Advances in Applied Mathematics and Mechanics DOI: 10.4208/aamm.

09-m09S01
Adv. Appl. Math. Mech., Vol. 1, No. 6, pp. 750-768 December 2009

A Revisit on the Derivation of the Particular

Solution for the Differential Operator ∆2 ± λ2
Guangming Yao1 , C. S. Chen1, ∗ and Chia Cheng Tsai2
1 Department of Mathematics, University of Southern Mississippi, Hattiesburg, MS
39406, USA
2 Department of Marine Environment Engineering, National Kaohsiung Marine

University, Taiwan
Received 28 February 2009; Accepted (in revised version) 25 August 2009
Available online 18 November 2009

Abstract. In this paper, we applied the polyharmonic splines as the basis functions
to derive particular solutions for the differential operator ∆2 ± λ2 . Similar to the
derivation of fundamental solutions, it is non-trivial to derive particular solutions
for higher order differential operators. In this paper, we provide a simple algebraic
factorization approach to derive particular solutions for these types of differential
operators in 2D and 3D. The main focus of this paper is its simplicity in the sense
that minimal mathematical background is required for numerically solving higher
order partial differential equations such as thin plate vibration. Three numerical
examples in both 2D and 3D are given to validate particular solutions we derived.
AMS subject classifications: 35J05, 35J25, 65D05, 65D15
Key words: The method of fundamental solutions, radial basis functions, meshless methods,
polyharmonic splines, the method of particular solutions.

1 Introduction
The idea of splitting a given partial differential equation into solving a homogeneous
equation and an inhomogeneous equation is well known. In recently years, such ap-
proach becomes very popular for various boundary meshless methods such as the
Trefftz method, the method of fundamental solutions [8, 9], and the boundary knot
method (BKM) [4], etc. By evaluating the particular solution, these boundary mesh-
less methods can be extended from solving only homogeneous equations to inho-
mogeneous equations and time-dependent problems [1, 9]. As a consequence, many

∗ Corresponding author.
URL: http://www.math.usm.edu/cschen/
Email: [email protected] (G. Yao), [email protected] (C. S. Chen), tsaichiacheng@mail.
nkmu.edu.tw (C. C. Tsai)

http://www.global-sci.org/aamm 750 c
°2009 Global Science Press
G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768 751

numerical techniques have been developed to evaluate particular solutions for vari-
ous types of partial differential equations. Chen and Rashed [2] were first to extend
the derivation of particular solutions to Helmholtz-type equations using thin plate
splines. Muleshkov et al. [11] further extended the concept to polyharmonic splines.
However, the derivation of particular solutions using the annihilator method and al-
gebraic techniques in [11] were too tedious to use for solving complicated differential
operators. Cheng [6] revisited the problem using the technique of fundamental so-
lutions so that particular solutions can be easily derived. Recently, Muleshkov and
Golberg [12], and Chen et al. [3] derived particular solution for more complicate dif-
ferential operators using radial basis functions and Chebyshev polynomials. In 2009,
Tsai et al. [13] extended the derivations of particular solutions to polyharmonic, poly-
Helmholtz operators and their products.
In contrast to the tedious derivation of particular solutions for Helmholtz-type
differential operators shown in [11] and its extension to general operators [13], we
propose a simple algebraic factorization approach to derive particular solutions for
the differential operators
∆2 ± λ2 ,
in 2D [5] and 3D using polyharmonic splines. On the other hand, Young et al. [14]
solved the homogeneous equation of plate vibration problem in which ∆2 −λ2 is the
differential operator. Coupled with the particular solutions derived in this paper, [14]
can be effectively extened to solving the arbitrarily loaded flexural vibrations of an
uniform thin plate.
This paper is organized as follows. In section 2, we derive particular solutions for
polyharmonic splines which includes two dimensional and three dimensional cases.
In Section 3, we derive particular solutions for the monomial term for ∆2 −λ2 . In Sec-
tion 4, numerical examples for two 2D examples and one 3D example are given. In
Section 5, we conclude this paper with opening issues and future applications.

2 Particular solutions for polyharmonic splines

Let us consider the following boundary value problem
¡ 2 ¢
∆ − λ2 u = f ( x ), x ∈ Ω ⊂ Rd , (2.1)
B1 u = g ( x ), x ∈ ∂Ω, (2.2)
B2 u = h ( x ), x ∈ ∂Ω, (2.3)

where λ is a non-zero constant, ∆ is the Laplacian, B1 and B2 are the boundary dif-
ferential operators, f , g and h are given functions, and Ω is an open bounded domain
in Rd , d=2, 3, with boundary ∂Ω. Note that x=( x, y) in 2D and x=( x, y, z) in 3D. For
d=2, (2.1)–(2.3) govern the loaded flexural vibrations of a uniform thin plate.
Let u p be a particular solution of the governing equation, then it satisfies
¡ 2 ¢
∆ − λ2 u p = f ( x ), (2.4)
752 G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768

but does not necessarily satisfy the boundary condition. If u p in (2.4) can be obtained,
then (2.1) –(2.3) can be reduced to the following homogeneous equation through the
variable substitution w=u − u p , i.e.,
¡ ¢
∆2 − λ2 w = 0, x ∈ Ω, (2.5)
B w = g(x) − u p , x ∈ ∂Ω. (2.6)

The above homogeneous equation can be easily solved using boundary methods such
as the Trefftz method [1], the method of fundamental solutions [8,9], and the boundary
knot method (BKM) [4], etc. The key problem to be considered is how to obtain an
approximation to the particular solution u p .
For arbitrary function f (x), it is difficult, if not impossible, to obtain u p . Therefore,
it is essential to approximate the source term f (x) by a series of basis functions [9].
In this section, we focus on the derivation of the closed-form approximate particular
solution of (2.4).
Let PQd −1 denote the space spanned by all d-variate polynomials of degree up to
q
Q−1 and pick a basis { p` }`=1 of this space. The dimension of q is
µ ¶
Q−1+d
.
d

To find approximate particular solutions u p to

¡ ¢
∆2 − λ2 u p = f ( x ), (2.7)

we approximate f by fˆ as follows

m q
fˆ(x) = ∑ α j φ (r j ) + ∑ β ` p ` ( x ), (2.8)
j =1 `=1

° ° © ªm
where r j =°x − x j ° , x j 1 is a unisolvent set of points for polynomial interpolation.
© ª
ϕ in (2.8) is radial basis functions. The coefficients α j and { β ` } can be determined
© ªm
by interpolating f by fˆ on x j 1 . The following additional conditions for (2.8) are
required
m
∑ α j p` (x) = 0, 1 ≤ ` ≤ q. (2.9)
j =1

As was shown by Duchon [7], (2.8)–(2.9) has a unique solution. Then, the approximate
particular solution can be obtained in the following way

m q
u p (x) ' û p (x) = ∑ a j Φ (r j ) + ∑ β ` χ ` ( x ), (2.10)
j =1 `=1
G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768 753

where
¡ ¢
∆2 − λ2 Φ (r j ) = φ (r j ), (2.11)
¡ 2 ¢
∆ − λ2 χ ` = p ` . (2.12)

In this paper we apply the following polyharmonic splines to approximate f (x) in

(2.7)
( 2n
r ln r, in R2 ,
(n)
φ (r ) : = ϕ (r ) = n ≥ 1.
r2n−1 , in R3 ,
The reason that we choose φ to be polyharmonic splines is due to its unique property
that ∆ L r2n ln(r )=0 for L sufficiently large [11].

2.1 Two dimensional case

In this subsection, we consider the derivation of the approximate particular solution
of the following equation in 2D
¡ ¢
∆2 − λ2 Φ(n) (r ) = r2n ln r, (2.13)

where λ>0, n is a positive integer and ∆ is the Laplacian where

1 d¡ d¢
∆= r .
r dr dr
For illustration of basic idea, let us first consider the case when n=2. We observe that
for r 6=0 and k ≥3

∆r4 ln r = 16r2 ln r + 8r2 , ∆2 r4 ln r = 64 ln r + 96, ∆k r4 ln r = 0. (2.14)

Since ∆4 r4 ln r =0, it follows that

³ ³ ∆2 ´2 ´
1− r4 ln r = r4 ln r. (2.15)
λ2
The differential operator on the the left hand side of (2.15) can be factored as follows:
µ ¶µ ¶
∆2 ∆2
1− 2 1+ 2 r4 ln r = r4 ln r. (2.16)
λ λ

By rearranging the above equation, we have

µ ¶µ ¶
¡ 2 2
¢ 1 ∆2
∆ −λ − 2 1+ 2 r4 ln r = r4 ln r. (2.17)
λ λ
| {z }
Φ (2) ( r )
754 G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768

Comparing (2.13) and (2.17), we can easily see the underbrace in (2.17) is a particular
solution Φ(2) , i.e.,
µ ¶µ ¶
(2) 1 ∆2 4
Φ (r ) = − 2 1 + 2 r ln r
λ λ
r4 ln r 64 ln r + 96
=− − , for r 6= 0. (2.18)
λ2 λ4
However, we notice that Φ(2) (r ) in (2.18) has a singularity at r =0, and the normal
derivative of Φ(2) (r ) also has singularity term 1/r2 at r =0 as follows

1 ∂Φ(2) (r ) r2 (4 ln r + 1) 64
=− − 4 2. (2.19)
r ∂r λ2 λ r
To remove √the singularities,
√ we use a similar technique to the one shown in [8]. We
denote K0 ( λr ) and Y0 ( λr ) as the first kind and second kind of Bessel functions
with order 0, respectively. For r 6=0,we have
√ √
(∆ − λ)K0 ( λr ) = 0, (∆ + λ)Y0 ( λr ) = 0,

which implies
√ √
(∆2 − λ2 )K0 ( λr ) = (∆ + λ)(∆ − λ)K0 ( λr ) = 0,
√ √
(∆2 − λ2 )Y0 ( λr ) = (∆ − λ)(∆ + λ)Y0 ( λr ) = 0.

We also notice that [10]

√
√ λ¢ ¡
K0 ( λr ) = −γ − ln − ln r + O(r ), (2.20)
2√
√ 2³ ¡ λ¢ ´
Y0 ( λr ) = γ + ln + ln r + O(r ), (2.21)
π 2
where γ'0.5772156649015328 is the Euler’s constant. We observe that K0 and Y0 con-
tain ln r which can be used to remove the singularities in (2.18) and (2.19). Let

r4 ln r 64 ln r + 96 √ √
Φ (2) ( r ) = − 2
− 4
+ aK0 ( λr ) + bY0 ( λr ). (2.22)
λ λ
Then,

1 ∂Φ(2) (r ) r2 64 a√ √ b√ √
= − 2 (4 ln r + 1) − 4 2 − λK1 ( λr ) − λY1 ( λr ), (2.23)
r ∂r λ λ r r r
where K1 and Y1 [10] are the first and second kind of Bessel functions with order 1,
respectively. To remove the singularity of ln r in (2.22), we have
64 ¡2¢
− 4
+ a(−1) + b = 0. (2.24)
λ π
G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768 755

To remove the singularities of 1/r2 and ln r in (2.23), since

√ √ √
λK1 ( λr ) λ λγ λ ¡ λ ¢ λ 1
− = − − ln − ln r − 2 + O(r ), (2.25)
√ r √ 4 2 2 2 2
√ r
λY1 ( λr ) 2 h λ λγ λ ¡ λ ¢ λ 1i
− =− − + + ln + ln r − 2 + O(r ), (2.26)
r π 4 2 2 2 2 r
we have that
64 2
− 4
+ a(−1) + b( ) = 0, (2.27)
λ π
and
λ 2 λ
a(− ) + b(− )( ) = 0. (2.28)
2 π 2
Notes that (2.24) and (2.27) are identical. This implies that

32 16π
a=− , and b= .
λ4 λ4
Thus,

 r4 ln r 64 ln r + 96 16 ³ √ √ ´
− − − 2K 0 ( λr ) − πY0 ( λr ) , for r 6= 0,
Φ(2) (r )= λ2 √ λ´4 λ4 (2.29)
 ³
 64 γ + ln( λ ) − 96 , for r = 0.
λ4 2 λ4
Note that [10]
√ √
dK0 ( λr ) √ √ dY0 ( λr ) √ √
= − λK1 ( λr ), = − λY1 ( λr ).
dr dr

By taking derivative of Φ(2) in (2.29) directly, we have

1 ∂Φ(2) (r ) r2 64
= − 2 (4 ln r + 1) − 4 2
r ∂r λ √ ³ λ r
16 λ √ √ ´
+ 4 2K1 ( λr ) − πY1 ( λr ) , for r 6= 0, (2.30a)
λ r
1 ∂Φ(2) (r )
= 0, for r = 0. (2.30b)
r ∂r
Using the following identities [10]
√
d(K1 ( λr )) √ √ 1 √
= − λK0 ( λr ) − K1 ( λr ), (2.31)
dr
√ r
d(Y1 ( λr )) √ √ 1 √
= λY0 ( λr ) − Y1 ( λr ), (2.32)
dr r
756 G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768

we can easily obtain the following


 8r2 16 ³ √ √ ´
(2)
∆Φ (r )= − 2
( 2 ln r + 1 ) − 2K 0 ( λr ) + πY0 ( λr ) , for r 6= 0, (2.33)
 0, λ λ3
for r = 0.

To find the general form of particular solutions Φ(n) (r ), we can follow the same pro-
cedure as shown above. We start with the following identity
³ ³ ∆ 2 ´ m +1 ´ n
1− r2n ln r = r2n ln r, m = [ ], (2.34)
λ2 2
where m means the largest integer not greater than n/2, and

∆k r2n ln r = 0, if k ≥ m + 1.

Similarly, (2.34) can be written as follows

³ −1 ´ m ³ ∆2 ´ i
( ∆2 − λ2 ) ∑ r2n ln r = r2n ln r. (2.35)
λ2 i =0
λ2

It follows that
1 m ³ ∆2 ´ i n
Φ (n ) (r ) = − ∑ r2n ln r, m = [ ]. (2.36)
λ2 i =0
λ2 2

The explicit form of Φ(n) (r ) in (2.36) can be obtained easily by direct differentiation or
using the symbolic software such as Mathematica or Maple.
For n even, a singularity will appear in ∆2m r2n ln r and must be removeed as in the
r ln r case. For n odd, there is no singularity in Φ(n) (r ), but the normal derivative of
4

Φ(n) (r ) will have a singularity which also comes from

1 ∂∆2m r2n ln r
.
r ∂r
It follows that for r ≥0,

1 m ³ ∆2 ´ i
(n)
Φ (r ) = − 2 ∑ r2n ln r
λ i =0
λ2
4n−1 (n!)2 ³ √ √ ´
− 2K0 ( λr ) + (−1)n+1 πY0 ( λr ) , (2.37)
λ n +2
and
Φ(n) (0) = lim Φ(n) (r ). (2.38)
r →0

In Appendix, we provide the explicit particular solutions and their derivatives for
n=3, 4 in 2D and 3D cases.
G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768 757

If (∆2 − λ2 ) in (2.13) is replaced by (∆2 + λ2 ), we can obtain a similar result as

follows:

1 m ³∆2 ´i 2n
Φ(n) = ∑ − r ln r
λ2 i =0
λ2
4n−1 (−1)m (n!)2 ³ √ n +1
√ ´
+ 2K 0 ( λr ) + (− 1 ) πY0 ( λr ) . (2.39)
λ n +2

2.2 Three dimensional case

A similar approach to the one seen above can be applied to find the closed form Φ(n) (r )
in (2.13) in the three-dimensional case. We consider

(∆2 − λ2 )Φ(n) (r ) = r2n−1 , n = 1, 2, 3, · · · . (2.40)

In 3D, the Laplacian is denoted as follows:

1 d³ 2d´
∆= r .
r2 dr dr
When n=1, it is easy to show that


 r 1 ³ √ √ ´
 − 2 + 3 e− λr − cos( λr ) , for r 6= 0,
λ λ r
Φ (1) ( r ) = (2.41)


 −5
−λ 2 , for r = 0.

Again, we first consider the case for n=2. We observe that for r 6=0

24
∆r3 = 12r, ∆2 r 3 = , ∆k r3 = 0, for k ≥ 3. (2.42)
r
We start with the following identity
³ ³ ∆2 ´2 ´
1− r3 = r3 . (2.43)
λ2
Similar to the 2D case, the left hand side of the above identity can be factored into the
form:
³ 1 ´³ ∆2 ´
( ∆2 − λ2 ) − 2 1 + 2 r 3 = r 3 . (2.44)
λ λ
Comparing (2.40) and (2.44), we have
³ 1 ´³ ∆2 ´ 3 r3 24
Φ (2) ( r ) = − 2
1 + 2
r = − 2
− 4 , for r 6= 0. (2.45)
λ λ λ λ r
758 G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768

Notice again we need to pay special attention to the singularity at r =0. We observe
that
√ √
³ e− λr ´ λ −√λr ³ e− λr ´ λ2 √
∆ = e , ∆2 = e− λr ,
r r r r
³ cos(√λr ) ´ λ √ ³ cos(√λr ) ´ λ2 √
∆ = − cos( λr ), ∆2 = cos( λr ).
r r r r
Then, √ √
2 2 e− λr cos( λr )2 2
(∆ − λ ) = 0, (∆ − λ ) = 0.
r r
√ √
This implies that we can add any multiples of e− λr /r and cos( λr )/r to Φ(2) (r ) in
(2.45) and Φ(2) (r ) is still a particular solution. As a result, we assume
√ √
(2) r3 24 e− λr cos( λr )
Φ (r ) = − 2 − 4 + a +b . (2.46)
λ rλ r r
Then, it follows that
√ √
1 ∂Φ(2) (r ) 3r 24 ³ √ e− λr e− λr ´
=− 2 + 4 3 +a − λ 2 − 3
r ∂r λ λ r r r
³ √ sin(√λr ) cos(√λr ) ´
+b − λ − , (2.47)
r2 r3
and √ √
(2) 12r e− λr cos( λr )
∆Φ (r ) = − 2 + aλ + b(−λ) . (2.48)
λ r r
By Taylor series expansion, we have the following expressions
√ √ 1 1 3
e− λr
= 1 − λr + λr2 − λ 2 r3 + O(r4 ), (2.49)
2 6
√ 1 2 1 2 4
cos( λr ) = 1 − λr + λ r + O(r6 ), (2.50)
2 24
√ √ 1 3 3
sin( λr ) = λr − λ 2 r + O(r5 ). (2.51)
6
From (2.49)–(2.51), to remove the singularity of 1/r in (2.46), we have
24
− + a + b = 0. (2.52)
λ4
To remove the singularities of 1/r, 1/r2 and 1/r3 in (2.47), we have that
√ √ λ λ
a(− λ)(− λ) + a(−1) − bλ + b(−1)(− ) = 0, (2.53)
√ √ 2 2
a(− λ) + a(−1)(− λ) = 0, (2.54)
24
− a − b = 0. (2.55)
λ4
G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768 759

To remove the singularity of 1/r in (2.48), we have

aλ bλ
− = 0. (2.56)
2 2
Notes that (2.52) and (2.55) are identical and (2.54) is true for any value a. Hence,
(2.52)–(2.56) have only one solution which is
12
a=b= .
λ4
Thus,


 −r 3 24 12 ³ −√λr √ ´
 − + e + cos ( λr ) , for r 6= 0,
Φ (2) ( r ) = λ2 rλ4 λ4 r (2.57)


 7
−12λ− 2 , for r = 0.
Therefore, we have
 √

 3r 24 12 λ ³ −√λr √ ´

 − + − e + sin ( λr )

 λ2 λ4 r 3 λ4 r2³ √ ´
( 2
1 ∂Φ (r )) 12 √
− λr + cos( λr ) ,
= − e for r 6= 0, (2.58)
r ∂r 
 λ4 r 3



 5
−4λ− 2 , for r = 0,
and 
 12r 12 ³ √ √ ´
 − 2 + 3 e− λr − cos( λr ) , for r 6= 0,
∆Φ(2) (r ) = λ λ r (2.59)

 5
−12λ− 2 , for r = 0.
In general, we can obtain the general form of Φ(n) (r ) using the same procedure. Since
n
∆m r2n−1 = 0, m = [ ],
2
it follows that
³ 1 ´ m ³ ∆2 ´i 2n−1
( ∆2 − λ2 ) − 2 ∑ r = r2n−1 . (2.60)
λ i =0 λ 2
Thus,
1 m ³ ∆2 ´ i n
Φ (n ) (r ) = − ∑ r2n−1 , m = [ ]. (2.61)
λ2 i =0
λ2 2
For n even, a singularity will appear in ∆2m r2n which must be removed as in the Φ(2)
case. For n odd, there is no singularity in Φ(n) (r ), but the normal derivative of Φ(n) (r )
will still have a singularity which also comes from
1 ∂∆2m r2n
.
r ∂r
760 G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768

It follows that if r >0,we can obtain the following

1 m ³ ∆2 ´ i (2n)! ³ −√λr √ ´
Φ (n ) (r ) = − ∑ r2n−1 + e + (− 1 ) n
cos ( λr ) , (2.62)
λ2 i =0
λ2 2λn+2 r

and
Φ(n) (0) = lim Φn (r ).
r →0

If the differential operator − ( ∆2 λ2 )

in (2.40) is replaced by (∆2 + λ2 ), a similar
result follows:
1 m ³ ∆2 ´ i (−1)m+1 (2n)! ³ −√λr √ ´
Φ(n) (r ) = 2 ∑ − 2 r2n+1 + e + (− 1 ) n
cos ( λr ) . (2.63)
λ i =0 λ 2λn+2 r

3 Particular solution for monomial term

Next, we need to find the closed-form particular solutions for (2.12). We consider
particular solutions of the following equation with the monomial term as the forcing
term.
( ∆2 − λ2 ) χ = x k y l , (3.1)
where k and l are nonnegative integers. To find χ, a similar derivation for the Cheby-
shev polynomial can be found in [3]. We start with the following identity
³ ³ ∆ 2 ´ L +1 ´
1− x k yl = x k yl , (3.2)
λ2
where L is sufficiently large. The left hand side of (3.2) can be factored as follows:
³ −1 ´ L ³ ∆2 ´ i
( ∆2 − λ2 ) ∑ x k yl = x k yl . (3.3)
λ2 i =0
λ2

Again, comparing (3.3) with (3.1), we obtain

−1 L ³ ∆2 ´ i k l
χ= 2 ∑ x y. (3.4)
λ i =0 λ 2

A similar conclusion can be obtained for the case ∆2 + λ2 in the 3D case. For illustra-
tion, we consider the following equation

( ∆2 − λ2 ) χ = x 3 y2 .

Using (3.4), we obtain

−1 ³ ∆2 ∆4 ´
3 2 − x 3 y2
χ= 1 + + + ... x y = .
λ2 λ2 λ4 λ2
G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768 761

4 Numerical examples
To validate the efficacy of the proposed method, we consider three examples in this
section. All of the numerical results compare well with the analytical solutions. We
will examine the influence of the number of boundary, interior nodes, and order of
polyharmonic splines on the accuracy of numerical results.
The root-mean-square error (RMSE) and the root-mean-square error of the deriva-
tive with respect to x (RMSEx) are used in this section to demonstrate the accuracy of
the solutions and they are defined as follows:
v
u q
u1
RMSE = t ∑ (ubj − u j )2 , (4.1)
q j =1
v
u q
u1 ∂ubj ∂u j 2
RMSEx = t ∑ ( − ) , (4.2)
q j=1 ∂x ∂x

where q is the number of testing nodes located uniformly within the domain. ubj de-
notes the numerical solution at the jth node. According to the following numerical
tests, we found that the RMSEy is similar to RMSEx. Therefore, we only show the
RMSE and RMSEx in the following tests.
To approximate the homogeneous solution, we apply the method of fundamen-
tal solutions (MFS) [8, 9]. In the MFS, the source points are located on the fictitious
boundary outside the computational domain. There are various ways to determine
the location of source points. In this paper, we apply the following approach to deter-
mine how to place the source points

xs = xb + d (xb − xc ), (4.3)

where xs , xb and xc denote the source, boundary, and central nodes. d determines
how far away of the source points to the boundary. The search of optimal d is still
an outstanding research problem. In the following examples, we find little difference
using different values of d.
Example 1. Consider the following problem in thin plate vibration

(∆2 − 2000)u( x, y) = f ( x, y), ( x, y) ∈ Ω, (4.4)

2 2
u( x, y) = sin(y + x ) − cos(y − x ), ( x, y) ∈ ∂Ω, (4.5)
∂u( x, y) ³ ´
= ∇ sin(y2 + x ) − cos(y − x2 ) · n, ( x, y) ∈ ∂Ω, (4.6)
∂n
where n is the normal vector along the boundary.
The non-homogeneous term f ( x, y) is given based on the analytical solution, which
is as follows:
u( x, y) = sin(y2 + x ) − cos(y − x2 ), ( x, y) ∈ Ω. (4.7)
762 G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768

Table 1: RMSE and RMSEx obtained by different number of boundary nodes nb , interpolation nodes ni
and order of polyharmonic splines n.

nb 80 120 160
ni 101 202 318
n=1 RMSE 1.33E − 3 3.73E − 4 1.43E − 4
RMSEx 1.09E − 2 3.35E − 3 1.62E − 3
n=2 RMSE 1.94E − 4 1.67E − 5 1.02E − 5
RMSEx 1.47E − 3 1.81E − 4 9.21E − 5
n=3 RMSE 6.23E − 4 2.48E − 5 4.05E − 6
RMSEx 4.76E − 3 2.74E − 4 4.69E − 5

Figure 1: Interior points (∗) and boundary points (•) (left) and the profile of solution in the extended
domain.

The computational domain is defined by the following parametric equation:

Ω = {( x, y) | x = ρ cos θ, y = ρ sin θ, 0 ≤ θ ≤ 2π }, (4.8)
where ³ q ´
ρ= cos(2θ ) + 1.1 − sin2 (2θ ) . (4.9)
The computational domain, distribution of interior and boundary nodes, and the pro-
file of solution in the extended domain are shown in Fig. 1. The number of testing
nodes q=212 is used for calculating the RMSE and RMSEx. Note that at each bound-
ary nodes there are two boundary conditions imposed. Hence, we have an overdeter-
mined system of equations. In this case, we have to use the linear least-square solver.
The RMSE and RMSEx are shown in Table 1 where the number of boundary nodes
and interior nodes are chosen as (80, 101), (120, 202), and (160, 318). The order of the
polyharmonic splines (r2n ln r ) is used for n=1, 2, 3. To obtain the homogeneous solu-
tion, we applied the MFS. To locate the source points, we choose d=6.
As shown in Table 1, the results improved with the increases of the number of
boundary, interior nodes, and the order of polyharmonic splines.
G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768 763

Example 2. In this example, we consider the similar problem as the previous example
with different boundary conditions and domain

(∆2 − 100)u( x, y) = f ( x, y), ( x, y) ∈ Ω, (4.10)

u( x, y) = sin(πx ) cosh(y) − cos(πx ) sinh(y), ( x, y) ∈ ∂Ω, (4.11)
³ ´
∆u( x, y) = ∆ sin(πx ) cosh(y) − cos(πx ) sinh(y) , ( x, y) ∈ ∂Ω. (4.12)

The non-homogeneous term f ( x, y) is based on the analytical solution, which is shown

as follows:

u( x, y) = sin(πx ) cosh(y) − cos(πx ) sinh(y), ( x, y) ∈ Ω. (4.13)

The computational domain is defined by the following parametric equation:

Ω = {( x, y) | x = ρ cos θ, y = ρ sin θ, 0 ≤ θ ≤ 2π }, (4.14)

where
³ q ´ 13
ρ= cos(3θ ) + 2 − sin2 (3θ ) . (4.15)

The computational domain, distribution of nodes, and profile of solution are shown
in Fig. 2. We choose q=230 testing nodes for calculating the RMSE and RMSEx.
Similar to the previous example, we use the linear least-square solver to solve the
overdetermined system. To compute RMSE and RMSEx, the number of boundary
and interior nodes are chosen as (80, 126), (140, 280) and (200, 374) with the order of
the ployharmonic splines n=1, 2, 3. For the MFS, we choose d=4.
From Table 2, we observe that even with a small amount of boundary and interior
nodes, we can still achieve good accuracy. This further confirms that once we obtain
the closed-form particular solution and fundamental solution of the given differential
equation, the problem can be solved efficiently.

Table 2: RMSE and RMSEx obtained by different number of boundary nodes, nb , interpolation nodes, ni ,
and order of polyharmonic splines, n.

nb 80 140 200
ni 126 208 374
n=1 RMSE 7.34E − 5 3.31E − 5 4.16E − 5
RMSEx 7.20E − 4 4.36E − 4 4.19E − 4
n=2 RMSE 4.52E − 5 1.58E − 5 7.84E − 6
RMSEx 2.04E − 4 1.32E − 4 8.39E − 5
n=3 RMSE 2.44E − 5 5.89E − 6 1.64E − 6
RMSEx 8.59E − 5 3.48E − 5 1.71E − 5
764 G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768

Figure 2: Computational domain (left) and profile of solution (right) in the extended domain.

Example 3. In this example, we consider the three dimensional problem

(∆2 − 1000)u( x, y, z) = f ( x, y, z), ( x, y, z) ∈ Ω, (4.16)

1
u( x, y, z) = ( x 5 + y5 + z5 ), ( x, y, z) ∈ ∂Ω, (4.17)
120
∆u( x, y, z) = g( x, y, z), ( x, y, z) ∈ ∂Ω, (4.18)

where f ( x, y, z) and g( x, y, z) are based on the analytical solution, which is shown as

follows:
1
u( x, y, z) = ( x 5 + y5 + z5 ), ( x, y, z) ∈ Ω. (4.19)
120
The computational domain, unit sphere, and distribution of boundary nodes are
shown in Fig. 3. We choose q=93 randomly distributed interior testing nodes for
calculating the RMSE and RMSEx. Note that the number of boundary and interior
nodes are chosen as (20, 27), (80, 93), (140, 160), (200, 251) and (260, 360) with the or-
der of the polyharmonic splines n=1, 2, 3, 4. The numerical results are shown in Tables

Figure 3: Computational domain and boundary points.

G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768 765

Table 3: RMSE error using various order of polyharmonic spliens.

nb 20 80 140 200 260

ni 27 93 160 251 360
n RMSE
1 1.23E − 3 8.96E − 4 9.20E − 4 9.77E − 4 1.14E − 3
2 1.27E − 3 8.37E − 4 9.19E − 4 9.96E − 4 1.07E − 3
3 1.20E − 3 6.76E − 4 5.01E − 4 4.12E − 4 3.11E − 4
4 1.88E − 3 7.77E − 4 3.17E − 4 2.65E − 4 2.15E − 4

Table 4: RMSE errors using various order of polyharmonic splines.

nb 20 80 140 200 260

ni 27 93 160 251 360
n RMSEx
1 4.35E − 3 3.24E − 3 2.74E − 3 2.34E − 3 2.47E − 3
2 4.51E − 3 2.40E − 3 2.10E − 3 2.20E − 3 2.41E − 3
3 4.30E − 3 2.05E − 3 1.37E − 3 1.12E − 3 9.18E − 4
4 6.43E − 3 2.51E − 3 1.33E − 3 9.70E − 4 8.76E − 4

Table 5: The optimal distance d between boundary and source points corresponding Tables 3 and 4.

nb 20 80 140 200 260

ni 27 93 160 251 360
n dopt
1 1.0765 1.0254 1.0097 1.0061 1.0042
2 1.0796 1.0161 1.0051 1.0019 1.0005
3 1.0769 1.0146 1.0076 1.0048 1.0068
4 1.0811 1.0178 1.0157 1.0124 1.0193

3 and 4. When we increase the order of polyharmonic spline and the number of bound-
ary points and interior points, the RMSE and RMSEx become smaller. The optimal d
for the source points in the MFS corresponding to Tables 3 and 4 are shown in Table
5. Since our main focus in this paper is the derivation of particular solutions, we will
not further elaborate how we choose the optimal d.
In general, it is nontrivial to solve a 3D problem using traditional numerical meth-
ods due to the complication of domain and surface meshing. In our approach, once
particular solutions are derived, a 3D problem can be handled as easily as a 2D prob-
lem.

5 Conclusions
Using a very simple algebraic factorization approach, we derive a closed-form par-
ticular solution in which polyharmonic splines have been used as basis functions to
approximate the non-homogeneous term of a given differential operator of the form
766 G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768

∆2 ± λ2 . Instead of tedious derivation of particular solutions as shown in [12, 13], our

derivation of particular solutions is simple and direct. Our approach for the derivation
of particular solutions can be easily extended to product of poly-Helmholtz operators.

Acknowledgement
The authors thank Dr. C. M. Fan for his helpful comments and suggestions.

Appendix
In the 2D case, for n=3, 4, particular solutions for (2.13) is given by

 r6 ln r r2 576 ³ √ √ ´
Φ (3) ( r ) = − 2
− ( 576 ln r + 480 ) − 2K 0 ( λr ) + πY0 ( λr ) , for r 6= 0,
 0, λ λ4 λ5
for r = 0.


 r4 384

 − 2
(6 ln r + 1) − 4 (3 ln r + 4)

 λ λ √
1 ∂Φ(3) (r ) 576 λ ³ √ √ ´
= + 2K 1 ( λr ) + πY1 ( λr ) , for r 6= 0,
r ∂r 
 5r
λ√

 192 ³ ´

 λ
4
11 − 6γ − 6 ln( ) , for r = 0.
λ 2


 384 12r4

 − 4 (6 ln r + 11) − 2 (3 ln r + 1)

 λ λ
(3) 576 ³ √ √ ´
∆Φ (r ) = − 4 2K0 ( λr ) − πY0 ( λr ) , for r 6= 0,

 λ√
 384 ³
 λ ´

 − 11 − 6γ − 6 ln ( ) , for r = 0.
λ4 2



 r8 ln r 192r4 12288

 − − (12 ln r + 7) − (12 ln r + 25)

 λ 2 4
λ ³ λ6 ´
36864 √ √
Φ (4) ( r ) = − 6 2K0 ( λr ) − πY0 ( λr ) , for r 6= 0,

 λ √ ´

 ³
 12288 − 25 + 12γ + 12 ln( λ ) ,

for r = 0.
λ6 2


 r6 1536r2 147456

 − ( 8 ln r + 1 ) − (6 ln r + 5) − 2 6
( 4 )
1 ∂Φ (r )  λ 2
√ λ 4 r λ
= 36864 λ ³ √ √ ´
r ∂r 
 + 2K 1 ( λr ) − πY1 ( λr ) , for r 6= 0,

 λ6 r
0, for r = 0.
G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768 767



 16r6 3072r2

 − ( 4 ln r + 1 ) − (12 ln r + 13)

 λ4 ³λ4
36864 √ √ ´
∆Φ(4) (r ) = − 5 2K0 ( λr ) + πY0 ( λr ) , for r 6= 0,

 λ √ ´

 ³
 − 384 11 − 6γ − 6 ln( λ ) ,

for r = 0.
λ4 2
In the 3D case, for n = 3, 4, particular solutions for (2.13) is given by


 r5 360r 360 ³ √ √ ´
 − 2 − 4 + 5 e− λr − cos( λr ) , for r 6= 0,
Φ (3) ( r ) = λ λ λ r


 9
−360λ− 2 , for r = 0,
 √ ³
 5r 3 360 λ −√λr √ ´

 − − e − sin ( λr )
1 ∂Φ(3) (r )  λ2 λ5 r 2
360 ³
360 − λr √ √ ´
=
r ∂r 
 − − e − cos ( λr ) , for r 6= 0,

 7
rλ4 λ5 r 3
−120λ− 2 , for r = 0,


 30r3 720 360 ³ −√λr √ ´
− − + e + cos ( λr ) , for r 6= 0,
∆Φ(3) (r ) = λ2 7 rλ4 λ4 r (5.1)

−360λ− 2 , for r = 0.

 r7 1680r3 40320 20160 ³ −√λr √ ´
(4) − 2− − + e + cos ( λr ) , for r 6= 0,
Φ (r ) = λ114 rλ6 λ6 r (5.2)
 λ
−20160λ− 2 , for r = 0.


 7r5 5040r 40320 20160 ³ −√λr √ ´

 − − + − e + cos ( λr )
1 ∂Φ(4) (r )  λ2 λ4 λ6 r 3 √ λ6 r 3
= 20160 λ −√λr ³ √ ´ (5.3)
r ∂r 
 − e + sin ( λr ) , for r 6= 0,

 λ 6 r2
 9
−6720λ− 2 , for r = 0.
 ³ ´
 56r5 20160r 20160 −√λr √
(4) − 2 − + e − cos ( λr ) , for r 6= 0,
∆Φ (r ) = λ λ 4 λ 5r (5.4)
 9
−20160λ− 2 , for r = 0.

References
[1] H. A. C HO , M. A. G OLBERG , A. S. M ULESHKOV AND X. L I, Trefftz methods for time
dependent partial differential equations, Computers, Materials and Continua., 1 (2004), pp.
1–37.
[2] C. S. C HEN AND Y. F. R ASHED, Evaluation of thin plate spline based particular solutions for
Helmholtz-type operators for the DRM, Mech. Res. Commun., 25 (1998), pp. 195–201.
768 G. Yao, C. S. Chen, C. C. Tsai / Adv. Appl. Math. Mech., 6 (2009), pp. 750-768

[3] C. S. C HEN , S. L EE AND C. S. H UANG, Derivation of particular solutions using Chebyshev

polynomial based functions, Int. J. Comput. Meth., 4 (2007), pp. 15–32.
[4] W. C HEN, Symmetric boundary knot method, Eng. Anal. Bound. Elem., 26 (2002), pp. 489–
494.
[5] W. C HEN , Z. J. S HEN , L. J. S HEN AND G. W. Y UAN, General solutions and fundamental
solutions of varied orders to the vibrational thin, the Berge, and the Winkler plates, Eng. Anal.
Bound. Elem., 29 (2005), pp. 699–702.
[6] A. H. D. C HENG, Particular solutions of Laplacian, Helmholtz-type, and polyharmonic oper-
ators involving higher order radial basis functions, Eng. Anal. Bound. Elem., 24 (2000), pp.
531–538.
[7] J. D UCHON, Sur l’erreur d interpolation des fonctions de plusiers variables par les Dm -splines,
RAIRO Anal. Numer., 12 (1978), pp. 325–334.
[8] G. FAIRWEATHER AND A. K ARAGEORGHIS, The method of fundamental solutions for elliptic
boundary value problems, Adv. Comput. Math., 9 (1998), pp. 69–95.
[9] M. A. G OLBERG AND C. S. C HEN, The method of fundamental solutions for potential,
Helmholtz and diffusion problems, in Boundary Integral Methods - Numerical and Math-
ematical Aspects, ed. M.A. Golberg, Computational Mechanics Publications, (1998), pp.
103–176.
[10] A. M ILTON AND A. S. I RENE, Handbook of Mathematical Functions with Formulas,
Graphs and Mathematical Tables, Dover, New York, (1964).
[11] A. S. M ULESHKOV, M. A. G OLBERG AND C. S. C HEN, Particular solutions of Helmholtz-
type operators using higher order polyharmonic splines, Comput. Mech., 24 (1999), pp. 411–
419.
[12] A. S. M ULESHKOV AND M. A. G OLBERG, Particular solutions of the multi-Helmholtz-type
equation, Eng. Anal. Bound. Elem., 31 (2007), pp. 624–630.
[13] C. C. T SAI , A. H. D. C HENG AND C. S. C HEN, Particular solutions of splines and monomials
for polyharmonic and products of Helmholtz operators, Eng. Anal. Bound. Elem., 33 (2009), pp.
514–521.
[14] D. L. Y OUNG , C. C. T SAI , Y. C. L IN AND C. S. C HEN, The method of fundamental solutions
for eigenfrequencies of plate vibrations, J. Comp. Mater. Con., 4 (2006), pp. 1–10.