Error estimate for the Finite Volume scheme applied

arXiv:math/0509274v2 [[Link]] 12 Jun 2006

to the linear advection equation

Benoı̂t Merlet and Julien Vovelle

Summary. We study the convergence of a Finite Volume scheme for the linear advection
equation with a Lipschitz divergence-free speed in Rd . We prove a h1/2 -error estimate in
the L∞ (0, t; L1 )-norm for BV data. This result was expected from numerical experiments
and is optimal.
Keywords : scalar conservation laws, advection equation, Finite Volume method, error
estimate
Mathematics Subject Classification : 35L65, 65M15

1 Introduction
The Finite Volume method is well adapted to the computation of the solution of pdes
which are conservation (or balance) laws, for the reason that it respects the property of
conservation of the pde under study. The mathematical analysis of the application of the
Finite Volume method to hyperbolic first-order conservation laws can be dated from the
mid sixties (see [TS62] for example). Concerning the specific problem of the estimate of
the rate of convergence of the method, the first result is due to Kuztnetsov [Kuz76], who
proves that this rate of convergence in L∞ (0, t; L1 ) is of order h1/2 , where h is the size
of the mesh, provided that the initial data is in BV and that the mesh is a structured
cartesian grid. Ever since, several studies and results have come to supplement the error
estimate of Kuznetsov. Before describing them, let us emphasize two points:
1. The analysis of the speed of convergence of the Finite Volume method is distinct
from the analysis of the order of the method. In the analysis of the speed of convergence
of the method, general data (e.g. BV data) are considered. Indeed, here, the problem
is to show that the Finite Volume method behaves well regarding the approximation of
the continuous evolution problem in all his features (in particular the creation and the
transport of discontinuities). On the other hand, in the analysis of the order of the
method, restrictions on the regularity of the data are of no importance: see the recent
work of Bouche, Ghidaglia, Pascal [BGP05] on that purpose.
2. If, numerically, the speed of convergence of the Finite Volume method applied to first-
order conservation laws is observed to be (at least) of order h1/2 in the L∞ (0, t; L1 )-norm,

1
whether the mesh is structured or is not, the preexisting theoretical and rigorous proofs
of this result appears to be strongly related to the structure of the mesh.
Indeed, in the case where the mesh is unstructured (what we call a structured mesh is a
cartesian mesh with identical cells but this can be slightly relaxed [CGY98]), the result
of Kuznetsov has been extended, but to the price of a fall in the order of the error esti-
mate. Indeed, h1/4 error estimate in the L∞ (0, t; L1 )-norm for the Finite Volume method
applied to hyperbolic conservation laws on unstructured meshes has been proved by Cock-
burn, Coquel, Lefloch [CCL94], Vila [Vil94] and Eymard, Gallouët, Herbin [EGH00] for
the Cauchy Problem (h1/6 error estimate for the Cauchy-Dirichlet Problem [OV04]). We
emphasize the fact that numerical tests give an order h1/2 for structured as well as un-
structured meshes; still, concerning these latter, numerical analysis did not manage to
give the rigorous proof of the order h1/2 : there is an upper limit at the order h1/4 .
In this paper, we consider the case where the conservation law is linear. More precisely,
we consider the linear advection problem with a Lipschitz divergence-free speed.
In the case of H 1 -data, optimal (with respect to the order of the error estimate) results
already exist. Under a strict CFL condition, Després [Des04c] shows an h1/2 -error estimate
in the L∞ 2
t Lx -norm for the upwind Finite Volume method applied to the linear advection
equation with constant speed. (The author deals with the particular case of 2D triangular
meshes but extensions to higher dimension and general polyhedral meshes are harmless).
His technique is based on the study of the dissipation of the consistency error by the
scheme and the adaptation of the Lax Theorem (see also [Des04a]). √Roughly speaking,
the consistency error created at time nδt is of order 1 but of order 1/ q + 1 after q time
steps of the scheme.
On the other hand, we refer to the work of Vila and Villedieu [VV03], who prove, again
under a strict CFL condition, an h1/2 -error estimate in the L2loc space-time norm for the
approximation of Friedrichs hyperbolic systems with H 1 data by energy estimates (they
consider explicit in time Finite Volume schemes, for implicit schemes in the case of scalar
advection equation, see, as they underline it, the result of Johnson and Pitkäranta [JP86]
who show an h1/2 -error estimate in the L2 space-time norm for H 1 data).
For an initial data u0 ∈ BV (Rd ), we prove in Theorem 2 the expected h1/2 -error estimate
in the L∞ 1
t Lx norm, under a strict CFL condition in the general case and, under a sharp
CFL condition if the speed is independent of t. This result is optimal [TT95, Şab97] and
in the context of unstructured meshes and BV -data, this is the first optimal result.
In our proof of Theorem 2, we first show that it is sufficient to consider initial data which
are characteristic functions. Then we use the same ingredients as Vila and Villedieu
in [VV03]. The approximate solution satisfies the weak formulation of the problem up to
an error term corresponding to the consistency error of the scheme. Roughly speaking,
this error is bounded via the Cauchy-Schwarz inequality by a H 1 (Rd × [0, t])-like semi-
norm of the discrete solution denoted Eh (u0 , t). A large error means that the scheme is
very dissipative and consequently, that the approximate solution is smooth, so Eh (u0 , t)
is small. The h1/2 -estimate follows from the equilibrium between these contradictory
constraints.

2
The paper is divided into six parts: first we continue this introduction by describing the
linear advection problem, the Finite Volume scheme, our results and we give the main
lines of the proof. In Section 2 we recall some classical results on the Finite Volume
scheme. In Section 3 we introduce the weak formulation satisfied by the approximate
solution. Then we reduce the study to initial data which are characteristic functions and
we build the test function used in the weak formulation. In Section 4, we prove some
Energy estimates. The first and second part of Theorem 2 are proved in Sections 5, 6
respectively.

Notations
If (X, µ) is a measurable set with finite (positive) measure and φ ∈ L1 (X), we denote the
mean of φ over X by
1
Z Z
− φdµ := φdµ.
X µ(X) X

If X is a set, 1X (x) = 1 if x ∈ X, 0 if x ∈
/ X.
The set of functions with bounded variation in Rd is the set of L1 functions with bounded
Radon measures as derivatives:
( Z )
BV (Rd ) := u ∈ L1 (Rd ); sup u(x)divϕ(x)dx < +∞
ϕ∈Cc∞ (Rd ,Rd ),kϕk∞ ≤1 U

p
where ||ϕ||∞ := || ϕ21 + · · · + ϕ2d ||L∞ (Rd ) for ϕ ∈ Cc∞ (Rd , Rd) and div is the divergence
operator. The total variation of u ∈ BV (Rd ) is given by
Z
kukT V := sup u(x)divϕ(x)dx .
ϕ∈Cc∞ (Rd ,Rd ),kϕk∞ ≤1 Rd

1.1 The linear advection equation

We consider the linear advection problem in Rd :

 ut + div(V u) = 0, x ∈ Rd , t ∈ R+ ,
(1.1)
u(x, 0) = u0 (x), x ∈ Ω,


where we suppose that V ∈ W 1,∞ (Rd × R+ , Rd) satisfies

div V (·, t) = 0 ∀t ∈ R+ . (1.2)

The problem (1.1) has a solution for u0 ∈ L1loc (Rd ); for the purpose of the error estimates,
we will consider initial data in BV (Rd ).

3
Theorem 1. For every u0 ∈ L1loc (Rd ), the problem (1.1) admits a unique weak solution
u, in the sense that u ∈ L1loc (R+ × Rd ) and: for every φ ∈ Cc∞ (Rd × R+ ),
Z Z Z
u(φt + V · ∇φ)dxdt + u0 φ(x, 0)dx = 0. (1.3)
R+ Rd Rd

Moreover, we have

u(x, t) = u0 (X(x, t)), ∀(x, t) ∈ Rd × R+ , (1.4)

where X ∈ C 1 (Rd × R+ , Rd ) is such that, for every t ≥ 0, X(·, t) : Rd → Rd is one to

one and onto.
Let Y (·, t) := X(·, t)−1 , then for every T ≥ 0, X −IdRdx and Y −IdRdx belong to W 1,∞ (Rd ×
[0, T ]). More precisely, there exists C0 ≥ 1 only depending on T , kV kW 1,∞ and d, such
that, ∀x ∈ Rd , ∀t ∈ [0, T ],

kX(x, t) − xk ≤ C0 t, k∇X(x, t)k ≤ C0 , k∂t X(x, t)k ≤ C0 , (1.5)

kY (x, t) − xk ≤ C0 t, k∇Y (x, t)k ≤ C0 , k∂t Y (x, t)k ≤ C0 . (1.6)

Finally, for every t ≥ 0, X(·, t) preserves the Lebesgue measure λ on Rd , i.e:

λ(X(E, t)) = λ(E), for every Borel subset E of Rd . (1.7)

Proof of Theorem 1 All the results cited in the Theorem follow from the characteristic
formula: u(x, t) = u0 (X(x, t)) where X(·, t) = Y (·, t)−1 and Y solves the Cauchy Problem
∂t Y (x, t) = V (Y (x, t), t), Y (x, 0) = x. To prove the estimates on X, we notice that
X(x, t) = Z(0; x, t) where Z(τ ; x, t) denotes the solution of the Cauchy Problem
dZ


 (τ ; x, t) = V (Z(τ ; x, t), τ ), τ ∈ [0, t],
dτ

Z(t; x, t) = x.


We only give the sketch of the proof: global existences for Y and Z and the desired
estimates follow from the Cauchy-Lipschitz Theorem and the Gronwall Lemma. Besides,
by (1.2), the flow preserves the Lebesgue measure on Rd . Existence for (1.1) follows from
the characteristic formula and, by an argument of duality, uniqueness also.
If u0 ∈ L1loc (Rd ), the solution u belongs to C(R+ , L1loc (Rd )). If the derivatives of u0 are
bounded Radon measures, we have a more precise result. Namely, using the estimates
above, we obtain
Corollary 1.1. Let T ≥ 0, there exists a constant C0 ≥ 1 depending on kV kW 1,∞ , T and
d such that, if u0 ∈ BV (Rd ):

ku(·, s) − u(·, t)kL1 ≤ C0 ku0kT V |s − t|, 0 ≤ s, t ≤ T,

ku(·, t)kT V ≤ C0 ku0kT V , 0 ≤ t ≤ T.

4
The following result is a direct consequence of the conservative property (1.7). The case
f (v) = v 2 will be crucial in the proof of the main result of this paper.
Corollary 1.2. Let f : R → R be Za measurable function and u0 ∈ L1loc (Rd ) such that
f ◦ u0 ∈ L1 (Rd ). Then the quantity f (u(·, t)) is constant.
Rd

1.2 Finite Volume scheme

The Finite Volume scheme which approximates (1.1) is defined on a mesh T which is a
family of closed connected polygonal subsets with disjoint interiors covering Rd ; the time
half-line is meshed by regular cells of size δt > 0. We also suppose that the partition T
satisfies the following properties: the common interface of two control volumes is included
in an hyperplane of Rd and
αhd ≤ |K|,

there exists α > 0 such that (1.8)
|∂K| ≤ α−1 hd−1 , ∀K ∈ T ,

where h is the size of the mesh: h := sup{diam(K), K ∈ T }, |K| is the d-dimensional

Lebesgue measure of K and |∂K| is the (d − 1)-dimensional Lebesgue measure of ∂K. If
K and L are two control volumes having a common edge, we say that L is a neighbor of
K and denote (quite abusively) L ∈ ∂K. We also denote Kp L the common edge and nKL
the unit normal to Kp L pointing outward K. We set
Z Z (n+1)δt
n
VKL := − V · nKL
KpL nδt

We denote by Kn := K × [nδt, (n + 1)δt) a generic space-time cell, by M := T × N the

space-time mesh and by ∂Kn− the set

∂Kn− := {L ∈ ∂K, VKL

n
< 0}.

In the same way we set Kp Ln := Kp L × [nδt, (n + 1)δt).

We will assume that the so called Courant-Friedrich-Levy condition is satisfied:
X
n
δt|VKL | ≤ (1 − ξ)|K|, ∀Kn ∈ M, for some ξ ∈ [0, 1). (1.9)
−
L∈∂Kn

Remark 1.1. Under condition (1.8), the CFL condition (1.9) holds as soon as

kV k∞ δt ≤ (1 − ξ)α−2 h.

Notice that if kV k∞ is small, we may choose a large time step δt. In order to avoid the
occurrence of terms with factor δt h−1 in our estimates, we add to (1.9) the following
condition: there exists c0 ≥ 0 such that

δt ≤ c0 h. (1.10)

5
The Finite Volume scheme with explicit time-discretization is defined by the following set
of equations:
Z
0
uK = − u0 (x) dx , ∀K ∈ T , (1.11)
K

un+1
K − uK
n
1 X n
+ VKL (unL − unK ) = 0, ∀Kn ∈ M. (1.12)
δt |K| −
L∈∂Kn

We then denote by uh the approximate solution of (1.1) defined by the Finite Volume
scheme:
uh (x, t) = unK , ∀(x, t) ∈ Kn . (1.13)

Main results
From now on, we assume that V ∈ W 1,∞ (Rd × R+ , Rd ) satisfies (1.2). We fix a mesh T
of mesh-size h > 0 satisfying the uniformity conditions (1.8). We also fix a time step δt
such that the CFL conditions (1.9)-(1.10) hold.
We fix two timesP 0 ≤ tN +1
≤ T and we assume that t = (N + 1)δt for some N ∈ N (so
that uh (·, t) = K∈T uK 1K ). In order to avoid the occurrences of h t−1 terms in the
estimates we assume that

h ≤ c1 t, for some c1 ≥ 0. (1.14)

In the sequel, given an initial condition u0 ∈ BV (Rd ), the function u ∈ C(R+ , L1 (Rd ))
denotes the exact solution to (1.1) and uh its numerical approximation obtained by the
upwind Finite Volume method (1.11)-(1.12)-(1.13).
In the results and in the proofs, C0 ≥ 1 is the constant introduced in Theorem 1 and
Corollary 1.1; this constant only depends on T , kV kW 1,∞ and d. The letter C denotes
various constants which are non decreasing functions of α−1 , c0 , c1 , kV kW 1,∞ and d but
do not depend on h, δt, ξ, u0 , t or T .
Finally, before stating the main results of the paper, we set

MN := {Kn ∈ M : 0 ≤ n ≤ N},

and we introduce the quantities

X X X
Eh (u0 , t) := |K||un+1 n 2
K − uK | +
n
δt|VKL | |unL − unK |2 ,
Kn ∈MN Kn ∈MN L∈∂Kn
−

Eh (u0 , t) := kuh (·, 0)k2L2 − kuh (·, t)k2L2 .

Remark 1.2. Notice that h−2 Eh (u0 , t) is a discrete version of the quantity
Z d
!
X
|∂t uh (x, s)|2 + |Vi (x, s)||∂xi uh (x, s)|2 dxds.
Rd ×[0,t] i=1

6
We have the following error estimate:

Theorem 2. Let u0 in BV (Rd ). Under a strict CFL condition ((1.9) with ξ > 0), we
have :

ku(·, t) − uh (·, t)kL1 ≤ CC0d+2 ξ∗−1 ku0 kT V (t1/2 h1/2 + ξ∗1/2 th), (1.15)

with ξ∗ = ξ.
Moreover, if V does not depend on the time variable, the estimate is valid with ξ∗ = 1
uniformly in ξ ∈ [0, 1).

This result deserves some comments.

Remark
R s 1.3. If the speed does not depend on x: V (x, s) := V(s), we have X(x, s) =
x − 0 V(r)dr and we can choose C0 = d1/2 + kVk∞ (which does not depend on T ) in
Theorem 1 and Corollary 1.1.
Remark 1.4. In the proof of the first part of Theorem 2, the dependency on ξ only appears
in Lemma 5.1.
Remark 1.5 (On the BV -norm of the approximate solution). The possible irregularity of
the mesh is an obstacle to the existence of uniform (with respect to h) bound on the
L1t BVx -norm X X
δt|Kp L||unL − unK |
Kn ∈MN L∈∂Kn
−

of the approximate solution uh . A counter-example to such a result has been given by

B. Després in [Des04b]. However, a uniform bound on the weaker norm (notice the weights
n
VKL ) X X
n
Qh (u0 , t) := δt|VKL ||unL − unK |
Kn ∈MN L∈∂Kn
−

would be enough to prove the error estimate (1.15) by use of the techniques of Kuznetsov
[Kuz76] (and, therefore, regardless whether the conservation law under consideration is
linear, the remark here is true). This uniform bound on Qh (u0 , t) is observed in numerical
experiments and is probably true. However, in this paper, we do not prove this claim and
it remains an open problem, even in the linear setting.
Interpolating between the h1/2 -error estimate in the L∞ 2
t Lx -norm of [Des04c] and the
estimate of Theorem 2, we obtain, under a strict CFL condition, a h1/2 -error estimate in
the L∞ p
t Lx -norm for data in W
1,p
(Rd ), 1 ≤ p ≤ 2.
The analogue result for initial data in W 1,p (Rd ), 2 ≤ p ≤ ∞, requires different techniques
and will be addressed in a subsequent paper.

1.3 Sketch of the proof of Theorem 2

We only sketch the proof of the first part of the Theorem; besides we do not exhibit the
dependency on t and ξ of the error estimates (the constants C may depend on t, T or ξ).

7
To prove Theorem 2, we use the fact that the difference between the exact and the
approximate solutions satisfies the weak formulation (Lemma 3.1):
Z Z
(u − uh )(φt + V · ∇φ) + (u − uh )(·, 0)φ(x, 0)dx = (µh + νh )(φ), (1.16)
Rd ×R+ R

for every φ ∈ Cc1 (Rd ×R+ ), where the error term (µh +νh )(φ) encompasses the consistency
error of the Finite Volume approximation and depends on the approximate solution.
We then notice that, by linearity of the equations, and subsequent principle of superpo-
sition, we may suppose that the initial data u0 ∈ BV (Rd ) is the characteristic function
u0 = 1A of a set A with finite perimeter. The norm ku0kT V is equal to the perimeter of A.
In fact, we show that we can also reduce the study to the case where A does not contain
very thin parts so that, in particular, the volume of the h1/2 -neighborhood of ∂A

A0 := x ∈ Rd : d(x, ∂A) ≤ h1/2 .



is bounded by C|∂A|h1/2 . At this stage, we build a function φ0 ∈ C 1 (Rd , [−1, 1]) such
that φ0 ≡ 1 on A \ A0 , φ0 ≡ −1 on Ac \ A0 and k∇φ0 k∞ ≤ Ch−1/2 . Then we set
φ(x, s) = φ0 (X(x, s))1[0,t] (s) so that φt + V · φ = 0 on Rd × [0, t]. After a regularization
process, the weak formulation yields

k(uh − u)(·, t)kL1 ≤ C|∂A|h1/2 + |µh (φ)| + |νh (φ)|.

The term µh (φ) is

X Z Z 
µh (φ) = |K||un+1
K − unK | − φ − − φ(·, (n + 1)δt) .
K∈MN Kn K

The purpose of the reduction to the case u0 = 1A is an accurate estimation of the term
µ(φ). Indeed, by definition of φ, we have: for 0 ≤ s ≤ t, φ(x, s) = 1 if x ∈ X(A \ A0 , s)
and φ(x, s) = −1 if x ∈ X(Ac \ A0 , s). More precisely, there exists C > 0 such that, if

d(Kn , ∪0≤s≤t (X(∂A, s) × {s}) ≥ C(h1/2 + h + δt),

then φ is constant on the cell Kn . We denote by M1N the set of cells which do not satisfy
this property; we have
X Z Z 
n+1 n
µh (φ) = |K||uK − uK | − φ − − φ(·, (n + 1)δt) .
Kn K
K∈M1N

and the (d + 1)-dimensional volume of the cells of M1N is bounded by C|∂A|h1/2 (Lemma
3.5). We use the Cauchy-Schwarz inequality and the estimate k∂t φk∞ ≤ Ch−1/2 to get

|µh (φ)| ≤ CE(u0 , t)1/2 |∂A|1/2 h1/4 .

A similar technique yields the same bound on νh (φ).

8
Thanks to the Energy estimate (Lemma 5.1) Eh (u0 , t) ≤ CEh (u0 , t), we get

k(uh − u)(·, t)kL1 ≤ C |∂A|h1/2 Eh (u0 , t)1/2 + |∂A|1/2 h1/4 .

Finally, since the L2 -norm of the exact solution is conserved, we get (Lemma 4.1):

Eh (u0 , t) ≤ 2k(uh − u)(·, t)kL1 .

Together with the preceding inequality, this yields successively Eh (u0 , t) ≤ C|∂A|h1/2 and
k(uh − u)(·, t)kL1 ≤ C|∂A|h1/2 .

2 Classical results
By continuity in BV of the L2 -projection on the space of functions which are constant
with respect to the mesh T , we have:
Lemma 2.1. There exists a constant c ≥ 0 only depending on d such that for every
u0 ∈ BV (Rd ),
1 X
kuh (·, 0)kT V = |Kp L| |u0L − u0K | ≤ cα−3 ku0kT V ,
2 K∈M, L∈∂K
kuh (·, 0) − u0 kL1 ≤ cα−1 ku0kT V h.

Proof of Lemma 2.1: Using the weak density of Cc1 (Rd ) in BV (Rd ), we may suppose
that u0 ∈ Cc1 (Rd ). Let K ∈ T , L ∈ ∂K. Since |x − y| ≤ 2h for every (x, y) ∈ K × L, we
have
1
Z Z
0 0
|uK − uL | ≤ |u0 (x) − u0 (y)|dxdy
|K| |L| K L
Z Z Z 1
2h
≤ |∇u0 ((1 − r)x + ry)|drdxdy.
|K| |L| K L 0
Now we perform the change of variables: (x, y, r) 7→ (w = x − y, z = (1 − t)x + ty, r = r)
(of Jacobian determinant equal to 1). We have
Z 1 Z
2h
Z 
0 0
|uK − uL | ≤ |∇u0(z)| g(w, z, r)dwdr dz,
|K| |L| B(xK ,2h) 0 Rd

where xK ∈ Rd is the centroid of K, and g is defined by g(w, z, r) = 1 if z + rw ∈ K and

Rz − (1 − r)w ∈ L, and d
g(w, z, r) = 0 otherwise. For (z, r) ∈ B(xK , 2h) × [0, 1], we have
z, r)dw ≤ 2d |L| if r < 1/2. Finally, we
R
Rd
g(w, z, r)dw ≤ 2 |K| if r ≥ 1/2 and Rd
g(w,
obtain
2d+1 max(|K|, |L|)h
Z
0 0
|uK − uL | ≤ |∇u0 (z)|dz.
|K| |L| B(xK ,2h)

9
We deduce from the inequalities (1.8) that for every K ∈ T ,
X Z
0 0 d+1 −2
|Kp L||uK − uL | ≤ 2 α |∇u0 (z)|dz.
L∈∂K B(xK ,2h)

Summing on K ∈ T , we get
X X Z
|Kp L||u0K − u0L | ≤ 2 d+1
α −2
|∇u0(z)|M2h (z)dz,
K∈T L∈∂K Rd

where M2h (z) is the cardinal of the set {K : d(K, z) ≤ 2h}. By (1.8), we have

M2h (z)αhd ≤ K : d(K,z)≤2h |K| ≤ |B(z, 3h)|.

P

Thus M2h (z) ≤ cα−1 and we get the first part of the Lemma.
Let K ∈ T . Similarly, we have
1
Z Z
|uh (·, 0) − u0 | ≤ |u0 (x) − u0 (y)|dxdy
K |K| K×K
Z
d
≤ 2 h |∇u0(z)|dz.
B(xK ,h)

Summing on K ∈ T and using the fact that the cardinal of the set {K : d(K, z) ≤ h} is
bounded by cα−1 , we get the second estimate.
The divergence free assumption (1.2) leads to the following identity
Lemma 2.2. Under the hypothesis on the divergence of V (1.2), one has
X X
n n
VKL = VLK ∀ Kn ∈ M.
−
L∈∂Kn L : K∈∂L−
n

Monotony
Under a CFL condition the Finite Volume scheme is order-preserving:
Proposition 2.1. Under condition (1.9), the linear application L : (unK ) 7→ (un+1
K ) defined
∞
by (1.12) is order-preserving and stable for the L -norm.
Proof of Proposition 2.1: Eq. (1.12) gives
 
X V n δt X V n δt
KL  n KL
un+1
K =  1 + u K − unL (2.1)
−
|K| −
|K|
L∈∂Kn L∈∂Kn

i.e., under (1.9), un+1

K is a convex combination of unK , (unL )L∈∂Kn− . The stability result
follows from L(1) = 1.

10
L1-stability
From Lemma 2.2 it is not difficult to see that the quantity K∈T |K|unK is conserved. By
P
the Crandall-Tartar Lemma [CT80], this fact and Proposition 2.1 imply
Proposition 2.2. Under condition (1.9), the scheme L : (unK ) 7→ (un+1
K ) is stable for the
L1 -norm:
X X
|K||un+1
K | ≤ |K||unK |, ∀n ≥ 0.
K∈T K∈T

3 Weak Formulation
In the remainder of the paper, we assume that u0 ∈ BV (Rd ), that the function u ∈
C(R+ , L1 (Rd )) is the exact solution to (1.1) and uh is the numerical approximation given
by the scheme (1.11)-(1.12)-(1.13). We also assume that (1.8), the CFL conditions (1.9)-
(1.10) and the condition (1.14) are satisfied.
We intend to prove that uh satisfies (1.3) up to an error term.
Lemma 3.1. For every φ ∈ Cc∞ (Rd × [0, +∞)),
Z Z Z
uh (φt + V · ∇φ) + uh (·, 0)φ(·, 0) = µh (φ) + νh (φ), (3.1)
R+ Rd Rd

X
µh (φ) := |K|(un+1 n n
K − uK )(hφiK − hφiK ((n + 1)δt)), (3.2)
Kn ∈M
X X
νh (φ) := δt(unL − unK )(VKL
n
hφinK − |Kp L|hV · n φinKL ), (3.3)
Kn ∈M L∈∂Kn
−

where
Z Z Z
hφinK := − φ , hφiK (t) := − φ(·, t) , hV · n φinKL := − φV · nKL .
Kn K KpLn

Proof of Lemma 3.1: We develop the first term of the inequality and perform a discrete
integration by parts:
Z ∞Z X Z
n
uh (φt + V · ∇φ)dxdt = |K|δt uK − (φt + V · ∇φ)dxdt
0 Rd Kn ∈M Kn
!
X   X
= unK |K| hφiK ((n + 1)δt) − hφiK (nδt) + δt|Kp L|hV · n φinKL
Kn ∈M L∈∂K
!
X X
= |K|(unK − un+1
K )hφiK ((n + 1)δt) + δt|Kp L|hV · n φinKLunK
Kn ∈M L∈∂K
X
− |K|δtu0K hφiK (0). (3.4)
Kn ∈M

11
 R
The last sum is equal to Rd uh (·, 0)φ(·, 0). Rearranging the terms in the second sum and
using the identity hV · n φinKL + hV · n φinLK = 0, we compute: for every n ≥ 0
X X X X
|Kp L|hV · n φinKL unK = |Kp L|hV · n φinKL (unK − unL ).
K∈T L∈∂K K∈T L∈∂Kn
−

Plugging this identity and the definition of the scheme (1.12) in (3.4), we get (3.1).
We now assume that u0 ∈ BV (Rd ). Before building the test function φ, we show that we
may reduce the study to more simple initial data.
Let T ′ be the cartesian mesh
( d
)
Y
t1/2 h1/2 [pk , pk + 1] : (p1 , · · · , pd ) ∈ Zd .
k=1

The diameter of the cells is h′ = d1/2 t1/2 h1/2 ; this mesh is uniform: conditions (1.8) are
satisfied for α′ = min(d−d/2 , d(d−3)/2 /2). In particular, by Lemma 2.1, we have

kv0 kT V ≤ Cku0 kT V , kv0 − u0 kL1 ≤ Cku0 kT V t1/2 h1/2 ,

X Z 
where v0 is the projection defined by v0 := − u0 1K ′ . These estimates, the L1 -
K ′ ∈T ′ K′
stability of the scheme and of the evolution equation (1.1) imply that it is sufficient to
prove Theorem 2 for the initial data v0 .
Lemma 3.2. [Fed69, Bre84] Let v0 ∈ BV (Rd ). For every η ∈ R, define

 1 if 0 < η < v0 (x),
χv0 (x, η) := −1 if v0 (x) < η < 0,
0 in the other cases.

Z
We have v0 = χv0 (·, η)dη, almost everywhere. Moreover,
R
Z
kv0 kL1 = kχv0 (·, η)kL1 dη,
ZR
kv0 kT V = kχv0 (·, η)kT V dη,
R

where the second identity is a consequence of the co-area formula for BV functions.

Applying this decomposition, we deduce from the linearity of the scheme and of the initial
problem, that it is sufficient to prove the Theorem for any initial data u0 which is a finite
sum of characteristic functions 1K ′ .
Therefore, from now on, we suppose that u0 = 1A with A = ∪1≤j≤l Kj′ and K1′ , · · · , Kl′
are distinct elements of T ′ .

12
Let us point out that
ku0 kL1 = |A| = ltd/2 hd/2 ,
and since, the boundary ∂A is a finite union of distinct edges denoted K1′ p L′1 , · · · , Km
′
p L′m ,
we deduce from the definition of the Total Variation:
m
X
ku0 kT V = |∂A| = |Ki′ p L′i | = mt(d−1)/2 h(d−1)/2 .
i=1

With the notations of Theorem 1, we have u = 1B where B ⊂ Rd × R+ is defined by

[
B := Y (A, s) × {s}.
s≥0

We are now going to build a test function φ for (1.3) and (3.1). Let Γ ∈ C ∞ (R, R)
such that 0 ≤ Γ ≤ 1, Γ(R) = 0 for R ≤ 1/3 and Γ(R) = 1 for R ≥ 2/3. We define
Φ0 ∈ Lip(Rd , R) by

Γ(t−1/2 h−1/2 d(x, ∂A)) if


x ∈ A,
Φ0 (x) :=
−Γ(t−1/2 h−1/2 d(x, ∂A)) if x ∈ Ac .

Since d(·, ∂A) is a 1-Lipschitz continuous function, we have k∇Φ0 k∞ ≤ t−1/2 h−1/2 kΓ′ k∞ .
Let us introduce a mollifier ρ ∈ Cc∞ (Rd , R+ ) such that Rd ρ = 1 and supp ρ ⊂ B(0, 1/3),
R

we set φ0 := t−d/2 h−d/2 ρ(t−1/2 h−1/2 · ) ⋆ Φ0 . The function φ0 belongs to C 1 (Rd ) with the
bound:

k∇φ0 k∞ ≤ t−1/2 h−1/2 kΓ′ k∞ .

Moreover, we have

x ∈ A =⇒ φ0 (x) ≥ 0, and x ∈ Ac =⇒ φ0 (x) ≤ 0,

and if we split Rd in three disjoint subsets: Rd = φ−1 −1 −1

0 ({1}) ∪ φ0 ((−1, 1)) ∪ φ0 ({−1}) =:
A+ ∪ A0 ∪ A− , we have u0 ≡ 1 on A+ , u0 ≡ 0 on A− and

A0 ⊂ x : d(x, ∂A) < t1/2 h1/2 ⊂ ∪m 1/2 1/2


i=1 B(xi , 1/2 t h ),

where xi ∈ Rd is the centroid of Ki′p L′i , 1 ≤ i ≤ m. Consequently the Lebesgue measure

of A0 satisfies:

|A0 | ≤ Cmtd/2 hd/2 = Cku0kT V t1/2 h1/2 (3.5)

Finally, we define our test-function by φ(x, s) := φ0 (X(x, s))1[0,t] (s) for (x, s) ∈ Rd × R+ ,
where X is the mapping introduced in Theorem 1. Since divV (·, s) ≡ 0 for every s ≥ 0,
we have

φt + V · ∇φ = φt + div(V φ) = 0 in Rd × (0, t),

13
and from the estimate on ∇φ0 above and on ∇X and ∂t X (1.5) in Theorem 1, φ belongs
to C 1 (Rd × [0, t]) and satisfies

k∇φkL∞ (Rd ×[0,T ]) ≤ CC0 t−1/2 h−1/2 , (3.6)

k∂t φkL∞ (Rd ×[0,T ]) ≤ CC0 t−1/2 h−1/2 . (3.7)

Moreover, since u(x, s) = u0 (X(x, s)), φ inherits the properties of φ0 : for (x, s) ∈ Rd ×R+ ,

u(x, s) = 1 =⇒ φ(x, s) ≥ 0, and u(x, s) = 0 =⇒ φ(x, s) ≤ 0. (3.8)

Lemma 3.3. We have

Z Z
(u − uh )(·, t)φ(·, t) − (u − uh )(·, 0)φ(·, t) = µh (φ) + νh (φ). (3.9)
Rd Rd

Proof of Lemma 3.3: The test function φ is not compactly supported and not time
differentiable at the time t. So it can not be used directly in (1.3), (3.1). First, remark that,
since u0 is compactly supported, there exists R > 0 such that supp uh (·, s), supp u(·, s) ⊂
B(0, R) for every 0 ≤ s ≤ t + δt, thus φ may be replaced in (1.3), (3.1) by the compactly
supported function φ · χ where χ ∈ Cc∞ (Rd ) is such that χ ≡ 1 on B(0, R).
To overcome the non differentiability at t we introduce a mollifying sequence to ap-
proximate the function 1[0,t] . Let (ψq )q≥1 be a sequence of Cc∞ (R+ ) functions satis-
fying ψq ≡ 1 on [0, t], ψq ≡ 0 on [t + δt/q, +∞) and 0 ≥ ψq′ ≥ −2q/δt. We set
φq (x, s) := φ0 (X(x, s))ψq (s) for every (x, s) ∈ Rd × R+ and q ≥ 1. The weak formu-
lation (3.1) reads
Z t+δt Z  Z
′
uh (x, s)φ0 (X(x, s))dx ψq (s)ds + uh (·, 0)φ0(·, 0) = µh (φq ) + νh (φq ),
t Rd Rd

The sequence of Radon measures (ψq′ (s)ds)q converges to −δt , so the first term of the left
hand side converges towards Z
− uh (·, t)φ(·, t)
Rd
as q tends to +∞.
Clearly, for every Kn ∈ M and L ∈ ∂Kn− , we have (hφq iK )nq → hφinK , (hV · nφq inKL )q →
hV · nφinKL and (hφq iK (nδt))q → hφiK (nδt). Passing to the limit on q, we get
Z Z
− uh (·, t)φ(·, t) + uh (·, 0)φ0 = µh (φ) + νh (φ).
Rd Rd

Similarly, we obtain from (1.3)

Z Z
− u(·, t)φ(·, t) + u(·, 0)φ0 = 0.
Rd Rd

Subtracting these equalities, we get the result.

We relate the first term of the equality of Lemma 3.3 to the L∞ 1
t Lx -norm of the error:

14
Lemma 3.4. We have
Z
k(uh − u)(·, t)kL1 ≤ (u − uh )(·, t)φ(·, t) + Cku0 kT V t1/2 h1/2 . (3.10)
Rd

Proof of Lemma 3.4: Recall that from the monotony of the scheme, we have

0 ≤ uh (x, s) ≤ 1, ∀ (x, s) ∈ Rd × R+ . (3.11)

Let us consider the disjoint decomposition Rd = X(A+ , t)∪X(A0 , t)∪X(A− , t) and notice
that (3.8), (3.11) imply that (u(x, t) − uh (x, t))φ(x, t) ≥ 0 for every x ∈ Rd . In particular
Z
0 ≤ (u − uh )(·, t)φ(·, t).
X(A0 ,t)

Since 0 ≤ u, uh ≤ 1, we deduce from (3.5) and the conservation property (1.7) that
Z
|(uh − u)(·, t)| ≤ |X(A0 , t)| = |A0 | ≤ Cku0kT V t1/2 h1/2 .
X(A0 ,t)

Finally, using the identities φ(·, t) ≡ 1, u(·, t) ≡ 1 on X(A+ , t) and φ(·, t) ≡ −1, u(·, t) ≡ 0
on X(A− , t) and the bounds 0 ≤ uh ≤ 1, we have
Z Z Z
|uh (·, t) − u(·, t)| = (u − uh )(·, t) + (uh − u)(·, t)
X(A+ ∪A− ,t) X(A+ ,t) X(A− ,t)
Z
= (u − uh )(·, t)φ(·, t).
X(A+ ∪A− ,t)

Summing up this equality and the two preceding estimates, we get (3.10).
Lemmas 3.3 and 3.4 and the second estimate of Lemma 2.1 (and h ≤ c1 t) yield

k(uh − u)(·, t)kL1 ≤ µh (φ) + νh (φ) + Cku0 kT V t1/2 h1/2 . (3.12)

The first key idea to estimate µh (φ) and νh (φ) is to remark that most of the terms in the
sums of (3.2), (3.3) vanish.
Lemma 3.5. There exists M1N ⊂ MN such that
X
µh (φ) = |K|(un+1 n n
K − uK )(hφiK − hφiK ((n + 1)δt)),
Kn ∈M1N
X X
νh (φ) = δt(unL − unK )(VKL
n
hφinK − |Kp L|hV · n φinKL ).
Kn ∈M1N L∈∂Kn
−

with the estimate

X
δt|K| ≤ CC0d ku0 kT V t3/2 h1/2 . (3.13)
Kn ∈M1N

15
Proof of Lemma 3.5: Let us denote by B the open subset of Rd × [0, t] of all the points
(x, s) such that |φ(x, s)| < 1. Let Kn ∈ MN which does not intersect B. In this case, φ
is constant (φ ≡ 1 or φ ≡ −1) on Kn and we have

hφinK = hφiK ((n + 1)δt) and n

VKL hφinKL = |Kp L|hV · n φinKL, ∀ L ∈ ∂Kn− .

Consequently, the corresponding terms in the definitions of µh and νh vanish and the
equalities of the Lemma hold with

M1N := {Kn ∈ MN : Kn ∩ B =
6 ∅} ,

We now estimate the (d + 1)-dimensional Lebesgue measure of the union of the cells
Kn ∈ M1N . From the definition of φ, for every Kn ∈ M1N , there exist (x, s) ∈ Kn such
that d(X(x, s), ∂A) ≤ t1/2 h1/2 . Since ∂A = ∪1≤i≤m Ki′ p L′i , for every Kn ∈ M1N , there
exist (xK,n , sKn ) ∈ Kn and 1 ≤ i ≤ m such that d(X(xKn , sKn ), xi ) ≤ 2t1/2 h1/2 , where xi
is the centroid of Ki′ p L′i . Thus using the estimates of (1.6) on ∇Y and ∂t Y , we obtain

[ m [
[
(xKn , sKn ) ⊂ Y (B(xi , 2t1/2 h1/2 ), s) × {s}.
Kn ∈M1N i=1 0≤s≤t
m [
[ N
B Y (xi , nδt), 2C0t1/2 h1/2 + C0 δt) × [nδt, (n + 1)δt],


⊂
i=1 n=0

and
[ m [
[ N
B Y (xi , nδt), 2C0 t1/2 h1/2 + C0 δt + h) × [nδt, (n + 1)δt].


Kn ⊂
Kn ∈M1N i=1 n=0

Finally, using C0 ≥ 1, δt ≤ c0 h and h ≤ c1 t, we get

X
δt|K| ≤ CC0d mt(td/2 hd/2 + hd ) ≤ CC0d ku0 kT V t3/2 h1/2 .
Kn ∈M1N

4 Energy Estimates
The two next lemmas contain the second key idea of the paper. The first one shows that
the L1 -error at time t controls Eh (u0 , t) (the jump of the L2 -energy of the approximate
solution). The second one relates Eh (u0 , t) to the regularity of the approximate solution.

Lemma 4.1. We have

Eh (u0 , t) ≤ 2 k(uh − u)(·, t)kL1 .

16
Proof of Lemma 4.1: Using the conservation property of Corollary 1.2 with f (v) = v 2 ,
we have ku(·, t)k2L2 = ku0 k2L2 . Thus

Eh (u0 , t) = kuh (·, 0)k2L2 − ku0k2L2 + ku(·, t)k2L2 − kuh (·, t)k2L2 .

Since uh (·, 0) is the L2 -projection of u0 on the mesh T , the first term is non positive.
Finally, 0 ≤ u, uh ≤ 1 yields u2 (·, t) − u2h (·, t) ≤ 2 |uh (·, t) − u(·, t)|.
In particular, (3.12) implies

Eh (u0 , t) ≤ 2µh (φ) + 2νh (φ) + Cku0kT V t1/2 h1/2 . (4.1)

Lemma 4.2. We have the following identity

 
X VKM n
X X δt
Eh (u0 , t) = 1 + n
 |VKL |δt (unL − unK )2
K ∈M − −
|K|
n N L∈∂Kn M ∈∂Kn
n n
1 X X VKL VKM δt2 n
+ (uM − unL )2 . (4.2)
2 K ∈M −
|K|
n N L,M ∈∂Kn

Remark 4.1. We say that the identity of Lemma 4.2 is an Energy estimate, because of
the analogy with the a priori inequality
Z tZ
2 2
kuν (·, 0)kL2 − kuν (·, t)kL2 ≤ 2ν |∇uν (x, s)|2 dxds,
0 Rd

where uν satisfies the advection-diffusion equation ∂t uν + div(V uν ) − ν∆uν = 0. In our

case the diffusion is due to the scheme (and does not have the regularity and the isotropy
of the Laplace operator), and the corresponding parameter ν is of order h.
Proof of Lemma 4.2: Let n ≥ 0. Using Lemma 2.2 to change the index of summation
in the last sum of the right hand-side of the first equality, we have
 
X X X X X
|K|(unK )2 = |K| + n
VKL δt (unK )2 − n
VLK δt(unK )2
K∈T K∈T L∈∂K −
n
K∈T L : K∈∂L−
n
  
X X X 
n
= |K| + VKL δt (unK )2 − n
VKL δt(unL )2
 
K∈T L∈∂Kn− −
L∈∂Kn
X X
= |K| anKL (unL )2 ,
K∈T L∈Vn− (K)

where we have set for every Kn ∈ M, Vn− (K) := ∂Kn− ∪ {K}, and for every L ∈ ∂Kn− :
X V n δt n
VKL δt
KM
anKK := 1 + , anKL := − .
−
|K| |K|
M ∈∂Kn

17
With this notations, (1.12) reads
X
un+1
K = anKL unK ,
L∈Vn− (K)

for every Kn ∈ M, and we have

  2 
X X  X X 
|K| (unK )2 − (un+1 2 n n 2 n n


K ) = |K| a (u
KL L ) −  a u
KL L
 − 
K∈T K∈T L∈Vn (K) L∈Vn− (K)
 
X  X X 
n n n 2 n n n n
= |K| aKL (1 − aKL )(uL ) − aKL aKM uL uM .
 − 
K∈T L∈Vn (K) −
L,M ∈Vn (K), M 6=L

anKL = 1, we compute
P
Using the identity L∈Vn− (K)
  
X X  X X 
|K|((unK )2 − (un+1 2
K ) ) = |K| anKL unL  n n n 
aKM (uL − uM )
 
K∈T K∈T L∈Vn− (K) M ∈Vn− (K)

1X X
= |K| anKL anKM (unL − unM )2 .
2 K∈T
L,M ∈Vn− (K)

Summing on 0 ≤ n ≤ N, we get the Lemma.

5 Proof of the first part of Theorem 2

In this section, we assume that ξ > 0. We estimate successively µh (φ) and νh (φ). Applying
the Cauchy-Schwarz inequality to the first formula of Lemma 3.5, we obtain
 !2 1/2
X hφinK − hφiK ((n + 1)δt)
|µh (φ)| ≤ Eh (u0 , t)1/2  |K|δt2  .
δt
Kn ∈M1N

On the basis of (3.7) and the CFL condition (1.10) δt ≤ c0 h, we have

|hφinK − hφiK ((n + 1)δt)| ≤ k∂t φkL∞ δt ≤ CC0 δt t−1/2 h−1/2 ≤ CC0 h1/2 t−1/2

and therefore, by (3.13)

1+d/2 1/2
|µh (φ)| ≤ CC0 Eh (u0 , t)1/2 ku0 kT V t1/4 h1/4 .

18
We now write
X X
νh (φ) = δt(unL − unK )VKL
n
(hφinK − hφinKL)
Kn ∈M1N L∈∂Kn
−

X X
+ δt(unL − unK )|Kp L| (hV · ninKL hφinKL − hV · n φinKL) ,
Kn ∈M1N L∈∂Kn
−

=: I + II. (5.1)

From the Cauchy-Schwarz inequality, we have

 !2 1/2
X X hφinK − hφinKL
|I| ≤ Eh (u0 , t)1/2  n
|VKL |δth2  .
h
Kn ∈M1N L∈∂Kn
−

The conditions (1.8) implies

X
n
|VKL |h ≤ α−2 kV k∞ |K| ≤ C|K|, ∀ Kn ∈ M.
−
L∈∂Kn

Consequently, (3.6) and (3.13) lead to

1+d/2 1/2
|I| ≤ CC0 Eh (u0 , t)1/2 ku0 kT V t1/4 h1/4 . (5.2)

We now estimateZ the term II; notice that it vanishes in case V = Cst. Let Kn ∈ M1N and
L ∈ ∂Kn− , since − hV · ninKL − V · nKL = 0, we have
KpLn

Z
hV · ninKL hφinKL − hV · n φinKL= − φ(x, r) (hV · ninKL − V (x, r) · nKL ) dxdr
Z KpLn
= − (φ(x, r) − hφinKL) (hV · ninKL − V (x, r) · nKL ) dxdr.
KpLn

Using (3.6)-(3.7) and |hV · ninKL − V (x, r) · nKL | ≤ Ch for (x, r) ∈ Kp Ln , we get

|hV · ninKL hφinKL − hV · n φinKL| ≤ CC0 t−1/2 h3/2 .

Thus, we have
   
X X
|II| ≤ CC0   |Kp L|h δt t−1/2 h1/2 .
Kn ∈M1N −
L∈∂Kn
P
As above, for every Kn ∈ M we have −
L∈∂Kn |Kp L|h ≤ C|K| and from (3.13):

|II| ≤ CC0d+1 ku0kT V th. (5.3)

19
Summing up (5.1)-(5.2)-(5.3), we obtain
 
d/2+1 1/2 1/2 1/4 1/4 d+1
|µh (φ)| + |νh (φ)| ≤ C C0 Eh (u0 , t) ku0 kT V t h + C0 ku0 kT V th , (5.4)

and (4.1) yields

 
d/2+1 1/2
Eh (u0 , t) ≤ C C0 Eh (u0 , t)1/2 ku0 kT V t1/4 h1/4 1/2 1/2
+ ku0 kT V (t h + C0d+1 th) . (5.5)

The following Lemma will allow us to bound Eh (u0 , t) by Eh (u0 , t).

Lemma 5.1.

Eh (u0 , t) ≤ Cξ −1 Eh (u0 , t).

Proof of Lemma 5.1: From (1.12) and the Cauchy-Schwarz inequality, we have for
every Kn ∈ M,
 
X |V n |δt X
KL
|K||un+1
K − u n 2
K | ≤   n
|VKL |δt|unL − unK |2 .
−
|K| −
L∈∂Kn L∈∂Kn

n
P
Due to the CFL condition (1.9), we have L∈∂Kn− |VKL |δt ≤ |K|; consequently, summing
the preceding inequality on Kn ∈ MN , we obtain
X X
n
Eh (u0 , t) ≤ 2 |VKL |δt|unL − unK |2 .
Kn ∈MN L∈∂Kn
−

P
Using again the CFL condition, for every Kn ∈ M we have (1+ M ∈∂Kn− VKM δt/|K|) ≥ ξ.
Thus ξEh (u0, t)/2 is bounded by the first sum of the right hand side of (4.2).
Plugging this inequality in (5.5), we get
 
d/2+1 1/2
Eh (u0, t) ≤ C C0 Eh (u0 , t)1/2 ξ −1/2 ku0 kT V t1/4 h1/4 + ku0 kT V (t1/2 h1/2 + C0d+1 th) .

We then use the Young inequality ab ≤ 21 a2 + 12 b2 with

d/2+1 −1/2 1/2
a = C −1/2 Eh (u0 , t)1/2 , b = C 1/2 C0 ξ ku0 kT V t1/4 h1/4

to derive the estimate

Eh (u0 , t) ≤ CC0d+2 ku0 kT V (ξ −1 t1/2 h1/2 + th).

Finally, from Lemma 5.1 and the inequalities (5.4), (3.12), we compute successively

Eh (u0 , t) ≤ CC0d+2 ξ −2ku0 kT V ( t1/2 h1/2 + ξ th),

|µh (φ)| + |νh (φ)| ≤ CC0d+2 ξ −1ku0 kT V (t1/2 h1/2 + ξ 1/2 th),
k(uh − u)(·, t)kL1 ≤ CC0d+2 ξ −1ku0 kT V (t1/2 h1/2 + ξ 1/2 th).

20
6 Proof of the Second part of Theorem 2
We suppose that V does not depend on the time variable. If
X
sup |K|−1 0
|VKL |δt < 2/3,
K∈T
L∈∂K0−

then the proof P

of the first case allows us to conclude; therefore we can suppose that we
⋆
have |K | ≤ 2 L∈∂K ⋆,− |VK0 ⋆ L |δt for some K ⋆ ∈ T . This cell K ⋆ satisfying (1.8), we
0
deduce that the following inverse CFL condition holds:

h ≤ 2α−2kV k∞ δt. (6.1)

We now bound µh (φ) and νh (φ). From Lemma 3.5 and the estimate (3.7) on ∂t φ, we have
X
|µh (φ)| ≤ Cδt(t−1/2 h−1/2 ) |K||un+1 n
K − uK |.
Kn ∈M1N

Since V does not depend on t, we have VKn = VK0 , for every Kn ∈ M. Consequently the
n
discrete time derivative vK := δt−1 (un+1 n 1
K −uK ) satisfies the scheme (1.12). The L -stability
of the scheme and Lemma 2.1 imply
X t X X
|K| un+1 n
K − uK ≤ |K||u1K − u0K | ≤ t 0
|VKL ||u0K − u0L |
Kn ∈MN
δt K∈T K∈T
X
≤ tkV k∞ |Kp L||u0K − u0L | ≤ Cku0 kT V t.
K∈T

Thus (using the CFL condition δt ≤ c0 h), we have

|µh (φ)| ≤ Cku0 kT V t1/2 h1/2 . (6.2)

As in the proof of the first case (see (5.1)), we write νh (φ) = I + II. Recall that
X X
I= δt(unL − unK )VKL
n
(hφinK − hφinKL) ,
Kn ∈M1N L∈∂Kn
−

and that, from the estimate (3.6) on ∇φ, we have

X X
|I| ≤ CC0 h1/2 t−1/2 n
δt|VKL ||unL − unK |.
Kn ∈M1N L∈∂Kn
−

Using the definition of the scheme (1.12), for Kn ∈ M1N and L ∈ ∂Kn− , we have

|K| 1 X
unL − unK = P un+1 n n
δt(unL − unM ).


n K − uK + P n
VKM
− V δt − V δt
M ∈∂Kn KM M ∈∂Kn KM −
M ∈∂Kn

21
Therefore,

X
|I| ≤ CC0 h1/2 t−1/2 |K||un+1 n
K − uK |
Kn ∈M1N
!
X 1 X
n n
+ P n
VKL VKM δt2 |unM − unL |
−
M ∈∂Kn |VKM |δt
Kn ∈M1N −
L,M ∈∂Kn

= CC0 h1/2 t−1/2 (I1 + I2 ).

Using again the L1 -stability of the scheme, we have
|I1 | ≤ Cku0 kT V t.
For I2 , we use the Cauchy-Schwarz inequality to get
 1/2
X VKL n n 2
X VKM δt
|I2 | ≤  (unM − unL )2 
Kn ∈MN L,M ∈∂Kn
−
|K|
 1/2
n n 2
X X VKL VKM δt |K|
×
2 
n
P
− V δt
1
Kn ∈MN L,M ∈∂Kn − M ∈∂Kn KM
 1/2  1/2
X VKL n n 2
X VKM δt X
=  (unM − unL )2   |K| .
K ∈M −
|K| 1
n N L,M ∈∂Kn Kn ∈MN
√
Thanks to Lemma 4.2, the first term is bounded by 2Eh (u0 , t)1/2 and by (3.13), the
d/2 1/2
second term is bounded by CC0 ku0 kT V t3/4 h1/4 δt−1/2 . Finally, using the inverse CFL
d/2 1/2
condition (6.1), |I2 | is bounded by CC0 ku0kT V t1/4 h1/4 Eh (u0 , t)1/2 and we have
 
d/2+1 1/2
|I| ≤ CC0 ku0 kT V t1/2 h1/2 + Eh (u0 , t)1/2 ku0 kT V t1/4 h1/4 . (6.3)

The proof of the estimate (5.3) of the previous Section is still valid and we have |II| ≤
CC0d+1 ku0kT V th. Summing up this estimate, (6.2) and (6.3), we obtain

|µh (φ)| + |νh (φ)|

 
d/2+1 d/2 1/2
≤ CC0 ku0kT V (t1/2 h1/2 + C0 th) + Eh (u0 , t)1/2 ku0 kT V t1/4 h1/4 . (6.4)

Plugging (6.4) in (4.1) and using Young inequality to absorb the term Eh (u0 , t) in the left
hand side, we are led to
Eh (u0 , t) ≤ CC0d+2 ku0 kT V (t1/2 h1/2 + th).
The second part of the Theorem follows from the last estimate, (6.4) and (3.12).

22
References
[BGP05] D. Bouche, J.-M. Ghidaglia, and F. Pascal, Error estimate and the geometric
corrector for the upwind finite volume method applied to the linear advection
equation, preprint (2005).
[Bre84] Y. Brenier, Averaged multivalued solutions for scalar conservation laws, SIAM
J. Numer. Anal. 21 (1984), no. 6, 1013–1037.
[CCL94] B. Cockburn, F. Coquel, and P. LeFloch, An error estimate for finite vol-
ume methods for multidimensional conservation laws, Math. Comp. 63 (1994),
no. 207, 77–103.
[CGY98] B. Cockburn, P.-A. Gremaud, and J. X. Yang, A priori error estimates for
numerical methods for scalar conservation laws. III. Multidimensional flux-
splitting monotone schemes on non-Cartesian grids, SIAM J. Numer. Anal.
35 (1998), no. 5, 1775–1803 (electronic).
[CT80] M.G. Crandall and L. Tartar, Some relations between nonexpansive and order
preserving mappings, Proc. Amer. Math. Soc. 78 (1980), no. 3, 385–390.
[Des04a] B. Després, Lax theorem and finite volume schemes, Math. Comp. 73 (2004),
no. 247, 1203–1234 (electronic).
[Des04b] B. Després, An explicit a priori estimate for a finite volume approximation of
linear advection on non-Cartesian grids, SIAM J. Numer. Anal. 42 (2004), no. 2,
484–504 (electronic).
[Des04c] B. Després, Convergence of non-linear finite volume schemes for linear trans-
port, Notes from the XIth Jacques-Louis Lions Hispano-French School on Nu-
merical Simulation in Physics and Engineering (Spanish). (2004), 219–239.
[EGH00] R. Eymard, T. Gallouët, and R. Herbin, Finite volume methods, Handbook of
numerical analysis, Vol. VII, North-Holland, Amsterdam, 2000, pp. 713–1020.
[Fed69] H. Federer, Geometric measure theory, Die Grundlehren der mathematischen
Wissenschaften, Band 153, Springer-Verlag New York Inc., New York, 1969.
[JP86] C. Johnson and J. Pitkäranta, An analysis of the discontinuous Galerkin method
for a scalar hyperbolic equation, Math. Comp. 46 (1986), no. 173, 1–26.
[Kuz76] N. N. Kuznetsov, The accuracy of certain approximate methods for the compu-
tation of weak solutions of a first order quasilinear equation, Ž. Vyčisl. Mat. i
Mat. Fiz. 16 (1976), no. 6, 1489–1502, 1627.
[OV04] M. Ohlberger and J. Vovelle, Error estimate for the approximation of non-linear
conservation laws on bounded domains by the finite volume method, to appear
in Math. of Comp.

23
[Şab97] F. Şabac, The optimal convergence rate of monotone finite difference methods
for hyperbolic conservation laws, SIAM J. Numer. Anal. 34 (1997), no. 6, 2306–
2318.

[TS62] A. N. Tihonov and A. A. Samarskiı̆, Homogeneous difference schemes on irreg-

ular meshes, Ž. Vyčisl. Mat. i Mat. Fiz. 2 (1962), 812–832.
√
[TT95] T. Tang and Z. H. Teng, The sharpness of Kuznetsov’s O( ∆x) L1 -error esti-
mate for monotone difference schemes, Math. Comp. 64 (1995), no. 210, 581–
589.

[Vil94] J.-P. Vila, Convergence and error estimates in finite volume schemes for gen-
eral multidimensional scalar conservation laws. I. Explicit monotone schemes,
RAIRO Modél. Math. Anal. Numér. 28 (1994), no. 3, 267–295.

[VV03] J.-P. Vila and P. Villedieu, Convergence of an explicit finite volume scheme for
first order symmetric systems, Numer. Math. 94 (2003), no. 3, 573–602.

24