Stability Analysis of Quadrature Methods for Two-Dimensional Singular Integral Equations

Ibrahim Saad Abdel-Fattah Department of Mathematics Faculty of Science Mansoura University Egypt [email protected] April 9, 1997

1991 Mathematics Subject Classi cation. 45L10, 45Exx, 65N38, 65R20. Keywords. singular integral equation, two-dimensional manifold, quadrature method.

Abstract
In this paper we apply a quadrature method based on the tensor product trapezoidal rule to the solution of a singular integral equation over the two-dimensional torus. We prove that this method is stable if and only if a certain numerical symbol does not vanish. For a special kernel function, we present a plot of numerically computed symbol values and, for symmetric kernels (Mikhlin-Giraud kernels), we show that the symbol is di erent from zero if the singular integral operator is invertible. Finally, we prove the convergence of our method and present numerical tests.

1 Introduction
In the last two decades a lot of problems in elasticity, uid mechanics, acoustics, optics, electrostatics, and other elds of engineering have been tackled by boundary element methods (cf. e.g. the overview articles by Mazya 15] and Wendland 36]). These methods include the analysis of strongly singular boundary integral equations ~~ Au = (~I + K )~ = f; a ~u ~ (1.1) where ~I stands for the multiplication operator a (~I )~(s) : = a(s)~(s) a u ~ u ~ multiplying by a real valued function a and K for the integral operator ~
Z

(1.2)

~ u ~~ (1.3) (K u)(s) : = S k(s; t)~(t)dtS over the boundary manifold S . We suppose that the kernel k(s; t) is strongly singular (cf. Section 2). This means that the integral in (1.3) is to be understood in the sense of a Cauchy principal value. In order to get the unknown function u we solve (1.1) ~ numerically. Originally, in the boundary element method this was done by a nite element discretization of (1.1). However, nowadays p- and h-p-methods, collocation, and quadrature schemes are popular as well. Several monographs are devoted to the study of Equation (1.1) and its numerical solution. Let us mention here e.g. the books written by Mikhlin, Pro dorf 16], Mikhlin, Morozov, Paukshto 17], Muskhelishvili 19], Pro dorf, Silbermann 28], and Parton, Perlin 20]. The main objective of this paper is to analyze quadrature methods for the numerical solution of singular integral equations over two-dimensional boundary manifolds and to prove convergence results similar to those known for collocation. Note that using the concept of strong ellipticity (cf. Stephan and Wendland 35]), the analysis of Galerkin methods for strongly elliptic singular integral equations is easy. The realization of these Galerkin schemes, however requires the computation of two-fold integrals over the boundary and, thus, is very time consuming. To reduce these e orts, collocation methods are applied. In contrast to their successful implementation, the convergence analysis is done for very special situations, only (cf. Pro dorf and Schneider 26]). Moreover, the collocation still requires the computation of singular integrals, which is accomplished by using quadratures. The advantage of quadrature schemes in comparison to Galerkin and collocation methods is that all the integrals are discretized within one discretization step, 1

i.e., quadrature methods are so-called fully discrete schemes. The corresponding number of quadrature knots and therewith, the computation time is much less than for other discretization schemes. The draw back of the quadrature methods is the larger discretization error. Hence a quadrature method could be a good choice if the convergence of Galerkin and collocation schemes is slow due to the lack of smoothness of the right-hand side and the underlying manifold. Moreover, low order quadrature methods can be considered as a starting point for an analysis of higher order fully discrete methods with minimal numbers of quadrature points. We expect that the optimal methods are slight modi cations of our quadrature methods. The theory of one-dimensional spline collocation has been established by Pro dorf and Schmidt 24, 25], Arnold and Wendland 1, 2], Saranen and Wendland 31], and Schmidt 32, 33]. In the end of the 80-ies Hsiao, Pro dorf, and Schneider started to generalize these results to the case of multi-dimensional pseudo-di erential equations. Unfortunately, the technique of Arnold and Wendland 1] could be generalized only by a di cult technical modi cation (cf. Hsiao and Pro dorf 13]). The techniques of Fourier analysis (or circulant techniques) take over to the multi-dimensional case if the underlying manifold is a torus or an open subset of the plane (cf. Pro dorf and Schneider 26, 27]). Note that the restriction to the arti cial torus manifold means the following: The stability of collocation is a local property. Collocation is stable if and only if it is locally stable in the neighborhood of any point of the underlying manifold. The problem of local stability, however, is solved only for points where the mesh is regular, i.e., close to a rectangular mesh over a torus. E.g., if we consider a sphere and take a partition along the lines of constant longitude and latitude, then the resulting grid is regular at any point except the two poles. In other words, the local stability problem is solved at any point of the sphere but the poles. The local stability near the poles is not solved yet. Further investigations for collocation methods are due to Costabel and McLean 8], Dahmen, Pro dorf, and Schneider 9], and Hagen, Roch, and Silbermann 12]. Note that the authors of 9, 12] have even dealt with wavelet collocation methods. Similar to the analysis of collocation, we have to restrict our consideration for quadrature methods to the special case that the underlying manifold is di eomorphic to the torus. Suppose : 0; 1] 0; 1] ! S is a parametrization of S which is 1-periodic in each argument. Then Equation (1.1) takes the form

Au(s) := a(s)u(s) +
where

0;1] 0;1]

k(s; t)u(t)dt = g(s); s 2 0; 1]

0; 1];

(1.4) (1.5)

a(s) = a ( (s)) ; g(s) = g ( (s)) ; u(s) = u ( (s)) ; ~ ~ ~ ~ ( (s); (t)) j 0 (t)j: k(s; t) = k
Discretizing (1.4) with the help of the trapezoidal rule
Z

0;1] 0;1]

'(t)dt
2

1 n?1 '( i ; j ); n2 i;j=0 n n

X

(1.6)

we arrive at the quadrature method

i j i j 1 a ( n ; n ) un ( n ; n ) + n2

n?1 X
k; l = 0
(k; l ) = (i; j )

i j k l k l k ( n ; n ); ( n ; n ) un( n ; n )
!

i j = g ( n ; n ) ; i; j = 0; : : :; n ? 1: (1.7) Note that this method can be derived using the so-called singularity subtraction technique if the kernel function satis es certain symmetry conditions (cf. Section 2). Only if this symmetry assumption is ful lled, then (1.7) can be convergent. If the assumption is violated, then (1.7) needs to be modi ed. For the quadrature method (1.7), we rst investigate the stability, i.e., we check if the discretized integral operator is invertible and if the inverse discretized operator is uniformly bounded for su ciently small mesh size 1=n. Note that stability of the quadrature method implies that the linear systems arising after discretization are well conditioned and that the convergence order of the approximate solution is the same as that of the quadrature rules. In analogy to the collocation, it turns out that stability is a local property. This means, the quadrature method (1.7) for A in (1.4) is stable if and only if it is locally stable (cf. Sections 3 and 4) at any point of the boundary S . The local stability at a given point t on S , however, is the same as the stability of the quadrature method to a convolution operator de ned over the tangent plane, if this convolution operator coincides with A in the neighborhood of t. In other words, it is su cient to consider the stability of quadrature methods applied to singular convolution equations over the plane. The discretized convolution operator turns out to be a discrete convolution matrix. Its stability is determined by the generating symbol which is called numerical symbol of the quadrature method. As the rst main result of this paper we prove that the numerical symbol is bounded (cf. Proposition 3.2). It is invertible (cf. Theorem 3.1) if a simple symmetry assumption for the kernel k(s; t) is ful lled. Thus we derive a su cient condition for the local stability. As the second main result we show that quadrature method (1.7) is stable if and only if it is locally stable (cf. Theorem 4.1), i.e., if and only if the numerical symbol does not vanish over S . Unfortunately, the values of the numerical symbol are given in form of an in nite sum and can be computed by numerical methods, only. We give one example for such a numerical computation (cf. Subsection 3.2). However, for the special case of the integral operator corresponding to the oblique derivative boundary value problem (cf. Section 6 and 18]), the su cient condition for the local stability is ful lled and global stability can be proved. The third main result (cf. Theorem 5.1) concerns the convergence of the quadrature method (1.7). Using the just established stability, we prove that, for any Lipschitz continuous right-hand side f , the solution un of (1.7) tends to the exact solution u of (1.4) in L2, i.e., i ; j ) ? u ( i ; j ) 2 ! 0: un ( n n n n i;j =0 To con rm the theoretically obtained results, we present some numerical tests in Section 6. We consider the singular integral equation corresponding to the oblique derivative boundary value problem for Laplace's equation on an unbounded domain with a boundary
1 n 3
v u uX t

n?1

manifold di eomorphic to the torus. For this equation, we present the approximation errors of the quadrature method (1.7).

2 The Quadrature Method over the Torus

Collocation methods and Galerkin methods are so-called semi-discrete schemes. In fact, to compute the integrals contained in the de nition of the matrix entries one has to apply analytic formulas or quadrature schemes. We now like to give the quadratures in an optimal way (minimal number of quadrature knots) and to perform the stability analysis for the quadrature algorithm simultaneously. This can be done by considering quadrature discretization schemes right from the start. In the case of one-dimensional singular integral equations this is done by Belotserkovski, Lifanov, Pro dorf, Rathsfeld, Sloan, Silbermann 3, 23, 29, 28]. We shall try to generalize these results to two dimensions. Let us consider the singular integral equation (with a classical pseudo-di erential operator of order zero corresponding to the symbol function A(x; ) 2 S 0, cf. e:g. 7])

Au(x) = a(x)u(x) +

over the torus T 2 := R2=Z 2 , where T Z Here kS (x; x ? y) is de ned by

T T

k(x; y)u(y)dyT 2 = g(x); x 2 T 2 T T

(2.1) (2.2) (2.3)

k(x; y) = kS (x; x ? y) + kR(x; y): kS (x; z) =

Z

(x; )eiz d : R A
2

We may suppose that A is a positive homogeneous function in of degree zero with T A 2 C 1 (T 2 R2 nf0g) and that the kernel kS satis es the following conditions : a) kS (x; z) 2 C 1(T 2 R2nf0g). T b) kS (x; tz) = t?2k(x; z), t > 0, x 2 T 2, z 2 R2nf0g. T T c) S kS (x; z)d (z) = 02 k(x; ei )d = 0, x 2 T 2.
R R
1

The additional kernel kR(x; y) is supposed to be continuous and to generate a compact operator. (Note that, for a general classical pseudo-di erential operator of order zero, the kernel kR is weakly singular only. The corresponding operators and discrete operators should be treated in a similar manner as the singular operators. For the sake of simplicity, however, we suppose kR to be continuous). The integral in (2.1) is to be understood as
Z

T2 T

k(x; y)u(y)dyT 2 = T

1 x1 + 1 Z x2 + 2 2

x1 ? 1 2

= "lim0 ?! 4

x2 ? 1 2

k(x; y)u(y)dy1dy2 k(x; y)u(y)dy:

y: " jx?yj

For the computation of an integral over the square, we choose the tensor product trapezoidal rule. Setting N = n2 with n even, m = (m1; m2), and tN ;m = ( m ; m ) and m n n assuming x = (x1; x2) = tN = tN ;k , we write k k
1 2 1 2 1 2

x1 + 1 2 x1 ? 1 2

x2 + 1 2

x2 ? 1 2

h(y)dy1dy2 !l ;l :=
1 2

1 n2
8 > < > :

k+ k+ N N ) := 1 h(tl n2 l =k ? l =k ? h(tl ;l )!l ;l ; l 1 if jl1 ? k1j < n=2; jl2 ? k2j < n=2 1=4 if jl1 ? k1j = jl2 ? k2 j = n=2 (2.4) 1=2 else.
X

8 < :

2 X

2 X

9 = ;

n
2

n
2

1 2

1 2

Note that T 2 is the tensor product of the periodic interval 0,1] by itself. In this sense we T get tN n;l = tN;l = tN;l n . To set up a quadrature method for solving (2.1) numerically, l l l we consider (2.1) at x from the set of collocation points ftN ;k g and replace the integration k by the corresponding quadrature rule (2.4). Since the value k(x; x) is in nite, we have to modify the quadrature. We do this by dropping the term in the quadrature sum containing k(x; x). This way we arrive at the quadrature method
1 2 1 2 1 2 1 2

1 a(tN )uN (tN ) + n2 k k

l: l6=k

k(tN ; tN )uN (tN ) = g(tN ); k1; k2 = 0; : : : ; n ? 1: k l l k

(2.5)

Unfortunately, the method (2.5) is not convergent in the general case. Namely, if usual quadrature rules are applied to a singular integral, convergence cannot be expected. The remedy for this is the so-called singularity subtraction technique. Suppose we can compute (cf. (2.2))

b(tN ) k
Z

kS (tN ; tN k k T2 T

? y)dy T 2 T
=

Z k1
n k1 n

+1 Z 2

k2 n k2 n

+1 2

?1 2

?1 2

kS (tN ; tN ? y)dy1dy2 k k

(analytically or numerically with ner quadrature procedures). Then, we write

kS (tN ; tN k k T2 T

? y)u(y)dy T 2 T

b(tN )u(tN ) + 2 kS (tN ; tN k k k k T T

? y) u(y) ? u(tN )]dy T 2: T k

The last integral is weakly singular only and the usual quadratures converge for this weakly singular integral. Applying this step to (2.1), we arrive at the quadrature method 1 a(tN ) + b(tN )]uN (tN ) + n2 0kS (tN ; tN ? tN ) uN (tN ) ? uN (tN )] k k k k k l l k l: l6=k 1 + n2 0kR (tN ; tN )uN (tN ) = g(tN ); k1; k2 = 0; : : : ; n ? 1; (2.6) k l l k l: l6=k
X X

which is equivalent to
"

1 a(tN ) + b(tN ) ? n2 k k

1 + n2 0k(tN ; tN )uN (tN ) = g(tN ); k1; k2 = 0; : : : ; n ? 1: (2.7) k l l k l: l6=k

X

l: l6=k

kS (tN ; tN k k

? tN ) l

uN (tN ) k

E.g., if the kernel kS (x; x ? y) is odd with respect to the second variable z = x ? y (i.e. if it is a Mikhlin-Giraud kernel), then we get b(tN ) = 0 and k 1 0 N N N 2 n l: l6=k kS (tk ; tk ? tl ) = 0: Note that (2.8) is true also if instead of
X

(2.8) (2.9) (2.10) (2.11) (2.12)

kS (x; (z1; z2)) = ?kS (x; (?z1; ?z2))

one of the following symmetry properties is satis ed for the kernel:

kS (x; (z1; z2)) = ?kS (x; (?z1; z2)); kS (x; (z1; z2)) = ?kS (x; (z2; z1)):
Consequently,
"

1 b(tN ) ? n2 0kS (tN ; tN ? tN ) = 0 k k k l l: l6=k

X #

and the method (2.7) is equivalent to (2.5). Hence, the quadrature method (2.5) is useful if k(x; x ? y) is odd with respect to the second variable (cf. (2.9)) or if (2.10) or (2.11) is satis ed. In the quadrature methods (2.5) and (2.7), the unknown solution uN is a sequence of point values fuN (tN ;k ); k1; k2 = 0; : : : ; n ? 1g. We denote the matrix in the linear system k (2.5) and (2.7) by AN . However, we shall identify uN with a piecewise constant function and AN with an operator acting in the space of piecewise constant functions. To this reason, we introduce the characteristic function
1 2

N (x) = l1;l2 :

8 <

1 if lj =n xj < (lj + 1)=n; j = 1; 2 0 else

and denote the space of piecewise constant functions by S N , i.e.,

S N = spanf
1 2

N l1;l2

: l1; l2 = 0; : : : ; n ? 1g:

Then we identify fuN (tN;l ) : l1; l2 = 0; : : : ; n ? 1g with the piecewise constant interpolal tion n?1 uN (tN;l ) N;l uN = l l
X

and the matrix AN with the operator in L(S N ) whose matrix with respect to the basis f N;l : l1; l2 = 0; : : : ; n ? 1g is just AN . l We call the quadrature method stable if the operators AN are invertible for su ciently large N and if the inverse operators A?1 2 L(S N ) are uniformly bounded with respect to N N (i.e. the norms of A?1 2 L(S N ) induced by the L2 norm are uniformly bounded). The N quadrature method is called convergent if, for any right-hand side g such that
1 2

l1;l2 =0

1 2

1 2

n?1 X l1 ;l2=0

g(tN;l ) l
1 2 1 2

N ?g l1;l2

L2 (T 2 ) T

! 0;

there exist unique solutions fuN (tN;l )g of the quadrature equations (2.5) or (2.7) with l

uN =

n?1 X l1;l2 =0

uN (tN;l ) l
1 2

N l1;l2

tending in the L2-norm to the exact solution u. The Sections 3, 4, and 5 are devoted to the stability and convergence analysis of method (2.5). Method (2.7) can be treated with slight modi cations.

3 Localized Operators and Localized Quadrature Method on the Plane

Stability is a local property. Therefore it is necessary to introduce the quadrature scheme for the localized singular integral operator over the plane and to investigate the stability by analyzing the corresponding numerical symbol of the method. For singular kernels with a natural symmetry property, we shall prove the local stability.

3.1 The Operators and the Numerical Scheme over the Plane
In this subsection we introduce simple local problems over the plane which later will turn out to be the quadrature methods applied to the singular integral operators with frozen symbols. We consider the singular integral operator

Au(x) = au(x) + (Ku)(x); x 2 R2; k(x ? y)u(y)dy: (Ku)(x) =

Z

with a real constant a > 0 and the convolution kernel ( x?y k(x ? y) = fr2 ) ; r = jx ? yj; = jx ? yj : Moreover, we suppose f to be a Lipschitz function and

(3.1) (3.2)

f (z)d (z) = 0: (3.3) To de ne the quadrature method for singular integral equation (3.1), we rewrite (3.1) in the form
S1

Au(x) = au(x) +

k(x ? y) u(y) ? u(x) dy +

7

k(x ? y)dy u(x):

(3.4)

Since

)d = 0 (cf. (3.3)), we have k(x ? y)dy = 0 and get () Au(x) = au(x) + R jxf? yj2 u(y) ? u(x) dy : In order to evaluate the integral in Equation (3.5), we use the quadrature rule 1 h(t)dt h(tj ) n2 ; tj = ( j1 ; j2 ): n n R j 2Z Z
R

S1 f (

(3.5) (3.6)

Applying this to (3.5) and neglecting the term corresponding to j = k, we obtain

Au(tk)

au(tk) + au(tk) +

X
j Z2 Z j =k

f (tk ; tj ) 1 f (tk ; tj ) 1 u(tj ) n2 ? jtj ? tk j2 jtj ? tk j2 n2 u(tk ); 2 2 6 6 (tk ; tj ) = jtk ? tj j : (3.7) tk ? tj Now we shall show that the last sum vanishes under an additional assumption. To this end, we suppose that f can be split into f ( ) = f1( ) + f2( ) + f3( ), where (3.8) f1 (cos '; sin ') = ?f1 (? cos '; ? sin ') ; f2 (cos '; sin ') = ?f2 (? cos '; sin ') ; (3.9) (3.10) f3 (cos '; sin ') = ?f3 (sin '; cos ') : Similarly to (2.12), we obtain f (tk ; tj ) 1 (3.11) jtj ? tk j2 n2 = 0: 2
X h X i
j Z2 Z j =k j Z2 Z j =k

2 6

f (tk ; tj ) 1 2 u(tj ) ? u(tk ) n2 jtj ? tk j

h i

j Z2 Z j =k

Equation (3.7) takes the form

Au(tk) au(tk) +

X
j Z2 Z j =k

2 6

f (tk ; tj ) 1 2 u(tj ) n2 : jtj ? tk j

Hence, the quadrature method over the plane is de ned by

auN (tk) +

X
j Z2 Z j =k

Though this method (3.12) could be used as a numerical scheme for the plane equation, the application of (3.12) would require a further step of reduction to a nite linear system of equations. However, we are not interested in solving the plane equation. The method (3.12) serves us only as a tool in the stability analysis of the corresponding method over the torus. 8

2 6

f (tk ; tj ) 1 Z2 2 uN (tj ) n2 = g (tk ); k 2 Z : jtj ? tk j

(3.12)

3.2 Stability of the Quadrature Method over the Plane

The matrix of the system (3.12) is

k?j f jk ? j j f ( (tk ; tj )) 1 AN = (ak;j )k;j2Z ; ak;j = jtk ? tj j2 n2 = jj ? kj2 if j 6= k (3.13) Z a if j = k : Thus the entries of AN are independent of N = n2 and we get AN = A1. Moreover, the entries of A1 depend only on the di erence k ? j .
8 > > > > < > > > > :
2

m ?2 (3.14) ak;j = ak?j ; am = f jmj jmj if m 6= 0 a if m = 0 : We identify AN with the operator acting in the space of piecewise constant functions
(

S N (R2) = spanf N ;k : k1; k2 2 Z g Z k whose matrix with respect to the basis f N ; k 2 Z 2g is AN . Since Z k
1 2

1 2 (3.15) n k2Z j k j ; k2Z z Z the operator norm of AN induced by the L2 space is equivalent to the matrix norm of the space j k j2 < 1 : l2(Z 2) := = ( j )j2Z : Z Z
X
2

N k k L2 (R2 ) =

It is a well-known fact that each discrete convolution operator can be represented as (cf. e.g. 4]) AN = F ?1MF , where the unitary operators F : l2(Z 2) ! L2(T 2) and Z T F ?1 : L2(T 2) ! l2(Z 2) are de ned by T Z i2 j t ; F ?1 : f (t) 7! f g F : f j gj2Z 7! je j j 2Z ; Z Z
X
2

k2Z 2 Z

1 1 i2 s i2 s ?i2 s j ?i2 s j f (e 1 ; e 2 )e 1 1 e 2 2 ds1 ds2: 0 0 The operator M mapping L2(T 2) into L2(T 2), takes the form M f (t) = T T 2 ! R given by (cf. (3.14)) the continuous function : T T
j

j 2Z Z

:=

(t) f (t) with (3.16)

(t) =

Obviously, the inverse operator M ?1 mapping L2(T 2) to L2(T 2) is of the form M ?1 f (t) = T T ?1 (t) f (t).

k2Z Z

ak ei2 k t:

Proposition 3.1 1) There holds kA1k = kM kL(L (TT )) = ess sup j (t)j; kM ?1kL(L (TT )) = ess sup j ?1(t)j:
2 2

t2T 2 T

t2T 2 T

2) Operator A1 is invertible if and only if ess t2TT j (t)j > 0: inf

2

(3.17)

The function is called the symbol of the discrete convolution operator and the numerical symbol of the method (3.12). Now let us show that the sequence AN is uniformly bounded. In view of AN = A1 and of Proposition 3.1, we have to prove that : T 2 ! R de ned T by the formula (3.16) is bounded.

Proposition 3.2 For the function , we get sup j (t)j < 1. Proof. We shall utilize the Galerkin method with piecewise constant trial functions. Let Qn be the orthogonal projection onto the span of the system f j : j 2 Z 2g, where Z
j

:= 1. j

Qn :

L2(R2) ! spanf j : i 2 Z 2g; Qnf = Z

2

With respect to the basis f j gj2Z of im Qn the matrix Z (because kQnk = 1 and A is bounded). The matrix of f j gj2ZZ is de ned by AG = aG k;j2Z ; aG = A j ; k = A 0; k;j k;j n Z
2 2

(f; j ) j : j 2Z 2 Z of AG=QnAjimQn is bounded n AG with respect to the basis n

k?j

= aG?j : k

Then AG = aG k;j is a discrete convolution operator. Since AG is a bounded operator n k;j n 2 ), there exist a bounded mG : T 2 ! R, in l2(Z Z T mG(t) = aG ei2 k t k
X

such that AG = F ?1MG F , that MG is the operator of multiplication by mG, and that MG n is bounded (cf. Proposition 3.1). Now let ak?j k;j denote the matrix of the quadrature method and m = the corresponding symbol. We write m(t) = m(t) ? mG(t)] + mG (t) = (ak ? aG)ei2 k t + aG ei2 k t: k k
X X

k2Z 2 Z

In order to prove that m is bounded, it is su cient to prove that (m ? mG) is bounded. We prove this by showing jak ? aGj < 1: (3.18) k
X

k2Z 2 Z

k2Z 2 Z

Since ak = ak;0, we get

k2Z 2 Z
Z

aG = A 0; k
=
Z

k1 +1 Z k2 +1 k1 k2

k1 +1 Z k2 +1 Z 1 Z k2

k1
Z

0 0

aG ? ak = k

k1 +1 Z k2 +1 Z 1 Z k2

k1

A 0 (t)dt t?s 1 f jt ? sj ds2ds1 dt2dt1; 0 jt ? sj2 k t?s f jt ? sj f jkj 1 jt ? sj2 ? jkj2 ds2ds1dt2dt1: 0
2 6 6 6 4 3 7 7 7 5

(3.19)

10

For the integrand, we get k f jt ? sj f jkj 1 k 1 k t?s t?s ? f jkj jt ? sj2 ? jk1j2 + jt ? sj2 f jt ? sj ? f jkj : (3.20) jt ? sj2 jkj2 Estimating the rst term on the right-hand side, we easily conclude 1 f jkj jt ? sj2 ? jk1j2 C jk1j3 : (3.21) k To estimate the second term in (3.20), we observe that f is Lipschitz by assumption. Hence 1 f t?s ?f k 1 t?s k C jk1j3 : C jt ? sj2 jt ? sj ? jkj 2 jt ? sj jt ? sj jkj We arrive at jkj?3 < 1: jaG ? ak j < C k
X X

k2Z 2 Z

k Z2 Z k =(0;0)

function has limits along all rays starting at (0; 0).

1 Remark 3.1 It is not hard to see that is continuous on ? 2 ; 1 ]2nf(0; 0)g. At (0; 0) the 2

Next we turn to the stability of AN . Since AN = A1, we only have to prove the invertibility, i:e. (3.17). Unfortunately, we cannot prove stability for the general case or for the case of strongly elliptic singular integral equation either. Instead we prove stability for the special case of singular integral equation with Mikhlin-Giraud kernels and present a numerical stability proof for singular kernels with an operator for which the constant a is a complex number.

Theorem 3.1 Suppose the integral equation to which we apply (3.12) is given by (3.1) with constant a > 0 and a convolution kernel k(x; y)=f ( )r?2 such that f (? )=?f ( ).
Then the quadrature method (3.12) is stable.

Proof. We only have to show (3.17). Recall that (cf. (3.14) and (3.16))
(t) = a + #(t);
# (t) =
X
k Z2 Z k =(0;0)

ak ei2 k t; ak = f jk j jkj?2: k
X

Since f is an odd function, we get a?k =?ak as well as # (t) = a?k e?i2 k t = ? #(t): ak e?i2 k t = ?
X
k Z2 Z k =(0;0)

k Z2 Z k =(0;0)

Hence, #(t) is purely imaginary and

j (t)j = a2 +
11

#(t)

2 i ] a:

Finally, let us suppose there exist real constants ; with 2 + 2=1 and f ( ) = 21 f sin ' + cos 'g; = ei': The symbol of the corresponding singular operator is
A (x;

(3.22) (3.23)

) = a + if sin ' + cos 'g;

= ei':

In this case we get the numerical symbol (t) = a + if 1(t) + 2(t)g; where the numerical symbols 1 and 2 are real and correspond to the characteristics 1 1 2 2 i2 sin ' and i22 cos ', respectively. Numerical computations of 1 + 2 con rm (cf. Figure 2 1) that 1 + 2 1. Hence, ?1 p 1 + 2] 1 and we obtain: If f is given by (3.22) I I with real numbers , such that 2 + 2 = 1 and if a 2 C n fz 2 C : ?1 Im z 1g, then the quadrature method (3.12) is stable. Note that the condition a 2 C n fz 2 C : I I ?1 Im z 1g, is equivalent to the fact that A de ned by (3.23) is strongly elliptic at least after multiplication by a suitable constant.
Num.Symb.

0.5

0 0.5 -0.5 0 0.5 -0.5 0

2 2 Figure 1: The numerical symbol 1 + 2 .

12

4 Localization Principle
4.1 The Theorem
Let us start with a few historical remarks. Localization techniques (principle of freezing the coe cients) have been known and applied for a long time to the analysis of partial di erential operators or pseudo-di erential operators. Later on these techniques have been reformulated in an algebra language which has turned out to be useful in the analysis of several kind of operator classes (cf. Simonenko 34], Gohberg, Krupnik 11], and Douglas 10]). The rst one to apply these techniques to numerical methods was Kozak 14]. His ideas have been generalized and developed into a very nice abstract scheme by the school of Silbermann (for details cf. the corresponding chapters of 28]). Parallel to this, an abstract setting for the application to spline methods is due to Pro dorf 21]. We shall use the same localization techniques. However, instead of using the abstract schemes of e.g. Silbermann, we perform the corresponding steps of proof directly. This is possible because the local principle in our situation is not very complicated. To get a better feeling for the localization, we recommend the reader to study the corresponding sections of 11, 28]. Let us consider the quadrature method (2.5) applied to the singular integral equation (2.1) over the torus and suppose (2.9) is satis ed. To the corresponding singular integral operator and to this quadrature method, we introduce a localized singular integral operator and a localized quadrature method at any point 2 T 2. Thus let us x a 2 T 2. T T The localized operator is the singular integral operator over the tangent plane with the same values of the kernel function kS (x; x ? y) at x = . To get an operator over the plane, we freeze the local variable x and consider the convolution kernel kS ( ; x ? y). In other words the localized singular integral operator A at is the singular convolution operator over the plane R2 with the kernel function k (x ? y) = kS ( ; x ? y) ; and with the multiplication operator a(x) replaced by the constant a = a( ). Thus the localized equation corresponding to (2.1) is

a u(x) +

To this we apply the quadrature method (3.12). The resulting scheme is the localized quadrature method of (2.5). We denote the matrix (or the discretized operator of the quadrature method) by (A )N 2 L S N (R2) . With this notation the localization principle for the quadrature method can be formulated as follows:
gral equation (2.1) including the invertible operator A which is supposed to be a pseudodi erential operator of order zero and to posses a symbol from the class S 0 . Suppose the local operators A are de ned by the left-hand side of (4.1) and consider their quadrature approximation (A )N of the form (3.12). Then the method (2.5) is stable if and only if it is locally stable, i.e., if for any 2 T 2, the quadrature operators (A )N are stable. T

k (x ? y)u(y)dy = g(x):

(4.1)

Theorem 4.1 Let us consider the quadrature method (2.5) applied to the singular inte-

The stability of the quadrature methods (A )N has been investigated in Section 3. 13

4.2 Su ciency of Local Stability

In this subsection, we prove the su ciency of the local stability. We retain the notation S N for the space of piecewise constant functions (cf. Section 2) and denote the orthogonal projection onto S N by LN . For the stability of the sequence of operators AN it is su cient to prove a representation

AN BN = IN + DN + LN TCN ; (4.2) where IN 2 L(S N ) is the identity, kDN kL(S ) 1 , the operators CN ; BN 2 L(S N ) are 2 uniformly bounded with respect to N , and T 2 L(L2(T 2)) is compact. Indeed, from T (4.2), we get ?1 ?1 AN BN IN + DN = IN + LN TCN IN + DN : (4.3) and the stability of AN follows from the following lemma and the strong convergence AN LN ! A which will be proved in Section 5.
N

Lemma 4.1 (cf. e.g. 22]) Suppose A 2 L(L2 (T 2)) is invertible and AN LN ! A for T N ). Moreover, suppose EN , FN 2 L(S N ) are sequences of uniformly bounded AN 2 L(S operators and T 2 L(L2 (T 2)) is compact. Then T
AN EN = IN + LN TFN
implies that AN is stable. The same conclusion holds if there exist more than one term of the form LN TFN on the right-hand side.

Let us derive (4.2). To get BN , we introduce a nite set of points k 2 T 2, k = 1; : : : ; M . T 0 2 C 1 (T 2) in the neighborhood of k such that We choose cut o functions k ; k T i) The values of ii) There holds:
k
P

0 k; k

belong to 0; 1].
k

2 supp

0 ft 2 T 2 : k (t) 1g supp k ; T 0

0 k k

= k:

iii) Let f = M k . Then we suppose that f is a positive function with values less k=1 than 4. Moreover, we suppose that, for any t0 2 T 2, there exist at most four T 0 functions k not vanishing at t0. We introduce the piecewise constant interpolation projector by

KN h =

n?1 X l1 ;l2 =0

h(tN;l ) l
1 2

N : l1 ;l2

For a function g on T 2, we set gN := KN gjS . In other words, the matrix of gN with T respect to the basis f N g is l n?1 gN = g(tN ) i;j i;j=0; i 0 0 and we get ( k )N ( k )N = ( k k )N = ( k )N . Using all these de nitions, we choose the matrix operator BN for (4.2) as
N

14

BN =

k k where the operator BN is de ned as BN = (A )N and (A )N is the localized quadrature operator of Subsection 4.1 de ned for a xed 2 supp k . To explain the expression k 0 ( k )N (BN )?1 ( k )N , we note that, for xed k = ( k;1; k;2) 2 T 2, the torus T 2 can be T T identi ed with the periodic square 1 1 1 1 k;1 ? ; k;1 + ] k;2 ? ; k;2 + ] 2 2 2 2 0 and can be embedded into R2. The functions k ; k with 1 1 1 1 0 supp k ; supp k k;1 ? ; k;1 + k;2 ? ; k;2 + 2 2 2 2 can be considered as functions over R2. If KN stands for the interpolation projection onto S N (R2) (We use the same symbol as for the corresponding operator on T 2.), then we T can set hN = KN hjS (R ) for any function h over R2. In particular, we arrive at a second 0 de nition for ( k )N and ( k )N . These di erent operators, one over T 2 and the other over T 2, however, can be identi ed since for each piecewise constant basis function N over R l ;l T 2 with supp N;l \ supp k 6= ; there exists a unique basis function N;l0 over R2 with T 0 l l N 0 = N over ( k;1 ? 1 ; k;1 + 1 ) ( k;2 ? 1 ; k;2 + 1 ). Identifying these basis functions, 0 ;l l ;l l 2 2 2 2 k we can identify the two operators. In this sense the operator (BN )?1 over R2 multiplied 0 k 0 by ( k )N and ( k )N over R2 can be considered as an operator ( k )N (BN )?1 ( k )N over the torus. We conclude
N
2 1 2 1 2 1 2 1 2 1 2

k=1

k 0 ( k )N (BN )?1( k )N (f ?1)N ;

AN BN = AN
=
M

k=1

k 0 ( k )N (BN )?1 ( k )N (f ?1 )N
i

Xh

k=1 M Xh

k 0 0 0 AN ( k )N ? ( k )N AN ( k )N (BN )?1( k )N (f ?1)N k 0 0 0 k ( k )N AN ( k )N ? ( k )N BN ( k )N (BN )?1( k )N (f ?1)N

i h i

+ + =

k=1 M X k=1 M X

k 0 k k 0 ( k )N BN ( k )N ? ( k )N BN (BN )?1( k )N (f ?1)N

k=1 M h X

0 k k 0 ( k )N ( k )N BN (BN )?1( k )N (f ?1 )N
i

k=1 M X

0 0 k 0 ~ AN ( k )N ? ( k )N AN ( k )N (BN )?1( k )N (f ?1)N + TN

k 0 k k 0 ( k )N BN ( k )N ? ( k )N BN (BN )?1( k )N (f ?1)N + IN ;
h i

where ~ TN =

k=1

(4.4) (4.5)

Xh

k=1

k 0 k 0 0 ( k )N AN ( k )N ? ( k )N BN ( k )N (BN )?1( k )N (f ?1)N :

i

15

The representation (4.4) will imply (4.2) if we can show : a) The operator AN ( k )N ? ( k )N AN is the sum of an operator LN TCN with T compact and CN uniformly bounded plus an operator DN tending to zero in the operator norm. k k 0 b) The operator ( k )N BN ( k )N ? ( k )N BN is the sum of an operator LN TCN with T compact and CN uniformly bounded plus an operator DN tending to zero in the operator norm. 0 ~ c) The operator TN of (4.5) has a norm less than any prescribed > 0 if the k ; k ; k are chosen suitably.
h i h i

It remains to prove a), b), and c). We start with a). Let us consider the kernel ~ k(x; y) = k(x; y) k (x) ? k (y) ;
h i

(4.6) which is the weakly singular kernel of a compact integral operator T and which satis es It is not hard to see that ~ jk(x; y)j C jx ? yj?1: (4.7)

Consequently, it remains to prove that kTN ? LN T jS k ! 0: ~ ~ ~ We put k = k1 + k2, ~ ~ ~ ~ k1(x; y) = k(x; y) (jx ? yj); k2(x; y) = k(x; y) 1 ? (jx ? yj) ;
N

~ j k 1 AN ( k )N ? ( k )N AN = TN = (k(tN ; tN ) n2 )j;k :

(4.8)

where 2 C 1 is chosen such that supp (? ; ) and 1 on (? =2; =2) for a prescribed > 0. According to the splitting of the kernel, we get the splitting T = T 1 + T 2: Operator T 2 has a smooth kernel. For (4.8) it remains to prove that (T 2)N ? LN T 2jS ! 0; kLN T 1jS k C; C; (T 1)N
N N

(4.9) (4.10) (4.11)

where the constant C is independent of and . Let us prove (4.9). Since T 2 : L2 ! C is compact, since LN ; KN : C ! L2 are uniformly bounded, and since (KN ? LN ) tends to zero strongly, the operator (KN ? LN )T 2 tends to zero in operator norm. On the other hand, for the quadrature discretization 2 ~ j k 1 TN = k2(tN ; tN ) n2 j;k ; 16

we obtain

KN T 2jimL ? (T 2)N = b2 j;k For the di erence of the entries, we conclude

N

j;k

b2 = j;k
= Since we continue

Z h Z

~ j ~ j k 1 k2(tN ; y) N (y)dy ? k2(tN ; tN ) n2 k ~ j ~ j k k k2(tN ; y) ? k2(tN ; tN ) N (y)dy

i

N (y ) dy k

y2supp

sup

N k

~ j ~ j k k2(tN ; y) ? k2(tN ; tN ) : 1 C n;

~ j ~ j k k2(tN ; y) ? k2(tN ; tN )

C y ? tN k

b2

j;k

(b2 )j;k j;k

N

1 1 1 N (y ) dy Cn C n n2 ; k 1 1 1 1 C ( n n2 )j;k 2 k: jkj n n n
Z X

1 C n:

(4.12)

This implies KN T 2jS ? (T 2)N ! 0 for any xed > 0. And, together with (KN ? LN )T 2 ! 0, we obtain (4.9). Let us turn to (4.11) and estimate the entries b1 = k1(tN ; tN )=n2. j;k ~ j k 1 = 1 1 if jj ? kj C n 1 tN ? tN n2 jj ? kj n (4.13) C j k bj;k 0 otherwise ~ Here is the number used for supp (? ; ) in the splitting of k. By Young's inequality we conclude 11 C: C (b1 )j;k (4.14) j;k jj j n 6
8 > < > : X

jjj

j =0 C n

Hence, (4.11) is proved. Relation (4.10) follows analogously if instead of the entry of the discretized operator the kernel function of the integral operator T 1 is considered. The proof of (4.9), (4.10), and (4.11) nishes the proof of assertion a). Let us turn to the proof of b). This proof, however, is completely analogous to that of a). Indeed, instead of (4.6) we get 0 ~ k(x; y) = k (x)k(x; y) k(y) ? k (x)] (4.15) which satis es (cf. (4.7)) ~ jk(x; y)j
8 > < > :

C jx ? yj?1 if jyj ?1 0 if x 2 supp = C jx ? yj?2 else

17

0 k

0 for su ciently small > 0. Since the support of k is compact, the integral operator with kernel function (4.15) is compact. Using the function , we split T into T 1 and T 2, and, analogously to (4.12) we arrive at C n?3 if jkj ?1n jb2 j j;k C jkj?2 if jkj ?1n; C n?1 + C (b2 )j;k jkj?4 C n?1 + C : j;k
( s X

Thus we obtain k(T 2)N ? LN T 2jS (R )k and (4.14), we get 1 1 C jj ? kj n jb1 j j;k C jkj?2 0 C +C (b1 )j;k j;k
N
2

j;k: jkj ?1 n; j 2n

C for su ciently large n. Similarly to (4.13)

if jj ? kj C n; jkj
?1 n

8 > > > > < > > > > :

This means kLN T 1jS (R )k C and all these facts together prove that b) is valid. ? Now let us prove assertion c). We consider a vector = ( j )n=01 and arbitrary matrices j k FN . Then we get
N
2

j;k: jkj ?1 n; j 2n

if jkj ?1n; if jj ? kj C n; jkj?4 C + C :

k=1

2 0 k 0 k )N FN ( k )N

C C

k=1 M X k=1

0 k 0 ( k )N FN ( k )N 0 k 0 ( k )N FN ( k )N

2 2

( 0k )N
M 2X k=1

0 k 0 C sup ( k )N FN ( k )N
k=1;:::;M

( 0k )N

C sup (
k=1;:::;M

2 0 k 0 k )N FN ( k )N

(4.16)

0 Here 0k denotes the characteristic function of the support of k and satis es the relation 0 )N ( 0 )N = ( 0 0 )N = ( 0 )N . Moreover, the estimates corresponding to the second ( k k k k k and last line of (4.16) are correct since, for each j = (j1; j2) with 0 j1; j2 n ? 1, there 0 k 0 exist at most four vectors ( k )N FN ( k )N and at most four ( 0k )N such that the j-th 0 component does not vanish (cf. condition iii) for the de nition of the k ). Hence,
X

and, choosing we arrive at ~ TN

k=1
h

k 0 0 ( k ) N FN ( k ) N

k 0 0 C sup ( k )N FN ( k )N k=1;:::;M

k k k FN = ( 0k)N AN ( k )N ? ( 0k )N BN ( k )N (BN )?1(f ?1)N ;

i

C sup

0 0 k C sup ( k )N AN ( k )N ? ( k )N BN ( k )N :
k=1;:::;M

k=1;:::;M

k 0 k 0 0 ( k )N AN ( k )N ? ( k )N BN ( k )N (BN )?1( k )N (f ?1)N

i

18

0 0 k It remains to prove that ( k )N AN ( k )N ? ( k )N BN ( k )N is small provided that the 0 supports of k and k have a small diameter. First we consider the case that A is a multiplication operator. We get 0 0 k 0 0 ( k )N AN ( k )N ? ( k )N BN ( k )N = ( k )N aN ( k )N ? ( k )N (a( ))N ( k )N = k (tN ;j ) a(tN;j ) ? a( )] i;j i;j ; j j 0 0 k ( k )N AN ( k )N ? ( k )N BN ( k )N C sup ja(t) ? a( )j:
h i h i
1 2

t2supp
h

0 0 k Since is taken from supp k too, we obtain that ( k )N AN ( k )N ? ( k )N BN ( k )N is small for k with su ciently small support supp k . Now, in the second case, suppose that operator A is an integral operator with bounded k kernel function kR. For this A, the localized operator A is zero. Thus BN = 0 and we 0 )N AN ( k )N is small provided the functions k ; 0 have supports have to prove that ( k k with su ciently small diameter. However, due to the quadrature weight n?2, each entry of 0 0 ( k )N AN ( k )N is less than Cn?2. The dimension of the non-zero part of ( k )N AN ( k )N 2 if the diameter of the supports supp k and supp 0 is less than . is less than n] k Consequently, Young's inequality implies 0 Cn?2 C 2 ( k ) N AN ( k ) N
X

l2Z 2 : jlj n Z

k 0 0 and ( k )N AN ( k )N ? ( k )N BN ( k )N is small for a small diameter of supp k and 0 supp k . In the third and last case we suppose that A is the singular integral operator with kernel kS . Moreover, we may assume that ?y kS (x; x ? y) = b(x)f ( jx ? yj )jx ? yj?2: (4.17) x Indeed, the characteristic f (x; x ? y) = jx ? yj2kS (x; x ? y) is a smooth function for a pseudo-di erential operator with a symbol from the class S 0. We can approximate f in the Lipschitz norm by the truncated trigonometric series with respect to the second variable z = x ? y. The singular integral operator and its quadrature discretization corresponding to the approximated characteristic are close to the original singular operator and its quadrature discretization (cf. 5, 6] and Lemma 5.1 for the discretized operators). Hence, we can replace A by the operator corresponding to the truncated trigonometric series of its characteristic and can treat each term of the sum separately. This way we arrive at kernels of the form (4.17). However, operators with kernel (4.17) are products of a multiplication operator (multiplication by b) and a convolution operator G with kernel x?y (4.18) f ( jx ? yj )jx ? yj?2: Similarly, AN is the product of the diagonal matrix bN (discretized multiplication operak tor) and the discretized convolution operator GN , and BN the product of b( )IN and Gk N 2). We conclude (discretized convolution operator over R 0 0 k ( k )N AN ( k )N ? ( k )N BN ( k )N

19

0 0 = ( k )N bN GN ( k )N ? ( k )N b( )IN Gk ( k )N N 0 ) b ( 0 ) ? ( 0 ) b( )I ( 0 ) G ( ) = ( kNN kN N k N N k N k N 0 +( k )N b( )IN ( 0k )N GN ( k )N ? ( 0k )N GK ( k )N : N

h i h i

(4.19)

The last bracket is zero since the kernel of the frozen operator with kernel (4.18) is the same as (4.18). The rst bracket on the right-hand side of (4.19) is small by the proof for the case when A is a multiplication operator. This completes the proof of assertion c).

4.3 Necessity of Local Stability

Suppose fAN g is stable and x a 2 T 2. We have to show that (A )1, i.e., the quadrature T operator (A )N for N = 1 is invertible. We shall show that fAN g can be considered as a stable and convergent approximation method for operator (A )1 which implies that (A )1 is invertible. In order to simplify the notation we suppose = (0; 0). In the previous subsections we have identi ed the operator (A )N 2 L S N (R2) with its matrix. Now we consider (A )N = (A )1 to be the xed matrix operator acting in l2(Z 2). Z N ) with its matrix, we introduce the isomorphism of For the identi cation of AN 2 L(S S N and the nite l2-space explicitly. We consider the set and introduce EN : l2(Z N ) ! S N by Z2 1 Z Z N = fl 2 Z 2 : ? 2 Z2
2

lj < 1 ; j = 1; 2g Z 2 Z n 2
X

EN ( l)l2Z = Z
N

l2Z N Z2

N l l :

Clearly, EN is invertible. To each operator BN 2 L(S N ) there corresponds the matrix ? ~ ~ operator B N := EN1BN EN , i.e., B N is the matrix of BN with respect to the basis f N : l 2 g. Moreover l2ZN Z ~ BN L(S ) = BN L(l (Z )): Z
N

Z Z2 Now l2(Z N ) can be embedded into l2(Z 2) by identifying l2(Z N ) with Z2 f( l)l2ZZ 2 l2(Z 2) : l = 0 for l 2 Z 2nZ N g: Z Z Z2 We denote the orthogonal projection from l2(Z 2) to l2(Z N ) by PN . Clearly, PN tends Z Z2 2 (Z 2). Thus we can consider the operator AN 2 ~ strongly to the identity operator in l Z L(imPN ) corresponding to our quadrature operator AN as an approximate operator for (A )1 2 L(l2(Z 2)). We shall prove that Z ~ ~ (4.20) AN PN ! (A )1; AN PN ! (A )1 is true in strong operator topology. If this is done, then we conclude from the stability ~N kA?1k C (which means also kA?1k C ) that N ~ lim C ?1 PN C ?1k k; (4.21) (A )1 = lim AN PN
2

(A )1

N !1 C ?1k

N !1

(4.22)

20

holds for any 2 l2. Relation (4.21) implies that (A )1 has a trivial null space and that the image space of (A )1 is closed. The inequality (4.22) proves that the kernel of (A )1 is trivial, i.e., the cokernel of (A )1 is trivial, too. Hence, (A )1 is invertible. It remains to show (4.20). To prove the strong convergence we use the Banach-Steinhaus theorem. The uniform ~ boundedness of the operators AN (and hence also of the AN ) will be proved in Lemma 5.1. Thus it remains to prove that, for any xed em = ( j;m)j2Z , Z ~ ~ AN PN em ! (A )1em ; AN PN em ! (A )1em: (4.23)
2

~ ~ Moreover, the adjoint matrices AN , (A )1 are of the same structure as AN , (A )1 since they correspond to the adjoint singular integral operators. In other words, we only prove the rst part of (4.23). We observe that, for any cut o function which is equal to one in a small neighborhood of = 0, there holds ~N em = (tN ) i;j em = (tN )em = em m j for su ciently large N . We introduce a cut o function
0

such that

supp

ft 2 T 2 : T

1g

and write (Recall that the matrix (A )N is independent of N.) 0 ~ ~ AN PN em = (A )1em + ~N ? IN (A )N ~N em 0 0 ~ 0 ~ ~ + ~N AN ~N ? ~N (A )N ~N ~N em + IN ? ~N AN ~N em:
h i

The third term on the right-hand side is small if and 0 are suitably chosen. Indeed, the corresponding operators without the tilde have been shown to be small in the proof to assertion c) in Subsection 4.2. The smallness of the second and of the last term follows from the next lemma. In other words, for any > 0, we can choose appropriate and 0 such that ~ AN PN em ? (A )1em l < ~ for N su ciently large. Thus AN PN ! (A )1 and the necessity is proved.
2

Lemma 4.2 Suppose that supp fx 2 Z 2 : jxj < 1g Z

(
0 N

fx 2 Z 2 : jxj < 2g Z
C 1;
2

fx 2 Z 2 : 0(x) = 1g; Z
N

where 0 < 1 < 2 and 2 is much larger than 1. Then we get

? IN )(A )N

0 N

? IN )AN

C 1:
2

(4.24)

0 Proof. Let us consider the matrices of ( k )N , ( k )N , (A )N with respect to the basis f N : l 2 Z 2g. We get Z l 0 (A )N = bi;j i;j2Z ; ( k )N = ci ij i;j2Z ; IN ? ( k )N = di ij i;j2Z ; Z Z Z
2 2 2

21

where obviously

bij

jcij jdij

?2 1 C t N ? t N n2 i j 1 if tN 2 supp i 0 if tN 2 supp i = 0 if ji=nj 2 1 else .

( (

k k

C i?j ; 1 if ji=nj 0 else ,

(

?2

Consequently, the norm in the rst part of (4.24), i.e., the l2 matrix norm of the corresponding matrix with respect to the basis f N : l 2 Z 2g is less than the norm of the Z l matrix EN FN , where

FN = (fi ij ) 2 Z 2; fi = 1 if i , 1n Z 0 else ?2 EN = (ei;j )i;j2Z ; ei;j = ei?j = C ji ? j j if ji:? j j ( 2 ? 1)n Z 0 else

(
2

Applying EN to a vector = ( l)l2Z , we get from Young's inequality Z

2

EN

l2

jij ( 2? 1 )n

C 2jij?4 k kl

C ( 2 ? 1)n

?1

k kl :
1

Now we use the Cauchy-Schwarz inequality to get

EN FN

l2

C ( 2 ? 1)n
h

?1

FN
X
1

l1 n

C ( 2 ? 1)n
1
s X
1

?1

C ( 2 ? 1)n
Consequently,

?1 s

jij

jij

j i j2 :

jij

j ij

1 C ? k kl ; EN FN 2 1 The second estimate of (4.24) follows analogously.

E N FN

l2

1 C ? 2 1

C 1:
2

5 The Convergence of the Quadrature Method

This section is devoted to the convergence of the quadrature method. We shall show that the discretized operator AN is uniformly bounded with respect to N . Using a BanachSteinhaus argument, we shall prove the strong convergence of the discretized operator AN LN to the singular integral operator A. This together with the stability implies the convergence of the quadrature method.

22

stable and that the discretized operator AN LN corresponding to (2.5) converges strongly to the operator on the left-hand side of (2.1). Then the method (2.5) is convergent, i.e. for any right-hand side g such that

Theorem 5.1 (cf. e.g. 28]) Suppose the quadrature method (2.5) applied to (2.1) is
n?1 X j1 ;j2 =0

g(tN ;j ) j
1 2

N ? gk L2 j1 ;j2

! 0;

the equation (2.5) has a unique solution uN if N is su ciently large, and uN tends in L2 to the exact solution u of (2.1).

Now let us turn to the boundedness of the discretized operator AN de ned in Section 3.3.

Lemma 5.1 There exists a constant C independent of N and of the operator A de ned
kAN k C kAkL(L (TT )) + kakL1(TT ) + kf kLip + kkRkL1(TT TT ) : Here the Lipschitz norm kf kLip of the characteristic of kernel kS is de ned by 0 )?f 0 kf kLip = kf kL1 + sup jf (x; jx ? x0(jx ; )j + sup jf (x; j ) ? f0(jx; )j : ? x; x 2 T 2 T ; 2 S1
2 2 2 2 2
0 0

on the left-hand side of (2.1) such that the L2 -operator norm of AN (or equivalently the l2-matrix norm of AN ) is bounded as
n o

x0 6= x

S1

6 x2T 2 T
0

Proof. Let us consider the Galerkin method where the trial space is spanned by the orthonormal basis fn N : k1; k2 = 0; : : :; n ? 1g. For the entries aG of the Galerkin k j;k
matrix AG we get N

aG j;k

= hA n

N ]; k

N ]i = 2 j=n a(x)dx j;k n j (j ?1)=n

Z

Z k=n Z 2 j=n k(x; x ? y)dxdy: n (j ?1)=n (k?1)=n

We denote the corresponding entries of the matrix AN for the quadrature method by aj;k . Since kAG k = kLN AjimL k kAk; N we only have to show
N

(aj;k ? aG )j;k L(l ) C kf kLip + kakL1 + kkRkL1 (5.1) j;k Moreover, since the boundedness proofs for the multiplication operator and for the integral operator with bounded kernel function kR are straight forward, we suppose a 0 and kR 0. We shall estimate (aj;k ? aG )j;k in two steps. First we shall derive a bound j;k for the matrix with all entries corresponding to the indices i,j such that ji ? j j > 2 (\o diagonal" entries) and later we consider the matrix with the entries such that ji ? j j 2 (\almost diagonal" entries). Let us estimate the \o diagonal" entries.
2

23

j j k 1 k(x; x ? y)dxdy aj;k ? aG = k n ; n ? n n2 ? n2 j;k j j k = n2 k n ; n ? n ? k(x; x ? y) dxdy : j k If we put x = n + , y = n + , = ( 1; 2), and = ( 1; 2), then
Z j +1 Z k +1
n n j n k n

Z j +1 Z k +1
n n j n k n

aj;k ? aG = j;k

Z 1 Z 1 Z 1 Z 1
n n n

Putting l = j ? k, and = ? we get

j j k n2 k n ; n ? n ? 0 0 0 0 j j k k n + ; n + ? (n + ) d d :
n

aj;k ? aG j;k

Z 1 Z 1 Z
n n

2 2 1 ?n

Z
1

1 1 ?n

n2

"

2 1 1 ? C jnj3 : (5.4) jl=nj2 jl=n + j2 l Since f satis es a Lipschitz condition with respect to the second variable, we nd for the second term on the right-hand side of (5.3)

The function f is bounded and

For T1 we get l k l+k + + ;n ?k n f l+k + n l+k + f n 1 + jl=nj2

n2 T1 + T2 d d ; ? 0 0 ? l l T1 = k l + k + + ; n ? k l + k + + ; n + n n l + k; l ? k l + k + + ; l : T2 = k n n n n
n n
2 1

Z 1 Z 1 Z

l k l + k; n n l ?k l + k + + ; n + n
i

dd
(5.2)

l+k + + ; l + n n l 1 ?f l+k + + ; l + 1 + ; n jl=nj2 n n jl=n + j2 l 1 1 ? + ;n + jl=nj2 jl=n + j2 l l f l + k + + ; n ? f l + k + + ; n + : (5.3) n n

l l f l+k + + ; n ?f l+k + + ; n + n n
24

l=n l=n + C jl=nj ? jl=n + j

C : (5.5) jlj

Substitution of (5.4) and (5.5) into (5.3) provides us with jT1j Cn2jlj?3. For T2, we arrive at 1 l l l l k l + k ; n ? k l + k + + ; n = jl=nj2 f l + k ; n ? f l + k + + ; n n n n n 2 1 1 C n jl=nj2 C jnj3 : (5.6) l Substituting the estimates for T1 and T2 into (5.2), we obtain n4 d d C aj;k ? aG C jl1j3 : j;k 3 ? jlj ? 0 0 Young's inequality implies for the \o diagonal" part of AN ? AG that N 1 (aj;k ? aG )j;k L(l (Z )) C j;k 3 C: Z l2Z f1 + jljg Z
Z 1 Z 1 Z
n n
1

On the other hand, let us turn to the \almost diagonal" entries. For the Galerkin matrix AG we conclude N jaG j (aG )j;k L(l (ZZ )) = kAG k = kLnALn k C kAk C: j;k j;k N For the \almost diagonal" entries of the quadrature method, we get 12 kS j ; j ? k if jj ? kj > 0 aj;k = n n n n 0 if j = k: If l = j ? k 6= 0, then we obtain j l 1 f ( j ; l ) l ?2 ; ja j f(n; n) C: aj;k = n2 n n n j;k Hence, each \almost diagonal" entry aj;k ? aG ] is bounded. Consequently, the \almost j;k diagonal" part of AN ? AG is bounded, too. N
2 2

8 > > < > > :

di erential operator of order zero with a symbol from S 0 . Moreover, let AN stand for the discretized quadrature operator of (2.5). We suppose that (2.9) is satis ed. Then AN LN u tends to u in the L2 -norm for any u 2 L2(T 2). T

Lemma 5.2 Suppose that the operator A given by the left-hand side of (2.1) is a pseudo-

Proof. In Lemma 5.1 we have shown that AN is uniformly bounded. Hence, in view
of the Banach-Steinhaus theorem, we may suppose that f is smooth and have to prove kAN LN f ? Af k ! 0 for any smooth f . Since f is smooth, KN Af tends to Af if KN is the piecewise linear interpolation projector. It remains to prove kAN LN f ? KN Af k ! 0. Moreover, since KN f ! f and since AN is bounded, we conclude AN LN f ? AN KN f ! 0. It remains to prove kAN KN f ? KN Af k ! 0. This, however, is a consequence of kAN KN f ? KN Af kL1 ! 0 which is equivalent to (5.7) sup AN KN f (tN ) ? Af (tN ) ! 0: i i
i

Now we study the di erence Af (tN ) ? AN KN f (tN ) in three cases : i i 25

1) A is a multiplication operator 2) A is an integral operator with a continuous and bounded kernel kR 3) A is the singular integral operator with kernel kS .
Case 1) is very simple since Af (tN ) = AN KN f (tN ) holds for multiplication operators. The i i assertion for Case 2) is well known, too. Indeed, the quadrature rule used for AN KN f (tN ) i has non-negative quadrature weights. Hence, it converges on continuous functions and even uniformly over the compact set of functions y 7! kR(x; y)f (y). It remains to consider Case 3). The di erence Af (tN ) ? AN KN f (tN ) takes the form (cf. Section 2) i i

Af (tN ) ? AN KN f (tN ) i i

?
T1 = T2 =
Z

T T

kS (tN ; tN ? y) f (y) ? f (tN )]dy T 2 T i i i

l: l6=i

1 kS (tN ; tN ? tN ) f (tN ) ? f (tN )] n2 = T 1 + T 2; i i l l i

(5.8)

?
Z

T 2nB (tN ; ) T i
X

l: jtN ?tN i l
N i

B (t ; )

kS (tN ; tN ? y) f (y) ? f (tN )]dy T 2 T i i i

kS (tN ; tN ? y) f (y) ? f (tN )]dy T 2 T i i i 1 kS (tN ; tN ? tN ) f (tN ) ? f (tN )] n2 ; i i l l i j>

1 kS (tN ; tN ? tN ) f (tN ) ? f (tN )] n2 ; i i l l i

(5.9)

(5.10)

l: l6=i; jtN ?tN j i l

where the number stands for a xed positive real, and B (tN ; ) T 2 is the ball with T i center tN and radius . In a minute we will prove that T 2 ! 0 for ! 0. On the other i hand, the integral in T 1 is regular for xed > 0. Thus the same arguments as for Case 2) imply T 1 ! 0 for N ! 1. We conclude T 1; T 2 ! 0 for N ! 0, and, using (5.8), we get (5.7). It remains to show T 2 ! 0 for ! 0. We estimate the two terms in (5.9) separately. For the integral, we get
Z

B (tN ; ) i

T kS (tN ; tN ? y) f (y) ? f (tN )]dy T 2 i i i

B (tN ; ) i

C jtN ? yj?1dy T 2 T i
1 n l: l6=i; jl?ij
X

C:

The quadrature sum can be estimated as

X

l: l6=i; jt ?t j
N i N l

1 kS (tN ; tN ? tN ) f (tN ) ? f (tN )] n2 i i l l i

ji ? lj?1

C:

Hence jT 2j C and T 2 ! 0 for ! 0 is proved.

6 Numerical Tests
In order to check the convergence properties of our quadrature method, we consider the following oblique derivative problem. We de ne the two-dimensional surface S by the 26

parametrization

S = f (s; t); 0 s; t 1g; (6.1) (s; t) = 2 + cos(2 s)] cos(2 t) ; sin(2 s) ; 2 + cos(2 s)] sin(2 t) :
Clearly, S is homeomorphic to the torus. The space R2nS is the union of the bounded ring shaped domain ? and the unbounded exterior domain . For this domain , we solve the oblique derivative boundary value problem (cf. 18]) 4V = 0 in ; (6.2) @ V = g on S = @ ; f : S = @ ! R3: (6.3) @f The oblique direction vector f (P ) is de ned as where n(P ) is the normal vector of unit length at P 2 S pointing into ? . We represent the unknown potential V in the form of a Newton potential x(Q) d S; V x(P ) = 41 (6.5) Q S jP ? Qj where x(Q) denotes an unknown single layer surface density. We apply the boundary operator of oblique derivative, and, with the well-known jump relations for the Newton potential, we obtain the boundary integral equation g(P ) = @f@P ) (V x)(P ) ( f (P ) (Q ? P ) x(Q) d S : (6.6) = ? 1 f (P ); n(P ) x(P ) ? 41 Q 2 jP ? Qj3
Z D E Z

f (P ) = n(P ) + 1 (0; 0; 1); 2

(6.4)

This is a strongly singular integral equation of the second kind for the unknown function x(Q). Using the parametrization , we transform (6.6) into (1.4), where the kernel takes the form k(t; s) = f ( (jt))s)(?(s)t?j3 (t)) j 0(s)j; (6.7) ( () and where j 0(s)j = j@s (s) @s (s)j is the density of the surface measure. Note that operator A is strongly elliptic since hf; ni > 0. Moreover, the singular part kS of the kernel is a Mikhlin-Giraud kernel, i:e., it satis es (2.9). This equation (1.4) is solved numerically by the quadrature method (2.5). Before we solve the linear equations, we check whether the quadrature approximation of the singular integral operator converges. For this purpose, we consider the singular integral ;P ? v(P ) = ? 1 f (P ); n(P ) w(P ) + 41 S hf (jP )? Qj3 Qi w(Q)dQS; 2 P w ( (t)) = sin(2 t1) sin(2 t2); t = (t1; t2) 2 0; 1]2
1 2

27

together with its approximation vN given at the grid points tj = tj ;j by

1 2

vN ( (tj )) :=

n?1 X k1 ;k2 =0

aj;k w ( (tk)) ;

where AN = (aj;k )j;k is the matrix of the quadrature method. For several n = nl = 2l and Nl = n2, we compute the L2-Norm error l

kvN ? vN
l

l+1

1 k := n l

v u u u t

nl ?1 X

j1 ;j2 =0

jvN ( (tj )) ? vN ( (tj ))j2

l l+1

and the approximate convergence order log := log kvN ? vN k ? 2 kvN ? ? vN k : log The results are presented in Table 1. It turns out that the approximate operator AN converges with order 1. nl Degrees of Freedom: Nl kvN ? vN k N 4 16 5:35 10?2 8 64 2:00 10?2 1.42 16 256 8:48 10?3 1.24 ?3 32 1024 4:16 10 1.03 64 4096 2:08 10?3 1.00 ?3 128 16384 1:04 10 1.00 Table 1: Approximation order of the quadratures
Nl
l l+1 l 1 l l l+1 l

The discretized operators are stable by the Theorems 3.1 and 4.1. Stability means that the matrices AN together with their inverses A?1 are uniformly bounded with respect N to N . Though we have not computed the Euclidean matrix norms of AN and A?1, we N have an indicator for the uniform boundedness. Normally, for bounded norms kAN k and kA?1k, the iterative solution of the matrix equation requires a number of iteration steps N which is bounded independently of N . In Table 2 we present the number of GMRES iterations (cf. 30]) necessary to achieve an error less than 10?12. Indeed, these numbers seem to grow very slowly. nl Nl Number of GMRES iterations 2 4 4 4 16 12 8 64 22 16 256 25 32 1024 28 64 4096 32 Table 2: Numbers of GMRES iterations Next we compute an approximate solution from solving (2.5). After determining the solution uN of the quadrature method at the grid points tj ;j , j1; j2 = 0; : : : ; n ? 1, we
1 2

28

compute an approximate solution UN for the Laplace equation by discretizing the single layer representation (6.5). 1 n?1 uN ( (tj ;j )) j 0(t )j: U (x) UN (x) := 41 n2 (6.8) j ;j j ;j =0 j (tj ;j ) ? xj In our rst example, we take a known solution of (6.2), (6.3) given by U (P ) = jP ? (2; 0; 0)j?1: (6.9) The oblique derivative is given by
X
2 1 2 1 2 1 1 2

) ; ; @ (6.10) g(P ) = @f U (P ) = f (Pj(2; ((20)0?0) ? P ) : 0; P j3 For this right-hand side g, we have solved the quadrature equations (2.5) and computed the L2 errors 1 n ?1 ju ( (t )) ? u ( (t )) j2 kuN ? uN k := n N j N j l j ;j =0 and the approximate convergence orders log kuN ? uN k ? log kuN ? ? uN k : N := log 2 Moreover, we have computed the approximate values UN (P ) for P = (1; 0; 0) and P = (0:3; 0:2; 0:1), the relative errors jUN (P ) ? U (P )j=jU (P )j with U (P ) from (6.9), and the approximate convergence orders log jUN (P ) ? U (P )j ? log jUN ? (P ) ? U (P )j : N := log 2 The numerical results are presented in the Table 3. They show that our quadrature solutions converge to the exact solutions. The convergence orders are close to one. nl Nl kuN ? uN k N jUN (jP )(?)U (P )j N jUN (jP )(?)U (P )j N UPj UPj P = (1; 0; 0) P = (0:3; 0:2; 0:1) 2 4 1.49 1.95 4 16 0.87 0.0032 8.88 0.79 1.32 8 64 0.13 2.69 0.22 -6.12 0.13 2.62 16 256 0.04 1.76 0.16 0.49 0.028 2.24 32 1024 0.019 1.09 0.08 0.98 0.013 1.12 64 4096 0.01 0.81 0.04 1.00 0.0063 1.00
l X
2

l+1

v u u u t

l+1

l+1

l 1

l 1

l+1

Table 3: Convergence of the quadrature method for g(Q) =

@ ?1 @f (Q) jQ ? (2; 0; 0)j

In a second example we consider an oblique derivative g for which the exact solution is unknown. Since our quadrature method is a low order method, we choose g with a low degree of smoothness. In particular, we have taken 1 if s 1 g1 ( (s; t)) = 0 else < 2 ,
(

29

Note that g1 2 H (s) for < 1. Instead of the error jUN (P ) ? U (P )j=jU (P )j we now compute the error jUN (P ) ? UN (P )j and the corresponding convergence rates
l l l+1

j ? log = log jUN (P ) ? UN (P )log 2 jUN ? (P ) ? UN (P )j : The numerical results are presented in Table 4. They show that our quadrature method converges with order one even for solutions with low degree of smoothness.
Nl
l l+1 l 1 l

nl Nl jUN (P ) ? UN (P )j N 2 4 4 16 0.53 8 64 0.59 -0.14 16 256 0.31 0.93 32 1024 0.13 1.35 64 4096 0.049 1.30 Table 4: Convergence of the quadrature method for g1 and = 0.5 at P =(1,0,0)
l l+1 l

References
1] D. N. Arnold and W. L. Wendland, On the asymptotic convergence of collocation methods, Mathematics of Computation 41, No. 164 (1983), 349-381. 2] D. N. Arnold and W. L. Wendland, The convergence of spline collocation of strongly elliptic equations on curves, Numer. Math. 47 (1985), 317-341. 3] S. M. Belotserkovski and I. K. Lifanov, Numerical methods for singular integral equations, Nauka, Moscow 1985 (Russian). 4] A. Bottcher and B. Silbermann, Analysis of Toeplitz operators, Akademie-Verlag, Berlin (1989). 5] A. P. Calderon and A. Zygmund, On singular integrals, Am. J. Math. 78 (2), (1956), 289-309. 6] A. P. Calderon and A. Zygmund, On singular integrals with variable kernels, Appl. Anal. 7 (1978), 221-238. 7] G. Chen and J. Zhou, Boundary element methods, Computational Mathematics and Applications (eds. : J. R. Whiteman and J. H. Davenport), London, New York, Boston, Tokyo (1992). 8] M. Costabel and W. McLean, Spline collocation for strongly elliptic equations on the torus, Numer. Math. 62 (1992), 511-538. 9] W. Dahmen, S. Pro dorf and R. Schneider, Wavelet approximation methods for pseudo-di erential equations I: Stability and convergence, Math. Zeitschr. 215 (1994), 583-620. 30

10] R. G. Douglas, Banach algebra techniques in operator theory, Academic press, New York, (1972). 11] I. Gohberg and N. Krupnik, One-dimensional linear singular integral equations, Vols. I and II, Operator Theory : Advances and Applications, Vols. 53 and 54, Birkhauser, Verlag, Basel, (1992), originally in Russian, Shtiintsa, Kishinev (1973). 12] R. Hagen, S. Roch and B. Silbermann, Stability of spline approximation methods for multidimensional pseudodi erential operators, Integral Equations and Operator Theory 19 (1994). 13] G. C. Hsiao and S. Pro dorf, A generalization of the Arnold-Wendland lemma to collocation methods for boundary integral equations in Rn , Math. Nachr. 163 (1993), 133-144. 14] A. B. Kozak, A local principle in the theory of projection methods, DAN USSR 112, No. 6 (1973), 157-160 (Russian). 15] V. G. Mazya, Boundary integral equations, in: Encyclopaedia of mathematical sciences, Vol. 27, Analsis IV (eds. : V. G. Mazya and S. M. Nikolskii), Springer-Verlag, Berlin, Heidelberg, New York, London, Paris, Tokyo, Hong Kong (1991). 16] S. Mikhlin and S. Pro dorf, Singular integral operators, Akademie-Verlag, Berlin (1986). 17] S. G. Mikhlin, N. F. Morozov and M. V. Paukshto, The integral equation of the theory of elasticity, B. G. Teubner-Verlagsges., Stuttgart, Leipzig (1995). 18] C. Miranda, Partial di erential equations of elliptic type, Springer-Verlag, Berlin, Heidelberg, (1970). 19] N. I. Muskhelishvili, Singular integral equations, Noordho , Groningen (1953). 20] V. Z. Parton and P. I. Perlin, Integral equations in elasticity theory, Nauka, Moscow (1977) (Russian). 21] S. Pro dorf, Ein Lokalisierungsprinzip in der Theorie der Spline Approximationen und einige Anwendungen, Math. Nachr. 119 (1984), 239-255. 22] S. Pro dorf and A. Rathsfeld, A Spline collocation method for singular integral equations with piecewise continuous coe cients, Integral Equations and Operator Theory 7 (1984), 536-560. 23] S. Pro dorf and A. Rathsfeld, A quadrature and collocation methods for singular Integral Equations on curves with corners, Zeitschrift fur Analysis und ihre Anwendungen 8, No. 3, (1989), 197-220. 24] S. Pro dorf and G. Schmidt, A nite element collocation method for singular integral equations, Math. Nachr. 100 (1981), 33-60. 25] S. Pro dorf and G. Schmidt, A nite element collocation method for systems of singular integral equations, Preprint P-Math. 26/81, Inst. f. Math., Akad. d. Wiss. d. DDR, Berlin (1981). 31

26] S. Pro dorf and R. Schneider, A spline collocation method for multidimensional strongly elliptic pseudodi erential operators of order zero, Integral Equations and Operator Theory 14 (1991), 399-435. 27] S. Pro dorf and R. Schneider, Spline approximation method for multidimensional periodic pseudodi erential equations, Integral Equations and Operator Theory 15 (1992), 626-672. 28] S. Pro dorf and B. Silbermann, Numerical analysis for integral and related operator equations, Birkhauser, Basel (1991). 29] S. Pro dorf and I. H. Sloan, A quadrature method for singular integral equations on a closed curves, SIAM J. Numer. Anal. 61 (1992), 543-559. 30] Y. Saad and M. M. Schultz, A generalized minimal residual algorithm for solving nonsymmetric linear systems, SIAM J. Sci. Statist. Comput. 7,(1986), 856- 869. 31] J. Saranen and W. L. Wendland, On the asymptotic convergence of collocation methods with spline functions of even degree, Math. Comp. 45 , No. 171 (1985), 91 - 108. 32] G. Schmidt, On spline collocation methods for boundary integral equation in the plane, Mathem. Meth. Appl. Sci. 7, 74-89 (1985). 33] G. Schmidt, On collocation for pseudodi erential equations on closed curves, Math. Nachr. 126 (1986), 183-196. 34] I. B. Simonenko, A new general method for the investigation of linear operator integral equations, I. Izv. Akad. Nauk SSSR, Ser. Mat. 29:3, 586-586 (1965) (Russian). 35] E. P. Stephan and W. L. Wendland, Remarks to Galerkin and least squares methods with nite elements for general elliptic problems, Manuscripta Geodaetica 1 (1976), 93-123. 36] W. L. Wendland, Boundary element methods and their asymptotic convergence, In: Theoretical Acoustics and Numerical Treatments, ed. P. Filippi, CISM Courses and Lectures, No. 277, 135-216, Springer-Verlag, Wien, New York (1983).

32