Sie sind auf Seite 1von 53

Galerkin Methods for Parabolic Equations

Author(s): Jim Douglas, Jr. and Todd Dupont


Reviewed work(s):
Source: SIAM Journal on Numerical Analysis, Vol. 7, No. 4 (Dec., 1970), pp. 575-626
Published by: Society for Industrial and Applied Mathematics
Stable URL: http://www.jstor.org/stable/2949379 .
Accessed: 14/12/2012 11:34
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Society for Industrial and Applied Mathematics is collaborating with JSTOR to digitize, preserve and extend
access to SIAM Journal on Numerical Analysis.

http://www.jstor.org

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

SIAM J. NUMER. ANAL.


Vol. 7, No. 4, December 1970

GALERKIN METHODS FOR PARABOLIC EQUATIONS*


JIM DOUGLAS, JR.,

TODD DUPONTt

AND

This paper is dedicated to Alston S. Householder.


Abstract. Galerkin-type methods, both continuous and discrete in time, are considered for
approximating solutions of linear and nonlinear parabolic problems. Bounds reducing the estimation
of the error to questions in approximation theory are derived for the several methods studied. These
methods include procedures that lead to linear algebraic equations even for strongly nonlinear problems.
A number of computational questions related to these procedures are also discussed.
CONTENTS
1.
2.
3.
4.
5.
6.
7.
8.
9.
10
11.
12.

Introduction
Formulation of the variational problem and Galerkin approximations
A priori estimates for the continuous time Galerkin approximation
A priori estimate for the Crank-Nicolson-Galerkin approximation
A priori estimate for predictor-corrector version of Crank-Nicolson-Galerkin approximation
A priori estimate for Crank-Nicolson extrapolation approximation
Extensions to more general second order parabolic equations
Other boundary conditions
Change of basis
Parabolic systems
Some computational considerations
Application of approximation theory

575
578
580
589
593
597
600
609
612
613
618
623

1. Introduction. This paper deals with some of the theoretical and computational questions which are connected with the approximation of solutions of
linear and nonlinear parabolic equations by Galerkin-type procedures. Several
Galerkin-type procedures will be formulated and in each case an a priori estimate
will be derived which can be used to bound the error.
In ?? 2 through 6 we shall be considering Galerkin procedures for approximating solutions of
Lu=

(1.1)

auX,t)

at

n~ a I

au

aj,j(x,u)a)

=0,

u(x,t) = O,

xeQf,

x Ean,

u(x, 0) = uO(x),

t>0,
t > O,
x E Q,

where Q is a bounded domain in Rn and the matrix a(x, u) = (ai,j(x,u)) is positive


definite. In particular, in ? 2 we shall first reformulate (1.1) as a variational problem
in the space variables, x. We shall then define a Galerkin approximation U(x, t)
to the solution of (1.1) by requiring that U lie in a finite-dimensional space of
functions and be such that LU is orthogonal to this space for each t. This first
type of Galerkin approximation is determined by a system of ordinary differential
equations. Also in ? 2 we shall formulate several procedures which define the
* Received by the editors June 23, 1969, and in revised form June 22, 1970.
t Department of Mathematics, University of Chicago, Chicago, Illinois 60637.

575

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

576

JIM DOUGLAS, JR., AND TODD DUPONT

, tM. The first of these is analoapproximate solution at discrete times t = to


gous to the Crank-Nicolson finite difference equation. In order to avoid having
to solve nonlinear algebraic systems at each time step we shall formulate several
procedures which can be used to approximate the solution to the nonlinear
problem (1.1), but which only require the solution of linear algebraic equations.
These procedures are of two types: two level predictor-corrector or three level
extrapolation. In ? 3 we shall derive two types of error bounds for the Galerkin
approximation. The first error bound is a modification of the familiar energy
inequality for parabolic equations. This bound can be used to show that under
reasonable hypotheses the Galerkin approximation and its first order space
derivatives tend to the solution of (1.1) and its first order space derivatives with
the error bounded by a multiple of h2'-1, where the approximate solution is
required to lie in a Hermite space of piecewise polynomials of degree 2m - 1
based on a rectangular grid of mesh size h [4]. The second type of estimate in ? 3
bounds a fractional derivative of the error; those readers not familiar with Fourier
transforms of distributions will want to skip Theorems 3.2, 3.3 and Lemma 3.1.
In ? 4 we derive an error bound for the Crank-Nicolson-Galerkin approximation.
In ?? 5 and 6 we derive estimates for the error in the predictor-corrector and
extrapolation procedures, respectively. The estimates in ?? 4, 5 and 6 can be used
to show that, if the approximate solution lies in the space mentioned above, the
error and its first order x-derivatives are O(h2m- + (At)2), where At = ti 1-ti .
In ? 7 we extend the results of ?? 2 through 6 to more general second order
parabolic equations of the form
-

au(x, t)

at )-

(1.2)

La ,- Ai(x, t, u,V.u)

= f(x, t,u, Vxu), x E Q, t > 0,

u(x,t) = 0,
u(x, 0) = uO(x),

xe 'Q, t > 0,
x E Q.

In (1.2), Vxuis used to denote the vector of first order x-derivatives of u. In order to
extend the analysis of ?? 3 and 4 we require only mild regularity hypotheses onf
and the Ai's. We need to be more restrictive on the form off and the Ai's in order
to generalize the results of ?? 5 and 6. In ? 8 we consider boundary conditions for
t > 0 which are of a more general form than those of (1.1) and (1.2). In ? 9 we
indicate bounds for the errors which result if the finite-dimensional space in which
, Tp.In ? 10 we extend
the solution is required to lie is changed at times t = T1,
some of these results to systems of second order nonlinear parabolic equations.
We consider in ? 11 many of the computational questions connected with computing the approximations discussed above. In ? 12 we shall make some direct
applications of approximation theory to the results of the previous sections.
In practice finite difference methods have had much greater popularity than
variational methods for approximating solutions of parabolic problems. It is not
clear that this is a reasonable state of affairs. Galerkin procedures are somewhat
more difficult to program for a computer than difference procedures; however,
Galerkin procedures have advantages which in many situations should far outweigh this minor drawback. The strongest point in favor of Galerkin-type methods
for parabolic problems is that in many instances they yield answers which are

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

GALERKIN

METHODS FOR PARABOLIC

EQUATIONS

577

considerably more accurate than those which we can get with comparable computing effort using difference methods. Some indication of this is given by the
convergence results cited above which indicate that, if for example we use Hermite
cubics, the error goes to zero as h3. Another indication of this improved accuracy
can be found in reported empirical results [11], [12], [22]. Another advantage of
Galerkin-type procedures is the form in which the results are presented. In most
cases the solution will be a smooth function which is piecewise polynomial. This
means that it is easy to get smooth graphical representations of the results. It also
allows integrals and derivatives of the results to be computed by formula.
Variational methods have a long history and the reader is referred to [19],
[20] for discussions of the use of these methods for approximating the solutions
of elliptic boundary value and eigenvalue problems. These two references also
have extensive bibliographies. More recently these techniques have been applied
by Price, Varga, et al. [4], [6], [7], [8], [9], [22], [23] to the above problems and
some linear parabolic problems. Swartz and Wendroff [24] have given a treatment
of certain time-dependent equations by Galerkin methods. Their results are derived
in a neat, simple fashion and apply most naturally to first order equations. A particularly elegant treatment of Galerkin procedures for elliptic equations has been
given by Aubin [2]. Much work has been done on special Galerkin-type procedures
which use a triangular grid and piecewise linear functions on this grid. Such
procedures are called finite element methods; see the work of Zlamal [26] for an
example of recent work with these methods and for references. We have used an
analysis similar to that in ?? 3 and 4 to demonstrate rates of convergence of two
Galerkin approximations to the solutions of a particular degenerate nonlinear
parabolic system which arises from petroleum engineering [10]. Together with
H. H. Rachford, Jr., we have applied these techniques to approximate solutions
of petroleum engineering problems. Some of these problems are quite difficult
to solve using finite differencetechniques but are solved rather easily with variational methods. Harold Meyer [18] has extended some of the results of ?? 2 through 7
to higher order parabolic equations. It is interesting to note that Galerkin methods
have also been used to demonstrate existence theorems for nonlinear parabolic
equations; see [5], [27].
Although this paper does not deal with finite difference techniques for approximating solutions of parabolic problems, it will be clear to the reader familiar
with these procedures that the ideas used here can be applied to differencemethods
with only minor changes. For example, the process given by (11.11) can be used
with difference methods and the analysis given in ??4, 5 and 6 can be used to give
convergence theorems not previously known for difference approximations of
solutions of (1.1).
We want to point out the strong influence that the general point of view of the
work of Lions and Aubin [2], [3], [16] has had on this work. Although our results
differ from theirs quite strongly in detail, we have found many of their ideas quite
useful.
All functions we shall be considering here, with the sole exception of the
Fourier transforms in ? 3, will be real-valued. We have used C as a generic constant which is not necessarily the same at each occurrence. In the same spirit we
have used E as a generic symbol for small positive constants.

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

578

JIM DOUGLAS, JR., AND TODD DUPONT

2. Formulation of the variational problem and Galerkin approximations.


Consider the boundary value problem
n

au(x,t)

at

au(x, t)
u)
ai,j(x,
a-ax

i,E

0,
(x, t) E-Q x (O, T] ,

u(x, 0) = u0(x),

(2.1)

x E Q,

u(x, t) = O,

(x, t) c- M x (O, T],

where Q is a bounded domain in Rn and the n x n matrix (ai,j(x, u)) is symmetric


and uniformly positive definite. If u(x, t) and a j(x, u(x, t)) have sufficient regularity,
then (2.1) is equivalent to

au'v) + a(u; u, v) = O,
au,
Ku(,O),v>

(2.2)

t > O, v c-Ho(Q)

Ku0,v>,

veHo(Q),

u(x, t) c-Ho(Q)

t > O,1

where
<W,V> =

w(x)v(x) dx

and

n
a(w; u, v) =

au(x)aVb(x)d..
ax

jX w(x)) ax-

aJ

Also, H'(Q) is the closure of C'(Q), the set of infinitely differentiable functions
with compact support in Q, with respect to the norm given by
IIU12
H1(Q) =

aU

IUI 2

fu

L2(Q) +

(|L

It is well known that

au 12
IIU12 C n 11OU
2

UL2(Q)?CZ

uEH 0(

thus, it is useful to define the norm


IUIIHI(Q)

Fn-aU
xjI

Li 1

ax

]1/2

L2(Q)_

See [16], [21], [25] for details.


The system (2.2) is more general than (2.1); thus, it is sufficient in general to
approximate solutions of (2.2) in order to approximate solutions of (2.1). Again,
see [16], [21], [25] to find a discussion of the concepts of weak solutions of differential equations. We shall always suppose that the solution u to (2.2) is such that
u, au/axiand au/at are in L2(Q x (0, T)). Note that, since au/atE L2(Q x (0, T)),
u has a natural extension to t = 0, and it is assumed that uo is this extension.

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

579

GALERKIN METHODS FOR PARABOLIC EQUATIONS

(In ? 3 we shall consider a generalization of (2.2) in Theorems 3.2 and 3.3 at which
time we shall not need the above assumptions.)
We shall approximate the solution of (2.2) by requiring that u and v lie in a
, N.
finite-dimensional subspace of Ho(Q) for each t. Let v, E Ho(.Q)for I = 1,
Assume that the set v1, ..., VNis linearly independent. Denote by // the subspace
spanned by v1,
N,V. We shall approximate u of (2.2) by a function
N

U(x,t) =

(2.3)

(t)VX)

which satisfies

au V

(2.4a)

+?a(U;U,V)=0,

<U, V> = <uo, V>,

(2.4b)

t>0,

Ve(#,

t = 0,

VeA/'.

The system (2.4) is actually an initial value problem for the set of nonlinear ordinary
differential equations
Ccu'(t)+ A(a)oc= 0,
t > 0,
Ca(0) = (<Uo, V1>, * * *, <UO, VN>),

where C and A(ac)are the positive definite matrices given by


Ck,I =

Ak,J()

<Vk,Vl>,

n'
a jx
JE

aVk
- alVI dx.
E 0Cv(X))
-

of A(o)will be assured[10]by the reasonableassumptions


Thepositive-definiteness
exists
on (a1,j(x,u)) to be made in the next section.It then follows easily that c4(t)
>
for t 0.
We shall consider several schemes for approximatingthe solution of (2.4)
in whichthe variablet will be discretized.Thus, we shall referto (2.4)as the continuoustime Galerkinapproximation.
Let tm= mAt,where At = T/M and M is an integer. Let fm denotef(tm).
We shall analyzethe differenceequation
+

Um+1
UrnV
?a(Um+I?
At22

Um;Um++?Um,v)

(2.5)

=0,
VeJ,

m>0,
V E At,

<U0, V> = <u0, V>,

which we shall call the Crank-Nicolson-Galerkinapproximation.We shall also


treatthe generalizationof (2.5)given by
+

V)

+ a(2(1 + 0)Um+ 1 + {(1

(2.6)

l1 - 0)Um, V)

0)Um; {(1 + O))Um+ I

0,

=
<U0t V>
<uo1V>,

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

V eJ

m > 0,

580

JIM DOUGLAS,

JR.,

AND TODD DUPONT

where 0 < 0 ? 1. For 0 = 0, (2.6) is (2.5) and for 0 = 1, (2.6) is a backwarddifference approximation to (2.4). It can be easily shown using a fixed-point argument that (2.5) and (2.6) have solutions (possibly nonunique) for any At > 0; see
the proof of Lemma 7.1.
The solution of any of the differencesystems given by (2.6) requires the solution
of a nonlinear algebraic system at each time step. In order to avoid the necessity of
solving these nonlinear systems we shall introduce the following predictorcorrector approximation to (2.6):
,V

AM,t
-

Um+l

At

um V

+ a(Um;2(1 + 0)Wm+I +?(1

+
+ a({1? + 0)Wm + ?(1

(2.7)

+
?(1-0)Um,

- 0)UMr V) = O,

0)Um; {1?0)Um

V) =0

+i

for m > 0,

Ve X,

<UO, V> = <uo V>,

VeJ.

In (2.7), Wm+1 is required to be in-X. Note that (2.7) requires the solution at each
time step of two linear algebraic systems; thus, it is basically linear from a computational point of view.
We shall also consider a procedure for approximating u which requires only
the solution of one set of linear equations per time step instead of the two sets of
(2.7). In this procedure we extrapolate Um to predict Um+1 and then correct.
We shall call this procedure Crank-Nicolson extrapolation. It is defined as
follows:
Urn+i
A

UrnV

(U

?a(Un2 m - 2mU

vV)

1; R(UM+ 1 + Um), V)=O,

(2.8)
(2.)Ve,

m= 1,

,M-1.

Such a difference equation requires a different starting procedure than we have


used previously since it supposes that the solution is known at two time levels.
We shall use UOand U1 defined by (2.7) with 0 = 0. The procedures (2.7) can also
be approximated without solving two sets of linear equations. For 0 E (0, 1] we
shall consider
At+1

,V

a(Um;2(?

+ 0)UM+

+?(1

-0)UM,

V) =0,

(2.9)
<UO, V> = <uo, V>.

It is clear that (2.5)-(2.9) can be formulated using a variable time step


Atm= tmr+
1 - tm; it will be clear that the analysis of the corresponding systems
with constant time steps carries over in an essentially unaltered fashion to the
case of variable time steps.
3. A priori estimates for the continuous time Galerkin approximation. In this
section we shall present two types of bounds for the error induced by using (2.4)

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

GALERKIN

METHODS FOR PARABOLIC

581

EQUATIONS

to approximate solutions of (2.2). In each case the error bound is given in terms of
a norm of u - u7, where u is the solution to (2.2) and u7is the "best possible" approximation to u which has the form (2.3). Thus in each case we reduce the question
of bounding the error in the approximate solution to a question in approximation
theory. After demonstrating these bounds for the nonlinear equation, we shall make
some remarks on simplifications and improvements which can be made if the
equation is linear.
Suppose that w(x, t) satisfies
(watw,v> + a(w; w, v)= <f, v>,
(3.1)

<w(*,O),v>

v c-H'(Q),

Kw0,v>,

w(x, t) E-Ho(Q),
0~

0< t,

v E H(Q),
0 < t.t

~ ~~~~~~0

It is easy to see that


rT

(3.2) IIw(, T)12(Q) +

a(w(t);w(t),w(t))dt < c[[ f

11L2(Q x

(OT))

IIWo1L2(Q)]

by choosing v = w; see [16] for derivations of such a priori estimates. The first
type of error bound will be derived by a process which is analogous to the proof
of (3.2). For wo of (3.1) in L2(Q) there is an extension v of w which satisfies

(3.3) ||Dtv

L2(Q

X R)

a(v(t);v(t),v(t))dt ? c[ If

1L2(Q

R +) +

WO

2()]

where R+ = (0, oo). The second type of error bound will be derived by a process
analogous to the proof of (3.3).
Throughout the remainder of the paper we shall require that there exist
CO,q > 0, such that for all e R', x e?Q and se R,
n

<

(3.4)
i=l

< CO Z Q7.

aijx, s)jj

i,,j=l

i=l

We shall also suppose that there exists K such that for all x e Q, r, s e R and
i,j = 1
, n,
(3.5)
aj(x, r) - ai j(x, s) < Klr - sl.
In addition we shall suppose that the first derivatives of u with respect to the x
variables are bounded for all x and t. (This can be weakened slightly to a certain
Lp condition on these derivatives provided K and the measure of Q are sufficiently
small.)
For u(x, t) defined on Q x [0, T] define
||U |L2 xL2

|U

L2(Qx(O,T))

Iu

IX

L2

fT

u(

t)

H(Q)dt,

(3.6)

au(x, t)
sup

11L x L11VxU
U IIL2 x L?

(x,t)eQx

=
0

(0,T)

SUp
IIU( *t)
_ t_ T

sup
i=1l,

-,n

Oxj

1| L2(Q)-

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

582

JIM DOUGLAS, JR., AND TODD DUPONT

There exist constants c1 and 6 < 0, whichdependon T, n, diam Q,


L??,t1 and CO,such that,for u and U solutions to (2.2) and (2.4), respectively, and uaany function of the form (2.3),
THEOREM3.1.

K IIVU IIL

IIU_ U112
5 IU - U
2U?-L-+ +bIlU

< Cl

HxL2

u-u

UI 2
-L-X

(3.7)

?
a2

IIU U112

HIx L2

at

L2 x L2

Proof. In order to derive the estimate (3.2) from (3.1), w is used for the test
function v and the resulting expression is integrated with respect to time. Hence we
shall use u - U as a test function in the relation obtained by subtracting (2.4)
from (2.2); this will induce some additional terms since u - U s X for each t.
Then,
0(

,u-U\
=

(3.8)

+?a(U;u-U,u-U)

[a(U; u,a - U)
-

au

au-u a

a(u; u,a - U)]

+ a(U; u, u-u)

where u is of the form (2.3). This can be rewritten as


2 dtt u1
(3.9)

+ a(U;u - U,u

U11L2(Q))
-

[a(U;u,u
+?

U)

a(u;u,u

at(u-U),u-u

U)

U)] + a(U;u

U,u -)

Note that for each t e (0, T),


Ia(U;uu

U)

a(u; u,

U)

Qij =1

[ai,j(x, U)

ai,j(x, u)]

au 0a(k- U)d

ax

axi

(3.10)
_nK|IV.UIILXx

<

dxJll

n3/2 KIIVXUIILO
- UH1(Q)
xLOOIU - U 1L2(Q)a11

< CIIU-UL2(Q)[11a

- UIIHI(Q)
+ IIU-UIIHI(n)]

Also note that

ja(u;u

uu

u= IO(u1-

ai,j(X,

U)

U) c(u - U)dx

UAax

(3.11)
<

COIIU -

UIIH IIU -

1IH

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

METHODS FOR PARABOLIC

GALERKIN

583

EQUATIONS

If we now use the trivial inequality ab < ea2 + (4)-<'b2 on the right-hand sides
of (3.10) and (3.1 1) and then add the resulting bounds, we see that
a(U; u,a - U)

3
(3.12)

- U)I + Ia(U;u - Ulu -u)l

a(u; u,

+ C[IIU- UIIH1(n)
+ IIU- U1122(Q)1,
UIIH1(Q)

< ?||U -

lower bound in (3.4) implies that

where C depends on

E. The

(3.13)

ullu - UI121(Q)< a(U;u

U,u

U).

Using (3.9), (3.12) and (3.13), we see that


+ (11U- U112k(Q)
d-(lIU- U1L12(Q))
dt
(3.14)

< C[l|a

flU UII2H1()]+ 2 a

U1l12(n) +

,T

term with

Multiplying (3.14) by e-t and combining the Cllu -Ull2(Q)


(d/dt)(llu- Ull122(Q)),we see that
d
+ (3ctll
~~~~12
- U||L2b)) + e
d-(etllau

dit

(3.15)

Hu-U1UU12
-

a(u-

Ce-c&llu- UllH(g2)+ 2e'

at

u)

We can integrate (3.15) with respect to t and see that for each s e (0, T],
rs

||slla
-

2
U11 2(s)

lu

?C,J

(3.16)

eCt ll

Ull2 2(Q,(0) + 6

+ 2

Ilu -_ ll(Q,)e-cdt

s~~ra( - U)

U112

dt

,ec((- dut. z))

The last term can be integrated by parts:


Js

(d('

C7cKu

tau-)u

LJ; e-t(u-u))dt-e

fLec
) t u-<u,
Ae-ct
u -U

dtt

at U,u->jdts

6>Ce

c'< u-

Uu - i>]

dt.

Thus the last term in (3.16) can be dominated by


g[IfU

U 11
--lu
L2(Q)(S)+

+ C

IIU-

U |fL2(Q)(0) + f|u

UIIL2XL2+

a(

U || 22(Q
+

x (0,

S))]

L- ullL2(S) +

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

flu-

iL(so(0j.

584

JIM DOUGLAS, JR., AND TODD DUPONT

If we substitute this bound into (3.16), use the fact that IIW1L2

if w E Ho foreach t andnote that Iu -

U 11L2(Q)(O)<

IIu-1-

L2

1L2(Q)(O),

CIIWIIHI>L2

the conclusion

follows.
In error bounds for Galerkin approximations to elliptic equations there is
usually a symmetry in that the norm which is applied to the error u - U is the
same norm that is applied to the interpolation error u - ui; see, for example, [4].
Because of this symmetry you can in effect say that in some particular norm the
Galerkin approximation to the solution, up to a constant factor, approximates
the solution as well as is possible by functions in the finite-dimensional space in
which it is required to lie. The error bound above lacks this symmetry since we
requirethat 1N(U - 1)/at 1L2(n (O0T)) be small but do not includea corresponding
term for the error. The bound to be derived now has the symmetry mentioned
above.
For w(x, t) E L2(Q x R) such that IzI
'12w(x, z) E L2(Q x R), where the denotes
Fourier transform in the t-variable, we define (Dt12w)(x,t) [IzI
'2w(x, z)] (x, t),
where superscript ^denotes inverse Fourier transform in the r variable. In particular,
-

|| Dt 11 W112
|L2 (Q XR)

II12
=||1|/W

12
1lL2(QXR)

gwl(x )dxddl .

1,zl1,12(Xrd
QxR

Suppose u satisfies (2.2) and


zero for t < 0, then u satisfies
(3.17)

au/atis in

L2(Qpx (0, T)). If we extend u to be

d <u, v> + a(u(t); u(t), v) = K<uo,


v>,

v E Ho

where 6 is the Dirac measure at t = 0. In [16, Theorem 4.1., p. 59], it is shown in


a more general setting that, if uo E L2, there exists a unique u E Ho() x L2(R)
which is zero for t < 0 and satisfies (3.17). It is also the case that if an inhomogeneous term (in L2(Q x (0, oo))) is added to the right-hand side of (3.17) and if
uo = 0, then D1/2u E L2(Q x R); see [16, p. 55]. We shall show that D'u E L2(Qpx R)
2
for any oc< 2 even though uo0 0. (D'u eL2(2 x R) if I1l'7cL2(
x R).) We
shall use these arguments to establish bounds on the error in the continuous time
Galerkin approximation.
It is of interest to note that even in the case uo0 0 there exists a function w,
which agrees with u for t > 0, such that D112w e L2(Q x R). This function w is
just twice the even part of u; i.e., w(x, t) = u(x, t) + u(x, - t). Since WV(x,
I)
2Re a(x,r), one can show that 1D1'2wlL2(QXR) ? C
by taking the
Fourier transform of (3.17) with respect to t, using wV(x,
r) as the test function v
for each r, and integrating the imaginary part of the resulting relation with respect
to r. Although a similar set of remarks hold for the Galerkin approximation, we
are not able to estimate the half-derivative of the error in cases in which u( *, 0)
and U(-, 0) do not agree. In such cases we shall estimate the Li-derivativeof the
error, where 0 < oc< 2
In the proof of Theorem 3.1 it is clear that there is exponential growth with
T of the bound for the error. We could have made the constants independent of T
by first making a change of dependent variable. We shall need to do so here since
in discussing fractional derivatives it is convenient to consider the error for all
t > 0. Suppose u(x, t) is zero for t < 0 and satisfies (3.17) for t E R. Let e'tw(x, t)
-

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

585

GALERKIN METHODS FOR PARABOLIC EQUATIONS


-

u(x, t). Then w satisfies

d-<w,v>

(3.18)

aA(w;w,v)

veH ,

A<w,v> = 6<uO,v>,

where
aA(w;u, v) = E

(3.19)

aij(x, e2tw(x, t))

(x, t)

(x, t) dx.

It is clear the form a. has a dependence on t; however, we shall not write t explicitly
as an argument of a2. Note that for each t, ai,j(x, eAtw) is Lipschitz continuous
with respect to w with Lipschitz constant eAtK,where K is given in (3.5). Also
notethat,if IIVXUIILOXLO< oo,then foreacht, IIVxw( , t)11L-(Q)< e-`t||VXuI||LooxLoo
THEOREM3.2. Let u be the solution to (3.17) which vanishesfor t < 0. Suppose
that 1V.U 1L- L- is finite or that the constant K in (3.5) can be chosen to be zero.
Let U(x, t) be the solution to (2.4) for t > 0 and let U(x, t) = 0 for t < 0.
Let z = u - U. Suppose that uo ce #. Then there exist constants i and C, dependent
on n, diamQ, K 1VxU 1Lx,
q and COsuch that
>

D/(ez)

LHzI

(3.20)

x L2

? C[IlDt (e2t(u

))2XL2

where u is anyfunction such thatfor each t, u( ., t) E X.


Proof. Let w(x, t) = e- Atu(x, t) and W(x, t) = e (3.18) and W satisfies

L2](

t). Then w satisfies

tU(x,

-<WI V> + a,(W; W, V) + A<W,V> = K<uo,V>,


dt

(3.21)

Hox

VeJf.

We shall use
(p(T)

(2c) -

rt

dt
eitzq{o(t)

1/2 {
J00

as the Fourier transform for po- (R)(Y(R) is the set of rapidly decreasing
functions and f'(R) is its dual; see [15]). If we take the Fourier transform in t
of (3.18) and (3.21), as elements of f'(R), we obtain, for each T,
v> + E X

iT<WK(,),

aij(x, eAw)

(x, T)

+<W*,T),

iT<<KW(

(3.23)

,T),V>+EV>

T),

(x) dx
V>

e AtW)
7w(x,T)
ai,j(x,
Ox

<U0,

V>,

vc

Ho,

(x) dx

\"j,j-',

? ,{KW(.,z), V>

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

Ku0,V>,

Ve /.

586

JIM DOUGLAS, JR., AND TODD DUPONT

Note that
(iT +

K)1(

W)

{(iT + A)<(w -

(3.24)

I'

W) (., T), (W

)K<(W-

W(

W) (X,

ai,j(x, eA w)]

W)(,

T), (W
-

W)

a(w -W)

(X,T)dx

T)>

(x, T) a(

At
tai,j(x, e W

) (x, T)

t
e w
(W
)
ai,j(x,
eAtW)
(X,T) 0(f--W

([ai,j(x, e tW)

? (iT +

~
a(w -W
j ai,j(x,
eK)d

Z{ (ai,j(x,etw)
-

~~, fI

T)I |2

(.,

aw

T) dx

(X, T)dX

(X, T) a

W) (X, T) dx

W) ( ,'T)>

(x, T)

(x, T)dx,

d-

where Cvt(.
, t) e #. Using (3.22) and (3.23) we see that the term in braces on the righthand side is zero. If we integrate (3.24) with respect to Tand use Parseval's formula,
={fdT

fgdt

we can see that


+

?111W
W|HoXL2

<

Jn

41 W
t

+ A4W -

(3.25)

<

W)|L2

||DtI(w

||W -

+ CoIW

a(

?1W -

WL2XL2W

)(x, t) (

W |L2 XL211|W1 -

WL2

WWHAXL2

+
C[HL11W

- 1Dj'V(W
IW)

DID2(W

)(x, t)dtdx
HOXL2

WL2
x L2

X L2

WWIHIxL2IW

1CW
WOIXIL2

W L2xL2

W)||L2 X L2||DtI(W W)x-

(x, t)dtdx

WXIIL2

EJJai,j(x,eAtw)

eatw)]Ow(x, t)

XL2|D/(W

W11L2XL2 IIW

n312K||VxUL||L?

W2XL2

[aij(x, eAW)ai,j(x,

+ |Dt I(w

W)L2 x L2 + (4-i + C)IW -

W2

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

xL2X

2XL2]

587

GALERKIN METHODS FOR PARABOLIC EQUATIONS

where C is independent of X. Note that uo e k is needed in order that


- W)IIL2 be finite. Hence if we take ) > 4C, we see from (3.25) that
D1D"2(w
IIW- _W

x1L2

IIW -

W11L2xL2

C[IIW -

+ rI|D' I2(w -

(3.26)

Multiply (3.24) by

|JDt 2(w- W) l2XL2

W)ll

i sgn T and integrate with respect to


< C[IIW -

1Lw +

lVIHiXL2

11w-

WIIHAXL2IIW

+ || Dt I(w

IIDP2(W

1W)IL2XL2]

2XL2.

to see that

VIIL2XL2iIW-

"IIHAXL2

WIIHAXL2

W)I||L2XL2I||DtI (W -

W)
llL2 xL2

L2I
L2IW - WIIL2X
+ AIIw- 14"IL2X

(3.27)
<{1 |D1'2(w
+

L2
W)jl ~X

C[IIW -

Now use (3.26) and (3.27) to see that


11D

(W
w

C[||w

1|IlAxL2

+IllDtW(w

1)IIL2xL21.

W IIHO
XL2

L2 +
W)l112X11D
?

||W

I LXL2

WVI
1w -

11
WIIH XL2
+

(3.2!8)

WIIHXL2

- W1ll1x

W1W-

L2 + IWDt

1)llL2XL2]-

Replace w by e-iu, W by e-itU and w~by e-itu to obtain the conclusion.


It is worth noting that the above theorem remains valid if inhomogeneous
terms f(u),v> and Kf(U), V> are added to (3.17) and (2.4) respectively, where
f (u) = f (x, t, u), f is uniformly Lipschitz continuous in u and f (x, t, 0) e L x L2.
This generalization requires almost no change in the proof.
X we shall need the
In order to get a similar bound in the case in which u0 sM
following continuous dependence theorem for the differential problem.
LEMMA 3.1. Supposeu satisfies(3.17) and thatv satisfies(3.17) withu0 replaced
< ciz.
by v0. Supposethat u and v vanishfor t <0O. Supposethat IIVXUIILOOxLX
Thenfor 0-?
<}2 there exist constants C and X, which depend on ~, n, diam Q,
KIIVXuIILOxLX, 112and COsuch that
(3.29)

t(u
D11D(e

v))1L2XL2 +

H|e I(u- v)l|

OXL2 ? C1UO -

VOl12-

In case K can be taken to be zero, we can take i = O.


Proof.If we subtract the relations satisfied by u and v, we see that
uuv-

v, w> +a(u -v,w)

=31uo7

2 rw>,

where

(3.30)

j
(r, s) Lai,(x,
tufxor

ac

+
(x,t)nu and j dx,
ranin

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

weHer

JIM DOUGLAS, JR., AND TODD DUPONT

588

ai j(x, t)

(3.31)

v(x, t))

taij(x,u(x, t))aijtx,
u(x, t)-v(x,

t)

1)

O,otherwise.
01
Since ai,j is bounded, it follows from [16, Theorem 1.1, p. 46] that there exist
constants C and i such that
Ile-(u

(3.32)
The

_ CIIUO -

v)IHIxL2

VOIIL2

is nonnegative and is chosen so that

a(r,r) + 11r 2 >

(3.33)

reH0;

lrllH,

thus i depends only on n, i and K1 VPu L XXL*(K 1VXUIIL-xL LX)2.) Let w = e6-t(u - v) and wo

= ,'(n3 n
(In fact we can take
uO - vo. Then w satisfies

z e HO

Z>,
d-<W,Z> + a(w, Z) + i<W, Z> = K<WO,
dt

(3.34)

Take the Fourier transform in t of (3.34) to see that


z e HO,

*, T),z> + A(T, z) = <wO z>,


)<WK(

(iT +

(3.35)
where

(3.36) A(TZ)

Thus, for 0 < y<

au

+ wa
aij[ X V, (X,T)

W (X,T)

(x)dx.

1,

ITI-II '( , T)IIL2 _ 1+ ||

( ,T|L2

T)|

=<(1 +
W( , T))I + (1 ? {) vW( , z)
IT|)zK[IA(T

(3.37)

IA(T,

w(

,T))I

(1

A)(

A(

I<W0,

L2
T,)>I]

,T)(.)L2

+ ?(1 + ITI)l|IWO11L2 + (1

+ ITI)l

IIW(,(

)IIL2

From (3.37) we can conclude that


(3.38)

IzI'

it|W(,zT)
12 ?

C[IA(T
z(.,

T))I+

vt'( ,zT)L2

+ ? ITI)
(1

IIWO2

If we integrate (3.38) with respect to T,we see that


(3.39)

ID(1 -y2W

I2L2

?2

C[IIWI1HxL2

WO

2L.]

The conclusion follows from (3.32) and (3.39).


Note that as a special case of Lemma 3.1 we can see that, if u is a solution to
(3.17) which vanishes for t < 0, then Dyu e L x L for 0 < oc <2. This follows
from the lemma by taking vo = 0, since this implies v(x, t) = 0.
3.3. Let u be as in Theorem 3.2. Let v be the solution to (3.17) with uo
THEOREM
< 2. Let
replaced by vo. Suppose vo 0 # and that v = Ofor t < O. Suppose 0 _ 0C
U be the solution to (2.4) with uo replaced by vo; extend U to be zerofor t > 0. Then

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

589

GALERKIN METHODS FOR PARABOLIC EQUATIONS

there exist constants i, C and C1 such that


ID'(e-

(U -

U))IIL2XL2

Ile-(u

U)1 HI XL2

C
c VO- UO L2

(0

C1[

DDtI (e-'t(v

+ Ile -(v

V))
i) 1

L2XL2

xL2],

where v(., t) e J for each t. The constants i, C and C1 depend on oa,n, diam Q, j
and CO; in acdditioni and C depend on KIIVXUIILXXLX, and C1 depends on
KIIV,VIIL-XLX-

Proof. Note that


IDI(e t(u - U))||

L2

?l e t(u -

U) 11 XL2

< C[||D'(e

(3.41)

t(u -

V))|| 2XL2

+ IDt 12(e t(v - U))||

I|et(u

L2

v) HIIxL2

?l e t(v - U) 11HxL2

The conclusion follows from (3.41), Theorem 3.2 and Lemma 3.1.
The line of thought which led to Theorems 3.2 and 3.3 was suggested to us
by E. B. Fabes and N. M. Riviere.
We shall now consider the case in which (2.2) is linear; i.e., the case in which the
functions ai,j are independent of the value of u. In this case we no longer need the
hypothesis that 11VxU
is finite. The terms which were estimated using this
L- Lhypothesis are zero in the linear case; see (3.10) and (3.25). It is worth noting that
in this case the constants in Theorem 3.1 can be chosen independent of T, and the i
in Theorems 3.2, 3.3 and Lemma 3.1 can be taken to be zero. (This is one of the few
places in which we use the boundary conditions for t > 0 in a critical way; see
also ? 8.) In the linear case we can obtain an a posteriori error bound. This bound
has the advantages that it requires very little information about the solution u to
problem (2.2) and requires no use of approximation theory theorems.
THEOREM
3.4. Suppose that (2.2) is linear and thefunctions aij are continuously
differentiablein Q. Suppose that tY consists offunctions which are twice continuously
differentiable on Q. Let u be the solution to (2.2). Let U be the solution to (2.4).
Let z = u - U. Let
(3.42)

-f(x,

t) =

at

aaijxx)a,ux

Suppose O _ OC< .2 Then there exists a constant C such that


(3.43)

2 + IUo -

U('.,0)II 2].
=
Proof. Note that z satisfies (3.1) with wo uo U( , 0). Thus the conclusion
follows from (3.2) and a minor modification of (3.3).
In the linear case using a special space of splines, Price and Varga [23] obtain
a somewhat different bound by an argument based on the energy inequality (3.2).
Zi2XL-

Z HIXL2

11Dtz L2XL2

4. A priori estimate for the Crank-Nicolson-Galerkin approximation. In


this section we shall derive a bound for the error induced by using the Crank-

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

JIM DOUGLAS, JR., AND TODD DUPONT

590

Nicolson-Galerkin system (2.5)to approximate the solution of the original differential problem (2.2). The bound is similar to the first type of error bound derived in
? 3 (Theorem 3.1); however, in this case we shall require additional smoothness
of u in the variable t since t-derivatives have been replaced by differences. It
will be shown that, as expected, the Crank-Nicolson-Galerkin procedure is
second order correct in t. We shall also indicate the derivation of a similar bound
for the approximation (2.6) which for 0 = 0 is first order correct in t.
In order to state concisely the bound for the error induced by using (2.5)
we shall define some notation. Let U be the solution to (2.5) and u the solution to
(2.2). Let z = u - U be the error; this is defined at times t = t. = mAt. For a
functionf of (x, t) or (x, ti) define

(4.1)

fm+1/2

+
[2fm ?m+11;

please note that fm+1/2 does not mean f((m + 1/2)At) even though the latter may
be defined, as in the case of u(x, t).
THEOREM
4.1. Let u(x, t) be the solution to (2.2)for (x, t) E Q x [0, T]. Suppose
that IVXUI L- XL? is finite (see (3.6)). Suppose that for each x E Q?,au/at and au/ax,
are twice continuouslydifferentiablewith respect to t and supposethata3u(x,t)/atN(x,t)
and a3u(x,t)/(8t2ax) (x, t) are bounded uniformly by C2 for (x, t) e Q x (0, T)
and i = 1, --., n. Then there exist constants C, 61 > 0 and To > 0 which depend
on T, n, diam Q, KIIVxU1 L- xL-, t and COsuch thatfor At ? zo,
M-1

||ZM||L2(n) + 61
_

cL{?
LL

(4.2)

lZm+12 HIM()At
m0~~~~~~
m=o

=
m0

E1

l+

I(U
1(

+ ||(U
a)OII22(Q)

)+ 1/2

u)m +1I2 Ho()

+ I|(U -

(U

7)m-12

At

")M-1/2lL2(I2) +

(U -

Il(U -

'0l/2I2L2?

At
A

2
L2(n)

C2(At)

1
j

where T = MAt and im CEX, m = 0,..., M.


Proof. Note that we can write (2.2) at t = (m + ')At in the following fashion:
(4.3)

Um+1 -um

At

+ p?
Cm;?Um+12
V) + a(um+1/2 +

veH'(Q),

where pmand Vxkmare O(C2(At)2)pointwise and in L2. We use (2.5), (4.3), much
care and little else to see that
At

Zm 112

=
+

+ a(Um+ 1/2; Zm+ 12, Zm+ 12)


+ ?/2

+AtZ,m(u-17)m
At

At

Zi

(U -

(u7-

+ a(Um +l/2; Zm+ 1/2I, (U -)m

U)m+ 1/2

+ 1/2)

+ a(um+ 1/2 + Cm; Um+1/2 + Cm

U)m+1/2) - a(Um+1/2; Um+1/2,(0 - U)m+1/2)

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

METHODS FOR PARABOLIC

GALERKIN

<Pm,(k

a(um+ 1/2 + Cm; Um+ 1/2 + Cm,

U)m?+i2>}

- a(Um+ 1/2;Cm(7
<

,(U -

/\-

At

U)m+ 1/2)]

+ {O} + KIIVxUliL-xLol4Zm+l/2
?Co||

ICmlHoIl(

<Pm'0

+ CmIIL2II(U

+ I1PmlL2l(U

U)m+1/211Hi

U)m+1/2)

U)m+ 1/2)

U)m+ 1/2>

+ COIIZm+1I211HtII(U

17)m+i1/2

+ Cm, (U

[a(Um+1/2;Um+1/2
-

591

EQUATIONS

")m+1/211HI

U)m+l12I1HI

U)m?+1/21L2.

?12
The first term in the left-hand side of (4.4) is just [ zM+,
2zm12 2]/2At, and
the second is bounded below by hlIZm?iI2112. The right-hand side of (4.4) is
bounded by
Zm+1

zm

A
/\t

(u

?+

1)m?+ 1/2

lZm+?1/2|

C[0(u

i)m+1/2Hl

(4.5)
?

IIPrnIIL2

KmIIH?

I+lZm+1/2IL2],

where C depends on E.In showing that (4.5) bounds the upper bound in (4.4), we
have replaced each term la- Ull by ll - ull + llzl and zlz+ CIIby llzll + 11411
We have used ab ? ea2 + b2/48 several times and have also used the inequality
11
f

I1L2 <

112
IlZm
(4.6) ?

> 0 and some C,

C! f 1WHl. Thus,we concludethat for some y

2-<1C{Lm2
1122u112
+
?

IH<

YZm+1/2

+ 11ZmIZL2}+

!!Zm+ilL2

2z

11f2,H + lp
ll 2?+
PmL2

m?
U)m+1/2

1(U

-<

(u -

At

C~m
N1

We can conclude from (4.6) that (with C increased by one so as to include the last
term)
(1 - CAt)zm+ 1||L2

(4.7)

? CAt{ 1(u -

?2At

+(1

CAt) fZmLj2 + YAtIlzm+1/2I Hi

L)m+ 1/2t1Hi

Zm+l1

At

Zm(

?
-

LPmjjL2

lmi2
m12L2

ttomlfIO}

AtjIlZm
1 2.

Let g(At) = (1 - CAt)(1 + CAt)- 1. Multiply (4.7) by [1 + CAt]that for some y > 0 and some new constant C,
g(A\t)m+

(4.8)

<

IIZm+ l22

CAt{ II(u -

g(At)mz lZmIj2
17)m+ 1/2 IHoI ?
iA

to see

+ YAtIIzm+1/2L1
I[PmIIL2 ?

Zml-z

At Z+1
+

lg(At)m

At

m2

7AtIIZmIL2

iKmIHo}

g(At)m

1? CAt
A(U
I

1)m+ 1/2

In deriving (4.8) we have noted that for 0 ? mAt < T, g(At)mis bounded above and
below by positive numbers independent of At, for At sufficiently small. Finally we

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

JIM DOUGLAS, JR., AND TODD DUPONT

592

sum (4.8) on m to show that


Z
1Zm+

ZM IL2 + 6$1

m=0

112IAt

-M-1
? C
(4.9)
(49)|

_ =

[ 11
(U (U -

_=

i)m + 1/2 Ho ?

Pm

At ? ||(U-

)m-

)m+1/2 (
At
l|(U

H|m(|i]At

L2 +

L2

1)1/12 L2

(U -

U)M-1/2

L2]

In order to show that (4.9) follows from (4.8) we sum the last term on the righthand side of (4.8) by parts to get

M-1/2,g()

CA
+L~1

Zm,

g(At)

(
(U -

1/2> -

<ZO,(U
-

)m 1/2-

g(At)m(u-

")m + 1/2

2]

These terms are dominated by norms by using Cauchy's inequality and then, as
usual, by sums of squares of norms. In this process a small multiplier is placed in
front of Y7M--il I m L2 so that it is cancelled by the sum of the next-to-last term in
(4.8).
Finally, in order to get (4.2) from (4.9) we simply note that 11pmL2 and lSm1'2A

whereIQIis the measureof Q.


are boundedby C2(At)4IQI,
Since the Hl-norms of the error which appear in the bound (4.2) are not
but IZm+1/2 '112, we conjecture that it may be possible to get a better approximation to u by using Um+ 1/2(x) to approximate u(tm+112,x) or by using
+ 1Um + Um_i to approximate um than by
2{Um+ii2 + Ur-1/2) =-Um+1
using Umto approximate um. It is certainly the case that either of the suggested
approximations remains second order correct in time.
THEOREM 4.2. Let u(x, t) be the solution to (2.2)for (x, t) e Q x [0, T], and let
Um(X)be the solution of (2.6) for 0 < tm= mAt ? T and x e Q. Let z = u - U.
Suppose that for each x E Q, au/atand au/axi are continuously differentiable with
respect to t and that 02U/at2and 02u/ataxiare boundedby C2for (x, t) fe Q x (0, T)
XL? isfinite. Thenfor 0 E [0, 1] there exist
and i = 1, ..-, n. Suppose that IIVU11L??
constants C, 61 > 0 and T0 > 0 which depend on T, n, diamQ, K1 VxU L XLO, '1
and COsuch that for At < T0,
M-1
1Zm112,

ZML2 +

(4.10)

0
mr=

? C

ZmOkHAt
E

+
l(u - U)mr011
H2IAt

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

At -2

At

593

GALERKIN METHODS FOR PARABOLIC EQUATIONS


+

(U

u)o

(u

L2?2

i)11

L2

?+

(u

L2

U)m-1,01

C2(At)2]v

where T = MAt, UmeAX andfm,0 = ((1 + 0)/2)fm+1 + ((1 - 0)/2)fm.


Proof. As in the proof of Theorem 4.1 we write (2.2) at t = ((1 + 0)/2)tm?+
+ ((1 - 0)/2)tm in the form

(4.11)

Um?i -/

+ a(um,o+ cm;umo + Cm,v) = 0,

+ Pm,V

Unt

veH(Q);

here Pmand Cmare O(C2At)pointwise and in L2. We use (4.11) and (2.6) with a test
function v = Zm,oto see that
2A\t[iIzm+ 11122

+
IIZmIL22]

IIiZm,OIIH

1t0
112 2 -

2<

iIZmiL22]+

2A

lZm+

112L2 +

Zm

Lj2

2KZm?lzm,Zm>}

+ a(Um,o;Zm,0
I Zm,o)
zm+ 1

(4.12)

{<

Zm?

,ZmZ0

Zm

+ a(Um,o; Zm,0 Zm,0)

+ pI (

U)m,0

+ a(um,o + Cm;um,0 + Cm, (a

a(Um,o; Um,0,(U

U)m,0)

U)m,O)}

+ [a(Um,o;
Um,0+ cmI(a - U)m,0)
-

a(um,0+ &m;Um,0+ Cm,(-

<Pm.'
(ta- U)m,0>+

Zm?i

(a
U)m,0)]- a(Um,o;Cm,
-Zm

U)m,0)

,(u - u)m0

+ a(Um,o;zm, (u - 9)m,0).
The rest of the proof is essentially the same as that of Theorem 4.1.
5. A priori estimate for predictor-corrector version of Crank-NicolsonGalerkin approximation. Most of this section will be devoted to showing that the
approximate solution Umdefined by (2.7) with 0 = 0 (Crank-Nicolson predictorcorrector) is second order correct in At. Since (2.7) involves only linear systems of
equations, this says that we can preserve the full order of accuracy of CrankNicolson while avoiding having to solve nonlinear systems of equations. In order
to maintain this second order accuracy in At we shall need to make stronger
hypotheses on u and a = (ai,j) than we have made previously. We shall indicate
two somewhat weaker results that do not require these strong additional hypotheses. If (2.7) with 0 E (0, 1] is used, then we maintain first order accuracy in At.
Notice that there is no reason for using (2.7) for linear equations since the same
linear system of equations would merely be solved twice.

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

594

JIM DOUGLAS, JR., AND TODD DUPONT

THEOREM5.1. Let u be the solution to (2.2) for 0 _ t _ T, and let U. be the


solution to (2.7) with 0 = Ofor0 < tm< T. Let z = u - U. Supposethat 03u/(0t2aXi),
03u/8t3 and all of the second derivatives of u are continuous and are boundedby C2.
Suppose that aa/au, and 02a/(Oxjau)are continuous and boundedby C2'. Then there
XLl
exist constants C, 5 > 0, and To > 0, which depend on T, n, diam Q, KIIVxUILO
C and CO such that for At ? To,
M-

IIZMII2+
<

1
IIZm?ii2IIHAt

||(U

)
112 /
u)mn+1/2ftHiAt

C[ME 1)m+
_1 l(
<C[Z

(U

rn>
m= 1

m=O

1/2

At

(U

)1/2

At

tL2

(5.1)
+

II(U

11(u

a)0II2

)M-l/2II2

i)112IIL2 + (C2 + 1)4C2(At)41,

11(u-

Eirne
where T = MAt and am

m = 0, .., M.
Proof. We shall denote the error in the prediction of um by Ym= um - WM

The proof has two parts. The first part is an estimate of the form
(5.2)

IIYrn?IL2

CfllZ,,jL2

I|(U -

)M+

1/211Ho] +

O(At)

The second part consists of inserting (5.2) into the argument used in the proof of
Theorem 4.1. In particular we shall show, using (5.2), that Zm satisfies (4.6) with
an additional O((At)4)term on the right-hand side.
In order to demonstrate (5.2) we first write (2.2) at t = (m + ')At as
urn?

(5.3)
(5.3)

+ a(umn+ r;in12+~V)
+ 1m;
1/2 + CM

Um? PnV

A
pAVt
UmA+
-t

) = 0,

vvc-H',

where Pmand Cmare O(C2(At)2) and aXmis O(C2(At)), pointwise and in L2. To
simplify notation and emphasize the similarity to the computations in the first
part of Theorem 4.1 we shall use the notation Ym+ 1/2 = (Y2n? 1 ? Z1) and dm+112
2-( M+1 + Um). Proceeding as before using (5.3) and the first equation in (2.7)
with 0 = 0 with the test function v = rm+1/2' we see that
2ASt[I + 1 1L2

11Zr|L2] +

1/2 1IHO
11IlYrm+

ZM+ Prn'(U7 -

< fYM+

W)M+ii2

+ a(um + Lrm;Umr+1/2 + Cmr,(fU -

Yrn+1u -

At

<Pmn(a
a(um +

m(U

+ a(Urm Ym+112, u

Om?+ 112

W)m?+ 1/2>

+ [a(Um;

rm;Umr+l/2 + Cm, (a

a(Um;

Um+ 112 +

W)m-+12)]

Cm(a
-

m+
1/2V

a(u -

i0m+1/2)

W)m+1/2)

a(U.; Cm,(ia -

(5.4)

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

W)m +i12)

W)mr+1/2)

k)m + 1/2)}

GALERKIN

1[I
At
?

lY?1L2

lPmIL2[1(U

? C0Km

HoI(U

[[a(Um;Um+l/2

METHODS FOR PARABOLIC

Zm L2]1T(U -

lim?

IL

i)m+1/24L2

+ (m,(U

? [[a(Um; Um+ 1/2 +

(mr Ym+ 1/2) -

+ C IZmI2 +
+ CC||Pm||L2

a(um;um+l/2

Lm+1/2H1%H]

+ At||?

H'm|IH

1)m+l/2)

+ a(um; um+ 1/2 + (m 'Ym+ 1/2)


<

1H1II(U- U)m+1/211HA

11Ym+1/2L2]

)m+1/211HI +

+ C0Unm+12

l/2K2

595

EQUATIONS

a(um

1(u-

Um+ 1/2 +

1/2)]]

nYm+
rm

a(um + otm; um+ 1/2 +

mllL2)

+ (m,Ym+1/2)

a(um + LXm;Um+l/2 + (m,(1" Om+1/2)-]

m'Ym+1/2)

iOrn1/2Kj
+

21KYm+

1/2

HA

a(um + OCm;Um+1/2 + (mwYm+1/2)

In deriving the last inequality the terms in double brackets were estimated by
K11VxUL xL- [|LZm

+ OCm||L2 11U

<
?||Ym? 1/2||H1

U)m+1/211HI +

IIZmIIL2IIYm+1/2IIHi]

+ C(At) 1[j|Zm|22 + ||(u

u)m+1/21A]

+ CAtlloCm12

Note that

Ia(tim; Um+
=

1/2 + (m' Ym+ 1/2) -

a(um + Otm;Um+ 1/2 +

[ai,j(x, um)- ai,j(x Um+

4jn E

{[ai,j(x um)- ai,j(x, um+

m Ymn+1/2)1

dx

m
'm)]-

(X tm+1/2)}

Ym+l/2(X)dX:

? C(C2 + 1) C2AtLiTm+l/2lL2
?=

L2] + C(C2 + 1)4C2(At)3.


t [IKYm+12L + 1Zm1 L2
2

8At

Using (5.5) and (5.4), we see that

(5.6)

Ym+

L2 = C[

Zm L2 +II

(U -

1)m + 1/2 II
HO

(C2 ? 12)C(At)4]

The inequality (5.6) is the particular form of (5.2) we shall use.


We now use (4.3) and the second equation in (2.7) with 0 = 0 and proceed
exactly as in the derivation of (4.4). There is only one change. The change is that
a(Umr+1/2; w, V) is replaced by a(Wm4+
1/2; w, c). This alters the estimation of only
one expression:
a(WI +

1/2; Um+ 1/2 +

a(um+ 1/2 +

<

K||VxU11L-xL-[

(rm,(1U

U)m+

1/2)

m; Umr+1/2 + k,( (U
1Ym+1i L2 +

U)m+ 1/2)

1Zn, 1 2 +

KmlnL2]1

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

(a

-U)mr+1/2

11HA

596

JIM DOUGLAS, JR., AND TODD DUPONT

After using (5.6) to estimate


4.1, we conclude that

2
j, 1
IIym+,

C[f|(U

+ IIZMr+111L2+ IIZMII'2
+

'4m+1/2IHI

?2 Zm+i
+ 24
m+A
At

Hi

+ YIIZr+1I2

A[WmZ++111L2 -lZrnl2]

(5.8)

in (5.7) and using the estimations of Theorem

Zrnm
I(u

(C; + 1)4C2(At)4]

- Om+1/2.

The relation (5.8) is just (4.6) with a different (At)4 term. If we perform the summation of (5.8) in the same fashion as that of (4.6), the conclusion follows.
If we drop the additional smoothness assumptions on a made in Theorem 5.1
but retain (3.4) and (3.5), we can show that (2.7) with 0 = 0 has a time truncation
error of O((0t)312) instead of O((At)2). This will be shown in Theorem 5.2. In
Theorem 5.3 we shall show that if two passes are made through the corrector
equation, then the scheme becomes second order correct without the additional
assumptions of Theorem 5.1.
THEOREM 5.2. Let u be the solution of (2.2) and U be the solution of (2.7) with
0 = 0. Let z = u - U. Suppose that 11
VXU11L-XL- is finite and that 03u/at3, 02u/at2
and 03u/(at2axi) are boundedby C2. Then the conclusion of Theorem5.1 holds with
the term (C i)4C2(At)4 replaced by C2(At)3.
Proof. This theorem will be proved by the same process as Theorem 5.1 with
one minor change. Only the estimation (5.5) has to be changed:
Ia(um;Um+ 1/2

+ Cm, Yrm+1/2)

LOxL
0Kl|VXu

(5.9)

451IIYm+
-

a(um+

am; Um+1/2 + CmIYm+ 1/2)I

rlmL2WYr+1/2KIA

112H + CII1 m 11L2


1/21j12

451JYJJ

+ CC2(At)2.

Using (5.9) instead of (5.5) gives

(5.10)

1L22 < CL |ZrMn12 +


I1Ym+

||(U -

Ujm

1/21HA + C2(At) 3]

instead of (5.6). The rest of the proof goes just as that of Theorem 5.1.
Let us define a predictor-corrector form of the Crank-Nicolson-Galerkin
equation in which the corrector is used twice. Let
-

A
t+1

AtM+,

m V

V)

+ a(Um;

(JWm+, + U.),

V) = 0,

+ a(2U(Um + Wm4+i); {(Um +

Tm7+i), V) =

0,

V ei,

(5.11)
<Un+iUm,nV)

+a('

Um + Tm i);2(Um+l

At
<U0,

V>

<Uo,

V>,

where WmV+
1, T7+ and Umr+i are in m/.

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

+ Um), V)=

0,

.,

Ve-/

V c-

GALERKIN METHODS FOR PARABOLIC EQUATIONS

597

THEOREM
5.3. Let u be the solution to (2.2)for 0 ? t < T. Let U be the solution
to (5.11). Let z = u - U. Suppose that u satisfies the hypotheses of Theorem 5.2.
Then the conclusion of Theorem 4.1 holds.
Proof. Let vm = urn- Tm.It is clearly sufficient to show
Vm+i

(5.12)

<L

C[IZm11L2+

(U

+ C2(At)4],

2)m+i12||Ho

since we can use (5.12) just as we used (5.6) in the proof of Theorem 5.1. In order
to show (5.12)just use the argument used to show (5.10) twice.
The proof of Theorem 5.2 will carry over to the case of 0 E (0, 1] in (2.7).
THEOREM
5.4. Let u be as in Theorem 4.2 and let Umsatisfy (2.7)for somefixed
0 E [0, 1]. Let z = u - U. Then the conclusion of Theorem 4.2 holds.
The value of 0 can be made to depend on time without changing the conclusion
of Theorem 5.4, but it is doubtful that this generalization is of any particular
practical interest.
Proof Let ym = Um - Wm.The proof consists of showing that
IIyM+ IL2 < C

(5.13)

IZmi1L2 +

O((At)3)+ C I(U -

U)M O||2

and then using (5.13) in a computation that exactly parallels the proof of Theorem
4.2. In order to show (5.13) we first write (2.2) at t = {(1 + 0)tm+1 + (1- 0)tm as
(5.14)

Um+
PMrv\ + a(um+ ?xm;um,o+

'uM+,

M V)= 0,

veH 1,

where now Pm'~Xmand Cmare O(C2At). Write fm,O= 2(1 + 0)yM+1 + 2(1-0)Zm
= (1 + 0)Wm+1 + 1(1 - 0)Um. Use (5.14) and the first equation in (2.7)
and W,m0o
to see that

11Zm11L2] + p1IIf O11H2


~~~~Hi
L
0

[IIyM+ 1 11L2

2At

0
-<

t [ 112Ym+1L2 -1

2At

2A01YL

Zm12
L2] +

+At

1Ym+11122
L2

2+ 11
Zm112
r L2

(5.15)
-

At

2<ym+1 zm>} + a(Um;

Lm!,0o
fm,o)

+ a(Um;zm,o, zm,o).

,
XZrnO )

Using (5.15) and the computation used to derive (5.10), we see that
(5.1)
(5.16)

+lL2
1

<

C{ Zm L2 +

l(u -

17)m+1/2

+ (At)[om

|2L

2Hl

L2

IInm Ho]>

This relation is used in modifying the proof of Theorem 4.2 to show the conclusion
in exactly the same way that (5.6) is used in modifying the proof of Theorem 4.1.
6. A priori estimate for Crank-Nicolson extrapolation approximation. While
the predictor-corrector Crank-Nicolson-Galerkin difference equation maintains
second order accuracy in At and requires only the solution of linear algebraic
systems of equations, carrying out the procedure requires twice as much work
as the solution of Crank-Nicolson-Galerkin in the linear case for each time step.

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

598

JIM DOUGLAS, JR., AND TODD DUPONT

In this section we shall show that (2.8) maintains second order accuracy in At
while requiring only half the work of (2.7) per time step. A word of caution is
necessary here. Since the error estimates involve constants that are related to the
solution of (2.2), it can easily be the case that more than twice as many time steps
could be required by the extrapolation method than by the predictor-corrector
method to obtain the same accuracy. This is qualitatively apparent in problems
for which a region of rapidly changing values of u is being translated in time. We
shall also show that (2.9) maintains first order accuracy in At.
THEOREM6.1. Let u be as in Theorem4.1. Let U satisfy (2.8), and let z = u - U.
Then there exist positive constants C, 61, co (determinedby the quantities listed in
Theorem4.1) such thatfor At < -co,
M-1
[M-1
1
||ZMII2 + 61
1Zm+i12jHAt < C
12(U iA)m+ii2iL\t
-

M=o
Y
Ilm-11(u

-o

Proof. For m ? 1 write (2.2) at


Um+ 1t

17)m112 2

m- At/

At

+ ll(U -

(6.2)

9)m 112 (U

+E

(6.1)

_M=o(U

iYm+/

tm+ 1/2

+ a(

+ Pm9V)

At ++ ll(U

+ I(U

1U)M- 1/2i2

ii)0112
L2

+ C2(At)4]

z)3,2I2

as

um - 2-Um-1 + Om; Um+ 1/2 +

m, v) =,

v eHo
Using (6.2) and (2.8), we see that for m ? 1,

where Pm,9mand Smare O(C2(At)2).


1

t[|Z+l||L2
2At

L]ZmL21

-<KZM+
=

{ <

Zm

+ a(Zum
-Umi Z2)
+mUm+l/2
(4Um

(6.3)

2Um-1;

+ a(4Um
-

a(3um

a(4Um

Z+

? 8iZm+U/2
+

l/2)

-2
-

Um- 1; Zm+ 1/2, Zm+ 1/2)

U)m+2 2/

+ 'm,(u

PM
Um+ (

U2)}
u

U)m+ii2)

um-1 + 1m; um+ 1/2 + Sm, (1


2Umi;t(m,(u

U)m+

AZt2 + I(ZU 1l2m +

U)m+ 1/21Ii

Z+a1

U)m+ 1/2)

KPm,

1/2)-

ui)m+1/2) + a(1Um

+ m C[j9(u

U)m+

U)m+i

+ 'm,(U~

2Umi1;um+ii2

Zm ,(U-l

a(4Um

+ Pm,( (

\At

HI

IllZm+1i2L

U)m+ii2>

1UmI;zm+i22(u

U)Pmm2+

12

Zm(U - u)m+ 1/2)1

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

-)m+ii2)
hmiL2

METHODS FOR PARABOLIC

GALERKIN

599

EQUATIONS

The summation of (6.3) is carried out in a manner quite similar to that of (4.6).
First we conclude from (6.3) that for some C3 and some y > 0,
(1

(1 + C3At) Zm
II2
+ CAt[lZm+ 12 + WZmW1j2 |Zm-||L2]

C3At)|Izm+I1I|22

+ YAtIIZm+1/2Hi

(6.4)

?<AtC[11(u-

21)m+iI2Wi1i

11Pm 122 +

KlCmWII

I'9mWIL2]

+ + 2Atzm ~A
1-zm, (U )m+ 1/2
Next we divide (6.4) by 1 + C3At and then multiply by g(At)m= [(1 - C3At)
(1 + C3At)-1]m to see that
1

g(At)m

IZm+

L2 - g(At)
11

11Zm1122

CAt112

(6.5)

?< CAt[II(u -

1 ZmL2
12

1L2

C3Atg(At)m[|zm+

U7)m+ii2W o +

2At g(At)m /ZM+


1 + C3At

1Z
1212A
+
ZmlL2]

PmL2

Z+1/2

HomWl+ h9mWL2]

am12

At

In deriving (6.5) we have used the boundedness of g(At)'"above and below by


positive numbers independent of At for small At. We now sum (6.5) from m = 1
to m = M - 1 and perform the usual summation by parts on the last term on the
, M - 2 arising
right-hand side. Note that the coefficient of 11Zm
122 for m = 2, *
from the sum of the bracketed terms on the left-hand side is
(6.6)

CAtA1
?

1[1 + g(At)

(t)m

g(At)2]

C4At

for some C4 > 0, C4 independent of At for At small. Hence, we may conclude that
M-1
IIZM1|22?+ 61

12lAt

IIZm+

m= I

< c [|(U

M-

(6.7)

>

i)1L22

{ (U

ll(U -

a)m+12IH

a)0112

IIPmIL2

?+

Km112

h9mflLj2}At

m= 1

(U -

M-1

a)m+1/2

(U

a)m1/2

A1t +

|(U -

)M/22

At

m=2

K1(U-

a)l+1/2

Using the proof of Theorem 5.2 we can easily see that

(6.8)

IIZ1II2

yAt

iz

H
12L

<

C[IIZO1I22

1(u -

i)112HAt

+ C2(At)4].

If (6.8) is used with (6.7), the conclusion (6.1) follows immediately.

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

600

JIM DOUGLAS, JR., AND TODD DUPONT

THEOREM6.2. Let u be as in Theorem 4.2. Let U satisfy (2.9) and let z = u - U.


Then there exist positive constants C, 51, z0 (determinedby the quantities listed in
Theorem4.2) such thatfor At < z0,
M-1

ZM 22 +

Z17H01At
ZM

51
m= 0
M-1

?CI3

M-1

H2At +

lt(UL)m,o

Lm=0o=

(U)m,0(U-

+ ll(U-_I)012 + (U - i7)0011L2+ I(U-

At

At

At

)m-1

U)M-l,O 1L2

C2(At)2].

The proof of Theorem 6.2 is a simplification of the proof of Theorem 5.2


and will be omitted.
7. Extensions to more general second orderparabolicequations. The arguments
of the previous sections can be applied in more general situations than we have
discussed thus far. In this section we shall show how these arguments can be used
to derive estimates in the case of more general second order parabolic equations.
Suppose that A(x, t, u, p) (A1(x, t, u, p), * An(x,t, u, p)) is a measurable
function from Q x [0, T] x R x Rninto Rn. Suppose that the matrix
,

OA (aAi
P
ap
pj/
is symmetric and positive definite with its spectrum bounded above and below by
positive numbers COand Cj,respectively, which are independent of the arguments
(x, t, u, p). Letf(x, t, u, p) be a measurable real-valued function of its arguments for
(x, t, u, p) EcQ x [0, T] x R x Rn. Further suppose that there exists a constant
, n, then A(x, t, w(x), p(x)) and
C such that, if w and pi are in L2(Q) for i = 1,
in
and
are
L2(Q)
p(x))
t,
w(x),
f(x,
(7.1) IA(x,t, w(x),

p(x)) ||L2 +

If(x,

+I p(X)IL2 + 1],
t, w(x), p(X))IL2 < C[IIW(X)11L2

for each t E [0, T]. (This assumption will be used in demonstrating that the CrankNicolson-Galerkin procedure possesses a solution at each time step.) Let
uo E L2(Q

Let u(x, t) be defined on Q x [0, T] with u(x, 0) = uo. Suppose that u, au/at
and Vxu = (au/ax1, u, xu/,x) are in L2(Q x (0, T)). Also suppose that, for
each t E (0, T), u E Ho(Q) and satisfies

au

a' v

(7.2)
where

f(u)
a(u, v) = f

v E H'(Q),

+ a(u, v) = Kf(u), v>,


f(x, t, u(x, t), VXu(x,t)),

A(x, t, u(x, t), Vxu(x, t)) Vxv(x) dx

(7.3)

=E
Zi j Ai(x,

t, u(x, t), Vxu(x, t)) -(x) dx.

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

a~~~~~~xj

601

GALERKIN METHODS FOR PARABOLIC EQUATIONS

We can define the continuous time Galerkin approximation by

a V
atV
(7.4)

+ a(U,V) = <f(U),V>,

< U, Y> =

0, V>,

VeJI,

t> 0,

V c,

t=O0,

where U is of the form (2.3). If A and f have sufficient regularity, then (7.4) will
have a solution U. Similarly we can define a Crank-Nicolson-Galerkin approximation scheme by
Um

At

Urm,

+ a(Umr+ 1/2 V)

<f(Um+

1/2), V>,

VJ,(

(7.5)
<U0, V>

m > 0,
Ve X.

<u0, V>,

In (7.5), t should be evaluated at (m + 2) At in both a(U, V) andf(U). It is clear


that (7.5) can be generalized to an analogue of (2.6), but we shall not do so here.
We shall show using a simple fixed-point argument that (7.5) has a solution
for At sufficiently small (independent of the space X#). We shall then indicate
how Theorems 3.1 and 4.1 can be generalized to include equations of the form
(7.2).
Let
a'(w, y, u, v)

(7.6)

= T

gA(x, t, l(w + y), {V(w + sy))ds(lV(u-y)(x))j

[T

Vxv(x)dx.

LEMMA 7.1. Suppose that, for each U, V and W in X, a('(W + Y), V),
a'(W, Y, U, V) and <f(f(W + Y)), V> are continuous as functions of Ye JI.
Thenfor At sufficientlysmall (7.5) has a solution.
Proof. Define W = JT(Y) as a map of X into ,X given by

W-

(7.7)

Urn V

At

+ a(2(Um + Y), V) + a'(Um, Y, W, V)


=

f(4(Um + Y)), V>,

V ES.

It is easily seen that such a W exists for each Y by noting that (7.7) is equivalent
to an equation of the form
(C + AtB)W= G,
where C and B are positive definite operators. In view of the hypotheses made on
a, a' andf, it is clear that 9- is continuous. It is also clear that if W is a fixed point
of Y then Um+I = W is a solution to (7.5).
Using V W in (7.7), we see that
WL2

(7.8)

+ At[a({(Y + Um), W) + a'(UM,Y, W, W)]


<{At)211f(UrT

Y))112

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

iIW11[i2

+ 211

602

JIM DOUGLAS, JR., AND TODD DUPONT

Note that
a({(Y +

Um), W)

A(x, t, (Y + Um), VxUm)SVxWdx


+ a'(Um, Y, 2Y, W).

Now,
a(Um,

Y, W, W) + a'(Um, Y, 2Y, W)

= JQLJOga(xt 1(Um

+ Y)s 2vx(u

+ sY))ds{VxW]

VxW(x)dx.

Since the spectrum of aA/apis bounded below by j, we conclude that


211W |2+

1A(x,t,(Y+

W |_At

t|

VxW11L
Um),2VxUm)11L21

12
+ (At)2 lf(1(Um + Y))IL2 + 2 Umr

<

?1

+ CAtjjA(x, t, l(Um + Y), VxU )

AtlWl|

+ (At)2 f(f(UUm + Y))f L2 +

L2

211
UMrn2.

From (7.1) and (7.10) we see that for At sufficiently small there is some ball in /
which is mapped into itself by iT. Hence by Brouwer's fixed-point theorem iT
has a fixed point.
In order to derive a priori error estimates for (7.4) and (7.5) we shall need the
following regularity hypotheses on A and f. Suppose that f = f1 + f2 and there
exists Ko such that for (x, t) e-Q x (0, T), r and s in R, and p and q in R',
(7.1la)

IA(x, t, r, Vxu(x, t)) - A(x, t, s, Vxu(x, t))#< Kor

(7.1 ib)

lJf(x, t, u(x, t), p) - f1(x, t, u(x, t), q)l

(7.1lc)

IAl(x,t, r, p) - J(x, t, s,p)l < Kor -sl,

(7.1ld)

Ij2(x, t, r, Vxu(x, t)) -f2(x,

(7.1Ie)

lJ (x, t, r, p) -f2(x,

sl,

Kolp -ql,

t, s, Vxu(x,t)) < Kolr - sl,

t, r, q)l < Kolp -ql.

There exist constants C and 3 > 0, whichdependon T, n, diam Q2,


such
that,for u and U solutions to (7.2) and (7.4), respectively, and u7
and
CO,
Ko, q
anyfunction of theform (2.3),
THEOREM7.1.

IZL2XL-

HlIZIiXL2

(7.12)
_

IIU
CL~~u
-u122

a
-

uIInAxL2 2 +

-(U
at
~~

2
-

U)

L2xL2j

1
],

u - U. (See (3.6) for notation.)


Let u be the solution to (7.2) and let U be a solution to (7.5).
Suppose thatfor each x e Q2,au/atand Vxuare twice continuouslydifferentiablewith
respect to t, and suppose that 03u(x, t)/at3 and 03u(x, t)/(at2axi) are boundedby C2
, n. Then there exist constants C, 3 > O,z0 > 0,
for (x, t) e-Q x (0, T) and i = 1,
where z

THEOREM 7.2.

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

GALERKIN

METHODS FOR PARABOLIC

603

EQUATIONS

whichdepend on T, n, diam Q2,Ko, q and COsuch that for At < zo,


M-1

+
11ZM11L2

IJZm+12IIkiAt

m=O

2tAt
E II(u- i7)m? A2II

_ C[

(7.13)

m=O

II(U -

i7)01122

II(U

(U -

M-1

2)M-l 2IL2

At

||(U

U)1/2II

(U -

Ut)m1/2

L2

C2(At)j4

where T = MAt, Umrcn/e/ m = 0,


, M.
The proofs of the above results are very similar to the proofs of Theorems
3.1 and 4.1. We shall write out the fundamental estimate in the case of Theorem 7.1
and leave the rest to the reader. It is easy to see that
2

--(Wu-u2)
2 dt(
llZ 12)
++ q llZI|
L
2 dt
o

<

au

-)

+J [A(x, t, U, V u)

(,

A(x, t, U, VxU)] Vxz(x)dx

az
+

at'

U-U

[A(x, t, U, Vxu) - A(x, t, u, Vxu)]- Vx(-

U)dx

az
+

at

f[A(x, t, U, Vxu) -

A(x, t, U, V, U)] V,x(u- i7)dx


.

(7.14)Q

{Kf(U)

f(U), ii

U>}

[A(x, t, U, Vxu) - A(x, t, u, Vxu)]- Vx(

U) dx

az
at

[A(x, t, U, Vxu) - A(x, t, U, VxU)] Vx(u .

i) dx

_ C[KO[IIZIIL2
+ IIZIIHK111U
- U[1L2 + KOIIZIIL2[11ZIIHO
+ IIU
+

<<_

COI|ZIIL2I[U
IZ112, +
+
Hzj~

az
1u1HI] +

a<t U -

_ U|112 +
CIIZI1z2
+
jzj2 + CIIU
CJ-u[HA

U
,z
- u7
-,u

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

-uIIH1H

604

JIM DOUGLAS, JR., AND TODD DUPONT

In deriving (7.14) we have used the uniform positive definiteness of aA/ap and
have applied (7.11) several times. It is clear that the conclusion follows from (7.14)
by the argument used to prove Theorem 3.1 since (7.14) implies (3.14) which gives
the conclusion when properly integrated.
The predictor-corrector and extrapolation schemes (2.7), (2.8) and (2.9) can
be formulated in a rather general setting. However, in order to derive an a priori
estimate by a reasonably direct extension of the arguments of ?? 5 and 6 we
shall have to be much more restrictive than in the first part of this section. Suppose
that

A(x, t, u, Vxu) = -(x, t, u)Vxu+ e(x, t, u)

(7.15)

and thatf(x, t, u, Vxu) f(x, t, u). We shall suppose that sl, e andf are measurable
functions of their arguments which are uniformly Lipschitz continuous in u with
Lipschitz constant K1. The coefficient a? is an n x n symmetric positive definite
matrix such that its spectrum is uniformly bounded above and below by positive
numbers COand C, respectively. The function 6 clearly has values in Rn. Let
a(w; u, v) =

[s(x, t, w(x))Vxu(x)]*Vxv(x)dx,

(7.16)
b(u, v) = J(x,

t, u(x)). Vxv(x)dx.

The forms a and b can have a dependence on t, but we shall suppress writing it out.
Define UmEX for each tmby
(7.17a)

mYfl+l

At2

Ur, V

+ a(Um;}(Um + Wm+,), V) + b(Um, V)


<f (Um),

(7.17b)

U t+i

Urn V

+ a(4(Um +WmJ+i);

-<f

(7.17c)

(2(UM

V>,

Umn+1/2, V)+

Wm+ 1)), V>,

<U0, V> = <u0, V>,

(2(U?Wm+i),V)

V E 1#1

Ve . X/.

where Wm
lI e X and f(w) denotesf(x, t, w(x)). In (7.17), t = tm+ 1/2 in a, b andf
We shall call U the predictor-corrector Crank-Nicolson approximation.
THEOREM
7.3. Let u be the solution to (7.2) and let U be the solution to (7.17).
=
Let z u U. Suppose that d3u/(at2axi),03u/at3and all the second derivatives oJ
u are continuousandboundedby C2. Supposethat a,4/au, 02S/au2, a2 A/(Oxiau).
are continuousforeach t andboundedby C'2. Then
8el/U,
a2/laU2 and a26/(axjau)
5
>
-r
constants
exist
there
C,
0, > 0, whichdependon T,n, diam Q, K1 fIVXUIILOO
XLI-,

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

GALERKIN

METHODS FOR PARABOLIC

605

EQUATIONS

and CO, such that for At < co,

'

M-1

+ (

||ZM||2

At

m= 1
M -- I

< CF

EIl(u Ut)m

mn=O

(7.8)

2
i/211 At

a ___

)m?12

At ? KlU-)12L

I~

1t

in =1

+ II(U -

U)o I 22 +

where T MAt,hmeKm = 0,
Proof. Write (7.2) at t =tm+
ways:
At

1/2

,M.
using the forms a and b in the following two
+ a(un1+

+ p,1,

U)M- 1/2 112 + (C'2 + 1)4C2(At)4{

II(U

0m; Um+ t/2

'm,t)

b(um

Cxm,V)

(7.19a)

I
Utti~~~~~~~~~
A ?-

<. (lJ
?(wtV\+ Va,
(u(

+ a(uin+1/2

Pm'

ii

(7.1l9b)

+ b(um+12 +

+ "mMUm + 1,2 + K,m,t)

Kf(Umfl+1/2+? m),V>I

<m,t')=

veH'O

where Pmand twm


are O(C2(At)2)
and xmis O(C2At). Let yn, = Urm- Wm.We shall
? Wm+i)
use the notation 3m +1/2 =-(
+ y
2()and Wm?112 =(U7
Using (7.17a) and (7.19a) we see that
2t

[WJYm+11112

Zm11L2]

?(--i utIX
At

12m+12

m+ 1/2

IHo

m+

;Y;m++

1/2

in

1/2)

I
m + Pm 0( -

=At
+ a(um

+ a(Un1;
-

-<

ln+

Pm,((-1
Im+

{b(U 1 (U--L)m
+

1/2

b(UmJ ,l+

+ b(um,Jfm+ii2)-

Cm (I

a(UmW;

(1i

+l/2

W)m?l4 1/2)j

+ 1/2)
+

W)m+I/2 >
1/2'

W)m + 1/2) -

+
?m

a(um+ m; lmI/12
+

+ a(U,;
?

1/2 +

?(m;Um+

W)m + 112

m
+

W)m+ 1/2)

m lUn -

1/2)

+ ome (k

1/2)-b(Uim

a(Um; m, (a

At

i)m+
n

1/2) -b(um,

-u

M+ 1/2)

Y l+ 112)

b(u01 +?

?m'Ym+1/2)

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

1 2

W)m+ 1i12)

606

JIM DOUGLAS, JR., AND TODD DUPONT

+ Kf(Um+ ?rM) f(Um), (a


+

+ Cm
r(a

a(Um;Umr+1/2

+ a(Um; um+1/2

W)m+1/2>}

U)m+1/2) a(Um + Lm;Um++1/2+ Cm(


a(um; um+1/2 + CmIYm+

Cm, Yrm+1/2)

+ a(um; Um+ 1/2 +


(7.20)

1/2)

a(um + Otm;Um+ 1/2 + Cm, Ym+ 1/2)

Cm, Ym+ 1/2)

- U)m+1/211HI + IlYm+1/2IIH4
+ C0IICmIIH1JI1(U
+

IlPmIIL2[11(U "U)m+l/2IIL2

+ yI[IIYm++1IL2

IYrm+11211L2]
-

IlZmIIL2]I(U

COIIYm+1/21H1I1(U

? Kl{Izm

tOm+12IIL2

-i)r+1/211H1

2IL211(U

)m+1211HI

- W)m +1/21IL2}
IZm + CamIIL21I(u

Kl|1VxUIILR xLoIlZm +

+ K
+

IYVm+1/2ItH1] +

amIIL2I1(U- )m?+1/211NO

IIPmIIL2[ll(U

+ a(um;Um+ 1/2 +

)m+112fIL2

lIZmIIL2]Jt(U
-

+ COiYm+ii21111I(U

+ b(umr
nm+1/2)

-b(um

IIm1121+
tmr+ 1/2,

-4?(x,

a(um + cc; Um+1/2 +

tmr

1)4C2(At)4

CmnYrm+n1/2)

i)r+1/2iIks)1
+ 2tl1IIY+1/2IHI

rL2At)
II?(M

Um)

l/2,

Um + (Xrm)]Vx(Um +

Xm))lmY+ 1/2

dx

b(x ,tm+ 1i2, Um + Ymr)])Ym+


1/2 dX

< ~[ilyvr+ii2~
+ || (U'
(lZlL2
+112
At 3 |Y+1|L2 ++ C(IIZm112
+

IIZmfIL2]

LXmrYrm+l/2)

(Vx 4[b(x, t+ 1/2 um) -

+ C(C2

IlYm+11IL2 +

)m+1/21H,0

a)m+r1/21IH6

'mrnYm+n1/2) -

[Vx * ([1(X,

)mr+1/2IIL2

t [ij'rn+ Iii22+ C(IlZmII12


+ 1j(u+ C(IiPmrnI22+

IIZm.IL2IIYm+1/2IIHI

IIZmIIL21Y1M+1/211I16 + COI1CmrIHj1II(U

IIVxUIIL XL

+ AtIIYm+1IIL2+

O)m+ 1f2 IHd)

ThIIYm +i2IIHi

It followsfrom(7.20)that
(7.21)

IIYm+rI
1I2 < C[IIZMrnj2+ ||(U

U)r

1/2II6 + (C2 + 1)4(At)4].

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

U)m+1/2)

607

GALERKIN METHODS FOR PARABOLIC EQUATIONS

Using (7.17b) and (7.19b), we see that


1

11 Z.+n1i2 Ho
L
IIZrnI2]
L
2At[IZ +1 1L2
< {b(Wm+1/2, (a - U)m+ 1/2) - b(umr+1/2+
+ Kf(Wmr+ 1/2)

ff(Um+ 1/2 + Cm), (

Cmr,(a

1/2)

U)m+

U)m+ I/2>}

+ a(Wm+1/2;um+1/2 + Cm,(u - U)m+1/2)

(7.22)

a(um+ 1/2 +

a(Wm+1/2;;Cm, (1u

<{1

+l(u fZy

h)m+?1/2?112

Zm

Zm+r1

(U -

1/2)

<Pm, (a -

U)m+ 1/2>
-i)m

1/2)

+ IlPmIIL2
KrmnIHA

|(U

Zm

m+1A

+ IlYm+r1L2] +

+ C[IlZrnlL2

11I2j1H

U)m+

+ a(Vm'+ l/2; Zm+ l/2 (U

n/2

112IJA + C[IlZmII22 +

lZrnl+

Cm,

U)m+ 1/2) -

U)rm
(U -

Zmn1

<-{1J

m;Umr+nl/2 +

U)rn?ii21Ho

)ri

+ (C2 +

1)4C2(At)4]

U)m+1/2

The conclusion follows from (7.22) as in the proof of Theorem 4.1.


THEOREM7.4. Let u be as in Theorem 7.2. Suppose in addition that 02u/at2 is
bounded by C2 and that 11VxU
1L- >L- is finite. Let U be a solution of (7.17) and let
z = u - U. Then the conclusion of Theorem 7.3 holds with (C'2+ 1)4(At)4replaced
by C2(At)3.
This theorem is proved by modifying the derivation of (7.21) in the same way
that Theorem 5.2 is proved by modifying the derivation of (5.6).
defined by
Consider Umi+1 E
Wm+At

Um V

+ a(Um;'(Wm
I4K

+ Um), V) + b(Um, V)

(7.23a)
< f(Urn), V>,
7

Urn

Atml Um,V

+ a('(Um

+ Wm+V+);(Tm7+

+ Um),V)

(7.23b)
+ b(4(Wm+?

Am+t

Um V

+ Um), V) = <f(4(Um

+ a({(TMI+1
+ Um); Umr+1/2,V) + b(P(Tm+ + Um), V)

(7.23c)
(7.23d)

+ Wrm+i)), V>,

f (21
(Um +

+1))I V),
Tm

<U0, V> = <uo, V>,

where (7.23a) through (7.23d) are required to hold for all Ve I and Wmand Tm
are in X.

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

608

JR., AND TODD

JIM DOUGLAS,

DUPONT

to (7.23) anidlet
THEOREM 7.5. Let u be as in Theorem 7.4. Let U be a sollutiont
z = u - U. Then the conclusion of Theorem 7.2 holds.
The proof of Theorem 7.5 is parallel to that of Theorem 5.3 and will be omitted.
It is clear that we can extend Theorem 6.1 in precisely the same way that we
have extended Theorems 5.1, 5.2 and 5.3. Consider U,m,+
1 defined by
IAm+t

V+ +a(Um-{Um;

(7.24)

Um+i2

V) + b(3Urn{Urni
VE,

Urni),V>,

<Kf(2Urn

V)

in> 1,

with UOand U1 defined by (7.17).


THEOREM 7.6. Let U be defined by (7.24) and let u be as in Theorem 7.4. Let
z = u - U. Then there exist constants C, 6 > 0, z0 > 0 determined by the quantities listed in Theorem 7.3 such that, for At < To, (6.1) holds.
We now return to considerations of the more general operators introduced in
the first part of this section. It is worth noting that a class of second order parabolic
equations which are of apparently more general form than (7.2) can be transformed
into a class of the form (7.2). Specifically, consider
(7.25)

$p(x,t, u) = V A(x, t, u, V.u) + f (x, t, u, Vxu),

where Spu= hpo/auis bounded above and below by positive numbers which are
independent of the arguments. Suppose u(x, t) E Ho(Q) for t fixed. If Soand u are
sufficiently smooth, (7.25) is equivalent to
<ut, V> +

j(x, t,

u) - 'A(x, t, u, Vxu) VxV(x) dx

(7.26)

where
(7.27)

[f

V
(X, t, u)).
(pt]- A Vx((p

Equation (7.26) is of the same type as (7.2), provided that Sp,A andfare sufficiently
regular. This transformation can be easy to carry out in practice (say if (p(x,t, u)
= q5(x)u) or quite difficult. Note that in the case (p(x,t, u) = (7(x)u where
0 < 1 < -< 62, the arguments we have used repeatedly can be used without
transforming the equation.
There is another direction in which the arguments we have used here can be
extended. We shall mention this extension but shall not carry it out in detail here.
We have always assumed the special part of the equation Vx A(x, t, u, Vxu)to be
uniformly elliptic. A certain type of degeneracy can be allowed without fundamental
change in the arguments. Suppose that
.

A(x, t, w, p) =

ao(x, t)al(x,

w)p,

where ao(x, t) is a bounded, nonnegative, continuous multiple of the identity


matrix; and suppose that a1(x, w) is an n x n matrix which is uniformly bounded

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

GALERKIN METHODS FOR PARABOLIC EQUATIONS

609

and positive definite. Then the arguments of this section can be carried out using
lIwll t

[ao(x, t)Vxw(x)] Vxw(x)dx

as a substitute for IWll2 ,.The boundary conditions may have to be changed since
we can no longer require that u = 0 on MQ,even in some generalized sense such
as u E H', if a(x, t) vanishes on MQ.This modification of the boundary condition
is related to ideas used in ? 8.
Extension of most of the results of ?? 2 through 7 to parabolic equations with
higher order operators in the x variables has been carried out by H. Meyer [18].
8. Other boundary conditions. In the previous sections we have used the
condition u = 0 on MQfor t > 0. We shall indicate in this section that the arguments apply as well if homogeneous boundary conditions u= 0 on F1 and
=
and Oula/vis the derivative
au/av= 0 on F2, t> 0, are used, where F1 U F2 -Q
of u in the direction conormal to MQ.Inhomogeneous boundary conditions can be
used by transforming the differential problem, provided the data is smooth and
that the boundary condition is linear.
Let Q F1 U t'2, F1 A F2 = 0. Define Hjl,(Q) to be the closure of the
set of the C"(R')-functions which vanish in a neighborhood of F1 with respect to
the norm
=

lu12

(8.1)

= 11U
1122(n)?

i=1

ax
ai

2
L2(nif

We shall use the (semi-)norm

aXi lL2(n)
~~~~~~~~~i=1C

(8.2)
on functions in these H'r1(Q)spaces.
Consider the solution u to

au -

at-V,,

(8.3)

u(x) = 0,

(a(x, u)Vxu)=f (x, t),

x EFI;

au(x)

o,

EF 2,

U(X,0) = UO.

If F1 and F2 are sufficiently smooth and u is sufficiently regular, then (8.3) is


equivalent to
(8.4)

?
+

a(x,u)Vxu VXv(x)dx = Kf(x,t),v>,

veH1c1H

Since (8.4) is more general than (8.3), it suffices to be able to approximate solutions
to (8.4). More generally, for sufficiently regular functions and Fi the following

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

610

JIM DOUGLAS,

JR., AND TODD DUPONT

two problems are equivalent:


Vx A(x, t, u, Vxu)

a_

(8.5)

u=0

onFT1;

f (x, t, U, Vxu),

A(x,t,u,Vxu)

n= 0

on F2,

u(x, 0) = uO,

where n is the normal to OQ,and

au

(e

JC A(x,t,u,Vxu)

(8.6)

Vxvdx = <f(x,t,u,Vxu),v>,
(

ve

H1'1(),

.
u(x,0O) == uO0.

We shall continue to consider only those functions u such that u, Vxu and u, are
in L2(Q x (0, T)).
The various Galerkin approximations formulated in ? 2 and generalized in
? 7 can be defined here. Let X be a finite-dimensional subspace of H,1(Q). The
Galerkin approximations are formulated by requiring that the approximate
solution U and the test functions V lie in the space X#. Let v1,
, VN be a basis
for X.
The continuous time Galerkin approximation to problem (8.6) is a solution
of the initial value problem
au V

A(x, t, U, V U) - VxV dx= <f(x, t, U, V U), V>,

VE,

(8.7)
<U(0), V> = <uo0V>,

Ve.7/,

where
N

(8.8)

U(x, t) =

cX1(t)v1(x).

Notice that (8.6) is the same as (7.4); the only differenceis in the interpretation since
in (7.4), X c Hl instead of H1'. If we formulated the other Galerkin approximations, they also would read exactly the same as those of ? 7. Also all of the theorems
of ? 7, except Theorem 7.3, hold with only minor modifications when u is the solution to (8.6) and U is the corresponding Galerkin approximation. The only change
necessary to make the theorems of ? 7 valid for the approximations to (8.6) is that,
in the conclusions, 1(u - l)m+ 1/2W (o) should be replaced by 1(u - a)m+l/2 12I(Q)
(in the continuous time case IIU
-U kII212
U _
OkXL2 should be replaced by IIU
XL2
=
u- lLuxL2+ 1The
of
HxL2).
Theorem
7.3
does
not
proof
generalize
o
to these boundary conditions since we would have nonzero boundary terms that
result from the integrations by parts that were done in (7.20).
The inhomogeneous Dirichlet problem is most easily attacked by subtracting
out the boundary values by extending them to a function defined in all of Q x [0, T]
and changing the dependent variable. If the boundary values for u are given by

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

GALERKIN

METHODS FOR PARABOLIC

611

EQUATIONS

g(x, t), consider


(8.9a)

au v

Vvdx =<f(x,t,u,V.u),v>,

A(x,t,u,V.u)

t > 0,

v E Ho,

(8.9b)

<u(x, 0), v> = <uO,v>,

v E Ho,

(8.9c)

u(x, t) - g(x, t) E H'(Q),

t > 0.

It is assumed that g(x, t) E H'(Q) for each t > 0. Also we shall assume that u, g,
VXu, Vxg, u, and g, are in L2(Q x (0, T)). The condition (8.9c) is equivalent to
u(x, t) = g(x, t) for (x, t) E M x (0, T) if u and g are smooth. We shall approximate
u by approximating w = u - g using the procedures of ? 7. Notice that w satisfies
(8.10)

a
at

+
J

A(x, t, w, Vxw) Vxvdx = <f(x, t, w,Vxw), v>,


X

<w(x,0),v> = <wo,v>,

veH ,
veH'O,

where
A(x, t, r, p) = A(x, t, r + g(x, t), p + Vxg(x, t)),
(8.11)

J(x, t, r, p) = f(x, t, r + g(x, t), p + Vxg(x, t)) - ag(X, t),


wo(x) = uo(x) - g(x, 0).

We shall indicate that the functions A and f retain the properties of A and f which
we used in ? 7, provided A and f had these properties. Recall that we supposed in
? 7 that aA(x, t, r, p)/ap is symmetric and positive definite with its spectrum bounded
above and below by C0 and i, respectively. Note that aA(x, t, r, p)/p
= aA(x, t, r + g(x, t), p + V.g(x, t))/ap and hence also has these properties.
Similarly, it is easily seen that if the conditions given in (7.11) hold, then the corresponding conditions, obtained by using w instead of u, A instead of A andf instead
of f, hold. If A(x, t, u, Vxu) is of the form (7.15), then A is of the same form. If
1x L is finite, the additional properties of A andfassumed in ? 7 are retained
11Vx9g
by A and f (with different constants). We can conclude that if v satisfies the
hypotheses of Theorems 7.3, 7.4, 7.5 or 7.6 and w is as smooth as u is required to be
in the corresponding theorem, then we can successfully use the corresponding
predictor-corrector or extrapolation process to approximate w.
The technique just applied to solve the problem with inhomogeneous
Dirichlet-type boundary conditions for t > 0 may be applied to certain more
general problems. Suppose that F1 and F2 (see the first part of this section for
notation) are smooth and that u is a smooth solution to

au
_

Vx. A(x, t, u, VXu)= f(x, t, u, V u),

(8.12)
A(x,t,u,Vxu)

u = g1

on Fl,

t > 0,

n=g2

onF2,

t > 0,

u(x, 0) = uO,

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

t = 0.

612

JIM DOUGLAS,

JR., AND TODD DUPONT

If A(x, t, u, Vxu)is linear on F2 in the sense that on F2,


A(x, t, r, p) = V(x, t)p + b(x, t)r,
then (8.12) is equivalent to u = w + g, where
<?

at

- Vx A(x, t, w, Vxw)= f (x, t,Iw, V w),


t > O,

onF1,

w=O
(8.13)

A(x, t, w, Vxw) n = A(x, t, w, Vxw) n

0 on F2,

t > 0,

t = 0,

w(x, t) = Wo,
provided that there exists some smooth function g such that
g(x,t)=g1
n=g2

A(x,t,g,Vxg)

onF1,

t>O,

onF2,

t>0.

The function w can then be approximated by the procedure indicated in the first
part of this section.
9. Change of basis. Since in particular problems the character of the solution
u of a problem of the form (2.2) can change with t, it could be advantageous to
change the space X at times t = T1, **-, T,. The methods we have used can be
applied in this case to bound the error of the resulting approximation. (In the
discrete t cases take Tp= k,At, kp integer.)
Let To = 0. Suppose that for t in the interval [Ti_ 1, Ti], we are using a finitedimensional function space X&i.Suppose that U'(Ti)is the calculated approximation at Tiin the space Xi. Let Ui+ l(Ti)be the L2-projection of U1(Ti)onto J/+?1 i.e.,
V EiIi+ 1

V>,
<Ui+?(Ti), V> = KUW(T1),

(9.1)

The function Ui+ 1(Ti)can then be used as initial values for restarting the Galerkin
process, either in continuous or discrete form. Note that
17, Ti) -

(9.2)

Ui+?l(Ti)1L2

<

JIu( ,

?< Iu( *,

Ti) -

Ui(Ti)L2

Ti)- U(1)L2

? 211u( , Ti) -

U(T1)DIL2

U(D)

Ui+ l(Ti)

+ 11OD

17i+
l(Ti)IIL2

+11

ll(u-i+

1L2

l)(T) 11L2,

where ui+ 1(Ti) is in Xi+ 1. If the arguments used previously are applied on the
intervals [Ti, Ti+7] and (9.2) is used to bound the error (u - U+1)(Ti), we can
conclude that the same types of error bounds obtain on each interval [T1,T+ 1]i.
Consequently, the error estimate for the full time interval [0, T] is increased by a
term of the form
-

xL2xLo.

It should also be noted that in the extrapolation procedures of the form (7.24),
the starting procedure should be used at each Ti to obtain the approximation at
Ti + At.

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

GALERKIN

METHODS FOR PARABOLIC

613

EQUATIONS

10. Parabolic systems. In this section we shall discuss the direct extensions
of the arguments of the previous sections to some parabolic systems. We shall
limit our discussion to the notationally simpler quasi-linear case, rather than discussing the more nonlinear forms. We shall discuss only homogeneous Dirichlettype boundary conditions for t > 0. There are many aspects of the theory of
parabolic systems which require much more detailed treatment than we shall
present here. In particular, see [10] for a discussion of the approximation of the
solutions of a degenerate parabolic system, arising in petroleum engineering,
by techniques similar to those used here.
We shall be dealing with vectors v = (vj,j) indexed by (l,j) e S, a finite set.
We shall also be considering matrices B = (bl,j;kji) operating on such vectors.
The notation Bv = f means that for (1,j) E S,

Z bl,j;k,iVk,i.

fij=

ksieS

One of the sets S which we shall use is S0 = {(l,j): 1 <I _ Li 1 < j _ n}. We
shall use two "inner products" on these vectors. Suppose that S = {(l,j): 1 ? I < L,
j E Sj}. Then for v = (vj,j)and w = (wl,j)define [v, w] to be the L-tuple given by
[V,W]

(10.1)

([V,W]1,

[V, w]1 =

*--

[,w]WL),

jeS1

Let
(10.2)

[[v,

WI]=

EVjW

j.

Consider the parabolic system


, at

u)Vxu] = 0,

[V,(x,
u=0

(10.3)

t > 0,
t > 0,

onaQ,

t = 0,

U = U0,
, uAx, t)). In (10.3), au/at = (au1/at,
where u = u(x, t) = (u1(x, t),
[Vx, V(x, u)Vxu]denotes an L-tuple whose Ith component is

(10.4)

ZLI
~ a
k= 1 i,j= 1 eXj

a1j;kj(x,u)

auL/at);

aUk
;
ax, 1

, L. We shall assume that the matrix


and u = u0 means uk(x, 0) = UO,k, k = 1,
v = (al,j;k,i) iS uniformly positive definite and bounded in the sense that there
exist tj > 0 and C0 such that for all 4 =

(10.5)

I2 <

where we have used the norm I

< Co g

al, j;k,iljXk

[[4,

4]]

1/2.

If u and v are sufficiently regular,

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

JIM DOUGLAS, JR., AND TODD DUPONT

614

(10.3) is equivalent to

e,v)

atI

u,

a(u;

t > 0, vc-Ho,

= o,

v)

t > 0,

u E-Ho1(Q),

(10.6)

u = a0,

t= 0,

where
<W, V> = (<W,

>1,

<W, V>L),

<W,V>1= fwl(x)vi(x) dx,

(10.7)

[a(x, w(x))Vxu(x),Vxv(x)]dx.

a(w; u, v) =

The notation u E Ho means each component is in H'(Q). We shall assume that


each component of u, Vxu and au/at is in L2(Q x (0, T)) and that u(x, 0) is the
natural extension of u to t = 0. The system (10.6) is more general than (10.3) so that
it suffices to approximate solutions to (10.6). In analogy with ??2-6 we shall
assume that a(x, u) is uniformly Lipschitz continuous in u with constant K in the
sense that for each x E Q, r, s E RL,
(10.8)

1s1(x, r) - V(x, s)I .< Kir - sl,

here is the matrix norm relative to the vector norm IuII= [[u, u]]'12
where
We shall approximate the solutions to (10.6)by requiring that the approximate
x ..x
...
solution U and test functions V lie in X =
XL, where Xfi is a finitedimensional subspace of Ho. Note that we are not requiring that each component
of u be approximated in the same finite-dimensional space. The continuous time
Galerkin approximation of (10.6) is defined as

au
at V
(10.9)

+ a(U; U, V) =0,

<U,V>=<uo,V>,

t > 0,

Ve k,

t=0,

Ve-'I,

where U E XWfor each t ? 0. Suppose that {I 11,' I. , 1N} is a basis for Xl. Let
, ,Jx)
v, j be the natural injection of v1,jinto X, in the sense that vl,j(x) = (0,
1 ?j ? Nl. If we write
0,.,
{l,j: 1 _ l_L,
0). Let S()=
.

(10.10)

U(x, t) = Ex#)v

JX),

S(X)

then (10.9) becomes


ax' + A(o)o = 0,
x0

(10.11)

t > 0,

cx(O) = xCO

where C and A(x) are linear operators on vectors x


If C = (Cl,j;k,i) and A(ox)= (AIj;k,i(J0)), then
(10.12a)
(10.12b)

Cl,j;k,i =

Kv1,j Vk,i),

Al1j;kJi(X) = <KKW(X,U(X))VXVk,i,

VxVIJ>,

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

(al,j), 1, e S(J1').

615

GALERKIN METHODS FOR PARABOLIC EQUATIONS

where U(x)

Zxr,rnvr,r

and we have used the notation


<v, w? =

(10.13)

[[v(x), w(x)]] dx.

(In (10.12a) the functions are vector-valued in a space in which the vectors are
indexed by a one-dimensional set; if this is considered a special case of twodimensional indexing, then (10.13) gives the correct formula for Cl,j;k,ji) Suppose
that W =
xlj,vi,j and V =
l,jv, j. Then

1W(X)12 dx

[[Cocc]]

(10.14)

[[V(x, V)VxW,VxW]]dx

[[A(fl3, a]j.

These relations make it clear that C and A(f) are positive definite. Hence, (10.11)
has a solution for all t > 0.
The Crank-Nicolson-Galerkin approximation of (10.6) is defined by
At

(10.15)

+a(Um+?12;Um+1/2,V)=0

<U0, V>

,x

Ku0, V>,

where Ume X, m = 0, 1,
, is the approximation to u( , mAt). It is easily seen
that under the assumptions made (10.15) has a solution for At > 0. (See the proof
of Lemma 7.1.)
We shall not formulate the various predictor-corrector and extrapolation
schemes; however, it is clear that they can be easily formulated for these systems.
For functions u(x, t) = (u1(x, t), * *, UL(x, t)) defined on Q x [0, T], let
(t) = IIUI2=
IIUIL2(Q

| U ||

21 (t)

u(x, t)12

E au,(x, t)

JU

1
1U[H1=

dx,
2

dx,

rT
=

||U|L2XL2

ILUII 2(Q x (O,T))

(10.16)

|U(X,t)f||dx

dt,

ul2
Uf1

OQSo

IIUIIL

SUp

0<t<

U112

|W
VXUWII
LT XLR =

(X t) dx dt,

Za

iX
|UIL2(Q),

sup

au

SUPt)~OX(0,T)
(x,t)eQ X (0,T)
So OXj

Za(x, t)

THEOREM 10.1. Suppose that u is the solution to (10.6) on Q x [0, T] and that U
is the solution to the continuous time Galerkin approximation(10.9)for 0 < t ? T.
Let z = u - U. Suppose that IVXUIILXXLO is finite if K of (10.8) cannot be chosen
to be zero. Then there exist constants C and 5 > 0, which depend on T, n, diam Q,

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

616

JR., AND TODD DUPONT

JIM DOUGLAS,

KIIVXU

XLX, ij

|LZ|L2xL+

and CO,such that

Z H1xL2

CIIU

U lHlxL2

IIU -

U'

L2XLX +

L211

2I

(10.17)
where iuis any vector-valuedfunction which is in -&for each t e [0, T].
ProofJThis proof parallels that of Theorem 3.1. Let the components of the
form a(u; w, v) of (10.7) be denoted by al(u; w, v); i.e.,
a(u; w, v) = (a1(u; w, v),

(10.18)

v)).

, aL(u; w,

The fundamental estimate for the proof is


--||

Z||2

+ I11 ZI12

faz z
+
L2
+-

V1019

al(U; z,z)

at U-

L
I
+ I= [al(U;
u, i

[al(u; u,i-

U) - al(u;u, i

az

tu u

U)] +

lU

z, u

+ Y a(;

1 _~
< KIVUIL
( 0.9)
X
? [(d(x,
{0} + f pV(x,

az

at'u

+
? K

)-H
U
a+(

|||L||

U)

u))Vxu,Vx(

u e+ COazH0u
Th -e

Vxu1 Le -xLce1on

,u||||

UlIH10

U)]}

U)]] dx
lu

H10

+ Ciil

ZH1011U

U(IIH10

z~
at~~~~~~~~a

? Hz~ +C

L~2 +

Hu UlA l

tu

The relation (10.19) is integrated in the same way the corresponding relation was
integrated in the proof of Theorem 3.1, and the conclusion follows.
THEOREM 10.2. Let u be as in Theorem 10.1 and suppose that for each x e-Q,
a3
au/at and Vxu are twice continuously differentiable in t. Let Ou(x,
t)/at3
1 and
to
be
a
solution
Let
e
be
bounded
T).
t)
for
Q(O,
by
C2
(x,
t)/at2
1a2VXU(X,
(10.15) for 0 ? mAt ? T. Let z = u - U. Then there exist constants C, 3 > 0,

Um

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

METHODS FOR PARABOLIC

GALERKIN

617

EQUATIONS

At < To,

To > 0 such thatfor


M-1
ZM

2 +

Z2

m=0
M- M-1

< c

+ (

k)n11?2At + II)(U

m = 0 11|

)+

U/2(_)_/_|/\

2 2 +

At

n)0ii2

12

C(A\t)4],

where iume - for 0 ? m ? M = T/At.


Proof. Write (10.6) at t = tm+ 1/2 as
(10.21)

+ a(um+

Pm,V

1/2

bm; Um+ 1/2

,m,V

0,

where pm and 'm are O(C2(At)2).The fundamental estimate for this proof will be
derived using (10.15) and (10.21):
+
L
Zm
L21

2At [Zm+1|L2
<

|Hl

rIIZm+1/2

1
-

-[IIZm+jII22

2At

L=l

ZTM?1
A(t\

IIZmIlL2]

aj(Um+1/2;Zm+?1/2Zm+1/2)

Zm+
PPmI@

/\t

U)m+

1/2

Z [al(um+ 1/2
-

m, (U -

+ Cm; Um+ 1/2 +

1/2; Um+ 1/2, (i

a(Um+

U)m+ 1/2)

U)m+ 1/2)]}

Y [al(Um+ 1/2;

Um+1/2 + Cm,(U

U)m+ 1/2)

al(um?+1/2 + Cm;Um+ 1/2 + Cm, (J


al(Um + 1/2 ;Cm, (a -

<PMI(a

U)m+ 1/2)

U)m + 1/2)]
ZM?i1

>+

Zm (u

At

U)m12?

)m-

+ 1/2

(10.22)

? {0}

Y al(Um+

1/2; Zm+1/2, (U

+ K11VxUIIL-XLx)

C0lSmllHlll(u

+ CmIL2L(-

Zm+1/2

H1 +

U)m+1/21

,(U -

a)m+ 1/2)

17)m+1/2

<= 2111 Zm+12 112i + C[lIZm12

U)m+1i2
-

IlPmIL21(u

H1

U)m+1/21IL2

+ C0I Zm+1/2 IH1 (U IIZm+l

122

1(u -

+ IIPmllL2 + ||Km|HAl]Zl+

)m+1121IH1

1)m211H
? 1
1m)m11/2

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

H (U

618

JIM DOUGLAS,

JR., AND TODD DUPONT

The relation (10.22)is summed, exactly as (4.6) was summed in the proof of Theorem
4.1, to yield (10.20).
It is clear from the proofs of Theorems 10.1 and 10.2 that essentially all the
results of ?? 3 through 9 can be generalized to systems related to (10.6).
11. Some computational considerations. In this section we shall consider
some modifications of the procedures defined in ? 2 which might be used in order
to carry out these processes numerically. In addition to formulating some of these
modifications we shall indicate their effects on the estimates of the error. The
considerations of this section are motivated by the presumption that the space X'
in which our approximation lies is a space of splines, i.e., smooth functions which
are piecewise polynomial. Many of the results in the theory of approximating
functions by splines are based on estimates of u - iu, where u is the function to be
approximated and a is a function in X which interpolates u (and perhaps some
of its derivatives) at a set of points in the domain of the functions. We shall consider
the error induced if the initial data is interpolated into X rather than projected
in using the L2 projection. We shall indicate a scheme which can be used to evaluate
approximately the coefficients in the finite-dimensional equations which determine
the approximate solution U. In particular we shall propose a scheme for approximately finding the coefficients in the linear equations which arise in the CrankNicolson predictor-corrector process (2.7) with 0 = 0. This is an important point
since a substantial portion of the total computing effort will be spent evaluating
these coefficients. We shall also suggest a modification of the Crank-Nicolson
predictor-corrector procedure which requires only one evaluation of the coefficients per time step. The solution of the linear systems of equations which arise
will be discussed briefly. We shall also mention an alternating direction Galerkintype process for approximately solving parabolic equations.
In each of the procedures we have discussed thus far we have taken
U0 = U( , 0) to be the L2 projection of uo = u( *, 0) into X/. The result of this
for any ae #, and we have used this
choice is that 1Uo- uOI1L2< It1- UOIIL2
fact to estimate U0-UO 1L2in the derivation of each of the estimates. If we choose
UOby some other process, such as interpolating uo, then each of the error estimates
we have derived remains valid if CitUo - Ui22 is added to the right-hand side
in the conclusion. If X is a spline subspace, it is well known [4] that smoothness
of uo implies that this term can be made small. In ? 9 we discussed changing the
space X from Ai to Mi, 1 at t = Ti. We proposed using the L2 projection of U'(T1)
(the approximation in Xi to u at t = Ti)into the space Mi, 1 to obtain the starting
value for U in the interval [Ti, 17?]. If we interpolate U'(T1)into Xi+ l to get the
starting values for U on [1T,Ti+ 1], then we can no longer use the argument (9.2)
to bound the initial error. Hence the use of the L2 projection in this case is essential
to the argument.
We have shown in ?? 5 and 7 that in order to approximate the solutions to
certain strongly nonlinear parabolic equations it suffices to solve linear algebraic
equations at each time step. However, unless the nonlinearities in the parabolic
equations are particularly simple we shall not be able to find the coefficients in
the linear algebraic equations exactly. As an illustration of how this difficulty may
be handled we shall derive an error estimate for a process which approximates
-

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

619

GALERKIN METHODS FOR PARABOLIC EQUATIONS

the Crank-Nicolson predictor-corrector process ((2.7) with 0 = 0). Let UM,


m = 0,1,
,be defined by

(Il.la)

+ a(UM;{+(4?

Atn+ Ur V

Vel,

+ UM),V)=0

At
(ill.b)

Urn+1Ur

V>

Urn?i2,

V)=0?,

<U0, V> = <U0, V>,

(11.lc)
where

Wr?i);

+i((Urn+

Veil,

Veil,

rn e# for m = 0, 1, ... We shall assume that


Wm,
Um

(11.2)

il(w;u,v)

iij(x,w(x)) ux) ( dx,

where

a(x, w) = (aijx(x,w))
(11.3)

= (I + y?)a(x, w)

= (I + 7y)(ai,j(x, w)).
In (11.3) we shall assume that the matrix 0 has norm bounded by one and that
a is positive definite. The motivation for considering Umdefined by (11.1) is that
the function a(x, w) = (ai,j(x, w)) can be interpolated into il', some spline space,
which allows us to evaluate by formula the integrals that need to be calculated.
(We are considering truncation error, not rounding error, so that we assume that
after interpolating a(x, w) into /#' the arithmetic is exact.)
THEOREM 11.1. Let u be the solution to (2.2)for 0 ? t _ T, and let Um
be the
solution to (11.1) for 0 < m ? M = T/At. Let z = u - U. Suppose that u and
a = (ai,j)satisfy the hypotheses of Theorem5.1. Then there exist constants C, 3 > 0,
T0 > 0 and Y0 > 0, which depend on T, n, diam Q, K VXU L- XLO, i and COsuch
thatfor At < T0 and 171_ 70Y
2 +
IIZM112

M-1
1Z2l

||

iiAt

rn0

M-1

? C E ||(U(11.4)

i)m+1/2 IHot

f
1M-1
(U

(U

k)r?12

+ E
rn=1

?
Proof Take y0 such that for 171

?-jECm
i=1

Usethenotationym = um -

+ 17(u
~~~~~~~At

At

+ II(U
- 2)112IfL2
+ II(u-

(11.5)

)mn12

U)M-/2I|2

112

+ (C2 + 1)4C2(At)4 +

721.

m0and r e R we have
<

a,jx, r)dj.
i,j

Wm,Jym
+?1/2 = 2(Ym+1 + Zm), Wm +1/2

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

2(Wm +I + UM)

620

JIM DOUGLAS, JR., AND TODD DUPONT

The first estimate will be much like (5.4). Using (5.3), (11.la) and (1 1.5) we see that
11
1L2

[/\t

2At

< 0Ym+

IZm L2] + 2111WYm+112


Hi

mYm+

1At
1

+ a(Um; Ym+

1/2

m PM
,(im+ 1W)?

At

a(Um;

W)m+1/2)-

W)m+ 1/2> + a(um; Um+ 1/2 +

Wm?+i1/2,(-

<Pm, (a -

a(um + cXm;Um+ 1/2 + Cm Ym+ 1/2) + a(um; Um+ 1/2 + 4,m(U

a(um + O(m;um+1/2 + Cm,(j7

a(um; Um+ 1/2 + Cm, (a

(11.6)

Ym+ 1/2)

+ a(um + cLm;Um+ 1/2 + lm,(f

1/2,

Ym+

Zm,(U -

U)m+112)

1L2

(U

U)m+1/2)

-m(-

W)m+1/2)
W)m+ 1/2)

W)m+ 1/2)
1)m+1/2)

12

+ C(WPM112 + KmWIH1
+ T2) + 41 IL1m+12
+ C(C'2

m, (i

+ a(Um;ym+1/2,(u

1)m?+/2

C(I1Zm

+ a(Um;Um+1/2 +

W)m+ 1/2)] - a(Um;

?
1[llym+1112
At
A

m gm+ 1/2)

W)m+ 1/2) + [a(Uni; Um+ 1/2 + Cm,(a


-

a(Um; Um+ 1/2 + Cm (i

W)m+1/2)}

Ho

1)4C2 (At)3.

In the last step of (1 1.6) we estimated the term in brackets by


1}l l Um+ 1/2 +

(11.7)

CmW1HOI
(U-

< C(u'2 + f|(u

W)m+ 1/2 IIHl

U)m? 1/21kA)

1/21A
11IIm?

where C depends on 11Um+1/2 + CmHl. The rest of the terms in (11.6) are estimated
exactly as they were in the proof of Theorem 5.1. Hence we conclude that

(11.8) ||Ym?+
1II2 < C[IIZpII72
+ 1(u-

2,)m+i2Ii + (C2 + 1)4C2(At)4+

y2Att].

Now use (4.3), (l1.lb) and (11.5) to see that


2At[WZm?1||L2

ZmL2] + 211Zm+ 1/211HI

<

+ a(2 m0+1 + Um); Zm+1/2, Zm+ 1/2)

, Zm+ 1/2)
Z+-Z

+ Pm,(

(I11.9)

-1

U)m +1i2

+ a(um?+1/2+

Cm;Um?lI12

At
+

(,m,(u

U)m+i12)

- ({(WVV

+1

Um);

Um +1/2,(a

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

U)m +1/2)t

GALERKIN

(Kpm,(a

METHODS FOR PARABOLIC

U)m+ 1/2> +

t(21(Wm+1 + Um); um+ 112 + (rnm,(a


a('2(Wm+ 1 +
-

a(I(WMV+1 + Um); Cm
n(U

Z+i

621

EQUATIONS

Um); Um+1?2

U)m+ 1i2)

+ (rnm,(U U)m-

+ 12)]

U)m+ 112)

m(Ua-)m+l/2

(Y(Wrm+I

Um);Zmr+12,

(U

a)mr+1/2)

The bracketed term on the right-hand side of (11.9) is estimated as in (11.7). The
other terms are estimated as in the proof of Theorem 5.1. The result is summed as in
the proof of Theorem 4.1 to yield the conclusion.
The reader can easily note from the proof of Theorem 11.1 that the evaluation
of the coefficients need not be as accurate in the predictor equation as in the
corrector equation since the error induced by this error is multiplied by At.
In order to see that (1 1.1) is a useful computational process we need to argue
that it is possible to find a space /' such that a(x, Um)and a(x, 2((WM+1 + Um))
can be interpolated into /I' with relative error y, for any small y. First note that
L2 and 11
11Um+ 111
Wm+ 1 11
L2 are bounded by 11
UO11
L2 as long as we have a(x, Uk) and
< k < m. To see this use (1. la)
+
Uk))
positive
semidefinite
0
for
a(x, 2(Wk +1
with V= (1m+,1 + Um)and (l1.lb) with V= Um+1/2 to see that
= IU M112

IIW+11122

2AtO(Um;

1(WM+ 1 +

UM) I2(WM+ 1 +

Um))

(I 1.10)L

(11.10

U+1||L2

2 -

||Ur|
-<

2Ata({(Wn+

1 Um);

Umr+ 1/2

Um+

1/2)

m11L22-

In case we have an inhomogeneous term we can still obtain bounds for Wm+1 and
UMfor mAt < T, provided that the inhomogeneous term is reasonably behaved.
But since Um+ 1 and Km+ 1 are in a finite-dimensional space XW,a bound on the L2norm gives a bound on all the derivatives of Wand U up to the order of differentiability of the functions in X Hence we conclude that if a(x, u) is a smooth function
of its arguments we can easily find an .Al' and an interpolation process such that
a

is as close to a as we want.

We have remarked earlier that a major fraction of the total computing effort
in solving nonlinear parabolic equations will probably be spent evaluating coefficients for the linear systems, even after the simplification indicated above. If
in the predictor-corrector scheme the prediction of the approximate solution at
t = tmr+1 is made using the same coefficients used to make the correction of the

approximate solution at t = tm, the process remains second order correct in At.
This modification is worth consideration because it requires only one evaluation
of the coefficients of the linear systems per time step. Consider Umdefined by
(I11.ll1a)

Wn+,?

At

Um V) + a('(Wm + U n-W);{(
2
-1;2

+
U
+1+U)

)=0
V E

(11.I

b)

AUr+t

Un

a(Y(Wn+?I

Um);

Umr+1/2,

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

V)

Ve=',

,#

622

JIM DOUGLAS,

JR., AND TODD DUPONT

for m ? 1. Let UOand W1 be defined by


(ll.llc)

<U0, V> = <u0, V>,


-

(ll.lld)

+ a(U0;l(Wl + UO),V) = O,

?, V

A>

Vei,
Ve i.

Let U1 be defined using (11.lb). In the above, Umand Wmare in il for each m.
THEOREM 1 1.2. Let u and a(x, u) be as in Theorem 5.1. Let U be given by (11.11),
and let z = u - U. Then there exist C, 3 > 0, T0 > 0 such thatfor At < T0 we have
M-1
ZM

L2

Z IlZm+112

+ 6

Ho

m=0

_- C

LMf1
m
=

)l2
| (UU- )m1/2|
)m?112

+
H/
)m+l2 112Mt
HoA

|(U -

Y,

(U
-(UUi)m-1/2
At

At

+ ll(U - a)1/2,L2 + ll(U -)1+1/2


+ II(u - U)o11L2

(11.12)

+ |(U

L2

U)M-1/2m L2 + (C2 + 1)4C2(At)4]

where Um c i for each m = 0, 1,*, M = T/At.


Proof. Write (2.2) at t = tm+ 1/2 for m ? 1 as
(1.13)

Um?+

+Pm,v

a(umu

m)

)m; Um+ 1/2 +

m, V) =0,

v e Ho,
where Pm and Cmare
Yrm+1/2 =

and

O(C2(At)2)

O(C2At). Let Ym = um and (11.13) that

WOm=

Wm and

2(Ym+ 1 + Zm). It can be shown using (ll.lla)

Ym+ 1llL2

< C[

Zrmn1 2

(U

+ (C2

U)rn?1/21HA

(11.14)

1)4 C2(At)4

At(I|ymrn 2 +

11Zm_1 I1L2)].

The demonstration of (11.14) is very much like that of (5.6). Using our standard
argument we see that
2 Zt[mZ ++1

22

<= 2Z11Zm+

Zmr1L2]

1/2

Ho

112W

u- k)m+?

Cl(U

HA

1/2

lZrn

L2

(11.15)
+

Zm+1

, (u

7)m + 1/2

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

Yrn?1

L2

C2(At)4]

GALERKIN

METHODS FOR PARABOLIC

623

EQUATIONS

Using (11.14) and (11.15) we conclude that for some C,


(1 - CAt)
(11.16)

+
<

Zm+ 111L2

1] Zm+i/12

C[1(u

+ 2

ZM+

U)+

-(1

+ CAt)I1Zm112

2At + [l ym+?112
112,

Zm
, (U -

(C2

1)4C2(At)4

U)m+ 1/2

Zm1

Zm+i1122
+2
+

(11ym 1L2

2]At
+1

Zmi

22)At]At

At.

The relation (11.16) is summed much as we summed (6.4). This will be left to the
interested reader.
It is clear that theorems analogous to Theorem 5.2 and Theorem 5.3 can be
obtained for (1 1.11)and the modification of it in which two passes are made through
the corrector equation. It is also clear that the arguments concerning the interpolation of u0 and the modification of the coefficients apply.
Suppose that the space X# is a space of splines which are based on a rectangular
grid that has been placed on the region Q. It is then possible to choose a basis
{v}lJ=1 for X such that each v, has its support in a small number of grid blocks.
See, for example, [2], [E1]. If the basis is chosen in this fashion, the matrices which
are to be inverted in the linear methods, such as Crank-Nicolson predictorcorrector, are sparse. If the basis is ordered in a reasonable way, they will also
turn out to be band matrices; i.e., all the nonzero entries are in a relatively narrow
band about the main diagonal. Thus they may be factored into upper and lower
triangular band matrices in such a way as not to increase the width of the band
[17]. For particular cases of the space X the matrices which must be inverted are
quite similar to (or even identical to) those which arise when finite difference
techniques are used to approximate solutions to parabolic equations. These
questions will be treated for particular types of spaces X in another paper. In
particular we intend to show that approximate factorization techniques [13], [14]
can be extended to many of these situations.
On certain types of domains Q in R' alternating direction techniques can be
used to yield finite difference approximations to solutions to parabolic problems.
The advantage of these techniques is that they require the solution of equations
which correspond to one-space-dimensional problems. These techniques were
developed by Peaceman, Rachford and one of the authors and have been extended
by many.
Locally one-dimensional procedures can be formulated to yield a Galerkintype approximation for solutions to parabolic problems for certain types of
domains in R'. These procedures will be discussed elsewhere [12].
12. Application of approximationtheory. The error bounds we have produced
thus far are in terms of u - a, where u is the solution of the differential problem
and a is any function which is in X for each t. In this section we shall apply some
results of Birkhoff, Schultz and Varga [4] to indicate how the results of the previous
sections can be used with approximation theory to give rates of convergence.
The approximation theory results that we shall use here are for Hermite interpolation in two variables, but it is worth noting that the same orders of convergence
can be obtained with other piecewise polynomial functions of the same degree [1].

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

624

JIM DOUGLAS, JR., AND TODD DUPONT

Let Q = (0, 1) x (0, 1) and for a positive integer N let h = N -'. For an ordered
pair of nonnegative integers oc= (OC1,a2) define Daf(x) = (a/ax )al(a/ax2)f2f(x)

and lal= ac + a2. Let H(m)denote the set of real-valued functions g such that, for
all ca = (OC1,aC2)satisfying 0<

1, D'g is continuous

?a2 ?< m-

al,

on Q and such

that on each square (kh,(k + 1)h) x (nh,(n + 1)h), where k and n are integers
satisfying 0 < k, n < N - 1, g is a polynomial of the form E 0 ci,jx'x? . Suppose
f is a function on Q such that D(' 012)fis continuous on Q2for 0 _ Xl, a2 < m-1.
We say that fm,h is the H(m)-interpolate of f if D(a112)(f-fm,h)(kh, nh) = 0 for all
n, k, 1 anda2 integers such that O < n, k < N and 0 ? oc1, aC2< m - 1. It is easily
seen that such an interpolate exists and is unique. The following is a special case
of Theorem 5 of [4].
LEMMA12.1. Supposef is afunction definedon Q2such that Dafis continuouson Q
for lacl< 2m and Daf e L2(Q)for locl= 2m. Letfmh be the H(m)interpolate off Then
there exists a constant M which is independentof h such that
11D(f

(12.1)
for locl < 2m

< Mh 2m-I

-fm,h)11L2

1 andO _ a1, X2 ? m. Further,

(12.2)

M'

lal= 2m

IIDfIIL2(),

where M' is independentoff and h.


THEOREM 12.1. Let X = H(m)n Hl(Q). Let u, U and z be as in Theorem 7.1.
Suppose that for every t E[0, T], u and au/at satisfy the hypotheses of Lemma 12.1
and that
1
t) IIL2()<E
E 1D"u(
lal= 2m

(12.3)

(12-3)

Ja~4.t)
YIN(

at

ilc2m

$y()

L2(Q)

where E is a constant independentof t, p is afunction in L2(0, T), and, in the sums in


(12.3), Da denotes x-derivatives only. Then there is a constant C such that
2IZI2
IZII2XL

(12.4)

xL2 < Ch2(2m-1)

Proof To see that this result holds, simply let u(x, t) be the H(m)-interpolateof
u(x, t) for each t. Note that aa(x,t)/at is just the H(m)-interpolateof au/at since
with 0_ $Xl,
C2 < mM 1
D'u and D'u agree along the lines (kh, nh, t) for a = (a1, aC2)
and O < k n < N.
= H(m)
THEOREM 12.2. Let
n Hl(Q). Let u, U and z be as in Theorem 7.2.
the
au/at
that
u
and
satisfy
regularity hypotheses of Theorem 12.1. Then
Suppose
there exists a constant C such that
M-1
(12.5)

IIZM122

1Zk+l/2IIHAt

<

+ (At)4],

C[h2(2m-1)

k=O

where M = T/At and C is independentof h and At.


ProoJf Let u7(x,t) be the H(m)-interpolateof u(x, t) for t = kAt. Note that
(ik+1/2

ak- 1/2)/At

is the H(m)-interpolate of

(Uk+ 1/2

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

Uk

1/2)/At. In order to

GALERKIN

METHODS FOR PARABOLIC

625

EQUATIONS

bound the term


M-1

(U -

(12.6)

a)k+1/2

(U -

Ui)k-1/2

At

2At

note that we can decompose each term in the following way. From the definition
of (u - l)k+ 1/2 (see (4.1)) we see that
(U

17)k+ 1/2

(U

)k-1/2

At

F(~i~u

(12.7)

- Ua)k+I r|l~(U

(12.7)

1
2 Lll

(I

i)2
(U -

17)ki

At

||tl

)(
-

f)

17)k

(U1

1)k -

I|

At

L2

2]

12j

Then note that

k~I

||(U
(u- -1U)k+

((1l

At

7)k 2l
- a)k
2

|r(k+
----<
1Jx,

~~~L2

1 )At 0(U

(.12.8)

2(

\JAt

<
kAt

t) (It2l (At) -2
2

dtWh2m
(dtAt

J<vt
r(k+ I)At

- U7)

Ot

kA

(At)

(7t

g(t) dth

2m)

(At)

(k+ 1)At
_<

(t)2

d tt (t) -h4m.

kAt

Hence

(12.9)

A(t-

)k/2

--(U--)

At <

p(t) dt I"

By looking at [4] the reader will note that we have simplified the discussion by
taking Q to be a square and using a uniform grid. The results of [4] are presented
for regions that are unions of rectangles with sides parallel to the coordinate
axes and nonuniform grid spacing. It is clear from the proofs of Theorems 12.1
and 12.2 that we have approximated the terms au/atand (uk+ - uk)/At more
accurately than necessary to get the conclusions of these theorems. This indicates
that it is probable that we could improve these theorems by weakening the assumptions on au/atif we had slightly more precise interpolation estimates. The authors
hope to look at the relevant approximation theory questions in the near future.
Also note that these estimates ignore one of the principal strengths of these procedures. That is, we have used a class of functions which can approximate all
smooth functions well, whereas in a particular case we would like to include
functions in the basis of ,/ which we know will approximate the solution to that
problem well.
REFERENCES
[1] J. H.

AHLBERG, E. N. NILSON AND J. L.


Academic Press, New York, 1967.

WALSH,

The Theory of Splines and Their Applications.

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

626

JIM DOUGLAS, JR., AND TODD DUPONT

[2] J. P. AUBIN, Behavior of' the error of' the approximate solutions of' boundary value problemsfor
linear elliptic operators by Galerkin's and finite-difference methods, Ann. Scuola Norm.
Sup. Pisa, 21 (1967), pp. 599-637.

, Approximation des espaces de distributions et des operateurs differentiels, Mem. Soc.


Math. France, 12 (1967), pp. 1-139.
[4] G. BIRKHOFF,M. H. SCHULTZAND R. S. VARGA, Piecewise Hermite interpolation in one and two
[3]

[5]
[6]

[7]
[8]
[9]
[10]

uariables with applications to partial differential equations, Numer. Math., 11 (1968), pp.
232-256.
F. E. BROWDER, Strongly non-linear parabolic boundary value problems, Amer. J. Math., 86
(1964), pp. 339-357.
P. G. CIARLET,M. H. SCHULTZAND R. S. VARGA, Numerical methods of high-orderaccuracyfor
nonlinear boundary value problems. I: One dimensional problem, Numer. Math., 9 (1967),
pp. 399-430.
, Numerical methods of high-order accuracy for nonlinear boundary value problems. II:
Nonlinear boundary conditions, Ibid., 11 (1968), pp. 331-345.
, Numerical methods of high-order accuracy for non-linear boundary value problems. III:
Eigenvalue problems, Ibid., 12 (1968), pp. 120-133.
, Numerical methods of high-order accuracy for nonlinear boundary value problems. IV:
Periodic boundarv conditions, Ibid., 12 (1968), pp. 266-279.
J. DOUGLAS, JR., AND T. DUPONT, The numerical solution ofjwaterfloodingproblems in petroleum
engineering by variational methods, Studies
and Applied

Mathematics,

in Numerical

Analysis

2, Society

for Industrial

1970.

Philadelphia,

T. DUPONT AND H. H. RACHFORD, JR., The application of variational methods


to waterflooding problems, J. Canad. Petroleum Tech., 8 (1969), pp. 79-85.
[12] J. DOUGLAS, JR., AND T. DUPONT, Alternating-direction Galerkin methods on rectangles, Proc.
[1 1] J. DOUGLAS, JR.,

Symposium

on

Maryland,

College

Numerical
Park,

Solution

Partial

of

Differential

Equations.

II,

University

of

1970.

[13] T. DUPONT, A factorization procedurefor the solution of'elliptic difference equations, this Journal,
5 (1968), pp. 753-782.
[14] T. DUPONT, R. P. KENDALLAND H. H. RACHFORD,Jr., An approximatefactorization procedure
for solving self-adjoint elliptic difference equations, this Journal, 5 (1968), pp. 559-573.
[15] L. H6RMANDER, Linear Partial Differential Equations, Academic Press, New York, 1963.
[16] J. L. LIONS, Equations differentielles operationnelles et probleme aux limites, Springer-Verlag,
Berlin,

1961.

[17] R. S. MARTIN AND J. H. WILKINSON,Handbook series linear algebra. Solution of symmetric and
unsymmetric band equations and the calculation of eigenvectors of band matrices, Numer.
Math., 9 (1967), pp. 279-301.
[18]

H. MEYER,

Doctoral

thesis,

University

of Chicago,

1969.

Chicago,

[19] S. G. MIKHLIN, VariationalMethods in Mathematical Physics, Pergamon Press, New York, 1964.
[20] S. G. MIKHLIN AND K. L. SMOLITSKIY,Approximate Methods for Solution of Differential and
Integral Equations, American

Elsevier,

New

York,

1967.

[21] C. B. MORREY,Multiple Integrals in the Calculus of Variations, Springer-Verlag, New York, 1966.
[22] H. S. PRICE, J. C. CAVENDISHAND R. S. VARGA, Numerical methods of higher-order accuracy
for diffusion-convectionequations, Soc. Petroleum Engrg. J., 8 (1968), pp. 293-303.
[23] H. S. PRICEAND R. S. VARGA, Error boundsfor semi-discrete Galerkinapproximationsof parabolic
problems with applications to petroleum reservoir mechanics,Numerical Solution of Field
Problems

in Continuum

Physics,

AMS,

Providence,

1970,

pp.

74-94.

[24] B. SWARTZ AND B. WENDROFF,Generalizedfinite-difference schemes, Math. Comp., 23 (1969),


pp. 37-49.
[25] G. STAMPACCHIA,Equations elliptiques du second ordre d coefficients discontinus, Les Presses de
l'Universite

de Montr6al,

Montr6al,

1966.

[26] M. ZLAMAL, On thefinite-element method, Numer. Math., 12 (1968), pp. 394-409.


[27] S. FAEDO, Un nuovo methodo per l'analisi esistenziale e quantitativa dei problemi di propagozione,
Ann.

Scuola

Norm.

Sup.

Pisa,

1 (1949),

pp. 1-40.

This content downloaded on Fri, 14 Dec 2012 11:34:45 AM


All use subject to JSTOR Terms and Conditions

Das könnte Ihnen auch gefallen