Introduction to the boundary element method.A case study: the Helmholtz equation

FranciscoJavier Sayas

January 2006

Contents

1 The Helmholtz equation 2

2 Scattering 4

3 Singlelayer acoustic potentials 8

4 The Sobolev setting 11

5 Discretization 16

6 Theoretical aspects: the continuous case 19

7 Theoretical aspects: the discrete case 24

8 Discretization revisited 29

9 A word on the twodimensional case 35

10 A direct formulation 36

11 Soundhard scattering 3911.1 Doublelayer potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4011.2 Single layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4211.3 Direct formulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

12 Further reading 46

1

Introduction

These are classnotes for the

Tercera Escuela de Verano

Universidad de Concepcion (Chile).Although the course will be taught in Spanish, these notes have been prepared in Englishto reach a wider audience and to encourage Spanish speaking students to read scientifictexts in English.

These notes are prepared so that the reader gets a feeling of what Boundary ElementMethods are and to have a taste of their analysis, which is, let us put it simply, a bithard. I firmly believe in understanding things first with an example and then organisingyour ideas to have a wider view of the problem. If you come from a very Mathematicallyoriented background, you will need faith to follow some of the steps, which are justsketched. Make the effort and get yourself a general idea of what this is about. If thenyou want to learn more, go for the many books on the subject. At the end I give somereferences where you can go on.

Instead of doing the theory with Laplaces equation, Ive preferred to work with theHelmholtz equation as an excuse to introduce some topics (Fredholm theory, compactoperators) and because Helmholtz fields are steady state waves that mix mathematicsand physics in an attractive way. The style is easygoing and I propose several exercises.Most of them are quite simple and they are provided so that you are given some time tohandle with these mathematical objects.

1 The Helmholtz equation

Consider a complex number k such that Im k 0. The equationu+ k2u = 0

is called the Helmholtz equation. Usually k is called the wave number. Even if k2 Rwe will always look for complex valued solutions to this equation. Notice that therestriction to Im k 0 is immaterial, since what we are using is k2 and not k, so what weare doing is restricting (arbitrarily) to one of the square roots.

Remark. When k = r, we obtain the following equation (Ive changed signs to make it lookclearer):

u+ r2u = 0.This is a typical elliptic equation, similar as those one obtains in reactiondiffusion systems.This equation is usually called the Yukawa equation. Waveoriented people find it bad taste tocall this equation the Helmholtz equation. More on this later.

The most important appearance of the Helmholtz equation is in the study of timeharmonic waves. Consider the wave equation

c2v = vtt,

2

where c and are positive numbers (again, that c is not negative is irrelevant, but hasto be positive). A timeharmonic (or steady) solution to the wave equation is a solutionof the form:

v(x, t) = Re[u(x) e t

]= ure(x) cos( t) + uim(x) sin( t).

Remark. A physicist would write this in terms of a positive amplitude and a phase

v(x, t) = A(x) cos( (t (x)) ), A 0.

Written in this way, one sees clearly that the solution is oscillating at the same frequency inall points ( is the frequency), but there is a different amplitude (maximum oscillation) and adifferent phase (the time where oscillation reaches its maximum) in each point. Seeing thingsas linear combinations of a sine and a cosine leads to a simpler mathematical statement: norestriction on positivity for the amplitude; no angular redundancy for the phase; and, above all,linearity! Notice also that u is not a proper amplitude, since it is not a positive real function; uis called a complex amplitude.

If a timeharmonic wave solves the wave equation, necessarily

c2u = 2uthat is,

u+ k2u = 0, k :=

c,

and u has to satisfy the Helmholtz equation. Notice that the wave number k is real,proportional to frequency and inversely proportional to c (velocity of transmission in themedium). It is also proportional to the square root of (the density).

Exercise. Instead of the wave equation, consider the following equation:

c2v = vtt + vt,

where , > 0. If this equation has a timeharmonic solution, what steadystate equation hasto satisfy its amplitude? Write explicitly the wave number.

Remark. The equation in the exercise above is a wave equation with damping. The onedimensional case (one dimension in space) is called the telegraphers equation. When the wavenumber in the Helmholtz equation has a nontrivial imaginary part, the medium is said to beabsorbing. The case k R is usually called the acoustic case. Notice also that high frequenciescorrespond to high wave numbers: one usually speaks of high frequencies. However, when themedium transmits waves very fast (c is large), high frequencies correspond to moderate wavenumbers.

Exercise. Look for timeharmonic solutions to the heat equation: (, > 0)

v = vt

3

How is the wave number in this case?

Exercise. Consider the Helmholtz equation

u+ k2u = 0

in a bounded domain . Prove that any classical boundary value problem (Dirichlet, Neumann)can be given a variational formulation by means of the sesquilinear form

a(u, v) :=u v k2u v.

If k2 = + with > 0, prove that there exists and C > 0 such that

Re[ea(u, u)

] Cu21,, u H1().

(here 1, is the classical H1()norm).

2 Scattering

Let now be a bounded domain in R3 with Lipschitz boundary . Consider the exteriorof , denoted +. We fix a wave number k and consider the Helmholtz equation

u+ k2u = 0, in +.

Let us assume that we have a solution of the same equation in the whole of the space,that is, a function uinc : R3 C satisfying

uinc + k2uinc = 0, in R3.

The wavefield uinc is called an incident field, or more generally, the correspondingwave is called an incident wave (recall that the complex amplitude defines a wave aftermultiplication with the timeoscillating term).

+

+

For instance,uinc(x) := e

kdx, |d| = 1

4

is an incident wavefield, corresponding to a plane wave. Then we decompose the totalcomplex amplitude (the unknown of our problem) as

u = uinc + uscat.

By linearityuscat + k

2uscat = 0, in +,

that is, the scattered field is a solution to the Helmholtz equation with the same wavenumber. The scattering measures the influence of the presence of the obstacle in thewavefield uinc.

Exercise. Prove that plane waves are solutions to the Helmholtz equation. What is thecorresponding incident wave in the time domain?

Exercise. If x0 is a fixed point in R3, prove that the function

uinc(x) :=14pi

e k |xx0|

|x x0|satisfies the Helmholtz equation in R3\{x0}. These incident waves are called point sources. Ingeneral we will not demand that incident waves satisfy the Helmholtz equation in the whole

space (as plane waves do), but only on a volume including the obstacle (as point sources with

x0 + do).

+

x0

We then have to give a boundary condition. The obstacle is called soundsoft when

u| = 0or, seeing uscat as unknown

uscat| = uinc|.Soundhard bodies are those for which

u| = 0,or equivalently

uscat| = uinc|.Here is the normal derivative, where the normal vector points always outwards.

5

The partial differential equation plus a boundary condition for uscat are not enoughsince we are dealing with a problem in an unbounded domain. We will have to give somekind of condition at infinity. This is what comes now.

Let us consider again the point source (spherical wave) defined in the last exercise,taking the origin as source, i.e., the function

u =1

4pi

e k r

r, r = |x|.

Then

ru k u = 14pi

e k r

r2,

which means that the leading term of the radial derivative of this function is k u. In thetime domain this corresponds to

Re[ e k r4pi r

e t]=

1

4pi rcos(k r t) = 1

4pi rcos(k (r

kt)).

Notice that /k is proportional to c (the factor is related to the density ). Notice alsothat the process

(x) 7 (x c t)moves the (graph of the) function , c units to the right every unit of time, that is, ismoving rightwards at speed c.

ct

(x) (xct)

Coming back to1

4pi rcos(k (r

kt)),

this is a function moving radially outwards at speed /k c. Besides, the function isitself oscillating in space: it is a cosine with decreasing amplitude (amplitude decreasesproportional to distance to the origin), the wave number being the number of oscillationsper space unit. Again: c is related to how fast the waves travel (speed) and k is related tohow oscillating it is. If one stays in a point, what one sees is an oscillation with frequency, but if one moves with the oscillation, the perceived frequency is different. Complicated,isnt it?

If we consider instead the function

e k r

4 pi r,

6

or in the time domain,

Re[e k r4pi r

e t]=

1

4 pi rcos(k r + t),

we obtain an incoming wave, travelling from infinity to the origin. The outgoing wavesatisfies the condition

limr

r (ru k u) = 0.This limit is uniform in all directions. This is usually written

ru k u = o(r1), r (with the implicit understanding that a radial limit is uniform along all radii). Theincoming wave does not satisfy this condition. This condition at infinity is usually calledthe Sommerfeld radiation condition.

The full scattering problem is written in terms of the scattering amplitude (we call itsimply u uscat): we impose the Helmholtz equation in the exterior of the obstacle

u+ k2u = 0, in +,

the Sommerfeld radiation condition

ru k u = o(r1), r .,a Dirichlet boundary condition

u| = g0if the obstacle is soundsoft (g0 = uinc|) or a Neumann boundary condition

u| = g1if the obstacle is soundhard (g1 = uinc|).

Recall that the true unknown of the problem is the total complex amplitude u+ uinc,but we will always consider the amplitude of the scattered wave as unknown.

Remark. In scattering problems, typically righthand sides are smooth functions restrictedto the boundary (why solutions to the Helmholtz equation are smooth is a question we willdeal with later on). However, both theory and numerics are always worked out with generalrighthand sides. As incident waves one considers mainly plane waves and point sources placedon points outside the obstacle.

There are also some other interesting problems related to scattering with more com-plicated boundary or transmission conditions. One of them, of great relevance when wemove from acoustic to electromagnetic waves, is the use of impedance conditions, of theform

u+ u = g.

Helmholtz transmission problems stem from situations when the wave is transmitted insidethe obstacle, which has different material properties (density and speed of transmissionof waves).

7

3 Singlelayer acoustic potentials

Let us consider the following function:

(x,x0) :=e k |xx0|

4pi |x x0| .

This function (of 2 3 variables) is called the outgoing fundamental solution to theHelmholtz equation, because it satisfies

( ,x0) + k2( ,x0) = x0(we will propose a proof of this in a forthcoming exercise) and the Sommerfeld radiationcondition at infinity. If we place a finite number of point sources on (xj , j), withdifferent charges/masses j, its superposition

j

(x,xj)j =j

e k |xxj |

4pi |x xj| j,

also satisfies the Helmholtz equation in + together with the Sommerfeld radiation con-dition. The basic idea is to convert this sum into an integral by using a density : Cand proposing

(x,x0)(x0)d(x0) =

e k |xx0|

4pi |x x0|(x0)d(x0)

as a candidate to solve our scattering problem:

u+ k2u = 0, in +,

ru k u = o(r1), r u| = g0.

Notice (prove it!) that it satisfies automatically the differential equation and the radiationcondition. What is left to impose is the boundary condition. But first, some nomenclature:the function (or class of functions)

(S)(x) :=

(x,x0)(x0)d(x0) =

e k |xx0|

4pi |x x0|(x0)d(x0)

is called a singlelayer potential and is called the density.

Exercise. Let D(R3) and 0(x) := e k |x|/(4pi |x|). Prove thatR30(+ k2) = lim

0

B(0;)

(r0) 0(r).

Write down the explicit expression for

r0|B(0;) and 0|B(0;).

8

Use them to prove that

lim0

B(0;)

(r0) = (0)

and that

lim0

B(0;)

0(r) = 0.

Conclude that0 + k20 = 0, in D(R3).

Deduce that( ,x0) + k2( ,x0) = x0 , in D(R3).

Remark. Since we are solving the exterior problem, we could think of using a potentialgenerated with a density defined on the whole of the obstacle, not only on its boundary, that is,something of the form

(x,x0)(x0)dx0.

Electrostatics (formally the case k = 0) say clearly that we shouldnt do this: in the equilibriumcharges are placed on the boundary. Mathematically this is a bit more complicated and youhave to go to deep results of potential theory to prove that trying to solve our problem with apotential defined on the obstacle leads to a (distributional) solution supported on the boundary.Anyway, from our point of view, we are only trying with a class of functions. The fact that theclass will be large enough to include a solution will be sufficient for us.

Remark. Another possibility would be considering a surface (or a volume) 0 strictly con-tained in the obstacle and trying a potential from there:

S0.

Notice that on (the true boundary of the scatterer) this function is very smooth. Imposingthe boundary condition is extremely simple

0

(x,x0)(x0)d(x0) = g0(x), x .

The approach looks correct but leads to a very illposed problem. But a very interesting problemthough! If we change the point of view, assuming that 0 is the boundary of the scatterer andthat we have information on a surface surrounding the obstacle , then we are dealing witha problem of acoustic holography, one of the simpletostate but difficulttosolve illposedproblems of acoustics. We will not look further into this issue.

We go then back to the definition of the potential

u := S :=

( ,y)(y)d(y),

9

and notice again thatu C(R3\)

u+ k2u = 0, in + = R3\and that u is an outgoing wavefield (u satisfies the Sommerfeld condition at infinity). Tosimplify things, assume now that and are as smooth as you need them to be for whatis coming. A simple argument using Lebesgues theorem (the dominated convergencetheorem) proves that the following limit exists

limxx0

(S)(x) =

(x0,y)(y)d(y),

i.e., the same expression (now with an isolated singularity in the integrand; we are notintegrating something smooth anymore) gives the interior (denoted with the sign ) andexterior (+) limits of the potential on the boundary. If we take the difference of thenormal derivatives from inside and outside, we find this other fact ((x0) is the normalvector at x0 pointing outwards)

( S)(x0) (+ S)(x0) = limxx+0

S(x) (x0) limxx0

S(x) (x0) = (x0),

which means that is the jump of the normal derivative of the potential. Notice that thegradient is continuous (formally at least) in tangential directions, which means that thediscontinuity of the gradient of the potential when approaching the boundary is a normalfield proportional to density.

Remark. If we were dealing with electrostatics we would say that electric charges createcontinuous potentials but discontinuous electric fields.

Then we define the singlelayer operator

(V)(x) :=

(x,y)(y)d(y), x .

You can be wondering now: but was this not S? In a way it was, but there is animportant difference. Given a density : C, S is a function defined on the wholeof R3 (a priori, we are interested only in + or even only on +), whereas V is afunction defined on . In fact, V is the value on of S. We recall names again: Sis a singlelayer potential and V is a singlelayer operator. The important thing isnot the name or the letter we use (youll find all kind of possibilities in the literature),but the concept: one is the field defined in the space, the other is a function defined inthe same surface as the density. The operator

7 V :=

( ,y)(y)d(y)

is a linear integral operator. It is the first occurrence of what we call a boundary integraloperator. We finally close the circle. The problem: find a density such that

V = g0 in

10

and then defineu := S

is a boundary integral method. One first solves an integral equation and then inputs itssolution in a potential expression. An integral equation involving a boundary integraloperator is called a boundary integral equation (BIE for short). Notice that aftersolving this BIE, we have obtained a solution of the exterior Dirichlet problem for theHelmholtz equation and that this solution is defined by means of a potential. The fact thatwe used an artificial device as a singlelayer potential, where a nonphysical quantity (thedensity) is involved, justifies why this kind of approach is called an indirect method.

Exercise. Prove that if L1(), both S and V are well defined. Prove also thelimiting expressions on for S and its gradient.

4 The Sobolev setting

Remark. If you are an engineer, you would be willing to have some numbers (a discretization)as fast as possible. Skip this section and move on to finding a numerical scheme. But then comeback here, even if you have to ignore the difficult parts with all the Sobolev spaces, because thereare some nontrivial solvability questions to be taken into account. Nontrivial and important.

Recall first your Sobolev spaces. H1() is the classical Sobolev space of order one (itselements and their first distributional derivatives are in L2()). The norm of H1() isdenoted 1, and the one of L2() = H0() as 0,. H1/2() is the trace space.The trace operator

: H1() H1/2()is continuous and onto. Since the inclusion

H1/2() L2()is dense and continuous, we can add the dual space of H1/2() to the other side of thisinclusion and obtain something like

H1/2() L2() H1/2().In this way, H1/2() is the dual space to H1/2() (the set of continuous linear maps fromH1/2() C), seen in a way such that any g L2() defines an element of H1/2() bymeans of the expression

H1/2() 3 7

g(y)(y)d(y),

and this element (a bounded linear functional H1/2() C) is also called g. The dualityproduct will be denoted

g, .

11

A priori g H1/2() and H1/2(), but if g L2() (which is a proper dense subsetof H1/2() with the identification above), this is just

g, =

g(y)(y)d(y).

We are not going to dive very deep into this, since this course is introductory. Nevertheless,if you are really willing to understand the mathematics of boundary elements, there is noway round.

Remark. A triple V H V defined as above (the key is that the injection of V into H iscontinuous and dense) is called a Gelfand triple or Courant triple (or triad).

The weak normal derivative is defined as follows: if u H1() and u L2(), thenthe expression

u v +

u v

is a continuous function of the trace of v H1() (not of v itself). Thus, we define

u, :=

u v +

u v

where v is any element of H1() such that v = . Maybe it looks simpler if we write itas follows

u, v :=

u v +

u v, v H1().

Then u H1/2() and

u1/2, C[u0, + u0,

].

It is not very obvious at first sight, but to define u on we only need that u and uare in L2 in a region near the boundary , since distant parts of the domain do not playany role in the definition.

We are going to need a new set. We say that

u H1loc(+)

if the restriction of u to any bounded open subset of + (possibly touching the boundary) is in H1. This local space keeps functions with locally H1 behaviour, ignoring therate of decay at infinity necessary to have squareintegrability. Since the trace operatoris local, we can define the exterior trace

: H1loc(+) H1/2().

Also we can define the exterior normal derivative for functions inH1loc(+) whose laplacian

is in L2loc(+).

12

Remark. The exterior trace is the same as the interior trace operator, but for functionsdefined on the other side of the boundary. It is, let me emphasize this, the same operator. Thesame thing applies for the exterior normal derivative. The key to a good understanding of theseoperators is that, even if their definition looks global (it is done in an operational way), finallythe operators are local, taking into account what happens around the boundary and not farfrom it.

Remark. The notation H1loc(+) is the correct one. Id be tempted to write the simplerform H1loc(

+), but what this really means is local H1 behaviour strictly inside +, that is, notonly ignoring the infinity but also the limits on . A very smooth function with a very strongsingularity on belongs to H1loc(

+) but not to H1loc(+).

Theorem For arbitrary H1/2(),S H1()H1loc(+),

is an outgoing solution of u+ k2u = 0. The interior and exterior traces coincide

(S) = Vand the gradient has a jump in the normal direction:

S + S = .Finally,

V : H1/2() H1/2()

is linear and bounded.

Remark. Notice that for any x R3\, the potential can be defined by duality(S)(x) = , (x, ).

The smooth part of the proposition above (that this function satisfies the Helmholtz equationplus the Sommerfeld radiation condition at infinity) can be proven with very classical arguments.The part related to traces and normal derivatives requires a higher level of analysis.

Uniqueness for boundary value problems for the Helmholtz equation is not straight-forward. This is what can be obtained.

The exterior problem has uniqueness of radiating solution:u+ k2u = 0, in +,

ru k u = o(r1), r u| = 0

implies that u 0

13

There is a sequence of eigenvalues for which the interior problem has no uniquesolution, i.e.,

u+ k2u = 0, in ,

u| = 0

implies that u 0 if and only if k2 is not a Dirichlet eigenvalue of the Laplacianin .

The Dirichlet eigenvalues of the Laplace operator form a sequence of numbers

0 < k21 < k22 < . . . < k

2n < . . .

for which the problems

u = k2nu, in ,u| = 0

have nonunique solution. To each eigenvalue corresponds a finite dimensionaleigenspace.

Remark. One can say a lot more about the Dirichlet eigenvalues and eigenfunctions of theLaplacian. For instance, the first eigenvalue is simple. Also, taking an L2()orthonormal basisof each eigenspace we obtain a Hilbert basis of L2(). Upon normalization this basis is also aHilbert basis of H10 (). This is precisely the origin of the method of separation of variables thatcan be found in any classical text of differential equations.

Assume now thatV = 0.

Defining u := S, we have

u+ k2u = 0, in + ,ru k u = o(r1), r

u| = V = 0

and therefore u+ 0. If k2 is not a Dirichlet eigenvalue of the Laplacian (in the interiorof the obstacle), then u 0. Hence

= u + u = 0,

and we just proved that V is injective.If on the other hand, k2 is a Dirichlet eigenvalue of the Laplace operator, and we

take 0 6= such that

= k2, in ,| = 0,

14

then = 6= 0. The reason why a nontrivial Dirichlet eigenfunction of the Laplaceoperator cannot be at the same time a Neumann eigenfunction, i.e., why

= k2, in ,| = 0,

| = 0,= 0

will be derived from a result we will see in a forthcoming section. With precisely thatresult it is possible to see that V = 0.

The preceding (formal) argument proves that

V is injective k2 is not a Dirichlet eigenvalue of in .In particular, when k2 has nontrivial imaginary part (the medium is absorbing), V isinjective.

Remark. Invertibility of V can be trivially used to show existence of solution to the exteriorHelmholtz problem with Dirichlet condition. Then the remaining point is showing invertibility(surjectivity!) of V, which follows a more complicated path. Well sketch the proof in due time.Anyway, you will definitely not learn here all the details of the proofs and Ill be very careful tohide some complicated facts.

As mentioned, if V : H1/2() H1/2() is invertible we can solve the exterior

Helmholtz equation (the soundsoft scattering problem)

u+ k2u = 0, in +,

ru k u = o(r1), r u| = g

for arbitrary g H1/2() as follows: first we solve the integral equation on

V = g

and the we input in the potential expression

u = S

to obtain the solution of the scattering problem. By looking at the integral equation(formally; in truth is not a function)

(V)(x) :=

(x,y)(y)d(y) = g(x), x

we observe the very nonlocal character of the equation, that is, the value of in someregion influences the value of V in the whole of .

15

5 Discretization

We begin by showing the simplest Galerkin scheme for the integral equation

(V)(x) =

(x,y)(y)d(y) = g(x), x .

Let us decompose the boundary into a finite set of curved patches

= {i | i = 1, . . . , N}with trivial overlappings: i j is onedimensional for i 6= j. We point out that we arelooking for in the space H1/2() and that

L2() H1/2().This implies that we will not need any degree of continuity in the discrete space here.Consequently, we can work with basically any partition of the boundary in curved polygonswithout the classical rules of eitherfullsidesorvertices match of finite element spaces.A traditional triangulation of the surface is however the typical choice.

Remark. A highly nontrivial question when we are dealing with triangulations of surfacesis what do we consider as a triangulation and how can we manage the fact that we usuallylack a parameterization valid for the whole surface. The following is a theoretical setting whichcorresponds to what we usually find in practice. Related to the real boundary there is a closedpolyhedron and a bijection

: which is: (a) continuous; (b) globally Lipschitz with globally Lipschitz inverse ; (c) smooth(for instance C2) on each face of the polyhedron . The second hypothesis means that we canestimate distances on the surface (geodesic distances) by distances on the polyhedron or byjoining points through the threedimensional space. The estimate can be bad but does notdegenerate. The polyhedron is composed of M polygonal faces {P1, . . . , PM} and their images{(Pn)|n = 1, . . . ,M} define a level zero partition of . We do not plan to discretize here. Thenwe can define a triangulation of the polyhedral surface in the usual triangles or quadrilateralswith the common rules (not really necessary here, but...). These planar elements on the straightfaces are then mapped onto the surface and define the partition of the surface.

Instead of a generic density : R we take h : R such thath(x) j, x j,

that is, h is constant on each element. Then

(Vn)(x) =

(x,y)h(y)d(y)

=Nj=1

j

(x,y)h(y)d(y)

=Nj=1

[ j

(x,y)d(y)]j.

16

Obviously Vh will not match g everywhere. We only demand that the average value ofboth functions on each element i coincide, that is, we demand that

i

(Vh)(x)d(x) =

i

g(x)d(x), i = 1, . . . , N

or equivalently

Nj=1

[ i

[ j

(x,y)d(y)]d(x)

]j =

i

g(x)d(x), i = 1, . . . , N,

which is an NN linear system, whose unknowns are the values of the piecewise constantdensity on the elements.

To avoid an excess of parentheses we will always adopt the following conventionA

B

f(x,y)d(x)d(y) :=

A

(B

f(x,y)d(y)

)d(x),

that is, the variables and sets of integration are given in the same order.

Exercise. Let

aij :=i

j

e k |xy|

4pi|x y|d(x) d(y)

be the matrix of the linear system above. Prove that it is symmetric but nonhermitian in

general. For which values of k is this matrix hermitian? Notice that a priori aij 6= 0 for all i, j.

The method we just described can be easily understood as a Galerkin method.First, remark that

V = g

is equivalent toV, = g, , H1/2().

The angled brackets , represent the duality of H1/2() on H1/2(). In this case,the second component is the element of H1/2() (the dual space) and the first is theelement of H1/2() (the primal space).

We can then define a bilinear form a : H1/2()H1/2() C

a(, ) := V, and a linear form ` : H1/2() C

`() := g, and write the whole problem as a typical variational problem in the Hilbert spaceH1/2()

H1/2(), a(, ) = `(), H1/2().

17

The bilinear and the linear forms are continuous. Notice also that the bilinear form canbe formally written as

(x,y)(y)(x)d(x) d(y),

that is, it involves a double integral on .

Remark. This is not properly a variational formulation in the sense you will have found inelliptic PDE theory and finite element implementations, since we have not moved derivativesfrom the original equation to the tests. It is simply a straightforward testing of an integralequation. This is a common feature to most boundary integral equations. In some equations,however, after this straightforward testing, we apply some identities to rewrite the bilinear formwithout changing the regularity requirements of the unknown though.

If we define nowXh := {h : R |h|j P0, j}

where P0 is the space of constant functions (polynomials of degree zero), then we candefine a Galerkin scheme by

h Xh, a(h, h) = `(h), h Xh.

Exercise. Prove that this Galerkin scheme is equivalent to the method we derived before.Why can we properly write

a(h, h) =

(x,y)h(y)h(x) d(x) d(y)

without needing duality products?

Assume again that we approximate by h Xh. Instead of requiring that Vh andg have equal averages on the elements, we do the following: we choose a point xi i oneach element; we demand that

(Vh)(xi) = g(xi), i = 1, . . . , N.

The method thus derived is called a collocation method. The points {xi} are calledcollocation nodes.

Exercise. Write this method as an equivalent system of linear equations. Is the correspondingmatrix symmetric?

Remark. Generally speaking engineers prefer the collocation method to the Galerkin method.The essential reason is that we have simpler integrals to compute or approximate and that thededuction is somewhat simpler in the sense that we are neither averaging on the cells norconsidering an intermediate variational formulation.

There are however some aspects (pros and cons) to be taken into account:

18

Not every choice of collocation nodes is going to work. For instance, taking a vertexper triangle is not a judicious choice.

There is not a satisfactory theory for this method working even for smooth surfaces.This should not worry too much a practitioner of the method: most people areconvinced that the theory will arrive in due time. However, the Galerkin settinggives more confidence to mathematically (or theoretically) oriented users of theboundary element method.

If g has discontinuities, where one places nodes is also relevant. In some instances, practitioners of the method use many more collocation nodesthan elements and solve the (incompatible) equations by mean squares. This hasthe advantage that we have to solve a Hermitian positive definite system. If we aregoing to use conjugate gradient iterates we dont even have to create the normalmatrix AA. There is a big drawback to this nevertheless. The conditioning ofthe systems we have obtained is not very good (it is not dramatically bad) andthe normal equations have it squared, which makes everything worseoff by makingiterative methods far too slow. The other point in favour of taking many morecollocation points than elements (two or three per element) is that maybe a subsetof them is stable and compensates the bad choice of other sets. My point here isthen: dont do this!

Sometimes, one sees in texts the collocation method written in a variational form as

h Xh, Vh, i = g, i, i = 1, . . . , Nor even as a PetrovGalerkin method

h Xh, Vh, h = g, h, h Yh = span{i | i = 1, . . . , N}where Yh is the space of linear combinations of the Dirac delta distributions i on thecollocation nodes xi. One has to take this with more than a bit of salt. It is formallyokay and makes the method look a particular case of the PetrovGalerkin method, butthe angled brackets are not really correct and one is in fact using a far more complicatedtheoretical setting.

In Section 8 we will work a bit more on the matrices for the Galerkin and collocationmethod as well as proposing a method based on point sources.

6 Theoretical aspects: the continuous case

The theory of boundary integral formulations and boundary element methods is basedboth on the wellknown elliptic theory and in the Fredholm theory, which is in a way thebranch of functional analysis that studies the effect of compact perturbations to operatorequations. Let us first expose the main results concerning this theory at the continuouslevel. For the next section we leave the effect of compact perturbations at the discretelevel.

19

Let H1 and H2 be Hilbert spaces. A linear operator K : H1 H2 is compact if anyof the three following equivalent conditions hold:

(a) K transforms weakly convergent sequences into strongly convergent sequences.

(b) If (n) is a bounded sequence inH1, then the image sequence (Kn) has a convergentsubsequence.

(c) There exist two orthonormal sequences, (n) in H1 and (n) in H2, and a sequenceof positive numbers n 0 such that

K =n

n( , n)n,

where the bracket denotes the inner product in H1 and convergence of the seriesholds in operator norm.

In the last characterization, the sum can be limited to a finite number of terms, in whichcase the operator is called degenerate or of finite rank.

Remark. That these three characterizations are equivalent is not obvious at all and requiressome effort of applying wellknown but strong results of Hilbert space theory. It is not verydifficult to prove that (b) is equivalent to: The image of any bounded set is relatively compact,that is, has a compact closure. This formulation allows for extensions of the definition to Banachor even normed spaces, and also to nonlinear operators. Characterizations (a) and (c) areHilbertian in essence. The series decomposition in (c) is called the Singular Value Decompositionof K and shows very clearly the illposed character of equations of the form K = .

From the point of view of what we are doing, the most important result of this theoryis the following.

Theorem Let H be a Hilbert space and K : H H be compact. Then:I +K is injective I +K is surjective.

In this case, the inverse of I +K is bounded.

Remark. The last part of the theorem is of general application for any bounded invertibleoperator between Hilbert spaces: its inverse is always bounded. This result holds in Banachspaces, where it is often called the Banach isomorphism theorem, and it is a consequenceof a more general result called the Banach open mapping theorem. You can find this typeof result in any book of basic functional analysis.

Exercise. Let H1,H2 and H3 be Hilbert spaces. Prove the following results:

(a) If A : H1 H2 is bounded and K : H2 H3 is compact, then KA : H1 H3 iscompact. (Hint. Use characterization (b) of compactness). If A : H1 H2 is boundedand K : H3 H1 is compact, then AK : H3 H2 is compact.

20

(b) If A : H1 H2 is bounded and inversible and K : H1 H2 is compact, thenA+K is injective A+K is surjective.

Remark. The theorem we enounced above is the first part of a very important result infunctional analysis with huge relevant applications to elliptic theory, even if it does not look likethat at first sight. The theorem is called the Fredholm alternative and, in its simplest form,states the following. If K : H H is compact then:

either I +K is invertible with continuous inverse,

or ker(I +K) is finite dimensional and im(I +K)> has the same finite dimension.

In the second case, there exist 1, . . . , N linearly independent (a basis of the orthogonal com-plement of im(I +K) such that the equation

+K =

is solvable if and only if is orthogonal to i for all i.

The final definition of this theory is given also in a set of equivalent characterizations.Let again H1 and H2 be Hilbert spaces. A bounded linear operator V : H1 H2 is aFredholm operator (short for a Fredholm operator of index zero) if any of the followingequivalent conditions hold:

(a) dim(kerV ) = dim(imV )

V0 is elliptic, that is, there exists > 0 such thatV0 , 1/2,, H1/2().

V V0 is compact.Hence, V is Fredholm of index zero and we have that

V is injective V is surjective.Moreover, we can write V = V0 + (V V0) and therefore we have a decomposition

V = V0 +K,

where V0 is elliptic and K is compact. This will be relevant in the numerical approxima-tion.

We now can collect the pieces so far:

We had begun with the exterior scattering problemu+ k2u = 0, in +,

ru k u = o(r1), r u| = g.

A priori g = uinc on , but we had widened our interest to arbitrary data inthe trace space: g H1/2(). The problem with g = 0 only admitted the trivialsolution, so we have uniqueness.

Our proposal to solution was a singlelayer acoustic potential

u := S =

e k | y|

4 pi | y|(y)d(y).

This expression gives also a definition of solution inside the obstacle. (By the way,if you really can sound as if you know your business with boundary integral formu-lations, you dont give a proposal of solution; you give an ansatz, which is Germanfor proposal).

We have proved that u = S solves the scattering problem if and only if V = g,that is

e k |xy|

4 pi |x y|(y)d(y) = g(x), x .

(The equation has to be understood with the integral operator in a weak sense, H1/2() and holding almost everywhere in ).

The operator V is injective if and only if k2 is not a Dirichlet eigenvalue of theLaplace operator in the interior domain. Injectivity was a consequence of uniquenessof the exterior (always) and interior (with the exception of eigenvalues!) boundaryvalue problems.

22

Hence, except in the singular cases, V is invertible. Therefore we can give a solution(the unique solution) of the exterior scattering problem.

The preceding argument furnishes a proof of existence of solution of the exterior scat-tering problem except for some values that correspond to interior Dirichlet eigenvalues.The usual proof for existence uses this kind of argumentation, only with a more compli-cated potential (ansatz) that does not fail to be invertible for some wave numbers.

Before going any further, let us examine what happens with more than one obstacle.Two will be enough. Assume that the obstacle is not a single domain, but the union oftwo disjoint domains 1 and 2 with boundaries 1 and 2. All the integrals over canbe decomposed

=

1

+

2

.

Let gi := g|i H1/2(i). Consider the following operators

Vij :=

i

( ,y)(y)d(y) : j C, i, j = 1, 2.

The notation is taken so that Vij goes from i to j. The density of the potential isgenerated in i and observed in j. This is is very similar to what we had in numericalapproximation, but here 1 and 2 are closed disjoint surfaces.

Exercise. In the notations above, answer these questions:

(a) When are Vii inversible?

(b) Using that the inclusion of H1(i) into H1/2(i) is compact, prove that V12 and V21 arecompact.

(c) If i H1/2(i), prove thatu = S11 + S22

solves the soundsoft scattering problem exterior to 1 2 is and only if[V11 V12

V21 V22

] [1

2

]=

[g1

g2

].

(d) Prove that [V11 V12

V21 V22

]is Fredholm as an operator H1/2(1)H1/2(2) H1/2(1)H1/2(2).

(e) Prove that if V11 and V22 are invertible, then the whole matrix of operators is invertible.

This looks new but follows directly from the theory by admitting not to be a singlesurface but a set of boundaries of nonintersecting domains (one surface enclosed byanother gives a very different problem though!). The point is that Dirichlet eigenvaluesin 1 2 are those of 1 and those of 2, since the problems are completely decoupled.

23

One can arrange (theoretically) the system of boundary integral equations[V11 V12

V21 V22

] [1

2

]=

[g1

g2

]

by multiplying each row by V 1ii , obtaining thus,

1 + V111 V122 = V

111 g1,

V 122 V211 + 2 = V122 g2.

Iterations to this formal system (that is, Jacobi iterations to the first system) involvesolving scattering problems exterior to each domain, not to the group of both obstacles.This is the origin of a technique often used by physicists and called multiple scattering.Each object reacts to the scattering produced by others. You can always think in thefollowing terms: densities emit a timeharmonic wave from each i; the operators Vijwith i 6= j correspond to emissions from i heard in j; the operators V 1ii are thereactions of each obstacle to emissions (something is received in i and the boundaryemits something just to compensate). Iterations happen till an asymptotic equilibrium isreached. Even if you are not going to use this kind of iterations, acquiring this languagewill help you to make yourself better understood.

7 Theoretical aspects: the discrete case

The Galerkin method we exposed two sections ago fits into a very general frameworkof Galerkin methods for operator equations. Consider an invertible operator A : H H , where H is a Hilbert space and H is its dual. We want to study the numericalapproximation for the operator equation

A = g.

Let Hh be a sequence of finite dimensional subspaces of H, directed in the parameter h,that tends to zero.

Remark. The parameter h can be a geometrical parameter (as it happens often in finiteelement analysis) or not. With it we simply want to describe that there are several subspacesand that, in general, they become richer as h 0. From now on we will use a commonconvention when writing numerical mathematics: C > 0, with possible subscripts, will be aconstant independent of h and of any other quantity it is multiplied by. The constant can bedifferent in different occurrences, unless we state the opposite.

The Galerkin approximation is the solution (if it exists and if it is unique) to thediscrete problem:

h Hh, Ah, h = g, h, h Hh.Before going on, let us point out some simple facts:

24

(a) If 1, . . . , N is a basis of Hh, the discrete problem is equivalent to the linear system

Nj=1

Aj, ij = g, i, i = 1, . . . , N

where (1, . . . , N) is the vector of coefficients of h Hh in that basis:

h =Nj=1

j j.

(b) Therefore existenceanduniqueness (both together) do not depend on the righthand side. They only depend on the invertibility of the matrix

Aj, i, i, j = 1, . . . , N.

(c) If the operator A is selfadjoint

A, = A,, , H,then the corresponding matrix at the discrete level is hermitian. If the operator issymmetric

A, = A,, , H,the matrix is symmetric.

(d) The method is based on reduction to finite dimension of the operator equationA = g written in variational form

H, A, = g, , H.As we had already mentioned in the particular case of the singlelayer equation,there is no additional effort in obtaining this formulation. It simply states thatsince A and g are the same element of the dual space H , they coincide whenacting for arbitrary H.

Exercise. Assume that for a particular h, the exact solution belongs to Hh. Prove thath = for that value of h.

Assume that the system is invertible at least for h small enough. The method issaid to be stable if there exists C1 such that

h C1g.(Notice that the norm for h is that of H, but the norm for g is that of H

). Since A isbounded, then stability can be written as a bound of the discrete solutions in terms ofthe continuous solution

h C2,

25

by simply taking C2 = C1 A. Stability implies the existence of C3 > 0 such that

h C3 infhHh

h.

Let us prove this. Take a particular h Hh and let f := A( h). Then, the exactsolution to the equation

A = f

is = h and the Galerkin solution

h Hh, Ah, h = f, h, h Hhis h h. Then

h h+ h h (1 + C2) h,

and therefore h (1 + C2) h, h Hh.

Since Hh is finite dimensional, it can be proved that in the bound

h C3 infhHh

h

the infimum is a minimum. Anyway, the conventional wisdom is writing an infimum, andwe are not going to challenge common uses. In the world of Galerkin approximation toelliptic boundary value problems, this bound is known as Ceas lemma. As we will seelater on, for elliptic equations stability is for free (all Galerkin methods satisfy it). In thenonelliptic world, the bound is known as a Cea estimate. The interesting fact is thatthis bound implies stability, which follows the following very simple argument

h + h + C3 infhHh

hh=0 (1 + C3).

A subtler argument shows that in the inequalities

h C2

and h C3 inf

hHh h

one can always have C2 = C3. The Cea estimate moves the problem of studying theGalerkin error to one purely of approximation theory, namely to prove that

infhHh

h h0 0.

If this holds for arbitrary H, then we know that

h h0 0

26

for all possible righthand sides in the equation. This property is called convergence ofthe method.

Sinceinf

hHh h h,

convergence implies the approximation property

h h0 0, H.

By using a wellknown but nontrivial theorem of functional analysis (the principle ofuniform boundedness or the BanachSteinhaus theorem), it is possible to prove thatconvergence implies stability. Therefore, we have

Convergence = Stability + Approximation property.

Remark. It is also possible to prove that stability is equivalent to the following condition onthe discrete bilinear form: there exists C > 0

sup0 6=hHh

|Ah, h|h C h, h Hh.

This condition is usually called a discrete infsup condition or BabuskaBrezzi condition.In relation to the first stability inequality

h C1g

it is simple to prove that C1 = 1/C. An advantage of using the discrete infsup condition isthat it implies invertibility of the system of linear equations and therefore the definition of thediscrete solution.

Remark. The underlying idea of this kind of analysis of Galerkin methods by showingstability and the approximation property is by no means necessary to prove that this kind ofnumerical methods work properly. Notice that if we want to show convergence for arbitraryrighthand sides (arbitrary solutions), the paradigm

Convergence = Stability + Approximation property

is the correct one. Stability plus knowing that

h h0 0,

for our unknown solution is enough to prove convergence for our sequence of solutions. It couldhappen that the sequence of discrete spaces is correct from the point of view of stability and toapproximate some solutions, but not all.

Theres more, however. Convergence for some solutions can be attained without the re-quirement of stability. This is the key to many numerical methods based on adaptivity, wherethe sequence of subspaces is chosen progressively depending on the behaviour of the discretesolutions obtained. The sequence of spaces depends then on the particular righthand side, is

27

constructed to satisfy the approximation property for the concrete solution but can fail to sat-isfy the infsup condition (stability). The nice paradigm is thus broken into pieces. Things arehowever not moving fast enough to justify forgetting all the good old theories to have somethingnewer and better, so stability is here to stay, at least for a while.

Remark. There are many other possible generalizations of operator equations that fit in aframework suitable for Galerkin approximations. For instance one can work withA : H H andtest the equation either with H itself (via the inner product). The RieszFrechet representationtheorem says that it doesnt matter which of them you use. However, the way the operatorequation is written down allows to use a Galerkin method or not. The Galerkin methods forma family in the wider class of PetrovGalerkin or projection methods. Sometimes a cleverrewriting of the discrete and continuous equations allow to understand a discretization schemeas a projection method, even if it is not selfevident.

If the operator A is elliptic

ReA, 2, H,stability of Galerkin schemes follows readily: it is no longer a hypothesis. First, sinceellipticity is inherited by the discrete spaces

ReAh, h h2, h Hh,(the constant is the same), then the solution to the Galerkin equations

h Hh, Ah, h = g, h, h Hhis unique (this is very simple to prove). Also

h2 ReAh, h |Ah, h| = |g, h| g h,so we obtain

h (1/)gand consequently

h (A/)and the Cea estimate

h (A/) infhHh

h.

Convergence then follows from the approximation property.The following key result relates convergence (not stability!) of Galerkin approxima-

tions with the Fredholm theory.

Theorem Assume that the Galerkin method based on the sequence of spaces {Hh} isconvergent for the operator A. If K is compact and A+K is invertible (it is enough thatA + K is injective), then the method is convergent for A + K. Moreover, the stabilityconstant is proportional to that of the method applied to A multiplied by (A+K)1.

28

The theorem can be looked at with positive eyes and asserts that your method is goingto work well if it works well adding or taking compact terms. For instance, a Galerkinmethod for an operator A+K, where A is elliptic, is going to be stable if the approximationproperty holds. Nevertheless, the stability constant can be quite bad if (A + K)1 islarge and the asymptotic regime (the hs small enough where all the inequalities hold) canbe beginning somewhat late. This is a drawback of a certain relevance in some (but byno means all) applications to scattering problems. More on this in the following section.

8 Discretization revisited

In this section we are going to apply the theory above to our case and to examine with moredetail some aspects related to implementation of the Galerkin and collocation methods.

Let us return to the equation

H1/2(), V = gand its Galerkin discretization

h Xh, Vh, h = g, h, h Xhby means of the space

Xh := {h : C |h|j P0, j}.We know that

V = elliptic + compact

and that V is invertible if k2 is not an interior Dirichlet eigenvalue. In this case, for hsmall enough, the corresponding matrix is invertible, we have a stability bound

h1/2, C 1/2,and the Cea estimate

h1/2, C infhXh

h1/2,.

Remark. The constant C deserves some additional attention. It is a stability constant thatdepends on the ellipticity constant for V0 and on the norm of the inverse of V. The first of theseconstants depends only on the boundary , but the second depends on the wave number. Theconstant will blowup if we approach one of the Dirichlet eigenvalues. Also, when the asymptoticregime (the h small enough so that all inequalities hold) begins depends on the wave number,both on its size and on the proximity to a bad eigenvalue.

The problem with the proximity to a Dirichlet eigenvalue can be solved using a better integralequation. I emphasize that we are still on a first step in an introductory course. It takes timeto see all the difficulties and more time to solve them.

There is another question on the ability of the space Xh to approximate the exact solution.For high wave numbers, the solution can be highly oscillating, and the triangulation has to be

29

very fine to cope will all the oscillations of the exact solution. Keep in mind that high frequenciesrequire specific techniques!

Ifh := max{hj | j = 1, . . . , N}, hj := diam(j)

converges to zero and we assume a maximum flattening hypothesis on the triangulation,then

infhXh

h1/2, h0 0,no matter how singular is. To see that, first approximate by an element in L2()(which is dense in H1/2()) and then make the L2()orthogonal projection of thisfunction. Both terms can be made as small as desired.

If H1(), with some additional analytic effort, we can prove thatinf

hXh h1/2, Ch3/21,.

Notice that we have the norm of in the righthand side and not the typical seminormyou find in finite element analysis. This is not really important and sometimes the boundsfor approximation on surfaces can be worked out a little bit to make them resemble morelike finite element estimates. But one has to acknowledge that curvature of the surface isalways a bore and lower order derivatives are due to appear most of the time in precisebounds. This approximation estimate (which has nothing to do with the boundary integralequation; it is simply interpolation theory) together with the Cea estimate, give orderto the method

h1/2, Ch3/21,under the hypothesis of sufficient regularity of the solution.

The order estimate for approximation is not really easy to prove. We are going tosimply sketch how one can prove this kind of results in this very simple situation. Thefirst try is classical: we bound in H0() = L2() for H1()

infhXh

h0, Ch1,.

This is easy to do by means of the L2()orthogonal projection and classical analysis.Then one tries this other one

infhXh

h1, Ch21,.

This one looks somewhat more difficult to prove since the H1()norm is usually definedby duality, so we have to find a way of obtaining h Xh such that

sup06=H1()

| h, |1, Ch

21,.

Nevertheless, the point here is simply to take again h as the best L2() approximation

and h Xh as the one of an arbitrary H1(), and bound as follows:| h, | = | h, h|

h0, h0, Ch21, 1,.

30

A key point here is that we could find a way of approximating simultaneously in H0()and H1(), that is, the same approximation serves for both bounds. Then we apply thisinequality

h1/2, h1/20, h1/21, Ch3/21,,to obtain the desired bound.

Remark. The general bound

g1/2, g1/20,g1/21,, g H0()

belongs to a very wide class of inequalities by Sobolev norms on the boundary. These are aconsequence of the structure of the family of spaces Hr() (for 1 r 1 or in wider intervalsif the surface is smooth enough). This kind of inequalities are often referred to as interpolationinequalities, which has no relation at all with interpolation in the numerical sense, but on thevery deep fact that all inner spaces are interpolated from the extremes.

To avoid the repeated occurrence of the constant C, which may be different in eachcase, we will adopt the following convention. We write that

a . b

when there exists C > 0 (as usual independent on h and on particular data of the problem)such that

a C b.

Exercise. Prove that if the solution to V = g belongs to H0() we can bound

h1/2, . h1/20,.

It is also possible to obtain error bounds of the numerical solution in stronger norms.For instance, it is possible to prove a bound

h0, . h1,.

Nevertheless, this bound needs from more stringent conditions on the triangulation. If allthe elements are triangles and the size of the largest triangle is asymptotically controlledby that of the smallest one

max{hj} . min{hj}then the space Xh satisfies an inverse inequality

h0, . h1/2h1/2,, h Xh,

which gives a bound on the equivalence constant for the H0() and H1/2()normsrestricted to the discrete finitedimensional space Xh (where all norms are equivalent!).

31

The fact that the size of the smallest triangle controls that of the largest one is calledquasiuniformity.

Then, take to be the exact solution, h the numerical solution and h the H0()

orthogonal projection of on Xh, and follow the chain of inequalities

h0, h0, + h h0,. h0, + h1/2h h1/2,. h0, + h1/2

( h1/2, + h1/2,

). h1,

that prove the desired convergence property. Notice that we have used the inverse inequal-ity (which follows from quasiuniformity, but could hold in more general cases) and againthe simultaneous optimal approximation properties of the H0()orthogonal projectionin two different norms.

Similar but more sophisticated arguments can be used to prove convergence in strongerSobolev norms, although the lack of regularity of the discrete space Xh imposes that wecannot arrive to measure errors in H1/2(), since this space does not contain piecewisesmooth discontinuous functions.

Another line of analysis goes to prove convergence in even weaker norms. To thefinite element oriented person, this can seem quite awkward, given the fact that H1/2(),the natural space in our analysis, is already a very weak space, endowed with a dualnorm. Well see in a while why weak norms are interesting. First, the bound. With someregularity requirements on the boundary, it is possible to prove that

h2, . h3/2 h1/2,,so in the optimal case of H1(),

h2, . h31,,which means that we have been able to double the order of convergence by looking ata much weaker norm. This result is a consequence of the socalled AubinNitschetechnique, which follows a very similar path as the one we use when we want to proveL2()convergence of finite elements for elliptic problems of the second order.

But, why can this inequality be interesting for us? Assume that the boundary issmooth enough to avoid unnecessary complications. Fix a point x +. Then thefunction

(x, ) : Cis very smooth (the singularity doesnt happen, since x is outside and the function moveson the boundary), so (x, ) C2() H2(). Hence

|(S)(x) (Sh)(x)| = |(x, ), h| (x, )2, h2,since the duality product , represents also the duality of H2() and H2(). There-fore if

u := S, uh := Sh

32

(these are the quantities of interest for the problem; is an artificial unknown), then

|u(x) uh(x)| . h31, (x, )2,The rightmost constant can be uniformly bounded on compact sets not intersecting theboundary, which gives a very high order of convergence for the computed potentials, aslong as we dont approach the boundary, since then the constant (x, )2, blowsup.

For bounds near the boundary, we only can use the natural norm in H1/2() for thedensity to obtain local H1 bounds in the exterior of order at most h3/2.

Remark. In boundary element computations you always have to keep in mind what you wantto compute (in fact, you should be doing this all the time you compute something) and no whatyou are computing. For us the density is not a quantity of particular interest. In a forthcomingsection we will see that in some boundary integral formulations the unknown has an interest ofits own. Right now, our interest is limited to the scattering amplitude u and its approximation.If you are not interested in what happens very near the boundary, the optimal convergenceorder h3 looks extremely promising for the method. This kind of weak norm estimates is veryparticular of Galerkin methods. For collocation, right now there is not a very satisfactory erroranalysis in any norm, and the duality argument to move to weaker norms does not apply soeasily (actually the method does not exhibit numerically such good properties when we lookat potentials). This makes the additional effort of the Galerkin method look more desirable,although the gain needs regularity in the boundary, so it isnt universal for all kind of scatterers.

If we now turn our attention to the matrix terms of the Galerkin methodi

j

(x,y)d(x) d(y)

there are some details that deserve to be commented:

The diagonal terms have a singularity all along the line x = y. If one appliesnumerical integration, one has to be very careful to avoid using the same rule in i forthe inner and outer integrals, since we cannot evaluate (x,x). Also, the integrandis not smooth and one cannot expect a very good behaviour of simple quadraturerules. Usually some transformations of variables are performed to improve the aspectof the integrand. It is also possible to do some kind of singularity extraction andperform part of the work analytically.

When two elements share a side, the singularity is reduced to both arguments x andy belonging to that side. When elements share a vertex, the singularity is limited toa single point of the fourdimensional domain ij. Again, numerical quadraturecannot ignore these singularities.

In other cases, the integrand is very smooth, but can be close to a singularity if theelements i and j are near each other.

33

In any case, there is a necessary balance between having a qualified numerical inte-gration procedure and not wasting a lot of computational effort in computing almostexactly some integrals, when we are approximating an integral equation with a loworder method.

Exercise. Let us now try another method. We also consider the partition in elements1, . . . ,N and a choice of points xj j (for instance barycenters if the elements are triangles).One can devise a method by approximating the integral by a quadrature rule where the densitiesact as weights:

( ,x)(x)d(x)

Nj=1

( ,xj)j .

The quantities j approximate the value of , taking into account the size of the element, whichwould give the correct weight if was known (which is not the case). Then, we compare thedata function g with this discrete potential generated by N sources by matching Galerkinwisetheir averages over the elements i:

Nj=1

(i

(y,xj)d(y))j =

i

g(y)d(y), i = 1, . . . , N.

This is called a delta or pointsource method.

(a) Compare the matrix and the righthand side of this method with those of the collocationmethod.

(b) In comparison with collocation, can you decide whether this method needs more, less orthe same regularity on the righthand side to be applied with property?

(c) Compare the potentials obtained by the Galerkin, collocation and delta method. Whichis simpler?

As we explained in Section 5, assume that our surface is given by a series of patchesfrom a polyhedron and that we have collected the transformations from each face of thepolyhedron to the corresponding curved patch in a single map

: .

With some degree of notational freedom we can consider integrals over as integrals withrespect to the twodimensional Lebesgue measure, so we will write

f()d

in the understanding that we are parameterizing each face of by simply moving (rotatingand translating) it to the plane R2. The area element in can be defined from andthus

f(x)d(x) =

f(()) () d, := |1 2|

34

where the same convention as to what we understand by .Then, the elements in , which we have called {1, . . . ,N} are the images through

of elements in , {1, . . . , N}.Exercise. Prove that the space we used in the Galerkin method can be written equivalentlyas

Xh := {h : C | (h )|j P0, j}.Consider instead a new Galerkin method with the following discrete space

Yh := {h : C | (h )|j P0, j}.

Write down what is the formula for the elements of the coefficient matrix and righthand side

vector of this new method. Devise a collocation method in the same spirit.

Higher order methods can be constructed by mapping higher order polynomials on theflat elements j to the curved boundary . Since the method does not require continuityof the discrete space, one can go further employing discontinuous elements. However, thecomputational effort to increase order is much higher (the number of degrees of freedomincreases very fast) and it is customary to employ continuous piecewise polynomials toreduce the number of degrees of freedom, while retaining the higher order approximationproperties of the method.

9 A word on the twodimensional case

Most of what we have done so far works also in two dimensions. Now, boundaries arecurves, which are much simpler to handle that surfaces. The fundamental solution is

(x,y) :=

4H

(1)0 (k |x y|).

The function H(1)0 is defined in terms of Bessel functions as

H(1)0 (r) = J0(r) + Y0(r)

(J0 is the Bessel function of order zero; Y0 is the Bessel function of the second kind andorder zero, also called Neumann function of order zero) and called Hankel function ofthe first kind and order zero. As r 0, H(1)(r) has a logarithmic singularity, whichis also integrable.

The Sommerfeld condition is now

limr

r1/2 (ru k u) = 0.

One defines similarly singlelayer potentials, the operator V and proves the equivalentproperties. For the decomposition

V = V0 + compact

35

we have to use

V0 := 12pi

log |( y)|(y)d(y)

(d( ) is a length element, not area!) with < 1/diam(). The reason for this newparameter related to size is not entirely obvious but has a lot to do with the essentialdifference of electrostatics in two and three dimensions.

If the boundary is smooth and we have a smooth parameterization of the whole curve(not a collection of pieces of parameterizations), we can very easily use high order elementswithout increasing the number of degrees of freedom, such as splines and trigonometricpolynomials.

10 A direct formulation

At this point we are going to restart and work with new formulations, to obtain newintegral equations, which we will have to numerically solve and then (when possible)analyse the discrete method. So, hold your breadth ... Part of what we are going to doincludes new concepts, but a lot of it is going to be a more or less step by step repetitionof already studied techniques.

The departure point for our new formulation is the third Green formula adapted tothe Helmholtz equation. Assume that u is the solution of an exterior scattering problem(for which we dont give the boundary condition yet):

u+ k2u = 0, in +,

ru k u = o(r1), r .

Then,

u(x) =

(x,y)u(y)d(y) +

(y)(x,y)u(y) d(y), x +.

The formula holds with ordinary integrals if u is smooth enough near the boundary andwith a duality product for the first term if u is simply locally H1. The formula says a lotof things. Among them:

It says that the solution is C(+), and that lack of regularity is limited to theproximities to the boundary of the scatterer. We already had this from the indirectformulation, excepting at Dirichlet eigenfrequencies.

It also says that if we know both Cauchy data (Dirichlet and Neumann data), thesolution is determined via an explicit formula. Taking this point of view, we saythat this is a representation formula.

If we take the limit to x we find something new: if x is a smooth point of , then1

2u(x) =

(x,y)u(y)d(y) +

(y)(x,y) u(y) d(y).

36

This means that if we evaluate verbatim the righthand side formula we obtain u(x) whenx +, but only 1

2u(x) when x . Moreover,

0 =

(x,y)u(x)d(y) +

(y)(x,y) u(y) d(y), x .

Remark. The formula holds also true, but with opposite signs, for interior problems. If

u+ k2u = 0, in ,

thenu(x) =

(x,y)u(y)d(y)

(y)(x,y)u(y) d(y), x .

Notice that the changes of signs makes sense since we are keeping the sense of the normalvectors (always pointing inside +), which means that they are now pointing outwards whereasbefore they where pointing inwards, given the fact that the domain was +. Notice also thatthis formula doesnt contradict the last one for the exterior problem: the Cauchy data in thatformula where for the exterior problem and here they are for the interior problem.

Exercise. Using the representation formula for the interior problem, prove that a functioncannot be simultaneously a Dirichlet and a Neumann eigenfunction for the Laplace operator in

. (Notice that we needed this to prove that invertibility of the singlelayer operator V wasequivalent to k2 not being a Dirichlet eigenvalue.)

Let us simplify the aspect of all these formulas using a new potential

(D)(x) :=

(y)(x,y)(y) d(y), y + = R3\

and a new operator

(K)(x) :=

(y)(x,y)(y) d(y), y .

As before, potential and operator are given by the same formula, only the potential is afunction defined in R3\ and the operator gives a function defined on .

Remark. The potential D is called a doublelayer potential for reasons we will explorein the following section.

If again

u+ k2u = 0, in +,

ru k u = o(r1), r ,

and we denoteg := u|, := + u|,

37

then the representation formula says that

u = Dg S, in +.

I want to emphasize that the Sommerfeld condition has to be satisfied for this to be true,since the righthand satisfies it even if g and are not the Cauchy data of a solution ofthe Helmholtz equation. The limiting value of this formula is

12g = Kg V.

For the Dirichlet problem (soundsoft scattering), we can try to solve the equation

V = 12g +Kg

and then obtain u by means of the representation formula. I dont waste your time, solets give fast facts. Bullet points!

The equation has the same structure as that of the indirect method. The righthandside is more complicated though.

If k2 is not a Dirichlet eigenvalue in , we have a unique solution of the problem.We can solve it with the same Galerkin method as before.

Now the unknown is a physical quantity, not a density we have devised to solve theproblem. Notice that if you know u| and u|, you also have a first order Taylorknowledge of u near the boundary.

The exterior Dirichlet problem admits a unique solution. This fact can be provedindependently of the value of k. Therefore, the integral equation always has asolution, but it can fail to be unique when V is not invertible. The nullspace (alsocalled kernel) of V is the set of normal derivatives of the eigenfunctions

+ k2 = 0, in ,

= 0, on

Since (see below to check the formulas for interior problems)

0 = S D = S,

any solution of the integral equation works from the point of view of the representa-tion formula. The problem is that only one of these solutions is the normal derivativeof the solution, so if you are after the correct value of the normal derivative, youhave to be extra careful here!

Exercise. Using the same Galerkin method as exposed in Section 5, what is the form of therighthand side? And for the collocation method?

38

Exercise. Write down the boundary integral equation you would have to solve for thesoundhard scattering problem (Neumann problem):

u+ k2u = 0, in +,ru k u = o(r1), r ,

u| = .

For the interior problem, we have the same type of formulas: if

u+ k2u = 0, in

and g := u|, := u|, then we have the representation formulau = S Dg, in ,

the identity satisfied by Cauchy data

12g = V Kg

and0 = S Dg, in +.

Exercise. Write down the integral equations you would have to solve for the interior Dirichletand Neumann problems.

Exercise. Let

u+ k2u = 0, in + ,ru k u = o(r1), r .

Prove thatu = S( u + u)D(u u+), in + .

(Hint: use the representation formulas for the interior and exterior problems and also the iden-

tities satisfied by the righthand side in the complementary domain). What is the value of the

function defined by the righthand side when restricted to the boundary ?

11 Soundhard scattering

We end this course by showing more formulas and more techniques to numerically solveexterior scattering problems via boundary elements. Our new problem is the exteriorNeumann problem,

u+ k2u = 0, in +,

ru k u = o(r1), r ,u| =

39

for given H1/2(). A proof of existence and uniqueness of this problem can begiven by following similar steps to the Dirichlet problem: we prove uniqueness of thehomogeneous problem ( = 0); we propose a potential solution and obtain a relatedboundary integral equation and, finally, we use this equation and Fredholm theory toprove existence of solution.

We will give it several tries: an indirect method with a doublelayer potential; anindirect method with a single layer potential and a direct method.

11.1 Doublelayer potentials

The first possibility we explore is trying

u := D, H1/2().

The following exercise proves that in the boundary

+(D) = 12+K(D) = 12+K+ (D) = (D).

The last expression gives rise to a boundary integral operator

W := (D)

or, written more explicitly

W :=

(y)( ,y)(y) d(y).

The negative sign in front of the definition looks quite arbitrary. It is imposed to ensurea certain degree of positiveness, as we will see in a while.

Exercise. Take an arbitrary g H1/2(), solve the exterior Dirichlet problem and call := u. The representation formulas of Section 10 read then:

u = S+Dg, in +12g = V+Kg, on 0 = S+Dg, in .

(a) By writing Dg = u+ S in +, prove that

+(Dg) = 12g +Kg.

(b) Prove that(Dg) = V = 12g +Kg.

What is the discontinuity of the doublelayer potential across ?

40

(c) Using the identities

= + (S) + + (Dg),0 = (S) + (Dg), = (S) + (S),

(the last one was mentioned and used when studying the singlelayer potential) prove that

+ Dg = Dg

and therefore the normal derivative of the doublelayer potential is continuous.

Remark. We already mentioned that singlelayer acoustic potentials extend to the Helmholtzequation (timeharmonic acoustic fields) the idea of electrostatic charges in equilibrium. Simi-larly, doublelayer potentials extend the idea of dipoles. Now the potential is discontinuous (thedipole distribution on the surface creates the discontinuity) but the electric field is continuousin the normal direction across the boundary.

It is very tempting to introduce the derivative under integral sign and write somethinglike this

(W)(x) =

(x)(y)(x,y)(y) d(y).

But you have to be careful with this. By doing the two normal derivatives, the kernelfunction

(x)(y)(x,y)

fails to be integrable. In order that the integral makes full sense, we have to subtract thesingularity and integrate the regular part of the integrand. What we really do there is theHadamard finite part of the integral. Independently of how this is computed, you write

W = f.p.

( )(y)( ,y)(y) d(y).

The important points about W are:

W is bounded from H1/2() to H1/2(). W is invertible if and only if k2 is not a Neumann eigenvalue of the Laplaceoperator in .

There exists an operator W0 such that

,W0 21/2,, H1/2()

and W W0 is compact. Here is where the minus sign in the definition becomesrelevant.

41

Because of the last property, in the case where W is invertible, a Galerkin methodwith a sequence of subspaces of Yh H1/2() that satisfy

infhYh

h1/2, h0 0, H1/2()

defines a convergent method. This means that the method defined by

h Yh, h,Wh = h, , h Yhsatisfies

h1/2, h0 0. A piecewise smooth function defined on a partition of into elements has to becontinuous to be an element of H1/2(). Therefore, boundary elements for thisproblem have the same continuity requirements as the typical finite elements forelliptic problems of the second order.

For instance if we use a space of continuous piecewise linear elements (the classicalCourant finite elements) we obtain, with regularity assumptions on the solution, the bound

h1/2, . h3/22,.Again, a duality argument (an more assumptions on the regularity of the domain), canbe used to improve this bound in a weaker norm to obtain

h1, . h32,.This is the optimal bound we observe if we look at pointwise values of the approximatedpotential sufficiently far from the boundary.

11.2 Single layers

Another option for an indirect formulation is trying again with singlelayer potentials:

u = S, H1/2().To do that, we have first to know what is the normal derivative of a single layer potential,which is going to be different from the inside to the outside, since its difference is preciselythe density . With some effort (not a lot) it can be proved that

+ u = 1

2 +

( )( ,y)(y) d(y), on .

Defining

(Kt)(x) :=

(x)(x,y)(y) d(y),

the expression simplifies to+ S = 12 +Kt.

42

Therefore, to solve the Neumann problem, we only have to try to find a solution to thenew boundary integral equation

12 +Kt =

and then input in the singlelayer potential u = S to obtain the scattered wave field.Exercise. Prove that

S = 12 +Kt.

The operators K and Kt are transposed to each other, meaning that

Kt, = K, , H1/2(), H1/2().

Remark. K and Kt are not adjoint to each other, but transposed. The only differencebetween these two concepts is conjugation in the complex plane.

Remark. In an exercise in the previous section, you will have arrived to a very similar integralequation

12 g +Kg = Vfor the Neumann problem. In this equation the unknown is the Dirichlet datum (this is a directmethod; no potentials are involved), we have K instead of Kt and an integral operator in therighthand side. Independently on whether these integral equations are uniquely solvable or not(we will not deal with this now), the direct formulation has always a solution, even if it is notunique.

With some regularity assumptions, that preclude polyhedra of the setting, it is possibleto prove that

K : H1/2() H1/2() is compact

and therefore

Kt : H1/2() H1/2() is compact.

This means that equations with operator of the form

12I +Kt

satisfy always the Fredholm alternative. There is however a somewhat difficult pointhere when we want to do numerical approximation. The question is ellipticity: trying aninequality like

I , 21/2,

43

is not possible, since in the bracket one component has to be in H1/2() and the other inH1/2(). A solution would be using the inner product of H1/2(), where ellipticity istrivial

(I , )1/2, = 21/2,.However, the numerical approximation by a Galerkin scheme requires then using thevariational formulation

H1/2(), (12 +Kt, )1/2, = (, )1/2,, H1/2()

as a starting point and the inner product in H1/2() is not easy to handle.The simplest thing to do is trying a Galerkin scheme in the spirit of what we did in

previous sections. We define Xh to be the space of piecewise constant functions and lookfor

h Xh, 12

h h +

(Kth)h =

h, h Xh.

For this we need the additional assumption that the Neumann datum belongs to L2(),which is not true for the Neumann problem in full generality, but certainly is for thesoundhard scattering problem. This is tantamount to moving the integral equation

12 +Kt =

to be happening in H0() instead of in H1/2(). Here the identity is elliptic withrespect to the H0()inner product and therefore, assuming unique solvability, we obtainasymptotic stability plus the convergence bound

h0, . infhXh

h0,,

which, when H1(), yields the convergence rate

h0, . h 1,.

This rate can be improved a little in weaker norms, but is still far worse than the kindof bounds we obtained for the Dirichlet problem with singlelayer potentials or for theNeumann problem with doublelayer potentials. It must be said, however, that in thelast case, continuity was a requirement on the space, since the density of a double layeroperator belongs to H1/2(), where line discontinuities are forbidden.

Remark. The integral equation

12 +

( )( ,y)(y) d(y) = , on

(this would be for an interior Neumann problem) is basically where everything begun for integralequations. In fact, the operator was the Laplace operator and the problem was being set in twodimensions. The name of Ivar Fredholm is strongly related to this effort of dealing with boundaryintegral equations of the second kind (identity plus compact operator), at a moment where allthe functional analysis was still in diapers. The effort of Fredholm, as well as that of Hilbert, the

44

Riesz brothers, Schauder and many others, on developing abstract mathematical tools to dealwith linear operators in abstract spaces was originated in a big part from the effort of solving(showing that it has a solution and trying to get it in some cases) this integral equation as away to prove existence of solution of boundary value problems when the nice Sobolevspaceanddistributiontheory was still decades away.

11.3 Direct formulations

We have already mentioned the possibility of using the representation formula

u = Dg S, in +and the integral equation

12g +K g = V

as a way of solving the exterior Neumann problem. This problem should be solved inH1/2() but, as before, the principal part of the integral equation (minus one half timesthe identity operator) is elliptic in H0(). One can devise then a Galerkin method similarto the one for the singlelayer representation, but using a H1/2conforming boundary ele-ment space, i.e., one with continuity conditions over interfaces of elements. The analyticaltechniques developed in Section 8 can be used here to show convergence in H1/2() pro-vided that we can use some inverse inequalities, which hold under some severe restrictionson the triangulation.

If one is not that much interested in convergence of the Galerkin approximation of gto this function in the correct norm (the trace norm H1/2()), but only in the exteriorsolution sufficiently far from the obstacle, it is possible to benefit from AubinNitscheestimates in weak norms, which provide improved convergence rates.

Finally, since we know what the normal derivatives of single and double layer operatorslook like, we can go back to the representation formula

u = Dg S, in +and take the normal derivative from the exterior

= + u = Dg + S = Wg (12+Kt)to obtain the integral equation

Wg = 12Kt,which is always solvable but fails to be uniquely solvable when k2 is a Neumann eigen-value of the Laplace operator in the interior domain. The same kind of numerical tech-niques than those used for the doublelayer representation apply here.

Remark. Ive deliberately omitted any mention to collocation methods in this section. Ofcourse they also apply to this situation, at least formally. In the case of equations of the secondkind (for sufficiently smooth obstacles) it is even possible to do a proper analysis of the method,but this analysis takes place more naturally in Holder spaces than in Sobolev spaces. When thehypersingular operator W appears in the equation, things become trickier from the analyticalpoint of view and there is still a lot to be done.

45

12 Further reading

There are not many books dealing with the mathematical and numerical analysis of bound-ary element methods. The literature is however much more important in the engineeringworld, where you will be able to find many details on algorithms, implementation andespecially different problems where boundary element techniques apply.

The book

Boundary Element Methods by Goong Chen, Jingmin Zhou, Academic Press,1992

details the boundary integral formulations for the Laplace, Helmholtz, NavierLame (lin-ear elasticity) and biharmonic (Kirchhoff plate) equations. The Sobolev theory is ex-plained with care in the case of smooth interfaces. The fundamentals of Sobolev theoryand finite element theory are carefully explained. There are also some explanations onpseudodifferential operators, a theory that allows for a generalization of the behaviourof all the boundary integral operators for smooth boundaries. The section on numericalanalysis is not very long and right now it is not uptodate.

The whole theory on boundary integral formulations based on elliptic operators (allthe equations where the principal part is V or W) is explained with an immense careand taste for mathematical detail in

Strongly Elliptic Systems and Boundary Integral Equations byWilliamMcLean,Cambridge University Press, 2000.

This is a book of hard mathematics, where you will learn a lot but are asked to havepatience. It does not cover numerical analysis.

If you are not afraid of reading maths in German, this is a very good choice:

Randelementmethoden. Analyse, Numerik und Implementierung schneller Al-gorithmen by Stefan Sauter and Christoph Schwab, Teubner, 2004.

The book covers theoretical aspects, numerical analysis as well as many details on howto approximate the integrals of Galerkin discretizations and how to solve efficiently thelarge dense linear systems.

Three other books, with an increasing slope towards applications:

A Practical Guide to Boundary Element Methods with the Software LibraryBEMLIB, by C. Pozrikidis, CRC Press, 2002.

Programming the Boundary Element Method : An Introduction for Engineers,by Gernot Beer, John Wiley & Sons, 2001.

Boundary Element Methods for Engineers and Scientists, by Lothar Gaul,Martin Kogl, Marcus Wagner, Springer, 2003

46