Вы находитесь на странице: 1из 27

Digital Object Identier (DOI) 10.

1007/s002110100332
Numer. Math. (2002) 91: 577603
Numerische
Mathematik
Runge-Kutta methods without order reduction
for linear initial boundary value problems
Isaas Alonso-Mallo
Departamentode Matem atica Aplicada yComputaci on, Facultadde Ciencias, Universidadde
Valladolid, c/Doctor Mergelina s/n, 47005 Valladolid, Spain; e-mail: isaias@mac.cie.uva.es
Received July 10, 2000 / Revised version received March 13, 2001 /
Published online October 17, 2001 c Springer-Verlag 2001
Summary. It is well-known the loss of accuracy when a RungeKutta
method is used together with the method of lines for the full discretiza-
tion of an initial boundary value problem. We show that this phenomenon,
called order reduction, is caused by wrong boundary values in intermedi-
ate stages. With a right choice, the order reduction can be avoided and the
optimal order of convergence in time is achieved. We prove this fact for
time discretizations of abstract initial boundary value problems based on
implicit RungeKutta methods. Moreover, we apply these results to the full
discretization of parabolic problems by means of Galerkin nite element
techniques. We present some numerical examples in order to conrm that
the optimal order is actually achieved.
Mathematics Subject Classication (1991): 65M20, 65M12, 65M60, 65J10
1. Introduction
When applied to stiff systems of ordinary differential equations, the Runge
Kutta methods suffer fromreduction of order [13], even when the solution is
regular. Since the spatial semidiscretization of a partial differential equation
becomes stiffer when the spatial discretization is rened, it is natural to
observe the order reduction phenomenon when an evolutionary problem in
partial differential equations is solved by using the method of lines approach.
The order observedinthe applications is governedessentiallybythe stage
order q of the RungeKutta method rather than the classical order p [6, 17
19, 21] and it depends on several factors. However, it is well-known that the
The author has obtained nancial support from DGICYT PB95-705 and JCyL VA025/01
578 I. Alonso-Mallo
order reduction is avoided when we use the method of lines and the solution
of the partial differential equation satises certain boundary constraints [9,
21, 29]. The crucial point is that these constraints do not agree in general
with the boundary relations satised when the solution is regular.
There are several techniques used to avoid the order reduction when a
RungeKutta method is applied. It is possible to nd RungeKutta methods
with high stage order [7, 16], but this technique increases the number of
stages and therefore the computational cost. It is possible to use modied
versions of the classical implementation of a RungeKutta method [11, 14,
15], but the computational cost is also increased in some cases. Moreover,
some substantial modications of the usual implementation of RungeKutta
methods are required.
Since it seems that the order reduction is caused by the boundary values
of the solution, it is natural to avoid this phenomenon by using the boundary
values of the exact solution. The rst advantage is that the boundary values
can be computed with a smaller computational cost than making evaluations
in the whole domain. This idea has been succesfully used in [2, 3] to dene
polynomial and rational methods without order reduction. These methods
are strongly related to RungeKutta methods [9, 31]. However, there are
several inconveniences in [2, 3]. First, the RungeKutta methods reduce to a
polynomial or rational method only for the linear case with time independent
operator, which makes impossible the generalization to other problems, even
the linear ones. Second, the standard software is not written for the rational
formulation of a RungeKutta method, but for the classical formulation.
Finally, we remark that the hypotheses on the spatial discretization in [2, 3]
are not adequate for some standard spatial discretizations.
In this paper, we use the classical formulation of a RungeKutta method
and we obtain the classical order by modifying only the boundary values
usually assigned to the intermediate stages. A similar technique is succes-
fully used in the numerical experiments presented in [1, 27] in the context of
nonlinear hyperbolic equations discretized with explicit RungeKutta meth-
ods and nite difference spatial discretizations. Since in the present work
we use implicit RungeKutta methods for the time integration, the stages
are dened implicitly and their boundary values are not determined by the
RungeKutta method. We show that it is possible to avoid the order reduc-
tion with an adequate choice. For this, we use boundary values obtained
from a recurrence relation where we have taken the usual boundary values
as starting values. We remark that these boundary values are given only in
terms of the data and therefore, they are calculated without knowledge of
the exact solution.
We begin the study with the semidiscretization in time of an abstract
linear initial boundary value problem (IBVP) with implicit RungeKutta
Runge-Kutta methods without order reduction 579
methods. We prove consistency and convergence avoiding the order reduc-
tion. Previous semidiscrete analysis is used in order to derive estimates for
the error of the full discretizations. Although some parts of this paper are
valid for the hyperbolic case, we only consider examples given by parabolic
problems with Galerkin nite element techniques for the spatial discretiza-
tion.
We remark that the order reduction phenomenon is also present when
other time integration methods are used. For example, [22] analyzes the
case of Rosenbrock methods, showing a sharp bound for the order of con-
vergence. In [4], we use a similar technique to avoid the order reduction for
full discretizations with Rosenbrock methods and spectral methods for the
spatial discretization.
The organization of the paper is as follows. The notation and the Runge
Kutta methods used for the time discretizations are the principal issues of
Sect. 2. In Sect. 3 we study the error of the time discretization. Section 4 is
devoted to the full discretization. Section 5 presents some examples which
show the applicability of the previous theory. Some numerical experiments
conrm that the order reduction is avoided.
2. Notation and preliminaries
Let X and Y be two complex Banach spaces, and D(A) be a dense subspace
of X and A : D(A) X X, : D(A) X Y be a pair of linear
operators. We wish to study time discretizations of the well-posed abstract
non-homogeneous linear IBVP
x

(t) = Ax(t) +f(t), 0 t T,


x(0) = u
0
X,
x(t) = g(t) Y, 0 t T,
_
_
_
(2.1)
with an implicit RungeKutta method. We use the framework developed in
[3, 5, 26], which covers several parabolic and hyperbolic problems of practi-
cal interest. We present in this section the main features of this framework.
Let us start making the following assumptions on the linear operators Aand
,
(A1) The boundary operator : D(A) X Y is onto.
(A2) Ker() is dense and A
0
: D(A
0
) = Ker() X X, the restric-
tion of A to Ker(), is the innitesimal generator of a C
0
-semigroup
S(t)
t0
in X.
(A3) If z is a complex number with 1(z) greater than the type of
S(t)
t0
, then the eigenvalue problem
(z A)x = 0,
x = v,
_
580 I. Alonso-Mallo
possesses, for each v Y , a unique solution x = K(z)v. Moreover,
this solution satises
|K(z)v| L|v| ,
for some constant L > 0 independent of z, for z in a half-plane of the
form 1(z)
0
> .
Equivalently, we can impose (A12) and (A4) below, instead of (A3).
(A4) The operator (A, ) : D(A) X X Y is closed.
Since we are interested in approximations of high order, we suppose that
the solution of (2.1) is regular. We assume that there exists an integer number
r 1 such that
A
rj
u
(j)
C([0, T], X), 0 j r. (2.2)
We remark that the assumption (2.2) implies that the time derivatives of the
solution are regular in space, but without to impose any restriction on the
boundary values.
Theorem 3.1 in [3] shows that (2.2) is satised when the data u
0
, f and
g are regular and the boundary values u
0
, f(0) and g(0) satisfy certain
natural compatibility constraints. As a consequence, we obtain
A
j
u(t) = u
(j)
(t)
j1

i=0
A
ji1
f
(i)
(t), 0 j r, (2.3)
and by applying the boundary operator,
A
j
u(t) = g
(j)
(t)
j1

i=0
A
ji1
f
(i)
(t), 0 j r. (2.4)
Later, we will use these boundary values of the exact solution. Notice that
the right-hand side of (2.4) is given only in terms of the data of (2.1).
For the time discretization of (2.1) we consider an s-stages RungeKutta
method with Butcher array,
c /
b
T
(2.5)
where b = [b
1
, . . . , b
s
]
T
, c = [c
1
, . . . , c
s
]
T
, and
/ =
_

_
a
11
. . . a
1s
.
.
.
.
.
.
a
s1
. . . a
ss
_

_
.
Runge-Kutta methods without order reduction 581
Also, we denote 1 = [1, . . . , 1]
T
R
s
, c
l
= [c
l
1
, . . . , c
l
s
]
T
, for l 0, and
by J the s-dimensional identity matrix.
We suppose that (2.5) has classical order p. We remember [13] that the
order conditions
b
T
/
j
c
m
=
1
(j +m+ 1) . . . (m+ 1)
, 0 j +m p 1, (2.6)
are necessary and sufcient so that (2.5) has classical order p when ap-
plied to a linear non-homogeneous ordinary differential equation with time
independent coefcients similar to (2.1).
The convergence results of this paper depend on some rational func-
tions dened by using the previous parameters; the stability function of the
RungeKutta method is
r(z) = 1 +zb
T
(J z/)
1
1.
We suppose that the RungeKutta method given by (2.5) is A-stable, i.e.
/is regular and, for 1(z) 0, the matrix J z/is regular and [r(z)[ 1
(Notice that this denition is slightly stronger than the usual one).
Further, we consider the functions,
R
l,j
(z) = zb
T
(J z/)
1
/
j
(c
l
l/c
l1
), j 0, l 1.
The stage order of the RungeKutta method is denoted by q, thus q is the
highest positive integer number satisfying
C( q) : /c
l1
=
1
l
c
l
1 l q.
Notice that R
l,j
(z) 0 for 1 l q.
For n > 0 integer, we consider time stepsizes k > 0 and we take
t
n+1
= t
n
+ k < T. In what follows, we seek an approximation u
n
to the
exact solution u(t
n
) of (2.1) by applying the implicit RungeKutta method
given by (2.5). A direct implementation of this method give rise to the
following equations for the intermediate stages,
(J I /kA)U
n
= (1 I)u
n
+k(/I)F
n+c
, (2.7)
where U
n
= [U
1
n
, . . . , U
s
n
]
T
, the vector of intermediate stages, and F
n+c
=
[f(t
n
+c
1
k), . . . , f(t
n
+c
s
k)]
T
.
The RungeKutta method does not give all the information needed, since
the boundary values of the intermediate stages are not determined. As a
consequence, the solution to equations (2.7) is not unique. So, we have to
specify those values. By denoting G
n
= [U
1
n
, . . . , U
s
n
]
T
, we obtain the
following equations for the stages,
(J I /kA)U
n
= (1 I)u
n
+k(/I)F
n+c
,
U
n
= G
n
.
_
(2.8a)
582 I. Alonso-Mallo
Now, we use the second equation of the RungeKutta method to obtain,
u
n+1
= u
n
+ (b
T
kA)U
n
+k(b
T
I)F
n+c
. (2.8b)
In order to see that (2.8a) has a unique solution, it sufces to consider the
problems
(J I /kA)U
0
n
= (1 I)u
n
+k(/I)F
n+c
,
U
0
n
= 0,
_
, (2.9)
and
(J I /kA)U
b
n
= 0,
U
b
n
= G
n
.
_
(2.10)
We endow the space X
s
with the usual norm product. Then, the solv-
ability of (2.9) and (2.10) comes from the following two lemmas [6].
Lemma 2.1 There exists a constant C > 0 such that the operator (J I
/kA
0
) : D(A
s
0
) X
s
X
s
is boundedly invertible and
|(J I /kA
0
)
1
| C,
for k > 0 small enough.
Lemma 2.2 For k > 0 small enough and W = [W
1
, . . . , W
s
]
T
Y
s
, the
problem,
(J I /kA)V = 0,
V = W.
_
possesses a unique solution V := K((k/)
1
)W X
s
. Moreover, there
exists a constant C > 0 such that,
|K((k/)
1
)W| C|W|.
By using Lemma 2.1, we deduce that functions (b
T
kA
0
)(J I /
kA
0
)
1
, r(kA
0
) and R
l,j
(kA
0
) are well-dened and bounded for k > 0.
We also deduce the following result.
Lemma 2.3 The values U
n
and u
n
dened in (2.8a) and (2.8b) satisfy
U
n
= (J I /kA
0
)
1
((1 I)u
n
+k(/I)F
n+c
)
+K((k/)
1
)G
n
, (2.11)
and
u
n+1
= u
n
(b
T
/
1
I)((1 I)u
n
+k(/I)F
n+c
)
+(b
T
/
1
I)(J I /kA
0
)
1
((1 I)u
n
+k(/I)F
n+c
)
+(b
T
/
1
I)K((k/)
1
)G
n
+k(b
T
I)F
n+c
. (2.12)
Runge-Kutta methods without order reduction 583
Proof. The expression (2.11) is straightforward by using the notation intro-
duced in Lemmas 2.1 and 2.2. Now, to prove (2.12), we can write
u
n+1
= u
n
+ (b
T
kA)U
n
+k(b
T
I)F
n+c
= u
n
+ (b
T
kA
0
)(J I /kA
0
)
1
((1 I)u
n
+k(/I)F
n+c
)
+(b
T
kA)K((k/)
1
)G
n
+k(b
T
I)F
n+c
,
= u
n
+ (b
T
/
1
I)(/kA
0
)(J I /kA
0
)
1
((1 I)u
n
+k(/I)F
n+c
)
+(b
T
/
1
I)(/kA)K((k/)
1
)G
n
+k(b
T
I)F
n+c
= u
n
(b
T
/
1
I)((1 I)u
n
+k(/I)F
n+c
)
+(b
T
/
1
I)(J I /kA
0
)
1
((1 I)u(t
n
) +k(/I)F
n+c
)
+(b
T
/
1
I)K((k/)
1
)G
n
+k(b
T
I)F
n+c
.
.
3. Error analysis of the semidiscrete problem
This section is devoted to the study the behaviour of local and global errors
of the semidiscrete method dened by (2.8a) and (2.8b). We denote by
U
n
= [U
1
n
, . . . , U
s
n
]
T
the s-dimensional vector that satises
(J I /kA)U
n
= (1 I)u(t
n
) +k(/I)F
n+c
,
U
n
= G
n
,
_
(3.1a)
and we next dene
u
n+1
= u(t
n
) + (b
T
kA)U
n
+k(b
T
I)F
n+c
. (3.1b)
With the previous notation, the semidiscrete local truncation error in t
n
, is
dened by

n
= u(t
n
) u
n
, 1 n N. (3.2)
Obviously, these local truncation errors depend on the boundary values of
the intermediate stages G
n
. As the stages U
n
are approximations of U
n+c
:=
[u(t
n
+c
1
k), . . . , u(t
n
+c
s
k)]
T
, a natural choice is G
n
= G
n+c
:= [g(t
n
+
c
1
k), . . . , g(t
n
+c
s
k)]
T
. As we show later, this usual choice is the origin of
the order reduction. Inorder toavoidthis phenomenon, we dene recursively,
U
[0]
n
= U
n+c
,
U
[j+1]
n
= (1 I)u(t
n
) + (/kA)U
[j]
n
+k(/I)F
n+c
, j 0,
_

_
(3.3)
584 I. Alonso-Mallo
for each 1 n N.
By using a recursive argument, we deduce
U
[j]
n
=
j1

l=0
(/
l
1 k
l
A
l
)u(t
n
) + (/
j
k
j
A
j
)U
n+c
+k
j

l=1
(/
l
k
l1
A
l1
)F
n+c
. (3.4)
Now, we dene
G
[j]
n
= U
[j]
n
=
j1

l=0
(/
l
1 k
l
A
l
)u(t
n
) + (/
j
k
j
A
j
)U
n+c
+k
j

l=1
(/
l
k
l1
A
l1
)F
n+c
, (3.5)
for j 0, and 1 n N. When these boundary values are used to
compute the intermediate stages, the corresponding solutions of (3.1a) and
(3.1b) are denoted by U
[j]
n
and u
[j]
n
. In Sect. 4, we also need the following
boundary values
H
[j]
n
= AU
[j]
n
, j 0, 1 n N. (3.6)
Notice that the expressions (3.4) and (2.3) allow us to state (3.5) and (3.6)
by using only the data of (2.1).
Theorem 3.1 Let u be the solution of (2.1) satisfying (2.2) for r = p + 1.
We use the boundary values G
[j]
n
, 0 j p q, in (3.1a). Then the local
errors
[j]
n
= u(t
n
) u
[j]
n
, 1 n N, 0 j p q, satisfy
|
[j]
n
| Ck
q(j)+1
, for k > 0, (3.7)
where q(j) = min p, q +j, and constant C depends only on the deriva-
tives of u, the RungeKutta method and the operator A
0
.
Proof. First, we take j = 0 and we use G
n
= G
[0]
n
to get
(J I /kA)U
[0]
n
= (1 I)u(t
n
) +k(/I)F
n+c
,
U
[0]
n
= U
[0]
n
,
_
(3.8)
and we can write
(J I /kA)U
[0]
n
= (1 I)u(t
n
) +k(/I)F
n+c
+
[0]
n
, (3.9)
Runge-Kutta methods without order reduction 585
where, by expanding into Taylor series,
[0]
n
is given by

[0]
n
=
p

l=q+1
k
l
l!
(c
l
l/c
l1
)u
(l)
(t
n
) +O(k
p+1
).
Denoting
[0]
n
= U
[0]
n
U
[0]
n
, and subtracting (3.8) from (3.9),
(J I /kA)
[0]
n
=
[0]
n
,

[0]
n
= 0.
_
Therefore, by Lemma 2.1, we have

[0]
n
= (J I /kA
0
)
1

[0]
n
,
for k > 0. On the other hand, since the RungeKutta method has order p,
u(t
n
) + (b
T
kA)U
[0]
n
+k(b
T
I)F
n+c
= u(t
n
) + (b
T
kA)U
n+c
+k(b
T
I)F
n+c
= u(t
n
) +k(b
T
I)(AU
n+c
+F
n+c
)
= u(t
n
) +k(b
T
I)U

n+c
= u(t
n
) +
p1

l=0
k
l+1
l!
b
T
c
l
u
(l+1)
(t
n
) +O(k
p+1
)
= u(t
n
) +
p1

l=0
k
l+1
(l + 1)!
u
(l+1)
(t
n
) +O(k
p+1
)
= u(t
n+1
) +O(k
p+1
), (3.10)
where we have used the order conditions b
T
c
l
= 1/(l +1), l = 0, . . . , p1.
Subtracting (3.1b) from (3.10), we get

[0]
n+1
= (b
T
kA
0
)(J I /kA
0
)
1

[0]
n
=
p

l=q+1
k
l
l!
R
l,0
(kA
0
)u
(l)
(t
n
) +O(k
p+1
) = O(k
q(0)+1
).
Let us take 1 j p q. Then we have
(J I /kA)U
[j]
n
= (1 I)u(t
n
) +k(/I)F
n+c
,
U
[j]
n
= G
[j]
n
,
_
and
(J I /kA)U
[j]
n
= (1 I)u(t
n
) +k(/I)F
n+c
+
[j]
n
, (3.11)
586 I. Alonso-Mallo
whence, by subtracting (3.3) from (3.11),

[j]
n
= U
[j]
n
U
[j+1]
n
= (/kA)(U
[j1]
n
U
[j]
n
)
= (/kA)
j
(U
[0]
n
U
[1]
n
)
= (/kA)
j

[0]
n
=
p

l=q+1
k
l+j
l!
/
j
(c
l
/c
l1
)A
j
u
(l)
(t
n
) +O(k
p+j+1
). (3.12)
We dene
u
[j]
n
= u(t
n
) + (b
T
kA)U
[j]
n
+k(b
T
I)F
n+c
.
We deduce from (3.10) that u
[0]
n
= u(t
n+1
) +O(k
p+1
). Moreover, we have
u
[j+1]
n
= u(t
n
) + (b
T
kA)U
[j]
n
+k(b
T
I)F
n+c
(b
T
kA)(U
[j]
n
U
[j+1]
n
)
= u(t
n
) + (b
T
kA)U
[j]
n
+k(b
T
I)F
n+c
(b
T
kA)
[j]
n
= u(t
n
) + (b
T
kA)U
[j]
n
+k(b
T
I)F
n+c

l=q+1
k
l+j+1
l!
b
T
/
j
(c
l
/c
l1
)A
j+1
u
(l)
(t
n
) +O(k
p+j+1
)
= u
[j]
n
+O(k
p+j+1
),
where we have used the order conditions (2.6). By using a recursive argu-
ment, we obtain
u
[j]
n
= u(t
n+1
) +O(k
p+1
), 0 j p q. (3.13)
Now, we denote
[j]
n
= U
[j]
n
U
[j]
n
. For k > 0, we have

[j]
n
= (J I /kA
0
)
1

[j]
n
.
We subtract (3.1b) from (3.13), and we get,

[j]
n+1
= (b
T
kA
0
)(J I /kA
0
)
1

[j]
n
+O(k
p+1
)
=
p

l=q+1
k
l+j
l!
R
l,j
(kA
0
)u
(l)
(t
n
) +O(k
p+1
)
= O(k
q(j)+1
). (3.14)
.
Runge-Kutta methods without order reduction 587
We now prove some additional results which will be useful in the fol-
lowing section.
Lemma 3.2 With the notation and the hypotheses of Theorem 3.1, we have

[j]
n
= O(k
q(j)+1
), (3.15)
for 1 n N and 1 < j < p q.
Proof. From (3.14), we obtain

[j]
n+1
= (b
T
kA
0
)(J I /kA
0
)
1

[j]
n
+O(k
p+1
)
= (b
T
/
1
I)(/kA
0
J I +J I)
(J I /kA
0
)
1

[j+1]
n
+O(k
p+1
)
= (b
T
/
1
I)
[j]
n
+(b
T
/
1
I)(J I /kA
0
)
1

[j]
n
+O(k
p+1
).
Therefore, we deduce that

[j]
n+1
= (b
T
/
1
I)
[j]
n
+O(k
p+1
),
and we obtain the result from (3.12). .
Lemma 3.3 With the notation and the hypotheses of Theorem 3.1, we have
A
[j]
n
= O(k
q(j)
)
for 1 n N and 1 j p q.
Proof. From (3.14), we obtain
A
[j]
n+1
= (b
T
kA)A
0
(J I /kA
0
)
1

[j]
n
+O(k
p+1
)
= (b
T
kA)(k
1
/
1
I)(/kA
0
J I +J I)
(J I /kA
0
)
1

[j+1]
n
+O(k
p+1
)
= (b
T
/
1
A)
[j]
n
+(b
T
/
1
kA
0
)(J I /kA
0
)
1
k
1

[j]
n
+O(k
p+1
).
and we obtain the result from (3.12). .
Let us see two examples of RungeKutta methods used for the time
discretization of (2.1). First, let us consider the one stage Gauss method.
The corresponding equations can be written as
_
1
kA
2
_
U
n
= u
n
+
k
2
F
n+1/2
,
U
n
= G
n
,
_
_
_
588 I. Alonso-Mallo
and
u
n+1
= u
n
+
kA
2
U
n
+
k
2
F
n+1/2
.
This method has classical order p = 2 and stage order q = 1. Our results
prove order 1 when G
n
= G
[0]
n
= g(t
n+1/2
) and order 2 when G
n
= G
[1]
n
=
g(t
n
) + (k/2)g

(t
n+1/2
). Moreover, the values (3.6) are given by H
[0]
n
=
g

(t
n+1/2
) f(t
n+1/2
) and H
[1]
n
= g

(t
n
) f(t
n
) +
k
2
(g

(t
n+1/2
)
f

(t
n+1/2
)).
Second, let us consider the two stages SDIRKmethod with Butcher array
0
1 1 2
1
2
1
2
for = (3 +

3)/6, which is A-stable. The equations of this method are


given by
(1 kA)U
1
n
= u
n
+kf(t
n+
),
(1 kA)U
2
n
= u
n
+k(1 2)AU
1
n
+k(1 2)f(t
n+
) +kf(t
n+(1)
),
U
1
n
= G
1
n
,
U
2
n
= G
2
n
,
_

_
and
u
n+1
= u
n
+
kA
2
(U
1
n
+U
2
n
) +
k
2
(f(t
n+
) +f(t
n+(1)
)).
This method has classical order p = 3 and stage order q = 1. In this case,
our results prove order 1 when
G
n
= G
[0]
n
=
_
g(t
n+
)
g(t
n+(1)
)
_
,
order 2 when
G
n
= G
[1]
n
=
_
g(t
n
)
g(t
n
)
_
+k
_
0
(1 2)
_ _
g

(t
n+
)
g

(t
n+(1)
)
_
,
and order 3 when
G
n
= G
[2]
n
=
_
g(t
n
)
g(t
n
)
_
+k
_
(g

(t
n
) f(t
n
))
(1 )(g

(t
n
) f(t
n
))
_
+k
2
_

2
0
2(1 2)
2
_ _
g

(t
n+
) f(t
n+
)
g

(t
n+(12
) f(t
n+(1)
)
_
+k
_
0
(1 2)
_ _
f(t
n+
)
f(t
n+(1)
)
_
.
Runge-Kutta methods without order reduction 589
Now, the values (3.6) are given by
H
[0]
n
=
_
g

(t
n+
) f(t
n+
)
g

(t
n+(1)
) f(t
n+(1)
)
_
,
H
[1]
n
=
_
g

(t
n
) f(t
n
)
g

(t
n
) f(t
n
)
_
+k
_
0
(1 2)
_ _
g

(t
n+
) f

(t
n+
)
g

(t
n+(1)
) f

(t
n+(1)
)
_
,
H
[2]
n
=
_
g

(t
n
) f(t
n
)
g

(t
n
) f(t
n
)
_
+k
_
(g

(t
n
) f

(t
n
) Af(t
n
))
(1 )(g

(t
n
) f

(t
n
) Af(t
n
)))
_
+k
2
_

2
0
2(1 2)
2
_

_
g

(t
n+
) f

(t
n+
) Af

(t
n+
)
g

(t
n+(1)
) f

(t
n+(1)
) Af

(t
n+(1)
)
_
+k
_
0
(1 2)
_ _
Af(t
n+
)
Af(t
n+(1)
)
_
.
We end this section studying the convergence with optimal order p. For
this, we need an additional hypothesis on stability. We suppose that variable
time stepsizes k
n
> 0 are used, with t
n+1
= t
n
+k
n
< T. Then, we assume
the following stability condition,
there exists a constant C = C(T) such that
[[
n

m=1
r(k
m
A
0
)[[ C(T), (3.16)
for 0 n N.
For the case of A-stable RungeKutta methods, it is well-known that
condition (3.16) is nearly satised for constant time stepsizes. The bound
[[r
n
(k
n
A
0
)[[ C(T)N
1/2
, 0 n N, (3.17)
is proved in [8]. Bound (3.17) is the best possible for the general case and a
half order can be lost. However, (3.17) may be improved for some particular
methods. This happens for instance for the RadauIIA methods of s stages,
whose stability functions fulll the sufcient conditions in [8]. Moreover,
there are two cases for which the termN
1/2
can be dropped even for variable
time stepsizes, when X is a Hilbert space and A
0
is a dissipative operator
[13, 20] and when A
0
is the innitesimal generator of an analytic semigroup
and the RungeKutta method (2.5) is assumed to be A()-stable [23, 24].
We dene the semidiscrete global error in t
n
as
e
[j]
n
= u(t
n
) u
[j]
n
, 1 n N, 0 j p q. (3.18)
590 I. Alonso-Mallo
Theorem 3.4 Assume that the hypotheses of Theorem 3.1 and the stability
condition (3.16) are satised. Then the global errors (3.18) satisfy
|e
[j]
n
| C(T)
n

m=1
k
q(j)+1
m1
, (3.19)
for 1 n N, 0 j p q, where constant C depends only on the
derivatives of u, the RungeKutta method and the operator A
0
.
Proof. For simplicity, we drop the dependence on 0 j p q in the
notation. For n 1, U
n
and U
n
share the same boundary values. Thus,
subtracting (2.8a) and (3.1a), we obtain
(J I /k
n
A)(U
n
U
n
) = (1 I)e
n
,
(U
n
U
n
) = 0,
_
and subtracting (2.8b) and (3.1b)
u
n+1
u
n+1
= e
n
+ (b
T
k
n
A)(U
n
U
n
)
= e
n
+ (b
T
k
n
A
0
)(J I /k
n
A
0
)
1
(1 I)e
n
= r(k
n
A
0
)e
n
.
This equality allows us to obtain the recurrence relation for the global
error
e
n+1
= u(t
n+1
) u
n+1
+ u
n+1
u
n+1
=
n+1
+r(k
n
A
0
)e
n
,
and by induction, supposing e
0
= 0,
e
n
=
n

m=1
n1

l=j
r(k
l
A
0
)
m
.
Finally, from the stability condition (3.16) and Theorem 3.1 we obtain
(3.19). .
Theorems 3.1 and 3.4 are sufcient in order to prove that the order
reduction can be avoided when the boundary values G
n
= U
[j]
n
, 1
j p q, are used. However, we note that the estimate for the local
errors obtained in Theorem 3.1 is not optimal [6, 21]. Therefore, the order
of convergence obtained in the applications may be greater than the order
obtained in Theorems 3.1 and 3.4.
For instance, suppose that the IBVP can be written as the abstract initial
value problem
x

(t) = A
0
x(t) +f(t), 0 t T,
x(0) = u
0
X,
_
(3.20)
Runge-Kutta methods without order reduction 591
being A
0
: D(A
0
) X an innitesimal generator of a C
0
-semigroup.
This case agrees with the problem (2.1) if g 0, i.e. the boundary operator
applied to the solution vanishes. Equivalently, the solution u belongs to a
subspace D(A

0
), 1 (the number may be fractional). In this case, [21]
proves the estimate
e
n
= O(k
min(p,q+1+)
),
for the global errors when the time stepsize k is constant and A
0
is a self-
adjoint operator in a Hilbert space setting. With the techniques in [22], it is
possible to obtain the same result when A
0
is the innitesimal generator of
an analytic semigroup, 1, the stepsize k is constant, and r() ,= 1. All
these hypotheses are weakened in [6].
As a consequence, the one stage Gauss method applied to the abstract
initial value problem (3.20) (with solution vanishing on the boundary) has
order 2 and therefore, the order reduction is not present. Analogously, the
SDIRK method has order 2 when it is applied to (3.20). Moreover, an ex-
tra order of convergence is achieved for constant stepsizes and the order
reduction in the global errors is avoided in this case.
4. Full discretization
In this section we shall study the application of the above results when
discretization also takes place with regard to the space X. As we will see
in the examples of Sect. 5, the framework studied here includes problems
discretized in space by Galerkin nite element techniques.
Let us denote by h (0, h
0
] the parameter of the spatial discretization.
Let X
h
be a family of nite dimensional spaces, approximating X and we
suppose that X
h
= X
h,0
X
h,b
. The norm in X
h
is denoted by ||
h
. Let
us take a projection operator L
h
: X X
h,0
such that L
h
x is the best
approximation in X
h,0
to x X. We also suppose that there exists another
operator Q
h
: Y X
h,b
and we then dene P
h
:= (L
h
L
h
Q
h
) :
D(A) X
h,0
.
For the approximation of the operator A : D(A) X X, we assume
that we are given the operators A
h
: X
h
X
h,0
in such a way that A
h,0
:
X
h,0
X
h,0
, the restrictions of A
h
to the subspace X
h,0
, are approximation
of A
0
: D(A
0
) D(A) X (Remind that elements of D(A
0
) are regular
in space and vanish in the boundary. However, it is possible to consider
an element u X regular in space but with no zero boundary value, i.e.
u D(A)). Therefore, when x
h
= x
h,0
+ x
h,b
X
h,0
X
h,b
= X
h
, we
have A
h
x
h
= A
h,0
x
h,0
+A
h
x
h,b
.
592 I. Alonso-Mallo
We pose the following semidiscrete problem. Find u
h
(t) X
h,0
such
that
u

h
(t) +L
h
Q
h
g

(t) = A
h
(u
h
(t) +L
h
Q
h
g(t)) +L
h
f(t)
u
h
(0) +L
h
Q
h
g(0) = L
h
u(0),
_
(4.1)
or, equivalently,
u

h
(t) = A
h,0
u
h
(t) +A
h
Q
h
g(t) +L
h
Q
h
(f(t) g

(t)) +P
h
f(t)
u
h
(0) = P
h
u(0),
_
which results from the discretization in space of (2.1).
The subsequent analysis is carried out under the following hypotheses,
which are very close to those in [9].
(H1) The operators A
h,0
are invertible and generate uniformly bounded
C
0
-semigroups e
tA
h,0
on X
h
satisfying
[[e
tA
h,0
[[
h
M, (4.2)
where M 1 is a constant (notice that (4.2) implies the invertibility of
(I A
h,0
) for all > 0, i.e. the operators A
h,0
are nearly invertible).
(H2) For each x X,
|L
h
u|
h
C|u|, (4.3)
and for each v Y ,
|Q
h
v|
h

h
|v|. (4.4)
(H3) We dene the elliptic projection R
h
: D(A) X
h,0
as follows. If
u D(A), then R
h
u satisfy
A
h
(R
h
u +Q
h
u) = L
h
Au,
or, equivalently
R
h
u = A
1
h,0
(L
h
Au A
h
Q
h
u)
(In applications to Galerkin nite element methods, R
h
u is the dis-
cretized solution of the elliptic problem with exact solution u).
Now, we suppose that, for u D(A),
|R
h
u +L
h
Q
h
u|
h

h
|u| +
h
|Au| , (4.5)
where
h
is bounded and
h
is small with h. We also assume that
there exists a subspace Z of X, such that,
(a) for u Z, c C with 1(z) > 0 and k > 0,
(cI kA
0
)
1
u Z, K(c/k)u Z, (4.6)
Runge-Kutta methods without order reduction 593
(b) for u Z D(A),
|L
h
u R
h
u L
h
Q
h
u|
h
= |P
h
u R
h
u|
h

h
|u|
Z
, (4.7)
where
h
is small with h.
Remark 4.1 In order to prove the estimate( 4.5), it is possible to obtain
|P
h
u R
h
u|
h

h
|Au| ,
which is very similar to (4.7) (with Z = D(A) and
h
=
h
), and we have
|R
h
u +L
h
Q
h
u|
h
|R
h
u P
h
u|
h
+|L
h
u|
h

h
|Au| +C |u| .
We pose the followingfull discretizationfor the numerical approximation
of (2.1),
U
h,n
= (1 I
h
)u
h,n
+ (/kA
h,0
)U
h,n
+ (/kA
h
Q
h
)G
n
(/kL
h
Q
h
)H
n
+k(/I
h
)P
h
F
n+c
, (4.8a)
u
h,n+1
= u
h,n
+ (b
T
kA
h,0
)U
h,n
+ (b
T
kA
h
Q
h
)G
n
(b
T
kL
h
Q
h
)H
n
+k(b
T
I
h
)P
h
F
n+c
, (4.8b)
where G
n
= G
[pq]
n
and H
n
= H
[pq]
n
. We remember that, with these
boundary values, the order reduction is completely avoided in Sect. 3. The
method (4.8a), (4.8b) is obtained by applying the standard formulation of a
Runge-Kutta method given by (2.5), with modied boundary values, to the
ordinary differential system (4.1).
We denote by U
h,n
= [U
1
h,n
, . . . , U
s
h,n
]
T
the s-dimensional vector sat-
isfying
U
h,n
= (1 I
h
)R
h
u(t
n
) + (/kA
h,0
)U
h,n
+ (/kA
h
Q
h
)G
n
(/kL
h
Q
h
)H
n
+k(/I
h
)R
h
F
n+c
, (4.9)
and the value u
h,n+1
as
u
h,n+1
= R
h
u(t
n
) + (b
T
kA
h,0
)U
h,n
+ (b
T
kA
h
Q
h
)G
n
(b
T
kL
h
Q
h
)H
n
+k(b
T
I
h
)R
h
F
n+c
. (4.10)
Now, we dene the full local truncation error in t
n
as

h,n
= R
h
u(t
n
) u
h,n
, 1 n N. (4.11)
594 I. Alonso-Mallo
Theorem 4.2 Assume that the hypotheses of Theorem 3.1 as well as (H1)
(H3) are satised. Moreover, we suppose the solution u and the source term
of (2.1) satisfy A
j
u C([0, T], Z), 0 j p q, A
j
f C([0, T], Z),
0 j p q 1, for 0 t T. Then the method (4.8a), (4.8b) is
consistent and the estimate of the full local truncation error

h,n+1
= R
h
u(t
n+1
) u
h,n+1
= O(
h
k
p+1
+
h
k
p
+
h
k
p+1
+k
h
), (4.12)
holds for 0 n N 1.
Proof. We have
R
h
u(t
n+1
) u
h,n+1
= (R
h
u(t
n+1
) R
h
u
n+1
) + (R
h
u
n+1
u
h,n+1
)),
and we wish to estimate the right hand side.
From (H1), (H3) and Lemmas 3.2 and 3.3,
|R
h
u(t
n+1
) R
h
u
n+1
|
h
= |R
h

n+1
|
h
|R
h

n+1
+L
h
Q
h

n+1
|
h
+|L
h
Q
h

n+1
|
h

h
|
n+1
| +
h
|A
n+1
| +C
h
|
n+1
|
= O(
h
k
p+1
+
h
k
p
+
h
k
p+1
).
By using (4.6) and Lemma 2.3, we deduce that the values U
n
and u
n
dened in (3.1a) and (3.1b) belong to the subspace Z. By applying the
operator R
h
to (3.1a), we have
R
h
U
n
= (1 I
h
)R
h
u(t
n
) + (/kR
h
A)U
n
+k(/I
h
)R
h
F
n+c
= (1 I
h
)R
h
u(t
n
) + (/k(A
h,0
R
h
+A
h
Q
h
L
h
Q
h
A))U
n
+k(/I
h
)R
h
F
n+c
+ (/k(R
h
AP
h
A))U
n
= (1 I
h
)R
h
u(t
n
) + (/k(A
h,0
R
h
+A
h
Q
h
L
h
Q
h
A))U
n
+k(/I
h
)R
h
F
n+c
+ (/k((R
h
P
h
)A))U
n
. (4.13)
From (H1) and (H3), we can write
R
h
U
n
U
h,n
= (J I
h
/kA
h,0
)
1
(/k((R
h
P
h
)A)U
n
)
= O(k
h
).
Now, we apply the operator R
h
to (3.1b) to get
R
h
u
n+1
= R
h
u(t
n
) + (b
T
k(A
h,0
R
h
+A
h
Q
h
L
h
Q
h
A))U
n
+k(b
T
I
h
)R
h
F
n+c
+ (b
T
k((R
h
P
h
)A)U
n
). (4.14)
Runge-Kutta methods without order reduction 595
Therefore, by subtracting (4.10) from (4.14)
R
h
u
n+1
u
h,n+1
= (b
T
kA
h,0
)(R
h
U
n
U
h,n
)
+(b
T
k((R
h
P
h
)A)U
n
= O(k
h
), (4.15)
concluding the proof. .
As in the semidiscrete case, we need a stability condition to obtain con-
vergence. From now on, we assume that we use variable time stepsizes and
the following condition holds:
there exists a constant C = C(T) such that
|
n

m=1
r(k
m
A
h,0
)|
h
C(T), (4.16)
for each h > 0 with t
n+1
= k
n
+t
n
< T and 0 n N.
Notice that from hypothesis (H1), we have the following stability con-
dition [8],
|r
n
(kA
h,0
)|
h
Kn
1/2
,
for each h, k and n with xed t = nk, and K being a constant independent
of n, k and h .
Theorem 4.3 Assume that the hypotheses of Theorem 3.1 as well as (H1)
(H3) are satised. Moreover, we suppose the solution u and the source term
of (2.1) satisfy A
j
u C([0, T], Z), 0 j p q, A
j
f C([0, T], Z),
0 j p q 1, for 0 t T. Then the method (4.8a), (4.8b) is
convergent and the estimate of the full global error
e
h,n+1
= L
h
u(t
n+1
) u
h,n+1
L
h
Q
h
u(t
n+1
)
= P
h
u(t
n+1
) u
h,n+1
(4.17)
= O(
h
n+1

m=1
k
p+1
m1
+
h
n+1

m=1
k
p
m1
+
h
n+1

m=1
k
p+1
m1
+
h
), (4.18)
holds for 0 n N 1.
Proof. Since
|L
h
u(t
n+1
) u
h,n+1
L
h
Q
h
u(t
n+1
)|
h
= |P
h
u(t
n+1
) u
h,n+1
|
h
|P
h
u(t
n+1
) R
h
u(t
n+1
)|
h
+|R
h
u(t
n+1
) u
h,n+1
|
h
,
596 I. Alonso-Mallo
and, from (H3),
|P
h
u(t
n+1
) R
h
u(t
n+1
)|
h

h
|u(t
n+1
)|
Z
= O(
h
),
it sufces to estimate |R
h
u(t
n+1
) u
h,n+1
|
h
.
Then, we have
R
h
u(t
n+1
) u
h,n+1
= (R
h
u(t
n+1
) u
h,n+1
) + (u
h,n+1
u
h,n+1
),
and we wish to estimate the right hand side.
From Theorem 4.2,
R
h
u(t
n+1
) u
h,n+1
= O(
h
k
p+1
n
+
h
k
p
n
+
h
k
p+1
n
+k
n

h
).
Subtracting (4.8a) from (4.9),
U
h,n
U
h,n
= (J I
h
/k
n
A
h,0
)
1
((1 I
h
)(R
h
u(t
n
) u
h,n
))
+(J I
h
/k
n
A
h,0
)
1
(k
n
(/I
h
)(R
h
P
h
)F
n+c
),
and (4.8b) from (4.10),
u
h,n+1
u
h,n+1
= R
h
u(t
n
) u
h,n
+ (b
T
k
n
A
h,0
)(U
h,n
U
h,n
)
+k
n
(b
T
I
h
)(R
h
P
h
)F
n+c
= r(k
n
A
h,0
)(R
h
u(t
n
) u
h,n
)
+(b
T
k
n
A
h,0
)(J I
h
/k
n
A
h,0
)
1
(k
n
(/I
h
)(R
h
P
h
)F
n+c
)
+k
n
(b
T
I
h
)(R
h
P
h
)F
n+c
.
Thus, we have proved the expression
R
h
u(t
n+1
) u
h,n+1
= r(k
n
A
h,0
)(R
h
u(t
n
) u
h,n
) +F
h,n
where
F
h,n
= O(
h
k
p+1
n
+
h
k
p
n
+
h
k
p+1
n
+k
n

h
).
The proof ends by using a standard recurrence argument. .
Runge-Kutta methods without order reduction 597
5. Examples and numerical results
We show the feasibility of the previous theory for the discretization of
parabolic evolution problems. The spatial semidiscretization is achieved
with Galerkin nite element methods. We use here the notation introduced
in [28]. The references [10, 12, 3032] are also of interest. We consider a
parabolic problem with non-homogeneous Dirichlet boundary conditions,
but we remark that it is possible to consider other boundary conditions with
slight modications.
Let us suppose that is a bounded domain in R
d
(d = 2, 3) with a
Lipschitz continuous boundary . We consider the second order linear
operator
Lw :=
d

i,j=1
D
i
(a
ij
D
j
w) +
d

i=1
(D
i
(b
i
w) +c
i
D
i
w) +a
0
w,
where D
i
=

x
j
. We suppose that the coefcients a
ij
(x), b
i
(x), c
i
(x) and
a
0
(x) are real smooth functions on . Lis elliptic, i.e. there exists a constant

0
> 0 such that
d

i,j=1
a
ij

j

0
[[
2
for all R
d
, a.e. on .
The operator L is associated to the bilinear form
a(w, v) :=
_

_
_
d

i,j=1
a
ij
D
j
wD
i
v
d

i=1
(b
i
wD
i
v c
i
vD
i
w) +a
0
wv
_
_
.
For V = H
1
0
(), the bilinear forma(, ) is well dened and continuous
in V V , i.e.
[a(u, v)[ C[[u[[
V
[[v[[
V
, u, v V.
Moreover, under suitable conditions, a(, ) is coercive (see [28], p.164), i.e.
a(u, u) [[u[[
2
V
.
Then the variational problem
nd u V such that
a(u, v) = (f, v), v V, (5.1)
is uniquely solvable for f L
2
().
598 I. Alonso-Mallo
Under the previous assumptions, a smooth solution of (5.1) is also the
solution to the homogeneous Dirichlet problem
Lu = f,
u[

= 0.
_
In order to consider the non-homogeneous Dirichlet problem, we take
a boundary datum g H
1/2
(). We extend g in whole to a function
g H
1
(). Now, the variational problem is
nd u V such that
a(u + g, v) = (f, v), v V. (5.2)
As it is well-known (see Remark 7.1.4 in [28]), it is possible to con-
sider other formulations of the non-homogeneous Dirichlet problem. These
formulations are based in approximations by mixed methods and are not
considered here.
Take X = L
2
(), Y = H
3/2
() and denote D(A) = H
2
() X.
Consider the operators acting on D(A),
Au = Lu, u = u[

.
On the other hand, we consider the operator A
0
= A[
ker()
. Then, A
0
is the
innitesimal generator of a holomorphic semigroup e
tA
0
in X. Therefore,
with the notation above, the IBVP
u
t
= Lu +f, on [0, T],
u[
t=0
= u
0
, on ,
u = g, on [0, T],
_
_
_
(5.3)
can be tted into the theory of abstract IBVPs developed in [3, 5, 26].
Finite elements are used for the semidiscretization of the weak formula-
tion of (5.3),
nd u L
2
(0, T; V ) C([0, T], L
2
()) such that
d
dt
(u(t) + g(t), v) +a(u(t) + g(t), v) = (f, v), v V.
where u(0) + g(0) = u
0
, g(t) is a suitable extension of the boundary datum
g(t) in whole and f L
2
( (0, T)). Assuming that the boundary
datum g is regular, we follow the remark 6.2.2 in [28].
We suppose that T
h
is a partitionof
h
=

TT
h
T, a suitable subdomain
of . We take a family of nite dimensional spaces V
h

h>0
X with the
inherited normand made up of nite elements. We suppose that the partition
Runge-Kutta methods without order reduction 599

h
=

TT
h
T and the nite elements satisfy the assumptions in Theorem
6.2.1 in [28].
Let us take X
h
= V
h
and we write X
h
= X
h,0
X
h,b
where X
h,0
and
X
h,b
collect the internal and boundary nodes, respectively. We denote by
Q
h
u X
h,b
to the interpolation operator in the boundary of u D(A).
Therefore, in (4.4), the factor
h
depends on the approximation of domain
and boundary condition (see for example Sect. 4.4 in [30]). Let L
h
: X
X
h,0
be the orthogonal projection dened by
(L
h
u, ) = (u, ), u X = L
2
(), X
h,0
,
which satises (4.3).
We dene the operators A
h
: X
h
X
h,0
, through the relation
(A
h
u
h
, ) = a(u
h
, ), u
h
X
h
, X
h,0
,
and, as in Sect. 4, A
h,0
is the restriction of A
h
to X
h,0
. Then, the operators
A
h,0
are invertible and generate C
0
-semigroups on X
h
satisfying (H1) (see
[12], Sect. 6 and 7). We also introduce the Ritz (or elliptic) projection dened
by
(A
h
(R
h
u +Q
h
u), ) = (Au, ) = (L
h
Au, ),
for u D(A) and X
h,0
.
We assume that the solution u of the variational problem (5.2) is in the
space Z = H
r
(). Since D(A
s
) = H
2s
, s > 0, we deduce that (4.6) is
satised. Moreover, with suitable hypotheses on the nite elements used
(see e.g. Remark 6.2.2 in [28]), we can obtain (4.7), the second part of (H3),
with
h
= O(h
r1
).
Now, we obtain (4.5) by using Remark 4.1. By taking r = 2 and since the
function(A, ) : D(A) = H
2
() L
2
()H
3/2
() is anisomorphism,
we obtain (4.5), the rst part of (H3), with
h
= O(1) and
h
= O(h). The
estimate on the global error obtained in Theorem 4.3 is
O
_
(
h
+ 1)
n+1

m=1
k
p+1
m1
+h
n+1

m=1
k
p
m1
+h
r1
_
,
and with additional hypotheses (Theorem 6.2.2 in [28]), it is possible to
prove the estimate
O
_
(
h
+ 1)
n+1

m=1
k
p+1
m1
+h
2
n+1

m=1
k
p
m1
+h
r
_
.
The mild condition h
2
Ck
n
, C constant, allows us to obtain
O
_
(
h
+ 1)
n+1

m=1
k
p+1
m1
+h
r
_
.
600 I. Alonso-Mallo
10
2
10
1
10
0
10
10
10
9
10
8
10
7
10
6
10
5
10
4
10
3
10
2
10
1
timestep
l
o
c
a
l

e
r
r
o
r
Fig. 1. Comparison of efciency. Local errors
Remark 5.1 Since the semidiscretization in time in Sect. 3 holds for a gen-
eral Banach space framework, it is also possible to consider the maximum
norm. See for example [25].
5.1. Numerical experiments
We present some numerical experiments for the one-dimensional parabolic
example
u
t
(x, t) = u
xx
(x, t) +f(x, t), 0 x 1, 0 t 1,
u(x, 0) = u
0
(x), 0 x 1,
u(0, t) = g
0
(t), 0 t 1,
u(1, t) = g
1
(t), 0 t 1.
_

_
(5.4)
This example is suitable for the study of the order of convergence because the
boundary is very simple and certain sources of error, as a curved boundary,
are not present. Therefore, it is possible to eliminate the spatial error and
to nd the order of convergence in time. This problem can be tted into
the theory of abstract IBVPs developed in [26] taking X = L
2
(0, 1), Y =
C
2
and dening, for u D(A) = H
2
(0, 1), Au = u
xx
and u =
[u(0), u(1)].
The space discretization is carried out as follows [30]. For J a positive
integer, we denote h = 1/J, we introduce the uniform grid x
j
= jh,
1 j J in (0, 1), and we take X
h
the space of nite elements which are
Runge-Kutta methods without order reduction 601
10
2
10
1
10
0
10
10
10
9
10
8
10
7
10
6
10
5
10
4
10
3
10
2
10
1
g
l
o
b
a
l

e
r
r
o
r
timestep
Fig. 2. Comparison of efciency. Global errors
quadratic in each subinterval [x
j
, x
j+1
], j = 0, . . . , J 1, with the L
2
-norm
inherited fromX. The operators P
h
, Q
h
, and A
h
are dened in an analogous
way to the previous example. We deduce that
h
= O(h
1/2
) in (4.4). Let
us take Z = D(A) = H
2
(0, 1) together with the usual norm. If u Z, the
conditions (4.5) and (4.7) are obtained with
h
= O(1),
h
= O(h
3
) and

h
= O(h
3
).
For the time discretization of (5.4), we use the two stage SDIRK method
studied in Sect. 3.
The numerical experiment is carried out for the data u
0
(x) = x
2

2x + 0.75, g
0
(t) = 0.75 exp(t), g
1
(t) = 0.25 exp(t) and f(x, t) =
exp(t)(x
2
2x1.25), which correspond to the solution u(x, t) = exp(t)
(x
2
2x + 0.75). We take the very small stepsize h = 1/4096 in order to
annihilate the error due to the space discretization (notice that this strategy
may be disastrous when the RungeKutta method is explicit because we do
not have unconditional stability).
Figure 1 displays the local errors in front of the time stepsizes. The lines
with circles, crosses and asterisks correspond, respectively, to the use of the
boundary values G
[0]
n
and H
[0]
n
, G
[1]
n
and H
[1]
n
, G
[2]
n
and H
[2]
n
. The numerical
order of consistency observed is roughly 2, 3 and 5. Therefore the order
reduction is totally avoided in the third case. We also remark the drastic
decrease of the sizes of the local errors.
We have also computed the global errors in t = 1. The results are dis-
played in Fig. 2 with the same plot symbols as Fig. 1. The numerical order
of convergence is 2, 2 and 4 respectively and the order reduction is also
602 I. Alonso-Mallo
totally avoided with the last modication of the boundary values. We also
note that the order displayed for the rst case is 2 because the time stepsize
k is constant [6, 13, 21].
References
1. S. Abarbanel, D. Gottlieb, M. H. Carpenter: On the removal of boundary errors caused
by RungeKutta integration of nonlinear partial differential equations. SIAM J. Sci.
Comput. 17, 777782 (1996)
2. I. AlonsoMallo: Single step methods with optimal order of convergence for partial
differential equations. Appl. Num. Math. 31, 117131 (1999)
3. I. AlonsoMallo: Rational methods with optimal order of convergence for partial dif-
ferential equations. Appl. Num. Math. 35, 265292 (2000)
4. I. AlonsoMallo, B. Cano: Spectral/Rosenbrock discretizations without order reduction
for linear parabolic problems (submitted)
5. I. AlonsoMallo, C. Palencia: On the convolutions operators arising in the study of
abstract initial boundary value problems. Proc. Royal Soc. Edinburgh. 126A, 515539
(1996)
6. I. AlonsoMallo, C. Palencia: Optimal orders of convergence in RungeKutta methods
for linear, non-homogeneous PDEs with singular source terms (preprint)
7. C. Bendtsen: On implicit RungeKutta methods with high stage order. BIT37, 221226
(1997)
8. P. Brenner, V. Thom ee: On rational approximations of semigroups. SIAM J. Numer.
Anal. 16, 683694 (1979)
9. P. Brenner, M. Crouzeix, V. Thom ee: Single step methods for non-homogeneous linear
differential equations in Banach spaces. R.A.I.R.O. An. Num. 16, 526 (1982)
10. P.G. Ciarlet: Basic error Estimates for Elliptic Problems, Vol. II, Part 1, in Handbook
of Numerical Analysis. P. G. Ciarlet, J. L. Lions (eds.), Amsterdam: NorthHolland,
1991
11. M.P. Calvo, C. Palencia: Avoiding the order reduction of RungeKutta methods for
linear initial boundary value problems (preprint)
12. H. Fujita, T. Suzuki: Evolution Problems, Vol. II, Part 1, in Handbook of Numerical
Analysis, P. G. Ciarlet, J. L. Lions (eds.) Amsterdam: NorthHolland, 1991
13. E. Hairer, G. Wanner: Solving Ordinary Differential Equations II, Stiff and Differential-
Algebraic Problems. Berlin: Springer, 1991)
14. S. Keeling: Galerkin/RungeKutta discretizations for parabolic equations with time-
dependent coefcients. Math. Comput. 52, 561586 (1989)
15. S. Keeling: Galerkin/RungeKutta discretizations for semilinear parabolic equations.
SIAM J. Numer. Anal. 27, 394418 (1990)
16. T. Koto: Explicit RungeKutta schemes for evolutionary problems in partial differential
equations. Annals of Numer. Math. 1, 335346 (1994)
17. C. Lubich, A. Ostermann: RungeKutta methods for parabolic equations and convolu-
tion quadrature. Math. Comput. 60, 105131 (1993)
18. C. Lubich, A. Ostermann: RungeKutta approximation of quasilinear parabolic equa-
tions. Math. Comput. 64, 601627 (1995)
19. C. Lubich, A. Ostermann: Interior estimates for time discretizations of parabolic equa-
tions. App. Num. Math. 18, 241251 (1995)
20. B. Sz.-Nagy, C. Foias: Harmonic Analysis of Operators on Hilbert Spaces, Amsterdam:
North-Holland, 1970
Runge-Kutta methods without order reduction 603
21. A. Ostermann, M. Roche: RungeKutta methods for partial differential equations and
fractional orders of convergence. Math. Comput. 59, 403420 (1992)
22. A. Ostermann, M. Roche: Rosenbrock methods for partial differential equations and
fractional orders of convergence. SIAM J. Numer. Anal. 30, 10841098 (1993)
23. C. Palencia: Astability result for sectorial operators in Banach spaces. SIAMJ. Numer.
Anal. 30, 13731384 (1993)
24. C. Palencia: Stability of rational multistep approximations of holomorphic semigroups.
Math. Comput. 64 (1995), 591599
25. C. Palencia: Maximum norm analysis of completely discrete nite element methods
for parabolic problems banach spaces. SIAM J. Numer. Anal. 33, 16541668 (1993)
26. C. Palencia, I. AlonsoMallo: Abstract initial boundary values problems. Proc. Royal
Soc. Edinburgh 124A, 879908 (1994)
27. D. Pathria: The correct formulation of intermediate boundary conditions for Runge
Kutta time integration of initial boundary value problems. SIAM J. Sci. Comput. 18,
12551266 (1997)
28. A. Quarteroni, A. Valli: Numerical Approximation of Partial Differential Equations.
Berlin: Springer, 1994
29. J.M. Sanz-Serna, J. G. Verwer: Stability and convergence at the PDE/Stiff ODE inter-
face. Appl. Num. Math. 5, 117132 (1989)
30. G. Strang, G.J. Fix: An Analysis of the Finite Element Method. Cliffs, Englewood NJ:
Prentice-Hall, 1973
31. V. Thom ee: Galerkin nite element methods for parabolic problems. In Lecture notes
Math. Vol. 1054. Berlin: Springer, 1984
32. L.B. Wahlbin: Finite element methods for evolution equations, in Advances in Numer-
ical Analysis, Vol. 1, W. Light (ed.). Oxford: Clarendon Press, 1991

Вам также может понравиться