0, T )) for all p ∈ [0, ∞)).
 112

paranlanf
paranlani
There remains to show that u is a weak solution of Problem (4.33)-(4.35), which concludes the proof of
parnlcvgce
Theorem 4.2.
Let m ∈ IN. For the sake of simplicity, we shall use the notations T = Tm , h = size(T ) and k = km . Let
parnlschema parnlschema
ψ ∈ AT . We multiply (4.38) page 103 by kψ(xK , nk), and sum the result on n ∈ {0, . . . , Nk } and K ∈ T .
We obtain

T1m + T2m = T3m , (4.77) parnl1g

with
Nk X
X
T1m = m(K)(un+1
K − unK )ψ(xK , nk),
n=0K∈T

Nk
X X X  
T2m = − k τK|L ϕ(unL ) − ϕ(unK ) ψ(xK , nk),
n=0 K∈T L∈N (K)

and
Nk X
X
n
T3m = k ψ(xK , nk)m(K)fK .
n=0 K∈T

We first consider T1m .

Nk X
X  
T1m = m(K)unK ψ(xK , (n − 1)k) − ψ(xK , nk) +
n=1K∈T
X  
m(K) uNK
k +1
ψ(xK , kNk ) − u0K ψ(xK , 0) .
K∈T
parnlestinfNk +1
Performing one more step of the induction in Lemma 4.1, it is clear that |uK | < U +2T kf kL∞(Ω×(0,2T )) ,
for all K ∈ T .
Since 0 < T − Nk k ≤ k, there exists C1,ψ which only depends on ψ, T and Ω, such that |ψ(xK , Nk k)| ≤
kC1,ψ . Hence,
X Nk +1
m(K)uK ψ(xK , kNk ) → 0 as m → ∞.
K∈T

Since
X
k u0K 1K − u0 kL1 (Ω) → 0, as m → ∞,
K∈T

(where 1K (x) = 1 if x ∈ K, 0 otherwise), one has

X Z
m(K)u0K ψ(xK , 0) → u0 (x)ψ(x, 0)dx as m → ∞.
K∈T Ω

Since (um )m∈IN converges, as m → +∞, to u in L∞ (Ω × (0, T )), for the weak-? topology, and since
|uN
K | < U + T kf kL∞(Ω×(0,T )) , for all K ∈ T , the following property also holds:
k

Nk X
X   Z T Z
m(K)unK ψ(xK , (n − 1)k) − ψ(xK , nk) → − u(x, t)ψt (x, t)dxdt as m → ∞.
n=1 K∈T 0 Ω

Therefore,
 113

Z T Z Z
T1m → − u(x, t)ψt (x, t)dxdt − u0 (x)ψ(x, 0)dx, as m → ∞.
0 Ω Ω

We now study T2m . This term can be rewritten as

Nk
X X ψ(xK , nk) − ψ(xL , nk)
T2m = − k m(K|L)(ϕ(unL ) − ϕ(unK )) .
n=0
dK|L
K|L∈Eint

It is useful to introduce the following expression:

Nk Z
X (n+1)k Z
0
T2m = ϕ(uT ,k (x, t))∆ψ(x, nk)dxdt
n=0 nk Ω
Nk
X X Z
= k ϕ(unK ) ∆ψ(x, nk)dx
n=0 K∈T K
Nk
X X Z
= k (ϕ(unK ) − ϕ(unL )) ∇ψ(x, nk) · nK,L dγ(x).
n=0 K|L∈Eint K|L

The sequence (ϕ(um ))m∈IN converges to ϕ(u) in L1 (Ω × (0, T )); furthermore, it is bounded in L∞ so that
the integral between T and (Nk + 1)k tends to 0. Therefore:
Z T Z
0
T2m → ϕ(u(x, t))∆ψ(x, t)dxdt, as m → ∞.
0 Ω
0
The term T2m + T2m can be written as
Nk
X X
0
T2m + T2m = k m(K|L)(ϕ(unK ) − ϕ(unL ))RK,L
n
,
n=0 K|L∈E

with
Z
n 1 ψ(xL , nk) − ψ(xK , nk)
RK,L = ∇ψ(x, nk) · nK,L dγ(x) − .
m(K|L) K|L dK|L
n
Thanks to the regularity properties of ψ there exists Cψ , which only depends on ψ, such that |RK,L
parnlestimx
|≤
0
Cψ h. Then, using the estimate (4.54), we conclude that T2m + T2m → 0 as m → ∞. Therefore,
Z T Z
T2m → − ϕ(u(x, t))∆ψ(x, t)dxdt, as m → ∞.
0 Ω
Let us now study T3m .
n
Define fT ,k ∈ X(T , k) by fT ,k (x, t) = fK if (x, t) ∈ K × (nk, nk + k). Since fT ,k → f in L1 (Ω × (0, T )
∞
and since f ∈ L (Ω × (0, 2T ),
Z Z T
T3m → f (x, t)ψ(x, t)dtdx, as m → ∞.
Ω 0
parnl1g paranlanf
paranlani
Passing to the limit in Equationparnlcvgce
(4.77) gives that u is a weak solution of Problem (4.33)-(4.35). This
concludes the proof of Theorem 4.2.

Remark 4.16 This convergence proof is quite similar in the case of the implicit scheme, with the addi-
parnlcfl
tional condition that (km )m∈IN converges to zero, since condition (4.53) does not have to be satisfied.
 114

subsequence Remark 4.17 The above convergence result was shown for a subsequence only. A convergence theorem
is obtained for the full set of approximate solutions, if a uniqueness result up
is valid. Such a result can be
easily obtained in the case meshneuman
of a smooth boundary
meshneuman
and is given in section 4.3.6 below. For this case, an
extension to the definition 3.5 page 63 of admissible meshes is given hereafter.

parnlHT Definition 4.4 (Admissible meshes for regular domains) Let Ω be an open bounded connected
subset of IR d , d = 2 or 3 with a C 2 boundary ∂Ω. An admissible finite volume mesh of Ω is given by an
open bounded polygonal set Ω0 containing Ω, and an admissible mesh T 0 of Ω0 in the sense of Definition
meshneumanmeshneuman
3.5 page 63. The set of control volumes of the mesh of Ω are {K 0 ∩ Ω, K 0 ∈ T 0 such that md (K 0 ∩ Ω) > 0}
and the set of edges of the mesh is E = {σ ∩ Ω, σ ∈ E 0 such that md−1 (σ ∩ Ω) > 0}, where E 0 denotes the
set of edges of T 0 and mN denotes the N -dimensional Lebesgue measure.

Remark 4.18 For smooth domains Ω, the set of edges E of an admissible mesh of Ω does not contain
the parts of the boundaries of the control volumes which are included in the boundary ∂Ω of Ω.

4.3.5 Weak convergence and nonlinearities

nlparaws parnlcvgce
We show here a property which was used in the proof of Theorem 4.2.

parnlAF Theorem 4.3 Let U > 0 and ϕ ∈ C([−U, U ]) be a nondecreasing function. Let ω be an open bounded
subset of IR N , N ≥ 1. Let (un )n∈IN ⊂ L∞ (ω) such that
(i) −U ≤ un ≤ U a.e. in ω, for all n ∈ IN;
(ii) there exists u ∈ L∞ (ω) such that (un )n∈IN converges to u in L∞ (ω) for the weak-? topology;
(iii) there exists a function χ ∈ L1 (ω) such that (ϕ(un ))n∈IN converges to χ in L1 (ω).
Then χ(x) = ϕ(u(x)), for a.e. x ∈ ω.
parnlAF
Proof of Theorem 4.3
First we extend the definition of ϕ by ϕ(v) = ϕ(−U ) + v + U for all v < −U and ϕ(v) = ϕ(U ) + v − U
for all v > U , and denote again by ϕ this extension of ϕ which now maps IR into IR, is continuous and
nondecreasing. Let us define α± from IR to IR by α− (t) = inf{v ∈ IR, ϕ(v) = t} and α+ (t) = sup{v ∈
IR, ϕ(v) = t}, for all t ∈ IR.
Note that the functions α± are increasing and that
(i) α− is left continuous and therefore lower semi-continuous, that is

t = lim tn =⇒ α− (t) ≤ lim inf α− (tn ),

n→∞ n→∞

(ii) α+ is right continuous and therefore upper semi-continuous, that is

t = lim tn =⇒ α+ (t) ≥ lim sup α+ (tn ).

n→∞ n→∞

Thus, since we may assume, up to a subsequence, that ϕ(un ) → χ a.e. in ω,

   
α− (χ(x)) ≤ lim inf α− ϕ(un (x)) ≤ lim sup α+ ϕ(un (x)) ≤ α+ (χ(x)), (4.78) parnl1
n→∞ n→∞

for a.e. x ∈ ω.

A direct application of the definition of the functions α− and α+ gives

   
α− ϕ(un (x)) ≤ un (x) ≤ α+ ϕ(un (x)) . (4.79) parnl5
parnl5
Let L1+ = {ψ ∈ L1 (ω), ψ ≥ 0 a.e.}. Let ψ ∈ L1+ . We multiply (4.79) by ψ(x) and integrate over ω, it
yields
 115

Z   Z Z  
α− ϕ(un (x)) ψ(x)dx ≤ un (x)ψ(x)dx ≤ α+ ϕ(un (x)) ψ(x)dx. (4.80) plus
ω ω ω

Applying Fatou’s lemma to the sequences of L1 positive functions α− (ϕ(un ))ψ − α− (ϕ(−U ))ψ and
parnl1
α+ (ϕ(U ))ψ − α+ (ϕ(un ))ψ yields, with (4.78),
Z Z  
α− (χ(x))ψ(x)dx ≤ lim inf α− ϕ(un (x)) ψ(x)dx,
ω n→∞ ω
and
Z   Z
lim sup α+ ϕ(un (x)) ψ(x)dx ≤ α+ (χ(x))ψ(x)dx.
n→∞ ω ω
plus
Then, passing to the lim inf and lim sup in (4.80) and using the convergence of (un )n∈IN to u in L∞ (ω)
for the weak-? topology gives
Z Z Z
α− (χ(x))ψ(x)dx ≤ u(x)ψ(x)dx ≤ α+ (χ(x))ψ(x)dx.
ω ω ω

Thus, since ψ is arbitrary in L1+ , the following inequality holds for a.e. x ∈ ω:

α− (χ(x)) ≤ u(x) ≤ α+ (χ(x)),

parnlAF
which implies in turn that χ(x) = ϕ(u(x)) for a.e. x ∈ ω. This completes the proof of Theorem 4.3.

parnlAF
Remark 4.19 Another proof of Theorem 4.3 is possible by passing to the limit in the inequality
Z
0 ≤ (ϕ(un )(x) − ϕ(v(x)))(un (x) − v(x))dx, ∀v ∈ L∞ (ω),
ω

which leads to Z
0≤ (χ(x) − ϕ(v(x)))(u(x) − v(x))dx, ∀v ∈ L∞ (ω).
ω

From this inequality, one deduces that χ = ϕ(u) a.e. on ω.

nlwsdef
A third proof is possible by using the concept of nonlinear weak-? convergence, see Definition 6.3 page
nlwsdef
190.

4.3.6 A uniqueness result for nonlinear diffusion equations

up paranlanf
paranlani
The uniqueness of the weak solution to variations
MEI
of Problem (4.33)-(4.35) has been proved by several
authors. BKP
For precise references we refer to Meirmanov [1992].
GHP
Also rather similar proofs have been
given in Bertsch, Kersner and Peletier [1995] and Guedda, Hilhorst and Peletier [1997].
Recall that this uniqueness result allows to
parabch
obtain a convergence
parabc2ch
result on the whole set of finite volume
subsequence
approximate solutions to Problem (4.1)-(4.4) (see Remark 4.17).
paranlanf
paranlani
The uniqueness of the weak solution to Problem (4.33)-(4.35) immediately results from the following
property.
arnlthunicite Theorem 4.4 Let Ω be an open bounded subset of IR d with a C 2 boundary, and suppose thatparanlanf
parnlH
itemsparanlani
(ii),
(iii) and (iv) of Assumption 4.1 are
parnldefess
satisfied.
parnldefess
Let u 1 and u 2 be two solutions of Problem ( 4.33)-( 4.35)
in the sense of Definition 4.1 page 102, with initial conditions u0,1 and uparanlani 0,2 and source terms v1 and v2
respectively,
paranlanf
that is, for u 1 (resp. u 2 ), u 0 = u 0,1 (resp. u 0 = u 0,2 ) in ( 4.35) and f = v1 (resp. v2 ) in
(4.33).
 116

Then for all T > 0,

Z T Z Z Z T Z
|u1 (x, t) − u2 (x, t)|dxdt ≤ T |u0,1 (x) − u0,2 (x)|dx + (T − t) |v1 (x, t) − v2 (x, t)|dxdt.
0 Ω Ω 0 Ω
parnlthunicite
Before proving Theorem 4.4, let us first show the following auxiliary result.

The existence of regular solutions to the adjoint problem

parnlauxi Lemma 4.8 Let Ω be an open bounded subset of IR d with a C 2 boundary, and suppose that ϕ is a
nondecreasing locally Lipschitz-continuous function. Let T > 0, w ∈ Cc∞ (Ω × (0, T )) such that |w| ≤ 1,
and g ∈ C ∞ (Ω × [0, T ]) such that there exists r ∈ IR with 0 < r ≤ g(x, t), for all (x, t) ∈ Ω × (0, T ).

Then there exists a unique function ψ ∈ C 2,1 (Ω × [0, T ]) such that

ψt (x, t) + g(x, t)∆ψ(x, t) = w(x, t), for all (x, t) ∈ Ω × (0, T ), (4.81) parnllemme1

∇ψ · n(x, t) = 0, for all (x, t) ∈ ∂Ω × (0, T ), (4.82) parnllemme3

ψ(x, T ) = 0, for all x ∈ Ω. (4.83) parnllemme2
Moreover the function ψ satisfies

|ψ(x, t)| ≤ T − t, for all (x, t) ∈ Ω × (0, T ), (4.84) parnllemme4

and
Z T Z  2 Z T Z
g(x, t) ∆ψ(x, t) dxdt ≤ 4T |∇w(x, t)|2 dxdt. (4.85) parnllemme5
0 Ω 0 Ω
parnlauxi
Proof of Lemma 4.8
parnllemme5
It will be useful in the following to point out that the right side of (4.85) does not depend on g.
handparnllemme2
parnllemme1
Since the function g is bounded away from zero, equations (4.81)-(4.83) define a boundary value problem
for a usual heat equation with an initial condition, in which the time variable
LSU
is reversed. Since Ω, g and
w are sufficiently smooth, this problem has a unique solution ψ ∈ AT , see Ladyženskaja, Solonnikov
and Ural’ceva [1968]. Since |w| parnllemme3
≤ 1, the functions T parnllemme4
parnllemme1
− t and −(T LSU
− t) are respectively upper and
lower solutions of Problem (4.81)-(4.82). Hence we get (4.84) (see Ladyženskaja, Solonnikov and
Ural’ceva [1968]).
parnllemme5 parnllemme1
In order to show (4.85), multiply (4.81) by ∆ψ(x, t), integrate by parts on Ω × (0, τ ), for τ ∈ (0, T ]. This
gives
Z Z Z τZ  2
1 2 1 2
|∇ψ(x, 0)| dx − |∇ψ(x, τ )| dx + g(x, t) ∆ψ(x, t) dxdt =
2 ZΩ Z 2 Ω 0 Ω (4.86) parnllemme6
τ
− ∇w(x, t) · ∇ψ(x, t)dxdt.
0 Ω
parnllemme6
Since ∇ψ(·, T ) = 0, letting τ = T in (4.86) leads to
Z Z TZ  2
1 2
|∇ψ(x, 0)| dx + g(x, t) ∆ψ(x, t) dxdt =
2 ZΩ Z 0 Ω (4.87) parnllemme7
T
− ∇w(x, t) · ∇ψ(x, t)dxdt.
0 Ω
parnllemme6
Integrating (4.86) with respect to τ ∈ (0, T ) leads to
 117

Z T Z Z
1 T
|∇ψ(x, τ )|2 dxdτ ≤ |∇ψ(x, 0)|2 dx +
2 0 Ω 2Z Ω Z
T  2
T g(x, t) ∆ψ(x, t) dxdt + (4.88) parnllemme8
Z0 T ZΩ
T |∇w(x, t) · ∇ψ(x, t)|dxdt.
0 Ω
parnllemme7 parnllemme8
Using (4.87) and (4.88), we get
Z T Z Z T Z
1
|∇ψ(x, τ )|2 dxdτ ≤ 2T |∇w(x, t) · ∇ψ(x, t)|dxdt. (4.89) parnllemme9
2 0 Ω 0 Ω
parnllemme9
Thanks to the Cauchy-Schwarz inequality, the right hand side of (4.89) may be estimated as follows:
hZ T Z i2 Z T Z
|∇w(x, t) · ∇ψ(x, t)|dxdt ≤ |∇ψ(x, t)|2 dxdt
0 Ω 0Z ΩZ
T
× |∇w(x, t)|2 dxdt.
0 Ω
parnllemme9
With (4.89), this implies
hZ T Z i2 Z T Z
|∇w(x, t) · ∇ψ(x, t)|dxdt ≤ 4T |∇w(x, t) · ∇ψ(x, t)|dxdt
0 Ω Z 0
T Z Ω

× |∇w(x, t)|2 dxdt.

0 Ω
Therefore,
Z T Z Z T Z
|∇w(x, t) · ∇ψ(x, t)|dxdt ≤ 4T |∇w(x, t)|2 dxdt,
0 Ω 0 Ω
parnllemme7 parnllemme5
which, together with (4.87), yields (4.85).

Proof of the uniqueness theorem

paranlanw
Let u1 and u2 be two solutions of Problem (4.36), with initial conditions u0,1 and u0,2 and source terms
v1 and v2 respectively. We set ud = u1 − u2 , vd = v1 − v2 and u0,d = u0,1 − u0,2 . Let us also define, for all
ϕ(u1 (x, t)) − ϕ(u2 (x, t))
(x, t) ∈ Ω × IR ?+ , q(x, t) = if u1 (x, t) 6= u2 (x, t), else q(x, t) = 0. For all T ∈ IR ?+
u1 (x, t) −paranlanw
u2 (x, t)
and for all ψ ∈ AT , we deduce from (4.36) that
Z T Z h   i
ud (x, t) ψt (x, t) + q(x, t)∆ψ(x, t) + vd (x, t)ψ(x, t) dxdt +
Z0 Ω (4.90) parnlun1
u0,d (x)ψ(x, 0)dx = 0.
Ω

Let w ∈ Cc∞ (Ω× (0, T )), such that |w| ≤ 1. Since ϕ is locally Lipschitz continuous, we can define its
Lipschitz constant, say BM , on [−M, M ], where M = max{ku1 kL∞ (Ω×(0,T )) , ku2 kL∞ (Ω×(0,T )) } so that
0 ≤ q ≤ BM a.e. on Ω × (0, T ).

1
Using mollifiers, functions q1,n ∈ Cc∞ (Ω × (0, T )) may be constructed such that kq1,n − qkL2 (Ω×(0,T )) ≤ n
and 0 ≤ q1,n ≤ BM , for n ∈ IN? . Let qn = q1,n + n1 . Then
1 1
≤ qn (x, t) ≤ BM + , for all (x, t) ∈ Ω × (0, T ),
n n
 118

and
Z T Z  Z T Z (q (x, t) − q (x, t))2
(qn (x, t) − q(x, t))2 n 1,n
dxdt ≤ 2 dxdt +
0 Ω qn (x, t) Z0 T ZΩ q n (x, t)
(q1,n (x, t) − q(x, t))2 
dxdt ,
0 Ω qn (x, t)
which shows that
Z Z  T m(Ω)
T
(qn (x, t) − q(x, t))2 1
dxdt ≤ 2n + .
0 Ω qn (x, t) n2 n2
It leads to
qn − q
k √ kL2 (Ω×(0,T )) → 0 as n → ∞. (4.91) parnlun3
qn
parnlauxi parnlun1
Let ψn ∈ AT be given by lemma 4.8, with g = qn . Substituting ψ by ψn in (4.90), using (with g = qn
parnllemme1 parnllemme4
and ψ = ψn ) (4.81) and (4.84) give
Z TZ  
| ud (x, t) w(x, t) + (q(x, t) − qn (x, t))∆ψn (x, t) dxdt| ≤
Z 0T Z Ω Z (4.92) parnlun10
|vd (x, t)|(T − t)dxdt + T |u0,d (x)|dx.
0 Ω Ω
The Cauchy-Schwarz inequality yields
hZ T Z i2
|ud (x, t)||(q(x, t) − qn (x, t))∆ψn (x, t)|dxdt ≤ 4M 2
Z 0 Z Ω Z TZ (4.93)
q(x, t) − qn (x, t) 2  2
T parnlun11
p dxdt qn (x, t) ∆ψn (x, t) dxdt.
0 Ω qn (x, t) 0 Ω
parnllemme5 parnlun3 parnlun11
We deduce from (4.85) and (4.91) that the right hand side of (4.93) tends to zero as
parnlun11
n → ∞. Hence the
parnlun10
left hand side of (4.93) also tends to zero as n → ∞. Therefore letting n → ∞ in (4.92) gives
Z T Z Z T Z
| ud (x, t)w(x, t)dxdt| ≤ |vd (x, t)|(T − t)dxdt +
0 Ω 0Z Ω (4.94) parnlun12
T |u0,d (x)|dx.
Ω
parnlun12
Inequality (4.94) holds for any function w ∈ Cc∞ (Ω
× (0, T )), with |w| ≤ 1. Let us take as functions w
the elements of a sequence (wm )m∈IN such that wm ∈ Cc∞ (Ω × (0, T )) and |wm | ≤ 1 for all m ∈ IN, and
the sequence (wm )m∈IN converges to sign(ud (·, ·)) in L1 (Ω × (0, T )). Letting m → ∞ yields
Z T Z Z T Z Z
|ud (x, t)|dxdt ≤ |vd (x, t)|(T − t)dxdt + T |u0,d (x)|dx,
0 Ω 0 Ω Ω
parnlthunicite
which concludes the proof of Theorem 4.4.
 Chapter 5

Hyperbolic equations in the one

dimensional case
hyper1d
This chapter is devoted to the numerical schemes for one-dimensional hyperbolic conservation laws. Some
basics on the solution to linear or nonlinear godlewski-ellipses
hyperbolic equations with initial data and without boundary
godlewski-springer
conditions will first
kronerbook
be recalled.
leveque
We refer to Godlewski
serre
and Raviart [1991], Godlewski and Raviart
[1996], Kröner [1997], LeVeque [1990] and Serre [1996] for extensive studies of theoretical and/or
numerical aspects; we shall highlight here the finite volume point of view for several well known schemes,
comparing them with finite difference schemes, either for the linear and the nonlinear case. Convergence
results for numerical schemes are presented, using a “weak BV inequality” which will be used later in
the multidimensional case. We also recall the classical proof of convergence which uses a “strong BV
estimate” and the Lax-Wendroff theorem. The error estimates which can also be obtained will be given
hypmd
later in the multidimensional case (Chapter 6).

Throughout this chapter, we shall focus on explicit schemes. However, all the results which are presented
here can be extended to implicit schemes (this requires a bit of work). This will be detailed in the
estschemai estschemai
multidimensional case (see (6.9) page 150 for the scheme).

5.1 The continuous problem

hyppb
Consider the nonlinear hyperbolic equation with initial data:

ut (x, t) + (f (u))x (x, t) = 0 x ∈ IR, t ∈ IR + ,
(5.1) hyperbolic1D
u(x, 0) = u0 (x), x ∈ IR,

where f is a given function from IR to IR, of class C 1 , u0 ∈ L∞ (IR) and where the partial derivatives of
u with respect to time and space are denoted by ut and ux .

Example 5.1 (Bürgers equation) A simple flow model was introduced by Bürgers and yields the
following equation:
ut (x, t) + u(x, t)ux (x, t) − εuxx (x, t) = 0 (5.2) Burgeps
hyperbolic1D
Bürgers studied the limit case which is obtained when ε tends to 0; the resulting equation is ( 5.1) with
s2
f (s) = , i.e.
2
1
ut (x, t) + (u2 )x (x, t) = 0
2

119
 120

defclass1 Definition 5.1 (Classical solution) Let f ∈ C 1 (IR, IR) and u0 ∈ C 1 (IR, IR); a classical solution to
hyperbolic1D
Problem (5.1) is a function u ∈ C 1 (IR × IR + , IR) such that

ut (x, t) + f 0 (u(x, t))ux (x, t) = 0, ∀ x ∈ IR, ∀ t ∈ IR + ,
u(x, 0) = u0 (x), ∀ x ∈ IR.

Recall that in the linear case, i.e. f (s) = cs for all s ∈ IR, for some c ∈ IR, there exists (for u 0 ∈ C 1 (IR, IR))
a unique classical solution. It is u(x, t) = u0 (x − ct), for all x ∈ IR and for all t ∈ IR + . In the nonlinear
case, the existence of such a solution depends on the initial data u0 ; in fact, the following result holds:

nocr Proposition 5.1 Let f ∈ C 1 (IR, IR) be a nonlinear function, i.e. such that there exist s1 , s2 ∈ IR with
hyperbolic1D
f 0 (s1 ) 6= f 0 (s2 ); then there exists u0 ∈ Cc∞ (IR, IR) such that Problem (5.1) has no classical solution.
nocr
Proposition 5.1 is an easy consequence of the following remark.
hyperbolic1D
Remark 5.1 If u is a classical solution to (5.1), then u is constant along the characteristic lines which
are defined by
x(t) = f 0 (u0 (x0 ))t + x0 , t ∈ IR + ,
where x0 ∈ IR is the origin of the characteristic. This is the equation of a straight line issued from the
point (x0 , 0) (in the (x, t) coordinates). Note that if f depends on x and u (rather than only on u), the
characteristics are no longer straight lines.
hyperbolic1D
The concept of weak solution is introduced in order to define solutions of (5.1) when classical solutions
do not exist.

defweak1 Definition
hyperbolic1D
5.2 (Weak solution) Let f ∈ C 1 (IR, IR) and u0 ∈ L∞ (IR); a weak solution to Problem
(5.1) is a function u such that

 u ∈ L∞ (IR × IR ?+ ),
 Z Z
 Z Z Z
u(x, t)ϕt (x, t)dtdx + f (u(x, t))ϕx (x, t)dtdx + u0 (x)ϕ(x, 0)dx = 0, (5.3) h1dweak


 IR IR + IR IR + IR
1
∀ϕ ∈ Cc (IR × IR + , IR).

Remark 5.2
1. If u ∈ C 1 (IR × IR + , IR) ∩ L∞ (IR × IR ?+ ) then u is a weak solution if and only if u is a classical solution.
2. Note that in the above definition, we require the test function ϕ to belong to Cc1 (IR × IR + , IR), so that
ϕ may be non zero at time t = 0.
hyperbolic1D
One may show that there exists at least one weak solution to (5.1). In the linear case, i.e. f (s) = cs, for
all s ∈ IR, for some c ∈ IR, this solution is unique (it is u(x, t) = u0 (x − ct) for a.e. (x, t) ∈ IR × IR + ).
However, the uniqueness of this weak solution in the general nonlinear case is no longer true. Hence the
concept of entropy weak solution, for which an existence and uniqueness result is known.

defent1 Definition 5.3 (Entropy weak solution) Let f ∈ C 1 (IR, IR) and u0 ∈ L∞ (IR); the entropy weak
hyperbolic1D
solution to Problem (5.1) is a function u such that

 ∞ ?


 Z ∈ ZL (IR × IR + ),
u Z Z Z

 η(u(x, t))ϕt (x, t)dtdx + Φ(u(x, t))ϕx (x, t)dtdx + η(u0 (x))ϕ(x, 0)dx ≥ 0,
IR IR + IR IR + IR (5.4) h1dent



 ∀ϕ ∈ Cc1 (IR × IR + , IR + ),

for all convex function η ∈ C (IR, IR) and Φ ∈ C (IR, IR) such that Φ0 = η 0 f 0 .
1 1
 121

h1dent h1dweak
Remark
h1dent
5.3 The solutions of (5.4) are necessarily solutions of (5.3). This can be shown by taking in
(5.4) η(s) = s for all s ∈ IR, η(s) = −s, for all s ∈ IR, and regularizations of the positive and negative
parts of the test functions of the weak formulation.

hyp1dexistu 5.1 Let f ∈ C 1 (IR, IR), u0 ∈ L∞ (IR), then there exists a unique entropy weak solution to
Theoremhyperbolic1D
Problem (5.1).
Vo
The proof of this result was first given by Vol’pert in Vol’pertOl [1967], introducing the space BV (IR)
which is defined hereafter and assuming u0 ∈ BV (IR), see also Oleinik [1963] for the convex case. In
krushkov
Krushkov [1970], Krushkov proved the theorem of existence and uniqueness in the general case u 0 ∈
L∞ (IR), using a regularization of u0 in BV (IR), under the slightly stronger assumption f ∈ C 3 (IR, IR).
Krushkov also proved that the solution is in the space C(IR + , L1loc (IR)). Krushkov’s proof uses particular
entropies, namely the functions | · −κ| for all κ ∈ IR, which are generally referred to as “Krushkov’s
entropies”. The “entropy flux” associated to | · −κ| may be taken as f (·>κ) − f (·⊥κ), where a>b denotes
the maximum of a and b and a⊥b denotes the minimum of a and b, for all real values a, b (recall that
f (a>b) − f (a⊥b) = sign(a − b)(f (a) − f (b))).

bv Definition 5.4 (BV (IR)) A function v ∈ L1loc (IR) is of bounded variation, that is v ∈ BV (IR), if
Z
|v|BV (IR) = sup{ v(x)ϕx (x)dx, ϕ ∈ Cc1 (IR, IR), |ϕ(x)| ≤ 1 ∀x ∈ IR} < +∞. (5.5) defBV
IR

divbv1d Remark 5.4

1. If v : IR → IR is piecewise constant, that is if there exists an increasing sequence (xP i )i∈ZZ with IR =
∪i∈ZZ [xi , xi+1 ] and a sequence (vi )i∈ZZ such that v|(xi ,xi+1 ) = vi , then |v|BV (IR) = i∈ZZ |vi+1 − vi |.
2. If v ∈ C 1 (IR, IR) then |v|BV (IR) = kvx kL1 (IR) .
3. The space BV (IR) is included in the space L∞ (IR); furthermore, if u ∈ BV (IR) ∩ L1 (IR) then
kukL∞(IR) ≤ |u|BV (IR) .
4. Let u ∈ BV (IR) and let (xi+1/2 )i∈ZZ be an increasing sequence of real values such that IR =
∪i∈ZZ [xi−1/2 , xi+1/2 ]. For i ∈ ZZ , let Ki = (xi−1/2 , xi+1/2 ) and ui be the mean value of u over Ki .
Then, choosing conveniently ϕ in the definition of |u|BV (IR) , it is easy to show that
X
|ui+1 − ui | ≤ |u|BV (IR) . (5.6) bvz
i∈ZZ
bvz
Inequality (5.6) is used for the classical nlstabv
proof of “BV estimates” fornlstabvco
nlstabv
the approximate
nlstabvco
solutions given
by finite volume schemes (see Lemma 5.7 page 139 and Corollary 5.1 page 139).
bvz
Note that (5.6) is also true when ui is the mean value of u over a subinterval of Ki instead of the
mean value of u over Ki .

Krushkov used a characterization of entropy weak solutions which is given in the following proposition.

prentkr1 Proposition 5.2 (Entropy weak solution using “Krushkov’s entropies”) hyperbolic1D
Let f ∈ C 1 (IR, IR)
∞
and u0 ∈ L (IR), u is the unique entropy weak solution to Problem (5.1) if and only if u is such that
 ∞ ?
 u

 Z ∈ ZL (IR × IR + ),



 |u(x, t) − κ|ϕt (x, t)dtdx+
ZIR ZIR +   Z (5.7) h1dentkr

 f (u(x, t)>κ) − f (u(x, t)⊥κ) ϕ (x, t)dtdx + |u (x) − κ|ϕ(x, 0)dx ≥ 0,

 x 0


 IR IR + IR
∀ϕ ∈ Cc1 (IR × IR + , IR + ), ∀κ ∈ IR.
 122

h1dent
The result of existence of an entropy weak solution defined by (5.4) was Ol
already proved by passing to
the limit on the solutions of an appropriate numerical scheme, see e.g. Oleinik [1963], and maynlcv1dalso be
obtained
nlcv1d
by passing to the limit on finite volume approximations
eghbcewsoleghbcewsol
of the solution (see Theorem 5.2 page
136 in the one-dimensional case and Theorem 6.4 page 179 in the multidimensional case).

Remark 5.5 An entropy weak solution is sometimes defined as a function u satisfying:

 Z Z Z Z Z

 u(x, t)ϕ (x, t)dtdx + f (u(x, t))ϕ (x, t)dtdx + u0 (x)ϕ(x, 0)dx = 0,

 t x

 IR IR + IR IR + IR


 Z Z Z Z ∀ϕ ∈ Cc1 (IR × IR + , IR).
(5.8) h1dentf

 η(u(x, t))ϕt (x, t)dtdx + Φ(u(x, t))ϕx (x, t)dtdx ≥ 0,

 IR IR + IR IR +



 ∀ϕ ∈ Cc1 (IR × IR ?+ , IR + ),

for all convex function η ∈ C (IR, IR) and Φ ∈ C (IR, IR) such that Φ0 = η 0 f 0 .
1 1

The uniqueness of an entropy weak solution thus defined depends hyp1dexistu

on the functional space to which u is
chosen
h1dentf
to belong. Indeed, the uniqueness result given in Theorem 5.1 is no longer true with u defined by
(5.8) such that

u, f (u) ∈ L1loc (IR × IR + ), u ∈ L∞ (IR × (ε, ∞)), ∀ε ∈ IR ?+ . (5.9) h1space

h1space h1dentf h1space
h1dentf h1dent
Under Assumption (5.9), every term in (5.8) makes sense. Note h1dentf
that (h1space
5.9)-(5.8) is weaker than (5.4). An
easy counterexample to a uniqueness result of the solution to (5.8)-(5.9) is obtained
h1dentf
with f (s) = s2 for
h1space
all s ∈ IR and u0 (x) = 0 for a.e. x ∈ IR. In this case, a first solution to (5.8)-( 5.9)
h1dentf
is u(x, t) = 0 for
h1space
a.e. (x, t) ∈ IR × IR + (it is the entropy weak solution). A second solution to (5.8)-(5.9) is defined for a.e.
(x, t) ∈ IR × IR + by
√ √
− t or x √
u(x, t) = 0, if x < √ > t,
x
u(x, t) = 2t , if − t < x < t.
h1dent
This second solution is not an entropy weak solution: it does not satisfy (5.4). Also note that this second
solution is not in the space C(IR + , L1loc (IR)) nor in the space L∞ (IR×IR + ) (it belongs toh1dentf
L∞ (IR + , L1 (IR))).
∞ 1
Indeed, under the assumption u ∈ L (IR × IR + ) ∩ C(IR + , Lloc (IR)), the solution of (5.8) is unique.
hyperbolic1D
The entropy weak solution to (5.1) satisfies the following L∞ and BV stability properties:
hyperbolic1D
Proposition 5.3 Let f ∈ C 1 (IR, IR) and u0 ∈ L∞ (IR). Let u be the entropy weak solution to (5.1).
Then, u ∈ C(IR + , L1loc (IR)); furthermore, the following estimates hold:

1. ku(·, t)kL∞(IR) ≤ ku0 kL∞ (IR) , for all t ∈ IR + .

2. If u0 ∈ BV (IR), then |u(·, t)|BV (IR) ≤ |u0 |BV (IR) , for all t ∈ IR + .

5.2 Numerical schemes in the linear case

hyperbolic1D
We shall first introduce the numerical schemes in the linear case f (u) = u in (5.1). The problem considered
in this section is therefore

ut (x, t) + ux (x, t) = 0 x ∈ IR, t ∈ IR + ,
(5.10) lhy1D
u(x, 0) = u0 (x), x ∈ IR.
lhy1D defclass1
Assume that u0 ∈ C 1 (IR, IR); Problem (5.10) has a unique classical solution, as definedlhy1D
in Definition 5.1,
∈ IR × IR + . If u0 ∈ L∞ (IR), then Problem (5.10) has a unique
which is u(x, t) = u0 (x − t) for all (x, t)defweak1
weak solution, as defined in Definition 5.2, which is again u(x, t) = u0 (x − t) for a.e. (x, t) ∈ IR × IR + .
Therefore, if u0 ≥ 0, the solution u is also nonnegative. Hence, it is advisable for many problems that
the solution given by the numerical scheme should preserve the nonnegativity of the solution.
 123

5.2.1 The centered finite difference scheme

Assume u0 ∈ C(IR, IR). Let h ∈ IR ?+ and xi = ih for all i ∈ ZZ . Let k ∈ IR ?+ be the time step. With
the explicit Euler scheme for the time discretization (the implicit Euler scheme could also be used), the
centered finite difference scheme associated to points xi and k is

 un+1 − uni un − uni−1
i
+ i+1 = 0, ∀n ∈ IN, ∀i ∈ ZZ , (5.11) DFCl
 k 2h 0
ui = u0 (xi ), ∀i ∈ ZZ .
The
lhy1D
discrete unknown uni is expected to be an approximation of u(xi , nk) where u is the solution to
(5.10).
It is well known that this scheme should be avoided. In particular, for the following reasons:

1. it does not preserve positivity, i.e. u0i ≥ 0 for all i ∈ ZZ does not imply u1i ≥ 0 for all i ∈ ZZ ; take
for instance u0i = 0 for i ≤ 0 and u0i = 1 for i > 0, then u10 = −k/(2h) < 0;
2. it is not “L∞ -diminishing”, i.e. max{|u0i |, i ∈ ZZ } = 1 does not imply that max{|u1i |, i ∈ ZZ } ≤ 1;
for instance, in the previous example, max{|u0i |, i ∈ ZZ } = 1 and max{|u1i |, i ∈ ZZ } = 1 + k/(2h);
P P
3. it is not “L2 -diminishing”, i.e. 0 2
i∈ZZ (ui ) = 1 does not imply that
1 2
i∈ZZ (ui ) ≤ 1; take for
0 0 1 1 1
instance
P ui = 0 for i 6= 0 and ui = 1 for i = 0, then u0 = 1, u1 = k/(2h), u−1 = −k/(2h), so that
1 2 2 2
i∈ZZ (u i ) = 1 + k /(2h ) > 1;

4. it is unstable in the von Neumann sense: if the initial condition is taken under the form u0 (x) =
exp(ipx), where p is given in ZZ , then u(x, t) = exp(−ipt) exp(ipx) (i is, here, the usual complex
number, u0 and u take values in Cl). Hence exp(−ipt) can be seen as an amplification factor, and
its modulus is 1. The numerical scheme is stable in the von Neumann senseDFCl if the amplification
factor for the discrete solution is less than or equal to 1. For the scheme (5.11), we have u1j =
u0j − (u0j+1 − u0j−1 )k/(2h) = exp(ipjh)ξp,h,k , with ξp,h,k = 1 − (exp(iph) − exp(−iph))k/(2h). Hence
|ξp,h,k |2 = 1 + (k 2 /h2 ) sin2 ph > 1 if ph 6= qπ for any q in ZZ .

In fact, one can also show that there exists u0 ∈ Cc1 (IR, IR) such that the solution given by the numerical
scheme does not tend to the solution of the continuous problem when h and k tend to 0 (whatever the
relation between h and k).
DFCl
Remark 5.6 The scheme (5.11) is also a finite volume scheme with the (spatial) mesh T given by
meshhyp1d
xi+1/2 = (i + 1/2)h in Definition 5.5 below and with a centered choice for theVFDA approximation of
u(xi+1/2 , nk): the value of u(xi+1/2 , nk) is approximated by (uni + uni+1 )/2, see (5.14)VFDA
where anDFCl
up-
stream choice for u(xi+1/2 , nk) is performed. In fact, the choice of u0i is different in (5.14) and in (5.11)
but this does not change the unstability of the centered scheme.

5.2.2 The upstream finite difference scheme

Consider now a nonuniform distribution of points xi , i.e. an increasing sequence of real values (xi )i∈ZZ
such that limi→±∞ xi = ±∞. For all i ∈ ZZ , we set hi−1/2 = xi − xi−1 . The time discretization is
performed with the explicit Euler scheme with time step k > 0. Still assuming u0 ∈ C(IR, IR), consider
the upwind (or upstream) finite difference scheme defined by
 n+1 n n n
 ui − ui + ui − ui−1 = 0, ∀n ∈ IN, ∀i ∈ ZZ ,
k hi− 12 (5.12) DFDA

u0i = u0 (xi ), ∀i ∈ ZZ .
Rewriting the scheme as
 124

k k
un+1
i = (1 − )uni + uni−1 ,
hi− 21 hi− 21

it appears that if inf i∈ZZ hi−1/2 > 0 and if k is such that k ≤ inf i∈ZZDFDA hi−1/2 then un+1
i is a convex
n n
combination of ui and ui−1 ; by induction, this proves that the scheme (5.12) is stable, in the sense that
if u0 is such that Um ≤ u0 (x) ≤ UM for a.e. x ∈ IR, where Um , UM ∈ IR, then Um ≤ uni ≤ UM for any
i ∈ ZZ and n ∈ IN.
Moreover, if u0 ∈ C 2 (IR, IR)∩L∞ (IR) and u00 and u000 belong to L∞ (IR), it is easily shown that the scheme
is consistent in the finite difference sense, i.e. the consistency error defined by

u(xi , (n + 1)k) − u(xi , nk) u(xi , nk) − u(xi−1 , nk)

Rin = + (5.13) consisthypli
k hi− 21
n
is such that |R i | ≤ Ch, where h = supi∈ZZ hi and C ≥ 0 only depends on u0 (recall that u is the solution
lhy1D
to problem (5.10)). Hence the following error estimate holds:

eefd Proposition 5.4 (Error estimate for the upwind finite difference scheme)
Let u0 ∈ C 2 (IR, IR)∩L∞ (IR), such that u00 and u000 ∈ L∞ (IR). Let (xi )i∈ZZ be an increasing sequence of real
values such that limi→±∞ xi = ±∞. Let h = supi∈ZZ hi− 12 , and assume that h < ∞ and inf i∈ZZ hi−1/2 >
lhy1D
0. Let k > 0 such that k ≤ inf i∈ZZ hi−1/2 . Let u denote the unique solution to (5.10) and {uni , i ∈ ZZ ,
DFDA
n ∈ IN} be given by (5.12); let eni = u(xi , nk) − uni , for any n ∈ IN and i ∈ ZZ , and let T ∈]0, +∞[ (note
that u(xi , nk) is well defined since u ∈ C 2 (IR × IR + , IR)).
Then there exists C ∈ IR + , only depending on u0 , such that |eni | ≤ ChT, for any n ∈ IN such that nk ≤ T ,
and for any i ∈ ZZ .
eefd
Proof of Proposition 5.4
consisthyplin
Let i ∈ ZZ and n ∈ IN. By definition of the consistency error Rin in (5.13), the error eni satisfies

en+1 − eni en − eni−1

i
+ i = Rin .
k hi− 12
Hence
k k
en+1
i = eni (1 − )+ eni−1 + kRin .
hi− 21 hi− 12
Using |Rin | ≤ Ch (for some C only depending on u0 ) and the assumption k ≤ inf i∈ZZ hi−1/2 , this yields

|en+1
i | ≤ sup |enj | + Ckh.
j∈ZZ

Since e0i = 0 for any i ∈ ZZ , an induction yields

sup |eni | ≤ Cnkh

i∈ZZ

and the result follows.

Note that in the above proof, the linearity of the equation and the regularity of u 0 are used. The next
questions to arise are what to do in the case of a nonlinear equation and in the case u 0 ∈ L∞ (IR).
 125

5.2.3 The upwind finite volume scheme

Let us first give a definition of the admissible meshes for the finite volume schemes.

Definition 5.5 (One-dimensional admissible mesh) An admissible mesh T of IR is given by an

increasing sequence of real values (xi+1/2 )i∈ZZ , such that IR = ∪i∈ZZ [xi−1/2 , xi+1/2 ]. The mesh T is the
set T = {Ki , i ∈ ZZ } of subsets of IR defined by Ki = (xi−1/2 , xi+1/2 ) for all i ∈ ZZ . The length of Ki
is denoted by hi , so that hi = xi+1/2 − xi−1/2 for all i ∈ ZZ . It is assumed that h = size(T ) = sup{hi ,
meshhyp1d i ∈ ZZ } < +∞ and that, for some α ∈ IR ?+ , αh ≤ inf{hi , i ∈ ZZ }.
meshhyp1d
Consider an admissible mesh in the sense of Definition 5.5. Let k ∈ IR ?+ be the time step.lhy1D
Assume
u0 ∈ L∞ (IR) (this is a natural hypothesis for the finite volume framework). Integrating ( 5.10) on each
control volume of the mesh, approximating the time derivatives by differential quotients and using an
upwind choice for u(xi+ 12 , nk) yields the following (time explicit) scheme:

 un+1 − uni
 hi i + uni − uni−1 = 0, ∀n ∈ IN, ∀i ∈ ZZ ,
k Z (5.14) VFDA
 1
 u0i = u0 (x)dx, ∀i ∈ ZZ .
hi K i
lhy1D
The value uni is expected to be an approximation of u (solution to (5.10)) in Ki at time nk. It is
easily shown that this scheme is not consistent in the finite difference sense if uni is considered to be
an approximation of u(xi , nk) with, for instance, xi = (xi−1/2 + xi+1/2 )/2 for all i ∈ ZZ . Even if
consisthyplin
u0 ∈ Cc∞ (IR, IR), the quantity Rin defined by (5.13) does not satisfy (except in particular cases) |Rin | ≤ Ch,
with some C only depending on u0 .
It is however
DFDA
possible to interpret this scheme as another expression of the upwind finite difference
scheme (5.12) (except for the minor modification of u0i , i ∈ ZZ ). One simply needs to consider uni as
an approximation of u(xi+1/2 , nk) which leads to a consistency property in the finite difference sense.
consisthyplin
Indeed, taking xj = xj+1/2 (for j = i and i − 1) in the definition (5.13) of Rin yields |Rin | ≤ Ch, eefd
where C
only depends on u0 . Therefore, a convergence result for this scheme is given by the proposition 5.4. This
monflux monflux
analogy cannot be extended to the general case of “monotone flux schemes” (see Definition 5.6 page 131
below) for a nonlinear equation for which
vfordf
therevfordf
may be no value of xi (independant of u) leading to such
a consistencyup1D
property, see Remark 5.11 page 131 for a counterexample (the analogy holds however for
vfdf
the scheme (5.28), convenient for a nondecreasing function f , see Remark 5.13).

uT ,k may be defined on IR × IR + from the discrete unknowns uni ,

The approximate finite volume solution VFDA
i ∈ ZZ , n ∈ IN which are computed in (5.14):

uT ,k (x, t) = uni for x ∈ Ki and t ∈ [nk, (n + 1)k). (5.15) uapp

∞
The following L estimate holds:

linstab Lemma 5.1 (L∞ estimate in the linear case) Let u0 ∈ L∞ (IR) and Um , UM ∈ meshhyp1d IR such that Um ≤
u0 (x) ≤ UM for a.e. x ∈ IR. Let T be an admissible mesh in the sense of Definition 5.5 and let k ∈ IR ?+
satisfying the Courant-Friedrichs-Levy (CFL) condition

k ≤ inf hi .
i∈ZZ

(note that taking

VFDA
k ≤ αh implies the above condition). Let uT ,k be the finite volume approximate solution
uapp
defined by (5.14) and (5.15).
Then,

Um ≤ uT ,k (x, t) ≤ UM for a.e. x ∈ IR and a.e. t ∈ IR + .

 126

linstab
Proof of Lemma 5.1
The proof that U ≤ uniDFDA
DFDA m
≤ UM , for all i ∈ ZZ and n ∈ IN, as in theVFDA
case of the upwind finite difference
scheme (see (5.12) page 123), consists in remarking that equation (5.14) gives, under the CFL condition,
an expression of un+1
i as a linear convex combination of uni and uni−1 , for all i ∈ ZZ and n ∈ IN.

The following inequality will be crucial for the proof of convergence.

bvwhyplin Lemma 5.2 (Weak BV estimate, linear case) Let T be an admissible mesh in the sense of Defini-
meshhyp1d meshhyp1d
tion 5.5 page 125 and let k ∈ IR ?+ satisfying the CFL condition

k ≤ (1 − ξ) inf hi , (5.16) cflhypl

i∈ZZ

for some ξ ∈ (0, 1) (taking k ≤ (1 − ξ)αh implies this condition). VFDA

Let {uni , i ∈ ZZ , n ∈ IN} be given by the finite volume scheme (5.14). Let R ∈ IR ?+ and T ∈ IR ?+ and
assume h = size(T ) < R, k < T . Let i0 ∈ ZZ , i1 ∈ ZZ and N ∈ IN be such that −R ∈ K i0 , R ∈ K i1 and
T ∈ (N k, (N + 1)k] (note that i0 < i1 ).
Then there exists C ∈ IR ?+ , only depending on R, T , u0 , α and ξ, such that
i1 X
X N
k|uni − uni−1 | ≤ Ch−1/2 . (5.17) bvf
i=i0 n=0
bvwhyplin
Proof of Lemma 5.2
VFDA
Multiplying the first equation of (5.14) by kuni and summing on i = i0 , . . . , i1 and n = 0, . . . N yields
A + B = 0 with
i1 X
X N
A= hi (un+1
i − uni )uni
i=i0 n=0

and
i1 X
X N
B= k(uni − uni−1 )uni .
i=i0 n=0

Noting that
i1 XN i1
1X 1X
A=− hi (un+1
i − u n 2
i ) + hi [(uN
i
+1 2
) − (u0i )2 ]
2 i=i n=0 2 i=i
0 0
VFDA
and using the scheme (5.14) gives
i1 XN i1
1X k2 n 1X
A=− (ui − uni−1 )2 + hi [(uN
i
+1 2
) − (u0i )2 ];
2 i=i n=0 hi 2 i=i
0 0
cflhypl
therefore, using the CFL condition (5.16),
i1 XN i1
1X n n 2 1X
A ≥ −(1 − ξ) k(ui − ui−1 ) − hi (u0i )2 .
2 i=i n=0 2 i=i
0 0

We now study the term B, which may be rewritten as

i1 XN N
1X 1X
B= k(uni − uni−1 )2 + k[(uni1 )2 − (uni0 −1 )2 ].
2 i=i n=0 2 n=0
0
 127

linstab linstab
Thanks to the L∞ estimate of Lemma 5.1 page 125, this last equality implies that
i1 XN
1X
B≥ k(uni − uni−1 )2 − T max{−Um , UM }2 .
2 i=i n=0
0
P i1
Therefore, since A + B = 0 and i=i0 hi ≤ 4R, the following inequality holds:
i1 X
X N
0≥ξ k(uni − uni−1 )2 − (4R + 2T ) max{−Um , UM }2 ,
i=i0 n=0

which, in turn, gives the existence of C1 ∈ IR ?+ , only depending on R, T , u0 and ξ such that
i1 X
X N
k(uni − uni−1 )2 ≤ C1 . (5.18) bvfsq
i=i0 n=0

Finally, using
i1
X Xi1
hi 4R
1≤ ≤ ,
i=i0 i=i
αh αh
0

the Cauchy-Schwarz inequality leads to

i1 X
X N
4R
[ k|uni − uni−1 |]2 ≤ C1 2T ,
i=i0 n=0
αh

which concludes the proof of the lemma.

Contrary to the discrete H01 estimates

ellestimx
which were obtained on
bvf
the approximate finite volume solutions
of elliptic equations, see e.g. (3.24), the weaklhy1D
BV estimate (5.17) is not related to an a priori estimate
on the solution to the continuous problem (5.10). It does not give any compactness property in the
space L1loc (IR) (there are some counterexamples); such a compactness property is obtained thanks
helly
to a
“strong BV estimate” (with, for instance, an L∞ estimate) as it is recalled below (see Lemma 5.6). In the
one-dimensional case which is studied here such a “strong BV estimate” can be obtained if u 0 ∈ BV (IR),
nlstabvco
see Corollary 5.1; this is no longer true in the multidimensional case with general meshes, for which only
the above weak BV estimate is available.

rbvweff Remark 5.7 The weak BV estimate is a crucial point for the proof of convergence. Indeed, the property
convhyplin
which is
bvwhyplin
used in the proof of convergence (see Proposition 5.5 below) is, with the notations of Lemma
5.2,
i1 X
X N
h k|uni − uni−1 | → 0, as h → 0, (5.19) bvweff
i=i0 n=0

for R, T , u0 , α and ξ fixed.

uapp
If a piecewise
VFDA
constant function uT ,k , such as given by (5.15) (with some uni in IR, not necessarily given
by (5.14)), is bounded in (for instance) L∞ (IR × IR + ) and converges in L1loc (IR × IR + ) as h → 0 and
bvweff
→ 0 (with a possible relation between k and h) then (5.19) holds. This proves that the hypothesis
kbvweff
(laxw
5.19) is included
laxw
in the hypotheses
bvweff
of the classical
bvf
Lax-Wendroff theorem of convergence
bvf
(see Theorem
5.3 page 140); note that (5.19) is implied by (5.17) and that it is weaker than (5.17)).

We show in the following remark how the “ weak” and “ strong” BV estimates may “formally” be
obtained on the “continuous equation”; this gives a hint of the reason why this estimate may be obtained
belong to the space BV (IR × IR + ). A similar remark also holds in the
even if the exact solution does nothyperbolic1D
nonlinear case (i.e. for Problem (5.1)).
 128

Remark 5.8 (Formal derivations of the strong and weak

lhy1D VFDA
BV estimates) When approximat-
ing the solution to (5.10) by the finite volume scheme (5.14) (with hi = h for all i, for the sake of
simplicity), the equation to which an approximation of a solution is sought is “close” to the equation

ut + ux − εuxx = 0 (5.20) rem

cflhypl
where ε = h−k
2 is positive under the CFL condition (5.16), which ensures that the scheme is diffusive.
We assume that u is regular enough, with null limits for u(x, t) and its derivatives as x → ±∞.

(i) “Strong” BV estimate.rem

Derivating the equation (5.20) with respect to the variable x, multiplying by signr (ux (x, t)), where signr
denotes a nondecreasing regularization of the function sign, and integrating over IR yields
Z Z Z


φr (ux (x, t))dx t + uxx (x, t)signr (ux (x, t))dx = −ε sign0r (ux (x, t))(uxx (x, t))2 dx ≤ 0,
IR IR IR

where φ0r = signr and φr (0) = 0. Since

Z Z
uxx (x, t)signr (ux (x, t))dx = (φr (ux (x, t)))x dx = 0,
IR IR

this yields, passing to the limit on the regularization, that kux (·, t)kL1 (IR) is nonincreasing with respect to
t. Copying this formal proof on the numerical scheme yields a strong BV estimate, which
nlstabv
is an a priori
nlstabvco helly
1
estimate giving
helly
compactness properties in L loc (IR × IR + ), see Lemma 5.7, Corollary 5.1 and Lemma 5.6
page 138.

(ii) “Weak” BV
rem
estimate
Multiplying (5.20) by u and summing over IR × (0, T ) yields
Z Z Z T Z
1 1
u2 (x, T )dx − u2 (x, 0)dx + εu2x (x, t)dxdt = 0,
2 IR 2 IR 0 IR
which yields in turn
Z T Z
1
ε u2x (x, t)dxdt ≤ ku0 k2L2 (IR) .
0 IR 2
bvfsq cflhypl
This is the continuous analogous of (5.18). Hence if h − k = ε ≥ ξh (this is Condition (5.16), note that
this condition is more restrictive than the usual CFL condition
bvfsq bvf
required for the L∞ stability), the discrete
equivalent of this formal proof yields (5.18) (and then (5.17)).

In the first case, we derivate the equation and we use some regularity on u0 (namely u0 ∈ BV (IR)). In
the second case, it is sufficient to have u0 ∈ L∞ (IR) but we need the diffusion term to be large enough
rem
in order to obtain the estimate which, by the way, does not yield any estimate on the solution of ( 5.20)
with ε = 0. This formal derivation may be carried out similarly in the nonlinear case.
VFDA
Let us now give a convergence result for the scheme (5.14) in L∞ (IR × IR ?+ ) for the weak-? topology.
Recall that a sequence (vn )n∈IN ⊂ L∞ (IR × IR ?+ ) converges to v ∈ L∞ (IR × IR ?+ ) in L∞ (IR × IR ?+ ) for the
weak-? topology if
Z Z
(vn (x, t) − v(x, t))ϕ(x, t)dxdt → 0 as n → ∞, ∀ϕ ∈ L1 (IR × IR ?+ ).
IR + IR
hyp1dnl
A stronger convergence result is available, and comes from the nonlinear study given in Section 5.3.
 129

convhyplin Proposition 5.5lhy1D (Convergence

lhy1D
in the linear case) Let u 0 ∈ L∞
defweak1
(IR) and u be the unique weak solu-
defweak1
tion to Problem (5.10) page 122 in the sense of Definition 5.2 page 120, with f (s) = s for all s ∈ IR. Let
meshhyp1d meshhyp1d
ξ ∈ (0, 1) and α > 0 be given. Let T be an cflhypl
admissible cflhypl
mesh in the sense of Definition 5.5 page 125 and
let k ∈ IR ?+ satisfying the CFL condition (5.16) page 126 (taking k ≤ (1 − ξ)αh implies this condition,
note that ξ and α do not depend on T ). VFDA uapp
Let uT ,k be the finite volume approximate solution defined by (5.14) and (5.15). Then uT ,k → u in
L∞ (IR × IR ?+ ) for the weak-? topology as h = size(T ) → 0.
convhyplin
Proof of Proposition 5.5
convhyplin
Let (Tm , km )m∈IN be a sequence of meshes and time steps satisfying the hypotheses of Proposition 5.5
that size(Tm ) → 0 as m → ∞.
and suchlinstab
Lemma 5.1 gives the existence of a subsequence, still denoted by (Tm , km )m∈IN , and of a function u ∈
L∞ (IR × IR ?+ ) such that uTm ,km → u in h1dweak
L∞ (IR × IR ?+ ) for the weak-? topology, as m → +∞. There
u is the solution of (5.3) (with f (s) = s for all s ∈ IR). The uniqueness of the weak
remains to show that lhy1D
solution to Problem (5.10) will then imply that the full sequence converges to u.
Let ϕ ∈ Cc1 (IR × IR , IR). Let m ∈ IN and T = Tm , k = km and h = size(T ). Let us multiply the first
VFDA +
equation of (5.14) by (k/hi )ϕ(x, nk), integrate over x ∈ Ki and sum for all i ∈ ZZ and n ∈ IN. This
yields

Am + B m = 0
with
X X Z
Am = (un+1
i − uni ) ϕ(x, nk)dx
i∈ZZ n∈IN Ki

and
X X Z
1
Bm = k(uni − uni−1 ) ϕ(x, nk)dx.
i∈ZZ n∈IN
hi Ki

Let us remark that Am = A1,m − A01,m with

Z ∞Z Z
A1,m = − uT ,k (x, t)ϕt (x, t − k)dxdt − u0 (x)ϕ(x, 0)dx
k IR IR
and
X Z Z
A01,m = u0i ϕ(x, 0)dx − u0 (x)ϕ(x, 0)dx.
i∈ZZ Ki IR
P
Using the fact that i∈ZZ u0i 1Ki → u0 in L1loc (IR) as m → ∞, we get that A01,m → 0 as m → ∞. (Recall
that 1Ki (x) = 1 if x ∈ Ki and 1Ki (x) = 0 if x ∈ / Ki .)
Therefore, since uT ,k → u in L∞ (IR ×IR ?+ ) for thecflhypl
weak-? topology as m → ∞, and ϕt (·, ·−k)1IR×(k,∞) →
ϕt in L1 (IR × IR ?+ ) (note that k → 0 thanks to (5.16)),
Z Z Z
lim Am = lim A1,m = − u(x, t)ϕt (x, t)dxdt − u0 (x)ϕ(x, 0)dx.
m→+∞ m→+∞ IR + IR IR

Let us now turn to the study of Bm . We compare Bm with

XZ (n+1)k Z
B1,m = − uT ,k (x, t)ϕx (x, nk)dxdt,
n∈IN nk IR
R R
which tends to − IR + IR
u(x, t)ϕx (x, t)dxdt as m → ∞. The term B1,m can be rewritten as
 130

X X
B1,m = k(uni − uni−1 )ϕ(xi− 21 , nk).
i∈ZZ n∈IN

IR ?+ , only
Let R > 0 and T > 0 be such that ϕ(x, t) = 0 if |x| ≥ R or t ≥ T . Then, there exists C ∈ cflhypl
depending on ϕ, such that, if h < R and k < T (which is true for h small enough, thanks to (5.16)),
i1 X
X N
|Bm − B1,m | ≤ Ch k|uni − uni−1 |, (5.21) eqcvhl2
i=i0 n=0

∈ ZZ , i1 ∈ ZZ and
where i0eqcvhl2 N ∈ IN are such that −R ∈ K i0 , R ∈ K i1 and T ∈ (N k, (N + 1)k].
bvwhyplin
Using (5.21) and Lemma 5.2, we get that Bm − B1,m → 0 and then
Z Z
Bm → − u(x, t)ϕx (x, t)dxdt as m → ∞,
IR + IR
lhy1D lhy1D
which completes the proof that u is bvweff
bvwhyplin
the weak solution to Problem (5.10) page 122 (note that here the
useful consequence of lemma 5.2 is (5.19)).

convhyplin
Remark 5.9 In Proposition 5.5, a simplerVFDA proof of convergence could be achieved, with ξ = 0, using
a multiplication of the first equation of (5.14) by (k/hi )ϕ(xi−1/2 , nk). However, this proof does not
generalize to the general case of nonlinear hyperbolic problems.
DFDA
Remark 5.10 Proving the convergence of the finite difference method (with the scheme ( 5.12)) with
u0 ∈ L∞ (IR) can be done using the same technique as the proof of the finite volume method (that is
considering the finite difference scheme as a finite volume scheme on a convenient mesh).

5.3 The nonlinear case

hyp1dnl
hyperbolic1D
In this section, finite volume schemes for the discretization of Problem (5.1) are presented and a theorem
nlcv1d
of convergence is given (Theorem 5.2) which will be generalized to the multidimensional case in the next
chapter. We also recall the classical proof of convergence which uses a “strong BV estimate” and the
Lax-Wendroff theorem. This proof, however, does not seem to extend to the multidimensional case for
hyperbolic1D
general meshes. The following properties are assumed to be satisfied by the data of problem ( 5.1).

datah1d Assumption 5.1 The flux function f belongs to C 1 (IR, IR), the initial data u0 belongs to L∞ (IR) and
Um , UM ∈ IR are such that Um ≤ u0 ≤ UM a.e. on IR.

5.3.1 Meshes and schemes

ms1d meshhyp1d meshhyp1d
Let T be an admissible mesh in the sense of Definition 5.5 page 125 and k ∈ IR ?+ be hyperbolic1D
the time step. In the
hyperbolic1D
general nonlinear case, the finite volume scheme for the discretization of Problem ( 5.1) page 119 writes
 h
 i
 (un+1 − uni ) + fi+1/2
n n
− fi−1/2 = 0, ∀n ∈ IN, ∀i ∈ ZZ ,
k i Z
1 xi+1/2 (5.22) h1D

 u0i = u0 (x)dx, ∀i ∈ ZZ ,
hi xi−1/2
where uni is expected to be an approximation of u at time tn = nk in cell Ki . The quantity fi+1/2 n
is
often called the numerical flux at point xi+1/2 and time tn (it is expected to be an approximation of f (u)
n
at point xi+1/2 and time tn ). Note that a common expression of fi+1/2 is used for both equations i and
h1D h1D
i + 1 in (5.22); therefore the scheme (5.22) satisfies the property of conservativity, common to all finite
volume schemes. In the case of a so called “scheme with 2p + 1 points” (p ∈ IN ? ), the numerical flux may
be written
 131

n
fi+1/2 = g(uni−p+1 , . . . , uni+p ), (5.23) nf
where g is the numerical flux function, which determines the scheme. It is assumed to be a locally
Lipschitz continuous function.
uapp uapp
As in the linear case (5.15) page 125, the approximate finite volume solution is defined by

uT ,k (x, t) = uni for x ∈ Ki and t ∈ [nk, (n + 1)k). (5.24) nluapp

h1D nf
The property of consistency for the finite volume scheme (5.22), (5.23) with 2p + 1 points, is ensured by
writing the following condition:

g(s, . . . , s) = f (s), ∀s ∈ IR. (5.25) consflux

This condition is equivalent to writing the consistency of the approximation of the flux (as in the elliptic
Diric1D
and parabolic cases, which were described in the previous chapters, see e.g. Section 2.1).

vfordf Remark 5.11 consflux

(Finite volumes and finite differences) We can remarkh1D that, as in the elliptic case,
the condition (5.25) does not generally give the consistency of the scheme (5.22) when it is considered as
a finite difference scheme. For instance, assume f (s) = s2 for all s ∈ IR, p = 1 and g(a, b) = f1 (a) + f2 (b)
f1 (s) = max{s, 0}2, f2 (s) = min{s, 0}2 (which is shown below to be a “good
for all a, b ∈ IR withflumono
choice”, see Example 5.2). Assume also h2i = h and h2i+1 = h/2 for all i ∈ ZZ . In this case, there is
n n
no choice of points xi ∈ IR such that the quantity (fi+1/2 − fi−1/2 )/hi is an approximation of order 1 of
n
(f (u))x (xi , nk), for any regular function u, when ui = u(xi , nk) for all i ∈ ZZ . Indeed, up to second order
terms, this property of consistency is achieved if and only if f20 (a)|xi+1 − xi | + f10 (a)|xi−1 − xi | = f 0 (a)hi
for all i ∈ ZZ and for all a ∈ IR. Choosing a > 0 and a < 0, this condition leads to |xi+1 − xi | = hi and
|xi+1 − xi | = hi+1 for all i ∈ ZZ , which is impossible.

Examples of convenient choices for the function g will now be given. An interesting class of schemes is
the class of 3-points schemes with a monotone flux, which we now define.
datah1d h1D
monflux Definition
nf
5.6 (Monotone flux schemes) Under Assumption 5.1, the finite volume scheme (5.22)--
(5.23) is said to be a “monotone flux scheme” if p = 1 and if the function g, only depending on f , U m
and UM , satisfies the following assumptions:

• g is locally Lipschitz continuous from IR 2 to IR,

• g(s, s) = f (s), for all s ∈ [Um , UM ],
• (a, b) 7→ g(a, b), from [Um , UM ]2 to IR, is nondecreasing with respect to a and nonincreasing with
respect to b.

The monotone flux schemes are worthy of consideration for they are consistent in the finite volume
sense, they are L∞ -stable under a condition (the so called Courant-Friedrichs-Levy
linfstah1d linfstah1d
condition) of the type
k ≤ C1 h, where C1 only depends on g and u0 (see Section 5.3.2 page 132 below), and they are “consistent
with the entropy inequalities” also under a condition of the type k ≤ C h, where C2 only depends on g
discenth1d discenth1d2
and u0 (but C2 may be different of C1 , see Section 5.3.3 page 133).

schmono Remark 5.12 A monotone flux scheme is a monotone scheme, under a Courant-Friedrichs-Levy condi-
tion, which means that the scheme can be written under the form

un+1
i = H(uni−1 , uni , uni+1 ),

with H nondecreasing with respect to its three arguments.

 132

godlewski-springer
flumono Example
leveque
5.2 (Examples of monotone flux schemes) (see also Godlewski and Raviart [1996],
datah1d
LeVeque [1990] and references therein).h1D
Under
nf
Assumption 5.1, here are some numerical flux functions
gmonflux
for which the finite volume scheme (5.22)-(5.23) is a monotone flux scheme (in the sense of Definition
5.6):

• the flux splitting scheme: assume f = f1 + f2 , with f1 , f2 ∈ C 1 (IR, IR), f10 (s) ≥ 0 and f20 (s) ≤ 0
for all s ∈ [Um , UM ] (such a decomposition for f is always possible, see the modified Lax-Friedrichs
scheme below), and take

g(a, b) = f1 (a) + f2 (b).

Note that if f 0 ≥ 0, taking f1 = f and f2 = 0, the flux splitting scheme boils down to the upwind
scheme, i.e. g(a, b) = f (a).
Go
• the Godunov scheme: the Godunov scheme, which was introduced in Godunov [1976], may be
summarized by the following expression.

min{f (ξ), ξ ∈ [a, b]} if a ≤ b,
g(a, b) = (5.26) fluxgodunov
max{f (ξ), ξ ∈ [b, a]} if b ≤ a.

• the modified Lax-Friedrichs scheme : take

f (a) + f (b)
g(a, b) = + D(a − b), (5.27) fluxLF
2
with D ∈ IR such that 2D ≥ max{|f 0 (s)|, s ∈ [Um , UM ]}. Note that in this modified version of
the Lax-Friedrichs scheme, the coefficient D only depends on f , Um and UM , while the original
Lax-Friedrichs scheme consists in taking D = h/(2k), in the case hi = h for all i ∈ IN, and therefore
monflux
satisfies the three items of Definition 5.6 under the condition h/k ≥ max{|f 0 (s)|, s ∈ [Um , UM ]}.
However, an inverse CFL condition appears
estext
to be necessary for the convergence of the original Lax-
estext
Friedrichs scheme (see remark 6.11 page 181); such a condition is not necessary for the modified
version.
Note also that the modified Lax-Friedrichs scheme consists in a particular flux splitting scheme
with f1 (s) = (1/2)f (s) + Ds and f2 (s) = (1/2)f (s) − Ds for s ∈ [Um , UM ].

vfdf Remark 5.13 In the case of a nondecreasing (resp. nonincreasing) function f , the Godunov monotone
fluxgodunov
flux scheme (5.26)
h1D
reduces
nf
to g(a, b) = f (a) (resp. f (b)). Then, in the case of a nondecreasing function
f , the scheme (5.22), (5.23) reduces to

un+1 − uni
hi i
+ f (uni ) − f (uni−1 ) = 0, (5.28) up1D
k
up1D
i.e. the upstream (or upwind) finite volume scheme. The scheme (5.28) is sometimes called “upstream
finite difference” scheme. In that particular case (f monotone and 1D) it is possible to find points x i in
order to obtain a consistent
VFDA
scheme
VFDA
in the finite difference sense (if f is nondecreasing, take x i = xi+1/2
as for the scheme (5.14) page 125).

5.3.2 L∞ -stability for monotone flux schemes

linfstah1d datah1d
nlh1dstab Lemma 5.3 (L∞ estimate inmeshhyp1dthe nonlinear
meshhyp1d
case) Under Assumption 5.1, let T be an admissible
mesh in the sense of definition 5.5 page 125 and let k ∈ IR ?+ be the time
h1D
step.
nluapp
Let uT ,k be the finite volume approximate solution defined by ( 5.22)-(
monflux monflux
5.24) and assume that the scheme is
a monotone flux scheme in the sense of definition 5.6 page 131. Let g1 and g2 be the Lipschitz constants
of g on [Um , UM ]2 with respect to its two arguments.
 133

Under the Courant-Friedrichs-Levy (CFL) condition

inf i∈ZZ hi
k≤ , (5.29) cflhypnl
g1 + g 2
cflhypnl
(note that taking k ≤ αh/(g1 + g2 ) implies (5.29)),
the approximate solution uT ,k satisfies

Um ≤ uT ,k (x, t) ≤ UM for a.e. x ∈ IR and a.e. t ∈ IR + .

nlh1dstab
Proof of Lemma 5.3
Let us prove that

Um ≤ uni ≤ UM , ∀i ∈ ZZ , ∀n ∈ IN, (5.30) induc

induc
by induction
h1D
on n,h1D
which proves the lemma. Assertion (5.30) holds for n = 0 thanks to the definition of
u0i in (5.22) page 130. Suppose
h1D
that it holds for n ∈ IN.
nf
For all i ∈ ZZ , scheme (5.22), (5.23) (with p = 1) gives

un+1
i = (1 − bni+ 1 − ani− 1 )uni + bni+ 1 uni+1 + ani− 1 uni−1 ,
2 2 2 2

with
 n n n
 k g(ui , ui+1 ) − f (ui )
if uni 6= uni+1 ,
bni+ 1 = hi uni − uni+1
2 
0 if uni = uni+1 ,
and
 n n n
 k g(ui−1 , ui ) − f (ui )
if uni 6= uni−1 ,
ani− 1 = hi uni−1 − uni
2 
0 if uni = uni−1 .
Since f (uni ) = g(uni , uni ) and thanks to the monotonicity of g, 0 ≤ bni+ 1 ≤ g2 k/hi and 0 ≤ ani− 1 ≤ g1 k/hi ,
cflhypnl 2 2
the value un+1
for all i ∈ ZZ . Therefore, under condition (5.29), induc i may be written as a convex linear
n n
combination of the values ui and ui−1 . Assertion (5.30) is thus proved for n + 1, which concludes the
proof of the lemma.

5.3.3 Discrete entropy inequalities

discenth1d datah1d
nlh1dentdis Lemma 5.4 (Discretemeshhyp1dentropy inequalities) Under Assumption 5.1, let T be an admissible mesh in
meshhyp1d
the sense of definition 5.5 page 125 and let k ∈ IR ?+ be the timeh1D
step.nluapp
Let uT ,k be the finite volume approximate solution defined by ( 5.22)-(
monflux monflux
5.24) and assume that the scheme is
a monotone flux scheme in the sense of definition 5.6 page 131. Let g1 and g2 be thecflhypnl
Lipschitz constants
of g on [Um , UM ]2 with respect to its two arguments. Under the CFL condition (5.29), the following
inequation holds:
hi  n+1 
|ui − κ| − |uni − κ| +
k (5.31) eqnlh1de
g(uni >κ, uni+1 >κ) − g(uni ⊥κ, uni+1 ⊥κ) − g(uni−1 >κ, uni >κ) + g(uni−1 ⊥κ, uni ⊥κ) ≤ 0,
∀ n ∈ IN, ∀ i ∈ ZZ , ∀ κ ∈ IR.
Recall that a>b (resp. a⊥b) denotes the maximum (resp. the minimum) of the two real numbers a and b.
 134

nlh1dentdis
Proof of Lemma 5.4
cflhypnl schmono
Thanks to the monotonicity properties of g and to the condition (5.29) (see remark 5.12),

un+1
i = H(uni−1 , uni , uni+1 ), ∀i ∈ ZZ , ∀n ∈ IN,
where H is a function from IR 3 to IR which is nondecreasing with respect to all its arguments and such
that κ = H(κ, κ, κ) for all κ ∈ IR.
Hence, for all κ ∈ IR,

un+1
i ≤ H(uni−1 >κ, uni >κ, uni+1 >κ),
and

κ ≤ H(uni−1 >κ, uni >κ, uni+1 >κ),

which yields

un+1
i >κ ≤ H(uni−1 >κ, uni >κ, uni+1 >κ).
In the same manner, we get

un+1
i ⊥κ ≥ H(uni−1 ⊥κ, uni ⊥κ, uni+1 ⊥κ),
and therefore, by substracting the last two equations,

|un+1
i − κ| ≤ H(uni−1 >κ, uni >κ, uni+1 >κ) − H(uni−1 ⊥κ, uni ⊥κ, uni+1 ⊥κ),
eqnlh1de
that is (5.31).
h1D nf
In the two next sections,
5pts1
we study the convergence
5pts2 hyperhos
of the schemes defined by (5.22), (5.23) with p = 1
(see the remarks 5.14 and 5.15 and Section 5.4 for the schemes with 2p + 1 points).
We first develop a proof
bvf
of convergence for the monotone flux schemes;
convhyplinconvhyplin
this proof is based on a weak BV
estimate similar to (5.17) like the proof of proposition 5.5
hypmd
page 129 in the linear case. It will be generalized
in the multidimensional case studied in Chapter 6. We then briefly describe the BV framework which
gave the first convergence results; its generalization to the multidimensional case is not so easy, except
in the case of Cartesian meshes.

5.3.4 Convergence of the upstream scheme in the general case

A proof of convergence similar to the proof of convergence given in the linear case can be developed. For
the sake of simplicity, we shall consider only the case of a nondecreasing function f and of the classical
upstream
hypmd
scheme (the general case for f and for the monotone flux schemes being handled in Chapter
6). We shall first prove a “weak BV ” estimate.
datah1d
nlwb1d Lemma 5.5 (Weak BV estimate for the nonlinear case) Under Assumption 5.1, assume that fmeshhyp1d is
nondecreasing.
meshhyp1d
Let ξ ∈ (0, 1) be a given value. Let T be an admissible mesh in the sense of definition 5.5
page 125, let M be the Lipschitz constant of f in [Um , UM ] and let k ∈ IR ?+ satisfying the CFL condition
inf i∈ZZ hi
k ≤ (1 − ξ) . (5.32) cflhypnlc
M
(The condition k ≤ (1 − ξ)αh/M
h1D
implies the above condition.) Let {u ni , i ∈ ZZ , n ∈ IN} be given by
nf
the finite volume scheme (5.22), (5.23) with p = 1 and g(a, b) = f (a). Let R ∈ IR ?+ and T ∈ IR ?+ and
assume h < R and k < T . Let i0 ∈ ZZ , i1 ∈ ZZ and N ∈ IN be such that −R ∈ K i0 , R ∈ K i1 ,and
T ∈ (N k, (N + 1)k]. Then there exists C ∈ IR ?+ , only depending on R, T , u0 , α, f and ξ, such that
 135

i1 X
X N
k|f (uni ) − f (uni−1 )| ≤ Ch−1/2 . (5.33) bvfnl
i=i0 n=0
nlwb1d
Proof of Lemma 5.5
h1D
We multiply the first equation of (5.22) by kuni , and we sum on i = i0 , . . . , i1 and n = 0, . . . , N . We get
A + B = 0, with
i1 X
X N
A= hi (un+1
i − uni )uni ,
i=i0 n=0

and
i1 X
X N  
B= k f (uni ) − f (uni−1 ) uni .
i=i0 n=0

We have
i1 XN i1
1X 1X
A=− hi (uin+1 − uni )2 + hi [(uN
i
+1 2
) − (u0i )2 ].
2 i=i n=0 2 i=i
0 0
h1D
Using the scheme (5.22), we get

k2  2 1 X
i1 XN i1
1X
A=− n n
f (ui ) − f (ui−1 ) + hi [(uN
i
+1 2
) − (u0i )2 ],
2 i=i n=0 hi 2 i=i
0 0
cflhypnlc
and therefore, using the CFL condition (5.32),

1 Xi1 XN  2 1 Xi1
A≥− (1 − ξ) k f (uni ) − f (uni−1 ) − hi (u0i )2 . (5.34) eqbvwhnl
2M i=i n=0
2 i=i
0 0

We now study the term B. Ra

Denoting by Φ the function Φ(a) = Um sf 0 (s)ds, for all a ∈ IR, an integration by parts yields, for all
(a, b) ∈ IR 2 ,
Z b
Φ(b) − Φ(a) = b(f (b) − f (a)) − (f (s) − f (a))dx.
a
petitlemmepetitlemme Rb 1
Using the technical lemma 4.5 page 107 which states a (f (s) − f (a))dx ≥ 2M (f (b) − f (a))2 , we obtain
1
b(f (b) − f (a)) ≥ (f (b) − f (a))2 + Φ(b) − Φ(a).
2M
The above inequality with a = uni−1 and b = uni yields

1 X
i1 XN  2 XN
B≥ k f (uni ) − f (uni−1 ) + k[Φ(uni1 ) − Φ(uni0 −1 )].
2M i=i n=0 n=0
0
linstab linstab
Thanks to the L∞ estimate of Lemma 5.1 page 125, there exists C1 > 0, only depending on u0 and f
such that

1 X
i1 XN  2
B≥ k f (uni ) − f (uni−1 ) − T C1 .
2M i=i n=0
0
 136

P i1
Therefore, since A + B = 0 and i=i0 hi ≤ 4R, the following inequality holds:
i1 X
X N  2
0≥ξ k f (uni ) − f (uni−1 ) − 4RM max{−Um , UM }2 − 2M T C1 ,
i=i0 n=0

which gives the existence of C2 ∈ IR ?+ , only depending on R, T , u0 , f and ξ such that

i1 X
X N  2
k f (uni ) − f (uni−1 ) ≤ C2 .
i=i0 n=0

The Cauchy-Schwarz inequality yields

hX
i1 X
N i2 4R
k|f (uni ) − f (uni−1 )| ≤ C2 2T ,
i=i0 n=0
αh

which concludes the proof of the lemma.

We can now state the convergence theorem.

datah1d
nlcv1d Theorem 5.2 (Convergence in the nonlinear case) Assume Assumption 5.1 and f nondecreasing.
Let ξ ∈ (0, 1) and α > 0 be given. meshhyp1d
Let M bemeshhyp1d
the Lipschitz constant of f in [Um , UM ]. For an admissible
mesh T in
cflhypnlc
the sense of Definition 5.5 page 125 and for a time step k ∈ IR ?+ satisfying the CFL condition
(5.32) (taking k ≤ (1 − ξ)αh/M is a sufficient condition,
h1D
that ξ and α do not depend of T ), let u T ,k
notenluapp
be the finite volume approximate solution defined by (5.22)-(5.24) with p =hyperbolic1D
1 and g(a, b) = f (a).
hyperbolic1D
Then the function uT ,k converges to the unique entropy weak solution u of (5.1) page 119 in L1loc (IR ×IR + )
as size(T ) tends to 0.

Proof
h1D nluapp
Let Y be the set of approximate solutions, that is the set of uT ,k , defined by (5.22)-(5.24) with p = 1 and
meshhyp1d meshhyp1d
g(a, b) = f (a), for all (T , k) where Tcflhypnlc
is an admissible mesh innlh1dstab
the sense of Definition 5.5 page 125 and
k ∈ IR ?+ satisfies the CFL condition (5.32). Thanks to Lemma 5.3, the set Y is bounded in L∞ (IR × IR + ).
nlcv1d
The proof of Theorem 5.2 is performed in three steps. In the first step, a compactness result is given for
Y , only using the boundeness of Y in L∞ (IR × IR + ). In the second step, it is proved that the eventual
limit (in a convenient sense) of a sequence of approximate solutions is a solution (in a convenient sense)
hyperbolic1D
of problem hypmd
(5.1). In the third step a uniqueness result yields the conclusion. For steps 1 and 3, we refer
to chapter 6 for a complete proof.
Step 1 (compactness result) nlwsprop nlwsprop
Let us first use a compactness result in L∞ (IR ×IR + ) which is stated in Proposition 6.4 page 191. Since Y
is bounded in L∞ (IR × IR + ), for any sequence (um )m∈IN of Y there exists a subsequence, still denoted by
(um )m∈IN , and there exists µ ∈ L∞ (IR ×IR + ×(0, 1)) such that (um )m∈IN converges to µ in the “nonlinear
weak-? sense”, that is
Z Z Z Z Z 1
θ(um (x, t))ϕ(x, t)dtdx → θ(µ(x, t, α))ϕ(x, t)dαdtdx, as m → ∞,
IR IR + IR IR + 0

for all ϕ ∈ L1 (IR × IR + ) and all θ ∈ C(IR, IR). In other words, for any θ ∈ C(IR, IR),

θ(um ) → µθ in L∞ (IR × IR + ) for the weak-? topology as m → ∞, (5.35) cvmug

where µθ is defined by
Z 1
µθ (x, t) = θ(µ(x, t, α))dα, for a.e. (x, t) ∈ IR × IR + .
0
 137

Step 2 (passage to the limit)

Let (um )m∈IN be a sequence of Y . Assume that (um )m∈IN converges to µ in the nonlinear weak-? sense
and that um = uTm ,km (for all m ∈ IN) with size(Tm ) → 0 as m → ∞ (note that km → 0 as m → ∞,
cflhypnlc
thanks to (5.32)). hyperbolic1D
Let us prove that µ is a “solution” to hyperbolic1D
problem (5.1) in the following sense (we shall say that µ is “an
entropy process solution” to problem (5.1)):



 µ ∈ L∞ (IR × IR + × (0, 1)),

 Z Z Z 1 

|µ(x, t, α) − κ|ϕt (x, t) + (f (µ(x, t, α)>κ) − f (µ(x, t, α)⊥κ))ϕx (x, t) dαdtdx
(5.36) h1deps
 IR IR + 0
 Z


 + |u0 (x) − κ|ϕ(x, 0)dx ≥ 0, ∀ϕ ∈ Cc1 (IR × IR + , IR + ), ∀κ ∈ IR.
IR

Let κ ∈ IR. Setting

Z 1
v(x, t) = |µ(x, t, α) − κ|dα, for a.e. (x, t) ∈ IR × IR +
0
and
Z 1 
w(x, t) = f (µ(x, t, α)>κ) − f (µ(x, t, α)⊥κ) dα, for a.e. (x, t) ∈ IR × IR + ,
0
h1deps
the inequality in (5.36) writes
Z Z Z
[v(x, t)ϕt (x, t) + w(x, t)ϕx (x, t)]dtdx + |u0 (x) − κ|ϕ(x, 0)dx ≥ 0,
IR IR + IR (5.37) cvh1dnl1
∀ϕ ∈ Cc1 (IR × IR + , IR + ).
cvh1dnl1
Let us prove that (5.37) holds; for m ∈ IN we shall denote by T = Tm and k = km . We use the result of
nlh1dentdis
Lemma 5.4, which writes in the present particular case f 0 ≥ 0,

vin+1 − vin
hi + win − wi−1
n
≤ 0, ∀i ∈ ZZ , ∀n ∈ IN,
k
where vin = |uni − κ| and win = f (uni >κ) − f (uni ⊥κ) = |f (uni ) − f (κ)|.
The functions vTm ,km and wTm ,km are defined in the same way as the function uTm ,km , i. e. with constant
values vin and win in each control volume Ki during each time step (nk, (n + 1)k). Choosing θ equal to
cvmug
the continuous functions | · −κ| and |f (·) − f (κ)| in (5.35) yields that the sequences (vTm ,km )m∈IN and
(wTm ,km )m∈IN converge to v and w in L∞ (IR × IR ?+ ) for the weak-? topology.
convhyplinconvhyplin
Applying the method which was used in the proof of Proposition 5.5 page 129, taking vil instead of uli in
= n and n + 1) and wjn instead of unj in the definition of Bm (for j = i and
the definition of Am (for lcvh1dnl1
i − 1), we conclude that (5.37) holds. bvf bvf
n n
Indeed, a weak BV inequality holds on thenlwb1d
values w i (that is (5.17) page 126 holds with wj instead of
nlwb1d
n
uj for j = i and i − 1), thanks to Lemma 5.5 page 134 and the relation
 
|f (un ) − κ| − |f (un ) − κ| ≤ |f (un ) − f (un )|, ∀i ∈ ZZ , ∀n ∈ IN.
i i−1 i i−1
bvweff
(Note
bvweff
that here, as in the linear case, the useful consequence of the weak BV inequality, is ( 5.19) page
127 with wjn instead of unj for j = i and i − 1.)
This concludes Step 2.
h1deps
Step 3 (uniqueness result for (5.36) and conclusion)
ewsuniciteewsunicite h1deps
Theorem 6.3 page 175 states that there exists at most one solution to (5.36) and that there exists u ∈
h1deps
L∞ (IR ×IR + ) such that µ solution to (5.36) implies µ(x, t, α) = u(x, t) for a.e. (x, t, α) ∈ IR ×IR + ×(0, 1).
hyperbolic1D
Then, u is necessarily the entropy weak solution to (5.1).
 138

Furthermore, if (um )m∈IN converges to u in the nonlinear weak-? sense, an easy argument shows that
(u m )m∈IN converges
cvforte cvforte
to u in L1loc (IR × IR + ) (and even in Lploc (IR × IR + ) for all 1 ≤ p < ∞), see Remark
(6.15) page 193. nlcv1d
Then, the conclusion of Theorem 5.2 follows easily from Step 2 and Step 1 by way of contradiction (in
order to prove the convergence of a sequence uTm ,km ⊂ Y to u, if size(Tm ) → 0 as m → ∞, without any
extraction of a “subsequence”).

nlcv1d
5pts1 Remark 5.14 In Theorem 5.2, we only consider the case f 0 ≥ 0 and the so called “upstream scheme”.
It is quite easy to generalize the result for any f ∈ C 1 (IR, IR) and any monotone flux scheme (see the
following chapter).
hyperhos
It is also possible to consider other schemes (for instance, some 5-points schemes, as
in Section 5.4). For a given scheme, the proof of convergence of the approximate solution towards the
entropy weak solution contains 2 steps:

1. prove an L∞ estimate on the approximate

nlcv1d
solutions, which allows to use the compactness result of
Step 1 of the proof of Theorem 5.2,

2. prove a “weak BV ” estimate and some “discrete entropy inequality” in order to have the following
property:
If (um )m∈IN is a sequence of approximate solutions which converges in the nonlinear weak-? sense,
then
Z Z  
lim |um (x, t) − κ|ϕt (x, t) + (f (um (x, t)>κ) − f (um (x, t)⊥κ))ϕx (x, t) dtdx
m→IN IR IR + Z
+ |u0 (x) − κ|ϕ(x, 0)dx ≥ 0, ∀ϕ ∈ Cc1 (IR × IR + , IR + ), ∀κ ∈ IR.
IR

5.3.5 Convergence proof using BV

bvmethod
We now give the details of the classical proof of convergence (considering only 3 points schemes), which
requires regularizations of u0 in BV (IR). It consists in using Helly’s compactness theorem (which may
also be used in the linear case to obtain a strong convergenceKolm
of uT ,k to u in L1loc (IR ×IR + )). This theorem
Kolm
is a direct consequence of Kolmogorov’s theorem (theorem 3.9 page 93). We give below the
defBV
definition of
defBV
BV (Ω) where Ω is an open subset of IR p (Ω), p ≥ 1 (already given in Definition 5.5 page 121 for Ω = IR)
and we give a straightforward consequence of Helly’s theorem for the case of interest here.

bvrd Definition 5.7 (BV (Ω)) Let p ∈ IN ? and let Ω be an open subset of IR p . A function v ∈ L1loc (Ω) has a
bounded variation, that is v ∈ BV (Ω), if |v|BV (Ω) < ∞ where
Z
|v|BV (Ω) = sup{ v(x)divϕ(x)dx, ϕ ∈ Cc1 (Ω, IR p ), |ϕ(x)| ≤ 1, ∀x ∈ Ω}. (5.38) defBVRp
Ω

helly Lemma 5.6 (Consequence of Helly’s theorem) Let A ⊂ L∞ (IR 2 ). Assume that there exists C ∈
IR + and, for all T > 0, there exists CT ∈ IR + such that

kvkL∞ (IR 2 ) ≤ C, ∀v ∈ A,
and
|v|BV (IR×(−T,T )) ≤ CT , ∀v ∈ A, ∀T > 0.
Then for any sequence (vn )n∈IN of elements of A, there exists a subsequence, still denoted by (vn )n∈IN ,
and there exists v ∈ L∞ (IR 2 ), with kvkL∞(IR 2 ) ≤ C and |v|BV (IR×(−T,T )) ≤ CT for all T > 0, such that
R
vn → v in L1loc (IR 2 ) as n → ∞, that is ω̄ |vn (x) − v(x)|dx → 0, as n → ∞ for any compact set ω̄ of IR 2 .
 139

helly
In order to use Lemma 5.6, one first shows the following BV stability estimate for the approximate
solution:
datah1d
nlstabv Lemma 5.7 (Discrete space BV estimate) Under Assumption 5.1, assume that u0 ∈ BV (IR); let T
meshhyp1d meshhyp1d
sense ofnfDefinition 5.5 page 125 and let k ∈ IR ?+ be the time step. Let {uni ,
be an admissible mesh in the h1D
i ∈ ZZ , n ∈ IN} be given by (5.22), (5.23) and assume that the scheme is a monotone flux scheme¡ in the
monflux monflux
sense of Definition 5.6 page 131. Let g1 and g2 be the Lipschitz
cflhypnl
constants of g on [Um , UM ]2 with respect
to its two arguments. Then, under the CFL condition (5.29), the following inequality holds:
X X
|un+1
i+1 − ui
n+1
|≤ |uni+1 − uni |, ∀ n ∈ IN. (5.39) eqnlstabv
i∈ZZ i∈ZZ
nlstabv
Proof of Lemma 5.7
P divbv1d divbv1d
First remark that, for n = 0,h1D i∈ZZnf|u0i+1 − u0i | ≤ |u0 |BV (IR) (see Remark 5.4 page 121).
For all i ∈ ZZ , the scheme (5.22), (5.23) (with p = 1) leads to

un+1
i = uni + bni+ 1 (uni+1 − uni ) + ani− 1 (uni−1 − uni ),
2 2

and

un+1 n n n n n n n
i+1 = ui+1 + bi+ 3 (ui+2 − ui+1 ) + ai+ 1 (ui − ui+1 ),
2 2
nlh1dstab nlh1dstab
where ai+1/2 and bi+1/2 are defined (for all i ∈ ZZ ) in Lemma 5.3 page 132. Substracting one equation
to the other leads to

un+1 n+1
i+1 − ui = (uni+1 − uni )(1 − bni+ 1 − ani+ 1 ) + bni+ 3 (uni+2 − uni+1 ) + ani− 1 (uni − uni−1 ).
2 2 2 2
cflhypnl
Under the condition (5.29), we get

|un+1 n+1
i+1 − ui | ≤ |uni+1 − uni |(1 − bni+ 1 − ani+ 1 ) + bni+ 3 |uni+2 − uni+1 | + ani− 1 |uni − uni−1 |.
2 2 2 2
eqnlstabv
Summing the previous equation over i ∈ ZZ gives (5.39).

datah1d
nlstabvco Corollary 5.1 (Discrete BV estimate) Under assumption 5.1, let u0 ∈ BV (IR); let T be an admis-
meshhyp1d meshhyp1d
sible mesh in the sense of Definition 5.5 page h1D 125 and let k ∈ IR ?+ be the time step. Let uT ,k be the
nluapp
finite volume approximate solution defined by (5.22)-(5.24) and assume that the scheme is a monotone
monflux monflux
flux scheme in the sense of Definition 5.6 page 131. Let g1 and g2 be the Lipschitz constants of g on
cflhypnl
[Um , UM ]2 with respect to its two arguments and assume that k satisfies the CFL condition ( 5.29). Let
uT ,k (x, t) = u0i for a.e. (x, t) ∈ Ki × IR − , for all i ∈ ZZ (hence uT ,k is defined a.e. on IR 2 ). Then, for
any T > 0, there exists C ∈ IR ?+ , only depending on u0 , g and T such that:

|uT ,k |BV (IR×(−T,T )) ≤ C. (5.40) conlstabv

nlstabvco
Proof of Corollary 5.1
nlstabv P
As in Lemma 5.7, remark that i∈ZZ |u0i+1 − u0i | ≤ |u0 |BV (IR) .
Let us first assume that T ≤ k. Then, the BV semi-norm of uT ,k satisfies
X
|uT ,k |BV (IR×(−T,T )) ≤ 2T |u0i+1 − u0i |.
i∈ZZ
conlstabv
Hence the estimate (5.40) is true for C = 2T |u0 |BV (IR) .
Let us now assume that k < T . Let N ∈ IN ? such that N k < T ≤ (N + 1)k. The definition of
| · |BV (IR×(−T,T )) yields
 140

P
|uT ,k |BV (IR×(−T,T )) ≤ T i∈ZZ |u0i+1 − u0i |+
NX−1 X X N
X −1 X
(5.41) bvtot
k|uni+1 − uni | + (T − N k) |uN
i+1 − u N
i | + hi |un+1
i − uni |.
n=0 i∈ZZ i∈ZZ n=0 i∈ZZ
nlstabv P n
Lemma 5.7 gives i∈ZZ |ui+1 − uni | ≤ |u0 |BV (IR) for all n ∈ IN, and therefore,
N
X −1 X X
k|uni+1 − uni | + (T − N k) |uN N
i+1 − ui | ≤ T |u0 |BV (IR) . (5.42) bvtot1
n=0 i∈ZZ i∈ZZ
bvtot h1D
In order to bound the last term of (5.41), using the scheme (5.22) yields, for all i ∈ ZZ and all n ∈ IN,
k k
|un+1
i − uni | ≤ g1 |uni − uni−1 | + g2 |uni − uni+1 |.
hi hi
Therefore,
X X
hi |un+1
i − uni | ≤ k(g1 + g2 ) |uni − uni+1 |, for all n ∈ IN,
i∈ZZ i∈ZZ

which yields, since N k < T ,

N
X −1 X
hi |un+1
i − uni | ≤ T (g1 + g2 )|u0 |BV (IR) . (5.43) bvtot2
n=0 i∈ZZ
conlstabv bvtot bvtot1 bvtot2
Therefore Inequality (5.40) follows from (5.41), (5.42) and (5.43) with C = T (2 + g1 + g2 )|u0 |BV (IR) .

Consider a sequence of admissible meshes and time steps verifying the CFL condition,nlstabvco
and the associated
nlh1dstab
sequence
nlh1dstab
of approximate solutions
nlstabvco
(prolonged on IR × IR − as in Corollary 5.1). By Lemma 5.3 page
helly
132 and
helly
Corollary 5.1, the sequence of approximate solutions satisfies the hypotheses of Lemma 5.6 page
138. It is therefore possible to extract a subsequence which converges in L 1loc (IR × IR + ) to a function
× IR ?+ ). It must still be shown that the function u is the unique weak
u ∈ L∞ (IRhyperbolic1D entropy solution of
eqnlh1de
Problem (5.1). This
eqnlstabv
may be proven by using the discrete entropy inequalities ( 5.31) and the strong BV
estimate (5.39) or the classical Lax-Wendroff theorem recalled below.
datah1d
laxw Theorem 5.3 (Lax-Wendroff ) Under Assumption 5.1, let α > 0 be given and let (Tm )m∈IN be a
meshhyp1d meshhyp1d
of Definition 5.5 page 125 (note that, for all m ∈ IN, the mesh
sequence of admissible meshes in the sensemeshhyp1d
Tm satisfies the hypotheses of Definition 5.5 where T = Tm and α is independent of m). Let (km )m∈IN
be a sequence of (positive) time steps. Assume that size(Tm ) → 0 and km → 0 as m → ∞.
h1D nluapp
For m ∈ IN, setting T = Tm and k = km , let um = uT ,k be the solution of (5.22)-(5.24) with p = 1
and some g from IR 2 to IR, only depending on f and u0 , locally Lipschitz continuous and such that
g(s, s) = f (s) for all s ∈ IR.
Assume that (um )m∈IN is bounded in L∞ (IR × IR +h1dweak
hyperbolic1D
) and that um → u a.e. on IR × IR + . Then, u is a
weak solution to problem (5.1) (that is u satisfies (5.3)).
Furthermore, assume that for any κ ∈ IR there exists some locally Lipschitz continuous function G κ from
IR 2 to IR, only depending on f , u0 and κ, such that Gκ (s, s) = f (s>κ) − f (s⊥κ) for all s ∈ IR and such
that for all m ∈ IN
1 n+1 1
(|u − κ| − |uni − κ|) + (Gκ (uni , uni+1 ) − Gκ (uni−1 , uni )) ≤ 0, ∀i ∈ ZZ , ∀n ∈ IN, (5.44) entlw
k i hi
h1D nf
where {uni , i ∈ ZZ , n ∈ IN} is the hyperbolic1D
solution to (5.22)-(5.23) for T = Tm and k = km . Then, u is the
h1dent
entropy weak solution to Problem (5.1) (that is u is the unique solution of (5.4)).
 141

laxw
Proof of Theorem 5.3
Since (um )m∈IN is bounded in L∞ (IR × IR + ) and um → u a.e. on IR × IR + , the sequence (um )m∈IN
converges
Kolm
to u in L1loc (IR × IR + ). This implies in particular (from Kolmogorov’s theorem, see Theorem
3.9) that, for all R > 0 and all T > 0,
Z 2T Z 2R
sup |um (x, t) − um (x − η, t)|dxdt → 0 as η → 0.
m∈IN 0 −2R

Then, taking η = α size(Tm ) (for m ∈ IN) and letting m → ∞ yields, in particular,

Z 2T Z 2R
|um (x, t) − um (x − α size(Tm ), t)|dxdt → 0 as m → ∞. (5.45) astuce
0 −2R
h1D nf
For m ∈ IN, let {uni , i ∈ ZZ , n ∈ IN} be the solution to (5.22)-(5.23) for T = Tm and k = km (note that
uni depends on m, even though this dependency is not so clear in the notation). We also set k m = k and
size(Tm ) = h, so that k and h are depending on m (but recall that α is not depending on m).
Let R > 0 and T > 0. Let i0 ∈ ZZ , i1 ∈ ZZ and N ∈ IN be such that −R ∈ K i0 , R ∈ K i1 and
T ∈ (N k, (N + 1)k]. Then, for h < R and k < T (which is true for m large enough),
i1 X
X N Z 2T Z 2R
αh k|uni − uni−1 | ≤ |um (x, t) − um (x − αh, t)|dxdt.
i=i0 n=0 0 −2R
astuce bvweff
Therefore, Inequality (5.45) leads to (5.19), that is
i1 X
X N
h k|uni − uni−1 | → 0 as m → ∞. (5.46) bvwelw
i=i0 n=0
bvwelw laxw convhyplin
Usingconvhyplin
(5.46), the remainder of the proof of Theorem 5.3 isnlcv1d
nlcv1d
very similar to the proof of Proposition 5.5
bvwelw
page 129 and to Step 2 in the proof of Theorem 5.2 page 136 (Inequality (5.46) replaces the weak BV
inequality).
h1dweak h1D
In order to prove that u is solution to (5.3), let us multiply the first equation of (5.22) by (k/hi )ϕ(x, nk),
integrate over x ∈ Ki and sum for all i ∈ ZZ and n ∈ IN. This yields

Am + B m = 0
with
X X Z
Am = (un+1
i − uni ) ϕ(x, nk)dx
i∈ZZ n∈IN Ki

and
X X Z
1
Bm = k(g(uni , uni+1 ) − g(uni−1 , uni )) ϕ(x, nk)dx.
hi Ki
i∈ZZ n∈IN
convhyplin
As in the proof of Proposition 5.5, one has
Z Z Z
lim Am = − u(x, t)ϕt (x, t)dxdt − u0 (x)ϕ(x, 0)dx.
m→+∞ IR + IR IR

Let us now turn to the study of Bm . We compare Bm with

XZ (n+1)k Z
B1,m = − f (uT ,k (x, t))ϕx (x, nk)dxdt,
n∈IN nk IR
 142

R R
which tends to − IR + IR f (u(x, t))ϕx (x, t)dxdt as m → ∞ since f (uT ,k ) → f (u) in L1loc (IR × IR + ) as
m → ∞.
The term B1,m can be rewritten as
X X
B1,m = k(f (uni ) − f (uni−1 ))ϕ(xi−1/2 , nk),
i∈ZZ n∈IN

which yields, introducing g(uni−1 , uni ),

X X
B1,m = k(f (uni ) − g(uni−1 , uni ))ϕ(xi−1/2 , nk)
X i∈ZX
Z n∈IN
+ k(g(uni−1 , uni ) − f (uni−1 ))ϕ(xi−1/2 , nk).
i∈ZZ n∈IN

Similarly, introducing f (uni ) in Bm ,

X X Z
1
Bm = k(f (uni )
− g(uni−1 , uni ))
ϕ(x, nk)dx
hi K i
i∈ZZ n∈IN
X X Z
1
+ k(g(uni , uni+1 ) − f (uni )) ϕ(x, nk)dx.
hi K i
i∈ZZ n∈IN

In order to compare Bm and B1,m , let R > 0 and T > 0 be such that ϕ(x, t) = 0 if |x| ≥ R or t ≥ T . Let
A > 0 be such that kum kL∞ (IR×IR + ) ≤ A for all m ∈ IN. Then there exists C > 0, only depending on ϕ
and the Lipschitz constants on g on [−A, A]2 , such that, if h < R and k < T (which is true for m large
enough),
i1 X
X N
|Bm − B1,m | ≤ Ch k|uni − uni−1 |, (5.47) bazar
i=i0 n=0

∈ ZZ , i1 ∈bvwelw
where i0bazar ZZ and N ∈ IN are such that −R ∈ K i0 , R ∈ K i1 and T ∈ (N k, (N + 1)k].
Using (5.47) and (5.46), we get |Bm − B1,m | → 0 and then
Z Z
Bm → − f (u(x, t))ϕx (x, t)dtdx as m → ∞,
IR IR +
h1dweak
which completes the proof that u is a solution to problem 5.3.
entlw h1dentkr h1dentkr
Under the additional assumption
h1dent
that um satisfies (5.44), one proves that u satisfies (5.7) page 121 (and
hyperbolic1D
therefore that u satisfies (5.4)) and is the entropy weak solution to Problem (5.1) by a similar method.
Indeed, let κ ∈ IR. One replaces uli by |uli − κ| in Am (for l = n and n + 1) and one replaces entlw
g by Gκ in
Bm . Then, passing to the limit in Am + Bm ≤ 0 (which is a consequence of the inequation (5.44)) leads
the desired result. laxw
This concludes the proof of Theorem 5.3

laxw
5pts2 Remark 5.15 Theorem 5.3 still holds with (2p + 1)-points schemes
laxw h1dweak
(p > 1). The generalization of the
first part laxw
of Theorem 5.3 (the proof that u is a solution to (5.3)) is quite easy. For the second part of
Theorem 5.3 (entropy inequalities) the discrete entropy inequalities may be replaced by some weaker ones
(in order to handle interestinglaxwschemes such as those which are described in the following section).
However, the use of Theorem 5.3 needs a compactness property of sequences of approximate solutions in
the space L1loceqnlstabv
(IR ×IR + ). Such a compactness property is generally achieved withHAa “strong BV estimate”
(similar to (5.39)). Hence an extensive literature on
godlewski-ellipses
“TVD schemes” (see Harten [1983]), “ENO
godlewski-springer
schemes”. . . (see Godlewski and Raviart [1991], Godlewski and Raviart [1996] and references
therein). The generalization of this method in the multidimensional case (studied in the following chapter)
does not seem so clear except in the case of Cartesian meshes.
 143

5.4 Higher order schemes

hyperhos monflux monflux
Consider a monotone flux scheme in the sense of Definition 5.6 page 131. By definition, the considered
scheme is a 3 points scheme; recall that the numerical flux function is denoted by g. Thehyperbolic1D
approximate
hyperbolic1D
solution obtained with this scheme converges to the entropy weak solution of Problem ( 5.1) page 119
nlcv1d
as the mesh size tends to 0 and under a so called CFL condition (it is proved in Theorem 5.2 for a
particular case and in the next chapter for the general case). However, 3-points schemes are known to
be diffusive, so that the approximate solution is not very precise near the discontinuities. An idea to
reduce the diffusion is to go to a 5-points scheme by introducing “slopes” on each discretization cell and
limiting the slopes in order for the scheme to remain stable. A classical way to do this is the “MUSCL”
vanleer3
(Monotonic Upwind Scheme for Conservation Laws, see Van Leer [1979]) technique .
ms1d godlewski-ellipses
We briefly describe, with the notations of Section 5.3.1, an example of such a scheme, see e.g. Godlewski
godlewski-springer
and Raviart [1991] and Godlewski and Raviart [1996] for further details. Let n ∈ IN.

• Computation of the slopes

uni+1 − uni−1
p̃ni = hi−1 hi+1
, i ∈ ZZ .
hi + 2 + 2

• Limitation of the slopes

pni = αni p̃ni , i ∈ ZZ , where αni is the largest number in [0, 1] such that

hi n n hi
uni + αi p̃i ∈ [uni ⊥uni+1 , uni >uni+1 ] and uni − αni p̃ni ∈ [uni ⊥uni−1 , uni >uni−1 ].
2 2
In practice, other formulas giving smaller values of αni are sometimes needed for stability reasons.
• Computation of un+1
i for i ∈ ZZ
nf
One replaces g(uni , un+1
i ) in (5.23) by :

hi n n hi+1 n+1
g(uni−1 , uni , uni+1 , uni+2 ) = g(uni + pi , ui+1 − p ).
2 2 i

The scheme thus constructed is less diffusive than the original one and it remains stable thanks to the
limitation of the slope. Indeed, if the limitation of the slopes is not active (that is αni = 1), the space
diffusion term disappears from this new scheme, while the time “antidiffusion” term remains. Hence it
seems appropriate to use a higher order scheme for the time discretization. This may be done by using,
for instance, an RK2 (Runge Kutta order 2, or Heun) method for the discretization of the time derivative.
The MUSCL scheme may be written as

U n+1 − U n
= H(U n ) for n ∈ IN,
k
where U n = (uni )i∈ZZ ; hence it may be seen as the explicit Euler discretization of

Ut = H(U );
therefore, the RK2 time discretization yields to the following scheme:

U n+1 − U n 1 1 
= H(U n ) + H(U n + kH(U n )) for n ∈ IN.
k 2 2
Going to a second order discretization in time allows larger time steps, without loss of stability.
 144

Results of convergence are possible with these new schemes (with eventually some adaptation of the
vilathese
slope limitations to obtain convenient discrete entropy inequalities, see Vila [1986]. It is also possible
to obtain errorChainais1
estimates in the spirit of those given in the following
Noelle kroner2
chapter, in the multidimensional
case, see e.g. Chainais-Hillairet [1996], Noëlle [1996], Kröner, Noelle and Rokyta [1995].
However these error estimates are somewhat unsatisfactory since they are of a similar order to that of the
original 3-points scheme (although these schemes are numerically more precise that the original 3-points
schemes).
hyperbolic1D
hyperbolic1D
The higher order schemes are nonlinear even if Problem (5.1) page 119 is linear, because of the limitation
of the slopes.

Implicit versions of these higher order schemes are more or less straightforward. However, the numerical
implementation of these implicit versions requires the solution of nonlinear systems. In many cases, the
solutions to these nonlinear systems
pfertzel
seem impossible to reach for large k; in fact, the existence of the
solutions is not so clear, see Pfertzel [1987]. Since the advantage of implicit schemes is essentially
the possibility to use large values of k, the above flaw considerably reduces the opportunity of their use.
Therefore, although implicit 3-points schemes are veryturbo
turbo
diffusive, they remain the basic schemes in several
industrial environments. See also Section 7.1.3 page 203 for some clues on implicit schemes applied to
complex industrial applications.
 Chapter 6

Multidimensional nonlinear
hyperbolic equations
hypmd
The aim of this chapter is to define and study finite volume schemes for the approximation of the
solution to a nonlinear scalar hyperbolic problem in several space dimensions. Explicit and implicit
time discretizations are considered. We prove the convergence of the approximate solution towards the
entropy weak solution of the problem and give an error estimate between the approximate solution and
the entropy weak solution with respect to the discretization mesh size.

6.1 The continuous problem

We consider here the following nonlinear hyperbolic equation in d space dimensions (d ≥ 1), with initial
condition

ut (x, t) + div(vf (u))(x, t) = 0, x ∈ IR d , t ∈ IR + , (6.1) estpbcont

u(x, 0) = u0 (x), x ∈ IR d , (6.2) estcondini

where ut denotes the time derivative of u (t ∈ IR + ), and div the divergence operator with respect to the
space variable (which belongs to IR d ). Recall that |x| denotes the euclidean norm of x in IR d , and x · y
the usual scalar product of x and y in IR d .

The following hypotheses are made on the data:

estdonnees Assumption 6.1

(i) u0 ∈ L∞ (IR d ), Um , UM ∈ IR, Um ≤ u0 ≤ UM a.e.,

(ii) v ∈ C 1 (IR d × IR + , IR d ),
(iii) divv(x, t) = 0, ∀(x, t) ∈ IR d × IR + ,
(iv) ∃V < ∞ such that |v(x, t)| ≤ V, ∀(x, t) ∈ IR d × IR + ,
(v) f ∈ C 1 (IR, IR).
estdonnees
baptisee Remark 6.1 Note that part (iv) of Assumption 6.1 is crucial. It ensures the property of “propagation
estkruzkov
in finite time” which is needed for the uniqueness of the solution of (6.3) and for the stability (under
a “Courant-Friedrichs-Levy”
estdonnees
(CFL) condition) of the time explicit numerical scheme. Part (iii) of As-
sumption 6.1, on the other hand, is only considered for the sake of simplicity; the results of existence and
uniqueness
esth14
of the esth14i
esth14
entropy weak solution and convergence (including error estimates as in the theorems
esth14i
6.5 page 180 and 6.6 page 181) of the numerical schemes presented below may be extended to the case
estdonnees
divv 6= 0. However, part (iii) of Assumption 6.1 is natural in many “applications” and avoids several

145
 146

technical complications. Note, in particular,

estkruzkov
divv 6= 0, the L∞ -bound on the solution
that, for instance, ifesteu
estinftyest
∞
of (6.3) and the L estimate (in Lemma 6.1 and Proposition 6.1) on the approximate solution depend
on v and T . The case F (x, t, u) instead
Chainais1
of v(x, t)f (u) is also feasible, but somewhat more technical, see
Chainais2
Chainais-Hillairet [1996] and Chainais-Hillairet [1999].
estpbcont
estcondini
Problem (6.1)-(6.2) has a unique entropy weak solution, which is the solution to the following equation
defent1 defent1
(which is the multidimensional extension of the one-dimensional definition 5.3 page 120).


 u ∈ LZ∞ (IR d × IR ?+ ),
 Z
 h i


 η(u(x, t))ϕt (x, t) + Φ(u(x, t))v(x, t) · ∇ϕ(x, t) dxdt +
ZIR + IR d



 η(u0 (x))ϕ(x, 0)dx ≥ 0, ∀ϕ ∈ Cc∞ (IR d × IR + , IR + ),

 d
 IR
∀η ∈ C 1 (IR, IR), convex function, and Φ ∈ C 1 (IR, IR) such that Φ0 = f 0 η 0 ,
where ∇ϕ denotes the gradient of the function ϕ with respect to the space variable (which belongs to
IR d ). Recall that Ccm (E, F ) denotes the set of functions C m from E to F , with compact support in E.
prentkr1 prentkr1
The characterization of the entropy weak solution by the Krushkov entropies (proposition 5.2 page 121)
still holds in the multidimensional case. Let us define again, for all κ ∈ IR, the Krushkov entropies (|·−κ|)
for which the entropy flux is f (·>κ) − f (·⊥κ) (for any pair of real values a, b, we denote again by a>b
the maximum of a and b, and by a⊥b the minimum of a and b). The unique entropy weak solution is
also the unique solution to the following problem:

 d


∞
Z ∈ LZ (IR
u × IR ?+ ),

 h   i

|u(x, t) − κ|ϕt (x, t) + f (u(x, t)>κ) − f (u(x, t)⊥κ) v(x, t) · ∇ϕ(x, t) dxdt +
d (6.3) estkruzkov
 ZIR + IR



 |u0 (x) − κ|ϕ(x, 0)dx ≥ 0, ∀κ ∈ IR, ∀ϕ ∈ Cc∞ (IR d × IR + , IR + ).
IR d
hyp1dexistu
hyp1dexistu
As in the one-dimensional case (Theorem 5.1estpbcont
pageestcondini
121), existence and uniqueness results are also known
for the entropy weak
estdonnees
solution
krushkov
to Problem ( 6.1)-(
Vo
6.2) under assumptions which differ slightly from as-
sumption 6.1 (see e.g. Krushkov [1970], Vol’pert [1967]). In particular, these results are obtained
with a nonlinearity F (in our case F = vf ) of class C 3 . We recall that the methods which were used in
krushkov
Krushkov [1970] require a regularization in BV (IR d ) of the function u0 , in order to hellytake advantage,
helly
for any T > 0, of compactness properties which are similar to those given in Lemma 5.6 page 138 for
p
the case d bvrd
= 1. Recall
bvrd
that the space BV (Ω) where Ω is an open subset of IR , p ≥ 1, was defined in
Definition 5.7 page 138; it will be used later with Ω = IR d or Ω = IR d × (−T, T ).
estpbcont
estcondini
The existence of solutions to similar problems to (6.1)-(CS 6.2) was already proved by passing to the limit
on
CS
solutions of an appropriate numerical scheme, see Conway and Smoller [1966]. The work of
Conway and Smoller [1966] uses a finite difference scheme on a uniform rectangular grid, in two
space dimensions, and requires that the initial condition u0 belongs to BV (IR d ) (and thus, the solution
estpbcont
estcondini
to Problem (6.1)-(6.2) also hashelly
helly
a locally bounded variation). These assumptions (on meshes and on u 0 )
yield, as in Lemma 5.6 page 138, a (strong) compactness property in L1loc (IR d × IR + ) on a family of
approximate solutions. In the following, however, we shall only require that u0 ∈ L∞ (IR d ) and we shall
be able to deal with more general meshes. We may use, for instance, a triangular mesh in the case of
two space dimensions. For each of these reasons, the BV framework may not be used and a (strong)
compactness property in L1loc on a family of approximate solutions is not easy to obtain (although this
compactness property
estpbcont
does hold and results from this chapter). In order to prove the existence of
estcondini
a solution to (6.1)-(6.2) by passing to the limit on the approximate solutions given by finite volume
schemes on general meshes (in the sense used below) in two or three space dimensions, we shall work
nlwsprop
∞
with some “weak” compactness result in L
nlwsdef nlwsdef
, namely Proposition 6.4, which yields the “nonlinear weak-?
convergence” (see Definition 6.3 page 190) of a family of approximate solutions. When doing so, passing
 147

to the limit with

estpbcont
the approximate estepsol
estcondini
solutions estepsol
will give the existence of an “entropy process solution” to
Problem (6.1)-( 6.2),
estpbcont
see
estcondini
Definition 6.2 page 173. A uniqueness result for the entropy process solution to
Problem (6.1)-(6.2) is then proven. This uniqueness result proves that the entropy process solution is
indeed the entropy weak solution, hence the existence and uniqueness of the entropy weak solution. This
uniqueness result also allows us toestschemaz
estschema
conclude to the convergence of the approximate solution given by the
estpbcont
estcondini
numerical scheme (that is (6.7), (6.5)) towards the entropy weak solution to (6.1)-(6.2) (this convergence
holds in Lploc (IR d × IR + ) for any 1 ≤ p < ∞).
estpbcont
estcondini
Note that uniqueness results for DP
“generalized” solutions
Sz
(namely measure
GH
valued solutions) to ( 6.1)-(6.2)
have recently been proved (see DiPerna [1985], Szepessy [1989], Gallouët and Herbin [1994]).
The proofs of these results rely on the one hand on the concept of measure valued solutions and on the
other hand on the existence of an entropy weak solution. The direct proof of the uniqueness of a measure
valued solution (i.e. without assuming any existence result of entropy weak solutions) leads to a difficult
problem involving the application of the theorem of continuity in mean. This difficulty is easier to deal
within the framework of entropy process solutions (but in fact, measure valued solutions and entropy
process solutions are two presentations of the same concept).

Developing the above analysis gives a (strong) convergence result of approximate solutions towards the
entropy weak solution. But moreover, we also derive some error estimates depending on the regularity of
u0 .

In the case of a Cartesian grid, the convergence and error analysis reduces essentiallyKuznetsov to a one-dimensional
discretization
CM
problem for whichS
results were proved some time ago, see e.g. Kuznetsov [1976],
Crandall and Majda [1980], Sanders [1983]. In the case of general meshes, the numerical schemes
are not generally “TVD”
bvmethod
(Total Variation Diminushing) and therefore the classical framework of the 1D
bvmethod
case (see Section 5.3.5 page 138) may not be used. More recent works dealCoquel with several convergence
results and error estimates
cgh
for time explicit finite volume schemes,Vi
see e.g. Cockburn,
kroner1
Coquel and
LeFloch [1994],
kroner2
Champier, Gallou ët and Herbin
kronerbook
[1993], Vila [1994], Kr öner and Rokyta
[1994], Kröner, Noelle and Rokyta [1995], Kr Sz
öner [1997]: following Szepessy’s work on the con-
vergence of the streamline DP
diffusion method (see Szepessy [1989]), most
GH
of these works use DiPerna’s
uniqueness
EGH2
theorem, see DiPerna [1985] (or an adaptation of it, see Gallouët and HerbinKuznetsov [1994] and
Eymard, Gallouët and Herbin [1995]), cgh
and the error estimates generalize the work
eggh
by Kuznetsov
[1976]. Here we use the framework of Champier, Gallouët and Herbin [1993], Eymard, Gallouët,
Ghilani and Herbin [1998]; we prove directly that any monotone flux scheme (defined below) satisfies
estwbvest estwbvestestwbvestiestwbvesti
a “weak BV ” estimate (see lemmata 6.2 page 153 and 6.3 page 159). This inequality appears to be a key
for the proof of convergence and for the error estimate. Some convergence results and error estimates are
also possible with some so called “higher order schemes” which are not monotone flux schemesNoelle
hyperhos hyperhos
(briefly
presented for the 1D
Chainais1
case in section 5.4 page 143). These results are not presented here, see No ëlle
[1996] and Chainais-Hillairet [1996] for some of them.

Note that the nonlinearity considered here is of the form v(x, t)f (u). This kind of flux is often encountered
in porous medium modelling, EG
where the hyperbolic equationVignal1
may then be coupled
Vignal2
with an elliptic
HL
or
parabolic equation (see e.g. Eymard and Gallouët [1993], Vignal [1996a], Vignal [1996b], Herbin
and Labergerie [1997]). It adds an extra difficulty to the case F (u) because of the dependency on x
baptisee
and t. Note again (see Remark 6.1) that the method which we present here for a nonlinearity of the form
v(x, t)f (u) also yields the same results in the case of a nonlinearity of the form F (x, t, u), see the recent
Chainais2
work of Chainais-Hillairet [1999].

The time implicit discretization

esteu
adds the extra difficulties of proving the existence of the approximate
esteu estlbvt
solution (see
estlbvt
Lemma 6.1 page 157) and proving a so called “strong time BV estimate” (see Lemma 6.5
1/4
page 162) in order to show that
√ the error estimate for the implicit scheme may still be of order h even
if the time step k is of order h, at least in particular cases.
eghbns
We first describe in section 6.2 finite volume schemes using a “general” mesh for the discretization of
 148

estpbcont
estcondini stabe stabi
(6.1)-(6.2). In sections 6.3 and 6.4entineq
some estimates on the approximate solution given by the numerical eghcvsol
schemes are shown and in Section 6.5 some entropy inequalities are proven. We then prove in section 6.6
the convergence of convenient subsequences of sequences of approximate solutions towards an entropy
process solution, by passing to the limit when the mesh size and theestcondini
time step go to 0. estepsol
estpbcont
A byproduct
estepsol
of this result is the existence of an entropy process solution to ( 6.1)-(
estpbcont
6.2) (see
estcondini
Definition 6.2 page 173).
The uniqueness of the entropy process solution to problem (6.1)-(6.2) is then proved; we can therefore
p
conclude to the existence and uniqueness of the entropy weak solution and also to the Leghcvsol loc convergence
for any finite p of the approximate solution towards the entropy weak solution (Section esterr
6.6). Using the
existence of the entropy weak solution, an error estimate result is given in Section 6.7 (which also yields
the convergence result). Therefore the mainestpbcont
interest of this convergence result is precisely to prove the
estcondini
existence of the entropy
nlwscv
weak solution to ( 6.1)-( 6.2) without any regularity assumption on the initial
data. Section 6.8 describeseghcvsol
the notion of nonlinear weak-? convergence, which is widely used in the proof
of convergence of section 6.6.
fve
Section 6.9 is not related to the previous sections. It describes a finite volume approach which may be
used to stabilize finite element schemes for the discretization of a hyperbolic equation (or system).

6.2 Meshes and schemes

eghbns
Let us first define an meshhyp1d
admissible mesh of IR d as a generalization of the notion of admissible mesh of IR as
meshhyp1d
defined in definition 5.5 page 125.

meshypmd Definition 6.1 (Admissible meshes) An admissible finite volume mesh of IR d , with d = 1, 2 or 3
estpbcont
estcondini
(for the discretization of Problem (6.1)-(6.2)), denoted by T , is given by a family of disjoint polygonal
connected subsets of IR d such that IR d is the union of the closure of the elements of T (which are called
control volumes in the following) and such that the common “interface” of any two control volumes is
included in a hyperplane of IR d (this is not necessary but is introduced to simplify the formulation).
Denoting by h = size(T ) = sup{diam(K), K ∈ T }, it is assumed that h < +∞ and that, for some α > 0,

αhd ≤ m(K),
(6.4) estmaillage
m(∂K) ≤ α1 hd−1 , ∀K ∈ T ,

where m(K) denotes the d-dimensional Lebesgue measure of K, m(∂K) denotes the (d − 1)-dimensional
Lebesgue measure of ∂K (∂K is the boundary of K) and N (K) denotes the set of neighbours of the
control volume K; for L ∈ N (K), we denote by K|L the common interface between K and L, and by
nK,L the unit normal vector to K|L oriented from K to L. The set of all the interfaces is denoted by E.

Note that, in this definition, the terminology is “mixed”. For d = 3,meshypmd

“polygonal” stands for “polyhedral”
meshhyp1d
and, for
meshhyp1d
d = 2, “interface” stands for “edge”. For d = 1 definition 6.1 is equivalent to definition 5.5 page
125.

In order to define the numerical flux, we consider functions g ∈ C(IR 2 , IR) satisfying the following
assumptions:
estdonnees
estcondflu Assumption 6.2 Under Assumption 6.1 the function g, only depending on f , v, Um and UM , satisfies

• g is locally Lipschitz continuous from IR 2 to IR,

• g(s, s) = f (s), for all s ∈ [Um , UM ],
• (a, b) 7→ g(a, b), from [Um , UM ]2 to IR, is nondecreasing with respect to a and nonincreasing with
respect to b.
 149

Let us denote by g1 and g2 the Lipschitz constants of g on [Um , UM ]2 with respect to its two arguments.

The hypotheses on gmonflux

are themonflux
same as those presented for monotone flux schemes in the one-dimensional
case (see
deflunum
definition 5.6 page 131); the function g allows the construction of a numerical flux, see Remark
6.3 below.
estcondflu
Remark 6.2 In Assumption 6.2, the third item will ensure some stability properties of the schemes
estschema
defined below. In particular, in the case of the “explicit scheme”
estcfl
(see ( 6.7)), it yields the monotonicity
of the scheme under a CFL condition (namely, condition (6.6) with ξ = 0). The second item is essentialflumono
since it ensures the consistency of the fluxes. All the examples of functions g given in Examples 5.2 page
flumono
132 satisfy these assumptions. We again give the important example of the “generalized 1D Godunov
scheme” obtained
Coquel
with a one-dimensional Godunov scheme Vi
for each interface (see e.g., for the explicit
scheme, see Cockburn, Coquel and LeFloch [1994], Vila [1994]),

max{f (s), b ≤ s ≤ a} if b ≤ a
g(a, b) =
min{f (s), a ≤ s ≤ b} if a ≤ b,
and also the framework of some “flux splitting” schemes:

g(a, b) = f1 (a) + f2 (b),

1
with f1 , f2 ∈ C (IR, IR), f = f1 + f2 , f1 nondecreasing and f2 nonincreasing (this framework is consider-
ably more simple that the general framework, because it reduces the study to the particular case of two
monotone nonlinearities).
estcondflu
Besides, it is possible to replace Assumption 6.2 on g by some slightly more general assumption,
estext
in order
to handle, in particular, the case of some “Lax-Friedrichs type” schemes (see Remark 6.11 below).

describe the numerical schemes considered here, let T be an admissible mesh in the sense of
In order tomeshypmd
Definition 6.1 and k > 0 be the time step. The discrete unknowns are unK , n ∈ IN ? , K ∈ T . The set {u0K ,
K ∈ T } is given by the initial condition,
Z
0 1
uK = u0 (x)dx, ∀ K ∈ T . (6.5) estschemaz
m(K) K
The equations satisfied by the discrete unknowns, unK , n ∈ IN? , K ∈ T , are obtained by discretizing
estpbcont
equation (6.1). We now describe the explicit and implicit schemes.

6.2.1 Explicit schemes

estcondflu
We present here the “explicit scheme” associated
estinftyest
to a function g satisfying Assumption 6.2. In this case,
estwbvest
for stability reasons (see lemmata 6.1 and 6.2), the time step k ∈ IR ?+ is chosen such that

α2 h
k ≤ (1 − ξ) , (6.6) estcfl
V (g1 + g2 )
where ξ ∈ (0, 1) is a given real value; recall that g1 and g2 are the Lipschitz constants of g with respect
to the first and second variables on [Um , UM ]2 and that Um ≤ u0 ≤ UM a.e. and |v(x, t)| ≤ V < +∞,
for all (x, t) ∈ IR d × IR + . Consider the following explicit numerical scheme:

un+1 − unK X

n
m(K) K
+ vK,L g(unK , unL ) − vL,K
n
g(unL , unK ) = 0, ∀K ∈ T , ∀n ∈ IN, (6.7) estschema
k
L∈N (K)

where
Z (n+1)k Z
n 1
vK,L = (v(x, t) · nK,L )+ dγ(x)dt
k nk K|L
 150

and
Z Z
1 (n+1)k
n
vL,K = (v(x, t) · nL,K )+ dγ(x)dt
k nk K|L
Z Z
1 (n+1)k
= (v(x, t) · nK,L )− dγ(x)dt.
k nk K|L

Recall that a+ = a>0 and a− = −(a⊥0) for all a ∈ IR and that dγ is the integration symbol for the
(d − 1)-dimensional Lebesgue measure on the considered hyperplane.

deflunum Remark 6.3 (Numerical fluxes) The numerical flux at the interface between the control volume K
n
and the control volume L ∈ N (K) is then equalmonflux
to vK,L g(unK , unL )−vL,K
monflux
n
g(unL , unK ); this expression yields
a monotone flux such as defined in definition 5.6 page 131, given in the one-dimensional case. However,
in the multidimensional case, the expression of the numerical flux depends on the considered interface;
this was not so in the one-dimensional case for which the numerical flux is completely defined by the
function g.

The approximate solution, denoted by uT ,k , is defined a.e. from IR d × IR + to IR by

uT ,k (x, t) = unK , if x ∈ K, t ∈ [nk, (n + 1)k), K ∈ T , n ∈ IN. (6.8) estuapp

6.2.2 Implicit schemes

The use of implicit schemes is steadily increasing in industrial codes for reasons such as robustness and
computational cost. Hence we consider in our analysis the following implicit numerical estcondflu
estcfl
scheme (for which
condition (6.6) is no longer needed) associated to a function g satisfying Assumption 6.2:

un+1 − unK X
m(K) K
+ n
(vK,L g(un+1 n+1 n n+1 n+1
K , uL ) − vL,K g(uL , uK )) = 0, ∀K ∈ T , ∀n ∈ IN. (6.9) estschemai
k
L∈N (K)

estschemaz
where {u0K , K ∈ T } is still determined by (6.5). The implicit approximate solution uT ,k , is defined now
a.e. from IR d × IR + to IR by

uT ,k (x, t) = un+1
K , if x ∈ K, t ∈ (nk, (n + 1)k], K ∈ T , n ∈ IN. (6.10) estuappimp

6.2.3 Passing to the limit

eghcvsol eghcvsol
We show in section 6.6 page 173 the convergence of the approximate solutions uT estcondini
,k (given by the numerical
estpbcont
schemes above described) towards the unique entropy weak solution u to (6.1)-(6.2) in an adequate sense,
when size(T ) → 0 and k → 0 (with, possibly, a stability condition). In order to describe the generalestschemaz line
ofestschema
thought estuapp
leading to this convergence result, we shall simply consider the explicit scheme (that is ( 6.5),
(6.7) and (6.8)) (the
stabe
implicit scheme will also be fully investigated later).
First, in section 6.3, by writing un+1
K as a convex combination of unK and (unL )L∈N (K) , the L∞ stability is
estcfl
easily shown under the CFL condition (6.6) (uT ,k is proved to be bounded in L∞ (IR d ×IR ?+ ), independently
of size(T ) and k).
By a classical argument, if any possible limit of a family of approximate
meshypmd meshypmd
solutions u T ,k (where T is an
estcfl
admissible mesh in the
estpbcont
sense
estcondini
of Definition 6.1 page 148 and k satisfies ( 6.6)) is the entropy weak solution
∞ d ?
to problem (6.1)-(6.2) then uT ,k converges
estcfl
(in L (IR × IR + ) for the weak-? topology, for instance),
estpbcont
h = size(T ) → 0 (and k satisfies (6.6)), towardsstabe
asestcondini the unique entropy weak solution to problem (6.1)-
(6.2). Unfortunately, the L∞ estimate of section 6.3 does not yield that any possible limit of a family
 151

estpbcont
estcondini
of approximate is solution hyper1d
to problem (6.1)-(6.2), even in the linear case (f (u) = u) (see the proofs
of convergence of Chapter 5). The “BV stability” can be used (combinedbvmethod
the L∞ stability) to
withbvmethod
show the convergence in the case of one space dimension (see section 5.3.5 page 138) and in the case of
Cartesian meshes in two or three space dimensions. Indeed, in the case of Cartesian meshes, assuming
u0 ∈ BV (IR d ) and assuming (for simplicity) v to be constant (a generalization is possible for v regular
enough), the following estimate holds, for all T ≥ k:
NT,k
X X
k m(K|L)|unK − unL | ≤ T |u0 |BV (IR d ) ,
n=0 K|L∈E
estschemaz
where NT,k ∈ IN is such that (NT,k + 1)k ≤ T < (NT,k + 2)k, and the values unK are given by (6.5) and
estschema meshypmd
(6.7). meshypmd
Such an estimate is wrong in the general case of admissible meshes in the sense of Definition 6.1
page 148, as it can be shown with easy counterexamples. It is, however, not necessary for the proof of
convergence.
nlwb1d
A weaker inequality, which is called “weak BV ” as in the one-dimensional case (see lemma
nlwb1d
5.5 page estwbvest
134) will estwbvest
be shown in the multidimensional case for both explicit and implicit schemes (see
estwbvestiestwbvesti
lemmata 6.2 page 153 and 6.3 page 159); the weak BV estimate yields the convergence of the scheme in
the general case. As an illustration,
estbvfx
consider
estbvfx
the case f 0 ≥ 0; using an upwind scheme, i.e. g(a, b) = fnlwb1d
nlwb1d
(a),
the weak BV inequality (6.16) page 153, which is very close to that of the 1D case (lemma 5.5 page 134),
writes
NT,k
X X C
n n
k (vK,L + vL,K )|f (unK ) − f (unL )| ≤ √ , (6.11) BVfaible
n
n=0 (K,L)∈ER h
n
where ER = {(K, L) ∈ T 2estwbvest
, L ∈ N (K), K|L ⊂ B(0, R) and unK > unL } and C only depends on v, g, u0 , α,
ξ, R and T (see Lemma 6.2).
BVfaible
We say that Inequality (6.11) is “weak”, but it is in fact “three times weak” for the following reasons:

1. the inequality is of order √1 , and not of order 1.

h
BVfaible n
2. In the left hand side of (6.11), the quantity which is associated to the K|L ∈ ER interface is zero
n n
if f is constant on the interval to which the values uK and uL belong; variations of the discrete
unknowns in this interval are therefore not taken into account.
BVfaible n n
3. The left hand side of (6.11) involves terms (vK,L + vL,K ) which are not uniformly bounded from
below by C m(K|L) with some C > 0 only depending on the data (that is v, u0 and g) and not on
n n
T (note that, for instance, vK,L = vL,K = 0 if v · nK,L = 0).
eghbcewsoleghbcewsol BVfaible
For the convergence result (namely Theorem 6.4 page 179) the useful consequence of (6.11) is
NT,k
X X
n n
h k (vK,L + vL,K )|f (unK ) − f (unL )| → 0 as h → 0,
n=0 n
(K,L)∈ER
nlcv1d nlcv1d esth14 esth14
as in
√ the 1D case, see Theorem 5.2 page 136. For the error estimate in Theorem 6.5 page 180, the bound
BVfaible
C/ h in (6.11) is crucial. Note
√ that a “twice weak BV ” inequality
BVfaible
in the sense (ii) and (iii), but of order
1 (that is C instead of C/ hesteqcorol
in the right hand side
esteqcorol
of ( 6.11)), would yield a sharp error estimate, i.e.
Ce h1/2 instead of Ce h1/4 in (6.94) page
BVfaible
180. estcfl
Note that, in order to obtain (6.11), ξ > 0 is crucial in the CFL condition (6.6).
BVfaible
Recall also that (6.11) together with the L∞ (IR d × IR ?+ ) bound does not yield any (strong) compactness
property in L1loc (IR d × IR + ) on a family of approximated solutions.
BVfaible
In the linear case (that is f (s) = cs for all s ∈ IR, for some c in IR), the inequality ( 6.11) is used in
the same mannerestpbcont
as inestcondini
the previous chapter; one proves that the approximate solution satisfies the weak
estkruzkov
formulation to (6.1)-(6.2) (which is equivalent to (6.3)) with an error which goes to 0 as h → 0, under
 152

estcfl estcfl
condition (6.6). We deduce from this the convergence of uT ,k (as h → 0 and under condition (6.6))
estpbcont
estcondini
towards the unique weak solution of (6.1)-(6.2) in L∞ (IR d × IR ?+ ) for the weak-? topology. In fact, the
convergence holds in Lploc (IR d × IR + ) (strongly) for any 1 ≤ p < ∞, thanks to the argument developped
for the study of the nonlinear case.
The nonlinear case adds an extra difficulty, as in the 1D case; it will be handled in detail in the present estcfl
chapter. This difficulty arises from the fact that, if uT ,k converges to u (as h → 0, under condition (6.6))
and f (uT ,k ) to µf , in L∞ (IR d × IR ?+ ) for the weak-? topology,
estpbcont
there remains to show thatBVfaible
estcondini
µf = f (u) and
that u is the entropy weak solution to problem (6.1)-(6.2). The weak BV inequality (6.11) is used to
show that, for any “entropy” function η, i.e. convex function of class C 1 from IR to IR, with associated
entropy flux φ, i.e. φ such that φ0 = f 0 η 0 , the following entropy inequality is satisfied:

 Z Z   Z
 µη (x, t)ϕt (x, t) + µφ (x, t)v(x, t) · ∇ϕ(x, t) dxdt + η(u0 (x))ϕ(x, 0)dx ≥ 0,
IR + IR d IR d (6.12) entropy

∀ϕ ∈ Cc∞ (IR d × IR + , IR + ),

L∞ (IR d × IR ?+ ) for the weak-? topology

where µη (resp. µφ ) is the limit of η(uT ,k ) (resp. φ(uT ,k )) in entropy
(the existence of these limits can indeed be assumed). From (6.12), it is shown that uT ,k converges to
estcfl
in L1loc
uestpbcont (IR d × IR ) (as h → 0, k satisfying (6.6)), and that u is the entropy weak solution to problem
estcondini +
(DP
6.1)-(6.2). This lastGH
result uses a generalization of a result on measure valued solutions of DiPerna (see
eghcvsol eghcvsol
DiPerna [1985], Gallouët and Herbin [1994]), and is developped in section 6.6 page 173.

6.3 Stability results for the explicit scheme

stabe
6.3.1 L∞ stability
estdonnees meshypmd
estinftyest Lemma 6.1 Under Assumption 6.1, let T estcondflube an admissible meshestcfl
in the sense of Definition 6.1 estuapp
and
2
kestschema
> 0, let g ∈ C(IR
estschemaz
, IR) satisfy Assumption 6.2 and assume that ( 6.6) holds; let u T ,k be given by ( 6.8),
(6.7), (6.5); then,
Um ≤ unK ≤ UM , ∀n ∈ IN, ∀K ∈ T , (6.13) estinduction
and

kuT ,k kL∞(IR d ×IR ?+ ) ≤ ku0 kL∞ (IR d ) . (6.14) eststabinf

estinftyest
Proof of Lemma 6.1
eststabinf estinduction
Note that (6.14) is a straightforward consequence
estinduction
of (6.13), which will be proved by induction. For n = 0,
estschemaz
since Um ≤ u0 ≤ UM a.e., (6.13) follows from (6.5).
n
Xn ∈ IN, assume that Um ≤ uK ≤ UM estschema
Let for all K ∈ T . Using the fact that divv = 0, which yields
n n
(vK,L − vL,K ) = 0, we can rewrite (6.7) as
L∈N (K)
un+1 − unK X  
n
m(K) K
+ vK,L (g(unK , unL ) − f (unK )) − vL,K
n
(g(unL , unK ) − f (unK )) = 0. (6.15) estschemab
k
L∈N (K)

Set, for unK 6= unL ,

n n g(unK , unL ) − f (unK ) n g(unL , unK ) − f (unK )

τK,L = vK,L n n − vL,K ,
uK − u L unK − unL
n
and τK,L if unK = unL .
= 0estcondflu estdonnees estschemab
n
Assumption 6.2 on g and Assumption 6.1 yields 0 ≤ τK,L ≤ V m(K|L)(g1 + g2 ). Using (6.15), we can
write
 153

 k X  k X
un+1
K = 1− n
τK,L unK + n
τK,L unL ,
m(K) m(K)
L∈N (K) L∈N (K)
estcfl
which gives, under condition (6.6), inf unL ≤ un+1
K ≤ sup unL , for all K ∈ T . This concludes the proof of
L∈T L∈T
estinduction eststabinf
(6.13), which, in turn, yields (6.14).
eststabinf estcfl
Remark 6.4 Note that the stability result (6.14) holds even if ξ = 0 in (6.6). However, we shall need
ξ > 0 for the following “weak BV ” inequality.

6.3.2 A “weak BV ” estimate

In the following lemma, B(0, R) denotes the ball of IR d of center 0 and radius R (IR d is always endowed
with its usual scalar product).
estdonnees meshypmd
estwbvest Lemma 6.2 Under Assumption 6.1, let T be an admissible meshestcfl
estcondflu
in the sense of Definition 6.1 estuapp
and
2
kestschema
> 0. Let g ∈ C(IR
estschemaz
, IR) satisfy Assumption 6.2 and assume that ( 6.6) holds. Let u T ,k be given by ( 6.8),
(6.7), (6.5).
n
Let T > 0, R > 0, NT,k = max{n ∈ IN, n < T /k}, TR = {K ∈ T , K ⊂ B(0, R)} and ER = {(K, L) ∈
2 n n
T , L ∈ N (K), K|L ⊂ B(0, R) and uK > uL }.
Then there exists C ∈ IR, only depending on v, g, u0 , α, ξ, R, T such that, for h < R and k < T ,
NT,k h  
X X
n
k vK,L max (g(q, p) − f (q)) + max (g(q, p) − f (p)) +
un
L
≤p≤q≤uK n un
L
≤p≤q≤uK n
n
n=0 (K,L)∈ER
 i
n
vL,K max (f (q) − g(p, q)) + max (f (p) − g(p, q)) (6.16) estbvfx
un
L
≤p≤q≤uK n un
L
≤p≤q≤uKn

C
≤ √ ,
h
and
NT,k
X X C
m(K)|un+1
K − unK | ≤ √ , (6.17) estbvft
n=0 K∈TR h
estwbvest
Proof of Lemma 6.2
we shall denote by Ci (i ∈ IN) various quantities only depending on v, g, u0 , α, ξ, R, T .
In this proof, estschemab
Multiplying (6.15) by kunK and summing the result over K ∈ TR , n ∈ {0, . . . , NT,k } yields

B1 + B2 = 0, (6.18) est1bv
with
NT,k
X X
B1 = m(K)unK (un+1
K − unK ),
n=0 K∈TR

and
NT,k
X X X  
n
B2 = k vK,L (g(unK , unL ) − f (unK ))unK − vL,K
n
(g(unL , unK ) − f (unK ))unK .
n=0 K∈TR L∈N (K)

Gathering the last two summations by edges in B2 leads to the definition of B3 :

 154

NT,k h  
X X
n
B3 = k vK,L unK (g(unK , unL ) − f (unK )) − unL (g(unK , unL ) − f (unL )) −
n
n=0 (K,L)∈ER
 i
n
vL,K unK (g(unL , unK ) − f (unK )) − unL (g(unL , unK ) − f (unL )) .

The expression |B3 −B2 | can be reduced to a sum of terms each of which corresponds to the boundary of a
control volume which is included in B(0, R+h)\B(0, R−h); since the measure of B(0, R+h)\B(0, R−h)
less than C2 h, the number of such terms is, for n fixed, lower than (C2 h)/(αhd ) = C3 h1−d . Thanks to
iseststabinf
(6.14), using the fact that m(∂K) ≤ (1/α)hd−1 , that |v(x, t)| ≤ V , that g is bounded on [Um , UM ]2 , and
that g(s, s) = f (s), one may show that each of the non zero term in |B3 − B2 | is bounded by C1 hd−1 .
Furthermore, since (NT,k + 1)k ≤ 2k, we deduce that

|B3 − B2 | ≤ C4 . (6.19) est5bv

2
Denoting by Φ a primitive of the function (·)f 0 (·), an integration by parts yields, for all (a, b) ∈ IR ,
Z b Z b
0
Φ(b) − Φ(a) = sf (s)ds = b(f (b) − g(a, b)) − a(f (a) − g(a, b)) − (f (s) − g(a, b))ds. (6.20) est6bv
a a
est6bv
Using (6.20), the term B3 may be decomposed as

B3 = B 4 − B 5 ,

where
NT,k Z Z !
X X un
L un
K
n
B4 = k vK,L (f (s) − g(unK , unL ))ds + n
vL,K (f (s) − g(unL , unK ))ds
n
n=0 (K,L)∈ER un
K
un
L

and
NT,k  
X X
n n
B5 = k (vK,L − vL,K ) Φ(unK ) − Φ(unL ) .
n
n=0 (K,L)∈ER

The term B5 is again reduced to a sum of terms corresponding

est5bv
to control volumes included in B(0, R +
h) \ B(0, R − h), thanks to divv = 0; therefore, as for (6.19), there exists C5 ∈ IR such that

B5 ≤ C 5 .
Let us now turn to an estimate of B4 . To this purpose, let a, b ∈ IR, define C(a, b) = {(p, q) ∈ [a⊥b, a>b]2;
(q − p)(b − a) ≥ 0}. Thanks to the monotonicity properties of g (and using the fact that g(s, s) = f (s)),
the following inequality holds, for any (p, q) ∈ C(a, b):
Z b Z d Z q
(f (s) − g(a, b))ds ≥ (f (s) − g(a, b))ds ≥ (f (s) − g(p, q))ds ≥ 0. (6.21) est7bv
a c p
petitlemmepetitlemme
The technical lemma 4.5 page 107 can then be applied. It states that
Z q
1
| (θ(s) − θ(p))ds| ≥ (θ(q) − θ(p))2 , ∀p, q ∈ IR,
p 2G
for all monotone, Lipschitz continuous function θ : IR → IR, with a Lipschitz constant G > 0.
petitlemme
From Lemma 4.5, we can notice that
Z q Z q
1
(f (s) − g(p, q))ds ≥ (g(p, s) − g(p, q))ds ≥ (f (p) − g(p, q))2 , (6.22) est8bv
p p 2g 2
 155

and
Z q Z q
1
(f (s) − g(p, q))ds ≥ (g(s, q) − g(p, q))ds ≥ (f (q) − g(p, q))2 . (6.23) est9bv
p p 2g1
est8bv est9bv
Multiplying (6.22) (resp. (6.23)) by g2 /(g1 + g2 ) (resp. g /(g1 + g2 )), taking the maximum for (p, q) ∈
est7bv 1
C(a, b), and adding the two equations yields, with (6.21),

Z b  
1
(f (s) − g(a, b))ds ≥ max (f (p) − g(p, q))2 + max (f (q) − g(p, q))2 . (6.24) est9bvbis
a 2(g1 + g2 ) (p,q)∈C(a,b) (p,q)∈C(a,b)
est9bvbis
We can then deduce, from (6.24):
NT,k
1 X X h
B4 ≥ k
2(g1 + g2 ) n=0 n
(K,L)∈ER
 
n 2 2 (6.25) est10bv
vK,L n
max n
(g(q, p) − f (q)) + max (g(q, p) − f (p)) +
un ≤p≤q≤un
uL ≤p≤q≤uK L K i
n
vL,K n
max n (f (q) − g(p, q))2 + n max n (f (p) − g(p, q))2 .
uL ≤p≤q≤uK uL ≤p≤q≤uK

This gives a bound on B2 , since (with C6 = C4 + C5 ):

B2 ≥ B 4 − C 6 . (6.26) est10bva

Let us now turn to B1 . We have

NT,k    2
1X X 1 X NT,k +1 2 1 X
B1 = − m(K)(un+1
K − u n 2
K ) + m(K) u K − m(K) u0K . (6.27) est11bv
2 n=0 2 2
K∈TR K∈TR K∈TR
estschemab
Using (6.15) and the Cauchy-Schwarz inequality yields the following inequality:

(un+1
K − unK )2 ≤
k 2 X X h  2  2 i
n n n
2
(vK,L + vL,K ) vK,L g(unK , unL ) − f (unK ) + vL,K
n
g(unL , unK ) − f (unK ) .
m(K)
L∈N (K) L∈N (K)
estcfl meshypmd estdonnees
Then, using the CFL condition (6.6), Definition 6.1 and part (iv) of Assumption 6.1 gives

m(K)(un+1K − unK )2 ≤
1−ξ X h  2  2 i
k n
vK,L n
g(unK , unL ) − f (unK ) + vL,K g(unL , unK ) − f (unK ) . (6.28) est13bv
g1 + g 2
L∈N (K)
est13bv
Summing equation (6.28) over K ∈ TR and over n = 0, . . . , NT,k , and reordering the summation leads to
NT,k NT,k
1X X 1−ξ X X h
m(K)(un+1K − u n 2
K ) ≤ k
2 n=0 2(g1 + g2 ) n=0 n
K∈TR (K,L)∈ER
  (6.29) est13bvbis
n n n n 2 n n n 2
vK,L (g(uK , uL ) − f (uK )) + (g(uK , uL ) − f (uL )) +
 i
n
vL,K (f (unK ) − g(unL , unK ))2 + (f (unL ) − g(unL , unK ))2 + C7 ,

where C7 accounts for the interfaces K|L ⊂ B(0, R) such that K ∈

/ TR and/or L ∈
/ TR (these control
volumes are included in B(0, R + h) \ B(0, R − h)).
 156

est13bvbis est10bv est1bv est10bva

that the right hand side of (6.29) is bounded by (1 − ξ)B4 + C7 (from (6.25)). Using (6.18), (6.26)
Note est11bv
and (6.27) gives

NT,k h  
ξ X X
k n
vK,L max (g(q, p) − f (q))2 + max (g(q, p) − f (p)) 2
+
2(g1 + g2 ) n=0 n
un
L
≤p≤q≤uK n un
L
≤p≤q≤uKn
(K,L)∈ER
 i
n
vL,K max (f (q) − g(p, q))2 + n max n (f (p) − g(p, q))2 (6.30) est14bv
un ≤p≤q≤un uL ≤p≤q≤uK
1 X
L K
 2
0
≤ m(K) uK + C6 + C7 = C8 .
2
K∈TR
estbvfx est14bv
Applying the Cauchy-Schwarz inequality to the left hand side of (6.16) and using (6.30) yields
NT,k h  
X X
n
k vK,L max (g(q, p) − f (q)) + max (g(q, p) − f (p)) +
n
un
L
≤p≤q≤uK n un
L
≤p≤q≤uK n
n=0 (K,L)∈ER
 i
n
vL,K max (f (q) − g(p, q)) + max (f (p) − g(p, q)) (6.31) est15bv
un
L
≤p≤q≤uK n un
L
≤p≤q≤uK n

NT,k
X X  12
n n
≤ C9 k (vK,L + vL,K ) .
n
n=0 (K,L)∈ER

Noting that
X X 1 d−1 m(B(0, R + h)) C10
n n
(vK,L + vL,K )≤ V m(∂K) ≤ V h d
=
n
α αh h
(K,L)∈ER K∈TR+h
estbvfx est15bv
and (NT,k + 1)k ≤ 2T , one obtains (6.16) from (6.31).
estschemab
Finally, since (6.15) yields
X  
m(K)|un+1
K − unK | ≤ k n
vK,L |g(unK , unL ) − f (unK )| + vL,K
n
|g(unL , unK ) − f (unK )| ,
L∈N (K)
estbvft estbvfx estwbvest
Inequality (6.17) immediately follows from (6.16). This completes the proof of Lemma 6.2.

6.4 Existence of the solution and stability results for the implicit
scheme
stabi
estschemaiestschemaz
This section is devoted to the time implicit scheme (given estschemaz
by (6.9) and (6.5)). We first prove the existence
estschemai
and uniqueness of the solution {unK , n ∈ IN, K ∈ T } of (6.5), (6.9) and such that unK ∈ [Um , UM ] for all
∈ T and all n ∈ IN. Then, one gives a “weak space BV ” inequality (this is equivalent
Kestbvfx estbvt
to the inequality
(6.16) for the explicit scheme) and a “(strong) time BV ” estimate (Estimate ( 6.45) below). This last
estimate requires that
esteqcoroli
v does not depend on t (and it leads to the term “k” in the right hand side
esth14i
of
estexti
(6.95) in Theorem 6.6). The error estimate, in the case where v depends on t, is given in Remark 6.12.

6.4.1 Existence, uniqueness and L∞ stability

estschemaz
estschemai
The following proposition gives an existence and uniqueness result of the solution to ( 6.5), (6.9). In this
proposition, v may depend on t and one does not need to assume u0 ∈ BV (IR d ).
 157

estdonnees meshypmd
esteu Proposition 6.1 Under Assumption 6.1, let Testcondflu
be an admissible mesh in the sense of Definition 6.1 and
k > 0. Let g ∈ C(IR 2 , IR) satisfy Assumption 6.2. estschemaz
estschemai
Then there exists a unique solution {unK , n ∈ IN, K ∈ T } ⊂ [Um , UM ] to (6.5),(6.9).
esteu
Proof of Proposition 6.1
esteu estschemaz
One proves Proposition 6.1 by induction. Indeed, {u0K , K ∈ T } is uniquely defined by (6.5) and one has
uK ∈ [Um , UM ], for all K ∈ T , since Um ≤ u0 ≤ UM a.e.. Assuming that, for some n ∈ IN, the set {unK ,
0

, the existence and uniqueness of {un+1

K ∈ T } is given and that unK ∈ [Um , UM ], for all K ∈ T estschemai K , K ∈ T },
n+1
such that uK ∈ [Um , UM ] for all K ∈ T , solution of (6.9), must be shown.
estschemai
Step 1 (uniqueness of {un+1
K , K ∈ T }, such that uK
n+1
∈ [Um , UM ] for all K ∈ T , solution of (6.9))
estschemai
Recall that n ∈ IN and {unK , K ∈ T } are given. Let us consider two solutions of (6.9), respectively
denoted by {uK , K ∈ T } and {wK , K ∈ T }; therefore, {uK , K ∈ T } and {wK , K ∈ T } satisfy {uK ,
K ∈ T } ⊂ [Um , UM ], {wK , K ∈ T } ⊂ [Um , UM ],
uK − unK X
n n
m(K) + (vK,L g(uK , uL ) − vL,K g(uL , uK )) = 0, ∀K ∈ T , (6.32) estschemaiut
k
L∈N (K)

and
wK − unK X
n n
m(K) + (vK,L g(wK , wL ) − vL,K g(wL , wK )) = 0, ∀K ∈ T . (6.33) estschemaiwt
k
L∈N (K)
estschemaiwt
estschemaiut
Then, substracting (6.33) to (6.32), for all K ∈ T ,

m(K) X
n
(uK − wK ) + vK,L (g(uK , uL ) − g(wK , uL ))
k
L∈N (K)
X X
n n
+ vK,L (g(wK , uL ) − g(wK , wL )) − vL,K (g(uL , uK ) − g(wL , uK )) (6.34) estschemaiu1
L∈N (K) L∈N (K)
X
n
− vL,K (g(wL , uK ) − g(wL , wK )) = 0
L∈N (K)
estschemaiu1t
thanks to the monotonicity properties of g, (6.34) leads to

m(K) X
n
|uK − wK | + vK,L |g(uK , uL ) − g(wK , uL )|
k
L∈N (K)
X X
n n
+ vL,K |g(wL , uK ) − g(wL , wK )| ≤ vK,L |g(wK , uL ) − g(wK , wL )| (6.35) estschemaiu2
L∈N (K) L∈N (K)
X
n
+ vL,K |g(uL , uK ) − g(wL , uK )|.
L∈N (K)

Let ϕ : IR 7→ IR ?+ be defined by ϕ(x) = exp(−γ|x|), for some positive γ which will be specified later.
d
estmaillage
For K ∈P T , let ϕK be the mean value of ϕ on K. Since ϕ is integrable over IR d (and thanks to (6.4)),
one has K∈T ϕK ≤ (1/(αhd ))kϕkL1 (IR d ) < ∞. Therefore the series

X X X X
n n
ϕK ( vK,L |g(wK , uL ) − g(wK , wL )|) and ϕK ( vL,K |g(uL , uK ) − g(wL , uK )|)
K∈T L∈N (K) K∈T L∈N (K)

estmaillage
(thanks to (6.4) and the boundedness of v on IR d and g on [Um , UM ]2 ).
are convergentestschemaiu2t
Multiplying (6.35) by ϕK and summing for K ∈ T yields five convergent series which can be reordered
in order to give
 158

X m(K) X X
n
|uK − wK |ϕK ≤ vK,L |g(uK , uL ) − g(wK , uL )||ϕK − ϕL |
k
K∈T K∈T L∈N (K)
X X
n
+ vL,K |g(wL , uK ) − g(wL , wK )||ϕK − ϕL |,
K∈T L∈N (K)

from which one deduces

X X
aK |uK − wK | ≤ bK |uK − wK |, (6.36) estschemaiu4
K∈T K∈T

m(K)
X
n n
with, for all K ∈ T , aK = k ϕK and bK = (vK,L g1 + vL,K g2 )|ϕK − ϕL |.
L∈N (K)
For K ∈ T , let xK be an arbitrary point of K. Then,
1 d
aK ≥ αh inf{ϕ(x), x ∈ B(xK , h)}
k
and

2V (g1 + g2 ) d
bK ≤ h sup{|∇ϕ(x)|, x ∈ B(xK , 2h)}.
α
Therefore, taking γ > 0 small enough in order to have

inf{ϕ(y), y ∈ B(x, h)} > C sup{|∇ϕ(y)|, y ∈ B(x, 2h)}, ∀x ∈ IR d (6.37) choixg

2
estschemaiu4t
yields aK > bK for all K ∈ T . Hence (6.36) gives uK = wK , for all K ∈ T .
with C = (2kV (g1 + g2 ))/α ,choixg
A choice of γ > 0 verifying (6.37) is always possible. Indeed, since |∇ϕ(z)| = γ exp(−γ|z|), taking γ > 0
such that γ exp(3γh) < 1/C is convenient.
This concludes Step 1.
estschemai
Step 2 (existence of {un+1
K , K ∈ T }, such that uK
n+1
∈ [Um , UM ] for all K ∈ T , solution of (6.9)).
n
Recall that n ∈ IN and {uK , K ∈ T } are given. For r ∈ IN ? , let Br = B(0, r) = {x ∈ IR d , |x| < r} and
estwbvest
Tr = {K ∈ T , K ⊂ Br } (as in Lemma 6.2). Let us assume that r is large enough, say r ≥ r0 , in order to
have Tr 6= ∅.
(r) (r)
If K ∈ T \ Tr , set uK = unK . Let us first prove that there exists {uK , K ∈ Tr } ⊂ [Um , UM ], solution to
(r) X
uK − unK n (r) (r)
n (r) (r)
m(K) + (vK,L g(uK , uL ) − vL,K g(uL , uK )) = 0, ∀K ∈ Tr . (6.38) estschemair
k
L∈N (K)

that passing to the limit as r → ∞ (up to a subsequence) leads to a solution

Then, we will proveestschemai
{un+1 n+1
K , K ∈ T } to (6.9) such that uK ∈ [Um , UM ] for all K ∈ T .
(r) estschemair
For a fixed r ≥ r0 , in order to prove the existence of {uK deimling
, K ∈ Tr } ⊂ [Um , UM ] solution to (6.38),
a “topological degree” argument is used (see, for instance, Deimling [1980] for a presentation of the
degree).
(r) estschemair
Let Urn = {unK , K ∈ Tr } and assume that Ur = {uK , K ∈ Tr } is a solution of (6.38). The families Ur
estschemair
and Urn may be viewed as vectors of IR N , with N = card(Tr ). Equation (6.38) gives

(r) k X (r) (r) (r) (r)

n n
uK + (vK,L g(uK , uL ) − vL,K g(uL , uK )) = unK , ∀K ∈ Tr ,
m(K)
L∈N (K)

which can be written on the form

Ur − Gr (Ur ) = Urn , (6.39) estschemairb

 159

where Gr is a continuous map from IR N into IR N .

One may assume that g is nondecreasing with respect to its first argument and nonincreasing with
second argument on IR 2 (indeed, thanks to the monotonicity properties of g given by
respect to itsestcondflu
Assumption 6.2, it is sufficient to change, if necessary, g on IR 2 \[Um , UM ]2 , setting, for instance, g(a, b) =
(r)
g(Um >(UM ⊥a), Um >(UM ⊥b))). Then, since unK ∈ [Um , UM ], for all K ∈ T , and uKestschemairb = unK ∈ [Um , UM ],
for all K ∈ T \ Tr , it is easy to show (using div(v) = 0) that if Ur satisfies (6.39), then one has
(r)
UM ], for all K ∈ Tr . Therefore, if Cr is a ball of IR N of center 0 and of radius great enough,
uK ∈ [Um ,estschemairb
Equation (6.39) has no solution on the boundary of Cr , and one can define the topological degree of
the application Id − Gr associated to the set Cr and to the point Urn , that is deg(Id − Gr , Cr , Urn ).
Furthermore, if λ ∈ [0, 1], the same argument allows us to define deg(Id − λGr , Cr , Urn ). Then, the
property of invariance of the degree by continuous transformation asserts that deg(Id − λG r , Cr , Urn ) does
not depend on λ ∈ [0, 1]. This gives

deg(Id − Gr , Cr , Urn ) = deg(Id, Cr , Urn ).

But, since Urn ∈ Cr ,

deg(Id, Cr , Urn ) = 1.
Hence

deg(Id − Gr , Cr , Urn ) 6= 0.
estschemairb
This proves that there exists a solution Ur ∈ Cr to (6.39). Recall also that we already proved that the
components of Ur are necessarily in [Um , UM ].
estschemai
In order to prove the existence of {un+1
K , K ∈ T } ⊂ [Um , UM ] solution to (6.9), let us pass to the limit as
(r) estschemair (r)
r → ∞. For r ≥ r0 , let {uK , K ∈ T } be a solution of (6.38) (given by the previous proof). Since {uK ,
r ∈ IN} is included in [Um , UM ], for all K ∈ T , one can find (using a “diagonal process”) a sequence
rl
(rl )l∈IN , with rl → ∞, as l → ∞, such that (u K )l∈IN converges (in [Um , UM ]) for all K ∈ T . One sets
estschemair
n+1 rl
K ∈ T , this equation is
uK = liml→∞ uK . Passing to the limit in (6.38) (this is possible since for allestschemai
satisfied for all l ∈ IN large enough) shows that {un+1
K , K ∈ T } is solution to ( 6.9).
estschemai (r)
Indeed, using the uniqueness of the solution of (6.9), one can show that uK → un+1 K , as r → ∞, for all
K ∈T. esteu
This completes the proof of Proposition 6.1.

6.4.2 “Weak space BV ” inequality

estbvfx
One gives here an inequality similar to Inequality (6.16) (proved for the explicit
estbvft
scheme). This inequality
d
does not make use
estlbvt
of u 0 ∈ BV (IR ) and v can depend on t. Inequality ( 6.17) also holds but is improved
in Lemma 6.5 when u0 ∈ BV (IR d ) and v does not depend on t.
estdonnees meshypmd
estwbvesti Lemma 6.3 Under Assumption 6.1, let T beestcondflu
an admissible mesh in the sense of Definition 6.1 estschemai
and
> 0. Let g ∈ C(IR 2 , IR) satisfy Assumption 6.2 and let {unK , n ∈ IN, K ∈ T } be the solution of (6.9),
kestschemaz
(6.5) such that un+1
K ∈ [Um , UMesteu
] for all K ∈ T and all n ∈ IN (existence and uniqueness of such a
solution is given by Proposition 6.1).
n
Let T > 0, R > 0, NT,k = max{n ∈ IN, n < T /k}, TR = {K ∈ T , K ⊂ B(0, R)} and ER = {(K, L) ∈
2 n n
T , L ∈ N (K), K|L ⊂ B(0, R) and uK > uL }.
Then there exists Cv ∈ IR, only depending on v, g, u0 , α, R, T such that, for h < R and k < T ,
 160

NT,k h  
X X
n
k vK,L max (g(q, p) − f (q)) + max (g(q, p) − f (p)) +
n=0 (K,L)∈E n+1 un+1
L
≤p≤q≤uKn+1
un+1
L
≤p≤q≤uKn+1

R  i
n
vL,K max (f (q) − g(p, q)) + max (f (p) − g(p, q)) (6.40) estbvfxi
un+1
L
≤p≤q≤un+1
K
un+1
L
≤p≤q≤un+1
K
Cv
≤ √ .
h
estbvft estbvft
Furthermore, Inequality 6.17 page 153 holds.
estwbvesti
Proof of Lemma 6.3
estschemai
We multiply (6.9) by kun+1 K , and sum the result over K ∈ TR andest11bv
estwbvest estwbvest
n ∈ {0, . . . , NT,k }. We can then follow,
step by step, the proof of Lemma 6.2 page 153 until Equation (6.27) in which the first term of the est14bv right
hand side appears with the opposite sign. We can thenestbvfxi
directly concludeestbvft
an inequality similar
estbvft
to ( 6.30),
which is sufficient to conclude the proof of Inequality (6.40). Inequality 6.17 page 153 follows easily from
estbvfxi
(6.40).

6.4.3 “Time BV ” estimate

estbvt
This section gives a so called “strong time BV estimate” (estimate (6.45)). For this estimate, the fact that
u0 ∈ BV (IR d ) and that v does not depend on t is required. Let us begin this section with a preliminary
lemma on the space BV (IR d ).
meshypmd meshypmd
divbv Lemma 6.4 Let defBVRp
admissible mesh in the sense of Definition 6.1 page 148 and let u ∈ BV (IR d )
T be andefBVRp
(see Definition 5.38 page 138). For K ∈ T , let uK be the mean value of u over K. Then,
X C
m(K|L)|uK − uL | ≤ |u| d , (6.41) bvzd
α4 BV (IR )
K|L∈E

where C only depends on the space dimension (d = 1, 2 or 3).

divbv
Proof of Lemma 6.4
divbv bvzd
Lemma 6.4 is proven bvzd
that if (6.41) holds for all u ∈ BV (IR d ) ∩
in two steps. In the first step, it is provedbvzd
C 1 (IR , IR) then (6.41) holds for all u ∈ BV (IR ). In Step 2, (6.41) is proved to hold for u ∈ BV (IR d ) ∩
d d

C 1 (IR d , IR).
Step 1 (passing from BV (IR d ) ∩ C 1 (IR d , IR) to BV (IR d ))
Recall that BV (IR d ) ⊂ L1loc (IR d ). Let u ∈ BV (IR d ), let us regularize u by a sequence of mollifiers.
R
Let ρ ∈ Cc∞ (IR d , IR + ) such that IR d ρ(x)dx = 1. Define, for all n ∈ IN ? , ρn by ρn (x) = nd ρ(nx) for all
x ∈ IR d and un = u ? ρn , that is
Z
un (x) = u(y)ρn (x − y)dy, ∀x ∈ IR d .
IR d

It is well known that (un )n∈IN ? is included in C ∞ (IR d , IR) and converges to u in L1loc (IR d ) as n → ∞.
bvzd
Then, the mean value of un over K converges, as n → ∞, to uK , for all K ∈ T . Hence, if (6.41) holds
with un instead
bvzd
of u (this will be proven in Step 2) and if |un |BV (IR d ) ≤ |u|BV (IR d ) for all n ∈ IN? ,
Inequality (6.41) is proved by passing to the limit as n → ∞.
In order to prove |un |BV (IR d ) ≤ |u|BV (IR d ) for all n ∈ IN ? (this will conclude step 1), let n ∈ IN ? and
ϕ ∈ Cc∞ (IR d , IR d ) such that |ϕ(x)| ≤ 1 for all x ∈ IR d . A simple computation gives, using Fubini’s
theorem,
 161

Z Z Z


un (x)divϕ(x)dx = u(x − y)divϕ(x)dx ρn (y)dy ≤ |u|BV (IR d ) , (6.42) bba
IR d IR d IR d

since, setting ψy = ϕ(y + ·) ∈ Cc∞ (IR d , IR d ) (for all y ∈ IR d ),

Z Z
u(x − y)divϕ(x)dx = u(z)divψy (z)dz ≤ |u|BV (IR d ) , ∀y ∈ IR d ,
IR d IR d
and
Z
ρn (y)dy = 1.
IR d
bba
Then, taking in (6.42) the supremum over ϕ ∈ Cc∞ (IR d , IR d ) such that |ϕ(x)| ≤ 1 for all x ∈ IR d leads to
|un |BV (IR d ) ≤ |u|BV (IR d ) .
bvzd
Step 2 (proving (6.41) if u ∈ BV (IR d ) ∩ C 1 (IR d , IR))
Recall that B(x, R) denotes the ball of IR d of center x and radius R. Since u ∈ C 1 (IR d , IR),
Z Z
u(x)divϕ(x)dx = − ∇u(x) · ϕ(x)dx.
IR d IR d
bvzd
Then |u|BV (IR d ) = k(|∇u|)kL1 (IR d ) and we will prove (6.41) with k(|∇u|)kL1 (IR d ) instead of |u|BV (IR d ) .
Let K|L ∈ E, then K ∈ T , L ∈ N (K) and
Z Z
1
uK − u L = (u(x) − u(y))dxdy.
m(K)m(L) L K
For all x ∈ K and all y ∈ L,
Z 1
u(x) − u(y) = ∇u(y + t(x − y)) · (x − y)dt.
0
Then,
Z Z Z 1
m(K)m(L)|uK − uL | ≤ ( |∇u(y + t(x − y))||x − y|dtdx)dy
Z LZ K
1 Z 0

≤ ( |∇u(y + t(x − y))||x − y|dxdt)dy.

L 0 K

Using |x − y| ≤ 2h and changing the variable x in z = x − y (for all fixed y ∈ L and t ∈ (0, 1)) yields
Z Z 1 Z
m(K)m(L)|uK − uL | ≤ 2h ( |∇u(y + tz)|dzdt)dy,
L 0 B(0,2h)

which may also be written (using Fubini’s theorem)

Z Z 1 Z
m(K)m(L)|uK − uL | ≤ 2h ( |∇u(y + tz)|dydt)dz. (6.43) oufa
B(0,2h) 0 L

For all K ∈ T , let xK be an arbitrary point of K. oufa

Then, changing the variable y in ξ = y + tz (for all fixed z ∈ L and t ∈ (0, 1)) in (6.43),
Z Z 1 Z
m(K)m(L)|uK − uL | ≤ 2h ( |∇u(ξ)|dξdt)dz,
B(0,2h) 0 B(xL ,3h)
estmaillage
estmaillage
which yields, since T is an adimissible mesh in the sense of Definition 6.4 page 148,
 162

Z
2hd
m(K|L)|uK − uL | ≤ m(B(0, 2h)) |∇u(ξ)|dξ.
α3 h2d B(xL ,3h)

Therefore there exists C1 , only depending on the space dimension, such that
Z
C1
m(K|L)|uK − uL | ≤ 3 |∇u(ξ)|dξ, ∀K|L ∈ E. (6.44) oufb
α B(xL ,3h)
Let us now remark that, if M ∈ T and L ∈ T , M ∩ B(xL , 3h) 6= ∅ implies L ⊂ B(xM , 5h). Then, for a
fixed M ∈ T , the number of L ∈ T such that M ∩ B(xL , 3h) 6= ∅ is less or equal to m(B(0, 5h))/(αhd )
that is less or equal C /α where C2 only depends on the space dimension.
oufb 2
Then, summing (6.44) over K|L ∈ E leads to
X Z
C1 C2 X C1 C2
m(K|L)|uK − uL | ≤ 4
|∇u(ξ)|dξ = k(|∇u|)kL1 (IR d ) ,
α M α4
K|L∈E M ∈T
bvzd
that is (6.41) with C = C1 C2 .
divbv bvzd
Note that, in Lemma 6.4 the bvz
estimate (6.41) depends on α. This dependency on α is not necessary in the
divbv1d
one dimensinal case (see (5.6) in Remark 5.4) and for particular meshes in the two and three dimensional
cases. Recall also that, except if d = 1, the space BV (IR d ) is not includeddivbv1d
in L∞ (IR d ). In particular, it
is then quite easy tobvzd
prove that, contrary to the 1D case given in Remark 5.4, it is not possible, for d = 2
or 3, to replace, in (6.41), uK by the mean value of u over an arbitrary ball (for instance) included in K.
Let us now give the “strong time BV estimate”.
estdonnees meshypmd
estlbvt Lemma 6.5 Under Assumption 6.1, let T be an admissible mesh in the sense of Definition 6.1 and
estcondflu
k > 0. Let g ∈ C(IR 2 , IR) satisfy Assumption 6.2. Assume that u0 ∈ BV (IR d ) and that v does not
depend on t. estschemai
estschemaz
Let {unK , n ∈ IN, K ∈ T } be the solution of (6.9), (6.5) such that unK ∈ [Umesteu all K ∈ T and all
, UM ] foresteu
n ∈ IN (existence and uniqueness of such a solution is given by Proposition 6.1 page 157).
Then, there exists Cb , only depending on v, g, u0 and α such that
X m(K)
|un+1
K − unK | ≤ Cb , ∀n ∈ IN. (6.45) estbvt
k
K∈T
estlbvt
Proof of lemma 6.5
n
Since v does not depend on t, one denotes vK,L = vK,L , for all K ∈ T and all L ∈ N (K).
For n ∈ IN, let
X |un+1 − unK |
K
An = m(K)
k
K∈T

and
X X
Bn = | [vK,L g(unK , unL ) − vL,K g(unL , unK )]|.
K∈T L∈N (K)

Since u0 ∈ BV (IR d ) and divv = 0, there exists Cb > 0, only depending on v, g, u0 and α, such that
B0 ≤ Cb . Indeed,
X X
B0 ≤ V (g1 + g2 )m(K|L)|u0K − u0L |.
K∈T L∈N (K)
 163

divbv
Thanks to lemma 6.4, B0 ≤ Cb with Cb = 2V (g1 + g2 )C(1/α4 )|u0 |BV (IR d ) , where C only depends on the
space dimension (d = 1, 2 or 3).
estschemai estlbvt
From (6.9), one deduces that Bn+1 ≤ An , for all n ∈ IN. In order to prove Lemma 6.5, there only remains
to prove that An ≤ Bn for all n ∈ IN (and to conclude by induction).
estschemai
Let n ∈ IN, in order to prove that An ≤ Bn , recall that the implicit scheme (6.9) writes

un+1 − unK X  
m(K) K
+ vK,L g(un+1 n+1 n+1 n+1
K , uL ) − vL,K g(uL , uK ) = 0. (6.46) estschemir
k
L∈N (K)
estschemir
From (6.46), one deduces, for all K ∈ T ,

un+1 − unK X
m(K) K
+ vK,L (g(un+1 n+1 n n+1
K , uL ) − g(uK , uL ))
k
X  L∈N (K)  X  
+ vK,L g(unK , un+1
L ) − g(u n
K , u n
L ) − v L,K g(u n+1
L , u n+1
K ) − g(u n
L , u n+1
K )
L∈N (K)   L∈N (K)
X
− vL,K g(unL , un+1
K ) − g(unL , unK )
L∈N (K)
X X
=− vK,L g(unK , unL ) + vL,K g(unL , unK ).
L∈N (K) L∈N (K)

Using the monotonicity properties of g, one obtains for all K ∈ T ,

|un+1 − unK | X
m(K) K
+ vK,L |g(un+1 n+1 n n+1
K , uL ) − g(uK , uL )|
k
L∈N (K)
X
+ vL,K |g(uL , un+1
n n n
K ) − g(uL , uK )|
L∈N (K) (6.47)
X X estschemirb
≤|− vK,L g(unK , unL ) + vL,K g(unL , unK )|
L∈N (K) L∈N (K)
X X
+ vK,L |g(unK , un+1
L ) − g(u n n
K , uL )| + vL,K |g(un+1 n+1 n n+1
L , uK ) − g(uL , uK )|.
L∈N (K) L∈N (K)
esteu
In order to deal with convergent series, let us proceed as in the proof of proposition 6.1. For 0 < γ < 1,
let ϕγ : IR d 7→ IR ?+ be defined by ϕγ (x) = exp(−γ|x|). esteu
For K ∈ T , let ϕγ,K be the mean value of ϕγ on K. As in Proposition 6.1, since ϕγ is integrable over
P estschemirb
IR d , K∈T ϕγ,K < ∞. Therefore, multiplying (6.47) by ϕγ,K (for a fixed γ) and summing over K ∈ T
yields six convergent series which can be reordered to give
X |un+1 − unK |
m(K) K ϕγ,K
k
K∈T
X X X
≤ |− vK,L g(unK , unL ) + vL,K g(unL , unK )|ϕγ,K
K∈T L∈N (K) L∈N (K)
X X
+ vK,L |g(un+1
K , u n+1
L ) − g(unK , un+1
L )||ϕγ,K − ϕγ,L |
K∈T L∈N (K)
X X
+ vL,K |g(unL , un+1 n n
K ) − g(uL , uK )||ϕγ,K − ϕγ,L |.
K∈T L∈N (K)

For K ∈ T , let xK ∈ K be such that ϕγ,K = ϕγ (xK ). Let K ∈ T and L ∈ N (K). Then there exists
s ∈ (0, 1) such that ϕγ,L − ϕγ,K = ∇ϕγ (xK + s(xL − xK )) · (xL − xK ). Using |∇ϕγ (x)| = γ exp(−γ|x|),
this yields |ϕγ,L − ϕγ,K | ≤ 2hγestdonnees
exp(2hγ)ϕ γ,K ≤ 2hγ exp(2h)ϕγ,K .
estcondflu
Then, using the assumptions 6.1 and 6.2, there exists some a only depending on k, V , h, α, g1 and g2
such that
 164

X |un+1 − unK |
m(K) K ϕγ,K (1 − γa)
k
K∈T
X X X
≤ |− vK,L g(unK , unL ) + vL,K g(unL , unK )|ϕγ,K ≤ Bn .
K∈T L∈N (K) L∈N (K)

Passing
estlbvt
to the limit in the latter inequality as γ → 0 yields An ≤ Bn . This completes the proof of Lemma
6.5.

6.5 Entropy inequalities for the approximate solution

entineq estentest
In this section, an entropy estimate on the approximate solution is proved (Theorem 6.1), which will
be used in the proofs of convergence and error estimate of the numerical scheme. In order to obtain
this entropy estimate, some discrete entropy inequalities satisfied by the approximate solution are first
derived.

6.5.1 Discrete entropy inequalities

estschema
In the case of the explicit scheme, the following lemma asserts that the scheme (6.7) satisfies a discrete
godlewski-ellipses
entropy condition
godlewski-springer
(this is classical in the study of 1D schemes, see e.g. Godlewski and Raviart [1991],
Godlewski and Raviart [1996]).
estdonneesestdonnees meshypmd
estldei Lemma 6.6 Under assumption 6.1 page 145, let T be an admissible mesh estcfl
estcondflu
in the sense of Definition 6.1
2
and k > 0. Let g ∈ C(IR , IR) satisfying
estuappestschema
assumption
estschemaz
6.2 and assume that ( 6.6) holds.
Let uT ,k be given by (6.8), (6.7), (6.5); then, for all κ ∈ IR, K ∈ T and n ∈ IN, the following inequality
holds:

|un+1 − κ| − |unK − κ| X h  
K n
m(K) + vK,L g(unK >κ, unL>κ) − g(unK ⊥κ, unL⊥κ) −
k (6.48) estschemak
L∈N (K)  i
n
vL,K g(unL >κ, unK >κ) − g(unL ⊥κ, unK ⊥κ) ≤ 0.
estldei
Proof of lemma 6.6
estschema
From relation (6.7), we express un+1
K as a function of unK and unL , L ∈ N (K),
k X
un+1
K = unK + n
(vL,K g(unL , unK ) − vK,L
n
g(unK , unL )).
m(K)
L∈N (K)

The right hand side is nondecreasing with respect to unL , Lestcfl

∈ N (K). It is also nondecreasing with respect
to unK , thanks to the Courant-Friedrichs-Levy condition (6.6), and the Lipschitz continuity of g.
Therefore, for all κ ∈ IR, using divv = 0, we have:
k X h i
un+1 n
K >κ ≤ uK >κ +
n
vL,K g(unL >κ, unK >κ) − vK,L
n
g(unK >κ, unL >κ) (6.49) est2ent
m(K)
L∈N (K)

and
k X
un+1 n
K ⊥κ ≥ uK ⊥κ +
n
(vL,K g(unL ⊥κ, unK ⊥κ) − vK,L
n
g(unK ⊥κ, unL⊥κ)). (6.50) est3ent
m(K)
L∈N (K)
est2ent est3ent estschemak
The difference between (6.49) and (6.50) leads directly to (6.48). Note that using divv = 0 leads to
 165

|un+1 − κ| − |unK − κ|
m(K) K +
X h  k 
n
vK,L g(unK >κ, unL>κ) − f (unK >κ) − g(unK ⊥κ, unL ⊥κ) + f (unK ⊥κ) − (6.51) estschemakb
L∈N (K)  i
n
vL,K g(unL >κ, unK >κ) − f (unK >κ) − g(unL ⊥κ, unK ⊥κ) + f (unK ⊥κ) ≤ 0.

For the implicit scheme, one obtains the same kind of discrete entropy inequalities.
estdonneesestdonnees meshypmd
estldeii Lemma 6.7 Under assumption 6.1 page 145, let T be an admissible
meshypmd
mesh in the sense of Definition 6.1
estcondflu
page 148 and k > 0. Let g ∈ C(IR 2 , IR) satisfying assumption 6.2.
estschemai
estschemaz
Let {unK , n ∈ IN, K ∈ T } ⊂ [Um , Uesteu
M ] be the solution of (6.9),(6.5) (the existence and uniqueness of such
a solution is given by Proposition 6.1). Then, for all κ ∈ IR, K ∈ T and n ∈ IN, the following inequality
holds:

|un+1 − κ| − |unK − κ| X h  
m(K) K
+ n
vK,L g(un+1
K >κ, u n+1
L >κ) − g(u n+1
K ⊥κ, u n+1
L ⊥κ)
k (6.52) estschemaki
 L∈N (K) i
n
−vL,K g(un+1 n+1 n+1
L >κ, uK >κ) − g(uL ⊥κ, uK ⊥κ)
n+1
≤ 0.
estldeii
Proof of lemma 6.7
estschemai
Let κ ∈ IR, K ∈ T and n ∈ IN. Equation (6.9) may be written as
k X
un+1
K = unK − n
(vK,L g(un+1 n+1 n n+1 n+1
K , uL ) − vL,K g(uL , uK )).
m(K)
L∈N (K)

The right hand side of this last equation is nondecreasing with respect to unK and with respect to un+1
L
for all L ∈ N (K). Thus,
k X
un+1
K ≤ unK >κ − n
(vK,L g(un+1 n+1 n n+1 n+1
K , uL >κ) − vL,K g(uL >κ, uK )).
m(K)
L∈N (K)

k X
n n
Writing κ = κ − (vK,L g(κ, κ) − vL,K g(κ, κ)), one may remark that
m(K)
L∈N (K)
k X
κ ≤ unK >κ − n
(vK,L g(κ, un+1 n n+1
L >κ) − vL,K g(uL >κ, κ)).
m(K)
L∈N (K)

Therefore, since un+1

K >κ = un+1
K or κ,

k X
un+1 n
K >κ ≤ uK >κ −
n
(vK,L g(un+1 n+1 n n+1 n+1
K >κ, uL >κ) − vL,K g(uL >κ, uK >κ)). (6.53) impt
m(K)
L∈N (K)

A similar argument yields

k X
un+1 n
K ⊥κ ≥ uK ⊥κ −
n
(vK,L g(un+1 n+1 n n+1 n+1
K ⊥κ, uL ⊥κ) − vL,K g(uL ⊥κ, uK ⊥κ)). (6.54) impb
m(K)
L∈N (K)
impb impt estschemaki
Hence, substracting (6.54) to (6.53) gives (6.52).
 166

6.5.2 Continuous entropy estimates for the approximate solution

For Ω = IR d or IR d × IR + , we denote by M(Ω) the set of positive measures on Ω, that is of σ-additive
R
applications from the Borel σ-algebra of Ω in IR + . If µ ∈ M(Ω) and ψ ∈ Cc (Ω), one sets hµ, ψi = ψdµ.
The following theorems investigate the entropy inequalities which are satisfied by the approximate so-
estentest
lutions uT ,k in theestentesti
case of the time explicit scheme (Theorem 6.1) and in the case of the time implicit
scheme (Theorem 6.2).
estdonnees meshypmd
estentest Theorem
meshypmd
6.1 Under assumption 6.1, let T be an admissible estcondflu
mesh in the senseestcfl
of Definition 6.1 page
148 and k > 0. Let g ∈estuappC(IR 2estschema
, IR) satisfy assumption
estschemaz
6.2 and assume that ( 6.6) holds.
Let uT ,k be given by (6.8), (6.7), (6.5); then there exist µT ,k ∈ M(IR d × IR + ) and µT ∈ M(IR d ) such
that
 Z Z 

 |uT ,k (x, t) − κ|ϕt (x, t)+



 IR + IR d 



 (f (u (x, t)>κ) − f (u (x, t)⊥κ))v(x, t) · ∇ϕ(x, t) dxdt +

 Z
T ,k T ,k


|u0 (x) − κ|ϕ(x, 0)dx ≥ (6.55) kruzkovap

 Zd
IR   Z



 − |ϕt (x, t)| + |∇ϕ(x, t)| dµT ,k (x, t) − ϕ(x, 0)dµT (x),

 IR d ×IR + IR d






∀κ ∈ IR, ∀ϕ ∈ Cc∞ (IR d × IR + , IR + ).
The measures µT ,k and µT verify the following properties:

1. For all R > 0 and T > 0, there exists C depending only on v, g, u0 , α, ξ, R and T such that, for
h < R and k < T ,
√
µT ,k (B(0, R) × [0, T ]) ≤ C h. (6.56) estmapp

2. The measure µT is the measure of density |u0 (·) − uT ,0 (·)| with respect to the Lebesgue measure,
where uT ,0 is defined by uT ,0 (x) = u0K for a.e. x ∈ K, for all K ∈ T .
If u0 ∈ BV (IR d ), then there exists D, only depending on u0 and α, such that

µT (IR d ) ≤ Dh. (6.57) estmapz

estentropiex Remark 6.5

estpbcont
estcondini kruzkovap
1. Let u be the weak entropy solution to (6.1)-(6.2). Then (6.55) is satisfied with u instead of uT ,k
and µT ,k = 0 and µT = 0.

2. Let BVloc (IR d ) be the set of v ∈ L1loc (IR d ) such that the restriction of v to Ω belongs to BV (Ω) for
all open bounded subset Ω of IR d .
following proof gives that if u0 ∈ BVloc (IR d ) instead of BV (IR d ) (in the
An easy adaptation of theestentest
second item of Theorem 6.1) then, for all R > 0, there exists D, only depending on u0 , α and R,
such that µT (B(0, R)) ≤ Dh.
estentest
Proof of Theorem 6.1
Let ϕ ∈ Cc∞ (IR d × IR + , IR + ) and κ ∈ IR.
estschemakb R (n+1)k R
Multiplying (6.51) by kϕnK = (1/m(K)) nk K
ϕ(x, t)dxdt and summing the result for all K ∈ T and
n ∈ IN yields

T1 + T2 ≤ 0,
 167

with
X X |un+1 − κ| − |un − κ| Z (n+1)k Z
K K
T1 = ϕ(x, t)dxdt, (6.58) est1kap
k nk K
n∈IN K∈T
and
X X h
T2 = k
n∈IN (K,L)∈En
 
n
vK,L ϕnK g(unK >κ, unL>κ) − f (unK >κ) − g(unK ⊥κ, unL ⊥κ) + f (unK ⊥κ)
 
n
−vK,L ϕnL g(unK >κ, unL>κ) − f (unL >κ) − g(unK ⊥κ, unL ⊥κ) + f (unL ⊥κ) (6.59) est2kap
 
n
−vL,K ϕnK g(unL >κ, unK >κ) − f (unK >κ) − g(unL ⊥κ, unK ⊥κ) + f (unK ⊥κ)
 i
n
+vL,K ϕnL g(unL >κ, unK >κ) − f (unL >κ) − g(unL ⊥κ, unK ⊥κ) + f (unL ⊥κ) ,

where En = {(K, L) ∈ T 2 , unK > unL }.

One has to prove
Z   Z
T10 + T20 ≤ |ϕt (x, t)| + |∇ϕ(x, t)| dµT ,k (x, t) + ϕ(x, 0)dµT (x), (6.60) estentcont
IR d ×IR + IR d

for some convenient measures µT ,k and µT , and T10 , T20 defined as follows
Z Z Z
T10 = − |uT ,k (x, t) − κ|ϕt (x, t)dxdt − |u0 (x) − κ|ϕ(x, 0)dx,
IR + IR d IR d
Z Z  
T20 = − (f (uT ,k (x, t)>κ) − f (uT ,k (x, t)⊥κ))v(x, t) · ∇ϕ(x, t) dxdt. (6.61) est7kap
IR + IR d
estentcont
In order to prove (6.60), one compares T1 and T10 (this will give µT , and a part of µT ,k ) and one compares
T2 and T20 (this will give another part of µT ,k ).
estbvft estbvfx
Inequality (6.17) (in the comparison estmapp
of T1 and T10 ) and Inequality (6.16) (in the comparison of T2 and
T20 ) will be used in order to obtain (6.56).
Comparison of T1 and T10
Using the definition of uT ,k and introducing the function uT ,0 (defined by uT ,0 (x) = u0K , for a.e. x ∈ K,
for all K ∈ T ) yields
X X |un+1 − κ| − |un − κ| Z (n+1)k Z
K K
T10 = ϕ(x, (n + 1)k)dxdt +
k nk K
Z n∈IN K∈T

(|uT ,0 (x) − κ| − |u0 (x) − κ|)ϕ(x, 0)dx.

IR d
The function | · −κ| is Lipschitz continuous with a Lipschitz constant equal to 1, we then obtain
X X |un+1 − un | Z (n+1)k Z
K K
|T1 − T10 | ≤ |ϕ(x, (n + 1)k) − ϕ(x, t)|dxdt +
k nk K
Z
n∈I
N K∈T

|u0 (x) − uT ,0 (x)|ϕ(x, 0)dx,

IR d
which leads to
XX Z (n+1)k Z
|T1 − T10 | ≤ |un+1
K − unK | |ϕt (x, t)|dxdt +
nk K (6.62)
Z
n∈IN K∈T est5kapb
|u0 (x) − uT ,0 (x)|ϕ(x, 0)dx.
IR d
 168

est5kapb
Inequality (6.62) gives
Z Z
|T1 − T10 | ≤ |ϕt (x, t)|dνT ,k (x, t) + ϕ(x, 0)dµT (x), (6.63) estmajtime
IR d ×IR + IR d

where the measures µT ∈ M(IR d ) and νT ,k ∈ M(IR d × IR + ) are defined, by their action on Cc (IR d ) and
Cc (IR d × IR + ), as follows
Z
hµT , ψi = |u0 (x) − uT ,0 (x)|ψ(x)dx, ∀ψ ∈ Cc (IR d ),
IR d

X X Z (n+1)k Z
hνT ,k , ψi = |un+1
K − unK | ψ(x, t)dxdt,
n∈IN K∈T nk K
∀ψ ∈ Cc (IR d × IR + ).
The measures µT and νT ,k are absolutely continuous with P respect
P to the Lebesgue measure. Indeed,
one has dµT (x) = |u0 (x) − uT ,0 (x)|dx and dνT ,k (x, t) = ( n∈IN K∈T |un+1 K − unK |1K×[nk,(n+1)k) )dxdt
(where 1Ω denotes the characteristic function of Ω for any Borel subset Ω of IR d+1 ).
estmapz
If u0 ∈ BV (IR d ), the measure µT verifies (6.57)
divbvd
with some D only depending on |u0 |BV (IR d ) and α (this
is classical result which is given in
estmapp
Lemma 6.8 below for the sake of completeness).
estbvft estcfl
The measure νT ,k satisfies (6.56), with νT ,k instead of µT ,k , thanks to (6.17) and condition (6.6). Indeed,
for R > 0 and T > 0,
Z T Z X X
νT ,k (B(0, R) × [0, T ]) = |un+1
K − unK |1K×[nk,(n+1)k) dxdt,
0 B(0,R) n∈IN K∈T

which yields, with T2R = {K ∈ T , K ⊂ B(0, 2R)} and NT,k k < T ≤ (NT,k + 1)k, h < R and k < T ,
NT,k
X X kC1
νT ,k (B(0, R) × [0, T ]) ≤ k m(K)|un+1
K − unK | ≤ √ ,
n=0 K∈T2R
h
estwbvest
where C is given by lemma 6.2 and only depends on v, g, u0 , α, ξ, R, T . Finally, since the condition
estcfl 1
(6.6) gives k ≤ C2 h, where C2 only depends on v, g, u0 , α, ξ, the last inequality yields, for h < R and
k < T,
√
νT ,k (B(0, R) × [0, T ]) ≤ C3 h, (6.64) boundnu
with C3 = C1 C2 .
Comparison of T2 and T20 est7kap
Using divv = 0, and gathering (6.61) by interfaces, we get
X X h  
T20 = − (f (unK >κ) − f (unK ⊥κ)) − (f (unL >κ) − f (unL ⊥κ))
n∈IN (K,L)∈En
Z Z (n+1)k   i (6.65) est8kap
v(x, t) · nK,L ϕ(x, t) dγ(x)dt .
K|L nk

Define, for all K ∈ T , all L ∈ N (K) and all n ∈ IN,

Z (n+1)k Z
1
(vϕ)n,+
K,L = (v(x, t) · nK,L )+ ϕ(x, t)dγ(x)dt
k nk K|L

and
Z (n+1)k Z
1
(vϕ)n,−
K,L = (v(x, t) · nK,L )− ϕ(x, t)dγ(x)dt.
k nk K|L
 169

est8kap
Note that (vϕ)n,+ n,−
K,L = (vϕ)L,K . Then, (6.65) gives
X X h
T20 = k
n∈IN (K,L)∈En
 
n,+
(vϕ)K,L g(unK >κ, unL >κ) − f (unK >κ) − g(unK ⊥κ, unL⊥κ) + f (unK ⊥κ)
 
−(vϕ)n,− n n n n n n (6.66) est8bkap
L,K g(uK >κ, uL >κ) − f (uL >κ) − g(uK ⊥κ, uL ⊥κ) + f (uL ⊥κ)
 
−(vϕ)n,− g(u n
>κ, u n
>κ) − f (u n
>κ) − g(u n
⊥κ, u n
⊥κ) + f (u n
⊥κ)
K,L
 L K K L K K
i
+(vϕ)n,+
L,K g(u n
L >κ, u n
K >κ) − f (u n
L >κ) − g(u n
L ⊥κ, u n
K ⊥κ) + f (u n
L ⊥κ) .

Let us introduce some terms related to the difference between ϕ on K ∈ T and K|L ∈ E,
n,+
rK,L n
= |vK,L ϕnK − (vϕ)n,+
K,L |

and
n,−
rK,L n
= |vL,K ϕnK − (vϕ)n,−
K,L |.
est2kap est8bkap
Then, from (6.59) and (6.66),
X X h
|T2 − T20 | ≤ k
n∈IN (K,L)∈En
 
n,+
rK,L g(unK >κ, unL >κ) − f (unK >κ) + g(unK ⊥κ, unL ⊥κ) − f (unK ⊥κ) +
 
n,−
rL,K g(unK >κ, unL >κ) − f (unL >κ) + g(unK ⊥κ, unL ⊥κ) − f (unL ⊥κ) + (6.67) est9kap
 
n,− n n n n n n
rK,L f (uK >κ) − g(uL >κ, uK >κ) + f (uK ⊥κ) − g(uL ⊥κ, uK ⊥κ) +
 i
n,+
rL,K f (unL >κ) − g(unL >κ, unK >κ) + f (unL ⊥κ) − g(unL ⊥κ, unK ⊥κ) .

For all (K, L) ∈ En , the following inequality holds:

0 ≤ g(unK >κ, unL >κ) − f (unK >κ) ≤ max (g(q, p) − f (q)),

un
L
≤p≤q≤un
K

more precisely, one has g(unK >κ, unL>κ) − f (unK >κ) = 0, if κ ≥ unK , and one has g(unK >κ, unL >κ) −
f (unK >κ) = g(q, p) − f (q) with p = κ and q = unK if κ ∈ [unL , unK ], and with p = unL and q = unK if κ ≤ unL .
In the same way, we can assert that

0 ≤ g(unK ⊥κ, unL ⊥κ) − f (unK ⊥κ) ≤ max (g(q, p) − f (q)).

un
L
≤p≤q≤un
K
est9kap
The same analysis can be applied to the six other terms of (6.67).
n,±
To conclude the estimate on |T2 − T20 |, there remains to estimate the two quantities rK,L . This will be
n,+
done with convenient measures applied to |∇ϕ| and |ϕt |. To estimate rK,L , for instance, one remarks
that
Z (n+1)k Z (n+1)k Z Z
n,+ 1
rK,L ≤ 2
|ϕ(x, t) − ϕ(y, s)|(v(y, s) · nK,L )+ dγ(y)dxdtds.
k m(K) nk nk K K|L

Hence
Z (n+1)k Z (n+1)k Z Z Z 1
n,+ 1
rK,L ≤ |∇ϕ(x + θ(y − x), t + θ(s − t)) · (y − x)+
k 2 m(K) nk nk K K|L 0
ϕt (x + θ(y − x), t + θ(s − t))(s − t)|(v(y, s) · nK,L )+ dθdγ(y)dxdtds
which yields
 170

Z (n+1)k Z (n+1)k Z Z Z 1
n,+ 1
rK,L ≤ 2 h|∇ϕ(x + θ(y − x), t + θ(s − t))|+
k m(K) nk nk  K K|L 0
k|ϕt (x + θ(y − x), t + θ(s − t))| (v(y, s) · nK,L )+ dθdγ(y)dxdtds.

This leads to the definition of a measure µn,+ d

K,L , given by its action on Cc (IR × IR + ):
Z (n+1)k Z (n+1)k Z Z Z 1 
2
hµn,+
K,L , ψi = 2
(h + k)ψ(x + θ(y − x), t + θ(s − t))
k m(K) nk nk K K|L 0
(v(y, s) · nK,L )+ dθdγ(y)dxdtds, ∀ψ ∈ Cc (IR d × IR + ),
n,+
in order to have 2rK,L ≤ hµn,+K,L , |∇ϕ| + |ϕt |i.
We define in the same way µn,− K,L , changing (v(y, s) · nK,L )
+
in (v(y, s) · nK,L )− . We finally define the
measure ν̃T ,k by
X X h 
hν̃T ,k , ψi = k n
max n
(g(q, p) − f (q)) hµn,+
K,L , ψi
uL ≤p≤q≤uK
n∈IN (K,L)∈En  
+ (g(q, p) − f (p)) hµn,−
max L,K , ψi
un ≤p≤q≤un
 L K  (6.68) est10kap
+ n max n (f (q) − g(p, q)) hµn,− K,L , ψi
 uL ≤p≤q≤uK  i
+ n max n (f (p) − g(p, q)) hµn,+L,K , ψi .
uL ≤p≤q≤uK

n,± est9kap est10kap

Since 2rK,L ≤ hµn,±
K,L , |∇ϕ| + |ϕt |i, estmajtime
(6.67) and (6.68) leads to |T2 − T20 | ≤ hν̃T ,k , |∇ϕ| + |ϕt |i. Therefore,
setting µT ,k = νT ,k + ν̃T ,k , using (6.63) and T1 + T2 ≤ 0,
Z   Z
T10 + T20 ≤ |ϕt (x, t)| + |∇ϕ(x, t)| dµT ,k (x, t) + ϕ(x, 0)dµT (x),
IR d ×IR + IR d
estentcont kruzkovap
which is (6.60) and yields (6.55).
estmapp
There remains to prove (6.56).
For all K ∈ T , let xK be an arbitrary point of K. For all K ∈ T , all K ∈ N (K) and all n ∈ IN, the
supports of the measures µn,±
K,L are included in the closed set B̄(xK , h) ∩ [nk, (n + 1)k]. Furthermore,

µn,+ d n n,− d n
K,L (IR × IR + ) ≤ 2vK,L (h + k) and µK,L (IR × IR + ) ≤ 2vL,K (h + k).

Then, for all R > 0 and T > 0, the definition of µT ,k (i.e. µT ,k = νT ,k + ν̃T ,k )) leads to
√
µT ,k (B(0, R) × [0, T ]) ≤ C3 h
NT,k
X X h  
n
+2(h + k) k vK,L n
max n
(g(q, p) − f (q)) + n
max n
(g(q, p) − f (p))
n
uL ≤p≤q≤uK uL ≤p≤q≤uK
n=0 (K,L)∈E2R
 i
n
+vL,K max (f (q) − g(p, q)) + max (f (p) − g(p, q)) ,
un
L
n
≤p≤q≤uK un
L
≤p≤q≤uK n

√ boundnu
for h < R and k < T , where
estwbvest
C 3 h is the bound of νT ,k (B(0, R) × [0, T ]) given in (6.64). Therefore,
thanks to Lemma 6.2,
√ C4 √
µT ,k (B(0, R) × [0, T ]) ≤ C3 h + (1 + C2 )h √ = C h,
h
estentest
where C only depends on v, g, u0 , α, ξ, R and T . The proof of Theorem 6.1 is complete.

The following theorem investigates the case of the implicit scheme.

 171

estdonnees meshypmd
estentesti Theorem 6.2 Under Assumption 6.1, let T estcondflu be an admissible mesh in the sense of Definition 6.1 and
k > 0. Let g ∈ C(IR 2 , IR) satisfy Assumption 6.2.
n n
Let {uestschemaz
K , n ∈ IN, K ∈ T }, such that uK ∈ [Um , UM ] for all K ∈ T and n ∈ esteu
estschemai
IN, be the solution of
(6.9),(6.5) (existence and uniqueness of such a solution are given by Proposition 6.1). Let uT ,k be given
estuapp
by (6.8). Assume that v does not depend on t and that u0 ∈ BV (IR d ).
Then, there exist µT ,k ∈ M(IR d × IR + ) and µT ∈ M(IR d ) such that
 Z Z 

 |uT ,k (x, t) − κ|ϕt (x, t)+



 IR + IR d 



 (f (uT ,k (x, t)>κ) − f (uT ,k (x, t)⊥κ))v(x, t) · ∇ϕ(x, t) dxdt +

 Z


|u0 (x) − κ|ϕ(x, 0)dx ≥ (6.69) kruzkovapi

 Zd
IR   Z



 − |ϕt (x, t)| + |∇ϕ(x, t)| dµT ,k (x, t) − ϕ(x, 0)dµT (x),

 IR d ×IR + IR d






∀κ ∈ IR, ∀ϕ ∈ Cc∞ (IR d × IR + , IR + ).
The measures µT ,k and µT verify the following properties:

1. For all R > 0 and T > 0, there exists C, only depending on v, g, u0 , α, R, T such that, for h < R
and k < T , √
µT ,k (B(0, R) × [0, T ]) ≤ C(k + h). (6.70) estmappi

2. The measure µT is the measure of density |u0 (·) − uT ,0 (·)| with respect to the Lebesgue measure and
there exists D, only depending on u0 and α, such that

µT (IR d ) ≤ Dh. (6.71) estmapzi

estentesti
Proof of Theorem 6.2
estentest
Similarly to the proof of Theorem 6.1, we introduce a test function ϕ ∈ Cc∞ (IR d × IR + , IR + ) and a real
estschemaki R (n+1)k R
number κ ∈ IR. We multiply (6.52) by (1/m(K)) nk K
ϕ(x, t)dxdt, and sum for all K ∈ T
the resultest2kap
est1kap
and n ∈ IN. We then define T1 and T2 such that T1 + T2 ≤ 0 estmajtime
using equations (6.58) and (6.59) in which
we replace unK by un+1
K and unL by un+1
L . Therefore we get (6.63), where the measure νT ,k is such that
for all T > 0, there exists C1 only depending on v, g, u0 α and T , such that, for k < T ,

νT ,k (IR d × [0, T ]) ≤ C1 k,
estlbvt estlbvt
using Lemma 6.5 page 162, which is available if v does not depend on t (and for which one needs that
u0 ∈ BV (IR d )).
estentest
The treatment of T2 is very similar to that of Theorem 6.1, replacing unK by un+1
K and unL by un+1
L . But,
n,±
since v does not depend on t, the bounds on rK,L are simpler. Indeed,
Z (n+1)k Z Z
n,± 1
rK,L ≤ |ϕ(x, t) − ϕ(y, t)|(v(y) · nK,L )± dγ(y)dxdt.
km(K) nk K K|L
n,±
Now 2rK,L ≤ hµn,±
K,L , |∇ϕ|i where µn,±
K,L is defined by
Z (n+1)k Z Z Z 1 
2
hµn,±
K,L , ψi = h ψ(x + θ(y − x), t)
km(K) nk K K|L 0
(v(y) · nK,L )± dθdγ(y)dxdt, ∀ψ ∈ Cc (IR d × IR + ).
est10kap
With this definition of µn,± n n+1
K,L , the bound on ν̃T ,k (defined by (6.68), replacing uK by uK and unL by
estwbvesti
estwbvesti
un+1
L ) becomes, thanks to Lemma 6.3 page 159,
 172

√
ν̃T ,k (B(0, R) × [0, T ]) ≤ C2 h,
for h < R and k < T , where C2 estentest
only depends on v, g, u0 , α, R and T .
Hence, defining (as in Theorem 6.1) µT ,k = νT ,k + ν̃T ,k , for all R > 0 and all T > 0 there exists C, only
depending on v, g, u0 , α, R, T such that, for h < R and k < T ,
√
µT ,k (B(0, R) × [0, T ]) ≤ C(k + h),
estmappi estentesti
which is (6.70) and concludes the proof of Theorem 6.2.

estlbvt
Remark 6.6 In the case where v depends on t, Lemma
estentest
6.5 cannot be
estentesti
used. However, it is easy to show
estmappi
(the proof follows that of Theorem 6.1) that Theorem 6.2 is true if (6.70) is replaced by
k √
µT ,k (B(0, R) × [0, T ]) ≤ C( √ + h), (6.72) estmappit
h
estexti estmappit
which leads to the result given in Remark 6.12. The estimate (6.72) may be obtained without assuming
that u0 ∈ BV (IR d ) (it is sufficient that u0 ∈ L∞ (IR d )).

For the sake of completeness we now prove a lemma which gives the bound on the measure µ T in the
two last theorems.
meshypmd meshypmd
divbvd T be an admissible
Lemma 6.8 LetdefBVRp defBVRp
mesh in the sense of Definition 6.1 page 148 and let u ∈ BV (IR d )
(see Definition 5.38 page 138). For K ∈ T , let uK be the mean value of u over K. Define uT by
uT (x) = uK for a.e. x ∈ K, for all K ∈ T . Then,
C
ku − uT kL1 (IR d ) ≤
h|u|BV (IR d ) , (6.73) bvzdd
α2
where C only depends on the space dimension (d = 1, 2 or 3).
divbvd
Proof of Lemma 6.8
divbv divbv
The proof is very similar to that of Lemma 6.4 and we will mainly refer to the proof of Lemma 6.4.
bvzdd bvzdd
First, remark that if (6.73) holds for all u ∈ BV (IR d ) ∩ C 1 (IR d , IR) then (divbv
6.73) holds for all u ∈ BV (IR d ).
Indeed, let u ∈ BV (IR d ), it is proven in Step 1 of the proof of Lemma 6.4 that there exists a sequence
(un )n∈IN ⊂ C ∞ (IR d , IR) such that un → u in L1loc (IR d ), as n → ∞, and kun kBV (IR d ) ≤ kukBV (IR d ) for all
d bvzdd
n ∈ IN. One may also assume, up to a subsequence,
bvzdd
that un → u a.e. on IRbvzdd . Then, if (6.73) is true with
un instead of u, passing to the limit in (6.73) (for un ) as n → ∞ leads to (6.73) (for u) thanks to Fatou’s
lemma.
bvzdd divbvd
Let us now prove (6.73) if u ∈ BV (IR d ) ∩ C 1 (IR d , IR) (this concludes the proof of Lemma 6.8). Since
u ∈ C 1 (IR d , IR),
|u|BV (IR d ) = k(|∇u|)kL1 (IR d ) ;
bvzdd
hence we shall prove (6.73) with k(|∇u|)kL1 (IR d ) instead of |u|BV (IR d ) .
For K ∈ T ,
Z Z Z
1
|u(x) − uK |dx ≤ ( |u(x) − u(y)|dx)dy.
K m(K) K K
divbv
Then, following the lines of Step 2 of Lemma 6.4,
Z Z Z 1 Z
1
|u(x) − uK |dx ≤ h ( |∇u(y + tz)|dydt)dz. (6.74) oufc
K m(K) B(0,h) 0 K

For all K ∈ T , let xK be an arbitrary point of K.

 173

oufc
Then, changing the variable y in ξ = y + tz (for all fixed z ∈ K and t ∈ (0, 1)) in (6.74),
Z Z Z 1 Z
1
|u(x) − uK |dx ≤ h ( |∇u(ξ)|dξdt)dz,
K m(K) B(0,h) 0 B(xK ,2h)
estmaillage
estmaillage
which yields, since T is an admissible mesh in the sense of Definition 6.4 page 148,
Z Z
1
|u(x) − uK |dx ≤ m(B(0, h))h |∇u(ξ)|dξ.
K αhd B(xK ,2h)

Therefore there exists C1 , only depending on the space dimension, such that
Z Z
C1
|u(x) − uK |dx ≤ h |∇u(ξ)|dξ, ∀K ∈ T . (6.75) oufd
K α B(xK ,2h)
divbv
As in Lemma 6.4, for a fixed M ∈ T , the number of K ∈ T such that M ∩ B(xK , 2h) 6= ∅ is less or equal
d
to m(B(0, 4h))/(αh
oufd
) that is less or equal to C2 /α where C2 only depends on the space dimension.
Then, summing (6.75) over K ∈ T leads to
XZ C1 C2 X
Z
C1 C2
|u(x) − uK |dx ≤ 2
h |∇u(ξ)|dξ = hk(|∇u|)kL1 (IR d ) ,
K α M α2
K∈T M ∈T
bvzdd
that is (6.73) with C = C1 C2 .

6.6 Convergence of the scheme

eghcvsol
This section is devoted to the proof of the existence and uniqueness of the entropy weak solution and of
the convergence of the approximate solution towards the entropy weak solution as the mesh size and time
eghcvepsol
step tend to 0. This proof will be performed in two steps. We first prove in section 6.6.1 the convergence
estepsol
of the approximate solution towards an entropy process solution which is defined in Definition 6.2 below
(note that the convergence also yields the existence of an entropy process solution).
estpbcont
estcondini
estepsol Definition 6.2 A function µ is an entropy process solution to problem ( 6.1)-(6.2) if µ satisfies


 µ ∈ L∞ (IR d × IR ?+ × (0, 1)),

 Z Z +∞ Z 1  



 η(µ(x, t, α))ϕt (x, t) + Φ(µ(x, t, α))v(x, t) · ∇ϕ(x, t) dαdtdx
 IRZd 0 0
(6.76) ews3

 + η(u 0 (x))ϕ(x, 0)dx ≥ 0,

 IR d



 for any ϕ ∈ Cc1 (IR d × IR + , IR + ),

for any convex function η ∈ C 1 (IR, IR), and Φ ∈ C 1 (IR, IR) such that Φ0 = f 0 η 0 .
estpbcont
estcondini
ewsews Remark 6.7 From an entropy weak estpbcont solution u to problem (6.1)-(6.2), one may easily construct an
estcondini
entropy process solution to problem (6.1)-(6.2) by setting µ(x, t, α) = u(x, for a.e. (x, t, α) ∈ IR d ×
t) estcondini
estpbcont
?
IR + × (0, 1). Reciprocally, if µ is an entropy process solution to problem (6.1)-(6.2) such that there exists
u ∈ L∞ (IR d × IR ?+ ) such that µ(x,estpbcont = u(x, t), for a.e. (x, t, α) ∈ IR d × IR ?+ × (0, 1), then u is an
t, α) estcondini
entropy weak solution to problem (6.1)-(6.2).
ewsp ewsews
In section 6.6.2, we show the uniqueness of the entropy process solution, which, thanks to remark 6.7,
also yields eghbscewsol
the existence and uniqueness of the entropy weak solution. This allows us to state and prove,
in section 6.6.3, the convergence of the approximate solution towards the entropy weak solution.
We now give a useful characterization of an entropy process solution in terms of Krushkov’s entropies (as
for the entropy weak solution).
 174

estpbcont
estcondini
sanslabel Proposition 6.2 A function µ is an entropy process solution of problem ( 6.1)-(6.2) if and only if,


 µ ∈ L∞ (IR d × IR ?+ × (0, 1)),

 Z Z +∞ Z 1  


 |µ(x, t, α) − κ|ϕt (x, t) + Φ(µ(x, t, α), κ)v(x, t) · ∇ϕ(x, t) dαdtdx
IR d 0 0 Z (6.77) ews3k



 + |u 0 (x) − κ|ϕ(x, 0)dx ≥ 0,

 IR d

∀κ ∈ IR, ∀ϕ ∈ Cc1 (IR d × IR + , IR + ),
where we set Φ(a, b) = f (a>b) − f (a⊥b), for all a, b ∈ IR.
sanslabel
Proof of Proposition 6.2
The proof of this result is similar
ews3k ews3
to the case of classical entropy weak solutions. The characterization
ews3
(6.77) can be obtained from
ews3k
( 6.76), by using regularizations of the function | · −κ|. Conversely, (6.76)
may be obtained from (6.77) by approximating any convex function η ∈ C 1 (IR, IR) by functions of the
Xn
(n) (n) (n)
form: ηn (·) = αi | · −κi |, with αi ≥ 0.
i=1

6.6.1 Convergence towards an entropy process solution

eghcvepsol
Let α > 0 and 0 < ξ < 1. Let (Tm , km )m∈IN be a sequence of admissible meshes in the sense of Definition
meshypmd meshypmd
6.1 page 148 and time steps. Note that Tm is admissible with α independent of m. Assume that km
estcfl
satisfies (6.6), for T estinftyest
= Tm and k = km , and that size(Tm ) → 0 as m → ∞.
estinftyest
By Lemma 6.1 page 152, the sequence (uTm ,km )m∈IN of approximate solutions defined by the finite volume
estschemaz estschema estschema
scheme (6.5) and (6.7) page 149, with T = Tm and k = km , is bounded in L∞ (IR d × IR ?+ ); therefore,
there exists µ ∈ L∞ (IR d × IR ?+ × (0, 1))nlwsdef
such that uTm ,km converges, nlwsprop
nlwsdef
as m tends to ∞, towards µ in the
nlwsprop
nonlinear weak-? sense (see Definition 6.3 page 190 and Proposition 6.4 page 191), that is:
Z Z Z Z Z 1
lim θ(uTm ,km (x, t))ϕ(x, t)dtdx = θ(µ(x, t, α))ϕ(x, t)dαdtdx,
m→∞ IR d IR + IR d IR + 0 (6.78) eghbconvy
∀ϕ ∈ L (IR ×1
IR ?+ ), ∀θ ∈ C(IR, IR).
eghbconvy
Taking for θ, in (6.78), the Krushkov entropies (namely θ = | · −κ|, for all κ ∈ IR) and the associated estentest
functions defining the entropy fluxes (namely θ = f (·, κ) = f (·>κ) − f (·⊥κ)) and using Theorem 6.1
kruzkovap
(that is passing to the limit, as m → ∞, in (6.55) written with uT ,k = uTm ,km ) yields that µ is an entropy
process solution. Hence the following result holds:
estdonnees
eghbcepsol Proposition 6.3 Under assumptions 6.1, let meshypmd
α > 0 and 0 < ξ < 1. Let (Tm , km )m∈IN be a sequence of
meshypmd
admissible meshes in the sense of Definition 6.1 page 148
estcfl
and time steps. Note that Tm is admissible
with α independent of m. Assume that km satisfy (6.6), for T = Tm and k = km , and that size(Tm ) → 0
as m → ∞.
Then there exists a subsequence, still denoted by (Tm , km )m∈IN , and a function µ ∈ L∞ (IR d × IR ?+ × (0, 1))
such that
estschema
estschemaz estuapp
with T = Tm and k = km , that is uTm ,km ,
1. the approximate solution defined by (6.7), (6.5) and (6.8) eghbconvy
converges towards µ in the nonlinear weak-? sense, i.e. (6.78) holds,
estpbcont
estcondini
2. µ is an entropy process solution of (6.1)-(6.2).
estcfl
Remark 6.8 The same theorem can be proved for the implicit scheme without condition (6.6) (and thus
without ξ).
eghbcepsol
eghbexepsol Remark 6.9estpbcont
Noteestcondini
that a consequence of Proposition 6.3 is the existence of an entropy process solution
to Problem (6.1)-(6.2).
 175

6.6.2 Uniqueness of the entropy process solution

ewsp
In order to show the uniqueness of an entropy process solution, we shall use the characterization of an
sanslabel
entropy process solution given in proposition 6.2.
estdonnees estpbcont
estcondini
ewsunicite Theorem 6.3estepsol
Under Assumption
estepsol
6.1, the entropy process solution µ of problem (6.1),(6.2), as defined
in Definition 6.2 page 173, is unique. Moreover, there exists a function u ∈ L∞ (IR d × IR ?+ ) such that
eghbcepsol ewsews
u(x, t) = µ(x, t, α), for a.e. (x, t, α) ∈ IR d × IR ?+ × (0, 1).estpbcont
(Hence, with Proposition 6.3 and Remark 6.7,
estcondini
there exists a unique entropy weak solution to Problem (6.1)-(6.2).)
ewsunicite
Proof of Theorem 6.3
estpbcont
estcondini
Let µ and ν be two entropy process solutions to Problem (6.1)-(6.2). Then, one has µ ∈ L∞ (IR d × IR ?+ ×
(0, 1)), ν ∈ L∞ (IR d × IR ?+ × (0, 1)) and
Z Z +∞ Z 1
|µ(x, t, α) − κ|ϕt (x, t)
IR d 0 0 
+(f (µ(x, t, α)>κ) − f (µ(x, t, α)⊥κ))v(x, t) · ∇ϕ(x, t) dαdtdx (6.79) ewsmu
Z
+ |u0 (x) − κ|ϕ(x, 0)dx ≥ 0, ∀κ ∈ IR, ∀ϕ ∈ Cc1 (IR d × IR + , IR + ),
IR d
Z Z +∞ Z 1
|ν(y, s, β) − κ|ϕs (y, s)
IR d 0 0 
+(f (ν(y, s, β)>κ) − f (ν(y, s, β)⊥κ))v(y, s) · ∇ϕ(y, s) dβdsdy (6.80) ewsnu
Z
+ |u0 (y) − κ|ϕ(y, 0)dy ≥ 0, ∀κ ∈ IR, ∀ϕ ∈ Cc1 (IR d × IR + , IR + ).
IR d
ewsunicite
The proof of Theorem 6.3 contains 2 steps. In Step 1, it is proven that
Z 1 Z 1 Z Z h
|µ(x, t, α) − ν(x, t, β)|ψt (x, t)
0 0 IR + IR d  i (6.81) step1
+ f (µ(x, t, α)>ν(x, t, β)) − f (µ(x, t, α)⊥ν(x, t, β)) v(x, t) · ∇ψ(x, t) dxdtdαdβ ≥ 0,
∀ψ ∈ Cc1 (IR d × IR + , IR + ).
In Step 2, it is proven that µ(x, t, α) = ν(x, t, β) for a.e. (x, t, α, β) ∈ IR d × IR ?+ × (0, 1) × (0, 1). We
then deduce that there exists u ∈ L∞ (IR d × IR ?+ ) such that µ(x, t, α) = u(x, t) for a.e. (x, t, α) ∈
estpbcont
estcondini
IR d × IR ?+ × (0, 1) (therefore u is necessarily the unique entropy weak solution to (6.1)-(6.2)).
step1
Step 1 (proof of relation (6.81))
step1
In order to prove relation (6.81), a sequence of mollifiers in IR and IR d is introduced .
Let ρ ∈ Cc∞ (IR d , IR + ) and ρ̄ ∈ Cc∞ (IR, IR + ) be such that

{x ∈ IR d ; ρ(x) 6= 0} ⊂ {x ∈ IR d ; |x| ≤ 1},

{x ∈ IR; ρ̄(x) 6= 0} ⊂ [−1, 0] (6.82) rhodec

and
Z Z
ρ(x)dx = 1, ρ̄(x)dx = 1.
IR d IR

For n ∈ IN ? , define ρn = nd ρ(nx) for all x ∈ IR d and ρ̄n = nρ̄(nx) for all x ∈ewsmu IR.
Let ψ ∈ Cc1 (IR d ×IR + , IR + ). For (y, s, β) ∈ IR d ×IR + ×(0, 1), let us take, in (6.79), ϕ(x, t) = ψ(x, t)ρn (x−
y)ρ̄n (t − s) and κ = ν(y, s, β). Then, integrating the result over IR d × IR + × (0, 1) leads to
 176

A1 + A2 + A3 + A4 + A5 ≥ 0, (6.83) ewscin
where
Z 1 Z 1 Z ∞ Z Z ∞ Z h
A1 = |µ(x, t, α) − ν(y, s, β)|
0 0 0 IR d i0 IR d
ψt (x, t)ρn (x − y)ρ̄n (t − s) dxdtdydsdαdβ,
Z 1 Z 1 Z ∞ Z Z ∞ Z h
A2 = |µ(x, t, α) − ν(y, s, β)|
0 0 0 IR d 0 IR d i
ψ(x, t)ρn (x − y)ρ̄0n (t − s) dxdtdydsdαdβ,
Z 1 Z 1 Z ∞ Z Z ∞ Z h 
A3 = f (µ(x, t, α)>ν(y, s, β)) − f (µ(x, t, α)⊥ν(y, s, β))
0 0 0 IR d 0 IR d i
v(x, t) · ∇ψ(x, t)ρn (x − y)ρ̄n (t − s) dxdtdydsdαdβ,
Z 1 Z 1 Z ∞ Z Z ∞ Z h 
A4 = f (µ(x, t, α)>ν(y, s, β)) − f (µ(x, t, α)⊥ν(y, s, β))
0 0 0 IR d 0 IR d i
v(x, t) · ∇ρn (x − y)ψ(x, t)ρ̄n (t − s) dxdtdydsdαdβ
and
Z 1 Z Z ∞ Z
A5 = |u0 (x) − ν(y, s, β)|ψ(x, 0)ρn (x − y)ρ̄n (−s)dydsdxdβ. a5
0 IR d 0 IR d
ewscin ewsnu step1
Passing to the limit in (6.83) as n → ∞ (using (6.80) for the study of A2 + A4 and A5 ) will give (6.81).
rhodec
Let us first consider A1 and A3 . Note that, using (6.82),
Z Z ∞
ρn (x − y)ρ̄n (t − s)dsdy = 1, ∀x ∈ IR d , ∀t ∈ IR + .
IR d 0
Then,
Z 1Z 1Z Z h i
|A1 − |µ(x, t, α) − ν(x, t, β)|ψt (x, t) dxdtdαdβ|
d
Z 1 Z0∞ Z0 IR
Z +∞ IZR h i
≤ |ν(x, t, β) − ν(y, s, β)||ψt (x, t)|ρn (x − y)ρ̄n (t − s) dxdtdydsdβ
0 0 IR d 0 IR d
≤ kψt kL∞(IR d ×IR ?+ ) ε(n, S),

with S = {(x, t) ∈ IR d × IR + ; ψ(x, t) 6= 0} and

1 1
ε(n, S) = sup{kν − ν(· + η, · + τ, ·)kL1 (S×(0,1)) ; |η| ≤ , 0 ≤ τ ≤ }.
n n
Since ν ∈ L1loc (IR d × IR + × [0, 1]) and S is bounded, one has ε(n, S) → 0 as n → ∞. Hence,
Z 1 Z 1 Z Z h i
A1 → |µ(x, t, α) − ν(x, t, β)|ψt (x, t) dxdtdαdβ, as n → ∞. (6.84) lima1
0 0 IR + IR d

Similarly, let M be the Lipschitz constant of f on [−D, D] where D = max{kµk∞ , kνk∞ }, with k·k∞ =
k·kL∞(IR d ×IR ?+ ×(0,1)) ,
Z 1Z 1Z Z  
|A3 − f (µ(x, t, α)>ν(x, t, β)) − f (µ(x, t, α)⊥ν(x, t, β))
0 0 IR + IR d
v(x, t) · ∇ψ(x, t)dxdtdαdβ| ≤ 2M V k(|∇ψ|)kL∞ (IR d ×IR ?+ ) ε(n, S),
 177

which yields
Z 1 Z 1 Z Z  
A3 → f (µ(x, t, α)>ν(x, t, β)) − f (µ(x, t, α)⊥ν(x, t, β))
0 0 IR + IR d (6.85) lima3
v(x, t) · ∇ψ(x, t)dxdtdαdβ, as n → ∞.

Let us now consider A2 + A4 .

For (x, t, α) ∈ IR d × IR + × (0, 1), let us take ϕ(y, s) = ψ(x, t)ρn (x − y)ρ̄n (t − s) and κ = µ(x, t, α) in
ewsnu
(6.80). Integrating the result over IR d × IR + × (0, 1) leads to

−A2 − B4 ≥ 0, (6.86) a2b4

with
Z 1 Z 1 Z ∞ Z Z ∞ Z h 
A4 − B 4 = f (µ(x, t, α)>ν(y, s, β)) − f (µ(x, t, α)⊥ν(y, s, β))
0 0 0 IR d 0 IR d i
(v(x, t) − v(y, s)) · ∇ρn (x − y)ψ(x, t)ρ̄n (t − s) dxdtdydsdαdβ.
lima24
Note that B4 = A4 if v is constant (and one directly obtains (6.88) below). In the general case, in order
lima24
to prove that A4 − B4 → 0 as n → ∞ (which then gives (6.88)), let us remark that, using divv = 0,
Z 1 Z 1 Z ∞ Z Z ∞ Z h 
f (µ(x, t, α)>ν(x, t, β)) − f (µ(x, t, α)⊥ν(x, t, β))
0 0 0 IR d 0 IR d i (6.87) abc4
(v(x, t) − v(y, s)) · ∇ρn (x − y)ψ(x, t)ρ̄n (t − s) dxdtdydsdαdβ = 0.

Indeed, the latter equality

abc4
follows from an integration by parts for the variable y ∈ IR d . Then, substracting
the left hand side of (6.87) to A4 − B4 and using the regularity of v, there exists C1 , only depending on
a2b4
M , v and ψ, such that |A4 − B4 | ≤ C1 ε(n, S). This gives A4 − B4 → 0 as n → ∞ and, thanks to (6.86),

lim sup(A2 + A4 ) ≤ 0. (6.88) lima24

n→∞

Finally, let us consider A5 . R∞ ewsnu

For x ∈ IR d , let us take ϕ(y, s) = ψ(x, 0)ρn (x − y) s ρ̄n (−τ )dτ and κ = u0 (x) in (6.80). Integrating the
resulting inequality with respect to x ∈ IR d gives

−A5 + B5a + B5b ≥ 0, (6.89) a5c

with
Z 1 Z ∞ Z Z Z ∞
B5a = − (f (ν(y, s, β)>u0 (x)) − f (ν(y, s, β)⊥u0 (x)))
0 0 IR d IR d s
v(y, s) · ∇ρn (x − y)ψ(x, 0)ρ̄n (−τ )dτ dydxdsdβ,
Z Z
B5b = ψ(x, 0)ρn (x − y)|u0 (x) − u0 (y)|dydx.
IR d IR d

Let S0 = {x ∈ IR d ; ψ(x, 0) 6= 0} and

Z
1
ε0 (n, S0 ) = sup{ |u0 (x) − u0 (x + η)|dx; |η| ≤ },
S0 n
so that B5b ≤ kψ(·, 0)kL∞(IR d ) ε0 (n, S0 ).
Since u0 ∈ L1loc (IR d ) and since S0 is bounded, one has ε0 (n, S0 ) → 0 as n → ∞. Then, B5b → 0 as
n → ∞.
 178

a5c lima5
Let us now prove that B5a → 0 as n → ∞ (then, (6.89) will give (6.90) below). Note that B5a =
−B5c + (B5a + B5c ) with
Z 1 Z ∞ Z Z Z ∞
B5c = (f (ν(y, s, β)>u0 (y)) − f (ν(y, s, β)⊥u0 (y)))
0 0 IR d IR d s
v(y, s) · ∇ρn (x − y)ψ(x, 0)ρ̄n (−τ )dτ dydxdsdβ.
Integrating by parts for the x variable yields
Z 1Z ∞Z Z Z ∞
B5c = (f (ν(y, s, β)>u0 (y)) − f (ν(y, s, β)⊥u0 (y)))
0 0 IR d IR d s
v(y, s) · ∇ψ(x, 0)ρn (x − y)ρ̄n (−τ )dτ dydxdsdβ.
Noting that the integration with respect to s is reduced to [0, 1/n], B5c → 0 as n → ∞.
There remains to study B5a + B5c . Noting that |f (a>b) − f (a>c)| ≤ M̄|b − c| and |f (a⊥b) − f (a⊥c)| ≤
M̄ |b − c| if b, c ∈ [−D̄, D̄], where D̄ = ku0 kL∞ (IR d ) and M̄ is the Lipschitz constant to f on [−D̄, D̄],
Z ∞Z Z Z ∞
|B5a + B5c | ≤ 2M̄V |u0 (x) − u0 (y)||∇ρn (x − y)|ψ(x, 0)ρ̄n (−τ )dτ dydxds,
0 IR d IR d s

which yields the existence of C2 , only depending on M̄, V and ψ, such that
Z 1
n
Z Z
|B5a + B5c | ≤ C2 |u0 (x) − u0 (x − z)|nd+1 dzdxds.
1
0 S0 B(0, n )

Therefore, |B5a + B5c | ≤ C3 ε0 (n, S0 ), with some C3 only depending on M̄ , V and ψ. Since
a5c
ε0 (n, S0 ) → 0
as n → ∞, one deduces |B5a + B5c | → 0 as n → ∞. Hence, B5a → 0 as n → ∞ and (6.89) yields

lim sup A5 ≤ 0. (6.90) lima5

n→∞
ewscin lima1 lima3
Itlima24
is now possible
lima5
to conclude
step1
Step 1. Passing to the limit as n → ∞ in (6.83) and using (6.84), (6.85),
(6.88) and (6.90) yields (6.81).
Step 2 (proof of µ = ν and conclusion) estdonnees
Let R > 0 and T > 0. One sets ω = V M (recall that V is given in Assumption 6.1 and that M is given
in Step 1).
Let ϕ ∈ Cc1 (IR + , [0, 1]) be a function such that ϕ(r) = 1 if r ∈ [0, R + ωT ], ϕ(r) = 0 if r ∈ [R + ωT + 1, ∞)
and ϕ0 (r) ≤ 0, step1
for all r ∈ IR + .
One takes, in (6.81), ψ defined by

ψ(x, t) = ϕ(|x| + ωt) TT−t , for x ∈ IR d and t ∈ [0, T ],
ψ(x, t) = 0, for x ∈ IR d and t ≥ T.
The function ψ is not in Cc∞ (IR d × IR + , IR + ), but, using a usual regularization step1
step1
technique, it may be
proved that such a function can be considered in (6.81), in which case Inequality (6.81) writes

Z 1 Z 1 Z T Z h T − t 
1
ωϕ0 (|x| + ωt) − ϕ(|x| + ωt) +
|µ(x, t, α) − ν(x, t, β)|
0 0 0 IR d T T − t T
xi
f (µ(x, t, α)>ν(x, t, β)) − f (µ(x, t, α)⊥ν(x, t, β)) ϕ0 (|x| + ωt)v(x, t) · dxdtdαdβ ≥ 0.
T |x|

Since ω = V M and ϕ0 ≤ 0, one has (f (a>b)−f (a⊥b))ϕ0(|x|+ωt)v(x, t)·(x/|x|) ≤ |a−b|ω(−ϕ0 (|x|+ωt)),

for a.e. (x, t) ∈ IR d × IR ?+ and all a, b ∈ [−D, D] (D is defined in Step 1). Therefore, since ϕ(|x| + ωt) = 1
if (x, t) ∈ B(0, R) × [0, T ], the preceding inequality gives
 179

Z 1 Z 1 Z T Z
|µ(x, t, α) − ν(x, t, β)|dxdtdαdβ ≤ 0,
0 0 0 B(0,R)

which yields, since R and T are arbitrary, µ(x, t, α) = ν(x, t, β) for a.e. (x, t, α, β) ∈ IR d × IR ?+ × (0, 1) ×
(0, 1).
Let us now deduce also from this uniqueness result that there exists u ∈ L∞ (IR d × IR ?+ ) such that
estepsol
µ(x, t, α) = u(x, t), for a.e. (x, t, α) ∈ IR d × estpbcont
IR ?+ ×estcondini
(0, 1) (then it is easy to see, with Definition 6.2, that
u is the entropy weak solution to Problem (6.1)-(6.2)). eghbcepsol
Indeed, it is possible to take, in the preceeding proof, µ= ν (recall that the
estpbcont
estcondini
proposition 6.3 gives the
eghbexepsol
existence of an entropy process solution to Problem (6.1)-(6.2), see Remark 6.9). This yields µ(x, t, α) =
µ(x, t, β) for a.e. (x, t, α, β) ∈ IR d × IR ?+ × (0, 1) × (0, 1). Then, for a.e. (x, t) ∈ IR d × IR ?+ , one has

µ(x, t, α) = µ(x, t, β) for a.e. (α, β) ∈ (0, 1) × (0, 1)

and, for a.e. α ∈ (0, 1),

µ(x, t, α) = µ(x, t, β) for a.e. β ∈ (0, 1).

d
Thus, defining u from IR × IR ?+ to IR by
Z 1
u(x, t) = µ(x, t, β)dβ,
0

one obtains µ(x, α) = u(x, t), for a.e. (x, t, α) ∈ IR d × IR ?+ ×ewsunicite

t, estcondini
estpbcont
(0, 1), and u is the entropy weak solution
to Problem (6.1)-(6.2). This completes the proof of Theorem 6.3.

6.6.3 Convergence towards the entropy weak solution

eghbscewsol estpbcont
estcondini estcondini
We now know that there exists a unique entropy process solution
estpbcont
to problem ( 6.1)-(6.2) page 145, which
estcondini
is identical to the entropy weak solution of problem (6.1)-(estschema
6.2); we may now prove the convergence of the
estschemaz estuapp
approximate solution given by the finite volume scheme (6.7), (6.5) and (6.8) towards the entropy weak
solution as the mesh size tends to 0.
estdonneesestdonnees
eghbcewsol Theorem 6.4 Under Assumptions meshypmd 145, let α ∈ IR ?+ and ξ ∈ (0, 1) estcfl
6.1 page meshypmd be given. For an admissible
mesh T in the sense of Definitionestschema 6.1 page 148 and for
estschemaz estuapp
k > 0 satisfying ( 6.6) (note that α and ξ are
fixed), let uT ,k be the solution to (6.7), (6.5) and (6.8).
Then, uT ,k → Lploc (IR d estcondini
u inestcondini
estpbcont
× IR + ) for all p ∈ [1, ∞), as h = size(T ) → 0, where u is the entropy weak
solution to (6.1)-(6.2) page 145.
eghbcewsol
Proof of Theorem 6.4
In order to prove that uT ,k → u (in Lploc (IR d × IR + ) for all p ∈ [1, ∞), as h = size(T ) → 0), let us proceed
by a classical
estpbcont
way of contradiction which uses the uniqueness of the entropy process solution to Problem
estcondini estcondini
(6.1)-(6.2) page 145. Assume that there exists 1 ≤ p0 < ∞, ε > 0, ω̄ a compact subset of IR d , Testcfl > 0 and
a sequence ((Tm , km ))m∈IN such that, for any m ∈ IN, Tm is an admissible mesh, km satisfies (6.6) (with
T = Tm and k = km , note that α and ξ are independent of m), size(Tm ) → 0 as m → ∞ and
Z T Z
|uTm ,km − u|p0 dxdt ≥ ε, ∀m ∈ IN, (6.91) eghbcontra
0 ω̄
estschema
estschemaz estuapp
where uTm ,kmestpbcont
is the solution to (6.7), (6.5) and (6.8) with T = Tm and k = km and u is the entropy weak
estcondini
solution to (6.1)-(6.2).
eghbcepsol
Using Proposition 6.3, there exists a subsequence of the sequence ((Tm , km ))m∈IN , still denoted by ((Tm ,
km ))m∈IN , and a function µ ∈ L∞ (IR d × IR ?+ × (0, 1)) such that
 180

1. uTm ,km → µ, as m → ∞, in the nonlinear weak-? sense, that is:

Z ∞ Z Z 1 Z ∞ Z
lim θ(uTm ,km (x, t))ϕ(x, t)dxdt = θ(µ(x, t, α))ϕ(x, t)dxdtdα,
m→∞ 0 IR d 0 0 IR d (6.92) eghbww
∀ϕ ∈ L (IR d × IR ?+ ), ∀θ ∈ C(IR, IR),
1

estpbcont
estcondini
2. µ is an entropy process solution to (6.1)-(6.2).
ewsuniciteewsunicite
By Theorem
estpbcont
·, α) = u, for a.e. α ∈ [0, 1] (and u is the entropyeghbww
6.3 page 175, one has µ(·,eghbww
estcondini
weak solution to
(6.1)-(6.2)). Taking first θ(s) = s2 in (6.92) and then θ(s) = s and ϕu instead of ϕ in (6.92) one obtains:
Z ∞Z
(uTm ,km (x, t)) − u(x, t))2 ϕ(x, t)dxdt → 0, as m → ∞, (6.93) eghbfort
0 IR d
eghbfort
for any function ϕ ∈ L1 (IR d × (0, T )). From (6.93), and thanks to the L∞ -bound on (uTm ,km )m∈IN , one
deduces the convergence of (u )
eghbcontra Tm ,km m∈IN
towards u in Lploc (IR d × IR + ) for all p ∈ [1, ∞), which is in
contradiction with (6.91).
This completes the proof of our convergence theorem.

Remark 6.10
eghbcewsol estschemai
1. Theorem 6.4 is also
estuappimp
true with the implicit scheme insteadestcfl
estschema estuapp
of the explicit scheme (that is (6.9) and
(6.10) instead of (6.7) and (6.8)) without the condition (6.6) (and thus without ξ).
2. The following section improves this convergence result and gives an error estimate.

6.7 Error estimate

esterr
6.7.1 Statement of the results
This section is devoted to the proof of an error estimate of time explicit and time implicit finite volume
estpbcont
estcondini estcondini
approximations to the solution u ∈ L∞ (IR d × IR ?+ ) of Problem (6.1)-(6.2) page 145. Assuming that
u0 ∈ BV (IR d ), a “h1/4 ” error estimate is shown for eghbns
a large variety
eghbns
of finite volume monotone flux
schemes such as those which were presented in Section 6.2 page 148.
estdonneesestdonnees meshypmd meshypmd
Under Assumption 6.1 page 145, let T be an admissible mesh in the sense of Definition 6.1 page 148 and
estcondflu
k > 0. Let g ∈ C(IR 2 , IR) satisfying Assumption
estpbcont
6.2.
estcondini
Let u be
estschema
the entropy
estschemaz
estuapp
weak solutionestcfl
of (6.1)-(6.2) and let uT ,k be the solution of the time explicit scheme
estschemai
(6.7), (estuappimp
6.5), (6.8), assuming that (6.6) holds, or uT ,k be the solution of the time implicit scheme (6.9),
estschemaz
(6.5), (6.10). Our aim is to give an error estimate between u and uT ,k .

In the case of the explicit scheme, one proves, in this section, the following theorem.
estdonneesestdonnees
esth14 Theorem 6.5 Under Assumption 6.1 page 145, let T be an admissible
meshypmd meshypmd estcondflu
mesh in the sense of estcfl
Definition
6.1 page 148 and k > 0. Let g ∈ C(IR 2 , IR) satisfy Assumption
estpbcont
estcondini
6.2 and assume that condition
estuappestschema
( 6.6) holds.
estschemaz
Let u be the unique entropy weak solution of (6.1)-(6.2) and uT ,k be given by (6.8), (6.7), (6.5). Assume
u0 ∈ BV (IR d ). Then, for all R > 0 and all T > 0 there exists Ce ∈ IR + , only depending on R, T , v, g,
u0 , α and ξ, such that the following inequality holds:
Z TZ
1
|uT ,k (x, t) − u(x, t)|dxdt ≤ Ce h 4 . (6.94) esteqcorol
0 B(0,R)

(Recall that B(0, R) = {x ∈ IR d , |x| < R}.)

 181

esth14 R
In Theorem 6.5, u0 is assumed to belong to BV (IR d ) (recall that u0 ∈ BV (IR d ) if sup{ u0 (x)divϕ(x)dx,
Cc∞ (IR d , IR d ); |ϕ(x)| ≤ 1, ∀x ∈ IR d } < ∞). This assumption allows us to obtain an h1/4 estimate
ϕ ∈esteqcorol
in (6.94). If u0 6∈ BV (IR d ) (but u0 still belongs to L∞ (IRestepsrk
d
)), one canestepsork
also give an error estimate which
depends on the functions ε(r, S) and ε0 (r, S) defined in (6.109) and (6.116).
esth14 esth14i
A slight improvement of Theorem 6.5 (and also Theorem 6.6 below) is esth14 possible. Using the fact that
u ∈ C(IR + , L1loc (IR d )) and thus u(·, t) is defined for all t ∈ IR + , Theorem 6.5 remains true with
Z
|uT ,k (x, t) − u(x, t)|dx ≤ Ce h1/4 , ∀t ∈ [0, T ],
B(0,R)
esteqcorol
instead of (6.94). The proof of such a result may be handled
EGH2
with an adaptation of the proof a uniqueness
Vi
of the entropy process
Coquel
solution given for instance in Eymard, Gallouët and Herbin [1995], see Vila
[1994] and Cockburn, Coquel and LeFloch [1994] for some similar results.
esth14
In some cases, it is possible to obtain h1/2 , instead of h1/4 , in Theorem 6.5. This is the case, for instance,
when the mesh T is composed of rectangles (d = 2) and when v does not depend on (x, t), since, in
estbvfx
this case, one obtains a “BV estimate” on uT ,k √ . In this case, the right hand sides of inequalities (6.16)
estbvft estmapp
and (6.17), proven √ above, are changed from C/1/2h to C, so that the right hand side of (6.56) becomes
esteqcorol
Ch instead of C h, which in turn yields Ce h in (6.94) instead of Ce h1/4 . It is, however, still an
open problem
esth14
obtain an error estimate with h1/2 , instead of h1/4 , in
to know whether it is possible to esth14
Theorem
C
6.5 (under the hypotheses of Theorem 6.5), even in the case where v does not depend on (x, t)
(see Cockburn and Gremaud [1996] for an attempt in this direction).
esth14 esth14i
estext Remark 6.11 Theorem 6.5 (and also Theorem 6.6) remains true with some slightly
estcondflu
more general as-
estschema
sumption on g, instead of 6.2, in order to allow g to depend on T and k. Indeed, in (6.7), one can replace
g(unK , unL ) (and g(unL , unK )) by gK,L (unK , unL , T , k) (and gL,K (unL , unK , T , k)). Assume that, for all K ∈ T
and all L ∈ N (K), the function (a, b) 7→ gK,L (a, b, T , k), from [Um , UM ]2 to IR, is nondecreasing with
respect to a, nonincreasing with respect to b, Lipschitz continuous uniformly with respect to K and L
d
K,L (a, a, T , k) = f (a) for all a ∈ [Um , UM ] (recall that Um ≤ u0 ≤ UM a.e. on IR ). Then
and that gesth14
Theorem 6.5 remains true.
estcfl esteqcorol esth14
However, note that condition (6.6) and Ce in the estimate (6.94) of Theorem 6.5 depend on the Lipschitz
constants of gK,L (·, ·, T , k) on [Um , UM ]2 . An interesting form for gK,L is gK,L (a, b, T , k) = cK,L (T , k)f (a)
+ (1−cK,L (T , k)) f (b) + DK,L (T , k) (a−b), with some cK,L (T , k) ∈ [0, 1] and DK,L (T , k) ≥ 0. In order to
obtain the desired properties on gK,L , it is sufficient to take max{|f 0 (s)|, s ∈ [Um , UM ]} ≤ DK,L (T , k) ≤ D
(for all K, L), with some D ∈ IR. The Lipschitz constants of gK,L on [Um , UM ]2 only depend on D, f ,
Um and UM .
For instance, a “Lax-Friedrichs type” scheme consists, roughly speaking, in taking D K,L (T , k) of order
“h/k”. The desired properties on gK,L are satisfied, provided that k/h ≤ C, with some C depending on
max{|f 0 (s)|, s ∈ [Um , UM ]}. Note, however, that the condition k/h ≤ C is not sufficient to give a real
esteqcorol
“h1/4 ” estimate, since the coefficient Ce in (6.94) depends on D. Taking, for example, k of order “h2 ” leads
to an estimate “Ce h1/4 ” which do not goes to 0 as h goes to 0 (indeed, it is known, in estcondini
this case, that the
estpbcont
approximate solution does not converge towards the entropy weak solution to ( 6.1)-(6.2)). One obtains
a real “h1/4 ” estimate, in the case of that “Lax-Friedrichs type” scheme, by taking C1 ≤ (k/h) ≤ C2 . In
order to avoid the condition C1 ≤ (k/h) (note that (k/h) ≤ C2 is imposed by the Courant-Friedrichs-Levy
estcfl
condition 6.6), a possibility is to take DK,Lflumono D = max{|f 0 (s)|, s ∈ [Um , UM ]} (this is related to
(T , k) =flumono
the “modified Lax-Friedrichs ” of Example 5.2 page 132 in the 1D case). Then D only depends on f and
esth14
u0 and, in the estimate “Ce h1/4 ” of Theorem 6.5, Ce only depends on R, T , v, f , u0 , α and ξ, which
leads to a convergence result at rate “h1/4 ” as h → 0 (with fixed α and ξ).
In the case of the implicit scheme, one proves the following theorem.
estdonneesestdonnees
esth14i Theorem 6.6 Under Assumption 6.1 page 145, let T be an admissible
meshypmd meshypmd
mesh in the sense of Definition
estcondflu
k > 0. Let g ∈ C(IR 2 , IR) satisfy Assumption 6.2. Let u be the unique entropy weak
6.1 page 148estpbcont
and estcondini
solution of (6.1)-(6.2). Assume that u0 ∈ BV (IR d ) and that v does not depend on t.
 182

estschemai estschemaz
Let {unK , n ∈ IN, K ∈ T } be the unique solution to (6.9) and (6.5) such that unK ∈ esteu
[Um , UM ] for all
K ∈ T and estuappimp
n ∈ IN (existence and uniqueness of such a solution is given by Proposition 6.1). Let uT ,k be
defined by (6.10).
Then, for all R > 0 and T > 0, there exists Ce , only depending on R, T , v, g, u0 and α, such that the
following inequality holds:
Z T Z
1 1
|uT ,k (x, t) − u(x, t)|dxdt ≤ Ce (k + h 2 ) 2 . (6.95) esteqcoroli
0 B(0,R)
esth14i
estexti Remark 6.12 Note that, in Theorem 6.6, there is no restriction on k (this is usual for an implicit
scheme),
esth14i
and one obtains an “h1/4 ” error estimate for some “large” k, namely if k ≤ h1/2 . In Theorem
6.6, if v depends on t and u0 ∈ L∞ (IR d ) (but u0 not necessarily in BV (IR d )), one can also give an error
estimate. Indeed one obtains
Z T Z
k 1 1
|uT ,k (x, t) − u(x, t)|dxdt ≤ Ce ( 1 + h2 )2 ,
0 B(0,R) h 2

which yields an “h1/4 ” error estimate if k is of order “h”.

esth14 esth14i estentest estentesti
Theorem 6.5 (resp. Theorem 6.6) is an easy consequence of Theorem 6.1 (resp.
estpbcont
6.2) and of a quite
estcondini
general theorem of comparison between the entropy weak
esterrest
solution to ( 6.1)-( 6.2) and an approximate
solution. This theorem of comparison (Theorem
estpbcont
6.7)
estcondini
may be used in other frameworks (for instance, to
compare the entropy weak
estpbcont
solution to ( 6.1)-( 6.2) and the approximate
compthm
solution obtained with aesth14
parabolic
esth14i
regularization of (6.1)). It is stated and
sprelim
proved in Section 6.7.3 where the proofs of theorems 6.5 and 6.6
estlemma
are also given. First, in Section 6.7.2, two preliminary lemmata are given. Indeed, Lemma 6.10 is the
crucial part of the two following sections.

6.7.2 Preliminary lemmata

sprelim
Let us first give a classical lemma on the space BV .

bvtrans Lemma 6.9 Let u ∈ BVloc (IR p ), p ∈ IN? , that is u ∈ L1loc (IR p
) and the
defBVRp
restriction of u to Ω belongs to
defBVRp
p
BV (Ω) for all open bounded subset Ω of IR (see Definition 5.38 page 138 for the definition of BV (Ω)).
Then, for all bounded subset Ω of IR p and for all a > 0,

ku(· + η) − ukL1 (Ω) ≤ |η||u|BV (Ωa ) , ∀η ∈ IR p , |η| ≤ a, (6.96) bvdtrans

p
where Ωa = {x ∈ IR ; d(x, Ω) < a} and d(x, Ω) = inf{|x − y|, y ∈ Ω} is the distance from x to Ω.
bvtrans
Proof of Lemma 6.9
Let Ω be a bounded subset of IR p and η ∈ IR p . The following equality classically holds:
Z
ku(· + η) − ukL1 (Ω) = sup{ (u(x + η) − u(x))ϕ(x)dx, ϕ ∈ Cc∞ (Ω, IR), kϕkL∞ (Ω) ≤ 1}. (6.97) defl1
Ω

Let ϕ ∈ Cc∞ (Ω, IR) such that kϕkL∞(Ω) ≤ 1.

Since ϕ(x) = 0 if x ∈ Ω|η| \ Ω (recall that Ω|η| = {x ∈ IR p ; d(x, Ω) < η}),
Z Z
u(x)ϕ(x)dx = u(x)ϕ(x)dx.
Ω Ω|η|

Similarly, using an obvious change of variables,

Z Z
u(x + η)ϕ(x)dx = u(x)ϕ(x − η)dx.
Ω Ω|η|
 183

Therefore,

Z Z Z Z 1
(u(x + η) − u(x))ϕ(x)dx = u(x)(ϕ(x − η) − ϕ(x))dx = − u(x)( ∇ϕ(x − sη) · ηds)dx
Ω Ω|η| Ω|η| 0

and, with Fubini’s theorem,

Z Z 1 Z
(u(x + η) − u(x))ϕ(x)dx = ( u(x)∇ϕ(x − sη) · ηdx)ds. (6.98) vtt
Ω 0 Ω|η|

For all s ∈ (0, 1), Define ψs ∈ Cc∞ (Ω|η| , IR p ) by ψs (x) = ϕ(x − sη)η; since ψs ∈ Cc∞ (Ω|η| , IR p ) and
|ψs (x)| ≤ |η| for all x ∈ IR p , the definition of |u|BV (Ω|η| ) yields
Z Z
u(x)∇ϕ(x − sη) · ηdx = u(x)divψs (x)dx ≤ |η||u|BV (Ω|η| ) .
Ω|η| Ω|η|
vtt
Then, (6.98) gives
Z
(u(x + η) − u(x))ϕ(x)dx ≤ |η||u|BV (Ω|η| ) . (6.99) vttt
Ω
vttt defl1
Taking in (6.99) the supremum over ϕ ∈ Cc∞ (Ω, IR) such that kϕkL∞(Ω) ≤ 1 yields, thanks to (6.97),

ku(· + η) − ukL1 (Ω) ≤ |η||u|BV (Ω|η| ) , ∀η ∈ IR p ,

bvdtrans
and (6.96) follows, since Ω|η| ⊂ Ωa if |η| ≤ a.

bvtrans
finvacances Remark 6.13 Let us give an application of the lemma 6.9 which will be quite useful R further on. Let
u ∈ BVloc (IR p ), p ∈ IN? . Let ψ, ϕ ∈ Cc (IR p , IR + ), a > 0 and 0 < ε < a such that IR p ϕ(x)dx = 1 and
ϕ(x) = 0 for all x ∈ IR p , |x| > ε. Let S = {x ∈ IR p , ψ(x) 6= 0}.
Then,
Z Z
|u(x) − u(y)|ψ(x)ϕ(x − y)dydx ≤ εkψkL∞(IR p ) |u|BV (Sa ) , (6.100) bvdvar
IR p IR p

where Sa = {x ∈bvtrans
IR p , d(x, S) < a}.
Indeed, Lemma 6.9 gives

ku(· + η) − ukL1 (S) ≤ |η||u|BV (Sa ) , ∀η ∈ IR p , |η| ≤ a. (6.101) ousuisje

bvdvar
Using a change of variables in the left hand side of (6.100),
Z Z Z Z
|u(x) − u(y)|ψ(x)ϕ(x − y)dydx ≤ kψkL∞(IR p ) ( |u(x) − u(x − z)|dx)ϕ(z)dz.
IR p IR p B(0,ε) S
ousuisje
Then, (6.101) yields
Z Z Z
|u(x) − u(y)|ψ(x)ϕ(x − y)dydx ≤ εkψkL∞(IR p ) |u|BV (Sa ) ϕ(z)dz,
IR p IR p IR p
bvdvar
which gives (6.100).
 184

estdonnees
estlemma Lemma 6.10 Under assumption 6.1, let u0 ∈ BV (IR d ) and ũ ∈ L∞ (IR d × IR ?+ ) such that Um ≤ ũ ≤ UM
a.e. on IR d × IR ?+ . Assume that there exist µ ∈ M(IR d × IR + ) and µ0 ∈ M(IR d ) such that
 Z Z 

 |ũ(x, t) − κ|ϕt (x, t)+



 IR + IR d 



 (f (ũ(x, t)>κ) − f (ũ(x, t)⊥κ))v(x, t) · ∇ϕ(x, t) dxdt +
 Z



|u0 (x) − κ|ϕ(x, 0)dx ≥ (6.102) est1est

 Zd
IR   Z



 − |ϕt (x, t)| + |∇ϕ(x, t)| dµ(x, t) − |ϕ(x, 0)|dµ0 (x),

 IR d ×IR + IR d






∀κ ∈ IR, ∀ϕ ∈ Cc∞ (IR d × IR + , IR + ).
estpbcont
estcondini
u be the unique entropy weak solution of (6.1)-(6.2) (i.e. u ∈ L∞ (IR d × IR ?+ ) is the unique solution
Let est1est
to (6.102) with u instead of ũ and µ = 0, µ0 = 0).
Then for all ψ ∈ Cc∞ (IR d × IR + , IR + ) there exists C only depending on ψ (more precisely on kψk∞ ,
kψt k∞ , k∇ψk∞ , and on the support of ψ), v, f , and u0 , such that
 Z Z h

 |ũ(x, t) − u(x, t)|ψt (x, t) +

 IR + IR d
  i
(6.103) est3est

 f (ũ(x, t)>u(x, t)) − f (ũ(x, t)⊥u(x, t)) (v(x, t) · ∇ψ(x, t)) dxdt ≥

 1
−C(µ0 ({ψ(·, 0) 6= 0}) + (µ({ψ 6= 0})) 2 + µ({ψ 6= 0})),
where {ψ 6= 0} = {(x, t) ∈ IR d × IR + , ψ(x, t) 6= 0} and {ψ(·, 0) 6= 0} = {x ∈ IR d , ψ(x, 0) 6= 0}. (Note
that k·k∞ = k·kL∞ (IR d ×IR ?+ ) .)
estlemma
Proof of Lemma 6.10
estlemma ewsunicite
The proof of Lemma 6.10 is close to that of step 1 in the proof of Theorem 6.3. Let us first define mollifiers
in IR and IR d . For p = 1 and p = d, one defines ρp ∈ Cc∞ (IR p , IR) satisfying the following properties:

supp(ρp ) = {x ∈ IR p ; ρp (x) 6= 0} ⊂ {x ∈ IR p ; |x| ≤ 1},

ρp (x) ≥ 0, ∀x ∈ IR p ,
Z
ρp (x)dx = 1
IR p
and furthermore, for p = 1,

ρ1 (x) = 0, ∀x ∈ IR + . (6.104) estrhodec

p
For r ∈ IR, r ≥ 1, one defines ρp,r (x) = rp ρp (rx), for all x ∈ IR . est1est
Using the mollifiers ρp,r will allow to choose convenient test functions in (6.102) (which are the inequalities
satisfied by ũ) and in the analogous inequalities satisfied by u which are

 Z Z h   i


 |u(y, s) − κ|ϕs (y, s) + f (u(y, s)>κ) − f (u(y, s)⊥κ) v(y, s) · ∇ϕ(y, s) dyds+
ZIR + IR d (6.105) estkruzkovys


 |u0 (y) − κ|ϕ(y, 0)dy ≥ 0, ∀κ ∈ IR, ∀ϕ ∈ Cc∞ (IR d × IR + , IR + ).
IR d
est1est estkruzkovys
Indeed, the main tool is to take κ = u(y, s) in (6.102), κ = ũ(x, t) in (6.105) and to introduce mollifiers
in order to have y close to x and s close to t.
 185

Let ψ ∈ Cc∞ (IR d × IR + , IR + ), and let ϕ : (IR d × IR + )2 → IR + be defined by:

ϕ(x, t, y, s) = ψ(x, t)ρd,r (x − y)ρ1,r (t − s).

Note that, for any (y, s) ∈ IR d × IR + , one has ϕ(·, ·, y, s) ∈ Cc∞ (IR d × IR + , IR + ) and, for any (x, t) ∈
est1est
IR d × IR + , one has ϕ(x, t, ·, ·) ∈ Cc∞ (IRestkruzkovys
d
× IR + , IR + ). Let us est1est
take ϕ(·, ·, y, s) as test function ϕ in (estkruzkovys
6.102)
and ϕ(x, t, ·, ·) as test function ϕ in (6.105). We take, in (6.102), estkruzkovys
est1est
κ = u(y, s) and we take, in (6.105),
κ = ũ(x, t). We then integrate (6.102) for (y, s) ∈ IR d × IR + , and (6.105) for (x, t) ∈ IR d × IR + . Adding
the two inequalities yields

E11 + E12 + E13 + E14 ≥ −E2 , (6.106) estcin

where
Z ∞ Z Z ∞ Z h i
E11 = |ũ(x, t) − u(y, s)|ψt (x, t)ρd,r (x − y)ρ1,r (t − s) dxdtdyds,
0 IR d 0 IR d
Z ∞Z Z ∞ Z h 
E12 = f (ũ(x, t)>u(y, s)) − f (ũ(x, t)⊥u(y, s))
0 IR d 0 IR d i
v(x, t) · ∇ψ(x, t)ρd,r (x − y)ρ1,r (t − s) dxdtdyds,
Z ∞ Z Z ∞ Z  
E13 = − f (ũ(x, t)>u(y, s)) − f (ũ(x, t)⊥u(y, s)) ψ(x, t)
0 IR d 0 IR d
(v(y, s) − v(x, t)) · ∇ρd,r (x − y)ρ1,r (t − s)dxdtdyds,
Z Z ∞ Z
E14 = |u0 (x) − u(y, s)|ψ(x, 0)ρd,r (x − y)ρ1,r (−s)dydsdx
IR d 0 IR d
and
Z ∞ Z Z 
E2 = |ρd,r (x − y)(ψt (x, t)ρ1,r (t − s) + ψ(x, t)ρ01,r (t − s))|
0 IR d IR d ×IR + 
+|ρ1,r (t − s)(∇ψ(x, t)ρd,r (x − y) + ψ(x, t)∇ρd,r (x − y))| dµ(x, t)dyds (6.107) est5est
Z ∞Z Z
+ |ψ(x, 0)ρd,r (x − y)ρ1,r (−s)|dµ0 (x)dyds.
0 IR d IR d
estcin
One may be surprised by the fact that the inequation (6.106) isestcondini
obtained without using the initial condition
estpbcont
which is satisfied by the entropy weak solutionestkruzkovys
u of (6.1)-(6.2). Indeed, this initial condition appears
only in the third term of the left hand side of (6.105); since ϕ(x, t, ·, 0) = 0 for all (x, t) ∈ IR d × IRestkruzkovys
estkruzkovys + , the
t, ·, ·)estcondini
third term of the left hand side of (6.105) is zero when ϕ(x,estpbcont is chosen as a test function in (6.105).
However, the fact that u satisfies the initial condition of (6.1)-(6.2) will be used later in order to get a
bound on E14 .
estcin
Let us now study the five terms of (6.106). One sets S = {ψ 6= 0} = {(x, t) ∈ IR d × IR + ; ψ(x, t) 6= 0}
and S0 = {ψ(·, 0) 6= 0} = {x ∈ IR d ; ψ(x, 0) 6= 0}. In the following, the notation Ci (i ∈ IN) will refer to
various real quantities only depending on kψk∞ , kψt k∞ , k∇ψk∞ , S, S0 , v, f , and u0 .
est5est
Equality (6.107) leads to

E2 ≤ (r + 1)C1 µ(S) + C2 µ0 (S0 ). (6.108) est6est

d estrhodec
Let us handle the term E11 . For all x ∈ IR and for all t ∈ IR + , one has, using (6.104),
Z Z ∞
ρd,r (x − y)ρ1,r (t − s)dsdy = 1.
IR d 0
Then,
 186

Z Z h i
|E11 − |ũ(x, t) − u(x, t)|ψt (x, t) dxdt| ≤
Z ∞Z Z+ ∞IRZd h
IR
i
|u(x, t) − u(y, s)||ψt (x, t)|ρd,r (x − y)ρ1,r (t − s) dxdtdyds ≤ kψt k∞ ε(r, S),
0 IR d 0 IR d
with
1 1
ε(r, S) = sup{ku − u(· + η, · + τ )kL1 (S) , |η| ≤ , 0 ≤ τ ≤ }. (6.109) estepsrk
r r
estpbcont
estcondini
Since u0 ∈ BV (IR d ), the function u (entropy weak solution tokrushkov
(6.1)-(6.2)) belongs to BV (IR d × (−T, T )),
Chainais2
for all T > 0, setting, for instance, u(., t) = u0 for t < 0 (see Krushkov [1970] or Chainais-Hillairet
[1999] where this result is proven passing to the limit on numerical schemes).
bvtrans √ bvtrans
Then, Lemma 6.9 gives, since r ≥ 1, (taking p = d + 1, Ω = S and a = 2 in Lemma 6.9,)
C3
ε(r, S) ≤ . (6.110) estubv
r
Hence,
Z Z h i C4
|E11 − |ũ(x, t) − u(x, t)|ψt (x, t) dxdt| ≤ . (6.111) est7estb
IR + IR d r
In the same way, using |f (a>b) − f (a>c)| ≤ M |b − c| and |f (a⊥b) − f (a⊥c)| ≤ M |b − c| for all a, b,
c ∈ [Um , UM ] where M is the Lipschitz constant of f in [Um , UM ],
Z Z  
|E12 − f (ũ(x, t)>u(x, t)) − f (ũ(x, t)⊥u(x, t))
IR + IR d (6.112) est8est
(v(x, t) · ∇ψ(x, t))dxdt| ≤ C5 ε(r, S) ≤ Cr6 .
Let us now turn to E13 . We compare this term with
Z ∞Z Z ∞Z  
E13b = − f (ũ(x, t)>u(x, t)) − f (ũ(x, t)⊥u(x, t)) ψ(x, t)
0 IR d 0 IR d
(v(y, s) − v(x, t)) · ∇ρd,r (x − y)ρ1,r (t − s) dxdtdyds.
Since div(v(·, s) − v(x, t)) = 0 (on IR d ) for all x ∈ IR d , t ∈ IR + and s ∈ IR + , one has E13b = 0. Therefore,
substracting E13b from E13 yields
Z ∞Z Z ∞Z
E13 ≤ C7 |u(x, t) − u(y, s)|ψ(x, t)
0 IR d 0 IR d (6.113) est8estb
|(v(y, s) − v(x, t)) · ∇ρd,r (x − y)|ρ1,r (t − s) dxdtdyds.
est8estb
The right hand side of (6.113) is then smaller than C8 ε(r, S), since |(v(y, s) − v(x, t)) · ∇ρd,r (x − y)| is
estubv
bounded by C9 rd (noting that |x − y| ≤ 1/r). Then, with (6.110), one has
C10
E13 ≤ . (6.114) est8estc
r
estkruzkovys
In order to estimate E14 , let us take in (6.105), for x ∈ IR d fixed, ϕ = ϕ(x, ·, ·), with
Z ∞
ϕ(x, y, s) = ψ(x, 0)ρd,r (x − y) ρ1,r (−τ )dτ,
s

and κ = u0 (x). Note that ϕ(x, ·, ·) ∈ Cc∞ (IR d × IR + , IR + ). We then integrate the resulting inequality
with respect to x ∈ IR d . We get

−E14 + E15 + E16 ≥ 0,

 187

with
Z ∞ Z Z Z ∞
E15 = − (f (u(y, s)>u0 (x)) − f (u(y, s)⊥u0 (x)))
0 IR d IR d s
v(y, s) · (ψ(x, 0)∇ρd,r (x − y))ρ1,r (−τ )dτ dydxds,
Z Z Z ∞
E16 = ψ(x, 0)ρd,r (x − y)ρ1,r (−τ )|u0 (x) − u0 (y)|dτ dydx.
IR d IR d 0
To bound E15 , one introduces E15b defined as
Z ∞Z Z Z ∞
E15b = (f (u(y, s)>u0 (y)) − f (u(y, s)⊥u0 (y)))
0 IR d IR d s
(v(y, s) · ∇ρd,r (x − y))ψ(x, 0)ρ1,r (−τ )dτ dydxds.
Integrating by parts for the x variable yields
Z ∞Z Z Z ∞
E15b = − (f (u(y, s)>u0 (y)) − f (u(y, s)⊥u0 (y)))
0 IR d IR d s
(v(y, s) · ∇ψ(x, 0))ρd,r (x − y)ρ1,r (−τ )dτ dydxds.
Then, noting that the time support of this integration is reduced to s ∈ [0, 1/r], one has
C11
E15b ≤ . (6.115) est9estd
r
Furthermore, one has

Z ∞ Z Z Z ∞
|E15 + E15b | ≤ C12 |u0 (x) − u0 (y)||v(y, s) · ∇ρd,r (x − y)|ψ(x, 0)ρ1,r (−τ )dτ dydxds,
0 IR d IR d s

which is bounded by C13 ε0 (r, S0 ), since the time support of the integration is reduced to s ∈ [0, 1/r],
where ε0 (r, S0 ) is defined by
Z
1
ε0 (r, S0 ) = sup{ |u0 (x) − u0 (x + η)|dx; |η| ≤ }. (6.116) estepsork
S0 r
bvtrans est9estd
Since u0 ∈ BV (IR d ), one has (thanks to Lemma 6.9) ε0 (r, S0 ) ≤ C14 /r and therefore, with (6.115),
E15 ≤ C15 /r.
bvtrans finvacances
Since u0 ∈ BV (IR d ), again thanks to Lemma 6.9, see remark 6.13, the term E16 is also bounded by
C16 /r.
Hence, since E14 ≤ E15 + E16 ,
C17
. E14 ≤ (6.117) est9estf
r
estcin est6est est7estb est8est est8estc est9estf
Using (6.106), (6.108), (6.111), (6.112),(6.114), (6.117), one obtains
Z Z h
|ũ(x, t) − u(x, t)|ψt (x, t) +
IR + IR d   i
f (ũ(x, t)>u(x, t)) − f (ũ(x, t)⊥u(x, t)) (v(x, t) · ∇ψ(x, t)) dxdt ≥
C18
−C1 (r + 1)µ(S) − C2 µ0 (S0 ) − r ,
p est3est
≤ 1 (r → ∞ if µ(S) = 0 and r = 1 if µ(S) > 1), gives (6.103).
which, taking r = 1/ µ(S) if 0 < µ(S) estlemma
This concludes the proof of the lemma 6.10.
 188

6.7.3 Proof of the error estimates

compthm estpbcont
estcondini
Let us now prove a quite general theorem of comparison between the entropy weak solution to ( 6.1)-(6.2)
esth14 esth14i
and an approximate solution, from which theorems 6.5 and 6.6 will be deduced.
estdonnees
esterrest Theorem 6.7 Under assumption 6.1, let u0 ∈ BV (IR d ) and ũ ∈ L∞ (IR d ×IR ?+ ) such that Um ≤ ũ ≤ UM
est1est
a.e. on IR d × IR ?+ . Assume that there exist µ ∈ M(IR d × IR + ) and µ0 ∈ M(IR d ) such that (6.102) holds.
estpbcont
estcondini
Let u be the unique entropy weak solution of (6.1)-(6.2) (note that u ∈ L∞ (IR d × IR ?+ ) is solution to
est1est
(6.102) with u instead of ũ and µ = 0, µ0 = 0).
Then, for all R > 0 and all T > 0 there exists Ce and R̄, only depending on R, T , v, f and u0 , such that
the following inequality holds:
RT R 1
0 B(0,R)
|ũ(x, t) − u(x, t)|dxdt ≤ Ce (µ0 (B(0, R̄)) + [µ(B(0, R̄) × [0, T ])] 2
+µ(B(0, R̄) × [0, T ])).
Recall that B(0, R) = {x ∈ IR d ; |x| < R}.
esterrest
Proof of Theorem 6.7
esterrest ewsunicite estlemma
The proof of Theorem 6.7 is close to that of Step
estlemma
2 in the proof of Theorem 6.3. It uses Lemma 6.10
sprelim
page 184, the proof of which is given in section 6.7.2 above. estdonnees
Let R > 0 and T > 0. One sets ω = V M , where V is given in Assumption 6.1 and M is the Lipschitz
constant of f in [Um , UM ] (indeed, since f ∈ C 1 (IR, IR), one has M = sup{|f 0 (s)|; s ∈ [Um , UM ]}).
Let ρ ∈ Cc1 (IR + , [0, 1]) be a function such that ρ(r) = 1 if r ∈ [0, R + ωT ], ρ(r) = 0 if r ∈ [R + ωT + 1, ∞)
and ρ0 (r) ≤ 0, for all r ∈ IR + (ρ only depends on R, T , v, f and u0 ).
est3est
One takes, in (6.103), ψ defined by

ψ(x, t) = ρ(|x| + ωt) TT−t , for x ∈ IR d and t ∈ [0, T ],
ψ(x, t) = 0, for x ∈ IR d and t ≥ T.
Note that ρ(|x| + ωt) = 1, if (x, t) ∈ B(0, R) × [0, T ].
The function ψ is not in Cc∞ (IR d × IR + , IR + ), but,est3est
using a usual regularization technique,
est3est
it may be
proved that such a function can be considered in (6.103), in which case Inequality (6.103) writes, with
R̄ = R + ωT + 1,
Z T Z h T − t 
1
ωρ0 (|x| + ωt) − ρ(|x| + ωt) +
|ũ(x, t) − u(x, t)|
0 IR d T  T i
T −t 0 x
f (ũ(x, t)>u(x, t)) − f (ũ(x, t)⊥u(x, t)) T ρ (|x| + ωt)(v(x, t) · |x| ) dxdt ≥
1
−C(µ0 (B(0, R̄)) + (µ(B(0, R̄) × [0, T ])) 2 + µ(B(0, R̄) × [0, T ])),
where C only depends on R, T , v, f and u0 .
Since ω = V M and ρ0 ≤ 0, one has
  T −t x 
f (ũ(x, t)>u(x, t)) − f (ũ(x, t)⊥u(x, t)) ρ0 (|x| + ωt)(v(x, t) · ) ≤
T |x|
|ũ(x, t) − u(x, t)| TT−t ω(−ρ0 (|x| + ωt)),
and therefore, since ρ(|x| + ωt) = 1, if (x, t) ∈ B(0, R) × [0, T ],

Z T Z
1
|ũ(x, t) − u(x, t)|dxdt ≤ CT (µ0 (B(0, R̄)) + (µ(B(0, R̄) × [0, T ])) 2 + µ(B(0, R̄) × [0, T ])).
0 B(0,R)
esterrest
This completes the proof of Theorem 6.7.
 189

esth14 esth14
Let us now conclude with the proofs
estschema
of theorems 6.5 page 180 (which gives an error estimate for the time
estschemaz estschemazesth14i esth14i
explicit numerical scheme (6.7), (estschemai
6.5) page 149) and 6.6 page 181 (which gives an error estimate estentest
estschemaz estschemaz
for the
time implicit
estentesti
numerical scheme
esterrest
( 6.9), ( 6.5) page 149). There are easy consequences of theorems 6.1 and
6.2 and of Theorem 6.7.
esth14
Proof of Theorem 6.5
esth14 estinftyest
Under the assumptions L∞ estimate on uT ,k (Lemma 6.1)
of Theorem 6.5, let ũ = uT ,k . Thanks to theesterrest
estentest
and to Theorem 6.1, ũ = uT ,k satisfies the hypotheses
estentest
of Theorem 6.7 with µ = µT ,k and µ0 = µT (the
measures µT ,k and µT are given in Theorem 6.1).
esterrest
Let R > 0 and T > 0. Then, Theorem 6.7 gives the existence of C1 and R̄, only depending on R, T , v,
f and u0 , such that
RT R 1
0 B(0,R)
|uT ,k (x, t) − u(x, t)|dxdt ≤ C1 (µT (B(0, R̄)) + [µT ,k (B(0, R̄) × [0, T ])] 2
(6.118) ouvaisje
+µT ,k (B(0, R̄) × [0, T ])).
estcfl
For h small enough, say h ≤ R0 , one has h < R̄ and k < T (thanks to condition 6.6, note that R0 only
depends on R, T , v, g, u0 , αestentest
and ξ). ouvaisje
Then, for h < R0 , Theorem 6.1 gives, with (6.118),
Z T Z √ 1 √ 1
|uT ,k (x, t) − u(x, t)|dxdt ≤ C1 (Dh + Ch 4 + C h) ≤ C2 h 4 ,
0 B(0,R)

where C2 only depends on R, T , esteqcorol

v, g, u0 , α and ξ.esth14
This gives the desired estimate (6.94) of Theorem 6.5 for h < R0 .
There remains the case h ≥ R0 . This case is trivial since, for h ≥ R0 ,

Z T Z
1 1
|uT ,k (x, t) − u(x, t)|dxdt ≤ 2 max{−Um , UM }m(B(0, R) × (0, T )) ≤ C3 (R0 ) 4 ≤ C3 h 4 ,
0 B(0,R)

for some C3 only depending on R, T , v, g, u , α and ξ.

esth14 0
This completes the proof of Theorem 6.5.
esth14i
Proof of Theorem 6.6
esth14i esth14 esth14
The proof of Theorem 6.6 is very similar to that of Theorem
esth14i
6.5 and we follow theestentest
estentesti
proof of Theorem 6.5.
6.6, using Theorem 6.2 instead of Theorem 6.1 gives that ũ = uT ,k
Under the assumptions of Theorem esterrest
satisfies the hypotheses
estentesti
of Theorem 6.7 with µ = µT ,k and µ0 = µT (the measures µT ,k and µT are given
in Theorem 6.2). esterrest
Let R > 0 and ouvaisje
T > 0. Theorem 6.7 gives the existence of C1 and R̄, only depending on R, T , v, f and
u0 , such that (6.118) holds. estentest ouvaisje
For h < R̄ and k < T Theorem 6.1 gives with (6.118),
Z T Z √ 1 1 1 1 1
|uT ,k (x, t) − u(x, t)|dxdt ≤ C1 (Dh + C(k + h 2 ) 2 + C(k + h 2 )) ≤ C2 (k + h 2 ) 2 ,
0 B(0,R)

where C2 only depends on R, T , esteqcoroli

v, g, u0 , α. esth14i
This gives the desired estimate (6.95) of Theorem 6.6 for h < R̄ and k < T .
There remains the cases h ≥ R̄ and k ≥ T . These cases are trivial since
Z T Z
1 1
|uT ,k (x, t) − u(x, t)|dxdt ≤ 2 max{−Um , UM }m(B(0, R) × (0, T )) ≤ C3 inf{R̄ 4 , T 2 }
0 B(0,R)

for some C3 only depending on R, T , v, g, u .

esth14i0
This completes the proof of Theorem 6.6.
 190

6.7.4 Remarks and open problems

esth14 esth14
Theorem 6.5 page 180 gives an error estimate of order h1/4 for the approximate solution of a nonlinear
hyperbolic equation
estschema estschemaz
div(vf (u)) = 0, with initial data in L∞ ∩ BV by
of the form ut +estschemaz the explicit
estcfl estcfl
finite
volume scheme (6.7) and (6.5) page 149, under a usual CFL condition k ≤ Ch (see (6.6) page 149).

Note that, in fact, the same estimate holds if u0 is only locally BV . More generally, if the initial data u0
is only in L∞ , then one still obtains an error estimate in terms of the quantities
Z
1 1
ε(r, S) = sup{ |u(x, t) − u(x + η, t + τ )|dxdt; |η| ≤ , 0 ≤ τ ≤ }
S r r
and
Z
1
ε0 (r, S0 ) = sup{ |u0 (x) − u0 (x + η)|dx; |η| ≤ },
S0 r
estepsrk estepsrk estepsork estepsork estentest
see (6.109) page 186
estentest
and (6.116) page 187. This is again an obvious consequence of Theorem 6.1 page
esterrest esterrest
166 and Theorem 6.7 page 188.

We also considered the implicit schemes, which seem to be much more widely used in industrial codes in
order to ensure their robustness. The implicit case required additional work in order
(i) to prove the existence of the solution to theestbvt
finite volume scheme,
(ii) to obtain the “strong time BV ” estimate
estexti
( 6.45) if v does not depend on t.
For v depending on t, Remark 6.12 yields an estimate of order h1/4 if k behaves as h; however, in the
esth14i
case where v does √ not depend on t, then an estimate of order h1/4 is obtained (in Theorem 6.6) for
√a
behaviour of k as h; Indeed, recent numerical experiments suggest that taking k of the order of h
yields results of the same precision than taking k of the order of h, with an obvious reduction of the
computational cost.

NoteChainais1
that the method described here may also be extended to higher order schemes for the same equation,
see Chainais-Hillairet [1996]; other methods have been used for error estimates for higher order
Noelle
schemes with a nonlinearity of the form F (u), as in Noëlle [1996]. However, it is still an open problem,
to our knowledge, to improve the order of the error estimate in the case of higher order schemes.

6.8 Nonlinear weak-? convergence

nlwscv eghbscewsol
The notion of nonlinear weak-? convergence was used in Section 6.6.3. We give here the definition of this
type of convergence and we prove that a bounded sequence of L∞ converges, up to a subsequence, in the
nonlinear weak-? sense.

nlwsdef Definition 6.3 (Nonlinear weak-? convergence)

Let Ω be an open subset of IR N (N ≥ 1), (un )n∈IN ⊂ L∞ (Ω) and u ∈ L∞ (Ω × (0, 1)). The sequence
(un )n∈IN converges towards u in the “nonlinear weak-? sense” if
Z Z 1 Z
g(un (x))ϕ(x)dx → g(u(x, α))ϕ(x)dxdα, as n → +∞,
Ω 0 Ω
(6.119) ews5
∀ϕ ∈ L1 (Ω), ∀g ∈ C(IR, IR).

nlwsib Remark 6.14 Let Ω be an open subset of IR N (N ≥ 1), (un )n∈IN ⊂ L∞ (Ω) and u ∈ L∞ (Ω × (0, 1))
such that (un )n∈IN converges towards u in the nonlinear weak-? sense. Then, in particular, the sequence
(un )n∈IN converges towards v in L∞ (Ω), for the weak-? topology, where v is defined by
Z 1
v(x) = u(x, α)dα, for a.e. x ∈ Ω.
0
 191

Therefore, the sequence (un )n∈IN is bounded in L∞ (Ω) (thanks to the Banach-Steinhaus theorem). The
following proposition gives that, up to a subsequence, a bounded sequence of L∞ (Ω) converges in the
nonlinear weak-? sense.

nlwsprop Proposition 6.4 Let Ω be an open subset of IR N (N ≥ 1) and (un )n∈IN be a bounded sequence of
L∞ (Ω). Then there exists a subsequence of (un )n∈IN , which will still be denoted by (un )n∈IN , and a
function u ∈ L∞ (Ω × (0, 1)) such that the subsequence (un )n∈IN converges towards u in the nonlinear
weak-? sense.

Proof
This proposition is classical in the framework of “Young measures” and we only sketch the proof for the
sake of completeness.
Let (un )n∈IN be a bounded sequence of L∞ (Ω) and r ≥ 0 such that kun kL∞ (Ω) ≤ r, ∀n ∈ IN.
Step 1 (diagonal process)
Thanks to the separability of the set of continuous functions defined from [−r, r] into IR (this set is
endowed with the uniform norm) and the sequential weak-? relative compactness of the bounded sets of
L∞ (Ω) , there exists (using a diagonal process) a subsequence, which will still be denoted by (u n )n∈IN ,
such that, for any function g ∈ C(IR, IR), the sequence (g(un ))n∈IN converges in L∞ (Ω) for the weak-?
topology towards a function µg ∈ L∞ (Ω).
Step 2 (Young measure)
In this step, we prove the existence of a family (mx )x∈Ω such that

1. for all x ∈ Ω, mx is a probability on IR whose support is included in [−r, +r] (i.e. mx is a σ-additive
application from the Borel σ-algebra of IR in IR + such that mx (IR) = 1 and mx (IR \ [−r, r]) = 0),
R
2. µg (x) = IR g(s)dmx (s) for a.e. x ∈ Ω and for all g ∈ C(IR, IR).

The family m = (mx )x∈Ω is called a “Young measure”.

Let us first claim that it is possible to define µg ∈ L∞ (Ω) for g ∈ C([−r, r], IR) by setting µg = µf where
f ∈ C(IR, IR) is such that f = g on [−r, r]. Indeed, this definition is meaningful since if f and h are two
elements of C(IR, IR) such that f = g on [−r, r] then µf and µh are the same element of L∞ (Ω) (i.e.
µf = µh a.e. on Ω) thanks to the fact that −r ≤ un ≤ r a.e. on Ω and for all n ∈ IN.
For x ∈ Ω, let
Z
1
Ex = {g ∈ C([−r, r], IR); lim µg (z)dz exists in IR},
h→0 m(B(0, h)) B(x,h)

where B(x, h) is the ball of center x and radius h (note that B(x, h) ⊂ Ω for h small enough).
If g ∈ Ex , we set
Z
1
µ̄g (x) = lim µg (z)dz.
h→0 m(B(0, h)) B(x,h)

Then, we define Tx from Ex in IR by Tx (g) = µ̄g (x). It is easily seen that Ex is a vector space which con-
tains the constant functions, that Tx is a linear application from Ex to IR and that Tx is nonnegative (i.e.
g(s) ≥ 0 for all s ∈ IR implies Tx (g) ≥ 0). Hence, using a modified version of the Hahn-Banach theorem,
one can prolonge Tx into a linear nonnegative application T x defined on the whole set C([−r, r], IR). By
a classical Riesz theorem, there exists a (nonnegative) measure mx on the Borel sets of [−r, r] such that
Z r
T x (g) = g(s)dmx (s), ∀g ∈ C([−r, r], IR). (6.120) ewsTbar
−r

all s ∈ [−r, r], the function g belongs to Ex and µ̄g (x) = 1 (note that µg = 1 a.e. on Ω).
If g(s) = 1 forewsTbar
Hence, from (6.120), mx is a probability over [−r, r], and therefore a probability over IR by prolonging it
by 0 outside of [−r, r]. This gives the first item on the family (mx )x∈Ω .
 192

Let us prove now the second item on the family (mx )x∈Ω . If g ∈ C([−r, r], IR) then grudin ∈ Ex for a.e.
x ∈ Ω and µg (x) = µ̄g (x) for a.e. x ∈ Ω (this is a classical result, since µg ∈ L1loc (Ω), see Rudin [1987]).
Therefore, µg (x) = Tx (g) = T x (g) for a.e. x ∈ Ω. Hence,
Z r
µg (x) = g(s)dmx (s) for a.e. x ∈ Ω,
−r

for all g ∈ C([−r, r], IR) and therefore for all g ∈ C(IR, IR). Finally, since the support of m x is included
in [−r, r],
Z
µg (x) = g(s)dmx (s) for a.e. x ∈ Ω, ∀g ∈ C(IR, IR).
IR
This completes Step 2.
Step 3 (construction of u)
It is well known that, if m̄ is a probability on IR, one has
Z Z 1
g(s)dm̄(s) = g(u(α))dα, ∀g ∈ Mb , (6.121) distfct
IR 0
where Mb is the set of bounded measurable functions from IR to IR and with

u(α) = sup{c ∈ IR; m̄((−∞, c)) < α}, ∀α ∈ (0, 1).

Note that the function u is measurable, nondecreasing and left continuous. Furthermore, if distfct the support
of m̄ is included in [a, b] (for some a, b ∈ IR, a < b) then u(α) ∈ [a, b] for all α ∈ (0, 1) and (6.121) holds
for all g ∈ C(IR, IR).
Applying this result to the measures mx leads to the definition of u as

u(x, α) = sup{c ∈ IR; mx ((−∞, c)) < α}, ∀α ∈ (0, 1), ∀x ∈ Ω.

For all x ∈ Ω, the function u(x, ·) is measurable (from (0, 1) to IR), nondecreasing, left continuous and
takes its values in [−r, r]. Furthermore,
Z 1
µg (x) = g(u(x, α))dα for a.e. x ∈ Ω, ∀g ∈ C(IR, IR).
0
Therefore,
Z Z Z 1
g(un (x))ϕ(x)dx → ( g(u(x, α))dα)ϕ(x)dx, as n → ∞,
Ω Ω 0
∀ϕ ∈ L1 (Ω), ∀g ∈ C(IR, IR).
nlwsprop
In order to conclude the proof of Proposition 6.4, there remains to show that modifying u on a negligible
set leads to a function (still denoted by u) measurable with respect to (x, α) ∈ Ω × (0, 1). Indeed, this
rudin
mesurability is needed in order to assert for instance, applying Fubini’s Theorem (see Rudin [1987]),
that
Z Z 1 Z 1 Z
( g(u(x, α))dα)ϕ(x)dx = ( g(u(x, α))ϕ(x)dx)dα,
Ω 0 0 Ω

for all ϕ ∈ L1 (Ω) and for all g ∈ C(IR, IR).

For all g ∈ C(IR, IR), one chooses for µg (which belongs to L∞ (Ω)) a bounded measurable function from
Ω to IR.
Let us define E = {ga,b ; a, b ∈ Q
l , a < b} where ga,b ∈ C(IR, IR) is defined by
 193

ga,b (x) = 1 if x ≤ a,
ga,b (x) = x−b
a−b if a < x < b,
ga,b (x) = 0 if x ≥ b.
Since E is a countable subset of C(IR, IR), there exists a Borel subset A of Ω such that m(A) = 0 and
Z
µg (x) = g(s)dmx (s), ∀x ∈ Ω \ A, ∀g ∈ E. (6.122) partout
IR

Define for all α ∈ (0, 1) v(., α) by

v(x, α) = 0 if x ∈ A,
v(x, α) = sup{c ∈ IR, mx ((−∞, c)) < α} if x ∈ Ω \ A,
so that u = v on (Ω \ A) × (0, 1) (and then u = v a.e. on Ω × (0, 1)).
Let us now prove that v is measurable from Ω × (0, 1) to IR (this will conclude the proof of Proposition
nlwsprop
6.4).
Since v(x, .) is left continuous on (0, 1) for all x ∈ Ω, proving that v(., α) is measurable (from Ω to IR)
rudin
for all α ∈ (0, 1) leads to the mesurability of v on Ω × (0, 1) (this is also classical, see Rudin [1987]).
There remains to show the mesurability of v(., α) for all α ∈ (0, 1).
Let α ∈ (0, 1) (in the following, α is fixed). Let us set w = v(., α) and define, for c ∈ IR,

fc (x) = mx ((−∞, c)) − α, x ∈ Ω \ A,

v(x, α) = w(x) = sup{c ∈ IR, fc (x) < 0} for all x ∈ Ω \ A.
so that partout
Using (6.122) leads to

mx ((−∞, c)) = sup{µg (x), g ≤ 1(−∞,c) and g ∈ E}, ∀x ∈ Ω \ A.

Then, the function fc : Ω \ A → IR is measurable as the supremum of a countable set of measurable
functions (recall that µg is measurable for all g ∈ E).
In order to prove the measurability of w (from Ω to IR), it is sufficient to prove that {x ∈ Ω\A; w(x) ≥ a}
is a Borel set, for all a ∈ IR (recall that w = 0 on A).
Let a ∈ IR, since fc (x) is nondecreasing with respect to c, one has

{x ∈ Ω \ A; w(x) ≥ a} = ∩n>0 {x ∈ Ω \ A; fa− n1 (x) < 0}.

Then {x ∈ Ω \ A; w(x) ≥ a} is measurable, thanks to the measurability of fc for all c ∈ IR.
nlwsprop
This concludes the proof of Proposition 6.4.

cvforte Remark 6.15 Let Ω be an open subset of IR N (N ≥ 1), (un )n∈IN ⊂ L∞ (Ω) and u ∈ L∞ (Ω × (0, 1))
such that (un )n∈IN converges towards u in the nonlinear weak-? sense. Assume that u does not depend
on α, i.e. there exists v ∈ L∞ (Ω) such that u(x, α) = v(x) for a.e. (x, α) ∈ Ω × (0, 1). Then, it is easy
to prove that (un )n∈IN converges towards u in Lp (B) for
ews5
all 1 ≤ p < ∞ and all bounded subset B of Ω.
Indeed, let B be a bounded subset of Ω. Taking, in (6.119), g(s) = s2 (for all s ∈ IR) and ϕ = 1B and
also g(s) = s (for all s ∈ IR) and ϕ = 1B v leads to
Z
(un (x) − v(x))2 dx → 0, as n → ∞.
B

This proves that (un )n∈IN converges towards u in L2 (B). The convergence of (un )n∈IN towards unlwsib
in Lp (B)
∞
for all 1 ≤ p < ∞ is then an easy consequence of the L (Ω) bound on (un )n∈IN (see Remark 6.14).
 194

6.9 A stabilized finite element method

fve
In this section, we shall try to compare the finite element method to the finite volume method for the
discretization of a nonlinear hyperbolic equation. It is well known that the use of the finite element is not
straightforward in the case of hyperbolic equations, since the lack of coerciveness of the operator yields
a lack of stability of the finite element scheme. There are
Sel
several techniques to stabilize these schemes,
which are beyond the scope of this work. Here, as in Selmin [1993], we are interested in viewing the
finite element as a finite volume flumono
method, flumono
by writing it in a conservative form, and using a stabilization
as in the third item of Example 5.2 page 132.

Let F ∈ C 1 (IR, IR 2 ), consider the following scalar conservation law:

ut (x, t) + div(F (u))(x, t) = 0, x ∈ IR 2 , t ∈ IR + , (6.123) claw

2
with an initial condition. Let T be a triangular mesh of IR , well suited for the finite element method. Let
S denote the set of nodes of this mesh, and let (φj )j∈S be the classical piecewise bilinear shape functions.
Following the finite element principles, let us look for an approximation of u in the space spanned by
the shape functions φj ; hence, at time tn = nk (where k is the time step), we look for an approximate
solution of the form X
u(., tn ) = unj φj ;
j∈S
claw P P
then, multiplying (6.123) by φi , integrating over IR 2 , approximating F ( j∈S unj φj ) by j∈S F (unj )φj
and using the mass lumping technique on the mass matrix yields the following scheme (with the explicit
Euler scheme for the time discretization):
Z X Z
un+1
i − uni n
φi (x)dx − F (uj ) · φj (x)∇φi (x)dx = 0,
k IR 2 IR 2
j∈S
Z Z X
which writes, noting that φj (x)∇φi (x)dx = − φi (x)∇φj (x)dx and that ∇φj (x) = 0,
j∈S
Z X Z
un+1 − uni
i
φi (x)dx + (F (uni ) + F (unj )) · φi (x)∇φj (x)dx = 0.
k IR 2 IR 2
j∈S

This last equality may also be written

Z X
un+1
i − uni
φi (x)dx + Ei,j = 0,
k IR 2 j∈S

where Z
1
Ei,j = (F (uni ) + F (unj )) · (φi (x)∇φj (x) − φj (x)∇φi (x))dx.
2 IR 2
Note that Ej,i = −Ei,j .
n
This is a centered and therefore unstable scheme. One way to stabilize it is to replace E i,j by
n n
Ẽi,j = Ei,j + Di,j (uni − unj ),
where Di,j = Dj,i (in order for the scheme to remain “conservative”) and Di,j ≥ 0 is chosen large enough
n
so that Ei,j is a nondecreasing function of uni and a nonincreasing function of unj , which ensure the
stability
fluxLF
of the scheme, underestext
fluxLF
a so called CFL condition, and does not change the “consistency” (see
estext
(5.27) page 132 and Remark 6.11 page 181).
 195

6.10 Moving meshes

For some evolution problems the use of time variable control volumes is advisable, e.g. when the domain
of study changes with time. This is the case, for instance, for the simulation of a flow in a porous medium,
when the porous medium is heterogeneous and its geometry changes with time. In this case, the mesh is
required to move with the medium. The influence of the moving mesh on the finite volume formulation
can be explained by considering the following simple transport equation:

ut (x, t) + div(uv)(x, t) = 0, x ∈ IR 2 , t ∈ IR + , (6.124) cetv

where v depends on the unknown u (and possibly on other unknowns). Let k be the time step, and set
tn = nk, n ∈ IN. Let T (t) be the mesh at time t. Since the mesh moves, the elements of the mesh vary
in time. For a fixed n ∈ IN, let R(K, t) be the domain of IR 2 occupied by the element K (K ∈ T (tn )) at
time t, t ∈ [tn , tn+1 ], that is R(K, tn ) = K. Let vs (x, t) be the velocity of the displacement of the mesh
at point x ∈ IR 2 and for all t ∈ [tn , tn+1 ] (note that vs (x, t) ∈ IR 2 ). Let unK and un+1
K be the discrete
unknowns associated to element K at times tn and tn+1 (they can be considered as the approximations of
the cetv
mean values of u(·, tn ) and u(·, tn+1 ) over R(K, tn ) and R(K, tn+1 ) respectively). The discretization
of (6.124) must take into account the evolution of the mesh in time. In order to do so, let us first consider
the following differential equation with initial condition:
∂y
(x, t) = −vs (y(x, t), t), t ∈ [tn , tn+1 ],
∂t (6.125) edy
y(x, tn ) = x.
Under suitable assumptions on vs (assume for instance that vs is continuous, Lipschitz continuous with
respect to its first
edy
variable and that the Lipschitz constant is integrable with respect to its second variable),
the problem (6.125) has, for all x ∈ IR 2 , a unique (global) edy
solution. For x ∈ IR 2 , define the function
y(x, ·) from [tn , tn+1 ] to IR as the solution of problem (6.125). Let (ϕp )p∈IN ⊂ Cc1 (IR 2 , IR + ) such that
2

≤ ϕp (x) ≤ 1 for x ∈ IR 2 and for all p ∈ IN, and such that ϕp → 1K a.e. as p → +∞. Multiplying
0cetv
(6.124) by ψp (x, t) = ϕp (y(x, t)) and integrating over IR 2 yields
Z  
∂(uψp )
(x, t) + u(x, t)∇ϕp (y(x, t)) · vs (y(x, t), t) − (uv)(x, t) · ∇ψp (x, t) dx = 0. (6.126) upsit
IR 2 ∂t
upsit
Using the explicit Euler discretization in time on Equation (6.126) and denoting by un (x) a (regular)
approximate value of u(x, tn ) yields
Z
1  n+1 
u (x)ψp (x, tn+1 ) − un (x)ψp (x, tn ) dx+
ZIR 2 k
un (x)(vs (x, tn ) − v(x, tn )) · ∇ϕp (x)dx = 0,
IR 2

which also gives (noting that ψp (x, t) = ϕp (y(x, t)))

Z
1  n+1 
u (x)ϕp (y(x, tn+1 )) − un (x)ϕp (y(x, tn )) dx−
IR 2 k Z (6.127) upsinb
div(un (vs − v))(x, tn ) · ϕp (x)dx = 0.
IR 2

Letting
upsinb
p tend to infinity and noting that 1K (y(x, tn )) = 1R(K,tn ) (x) and 1K (y(x, tn+1 )) = 1R(K,tn+1 ) (x),
(6.127) becomes
Z Z  Z
1 n+1 n
u (x)dx − u (x)dx + div((v − vs )un )(x, tn )dx = 0,
k R(K,tn+1 ) R(K,tn ) R(K,tn )

which can also be written

 196

1 n+1
(u m(R(K, tn+1 )) − unK m(R(K, tn )))+
Z k K
(v − vs )(x, tn ) · nK (x, tn )un (x)dγ(x) = 0,
∂R(K,tn )
R R
where unK = [1/m(R(K, tn ))] R(K,tn ) un (x)dx and un+1
K = [1/m(R(K, tn+1 ))] R(K,tn+1 ) un+1 (x)dx. Re-
call that nK denotes the normal to ∂K, outward to K. The complete discretization of the problem uses
some additional equations (on v, vs . . . ).

Remark 6.16 The above considerations concern a pure convection equation. In the case of a convection-
meshdirichlet
diffusion equation, such a moving mesh may become non-admissible in the sense of definitions 3.1 page
meshdirichlet
meshneumanmeshneuman
37 or 3.5 page 63. It is an interesting open problem to understand what should be done in that case.
 Chapter 7

Systems
systemes ellundhypmd
In chapters 2 to 6, the finite volume was successively investigated for the discretization of elliptic,
parabolic, and hyperbolic equations. In most scientific models, however, systems of equations have
to be discretized. These may be partial differential equations of the same type or of different types, and
they may also be coupled to ordinary differential equations or algebraic equations.
The discretization of systems of elliptic equations byellund
the finite volume method is straightforward, following
ellmd
the principles which were introduced in chapters 2 and 3. Examples of the performance of the finite
volume method for systems of elliptic equations on rectangular meshes, with “unusual” source terms
(in
Angot
particular, with source terms located on theFH1
edges or interfaces of the mesh) may be found in e.g.
Angot [1989] (see also references therein), Fiard, Herbin [1994] (where a comparison to a mixed
finite element formulation is also performed). Parabolic systems are treated similarly as elliptic systems,
with the addition of a convenient time discretization.
A huge literature is devoted to the discretization of hyperbolic systems of equations, in particular to
systems related to the compressible Euler equations, using structured or unstructured meshes. We shall
systemh
give only a short insight on this subject in Section 7.1, without any convergence result. Indeed, very few
godlewski-springer
theoretical results of convergence of numerical schemes are known on this subject. We refer to Godlewski
and Raviart [1996] and references therein for a more complete description of the numerical schemes for
hyperbolic systems.
Finite volume methods are also well adapted to the discretization of systems of equations of different
types (for instance, an elliptic
nseq
or parabolic
nseq
equation
flowpm
coupled
flowpm
with hyperbolic equations). Some examples
are considered in sections 7.2 page 208 and 7.3 page 212. The classical case of incompressible Navier-
Stokes (for which, generally, staggered grids are used) and examples which arise in the simulation of a
multiphase flow in a porous medium are described. The latter example also serves as an illustration of
how to deal with algebraic equations and inequalities.

7.1 Hyperbolic systems of equations

systemh
Let us consider a hyperbolic system consisting of m equations (with m ≥ 1). The unknown of the system
is a function u = (u1 , . . . , um )t , from Ω × [0, T ] to IR m , where Ω is an open set of IR d (i.e. d ≥ 1 is the
space dimension), and u is a solution of the following system:

Xd
∂ui ∂Gi,j
(x, t) + (x, t) = gi (x, t, u(x, t)),
∂t ∂xj (7.1) rtgeq
j=1
x = (x1 , . . . , xd ) ∈ Ω, t ∈ (0, T ), i = 1, . . . , m,
where
Gi,j (x, t) = Fi,j (x, t, u(x, t)),

197
 198

and the functions Fj = (F1,j , . . . , Fm,j )t (j = 1, . . . , d) and g = (g1 , . . . , gm )t are given functions from
Ω×[0, T ]×IR m (indeed, generally, a part of IR m , instead of IR m ) to IR m . The function F = (F1 , . . . , Fd ) is
assumed to satisfy the usual hyperbolicity condition, that is, for any (unit) vector of IR d , n, the derivative
of F · n with respect to its third argument (which can be considered as an m × m matrix) has only real
eigenvalues and is diagonalizable.
Note that in real applications, diffusion
rtgeq
terms may also be present in the equations, we shall omit them
here. In order to complete System (7.1), an initial condition for t = 0 and adequate boundary conditions
for x ∈ ∂Ω must be specified.
godroe
In the first section (Section 7.1.1), we shall only briefly describe the general method of discretization
by finite volume and some classical schemes. In the subsequent sections, some possible treatments of
difficulties appearing in real simulations will be given.

7.1.1 Classical schemes

godroe rtgeq
Let us first describe some classical finite volume schemes for the discretization of ( 7.1) with
hypmd
initial and
boundary conditions, using the concepts and notations which
meshypmd meshypmd
were introduced in chapter 6. Let T be an
admissible mesh in the sense of Definition 6.1 page 148 and k be the time step, which is assumed to be
constant (the generalization to a variable time step is easy). We recall that the interface, K|L, between
any two elements K and L of T is assumed to be included in a hyperplane of IR d . The discrete unknowns
are the unK , K ∈ T , n ∈ {0, . . . , Nk + 1}, with Nk ∈ IN, (Nk + 1)k = T . For K ∈ T , let N (K) be the
set of its neighbours, that is the set of elements L of T such that the (d − 1) Lebesgue measure of K|L is
positive. For L ∈ N (K), let nK,L be the unit normal vector to K|L oriented from K to L. Let tn = nk,
for n ∈ {0, . . . , Nk + 1}.
A finite volume scheme writes

un+1 − unK X
K n n
m(K) + m(K|L)FK,L = m(K)gK ,
k (7.2) rtgschema
L∈N (K)
K ∈ T , n ∈ {0, . . . , Nk },
where

1. m(K) (resp. m(K|L)) denotes the d (resp. d − 1) Lebesgue measure of K (resp. K|L),
n
2. the quantity gK , which depends on unK (or un+1
K or unK and un+1
K ), for K ∈ T , is some “consistent”
approximation of g on element K, between times tn and tn+1 (we do not discuss this approximation
here).
n
3. the quantity FK,L , which depends on the set of discrete unknowns unM (or un+1
M or unM and un+1
M )
for M ∈ T , is an approximation of F · nK,L on K|L between times tn and tn+1 .

In order to obtain a “good” scheme, this approximation of F · nK,L has to be consistent, conservative
n n
(that is FK,L =-FL,K ) and must ensure some stability properties on the approximate solution given by
the scheme (indeed, one also needs some consistency with respect to entropies, when entropies exist. . . ).
Except in the scalar case, it is not so easy to see what kind of stability properties is needed. . . . Indeed, in
the scalar case, that is m = 1, taking g = 0 and Ω = IR d (for simplicity), it is essentially sufficient to have
an L∞ estimate (that is a bound on unK independent of K, n, and of the time and space discretizations)
hyper1dhypmd Chainais1
and a “touch” of “BV estimate” (see, for instance, chapters 5 and 6 and Chainais-Hillairet [1996] for
more precise assumptions). In the case m > 1, it is not generally possible to give stability properties from
which a mathematical proof of convergence could be deduced. However, it is advisable to require some
stability properties such as the positivity of some quantities depending on the unknowns; in the case of
flows, the required stability may be the positivity of the density, energy, pressure. . . ; the positivity of
these quantities may be essential for the computation of F (u) or for its hyperbolicity.
 199

n
The computation of FK,L is often performed, at each “interface”, by solving the following 1D (for the
space variable) system (where, for simplicity, the possible dependency of F with respect to x and t is
omitted):

∂u ∂fK,L (u)
(z, t) + (z, t) = 0, (7.3) fnkl
∂t ∂z
where fK,L (u)(z, t) = F ·nK,L (u(z, t)), for all z ∈ IR and trtgschema
∈ (0, T ), which gives consistency, conservativity
(and, hopefully, stability) of the final scheme (that is (7.2)). To be more precise, in the casefnkl of lower
n n
order schemes, FK,L may be taken as: FK,L = F.nK,L (w) where w is the solution for z = 0 of (7.3) with
initial conditions u(x, 0) = unK if x < 0 and u(x, 0) = unL if x > 0. Note that the variable
hypmd
z lies in IR, so
that the multidimensional problem has therefore been transformed (as in chapter 6) into a succession of
one-dimensional problems. Hence, in the following, we shall mainly keep to the case d = 1.

Let us describe two classical schemes, namely the Godunov scheme and the Roe rtgeq
scheme, in the case
d = 1, Ω = IR, F (x, t, u) = F (u) and g = 0 (but m ≥ 1), in which case System (7.1) becomes

∂u ∂F (u)
(x, t) + (x, t) = 0, x ∈ IR, t ∈ (0, T ). (7.4) rtgeqs
∂t ∂x
in order to complete this system, an initial condition must be specified, the discretization of which is
standard. meshhyp1d meshhyp1d
Let T be an admissible mesh in the sense of Definition 5.5 page 125, that is T = (Ki )i∈ZZ , with
Ki =(xi−1/2 ,xi+1/2 ), with xi−1/2 < xi+1/2 , i ∈ ZZ . One sets hi = xrtgschema
i+1/2 − xi−1/2 , i ∈ ZZ . The dis-
crete unknowns are uni , i ∈ ZZ , n ∈ {0, . . . , Nk + 1} and the scheme (7.2) then writes

un+1
i − uni n n
1 − F
hi
k
+ Fi+
2 i− 21 = 0, i ∈ ZZ , n ∈ {0, . . . , Nk }, (7.5) rtgschemas
n
where Fi+1/2 is a consistent approximation of F (u(xi+1/2 , tn ). This scheme is clearly conservative (in the
n
sense defined above). Let us consider explicit schemes, so that Fi+1/2 is a function of unj , j ∈ ZZ . The
Go n
principle of the Godunov scheme Godunov [1976] is to take Fi+1/2 = F (w) where w is the solution, for
x = 0 (and any t > 0), of the following (Riemann) problem

∂u ∂F (u)
(x, t) + (x, t) = 0, x ∈ IR, t ∈ IR + , (7.6) rtgrp
∂t ∂x

u(x, 0) = uni , if x < 0,

(7.7) rtgbcrp
u(x, 0) = uni+1 , if x > 0.
Then, w depends on uni , uni+1 and F .
The time step is limited by the so called “CFL condition”, which writes k ≤ Lhi , for all i ∈ ZZGo, where L
is given by F and the initial condition. The quantity un+1
i , given by the Godunov scheme,
rtgeqs
see Godunov
[1976], is, for all i ∈ ZZ , the mean value on Ki of the exact solution at time k of (7.4) with the initial
condition (at time t = 0) u0 defined, a.e. on IR, by u0 (x) = uni if xi−1/2 < x < xi+1/2 .

The Godunov scheme is an efficient scheme (consistent, conservative, stable), sometimes too diffusive
(especially if k is far from Lhi defined above), but easy improvements are possible, such as the MUSCL
hyperhos
technique, see below and Section 5.4. Its principal drawback is its difficult implementation for many
problems, indeed the computation of F (w) can be impossible or too expensive. For instance, this com-
putation may need a non trivial parametrization of the non linear waves. Note also that F is generally
not given directly as a function of u (the components of u are called “conservative unknowns”) but as
a function of some “physical” unknowns (for instance, pressure, velocity, energy. . . ), and the passage
from u to these physical unknowns (or the converse) is often not so easy. . . it may be the consequence of
expensive and implicit calculations, using, for instance, Newton’s algorithm.
 200

hlvl
Due to this difficulty of implementation, some “Godunov type” schemes were developed (see Harten,
Lax and Van Leer [1983]).rtgeqs
The idea is to take, for un+1
i , the mean value on Ki of an approximate
solution at time k of (7.4) with the initial condition (at time t = 0), u0 , defined by u0 (x) = uni , if
rtgschemas
xi−1/2 < x < xi+1/2 . In order for the scheme to be written under the conservative form (7.5), with a
consistent approximation of the fluxes, this approximate solution must satisfy some consistency relation
(another relation is needed for the consistency
roe
with entropies).
roe2
One of the best known of this family
of schemes is the Roe scheme (see Roe [1980] and Roe [1981]), where this approximate solution is
computed by the solution of the following linearized Riemann problems:

∂u(x, t) ∂u(x, t)
+ A(uni , uni+1 ) = 0, x ∈ IR, t ∈ IR + , (7.8) rtglrp
∂t ∂x

u(x, 0) = uni , if x < 0,

(7.9) rtgbclrp
u(x, 0) = uni+1 , if x > 0,
where A(·, ·) is an m×m matrix, continuously depending on its two arguments, with only real eigenvalues,
diagonalizable and satisfying the so called “Roe condition”:

A(u, v)(u − v) = F (u) − F (v), ∀u, v ∈ IR m . (7.10) rtgrc

rtgrc rtgschemas
Thanks to (7.10), the Roe scheme can be written as (7.5) with
n n − n n n n
Fi+ 1 = F (ui ) + A (ui , ui+1 )(ui − ui+1 )
2 (7.11) rtgfr
(= F (uni+1 ) + A+ (uni , uni+1 )(uni − uni+1 )),
where A± are the classical nonnegative and nonpositive parts of the matrix A: let A be a matrix with
only real eigenvalues, (λp )p=1,...,m , and diagonalizable, let (ϕp )p=1,...,m be a basis of IR m associated to
these eigenvalues. Then, the matrix A+ is the matrix which has the same eigenvectors as A and has
(max{λp , 0})p=1,...,m as corresponding eigenvalues. The matrix A− is (−A)+ .
Roe’s scheme was proved to be an efficient scheme, often less expensive than Godunov’s scheme, with,
more or less the same limitation on the time step, the same diffusion effect and some lack of entropy
consistency, which can be corrected. It has some properties of consistency
rtgrc
and stability. Its main drawback
is the difficulty of the computation of a matrix A(u, v) satisfying (7.10). For instance, when it is possible
to compute and diagonalize the derivative of F , DF (u), one can take A(u, v) = DF (u ? ), but the difficulty
rtgrc
?
is to find u such that (7.10) holds (note that this condition is crucial in order to ensure conservativity
of Roe’s scheme). In some difficult cases, the Roe matrix is computed approximately by using a “limited
expansion” with respect to some small parameter.

7.1.2 Rough schemes for complex hyperbolic systems

rough
The aim of this section is to present some discretization techniquesrtgeq
for “complex” hyperbolic systems.
In many applications, the expressions of g and F which appear in (7.1) are rather “complex”, and it is
difficult or impossible to use classical schemes such as the 1D Godunov or Roe schemes or their standard
extensions, for multidimensional problems, using 1D solvers on the interfaces of the mesh. This is the
case of gas dynamics (Euler equations) with real gas, for which the state law (pressure as a function of
density and internal energy) is tabulated or given by some complex analytical expressions. This is also
the case when modelling multiphase flows in pipe-lines: the function F is difficult to handle and highly
depends on x and u, because, for instance, of changes of the geometry and slope of the pipe, of changes
of the friction law or, more generally, of the varying nature of the flow. Most of the attempts given
below were developed for this last situation. Other interesting cases of “complexity” are the treatment of
hypbc
boundary conditions (mathematical literature is rather scarce on this subject, see Section 7.1.4 for a first
insight), and the way to handle the case where the eigenvalues (of the derivative of F · n with respect to
turbo
its third argument) are of very different magnitude, see Section 7.1.3. Another case of complexity is the
 201

h2
treatment of nonconservative terms in the equations. One refers, for instance, to Brun, Hérard, Leal
De Sousa and Uhlmann [1996] and references therein, for this important case.

Possible modifications of Godunov and Roe schemes (including “classical” improvements to avoid ex-
cessive artificial diffusion) are described now to handle “complex” systems. Because of the complexity
of the models, the justification of the schemes presented here is rather numerical than mathematical.
Many variations have also Ghidaglia
been developed, which are not presented here. Note that other approaches
are also possible, see e.g. Ghidaglia, Kumbaro and Le Coq [1996]. For simplicity,godroe one considers
the case d = 1, Ω = IR, F (x, t, u) = F (u) and g = 0 (but m ≥ 1) described in Sectionrtgschemas
7.1.1, with the
n
same notations. The Godunov and Roe schemes can both be written under the form (7.5) with Fi+1/2
godroe
computed as a function of uni and uni+1 ; both schemes are consistent (in the sense of Section 7.1.1, i.e.
n
consistency of the “fluxes”) since Fi+1/2 = F (u) if uni = uni+1 = u.

Going further along this line of thought yields (among other possibilities, see below) the “VFRoe” scheme
rtgschemas
which is (7.5), that is:

un+1
i − uni n n
1 − F
hi
k
+ Fi+
2 i− 21 = 0, i ∈ ZZ , n ∈ {0, . . . , Nk }, (7.12) rtgvfroe1

n
rtglrp rtgbclrp
with Fi+1/2 = F (w), where w is the solution of the linearized Riemann problem (7.8), (7.9), with
A(ui , uni+1 ) = DF (w? ), that is:
n

∂u(x, t) ∂u(x, t)
+ DF (w? ) = 0, x ∈ IR, t ∈ IR + , (7.13) rtgvfroe2
∂t ∂x

u(x, 0) = uni , if x < 0,

(7.14) rtgvfroe3
u(x, 0) = uni+1 , if x > 0,
where w? is some value between uni and uni+1 (for instance, w ? = (1/2)(uni + uni+1 )). In this scheme, the
rtgrc
Roe condition (7.10) is not required (note that it is naturally conservative, thanks to its finite volume
origin). Hence, the VFRoe scheme appears to be a simplified version of the Godunov and Roe schemes.
The study of the scalar case (m = 1) shows that, in order to have some stability, at least as much as
in Roe’s scheme, the choice of w ? is essential. In practice, the choice w ? = (1/2)(uni + uni+1 ) is often
adequate, at least for regular meshes.
rtgrc
Remark 7.1 In Roe’s scheme, the Roe condition (7.10) ensures conservativity. The VFRoe scheme is
“naturally” conservative, and therefore no such condition is needed. Also note that the VFRoe scheme
yields precise approximations of the shock velocities, without Roe’s condition.

Numerical tests show the good behaviour of the VFRoe scheme. Its two main flaws are a lack of entropy
consistency (as in Roe’s scheme) and a large diffusion effect (as in the Godunov and Roe schemes). The
first drawback
hh
can be corrected, as for Roe’s scheme,
mfg
with a nonparametric entropy correction inspired
from Harten, Hyman and Lax [1976] (see Masella, Faille, and Gallouët [1996]). rtgbclrp The two
drawbacks
rtgbclrp
can be corrected with a classical MUSCL technique, which consists in replacing, in ( 7.9) page
200, uni and uni+1 by uni+1/2,− and uni+1/2,+ , which depend on {unj , j = i−1, i, i+1, i+2} (see, for instance,
hyperhos hyperhos godlewski-springer leveque
Section 5.4 page 143 and Godlewski and Raviart [1996] or LeVeque [1990]). For stability reasons,
the computation of the gradient of the unknown (cell by cell) and of the “limiters” is performed on some
“physical” quantities (such as density, pressure, velocity for Euler equations) instead of u. The extension
of the MUSCL technique to the case d > 1 is more or less straightforward.

This MUSCL technique improves the space accuracy (in the truncation error) and the numerical results
are significantly better. However, stability is sometimes lost. Indeed, considering the linear scalar equa-
tion, one remarks that the scheme is antidiffusive when the limiters are not active, this might lead to a
loss of stability. The time step must then be reduced (it is reduced by a factor 10 in severe situations. . . ).
 202

In order to allow larger time steps, the time accuracy should be improved by using, for instance, an
order 2 Runge-Kutta scheme (in the severe situations suggested above, the time step is then multiplied
by a factor 4). Surprisingly, this improvement of time accuracy is used to gain stability rather than
precision. . .
mfg
Several numerical experiments (see Masella, Faille, and Gallouët [1996]) sod
were performed which
prove the efficiency of the VFRoe scheme, such as the classical Sod tests (Sod [1978]). The shock
velocities are exact, there are no oscillations. . . . For these tests, the treatment of the boundary conditions
is straightforward. Throughout these experiments, the use of a MUSCL technique yields a significant
improvement, while the use of a higher order time scheme is not necessary. In one of the Sod tests, the
entropy correction is needed.
A comparison between the VFRoe scheme and the Godunov scheme was performed by J. M. Hérard
(personal
rtgrc
communication) for the Euler equations on a Van Der Wals gas, for which a matrix satisfying
(7.10) seems difficult to find. The numerical results are better with the VFRoe schem, which rtgvfroe2
is also rtgvfroe3
much
cheaper computationally. An improvment of the VFRoe scheme is possible, using, instead of ( 7.13)-(7.14),
linearized
rtgeqs
Riemann problems associated
rtgeq
to a nonconservative form of the initial system, namelyrtgvfroe1
System
n
(7.4) or more generally System (7.1), for the computation of w (which gives the flux Fi+1/2 in (7.12) by
n
bgh
the formula Fi+1/2 = F (w)), see for instance Buffard, Gallouët and Hérard [1998] for a simple
example.
In some more complex cases, the flux F may also highly, and not continuously, depend on the space
variable x. In the space discretization, it is “natural” to set the discontinuities of F with respect to x on
the boundaries of the mesh. The function F may change drastically from Ki to Ki+1 . In this case, the
implementation of the VFRoe scheme yields two additional difficulties:
rtglrp rtgbclrp
(i) The matrix A(uni , uni+1 ) in the linearized Riemann problem (7.8), (7.9) now depends on x:
A(uni , uni+1 ) = Du F (x, w? ), where w? is some value between uni and uni+1 and Du F denotes the
derivative of F with respect to its “u” argument.
rtglrprtgbclrp
(ii) once the solution, w, of the linearized problem (7.8) (7.9), for x = 0 and any t > 0, is calculated,
n
the choice Fi+1/2 = F (x, w) again depends on x.
n n
The choice of Fi+1/2 (point (ii)) may be solved by remarking that, in Roe’s scheme, Fi+1/2 may be written
rtgrc
(thanks to (7.10)) as

n 1 1
Fi+ 1 = (F (uni ) + F (uni+1 )) + Ani+ 1 (uni − uni+1 ), (7.15) rtgfrc
2 2 2 2

where Ani+1/2 = |A(uni , uni+1 )|, and |A| = A+ + A− .

rtgfrc
Under this form, the second term of the right hand side of (7.15) appears to be a stabilization term,
which does not affect the consistency. Indeed, in the scalar case (m = 1), one has A ni+1/2 = |F (uni ) −
F (uni+1 )|/|uni − uni+1 |, which easily yields the L∞ stability of the scheme (but not the consistency with
respect to the entropies). Moreover, the scheme is stable and consistent with respect to the entropies,
n
under a Courant-Friedrichs-Levy (CFL) condition, if Fi+1/2 is nondecreasing with respect to uni and
nonincreasing with respect to ui+1 , which holds if Ai+1/2 ≥ sup{|F 0 (s)|, s ∈ [uni , uni+1 ] or [uni+1 , uni ]}.
n n

This remark rtgvfroe1

suggests a slightly different version of the VFRoescheme (closer to Roe’s scheme), which is
rtgvfroe3
the scheme (7.12)-(7.14), taking

n 1 1
Fi+1/2 = (F (uni ) + F (uni+1 )) + |DF (w? )|(uni − uni+1 ),
2 2
rtgvfroe1 n
in (7.12), instead of Fi+1/2 = F (w). Note that it is also possible to take other convex combinations of
F (uni ) and F (uni+1 ) in the latter expression of Fi+1/2
n
, without modifying the consistency of the scheme.

When F depends rtgfrc

on x, the discontinuities of F being on the boundaries of the control volumes, the
generalization of (7.15) is obvious, except for the choice of Ani+1/2 . The quantity F (uni ) is replaced by
 203

F (xi , uni ), where xi is the center of Ki . Let us now turn to the choice of a convenient matrix Ani+1/2 for
this modified VFRoe scheme, when F highly depends on x. A first possible choice is

Ani+1/2 = (1/2)(|Du F (xi , uni )| + |Du F (xi+1 , uni+1 )|).

The following slightly different choice for Ani+1/2 seems, however, to give better numerical results (see
fh
Faille and Heintzé [1999]). Let us define

Ai = Du F (xi , uni ), ∀ i ∈ ZZ
(i)
(for the determination of Ani+1/2 the fixed index n is omitted). Let (λp )p=1,...,m be the eigenvalues of Ai
(i) (i) (i)
(with λp−1 ≤ λp , for all p) and (ϕp )p=1,...,m a basis of IR m associated to these eigenvalues. Then, the
(−) (+)
matrix Ai+1/2 [resp. Ai+1/2 ] is the matrix which has the same eigenvectors as Ai [resp. Ai+1 ] and has
(i) (i+1)
(max{|λp |, |λp |})p=1,...,m as corresponding eigenvalues. The choice of Ani+1/2 is
λ (−) (+)
Ani+ 1 =
(A 1 + Ai+ 1 ), (7.16) rtggroomf
2 i+ 2 2 2

where λ is a parameter, the “normal” value of which is 1. Numerically, larger values of λ, say λ = 2 or
λ = 3, are sometimes needed, in severe situations, to obtain enough stability. Too large values of λ yield
too much artificial diffusion.
rtgvfroe1
rtgvfroe3
The new scheme is then (7.12)-(7.14), taking

n 1  1
Fi+1/2 = F (xi , uni ) + F (xi , uni+1 ) + Ani+ 1 (uni − uni+1 ). (7.17) rtgfrcb
2 2 2

rtggroomf
where Ani+1/2 is defined by (7.16). It has, more or less, the same properties as the Roe and VFRoe schemes
but allows the simulation of more complex systems. It needs a MUSCL technique to reduce diffusion
effects and order 2fhRunge-Kutta for stability. It was implemented for the simulation of multiphase flows
in pipe lines (see Faille and Heintzé [1999]). The other difficulties encountered in this case are the
treatment of the boundary conditions and the different magnitude of the eigenvalues, which are discussed
in the next sections.

7.1.3 Partial implicitation of explicit scheme

turbo
In the modelling of flows, where “propagation” phenomena and “convection” phenomena coexist, the
Jacobian matrix of F often has eigenvalues of different magnitude, the “large” eigenvalues (large meaning
“far from 0”, positive or negative) corresponding to the propagation phenomena and “small” eigenvalues
corresponding to the “convection” phenomena . Large and small eigenvalues may differ by a factor 10 or
100.

With the explicit schemes described in the previous sections, the time step is limited by the CFLgodroe
condition
corresponding to the large eigenvalues. Roughly speaking, with the notations of Section 7.1.1, this
condition is (for all i ∈ ZZ ) k ≤ |λ|−1 hi , where λ is the largest eigenvalue. In some cases, this limitation
can be unsatisfactory for two reasons. Firstly, the time step is too small and implies a prohibitive
computational cost. Secondly, the discontinuities in the solutions, associated to the small eigenvalues,
are not sharp because the time step is far from the CFL condition of the small eigenvalues (however,
this can be somewhat corrected with a MUSCL method). This is in fact a major problem when the
discontinuities associated to the small eigenvalues need to be computed precisely. It is the case of interest
here.

A first method to avoid the time step limitation is to take a “fully implicit” version of the schemes
n
developed in the previous sections, that is Fi+1/2 function of un+1
j , j ∈ ZZ , instead of unj , j ∈ ZZ (the
 204

fer
terminology “fully implicit” is by opposition to “linearly implicit”, see below and Fernandez [1989]).
However, in order to be competitive with explicit schemes, the fully implicit scheme is used with large
time steps. In practice, this prohibits the use of a MUSCL technique in the computation of the solution
at time tn+1 by, for instance, a Newton algorithm. This implicit scheme is therefore very diffusive and
will smear discontinuities.
A second method consists in splitting the system into two systems, the first one is associated with the
“small” eigenvalues, and the second one with the “large” eigenvalues (in the case of the Euler equations,
this splitting may correspond to a “convection” system and a “propagation” system). At each time step,
the first system is solved with an explicit scheme and the second one with an implicit scheme. Both use
the same time step, which is limited by the CFL condition of the small eigenvalues. Using a MUSCL
technique and an order 2 Runge-Kutta method for the first system yields sharp discontinuities associated
to the small eigenvalues. This method is often satisfactory, but is difficult to handle in the case of
severe boundary conditions, since the convenient boundary conditions for each system may be difficult
to determine. tur
Another method, developed by E. Turkel (see Turkel [1987]), in connexion with Roe’s scheme, uses a
change of variables in order to reduce the ratio between large and small eigenvalues.

Let us now describe a partially linearly fh

implicit method (“turbo” scheme) which was successfully
fer
tested
for multiphase flows in pipe lines (see Faille and Heintzé [1999]) and other cases (seerough Fernandez
[1989]). For the sakertgvfroe1
of simplicity,
rtgvfroe3
the method is described for the last rtggroomf
rtgfrcb
scheme of Section 7.1.2, i.e. the
n
scheme defined by (7.12)- (7.14), where Fi+ 1 is defined by (7.17) and (7.16) (recall that F may depend
2
on x).

Assume that I ⊂ {1, . . . , m} is the set of index of large eigenvalues (and does not depend on i). The aim
(−) (+)
here is to “implicit” the unknowns coresponding to the large eigenvalues only: let Ãi , Ãi+1/2 and Ãi+1/2
(−) (+)
be the matrix having the same eigenvectors as Ai , Ai+1/2 and Ai+1/2 , with the same large eigenvalues
(i.e. corresponding to p ∈ I) and 0 as small eigenvalues. Let
(−) (+)
Ãni+1/2 = (λ/2)(Ãi+1/2 + Ãi+1/2 ).
n
rtgschemas
n
Then, the partially linearly implicit scheme is obtained by replacing Fi+1/2 in (7.5) by F̃i+1/2 defined by
n
F̃i+ 1 = F
n
i+ 1
+ 12 (Ãi (un+1
i − uni ) + Ãi+1 (un+1 n
i+1 − ui+1 ))
2 2
+ 21 Ãni+ 1 (un+1
i − uni + uni+1 − un+1
i+1 ).
2

In order to obtain sharp discontinuities corresponding to the small eigenvalues, a MUSCL technique is
n
used for the computation of Fi+1/2 . Then, again for stability reasons, it is preferable to add an order
2 Runge-Kutta method for the time discretization. Although it is not so easy to implement, the order
2 Runge-Kutta method is needed to enable the use of “large” time steps. The time step is, in severe
situations, very close to that given by the usual CFL condition corresponding fh
to the small eigenvalues,
and can be considerably larger than that given by the large eigenvalues (see Faille and Heintzé [1999]
for several tests).

7.1.4 Boundary conditions

hypbc
In many simulations of real situations, the treatment of the boundary conditions is not easy (in kum
particular
in the case dlf
of sign change of eigenvalues). We give here a classical possible mean (see e.g. Kumbaro
[1992] and Dubois and LeFloch [1988]) of handling boundary conditions (a more detailed description
jmm
may be found in Masella [1997] for the case of multiphase flows in pipe lines).
rtgeqs
Let us consider now the system (7.4) where “x ∈ IR” is replaced by “x ∈ Ω” with Ω = (0, 1). In order for
the system to be well-posed, an initial condition (for t = 0) and some convenient boundary conditions
 205

for x = 0 and x = 1 are needed; these boundary conditions will appear later in the discretization (we do
not detail
serre
here the mathematical analysis of the problem of the adequacy of the boundary conditions, see
e.g. Serre [1996] and references therein). Let us now explain the numerical treatment of the boundary
condition at x = 0.
godroe PNT
With the notations of Section 7.1.1, the space mesh is given by {Ki , i ∈ {0, . . . , NT }}, with i=1
rtgschemas
hi = 1.
Using the finite volume scheme (7.5) with i ∈ {1, . . . , NT } instead of i ∈ ZZ needs, for the computation
of un+1
1 , with {uni , i ∈ {1, . . . , NT }} given, a value for F1/2
n
(which corresponds to the flux at point x = 0
and time t = tn ).

For the sake of simplicity, consider only the case of the Roe and VFRoe schemes. Then, the “interior
n
fluxes”, that is Fi+1/2 for i ∈ {1, . . . , NT − 1}, are determined by using matrices A(uni , uni+1 ) (i ∈
n
rtgfr rtgfrc
{1, . . . , NT − 1}). In the case of the Roe scheme, Fi+1/2 is given by (7.11) or (7.15) and A(·, ·) satisfies
rtgrc n
the Roe condition (7.10). In the case of the VFRoe scheme, Fi+1/2 is given through the resolution of
rtglrp rtgbclrp
the linearized Riemann problem (7.8), (7.9) with e.g. A(ui , ui+1 ) = DF ((1/2)(uni + uni+1 )). In order
n n
n
to compute F1/2 , a possibility is to take the same method as for the interior fluxes; this requires the
determination of some un0 . In some cases (e.g. when all the eigenvalues of Du F (u) are nonnegative), the
given boundary conditions at x = 0 are sufficient to determine the value un0 , or directly F1/2 n
, but this is
not true in the general case. . . . In the general case, there are not enough given boundary conditions to
determine un0 and missing equations need to be introduced. The idea is to use an iterative process. Since
A(un0 , un1 ) is diagonalizable and has only real eigenvalues, let λ1 , . . . , λm be the eigenvalues of A(un0 , un1 )
and ϕ1 , . . . , ϕm a basis of IR m associated to these eigenvalues. Then the vectors un0 and un1 may be
decomposed on this basis, this yields
m
X m
X
un0 = α0,i ϕi , un1 = α1,i ϕi .
i=1 i=1

Assume that the number of negative eigenvalues of A(un0 , un1 ) does not depend on un0 (this is a simplifying
assumption); let p be the number of negative eigenvalues and m − p the number of positive eigenvalues
of A(un0 , un1 ).
Then, the number of (scalar) given boundary conditions is (hopefully . . . ) m − p. Therefore, one takes,
for un0 , the solution of the (nonlinear) system of m (scalar) unknowns, and m (scalar) equations. The
m unknowns are the components of un0 and the m equations are obtained with the m − p boundary
conditions and the p following equations:

α0,i = α1,i , if λi < 0. (7.18) rtgbc

Note that the quantities α0,i depend on A(un0 , un1 ); the resulting system is therefore nonlinear and may
be solved with, for instance, a Newton algorithm.

Other possibilities around this method are possible. For instance, another possibility, perhaps rtgbc
more
natural, consists in writing the m − p boundary conditions on un1/2 instead of un0 and to take (7.18) with
rtglrp rtgbclrp
the components of un1/2 instead of those of un0 , where un1/2 is the solution at x = 0 of (7.8), (7.9) with
n
i = 0. With the VFRoe scheme, the flux at the boundary x = 0 is then F1/2 = F (un1/2 ). In the case of a
linear system with linear boundary conditions and with the VFRoe scheme, this method gives the same
n
flux F1/2 as the preceding method, the value un1/2 is completely determined although un0 is not completely
determined.
rough
In the case of the scheme described in the second part of Section 7.1.2, the following “simpler”
rtgfrc
possibility
rtggroomf
n
was implemented. For this scheme, Fi+1/2 is given, for i ∈ {1, . . . , NT − 1}, by (7.15) with (7.16). Then,
n
the idea is to take the same equation for the computation of F1/2 but to compute un0 as above (that is
rtgbc
with m − p boundary conditions and (7.18)) with the choice A(un0 , un1 ) = Du F (x1 , un1 ).
 206

This method of computation of the boundary fluxes gives good results but is not adapted to all cases
(for instance, if p changes during the Newton iterations or if the number of boundary conditions is not
equal to m − p. . . ). Some particular methods, depending on the problems under consideration, have to
be developped.

We now give an attempt for the justification of this treatment of the boundary conditions, at least for a
linear system with linear boundary conditions.
Consider the system

ut (x, t) + ux (x, t) = 0, x ∈ (0, 1), t ∈ IR + ,

(7.19) sl
vt (x, t) − vx (x, t) = 0, x ∈ (0, 1), t ∈ IR + ,
with the boundary conditions

u(0, t) + αv(0, t) = 0, t ∈ IR + ,
(7.20) slcl
v(1, t) + βu(1, t) = 0, t ∈ IR + ,
and the initial conditions

u(x, 0) = u0 (x), x ∈ (0, 1),

(7.21) slci
v(x, 0) = v0 (x), x ∈ (0, 1),
where
sl
∈ IR ? , β ∈ IR ? , u0 ∈ L∞ (Ω) and v0 ∈ L∞ (Ω) are given. It is well known that the problem
αslci
(7.19)-(7.21) admits a unique weak solution (entropy conditions are not necessary to obtain uniqueness
of the solution of this linear system).
sl slci
A stable numerical scheme for the discretization of the problem (7.19)-(7.21) will add some numerical
diffusion terms. It sl
seems quite natural to assume that this diffusion does not lead a coupling between the
two equations of (7.19). Then, roughly speaking, the numerical scheme will consist in an approximation
of the following parabolic system:

ut (x, t) + ux (x, t) − εuxx (x, t) = 0, x ∈ (0, 1), t ∈ IR + ,

(7.22) slp
vt (x, t) − vx (x, t) − ηvxx (x, t) = 0, x ∈ (0, 1), t ∈ IR + ,
for some ε > 0 and η > 0 depending on the mesh (and time step) and ε → 0, η → 0 as the space and
time steps tend to 0. slci
In order to be well posed, this parabolic system has to be completed with the initial conditions ( 7.21)
and (for all t > 0) four boundary conditions, i.e. two conditions at x = 0 and two conditions slp
at x = 1.
This is also the case for the numerical scheme
slcl
which may be viewed as a discretization of ( 7.22). There
are two boundary conditions given by (7.20). Hence two other boundary conditions must be found, one
at x = 0 and the other at x = 1.

Ifslcl
these slp
two additional conditions are, for instance, v(0, t) = u(1, t) = 0, then the (unique) solution to
(7.20)-(7.22)slwith these
slci
two additional conditions does not converge, as ε → 0 and η → 0, to the weak
solution of (7.19)-(7.21). This negative result is also true for a large choice of other additional boundary
conditions. However, if theslp
slcl
additional boundary conditions are (wisely) chosen to be v x (0, t) = ux (1, t) =
0,slthe solution
slci
to ( 7.20)-( 7.22) with these two additional conditions converges to the weak solution of
(7.19)-(7.21).

Theslcl
numerical treatment of the boundary conditions described above may be viewed as a discretization
of (7.20) and vx (0, t) = ux (1, t) = 0; this remark gives a formal justification to such a choice.
 207

7.1.5 Staggered grids

For some systems of equations it may be “natural” (in the sense that the discretization seems simpler) to
associate different grids to different unknowns of the problem. To each unknown is associated an equation
and this equation is integrated over the elements (which are the control volumes) of the corresponding
mesh, and then discretized by using one discrete unknown per control volume (and time step, for evolution
problems). This is the case, for instance, ofPatthe well known discretization sstokes
of the incompressible Navier-
Stokes equations with staggered grids, see Patankar [1980] and Section 7.2.2.
Let us now give an example in order to show that staggered grids should be avoided in the case of
nonlinear hyperbolic systems since they may yield some kind of “instability”. As an illustration, let us
consider the following “academic” problem:

ut (x, t) + (vu)x (x, t) = 0, x ∈ IR, t ∈ IR + ,

vt (x, t) + (v 2 )x (x, t) = 0, x ∈ IR, t ∈ IR + ,
(7.23) sg
u(x, 0) = u0 (x), x ∈ IR,
v(x, 0) = u0 (x), x ∈ IR,
where u0 is a bounded function from IR to [0, 1]. Taking u = v equal to the weak entropy solution of the
Bürgers
sg
equation (namely ut + (u2 )x = 0), with initial condition u0 , leads to a solution of the problem
(7.23). One would expect a numerical scheme to give an approximation of this solution. Note that the
solution of the Bürgers equation, with initial condition u0 , also takes its values in [0, 1], and hence, a
“good” numerical scheme can be expected to give approximate solutions taking values in [0, 1]. Let us
show that this property is not satisfied when using staggered grids.

Let k be the time step and h be the (uniform) space step. Let xi = ih and xi+1/2 = (i + 1/2)h, for
i ∈ ZZ . Define, for i ∈ ZZ , Ki = (xi−1/2 , xi+1/2 ) and Ki+1/2 = (xi , xi+1 ).
The mesh associated to u is {Ki , i ∈ ZZ } and the mesh associated to v is {Ki+1/2 , i ∈ ZZ }. Using the
principle of staggered grids, the discrete unknowns are uni , i ∈ ZZ , n ∈ IN? , and vi+1/2
n
, i ∈ ZZ , n ∈ IN? .
The discretization of the initial conditions is, for instance,
Z
1
u0i = u0 (x)dx, i ∈ ZZ ,
h KZi
0 1 (7.24) sgic
vi+ 1 = u0 (x)dx, i ∈ ZZ .
2 h Ki+ 1
2
sg
The second equation of (7.23) does not depend on u. It seems reasonable to discretize this equation with
the Godunov scheme,sgwhich is here the upstream scheme, since u0 is nonnegative. The discretization of
n
the first equation of (7.23) with the principle of staggered grids is easy. Since vi+1/2 is always nonnegative,
we also take an upstream value for u at the extremities of the cell Ki . Then, with the explicit Euler
scheme in time, the scheme becomes
1 n+1 1 n
(u − uni ) + (vi+ n
1 ui − v
n n
i− 21 ui−1 ) = 0, i ∈ ZZ , n ∈ IN,
k i h 2
(7.25) sgd
1 n+1 n 1 n 2 n 2
(v 1 − vi+ 1) + ((v 1 ) − (vi− 1 ) ) = 0, i ∈ ZZ , n ∈ IN.
k i+ 2 2 h i+ 2 2

It is easy to show that, whatever k and h, there exists u0 (function from IR to [0, 1]) such that sup{u1i , i ∈
1. In fact, it is possible to have, for instance, sup{u1i , i ∈ ZZ } = 1 + k/(2h).
ZZ } is strictly larger thansgd
In this sense the scheme (7.25) appears to be unstable. Note that the same phenomenon exists with the
implicit Euler scheme instead of the explicit Euler scheme . Hence staggered grids do not seem to be the
best choice for nonlinear hyperbolic systems.
 208

7.2 Incompressible Navier-Stokes Equations

nseq
The discretization of the stationary Navier-Stokes equations by the finite volume method is presented in
this section. We first recall the classical discretization on cartesian staggered grids. We then study, in
the linear case of the Stokes equations, a finite volume method on a staggered triangular grid, for which
we show, in a particular case, the convergence of the method.

7.2.1 The continuous equation

cstokes
Let us consider here the stationary Navier-Stokes equations:
d
X ∂u(i) ∂p
−ν∆u(i) (x) + u(j) (x) (x) + (x) = f (i) (x), x ∈ Ω, ∀i = 1, . . . , d,
j=1
∂xj ∂xi
d
(7.26) nstkc
X ∂u(i)
(x) = 0, x ∈ Ω.
i=1
∂xi
with Dirichlet boundary condition

u(i) (x) = 0, x ∈ ∂Ω, ∀i = 1, . . . , d, (7.27) stkb

under the following assumption:
stkH Assumption 7.1

(i) Ω is an open bounded connected polygonal subset of IR d , d = 2, 3,

(iii) ν > 0,
(iii) f (i) ∈ L2 (Ω), ∀i = 1, . . . , d.

In the above equations, u(i) represents the ith component of the velocity of a fluid, ν the kinematic
viscosity and p the pressure. The unknowns of the problem are u(i) , i ∈ {1, . . . , d} and p. The number
nstkc
of unknown functions from Ω to IR which are to be computed is therefore d + 1. Note that ( 7.26) yields
d + 1 (scalar) equations.

We shall also consider the Stokes equations, which are obtained by neglecting the nonlinear convection
term.
∂p
−ν∆u(i) (x) + (x) = f (i) (x), x ∈ Ω, ∀i = 1, . . . , d,
∂xi
d
X ∂u(i) (7.28) stkc
= 0, x ∈ Ω.
i=1
∂xi
nstkc stkb stkc stkb
There
temam
exist several convenient mathematical formulations of (7.26)-(7.27) and (7.28)-(7.27), see e.g.
Temam [1977]. Let us give one of them for the Stokes problem. Let
d
X ∂u(i)
V = {u = (u(1) , . . . , u(d) )t ∈ (H01 (Ω))d , = 0}.
i=1
∂xi
stkH
Under assumption 7.1, there exists a unique function u such that

u ∈ V,
d Z
X d Z
X
(i) (i) (7.29) stksolvar
ν ∇u (x) · ∇v (x)dx = f (i) (x)v (i) (x)dx, ∀v = (v (1) , . . . , v (d) )t ∈ V.
i=1 Ω i=1 Ω
stksolvar R
Equation (7.29) yields the existence of p ∈ L2 (unique if Ω p(x)dx = 0) such that
 209

∂p
−ν∆u(i) + = f (i) in D0 (Ω), ∀i ∈ {1, . . . , d}. (7.30) stkp
∂xi
nstkc stkb
Instkc
the following,
stkb
we shall study finite volume schemes for the discretization of Problem ( 7.26)-(7.27) and
(7.28)-(7.27). Note that the Stokes equations may also be successfully discretized by the finite element
giraultraviart
method, see e.g. Girault and Raviart [1986] and references therein.

7.2.2 Structured staggered grids

sstokes
The
Pat
discretization of the incompressible Navier-Stokes equations with staggered grids is classical (see
Patankar [1980]): the idea is to associate different control volume grids to the different unknowns. In
the two-dimensional case, the meshes consist in rectangles. Consider, for instance, the mesh, say T , for
the pressure p. Then, considering that the discrete unknowns are located at the centers of the elements of
their associated mesh, the discrete unknowns for p are, of course, located at the centers of the element of
T . The meshes are staggered such that the discrete unknowns for the x-velocity are located at the centers
of the edges of T parallel to the y-axis, and the discrete unknowns for the y-velocity are located at the
centers of the edges of T parallel to the x-axis. The two equations of “momentum” are associated to the
x and y-velocity (and integrated over the control volumes of the considered mesh) and the “divergence
free” equation is associated to the pressure (and integrated over the control volume of T ). Then the
discretization of all the terms of the equations is straightforward, except for the convection terms (in
the momentum equations) which, eventually, have to be discretized according to the Reynolds number
(upstream or centerednicol1
discretization. . . ). The convergence analysis of this so-called “MAC” (Marker and
nicol2
Cell) is performed in Nicolaides [1992] in the linear case and Nicolaides and Wu [1996] in the case
of the Navier-Stokes equations.

7.2.3 A finite volume scheme on unstructured staggered grids

ustokes
Let us now turn to the case of unstructured grids; the scheme we shall study uses the same control
volumes for all the components of the velocity. The pressure unknowns are located at the vertices, and a
Galerkin expansion is used for the approximation of the pressure. Note that other finite volume schemes
have been proposed bottahempel
for the discretization of the Stokes and incompressible Navier-Stokes equations on
unstructured grids (Botta and Hempel [1996]), but, to our knowledge, no proof of convergence has
been given yet.
meshdirichlet
meshdirichlet
We again use the notion of admissible mesh, introduced in Definition 3.1 page 37, in the particular case
of triangles, if d = 2, or tetrahedra, if d = 3. We limit the description below to the case d = 2 and to the
Stokes equations. Let Ω be an open bounded polygonal connected subset Ω of IR 2 . Let T be a mesh ciarlet
of Ω
consisting of triangles, satisfying the properties required for the finite element method (see e.g. Ciarlet,
P.G. [1978]), with acute angles only. Defining, for all K ∈ T , the point xK as the intersection of the
orthogonal meshdirichletof the sides of the triangle K yields that T is an admissible mesh in the sense of
bisectorsmeshdirichlet
Definition 3.1 page 37. Let ST be the set of vertices of T . For S ∈ ST , let φS be the shape function
associated to S in the piecewise linear finite element method for the mesh T . For all K ∈ T , let SK ⊂ ST
be the set of the vertices of K.
A possible finite volume scheme using a Galerkin expansion
meshdirichlet
for the pressure is defined by the following
meshdirichlet
equations, with the notations of Definition 3.1 page 37:
X (i) X Z
∂φS (i)
ν FK,σ + pS (x)dx =m(K)fK ,
K ∂xi (7.31) stkschu
σ∈EK S∈SK
∀K ∈ T , ∀i = 1, . . . , d,
(i) (i) (i)
FK,σ = τσ (uK − uL ), if σ ∈ Eint , σ = K|L, i = 1, . . . , d,
(i) (i) (7.32) stkflu
FK,σ = τσ uK , if σ ∈ Eext ∩ EK , i = 1, . . . , d,
 210

d
XX Z
(i) ∂φS
uK (x)dx = 0, ∀S ∈ ST , (7.33) stkschd
K ∂xi
K∈T i=1
Z X
pS φS (x)dx = 0, (7.34) stkschpp
Ω S∈S
T
Z
(i) 1
fK = f (x)dx, ∀K ∈ T . (7.35) stkschf
m(K) K
stkschustkschf (i)
The discrete unknowns of (7.31)-(7.35) are uK , K ∈ T , i = 1, . . . , d and pS , S ∈ ST .
The approximate solution is defined by
X
pT = p S φS , (7.36) spapp
S∈ST

(i) (i)
uT (x) = uK , a.e. x ∈ K, ∀K ∈ T , ∀i = 1, . . . , d. (7.37) suapp
The proof of the convergence of the scheme is not straightforward in the general case. We shall prove
instkschu
the following
stkschf
proposition the convergence of the discrete velocities given by the finite volume scheme
(7.31)-(7.35) in the simple case of a mesh consisting of equilateral triangles.
stkH
pstk Proposition 7.1 Under Assumption 7.1, let T be a triangular finite element mesh of Ω, with acute
angles only, and let, for all K ∈ T , xK be the intersection of the orthogonal bisectors
meshdirichlet
of the sides of the
meshdirichlet
triangle K (hence T is an admissible mesh in the sense of Definition 3.1 page 37). Then, there exists a
stkschu stkschf (i)
unique solution to (7.31)-(7.35), denoted by {uK , K ∈ T , i = 1, . . . , d} and {pS , S ∈ ST }. Furthermore,
if the elements of T are equilateral triangles, then uT → u in (L2 (Ω))d , as size(T ) → 0, where u is the
stksolvar (1) (d) suapp
(unique) solution to (7.29) and uT = (uT , . . . , uT )d is defined by (7.37).
pstk
Proof of Proposition 7.1.
Step 1 (estimate on uT )
pstk (i)
Let T be an admissible mesh, of Proposition 7.1, and {uK , K ∈ T , i = 1, . . . , d}, {pS ,
in the sense stkschf
stkschustkschd
S ∈ ST } be a solution of (7.31)-(7.33) with (7.35).
stkschu (i) stkschd
Multiplying the equations (7.31) by uK , summing over i = 1, . . . , d and K ∈ T and using (7.33) yields
d X
X d X
X (i) (i)
ν τσ (Dσ u(i) )2 = m(K)uK fK , (7.38) stkest1
i=1 σ∈E i=1 K∈T
(i) (i) (i)
with Dσ u(i) = |uL − uK | if σ ∈ Eint , σ = K|L, i ∈ {1, . . . , d} and Dσ u(i) = |uK | if σ ∈ Eext ∩ EK ,
i ∈ {1, . . . , d}.
stkschustkschf
In stepstkest1
2, the existence and the uniqueness of the solution of (7.31)-(7.35) will be essentially deduced
from (7.38).
ellinpoin stkest1
Using the discrete Poincaré inequality (3.13) in (7.38) gives an L2 estimate andellexistu
an estimate on the
ellexistu
“discrete H01 norm” on the component of the approximate velocities, as in Lemma 3.2 page 42, that is:
(i) (i)
kuT k1,T ≤ C, kuT kL2 (Ω) ≤ C, ∀i ∈ {1, . . . , d},
where C only depends on Ω, vu and f (i) , i = 1, . h1dtot
ellcvgce ellcvgce
. . , d. h1dtot kolmh10 kolmh10
As in Theorem 3.1 page 46 (thanks to Lemma 3.3 page 44 and Theorem 3.10 page 93), this estimate
gives the relative compactness in (L2 (Ω))d of pstk
the set of approximate solutions uT , for T in the set of
admissible meshes in the sense of Proposition 7.1. It also gives that if uTn → u in (L2 (Ω))d , as n → ∞,
where uTn is the solution associated to the mesh Tn , and size(Tn ) → 0 as n → ∞, then u ∈ (H 1 (Ω))d .
stksolvar 0
This will be used in Step 3 in order to prove the convergence of uT to the solution of (7.29).
 211

Step 2 (existence and uniqueness of uT and pT ) pstk stkschd

Let T be an admissible mesh, in the sense of Proposition 7.1. Replace, in the right hand side of (7.33),
(i) stkschustkschd
“0” by “gS ” with some {gS , S ∈ ST } ⊂ IR. Eliminating FK,σ , the system (7.31)-(7.33) becomes a linear
(i)
system with as many equations as unknowns. The sets of unknowns are {uK , K ∈ T , i = 1, . . . , d} and
{pS , S ∈ ST }. Ordering the equations and the unknowns yields a matrix, say A, defining this system.
(i)
Let us determine the kernel of A; let fK = 0 and gS = 0 for all K ∈ T , all S ∈ ST and all i ∈ {1, . . . , d}.
stkest1 (i) stkschu
leads to uK = 0 for all K ∈ T and all i ∈ {1, . . . , d}. Turning back to (7.31) yields that pT
Then, (7.38) spapp
(defined by (7.36)) is constant on K for all K ∈ T . Therefore, since Ω is connected, pT is constant on
Ω. Hence, the dimension of the kernel of A is 1 and so is the codimension of the range of A. In order to
determine the range of A, note that
X
ϕS (x) = 1, ∀x ∈ Ω.
S∈ST
stkschustkschd
Then, a necessary condition in order that the linear system (7.31)-(7.33) has a solution is
X
gS = 0 (7.39) csex
S∈ST

and, since the

csex
codimension of the range of A is 1, this condition is also sufficient. Therefore, under the
stkschustkschd
condition (7.39), the linear system (7.31)-(7.33) has a solution, this solution is stkschu
uniquestkschf
up to an additive
constant for pT . In the particular case gS = 0 for all S ∈ ST , this yields that (7.31)-(7.35) has a unique
solution.
Step 3 (convergence of uT to u)
In this step the convergence of uT towards u in (L2 (Ω))d as size(T ) → 0 is shown for meshes consisting of
pstk
equilateral triangles. Let (Tn )n∈IN be a sequence of meshes (such as defined in Proposition 7.1) consisting
of equilateral triangles and let (uTn )n∈IN be the associated solutions. Assume that size(Tn ) → 0 and
(L2 (Ω))d as n → ∞. Thanks to the compactness result of Step 1,
uTn → u in stksolvar proving that u is the
pstk
solution of (7.29) is sufficient to conclude this step and to conclude Proposition 7.1. stksolvar
By Step 1, u ∈ (H01 (Ω))d . Itstksolvar
remains to show that u ∈ V (which is the first part of (7.29)) and that u
satisfies the second part of (7.29).

For the sake of simplicity of the notations, let us omit, from now on, the index n in Tn and let h = size(T ).
Note that xK (which is the intersection of the orthogonal bisectors of the sides of the triangle K) is the
center of gravity of K, for all K ∈ T . Let ϕ = (ϕ(1) , . . . , ϕ(d) )t ∈ V and assume that the functions ϕ(i)
are regular functions with compact support in Ω, say ϕ(i) ∈ Cc∞ (Ω) for all i ∈ {1, . . . , d}. There exists
C > 0 only depending on ϕ such that
Z
1
|ϕ(i) (xK ) − ϕ(i) (x)dx| ≤ Ch2 , (7.40) stkapp
m(K) K
for all K ∈ T and i = 1, . . . , d. Let usellcvgce
proceedellcvgce
as in the proof of convergence of the finite volume scheme
for the Dirichlet problem (Theorem 3.1 page 46).
Assume that h is small enough so that ϕ(x) = 0 for all x such that x ∈ K, K ∈ T and EK ∩ Eext 6= ∅.
Note that (∂φS )/(∂xi ) is constant in each K ∈ T and that

Xd Z Z Xd
∂φS ∂ϕ(i)
(x)ϕ(i) (x)dx = − φS (x) (x)dx = 0.
i=1 Ω
∂xi Ω i=1
∂xi
Then,
d X X
X Z Z
∂φS 1
pS (x)dx ϕ(i) (x)dx = 0.
i=1 K∈T S∈SK K ∂xi m(K) K
 212

stkschu R
Therefore, multiplying the equations (7.31) by (1/m(K)) K
ϕ(i) (x)dx, for each i = 1, . . . , d, summing
the results over K ∈ T and i ∈ {i . . . , d} yields
d
X X Z Z
(i) (i) 1 1
ν τK|L (uL − uK )( ϕ(i) (x)dx − ϕ(i) (x)dx) =
i=1 K|L∈Eint
m(L) L m(K) K
d X Z (7.41) stkeq1
X (i) (i)
fK ϕ (x)dx.
i=1 K∈T K
stkeq1 stkapp
Passing to the limit in (7.41) as n → ∞ and using (7.40) gives, in thestksolvar
ellcvgce ellcvgce
same way as for the Dirichlet problem
∈ V ∩ (Cc∞ (Ω))d .
(see Theorem 3.1 page 46), that u satisfies the equation given in (7.29), at least for vlions
∞ d 1 d
Then, since V ∩ (Cc (Ω)) is dense (for the (H0 (Ω)) -norm) in V (see, for instance, Lions [1996] for a
stksolvar
proof of this result), u satisfies the equation given in (7.29).
stkschd
Since u ∈ (H01 (Ω))d , it remains to show that u is divergencePfree. Let ϕ ∈ Cc∞ (Ω). Multiplying (7.33) by
ϕ(S), summing over S ∈ ST and noting that the function S∈ST ϕ(S)φS converges to ϕ in H 1 (Ω), one
obtains that stksolvar
u is divergence free and then belongs to V . This
pstk
completes the proof that u is the (unique)
solution of (7.29) and concludes the proof of Proposition 7.1.

7.3 Flows in porous media

flowpm
7.3.1 Two phase flow
This section is devoted to the discretization of a system which may be viewed as an elliptic equation
coupled to a hyperbolic equation. This system appears in the modelling of a two phase flow in a porous
medium. Let Ω be an open bounded polygonal subset of IR d , d = 2 or 3, and let a and b be functions of
class C 1 from IR to IR + . Assume that a is nondecreasing and b is nonincreasing. Let g and u be bounded
functions from ∂Ω × IR + to IR, and u0 be a bounded function from Ω to IR. Consider the following
problem:

ut (x, t) − div(a(u)∇p)(x, t) = 0, (x, t) ∈ Ω × IR + ,

(1 − u)t (x, t) − div(b(u)∇p)(x, t) = 0, (x, t) ∈ Ω × IR + ,
∇p(x, t) · n(x) = g(x, t), (x, t) ∈ ∂Ω × IR + , (7.42) egeqeeh2D
u(x, t) = u(x, t), (x, t) ∈ ∂Ω × IR + ; g(x, t) ≥ 0,
u(x, 0) = u0 (x), x ∈ Ω,
where n is the normal to ∂Ω, outward to Ω. The unknowns
egeqeeh2D
of this system are the functions p and u (from
Ω×IR + to IR). Adding the two first equations of (7.42), this system may be viewed as an elliptic equation
with respect to the unknown p, for a given u (note that there is no time derivative in this equation), with
a Neumann condition, coupled to a hyperbolic equation with respect to the unknown u (for a given p).
Note that, for the elliptic problem with the Neumann condition, the compatibility condition on g writes
Z
M (u(x, t))g(x, t)dγ(x) = 0, t ∈ IR + ,
∂Ω
egeqeeh2D
where M = a + b. It is not known whether the system (7.42) has a solution, except in the simple
case where the function M is a positive constant (which is, however, already an interesting case for real
applications).
egeqeeh2D meshneumanmeshneuman
In order to discretize (7.42), let T be an admissible mesh of Ω in the sense of Definition 3.5 page 63 and
k > 0 be the time step. The discrete unknowns are pnK and unK for K ∈ T and n ∈ IN ? . The discretization
of the initial condition is
 213

Z
1
u0K = u0 (x)dx, K ∈ T .
m(K) K

In order to take into account the boundary condition on u, define, with tn = nk,
Z Z tn+1
n 1
uK = u(x, t)dγ(x)dt, K ∈ T , n ∈ IN.
k m(∂K ∩ ∂Ω) ∂K∩∂Ω tn
The scheme will use an “upstream choice” of a(u) and b(u) on each “interface” of the mesh, that is, for
all K ∈ T , L ∈ N (K),

(a(u))nK,L = a(unK ) if pn+1

K ≥ pn+1
L
(a(u))nK,L = a(unL ) if pK < pn+1
n+1
L ,
(b(u))nK,L = b(unK ) if pn+1
K ≥ pn+1
L
(b(u))nK,L = b(unL ) if pK < pn+1
n+1
L ,

The discrete equations are, for all K ∈ T , n ∈ IN,

un+1 − unK X
m(K) K
− τK|L (pn+1
L − pn+1
K )(a(u))K,L
n
k
L∈N (K)
Z Z tn+1 Z Z tn+1
a(unK ) a(unK )
− g + (x, t)dγ(x)dt + g − (x, t)dγ(x)dt = 0,
k ∂K∩∂Ω tn k ∂K∩∂Ω tn
un+1 − unK X
−m(K) K − τK|L (pn+1
L − pn+1
K )(b(u))K,L
n
k
Z Z tn+1L∈N (K) Z Z tn+1
b(unK ) b(unK )
− g + (x, t)dγ(x)dt + g − (x, t)dγ(x)dt = 0.
k ∂K∩∂Ω tn k ∂K∩∂Ω tn
meshdirichlet
meshdirichlet
Recall that g + (x, t) = max{g(x, t), 0}, g − = (−g)+ and τK|L = m(K|L)/dK|L (see Definition 3.1 page 37).
This finite volume scheme gives very good numerical results under a usual stability condition on the time
step with respect to the space mesh. It can be generalized to more complicated systems (in particular, for
the Esimulation of multiphase flows in porous medium such as the “black oil” case of reservoir engineering,
see Eymard [1992]). It is possible to prove the convergence of this scheme in the case where the function
M is constant and the function g does not depend on t. In this case, the scheme may be written as a finite
volume scheme for a stationary diffusion equation with respect to the unknown p (which does not depend
on t) and an upstream finite volume scheme for a hyperbolic equation with respect to the unknown u.
egcvgce
The proof of this convergence
Vignal1
is given below (Theorem 7.1) under the assumptions that a(u) = u and
b(u) = 1 − u (see also Vignal [1996a]). Note that the elliptic equation with respect to the pressure
may also be discretized with a finite element method, and coupled to the finite volume scheme F2
for the
hyperbolic equation. This coupling of finite elements and finite volumes ES
was introduced in Forsyth
[1991],
EG
where it is called “CVFE” (Control Volume Finite Element), in Sonier and Eymard [1993] and
in Eymard and Gallouët [1993], where the convergence of the finite element-finite volume scheme is
shown under the same assumptions.

7.3.2 Compositional multiphase flow

thermo
Let us now turn to the study of a system of partial differential equations which arises in the simulation of
aE multiphase flow in a porous medium (the so called “Black Oil” case in petroleum engineering, see e.g.
Eymard [1992]). This system consists in a parabolic equation coupled with hyperbolic equations and
algebraic equations and inequalities (these algebraic equations and inequalities are given by an assumption
of thermodynamical equilibrium). It may be written, for x ∈ Ω and t ∈ IR + , as:
∂
(ρ1 (p)u)(x, t) − div(f1 (u, v, c)∇p)(x, t) = 0, (7.43) triph1
∂t
 214

∂
(ρ2 (p, c)(1 − u − v)(1 − c))(x, t) − div(f2 (u, v, c)∇p)(x, t) = 0, (7.44) triph2
∂t
∂
(ρ2 (p, c)(1 − u − v)c + ρ3 (p)v)(x, t) − div(f3 (u, v, c)∇p)(x, t) = 0, (7.45) triph3
∂t
(v(x, t) = 0 and c(x, t) ≤ f (p(x, t)) or (c(x, t) = f (p(x, t)) and v(x, t) ≥ 0), (7.46) triph4
where Ω is a given open bounded polygonal subset of IR d (d = 2 or 3), f1 , f2 , f3 are given functions from
IR 3 to IR + , f , ρ1 , ρ3 are given functions from IR to IR + and ρ2 is a given function from IR 2 to IR + . The
problem is completed by initial and boundary conditions which are omitted here. The unknowns of this
problem are the functions u, v, c, p from Ω × IR + to IR.

In order to discretize this problem, let k be the time step (as usual, k may in fact be variable) and T be a
cartesian mesh of Ω. Following the ideas (and notations) of the previous chapters, the triph1 discrete unknowns
triph3
are unK , vKn
, cnK and pnK , for K ∈ T and n ∈ IN? and it is quite easy to discretize (7.43)-(7.45) with a
classical finite volume method. Note that the time discretization of the unknown p must generally be
implicit while the time discretization of the unknowns u, v, c may be explicit or implicit. The explicit
choice requires a usual restriction on triph4
the time step (linearly with respect to the space step). The only
new problem is the discretization of (7.46), which is now described.
Let n ∈ IN. The discrete unknowns at time tn+1 , namely un+1 n+1
K , vK , cK
n+1
and pn+1
K , K ∈ T , have to be
n n n n
computed from the discrete unknowns
triph1 triph3
at time t n , namely u K , v K , c K and p K , K ∈ T . Even if the time
discretization of (7.43)-(7.45) is explicit with respect to the unknowns u, v and c, the system of discrete
equations
triph4
(with unknowns un+1 n+1
K , vK , cK
n+1
and pn+1 K , K ∈ T ) is nonlinear, whatever the discretization
of (7.46). It can be solved by, say, a Newton process. Let l ∈ IN be the index of the “Newton iteration”,
n+1,l n+1,l n+1,l n+1,l
and uK , vK , cK and pK (K ∈ T ) be the computed unknowns at iteration l. As usual, these
triph4
unknowns are, for l = 0, taken equal to unK , vK n
, cnK and pnK . In order to discretize (7.46), a “phase index”
is introduced; it is denoted by inK , for all K ∈ T and n ∈ IN and it is defined by:

if inK = 0 then vK
n
= 0 ( and cnK ≤ f (pnK )),
if iK = 1 then cK = f (pnK ) ( and vK
n n n
≥ 0).
In the Newton process for the computation of the unknowns at time tn+1 , a “phase index”, denoted by
n+1,l n+1,0
iK is also introduced, with iK = inK . This phase index is used in the computation of un+1,l+1K ,
n+1,l+1 n+1,l+1 n+1,l+1 n+1,l+1 n+1,l n+1,l n+1,l n+1,l n+1,l
vK , cK , pK and iK (K ∈ T ), starting from uK , vK , cK , pK and iK .
n+1,l+1 n+1,l
Setting vK = 0 if iK = 0, and cn+1,l+1
K = f (p n+1,l+1
K ) if i n+1,l
K = 1, the computation of (inter-
n+1,l+1 n+1,l+1 n+1,l+1 n+1,l+1 triph1
mediate) values of uK
triph2 triph3
, vK , cK , pK is possible with a “Newton iteration” on (7.43),
(7.44), (7.45) (note that the number of unknowns is equal to the number of equations). Then, for each
K ∈ T , three cases are possible:

1. if cn+1,l+1
K
n+1,l+1
≤ f (pK n+1,l+1
) and vK ≥ 0, then set in+1,l+1
K
n+1,l
= iK ,

2. if cn+1,l+1
K > f (pn+1,l+1
K
n+1,l
) (and necessarily iK n+1,l+1
= 0), then set cK = f (pn+1,l+1
K ) and
in+1,l+1
K = 1,
n+1,l+1 n+1,l n+1,l+1
3. if vK < 0 (and necessarily iK = 1), then set vK = 0 and in+1,l+1
K = 0.

This yields the final values of un+1,l+1

K
n+1,l+1
, vK , cn+1,l+1
K , pn+1,l+1
K
n+1,l+1
and iK (K ∈ T ).

When the “convergence” of the Newton process is achieved, say at iteration l ? , the values of the unknowns
at time tn+1 are found. They are taken equal to those indexed by (n + 1, l ? ) (for u, v, c, p, i). It can be
proved, under convenient hypotheses on the function f (which are realistic in the applications), that there
is noEG1
“oscillation” of the “phase index” during the Newton iterations performed from time t n to time tn+1
(see Eymard and Gallouët [1991]). This method, using the phase index, HM
was also successfully adapted
for the treatment of the obstacle problem and the Signorini problem, see Herbin and Marchand [1997].
 215

7.3.3 A simplified case

casimple
The aim of this section and of the following sections is the study of the convergence of two coupled finite
volume schemes, for the system of equations ut − div(u∇p) = 0 and ∆p = 0, defined on an open set
Ω. A finite volume mesh T is used for the discretization in
Vignal1
space, together with an explicit Euler time
vigver
discretization. Similar results are in Vignal [1996a] and Vignal and Verdière [1998] where the case
of different space meshes for the two equations is also studied.

We assume that the following assumption is satisfied.

eghyp1 Assumption 7.2 Let Ω be an open polygonal bounded connected subset of IR d , d = 2 or 3, and ∂Ω its
boundary. We denote by n the normal vector to ∂Ω outward to Ω.
Let g ∈ L2 (∂Ω) be a function such that
Z
g(x)dγ(x) = 0,
∂Ω

and let ∂Ω ={x ∈ ∂Ω, g(x) ≥ 0}, Ω =Ω ∪ ∂Ω+ and ∂Ω− ={x ∈ ∂Ω, g(x) ≤ 0}. Let u0 ∈ L∞ (Ω)
+ +

and ū ∈ L∞ (∂Ω+ × IR ?+ ) represent respectively the initial condition and the boundary condition for the
unknown u.

The set
D(Ω+ × IR + ) = {ϕ ∈ Cc∞ (IR d × IR, IR), ϕ = 0 on ∂Ω− × IR + }
egeq1
will be the set of test functions for Equation (7.51) in the weak formulation of the problem, which is
given below.

Definition 7.1 A pair (u, p) ∈ L∞ (Ω × IR ?+ ) × H 1 (Ω) (u is the saturation, p is the pressure) is a weak
solution of


 ∆p(x) = 0, ∀x ∈ Ω,


 ∇p(x) · n(x) = g(x), ∀x ∈ ∂Ω,
ut (x, t) − div(u∇p)(x, t) = 0, ∀x ∈ Ω, ∀t ∈ IR + , (7.47) upfort



 u(x, 0) = u 0 (x), ∀x ∈ Ω,

u(x, t) = ū(x, t), ∀x ∈ ∂Ω+ , ∀t ∈ IR + .
if it verifies
p ∈ H 1 (Ω), (7.48) egeq02

u ∈ L∞ (Ω × IR ?+ ), (7.49) egeq01

Z Z
∇p(x) · ∇X(x)dx − X(x)g(x)dγ(x) = 0, ∀X ∈ H 1 (Ω). (7.50) egeq2
Ω ∂Ω
and Z Z Z
u(x, t)(ϕt (x, t) − ∇p(x) · ∇ϕ(x, t)dxdt + u0 (x)ϕ(x, 0)dx+
ZIR + ZΩ Ω
(7.51) egeq1
ū(x, t)ϕ(x, t)g(x)dγ(x)dt = 0, ∀ϕ ∈ D(Ω+ × IR + ).
IR + ∂Ω+
eghyp1
Under Assumption 7.2, a classical result gives the existence of p ∈ H 1 (Ω) and the uniqueness of ∇p where
egeq02 egeq2
p is the solution of (7.48),(7.50), which is a variational formulation of the classical Neumann problem.
Additional hypotheses
egeq1
on the function g are necessary to get the uniqueness of u ∈ L ∞ (IR d × IR ?+ )
solution of (7.51). The existence of u results from the convergence of the scheme, but not its uniqueness,
 216

which could be obtained thanks to regularity properties of ∇p. We shall assume such regularity, which
ensures the uniqueness of the function u and allows an error estimate between the finite volume scheme
approximation of the pressure and the exact pressure. In fact, for the sake of simplicity, we assume (in
eghyp4
Assumption 7.3 below) that p ∈ C 2 (Ω). This is a rather “strong” assumption which can be weakened.
egcvgce
However, a convergence result (such as in Theorem 7.1) with the only assumption p ∈ H 1 (Ω) seems
not easy to obtain. Note also that similar results of convergence (for the “pressure scheme” and for the
“saturation scheme”) are possible with an open bounded connected subset of parnlHTIR d withparnlHT
a C 2 boundary
d
(instead of an open bounded connected polygonal subset of IR ) using Definition 4.4 page 114 of admissible
meshes.
egeq2
eghyp4 Assumption 7.3 The pressure p, weak solution in H 1 (Ω) to (7.50), belongs to C 2 (Ω).
egeq02 egeq1
Remark 7.2 The solution (u, p) of (7.48)-(7.51) is also a weak solution of

(1 − u)t (x, t) − div((1 − u)∇p)(x, t) = 0.

Remark 7.3 The finite volume scheme will ensure the conservation of each of the quantities u and
1 − u. It can be extended
thermo
to more complex phenomena such as compressibility, thermodynamic equilib-
rium. . . (see Section 7.3.2)

Remark 7.4 The proof which is given here can easily be extended to the case of the existence of a source
term which writes

−∆p(x) = v(x), x ∈ Ω,
∇p(x) · n(x) = g(x), x ∈ ∂Ω,
ut (x, t) − div(u∇p)(x, t) + u(x, t)v − (x) = s(x, t)v + (x), x ∈ Ω, t ∈ IR + ,
u(x, 0) = u0 (x), x ∈ Ω,
u(x, t) = ū(x, t), x ∈ ∂Ω+ , t ∈ IR + ,
Z Z
where v ∈ L2 (Ω) with g(x)dγ(x) + v(x)dx = 0 and s ∈ L∞ (Ω × IR ?+ ). All modifications which are
∂Ω Ω
connected to such terms will be stated in remarks.

7.3.4 The scheme for the simplified case

Let Ω be an open polygonal bounded connected subset of IR d . Let T be an admissible mesh, in the sense
meshneumanmeshneuman
of Definition 3.5 page 63, and let h = size(T ). Assume furthermore that, for some α > 0, dσ ≥ αh for all
σ ∈ Eint .

The pressure finite volume scheme

nll nll
We first define the
nllcondbor
approximate pressure, using the finite volume scheme defined in section 3.2 page 63
nllschemaO
(that is (3.85)-(3.87)).
(i) The values GK , for K ∈ T , are defined by
Z
GK = g(x)dγ(x) if m(∂K ∩ ∂Ω) 6= 0,
∂K∩∂Ω (7.52) egschemapb
GK = 0, if m(∂K ∩ ∂Ω) = 0.

(ii) The scheme is defined by

X  
− τK|L pL − pK = GK , ∀K ∈ T , (7.53) egschemap
L∈N (K)
 217

and
X
m(K)pK = 0. (7.54) egschemap0
K∈T
nllexistu nllexistu
We recall that, from lemma 3.6 page 64, there exists a unique function ∈ X(T ) defined by pT (x) = pnllesterr
pT egschemap0
egschemapb K
for a.e. x ∈ K,
nllesterr
for all K ∈ T , where (p )
K K∈T satisfy equations ( 7.52)-( 7.54). Then, using Theorem 3.5
page 69, there exist C1 and C2 , only depending on p and Ω, such that

kpT − pkL2 (Ω) ≤ C1 h (7.55) egllesterr

and
X Z
pL − p K 1
2
m(K|L)dK|L − ∇p(x) · nK,L dγ(x) ≤ (C2 h)2 . (7.56) egllesterr1
dK|L m(K|L) K|L
K|L∈Eint
nllest nllest
Last but not least, using lemma 3.11 page 74, there exists C3 , only depending on g and Ω, such that
X
τK|L (pL − pK )2 ≤ (C3 )2 . (7.57) egllesterr2
K|L∈Eint

The saturation finite volume scheme

egeq1
Let us now turn to the finite volume discretization of the hyperbolic equation (7.51). In order to write
the scheme, let us introduce the following notations: let
Z Z
(+) (−)
GK = g + (x)dγ(x) and GK = g − (x)dγ(x),
∂K∩∂Ω ∂K∩∂Ω
(+) (−)
so that GK − GK = GK . Let
Z X (+)
G(+) = g + (x)dγ(x) = GK
∂Ω K∈T
egschemap
(note that G(+) does not depend on T ). The scheme (7.53) may also be written
X  
(+) (−)
τK|L pL − pK + GK − GK = 0, ∀K ∈ T . (7.58) egeq3
L∈N (K)
egschemap
Remark 7.5 In the case of the problem with source terms, the right hand side of the equation ( 7.53) is
(+) (−)
replaced by GK + VK − VK with
Z
(±)
VK = v ± (x)dx.
K
egeq3 (±) (±) (±)
Then, in the equation (7.58) the quantities GK are replaced by GK + VK .

Let ξ ∈ (0, 1). Given an admissible mesh T , the time step is defined by a real value k > 0 such that

m(K) (1 − ξ)
k ≤ inf X (+)
. (7.59) egeqS
K∈T τK|L (pL − pK )+ + GK
L∈N (K)
 218

egeqS
Remark 7.6 Since the right hand side of (7.59) has a strictly positive lower bound,
egeqS egeqS
it is always possible
to find values k > 0 which satisfy (7.59). Roughly speaking, the condition (7.59) is a linear condition
between the time step and the size of the mesh. Let us explain this point in more detail: in most practical
cases, function g is regular enough so that |pL − pK |/dK|L is bounded by some C only meshneuman
depending on g
meshneuman
and Ω. Assume furthermore that the mesh T is admissible in the sense of Definition 3.5 page 63 and
that, for some α > 0, dK,σ ≥ αh, for all K ∈ T and σ egeqS ∈ E. Then the condition k ≤ Dh, with
D = ((1 − ξ)α)/(d(C + kgkL∞(∂Ω) )), implies the condition (7.59). Note also that for all g ∈ L2 (∂Ω) we
already have a bound
egeqS
for |pT |1,T (but this does not yield a bound on |pL − pK |/dK|L ). Finally, note
that condition (7.59) is easy to implement in practise, since the values τK|L and pK are available by the
pressure scheme.
egeqS
Remark 7.7 In the problem with source terms, the condition (7.59) will be modified as follows:
m(K) (1 − ξ)
k ≤ inf X (+) (+)
.
K∈T τK|L (pL − pK )+ + GK + VK
L∈N (K)

The initial condition is discretized by:

Z
1
u0K = u0 (x)dx, ∀K ∈ T . (7.60) egschema0
m(K) K

We extend the definition of ū by 0 on ∂Ω− × IR + , and we define ūnK , for K ∈ T and n ∈ IN, by
Z (n+1)k Z
1
ūnK = ū(x, t)dγ(x)dt, if m(∂K ∩ ∂Ω) 6= 0,
k m(∂K ∩ ∂Ω) nk ∂K∩∂Ω
(7.61) egschemab
ūnK = 0, if m(∂K ∩ ∂Ω) = 0.
Hence the following function may be defined on ∂Ω × IR + :

ūT ,k (x, t) = ūnK , ∀x ∈ ∂K ∩ ∂Ω, ∀K ∈ T , ∀t ∈ [nk, (n + 1)k), n ∈ IN.

egeq1
The finite volume discretization of the hyperbolic equation (7.51) is then written as the following relation
between un+1
K and all unL , L ∈ T .

h X i
(+) (−)
m(K)(un+1 n
K − uK ) − k τK|L unK,L (pL − pK ) + ūnK GK − unK GK = 0, ∀K ∈ T , ∀n ∈ IN, (7.62) egschema
L∈N (K)

in which the upstream value unK,L is defined by

unK,L = unK , if pK ≥ pL ,
(7.63) egschemau
unK,L = unL , if pL > pK .

The approximate solution, denoted by uT ,k , is defined a.e. from Ω × IR + → to IR by

uT ,k (x, t) = unK , ∀x ∈ K, ∀K ∈ T , ∀t ∈ [nk, (n + 1)k), ∀n ∈ IN. (7.64) eguapp

Remark 7.8 In the case of source terms, the following term is defined:
Z (n+1)k Z
1
snK = s(x, t)dxdt
m(K)k nk K

(+) (−) egschema

and the term k(snK VK − unK VK ) is added to the right hand side of (7.62) .
 219

7.3.5 Estimates on the approximate solution

Estimate in L∞ (Ω × IR ?+ )
eghyp1 eghyp4
egestlinf Lemma 7.1 Under the assumptions
meshneumanmeshneuman
7.2 and 7.3, let T be an admissible mesh inegschemapb
egeqS
the sense of Definition
egschemap0 egschema0
3.5 page
eguapp
63 and k > 0 satisfying ( 7.59). Then, the function u T ,k defined by ( 7.52)-( 7.54) and (7.60)-
(7.64) satisfies

kuT ,k kL∞(Ω×IR ?+ ) ≤ max{ku0 kL∞(Ω) , kūkL∞ (∂Ω+ ×IR ?+ ) }. (7.65) egeqestlinf

egestlinf
Proof of Lemma 7.1
egschema
Relation (7.62) can be written as
 k X (−)

un+1
K = unK 1 − τK|L (pK − pL )− + GK +
m(K)
L∈N (K)
k X (+)

τK|L unL (pL − pK )+ + GK ūnK .
m(K)
L∈N (K)

Using
X (+)
X (−)
τK|L (pL − pK )+ + GK = τK|L (pK − pL )− + GK ,
L∈N (K) L∈N (K)
egeqS
and Inequality (7.59), the term un+1K may be expressed
egeq3
as a linear combination of terms unL , L ∈ T , and
n
ūK , with positive coefficients. Thanks to relation (7.58), the sum of these coefficients is equal to 1. The
egeqestlinf
estimate (7.65) follows by an easy induction.

egestlinf
Remark 7.9 In the case of source terms, Lemma 7.1 remains true with the following estimate instead
egeqestlinf
of (7.65):

kuT ,k kL∞ (Ω×IR ?+ ) ≤ max{ku0 kL∞ (Ω) , kūkL∞ (∂Ω+ ×IR ?+ ) , kskL∞(Ω×IR ∗+ ) }.

Weak BV estimate
eghyp1 eghyp4
egestbvw Lemma 7.2 Under the assumptions 7.2 and 7.3, let T be an admissible mesh in the sense of Definition
meshneumanmeshneuman egeqS
3.5 page 63. Let h = size(T ) and α > 0 be suchegschema0σ ≥ αh for all σ ∈ Eint . Let k > 0 satisfying
that degschemau egschemapb
(7.59).
egschemap0
Let {unK , K ∈ T , n ∈ IN} be the solution to (7.60)-(7.63) with {pK , K ∈ T } given by (7.52)-(7.54). Let
T > k be a given real value, and let NT,k be the integer value such that NT,k k < T ≤ (NT,k + 1)k. Then
there exists H, which only depends on T , Ω, u0 , ū, g, α and ξ, such that the following inequality holds:
NT,k NT,k
X X XX (+) H
k τK|L |pK − pL ||unK − unL | +k GK |unK − ūnK | ≤ √ . (7.66) egeqbvw
n=0 K|L∈Eint n=0 K∈T h
egestbvw
Proof of Lemma 7.2
egschema
For n ∈ IN and K ∈ T , multiplying (7.62) by unK yields
X (+) (−)
m(K)(un+1 n n n
K uK − uK uK ) − k( τK|L unK,L unK (pL − pK ) + ūnK unK GK − (unK )2 GK ) = 0. (7.67) egeq9
L∈N (K)

egeq9
Writing un+1 n n n 1 n+1
K uK − uK uK = − 2 (uK − unK )2 − 21 (unK )2 + 12 (un+1 2
K ) and summing (7.67) on K ∈ T and
n ∈ {0, . . . , NT,k } gives
 220

NT,k
1X X 1X N +1
− m(K)(un+1
K − unK )2 + m(K)((uKT,k )2 − (u0K )2 )
2 n=0 2
K∈T K∈T
NT,k (7.68) egeq10
XX X (+) (−)
−k ( τK|L unK,L unK (pL − pK ) + ūnK unK GK − (unK )2 GK ) = 0.
n=0 K∈T L∈N (K)
egschemau
Using (7.63) gives, for all K ∈ T ,

X X X
− τK|L unK,L unK (pL − pK ) = τK|L (unK )2 (pK − pL )+ − τK|L unL unK (pL − pK )+ .
L∈N (K) L∈N (K) L∈N (K)

Then,
X X X X
− τK|L unK,L unK (pL − pK ) = τK|L ((unK )2 − unL unK )(pK − pL )+ .
K∈T L∈N (K) K∈T L∈N (K)

1
Therefore, since (unK )2 − unK unL = − unL )2 + 12 ((unK )2 − (unL )2 ),
n
2 (uK
X X X X
− τK|L unK,L unK (pL − pK ) = 12 τK|L (unK − unL )2 (pK − pL )+
K∈T L∈N (K) K∈T L∈N (K)
X X
+ 21 τK|L (unK )2 (pK − pL )+
K∈T L∈N (K)
X X
− 21 τK|L (unL )2 (pK − pL )+
K∈T L∈N (K)
X X
= 1
2 τK|L (unK − unL )2 (pK − pL )+
K∈T L∈N (K)
X X
+ 21 τK|L (unK )2 (pK − pL )
K∈T L∈N (K)
egeq3
and, using (7.58),
X X X X
− τK|L unK,L unK (pL − pK ) = 1
2 τK|L (unK − unL )2 (pK − pL )+
K∈T L∈N (K) K∈T L∈N (K)
X (+) 1 X (−) n 2
+ 21 GK (unK )2 − GK (uK ) .
2
K∈T K∈T

Hence
NT,k
XX X (+) (−)
−k ( τK|L unK,L unK (pL − pK ) + ūnK unK GK − (unK )2 GK ) =
n=0 K∈T L∈N (K)
NT,k
X X X (+)
1
2k ( τK|L |pK − pL |(unK − unL )2 + GK (unK − ūnK )2 ) − (7.69) egeq11
n=0 K|L∈Eint K∈T
NT,k
X X (+) (−)
1
2k (GK (ūnK )2 − GK (unK )2 ).
n=0 K∈T
egschema
Using (7.62), we get

NT,k NT,k
XX XX k2 X (+) (−)
2
m(K)(un+1
K − unK )2 = τK|L unK,L (pL − pK ) + ūnK GK − unK GK .
n=0 K∈T n=0 K∈T
m(K)
L∈N (K)
 221

egeq3 egschemau
Then, for all K ∈ T , using again (7.58) and the definition (7.63),
NT,k
XX
m(K)(un+1
K − unK )2 =
n=0 K∈T
NT,k
XX k2 X (+)
2
τK|L (unL − unK )(pL − pK )+ + GK (ūnK − unK ) .
n=0 K∈T
m(K)
L∈N (K)

The Cauchy-Schwarz inequality yields

NT,k NT,k
XX k2  X
XX (+)

m(K)(un+1
K − unK )2 ≤ τK|L (pL − pK )+ + GK
m(K)
n=0X
K∈T n=0 K∈T L∈N (K) 
(+)
τK|L (pL − pK )+ (unL − unK )2 + GK (ūnK − unK )2 .
L∈N (K)
egeqS
Using the stability condition (7.59) and reordering the summations gives
NT,k NT,k
XX X
m(K)(un+1
K − unK )2 ≤ k (1 − ξ)
n=0 K∈T
X n=0
X  (7.70) egeq12
(+)
τK|L |pL − pK |(unL − unK )2 + GK (ūnK − unK )2 .
K|L∈Eint K∈T
egeq10 egeq11 egeq12
Using (7.68), (7.69) and (7.70), we obtain
X N +1
m(K)((uKT,k )2 − (u0K )2 )
K∈T
NT,k  
X X X (+)
+ξk τK|L |pK − pL |(unK − unL )2 + GK (unK − ūnK )2 (7.71) egeq13
n=0 K|L∈Eint K∈T
NT,k
XX (+) (−)
−k (GK (ūnK )2 − GK (unK )2 ) ≤ 0.
n=0 K∈T

Then, setting C4 = m(Ω)ku0 k2L∞(Ω) + 2T G(+) kūk2L∞ (∂Ω+ ×IR ? ) which only depends on Ω, u0 , T , g and ū,
+

NT,k
X N +1 2
XX (−)
m(K)(uKT,k ) +k GK (unK )2 ≤ C4
K∈T n=0 K∈T

(this inequality will not be used in the sequel) and

NT,k NT,k
X X XX (+) C4
k τK|L |pK − pL |(unK − unL )2 +k GK (unK − ūnK )2 ≤ . (7.72) egeq17
n=0 K|L∈Eint n=0 K∈T
ξ

The Cauchy-Schwarz inequality yields

NT,k NT,k
X X XX (+)
k τK|L |pK − pL ||unK − unL | +k GK |unK − ūnK | ≤
n=0 K|L∈Eint n=0 K∈T
NT,k NT,k
X X XX (+) 1
(k τK|L |pK − pL |(unK − unL )2 + k GK (unK − ūnK )2 ) 2 (7.73) egeq18
n=0 K|L∈Eint n=0 K∈T
 X NT,k
X X (+)  12
k ( τK|L |pK − pL | + GK )
n=0 K|L∈Eint K∈T
 222

X
The expression W , defined by W = τK|L |pK − pL |, verifies
K|L∈Eint
X 1
X 1
X 1
W ≤( τK|L ) 2 ( τK|L (pK − pL )2 ) 2 ≤ C3 ( τK|L ) 2 (7.74) egeq19
K|L∈Eint K|L∈Eint K|L∈Eint
egllesterr2
using (7.57). Recall that C3 only depends on g and Ω.
Since
X X 1 dm(Ω)
τK|L ≤ ( m(K|L)dK|L ) ≤ 2 2 (7.75) sumtau
α 2 h2 α h
K|L∈Eint K|L∈Eint

and
X Z
(+)
GK = g + (x)dγ(x),
K∈T ∂Ω
egeqbvw
we finally conclude that (7.66) holds.
NT,k
XX (+)
Remark 7.10 In the case of source terms, one adds the term k VK |unK − snK | in the left hand
egeqbvw n=0 K∈T
side of (7.66) (and H also depends on v and s).

7.3.6 Theorem of convergence

casimplec nll nll
We already know, by the results of section 3.2 page 63, that the pressure scheme converges.
egschema egeqestlinf
Let us now
convergence of the saturation scheme (7.62). Thanks to the estimate (7.65) in L∞ (Ω × IR ?+ )
prove theegestlinf
(Lemma 7.1), for any sequence of meshes and time steps, such that the size of the mesh tends to 0, we can
extract a subsequence such that the approximate saturation converges to a egeq01 function u in L ∞ (Ω × IR ?+ )
egeq1
for the weak-? topology. We have to show that u is the (unique) solution of ( 7.49), (7.51) (the uniqueness
eghyp4
of the solution is given by Assumption 7.3).
eghyp1 eghyp4
egcvgce Theorem 7.1 Under assumptions 7.2 and 7.3, let ξ ∈ (0, 1) and α > 0 be given. For an admissible mesh
meshneumanmeshneuman
T , in the sense ofegeqS
Definition 3.5 page 63, such that ≥ α size(Tegschema0
dσegschemap0
egschemapb
) for all σ ∈ E and for a time step
eguapp int
k > 0 satisfying (7.59), let u
egeq01 egeq1 T ,k be defined by ( 7.52)-( 7.54) and ( 7.60)-( 7.64). Then uT ,k converges to
the solution u of (7.49), (7.51) in L∞ (Ω × IR ?+ ) for the weak-? topology, as size(T ) → 0.
egcvgce
Proof of Theorem 7.1
In the case g(x) = 0 for a.e. (for the (d−1)-dimensional Lebesgue measure) x ∈ ∂Ω, the proof of Theorem
egcvgce
7.1 is easy. Indeed, ∇p(x) = 0 for a.e. x ∈ Ω and, for any mesh and time step, pK − pL = 0 for all K,
L ∈ T . Then, unK = u0K for all K ∈ T and all n ∈ IN. Therefore, it is easy to prove that the sequence
uT ,k converges, as size(T ) → 0 (for any k. . . ), to u, defined by u(x, t) = u0 (x) for a.e. (x, t) ∈ Ω × IR + ;
egeq01 egeq1
note that u is the unique solution to (7.49), (7.51).
Let us now assume that g is not the null function in L2 (∂Ω).
Let (Tm , km )m∈IN be a sequence of space meshes and time steps. For all m ∈ IN, assume that Tm is an
meshneuman
mesh in the sense of Definition 3.5, that dσ ≥ αsize(Tm ) for all σ ∈ Eint and that km > 0
admissibleegeqS
satisfies (7.59) (with k = km and T = Tm ).egschemapb
Assume m ) → 0 as m → ∞.
also that size(Teguapp
egschemap0 egschema0
Let um be the function
egestlinf
u T ,k defined by ( 7.52)-( 7.54) and ( 7.60)-( 7.64), for T = Tm and k = km . By
Lemma 7.1, the sequence (um )m∈IN is bounded in L∞ (Ω × IR ?+ ). In order to prove that the sequence
egeq01 egeq1
(um )m∈IN converges in L∞ (Ω × IR ?+ ) for the weak-? topology to the solution of (7.49), (7.51), using a
classical contradiction argument, it is sufficientegeq01
to prove that if um → u in L∞ (Ω × IR ?+ ) for the weak-?
egeq1
topology then the function u is a solution of (7.49), (7.51).
 223

Let us proceed in two steps. In the first step, it is proved that km → 0 as m → ∞. Then, in the second
egeq01 egeq1
step, it is proved that the function u is a solution of (7.49), (7.51).
From now on, the index “m” is omitted.
Step 1 (proof of k → 0 as m → ∞) egeqS
The proof that k → 0 (as m → ∞) uses (7.59) and the fact that size(T ) → 0.Indeed, define
X
AT = m(K|L)|pK − pL |,
K|L∈Eint

and, for σ ∈ Eint , define χσ from Ω × Ω to {0, 1} by

χσ (x, y) = 1, if σ ∩ [x, y] 6= ∅,

χσ (x, y) = 0, if σ ∩ [x, y] = ∅.
d
Let η ∈ IR \ {0} and ω̄ ⊂ Ω be a compact set such that d(ω̄, Ωc ) ≥ η. Recall that pT is defined by
pT (x) = pK for a.e. x ∈ K and all K ∈ T . For a.e. x ∈ ω̄ one has
X
|pT (x + η) − pT (x)| ≤ χσ (x, x + η)|pK − pL |,
σ=K|L∈Eint
R
integrating this inequality over ω̄ yields, using ω̄ χσ (x, x + η)dx ≤ |η|m(σ),

kpT (· + η) − pT kL1 (ω̄) ≤ |η|AT . (7.76) atau

atau
Assume AT → 0 as m → ∞. Then, since pT → p in L1 (Ω), one deduces from (7.76) that ∇p = 0 a.e. on
Ω which is impossible (since g is not the null function in L2 (∂Ω)). By the same way, it is also impossible
that AT → 0 for a subsequence. Then there exists a > 0 (only depending on the sequence (p T )m∈IN ,
recall that pT = pTm since
X weX omit the index m) such that AT ≥ a for all m ∈ IN.
Therefore, since AT = m(K|L)(pL − pK )+ ≥ a, there exists K ∈ T such that
K∈T L∈N (K)
X m(K)
m(K|L)(pL − pK )+ ≥ a ,
m(Ω)
L∈N (K)

Then, since τK|L = m(K|L)/dK|L and dK|L ≤ 2h,

X m(K)
τK|L (pL − pK )+ ≥ a ,
2hm(Ω)
L∈N (K)
egeqS
which yields, using (7.59),
2
k ≤ (1 − ξ)m(Ω) h.
a
Hence k → 0 as m → ∞ (since h → 0 as m → ∞). This concludes Step 1.
egeq1
Step 2 (proof of u solution to (7.51))
Let ϕ ∈ D(Ω+ × IR + ). Let T > 0 such that, for all t > T − 1 and all x ∈ Ω, ϕ(x, t) = 0. Let m ∈ IN such
that h < 1 and k < 1 (thanks to Step 1, this is true for m large enough). Recall that we denote T = T m ,
= size(Tm ) and k = km . Let NT,k ∈ IN be such that NT,k k < T ≤ (NT,k + 1)k. Multiplying equation
hegschema
(7.62) by ϕ(xK , nk) and summing the result on K ∈ T and n ∈ IN yields

E1,m + E2,m = 0,

with
NT,k
X X
E1,m = m(K)(un+1
K − unK )ϕ(xK , nk)
n=0 K∈T
 224

and
NT,k
X X X (+) (−)

E2,m = − k τK|L unK,L (pL − pK ) + GK ūnK − GK unK ϕ(xK , nk).
n=0 K∈T L∈N (K)

It is shown below that

lim E1,m = T1 , (7.77) egcvgce1

m→∞

where
Z Z Z
T1 = − u(x, t)ϕt (x, t)dxdt − u0 (x)ϕ(x, 0)dx,
IR + Ω Ω

and that

lim E2,m = T2 , (7.78) egcvgce2

m→∞

where
Z Z Z Z
T2 = u(x, t)∇p(x) · ∇ϕ(x, t)dxdt − ū(x, t)ϕ(x, t)g(x)dγ(x)dt.
IR + Ω IR + ∂Ω
egeq01 egeq1
Then, passing to the limit in E1,m + E2,m = 0 proves that u is the (unique) solution of (7.49), (7.51) and
egcvgce
concludes the proof of Theorem 7.1.
egcvgce1
Let us first prove (7.77). Writing E1,m in the following way:
NT,k
X X ϕ(xK , (n − 1)k) − ϕ(xK , nk) n X
E1,m = m(K) uK − m(K)u0K ϕ(xK , 0),
n=1 K∈T
k
K∈T
egcvgce1 parnlcvgce
the assertion (7.77) is easily proved, in the same way as, for instance, in the proof of Theorem 4.2 page
parnlcvgce
111.
egcvgce2
Let us prove now (7.78). To this purpose, we need auxiliary expressions, which make use of the conver-
gence of the approximate pressure to the continuous one. Define E3,m and E4,m by
NT,k
X X Z
pL − p K
E3,m = k (unK − unL ) ϕ(x, nk)dγ(x)
n=0 K|L∈Eint
dK|L K|L
NT,k
X X Z
+ k (unK − ūnK ) g(x)ϕ(x, nk)dγ(x)
n=0 K∈T ∂K∩∂Ω

and
XZ (n+1)k Z Z


E4,m = uT ,k (x, t)∇p(x) · ∇ϕ(x, nk)dx − ūT ,k (x, t)ϕ(x, nk)g(x)dγ(x) dt.
n∈IN nk Ω ∂Ω

We have E4,m → T2 as m → ∞ thanks to the convergence of uT ,k to u in L∞ (Ω × IR) for the weak-?

topology and to the convergence of ūT ,k to ū in L∞ (∂Ω+ × IR + ) for the weak-? topology (the latter
convergence holds also in Lp (∂Ω+ × (0, S)) for all 1 ≤ p < ∞ and all 0 < S < ∞). Let us prove that
|E3,m − E4,m | → 0 as m → ∞ (which gives E3,m → T2 as m → ∞).
using the equation satisfied by p leads to
 225

NT,k
X X Z
E4,m = k (unK − unL ) ϕ(x, nk)∇p(x) · nK,L dγ(x)
n=0 K|L∈Eint K|L
NT,k
X X Z
+ k (unK − ūnK ) g(x)ϕ(x, nk)dγ(x).
n=0 K∈T ∂K∩∂Ω

Therefore,
NT,k
X X Z
pL − p K
E3,m − E4,m = k (unK − unL ) ( − ∇p(x) · nK,L )ϕ(x, nk)dγ(x)
n=0 K|L∈Eint K|L dK|L
NT,k
X X  X Z 
pL − p K
= k unK ( − ∇p(x) · nK,L )ϕ(x, nk)dγ(x) .
dK|L
n=0 K∈T L∈N (K) K|L
upfort egschemap
Using the equation satisfied by the pressure in (7.47) and the pressure scheme (7.53) yields
NT,k
X X  X Z 
pL − p K
E3,m − E4,m = k unK ( − ∇p(x) · nK,L )(ϕ(x, nk) − ϕ(xK , nk))dγ(x) .
n=0 K∈T K|L dK|L
L∈N (K)

Thanks to the regularity of ϕ and p, there exists C5 > 0, only depending on p, and C6 , only depending
on ϕ, such that, for all K|L ∈ Eint ,
Z
pL − p K pL − p K 1
| − ∇p(x) · nK,L )| ≤ | − ∇p(x) · nK,L dγ(x)| + C5 h, ∀x ∈ K|L
dK|L dK|L m(K|L) σ
and, for all K ∈ T ,

|ϕ(x, nk) − ϕ(xK , nk)| ≤ C6 h, ∀x ∈ K, ∀n ∈ IN.

Thus,
NT,k
X X  X Z 
|E3,m − E4,m | ≤ k |unK | |τK|L (pL − pK ) − ∇p(x) · nK,L dγ(x)| C6 h
n=0 K∈T L∈N (K) K|L
NT,k
X X X
+ k |unK |( m(K|L)C6 C5 h2 ),
n=0 K∈T L∈N (K)
egllesterr1
sumtau
which leads to |E3,m − E4,m | → 0 as m → ∞, using (7.56), (7.75) and the Cauchy-Schwarz inequality.
egcvgce
In order to prove that E2,m → T2 as m → ∞ (which concludes the proof of Theorem 7.1), let us show
that |E2,m − E3,m | → 0 as egschemau
egeq3
m → ∞.
We get, using (7.58) and (7.63)
NT,k
X X
E2,m = − k τK|L (unL − unK )(pL − pK )ϕ(xK , nk)
n=0 K|L∈Eint
NT,k
X X (+)
− k (ūnK − unK )GK ϕ(xK , nk).
n=0 K∈T

This yields
 226

NT,k
X X
E3,m − E2,m = k τK|L (unK − unL )(pL − pK )φnK,L +
n=0 K|L∈Eint
NT,k
(7.79) egeq28
X X (+)
k (unK − ūnK )GK φnK ,
n=0 K∈T

where
Z
1
φnK,L = ϕ(x, nk)dγ(x) − ϕ(xK , nk), ∀K ∈ T , ∀L ∈ N (K)
m(K|L) K|L

and
Z
(+) (+)
GK φnK = ϕ(x, nk)g(x)dγ(x) − GK ϕ(xK , nk).
∂K∩∂Ω

We recall that, for all x ∈ ∂Ω, ϕ(x, nk)g + (x) = ϕ(x, nk)g(x), by definition of D(Ω+ × IR + ). Therefore,
(+) (+)
there exists C7 , which only depends on ϕ, such that |φnK,L | ≤ C7 h and GK |φnK | ≤ GK C7 h, for all
egestbvw
K ∈ T , L ∈ N (K) and all n ∈ IN. Therefore, using Lemma 7.2, we get |E3,m − E2,m | ≤ C7 h √Hh which
egcvgce
yields |E2,m − E3,m | → 0 and then E2,m → T2 as m → ∞. This concludes the proof of Theorem 7.1.

egcvgce
Remark 7.11 In the case of source terms, the convergence theorem 7.1 still holds. There are some minor
modifications in the proof. The definitions of E2,m , E3,m and E4,m change. In the definition of E2,m , the
(+) (−) (+) (−) (+) (−)
quantity GK ūnK − GK unK is replaced by GK ūnK − GK unK + VK snK − VK unK . In the definition of
E3,m one adds
NT,k
X X Z
n n
k (uK − sK ) v + (x)ϕ(x, nk)dx.
n=0 K∈T K

The quantity E3,m − E4,m does not change and in order to prove E3,m − E2,m → 0 it is sufficient to
remark that there exists C8 , only depending on ϕ, such that
Z
(+) (+)
| ϕ(x, nk)v + (x)dx − VK ϕ(xK , nk)| ≤ VK C8 h.
K
 Bibliography

angermann Angermann, L. (1996), Finite volume schemes as non-conforming Petrov-Galerkin approximations

of primal-dual mixed formulations, Report 181, Institüt für Angewandte Mathematik, Universtät
Erlangen-Nurnberg.

AG Amiez, G. and P.A. Gremaud (1991), On a numerical approach to Stefan-like problems, Numer.
Math. 59, 71-89.
Angot Angot, P. (1989), Contribution à l’etude des transferts thermiques dans les systèmes complexes, appli-
cation aux composants électroniques, Thesis, Université de Bordeaux 1.
ABMO Agouzal, A., J. Baranger, J.-F.Maitre and F. Oudin (1995), Connection between finite vol-
ume and mixed finite element methods for a diffusion problem with non constant coefficients, with
application to Convection Diffusion, East-West Journal on Numerical Mathematics., 3, 4, 237-254.
arbogast Arbogast, T., M.F. Wheeler and I. Yotov(1997), Mixed finite elements for elliptic problems with
tensor coefficients as cell-centered finite differences, SIAM J. Numer. Anal. 34, 2, 828-852.
ATT Atthey, D.R. (1974), A Finite Difference Scheme for Melting Problems, J. Inst. Math. Appl. 13, 353-
366.
BR R.E. Bank and D.J. Rose (1986), Error estimates for the box method, SIAM J. Numer. Anal., 777-790
BMO Baranger, J., J.-F. Maitre and F. Oudin (1996), Connection between finite volume and mixed
finite element methods, Modél. Math. Anal. Numér., 30, 3, 4, 444-465.
Barth Barth, T.J. (1994), Aspects of unstructured grids and finite volume solvers for the Euler and Navier-
Stokes equations, Von Karmann Institute Lecture.
belmouhoub Belmouhoub R. (1996) Modélisation tridimensionnelle de la genèse des bassins sédimentaires, Thesis,
Ecole Nationale Supérieure des Mines de Paris, 1996.
BBR Berger, A.E., H. Brezis and J.C.W. Rogers (1979), A Numerical Method for Solving the Problem
ut − ∆f (u) = 0, RAIRO Numerical Analysis, 13, 4, 297-312.
BKP Bertsch, M., R. Kersner and L.A. Peletier (1995), Positivity versus localization in degenerate
diffusion equations, Nonlinear Analysis TMA, 9, 9, 987-1008.
bottahempel Botta, N. and D. Hempel (1996), A finite volume projection method for the numerical solution of the
incompressible Navier-Stokes equations on triangular grids, in: F. Benkhaldoun and R. Vilsmeier eds,
Finite volumes for complex applications, Problems and Perspectives (Hermes, Paris), 355-363.
bre Brezis, H. (1983), Analyse Fonctionnelle: Théorie et Applications (Masson, Paris).
h2 Brun, G., J. M. Hérard, L. Leal De Sousa and M. Uhlmann (1996), Numerical modelling of
turbulent compressible flows using second order models, in: F. Benkhaldoun and R. Vilsmeier eds,
Finite volumes for complex applications, Problems and Perspectives (Hermes, Paris), 338-346.

227
 Bibliography 228

bgh Buffard, T., T. Gallouët and J. M. Hérard (1998), Un schéma simple pour les équations de
Saint-Venant, C. R. Acad. Sci., Série I,326, 386-390.
Cai Cai, Z. (1991), On the finite volume element method, Numer. Math., 58, 713-735.
CMMC Cai, Z., J. Mandel and S. Mc Cormick (1991), The finite volume element method for diffusion
equations on general triangulations, SIAM J. Numer. Anal., 28, 2, 392-402.
Chainais1 Chainais-Hillairet, C. (1996), First and second order schemes for a hyperbolic equation: convergence
and error estimate, in: F. Benkhaldoun and R. Vilsmeier eds, Finite volumes for complex applications,
Problems and Perspectives (Hermes, Paris), 137-144.
Chainais2 Chainais-Hillairet, C. (1999), Finite volume schemes for a nonlinear hyperbolic equation. Convergence
towards the entropy solution and error estimate, Modél. Math. Anal. Numér., 33, 1, 129-156.
cg Champier, S. and T. Gallouët (1992), Convergence d’un schéma décentré amont pour une équation
hyperbolique linéaire sur un maillage triangulaire, Modél. Math. Anal. Numér. 26, 7, 835-853.
cgh Champier, S., T. Gallouët and R. Herbin (1993), Convergence of an Upstream finite volume Scheme
on a Triangular Mesh for a Nonlinear Hyperbolic Equation, Numer. Math. 66, 139-157.
CJ Chavent, G. and J. Jaffré (1990), Mathematical Models and Finite Element for Reservoir Simulation,
Studies in Mathematics and its Applications (North Holland, Amsterdam).
Che Chevrier, P. and H. Galley (1993), A Van Leer finite volume scheme for the Euler equations on
unstructured meshes, Modél. Math. Anal. Numér., 27, 2, 183-201.
ciarlet Ciarlet, P.G. (1978), The Finite Element Method for Elliptic Problems (North-Holland, Amsterdam).
finitelement Ciarlet, P.G. (1991), Basic error estimates for elliptic problems in: Handbook of Numerical Analysis
II (North-Holland,Amsterdam) 17-352.
CIA Ciavaldini, J.F. (1975), Analyse numérique d’un problème de Stefan à deux phases par une méthode
d’éléments finis, SIAM J. Numer. Anal., 12, 464-488.
Coquel Cockburn, B., F. Coquel and P. LeFloch (1994), An error estimate for finite volume methods for
multidimensional conservation laws. Math. Comput. 63, 207, 77-103.
Coquelconv Cockburn, B., F. Coquel and P. LeFloch (1995), Convergence of the finite volume method for
multidimensional conservation laws, SIAM J. Numer. Anal. 32, 687-705. s
C Cockburn, B. and P. A. Gremaud (1996), A priori error estimates for numerical methods for scalar
conservation laws. I. The general approach, Math. Comput. 65,522-554.
cgr Cockburn, B. and P. A. Gremaud (1996), Error estimates for finite element methods for scalar
conservation laws, SIAM J. Numer. Anal. 33, 2, 522-554.
CS Conway E. and J. Smoller (1966), Global solutions of the Cauchy problem for quasi-linear first-order
equations in several space variables, Comm. Pure Appl. Math., 19, 95-105.
clfmuscl Coquel, F. and P. LeFloch (1996), An entropy satisfying MUSCL scheme for systems of conservation
laws, Numer. Math. 74, 1-33.
Putti Cordes C. and M. Putti (1998), Finite element approximation of the diffusion operator on tetrahedra,
SIAM J. Sci. Comput, 19, 4, 1154-1168.
sobolev Coudière, Y., T. Gallouët and R. Herbin (1998), Discrete Sobolev Inequlities and Lp error estimates
for approximate finite volume solutions of convection diffusion equations, submitted.
 Bibliography 229

rouencoudiere Coudière, Y., J.P. Vila, and P. Villedieu (1996), Convergence of a finite volume scheme for a
diffusion problem, in: F. Benkhaldoun and R. Vilsmeier eds, Finite volumes for complex applications,
Problems and Perspectives (Hermes, Paris), 161-168.
coudiere Y. Coudière, J.P. Vila and P. Villedieu (1999), Convergence rate of a finite volume scheme for a
two-dimensional convection diffusion problem, Math. Model. Numer. Anal. 33 (1999), no. 3, 493–516.
croisille Courbet, B. and J. P. Croisille (1998), Finite-volume box-schemes on triangular meshes, M2AN, vol
32, 5, 631-649, 1998.
CM Crandall, M.G. and A. Majda (1980), Monotone Difference Approximations for Scalar Conservation
Laws, Math. Comput. 34, 149, 1-21.
DB Dahlquist, G. and A. Björck (1974), Numerical Methods, Prentice Hall Series in Automatic Compu-
tation.
deimling Deimling, K. (1980), Nonlinear Functional Analysis, (Springer, New York).
DP DiPerna, R. (1985), Measure-valued solutions to conservation laws, Arch. Rat. Mech. Anal., 88, 223-270.
dlf Dubois, F. and P. LeFloch (1988), Boundary conditions for nonlinear hyperbolic systems of conser-
vation laws, Journal of Differential Equations, 71, 93-122.
E Eymard, R. (1992), Application à la simulation de réservoir des méthodes volumes-éléments finis;
problèmes de mise en oeuvre, Cours CEA-EDF-INRIA.
EG Eymard, R. and T. Gallouët (1993), Convergence d’un schéma de type eléments finis - volumes finis
pour un système couplé elliptique - hyperbolique, Modél. Math. Anal. Numér. 27, 7, 843-861.
EG1 Eymard, R. and T. Gallouët (1991), Traitement des changements de phase dans la modélisation de
gisements pétroliers, Journées numériques de Besançon.
eggh Eymard, R., T. Gallouët, M. Ghilani and R. Herbin (1997), Error estimates for the approximate
solutions of a nonlinear hyperbolic equation given by finite volume schemes, IMA Journal of Numerical
Analysis, 18, 563-594.
EGH2 Eymard, R., T. Gallouët and R. Herbin (1995), Existence and uniqueness of the entropy solution
to a nonlinear hyperbolic equation, Chin. Ann. of Math., 16B 1, 1-14.
cvnl Eymard, R., T. Gallouët and R. Herbin, Convergence of finite volume approximations to the
solutions of semilinear convection diffusion reaction equations, Numer. Math., 82, 91-116.
EG2 Eymard R., M. Ghilani (1994), Convergence d’un schéma implicite de type éléments finis-volumes
finis pour un système formé d’une équation elliptique et d’une équation hyperbolique, C.R. Acad. Sci.
Paris Serie I,, 319, 1095-1100.
Fa1 Faille, I. (1992), Modélisation bidimensionnelle de la genèse et la migration des hydrocarbures dans un
bassin sédimentaire, Thesis, Université de Grenoble.
Fa2 Faille, I. (1992), A control volume method to solve an elliptic equation on a two-dimensional irregular
meshing, Comp. Meth. Appl. Mech. Engrg., 100, 275-290.
fh Faille, I. and Heintzé E. (1998), A rough finite volume scheme for modeling two-phase flow in a
pipeline Computers and Fluids 28, 213 - 241.
FFL1 Feistauer M., J. Felcman and M. Lukacova-Medvidova (1995), Combined finite element-finite
volume solution of compressible flow, J. Comp. Appl. Math. 63, 179-199.
 Bibliography 230

FFL2 Feistauer M., J. Felcman and M. Lukacova-Medvidova (1997)], On the convergence of a combined
finite volume-finite element method for nonlinear convection- diffusion problems, Numer. Meth. for
P.D.E.’s, 13, 163-190.
fer Fernandez G. (1989), Simulation numérique d’écoulements réactifs à petit nombre de Mach, Thèse de
Doctorat, Université de Nice.
La Fezoui L., S. Lanteri, B. Larrouturou and C. Olivier (1989), Résolution numérique des équations
de Navier-Stokes pour un Fluide Compressible en Maillage Triangulaire, INRIA report 1033.
thesejmf Fiard, J.M. (1994), Modélisation mathématique et simulation numérique des piles au gaz naturel à
oxyde solide, thèse de Doctorat, Université de Chambéry.
FH1 Fiard, J.M., R. Herbin (1994), Comparison between finite volume finite element methods for the
numerical simulation of an elliptic problem arising in electrochemical engineering, Comput. Meth.
Appl. Mech. Engin., 115, 315-338.
F1 Forsyth, P.A. (1989), A control volume finite element method for local mesh refinement, SPE 18415,
85-96.
F2 Forsyth, P.A. (1991), A control volume finite element approach to NAPL groundwater contamination,
SIAM J. Sci. Stat. Comput., 12, 5, 1029-1057.
FS Forsyth, P.A. and P.H. Sammon (1988), Quadratic Convergence for Cell-Centered Grids, Appl. Num.
Math. 4, 377-394.
rouentg Gallouët, T. (1996), Rough schemes for systems for complex hyperbolic systems, in: F. Benkhaldoun
and R. Vilsmeier eds, Finite volumes for complex applications, Problems and Perspectives (Hermes,
Paris), 10-21.
GH Gallouët, T. and R. Herbin (1994), A Uniqueness Result for Measure Valued Solutions of a Nonlinear
Hyperbolic Equations. Int. Diff. Int. Equations., 6,6, 1383-1394.
ghv Gallouët, T., R. Herbin and M.H. Vignal, (1999), Error estimate for the approximate finite volume
solutions of convection diffusion equations with general boundary conditions, SIAM J. Numer. Anal.,
37, 6, 1935-1972, 2000.
iraultraviart Girault, V. and P. A. Raviart(1986), Finite Element Approximation of the Navier-Stokes Equations
(Springer-Verlag).
Ghidaglia Ghidaglia J.M., A. Kumbaro and G. Le Coq (1996), Une méthode “volumes finis” à flux car-
actéristiques pour la résolution numérique de lois de conservation, C.R. Acad. Sci. Paris Sér. I 332,
981-988.
wski-ellipses Godlewski E. and P. A. Raviart (1991), Hyperbolic systems of conservation laws, Ellipses.
wski-springer Godlewski E. and P. A. Raviart (1996), Numerical approximation of hyperbolic systems of conser-
vation laws, Applied Mathematical Sciences 118 (Springer, New York).
Go Godunov S. (1976), Résolution numérique des problèmes multidimensionnels de la dynamique des gaz
(Editions de Moscou).
GHP Guedda, M., D. Hilhorst and M.A. Peletier (1997), Disappearing interfaces in nonlinear diffusion,
Adv. Math. Sciences and Applications, 7, 695-710.
guo Guo, W. and M. Stynes (1997), An analysis of a cell-vertex finite volume method for a parabolic
convection-diffusion problem, Math. Comput. 66, 217, 105-124.
 Bibliography 231

HA Harten A. (1983), On a class of high resolution total-variation-stable finite-difference schemes, J. Com-

put. Phys. 49, 357-393.
hlvl Harten A., P. D. Lax and B. Van Leer (1983), On upstream differencing and Godunov-type schemes
for hyperbolic conservations laws, SIAM review 25, 35-61.
hh Harten A., J. M. Hyman and P. D. Lax (1976), On finite difference approximations and entropy
conditions, Comm. Pure Appl. Math. 29, 297-322.
Heinrich Heinrich B. (1986), Finite diference methods on irregular networks, I.S.N.M. 82(Birkhauser).
VF4 Herbin R. (1995), An error estimate for a finite volume scheme for a diffusion-convection problem on a
triangular mesh, Num. Meth. P.D.E. 11, 165-173.
rouenrh Herbin R. (1996), Finite volume methods for diffusion convection equations on general meshes, in Finite
volumes for complex applications, Problems and Perspectives, F. Benkhaldoun and R. Vilsmeier eds,
Hermes, 153-160.
HL Herbin R. and O. Labergerie (1997), Finite volume schemes for elliptic and elliptic-hyperbolic prob-
lems on triangular meshes, Comp. Meth. Appl. Mech. Engin. 147,85-103.
HM Herbin R. and E. Marchand (1997), Numerical approximation of a nonlinear problem with a Signorini
condition , Third IMACS International Symposium On Iterative Methods In Scientific Computation,
Wyoming, USA, July 97.
KAM Kamenomostskaja, S.L. (1995), On the Stefan problem, Mat. Sb. 53, 489-514 (1961 in Russian).
keller Keller H. B. (1971), A new difference sheme for parabolic problems, Numerical solutions of partial
differential equations, II, B. Hubbard ed., (Academic Press, New-York) 327-350.
kronerbook Kröner D. (1997) Numerical schemes for conservation laws in two dimensions, Wiley-Teubner Se-
ries Advances in Numerical Mathematics. (John Wiley and Sons, Ltd., Chichester; B. G. Teubner,
Stuttgart).
kroner1 Kröner D. and M. Rokyta (1994), Convergence of upwind finite volume schemes on unstructured
grids for scalar conservation laws in two dimensions, SIAM J. Numer. Anal. 31, 2, 324-343.
kroner2 Kröner D., S. Noelle and M. Rokyta (1995), Convergence of higher order upwind finite volume
schemes on unstructured grids for scalar conservation laws in several space dimensions,Numer. Math.
71, 527-560,.
krushkov Krushkov S.N. (1970), First Order quasilinear equations with several space variables, Math. USSR. Sb.
10, 217-243.
kum Kumbaro A. (1992), Modélisation, analyse mathématique et numérique des modèles bi-fluides
d’écoulement diphasique, Thesis, Université Paris XI Orsay.
Kuznetsov Kuznetsov (1976), Accuracy of some approximate methods for computing the weak solutions of a first-
order quasi-linear equation, USSR Comput. Math. and Math. Phys. 16, 105-119.
LSU Ladyženskaja, O.A., V.A. Solonnikov and Ural’ceva, N.N. (1968), Linear and Quasilinear Equa-
tions of Parabolic Type, Transl. of Math. Monographs 23.
rouenlazarov Lazarov R.D. and I.D. Mishev (1996), Finite volume methods for reaction diffusion problems in: F.
Benkhaldoun and R. Vilsmeier eds, Finite volumes for complex applications, Problems and Perspectives
(Hermes, Paris), 233-240.
lazarovsiam Lazarov R.D., I.D. Mishev and P.S Vassilevski (1996), Finite volume methods for convection-
diffusion problems, SIAM J. Numer. Anal., 33, 1996, 31-55.
 Bibliography 232

leveque LeVeque, R. J. (1990), Numerical methods for conservation laws (Birkhauser verlag).
lions Lions, P. L. (1996), Mathematical Topics in Fluid Mechanics; v.1: Incompressible Models, Oxford
Lecture Series in Mathematics & Its Applications, No.3, Oxford Univ. Press.
mackenzie Mackenzie, J.A., and K.W. Morton (1992), Finite volume solutions of convection-diffusion test
problems, Math. Comput. 60, 201, 189-220.
WM Manteuffel, T., and A.B.White (1986), The numerical solution of second order boundary value
problem on non uniform meshes,Math. Comput. 47, 511-536.
jmm Masella, J. M. (1997), Quelques méthodes numériques pour les écoulements diphasiques bi-fluide en
conduites petrolières, Thesis, Université Paris VI.
mfg Masella, J. M., I. Faille, and T. Gallouët (1996), On a rough Godunov scheme, accepted for
publication in Intl. J. Computational Fluid Dynamics.
MEI Meirmanov, A.M. (1992), The Stefan Problem (Walter de Gruyter Ed, New York).
MEY Meyer, G.H. (1973), Multidimensional Stefan Problems, SIAM J. Num. Anal. 10, 522-538.
mishev Mishev, I.D. (1998), Finite volume methods on Voronoı̈ meshes Num. Meth. P.D.E., 14, 2, 193-212.
morton Morton, K.W. (1996), Numerical Solutions of Convection-Diffusion problems (Chapman and Hall,
London).
mortonsuli Morton, K.W. and E. Süli (1991), Finite volume methods and their analysis, IMA J. Numer. Anal.
11, 241-260.
mortonss Morton, K.W., Stynes M. and E. Süli (1997), Analysis of a cell-vertex finite volume method for a
convection-diffusion problems, Math. Comput. 66, 220, 1389-1406.
necas Nečas, J. (1967), Les méthodes directes en théorie des é quations elliptiques (Masson, Paris).
nicol1 Nicolaides, R.A. (1992) Analysis and convergence of the MAC scheme I. The linear problem, SIAM
J. Numer. Anal.29, 6, 1579-1591.
nicolbook Nicolaides, R.A. (1993) The Covolume Approach to Computing Incompressible Flows, in: Incompress-
ible computational fluid dynamics M.D Gunzburger, R. A. Nicolaides eds, 295–333.
nicol2 Nicolaides, R.A. and X. Wu (1996) Analysis and convergence of the MAC scheme II. Navier-Stokes
equations ,Math. Comput.65, 213, 29-44.
Noelle Noëlle, S. (1996) A note on entropy inequalities and error estimates for higher-order accurate finite
volume schemes on irregular grids, Math. Comput. 65, 1155-1163.
oden Oden, J.T. (1991), Finite elements: An Introduction in: Handbook of Numerical Analysis II (North-
Holland,Amsterdam) 3-15.
Ol2 Oleinik, O.A. (1960), A method of solution of the general Stefan Problem, Sov. Math. Dokl. 1, 1350-
1354.
Ol Oleinik, O. A. (1963), On discontinuous solutions of nonlinear differential equations,Am. Math. Soc.
Transl., Ser. 2 26, 95-172.
O Osher S. (1984), Riemann Solvers, the entropy Condition, and difference approximations, SIAM J.
Numer. Anal. 21, 217-235.
Pat Patankar, S.V. (1980), Numerical Heat Transfer and Fluid Flow, Series in Computational Methods in
Mechanics and Thermal Sciences, Minkowycz and Sparrow Eds. (Mc Graw Hill).
 Bibliography 233

pfertzel Pfertzel, A. (1987), Sur quelques schémas numériques pour la résolution des écoulements diphasiques
en milieux poreux, Thesis, Université de Paris 6.
dfinitelement Roberts J.E. and J.M. Thomas (1991), Mixed and hybrids methods, in: Handbook of Numerical
Analysis II (North-Holland,Amsterdam) 523-640.
roe Roe P. L. (1980), The use of Riemann problem in finite difference schemes, Lectures notes of physics
141, 354-359.
roe2 Roe P. L. (1981), Approximate Riemann solvers, parameter vectors, and difference schemes, J. Comp.
Phys. 43, 357-372.
rudin Rudin W. (1987), Real and Complex analysis (McGraw Hill).
samarski1 Samarski A.A. (1965), On monotone difference schemes for elliptic and parabolic equations in the case
of a non-selfadjoint elliptic operator, Zh. Vychisl. Mat. i. Mat. Fiz. 5, 548-551 (Russian).
samarski2 Samarskii A.A. (1971), Introduction to the Theory of Difference Schemes, Nauka, Moscow (Russian).
samarskilaz Samarskii A.A., R.D. Lazarov and V.L. Makarov (1987), Difference Schemes for differential equa-
tions having generalized solutions, Vysshaya Shkola Publishers, Moscow (Russian).
S Sanders R. (1983), On the Convergence of Monotone Finite Difference Schemes with Variable Spatial
Differencing, Math. Comput. 40, 161, 91-106.
Sel Selmin V. (1993), The node-centered finite volume approach: Bridge between finite differences and finite
elements, Comp. Meth. in Appl. Mech. Engin. 102, 107- 138.
serre Serre D. (1996), Systèmes de lois de conservation (Diderot).

shashkov Shashkov M. (1996), Conservative Finite-difference Methods on general grids, CRC Press, New York.
sod Sod G. A. (1978), A survey of several finite difference methods for systems of nonlinear hyperbolic
conservation laws, J. Comput. Phys., 27, 1-31.
ES Sonier F. and R. Eymard (1993), Mathematical and numerical properties of control-volume finite-
element scheme for reservoir simulation, paper SPE 25267, 12th symposium on reservoir simulation,
New Orleans.
suli Süli E. (1992), The accuracy of cell vertex finite volume methods on quadrilateral meshes, Math. Comp.
59, 200, 359-382.
Sz Szepessy A. (1989), An existence result for scalar conservation laws using measure valued solutions,
Comm. P.D.E. 14, 10, 1329-1350.
temam Temam R. (1977), Navier-Stokes Equations (North Holland, Amsterdam).
samarskitich Tichonov A.N. and A.A. Samarskii (1962), Homogeneous difference schemes on nonuniform nets Zh.
Vychisl. Mat. i. Mat. Fiz. 2, 812-832 (Russian).
T Thomas J.M. (1977), Sur l’analyse numérique des méthodes d’éléments finis hybrides et mixtes, Thesis,
Université Pierre et Marie Curie.
itedifference Thomée V. (1991), Finite differences for linear parabolic equations, in: Handbook of Numerical Analysis
I (North-Holland,Amsterdam) 5-196.
tur Turkel E. (1987), Preconditioned methods for solving the imcompressible and low speed compressible
equations, J. Comput. Phys. 72, 277-298.
 Bibliography 234

vanleer1 Van Leer B. (1974), Towards the ultimate conservative difference scheme, II. Monotonicity and conser-
vation combined in a second-order scheme J. Comput. Phys. 14, 361-370.
vanleer2 Van Leer B. (1977), Towards the ultimate conservative difference scheme, IV. a new approach to
numerical convection. J. Comput. Phys. 23, 276-299.
vanleer3 Van Leer B. (1979), Towards the ultimate conservative difference scheme, V, J. Comput. Phys. 32,
101-136.
Vanselow Vanselow R. (1996), Relations between FEM and FVM, in: F. Benkhaldoun and R. Vilsmeier eds,
Finite volumes for complex applications, Problems and Perspectives (Hermes, Paris), 217-223.
vass Vassileski, P.S., S.I. Petrova and R.D. Lazarov (1992), Finite difference schemes on triangular
cell-centered grids with local refinement, SIAM J. Sci. Stat. Comput. 13, 6, 1287-1313.
Vignal1 Vignal M.H. (1996), Convergence of a finite volume scheme for a system of an elliptic equation and a
hyperbolic equation Modél. Math. Anal. Numér. 30, 7, 841-872.
Vignal2 Vignal M.H. (1996), Convergence of Finite Volumes Schemes for an elliptic hyperbolic system with
boundary conditions,in: F. Benkhaldoun and R. Vilsmeier eds, Finite volumes for complex applications,
Problems and Perspectives (Hermes, Paris), 145-152.
vigver Vignal M.H. and S. Verdière (1998), Numerical and theoretical study of a dual mesh method using
finite volume schemes for two phase flow problems in porous media, Numer. Math., 80, 4, 601-639.
vilathese Vila J.P. (1986), Sur la théorie et l’approximation numérique de problèmes hyperboliques non linéaires.
Application aux équations de saint Venant et à la modélisation des avalanches de neige dense, Thesis,
Université Paris VI.
Vi Vila J.P. (1994), Convergence and error estimate in finite volume schemes for general multidimensional
conservation laws, I. explicit monotone schemes, Modél. Math. Anal. Numér. 28, 3, 267-285.
Vo Vol’pert A. I. (1967), The spaces BV and quasilinear equations, Math. USSR Sb. 2, 225-267.
weiser Weiser A. and M.F. Wheeler (1988), On convergence of block-centered finite-differences for elliptic
problems, SIAM J. Numer. Anal., 25, 351-375.
Index

BV Dirichlet boundary condition, 12–14, 16–21, 32–

– Space, 121, 138, 166, 182–183 37
Initial condition in –, 162, 166, 181, 184 Discontinuous coefficients, 8, 78
Strong – estimate, 128, 138, 139 Discrete
Strong time – estimate, 160–164 – L2 (0, T ; H 1 (Ω)) seminorm, 105
Weak – estimate, 126–128, 134–136, 153– – H01 norm, 40
156, 159–160, 219 – Poincaré inequality, 40, 42, 62, 65–69, 74,
82
CFL condition, 125, 133 – Sobolev inequality, 60–62
Compactness – entropy inequalities, 133, 164–165
– of the approximate solutions in L1 , 142 – maximum principle, 44, 104
– of the approximate solutions in L1loc , 127
– of the approximate solutions in L2 , 46, 74, Edge, 5, 38
75 Entropy
– results in L2 , 92–94 – function, 152
Discrete – results, 29–30 – process solution, 137, 147, 148, 173–179
Helly’s – theorem, 138 – weak solution, 120–122, 146, 173
Kolmogorov’s – theorem, 93 Discrete – inequalities, 164–165
Nonlinear weak – in L∞ , 191 Continuous – inequalities, 120
Weak – in L∞ , 191 Discrete – inequalities, 133
Conservativity, 4, 8, 35, 83, 201 Equation
Consistency Conservation –, 6
– error, 53, 56, 59, 70, 124 Convection-diffusion –, 19, 32, 63, 78, 95,
– in the finite difference sense, 10, 15, 124, 98, 196
125 Diffusion –, 5, 33, 102
– of the fluxes, 6, 13, 24, 25, 35, 36, 202 Hyperbolic –, 119, 145, 190
Weak –, 21, 37 Transport –, 4, 122
Control volume, 4, 5, 12, 33, 37, 39 Error estimate
Control volume finite element method, 9, 85, 213 – for a hyperbolic equation
Convection term, 41, 63, 95, 97, 203, 208, 209 – in the general case, 180, 181
Convergence – in the one dimensional case, 124
– for the weak star topology, 21, 114, 128 – for a parabolic equation, 99
Nonlinear weak star –, 146, 190 – for an elliptic equation
Convergence of the approximate solutions – in the general case, 34, 52, 56, 62, 69,
– towards the entropy process solution, 174 71, 81, 92
– towards the entropy weak solution, 179 – in the one dimensional case, 16, 21, 23
Convergence of the finite volume method Estimate on the approximate solution
– for a linear hyperbolic equation, 129 – for a hyperbolic equation, 125, 128, 132,
– for a nonlinear hyperbolic equation, 136 139, 152, 156, 160, 219
– for a nonlinear parabolic equation, 111 – for a parabolic equation, 99, 104–111
– for a semilinear elliptic equation, 30–31 – for an elliptic equation, 28, 42, 74
– for an elliptic equation, 46, 51, 74 Euler equations, 7, 11, 200–202, 204
Delaunay condition, 39, 53, 85–87, 89

235
Index 236

Finite difference method, 6, 8, 9, 14, 15, 19, 21, Neumann boundary conditions, 63–78
98 Newton’s algorithm, 199, 204–206, 214
Finite element method, 9, 15–16, 37, 53, 84–89,
98, 194, 197, 209, 213 Poincaré
Finite volume finite element method, 88–90 Discrete – inequality, 40, 62, 65–69, 74, 82
Finite volume principles, 4, 6 Poincaré inequality, Neumann boundary condi-
Finite volume scheme tions65
– for elliptic problems Poincaré inequality
– in the one dimensional case, 28 – Dirichlet boundary conditions, 40
– in one space dimension, 13–14, 23, 26 Poincaré inequality , Dirichlet boundary con-
– in two or three space dimensions, 34–37, ditions69, Neumann boundary condi-
41–42, 64, 79–80, 82–84 tions69
– for hyperbolic problems
– in one space dimension, 125, 130 Roe scheme, 200
– in two or three space dimensions, 5, 149,
Scheme
150
Explicit Euler –, 5, 7, 143, 194
– for hyperbolic systems, 200–207
Higher order –, 143–144
– for multiphase flow problems, 212–213, 216–
Implicit – for hyperbolic equations, 147, 156–
218
164
– for parabolic problems, 97–98, 103
Implicit – for parabolic equations, 97, 103,
– for the Stokes system, 209
108, 110
Galerkin expansion, 9, 16, 88, 98, 209 Implicit Euler –, 8, 98
Lax-Friedrichs, 132
Helly’s theorem, 138 Lax-Friedrichs –, 181
Monotone flux –, 131–132, 139
Kolmogorov’s theorem, 93 MUSCL –, 143
Krushkov’s entropies, 121 Roe –, 200
Van Leer –, 143
Lax-Friedrichs scheme, 132, 181 VFRoe –, 201–203, 205
Lax-Wendroff theorem, 140 Singular source terms in elliptic equations, 90–92
Sobolev
Mesh Discrete – inequality, 60, 62
– Refinement, 92 Stability, 15, 20, 22, 28, 44, 74, 96, 104, 122,
Admissible – 132–133, 139, 149, 152, 156
–for Dirichlet boundary conditions, 37 Stabilization of a finite element method, 194
–for Neumann boundary conditions, 63 Staggered grid, 11, 197, 207–209
–for a general diffusion operator, 79 Stokes equations, 209
–for hyperbolic equations, 125, 148
–for regular domains, 114 Transmissibility, 38, 39, 85–87, 89
–in the one-dimensional elliptic case, 12 Two phase flow, 212
Restricted – for Dirichlet boundary con-
ditions, 56 Uniqueness
Restricted – for Neumann boundary con- – of the entropy process solution to a hyper-
ditions, 71 bolic equation, 175
Dual –, 84 – of the solution to a nonlinear diffusion
Moving –, 195–196 equation, 115
Rectangular –, 33 Upstream, 5, 22, 41, 80, 97, 123, 132, 134, 138,
Structured –, 33–35 207, 213, 218
Triangular –, 39
Voronoı̈ –, 39, 86 Van Leer scheme, 143
Voronoı̈, 39, 86
Navier-Stokes equations, 11, 197, 207, 208