- ESAIM: Mathematical Modelling and Numerical

Comments

Transcription

- ESAIM: Mathematical Modelling and Numerical
Mathematical Modelling and Numerical Analysis
ESAIM: M2AN
Vol. 35, No 2, 2001, pp. 239–269
Modélisation Mathématique et Analyse Numérique
CONVERGENCE OF FINITE DIFFERENCE SCHEMES FOR VISCOUS
AND INVISCID CONSERVATION LAWS WITH ROUGH COEFFICIENTS
Kenneth Hvistendahl Karlsen 1 and Nils Henrik Risebro 2
Abstract. We consider the initial value problem for degenerate viscous and inviscid scalar conservation laws where the flux function depends on the spatial location through a “rough” coefficient function
k(x). We show that the Engquist-Osher (and hence all monotone) finite difference approximations converge to the unique entropy solution of the governing equation if, among other demands, k0 is in BV ,
thereby providing alternative (new) existence proofs for entropy solutions of degenerate convectiondiffusion equations as well as new convergence results for their finite difference approximations. In the
inviscid case, we also provide a rate of convergence. Our convergence proofs are based on deriving a
series of a priori estimates and using a general Lp compactness criterion.
Mathematics Subject Classification. 65M06, 35L65, 35L45, 35K65.
Received: September 6, 2000.
1. Introduction
The main subject of this paper is finite difference schemes for computing the entropy solution of scalar viscous and inviscid conservation laws where the transport term depends explicitly on the spatial location. Such
equations are of the form
ut + divf (k, u) = ∆A(u),
u(x, 0) = u0 (x),
(x, t) ∈ ΠT = Rd × (0, T ),
(1.1)
where the flux function f (k, u) = (f1 (k 1 , u), . . . , fd (k d , u)) depends on the spatial location through the coefficient k = k(x),
k(x) = (k 1 (x), . . . , k d (x)).
For the initial value problem (1.1) to be well-posed, we must require that the nonlinear elliptic operator
u 7→ ∆A(u) satisfies the degenerate ellipticity condition
A(·) nondecreasing with A(0) = 0.
(1.2)
Keywords and phrases. Conservation law, degenerate convection-diffusion equation, entropy solution, finite difference scheme,
convergence, error estimate.
1
Department of Mathematics, University of Bergen, Johs. Brunsgt. 12, 5008 Bergen, Norway. e-mail: [email protected];
URL: http://www.mi.uib.no/~kennethk/
2 Department of Mathematics, University of Oslo, P.O. Box 1053, Blindern, 0316 Oslo, Norway. e-mail: [email protected];
URL: http://www.math.uio.no/~nilshr/
c EDP Sciences, SMAI 2001
240
K.H. KARLSEN AND N.H. RISEBRO
Note that (1.2) implies that many well known nonlinear partial differential equations are special cases of (1.1).
In particular (1.2) includes as special cases the inviscid conservation law, the heat equation, one-point degenerate porous medium type equations [43], two-point degenerate oil reservoir flow equations [15], and strongly
degenerate convection-diffusion equations of the type arising in the theory of sedimentation-consolidation processes [4].
We recall that if (1.1) is allowed to degenerate at certain points, that is, A0 (s) = 0 for some values of s,
solutions are not necessarily smooth (but typically continuous) and weak solutions must be sought. On the other
hand, if A0 (s) is zero on an interval [α, β], (weak) solutions may be discontinuous and they are not uniquely
determined by their initial data. Consequently, an entropy condition must be imposed to single out the physically correct solution. Roughly speaking, we call a function u ∈ L1 ∩ L∞ an entropy solution of (1.1) if


∂t |u − c| + div [sign (u − c) (f (k, u) − f (k, c))] − ∆ |A(u) − A(c)|
(i)
(1.3)
+sign (u − c) divf (k, c) ≤ 0 in D0 ∀c ∈ R,


2
(ii)
∇A(u) belongs to L .
We refer to Section 2.1 for a more precise statement of the definition of an entropy solution as well as precise conditions on u0 , f, k, A ensuring that this definition makes sense. Relevant mathematical (existence and
uniqueness) theory for entropy solutions can be found in [3, 5, 28, 42].
For the hyperbolic equation, the convergence analysis of numerical schemes has very long traditions and
goes back to the 1950s. Being extremely selective, we mention only a few references related to finite difference
and finite volume approximations. The case of finite difference schemes have been treated by Oleı̆nik [40],
Harten et al. [24], Kuznetsov [36], Crandall and Majda [12], Sanders [44], Lucier [37], Osher and Tadmor [41],
Cockburn and Gremaud [10], and many others. The study of finite volume methods is more recent and have been
conducted by Champier et al. [7], Vila [48], Cockburn et al. [8, 9], Kröner and Rokyta [32], Kröner et al. [31],
Noelle [38], Eymard et al. [21], and Chainais-Hillairet [6], as well as many others. Among the cited papers,
only [6, 7, 21, 40] treat equations where the nonlinearity f depends on the spatial position x (and time t).
Although there seems to be an increasing interest in the (analysis of) numerical approximation of entropy
solutions of degenerate parabolic equations, the amount of literature on the subject is at the moment modest.
The (very recent) literature include papers by Evje and Karlsen [18], Holden et al. [25], and Holden et al. [26]
on operator splitting methods (see also the lecture notes by Espedal and Karlsen [15]); Evje and Karlsen [16,
17, 19, 20] on upwind difference schemes; Kurganov and Tadmor [35] on central difference schemes; Bouchut
et al. [2] on kinetic BGK schemes; Afif and Amaziane [1] and Ohlberger [39] on finite volume methods; and
Cockburn and Shu [11] on the local discontinuous Galerkin method. Strictly speaking, the authors of [1, 11, 35]
do not analyze their numerical methods within an entropy solution framework.
It is somewhat surprising that there have been few attempts up to very recently (confer the list of references
given above) to develop a systematic treatment of mixed hyperbolic-parabolic partial differential equations
within a unified mathematical (entropy solution) framework. In fact, the construction and analysis of numerical
methods for first order hyperbolic and second order parabolic equations are usually considered as separate
subject areas. In this work we demonstrate that it is possible to give a coherent treatment of numerical
methods for such large class of nonlinear partial differential equations. Our main long-term goal is indeed to
develop a consistent (mathematical/numerical) framework which is the same whether we are working with the
hyperbolic case (A0 ≡ 0), the parabolic case (A0 > 0), or with the mixed hyperbolic-parabolic case (A0 ≥ 0). In
the present paper (see also [16,17,19,20]), we are concerned with finite difference schemes and their convergence
analysis. For related work on other numerical methods for strongly degenerate parabolic equations, see the list
of references given above.
To illustrate the results of this paper, we now state them in the one dimensional case (i.e., d = 1). For the
general results, we refer to Sections 3 and 4. As a model difference scheme for (1.1), we consider the generalized
upwind (Engquist-Osher) scheme
f EO ki+1/2 , uni , uni+1 − f EO ki−1/2 , uni−1 , uni
A uni+1 − 2A (uni ) + A uni−1
− uni
un+1
i
+
=
,
∆t
h
h2
DIFFERENCE SCHEMES WITH ROUGH COEFFICIENTS
241
where the so-called Engquist-Osher numerical flux [14] takes the form
Z
f EO (k, u1 , u2 ) =
u1
Z
fu (k, s) ∨ 0 ds +
0
u2 fu (k, s) ∧ 0 ds + f (k, 0).
0
Here and in the following, a ∨ b = max {a, b} and a ∧ b = min {a, b}. Note that k and u are discretized on grids
that are staggered with respect to each other. Concretely, we set
u∆t (x, t) =
X
χ[xi−1/2 ,xi+1/2 i×[tn ,tn+1 i uni ,
k∆t =
i,n
X
χ[xi ,xi+1 i ki+1/2 ,
i
where ki+1/2 = k xi+1/2 and xi = ih.
We have chosen to analyse the above scheme because of its simplicity. One can, however, adopt the method
of proof developed in this paper and obtain similar results for other schemes (e.g., all monotone schemes) as
well as more general equations. For example, under the assumption that k ≡ 1, Evje and Karlsen [20] have
studied high order difference schemes (based on the MUSCL idea) for degenerate parabolic equations with
source terms. One can easily combine the ideas in the present paper with those in [20] and obtain high order
difference schemes for degenerate parabolic equations with source terms and k = k(x) non-constant. Moreover,
one can easily treat the case where k possesses also a temporal dependence, i.e., k = k(x, t). Although we
consider only explicit schemes in this paper, one can adopt the techniques used herein to analyse semi-implicit
and implicit schemes (the details will be presented elsewhere).
Assuming that u0 , k and k 0 are in BV , we are able to show that the approximate solutions {u∆t}∆t>0
generated by our scheme converge strongly in L1loc as ∆t ↓ 0 to the unique entropy solution. Furthermore, in
the hyperbolic case, we show that this convergence has a rate. More precisely, we prove
Theorem 1.1. Let u∆t denote the function generated by the Engquist-Osher scheme. We assume that the
time step ∆t is related to the spatial step h through an appropriate CFL condition. If u0 , k and k 0 are in
BV ∩ L1 ∩ L∞ , then
u = lim u∆t
∆t↓0
is the unique entropy solution to (1.1). Furthermore, if A0 ≡ 0, then
ku(·, t) − u∆t (·, t)kL1 (R) = O
√ ∆t .
We remark that Theorem 1.1 provides an existence result for entropy solutions of strongly degenerate parabolic
equations which complements those in [5, 50]. We also remark that the question of a convergence rate for the
difference approximations to degenerate parabolic equations will be addressed elsewhere.
We now relate our results to the ones obtained by Evje and Karlsen [17,19], who analyse monotone difference
approximations of (1.1) in the special case k ≡ 1. In this case, the authors gave a fairly complete analysis for
the one-dimensional equation under certain smoothness assumptions on the initial function u0 , in which case
it actually holds that A(u) belongs to the Hölder space C 1,1/2 (R × [0, T ]), and not merely L2 (0, T ; H 1 (R)) as
follows from our analysis. We mention also the work [16] which generalizes the analysis in [17, 19] to the more
difficult case of doubly nonlinear degenerate parabolic equations. In the present paper, we dispense with most of
the smoothness assumptions on u0 used in [17, 19]. Moreover, in the multidimensional case, the authors of [19]
do not prove that the limit u of their monotone difference approximations satisfies (ii) in (1.3), a result that
can be easily established by adopting the techniques developed in the present paper.
We continue with a few words about the proof of Theorem 1.1. The proof of the first part of Theorem 1.1
is based on deriving uniform L∞ , L1 , and BV bounds on the approximate solution u∆t . Equipped with the
BV bound, we use the difference scheme itself and Kružkov’s interpolation lemma [33] to show that u∆t is
uniformly L1 continuous in time. Kolmogorov’s compactness criterion then immediately gives L1loc convergence
242
K.H. KARLSEN AND N.H. RISEBRO
(along a subsequence) of {u∆t }∆t>0 to a function u ∈ L1 ∩ L∞ . Uniqueness of the entropy solution [5, 28] (see
also Th. 2.1 herein) will imply that the whole sequence {u∆t}∆t>0 converges and not just a subsequence.
To ensure that the limit u is the (unique) entropy solution in the sense of (1.3), we first prove that the
difference scheme satisfies a so-called discrete (or cell) entropy inequality and hence it follows, by arguments
analogous to the ones used to prove the Lax-Wendroff theorem, that the entropy condition (i) in (1.3) holds
true for the limit u. In passing, we mention that the BV regularity and the cell entropy inequalities are used
to derive the error estimate in the hyperbolic case. In doing so, we follow Kuznetsov [36] and Kružkov [34].
Finally, we show that the limit u satisfies (ii) in (1.3). The arguments needed to prove (ii) are rather involved
and based on deriving a space estimate that is resemblant of the so-called weak BV estimates employed by
Champier et al. [7] and Eymard et al. [21] to prove convergence of finite volume methods on unstructured grids
for the hyperbolic equation, see also [1, 22] for the diffusion equation. Equipped with our weak BV estimate
and an appropriate time estimate, Kolmogorov’s compactness criterion implies strong L2loc convergence (along
a subsequence) of {A(u∆t )}∆t>0 to A(u) and ∇A(u) ∈ L2 .
Throughout this paper the coefficient k(x) is not allowed to be discontinuous. In the one-dimensional hyperbolic case (A0 ≡ 0) with k(x) depending discontinuously on x, the equation (1.1) is often written as the
following 2 × 2 system:
ut + f (k, u)x = 0,
kt = 0.
(1.4)
If ∂f /∂u changes sign, then this system is non-strictly hyperbolic. This complicates the analysis, and in order
to prove compactness of approximated solutions a singular transformation Ψ(k, u) has been used by several
authors [23, 29, 30, 45]. In these works convergence of the Glimm scheme and of front tracking was established
in the case where k may be discontinuous. Under weaker conditions on k, e.g., k 0 ∈ BV , and for f convex in
u, convergence of the one-dimensional Godunov method for (1.4) (not for (1.1)) was shown by Isaacson and
Temple in [27]. Recently, convergence of the one-dimensional Engquist-Osher method for (1.1) was shown by
Towers [46,47] in the case where k is piecewise continuous. In this case, the Kružkov entropy condition (1.3) no
longer applies, and in [30] a wave entropy condition analogous to the Oleinik entropy condition introduced in [40]
was used to obtain uniqueness. We intend to study the degenerate parabolic case (1.1) with a discontinuous
k(x) in future work.
The rest of this paper is organized as follows: in the next section we introduce (precisely) the notion of an
entropy solution, and state the theorem regarding uniqueness and the L1 contraction property of the solution
operator to (1.1). We then proceed to show convergence and convergence rates of difference schemes for the
hyperbolic equation. In the last section we show convergence of difference schemes for the degenerate parabolic
equation.
Throughout this paper we denote by C a generic constant, not depending on our discretization parameter
∆t. Note that the actual value of C may change from one line to the next during a calculation.
2. Preliminaries
In this section we first give a precise definition of an entropy solution, and then present some technical tools
that we shall use.
2.1. Definition of the entropy solution
Throughout this paper we let fi (k i , u) be smooth functions R×R → R, and set f (k, u) = f1 (k 1 , u), . . . , fd (k d , u) .
We assume that A : R → R is a function that satisfies
A ∈ Liploc (R) and A(·) is nondecreasing with A(0) = 0.
(2.1)
Concerning the flux function f : Rd × R → Rd , we assume that f ∈ C 3 Rd × R; Rd and that
fi , ∂u fi , ∂k fi , ∂uk fi , ∂ukk fi ∈ Lip(R × R; R),
for i = 1, . . . , d.
(2.2)
DIFFERENCE SCHEMES WITH ROUGH COEFFICIENTS
243
Furthermore, we assume that the relevant Lipschitz constants are bounded by
|∂u fi | ≤ Lu ,
|∂k fi | ≤ Lk ,
|∂uk fi | ≤ Luk ,
and so on,
for all i and for some constants Lu , Lk , Luk . Without explicitly mentioning this any more, we will always
assume in this paper that f (k, 0) = 0 for all k.
Regarding the coefficient k we assume that
(2.3)
k i ∈ C Rd ∩ BV Rd , ∂xj k i ∈ L∞ Rd ∩ BV Rd for all i and j.
Under the above assumptions we shall study difference approximations to (1.1). Following [28] (see also
Carrillo [5]), an entropy solution is defined as follows:
Definition 2.1. An entropy solution of (1.1) is a measurable function u = u(x, t) satisfying:
D.1 u ∈ L1 (ΠT ) ∩ L∞ (ΠT ) ∩ C(0, T ; L1 (Rd )).
D.2 For all c ∈ R, and all non-negative test functions ϕ in C0∞ (ΠT ) the following entropy inequality holds:
ZZ |u − c|ϕt + sign (u − c) f (k(x), u) − f (k(x), c) · ∇ϕ + |A(u) − A(c)|∆ϕ
ΠT
− sign (u − c) divf (k(x), c)ϕ dt dx ≥ 0.
D.3 A(u) ∈ L2 (0, T ; H 1 (Rd )).
D.4 Essentially as t → 0+,
Z
Rd
(2.4)
u(x, t) − u0 (x) dx → 0.
Remark 2.1. (i) Observe that when A0 ≡ 0, (2.4) reduces to the well known entropy inequality for scalar
conservation laws introduced by Kružkov [34] and Vol’pert [49].
(ii) Condition (D.4), i.e., that the initial datum u0 should be taken by continuity, motivates the requirement
of continuity with respect to t in condition (D.1).
The following theorem from [28] shows that the initial value problem (1.1) is well posed:
Theorem 2.1. Assume that (2.1, 2.2) and (2.3) hold. Let v, u ∈ L∞ (0, T ; BV (Rd )) be entropy solutions
of (1.1) with initial data v0 , u0 ∈ L1 (Rd ) ∩ L∞ (Rd ) ∩ BV (Rd ), respectively. Then for almost all t ∈ (0, T ),
kv(·, t) − u(·, t)kL1 (Rd ) ≤ kv0 − u0 kL1 (Rd ) .
(2.5)
In particular, there exists at most one entropy solution of the initial value problem (1.1).
Remark 2.2. At the expense of loosing (2.5), the assumption that v, u ∈ L∞ (0, T ; BV (Rd )) can be removed
and uniqueness still holds, see [28].
2.2. Some mathematical tools
In this section we present some mathematical tools that we shall
use in the analysis.
Let u : Rd ×(0, T ) → R be a function such that u(·, t) ∈ L1 Rd for all t ∈ (0, T ). By a modulus of continuity,
we mean a nondecreasing continuous function ν : R0+ → R0+ such that ν(0) = 0. We say that u has a spatial
modulus of continuity if
Z
sup
|u(x + ε, t) − u(x, t)| dx ≤ ν(y; u),
(2.6)
|ε|≤y
Rd
244
K.H. KARLSEN AND N.H. RISEBRO
(where ν may depend on t). We also say that u has a temporal modulus of continuity if there is a modulus of
continuity ω(·; u) such that for each τ ∈ (0, T ),
Z
sup
0≤ε≤τ
Rd
|u(x, t + ε) − u(x, t)| dx ≤ ω(τ ; u),
∀t ∈ (0, T − τ ).
(2.7)
Let θ(r) be a smooth non-negative function of a single variable r such that
Z
θ(r) = θ(−r),
θ(r) = 0,
for |r| ≥ 1, and
θ(r) dr = 1.
R
Let δε (x) = (1/εd)θ(|x|/ε), and, with a slight abuse of notation, δε (t) = (1/ε)θ(t/ε). Now define a test function
ϕ(x, y, t, s) by
ϕ(x, y, t, s) = δε (x − y)δε (t − s).
For a function u = u(x, t), set
ZZ λ(u, c) = −
|u − c| ϕt + sign (u − c) (f (k, u) − f (k, c))ϕx − sign (u − c) f (k, c)x ϕ dt dx
ΠT
Z
+
(2.8)
t=T
|u − c| ϕ t=0 dx.
R
For two functions u and v we define the functional Λε (u, v) as
ZZ
Λε (u, v) =
λ(u, v(y, s)) ds dy,
(2.9)
ΠT
where u = u(x, t) and v = v(y, s). In passing, we note that if A0 ≡ 0, and u is an entropy solution of (1.1), then
Λε (u, v) ≤ 0.
(2.10)
For two arbitrary functions u and v we have the following result:
Lemma 2.1 (Kuznetsov’s lemma). Assume that k = (k 1 , . . . , k d ) is in C(Rd ) and k 0 is in L∞ (Rd ) and that
both k and kxi have moduli of continuity for all i = 1, . . . , d. If u and v are in L1 (ΠT ) and have moduli of
continuity in space and time, then
ku(·, T ) − v(·, T )kL1 (Rd ) ≤ ku(·, 0) − v(·, 0)kL1 (Rd ) + Λε (u, v) + Λε (v, u)
1
[ν(u(·, 0); ε) + ν(v(·, 0); ε) + ν(u(·, T ); ε) + ν(v(·, T ); ε)]
2
1
+ [ω(u(·, T ); ε) + ω(v(·, T ); ε) + ω(u(·, 0); ε) + ω(v(·, 0); ε)]
2
+ kdivkkL∞ Lu T sup (ν(u(·, t); ε) + ω(u(·, t); ε))
+
(2.11)
0≤t≤T
+ CT kdivkkL∞ ν(k; ε) + max ν (kxi ; ε)) ,
i
where ω(·; ·) denotes a temporal modulus of continuity, and ν(·; ·) denotes a spatial modulus. The constant C
depends on fk and fkk .
DIFFERENCE SCHEMES WITH ROUGH COEFFICIENTS
245
Proof. We shall prove this lemma for d = 1, the general proof is completely analogous. Remember that
ϕx = −ϕy and ϕt = −ϕs . By adding Λε (u, v) and Λε (v, u) we find that
ZZ ZZ
Λε (u, v) + Λε (v, u) = −
sign (u − v) (f (k(x), u) − f (k(x), v)) ϕx − f (k(x), v)x ϕ
(2.12)
ΠT
ΠT
− (f (k(y), u) − f (k(y), v)) ϕx + f (k(y), u)y ϕ dt dx ds dy
ZZ Z
ϕ(x, y, T, s) |u(x, T ) − v(y, s)| dx ds dy
+
ΠT
R
ΠT
R
ZZ Z
ϕ(x, y, t, T ) |u(x, t) − v(y, T )| dy dt dx
+
ZZ Z
ϕ(x, y, 0, s) |u(x, 0) − v(y, s)| dx ds dy
−
ΠT
R
ZZ Z
−
ϕ(x, y, t, 0) |u(x, t) − v(y, 0)| dy dt dx.
ΠT
R
Using standard arguments, the four last terms will give the ku − vkL1 terms and the terms starting with 12 [. . .
in (2.11). Regarding the remaining term, we follow Kružkov [34]. Let
m(x, y, w) = fk (k(x), w)k 0 (x) − fk (k(y), w)k 0 (y).
Then we rewrite the square brackets in (2.12) as
(f (k(x), u) − f (k(x), v)) ϕx − f (k(x), v)x ϕ − (f (k(y), u) − f (k(y), v)) ϕx + f (k(y), u)y ϕ
= [(f (k(y), u) − f (k(x), u)) ϕy + f (k(y), u)y ϕ]
+ [(f (k(y), v) − f (k(x), v)) ϕx − f (k(x), v)x ϕ]
= fk (k(y), u)k 0 (y) [(y − x)ϕy + ϕ] + m(ξ1 , y, u)(x − y)ϕy
+ fk (k(y), v)k 0 (y) [(y − x)ϕx − ϕ]
+ m(y, x, v)ϕ + m(ξ2 , y, v)(y − x)ϕx
= −fk (k(y), u)k 0 (y) [(y − x)ϕ]x + fk (k(y), v)k 0 (y) [(y − x)ϕ]x
+ (m(ξ1 , y, u) + m(ξ2 , y, v)) ϕy + m(y, x, v)ϕ
= k 0 (y) (fk (k(y), v) − fk (k(y), u)) [(x − y)ϕ]x
+ (m(ξ1 , y, u) + m(ξ2 , y, v)) (x − y)ϕy + m(y, x, v)ϕ,
where |ξi − y| ≤ |x − y| for i = 1, 2. Let now
F (u, v) = sign (u − v) k 0 (y) (fk (k(y), v) − fk (k(y), u)) ,
and note that F is Lipschitz continuous, with Lipschitz constant given by kk 0 kL∞ Lu , in both arguments. Next,
we obviously have that
ZZ ZZ
F (u(y, s), v(y, s)) [(x − y)ϕ]x dt dx ds dy = 0.
ΠT
ΠT
246
K.H. KARLSEN AND N.H. RISEBRO
Thus, our troublesome term reads
ZZ ZZ
[F (u(x, t), v(y, s)) − F (u(y, s), v(y, s))] [(x − y)ϕ]x dt dx ds dy
ΠT
ΠT
ZZ ZZ +
ΠT
sign (u − v) (m(ξ1 , y, u) + m(ξ2 , y, v)) (x − y)ϕy + m(y, x, v)ϕ dt dx ds dy.
(2.13)
(2.14)
ΠT
Now
|m(x, y, u)| ≤ Lk |k 0 (x) − k 0 (y)| + kk 0 kL∞ Luk |k(x) − k(y)| .
Hence (2.14) is bounded by integrals of the form
ZZ ZZ
ΠT
|l(ξ) − l(y)| |(x − y)δε0 (x − y)| δε (t − s) dt dx ds dy
(2.15)
ΠT
and
ZZ ZZ
|l(ξ) − l(y)| δε (x − y)δε (t − s) dt dx ds dy,
ΠT
(2.16)
ΠT
where ξ = x, ξ1 or ξ2 , and l = k or k 0 . Since |x − y| ≤ ε we have that |ξ − y| ≤ ε in (2.15) and (2.16), so they
are both easily seen to be bounded by ν(l; ε). The rest of (2.13) is bounded as follows
ZZ ZZ F (u(x, t), v(y, s)) − F (u(y, s), v(y, s)) |((x − y)ϕ)x | dt dx ds dy
ΠT
ΠT
ZZ ZZ
≤ kk 0 kL∞ Lu
|u(x, t) − u(y, s)| |((x − y)ϕ)x | dt dx ds dy
kk 0 kL∞ Lu
≤
ε
ZZ Z
+
Z
T
ZZ
|u(x, t) − u(y, t)| dx dy dt
0
|x−y|≤ε
|u(y, t) − u(y, s)| dy dt ds
|t−s|≤ε
≤ kk 0 kL∞ Lu T sup (ν(u(·, t); ε) + ω(u(·, t); ε)) .
0≤t≤T
This concludes the proof of the lemma.
We need the following general L1 and L2 compactness criteria.
Lemma 2.2 (L1loc compactness lemma). Let {zh }h>0 be a sequence of functions defined on Rd × (0, T ) which
satisfies:
(1) There exists a constant C1 > 0 which is independent of h such that
kzh (·, t)kL1 (Rd ) and kzh (·, t)kL∞ (Rd ) ≤ C1 ,
∀t ∈ (0, T ).
DIFFERENCE SCHEMES WITH ROUGH COEFFICIENTS
247
(2) There exists a spatial modulus of continuity ν which is independent of h such that
kzh (· + y, t) − zh (·, t)kL1 (Rd ) ≤ ν(|y|; zh ) as y → 0,
∀t ∈ (0, T ).
(3) There exists a temporal modulus of continuity ω which is independent of h such that
kzh (·, t + τ ) − zh (·, t)kL1 (Rd ) ≤ ω(τ ; zh ),
∀t ∈ (0, T − τ ) whenever τ ∈ (0, T ).
Then {zh }h>0 is compact in the strong topology of L1loc (Rd × (0, T )). Moreover, any limit point of {zh }h>0
belongs to L1 (Rd × (0, T )) ∩ L∞ (Rd × (0, T )) ∩ C(0, T ; L1(Rd )).
Lemma 2.3 (L2loc compactness lemma). Let {zh }h>0 be a sequence of functions defined on Rd × (0, T ) which
satisfies:
(1) There exists a constant C1 > 0 which may depend on T , but not h, such that
kzh kL2 (Rd ×(0,T )) ≤ C1 .
(2) There exists a constant C2 > 0 which may depend on T but not h such that
kzh (· + y, ·) − zh (·, ·)kL2 (Rd ×(0,T )) ≤ C2 (|y| + h) for all y as h ↓ 0.
(3) There exists a constant C3 > 0 which may depend on T but not h such that
√
kzh (·, · + τ ) − zh (·, ·)kL2 (Rd ×(0,T −τ )) ≤ C3 τ + h for all τ > 0 as h ↓ 0.
Then {zh }h>0 is compact in the strong topology of L2loc (Rd × (0, T )). Moreover, any limit point of {zh }h>0
belongs to L2 (0, T ; H 1(Rd )).
To prove that the difference approximations possess some L1 time continuity, we shall use the following
lemma due to Kružkov [33].
Lemma 2.4 (Kružkov’s interpolation lemma [33]). Let z(x, t) be a bounded measurable function defined on
Rd × (0, T ). For t ∈ (0, T ) assume that z possesses a spatial modulus of continuity
Z z x + ε, t − z(x, t) dx ≤ ν(|ε| ; z),
(2.17)
Rd
where ν does not depend on t. Suppose that for any φ ∈ C0∞ (Rd ) and any t1 ,t2 ∈ (0, T ),
Z
X
α
cα kD φkL∞ (Rd ) · |t2 − t1 | ,
d (z (x, t2 ) − z (x, t1 )) φ(x) dx ≤ ConstT ·
R
(2.18)
|α|≤m
where α denotes a multi-index, and cα are constants not depending on φ or t. Then for any t1 , t2 ∈ (0, T ) and
all ε > 0


Z
X cα
|z(x, t2 ) − z(x, t1 )| dx ≤ C · |t2 − t1 |
+ ν(z; ε) .
(2.19)
ε|α|
Rd
|α|≤m
Proof. Let δε (x) denote the usual mollifying kernel of radius ε, and let d(x) = z(x, t2 ) − z(x, t1 ). For r > ε, set
(
sign (d(x)) for |x| ≤ r − ε,
β(x) =
0
otherwise,
248
K.H. KARLSEN AND N.H. RISEBRO
and set βε = β ∗ δε . Then βε is in C0∞ (Rn ), has support inside the ball Br , and we have the bound |Dα βε | ≤
Const/ε|α| . Also
Z
Z
Z
d(x) dx ≤
|d(x) − βε (x)d(x)| dx + βε (x)d(x) dx
Br
B
Br
Z Zr
X cα
≤
|d(x) − d(x − y)| δε (y) dx dy + C |t2 − t1 |
ε|α|
|α|≤m
Rd
X cα ≤ C ν(ε; z) + |t2 − t1 |
·
ε|α|
|α|≤m
Letting r ↑ ∞, we obtain (2.19).
We shall also need the technical result:
Lemma 2.5 (Crandall and Tartar [13]). Let (Ω, µ) be some measure space and let D be a subset of L1 (Ω).
Assume that if u and v are in D, then also u ∨ v is in D. Let T be a map D → D such that
Z
Z
T (u) dµ =
Ω
u dµ,
∀u ∈ D.
Ω
Then the following statements, valid for all u and v in D, are equivalent:
(i) IfR u ≤ v, then T (u) ≤ T (v).R
(ii) RΩ (T (u) − T (v)) ∨ 0 dµR≤ Ω (u − v) ∨ 0 dµ.
(iii) Ω |T (u) − T (v)| dµ ≤ Ω |u − v| dµ.
3. Difference approximations: The hyperbolic equation
In this section we analyze a difference approximation to the solution of the hyperbolic equation
ut + divf (k, u) = 0,
u(x, 0) = u0 (x),
(x, t) ∈ ΠT ,
(3.1)
where k and f satisfy (2.2) and (2.3) respectively. For simplicity, we shall assume that u0 has compact support,
which implies that all subsequent sums over I are finite. To obtain results in the general case, we can use the
stability result in Theorem 2.1 (these standard details will not be written out). We have chosen to analyze the
hyperbolic equation separately since the analysis parallels the general case but is simpler. In the next section,
where we consider the general case, we shall use several of the estimates obtained in this section. Furthermore,
we provide a convergence rate in the hyperbolic case.
As already mentioned in the introduction, we use the Engquist-Osher scheme to make the analysis more
concrete, but our methods can easily be adapted to general monotone schemes. For a scalar flux function
fi (k, u), the Engquist-Osher flux fiEO [14] can be written as
fiEO (k, u, v)
1
=
2
Z
fi (k, u) + fi (k, v) −
v
|∂u fi (k, s)| ds .
(3.2)
u
To define the scheme, let I be a multi-index I = (i1 , . . . , id ) and set ei to be a multi-index with zeros everywhere
except for a 1 at the ith place. Furthermore, we choose a time step ∆t such that N ∆t = T and a spatial
DIFFERENCE SCHEMES WITH ROUGH COEFFICIENTS
249
discretization parameter h > 0. Letting λ = ∆t/h, the Engquist-Osher scheme reads (d ≥ 2)
= unI − λ
un+1
I
d h
i
X
EO
¯EO
f¯I+e
−
f
I−ei /2
i /2
i=1
= unI − λ
d
X
i=1
h
i
i
n
n
EO
i
n
n
fiEO kJ+e
,
u
,
u
−
f
k
,
u
,
u
,
I
I+e
i
I−e
I
J−ei /2
i
i
i /2
X
1
2d−1
(3.3)
j6=i
J=I±ej /2
where
EO
f¯I+e
=
i /2
i
n
n
fiEO kJ+e
,
u
,
u
I
I+ei .
i /2
X
1
2d−1
(3.4)
j6=i
J=I±ej /2
For d = 1, the Engquist-Osher scheme reads
h
i
n
EO
n
n
EO
n
n
un+1
=
u
−
λ
f
k
,
u
,
u
−
f
k
,
u
,
u
.
i+1/2
i−1/2
i
i
i+1
i−1
i
i
The approximate solution u∆t is then defined as
u∆t (x, t) = unI ,
for (x, t) ∈ χI × [tn , tn+1 ),
(3.5)
where χI denotes the set
n
o
χI = x ∈ Rd xI−ei /2 i ≤ xi < xI+ei /2 i , i = 1, . . . , d
and xI = hI. We initialize the scheme by setting
u0I =
1
|χI |
Z
u0 (x) dx.
χI
Note that fiEO (k, u, v) is not continuously differentiable in the first variable but merely Lipschitz. Therefore we
introduce the following auxiliary numerical flux
fiEO,ε (k, u, v)
1
=
2
Z
fi (k, u) + fi (k, v) −
v
u
|∂u fi (k, s)|ε ds ,
where | · |ε is a smooth approximation to the absolute value function | · | such that
|a| − |a|ε ≤ ε and |a| = |a|ε
for |a| > ε.
250
K.H. KARLSEN AND N.H. RISEBRO
Note in particular that
EO
EO,ε
(k, u, v) ≤ |u − v| ε,
fi (k, u, v) − fi
∀k, u, v.
(3.6)
This scheme can be analyzed as follows. Set
X
d
i
=
X
1
2d−1
and ψi (k, u, v) =
j6=i
J=I±ej /2
∂fiEO,ε
(k, u, v).
∂k
Note that
|ψi (k, u, v)| ≤ Lk +
1
|u − v| Luk .
2
Then we define FI as the right hand side in (3.3), i.e.,
un+1
= FI (un ) .
I
(3.7)
Assuming the CFL condition
λ
d
X
i=1
it is easy to show that
U n = maxI |unI |, then
∂FI
∂un
J
max |∂u fi (k, u)| ≤ 1,
(3.8)
k,u
≥ 0 for all J. In other words, the Engquist-Osher scheme is monotone. Let
n+1 u
= |FI (un )| ≤ FI (U n )
I
= Un + λ
d X h i
X
d
i
n
i
n
fi kJ+e
,
U
−
f
k
,
U
i
/2
J−e
/2
i
i
i=1
i
d X d
1
X
X
i d i
i
n
kx ∞ d .
≤ U + Lk ∆t
kJ+ei /2 − kJ−ei /2 ≤ U + Lk ∆t
i L (R )
i
h
i=1
i=1
n
From this it follows that
ku∆t (·, T )kL∞ (Rd ) ≤ ku0 kL∞ (Rd ) + T Lk d max kxi i L∞ (Rd ) .
i
Next, by the Crandall and Tartar lemma (see Lem. 2.5) and the monotonicity of FI ,
X
I
|FI (un ) − FI (0)| ≤
X
I
|unI | .
(3.9)
251
DIFFERENCE SCHEMES WITH ROUGH COEFFICIENTS
Hence
hd
d X X
X
X
X
d i i
i
d
un+1 ≤ hd
|unI | + hd λ
,
0
≤
h
|unI | .
f kJ+ei /2 , 0 − f i kJ−e
/2
I
i
I
i=1
I
i
I
This means that
ku∆t (·, T )kL1 (Rd ) ≤ ku0 kL1 (Rd ) .
(3.10)
For any quantity XI defined on our grid let D` XI denote the upward difference
D` XI = XI+e` − XI .
(3.11)
To bound the total variation of u∆t we again use the Crandall-Tartar lemma, which in this case gives
X
FI+e
`
X n
n
u
un·+e` − FI+e` (un ) ≤
I+e` − uI .
I
I
Note that we have
n
un+1
I+e` = FI+e` u·+e`
and
Xh
i
X
X
D` un+1 ≤
|D` unI | + FI+e` (un ) − FI (un ) .
I
I
I
I
Before we start to estimate the difference on the right-hand side of (3.12), note that
1 ψi (k1 , u, v) − ψi (k2 , u, v) = (∂kk fi (η, u) + ∂kk fi (ν, v)) (k1 − k2 )
2
Z v
−
[signε (∂u fi (k1 , s)) − signε (∂u fi (k2 , s))] ∂uk fi (k1 , s)
u
+ signε (∂u fi (k2 , s)) ∂ukk fi (γ, s) ds
≤ max |∂kk fi (k, u)| |k1 − k2 | + max |∂uk fi (k, u)| |u − v|
k,u
k,u
1
+ max |∂ukk fi (k, u)| |u − v| |k1 − k2 | ,
2 k,u
where signε (·) denotes the derivative of | · |ε . Furthermore, we have
|ψi (k, u1 , v) − ψi (k, u2 , v)| ≤ max |∂ku fi (k, u)| |u1 − u2 | ,
k,u
|ψi (k, u, v1 ) − ψi (k, u, v2 )| ≤ max |∂ku fi (k, u)| |v1 − v2 | .
k,u
(3.12)
252
K.H. KARLSEN AND N.H. RISEBRO
i
i
Using the above estimates on ψi and (3.6), there exist numbers ξJ+e` /2±ei /2 between kJ±e
and kJ+e
i /2
` ±ei /2
such that
d X X
d
i
n
n
EO
i
n
n
fiEO kJ+e
,
u
,
u
−
f
k
,
u
,
u
FI+e` (u ) − FI (u ) = −λ
I
I+e
i
I
I+e
J+e
+e
/2
/2
i
i
i
i
`
n
n
−
i=1
fiEO
i
= O(ε) − λ
i
kJ+e
, unI−ei , unI
` −ei /2
−
fiEO
i
n
n
kJ−ei /2 , uI−ei , uI
d X X
d
i=1
i
i
ψi ξJ+e` /2+ei /2 , unI , unI+ei D` kJ+e
i /2
i
− ψi ξJ+e` /2−ei /2 , unI−ei , unI D` kJ−e
i /2
= O(ε) − λ
d X X
d
i=1
−
i
i
ψi ξJ+e` /2+ei /2 , unI , unI+ei Di D` kJ+e
i /2
Di ψi ξJ+e` /2−ei /2 , unI−ei , unI
= O(ε) − λ
i
D` kJ−e
i /2
d X X
d
i=1
i
i
ψi ξJ+e` /2+ei /2 , unI , unI+ei Di D` kJ+e
i /2
− ψi ξJ+e` /2+ei /2 , unI , unI+ei − ψi ξJ+e` /2−ei /2 , unI , unI+ei
+ ψi ξJ+e` /2−ei /2 , unI , unI+ei − ψi ξJ+e` /2−ei /2 , unI , unI
n
n
n
n
i
+ ψi ξJ+e` /2−ei /2 , uI , uI − ψi ξJ+e` /2−ei /2 , uI−ei , uI D` kJ−ei /2
d X X
d i
≤ O(ε) + λ
ψi ξJ+e` /2+ei /2 , unI , unI+ei Di D` kJ+e
/2
i
i=1
i
Lukk − Lkk Di ξJ+e` /2−ei /2 + Luk |Di unI | +
Di unI Di ξJ+e` /2−ei /2 2
i
+ Lku |Di unI | + Di unI−ei D` kJ−e
.
i /2
As the above inequality holds for any ε > 0, we can let ε ↓ 0 and obtain
|FI+e` (un ) − FI (un )| ≤ λ
d X X
d i
ψi ξJ+e` /2+ei /2 , unI , unI+ei Di D` kJ+e
i /2
i=1
i
Lukk − Lkk Di ξJ+e` /2−ei /2 + Luk |Di unI | +
Di unI Di ξJ+e` /2−ei /2 2
i
+ Lku |Di unI | + Di unI−ei D` kJ−e
.
i /2
253
DIFFERENCE SCHEMES WITH ROUGH COEFFICIENTS
Since we have that ψi is bounded, we find that
d X X
X
XX
d n
d−1
d−2
i
D` un+1 hd−1 ≤
D
|D
u
|
h
+
C∆t
h
D
k
` I
` i J+ei /2 I
I
I
I
+ hd−1 max kxi j i,j
L∞ (Rd )
i
i=1
d X XX
d i i
d−1
+
h
max
D` kJ−e
kxj i /2
I
i=1
i
i,j
L∞ (Rd )
d
XX
I
|Di unI | ,
i=1
for some constant C independent
of ∆t. The first and second sums inside {· · · } are bounded since we assume
that k i and kxi j are in BV Rd . By summing the above over ` = 1, . . . , d, we find that
|u∆t (·, tn+1 )|BV (Rd ) = hd−1
d X
X
D` un+1 ≤ (1 + C∆t) |u∆t (·, tn )|
I
`=1
I
+ C∆t |k|BV (Rd ) + max kxi j i,j
BV (Rd )
BV (Rd )
.
Consequently,
|u∆t (·, t)|BV (Rd ) ≤ C |u0 |BV (Rd ) + t ,
∀t ∈ (0, T ),
(3.13)
where C does not depend on ∆t.
Next, we shall use the scheme to show that u∆t ∈ C 0, T ; L1(Rd ) uniformly in ∆t. This is done as follows
d X
XX
d X EO i
n+1
n
d−1
i
n
n h
uI − uI ≤ ∆th
kJ+ei /2 , unI , unI+ei − fiEO kJ+e
,
u
,
u
fi
I−e
I
/2
i
i
d
I
i
I
i=1
EO ∂fi i
+ max Di kJ−ei /2 ∂k ≤ C∆t |u∆t |BV (Rd ) + |k|BV (Rd ) ,
from which we obtain
ku∆t (·, t + τ ) − u∆t (·, t)kL1 (Rd ) ≤ Cτ,
∀t ∈ [0, T − τ ].
(3.14)
By Lemma 2.2, we have that the sequence {u∆t }∆t>0 is compact in L1loc (ΠT ). Moreover, any limit point of this
sequence satisfies (D.1) and (D.4).
Next, we shall prove two cell entropy inequalities. The first one is based on Kružkov’s entropies and will be
used later to prove an error estimate. The second one is based on smooth (C 2 ) convex entropies and will be
used to show that the limit of any convergent subsequence of {u∆t }∆t>0 satisfies the entropy condition (D.2).
Let
wIn = unI ∨ c, vIn = unI ∧ c, wIn+1 = FI (wIn ) , vIn+1 = FI (vIn ) , cn+1
= FI (c),
I
where FI is defined in (3.7). Then we have that
≥ vIn+1 ,
wIn+1 ≥ un+1
I
This implies that
vIn+1 ≤ cn+1
≤ wIn+1 .
I
n+1
u
− cn+1 ≤ wn+1 − v n+1 .
I
I
I
I
254
K.H. KARLSEN AND N.H. RISEBRO
Now
wIn+1
−
vIn+1
=
|unI
d X X
d
i
n
n
− c| − λ
fiEO kJ+e
,
u
∨
c,
u
∨
c
I
I+e
/2
i
i
i
i=1
i
n
n
EO
i
n
n
,
u
∧
c,
u
∧
c
−
f
k
,
u
∨
c,
u
∨
c
−fiEO kJ+e
I
I+e
i
I−e
I
J−ei /2
i
i
i /2
i
n
n
−fiEO kJ−e
,
u
∧
c,
u
∧
c
.
I−ei
I
i /2
Denoting the numerical entropy flux by
qiEO (k, u, v) = fiEO (k, u ∨ c, v ∨ c) − fiEO (k, u ∧ c, v ∧ c)
and noting that
d X h i
X
n+1
n+1
d
n+1 i
i
u
fi kJ+ei /2 , c − fi kJ−ei /2 , c − cI
= uI − c + λ
I
i
i=1
!
d X h i
X
d
i
i
≥ sign un+1
− c un+1
−c+λ
fi kJ+e
, c − fi kJ−e
,c
I
I
i /2
i /2
i=1
i
d X h i
n+1
X
d
n+1
i
i
≥ uI − c + sign uI − c λ
fi kJ+e
, c − fi kJ−e
,c ,
i /2
i /2
i=1
i
we find that
d X X
n+1
d EO i
n
n
n
EO
i
n
n
u
− c ≤ |uI − c| − λ
qi
kJ+ei /2 , uI , uI+ei − qi
kJ−ei /2 , uI−ei , uI
I
i
i=1
− sign
un+1
I
d X
X
d
i
−c λ
Di fi kJ−e
,c .
i /2
i=1
(3.15)
i
This discrete entropy inequality will be used later to derive an error estimate.
Let U ∈ C 2 be a convex entropy and let q̃iEO (k, u, v) denote the associated numerical entropy flux defined by
Z
q̃iEO (k, u, v) =
u
Z
U 0 (s) ∂u fi (k, s) ∨ 0 ds +
0
v
U 0 (s) ∂u fi (k, s) ∧ 0 ds.
(3.16)
0
Let us write the scheme (3.3) as
un+1
= wIn − λ
I
d X h
i
X
d
i
n
n
EO
i
n
n
fiEO kJ+e
,
u
,
u
−
f
k
,
u
,
u
,
I
I+ei
i
I
I+ei
J−ei /2
i /2
i
i=1
(3.17)
where
wIn = unI − λ
d X h
i
X
d
i
i
fiEO kJ−e
, unI , unI+ei − fiEO kJ−e
, unI−ei , unI .
i /2
i /2
i=1
i
(3.18)
255
DIFFERENCE SCHEMES WITH ROUGH COEFFICIENTS
From (3.18), it follows in a standard way using monotonicity (i.e., the CFL condition) that
−
∆t
U (wIn )
U (unI )
+
n
n
EO
i
n
n
d X q̃ EO k i
−
q̃
k
,
u
,
u
,
u
,
u
X
i
I
I+ei
I−ei
I
J−ei /2
J−ei /2
d i
h
i
i=1
≤ 0.
(3.19)
Using (3.17) and convexity of U (·), it follows that
) + U 0 (un+1
)λ
U (wIn ) ≥ U (un+1
I
I
d X h
X
d
i
i=1
i
i
i
fiEO kJ+e
, unI , unI+ei − fiEO kJ−e
, unI , unI+ei .
i /2
i /2
Plugging this into (3.19), we get our desired cell entropy inequality
−
∆t
U (un+1
)
I
U (unI )
+
n
n
EO
i
n
n
d X q̃ EO k i
,
u
,
u
−
q̃
k
,
u
,
u
X
i
i
I
I+e
I−e
I
J+e
J−e
/2
/2
i
i
d
i
i
i=1
i
+ U 0 (un+1
)
I
n
n
EO
i
d X f EO k i
, unI , unI+ei
kJ−e
X
J+ei /2 , uI , uI+ei − fi
d i
i /2
i=1
−
d X
X
d
i=1
h
q̃iEO
h
i
i
n
n
EO
i
n
n
,
u
,
u
,
u
,
u
kJ+e
−
q̃
k
i
I
I+ei
I
I+ei
J−ei /2
i /2
h
i
(3.20)
≤ 0.
We now multiply (3.20) by a nonnegative test function, do summation by parts, and then subsequently send
∆t → 0. We conclude that a limit point u of {u∆t }∆t>0 satisfies the entropy condition
ZZ U (u)ϕt + q(k, u) · ∇ϕ − U 0 (u)fx (k, u) − qx (k, u) ϕ dt dx ≥ 0,
(3.21)
ΠT
where q = (q1 , . . . , qd ) satisfies
∂u q(k, u) = U 0 (u)∂u f (k, u).
Furthermore, fx , qx denote the functions obtained by taking the divergence of f (k(·), u), q(k(·), u), respectively.
Since (3.21) holds for any convex C 2 entropy U and corresponding entropy flux q, a standard approximation
argument applied to | · −c| will reveal that (2.4) (with A0 ≡ 0) holds for all c ∈ R. Hence, by the uniqueness of
the entropy solution, the whole sequence {u∆t }∆t>0 converges to the unique entropy solution.
Now we shall use the cell entropy inequality (3.15) and Kuznetsov’s lemma (Lem. 2.1) to show that u∆t
converges to the unique entropy solution at a rate of ∆t1/2 . For simplicity we shall restrict our proof to the case
of one space dimension, i.e., d = 1. However, with some effort, the calculations given below can be generalized
to the multi-dimensional case d > 1.
Let u(x, t) be the unique entropy solution to
ut + f (k, u)x = 0,
u(x, 0) = u0 (x)
(3.22)
where u0 and k 0 is of bounded variation, and let u∆t (x, t) and k∆t be as before. Now Λε (u, u∆t ) ≤ 0, and all
the continuity moduli in Kuznetzov’s lemma are linear in ε, therefore (2.11) reads
ku(·, T ) − u∆t (·, T )kL1 (R) ≤ ku0 − u∆t (·, 0)kL1 (R) + Λε (u∆t , u) + Cε,
(3.23)
256
K.H. KARLSEN AND N.H. RISEBRO
where the constant C does not depend on ∆t. We must estimate Λε (u∆t , u). Set
η(u) = |u − c| ,
q(k, u) = sign (u − c) (f (k, u) − f (k, c)).
Multiplying the cell entropy inequality (3.15) by positive numbers hϕni with ϕni = 0 for |i| large, and summing
over i and n = 0, . . . , N − 1 where N ∆t = T , we find that
Xh
l(u∆t , c) :=
ηin+1 − ηin ϕni h + q EO ki+1/2 , uni , uni+1 − q EO ki−1/2 , uni−1 , uni ϕni ∆t
i,n
n i
+ sign un+1
−
c
f
k
,
c
−
f
k
,
c
ϕi ∆t
i+1/2
i−1/2
i
(3.24)
=l1 + l2 + l3 ≤ 0.
We also find that
λε (u∆t , c) =
"
X
ηin+1
−
ηin
Z
xi+1/2
ϕ(x, tn ) dx
xi−1/2
i,n
+ q ki+1/2 , uni+1 − q ki−1/2 , uni
Z
(3.25)
tn+1
ϕ(xi , t) dt
tn
Z tn+1 ϕ xi+1/2 , t − ϕ (xi , t) dt
tn
#
Z
tn+1
n+1
+ sign ui − c f ki+1/2 , c − f ki−1/2 , c
ϕ (xi , t) dt
+ q ki+1/2 , uni+1 − q ki+1/2 , uni
(3.26)
(3.27)
(3.28)
tn
=: λ1 + λ2,1 + λ2,2 + λ3 .
Since l(u∆t , c) ≤ 0,
λε (u∆t , u) ≤ |λε (u∆t , u) − l (u∆t , u)| ≤ |λ1 − l1 | + |λ2,1 − l2 | + |λ3 − l3 | + |λ2,2 | .
First we note that
|λ2,2 | ≤ Lu
XZ
i,n
and
xi+1/2
Z
xi
tn+1
|δε0 (x − y)| δε (t − s) dt dx uni+1 − uni tn
ZZ
|λ2,2 | ds dy ≤
Lu T h
|u∆t |BV (R) .
ε
ΠT
We choose the numbers ϕni as
ϕni =
1
∆th
Z
xi+1/2
Z
tn+1
ϕ(x, t) dt dx.
xi−1/2
tn
Using this we find that
Z tn+1 Z xi+1/2
X
n+1
n 1
|λ1 − l1 | ≤
ηi
− ηi
|ϕ(x, t) − ϕ(x, tn )| dx dt
∆t tn
xi−1/2
i,n
Z
Z
X
tn+1 xi+1/2
η n+1 − ηin ≤
|ϕt | dx dt.
i
i,n
tn
xi−1/2
(3.29)
257
DIFFERENCE SCHEMES WITH ROUGH COEFFICIENTS
Therefore,
Z
X
n+1
n
|λ1 − l1 | ds dy ≤
ηi
− ηi
ZZ
Z
tn
i,n
ΠT
tn+1
xi+1/2
ZZ
|δε0 (t − s)| δε (x − y) ds dy dx dt
xi−1/2
X
η n+1 − ηin C ∆th
≤
i
ε
i,n
≤C
(3.30)
∆t
·
ε
We continue with
|λ2,1 − l2 | ≤
"
X Z
tn+1
ϕ(xi , t) dt × q ki+1/2 , uni+1 − q EO ki+1/2 , uni , uni+1
tn
i,n
− q ki−1/2 , uni − q EO ki−1/2 , uni−1 , uni #
Z tn+1
EO
n
n
n
EO
n
n +
ϕ(xi , t) dt − ϕi ∆t q
ki+1/2 , ui , ui+1 − q
ki−1/2 , ui−1 , ui
tn
"
Z
X n
EO
n
n
+
q ki+1/2 , ui+1 − q
ki+1/2 , ui , ui+1
tn
i,n
Z
tn+1
Z
xi+1/2
+
tn
" Z
X
≤
Lu
i,n
tn+1
tn+1
tn+1
|ϕx | dx dt
xi
#
EO
|ϕx | dx dt q
ki+1/2 , uni , uni+1 − q EO ki−1/2 , uni−1 , uni Z
xi+1
xi
Z
|ϕx | dx dt uni+1 − uni #
n
n
n
o
n n
|ϕx | dx dt ki+1/2 − ki−1/2 + ui − ui−1 + ui+1 − ui
·
xi+1/2
+C
tn
xi+1
xi−1/2
tn
Z
Z
xi−1/2
Consequently
ZZ
h
h
|λ2,1 − l2 | dy ds ≤ C |u∆t |BV (R) + |k|BV (R)
≤C ·
ε
ε
(3.31)
ΠT
Finally we estimate
"
Z tn+1 Z xi+1/2
X f ki+1/2 , c − f ki−1/2 , c 1
|λ3 − l3 | ≤
|ϕ(xi , t) − ϕ(x, t) + ϕ(x, t) − ϕ(x, t − ∆t)| dx dt
h tn
xi−1/2
i,n
Z ∆t
#
X
N −1
+
f ki+1/2 , c − f ki−1/2 , c
ϕ(xi , t) dt + ∆tϕi
i
"
Z
X ki+1/2 − ki−1/2 ≤ Lk
0
tn+1 Z
tn
i,n
h |ϕx | + ∆t |ϕt | dx dt
xi−1/2
Z
X
ki+1/2 − ki−1/2 + Lk
i
xi+1/2
0
∆t
ϕ(xi , t) dt +
∆tϕiN −1
#
·
258
K.H. KARLSEN AND N.H. RISEBRO
Therefore
ZZ
2
X
X
2
ki+1/2 − ki−1/2 h ∆t + h∆t +
ki+1/2 − ki−1/2 C∆t
ε
i,n
i
2
h + h∆t
≤ C |k|BV (R)
+ ∆t .
ε
|λ3 − l3 | ds dy ≤ Lk
ΠT
(3.32)
Collecting the bounds (3.29–3.31), and (3.32), we find that
h + ∆t + h2 + h∆t
,
Λε (u∆t , u) ≤ C ∆t +
ε
(3.33)
for some constant C not depending on ∆t or h. Since we assume that u0 has bounded variation,
ku(·, 0) − u∆t (·, 0)kL1 (R) ≤ Ch,
and using this in (3.23) as well as h = ∆t/λ we arrive at the inequality
∆t
ku(·, T ) − u∆t (·, T )kL1 (R) ≤ C ε +
,
ε
(3.34)
√
which is minimized by setting ε = ∆t.
The main result of this section is summed up in the following theorem, which is stated for multi-dimensional
equations:
Theorem 3.1. Assume that f and k satisfy (2.2) and (2.3), respectively. Moreover, assume that u0 is a
function in L1 (Rd )∩L∞ (Rd )∩BV (Rd ). Let u be the unique entropy solution of (3.1). If the CFL condition (4.16)
holds, then there exists a constant C, depending on k, kxi , u0 , f and T , but not on ∆t, such that
√
ku∆t (·, T ) − u(·, T )kL1 (Rd ) ≤ C ∆t,
(3.35)
where the Engquist-Osher approximate solution u∆t is build from (3.3) and (3.5).
Remark 3.1. The assumptions on k = (k 1 , . . . , k d ) in Theorem 3.1 are (slightly) less restrictive than those
used in [1, 6, 7, 21] for finite volume methods.
4. Difference approximations: The degenerate parabolic equation
In this section we analyse the Engquist-Osher scheme for the degenerate parabolic equation (1.1). Again we
shall assume that u0 has compact support so that all subsequent sums over I are finite. To obtain results for
the general case, we can use the stability result in Theorem 2.1.
Let λ = ∆t/h (as usual) and µ = ∆t/h2 , then this the scheme reads (d ≥ 2)
un+1
= unI − λ
I
d X
X
d
i=1
i
d
X
i
n
n
Di fiEO kJ−e
,
u
,
u
+
µ
Di2 A unI−ei
I−e
I
/2
i
i
i=1
n
=: GI (u ) ,
where Di denotes the usual upwind difference operator, see (3.11). For d = 1, the scheme reads
un+1
= uni − λDf EO ki−1/2 , uni−1 , uni + µD2 A(uni−1 ).
i
(4.1)
259
DIFFERENCE SCHEMES WITH ROUGH COEFFICIENTS
Let u∆t be the piecewise constant function defined by (4.1) and (3.5).
As a starting point we assume that the following CFL condition holds
CFL = λ
d
X
i=1
max |∂u fi (k, u)| + 2µd max A0 (u) ≤ 1.
(4.2)
u
k,u
Remark 4.1. The CFL condition (4.2) will be sufficient to establish the convergence of the sequence {u∆t}∆t>0
and moreover that a limit point u of this sequence satisfies (D.1), (D.2), and (D.4). However, to prove that u
obeys (D.3), we shall later need a slightly stronger CFL condition (see (4.16) below).
I
Now it is easy to see that the CFL condition (4.2) implies that ∂G
≥ 0 for all J and the scheme (4.1) is
∂un
J
monotone. In the same manner as the bounds (3.9, 3.10), and (3.13), we show that
ku∆t (·, t)kL∞ (Rd ) ≤ C,
ku∆t (·, t)kL1 (Rd ) ≤ C,
|u∆t (·, t)|BV (Rd ) ≤ C,
∀t ∈ (0, T ),
(4.3)
for some constant C not depending on ∆t. To show compactness of the scheme, we must also show that
u∆t ∈ C 0, T ; L1(Rd ) uniformly in ∆t. In order to do this, we use the Kružkov interpolation lemma (Lem. 2.4).
Let ϕ(x) be a test function and set ϕI = ϕ (xI ). Let Dt XIn = XIn+1 − XIn for any XIn . From (4.1) we find that
X
I
Dt unI ϕI hd =
d X
d
X X
X
d EO
λ
fi
kJ−ei /2 , unI−ei , unI Di ϕI−ei hd + µ
Di A (unI ) Di ϕI−ei hd
I
i
i=1
≤ C∆t k∇ϕkL∞ (Rd )
i=1
!
d
d
XX
XX
d X EO
n
n d
n
d−1
fi
kJ−ei /2 , uI−ei , uI h +
|Di A (uI )| h
.
I
i
i=1
I
i=1
In view of the uniform L1 and BV bounds in (4.3), an application of Lemma 2.4 gives
ku∆t (·, t1 ) − u∆t (·, t2 )kL1 (Rd ) ≤ C
p
|t1 − t2 |,
(4.4)
for some constant C not depending on ∆t. By Lemma 2.2, the sequence {u∆t }∆t>0 is compact in L1loc (ΠT ) and
any limit point will satisfy (D.1) and (D.4).
We next show that the entropy condition (D.2) holds. To this end, a cell entropy inequality for the
scheme (4.1) is established in the same way as in the hyperbolic case. A modification of the arguments leading
to equation (3.15) yields
X q̃i
) − U (unI ) X d
U (un+1
I
+
i
∆t
i=1
d
EO
i
i
kJ+e
, unI , unI+ei − q̃iEO kJ−e
, unI−ei , unI
i /2
i /2
h
n
n
EO
i
n
n
d
d X f EO k i
n
n
n
,
u
,
u
−
f
k
,
u
,
u
X
X
i
r(uI+ei ) − 2r(uI ) + r(uI−ei )
I
I+ei
I
I+ei
J+ei /2
J−ei /2
d i
−
+ U 0 (un+1
)
I
2
i
h
h
i=1
i=1
EO
i
n
n
EO
i
n
n
d
q̃
k
,
u
,
u
−
q̃
k
,
u
,
u
XX
i
i
I
I+e
I
I+e
J+e
J−e
/2
/2
i
i
d
i
i
−
≤ 0, (4.5)
i
h
i=1
260
K.H. KARLSEN AND N.H. RISEBRO
for any convex C 2 entropy U and corresponding numerical entropy fluxes q̃iEO defined in (3.16) and r defined
by r0 (u) = U 0 (u)A0 (u). Consequently, any limit point of {u∆t }∆t>0 satisfies
ZZ U (u)ϕt + q(k, u) · ∇ϕ + r(u)∆ϕ − U 0 (u)fx (k, u) − qx (k, u) ϕ dt dx ≥ 0,
(4.6)
ΠT
where q obeys ∂u q(k, u) = U 0 (u)∂u f (k, u) and fx , qx denote the divergence of f (k(·), u), q(k(·), u), respectively.
Since (4.6) holds for any convex C 2 entropy U and corresponding entropy fluxes q, r, the usual approximation
argument will show that (2.4) holds ∀c ∈ R.
It remains to show that a limit u of {u∆t }∆t>0 satisfies (D.3). This will be done by deriving a weak BV
estimate [1, 7, 21, 22]. Multiplying (4.1) by unI hd , and summing over I, we find that
d X
X
X
d n EO i
i
n
n
unI Dt unI hd + ∆t
kJ+ei /2 , unI , unI+ei − fiEO kJ+e
hd−1
u I fi
,
u
,
u
I−e
I
/2
i
i
i=1
I
+µ
i
d
X
d X
XX
d n EO i
Di A unI−ei Di unI−ei = −∆t
u I fi
kJ+ei /2 , unI−ei , unI
i=1
We can write
unI Dt unI =
Di A
i
n
n
d−1
,
u
,
u
h
.
− fiEO kJ−e
I−ei
I
i /2
1
2
2
Dt (unI ) − (Dt unI ) ,
2
and we also have that
unI−ei
i
i=1
I
Di unI−ei
2
Di A unI−ei
≥
maxu A0 (u)
since A0 (u) ≥ 0. Using these observations, we find that
XX
∆thd
0
maxu A (u)
i=1
d
I
!2
d X
XX
Di A unI−ei
d n EO i
+ ∆t
u I fi
kJ+ei /2 , unI , unI+ei
i
h
I i=1
i
−hd Xh
2
2
i
−fiEO kJ+e
, unI−ei , unI
hd−1 ≤
Dt (unI ) − (Dt unI )
i /2
2
I
− ∆t
d X
XX
d
I
i=1
i
i
i
unI fiEO kJ+e
, unI−ei , unI − fiEO kJ−e
, unI−ei , unI hd−1 . (4.7)
i /2
i /2
Before we proceed, we note that
fiEO (k, v, w) − fiEO (k, u, v) = fi− (k, w) − fi− (k, v) + fi+ (k, v) − fi+ (k, u) ,
where
fi− (k, u) =
Z u
0
∂u fi (k, s) ∧ 0 ds,
fi+ (k, u) =
Z u
∂u fi (k, s) ∨ 0 ds.
0
261
DIFFERENCE SCHEMES WITH ROUGH COEFFICIENTS
Thus we can write inequality (4.7) as
XX
∆thd
0
maxu A (u)
i=1
d
Di A unI−ei
h
I
!2
+ ∆th
d−1
+ unI
≤
i
i
unI fi− kJ+e
, unI+ei − fi− kJ+e
, unI
i /2
i /2
i
i=1
I
"
d X
XX
d
(4.8)
#
+
+
i
n
i
n
fi kJ+ei /2 , uI − fi kJ+ei /2 , uI−ei
(4.9)
d i
XX
−hd Xh
d−1
2
2
i
Dt (unI ) − (Dt unI ) + C∆t
Di kI−e
h , (4.10)
i /2
2
i=1
I
I
for some constant C. The sum in (4.9) can be analyzed further by introducing the functions
Fi± (k, u)
Z
u
=
s∂u f ± (k, s) ds.
0
Then an integration by parts reveals that
Fi± (k, b)
−
Fi± (k, a)
=b
fi± (k, b)
−
fi± (k, a)
Z b
−
fi± (k, s) − fi± (k, a) ds.
a
Therefore
+
+
+
i
n
i
n
i
n
i
n
unI fi+ kJ+e
,
u
,
u
,
u
,
u
−
f
k
=
F
k
−
F
k
I
I−ei
I
I−ei
J+ei /2
J+ei /2
J+ei /2
i
i
i
i /2
Z
un
I
+
un
I−e
(4.11)
+
i
i
n
fi+ kJ+e
,
s
−
f
k
,
u
ds,
I−ei
J+ei /2
i
i /2
i
−
i
n
i
n
i
i
unI fi− kJ+e
,
u
−
f
k
,
u
= Fi− kJ+e
, unI+ei − Fi− kJ+e
, unI
I+ei
I
J+ei /2
i
i /2
i /2
i /2
Z
−
un
I
un
I+e
(4.12)
−
i
i
n
fi− kJ+e
,
s
−
f
k
,
u
ds.
I+ei
J+ei /2
i
i /2
i
Consequently (4.9) can be written
d−1
∆th
d X
XX
d
I
i=1
"
Fi+
i
Z
+
i
n
i
n
kJ+ei /2 , uI − Fi kJ+ei /2 , uI−ei +
un
I
un
I−e
i
fi+ kJ+e
,
s
/2
i
i
−
fi+
i
i
i
kJ+e
, unI−ei ds + Fi− kJ+e
, unI+ei − Fi− kJ+e
, unI
i /2
i /2
i /2
Z
−
un
I
un
I+e
i
fi−
i
kJ+e
,s
i /2
−
fi−
#
i
n
kJ+ei /2 , uI+ei ds .
262
K.H. KARLSEN AND N.H. RISEBRO
We also have that Fi± is locally Lipschitz continuous in k as
Z
±
F (k1 , u) − F ± (k2 , u) = i
i
u
0
s∂u fi± (k1 , s) − fi± (k2 , s) ds ≤ max |∂uk fi | |u| |k1 − k2 | .
k,u
Hence we obtain
d X h
X X
i
d
±
±
i
n
i
n
d−1 Fi kJ+ei /2 , uI − Fi kJ−ei /2 , uI h i
I
i=1
≤ Luk max ku∆t (·, t)kL∞ (Rd )
t
d
XX
d−1
i
Di kJ−e
h ,
i
I
(4.13)
i=1
which is bounded uniformly in ∆t. To bound the terms involving integrals, we need the following technical
lemma (whose easy proof can be found in [22]):
Lemma 4.1. Let h : R → R be a monotone, Lipschitz continuous function, with a Lipschitz constant Lh . Then
we have
Z
b
1
2
(h(ξ) − h(a)) dξ ≥
(h(b) − h(a)) ,
a
2Lh
∀a, b ∈ R.
Applying this to fi± we find that
Z
un
I
un
I−e
+
i
i
n
fi+ kJ+e
,
s
−
f
k
,
u
ds
I−e
J+e
/2
/2
i
i
i
i
i
≥
2
1
i
i
fi+ kJ+e
, unI − fi+ kJ+e
, unI−ei
,
i /2
i /2
2 maxu |∂u f (k, u)|
Z unI −
i
i
n
−
fi− kJ+e
,
s
−
f
k
,
u
ds
I+e
J+e
/2
/2
i
i
i
i
un
I+e
i
≥
2
1
i
i
fi− kJ+e
, unI+ei − fi− kJ+e
, unI
.
i /2
i /2
2 maxu |∂u f (k, u)|
The above and (4.7) imply that
!2
d
d X XX
XX
Di A unI−ei
∆thd
∆thd−1
d
−
i
n
+
f
k
,
u
I+ei
J+ei /2
i
i
maxu A0 (u)
h
2 maxu,k |∂u f (k, u)|
I i=1
I i=1
2 2 +
+
i
n
i
n
i
n
−fi− kJ+e
,
u
+
f
k
,
u
−
f
k
,
u
I
I
I−ei
J+ei /2
J+ei /2
i
i
i /2
≤
i
−hd Xh
2
2
Dt (unI ) − (Dt unI ) + C∆t,
2
I
(4.14)
DIFFERENCE SCHEMES WITH ROUGH COEFFICIENTS
263
where the constant C does not depend on ∆t. Furthermore, using the definition of the scheme, (4.1), and the
inequality (a + b)2 ≤ 2(a2 + b2 ), we find that
d X 2
X
1
d
n 2
2
i
n
n
EO
i
n
n
(Dt uI ) ≤ 2λ d
fiEO kJ+e
,
u
,
u
−
f
k
,
u
,
u
I
I+e
i
I−e
I
J+e
/2
/2
i
i
i
i
i
2
i=1
2
i
i
+ fiEO kJ+e
, unI−ei , unI − fiEO kJ−e
, unI−ei , unI
i /2
i /2
+ 2µ2 d
d h
X
2 i
2
(Di A (unI )) + Di A unI−ei
i=1
≤ 4λ2 d
2
d − i
i
n
fi kJ+ei /2 , unI+ei − fi− kJ+e
,
u
I
i /2
d X
X
i=1
i
(4.15)
2 +
i
n
i
n
+ fi+ kJ+e
,
u
−
f
k
,
u
I
I−ei
J+ei /2
i
i /2
+ 2λ2 d
d X 2
X
d
i
n
n
EO
i
n
n
,
u
,
u
,
u
,
u
fiEO kJ+e
−
f
k
I−ei
I
i
I−ei
I
J−ei /2
i /2
i=1
+ 2µ2 d
i
d h
X
2 i
(Di A (unI ))2 + Di A unI−ei
.
i=1
In what follows, we assume the following strengthened CFL condition
CFLε := 8λd max |∂u fi (k, u)| + 4µd max A0 (u) ≤ 1 − ε,
u
i,k,u
(4.16)
where ε ∈ (0, 1) is given a real number. Note that if A0 ≡ 0, i.e., in the hyperbolic case, (4.16) implies
CFLε ∈ (0, 18 ), which should be compared with the usual CFL ∈ (0, 1), see (4.2). The new CFL condition
implies in particular that
∆t
∆t
∆t(1 − ε)
= 8λd max |∂u fi (k, u)|
≤
,
k,u,i
h
2h maxk,u,i |∂u fi (k, u)|
2h maxk,u,i |∂u fi (k, u)|
∆t
∆t
∆t(1 − ε)
2µ2 d = 2µd 2 = 4d max A0 (u) 2
≤ 2
,
u
h
2h maxu A0 (u)
2h maxu A0 (u)
4λ2 d = 4λd
and therefore
d X 2
X
1
∆t(1 − ε)
d
+
n 2
i
n
i
n
(Dt uI ) ≤
fi+ kJ+e
,
u
−
f
k
,
u
I
I−e
J+e
/2
/2
i
i
i
i
i
2
2h maxk,u,i |∂u fi (k, u)| i=1
2
−
i
n
i
n
+ fi− kJ+e
,
u
−
f
k
,
u
I+e
I
J+e
i
/2
/2
i
i
i
2
XX
∆t(1 − ε)
d i
+
const. Di kJ−e
/2
i
i
2h maxk,u,i |∂u fi (k, u)| i=1
d
Xh
2 i
∆t(1 − ε)
(Di A (unI ))2 + Di A unI−ei
.
2
0
2h maxu A (u) i=1
d
+
(4.17)
264
K.H. KARLSEN AND N.H. RISEBRO
Now we multiply the above inequality (4.17) by hd and sum over I and n = 0, . . . , N , and sum (4.14) over n,
and add the results to find that
∆tεhd X X
maxu A0 (u)
d
n,I i=1
!2
d X XX
Di A unI−ei
∆thd−1 ε
d
i
fi+ kJ+e
+
, unI
/2
i
i
h
2 maxi,k,u |∂u fi (k, u)|
n,I i=1
2 2 +
−
−
i
n
i
n
i
n
−fi kJ+ei /2 , uI−ei
+ fi kJ+ei /2 , uI+ei − fi kJ+ei /2 , uI
≤
hd X 0 2
1
uI + CT ≤ ku0 kL∞ (Rd ) ku0 kL1 (Rd ) + CT ≤ C,
2
2
(4.18)
I
for some constant C not depending on ∆t. We thus have the following bound
∆thd
2
d XX
Di A (un )
I
n,I i=1
h
≤ C,
(4.19)
for some constant C not depending on ∆t.
Next let I(x) be the multi-index such that x ∈ χI . Then we have that I(x + y) − I(x) =: J = (J1 , . . . , Jd ),
and |Jh| ≤ |y| + h. Set Ki = I + (J1 , . . . , Ji−1 , 0, . . . , 0). Using this notation we can write
Ji
d X
X
A (u∆t (x + y, t)) − A (u∆t (x, t)) = A unI(x+y) − A unI(x) =
Di A unKi −(j−1)ei .
i=1 j=1
By the Cauchy-Schwartz inequality
Ji d X
2
2
X
A (u∆t (x + y, t)) − A (u∆t (x, t)) ≤ |J|
Dj A unKi −(j−1)ej
,
i=1 j=1
where |J|
d
X
= |J i | .
i=1
Hence using (4.19),
ZZ 2
2
X A (u∆t (x + y, t)) − A (u∆t (x, t)) dt dx ≤ ∆thd
A unI(xI +y) − A unI(xI )
n,I
ΠT
≤ ∆th |J|
d
Ji d X
XX
2
Di A unKi −(j−1)ei
n,I i=1 j=1
≤ |J| ∆thd
Ji X
d d X
2
XX
D` A unKi −(j−1)ei
n,I i=1 j=1 `=1
≤ (|J|)2 d∆thd
d
XX
(Di A (unI ))2 ≤ C (|J| h)2 ,
n,I i=1
where the constant C does not depend on y or h. Noting that |J| h ≤ C (|y| + h), we find that
kA (u∆t (· + y, ·)) − A (u∆t (·, ·))kL2 (ΠT ) ≤ C(|y| + h).
(4.20)
DIFFERENCE SCHEMES WITH ROUGH COEFFICIENTS
265
Next, we will use the weak space estimate (4.19) and the difference scheme itself to show that A (u∆t ) is also
L2 continuous in time. Let n(t) denote the integer such that t ∈ [tn(t) , tn(t)+1 ). Then we have
ZZ 2
A (u∆t (x, t + τ )) − A (u∆t (x, t)) dt dx
ΠT −τ
≤ max A0 (u)
u
ZZ A (u∆t (x, t + τ )) − A (u∆t (x, t)) (u∆t (x, t + τ ) − u∆t (x, t)) dt dx
ΠT −τ
= max A0 (u)
u
Z
T −τ
hd
0
X n(t+τ ) n(t)
n(t+τ )
n(t)
− uI
A uI
− A uI
uI
dt.
I
We denote the above integrand by B(t) and write
n(t)+n(τ
X n(t+τ ) X )−1
n(t)
Dt unI .
B(t) = h
A uI
− A uI
d
I
(4.21)
n=n(t)
Then each term in the sum over n above equals
hd
X
n(t+τ )
n(t)
( A uI
− A uI
Dt unI
I
= −∆th
−
fi−
d−1
d X
XX
d
I
i=1
i
kJ+e
, unI
i /2
i
"
n(t+τ )
n(t)
i
n
A uI
− A uI
fi− kJ+e
,
u
I+e
/2
i
i
fi+
+
i
n
i
n
kJ+ei /2 , uI − fi kJ+ei /2 , uI−ei
+
n(t+τ )
n(t)
+ A uI
− A uI
#
i
i
× fiEO kJ+e
, unI−ei , unI − fiEO kJ−e
, unI−ei , unI
i /2
i /2
+ ∆thd−2
d
X
X n(t+τ ) n(t)
Di2 A unI−ei .
A uI
− A uI
I
(4.22)
(4.23)
(4.24)
i=1
We can do a partial summation in (4.22) to find that
d X
X
d
n(t+τ )
n(t)
EO
i
n
n |(4.22)| = ∆th Di A uI−ei
− A uI−ei fi
kJ+ei /2 , uI−ei , uI i
i=1

2 
2 
n(t+τ )
n(t)
d X
d X
A
u
A
u
D
D
i
i
I−e
I−e
∆th
i
i
d EO i

 +
 
≤
fi
kJ+ei /2 , unI−ei , unI 
.
i
2 i=1
h
h
d−1 (4.25)
266
K.H. KARLSEN AND N.H. RISEBRO
In (4.25), one should recall that we have uniform control over the L2 (ΠT ) norm of the discrete gradient of
A(u∆t ), see (4.19). Since k is of bounded variation,
|(4.23)| ≤ ∆tC |k|BV (Rd ) ,
(4.26)
where C does not depend on ∆t. Regarding (4.24) we have that
d
X X
n(t+τ
)
n(t)
|(4.24)| = ∆thd−2 −Di A uI−ei
Di A unI−ei + Di A uI−ei A unI−ei I i=1
"
#
d
2 1 2
XX
2
1
n(t+τ )
n(t)
d−2
n
.
≤ ∆th
Di A uI−ei
+
Di A uI−ei
+ Di A uI−ei
2
2
i=1
(4.27)
I
We can write
Z
Z
1
1
B1 (t) + B2 (t) + B3 (t) dt,
2
2
0
0
where C does not depend on ∆t and B1 , B2 , B3 are defined via (4.21) and (4.27). Now
Z
T −τ
B(t) dt ≤ (T − τ )C n(τ )∆t +
Z
T −τ
T −τ
B1 (t) dt =
0
∆thd−2
d n(t)+n(τ
XX
X )−1 0
i=1
I
N −n(τ )
=
X
T −τ ∆thd−2
m=0
2
n(t+τ )
Di A uI
dt
n=n(t)
d n(tm )+n(τ
XX
X )−1 2
n(t +τ )
Di A uI m
∆t
I
i=1
≤ (n(τ ) + 1)∆t∆thd−2
n=n(tm )
d
XXX
n
I
(Di A (unI ))2
i=1
≤ C(τ + ∆t).
(4.28)
Similarly
Z
T −τ
B2 (t) dt ≤ C (τ + ∆t) .
(4.29)
0
Finally
Z
T −τ
N −n(τ )
B3 (t) dt ≤
0
X
∆th
n=m
m=0
X
d
XX
I
(Di A (unI ))2 ∆t
i=1
N −n(τ )
n(τ )−1
= ∆t
X
m+n(τ )−1
d−2
∆thd−2
k=0
d
X XX
n=0
I
Di A un+k
I
2
i=1
≤ C(τ + ∆t),
where C does not depend on ∆t. Using the bounds (4.28–4.30), we find that
√
kA (u∆t (·, · + τ )) − A (u∆t (·, ·))kL2 (Rd ×(0,T −τ )) ≤ C ∆t + τ .
In view Lemma 2.3, we conclude that
A (u∆t ) → A strongly in L2loc (R × (0, T )) as ∆t ↓ 0 and A ∈ L2 (0, T ; H 1(R)).
(4.30)
(4.31)
DIFFERENCE SCHEMES WITH ROUGH COEFFICIENTS
267
Equipped with the strong convergence u∆t → u a.e., we conclude immediately that A = A(u) and thus (D.3)
holds.
We sum up our results in the following theorem:
Theorem 4.1. Assume that A, f , and k satisfy (2.1, 2.2), and (2.3), respectively. Furthermore, assume
that u0 is a function in L1 (Rd ) ∩ L∞ (Rd ) ∩ BV (Rd ). If the CFL condition (4.16) holds, then the piecewise
constant approximate solutions (3.5) generated by the Engquist-Osher scheme (4.1) converge to the unique
entropy solution of (1.1).
In the special case without coefficients, i.e., k = 1, we can use our techniques to prove existence of an
entropy solution without necessarily having initial data in BV (Rd ). This can be done as follows. Assuming
that u0 ∈ L∞ (Rd ) ∩ L1 (Rd ), we study the problem
ut + divf (u) = ∆A(u),
u(x, 0) = u0 (x),
(4.32)
where f and A are as before. We obtain the two first bounds in (4.3) as before. Fixing ε ∈ Rd , we have
ku∆t (· + ε, t) − u∆t (·, t)kL1 (Rd ) ≤ ku0 (· + ε) − u0 (·)kL1 (Rd ) ≤ ν(|ε| ; u0 ),
since any function in L∞ (Rd )∩L1 (Rd ) has some modulus of continuity and the scheme is now translation invariant. Then we use Kružkov’s interpolation lemma (Lem. 2.4) to show that u∆t also has a modulus of continuity
in time. Next, we use the L1 compactness lemma (Lem. 2.2) to show that {u∆t }∆t>0 has a subsequence that
converges strongly in L1 to a function that satisfies (D.1), (D.4), and the entropy condition (D.3). To finally
show that the limit satisfies (D.3), note that to obtain the crucial estimates (4.20) and (4.31) we did not use a
BV bound on u∆t . Thus we have shown the following theorem:
Theorem 4.2. Assume that the function u0 is in L∞ (Rd ) ∩ L1 (Rd ). If the CFL condition (4.16) holds, then
then the piecewise constant approximate solutions (3.5) generated by the Engquist-Osher scheme (4.1) converge
to the unique entropy solution of (4.32).
References
[1] M. Afif and B. Amaziane, Convergence of finite volume schemes for a degenerate convection-diffusion equation arising in
two-phase flow in porous media. Preprint (1999).
[2] F. Bouchut, F.R. Guarguaglini and R. Natalini, Diffusive BGK approximations for nonlinear multidimensional parabolic
equations. Indiana Univ. Math. J. 49 (2000) 723–749.
[3] R. Bürger, S. Evje and K.H. Karlsen, On strongly degenerate convection-diffusion problems modeling sedimentationconsolidation processes. J. Math. Anal. Appl. 247 (2000) 517–556.
[4] M.C. Bustos, F. Concha, R. Bürger and E.M. Tory, Sedimentation and thickening: Phenomenological foundation and mathematical theory. Kluwer Academic Publishers, Dordrecht (1999).
[5] J. Carrillo, Entropy solutions for nonlinear degenerate problems. Arch. Rational Mech. Anal. 147 (1999) 269–361.
[6] C. Chainais-Hillairet, Finite volume schemes for a nonlinear hyperbolic equation. Convergence towards the entropy solution
and error estimate. RAIRO-Modél. Math. Anal. Numér. 33 (1999) 129–156.
[7] S. Champier, T. Gallouët and R. Herbin, Convergence of an upstream finite volume scheme for a nonlinear hyperbolic equation
on a triangular mesh. Numer. Math. 66 (1993) 139–157.
[8] B. Cockburn, F. Coquel and P. Le Floch, An error estimate for finite volume methods for multidimensional conservation laws.
Math. Comp. 63 (1994) 77–103.
[9] B. Cockburn, F. Coquel and P.G. LeFloch, Convergence of the finite volume method for multidimensional conservation laws.
SIAM J. Numer. Anal. 32 (1995) 687–705.
[10] B. Cockburn and P.-A. Gremaud, A priori error estimates for numerical methods for scalar conservation laws. I. The general
approach. Math. Comp. 65 (1996) 533–573.
268
K.H. KARLSEN AND N.H. RISEBRO
[11] B. Cockburn and C.-W. Shu, The local discontinuous Galerkin method for time-dependent convection-diffusion systems. SIAM
J. Numer. Anal. 35 (1998) 2440–2463 (electronic).
[12] M.G. Crandall and A. Majda, Monotone difference approximations for scalar conservation laws. Math. Comp. 34 (1980) 1–21.
[13] M.G. Crandall and L. Tartar, Some relations between nonexpansive and order preserving mappings. Proc. Amer. Math. Soc.
78 (1980) 385–390.
[14] B. Engquist and S. Osher, One-sided difference approximations for nonlinear conservation laws. Math. Comp. 36 (1981)
321–351.
[15] M.S. Espedal and K.H. Karlsen, Numerical solution of reservoir flow models based on large time step operator splitting
algorithms, in Filtration in Porous media and industrial applications. Lect. Notes Math. 1734, Springer, Berlin (2000) 9–77.
[16] S. Evje and K.H. Karlsen, Discrete approximations of BV solutions to doubly nonlinear degenerate parabolic equations.
Numer. Math. 86 (2000) 377–417.
[17] S. Evje and K.H. Karlsen, Degenerate convection-diffusion equations and implicit monotone difference schemes, in Hyperbolic
problems: Theory, numerics, applications, Vol. I (Zürich, 1998). Birkhäuser, Basel (1999) 285–294.
[18] S. Evje and K.H. Karlsen, Viscous splitting approximation of mixed hyperbolic-parabolic convection-diffusion equations.
Numer. Math. 83 (1999) 107–137.
[19] S. Evje and K.H. Karlsen, Monotone difference approximations of BV solutions to degenerate convection-diffusion equations.
SIAM J. Numer. Anal. 37 (2000) 1838–1860 (electronic).
[20] S. Evje and K.H. Karlsen, Second order difference schemes for degenerate convection-diffusion equations. Preprint (in preparation).
[21] R. Eymard, T. Gallouët, M. Ghilani and R. Herbin, Error estimates for the approximate solutions of a nonlinear hyperbolic
equation given by finite volume schemes. IMA J. Numer. Anal. 18 (1998) 563–594.
[22] R. Eymard, T. Gallouët, D. Hilhorst and Y. Naı̈t Slimane, Finite volumes and nonlinear diffusion equations. RAIRO-Modél.
Math. Anal. Numér. 32 (1998) 747–761.
[23] T. Gimse and N.H. Risebro, Solution of the Cauchy problem for a conservation law with a discontinuous flux function. SIAM
J. Math. Anal. 23 (1992) 635–648.
[24] A. Harten, J.M. Hyman and P.D. Lax, On finite-difference approximations and entropy conditions for shocks. Comm. Pure
Appl. Math. XXIX (1976) 297–322.
[25] H. Holden, K.H. Karlsen and K.-A. Lie, Operator splitting methods for degenerate convection-diffusion equations I: Convergence and entropy estimates, in Stochastic processes, physics and geometry: New interplays. A volume in honor of Sergio
Albeverio. Amer. Math. Soc. (to appear).
[26] H. Holden, K.H. Karlsen, K.-A. Lie and N.H. Risebro, Operator splitting for nonlinear partial differential equations: An L1
convergence theory. Preprint (in preparation).
[27] E. Isaacson and B. Temple, Convergence of the 2 × 2 Godunov method for a general resonant nonlinear balance law. SIAM J.
Appl. Math. 55 (1995) 625–640.
[28] K.H. Karlsen and N.H. Risebro, On the uniqueness and stability of entropy solutions of nonlinear degenerate parabolic equations with rough coefficients. Preprint, Department of Mathematics, University of Bergen (2000).
[29] C. Klingenberg and N.H. Risebro, Stability of a resonant system of conservation laws modeling polymer flow with gravitation.
J. Differential Equations March (2000).
[30] C. Klingenberg and N.H. Risebro, Convex conservation laws with discontinuous coefficients. Existence, uniqueness and asymptotic behavior. Comm. Partial Differential Equations 20 (1995) 1959–1990.
[31] D. Kröner, S. Noelle and M. Rokyta, Convergence of higher order upwind finite volume schemes on unstructured grids for
scalar conservation laws in several space dimensions. Numer. Math. 71 (1995) 527–560.
[32] D. Kröner and M. Rokyta, Convergence of upwind finite volume schemes for scalar conservation laws in two dimensions. SIAM
J. Numer. Anal. 31 (1994) 324–343.
[33] S.N. Kružkov, Results on the nature of the continuity of solutions of parabolic equations, and certain applications thereof.
Mat. Zametki 6 (1969) 97–108.
[34] S.N. Kružkov, First order quasi-linear equations in several independent variables. Math. USSR Sbornik 10 (1970) 217–243.
[35] A. Kurganov and E. Tadmor, New high-resolution central schemes for nonlinear conservation laws and convection-diffusion
equations. J. Comput. Phys. 160 (2000) 241–282.
[36] N.N. Kuznetsov, Accuracy of some approximative methods for computing the weak solutions of a first-order quasi-linear
equation. USSR Comput. Math. Math. Phys. Dokl. 16 (1976) 105–119.
[37] B.J. Lucier, Error bounds for the methods of Glimm, Godunov and LeVeque. SIAM J. Numer. Anal. 22 (1985) 1074–1081.
[38] S. Noelle, Convergence of higher order finite volume schemes on irregular grids. Adv. Comput. Math. 3 (1995) 197–218.
[39] M. Ohlberger, A posteriori error estimates for vertex centered finite volume approximations of convection-diffusion-reaction
equations. Preprint, Mathematische Fakultät, Albert-Ludwigs-Universität Freiburg (2000).
[40] O.A. Oleı̆nik, Discontinuous solutions of non-linear differential equations. Amer. Math. Soc Transl. Ser. 2 26 (1963) 95–172.
[41] S. Osher and E. Tadmor, On the convergence of difference approximations to scalar conservation laws. Math. Comp. 50 (1988)
19–51.
DIFFERENCE SCHEMES WITH ROUGH COEFFICIENTS
269
[42] É. Rouvre and G. Gagneux, Solution forte entropique de lois scalaires hyperboliques-paraboliques dégénérées. C. R. Acad.
Sci. Paris Sér. I Math. 329 (1999) 599–602.
[43] A.A. Samarskii, V.A. Galaktionov, S.P. Kurdyumov and A.P. Mikhailov, Blow-up in quasilinear parabolic equations. Walter
de Gruyter & Co., Berlin (1995). Translated from the 1987 Russian original by Michael Grinfeld and revised by the authors.
[44] R. Sanders, On convergence of monotone finite difference schemes with variable spatial differencing. Math. Comp. 40 (1983)
91–106.
[45] B. Temple, Global solution of the Cauchy problem for a class of 2 × 2 nonstrictly hyperbolic conservation laws. Adv. in Appl.
Math. 3 (1982) 335–375.
[46] J. Towers, Convergence of a difference scheme for conservation laws with a discontinuous flux. Preprint, Available at the
URL http://www.math.ntnu.no/conservation/
[47] J. Towers, A difference scheme for conservation laws with a discontinuous flux - the nonconvex case. Preprint, Available at
the URL http://www.math.ntnu.no/conservation/
[48] J.-P. Vila, Convergence and error estimates in finite volume schemes for general multidimensional scalar conservation laws. I.
Explicit monotone schemes. RAIRO-Modél. Math. Anal. Numér. 28 (1994) 267–295.
[49] A.I. Vol’pert, The spaces BV and quasi-linear equations. Math. USSR Sbornik 2 (1967) 225–267.
[50] A.I. Vol’pert and S.I. Hudjaev, Cauchy’s problem for degenerate second order quasilinear parabolic equations. Math. USSR
Sbornik 7 (1969) 365–387.
To access this journal online:
www.edpsciences.org

Similar documents