c 1999 Society for Industrial and Applied Mathematics

SIAM J. NUMER. ANAL.

Vol. 37, No. 1, pp. 286318

FINITE ELEMENT APPROXIMATION OF THE CAHNHILLIARD

EQUATION WITH DEGENERATE MOBILITY
JOHN W. BARRETT , JAMES F. BLOWEY , AND HARALD GARCKE
Abstract. We consider a fully practical finite element approximation of the CahnHilliard
equation with degenerate mobility
u
t

= .( b(u) (u + 0 (u))),

where b() 0 is a diffusional mobility and () is a homogeneous free energy. In addition to

showing well posedness and stability bounds for our approximation, we prove convergence in one
space dimension. Furthermore, an iterative scheme for solving the resulting nonlinear discrete system
is analyzed. We also discuss how our approximation has to be modified in order to be applicable to
a logarithmic homogeneous free energy. Finally, some numerical experiments are presented.
Key words. fourth order degenerate parabolic equation, CahnHilliard, phase separation, finite
elements, convergence analysis
AMS subject classifications. 65M60, 65M12, 35K55, 35K65, 35K35, 82C26
PII. S0036142997331669

1. Introduction. The CahnHilliard equation

u
t

= .( b(u) (u + 0 (u))),

x , t > 0,

was introduced to model spinodal decomposition and coarsening phenomena (Ostwald

ripening) in binary alloys (cf. [10] and [12]). The quantity u is defined to be the
difference of the local concentrations cA , cB [0, 1] of the two components A and B
of the alloy and hence u is restricted to lie in the interval [1, 1]. The theory of Cahn
and Hilliard is based on a GinzburgLandau free energy of the form
Z

2
> 0.
E(u) :=
2 |u| + (u) dx,

The first term in the free energy penalizes large gradients and was introduced in
the theory of phase transitions to model capillary effects. The second term is the
homogeneous free energy, which contains a term describing the entropy of mixing and
a term taking into account the interaction between the two components. A mean field
model leads to the potential





1+u
1u

(1 + u) ln
+ (1 u) ln
+ F 0 (u),
(1.1)
(u) :=
2
2
2
where is the absolute temperature and F 0 is a smooth function on the interval
[1, 1]. A typical example is F 0 (u) := 2c (1 u2 ), giving rise to a double well form of
Received by the editors December 18, 1997; accepted for publication (in revised form) February
2, 1999; published electronically December 3, 1999. This work was partially supported by the DFG
through SFB256 Nichtlineare Partielle Differentialgleichungen.
http://www.siam.org/journals/sinum/37-1/33166.html
Department of Mathematics, Imperial College, London, SW7 2BZ, UK ([email protected]).
Department of Mathematical Sciences, University of Durham, DH1 3LE, UK (j.f.blowey@
durham.ac.uk).
Institut f
ur Angewandte Mathematik, Wegelerstrasse 6, 53115 Bonn, Germany ([email protected]).

286

THE CAHNHILLIARD EQUATION WITH DEGENERATE MOBILITY

287

if < c . But there are other reasonable choices of . If the temperature is below
the critical temperature c and the quench is shallow, i.e., 0  < c , one could take,
e.g.,
(u) := (u2 a2 )2 ,

a R.

This has the advantage of being smooth, but the disadvantage that physically nonadmissible values with |u| > 1 can be attained during the evolution. For low temperatures, an obstacle potential of the form
1


1 u2
if |u| 1,
(1.2)
(u) := 2

if |u| > 1
was suggested in [9]. This is formally the limit of the logarithmic potential, (1.1),
with F 0 (u) := 12 (1 u2 ) in the deep quench limit 0. The general feature is that
below a certain critical temperature precisely two global minima of exist. As these
minima are interpreted as phases, the potential Ris said to support two phases. If one
R, where u
lies between
minimizes E() subject to the integral constraint u = u
the two minima of , then the minimizer umin , roughly speaking, will give rise to
the following structure. The function umin divides the domain into three sets. On
two of these sets umin will be close to the minima of , whereas the third will be

an interfacial regime of thickness approximately proportional to dividing the two

phases. Generically the minima are realized as large time limits of the CahnHilliard
evolution with constant mobility.
To obtain the CahnHilliard equation one introduces a chemical potential w as
the variational derivative of E,
w :=

E
= u + 0 (u),
u

and defines a flux,

J := b(u)w.
Here b() is the nonnegative diffusional mobility, and in most of the literature on the
CahnHilliard equation b was assumed to be constant. But in the original derivation
of the equation a u-dependent mobility appeared ([10] and [24]), and in fact with
the diffusion in the interfacial region enhanced and hence stronger than in the pure
phases. This enhanced interfacial diffusion is, in particular, observed in experiments
at low temperatures.
It was suggested by many authors to take a mobility of the form b(u) := 1 u2 ;
but the main feature a mobility should have is that it is zero in the pure component,
i.e., when u = 1, and the mobility should be positive for |u| < 1. Having defined the
flux the CahnHilliard equation now follows from the equation u
t + J = 0, which
is a consequence of mass conservation. The system is completed by taking initial
conditions and the natural and no-flux boundary conditions u
= J = 0 on ,
where is normal to .
It is the aim of this work to develop an efficient numerical method for the Cahn
Hilliard equation with degenerate mobility. Besides the case in which the homogeneous free energy is smooth we want to be able to handle the cases of a logarithmic
free energy and of an obstacle potential. In the following we briefly describe what
is known for the CahnHilliard equation with a concentration dependent mobility.

288

JOHN W. BARRETT, JAMES F. BLOWEY, AND HARALD GARCKE

For an overview on the vast literature on the CahnHilliard equation with constant
mobility we refer to [17] and [26]. Existence results for the CahnHilliard equation
with degenerate mobility were obtained in [27] in one space dimension and in [19]
for space dimensions larger than one (see also [22] and [20]). One main feature of
these results is the fact that solutions with initial data where |u0 ()| 1 have the
property that |u(, t)| 1 for all later times t. This physically reasonable result is
true for all potentials and a degenerate mobility but cannot be guaranteed if one
takes (u) = (u2 a2 )2 and a constant mobility. We remark that so far no uniqueness
result for the CahnHilliard equation with degenerate mobility is known.
An important result which gave insight into the qualitative behavior of solutions
to the CahnHilliard equation with degenerate mobility was established in [11] (see
also [21]). They used the technique of formal asymptotic expansions to show that
with the scaling = 2 , t 2 t and in the deep quench limit 0, one can identify
a limit as 0, which is a geometric motion for the interface known as motion by
surface diffusion (cf. [13]). We remark that as 0 the interface thickness, which is

approximately proportional to , tends to zero and the interfacial region becomes

a sharp interface. The limiting motion is an evolution law for hypersurfaces and it
reads as
2

V = 16 S ,
where V is the normal velocity, S is the surface Laplacian, and is the mean
curvature of the interface. This is a purely local geometric motion for the interface
and is in contrast to the MullinsSekerka evolution, which is obtained with a constant
mobility. In the latter case, two interfaces which are a distance away from each other
are coupled through bulk terms. Whereas in the case of motion by surface diffusion,
two such interfaces would evolve independently of each other as long as they do not
intersect; for example, a collection of spheres which do not intersect each other are
stationary. On the level of the CahnHilliard equation with degenerate mobility this
property would correspond to a pinning effect (see [23], which reports on pinning
effects in spinodal decomposition of polymer mixtures). Also let us mention that the
time scale in the asymptotics of [11] is slower than the time scale which was used in
the asymptotics for a constant mobility. One should bear in mind these results when
studying the numerical simulations presented in section 5.
There are a number of papers on the CahnHilliard equation with constant mobility from the numerical analysis point of view. We refer to [17] for an overview.
Most numerical approaches are based on a splitting method, which uses the chemical
potential w as an unknown function and hence only requires continuous finite element
approximations for {u, w} (see [18]). To our knowledge there has been no numerical
analysis for the CahnHilliard equation with a degenerate mobility. However, in [4] an
error bound is proved for a fully practical piecewise linear finite element approximation of the CahnHilliard equation with a logarithmic free energy and a nondegenerate
concentration dependent mobility; see also [3] and [5] for extensions to the multicomponent case and the deep quench limit. Although the numerical approximation in
the above papers is well defined for a degenerate mobility, the authors were not able
to prove stability bounds and hence convergence of this approximation. The case of
a degenerate mobility with its possible lack of uniqueness requires a more delicate
approximation.
In what follows we state precisely the problem we wish to approximate numerically
and we make some general assumptions. Let be a bounded domain in Rd , d 3,

THE CAHNHILLIARD EQUATION WITH DEGENERATE MOBILITY

289

with a Lipschitz boundary . We consider the initial boundary value problem for
the CahnHilliard equation:
(P) Find u(x, t) such that
u
= .( b(u) (u + 0 (u)))
in T := (0, T ),
t
(1.3b)
x ,
u(x, 0) = u0 (x)
u

= b(u) (u + 0 (u)) = 0
(1.3c)
on (0, T ),

where is normal to and is a positive constant. The diffusional mobility b

C([1, 1]) is assumed to satisfy

(1.3a)

(1.4)

b(1) = b(1) = 0

and

b(s) > 0

s (1, 1).

The free energy C([1, 1]) is such that

(1.5)

(s) := 1 (s) + 2 (s) +

c
2 (1

s2 ),

where c is a nonnegative constant and 1 C 1 ((1, 1)) and 2 C 1 ([1, 1]) are
convex and concave, respectively. Clearly the third term can be absorbed into 2 ();
however, for later purposes we decompose () into this form. Obviously all examples
for given above can be written in the form (1.5). In particular the double obstacle
potential, (1.2), corresponds to the case 1 2 0 and c = 1 or 1 0, 2 (s) =
1s2
and c = 0. We point out that in the case of a degenerate mobility, the obstacle
2
in the potential (1.2) is not needed to describe the motion in the deep quench limit.
In particular, as was shown in [19], the evolution with respect to the potential (1.2)
is given by an equation instead of a variational inequality.
In [6] we considered a fully practical finite element approximation of the fourth
order nonlinear degenerate parabolic equation (1.3ac) with () 0 and b(u) := |u|p
for any given p (0, ). Such problems arise in lubrication approximations of thin
viscous films and have been studied extensively in the mathematics literature in recent
years. A key feature of this problem is that there is no uniqueness result. In addition
to establishing well posedness of our finite element approximation for all d 3, we
proved convergence in one space dimension to solutions using the very weak solution
concept introduced in [8] for this problem. This basically states that u is a solution if
Z
Z T
h u
,
i
dt

b(u)u dx dt = 0 L2 (0, T ; H 1 ()).

t
{|u|>0}

The restriction of convergence to one space dimension is due to the fact that our
a priori bounds on the finite element approximation only guarantee in the case of
d = 1 uniform boundedness and equicontinuity of the approximate solutions, which is
necessary to be able to pass to the limit in the discrete problem. For similar reasons,
the results of [8] were restricted to one space dimension. In this paper we extend the
techniques in [6] to the CahnHilliard equation with degenerate mobility, (1.3ac).
This paper is organized as follows. In section 2 we formulate a fully practical finite
element approximation, (Ph,t ), of problem (P) in the case of 1 , 2 C 1 ([1, 1]).
This obviously excludes the choice of as the logarithmic potential, (1.1). This
discretization is based on introducing the chemical potential w and writing the fourth
order parabolic equation as a system of equations
(1.6)

u
t

= .(b(u)w)

in T ,

u + 0 (u) = w,

290

JOHN W. BARRETT, JAMES F. BLOWEY, AND HARALD GARCKE

where the second equation holds on the set { |u| < 1 }; and as b(1) = b(1) = 0, see
(1.4), then w is only required in this region. Unfortunately, a naive finite element
approximation of problem (P) does not a priori guarantee that the discrete solution
fulfills |u| 1. Therefore we impose the physically reasonable property |u| 1 as a
constraint. This leads to a variational inequality which has to be solved at each time
step. We prove well posedness and stability bounds for our approximation, (Ph,t ),
of (P) for space dimensions 1, 2, and 3 and show convergence in one space dimension.
In section 3 we introduce and prove convergence of an iterative scheme, based on the
abstract splitting approach in [25], for solving the nonlinear discrete system arising in
(Ph,t ) at each time level. In section 4, we introduce a variation of the approximation
(Ph,t ), studied in section 2, to cope with the choice of a logarithmic free energy.
Furthermore, we extend the results of sections 2 and 3 to this case. Finally, in section
5 we report on some numerical experiments.
Notation and auxiliary results. We have adopted the standard notation for
Sobolev spaces, denoting the norm of W m,p () (m N, p [1, ]) by k km,p and
the seminorm by | |m,p . For p = 2, W m,2 () will be denoted by H m () with the
associated norm and seminorm written, respectively, as k km and | |m . Throughout,
(, ) denotes the standard L2 inner product over and h, i denotes the duality pairing

0

between H 1 () and H 1 ().

For later purposes, we recall the following well-known Sobolev interpolation results, e.g., see [1]: Let p [1, ], m 1,

if m dp > 0,

[p, ]
if m dp = 0,
r [p, )

[p,
d
d
m(d/p) ] if m p < 0,


d
1
1

and = m
p
r . Then there is a constant C depending only on , p, r, m such
that for all v W m,p () the inequality

kvk0,r Ckvk1
0,p kvkm,p

(1.7)

holds. It is convenient to introduce the inverse Laplacian operator G : F V such

that
(Gv, ) = hv, i H 1 (),

where F := v (H 1 ())0 : hv, 1i = 0 and V := {v H 1 () : (v, 1) = 0}. The well

posedness of G follows from the LaxMilgram theorem and the Poincare inequality
(1.8)

(1.9)

||0,p C(||1,p + |(, 1)|) W 1,p ()

and p [1, ].

One can define a norm on F by

(1.10)

kvk1 := |Gv|1 hv, Gvi 2

v F.

We note also for future reference that using a Youngs inequality yields for all > 0
that
(1.11)

hv, i = (Gv, ) kvk1 ||1

1
2
2 kvk1

+ 2 ||21

v F, H 1 ().

Throughout C denotes a generic constant independent of h and t, the mesh and

temporal discretization parameters. In addition C(a1 , . . . , aI ) denotes a constant
depending on the nonnegative parameters {ai }Ii=1 , such that for all C1 > 0 there
exists a C2 > 0 such that C(a1 , . . . , aI ) C2 if ai C1 for i = 1 I.

THE CAHNHILLIARD EQUATION WITH DEGENERATE MOBILITY

291

2. Finite element approximation. We consider the finite element approximation of (P) under the following assumptions on the meshes.
(A)

Let be a polyhedral domain. Let {T h }h>0 be a quasi-uniform family of

partitionings of into disjoint open simplices with h := diam() and
h := maxT h h , so that = T h .

Associated with T h is the finite element space

S h := { C() : | is linear T h } H 1 ().
We also introduce
K h := { S h : 1 1 in }.
Let J be the set of nodes of T h and {xj }jJ the coordinates of these nodes. Let
{j }jJ be the standard basis functions for S h ; that is, j K h and j (xi ) = ij
for all i, j J. We introduce h : C() S h , the interpolation operator, such that
h (xj ) = (xj ) for all j J. A discrete semi-inner product on C() is defined by
Z
X
(2.1)
h (1 (x) 2 (x)) dx
j 1 (xj ) 2 (xj ),
(1 , 2 )h :=

jJ
1

where j := (1, j ). The induced seminorm is then | |h := [( , )h ] 2 . We introduce

h : L2 () S h defined
the L2 projection Qh : L2 () S h and the more practical Q
by
h , )h = (, )
(Qh , ) = (Q

(2.2)

Sh.

Given N , a positive integer, let t := T /N denote the time step and tn := nt,
n = 1 N . Assuming that 1 , 2 C 1 ([1, 1]), we consider the following fully
practical finite element approximation of (P):
(Ph,t )
(2.3a)
(2.3b)

For n 1, find {U n , W n } K h S h such that



h

U n U n1
,
t

+ b(U n1 )W n , = 0 S h ,

(U n , ( U n )) + (10 (U n ) c U n , U n )h
(W n 20 (U n1 ), U n )h

K h,

where U 0 K h is an approximation of u0 K := { H 1 () : 1 1
h u0 . As we motivated
almost everywhere (a.e.) in }, e.g., U 0 h u0 (if d = 1) or Q
in the introduction the variational inequality (2.3b) is introduced because we wish
to impose the physically reasonable property |U n | 1, which is not automatically
guaranteed by a straightforward discretization of (P). It follows immediately from
(2.3b) that for n 1 and for all j J, either |U n (xj )| = 1 or |U n (xj )| < 1 and
(U n , j ) + (10 (U n ) c U n , j )h = (W n 20 (U n1 ), j )h .
Hence (2.3b) approximates u + 0 (u) = w in the region |u| < 1 as required;
see (1.6). We note that for a general degenerate
mobility b() satisfying (1.4), (2.3a)
R
is not fully practical as it assumes that b(U n1 ) dx can be calculated exactly. Obviously, one could consider using numerical integration on this term; e.g., replace (, )
by (, )h in (2.3a). However, for ease of exposition and as the model case b(s) := 1s2
can be easily dealt with, we consider (2.3a) in its present form.

292

JOHN W. BARRETT, JAMES F. BLOWEY, AND HARALD GARCKE

Below we recall some well-known results concerning S h .

(2.4)
(2.5)
(2.6)

||m,p2 Chd( p1 p2 ) ||m,p1 S h , 1 p1 p2 ,

||1,p Ch1 ||0,p S h , 1 p ;
lim k(I h )k0, = 0 C();

m = 0 or 1;

h0

(2.7) |(I Q )|0 + h|(I Qh )|1 Chm ||m H m (), m = 1 or 2;

(2.8)
||20 ||2h (d + 2)||20 S h ;
h
h
h
(2.9) |(v , ) (v , )| Ch1+m kv h km kk1 v h , S h , m = 0 or 1;
and if d = 1
(2.10) |(I h )|m,r Ch1m ||1,r W 1,r (), m = 0 or 1, any r [1, ];
(2.11)
lim k(I h )k1 = 0 H 1 ().
h0

If d = 1, then a simple consequence of (2.9) and (2.10) is that

|(v, )h (v, )| |( h v, h )h ( h v, h )| + |((I h )v, h )| + |(v, (I h ))|


v C(), H 1 ().
C |(I h )v|0 + h|v|0 kk1
(2.12)
h with Qh and noting (2.9), (2.5), and (2.7) yields that
Comparing Q
(2.13)

h )|1 Ch||1
h )|0 + h|(I Q
|(I Q

H 1 ().

It follows from (2.2) that

h )(xj )
(2.14) (Q

(, j )
(1, j )

jJ

h k0, kk0,
kQ

L ().

Similarly to (1.8), we introduce the operator Gh : F h V h such that

(2.15)

(Gh v, ) = (v, )h

Sh,

where V h := {v h S h : (v h , 1) = 0} and F h := {v C() : (v, 1)h = 0}. Similarly

to (1.11), we have for all > 0 that
(2.16) (v, )h (Gh v, ) |Gh v|1 ||1

1 h 2
2 |G v|1

+ 2 ||21

v F h, Sh.

We now follow the approach taken in [6]. To establish the existence of a solution
h,t
), we must introduce some notation. In particular weR define
{U n , W n }N
n=1 to (P
h
n1
sets V (U
) in which we seek the update U n U n1 . Given q h K h with q h :=
1
h
h
|| (q , 1) (1, 1), we define a set of passive nodes J0 (q ) J by
(2.17)

j J0 (q h ) (b(q h ), j ) = 0 b(q h ) 0 on supp(j ).

All other nodes we call active

SM nodes and they can be uniquely partitioned so that
J+ (q h ) := J \ J0 (q h ) m=1 Im (q h ), M 1, where Im (q h ), m = 1 M , are
mutually disjoint and maximally connected in the following sense: Im (q h ) is said

THE CAHNHILLIARD EQUATION WITH DEGENERATE MOBILITY

293

h
to be connected if for all j, k Im (q h ), there exist {` }L
`=1 T , not necessarily
distinct, such that

(a) xj 1 , xk L ,
(b) ` `+1 6= ,
6 0 on ` ,
(c) b(q h )

(2.18)

` = 1 L 1,
` = 1 L.

Im (q h ) is said to be maximally connected if there is no other connected subset of

J+ (q h ), which contains Im (q h ). Clearly J+ (q h ) is nonempty, since if (b(q h ), j ) = 0
h
h
h
j J then b(q h ) 0, and since
R h q S it follows that q 1 or 1 which
contradicts the assumption that q (1, 1). We then set
V h (q h ) := { v h S h : v h (xj ) = 0 j J0 (q h ) and (v h , m (q h ))h = 0, m = 1 M },
(2.19)
where for m = 1 M
m (q h ) :=

(2.20)

jIm

j .

(q h )

The space V h (q h ) consists of all those v h S h which are orthogonal, with respect to
the (, )h inner product, to j , for all j J0 (q h ), and to m (q h ), m = 1 M . We
note that for all q h K h , V h (q h ) V h and that |q h | < 1 = V h (q h ) V h . Another
immediate consequence of the above definitions is that on any T h either
(2.21)

b(q h ) 0

or

m? (q h ) 1 for some m? and m (q h ) 0 for m 6= m? .

For later reference we state that any v h S h can be written as

"Z
M
X
X
X
h
h
h
h
(2.22a) v
v (xj )j v +
v (xj )j +

jJ

jJ0 (q h )

m=1

where m (q h ) := { T h : m (q h )(x) = 1 x },
Z
(v h , m (q h ))h
, and
v h :=

(1, m (q h ))
m (q h )
(2.22b)
"
Z
M
X
X
h
h
v :=
v (xj )
m=1 jIm (q h )

#
m (q h )

m (q h ),

#
m (q h )

j V h (q h )

is the projection
with respect to the (, )h scalar product of v h onto V h (q h ). We
R
remark that m (qh ) v h is not the standard mean value on the set m (q h ).
InR order to express W n in terms of U n and U n1 we introduce for all q h K h
with q h (1, 1) the discrete anisotropic Greens operator Gqhh : V h (q h ) V h (q h )
such that
(2.23)

(b(q h )Gqhh v h , ) = (v h , )h

Sh.

To show the well posedness of Gqhh , we first note that choosing j , j J0 (q h ),

in (2.23) leads to both sides vanishing on noting (2.17) and (2.19). Similarly, choosing m (q h ), m = 1 M , in (2.23) leads to both sides vanishing on noting

294

JOHN W. BARRETT, JAMES F. BLOWEY, AND HARALD GARCKE

(2.21) and (2.19). Therefore for well posedness, it remains to prove uniqueness as
V h (q h ) has finite dimension. If there exist two solutions Zi V h (q h ), i = 1, 2, with
(b(q h )Zi , ) = (v h , )h for all S h , then Z := Z1 Z2 V h (q h ) satisfies, on
noting (2.21),
M Z
X
bmin (q h )

h
m=1 m (q )

|Z|2 dx

M Z
X
h
m=1 m (q )

b(q h )|Z|2 dx

b(q h )|Z|2 dx = 0,

(2.24a)
where (q h ) :=

SM

m=1

m (q h ) and
bmin (q h ) :=

(2.24b)

1
(q h ) ||
min

b(q h ) dx.

Hence it follows that Z is constant on each m (q h ). However as Z V h (q h ), it follows

that Z 0. Thus Gqhh is well posed.
For later purposes we note from (2.23) and Youngs inequality that, similarly to
(2.16), for any > 0,
(2.25)

|
v h |2h

1
h
h h |2
0
2 |b(q )Gq h v

+ 2 |
v h |21

vh V h (q h ).

Theorem
2.1. Let and T h be such that assumption (A) holds and let U 0 K h
R 0
with U (1, 1). In addition let b and fulfill the assumptions stated above. Then
h,t
).
for all t > 0 there exists a solution {U n , W n }N
n=1 to (P
2
n1
n1
:= kb(U n1 )k0, ,
If c bmax t < 4, where bmax maxn=1N bmax and bmax
n N
then {U }n=1 is unique. Furthermore, the following stability bounds hold:
max kU n k21 + (t)2

n=1N

(2.26)

N
X

|U

n=1

+ t

N
X

U n1 2
|1
t

1 h U
[bn1
|G [
max ]

n=1

+ t

N
X

|[b(U n1 )] 2 W n |20

n=1

U n1 2
]|1
t



C |U 0 |21 + 1 .

In addition W n is unique on m (U n1 ) if |U n (xj )| < 1 for some j Im (U n1 ),

m = 1 M, n = 1 N.
Proof. It follows from (2.3a) and (2.23) that for n 1, given U n1 K h , we seek
n
U K h (U n1 ), where
(2.27)

K h (U n1 ) := { K h : U n1 V h (U n1 ) }.

In addition a solution W n in (2.3a), (2.3b) can be expressed in terms of U n as (cf.

(2.23) and (2.22a,b))
(2.28)

n
n1
]+
W n GUh n1 [ U U
t

jJ0

nj j +

(U n1 )

M
X

nm m (U n1 ),

m=1

h,t
) can be restated as
where {nj }jJ0 (U n1 ) and {nm }M
m=1 are constants. Hence (P
n
h
n1
) and constant Lagrange multipliers
the following: For n 1, find U K (U

THE CAHNHILLIARD EQUATION WITH DEGENERATE MOBILITY

295

{nj }jJ0 (U n1 ) , {nm }M

m=1 such that
h

n
n1
0
n
n
n
]
+

(U
)

U
,

U
(U n , ( U n )) + GUh n1 [ U U
c
1
t

h
M
X
X
(2.29)
nj j +
nm m (U n1 ) 20 (U n1 ), U n
jJ0 (U n1 )

K h.

m=1

It follows from (2.29), (2.27), and (2.19) that U n K h (U n1 ) is such that

n
n1
v h U n )) + (GUh n1 [ U U
] + 10 (U n ) c U n , vh U n )h
(U n , (
t

(20 (U n1 ), vh U n )h

(2.30)

vh K h (U n1 ).

We note that (2.30) is the EulerLagrange variational inequality of the minimization

problem
n
1
1
E h (
v h ) := |
v h |21 + t
|[b(U n1 )] 2 GUh n1 (
v h U n1 )|20
min
v
h K h (U n1 )
o
+(21 (
v h ) c (
v h )2 + 220 (U n1 )
v h , 1)h .
As we minimize E h () on the compact set K h (U n1 ), we have existence of a solution to
(2.30). Existence of the Lagrange multipliers {nj }jJ0 (U n1 ) and {nm }M
m=1 , for fixed
n, follows from standard optimization theory; e.g., see [15]. Therefore, on noting
h,t
).
(2.28), we have existence of a solution {U n , W n }N
n=1 to (P
n,i
n,i
M
For fixed n 1, if (2.29) has two solutions { U , {j }jJ0 (U n1 ) , {n,i
m }m=1 },
n
i = 1, 2, then it follows from (2.30), the convexity of 1 (), and (2.25) that U :=
U n,1 U n,2 V h (U n1 ) satisfies
n

n
1
n1 12
)] GUh n1 U |20
t |[b(U
n
n
n
1
1
|U |21 + t
|[b(U n1 )] 2 GUh n1 U |20 + (10 (U n,1 ) 10 (U n,2 ), U )
n
n
n
1
2 bn1 t
1
|[b(U n1 )] 2 GUh n1 U |20 + c max
|U |21 .
c |U |2h t
4

|U |21 +

R
R
Therefore the uniqueness of U n follows from (1.9) and U n = U 0 under the stated
restriction on t. If |U n (xj )| < 1 for some j Im (U n1 ), then (1 (U n (xj ))2 )
m (U n1 (xj )) > 0 and choosing U n h [(1 (U n )2 ) m (U n1 )] 6 0 in (2.29)
for > 0 sufficiently small yields uniqueness of the Lagrange multiplier nm . Hence
the desired uniqueness result on W n follows from noting (2.28).
We now prove the stability bound (2.26). For fixed n 1 choosing W n in
(2.3a), U n1 in (2.3b), and combining yields that

(2.31)

n 2
n
n1 2
|1 2 |U n1 |21 + ((U n ), 1)h ((U n1 ), 1)h
2 |U |1 + 2 |U U
1
2c |U n U n1 |2h + t|[b(U n1 )] 2 W n |20
(U n , (U n U n1 )) + (10 (U n ) + 20 (U n1 ), U n U n1 )h
1
+t|[b(U n1 )] 2 W n |20 c (U n , U n U n1 )h 0,

where we have noted the identity

(2.32)

2s(s r) = s2 r2 + (s r)2

r, s R,

296

JOHN W. BARRETT, JAMES F. BLOWEY, AND HARALD GARCKE

and the convexity and concavity of 1 and 2 , respectively. Choosing t(U n

3
U n1 ) in (2.3a) and applying a Youngs inequality, similarly to (2.25), with = 2
c
and noting the stated restriction on t, it follows from (2.31) that

n 2
2 |U |1

+ 8 |U n U n1 |21 2 |U n1 |21
1

+((U n ), 1)h ((U n1 ), 1)h + 23 t|[b(U n1 )] 2 W n |20 0.

(2.33)

Summing from n = R1 m,R for m = 1 N , and noting the properties of i

(i = 1, 2), (1.9), and U n = U 0 yields the first three bounds in (2.26). Choosing
n
n1
) in (2.3a) and noting (2.15) yields for n 1 that
Gh ( U U
t
|Gh [ U

U n1 2
]|1
t

(2.34)

n
n1
U n1 h U n U n1 h
,G [
]) = (b(U n1 )W n , Gh [ U U
])
t
t
t
n1
n 2
n1
n1 12
n 2
)]W |0 bmax |[b(U
)] W |0 .
|[b(U

= (U

Summing (2.34) from n = 1 N and noting the third bound in (2.26) yields the
desired fourth bound in (2.26).
Remark. (i) Given a convex function 1 C([, 1, 1]), a concave function
C([1, 1]) and R+ , the free energy (s) := 1 (s)+(s)+ 2 (1s2 ) can be written
in the form (1.5) with either (a) 2 () () and c = or (b) 2 () ()+ 2 (1s2 )
and c = 0. We see from Theorem 2.1 that a time step restriction is required for the
well posedness of (Ph,t ) in case (a) but not in case (b).
(ii) As can be seen from (2.33) the finite element approximation has the property
that 2 |U |21 + ((U ), 1)h is a Lyapunov function for the discrete evolution.
Let
(2.35a)

U (t) :=

ttn1 n
t U

tn t n1
,
t U

t [tn1 , tn ],

n 1,

and
U + (t) := U n ,

(2.35b)

U (t) := U n1 ,

t (tn1 , tn ],

n 1.

We note for future reference that

U
U U = (t t
n ) t ,

(2.36)

t (tn1 , tn ),

n 1,

where t+
n := tn and tn := tn1 . Using the above notation and introducing analogous
notation for W , (2.3a,b) can be restated as the following.
Find {U, W } H 1 (0, T ; K h ) L2 (0, T ; S h ) such that

Z
(2.37a)
Z

T
0

(2.37b)

T
0

h
U
t ,

i
+ b(U )W + , dt = 0 L2 (0, T ; S h ),


(U + , ( U + )) + (10 (U + ) + 20 (U ) c U + , U + )h dt
Z

T
0

(W + , U + )h dt

L2 (0, T ; K h ).

R
Theorem 2.2. Let d = 1 and u0 K with u0 (1, 1). Let {T h , U 0 , t}h>0
be such that
(i) U 0 K h and U 0 u0 in H 1 () as h 0,
(ii) and {T h }h>0 fulfill assumption (A),

THE CAHNHILLIARD EQUATION WITH DEGENERATE MOBILITY

297

(iii) t 0 as h 0.
Then there exists a subsequence of {U, W }h and a function u L (0, T ; K)
1 1

2 8
(T ) and a w L2loc ({|u| < 1}) with
H 1 (0, T ; (H 1 ())0 ) Cx,t
such that as h 0,

w
x

L2loc ({|u| < 1})

uniformly on T ,
U, U u

weakly in L2 (0, T ; H 1 ()),

U, U u

(2.38)
(2.39)

W +
x

W + w,

(2.40)

w
x

weakly in L2loc ({|u| < 1}),

where {|u| < 1} := {(x, t) T : 1 < u(x, t) < 1 }.

Furthermore, u and w fulfill u(, 0) = u0 () and
Z T
Z

u
2
1
(2.41a)
h t , i dt +
b(u) w
x x dx dt = 0 L (0, T ; H ()),
{|u|<1}

(2.41b)

w = xu2 + 0 (u)

on the set

{|u| < 1}.

Proof. Without loss of generality we assume that c2 kbk0, t R< 4 and that h is
sufficiently small. We note that the assumption (i) implies that U 0 (1, 1) for
h sufficiently small. Hence the definition (2.35a,b), the first three bounds in (2.26),
(2.10), and (1.9) together with the fact that V h (q h ) V h imply

+ 2
1

(2.42) kU k2L (0,T ;H 1 ()) + t kU k2H 1 (0,T ;H 1 ()) + [b(U )] 2 W

C.
x 2
L (T )

Furthermore, we deduce from (2.36) and (2.42) that

(2.43)

2
kU U k2L2 (0,T ;H 1 ()) (t)2 k U
t kL2 (0,T ;H 1 ()) C t.

In the next step we show that the discrete solutions U are uniformly H
older
continuous. The first bound in (2.42) gives via a standard imbedding result
(2.44)

|U (y2 , t) U (y1 , t)| C |y2 y1 | 2

y1 , y2 ,

t 0.

In addition it follows from (1.7), (2.8), (2.16), (2.42), and (2.26) that
1

kU (, tb ) U (, ta )k0, C kU (, tb ) U (, ta )k02 kU (, tb ) U (, ta )k12

1

C |Gh (U (, tb ) U (, ta )) |14 kU (, tb ) U (, ta )k14


  14 
 Z tb
 34
 h

U



2
kU
k
C G
(,
t)
dt

1
L (0,T ;H ())
t

Z

C 

ta

tb

ta

(2.45)

 14
Z

1
h
U
G t (, t) dt  C(tb ta ) 8

ta

C(tb ta )

tb

1
8

 18


 h U 2
G t  dt
1

tb ta 0.

An immediate consequence of (2.45) is that

(2.46)

kU U kL (T ) C(t) 8 .
1 1

2 8
(T ) norm of U is bounded indeNow (2.42), (2.44), and (2.45) imply that the Cx,t
pendently of h, t, and T . Hence, under the stated assumptions on t, every sequence

298

JOHN W. BARRETT, JAMES F. BLOWEY, AND HARALD GARCKE

{U }h is uniformly bounded and equicontinuous on T for any T > 0. Therefore by

the Arzel`aAscoli theorem there exists a subsequence such that
1 1

(2.47)

2 8
(T )
U u Cx,t

uniformly on T as h 0

and |u| 1. Moreover (2.42) implies that this same subsequence is such that
(2.48)

weakly in L2 (0, T ; H 1 ()) as h 0.

U u

For any H 1 (0, T ; H 1 ()) we choose h in (2.37a) and now analyze the
subsequent terms. First, we have that
Z

U
t

(2.49) =

, h

h

dt
h

U, (t )

h

h

h
dt + U (, T ), h (, T ) U (, 0), h (, 0) .

Next we conclude using the regularity of , (2.12), and (2.47) that

Z
(2.50)

T
0

U, (t )

h

Z
dt

u,
t

dt

as h 0

for all as above.

In view of (2.42) we deduce that


Z


W +
h


b(U
)
(I

)
dx
dt
x x


T

h
2
2
1
k[b(U )] 2 kL (T ) k[b(U )] 2 W
x kL (T ) k(I )kL (0,T ;H ())

(2.51)

Ck(I h )kL2 (0,T ;H 1 ()) .

We now show the compactness of {W + }h on compact subsets of {|u| < 1}. For any
> 0, we set
D+ := { (x, t) T : |u(x, t)| < 1 }
(2.52)

and D+ (t) := { x : |u(x, t)| < 1 }.

For a fixed > 0, it follows from (2.47) and (2.46) that there exists an h0 () R+
such that for all h h0 ()
(2.53)

and

(x, t) 6 D+
1 2 |U (x, t)| 1
|U (x, t)| 1 18 (x, t) D+
.
4

On noting (2.53) and (2.42) we have that


Z


+


b(U ) W
dx
dt


x x

 T \D+

2
2
1
k[b(U )] 2 kL (T \D+ ) k[b(U )] 2 W
x kL (T ) kkL (0,T ;H ())

C[Bmax (2)] 2 kkL2 (0,T ;H 1 ())

(2.54)

L2 (0, T ; H 1 ()),

h h0 (),

299

THE CAHNHILLIARD EQUATION WITH DEGENERATE MOBILITY

and
Bmin ( 8 )
(2.55)
where

Z
D+

+
2
| W
x |

Bmax () :=

Z
dx dt

max

1|z|1

D+

b(z)

2
b(U )| W
x | dx dt C

h h0 (),

and Bmin () := min b(z).

|z|1

In what follows we want to relate W + to U + and U on the sets D+ . From (2.53)

we have that for all h h0 () and for almost every (a.e.) t (0, T )
(, t) U + (, t) 18

h (, t)
Kh
k h (, t)k0,
h L2 (0, T ; S h )

(2.56)

with supp( h ) D+
.
4

Choosing such in (2.37b) yields for all h h0 () that

Z
Z Th
i
+ h
0
+
0

+ h h
( U
dt
=
,
)
+
(
(U
)
+

(U
)

U
,

)
c
1
2
x
x
0

h L2 (0, T ; S h )

(2.57)

(W + , h )h dt

with supp( h ) D+
.
4

Next we derive a bound of W + locally on the set {|u| < 1}. For any t [0, T ], we
choose a cut-off function (, t) C0 (D+
(t)) such that
2

(2.58)

(, t) 1

on

D+ (t),

0 (, t) 1,

| x
(, t)| C 2 .

This last property can be achieved since for y1 , y2 such that |u(y1 , t)| 1 12
1
and |u(y2 , t)| 1 we have from (2.47) that 12 |u(y2 , t) u(y1 , t)| C|y2 y1 | 2 .
It follows from (2.14) and (2.58) that there exists an h1 () h0 () such that
h (2 W + )) D+
supp(Q

(2.59)

h h1 ().

It follows from (2.2), (2.59), (2.57), (2.13), the continuity properties of i (i = 1, 2),
|U | 1, and (2.58) that for all h h1 ()
Z T
Z
h (2 W + ))h dt
2 (W + )2 dx dt =
(W + , Q

T
0

i
+
h 2
+
0
+
0

+ h 2
+ h
( U
dt
x , x (Q ( W ))) + (1 (U ) + 2 (U ) c U , Q ( W ))

(2 W + )kL2 (T ) + Ck W + kL2 (T )
CkU + kL2 (0,T ;H 1 ()) k x


1
+
W +
+
C( ) (kU kL2 (0,T ;H 1 ()) + 1) k x kL2 (D+ ) + k W kL2 (T ) .

(2.60)
Applying Youngs inequality then gives
Z
(2.61)

+ 2

(W ) dx dt C(

"

) 1+

kU + k2L2 (0,T ;H 1 ())

+
2
k W
x kL2 (D + )

300

JOHN W. BARRETT, JAMES F. BLOWEY, AND HARALD GARCKE

Therefore combining (2.55), (2.61), and (2.42) we have that

(2.62)

kW + kL2 (0,T ;H 1 (D+ (t))) C( 1 )[Bmin ( 8 )]1 C( 1 )

h h1 ().

The last estimate implies the existence of a subsequence and a w L2 (0, T ; H 1 (D+ (t)))
such that
W + w,

(2.63)

W +
x

w
x

weakly in L2 (D+ ) as h 0.

Next noting the uniform continuity of b, (2.55), and (2.38) we conclude that

Z





W +
[b(U ) b(u)] x x dx dt 


 D+

2
1
+
kb(u) b(U )kL (T ) k W
x kL2 (D ) kkL (0,T ;H ())

(2.64)

C
1

[Bmin ( 8 )] 2

kb(u) b(U )kL (T ) kkL2 (0,T ;H 1 ())

will converge to 0 as h 0.
Combining (2.51), (2.64), and (2.63) and noting (2.11), (2.47), and (2.46) yields
that
Z
Z

W +
h
b(U ) x x ( ) dx dt
b(u) w
x x dx dt as h 0
D+

D+

L2 (0, T ; H 1 ()).

(2.65)

Moreover, by (2.11), (2.48), and (2.43) we have that

Z T
Z T

U +
h
(2.66)
( x , x ( )) dt
( u
x , x ) dt as h 0
0

L2 (0, T ; H 1 ()).

Using (2.1), (2.10), and (2.62) we deduce that

 Z
Z

 T

  

+
h h
+
h
+

(I )(W ) dx dt 
(W , ) (W , ) dt  


 0
T
Z

| x
(W + )| dx dt
Ch
T

Ch kW + kL2 (0,T ;H 1 (D+ (t))) kkL2 (0,T ;H 1 ())

C(

) h kkL2 (0,T ;H 1 ())

L (0, T ; H 1 ()) with supp() D+ .

2

(2.67)

Noting that 1 C 1 ([1, 1]) and using (2.1), (2.10), (2.38), (2.12), and (2.6) yields
that

Z

 T



(10 (U + ), h )h (10 (u), ) dt 


 0
Z T
Z T

 0 +

 0
0
h h

(1 (u), h )h (10 (u), ) dt
(1 (U ) 1 (u), ) dt +

(2.68)

as h 0

L2 (0, T ; H 1 ()).

THE CAHNHILLIARD EQUATION WITH DEGENERATE MOBILITY

301

Using a similar argument for the remaining terms and combining (2.67), (2.63),
and (2.68) implies that
Z T
Z T
(W + 10 (U + ) 20 (U ) + c U + , h )h dt
(w 0 (u), ) dt
0

(2.69)

L (0, T ; H ())

as h 0

with supp() D+ .

Combining (2.66) and (2.69) and noting (2.57) yields that

Z
h
i

0
u
x x + ( (u) w) dx dt = 0
D+

L2 (0, T ; H 1 ()) with supp() D+ .

(2.70)

This uniquely defines w in terms of u on the set D+ . Repeating (2.65) for all > 0
and noting (2.54), Bmax () 0 as 0 and (2.11) yield that
Z
Z
+

h
b(U ) W
(
)
dx
dt

b(u) w
x x
x x dx dt as h 0
D0+

L2 (0, T ; H 1 ()).

(2.71)

Combining (2.37a), (2.49), (2.50), (2.71) and arguing similarly as in (2.68) by using
(2.38), (2.12), (2.6), and assumption (i) we conclude that for all H 1 (0, T ; H 1 ())
Z T
Z

(2.72) (u(, T ), (, T )) (u0 (), (, 0))

(u,
)
dt
+
b(u) w
t
x x dx dt = 0.
0

The fact that

o
n
+
b(U ) W
x

h>0

D0+

is uniformly bounded in L2 (T ) implies that

+
2
1
1
0
b(u) w
x L (D0 ) and hence we conclude from (2.72) that u H (0, T ; (H ()) ).
Therefore combining the above results, repeating (2.70) for all > 0 yields that u
1 1

2 8
(T ) and w L2loc (D0+ ), with
L (0, T ; H 1 ()) H 1 (0, T ; (H 1 ())0 ) Cx,t
+
2
Lloc (D0 ), are such that u(, 0) = u0 () and

Z
(2.73a)


t , dt +

h
D0+

(2.73b)

Z
D0+

w
x

2
1
b(u) w
x x dx dt = 0 L (0, T ; H ()),

0
u
x x + ( (u) w)

dx dt = 0

L2 (0, T ; H 1 ()) with supp() D0+ .

Hence we have established the desired result (2.41a,b).

Remark. Theorem 2.2 also establishes existence of a solution to problem (P) and
yields the result of [27], where existence is proved in one space dimension, (see also
[19]). In addition we note that we assumed only continuity of the mobility b. All other
existence results for degenerate parabolic equations of fourth order in the literature
require at least H
older regularity for b.
3. Solution of the discrete variational inequality. We now consider an
algorithm for solving the variational inequality at each time level in (Ph,t ). This
is based on the general splitting algorithm of [25]; see also [16] and [2] where this
algorithm has been applied to solve (Ph,t ) with constant mobility.

302

JOHN W. BARRETT, JAMES F. BLOWEY, AND HARALD GARCKE

For n fixed, multiplying (2.3b) by > 0, a relaxation parameter, adding

(U n , U n )h to both sides, and rearranging on noting (2.3a) and (2.29), it follows
that {U n , W n } K h S h satisfy
h

(3.1a)
(U n + 10 (U n ), U n ) (Z n , U n )h K h ,
h
 n
U U n1
, + bn1 (W n , ) = ([bn1 b(U n1 )]W n , ) S h ,
t
(3.1b)
where Z n S h is such that


(Z n , )h := (U n , )h (U n , ) + (20 (U n1 ) c U n W n , )h
(3.1c)

Sh

n1
, bmax ] with bn1
and bn1 is chosen such that bn1 [bmax
max and bmax as defined in
n
h
Theorem 2.1. We introduce X S such that


Sh
(X n , )h := (U n , )h + (U n , ) + (20 (U n1 ) c U n W n , )h
(3.1d)

and note that X n = 2U n Z n . We use this as a basis for constructing our iterative
procedure: For n 1 set {U n,0 , W n,0 } {U n1 , W n1 } K h S h , where W 0 S h
is arbitrary if n = 1.
For k 0 we define Z n,k S h such that for all S h


(Z n,k , )h = (U n,k , )h (U n,k , ) + (20 (U n1 ) c U n,k W n,k , )h .
(3.2a)
1
Then find U n,k+ 2 K h such that
1

U n,k+ 2 (xj ) = U n1 (xj )

(U

n,k+ 12

(xj ) + 10 (U

(3.2b)

n,k+ 12

if j J0 (U n1 ),
1

(xj )) Z n,k (xj ))(r U n,k+ 2 (xj )) 0

r [1, 1] if j J+ (U n1 ),

and find {U n,k+1 , W n,k+1 } S h S h such that



h
U n,k+1 U n1
, + bn1 (W n,k+1 , ) = ([bn1 b(U n1 )]W n,k , )
t

(3.2c)
(U

n,k+1

Sh,


n,k+1
0
n1
n,k+1
, ) + (U
, ) + (2 (U
) c U
W n,k+1 , )h
h

= (X n,k+1 , )h

(3.2d)
1

Sh,

where X n,k+1 := 2U n,k+ 2 Z n,k . For j J+ (U n1 ) existence and uniqueness

1
of U n,k+ 2 (xj ) in the variational inequality (3.2b) follows from the monotonicity of
10 ().
It remains to show that (3.2c) and (3.2d) possess a unique solution {U n,k+1 ,
n,k+1
} S h S h . Let An,k V h be such that
W
(3.3)

(An,k , )h = (b(U n1 )W n,k , ) S h .

303

THE CAHNHILLIARD EQUATION WITH DEGENERATE MOBILITY

It then follows from (3.2c), (2.15), and (3.2d) with 1 that

R
n,k+1
U n1
+ An,k )
W n,k+1 = (I )W n,k [bn1 ]1 Gh ( U
t
R 1 n,k+1
(3.4)
X n,k+1 ) + h 20 (U n1 ) c U n,k+1 ).
+ ( (U
n,k+1
h
Sm
:= {v h
Therefore
R h (3.2c,d)
R 0 may be written equivalently as follows: find U
h
S : v = U } such that
R
(U n,k+1 , (I ))h
h

h i
h n,k+1 n1 i
R
U
n,k+1
+ (U n,k+1 , ) + [bn1 ]1 Gh U
(I

)U
,

c
t
R
n,k+1
n,k
0
n1
n1 1 h n,k
+ (W
2 (U
) [b
] G A ), (I ))h S h .
= (X
(3.5)
h
satisfying (3.5) follows, noting (2.16) and the time step
Existence of U n,k+1 Sm
2 n1
t < 4, since this is the EulerLagrange equation of the minimizarestriction c b
tion problem
h
i

1
h ( U n1 )|2 c ||2
|
G
min ||2h + ||21 + bn1
0
h
t
h
Sm

2(X n,k+1 + (W n,k 20 (U n1 ) [bn1 ]1 Gh An,k ), )h .

Uniqueness of U n,k+1 follows in a similar way to that of U n . Finally, W n,k+1 is
uniquely defined by (3.4). Hence the iterative procedure (3.2ad) is well defined.
Theorem 3.1. Let 3c2 bn1 t < 4. Then for all R+ and {U n,0 , W n,0 }
S h S h the sequence {U n,k , W n,k }k0 generated by the algorithm (3.2ad) satisfies
Z
U n,k U n and
(3.6)
b(U n1 )|(W n,k+1 W n )|2 dx 0 as k .

In addition, if 10 () is strictly monotone then U n,k+ 2 U n as k .

Proof. It follows from (3.1c), (3.1d), (3.2a), (3.2d), and by the definition of X n,k+1
that for k 0
U n = 12 (X n + Z n ),

(3.7)

U n,k = 12 (X n,k + Z n,k ),

U n,k+ 2 = 12 (X n,k+1 + Z n,k ).

h
, it follows from (3.2d), (3.1d), and (3.7) that
As U n,k+1 , U n Sm

|U n,k+1 U n |21 c |U n,k+1 U n |2h (W n,k+1 W n , U n,k+1 U n )h

1
= 4
(X n,k+1 X n Z n,k+1 + Z n , X n,k+1 X n + Z n,k+1 Z n )h
(3.8)

1
n,k+1
4 (|X

X n |2h |Z n,k+1 Z n |2h ).

Choosing U n,k+ 2 in (3.1a) and for j J+ (U n1 ) choosing U n (xj ) in

(3.2b), multiplying by j on recalling (2.1), and summing over j yields, on noting
1
that U n (xj ) = U n,k+ 2 (xj ) for j J0 (U n1 ),
1

|U n,k+ 2 U n |2h +(10 (U n,k+ 2 )10 (U n ), U n,k+ 2 U n )h (Z n,k Z n , U n,k+ 2 U n )h .

(3.9)
Combining (3.9) and (3.7) yields that

304

JOHN W. BARRETT, JAMES F. BLOWEY, AND HARALD GARCKE

1

(3.10) 4(10 (U n,k+ 2 ) 10 (U n ), U n,k+ 2 U n )h + |X n,k+1 X n |2h |Z n,k Z n |2h .

Using (3.2c), (3.1b), and (2.32) it follows that
(W n,k+1 W n , U n,k+1 U n )h
= (W n,k+1 W n , U n,k+1 U n1 )h (W n,k+1 W n , U n1 U n )h
1

= t|[b(U n1 )] 2 (W n,k+1 W n )|20

+t([bn1 b(U n1 )](W n,k+1 W n,k ), (W n,k+1 W n ))
h
1
1
n1
= t|[b(U n1 )] 2 (W n,k+1 W n )|20 + t
b(U n1 )] 2 (W n,k+1 W n )|20
2 |[b
i
1
1
|[bn1 b(U n1 )] 2 (W n,k W n )|20 + |[bn1 b(U n1 )] 2 (W n,k+1 W n,k )|20 .
(3.11)
Similarly to (3.11) and using Youngs inequality we have that for any (0, 1)
c |U n,k+1 U n |2h = c (U n,k+1 U n1 , U n,k+1 U n )h + c (U n1 U n , U n,k+1 U n )h
= c t(b(U n1 )(W n,k+1 W n ), (U n,k+1 U n ))
c t([bn1 b(U n1 )](W n,k+1 W n,k ), (U n,k+1 U n ))


1
2 bn1 t
1
n1
b(U n1 )] 2 (W n,k+1 W n,k )|20
c 4
2 + 1
|U n,k+1 U n |21 + t
2 |[b
1

+(1 )t|[b(U n1 )] 2 (W n,k+1 W n )|20 .

(3.12)
Combining (3.8), (3.10), (3.11), (3.12), and rearranging yields that



2 bn1 t
1
2 + 1
|U n,k+1 U n |21
c 4
1
+ 4
|Z n,k+1 Z n |2h +

t
n1
2 |[b

b(U n1 )] 2 (W n,k+1 W n )|20

+t|[b(U n1 )] 2 (W n,k+1 W n )|20 + (10 (U n,k+ 2 ) 10 (U n ), U n,k+ 2 U n )h

1
n,k
4 |Z

Z n |2h +

t
n1
2 |[b

b(U n1 )] 2 (W n,k W n )|20 .

(3.13)
Therefore noting the monotonicity of 10 () and the restriction on t we have that
1
1
n1
|Z n,k Z n |2h + t
b(U n1 )] 2 (W n,k W n )|20 }k0
for sufficiently small { 4
2 |[b
is a decreasing sequence which is bounded below and so has a limit. Therefore the
desired results (3.6) follow from this and (3.13).
Remark. We see from (3.2ad) and (3.5) that at each iteration one needs to solve
only (i) a fixed linear system with constant coefficients and (ii) a nonlinear equation
at each mesh point. On a uniform mesh (i) can be solved efficiently using a discrete
cosine transform; see [9, section 5], where a similar problem is solved.
4. Logarithmic free energy. In this section we modify our approximation
(Ph,t ) and the results in the previous two sections to cope with the logarithmic
free energy, that is


1s
(4.1)
1 (s) := 2 (1 + s) ln[ 1+s
2 ] + (1 s) ln[ 2 ] .
Here we have the additional difficulty that 0 () (see (1.5)) is not uniformly bounded
on (1, 1) with 10 (1) = .

THE CAHNHILLIARD EQUATION WITH DEGENERATE MOBILITY

305

Our modified approximation is the following.

e
(P

h,t

(4.2a)

For n 1, find {U n , W n } S h S h such that



h

U n U n1
,
t

+ b(U n1 )W n , = 0 S h ,

(U n , ) + (10 (U n ) c U n , )h = (W n 20 (U n1 ), )h
(4.2b)

Ve h (U n1 ),

where U 0 K h is an approximation of u0 and for q h K h we define



(4.3)
Ve h (q h ) := v h S h : v h (xj ) = 0 j J0 (q h ) .
Clearly (4.2b) implies that |U n (xj )| < 1 for all j J+ (U n1 ). Moreover, we will show
e h,t ) has the property that kU 0 k0, < 1 implies kU n k0, < 1 for all n 1.
that (P
We prove well posedness of this approximation via the regularization


 


(1 s)2 + 2 (1 s) ln 2
if s 1 ,
+ 4
2 (1 + s) ln 1+s
2
4
1, (s) := 1 (s)
 1s 
  if |s| 1 ,

2
(1

s)
ln
(1
+
s)
+
(1
+
s)
ln
if s 1 + .
+
2
2
4
2
2 4
(4.4)
Let us emphasize that we introduce 1, only to prove well posedness of problem
e h,t ) directly. We note that 1, (s) 1 (s) for all
e h,t ). In practice we solve (P
(P
|s| 1 and define 1, (s) := 1, (s) + 2c (1 s2 ) for all s R. The monotone function

(1 s) 2 ln
if s 1 ,
2 (1 + ln(1 + s)) 2
0
0
1 (s)
if |s| 1 ,
(4.5)
1, (s) =

(1 + s) + 2 ln if s 1 + ,
2 (1 + ln(1 s)) + 2
and the function 01, satisfy the following properties:
For all r, s R, on noting (2.32),
0
(s)(r s) c s(r s) 1, (r) 1, (s) + c s(s r)
01, (s)(r s) = 1,

(4.6)

= 1, (r) 1, (s) +

c
2 (r

s)2 .

For 1 and for all r, s R,

0
0
00
0
0
(r) 1,
(s))2 k1,
k0, (1,
(r) 1,
(s))(r s)
(1,

(4.7)

0
0
(r) 1,
(s))(r s).
(1,

It is a simple matter to show that 1, is bounded below for sufficiently small; e.g.,
if 0 := /(8c ), then

(4.8)
s R,
[s 1]2+ + [1 s]2+ c c
1, (s) 8
where []+ := max{, 0}; see [4] for details. In addition we introduce the concave
preserving extension e2 C 1 (R) of 2 C 1 ([1, 1]),

if s 1,
2 (1) + (s 1)20 (1)
if |s| 1,
2 (s)
(4.9)
e2 (s) :=

2 (1) + (s + 1)20 (1) if s 1,

306

JOHN W. BARRETT, JAMES F. BLOWEY, AND HARALD GARCKE

and then define

(4.10)

(s) := 1, (s) + e2 (s)

s R.

It follows immediately from (4.8) and (4.9) that is bounded below; e.g., if 0
then

(4.11)
[s 1]2+ + [1 s]2+ C C
s R.
(s) 16
Finally, we need a further restriction on the mesh in order to prove well posedness
e h,t ). We modify our assumption (A) to
of (P
In addition to the assumption (A), we assume for all h > 0 that T h is an acute
(A)
partitioning; that is, for (i) d = 2, the angle of any triangle does not exceed

2 , and (ii) d = 3, the angle between any two faces of the same tetrahedron
does not exceed 2 .
This acuteness assumption yields that
Z
(4.12)
i j dx 0, i 6= j, T h .

e that for all 1 and for all T h

In addition it follows from (4.7) and (A)
Z
Z
0
00
0
| h [1,
()]|2 dx 1,
(sup |(x)|) h [1,
()] dx
x

Z
0
(4.13)
h [1,
()] dx
Sh;

see, e.g., [14].

e holds and let U 0 K h
Theorem
4.1. Let and T h be such that assumption (A)
R
0
2
with U (1, 1). Then for all t > 0 such that c bmax t < 4, there exists a
e h,t ). Moreover, {U n }N is unique and the stability
solution {U n , W n }N
n=1 to (P
n=1
bounds (2.26) hold. In addition W n is unique on (U n1 ), n = 1 N .
Proof. Given U n1 K h with |U n1 |1 C, we prove existence of {U n , W n }
solving (4.2ab) by introducing a regularized version, as follows.
Find {Un , Wn } S h S h such that
 n n1 h


U U
(4.14a)
, + b(U n1 )Wn , = 0 S h ,
t
(4.14b) (Un , ) + (01, (Un ), )h = (Wn 20 (U n1 ), )h

Ve h (U n1 ).

Similarly to (2.28) we have that

(4.15)

Wn GUh n1

Un U n1
t

nj, j +

jJ0 (U n1 )

M
X

nm, m (U n1 ).

m=1

Existence of {Un , Wn }, uniqueness of Un , and uniqueness of Wn on m (U n1 ),

m = 1 M , follows as for {U n , W n } in the proof of Theorem 2.1 under the stated
time step restriction. Similarly to (2.33), on noting the convexity of 1, , the concavity
of e2 , and the assumptions on U n1 , we have that Un U n1 Ve h (U n1 ) is such
that

n 2
2 |U |1

(4.16)

+ 8 |Un U n1 |21 + ( (Un ), 1)h + 23 t|[b(U n1 )] 2 Wn |20

1

( (U n1 ), 1)h + 2 |U n1 |21 C.

THE CAHNHILLIARD EQUATION WITH DEGENERATE MOBILITY

307

From (4.16) and (4.11) we deduce that

(4.17)

|[Un 1]+ |h + |[Un 1]+ |h C.

Hence it follows from (4.17), (2.8), and (2.4) that

(4.18a)

k[Un 1]+ k0, + k[Un 1]+ k0, Ch 2 2

and from (4.4) and (4.18a) that

(4.18b)

k1, (Un )k0, max{ 1, 1, (kUn k0, ) } C(h1 ).

The next part of the proof is now concerned with establishing an independent
bound on |01, (Un )m (U n1 )|h . Due to the logarithmic term in 1 this then implies
that the 0 limits of subsequences of Un are less than one in magnitude on the
set J+ (U n1 ). Using a Poincare type inequality on m (U n1 ) it follows similarly to
(2.24a) by noting (2.22b) and (4.16) that
R
|[(I m (U n1 ) )Wn ] m (U n1 )|2h
Z
1
C(h )
|Wn |2 dx
m (U n1 )
Z
b(U n1 )|Wn |2 dx
C(h1 ) [bmin (U n1 )]1
(4.19)

n1

m (U n1 )
1

)] (t) .
C(h ) [bmin (U
R
We now bound m (U n1 ) Wn . Choosing m (U n1 ) in (4.14b) yields that
(4.20)

(Wn , m (U n1 ))h = (01, (Un ), m (U n1 ))h + ,m ,

where ,m := (Un , [m (U n1 )]) + (20 (U n1 ), m (U n1 ))h . It is convenient to

introduce the set m (U n1 ) := supp{m (U n1 )} \ m (U n1 ), and noting (4.16) we
have that
Z





Un [m (U n1 )] dx
|(Un , [m (U n1 )])| = 
 m (U n1 )

(4.21)

|m (U n1 )| 2 k[m (U n1 )]k0, |Un |1 Ch1 .

Noting that U n1 K h , 20 C([1, 1]), and (4.21), it follows that

|,m | C[1 + h1 ].

(4.22)

Similarly to (4.21), we have on noting (2.4), (1.9), and (4.16) that

R
(Un , [ h ([(I m (U n1 ) )Un ] m (U n1 ))])
Z
Z
R
n 2
|U | dx +
Un [ h ([(I m (U n1 ) )Un ] m (U n1 ))] dx
=
m (U n1 )

m (U n1 )
1

|Un |21 + C|Un |1 |m (U n1 )| 2 k[m (U n1 )]k0, kUn k0,

C[1 + h1 ]kUn k21 C(h1 ).
(4.23)

308

JOHN W. BARRETT, JAMES F. BLOWEY, AND HARALD GARCKE

R
Now choosing h ([(I m (U n1 ) )Un ] m (U n1 )) in (4.14b) it follows on
noting (4.6), 2 C 1 ([1, 1]), U n1 K h , (4.23), (4.17), the remark after (4.4),
(4.8), and (4.19) that for all [1, 1]
R
(01, (Un ) m (U n1 ), m (U n1 ) Un )h
= (01, (Un ) m (U n1 ), Un )h
R
+(Wn 20 (U n1 ), [(I m (U n1 ) )Un ] m (U n1 ))h
R
(Un , [ h ([(I m (U n1 ) )Un ] m (U n1 ))])
(1, () 1, (Un ), m (U n1 ))h + 2c |Un |2h
R
+|[(I m (U n1 ) )Wn ] m (U n1 )|2h + C[ 1 + |Un |2h + C(h1 ) ]
(4.24)

C([bmin (U n1 )]1 , h1 , (t)1 ).

Here and below we have used the notation C(a1 , . .R. , aI ) defined at the
R end of section 1.
Hence choosing = 1 in (4.24) and noting that m (U n1 ) Un = m (U n1 ) U n1
(1, 1), we deduce that
|(01, (Un ), m (U n1 ))h | C([bmin (U n1 )]1 , h1 , (t)1 ).
R
We note that the constant on the right-hand side depends on m (U n1 ) U n1 but is
independent of . Furthermore, combining (4.20), (4.25), and (4.22) it follows that
(4.25)

|(Wn , m (U n1 ))h | C([bmin (U n1 )]1 , h1 , (t)1 ).

(4.26)

Hence combining (4.19) and (4.26) yields that

|Wn m (U n1 )|h C([bmin (U n1 )]1 , h1 , (t)1 ).

(4.27)

Choosing h [01, (Un ) m (U n1 )] Ve h (U n1 ) in (4.14b) and rearranging

yields that
Z
0
Un h [1,
(Un )] dx + |01, (Un ) m (U n1 )|2h

m (U n1 )
Z
0
=
Un h [1,
(Un ) m (U n1 )] dx + c (Un , h [Un m (U n1 )])
m (U n1 )

+ (Wn 20 (U n1 ), 01, (Un ) m (U n1 ))h .

(4.28)

We now estimate the terms on the right-hand side of (4.28). For any simplex
d+1
xj }j=1
{xj }jJ and corresponding basis functions {e
j }d+1
m (U n1 ) with vertices {e
j=1
{j }jJ , we have on noting (4.12), (4.6), and (4.18b) that
Z
0
(Un ) m (U n1 )] dx
Un . h [1,

d+1
X

0
Un (e
xi ) 1,
(Un (e
xj )) m (U n1 )(e
xj )

i,j=1

d+1
X
i,j=1

[Un (e
xi )

e
i e
j dx

0
Un (e
xj )] 1,
(Un (e
xj )) m (U n1 )(e
xj )

e
i e
j dx

THE CAHNHILLIARD EQUATION WITH DEGENERATE MOBILITY

d+1
X

[1, (Un (e
xi )) 1, (Un (e
xj ))] m (U n1 )(e
xj )

309

i,j=1

e
i e
j dx

(4.29) C(h1 ).
Hence, by summing (4.29) over all m (U n1 ), we deduce that
Z
0
(4.30)
Un h [1,
(Un ) m (U n1 )] dx C(h1 ).

m (U n1 )

Similarly to (4.23), we have that

(4.31)

(Un , h [Un m (U n1 )]) C(h1 ).

It follows from (4.28), on noting (4.13), (4.30), (4.31), a Youngs inequality, (4.27),
U n1 K h , and 20 C([1, 1]), that
(4.32)

|01, (Un ) m (U n1 )|h C([bmin (U n1 )]1 , h1 , (t)1 ).

It follows from (4.16), the fact that (Un , 1)h = (U n1 , 1)h , (1.9), and (4.17) that
there exists U n K h and a subsequence {Un0 } such that Un0 U n as 0 0. As
Un U n1 Ve h (U n1 ), it follows that U n U n1 Ve h (U n1 ). It follows from
(4.32) and the above that there exists n S h and a subsequence {Un0 } such that
h [01,0 (Un0 )] n c U n on (U n1 ) as 0 0. Since 01,0 (Un0 (xj )) is uniformly
0
bounded in , noting (4.7) and using that for all s R [1,
]1 (s) [10 ]1 (s) as
0, we have that U n (xj ) = [10 ]1 (n (xj )) and therefore n (xj ) = 10 (U n (xj )) for
all j J+ (U n1 ). Hence we have that
(4.33)

|01 (U n ) m (U n1 )|2h C([bmin (U n1 )]1 , h1 , (t)1 ),

which immediately implies that |U n (xj )| < 1 for all j J+ (U n1 ). Finally, it follows
from (4.27) that there exists W n S h and a subsequence {Wn0 } such that Wn0 W n
on (U n1 ) as 0 0. Hence we may pass to the limit 0 0 in (4.14a,b), on noting
e h,t ).
(2.21), to prove existence of a solution {U n , W n }N
n=1 to (P
The uniqueness result follows as in Theorem 2.1 on noting that |U n (xj )| < 1 for
all j J+ (U n1 ). The stability bounds (2.26) follow as in Theorem 2.1 by choosing
W n S h in (4.2a) and U n U n1 Ve h (U n1 ) in (4.2b).
e h,t ), we have
Adopting the notation (2.35a,b) for the solution {U n , W n }N
n=1 of (P
the analogue of Theorem 2.2.
Theorem 4.2. Let the assumptions of Theorem 2.2 hold with (A) replaced by
e
(A), and now in particular with 1 assumed to be of the logarithmic form (4.1). Then
e h,t ) and a function
there exists a subsequence of solutions {U, W }h of problem (P
1 1

2 8
u L (0, T ; K) H 1 (0, T ; (H 1 ())0 ) Cx,t
(T ) and a w L2loc ({|u| < 1}) with
w
2
x Lloc ({|u| < 1}) such that as h 0, (2.38)(2.40) and (2.41a,b) hold.
Proof. The proof is the same as that of Theorem 2.2 with the following minor
changes. We mention only the modifications caused by the presence of the logarithmic
free energy which implies that 0 becomes unbounded. Clearly the inequality, the test
function U + , and K h in (2.37b) are replaced by equality, , and Ve h , respectively.
Although (2.56) is redundant, (2.57) still holds on noting the above and (2.53). It
follows from (2.59) and (2.53) that Ck W + kL2 (T ) on the right-hand side of the

310

JOHN W. BARRETT, JAMES F. BLOWEY, AND HARALD GARCKE

first inequality in (2.60) is replaced by C( 1 )k W + kL2 (T ) with the final bound of

(2.60) remaining the same. Clearly (2.68) remains true for all L2 (0, T ; H 1 ())
with supp() D+ on noting the technique used in (2.67). Hence (2.70) remains
true.
Finally, we modify the iterative algorithm in section 3 to solve the nonlinear
e h,t ). We have that {U n , W n } S h S h
algebraic system for {U n , W n } arising in (P
satisfy
(U n + 10 (U n ), )h = (Z n , )h Ve h (U n1 ),
U n (xj ) = U n1 (xj ) j J0 (U n1 )

(4.34)

in place of (3.1a) with (3.1bd) remaining the same. Hence we modify our iterative
1
procedure (3.2ad) by replacing (3.2b) by the following: Find U n,k+ 2 S h such that
1

U n,k+ 2 (xj ) = U n1 (xj )

(4.35)

n,k+ 12

(xj ) + 10 (U

n,k+ 12

(xj )) = Z n,k (xj )

if j J0 (U n1 ),
if j J+ (U n1 )

and keeping (3.2a,c,d) the same. For j J+ (U n1 ) existence and uniqueness of

1
U n,k+ 2 (xj ) follows from the monotonicity of 10 (). Hence this modified iterative
procedure is well defined.
Theorem 4.3. Let 3c2 bn1 t < 4. Then for all R+ and {U n,0 , W n,0 }
h
S S h the sequence {U n,k , W n,k }k0 generated by the modified algorithm, (3.2ad)
with (3.2b) replaced by (4.35), satisfies
Z
n,k
n,k+ 12
n
U
and
b(U n1 )|(W n,k+1 W n )|2 dx 0 as k .
U , U

(4.36)

Proof. The proof is just a simple modification of the proof of Theorem 3.1 to
take into account the changes (4.34) and (4.35) to (3.1a) and (3.2b), respectively. We
introduce the following modification of the discrete semi-inner product, (2.1):
(4.37)

(1 , 2 )hJ+ (U n1 ) :=

X
jJ+

j 1 (xj ) 2 (xj )

1 , 2 C((U n1 )).

(U n1 )

The only changes to the proof of Theorem 3.1 are the following: ( , )h on the left-hand
sides of (3.9), (3.10), and (3.13) is replaced by ( , )hJ+ (U n1 ) . The right-hand side of
1

(3.9) remains the same as U n,k+ 2 (xj ) = U n1 (xj ) = U n (xj ) for all j J0 (U n1 ).
Hence we obtain the desired convergence (4.36).

5. Numerical experiments. In this section we report on some numerical results with the intention of demonstrating the practicability of our method as well as
showing that in the case of a degenerate mobility a quite different qualitative behavior
is observed when compared to results obtained with constant mobility.
In order to avoid numerical difficulties we introduced approximative analogues of
the sets Im (q h ) denoted by Im (q h ), which were defined by replacing (c) in (2.18) by
(c) b(q h ) > tol1 := 106 at a vertex of l , l = 1 L. In addition, for each n we
adopted the following stopping criterion for (3.2ad): If kU n,k? U n,k? 1 k0, < tol
with tol = 107 then we set {U n , W n } {U

n,k?

, W n,k? }, where U

n,k?

K h was

THE CAHNHILLIARD EQUATION WITH DEGENERATE MOBILITY

defined by
(5.1)

n,k?
Uj

(
:=

311

n1

),
max{1, min{Ujn,k? , 1}} if j J+ (U n1 ) := M
m=1 Im (U
n1
n1
n1
if j J0 (U
) := J \ J+ (U
),
Uj
1

if 10 was not strictly monotone and {U n , W n } {U n,k? 2 , W n,k? } otherwise. Finally, we chose, from experimental evidence, the relaxation parameter h in
(3.2ad) in order to improve its convergence.
All computations were performed in double precision on a Sparc 20. The program
was written in Fortran 77 using the NAG subroutine C06HBF for calculating the
discrete cosine transform used in solving (3.5).
5.1. One space dimension. The computations were performed on a uniform
partitioning of = (0, 1) with mesh points xj = (j 1)h, j = 1 #J, where
h = 1/(#J 1). We note that the integral on the right-hand side of (3.3) can be
evaluated exactly using Simpsons rule if b() is quadratic.
Experiment 1. One characteristic feature of the discretizations (Ph,t ) and
e h,t ) is that
(P
U n1 (xj1 ) = U n1 (xj ) = U n1 (xj+1 ) = 1 = j J0 (U n1 ) and U n (xj ) = 1,
(5.2)
so that the free boundaries {|U n | = 1} can advance at most one mesh point locally
from one time level to the next. This implies that over a time interval of length T
h
T . To be able to track a
the free boundary can advance by at most a distance of t
free boundary which moves with a finite but a priori unknown speed, one needs to
h
. For a nonuniform mesh the above requirement
choose t and h such that t
min

T
on h and t has to be replaced by
. If we choose the time step too
t
h
large, e.g., if t 0, we obtain the existence of a solution in the limit as h, t 0
which would not spread at all for all initial data u0 K. Similar results hold for the
degenerate equation

ut + (up uxxx )x = 0

in T ,

ux = up uxxx = 0

on (0, T );

(cf. Lemma 5.1 in [7] and [6]).

As data we took = 0.01, 1 2 0, and c = 1, i.e., the deep quench limit,
b(u) := 1 u2 , and
(
 1

x 2

1

cos

1
if
|x

2
2 ,
u0 (x) =
1
otherwise.
We note that u0 6 C 1 ([0, 1]) may be considered to be a stationary solution of (P) on
2
noting (2.41a,b), since xu20 u0 = 1 for |u0 | < 1. We performed two separate sets
1
of experiments, one with t = 40.96h2 and h = 26l , the other with t = 0.08h 2
and h = 262l , both for l = 0, 1, 2, 3, and 4. In both cases we took bmax = bn1 = 1.
Note that the time step restriction of Theorem 3.1, and hence Theorem 2.1, holds. We
see in Figure 5.1 when t = 40.96h2 that our numerical solution in the limit h 0
appears to spread to the stationary C 1 ([0, 1]) solution:
( h
 1 i

x 2
1

1
+
cos
1 if |x 12 | ,

1
otherwise.

312

JOHN W. BARRETT, JAMES F. BLOWEY, AND HARALD GARCKE

0
h u0
#J =
65
129
257
513
1025

14
12
34

1
4

1
2

3
4

0
h u0
#J =
65
257
1025
4097
16385

14
12
34

1
4

1
2

3
4

1
1

Fig. 5.1. U (x, 0.2) plotted against x with t = 40.96h2 and t = 0.08h 2 for several values of h.
1

In contrast, when t = 0.08h 2 the solutions with #J = 65 and 256 are more or less
identical to the above. However, for h sufficiently small the region {|u(, t)| = 1}
cannot advance sufficiently fast to capture the apparent former solution. It certainly
appears that the limit h 0 yields that u0 is a stationary solution. This numerical
experiment also appears to indicate that as posed (P) does not have a unique solution.
In this context we refer to [19], where existence of a solution is proved with the
property that u L2 (0, T ; H 2 ()) for arbitrary initial data u0 H 1 (). Since
/ H 2 () this means that u0 as initial data would lead to a nonstationary solution.
u0
We conjecture that we compute the solutions constructed in [19] if we take our time
step small enough.
Experiment 2. In the second experiment we took 2 0 and c = 1 as in the
previous experiment, but we varied 1 . For the initial data we took

1
1
if 0 x 13 20
,

1
1
20( 1 x)
if |x 3 | 20 ,
3
u0 (x) =
41
1

20|x 41

50 | if |x 50 | 20 ,

1
otherwise,
with = 103 , h = 0.005, and t = 10h2 .

THE CAHNHILLIARD EQUATION WITH DEGENERATE MOBILITY

313

U (., 0)
.
U ( , 0.002)

1
2

0
1
2

1
4

1
2

3
4

U (., 0)
U (., 0.001)
U (., 0.002)
U (., 0.003)

1
2

0
1
2

1
4

1
2

3
4

U (., 0)
U (., 0.001)
U (., 0.002)
U (., 0.009)

1
2

0
1
2

1
4

1
2

3
4

U (., 0)
U (., 0.25)
U (., 0.75)
U (., 5)

0.5
0
0.5
1

1
4

1
2

3
4

Fig. 5.2. U (x, t) plotted against x at different times, where is given by (1.2) and (1.1) with
= 0.3 for constant and degenerate mobility.

314

JOHN W. BARRETT, JAMES F. BLOWEY, AND HARALD GARCKE

In Figure 5.2 the graphs are arranged as follows: the first two rows are for as
given in the previous experiment
limit
 (the
 deep quench
 1u
 (1.2)), and the last two rows
3
(1 + u) ln 1+u
+
(1

u)
ln
; the second and third rows
are for 1 (u) = 20
2
2
are for b(u) 1 (constant mobility) and the first and fourth are for b(u) := 1 u2
(degenerate mobility). In all cases we took bmax = bn1 = 1 and note that the time
step restriction of Theorem 3.1, and hence Theorem 2.1, holds. We make the following
remarks:
The algorithm (3.2ad) with bmax b 1 is precisely that described in [2]
to solve (Ph,t ) for n fixed and constant mobility.
To ensure that our computations were not dependent on h we repeated the
experiment with h = 0.0025 and obtained graphically indistinguishable pictures.
For constant mobility, regardless of which we take the second bump gets
drawn out to the left rather quickly. This is due to the fact that the mobility
is positive in the pure phases, i.e., at points where u is close to the minima
of .
With b(u) := 1 u2 and given by the double obstacle potential (1.2)
the second bump does not lose mass. However for the logarithmic , we
observe diffusion through the bulk although the time scale is greatly increased;
see [11]. As in the case of the constant mobility the final profile is given by one
transition layer. We remark that the minima for the logarithmic potential
are less than one in magnitude. For converging to zero the minima converge
to 1. This implies that the diffusion through the bulk becomes smaller and
smaller at low temperatures. Also we note that |U | < 1 in the last two rows
of Figure 5.2 (the case of the logarithmic potential).
5.2. Two space dimensions.
Experiment 3. We performed two numerical experiments in two spatial dimensions with = (0, 1) (0, 1). In the first experiment we took degenerate mobility,
b(u) := 1 u2 . In the second experiment we took exactly the same data, but now
with constant mobility, b(u) 1.
We took a uniform mesh consisting of squares e of length h = 1/256, each of
which was then subdivided into two triangles by its northeast diagonal. We used the
following discrete semi-inner product on C(),
(5.3)

(1 , 2 )h? :=

h (1 (x)2 (x)) dx,

in place of (2.1). Here h is the piecewise continuous bilinear interpolant on which

is affine linear for x1 (or x2 ) fixed and interpolates at the vertices on each square e.
Using (5.3) instead of (2.1) only changes the algorithm at the corners of the square
and has the advantage that one can then solve (3.5) using the discrete cosine
transform; see [9]. We note that similarly to (2.8), the induced norm from (5.3) on
S h is equivalent to the standard L2 norm. Therefore it is easy to adapt the proofs
to show that Theorems 2.1, 3.1, and 4.1 in this paper remain true with this choice of
discrete semi-inner product.
We took to be the deep quench limit (1.2) with the splitting 1 (u) 0, 2 (u) :=
1
(1
u2 ), and c = 0 (this allows us to take an arbitrarily large time step), =
2
3.2 104 , t = 1.6 103 and we relaxed our stopping criterion to be tol = 106 .
Once again we took bn1 = 1.

THE CAHNHILLIARD EQUATION WITH DEGENERATE MOBILITY

315

Fig. 5.3. U (, t) plotted for t = 0.04, 0.08, 0.12, 0.20, 0.24, 0.44, 0.48, and 2.76 when b(u) :=
1 u2 and is given by (1.2).

316

JOHN W. BARRETT, JAMES F. BLOWEY, AND HARALD GARCKE

Fig. 5.4. U (, t) plotted for t = 0, 0.04, 0.08, 0.16, 0.24, and 0.64 when b(u) 1 and is given
by (1.2).

For the above choices of b, An,k satisfying (3.3) can be evaluated exactly by
sampling at the midpoints of the sides over each triangle . The initial data were
taken to be U 0 = 0.4 h , where h S h with k h k0, 0.05. In Figures 5.3 and
5.4 we plot a grey scale grid plot of U (, t) at several times. The pictures are arranged
in a matrix format with time increasing to the right in rows then down columns. The
grey scale ranges from 0.9 to 0.9 in steps of 0.2 with pure black/white representing
values larger/smaller than 0.9/0.9. We note that there are approximately 10 mesh

THE CAHNHILLIARD EQUATION WITH DEGENERATE MOBILITY

317

points across each interface. The final numerical solution plotted in Figure 5.3 is a
stationary numerical solution, that is, the stopping criterion for the iterative procedure
is satisfied in a single step from one time level to the next. However, the final picture
in Figure 5.4 is not stationary.
In Figure 5.3, the case of degenerate mobility, after the early stages there is very
little interaction of regions which do not intersect and the evolution takes place locally
where the local mass is preserved. The final frame yields a numerical stationary solution consisting of many circles which do not intersect; this corresponds to a pinning
effect reported in [23] for spinodal decomposition of polymer mixtures. In Figure 5.4,
the case of constant mobility, we start with U 0 () U (, 0.04) from the first experiment. In contrast, there is evolution and growth of regions which do not intersect
(see Figure 5.4); moreover, circles which coexist are not stationary since there is a
coupling through bulk terms.
REFERENCES
[1] R.A. Adams and J. Fournier, Cone conditions and properties of Sobolev spaces, J. Math.
Anal. Appl., 61 (1977), pp. 713734.
[2] J.W. Barrett and J.F. Blowey, Finite element approximation of a model for phase separation
of a multi-component alloy with non-smooth free energy, Numer. Math., 77 (1997), pp. 134.
[3] J.W. Barrett and J.F. Blowey, Finite element approximation of a model for phase separation
of a multi-component alloy with a concentration dependent mobility matrix , IMA J. Numer.
Anal., 18 (1998), pp. 287328.
[4] J.W. Barrett and J.F. Blowey, Finite element approximation of the Cahn-Hilliard equation
with concentration dependent mobility, Math. Comp., 68 (1999), pp. 487517.
[5] J.W. Barrett and J.F. Blowey, Finite element approximation of a model for phase separation of a multi-component alloy with nonsmooth free energy and a concentration dependent
mobility matrix , Math. Models Methods Appl. Sci., 9 (1999), pp. 627663.
[6] J.W. Barrett, J.F. Blowey, and H. Garcke, Finite element approximation of a fourth order
nonlinear degenerate parabolic equation, Numer. Math., 80 (1998), pp. 525556.
[7] E. Beretta, M. Bertsch, and R. Dal Passo, Nonnegative solutions of a fourth-order nonlinear
degenerate parabolic equation, Arch. Rational Mech. Anal., 129 (1995), pp. 175200.
[8] F. Bernis and A. Friedman, Higher order nonlinear degenerate parabolic equations, J. Differential Equations, 83 (1990), pp. 179206.
[9] J.F. Blowey and C.M. Elliott, The Cahn-Hilliard gradient theory for phase separation with
non-smooth free energy II. Numerical analysis, European J. Appl. Math., 3 (1992), pp. 147
179.
[10] J.W. Cahn, On spinodal decomposition, Acta Metall., 9 (1961), pp. 795801.
[11] J.W. Cahn, C.M. Elliott, and A. Novick-Cohen, The Cahn-Hilliard equation with a concentration dependent mobility: Motion by minus the Laplacian of the mean curvature, European
J. Appl. Math., 7 (1996), pp. 287301.
[12] J.W. Cahn and J.E. Hilliard, Free energy of a non-uniform system. I. Interfacial free energy,
J. Chem. Phys., 28 (1958), pp. 258267.
[13] J.W. Cahn and J.E. Taylor, Surface motion by surface diffusion, Acta Metall., 42 (1994),
pp. 10451063.
[14] J.F. Cialvaldini, Analyse num
erique dun probl`
eme de Stefan a
` deux phases par une m
ethode
delements finis, SIAM J. Numer. Anal., 12 (1975), pp. 464487.
[15] P.G. Ciarlet, Introduction to Numerical Linear Algebra and Optimisation, Cambridge University Press, Cambridge, 1988.
[16] M.I.M. Copetti and C.M. Elliott, Numerical analysis of the Cahn-Hilliard equation with
logarithmic free energy, Numer. Math., 63 (1992), pp. 3965.
[17] C.M. Elliott, The Cahn-Hilliard model for the kinetics of phase transitions, in Mathematical
Models for Phase Change Problems, J.F. Rodrigues, ed., Internat. Ser. Numer. Math., Vol.
88, Birkh
auser-Verlag, Basel, 1989, pp. 3573.
[18] C.M. Elliott, D.A. French, and F.A. Milner, A second order splitting method for the CahnHilliard equation, Numer. Math., 54 (1989), pp. 575590.
[19] C.M. Elliott and H. Garcke, On the Cahn-Hilliard equation with degenerate mobility, SIAM
J. Math. Anal., 27 (1996), pp. 404423.

318

JOHN W. BARRETT, JAMES F. BLOWEY, AND HARALD GARCKE

[20] C.M. Elliott and H. Garcke, Diffusional phase transitions in multicomponent systems with
a concentration dependent mobility matrix , Phys. D, 109 (1997), pp. 242256.
[21] H. Garcke and A. Novick-Cohen, A singular limit for a system of degenerate Cahn-Hilliard
equations, Adv. Differential Equations, to appear.
n, Degenerate parabolic differential equations of fourth order and a plasticity model with
[22] G. Gru
nonlocal hardening, Z. Anal. Anwendungen, 14 (1995), pp. 541574.
[23] T. Hashimoto, M. Takenaka, and T. Izumitani, Spontaneous pinning of domain growth
during spinodal decomposition of off-critical polymer mixtures, J. Chem. Phys., 97 (1992),
pp. 679689.
[24] J.E. Hilliard, Spinodal decomposition, in Phase Transformations, H.I. Aaronson, ed., American
Society for Metals, Cleveland, OH, 1970, pp. 497560.
[25] P.-L. Lions and B. Mercier, Splitting algorithms for the sum of two nonlinear operators,
SIAM J. Numer. Anal., 16 (1979), pp. 964979.
[26] A. Novick-Cohen, The Cahn-Hilliard equation: Mathematical and modelling perspectives,
Adv. Math. Sci. Appl., 8 (1998), pp. 965985.
[27] J. Yin, On the existence of nonnegative continuous solutions of the Cahn-Hilliard equation, J.
Differential Equations, 97 (1992), pp. 310327.