Académique Documents
Professionnel Documents
Culture Documents
FUDZIAH ISMAIL
Universiti Putra Malaysia
Department of Mathematics
43400 UPM Serdang, Selangor
MALAYSIA
fudziah@math.upm.edu.my
Abstract: In this paper, the fifth order Improved Runge-Kutta method (IRK5) that uses just five function evaluations
per step is developed. The method proposed here are derived with only five stages which results in lower number
of function evaluations. Therefore, IRK5 has a lower computational cost than the classical fifth order Runge-Kutta
method (RK5). Here, the order conditions of the method are obtained up to order six and the coefficients of the
fifth order method are determined by minimizing the error norm of the sixth order method. Numerical examples
are given to illustrate the computational efficiency and accuracy of IRK5 in compared with RK5.
KeyWords: Improved Runge-Kutta methods; Two-step methods; Order conditions; Ordinary differential equations
1 Introduction
systems. Rabiei et al ([5]) constructed the IRK methods and obtained the order conditions of the methods
up to order five. The obtained methods were of order three and four, also the convergence and stability
region of the methods were also discussed.
The IRK methods arise from the classical RK
methods, can also be considered as a special class of
two-step methods. That is, the approximate solution
yn+1 is calculated using the values of yn and yn1 .
Our method introduces the new terms of ki , which
are calculated from ki , (i > 2) in the previous step.
The scheme proposed herein has a lower number of
function evaluations than the RK methods while maintaining the same order of local accuracy.
In section 2, we give a general idea of IRK method
and order conditions of the method up to order six are
given in section 3. In section 4, we derived the fifth
order methods with five stages. Finally the numerical
results and discussions are presented in section 5.
x [x0 X],
(1)
y(x0 ) = y0 .
One of the most common method for solving numerically (1) is Runge-Kutta (RK) method. Most efforts
to increase the order of RK method, have been accomplished by increasing the number of Taylors series terms used and thus the number of function evaluations. The RK method of order p has a local error
over the step size h of O(hp+1 ).
Many authors have attempted to increase the efficiency of RK methods with a lower number of function evaluations required. As a result, Goeken et.al
[1] proposed a class of Runge-Kutta method with
higher derivatives approximations for the third and
fourth-order method. Xinyuan [2] presented a class of
Runge-Kutta formulae of order three and four with reduced evaluations of function. Phohomsiri and Udwadia [3] constructed the Accelerated Runge-Kutta integration schemes for the third-order method using two
functions evaluation per step. Udwadia and Farahani
[4] developed the Accelerated Runge-Kutta methods
for higher orders. However most of the presented
methods are obtained for the autonomous system
while the Improved Runge-kutta methods (IRK) can
be used for autonomous as well as non-autonomous
ISBN: 978-1-61804-039-8
+h b1 k1 b1 k1 +
s
X
i=2
129
!
bi (ki ki ) ,
xn2
k1
k2
..
.
k(s1)
xn1
k1
k2
..
.
ks1
k1
k2
..
.
k(s1)
xn+1
xn
k1
k2
..
.
ks1
k1
k.2
..
k(s1)
!
s
X
yn+1 = yn + h b1 k1 b1 k1 +
bi (ki ki ) ,
xn+2
k1
k2
..
.
ks1
i=2
for 1 n N 1, where
k1 = f (xn , yn ),
for 1 n N 1, where
ki = f (xn + ci h, yn + h
k1 = f (xn , yn ),
ki
k1 = f (xn1 , yn1 ),
i1
X
= f (xn + ci h, yn + h
aij kj ),
i1
X
aij kj ).
j=1
aij kj ).
(3)
j=1
(2)
for 2 i s, c2 , . . . , cs [0 1] and f depends on
both x and y while ki and ki depend on the values
of kj and kj for j = 1, ..., i 1. Here s is the number of function evaluations performed at each step and
increases with the order of local accuracy of the IRK
method. In each step we only need to evaluate the values of k1 , k2 , . . . , while k1 , k2 , . . . are calculated
from the previous step Figure (1) shows this idea.
The accelerated Runge-Kutta method (see [4]), is
derived purposely for solving autonomous first order
ODEs where the stage or function evaluation involved
is of the form
0
c2
c3
..
.
a21
a31
..
.
a32
..
.
cs
b1
as1
b1
as2
b2
..
.
...
...
ass1
bs1
bs
where ki is a function of y only and the term involved ki1 . There are two improvement used here,
the first one is the function f is not autonomous, thus
the method is not specific for y 0 = f (y(x)) but it
can be used for solving both the autonomous equations as well as the more general differential equations y 0 = f (x, y(x)). The second improvement is
that the internal stages ki and ki contain more k vali1
P
aj kj for i = 1, ..., s,
ues which are defined as
j=1
aij kj ),
ki = f (xn1 + ci h, yn1 + h
i1
X
= f (xn1 , yn1 ),
j=1
j=1
ki = f (xn1 + ci h, yn1 + h
k1
i1
X
Order conditions
130
0
c2
c3
c4
c5
b1
method up to order five. Here, we developed the Taylor series expansion and obtained the order conditions
up to order six as presented in Table 2.
bi c4i =
a54
b4
b5
5
X
!
bi (ki ki ) ,
i=2
k1 = f (xn1 , yn1 ),
i1
X
= f (xn + ci h, yn + h
aij kj ),
k1 = f (xn , yn ),
ki
j=1
ki = f (xn1 + ci h, yn1 + h
i1
X
aij kj ).
j=1
(4)
for 2 i 5, ci [0 1] and ci =
i1
P
j=1
aij , i =
i,j
i,j,k,m
a43
a53
b3
yn+1 = yn + h b1 k1 b1 k1 +
1
5
1
bi c2i aij cj = 10
P
1
bi aij cj aik ck = 20
i,j,k
P
1
bi ci aij c2j = 15
i,j,k
P
1
bi aij c3j = 20
i,j
P
1
bi ci aij ajk ck = 30
i,j,k
P
1
bi aij cj ajk ck = 40
i,j,k
P
1
bi aij ajk c2k = 60
i,j,k
P
bi aij ajk akm cm =
a32
a42
a52
b2
i,j,k
sixth order
a21
a31
a41
a51
b1
1
46
, b1 = ,
45
45
b1 =
b2 =
1
(6a32 58a52 + 60 + 9a42
90 a32
+9a43 58a53 116a54 ),
b3 =
1
120
1
(58a52 + 60 + 9a42 + 9a43
90 a32
58a53 116a54 ),
1
29
, b5 = .
10
45
Substitute the values of b1 , b1 , b2 , b3 , b4 and b5
in the following reminder order conditions for the fifth
order method
b4 =
5 X
i1
X
bi ci aij cj =
i=3 j=2
5 X
i1
X
i=3 j=2
ISBN: 978-1-61804-039-8
131
bi aij c2j =
31
,
240
31
,
360
0.25
0.25
0.0082
0.2586
0.5
0.3860
0.5312
0.6444
0.75
0.2060
-0.9002
0.8918
0.5517
0.0222
1.0222
0.0403
0.1070
-0.1
j1
5 X
i1 X
X
i=4 j=3 k=2
0.6444
31
.
bi aij ajk ck =
720
a54 =
151
81
9
+
a43 a32 a53 a32 a53 ,
7424 3712
64
j1
5 X
i1 X
X
bi ci aij ajk ck
f8 =
j1
5 X
i1 X
X
Define
1
2
3
4
3
16
9
16
3
7
2
7
12
7
12
7
8
7
7
90
32
90
12
90
32
90
7
90
y(0) = 1,
y(x) = esin(x) .
Problem 2 (see [8])
y10 = 2y1 + y2 + 2sin(x),
y20 = y1 2y2 + 2(cos(x) sin(x)),
y1 (0) = 2, y2 (2) = 3,
1
,
30
Exact solution:
1
,
60
f62 + f82 .
Substituting the free parameters we obtained the values of other coefficients which are presented in table
3
ISBN: 978-1-61804-039-8
1
2
Exact solution:
q
=
1
8
y 0 = ycos(x),
16
.
29
1
8
To minimize the error norm of the sixth order method we substitute all the parameters into as
many as possible order conditions for the sixth order method. Here, we choose the second and third
equations from the order conditions of the sixth order
method as follows:
f6 =
1
4
29
31
9
a43 a32 a53 a32 a43 + ,
32
16
64
a54 =
1
4
5 Numerical examples
a42 =
1
4
132
methods have been developed for numerical integration of first order ordinary differential equations with
reduced number of function evaluations required per
step. The order conditions of the new method are derived up to order six and by satisfying the appropriate
order conditions we obtained the methods of order five
with the lower stages. The IRK5 method is almost
two-step in nature and is computationally more efficient compared to the existing classical RK5 method
.
12000
IRK5
RK5
10000
Function evaluations
8000
6000
References:
4000
2000
10
12
14
2.5
x 10
IRK5
RK5
Function evaluations
1.5
0.5
10
12
14
ISBN: 978-1-61804-039-8
133