1

A 2-D INTERPOLATION METHOD

THAT MINIMIZES AN ENERGY INTEGRAL

1
E. Rusyaman,
2
H. Gunawan,
1
A.K. Supriatna,
3
R.E. Siregar
Mathematics Dept
1
Padjadjaran University,
2
ITB,
3
Physics Dept Padjadjaran University
e-mail:
1
e_rusyaman@unpad.ac.id,
2
hgunawan@math.itb.ac.id


Abstract

This paper will present a method of interpolation which is used to construct a surface
-expressed as a continuous function of two variables- passing some arbitrary points on a
square domain. The function must minimize an energy integral of fractional order. To
construct such a function, the double Fourier sine series as well as the functional
analysis arguments are used. An iterative procedure to obtain the solution is also
presented.


Keywords : interpolation, energy integral, double Fourier sine series.

Presented at IICMA 2009 in Yogyakarta, October 12
th
- 13
th
, 2009 .




1. INTRODUCTION



Newton (1675), Lagrange (1795) and Hermite (1870) pioneered and developed the theory
and methods of interpolation. In this decade, interpolation methods continue to grow
rapidly. Some of them including the problem of energy minimizing interpolation can be
seen in [1, 5, 10] and other references.

In [7], a method to construct a surface u(x , y) on a square domain D = { (x , y)0 x 1 ,
0 y 1 }, passing the MN points (x
i
, y
j
, c
ij
), where 0 < x
1
< . . . < x
M
< 1, 0 < y
1
< . . .
< y
N
< 1 and u(0, y) = u(1 , y) = u(x , 0) = u(x , 1) = 0 has been proposed. The surface
must minimize the energy integral with fractional order:


=
1
0
1
0
2
2 /
) ( : ) ( dy dx u u E

.

For = 2, the integral is nothing but the total curvature of u on D. In reality, the available
data of (x
i
, y
j
, c
ij
) points in the domain, are not always homogeneous, but may scatter
randomly. This paper will develop an interpolation method in which the number of known
points is K and scatter randomly in region D.





2
2. FUNCTIONAL ANALYSIS FOR EXISTENCE OF SOLUTION


The problem of this research is how to construct a surface which is expressed as a two
variable continuous function z = u(x,y) such that it passes K given points (x
i
, y
i
, c
i
) where
0 x 1 , 0 y 1, with u(0 , y) = u(1 , y) = u(x , 0) = u(x , 1) = 0, and minimizes the
energy integral of fractional order E

(u) , where


=
1
0
1
0
2
2 /
) ( : ) ( dy dx u u E

. ... (1)

To find such a function, we will use the double Fourier sine series
u(x,y) = [ ]

= 1 1
sin . sin
m n
mn
y n x m a
.
Then, we have
E

(u) :=


1
0
1
0
2
2 /
) ( dxdy u

2
4
1
( )

=
+
1 1
2 2 2
m n
mn
n m a

.

Hence the coefficient of double Fourier sine series a
mn
must satisfy
( )

=
+
1 1
2 2 2
m n
mn
n m a

< .

To solve the problem, first we define
W

= { u(x,y) = [ ]

= 1 1
sin sin
m n
mn
y n x m a
( )

=
+
1 1
2 2 2
m n
mn
a n m

< }.

It has been shown in [4] that u(x,y) is continuous for > 1, and W

is Hilbert space with

the inner product:

= v u, ( )

=
+
1 1
2 2
m n
mn mn
n m b a

for u ,v e W

.

Furthermore we also define
V = { u(x,y) W

u (x
i
, y
i
) = 0 , i = 1 , . . . , K }

and
U = { u(x,y) W

u (x
i
, y
i
) = c
i
, i = 1 , . . . , K }

and obtain the following lemma.

Lemma: V is a closed subspace, while U is closed, convex, and not empty subset of W

.

Proof: The proof that V is closed, U closed and convex is similar to that in [7], so in this
paper we will only show that U is not an empty set.

Suppose that u(x,y) passes K given points: (x
1
, y
1
, c
1
), (x
2
, y
2
, c
2
), (x
3
, y
3
, c
3
), . . . ,
(x
K
, y
K
, c
K
). If x
i
x
j
and y
s
y
t
for each i, j, s, and t, (see Figure-1), then by
3
reordering x
i
and y
i
such that x
i
< x
i+1
and y
i
< y
i+1
, we have the fact that u(x,y) passes
the K
2
points (x
i
, y
j
,c
ij
) with i = 1,. . ., K and j = 1,. . ., K, (see Figure-2). Based on the
results in [8], the existence of a solution function u(x,y) passing the K
2
points is guaranteed
and this implies that this function also passes the given K points.


Figure-1 Figure-2


On the other hand, if x
i
= x
j
for some i and j, and or y
s
= y
t
for some s and t, (see Figure-3),
then by reordering x
i
and y
i
we will have M < K and N < K such that 0 < x
1
< x
2
< . . .< x
M

< 1 dan 0 < y
1
< y
2
< . . . < y
N
< 1, (see Figure-4). By referring to [8] then the existence of
u(x,y) passing MN points is guaranteed, and this implies that the function also passes the K
points.

Figure-3 Figure-4

From both cases above, the existence of u(x,y) U is guaranteed, and so that U is not
empty.

Theorem: The minimization problem (1) has a unique solution in W

, and this solution

is u = u
0
proj
V
(u
0
) with u
0
is an arbitrary element of U, and proj
V
(u
0
) is an
orthogonal projection of u
0
to V.


4
3. PROCEDURE FOR CONSTRUCTING THE FUNCTION: AN EXAMPLE

We shall describe the procedure to construct the function through an example.
Suppose that we want to find the function u(x,y) passing K = 3 given points :
A (x
1
, y
1
, c
1
) , B (x
2
, y
2
, c
2
) dan C (x
3
, y
3
, c
3
) . Based on lemma above, there is initial
function u
0
(x,y) passing K
2
= 9 points (x
i
, y
j
, c
ij
), namely:
u
0
(x,y) = [ ]

= =
3
1
3
1
sin . sin
m n
mn
y n x m a
= S
11
(x, y) + o
12
S
12
(x, y) + o
22
S
22
(x, y) +o
21
S
21
(x, y) +o
13
S
13
(x, y) +
o
23
S
23
(x, y) + o
33
S
33
(x, y) +o
32
S
32
(x, y) + o
31
S
31
(x, y).

where S
mn
(x, y) = sinmnx sinnny. This means that:

) , ( ... ) , ( ) , ( ) , ( ) , (
) , ( ... ) , ( ) , ( ) , ( ) , (
) , ( ... ) , ( ) , ( ) , ( ) , (
) , ( ... ) , ( ) , ( ) , ( ) , (
) , ( ... ) , ( ) , ( ) , ( ) , (
) , ( ... ) , ( ) , ( ) , ( ) , (
) , ( ... ) , ( ) , ( ) , ( ) , (
) , ( ... ) , ( ) , ( ) , ( ) , (
) , ( ) , ( ) , ( ) , ( ) , ( ) , ( ) , ( ) , ( ) , (
3 3 31 3 3 11 3 3 22 3 3 12 3 3 11
3 2 31 3 2 12 3 2 12 3 2 12 3 2 11
2 3 31 3 1 12 3 1 12 3 1 12 3 1 11
2 3 31 2 3 21 2 3 22 2 3 12 2 3 11
2 2 31 2 2 21 2 2 22 2 2 12 2 2 11
2 1 31 2 1 12 2 1 12 2 1 12 2 1 11
1 3 31 1 3 21 1 3 22 1 3 12 1 3 11
1 2 31 1 2 21 1 2 22 1 2 12 1 2 11
1 1 31 1 1 32 1 1 33 1 1 23 1 1 13 1 1 21 1 1 22 1 1 12 1 1 11
y x S y x S y x S y x S y x S
y x S y x S y x S y x S y x S
y x S y x S y x S y x S y x S
y x S y x S y x S y x S y x S
y x S y x S y x S y x S y x S
y x S y x S y x S y x S y x S
y x S y x S y x S y x S y x S
y x S y x S y x S y x S y x S
y x S y x S y x S y x S y x S y x S y x S y x S y x S
0

The following will show the existence of a more simple initial function u
0
(x,y) that only
contains three terms in accordance with the number of known points.

Note that the row vector-1, row vector-5, and row vector-9:
S
11
(x
1
, y
1
) S
12
(x
1
, y
1
) S
22
(x
1
, y
1
) S
21
(x
1
, y
1
) S
13
(x
1
, y
1
) S
23
(x
1
, y
1
) S
33
(x
1
, y
1
) S
32
(x
1
, y
1
) S
31
(x
1
, y
1
),
S
11
(x
2
, y
2
) S
12
(x
2
, y
2
) S
22
(x
2
, y
2
) S
21
(x
2
, y
2
) S
13
(x
2
, y
2
) S
23
(x
2
, y
2
) S
33
(x
2
, y
2
) S
32
(x
2
, y
2
) S
31
(x
2
, y
2
),
S
11
(x
3
, y
3
) S
12
(x
3
, y
3
) S
22
(x
3
, y
3
) S
21
(x
3
, y
3
) S
13
(x
3
, y
3
) S
23
(x
3
, y
3
) S
33
(x
3
, y
3
) S
32
(x
3
, y
3
) S
31
(x
3
, y
3
)

are linearly independent.
This mean that there are at least two columns ij and kl such that for these ij and kl we
have
S
]
(x
1
, y
1
) S
kI
(x
1
, y
1
) S
st
(x
1
, y
1
)
S
]
(x
2
, y
2
) S
kI
(x
2
, y
2
) S
st
(x
2
, y
2
) 0 for st ij and st kl.
S
]
(x
3
, y
3
) S
kI
(x
3
, y
3
) S
st
(x
3
, y
3
)
So there are at least 9 2 = 7 pairs with determinant 0.
In this case, the initial function u
0
(x , y) passing K = 3 points is a function with three terms:

u
0
(x , y) = o
]
sin inx. sin ]ny + o
kI
sin knx. sin lny + o
st
sin mnx. sin nny ...(*)

Furthermore, to determine the basis of
V = { u(x,y) W

u (x
1
, y
1
) = 0 , u (x
2
, y
2
) = 0 , u (x
3
, y
3
) = 0 },
5
take an arbitrary function: u(x , y) V :

u(x , y) = y x a sin . sin
11
+ y x a 2 sin . sin
12
+ y x a 2 sin . 2 sin
22
+ y x a sin . 2 sin
21
+
y x a 3 sin . sin
13
+ y x a 3 sin . 2 sin
23
+ y x a 3 sin . 3 sin
33
+ y x a 2 sin . 3 sin
32
+
y x a sin . 3 sin
31
+ y x a 4 sin . sin
14
+ y x a 4 sin . 2 sin
24

+ y x a 4 sin . 3 sin
34
+ . . .

Since u(x
i
, y
i
) = 0 for i = 1, 2, 3, then we have three equations:
0 =
i i
y x a sin . sin
11
+
i i
y x a 2 sin . sin
12
+
i i
y x a 2 sin . 2 sin
22
+
i i
y x a sin . 2 sin
21

+
i i
y x a 3 sin . sin
13
+
i i
y x a 3 sin . 2 sin
23
+
i i
y x a 3 sin . 3 sin
33
+
i i
y x a 2 sin . 3 sin
32

+
i i
y x a sin . 3 sin
31
+
i i
y x a 4 sin . sin
14
+
i i
y x a 4 sin . 2 sin
24

+ . . .

If the notation of
k k
y n x m sin . sin is simplified by
k
mn
s then we have
0 = ... . . . . . . . . .
1
31 31
1
32 32
1
33 33
1
23 23
1
13 13
1
21 21
1
22 22
1
12 12
1
11 11
+ + + + + + + + + s a s a s a s a s a s a s a s a s a
0 = ... . . . . . . . . .
2
31 31
2
32 32
2
33 33
2
23 23
2
13 13
2
21 21
2
22 22
2
12 12
2
11 11
+ + + + + + + + + s a s a s a s a s a s a s a s a s a
0 = ... . . . . . . . . .
3
31 31
3
32 32
3
33 33
3
23 23
3
13 13
3
21 21
3
22 22
3
12 12
3
11 11
+ + + + + + + + + s a s a s a s a s a s a s a s a s a

From (*), we assume that a
ij
, a
kl
, and a
st
are the dependent variables, while a
mn
are the
independent variables. Other steps such as finding the basis, orthogonalization process,
finding solution function, and the calculation of energy are similar to the steps described in
[7].

A concrete example to find the function is presented as follows.

Suppose that we wish to construct a function u(x, y) passing three points A = (1/3 , 1/3 , 1),
B = ( 1/2 , 2/3 , 2 ), and C = ( 2/3 , 1/2 , 1 ), with u(0 , y) = u(1 , y) = u(x , 0) = u(x , 1) = 0
and minimize (1) for = 1,25.

The steps undertaken are as follows:

Step-1: Determining initial function u
0
(x , y)

By using the above concept of the existence of the initial function with three terms, we can
determine the fixed coefficient of initial function u
0
(x , y).
Let these coefficients are a
11
, a
12
, and a
13
. So we take
u
0
(x,y) = y x a sin . sin
11
+ y x a 2 sin . sin
12
+ y x a 3 sin . sin
13

Since u
0
(x,y) passing A, B, and C, then we have
1 =
3
4
a
11
+
3
4
a
12
+ 0 a
13

2 =
1
2
S a
11
-
1
2
S a
12
+ 0 a
13

1 =
1
2
S a
11
-
1
2
S a
13
,

Whence : a
11
= 1.821367 , a
12
= - 0.488034 , and a
13
= 0.66667.

Hence the initial function is:
6
u
0
(x , y) = 1.821367 y x sin . sin - 0.488034 y x 2 sin . sin + 0.66667 y x 3 sin . sin

=

% # # # #
... 0 0 0 0
... 0 0 0 0
... 0 0 0 0
... 0 66667 . 0 488034 . 0 1.821367



Step-2: Determining Basis of V = {u(x,y)W
1.25
u (x
1
, y
1
) = u (x
2
, y
2
) = u (x
3
, y
3
) = 0 }.

Take arbitrary u(x , y) V where

u(x , y) = y x a sin . sin
11
+ y x a 2 sin . sin
12
+ y x a 2 sin . 2 sin
22
+ y x a sin . 2 sin
21
+
y x a 3 sin . sin
13
+ y x a 3 sin . 2 sin
23
+ y x a 3 sin . 3 sin
33
+ y x a 2 sin . 3 sin
32

+ y x a sin . 3 sin
31
+ y x a 4 sin . sin
14
+ y x a 4 sin . 2 sin
24
+ . . .

Since u( 1/3 , 1/3 ) = 0 , u( 1/2 , 2/3) = 0 , and u( 2/3 , 1/2) = 0 then
0 =
3
4
a
11
+
3
4
a
12
+
3
4
a
22
+
3
4
a
21
-
3
4
a
14
-
3
4
a
24
+
3
4
a
44
-
3
4
a
42
-
3
4
a
41
+ . . .
0 =
1
2
S a
11
-
1
2
S a
12
+
1
2
S a
32
-
1
2
S a
31
+
1
2
S a
14
-
1
2
S a
34
-
1
2
S a
15
+

. . .
0 =
1
2
S a
11
-
1
2
S a
21
-
1
2
S a
13
+
1
2
S a
23
-
1
2
S a
43
+
1
2
S a
41
+
1
2
S a
15
+ . . .

and since coefficients of initial function u
0
(x , y) are a
11
, a
12
, and a
13
, then we assume
a
11
, a
12
, and a
13
as the dependent variable such that we have :
a
11
+ a
12
= - a
22
- a
21
+ a
14
+ a
24
- a
44
+ a
42
+ a
41
- . . .
a
11
- a
12
= - a
32
+ a
31
- a
14
+ a
34
+ a
15
+

. . .
a
11
- a
13
= + a
21
- a
23
+ a
43
- a
41
- a
15
+ . . .

With elimination, we obtain:
a
11
= -
1
2
a
22
-
1
2
a
21
-
1
2
a
32
+
1
2
a
31
+
1
2
a
24
+
1
2
a
34
-
1
2
a
44
+
1
2
a
42
+
1
2
a
41
+
1
2
a
15
+

. . .
a
12
= -
1
2
a
22
-
1
2
a
21
+
1
2
a
32
-
1
2
a
31
+

a
14
+
1
2
a
24
-
1
2
a
34
-
1
2
a
44
+
1
2
a
42
+
1
2
a
41
-
1
2
a
15
+

. . .
a
13
= -
1
2
a
22
-
3
2
a
21 +
a
23
-
1
2
a
32
+
1
2
a
31
+
1
2
a
24
+
1
2
a
34
-
1
2
a
44
- a
43
+
1
2
a
42
+
3
2
a
41
+
3
2
a
15
+

. . .

In matrix equation, we can write:

% ... ... ... ...

...
...
...
...
44 43 42 41
34 33 32 31
24 23 22 21
14 13 12 11
a a a a
a a a a
a a a a
a a a a
= a
22

% ... ... ... ...
... 0 0 0 0
... 0 0 0 0
... 0 0 1 0
... 0 5 . 0 5 . 0 5 . 0

+ a
21

% ... ... ... ...
... 0 0 0 0
... 0 0 0 0
... 0 0 0 1
... 0 5 . 1 5 . 0 5 . 0

+ a
23

% ... ... ... ...

... 0 0 0 0
... 0 0 0 0
... 0 1 0 0
... 0 1 0 0
+ a
33

% ... ... ... ...

... 0 0 0 0
... 0 1 0 0
... 0 0 0 0
... 0 0 0 0
+ a
32

% ... ... ... ...
... 0 0 0 0
... 0 0 1 0
... 0 0 0 0
... 0 5 . 0 5 . 0 5 . 0

7
+ a
31

% ... ... ... ...
... 0 0 0 0
... 0 0 0 1
... 0 0 0 0
... 0 5 . 0 5 . 0 5 . 0

+ a
14

% ... ... ... ...

... 0 0 0 0
... 0 0 0 0
... 0 0 0 0
... 1 0 1 0
+ a
24

% ... ... ... ...

... 0 0 0 0
... 0 0 0 0
... 1 0 0 0
... 0 5 . 0 5 . 0 5 . 0


+ a
34

% ... ... ... ...
... 0 0 0 0
... 1 0 0 0
... 0 0 0 0
... 0 5 . 0 5 . 0 5 . 0
+ a
44

% ... ... ... ...
... 1 0 0 0
... 0 0 0 0
... 0 0 0 0
... 0 5 . 0 5 . 0 5 . 0

+ a
43

% ... ... ... ...

... 0 1 0 0
... 0 0 0 0
... 0 0 0 0
... 0 0 0 0
+ . . .

= a
22
v
1
+ a
21
v
2
+ a
23
v
3
+ a
33
v
4
+ . . .

such that the above matrix { v
1
, v
2
, v
3
, v
4
, . . . } is the basis of V.


Step-3: Orthogonalizing Basis of V

Since the above basis is not orthogonal, then we need to orthogonalize this basis to become
an orthogonal basis { w
1
, w
2
, w
3
, w
4
, . . . }. Here

w
1
= v
1
=

% ... ... ... ...
... 0 0 0 0
... 0 0 0 0
... 0 0 1 0
... 0 5 . 0 5 . 0 5 . 0


w
2
= v
2
-
1
1 1
1 2
.
,
,
w
w w
w v


= v
2
-
(2
1.2S
. + 5
1.2S
. + 10
1.2S
.)
(2
1.2S
. + 5
1.2S
. + 10
1.2S
. +8
1.2S
.1)
w
1


=

... ... ... ... ...
... 0 0 0 0
... 0 0 0 0
... 0 0 775929 . 0 1
... 0 11204 . 1 112036 . 0 112036 . 0


and for k N
w
k
= v
k
-
1
1 1
1
.
,
,
w
w w
w v
k


-
2
2 2
2
.
,
,
w
w w
w v
k


- . . . -
1
1 1
1
.
,
,

k
k k
k k
w
w w
w v


with = v u, ( )
mn mn
m n
b a n m .
1 1
25 . 1
2 2

=
+ is inner product on W
1.25
.


Step-4: Determining an Approximate Function: u
n


From step-1 we have initial function u
0
(x,y) U is
8

u
0
(x , y) =

% ... ... ... ...
... 0 0 0 0
... 0 0 0 0
... 0 0 0 0
... 0 66667 . 0 488034 . 0 1.821367


The first approximate function from the first iteration is an orthogonal projection of u
0
to
V
1
= span { w
1
}:
u
1
= u
0
-
1
1 1
1 0
,
,
w
w w
w u


=

% ... ... ... ...
... 0 0 0 0
... 0 0 0 0
... 0 0 308 . 0 0
... 0 513 . 0 642 . 0 1.667


The second approximate function from the second iteration is an orthogonal projection of
u
0
to V
2
= span { w
1
, w
2
}:
u
2
= u
0
-
1
1 1
1 0
,
,
w
w w
w u


-
2
2 2
2 0
,
,
w
w w
w u


= u
1
-
2
2 2
2 0
,
,
w
w w
w u


,
and so on.


Step-5: Determining The Energy Integral Value: E
1.25
(u)

From (1) with = 1.25 we have

=
1
0
1
0
2
625 . 0
25 . 1
) ( : ) ( dy dx u u E
=
5 .. 2
.
4
1
( )

=
+
1 1
25 .. 1
2 2 2
m n
mn
n m a ,
hence for solution function u
1
:
E
1.25
(u
1
) =
5 . 2
.
4
1
. [ (1+1)
1.25
.(1.667)
2
+ (1+4)
1.25
.(-0.642)
2
+ (1+9)
1.25
.(0.513)
2

+ (4+4)
1.25
.(0.308)
2
]

= 3.91109
2.5

Without writing the constant
2.5
, with a computer program, we obtain the results as
follows:

E
1.25
(u
2
) =2.74487 E
1.25
(u
3
) = 2.70349 E
1.25
(u
4
) = 2.70349 . . .
E
1.25
(u
1218
) =1.42688 E
1.25
(u
1219
) = 1.42682 E
1.25
(u
1220
) = 1.42680
E
1.25
(u
1221
) = 1.42672.

Finally if we take an error factor = 0.0001, then we can stop at iteration-1221, with the
last approximate function is u
1221
and the minimum energy is 1.4267.

The second example for = 1.1 produces an approximate function u
1220
, and its minimum
energy value is 1.01954 with an error factor = 0.0001.
The last example we take = 2.0 , and the approximate function is u
1222
, whose minimum
energy value is 4.8471 with an error factor = 0.0001.

The graphs of the approximate functions are:
9

For = 1.1 :

u

u
147
u
122



For = 1.25 :

u
1
u
25
u
1221

For = 2.0 :

u
1
u
8
u
1222



Note that from the three examples above, we see that different values of give different
function or surfaces that minimizes the energy. The larger the value of , the smoother the
surface.


10

Acknowledgement

We would like to thank Akmal, M.T, from Mathematics Department, Padjadjaran
University, for useful discussion concerning computer programming aspect. H. Gunawan is
supported by ITB Research Grant No. 252/2009.





Reference

[1] Atkinson.K & Han.W, Theorical Numerical Analysis A Functional Analysis Framework,
Springer Verlag, New York, 2001.

[2] Folland. G.B, Fourier Analysis And Its Applications, Wadsworth & Brooks/Cole Advanced
Books & Software, California. York, 1992.

[3] Gunawan.H, Pranolo.F, Rusyaman.E, An Interpolation Method That Minimizes an Energy
Integral of Fractional Order. Singapore: Proceeding of Asian Symposium of Computer
Mathematics, 2007.

[4] Kaw.A. & Kateltas, 2008, History of Interpolation, http://numericalmethods.eng.usf.edu

[5] Langhaar.H.L, Energy Methods in Applied Mechanics, John Wiley & Sons, New York, 1962.

[6] Rusyaman.E. and Gunawan.H, Permukaan yang meminimumkan Kurvatur, Mathematics
Nasional Conference, Palembang, 2008

[7] Rusyaman. E, Gunawan.H, and Ambarwati. L, Surfaces With Prescribed Nodes and Minimum
Energy of Fractional Order, 2009, (submitted).

[8] Rusyaman.E and Gunawan.H, Supriatna.AK, Siregar.R, Syarat Cukup Bagi Eksistensi Deret
Sinus Fourier Ganda Pada Masalah Interpolasi, Analysis Symposium, Brawijaya University,
Malang, 2009.

[9] Tolstov.G.P, Fourier Series, New York, Dover Publications, Inc, (1976).

[10] Wan.W.L, Chan.T.F, Barry Smith, An Energy-Minimizing Interpolation for Robust Multigrid
Methods, Siam.J.ScI Compute, vol.21 no.4, 2000.