) (
) (
)
(
) (
) (
)
(
) (
) (
)
Fig.1: 9-point Computational Network
2. The approximation based on cubic spline polynomial
We consider our region of interest, a rectangular domain = 0, 1 0, 1 . We choose grid
spacing > 0 and > 0 in the directions - and - respectively, so that the mesh points
(
= and
= ; = 0,1, , + 1; =
0,1, , + 1, where and are positive integers such that (N+1)h = 1 and ( + 1) = 1.
The mesh ratio parameter is denoted by ( / ) k h = . The notations
, ,
and
l m l m
u U are used for the
discrete approximation and the exact value of ( , ) u x y at the grid point ( , )
l m
x y , respectively.
At the grid point ( , )
l m
x y , we denote
a b
ab
a b
l m
W
W
x y
+
=
, W = U, D and g (3)
Let ( )
m
S x be the cubic spline interpolating polynomial of the function ( , )
m
u x y between the
grid point
1
( , )
l m
x y
and ( , )
l m
x y , and is given by
3 3 2 2
1 1
1, , 1, 1, , ,
( ) ( )
( ) ,
6 6 6 6
l l l l
m l m l m l m l m l m l m
x x x x x x x x h h
S x M M U M U M
h h h h
| | | | | | | |
= + + +
| | | |
\ \
\ \
1
; 1, 2,..., 1 ; 0,1, 2,..., 1
l l
x x x l N m M
= + = + (4)
which satisfies at mth-line parallel to x-axis the following properties:
(i) ( )
m
S x coincides with a polynomial of degree three on each
1
[ , ],
l l
x x
1, 2,..., 1; 1, 2, ... , l N m M = + =
(ii)
2
( ) [0,1],
m
S x C and
(iii)
,
( ) , 0,1, 2,..., 1; 1, 2,...,
m l l m
S x U l N m M = = + = .
The derivatives of cubic spline function ( )
m
S x are given by
2 2
, 1,
1
1, , , 1,
( ) ( )
( )
2 2 6
l m l m
l l
m l m l m l m l m
U U
x x x x h
S x M M M M
h h h
( = + +
, (5)
1
1, ,
( ) ( )
( )
l l
m l m l m
x x x x
S x M M
h h
= + (6)
where
, , , ,
,
( ) , 0,1, 2,..., 1; 1, 2,...,
l m m l xx yy x l m l m l m
l m
l
M S x U U U f l N j J
x
= = = = + = , (7)
, 1,
, 1, , 1 ,
( ) 2 ,
6
l m l m
l m m l x l m l m l l l m
U U
h
m S x U M M x x x
h
( = = = + +
, (8)
and replacing h by h, we get
1, ,
, 1, , 1 ,
( ) 2 ,
6
l m l m
l m m l x l m l m l l l m
U U
h
m S x U M M x x x
h
+
+
( = = = +
, (9)
Combining (8) and (9), we obtain
1, 1,
, 1, 1, ,
( )
2 12
l m l m
l m m l x l m l m l m
U U
h
m S x U M M
h
+
+
( = = =
, (10)
Further, from (8), we have
1, ,
1, 1 . 1, 1,
( ) 2
6
l m l m
l m m l x l m l m l m
U U
h
m S x U M M
h
+
+ + + +
( = = = + +
, (11)
and from (9), we have
, 1,
1, 1 , 1, 1,
( ) 2
6
l m l m
l m m l x l m l m l m
U U
h
m S x U M M
h
( = = = +
, (12)
We consider the following approximations:
( )
2 2
, 1 , , 1
, ,
2 /( ) ( )
yy l m l m l m yy
l m l m
U U U U k U O k
+
= + = + , (13.1)
( )
2 2
1, 1 1, 1, 1
1, 1,
2 /( ) ( )
yy l m l m l m yy
l m l m
U U U U k U O k
+ + + +
+ +
= + = + , (13.2)
( )
2 2
1, 1 1, 1, 1
1, 1,
2 /( ) ( )
yy l m l m l m yy
l m l m
U U U U k U O k
+
= + = + , (13.3)
( )
2
4
1, 1, 30 , ,
/(2 ) ( )
6
x l m l m x l m l m
h
U U U h U U O h
+
= = + + , (14.1)
( )
2
3
1, , 1, 30 1, 1,
3 4 /(2 ) ( )
3
x l m l m l m x l m l m
h
U U U U h U U O h
+ + +
= + = + , (14.2)
( )
2
3
1, , 1, 30 1, 1,
3 4 /(2 ) ( )
3
x l m l m l m x l m l m
h
U U U U h U U O h
+
= + = + , (14.3)
Since the derivative values of ( )
m
S x defined by (7), (10), (11) and (12) are not known at
each grid point ( , )
l m
x y , we use the following approximations for the derivatives of ( )
m
S x .
Let
, , ,
,
l m yy l x l m l m
l m
M U DU g = + + , (15.1)
1, 1 1, 1,
1,
l m yy l x l m l m
l m
M U D U g
+ + + +
+
= + + , (15.2)
1, 1 1, 1,
1,
l m yy l x l m l m
l m
M U D U g
= + + , (15.3)
1, ,
, 1, 1,
2
6
l m l m
x l m l m l m
U U
h
U M M
h
+
+ +
( = + +
, (16.1)
, 1,
, 1, 1,
2
6
l m l m
x l m l m l m
U U
h
U M M
h
( = +
, (16.2)
1, 1 1, 1, l m l x l m l m
F D U g
+ + + +
= + , (16.3)
1, 1 1, 1, l m l x l m l m
F D U g
= + (16.4)
1, 1, , ,
1, 1,
12 12
x x l m l m yy yy l m l m
l m l m
h h
U U F F U U
+
+
(
( = +
(16.5)
, , ,
l m l x l m l m
F DU g = + (16.6)
Then at each grid point ( , )
l m
x y , a cubic spline finite difference method of Numerov type
with accuracy of
2 4
( ) O k h + for the solution of differential equation (1) may be written as
2 2
2 2
, 1 1 1, 1, ,
1, 1, ,
10 10
12 12
x l m yy yy yy l x l x l x l m l m l m
l m l m l m
k k
U U U U D U D U DU
+ +
+
( (
+ + + = + +
2
1, 1, ,
10
12
l m l m l m
k
g g g
+
( + + +
+
, l m
T . (17)
where the local truncation error
4 2 4
,
( )
l m
T O k k h = + . Note that the method (17) is of
2 4
( ) O k h + for the numerical solution of equation (1). However the method (17) fail to
compute at 1 l = , when D(x) and/or g(x,y) contains the singular terms like
1
x
,
2
1
x
, etc. For
example, if D(x) =
1
x
, this implies
1
1
1
l
l x
D
= + (18.2)
2
3
1, , 10 20 2
( )
h
l m l m
g g hg g O h
+
= + + + , (18.3)
2
3
1, , 10 20 2
( )
h
l m l m
g g hg g O h
= + , (18.4)
Substituting the approximations (18.1)-(18.4) into (17) and neglecting higher order
terms and local truncation error, we get
( )( )
( )
2
2 2 2 2 2 2 2 2 2
, , 00 20 , 10 00 ,
(12 ) 12 12 2
2
x y l m x l m x x l m x l m
h
u u D h D u h D D u
+ + = + +
( ) ( )
2 2 2 2 2 00
10 , , 00 20 10 00 00 10
2 12
2
y l m y x x l m
hD
h D u u k g h g D g D g ( + + + +
(19)
where
( )
1 1
2 2
1
2 x l l l
u u u
+
= + and
( )
1 1
2 2
x l l l
u u u
+
= . Note that, the cubic spline method
(19) is of
2 4
( ) O k h + for the numerical solution of elliptic equation equation (1), which is
also free from the terms
1
1
l
x
( + ) and
= ( )
(24)
The matrix has block tri-diagonal form and hence is block consistently ordered (see Varga
[22] ).
It can be verified that
0
>0 and
j
<0 for = 1,2, . . ,8 provided 1 R < . One can also verify that
8
0
1
i
i
=
=
(25)
which implies that is weakly diagonally dominant. Since is reducible, we conclude that it is
an -matrix (see Varga [22]).
Applying the Jacobi iterative method to the scheme (21), we get the iterative scheme
1, 1 1, 1, 1
2
( ) 2 ( ) ( )
(1 ) 12 (1 ) 2(1 ) (1 )
3
l m l m l m
k k k
R
R u R R u R u
+ + + +
(
+ + +
(
, 1 , , 1 1, 1
2
( ) 2 ( 1) ( ) ( )
10 [20 24 (1 )] 10 (1 )
3
l m l m l m l m
k k k k
R
u u u R u
+ +
+
+ + + + + +
( )
1, 1, 1
2
2 ( ) ( )
12 1 2 1 (1 ) 0
3
l m l m
k k
R
R R u R u
( | |
+ + + + + + =
( |
\
. (26)
The propagating factors for the Jacobi iterative methods is given by
2
2
2
2
1
5cos( ) 6 1 (1 )(1 cos )
3
5 6 1
3
j
R
k p R R k
R
p
| |
| |
| = + + + +
|
|
| |
\
\ + +
|
\
2
2
6 1 (1 )(1 cos ) cos
3
R
p R R k h
( | |
| |
( | +
|
|
(
\
\
(27)
where (M+1)h=1, (N+1)k = 1. Consequently, the spectral radii of the Jacobi and Gauss-Seidel
matrices are related by
(
) = (
(28)
Hence, the associated iteration
()
=
()
+ (29)
converges for any initial guess, where G is either Jacobi or Gauss-Seidel iteration matrix. The
Jacobi and Gauss-Seidel splitting are regular for both the line and point versions, and hence they
converge for any initial guess.
4. Numerical Results
If we replace the partial derivatives in equation (1) by the central difference approximations at
the grid point ( , ) l m , we obtain a central difference scheme (CDS) which is of
2 2
( ) O k h + . We now
solve the following two benchmark problems whose exact solutions are known. The right-hand side
hogrneous function and boundary conditions may be obtained by using the exact solution as a test
procedure. We use block Gauss-Seidel iterative method (See [22-25]) to solve the proposed scheme
(19). In all cases, we have considered
(0)
= u 0 as the initial guess and the iterations were stopped
when the absolute error tolerance
( 1) ( ) 12
10
k k +
u u was achieved. In all cases, we have calculated
maximum absolute errors ( l
-norm) for different grid sizes. All computations were performed using
double precision arithmetic.
Example 1 (Convection-diffusion equation)
The problem is to solve (20) in the solution region 0 < , < 1 whose exact solution is
(, ) =
+ (1 ) , where
and 0 > .
The maximum absolute errors for are tabulated in Table 1.1 and Table 1.2.
Example 2 (Poissons equation in polar cylindrical coordinates)
2 2
2 2
1
( , )
u u u
f r z
r z r r
+ + =
, 0<r, z<1 (30)
The exact solutions are given by u(r, z) = r
2
sinhr coshz. The maximum absolute errors for
are tabulated in Table 2.1 and Table 2.2.
5. Concluding remarks
Available numerical methods based on spline approximations for the numerical solution of
2D Poissons equation are of
2 2
( ) O k h + accurate, which require nine grid points. In this article,
using the same number of grid points , we have discussed a new stable compact nine point cubic
spline finite difference method of
2 4
( ) O k h + accuracy for the solution of Poissons equation in
polar cylindrical coordinates. For a fixed parameter
2
k
h
= , the proposed method behaves like a
fourth order method, which is exhibited from the computed results.
References
[1] R.E. Lynch and J.R. Rice, High accuracy finite difference approximation to solutions
of elliptic partial differential equations, Proceedings of National Academy of Sciences,
U.S.A., vol. 75, pp. 2541-2572, (1978).
[2] R.F. Boisvert, Families of High order Accurate discretization of some elliptic problems,
SIAM. J. Sci. Stat. Comput., vol. 2, pp. 268-285, (1981).
[3] I. Yavneh, Analysis of Fourth-order compact scheme for convection diffusion equation, J.
Comput. Phys., vol. 133, pp. 361-373, (1997).
[4] M. K. Jain, R.K. Jain and R.K. Mohanty, A Fourth Order Difference method for elliptic
equations with Non Linear first derivative terms, Numer. Meth. Partial Diff. Eqs., vol. 5, pp.
87-95, (1989).
[5] M. K. Jain, R.K. Jain and R.K. Mohanty, Fourth Order Difference methods for the system of
2-D nonlinear elliptic partial differential equations, Numer. Meth. Partial Diff. Eqs., vol. 7,
pp. 227-244, (1991).
[6] R. K. Mohanty, Order
h
Proposed
2 4
( ) O k h + - method
4 4
( ) O k h + - method discussed in [8]
=5 =10 =15 =5 =10 =15
1
16
0.1808E-01
0.1636E-01
0.1494E-01
0.4408E-01
0.4212E-01
0.4018E-01
1
32
0.1129E-02
0.1026E-02
0.9396E-03
0.2692E-02
0.2554E-02
0.2482E-02
1
64
0.7054E-04
0.6419E-04
0.5881E-04
0.1645E-03
0.1566E-03
0.1512E-03
Table 2.1
Example 2: The maximum absolute errors
(h, k)
Proposed
2 4
( ) O k h + - method
2 2
( ) O k h + - method
( )
1 1
20 10
,
0.3986E-03
0.1133E-02
( )
1 1
40 20
,
0.1089E-03
0.3003E-03
( )
1 1
20 40
,
0.4117E-03
0.1132E-02
( )
1 1
80 40
,
0.2893E-04
0.7836E-04
( )
1 1
40 80
,
0.1116E-03
0.2984E-03
Table 2.2
Example 2: The maximum absolute errors
2
( 64)
k
h
=
h
Proposed
2 4
( ) O k h + - method
4 4
( ) O k h + - method discussed
in [8]
1
16
0.2711E-02
0.4842E-02
1
32
0.1654E-03
0.2988E-03
1
64
0.1016E-04
0.1811E-04