stability analysis via condition number and effective condition number for the first kind boundary...
TRANSCRIPT
Engineering Analysis with Boundary Elements 35 (2011) 667–677
Contents lists available at ScienceDirect
Engineering Analysis with Boundary Elements
0955-79
doi:10.1
$The
(108710� Corr
Yat-sen
E-m
(H.-T. H
journal homepage: www.elsevier.com/locate/enganabound
Stability analysis via condition number and effective condition number for thefirst kind boundary integral equations by advanced quadrature methods,a comparison$
Jin Huang a, Hung-Tsai Huang b, Zi-Cai Li c,d,�, Yimin Wei e,f
a College of Applied Mathematics, University of Electronic & Science Technology of China, ChengDu, Chinab Department of Applied Mathematics, I-Shou University, Kaohsiung County, Taiwanc Department of Applied Mathematics, National Sun Yat-sen University, Kaohsiung, Taiwand Department of Computer Science and Engineering, National Sun Yat-sen University, Kaohsiung, Taiwane School of Mathematical Sciences, Fudan University, Shanghai 200433, PR Chinaf Key Laboratory of Mathematics for Nonlinear Sciences, (Fudan University) Ministry of Education, Shanghai 200433, PR China
a r t i c l e i n f o
Article history:
Received 15 August 2010
Accepted 6 November 2010Available online 30 December 2010
Keywords:
Stability analysis
Condition number
Effective condition number
First kind boundary integral equation
Advanced (i.e., mechanical) quadrature
method
Numerical partial differential equations
97/$ - see front matter & 2010 Elsevier Ltd. A
016/j.enganabound.2010.11.006
work is supported by the National Natural
34).
esponding author at: Department of Applied
University, Kaohsiung, Taiwan.
ail addresses: [email protected] (J. Hu
uang), [email protected] (Z.-C. Li), ym
a b s t r a c t
In our previous study [Huang et al., 2008, 2009, 2010 [21,24,20]; Huang and Lu, 2004 [22,23]; Lu and
Huang, 2000 [38]], we have proposed advanced (i.e., mechanical) quadrature methods (AQMs) for solving
the boundary integral equations (BIEs) of the first kind. These methods have high accuracy O(h3), where
h¼max1pmpdhm and hm (m¼1,y,d) are the mesh widths of the curved edge Gm. The algorithms are
simple and easy to carry out, because the entries of discrete matrix are explicit without any singular
integrals. Although the algorithms and error analysis of AQMs are discussed in Huang et al. (2008, 2009,
2010) [21,24,20], Huang and Lu (2004) [22,23], Lu and Huang (2000) [38], there is a lack of systematic
stability analysis. The first aim of this paper is to explore a new and systematic stability analysis of AQMs
based on the condition number (Cond) and the effective condition number (Cond_eff) for the discrete
matrix Kh. The challenging and difficult lower bound of the minimal eigenvalue is derived in detail for the
discrete matrix of AQMs for a typical BIE of the first kind. We obtain Cond¼O(hmin�1 ) and Cond_eff¼O(hmin
�1 ),
where hmin ¼min1pmpdhm , to display excellent stability. Note that Cond_eff ¼ O(Cond) is greatly distinct
to the case of numerical partial differential equations (PDEs) in Li et al. (2007, 2008, 2009, 2010) [26,31–
37], Li and Huang (2008) [27–30], Huang and Li (2006) [19] where Cond_eff is much smaller than Cond.
The second aim of this paper is to explore intrinsic characteristics of Cond_eff, and to make a comparison
with numerical PDEs. Numerical experiments are carried out for three models with smooth and
singularity solutions, to support the analysis made.
& 2010 Elsevier Ltd. All rights reserved.
1. Introduction
The advanced (i.e., mechanical) quadrature methods (AQMs) areproposed in [20–24,38] for the first kind boundary integralequations (BIEs) of Laplace’s equation. The AQMs provide highaccuracy O(h3), accompanied with low computation complexity,where h¼max1rmrdhm and hm (m¼1,y,d) are the mesh widthsof the curved edge Gm. Note that the entries of discrete matrix Kh
are explicit without any singular integrals. Especially, for concavepolygons O, the solution at concave corners of @O has singularities,to heavily damage accuracy of numerical solutions. The accuracy of
ll rights reserved.
Science Foundation of China
Mathematics, National Sun
ang), [email protected]
[email protected] (Y. Wei).
Galerkin methods (GMs) [44,45] is only Oðh1þ eÞ ð0oeo1Þ and theaccuracy of collocation methods (CMs) [47] is even lower. Incontrast, the accuracy of AQMs for singularities is as high asO(h3). In fact, the quadrature method was first proposed for anintegral equation with a logarithmic kernel in Christiansen [11] in1971, called the modified quadrature method, and its analysis wasgiven in Saranen [41], to yield only the O(h2) convergence rate. Inprevious study [20–24,38], we propose the new quadraturemethods called the AQMs, to yield the high O(h3) convergencerate. Moreover, for AQMs, by extrapolations and splitting extra-polations methods (SEMs), the higher precision of numericalsolutions and a posteriori error estimates can be achieved.
This paper is devoted to its stability analysis. Consider the linearalgebraic equations:
Khx¼ b, ð1:1Þ
resulting from the first kind BIE, where the xARn and bARn are theunknown and known vectors, respectively. The condition number
J. Huang et al. / Engineering Analysis with Boundary Elements 35 (2011) 667–677668
is defined by
Cond¼jl1j
jlnj, ð1:2Þ
where li ði¼ 1, . . . ,nÞ are the eigenvalues of matrix KhARn�n in thedescending order in magnitude: jl1jZ � � �Z jlnj40. When thereoccurs a perturbation of b, the errors Dx of x also satisfy1
KhðxþDxÞ ¼ bþDb:
The values of Cond are used to measure the relative errors of x,given by
JDxJ
JxJr Cond �
JDbJ
JbJ, ð1:3Þ
where JxJ is the Euclidean norm. Note that the equality of (1.3)occurs only at very rare cases. In practical applications, the vector bvaries within a certain region, and the true relative errors from theperturbation of b or Kh may be smaller, or even much smaller thanCond given in (1.2).
Here we briefly provide the algorithms for the effective conditionnumber. Details are given in [26,27,31]. Let the matrix KhARn�n bereal symmetric. The eigenvectors ui satisfy Khui ¼ liui, where {ui} areorthogonal, with uT
i uj ¼ dij, where dij ¼ 1 if i¼ j and dij ¼ 0 if ia j.In [12,26,27,31] the effective condition number is defined by
Cond_eff ¼JbJ
jlnjJxJ¼
1
jlnj
JbJffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiPni ¼ 1
b2i
l2i
s , ð1:4Þ
where bi ¼ uTi b.
When the matrix Kh in (1.1) is singular with RankðKhÞ ¼ rrn,the corresponding traditional and the effective condition numbersare defined by
Cond¼jl1j
jlr j, Cond_eff ¼
JbJ
jlrjJxJ, jlrj40:
For the linear algebraic equations in (1.1), the traditional conditionnumber in (1.2) provides the stability for all b; but the effectivecondition number in (1.4) pursues the stability for the specificvector b given. A similar idea of the stability for the specific solutionx obtained is discussed in Chen et al. [9], called model participationfactor in engineering for source excitation.
Stability is a severe issue for numerical solutions of the first kindBIEs. Although stability analysis for the modified quadrature methodwas given in [12,13], it is important to provide a systematic stabilityanalysis for the AQMs. In this paper, the effective condition number isalso applied to the first kind BIEs for Laplace’s equation on arbitraryplane domains by AQMs, and the bounds of both Cond_eff and Condare derived in detail. In this paper, we obtain that Cond ¼ O(hmin
�1 ) andCond_eff¼O(hmin
�1 ), where hmin ¼min1rmrdhm. This displays anexcellent stability of AQMs for smooth and singularity problems.
In [20], the AQMs are applied to Laplace’s equation withdiscontinuity solutions, and the Cond ¼ O(h�1) remains. This isa remarkably advantage over the FEM, FDM, etc. for PDEs, wherelarge or even huge Cond always occur for corner and discontinuitysingularity.
On the other hand, the fact that Cond_eff ¼ O(Cond) indicatesthe insignificant improvements of Cond_eff over Cond for stability.However, in [26–37,19] for numerical PDEs, the Cond is large oreven huge, but the Cond_eff is much smaller than Cond. Why is theperformance of Cond_eff so different? What are the rationalesbehind? To find their answers, in this paper we will exploreintrinsic characteristics of Cond_eff for the first kind BIEs, andmake a comparison with numerical PDEs. Such comprehensiveanalysis and comparison may provide an objective and
1 More discussions are given in [27,31] for a perturbation of both b and Kh.
comprehensive evaluation on Cond_eff. In fact, the improvementsof Cond_eff for perturbation errors are also insignificant for solvinglinear algebraic equations by Gaussian elimination and QR factor-ization. Numerical examples were first given in Banoczi et al. [4] in1998, and a theoretical justification is supplied in [27] recently.
Let us briefly review the references of condition number andeffective condition number. The definition of condition numberwas first given in Wilkinson [46], and then often used (e.g., Goluband van Loan [16]). Other discussions of normwise distance forstructure perturbation are given in Rump [40]. The conditionnumber is used to provide bounds of relative errors from aperturbation of Kh and all b. However, in practical applications,the true relative errors may be smaller, or even much smaller thanthe worst Cond. Such a case was first studied in Rice [39] in 1981,and then in Chan and Foulser [5], Christiansen and Hansen [12],Christiansen and Saranen [13], Axelsson and Kaporin [2,3], and Liet al. in [26–37,19] for numerical PDEs recently. This paper providesa concise review on the effective condition number, and explainsthe fact that why the effective condition number is not developeduntil its applications to numerical PDEs.
This paper is organized as follows. In the next section, the bounds ofCond_eff and Cond are derived in detail for typical BIEs of the first kind.In Section 3, bounds of Cond_eff and Cond are provided for closedsmooth curvesG, curved polygonsG and open contoursG. In Section 4,analysis and comparisons of Cond_eff are made for between numericalBIEs and PDEs, and in Section 5, three numerical examples are reportedto support the stability analysis given, and to show significance ofAQMs. In the last section, a few remarks are made.
2. Typical BIEs of the first kind
There are the Galerkin methods (GMs) and the collocationmethods (CMs). Since the AQMs are superior to GMs and CMs, thestability analysis of the AQMs may be regarded as a representationof that for numerical first kind BIE. To catch the intrinsic char-acteristics of stability easily, we stat with typical BIEs.
2.1. Description of algorithms
We have proposed the AQMs in [38] for solving
Av¼ f , ð2:1Þ
where the unknown v(t) and known f(t) are smooth periodicfunctions on ½0,2p� with the period 2p, and the boundary integraloperator A is defined by
ðAvÞðtÞ ¼
Z 2p
0aðt,tÞvðtÞ dt, tA ½0,2p�, ð2:2Þ
where the integral kernel aðt,tÞ ¼�1=2plnj2e�1=2sinðt�tÞ=2j.Denote h¼ 2p=n,nAN and let ftj ¼ jh,j¼ 1, . . . ,ng be the mesh
set. Using the quadrature rules [43], the Nystrom approximateoperator is given by
ðAhvÞðtÞ ¼ hXn
j ¼ 1t a tj
aðt,tjÞvðtjÞ�1
2p lne�1=2h
2p
� �vðtÞh, ð2:3Þ
with the errors
EnðAÞ ¼ 2X‘�1
m ¼ 1
1
ð2mÞ! xuð�2mÞvð2mÞðtÞh2mþ1þOðh2‘Þ as h-0, ð2:4Þ
where
EnðAÞ ¼ ðAhvÞðtÞ�ðAvÞðtÞ, vA ~C2‘½0,2p�,
and xðtÞ is the Riemann zeta function. The notation vA ~C2‘½0,2p�
denotes the periodic function v with period 2p having 2‘ order
J. Huang et al. / Engineering Analysis with Boundary Elements 35 (2011) 667–677 669
derivatives, defined by
vA ~C2‘½0,2p� ¼ fvðtÞjvðmÞðtÞAC½0,2p�
and
vðmÞðtþ2pÞ ¼ vðmÞðtÞ, m¼ 0,1, . . . ,2‘g:
Using quadrature rules (2.3), we obtain the linear algebraic equations
ðAhvhÞðtiÞ ¼ f ðtiÞ, i¼ 1, . . . ,n: ð2:5Þ
In [38] we have proved that there exist the unique solutions for(2.5) such that
jvhðtiÞ�vðtiÞj ¼Oðh3Þ, i¼ 1, . . . ,n as vðtÞA ~C4½0,2p�:
Moreover, we have derived in [38]
vhðtÞ�vðtÞ ¼X2
m ¼ 1
wmðtÞh2mþ1þOðh6Þ, tAftig as vðtÞA ~C
6½0,2p�,
where wmðtÞA ~C ½0,2p� ðm¼ 1,2Þ are independent of h, and vh(t) andv(t) are the solutions of (2.5) and (2.1) at tAfti,i¼ 1,2, . . . ,ng,respectively. Hence, for AQMs the superconvergence O(h6) canbe achieved by Richardson’s extrapolations [38].
2.2. Condition number and effective condition number
From (2.3) and (2.5), the discrete matrix Ah is symmetric andcirculant, with the entries
a0 ¼�1
nln
e�1=2
n, aj ¼�
1
nln 2e�1=2sin
jpn
��������, j¼ 1, . . . ,n�1,
where h¼ 2p=n. Based on the theory of circulant matrix [14],eigenvalues lk of matrix Ah can be expressed by
lk ¼Xn�1
j ¼ 0
ajejk, ek ¼ expð2kp
ffiffiffiffiffiffiffi�1p
=nÞ, k¼ 0,1, . . . ,n�1: ð2:6Þ
We have the following theorem.
Theorem 2.1. Let the discrete matrix Ah in (2.5) from (2.3) be
symmetric and circulant. Then the condition number has the bound
Cond¼maxkjlkj
minkjlkjr
1
n¼ Oðh�1Þ: ð2:7Þ
Proof. First, we estimate the upper bound of lk. For the leadingeigenvalue of matrix Ah, we have from (2.6),
l0 ¼�1
nln
e�1=2
nþXn�1
j ¼ 1
ln 2e�1=2sinjpn
��������
8<:
9=;
¼�1
n�
1
2�ln nþðn�1Þ ln 2�
1
2
� �þ ln Pn�1
j ¼ 1sinjpn
� �� �: ð2:8Þ
From the equality in [24]
Pn�1j ¼ 1sin
jpn¼
n
2n�1, ð2:9Þ
we have
ln Pn�1j ¼ 1sin
jpn
� �¼ ln n�ðn�1Þln 2: ð2:10Þ
Combining (2.8) and (2.10) gives
l0 ¼12: ð2:11Þ
Next, from (2.6) other eigenvalues lkð ¼ 1,2, . . . ,n�1Þof matrix Ah
are given by
lk ¼�1
nln
e�1=2
nþXn�1
j ¼ 1
cos2kpj
nln 2e�1=2sin
jpn
��������
8<:
9=;,
k¼ 1,2, . . . ,n�1: ð2:12Þ
We have from (2.9)
jlkjr1
nln
e�1=2
n
��������þ lnPn�1
j ¼ 1 2e�1=2sinjpn
��������
� �
¼1
n
1
2þ ln nþ
1
2þ ln n�
n
2
��������
� �¼
1
2: ð2:13Þ
Combining (2.13) and (2.11) gives
maxkjlkj ¼
12: ð2:14Þ
Below, we derive the lower bound of lk. Since
Xn�1
j ¼ 1
cos2kpj
n¼�1, ð2:15Þ
we have from (2.12)
lk ¼1
nln n�
Xn�1
j ¼ 1
cos2kpj
nln 2sin
jpn
��������
8<:
9=;: ð2:16Þ
Let us introduce the cðzÞ special function, defined by the logarith-
mic derivative of the Gamma function (see Gradshteyan and Ryzhik
[17], p. 943 (item 1 in 8.362)),
cðzÞ ¼d
dzlnGðzÞ ¼�g�1
zþzX1j ¼ 1
1
jðjþzÞ, za0,1,2, . . . , ð2:17Þ
where gð ¼ 0:5772 . . .Þ is the Euler constant. Moreover, there exist
other expansions for cðzÞ (see Remark 3.1 below),
ck
n
� �¼�g�ln n�
p2
cotkpn
� �þXn�1
j ¼ 1
cos2kpj
nln 2sin
jpn
��������: ð2:18Þ
Then we have from (2.17) and (2.18)
Xn�1
j ¼ 1
cos2kpj
nln 2sin
jpn
��������¼ ln n�
n
kþp2
cotkpn
� �þ
k
n
X1j ¼ 1
1
j jþk
n
� � :ð2:19Þ
Hence Eq. (2.16) leads to
lk ¼1
k�
p2n
cotkpn
� ��
k
n2
X1j ¼ 1
1
j jþk
n
� � : ð2:20Þ
Moreover, there exists the series expansion for cotx in [17], p. 35
(item 7 in 1.411),
cotx¼1
x�
x
3� � � ��
22jjB2jj
ð2jÞ!x2j�1� � � � , jxjop, ð2:21Þ
where Bj is the Bernoulli number. Substituting (2.21) into (2.20)
gives
lk ¼1
k�
p2n
n
kp�kp3n� � � � �
22jjB2jj
2j!
kpn
� �2j�1
� � � �
( )�
k
n2
X1j ¼ 1
1
j jþk
n
� �
Z1
k�
1
2kþ
kp2
6n2�
k
n2
X1j ¼ 1
1
j jþk
n
� �
¼1
2kþ
k
n2
X1j ¼ 1
1
j2�
1
j jþk
n
� �8>><>>:
9>>=>>;Z
1
2kZ
1
2n, k¼ 1,2, . . . ,n�1,
ð2:22Þ
2 For the mixed (i.e., Robin) boundary condition, the AQMs with extrapolation
techniques are explored in Huang et al. [21].
J. Huang et al. / Engineering Analysis with Boundary Elements 35 (2011) 667–677670
where we have used the bound, p2=6¼P1
j ¼ 1 1=j2. Furthermore,
combining (2.11) and (2.22) gives
minkjlkjZ
1
2n: ð2:23Þ
The desired result (2.7) follows from (2.14) and (2.23). This
completes the proof of Theorem 2.1. &
Below, we derive the bounds of effective condition number.From [1,44,48], when the solution vðtÞA ~C
4½0,2p� of (2.1), we have
jvhðtiÞ�vðtiÞj ¼Oðh3Þði¼ 1, . . . ,nÞ, to give
JxJ2 ¼Xn
i ¼ 1
v2hðtiÞ
( )1=2
�Xn
i ¼ 1
v2ðtiÞ
( )1=2
� Oðh�1=2Þ, ð2:24Þ
where the notation a � bða � OðbÞÞ,b40, denotes that there existtwo constants c1 and c2 independent of n such that c1br jajrc2b.Next, from [1,44,48], if vðtÞA ~C
k½0,2p�, then f ðtÞA ~C
kþ1½0,2p�.
We have
JbJ2 ¼Xn
j ¼ 1
½f ðtjÞ�2
8<:
9=;
1=2
¼ h�1=2Xn
j ¼ 1
h½f ðtjÞ�2
8<:
9=;
1=2
Ch�1=2
Z 2p
0f 2ðtÞdt
( )1=2
� O h�1=2�
: ð2:25Þ
From (2.24), (2.25) and (2.23), we have
Cond_eff ¼JbJ
minkjlkjJxJ¼Oðh�1Þ, ð2:26Þ
to give the following theorem.
Theorem 2.2. Let the discrete matrix Ah in (2.5) from (2.3) be
symmetric and circulant. Then the effective condition number has
the bound (2.26).
Remark 2.1. For the lower bound (2.23) of lk, Eq. (2.18) is a keyexpansion we need to double-check. We cite from [17], p. 944 (item6 in 8.363),
ck
n
� �¼�C�ln n�
p2
cotkpn
� �þ2
XEððnþ1Þ=2Þ�1
j ¼ 1
cos2kpj
nln sin
jpn
��������, ð2:27Þ
where C is a constant, and E(x) is the integer part of real x. We havefrom (2.15)
2XEððnþ1Þ=2Þ�1
j ¼ 1
cos2kpj
nln sin
jpn
��������¼Xn�1
j ¼ 1
cos2kpj
nln sin
jpn
��������
¼Xn�1
j ¼ 1
cos2kpj
nln 2sin
jpn
���������ln 2
Xn�1
j ¼ 1
cos2kpj
n
¼ ln 2þXn�1
j ¼ 1
cos2kpj
nln 2sin
jpn
��������: ð2:28Þ
Combining (2.27) and (2.28) gives
ck
n
� �¼�Cþ ln 2�ln n�
p2
cotkpn
� �þXn�1
j ¼ 1
cos2kpj
nln 2sin
jpn
��������: ð2:29Þ
Compared (2.29) with (2.18), we find the constant
C ¼ gþ ln 2: ð2:30Þ
There may have the other choice of constant as C ¼ g. By using somespecial values of cðzÞ in [17], p. 945 (8.366),
c1
2
� �¼�g�2ln 2¼�1:963510026, c
1
4
� �¼�g�p
2�3ln 2, ð2:31Þ
we confirm (2.18) as well as the constant (2.30).
3. Dirichlet’s problems of Laplace’s equation
3.1. Closed smooth curve G
By the layer potential theory, Dirichlet’s problems of Laplace’sequation2:
Du¼ 0 in O,
u¼ f on G¼ @O,
uðxÞ ¼OðlnjxjÞ as jxj-1,
8><>: ð3:1Þ
are converted into the first kind BIEs [1,44]
�1
2p
ZG
vðxÞlnjx�yj dsx ¼ f ðyÞ, yAG, ð3:2Þ
where O� R2 is a bounded domain with a closed smooth edge G, and
jx�yj ¼ fðx1�y1Þ2þðx2�y2Þ
2g1=2:
In (3.2) the unknown function
vðxÞ ¼@uðxÞ
@n� �@uðxÞ
@nþ ,
where n is a unit outward normal at a point xAG. From the knownresults [1,44,47], when the logarithmic capacity (transfinite diameter)CGa1, there exists a unique solution of (3.2). As soon as v(x) is solvedfrom (3.2), the solutions of (3.1) at interior or exterior points can beevaluated by
uðyÞ ¼�1
2p
ZG
vðxÞlnjx�yj dsx, yAR2\G:
Assume that CGa1 andG can be described by the parameter mapping
xðtÞ ¼ ðx1ðtÞ,x2ðtÞÞA ~C‘½0,2p� : ½0,2p�-G
with
�mZ jxuðtÞj2 ¼ jxu1ðtÞj2þjxu2ðtÞj
2Zm40,
where �m and m are two constants. Define the boundary integraloperator
ðKvÞðtÞ ¼
Z 2p
0kðt,tÞvðtÞ dt, tA ½0,2p�,
where kðt,tÞ ¼ � 12p lnjxðtÞ�xðtÞj and vðtÞ ¼ vðxðtÞÞjxuðtÞj. Then Eq. (3.2)
is converted into
Kv¼ AvþBv¼ f , ð3:3Þ
where B¼K�A and ðBvÞðtÞ ¼R 2p
0 bðt,tÞvðtÞ dt with
bðt,tÞ ¼�
1
2plne1=2ðxðtÞ�xðtÞÞ
2sint�t
2
� ���������
��������for t�t =2 2pZ,
�1
2pln e1=2xuðtÞ�� �� for t�tA2pZ,
8>>>>>>><>>>>>>>:
where Z ¼ f0,71,72, . . .g.Using the trapezoidal or the midpoint rule [15], Nystrom’s
approximate operator Bh of B is given by
ðBhvÞðtÞ ¼Xn
tj a t
j ¼ 1
hbðt,tjÞvðtjÞþ�h
2p ln e1=2xuðtÞ�� ��vðtÞ:
Hence, we obtain the approximate equations of (3.3)
AhvhðtiÞþBhvhðtiÞ ¼ f ðtiÞ, i¼ 1, . . . ,n: ð3:4Þ
J. Huang et al. / Engineering Analysis with Boundary Elements 35 (2011) 667–677 671
Lemma 3.1. Let G with CGa1 be an arbitrarily closed smooth curve.
Assume that vðtÞAC6½0,2p� and kðt,tÞvðtÞ is periodic with period 2p,and has six order differentiable on ð�1,1Þ\ftþ2pmg1m ¼ �1.
(1)
There exists a unique solution in (3.4) with the errorsjvhðtiÞ�vðtiÞj ¼ Oðh3Þ, i¼ 1, . . . ,n:
(2)
There exist the functions wmðtÞA ~C ½0,2p� ðm¼ 1,2Þ independent ofh such that
vhðtÞ�vðtÞ ¼X2
m ¼ 1
wmðtÞh2mþ1þOðh6Þ, tAftig,
where vh(t) and v(t) are the solutions of (3.4) and (3.3) at t¼tj,respectively.
Lemma 3.1 implies that for the closed smooth curve G with CGa1,the superconvergence O(h6) can also be achieved by Richardson’sextrapolations.
3.2. Closed curved polygons G or open contours G
Let G¼Sd
m ¼ 1 Gmðd41Þ be curved polygons or open contourswith CGa1, and Gm be a piecewise smooth curve. Define theboundary integral operators on Gm,
ðKqmvmÞðyÞ ¼�1
2p
ZGm
vmðxÞlnjy�xj dsx, yAGq, m,q¼ 1, . . . ,d,
ð3:5Þ
where vmðxÞ ¼ @umðxÞ=@n��@umðxÞ=@nþ . Then Eq. (3.2) can be con-verted into a matrix operator equation
Kv¼ F, ð3:6Þ
where K¼ ½Kqm�dq,m ¼ 1, v¼ ðv1ðxÞ, . . . ,vdðxÞÞ
T and F¼ ðf1ðyÞ, � � � ,fdðyÞÞT .
Here, let K be symmetric operators.Assume that Gm can be described by the parameter mapping
xmðsÞ ¼ ðxm1ðsÞ,xm2ðsÞÞ : ½0,1�-Gm
with
�mZ jxumðsÞj ¼ ½jxum1ðsÞj2þjxum2ðsÞj
2�1=2Zm40,
where �m and m are two constants. Using the sinp�transformation in
[42]
s¼jpðtÞ : ½0,1�-½0,1�, pAN, ð3:7Þ
with jpðtÞ ¼ WpðtÞ=Wpð1Þ and WpðtÞ ¼R t
0ðsinpyÞpdy, then the integraloperators (3.5) can be converted into integral operators on [0,1] asfollows:
ðAqqwqÞðtÞ ¼
Z 1
0aqðt,tÞwqðtÞ dt, tA ½0,1�, ð3:8Þ
ðBqmwmÞðtÞ ¼
Z 1
0bqmðt,tÞwmðtÞ dt, tA ½0,1�, ð3:9Þ
where
aqðt,tÞ ¼ � 1
2pln 2e�1=2sinpðt�tÞ�� ��, ð3:10Þ
wmðtÞ ¼ vmðxmðjpðtÞÞÞjxumðjpðtÞÞjjupðtÞ, ð3:11Þ
bqmðt,tÞ ¼�
1
2pln
xqðtÞ�xqðtÞ2e�1=2sinpðt�tÞ
�������� for q¼m,
�1
2plnjxqðtÞ�xmðtÞj for qam:
8>>><>>>:
ð3:12Þ
In (3.8) and (3.9),
xmðtÞ ¼ ðxm1ðjpðtÞÞ, xm2ðjpðtÞÞÞ ðm¼ 1, . . . ,dÞ,
jxqðtÞ�xmðtÞj ¼ ½ðxq1ðtÞ�xm1ðtÞÞ2þðxq2ðtÞ�xm2ðtÞÞ2�1=2:
Hence Eq. (3.6) becomes
ðAþBÞW¼ G, ð3:13Þ
where A¼ diagðA11, . . . ,AqqÞ and B¼ ½Bqm�dq,m ¼ 1 are symmetric
operators, and W¼ ðw1, . . . ,wdÞT and G¼ ðg1, . . . ,gdÞ
T withgmðtÞ ¼ fmðxmðtÞÞ. Let
hm ¼1
nm, nmAN, m¼ 1, . . . ,d,
be mesh widths for the nodes,
tj ¼ tj ¼ ðj�12Þhm, j¼ 1, . . . ,nm:
By the trapezoidal or the midpoint rule [15] we have Nystrom’sapproximate operator Bqm
h of Bqm. For the weakly singular operatorsAmm, by the quadrature formula (2.3) (see [43]), we can also havethe Nystrom approximate operator Aqq
h . Setting t¼ti ði¼ 1, . . . ,nqÞ,we obtain the following approximate equations of (3.13)
KhWh ¼ ðAhþBhÞWh ¼Gh, ð3:14Þ
where
Wh ¼ ðwh1ðt1Þ, . . . ,w
h1ðtn1Þ, . . . ,wh
dðt1Þ, . . . ,whdðtndÞÞ
T ,
Ah ¼ diagðAh11, . . . ,Ah
ddÞ, Ahqq ¼ ½aqðtj,tiÞ�
nq
j,i ¼ 1,
Bh ¼ ½Bhqm�
dq,m ¼ 1, Bh
qm ¼ ½bqmðtj,tiÞ�nq ,nm
j,i ¼ 1,
Gh ¼ ðg1ðt1Þ, . . . ,g1ðtn1Þ, . . . ,gdðt1Þ, . . . ,gdðtnd
ÞÞT ,
and
aqðtj,tiÞ ¼
�hqlnj2e�1=2sinpðti�tjÞj
2p as ia j,
�hqlnje�1=2hqj
2pas i¼ j:
8>>><>>>:
ð3:15Þ
Obviously, (3.14) is a linear system with n-unknowns, wheren¼ n1þ � � � þnd. We cite the results of [23,24] as a lemma.
Lemma 3.2. Assume that Gmðm¼ 1, . . . ,dÞ are smooth curves, G¼[d
m ¼ 1Gm with CGa1, and h¼max1rmrdhm is sufficiently small. Also
let u0AC4ðGÞ � C4ðGÞ and pZ6 in (3.7). Then there exists a unique
solution Wh of (3.14) such that
Wh�W¼ diagðh31, . . . ,h3
dÞfþOðh4Þ, ð3:16Þ
at node points, where a vector functionf¼ ðf1, . . . ,fdÞT AðC0½0,1�Þd is
independent of h¼ ðh1, . . . ,hdÞT , and the subspace
C0½0,1� ¼ fvðtÞAC½0,1� :vðtÞ
sin2ðptÞ
AC½0,1�g,
with the norm JvJ ¼max0r tr1jvðtÞ=sin2ðptÞj.
From Lemma 3.2 we can obtain the approximate solutions witha higher order accuracy o(h3) by solving some coarse grid discreteequations in parallel. The algorithms of the splitting extrapolationalgorithms (SEAs) are described as follows.
Step 1. Choose h(0)¼(h1,y,hd) and hðmÞ ¼ ðh1, . . . ,hm=2, . . . ,hdÞ,
solve (3.14) for h(m) (m¼1,y,d) in parallel, and then obtain thesolutions WhðmÞ ðtiÞ.
Step 2. Compute the h3-Richardson extrapolation, based on thenumerical solutions on coarse grids
WðtiÞ ¼8
7
Xd
m ¼ 1
WhðmÞ ðtiÞ� d�7
8
� �Whð0Þ ðtiÞ
" #, ð3:17Þ
Table 1The errors en and en
E, Cond and Cond_eff for (5.1).
n en enE
jl1j jlnj Cond Cond_eff
23 3.820(�2) 3.156 1.088 2.899 1.500
24 4.737(�3) 4.361(�5) 3.143 5.444(�1) 5.774 2.899
25 5.885(�4) 4.086(�6) 3.141 2.722(�1) 1.154(+1) 5.7742
26 7.342(�5) 1.695(�7) 3.141 1.361(�1) 2.308(+1) 1.154(+1)
27 9.172(�6) 5.955(�9) 3.141 6.805(�2) 4.616(+1) 2.308(+1)
28 1.146(�6) 1.964(�10) 3.141 3.402(�2) 9.233(+1) 4.616(+1)
29 1.432(�7) 6.293(�12) 3.141 1.701(�2) 1.846(+2) 9.233(+1)
210 1.791(�8) 2.978(�14) 3.141 8.506(�3) 3.693(+2) 1.846(+2)
Table 2
The values of lk , bk and yk at n¼25.
k lk bk yk k lk bk yk
1 3.141 �1.001(�15) �1.570 17 0.408 �8.081(�16) �1.570
2 3.141 �2.053(�15) �1.570 18 0.369 6.993(�16) 1.570
3 3.141 3.162(�15) 1.570 19 0.369 �2.224(�16) �1.570
4 1.571 �6.249 �0.102 20 0.339 �1.055(�15) �1.570
5 1.571 �6.456(�1) 1.467 21 0.339 3.333(�16) 1.570
J. Huang et al. / Engineering Analysis with Boundary Elements 35 (2011) 667–677672
and then obtain uh(y) ðyAO\GÞ by
uhðyÞ ¼�1
2pXd
m ¼ 1
Xnm
i ¼ 1
hmflnjy�xmðtiÞjgjxumðtiÞjwhmðtiÞ: ð3:18Þ
Step 3. From (3.17) we have
WðtiÞ�1
d
Xd
m ¼ 1
WhðmÞ ðtiÞ
����������r WðtiÞ�
8
7
Xd
m ¼ 1
WhðmÞ ðtiÞ� d�7
8
� �Whð0Þ ðtiÞ
" #����������
þ8d�7
7
� �1
d
Xd
m ¼ 1
WhðmÞ ðtiÞ�Whð0Þ ðtiÞ
����������
r8d�7
7
� �1
d
Xd
m ¼ 1
WhðmÞ ðtiÞ�Whð0Þ ðtiÞ
����������þoðh3
0Þ: ð3:19Þ
Note that the most right-hand side of (3.19) provides the aposteriori error estimates.
3.3. Condition number and effective condition number
Take the first kind BIEs in Section 3.2 for example. Under theabove assumptions, the operators A and B in (3.13) are symmetric,and the matrices Ah and Bh in (3.14) are also symmetric. FollowingSection 2.2, we can obtain the similar bounds of condition numberand effective condition number (see [24]).
Theorem 3.1. Assume that Gm ðm¼ 1, . . . ,dÞ is smooth curve and
G¼Sd
m ¼ 1 Gm with CGa1. Let Ah be defined by rules (3.15) and Bh be
defined by the trapezoidal or the midpoint rule [15]. Then the condition
number for (3.14) has the bound
Cond¼ Oðh�1minÞ, hmin ¼ min
1rmrdhm, ð3:20Þ
and effective condition number for (3.14)
Cond_eff ¼JbJ
minkjlkjJxJ¼Oðh�1
minÞ: ð3:21Þ
Eqs. (3.20) and (3.21) imply
Cond_eff ¼OðCondÞ: ð3:22Þ
Hence, the numerical stability of the AQMs is excellent for the firstkind BIEs, which also agrees with [13]. The new stability analysis inthis paper enhances the AQMs, whose error analysis has alreadybeen explored in [20–24,38]. Note that the improvements ofCond_eff for stability are insignificant, since the Cond itself isnot large in practical applications, due to not small h in computa-tion. However, when CG-1 (see [13]) the values of Cond are large,and the Cond_eff may be much smaller than Cond.
Note that the analysis in this paper is under the condition of thelogarithmic capacity (transfinite diameter) CGa1, where CG is just thediameter of the circular boundaryG. Interestingly, even for the Dirichletproblem (3.1), when CG ¼ 1 the solutions may not exist, or not unique ifexisting, to cause a singularity of the discrete algebraic equations. Theproblem with CG ¼ 1 in the BEM is called the degenerate scaleproblems; the treatments for such a difficulty in computation arereported for circular domains with circular holes in Chen et al. [10].
6 1.049 �2.155(�15) �1.570 22 0.316 �7.770(�16) �1.570
7 1.049 �1.836(�15) �1.570 23 0.316 �4.167(�17) �1.570
8 0.789 �1.199(�15) �1.570 24 0.299 2.677(�16) 1.570
9 0.789 3.882(�15) 1.570 25 0.299 �4.991(�15) �1.570
10 0.634 �3.711(�16) �1.570 26 0.287 �4.814(�16) �1.570
11 0.634 1.119(�16) 1.570 27 0.287 1.118(�16) 1.570
12 0.532 1.944(�15) 1.570 28 0.278 8.322(�16) 1.570
13 0.532 3.536(�16) 1.570 29 0.278 2.778(�16) 1.570
14 0.460 1.117(�15) 1.570 30 0.273 �3.600(�16) �1.570
15 0.460 1.278(�15) 1.570 31 0.273 �5.557(�17) �1.570
16 0.408 �1.668(�16) �1.570 32 0.272 �9.434(�16) �1.570
4. Analysis of effective condition number
In [26,27], the effective condition number is applied to the finitedifference method (FDM) for solving Poisson’s equation withDirichlet boundary conditions,
�Du¼ f in O,
u¼ g on G,
(ð4:1Þ
where D¼ @2=@x2þ@2=@y2, and O is a polygon with the boundary@O¼G. The traditional condition number of the difference matrixis well known as
Cond¼Oðh�2minÞ, ð4:2Þ
where hmin is the minimal meshspacing of difference grids. In [26],the bounds of Cond_eff were derived, to give
Cond_effrc Jf J0,Oþh�1=2h�1minJgJ0,G
n o, ð4:3Þ
where Jf J0,O and JgJ0,G are the Sobolev norms, h is the maximalmeshspacing, and c is a constant independent of h. Evidently,Cond_eff in (4.3) is smaller than Cond in (4.2). In particular, whenthe boundary conditions are homogeneous, i.e., g¼0, we obtain
Cond_eff ¼Oð1Þ, ð4:4Þ
which is significantly smaller than Cond in (4.2).In [31,36], the effective condition number is applied to the
collocation Trefftz method (CTM) for Laplace’s and biharmonicequations with crack singularities. The condition number growsexponentially with respect to the number of singular solutions used,but the effective condition number grows polynomially (even linearly[31]). The improvements of Cond_eff for stability are most significant.
In Sections 2 and 3, for numerical BIEs of the first kind by AQMs,when CGa1, Cond_eff and Cond have the same growth rates ash-0 (see (3.22)). Also from the data in Tables 1–3 given in Section 5later, we can see
Cond
Cond_eff¼ cAð1:5,3:5Þ:
Hence, the improvements of Cond_eff over Cond are insignificantfor stability analysis. For numerical partial differential equations(PDEs) and numerical BIEs, why is the performance of Cond_eff so
Table 3The errors en and en
E, Cond and Cond_eff for (5.7).
n en enE
jl1j jlnj Cond Cond_eff
24 2.477(�3) 2.458(+1) 5.474(�1) 4.489(+1) 2.929(+1)
25 2.441(�4) 7.481(�5) 2.458(+1) 2.722(�1) 9.030(+1) 5.879(+1)
26 3.045(�5) 7.296(�8) 2.458(+1) 1.361(�1) 1.806(+2) 1.173(+2)
27 3.810(�6) 4.061(�9) 2.458(+1) 6.805(�2) 3.612(+2) 2.346(+2)
28 4.765(�7) 1.529(�10) 2.458(+1) 3.402(�2) 7.224(+2) 4.693(+2)
29 5.956(�8) 5.048(�12) 2.458(+1) 1.701(�2) 1.444(+3) 9.387(+2)
210 7.445(�9) 1.593(�13) 2.458(+1) 8.506(�3) 2.889(+3) 1.877(+3)
J. Huang et al. / Engineering Analysis with Boundary Elements 35 (2011) 667–677 673
different? What are the rationales behind? Below, we will exploresuch distinct characteristics of Cond_eff, based on matrix analysis.
4.1. Effective condition number for first kind BIEs
Denote the real eigenvalues li of the discrete stiffness matrix Kh
(or Ah) in Sections 2 and 3 as a descent order,
jl1j4 jl2jZ � � �Z jlnj40: ð4:5Þ
Based on the analysis of Sections 2 and 3 (see (2.14) and (2.23)),we assume that
jl1j � Oð1Þ, jlnj � OðhpÞ, p40, ð4:6Þ
where p¼1. Hence, the traditional condition number has the bound
Cond¼jl1j
jlnj� Oðh�pÞ:
Based on Sections 2 and 3 (see (2.24) and (2.25)), there exist the bounds
JxJ � JbJ � O h�1=2�
: ð4:7Þ
We have the following theorem.
Theorem 4.1. Let (4.6) and (4.7) hold. Then there exists the
equivalence
Cond_eff � OðCondÞ:
Proof. We have from (4.5) and (4.7)
Cond_eff ¼JbJ
jlnjJxJ� O
1
jlnj
� �� O
jl1j
jlnj
� �� OðCondÞ: &
Denote the eigenpairs ðli,uiÞ of Kh, and the angles yi between band ui by
cosyi ¼ cosðb,uiÞ ¼ðb,uiÞ
JbJ¼
bi
JbJ, ð4:8Þ
where bi ¼ ðb,uiÞ ¼ uTi b. We have the following lemmas.
Lemma 4.1. Let ðli,uiÞ be eigenpairs of matrix Kh. Suppose that matrix
Kh is symmetric and nonsingular, and that from (4.7)
JxJ � JbJ: ð4:9Þ
Then there exists the bound,
Xn
i ¼ 1
cos2yi ¼ 1, ð4:10Þ
and
Xn
i ¼ 1
cos2yi
l2i
� Oð1Þ, ð4:11Þ
where the angles yi are given in (4.8).
Proof. We have b¼Pn
i ¼ 1 biui. Then we obtain from (4.8)
JbJ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiXn
i ¼ 1
b2i
vuut ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXn
i ¼ 1
cos2yi
vuut8<:
9=;JbJ,
to give the first desired result (4.10). Next, from x¼K�1h b we have
x¼Pn
i ¼ 1bi
liui. Then there exists an equality,
JxJ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiXn
i ¼ 1
b2i
l2i
vuut ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXn
i ¼ 1
cos2yi
l2i
vuut JbJ:
We obtain from (4.9)
Xn
i ¼ 1
cos2yi
l2i
¼JxJ2
JbJ2� Oð1Þ: ð4:12Þ
This is the second result (4.11), and completes the proof of Lemma4.1. &
From (4.5) and (4.6), we may denote the eigenvalues by
jlij ¼ cihpi � Oðhpi Þ, ð4:13Þ
where ci are positive constants, and the powers pi are given by
0¼ p1rp2o � � �rpn ¼ p:
Lemma 4.2. Let the conditions in Lemma 4.1 and (4.13) hold. Then
there exist the bounds
jcosyij ¼ Oðhpi Þ ¼OðjlijÞ, ð4:14Þ
where yi are defined in (4.8).
Proof. We have from Lemma 4.1
jcosyij ¼ jlij
ffiffiffiffiffiffiffiffiffiffiffiffiffifficos2yi
l2i
sr jlij
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiXn
i ¼ 1
cos2yi
l2i
vuut rcjlijrchpi ,
to give the desired result (4.14). This completes the proof ofLemma 4.2. &
In fact, u1 in eigenpair ðl1,u1Þ is the lowest frequency eigen-vector, and un in eigenpair ðln,unÞ is the highest frequencyeigenvector. The ui is said the low frequency eigenvector if itscorresponding eigenvalue li satisfying jlij � Oðjl1jÞ � Oð1Þ, or thehigh frequency eigenvector if jlij � OðjlnjÞ. Also b has a portioneddistribution on ui if
jcosyij ¼ jcosðb,uiÞjZc040, ð4:15Þ
where c0 is a constant independent of h. Hence, Lemma 4.2 impliesthat b must not have a portioned distribution on high frequencyeigenvectors, because jcosyij-0 when h is small.
Lemma 4.3. Let the conditions in Lemma 4.1 hold. Then the solution
vector x has a portioned distribution on low frequency eigenvector uk if
and only if b has a portioned distribution on uk.
Proof. First suppose that b has a portioned distribution on the lowfrequency eigenvector uk with jlkj � Oðjl1jÞ � Oð1Þ, where k is a smallinteger. Eq. (4.15) gives jcosðb,ukÞjZc040. Since JbJ=JxJZc140from (4.9), where c1 is a constant independent of h, we have
jcosðx,ukÞj ¼jðx,ukÞj
JxJ¼jðK�1
h b,ukÞj
JxJ¼jðb,K�1
h ukÞj
JxJ
¼jðb,ukÞj
jlkjJxJ¼jcosðb,ukÞj
jlkj
JbJ
JxJZ
c0c1
jlkj¼ c140, ð4:16Þ
where we have used jlkj � Oð1Þ, and c1 is a constant independent of h.This implies that x has a portioned distribution on uk.
On the other hand, suppose that jcosðx,ukÞjZc040, we have
jcosðb,ukÞj ¼jðb,ukÞj
JbJ¼jðKhx,ukÞj
JbJ¼jðx,KhukÞj
JbJ
¼jlkjjðx,ukÞj
JbJ¼ jlkjjcosðx,ukÞj
JxJ
JbJZ jlkj
c0
c2Zc240,
where we have used the bound, JxJ=JbJZc2 from (4.9), where c2
and c2 are constants independent of h. This implies that b
J. Huang et al. / Engineering Analysis with Boundary Elements 35 (2011) 667–677674
has a portioned distribution on uk, and completes the proof of
Lemma 4.3. &
Theorem 4.2. Let all conditions in Lemma 4.1 hold. Suppose that x has
a portioned distribution on a low frequency eigenvector uk. Then there
exists a constant c0 with 0oc0r1 independent of h such that
Cond_effZc0Cond: ð4:17Þ
Proof. By the assumption, we have jcosðx,ukÞjZc040. Denote thesolution vector x¼ x1þx2, where x1 ¼ ðx,ukÞuk. Then we have
Jx1J¼ jðx,ukÞj ¼ jcosðx,ukÞjJxJZc0JxJ,
to give
JxJr1
c0Jx1J: ð4:18Þ
Since b¼Khx, we have b¼ b1þb2, where b1 ¼Khx1 ¼ lkx1. Hence,we obtain
Jx1J¼1
jlkjJb1J: ð4:19Þ
Combining (4.18) and (4.19) yields
JxJr1
c0
1
jlkjJb1J:
Now we obtain
Cond�eff ¼JbJ
jlnjJxJZc0jlkj
jlnj
JbJ
Jb1J: ð4:20Þ
Since JbJZJb1J and jlkj=jl1jZgk40 with gk � Oð1Þ. Eq. (4.20)leads to
Cond_effZc0jlkj
jlnjZc0gk
jl1j
jlnj¼ c0gkCond¼ c0Cond,
where c0 ¼ c0gk. This is the desired result (4.17), and completes theproof of Theorem 4.2. &
Remark 4.1. Theorems 4.1 and 4.2 can be applied to manynumerical methods for BIEs of the first kind in [13] and the secondkind in Atkinson and Han [1], where the pesudodifferentialoperator is of the boundary integral type, and other boundaryintegration equation methods (BIEM) [18,25]. The stability analysisin this paper may also be extended to the BIEs with dual boundaryintegral equations (DBIEs) in Chen et al. [6–8]. Moreover, theconclusions in this subsection are also valid for the boundaryelement method (BEM).
4.2. Effective condition number for numerical PDEs
Now, we turn to study Cond_eff for numerical PDEs. The discretealgebraic equations of (4.1) from the finite difference method(FDM) or the finite element method (FEM), and are denoted by thematrix form
Khx¼ b, ð4:21Þ
where the matrix Kh is symmetric and positive definite. Also denoteðli,uiÞ the eigenpairs of matrix Kh, and the eigenvalues are alsogiven in a descent order [26,27]
c1h�p ¼ l14l2Z � � �Zln ¼ c040, ð4:22Þ
where pZ2,c1 and c0 are two constants independent of h. For thehomogeneous boundary conditions (i.e., g¼0), the following rela-tions are satisfied (see [26,27,30,32–34]):
JxJ � JbJ � Oðh�1Þ: ð4:23Þ
We can derive the following theorems by following the argumentsof Theorems 4.1 and 4.2.
Theorem 4.3. Let (4.22) and (4.23) be given. Suppose that x has a
portioned distribution on a low frequency eigenvector uk. Then
Eq. (4.4) holds.
From Theorems 4.1–4.3, the effective condition number may besignificantly smaller than Cond only for numerical PDEs, but not fornumerical BIEs whose pseudodifferential operator is of boundaryintegral type as given. The intrinsic characteristics result from thefollowing facts. The differential operator �D is unbounded, so that foreigenvalues of the discrete matrix have the bounds in (4.22)
ln � Oð1Þ, l1 � Oðh�pÞ, pZ2: ð4:24Þ
On the other hand, the integral compact operator is bounded, so thateigenvalues of the discrete matrix have the following different boundsin (4.6):
jl1j � Oð1Þ, jlnj � OðhpÞ, pZ1: ð4:25Þ
Hence, we conclude that the effective condition number is significantfor numerical PDEs, but not for numerical BIEs. For numerical BIEs, thetraditional Cond ¼ O(h�1) is not large, and the Cond_eff is insignificantfor stability. This fact also displays that the Cond_eff is really significant
when the traditional Cond is large for numerical PDEs.For numerical PDE by the Trefftz method (TM) and the colloca-
tion Trefftz method (CTM), Eq. (4.21) is obtained with matrixKhARm�nðmZnÞ, the above analysis and conclusions are also valid(see [28,29,31,35–37]).
5. Numerical experiments
We carry out three experiments by AQMs and h3-Richardson’sextrapolation or splitting extrapolation methods (SEM), and verifythe stability and the error analysis made in the above sections.
Example 1. Let G be a circle with radius e�1=2. Based on [1,44,47],CG ¼ e�1=2a1. Consider the typical BIEs of the first kind
�
Z p
�pln 2e�1=2sin
t�t2
��������wðtÞ dt¼ p
2cos2t, ð5:1Þ
where the true solution w(t)¼cos 2t. In Table 1, we list Cond,Cond_eff and the errors
en ¼ max1r irn
jwðtiÞ�whðtiÞj, eEn ¼ max
1r irnjwðtiÞ�wEðtiÞj,
where Richardson’s interpolation
wEðtiÞ ¼ ð8wh=2ðtiÞ�whðtiÞÞ=7:
Table 2 lists Zk and yk, to indicate v(t) having a proportioneddistribution on uk, discussed later.
Compared the kernel in (5.1) with that in (2.2), there is a factor2p of differences. We have from (2.14) and (2.23)
jl1j ¼maxijlij ¼ p, ð5:2Þ
jlnj ¼minijlij ¼
pn: ð5:3Þ
When n¼210¼1024, the theoretical value of jlnj is given by
jlnj ¼p
1024¼ 3:068ð�3Þ: ð5:4Þ
On the other hand, from Table 1 we can see the numerical data for n
¼ 210,
jl1j ¼ 3:141, ð5:5Þ
jlnj ¼ 8:506ð�3Þ: ð5:6Þ
Eqs. (5.2) and (5.5) coincide with each other perfectly. The value ofjlnj in (5.4) is smaller but close to that in (5.6), to verify our lowerestimation in Theorem 2.1 very well.
Table 5Errors, Cond and Cond_eff for (5.8).
(n1,n2) juh�uj jlnj jl1j Cond Cond_eff
(4,4) 4.413(�2) 0.104 4.287 40.865 12.042
(8,4) 2.166(�2) 0.113 4.147 36.420 11.979
(4,8) 2.166(�2) 0.113 4.147 36.420 12.049
eE 7.229(�3)
(8,8) 1.738(�3) 0.055 4.374 78.474 22.395
(16,8) 9.452(�4) 0.056 4.312 76.007 23.820
(8,16) 9.452(�4) 0.056 4.312 76.007 23.856
eE 7.495(�5)
(16,16) 1.383(�4) 0.028 4.378 154.828 44.046
(32,16) 7.805(�5) 0.028 4.357 153.680 47.609
(16,32) 7.805(�5) 0.028 4.357 153.680 47.618
eE 5.184(�7)
(32,32) 1.725(�5) 0.014 4.375 308.348 87.754
(64,32) 9.703(�6) 0.014 4.369 308.234 95.208
(32,64) 9.703(�6) 0.014 4.369 308.234 95.210
eE 1.350(�9)
Table 6
The values of lk , bk and cosyk at (n1, n2) ¼(8,8).
k lk bk jcosykj k lk bk jcosykj
1 4.374 3.006 0.752 9 0.280 �3.608(�16) 6.123(�17)
2 2.309 1.110(�16) 6.123(�17) 10 0.225 �0.021 5.334(�3)
3 2.076 2.542 0.636 11 0.218 8.326(�17) 6.123(�17)
4 0.941 0.653 0.163 12 0.187 0.014 3.496(�3)
5 0.551 1.110(�16) 6.123(�17) 13 0.184 4.718(�16) 6.123(�17)
6 0.532 0.258 0.065 14 0.162 5.153(�3) 1.288(�3)
7 0.396 2.775(�16) 6.123(�17) 15 0.090 �2.775(�16) 6.123(�17)
8 0.308 0.078 0.019 16 0.055 3.665(�17) 6.123(�17)
J. Huang et al. / Engineering Analysis with Boundary Elements 35 (2011) 667–677 675
Example 2 (Lu and Huang [38], Sidi and Israrli [43]). Let G bexðtÞ ¼ c0ðe
ffiffiffiffiffi�1p
tþc1e�ffiffiffiffiffi�1p
tÞ,tA ½0,2p�, which is an elliptic curve,where c0¼50 and c1¼0.5. Since CGa1, the boundary integralequation:Z 2p
0lnjxðtÞ�xðtÞjwðtÞ dt¼ 2plnjxðtÞj, ð5:7Þ
has the unique solution
wðtÞ ¼ 1þ4X1k ¼ 1
ð�1Þkck
1
1þc2k1
cosð2ktÞ:
The computed results are listed in Tables 3 and 4. A comparison ofTable 3 with Table 1 is given below.
Example 3 (Sloan and Spence [45]). Let G be an open contour oflength 2, in the form of a right-angled wedge:
G¼ fðx1,0Þ : 0rx1r1g [ fð0,x2Þ : 0rx2r1g :¼ G1 [ G2:
The integral equation is chosen as
�
ZG
lnjy�xjvðxÞ dsx ¼ 1 for ðy1,y2ÞAG: ð5:8Þ
We compute the numerical solution of
uðyÞ ¼ �
ZG
lnjy�xjvðxÞ dsx
at (0.5,0.5), whose true value u(0.5,0.5) ¼ 0.621455343.
From [45], although the true solution vðxÞ ¼ Oðjx�x0j�1=3Þ singu-
larity at the right-angled corner, the dominant singularities in v(x)occur at two ends, as u¼Oðjx�x0j
�1=2Þ. Based on [23,24], usingj6ðtÞ inthe periodic transformation (3.7), we obtain the numerical results atQ¼(0.5,0.5) by AQMs and list Cond and Cond_eff in Table 5. Let n1 andn2 be the numbers of uniform partition on [0,1] for G1 and G2,respectively. Based on (3.16), we can obtain the splitting extrapolationerrors eEðQ Þ ¼ juEðQ Þ�uðQ Þj, where the splitting extrapolation values
uEðQ Þ ¼8
7
Xd
m ¼ 1
uhðmÞ ðQ Þ� d�7
8
� �uhð0Þ ðQ Þ
" #, d¼ 2:
The errors juhðQ Þ�uðQ Þj and the splitting extrapolation errors eE(Q) arealso listed in Table 5.
Now, let us examine more the results in Tables 1 and 3. We cansee numerically,
e¼ Oðh3Þ, eE ¼ Oðh5Þ,
and
jl1j � c, jlnj ¼Oðh�1Þ: ð5:9Þ
Table 4
The values of lk , bk and yk at n¼25.
k lk bk yk k lk bk yk
1 24.58 1.534(+2) 0.000 17 0.406 �2.365(�14) �1.570
2 4.712 �2.931(�14) �1.570 18 0.369 �8.881(�15) �1.570
3 1.964 �1.964(�14) �1.570 19 0.368 1.159(�14) 1.570
4 1.571 �2.086(�14) �1.570 20 0.339 8.881(�15) 1.570
5 1.180 4.440(�15) 1.570 21 0.339 2.471(�14) 1.570
6 1.179 �9.014(�15) �1.570 22 0.316 �8.881(�15) �1.570
7 0.918 �1.536(�14) �1.570 23 0.316 1.366(�14) 1.570
8 0.838 8.881(�14) �1.570 24 0.299 �1.332(�14) �1.570
9 0.740 1.831(�14) 1.570 25 0.299 1.233(�15) 1.570
10 0.653 6.217(�15) 1.570 26 0.287 3.908(�14) 1.570
11 0.614 1.222(�14) 1.570 27 0.287 �1.411(�14) �1.570
12 0.540 0.000 1.570 28 0.278 0.000 1.570
13 0.523 �2.735(�14) �1.570 29 0.278 6.920(�15) 1.570
14 0.464 �3.730(�14) �1.570 30 0.273 �8.881(�15) �1.570
15 0.457 �3.193(�14) �1.570 31 0.273 2.418(�14) 1.570
16 0.409 �1.687(�14) �1.570 32 0.272 �5.329(�15) �1.570
Eqs. (5.9) coincide with the theoretical estimates of jl1j and jlnj
given in Theorem 2.1. Next, from Tables 1 and 3 we find
Cond¼Oðh�1Þ, Cond_eff ¼Oðh�1Þ, ð5:10Þ
which are also consistent with Theorems 2.1, 2.2 and 3.1. Moreover,from Table 5 we have
juh�ujð16,16Þ
juh�ujð32,32Þ
¼1:383ð�4Þ
1:725ð�5Þ¼ 8:01, ð5:11Þ
and
eEð16,16Þ
eEð32,32Þ
¼5:184ð�7Þ
1:350ð�9Þ¼ 384: ð5:12Þ
Eq. (5.11) displays the O(h3) convergence rate, and Eq. (5.12) alsodisplays the O(h6) convergence rate since 384Z26
¼ 64. Hence, theSEMs can provide more accurate solutions. Note that from Table 5with the total number n ¼ n1 + n2 ¼32 and 64, the errors of SEMsare 5.184(�7) and 1.350(�9), respectively. In contrast, whenn¼256 the solution uh
¼0.62125 is given in [45] by Galerkinmethods. This fact displays the efficiency of AQMs and SEMs.
Finally, to scrutinize the spectral distribution of vector b in uk,we compute all bk ¼ uT
k b and yk ¼ arccosðbk=JbJÞ, and list them inTables 2 and 4 in the descending order of jlkj. We can see thatyk � 7p=2 except y4 ¼�0:102 and y1 ¼ 0:000 from Tables 2 and 4,respectively. This fact coincides with Theorem 4.2. In Table 6, welist all bk and cosyk ¼ bk=JbJ, to find the dominant distribution ofbk. We can see that cosy1 ¼ 0:752, cosy3 ¼ 0:636 and cosy4 ¼ 0:163.Since cos2y1þcos2y3þcos2y4 ¼0:996569� 1, the dominate vec-tors (i.e., the proportional distribution) of b happen just at the threelow frequency eigenvectors u1, u3 and u4, also to support theassumptions in Theorems 4.2 and 4.3.
J. Huang et al. / Engineering Analysis with Boundary Elements 35 (2011) 667–677676
6. Concluding remarks
1.
A new and systematic stability analysis is made for the advanced(i.e., mechanical) quadrature methods (AQMs) for the first kindBIEs in [20–24,38], based on Cond and Cond_eff. For continuous,singularity and discontinuous problems [20], there exist theboundsCond_eff � Cond � Oðh�1Þ: ð6:1Þ
Eq. (6.1) display an excellent stability of the AQMs. For discretematrix Ah in (2.5) from (2.3), the bound as (2.23) in Theorem 2.1is derived, by using the special cðzÞ function.
2.
Numerical experiments are carried out for the arbitrary boundaryG with CGa1 by AQMs and SEAs, and the computed results inExamples 1–3 and those in [20] agree with (6.1) perfectly. Also, thenumerical condition numbers show the optimal O(h�1).3.
Let us compare the stability analysis in this paper with that inChristiansen and Saranen [13] in more detail. In [13, p. 48], thealgorithms (1.4) were also proposed, where the local conditionnumber was called. In this paper, we discuss three types of thefirst kind BIEs in Sections 2 and 3. Although the error analysishas been made in [20–24,38], the excellent stability as (6.1) isconfirmed for the AQMs, to enhance its application.4.
In Section 4, based on matrix analysis, we explore the intrinsiccharacteristics, to give a theoretical justification to explain the factthat Cond and Cond_eff have the same growth rate for numericalBIEs. Note that the improvements of stability by the Cond_eff arenot as significant as those for numerical PDEs in [26–37,19]. Therationale is that the operators of BIEs and PDEs are bounded andunbounded, respectively. Hence, the eigenvalues of their discretematrices have different bounds in (4.24) and (4.25).5.
The analysis in Section 4 also explains that why the effectivecondition number is not developed until its applications tonumerical PDEs in [26–37,19]. Because the effective conditionnumber is significantly smaller than Cond for numerical PDEs,but not for numerical BIE in this paper and numerical linearalgebra [27]. This provides an important and complete knowl-edge of effective condition number, as a concise review oneffective condition number in [26–37,19].Acknowledgements
Authors are grateful to Prof. J.T. Chen and the reviewers for theirvaluable comments and helpful suggestions. This work is sup-ported by the National Natural Science Foundation of China underGrant 10871051, Doctoral Program of the Ministry of Educationunder Grant 20090071110003 and 973 Program Project underGrant 2010CB327900, and Shanghai Science & TechnologyCommittee under Grant 09DZ2272900 and Shanghai EducationCommittee (Dawn Project).
References
[1] Atkinson KE. The numerical solution of integral equations of the second kind.Cambridge: Cambridge University Press; 1997.
[2] Axelsson O, Kaporin I. On the sublinear and superlinear rate of convergence ofconjugate gradient methods. Numer Algorithms 2000;25:1–22.
[3] Axelsson O, Kaporin I. Error norm estimation and stopping criteria inpreconditioned conjugate gradient iterations. Numer Linear Algebra Appl2001;8:265–86.
[4] Banoczi JM, Chiu M, Cho GE, Ipsen CF. The lack of influence of right-hand side onaccuracy of linear system solution. SIAM J Sci Comput 1998;20:203–27.
[5] Chan TF, Foulser DE. Effectively well-conditioned linear systems. SIAM J Sci StatComput 1988;9:963–9.
[6] Chen JT, Chiu YP. On the pseudo-differential operators in the dual boundaryintegral equations using degenerate kernels and circulants. Eng Anal BoundaryElem 2002;26:41–53.
[7] Chen JT, Hong HK. Review of dual boundary elements with emphasis onhypersingular integrals and divergent series. Appl Mech Rev ASME 1999;52:17–33.
[8] Chen JT, Wu CS, Chen KH. A study of free terms for plate problems in the dualboundary integral equations. Eng Anal Boundary Elem 2005;29:435–46.
[9] Chen JT, Chen KH, Chen IL, Liu LW. A new concept of modal participation factorfor numerical instability in the dual BEM for exterior acoustics. Mech ResCommun 2003;26(2):161–74.
[10] Chen JT, Lee CF, Chen IL, Lin JH. An alternative method for degenerate scaleproblems in boundary element methods for the two-dimensional Laplaceequation. Eng Anal Boundary Elem 2002;26:559–69.
[11] Christiansen S. Numerical solution of an integral equation with a logarithmickernel. BIT 1971;11:276–87.
[12] Christiansen S, Hansen PC. The effective condition number applied to erroranalysis of certain boundary collocation methods. J Comput Appl Math1994;54(1):15–36.
[13] Christiansen S, Saranen J. The conditioning of some numerical methods for firstkind boundary integral equations. J Comput Appl Math 1996;67:43–58.
[14] Davis P. Circulant matrices. New York: John Wiley & Sons; 1979.[15] Davis PJ, Rabinowitz P. Methods of numerical integration. second ed..
Academic Press; 1984. p. 315.[16] Golub GH, van Loan CF. Matrix computations. third ed.. Baltimore and London:
The Johns Hopkins; 1996.[17] Gradshteyan IS, Ryzhik ZM. Tables of integrals, series and products. New York:
Academic Press; 1980.[18] Hong HK, Chen JT. Derivations of integral equations of elasticity. J Eng Mech
ASCE 1988;114(6):1028–44.[19] Huang HT, Li ZC. Effective condition number and superconvergence of the
Trefftz method coupled with high order FEM for singularity problems. Eng AnalBoundary Elem 2006;30:270–83.
[20] Huang J, Li ZC, Chen IL, Cheng AH-D. Advanced quadrature methods andsplitting extrapolation algorithms for first kind boundary integral equations ofLaplace’s equation with discontinuity solutions. Eng Anal Boundary Elem2010;34(2010):1003–8.
[21] Huang J, Li ZC, Lu T, Zhu R. Splitting extrapolations for solving boundaryintegral equations of mixed boundary conditions of mechanical quadraturemethods. Taiwanese J Math 2008;12:2341–61.
[22] Huang J, Lu T. Mechanical quadrature methods and their extrapolation forsolving boundary integral equation of Steklov eigenvalue problem. J ComputMath 2004;22:719–26.
[23] Huang J, Lu T. The mechanical quadrature methods and their splittingextrapolations for first-kind boundary integral equation on polygonal regions(in Chinese). Math Numer Sin 2004;26:51–9.
[24] Huang J, Lu T, Li ZC. Mechanical quadrature methods and their splittingextrapolations for boundary integral equations of first kind on open contours.Appl Numer Math 2009;59:2908–22.
[25] Kuo SR, Chen JT, Huang CX. Analytical study and numerical experiments fortrue and spurious eigensolutions of a circular cavity using the real-part dualBEM. Int J Numer Methods Eng 2000;48(9):1401–22.
[26] Li ZC, Chien CS, Huang HT. Effective condition number for finite differencemethod. J Comput Appl Math 2007;198:208–35.
[27] Li ZC, Huang HT. Effective condition number for numerical partial differentialequations. Numer Linear Algebra Appl 2008;15:575–94.
[28] Li ZC, Huang HT. Effective condition number for simplified hybrid Trefftzmethods. Eng Anal Boundary Elem 2008;27:757–69.
[29] Li ZC, Huang TH. Study on effective condition number for collocation methods.Eng Anal Boundary Elem 2008;32:839–48.
[30] Li ZC, Huang HT. Effective condition number for finite element method usinglocal mesh refinements. Appl Numer Math 2009;59:1779–95.
[31] Li ZC, Huang HT, Chen JT, Wei Y. Effective condition number and its application.Computing 2010;89:87–112.
[32] Li ZC, Huang HT, Hunag J. Stability analysis and superconvergence for penaltyTrefftz method coupled with FEM for singularity problems. Eng Anal BoundaryElem 2007;31:631–45.
[33] Li ZC, Huang HT, Huang J. Effective condition number of Hermite finite elementmethods for biharmonic equations. Appl Numer Math 2008;58:1291–308.
[34] Li ZC, Huang HT, Huang J. Superconvergence and stability for boundary penaltytechniques of finite difference methods. Numer Methods PDEs 2008;24:972–90.
[35] Li ZC, Huang HT, Huang J, Ling L. Stability analysis and superconvergence forthe penalty plus hybrid and the direct Trefftz method for singularity problems.Eng Anal Boundary Elem 2007;31:163–75.
[36] Li ZC, Lu TT, Wei Y. Effective condition number of Trefftz method forbiharmonic equations with crack singularities. Numer Linear Algebra Appl2009;16:145–71.
[37] Lu TT, Chang CM, Huang HT, Li ZC. Stability analysis of Trefftz methods for thestick–slip problem. Eng Anal Boundary Elem 2009;33:474–84.
[38] Lu T, Huang J. Quadrature methods with high accuracy and extrapolation forsolving boundary integral equations of the first kind (in Chinese). Math NumerSin 2000;22:59–72.
[39] Rice JR. Matrix computations and mathematical software. New York: McGraw-Hill Book Company; 1981.
[40] Rump SM. Structure perturbation, part I: normwise distances. SIAM J MatrixAnal Appl 2003;25:1–30.
[41] Saranen J. The modified discrete spectral collocation method for first kindintegration with logarithmic kernel. J Integr Equation Appl 1993;5:547–67.
[42] Sidi A. A new variable transformation for numerical integration. Int Ser NumerMath 1993;112:359–73.
J. Huang et al. / Engineering Analysis with Boundary Elements 35 (2011) 667–677 677
[43] Sidi A, Israrli M. Quadrature methods for periodic singular Fredholm integralequation. J Sci Comput 1988;3:201–31.
[44] Sloan IH, Spence A. The Galerkin method for integral equations of first-kindwith logarithmic kernel: theory. IMA J Numer Anal 1988;8:105–22.
[45] Sloan IH, Spence A. The Galerkin method for integral equations of first-kindwith logarithmic kernel: applications. IMA J Numer Anal 1988;8:123–40.
[46] Wilkinson JH. The algebraic eigenvalue problem. Oxford: Clarendon Press;1965.
[47] Yan Y. The collocation method for first kind boundary integral equations onpolygonal domains. Math Comput 1990;54:139–54.
[48] Yan Y, Sloan IH. On integral equations of the first kind with logarithmic kernels.J Integr Equation Appl 1988;1:517–48.