An Analytical Method For Linear Elliptic Pdes and Its Numerical Implementation

Available online at www.sciencedirect.
com
Journal of Computational and Applied Mathematics 167 (2004) 465483
www.elsevier.com/locate/cam
An analytical method for linear elliptic PDEs and its
numerical implementation
S.R. Fulton
, A.S. Fokas
1
, C.A. Xenophontos
2
Department of Mathematics and Computer Science, Clarkson University, Potsdam, NY 13699-5815, USA
Received 29 October 2001; received in revised form 17 September 2003
Abstract
A new numerical method for solving linear elliptic boundary value problems with constant coecients
in a polygonal domain is introduced. This method produces a generalized DirichletNeumann map: given
the derivative of the solution along a direction of an arbitrary angle to the boundary, the derivative of the
solution perpendicular to this direction is computed without solving on the interior of the domain. If desired,
the solution on the interior can then be computed via an integral representation.
The key to the method is a global condition which couples known and unknown components of the
derivative on the boundary and which is valid for all values of a complex parameter k. This condition has
been solved recently analytically for several equations on simple domains. In this paper, rst the previous
analytical result is strengthened, and then a numerical method is introduced for solving the global condition for
the Laplace equation on an arbitrary bounded convex polygon. Numerical results demonstrate the applicability
and convergence of the method; however, a rigorous proof of convergence remains open. Extensions to other
problems are also discussed.
c _ 2003 Elsevier B.V. All rights reserved.
MSC: 35J25; 65N35; 65N38; 65N99
Keywords: Elliptic partial dierential equations; DirichletNeumann map; Global condition; Spectral method
Corresponding author.
E-mail addresses: [email protected] (S.R. Fulton), [email protected] (A.S. Fokas),
[email protected] (C.A. Xenophontos).
1
Permanent address: Department of Applied Mathematics and Theoretical Physics, University of Cambridge, Cambridge
CB3 9EW, UK.
2
Current address: Department of Mathematical Sciences, Loyola College, Baltimore, MD 21210-2699, USA.
0377-0427/$ - see front matter c 2003 Elsevier B.V. All rights reserved.
doi:10.1016/j.cam.2003.10.012
466 S.R. Fulton et al. / Journal of Computational and Applied Mathematics 167 (2004) 465483
1. Introduction
Solving boundary value problems for partial dierential equations (PDEs) is a central topic in
applied mathematics. Analytical methods (e.g., separation of variables and transform techniques) are
valued for their exactness and the insight they provide; however, the range of problems they solve
is limited. Numerical methods (e.g., nite element, nite dierence, and spectral methods) solve a
much wider range of problems, albeit only approximately. Some methods (e.g., boundary integral
methods) combine specic analytical information about the solution with numerical approximations.
This paper describes a new method which ts into this latter class.
Recently, a new analytical method for studying boundary value problems for integrable PDEs in
two dimensions has been introduced in the literature, see for example [1,2]. This method has been
applied to linear elliptic PDEs in convex polygonal domains [2], yielding analytical solutions in
cases not amenable to treatment by standard transform methods. A key to this method is the global
condition [cf. (7)] which couples specied and unknown values of the solution or its derivatives
on the boundary. In cases where this condition can be solved analytically, the method yields the
solution in closed form, generalizing standard transform methods (e.g., Fourier, Bessel, etc.).
In this paper, we extend this method in two ways. First, we strengthen the analytical results of
[2]: we show that the global condition is not only a necessary but also a sucient condition for
existence. This reduces the problem of solving Laplaces equation to the problem of solving the
global condition. Second, we introduce an approach to solving the global condition numerically. The
result is a new method for solving PDEs in two dimensions which couples the analytical information
obtained by the method of [2] with the numerical solution of a one-dimensional problem. Applied to
the Laplace problem on a convex polygonal domain, the method provides a generalized Dirichlet
Neumann map: given the derivative of the solution in the direction of some (arbitrary) angle to the
boundary, the method yields the derivative of the solution perpendicular to this direction without
solving on the interior of the domain.
This paper is organized as follows: Section 2 outlines the analytical method of [2] and gives
an overview of the numerical implementation introduced here. In that section, we also discuss the
relation of our method to a boundary integral method (the boundary element method). In Section
3 we extend the analytical result of [2], proving (as mentioned above) that solving the Laplace
equation is equivalent to solving the global condition. The main improvement as compared with [2]
is that now the result is proven without assuming the existence of the solution. Section 4 gives
the details of the numerical solution of the global condition. In Section 5 we apply the method
to the Laplace equation in a variety of domains with dierent boundary conditions and present
numerical results demonstrating its convergence. Section 6 summarizes our conclusions and discusses
extensions.
2. Overview
In this section, we review the analytical method of [2], outline the numerical solution of the
resulting global condition, and compare our approach with boundary integral methods, with which
it shares some similarities.
S.R. Fulton et al. / Journal of Computational and Applied Mathematics 167 (2004) 465483 467
2.1. The method of Fokas
The method introduced by Fokas [1,2] solves integrable PDEs in two dimensions. An equation in
two dimensions (x, ,) is called integrable if and only if it can be expressed as the condition that a
certain associated dierential 1-form W(x, ,, k), k C, is closed, i.e., dW =0. Examples of integrable
equations are linear PDEs with constant coecients and the usual integrable nonlinear PDEs such
as the nonlinear Schr ondinger and the Kortewegde Vries equations. In what follows, we formulate
the global condition without using the language of dierential forms.
For elliptic equations it is convenient to replace the usual Cartesian coordinates (x, ,) with the
complex coordinates (:, :) = (x + i,, x i,). For example, the Helmholtz equation
q
xx
+ q
,,
+ 4:q = 0, : constant, (1)
may be written, using q
:
=
1
2
(q
x
iq
,
) and q
:
=
1
2
(q
x
+ iq
,
), as
q
: :
+ :q = 0. (2)
This equation can be rewritten in the form
(e
ik:i: :}k
q
:
)
:
+
i:
k
(e
ik:i: :}k
q)
:
= 0, k C. (3)
It is emphasized that (2) is equivalent to (3) for an arbitrary complex parameter k.
Suppose that q(:, :) satises (2) in a simply connected bounded domain D with boundary 9D.
Then the complex form of Greens theorem implies
_
9D
W(:, :, k) = 0, k C, (4)
where
W(:, :, k) = e
ik:i: :}k
_
q
:
d:
i:
k
q d :
_
. (5)
Likewise, the Laplace equation, i.e., (2) with : = 0, is equivalent to
(e
ik:
q
:
)
:
= 0, k C (6)
so in analogy with (4) we now nd
_
9D
e
ik:
q
:
d: = 0, k C. (7)
Following [2] we shall refer to (4) and (7) as the global conditions associated with the Helmholtz
and Laplace equations, respectively.
For example, suppose that q(:, :) satises the Laplace equation in a convex bounded polygon
with vertices :
1
, :
2
, . . . , :
n
(indexed counterclockwise, modulo n) and interior D as in Fig. 1. Then
the global condition (7) becomes
n
)=1
j
)
(k) = 0, k C, (8)
Fig. 1. Part of the bounded convex polygon with vertices :
)
, sides S
)
, and interior D.
where the functions j
)
(k) are dened by the line integrals
j
)
(k) =
_
S
)
e
ik:
q
:
(:) d:, k C, ) = 1, . . . , n (9)
with S
)
being the side from :
)
to :
)+1
(not including the endpoints).
It was shown in [2] that the global condition plays a crucial role in the analysis of boundary
value problems. For example, consider the Neumann problem for the Laplace equation in the above
polygon. Let q
( ))
s
and q
( ))
n
denote the tangential and (outward) normal components of q
:
along the
side S
)
. Then on this side
q
:
=
1
2
e
i:
)
(q
( ))
s
+ iq
( ))
n
), :
)
= arg(:
)+1
:
)
). (10)
Substituting (10) into (9), the global condition becomes
n
)=1
e
i:
)
)
(k) = G(k), k C, (11)
where
)
(k) denotes the unknown line integral
)
(k) =
_
S
)
e
ik:
q
( ))
s
(:) d:, k C, ) = 1, . . . , n (12)
and G(k) can be computed in terms of the given boundary data q
( ))
n
. Eq. (11) is only one equation
for the n unknown functions
)
(k). In spite of this ominous-looking situation, it is possible using the
global condition (11) to determine all the unknown functions
)
(k). This is a consequence of the
fact that (11) is valid for all complex values k. The analytical investigation of the global condition
is discussed in [2] in general, and in [3] for the Laplace equation in particular. One of the main
goals of this paper is to introduce a numerical algorithm for solving the global condition.
Eq. (11) indicates that the global condition determines the integrals
)
(k) and not the functions
q
( ))
s
. This suggests that it would be desirable to express q(:, :) in terms of j
)
(k) and not in terms
of the boundary values of q. For linear PDEs such formulae have recently been derived using the
spectral method introduced in [2]. For example, for the Laplace equation the following result from
[2] is valid:
Proposition 1. Consider the Laplace equation in a convex bounded polygon with vertices :
1
,
:
2
, . . . , :
n
(indexed counterclockwise, modulo n) and interior D. Assume that appropriate boundary
conditions are prescribed such that there exists a solution q(:, :) which is smooth all the way to
the boundary. Then
q
:
=
1
2
n
)=1
_
)
e
ik:
j
)
(k) dk, : D, (13)
where j
)
(k) are the functions dened by (9) and
)
are the rays in the complex k-plane oriented
away from the origin dened by
)
=k C: arg(k) =arg(:
)+1
:
)
). (14)
In Section 2 we will prove a stronger version of this proposition (cf. Proposition 2). This result
reduces the solution of a given boundary value problem for the Laplace equation to the following
problem: Use the global condition (8) to determine j
)
(k) in terms of the given boundary data.
In this paper, we will study the Laplace equation for an arbitrary convex bounded polygon with
an arbitrary component of the derivative specied on each side. Specically, on each side S
)
we will
specify the derivative in the direction given by the angle o
)
relative to the positive real axis (angle
[
)
= :
)
o
)
measured outward from the side S
)
) leading to the mixed boundary condition
cos([
)
)q
( ))
s
+ sin([
)
)q
( ))
n
= q
())
, : S
)
, (15)
where q
())
is a given smooth function. Dirichlet and Neumann conditions correspond to the special
cases [ = 0 and }2, respectively. For this problem the relevant unknown is the derivative in the
direction normal to o
)
, i.e., the function [
())
dened by
sin([
)
)q
( ))
s
+ cos([
)
)q
( ))
n
= [
())
, : S
)
. (16)
Solving (15) and (16) for q
( ))
s
and q
( ))
n
and substituting into (10) yields
q
( ))
:
=
1
2
e
io
)
(q
())
+ i[
())
). (17)
Replacing q
:
in (9) by the above expression, it follows that j
)
(k) involves the unknown integral
)
(k) =
_
S
)
e
ik:
[
())
(:) d:. (18)
2.2. Numerical solution
In order to determine these unknown integrals, we use a collocation projection of the global
condition in the complex k-plane (see Section 4.4):
(1) For each side S
)
, set F
())
= [
())
[
( ))
, where the linear function [

( ))
is chosen so that F
())
vanishes at the endpoints :
)
and :
)+1
. This is made possible by using the continuity of q
:
at
the n vertices :
1
, . . . , :
n
to determine the values of [
())
at the endpoints.
(2) Approximate F
())
by
F
( ))
N
(s) =
N
r=1
c
( ))
r

r
(s), (19)
where s is a local parameter along the side S
)
,
r
N
r=1
are appropriate basis functions, and
N is even. This approximation introduces n N unknown (real) constants c
( ))
r
, 1 6r 6N,
1 6) 6n.
(3) Evaluate the global condition (8) at M = N}2 collocation points on each ray

)
, where

)
is
the continuation of
)
, i.e.,
)
=k C: arg(k) = :
)
, ) = 1, . . . , n. (20)
The reason for this choice of the rays

)
is explained in Section 4.2. This yields n M
(complex) linear equations.
(4) Solve the resulting linear system to obtain the constants c
( ))
r
, which in turn yield an approxi-
mation to [
())
(s) and thus j
)
(k), ) = 1, . . . , n.
Note that while this method uses a collocation projection and a spectral representation, it is not a
typical spectral collocation method: the collocation takes place in the complex k-plane, and the
spectral representation used refers not to the basis functions in the numerical approximation but to
the underlying analytical representation.
2.3. Discussion
Since the method treated here shares some similarities with boundary integral methods, it is ap-
propriate to discuss their relation. We will restrict our comments to the boundary element method
(BEM) [8], which is a well-established method for the numerical solution of boundary value prob-
lems; its applicability and underlying theory have been (and still are) studied extensively (see, e.g.,
[10] and the references therein). When using the BEM to approximate the solution to elliptic bound-
ary value problems, one starts with a fundamental solution and converts the given PDE into an
integral equation posed on the boundary of the domain. The resulting equation is then discretized
and solved numerically. This procedure reduces the dimension of the problem by one, hence keeping
the computational cost low. The use of a fundamental solution can be viewed as a disadvantage,
since its availability and/or simplicity is not always guaranteed. As a result, some researchers have
combined the BEM with other methods to bypass this step, while still solving a problem on the
boundary of the domain (e.g., [6,11]).
Like the BEM, the method presented in this paper combines analytical information with a numer-
ical approximation and reduces the numerical work to solving a one-dimensional problem posed on
the boundary of the domain. However, this is the only feature these methods share. In the present
method, the one-dimensional problem to be solved (the global condition) comes from a dierent
source and is solved dierently. Furthermore, the resulting functions j
)
(k) then provide the solu-
tion of the original PDE in a dierent form (the spectral representation (13)). Also, while strictly
DirichletNeumann maps have been studied in the context of the BEM (e.g., [4,9]), we believe that
the ability of the present method to produce automatically a generalized DirichletNeumann map is
not present in the BEM, nor in boundary integral methods in general.
3. Spectral solution of the Laplace equation
Here, we state and prove a stronger version of Proposition 1 for the solution of the Laplace
equation. As above, we let :
1
, . . . , :
n
denote the vertices of a convex bounded polygon in the complex
plane (indexed counterclockwise, modulo n) with interior D; S
)
denotes the side from :
)+1
to :
)
(not including the endpoints), and :
)
:= arg(:
)+1
:
)
) denotes the angle between side S
)
and the
positive real axis. Note that the boundary 9D of D consists of S :=

n
)=1
S
)
, together with the
vertices :
1
, . . . , :
n
. The following proposition allows for the case of singularities at the vertices, and
establishes (rather than assumes) the existence of the solution.
Proposition 2. For each ) = 1, . . . , n let r
())
H
1}2+c
(S
)
) for c 0 with r
())
(:
)+1
) = r
()+1)
(:
)+1
) and
dene j
)
(k) by the line integral
j
)
(k) =
_
S
)
e
ik:
r
())
(:) d:, k C (21)
along that side. Assume that the functions j
)
satisfy the global condition (8). Then the function
r(:) :=
1
2
n
)=1
_
)
e
ik:
j
)
(k) dk (22)
and its antiderivative q(:) are continuous on DS and analytic on D, Re(q) satises the Laplace
equation on D, and on each side S
)
() = 1, . . . , n), q
:
= r = r
())
.
Thus, solving the Laplace equation is equivalent to solving the global condition: given appropriate
boundary data, amounting to half of q
:
, if one can nd the other half by the requirement that
the functions j
)
dened by (21) satisfy the global condition, then the function r(:) dened by (22)
solves the Laplace equation and satises the given boundary conditions.
The key to the proof is the fact that for certain values of : and k, the integrand in (22) decays
exponentially as [k[ . More precisely, we have:
Lemma 1. For k
)
and : (D S) S
)
, the function
e
ik:
j
)
(k) =
_
S
)
e
ik(::
)
r
())
(:
) d:
(23)
decays exponentially as [k[ .
Proof. By convexity (see Fig. 2),
:
)
arg(: :
)
) arg(: :
) arg(: :
)+1
) :
)
+ , (24)
so there exists some c (0, }2) such that
:
)
+ c 6arg(: :
) 6:
)
+ c :
S
)
. (25)
Fig. 2. Geometry for Lemma 1.
Since arg(k) =:
)
, we have
c 6arg[k(: :
)] 6 c :
S
)
. (26)
Thus, Im[k(: :
)] [k[p sin(c), where p := inf

:
S
)
[: :
[ 0, hence
[e
ik(::
)
[ 6e
|k|p sin(c)
:
S
)
. (27)
This gives the bound
[e
ik:
j
)
(k)[ =
_
S
)
e
ik(::
)
r
())
(:
) d:
6e
|k|p sin(c)
|r
())
|
L
1
(S
)
)
. (28)
Likewise, for : on a side other than S
)
, the exponential decay of e
ik:
j
)
(k) between two associated
rays allows us to change paths of integration as follows:
Lemma 2. If : S
with ,= ), then
_
e
ik:
j
)
(k) dk =
_
)
e
ik:
j
)
(k) dk. (29)
Proof. When [:
)
:
[ = (i.e., sides S
)
and S
are parallel) there is nothing to prove, since the

rays
)
and

coincide. Thus, we consider the case 0 :
:
)
as shown in Fig. 3(a); the
case 0 :
)
:
can be treated similarly. Note that if ,= ) + 1 then the integrand decays

exponentially for all k on and between the rays
)
and

and the proof is straightforward. Therefore,

Fig. 3(a) depicts the case = ) + 1, which requires the more involved argument outlined below.
Consider the contour C
1
C
2
C
3
C
4
in the complex k-plane as shown in Fig. 3(b), where
c (0, :
+:
)
). Since from (21) and the Schwarz inequality j
)
(k) is an entire function of k, by
the CauchyGoursat theorem
_
C
1
e
ik:
j
)
(k) dk =
_
C
2
e
ik:
j
)
(k) dk +
_
C
3
e
ik:
j
)
(k) dk +
_
C
4
e
ik:
j
)
(k) dk. (30)
In the limit as R , (30) reduces to (29) provided that the last two integrals vanish. To bound
the integral along C
3
, note that the integrand satises a uniform bound of the form (28) as in
Lemma 1, with [k[ = R for k on C
3
. Thus, the integral along C
3
vanishes in the limit as R .
Fig. 3. Geometry for Lemma 2 in (a) the :-plane and (b) the k-plane.
To bound the integral along C
4
, we write
_
C
4
e
ik:
j
)
(k) dk =
_
C
4
e
ik(::
)+1
)
q(k) dk, (31)
where
q(k) := e
ik:
)+1
j
)
(k) =
_
S
)
e
ik(:
)+1
:
)
r
())
(:
) d:
. (32)
For c suciently small we can establish the uniform bound
[q(k)[ 6B
R
:= |e
R|:
)+1
:
| sin(c)
|
L
2
(S
)
)
|r
())
|
L
2
(S
)
)
(33)
and then (as in the proof of the Jordan Lemma) we have
_
C
4
e
ik:
j
)
(k) dk
6
B
R
r
(1 e
rR
), (34)
so the integral along C
4
vanishes in the limit as R since B
R
0.
Proof of Proposition 2. First, for k
)
, j
)
(k) is a scaled version of the Fourier transform of r
())
.
To see this, parameterize : S
)
as : = m
)
+ sh
)
, s , where m
)
:= (:
)
+ :
)+1
)}2 and h
)
:=
(:
)+1
:
)
)}(2), and set
[(s) =
_
r
())
(m
)
+ sh
)
), s ,
0 otherwise.
(35)
Likewise, parameterize k
)

)
0 as k = t}h
)
, t R, so positive and negative t correspond to
k
)
and k

)
, respectively. For these k values, (21) reduces to
j
)
(k) = j
)
(t}h
)
) = 2h
)
e
itm
)
}h
)

[(t) t R, (36)
where
[(t) :=
1
2
_

e
its
[(s) ds (37)
is the Fourier transform of [. The inverse transform is
[(s) =
_

e
its

[(t) dt, (38)
which converges for all s (, ). This may be written as
r
())
(:) =
1
2
_
_
)
e
ik:
j
)
(k) dk
_
)
e
ik:
j
)
(k) dk
_
: S
)
(39)
with both integrals nite.
Now since for each ) = 1, . . . , n the integral in (22) is nite for : (D S) S
)
by Lemma 1
and for : S
)
by (39), the function r(:) is dened for all : D S. By subtracting a polynomial
we may assume that r
())
= 0 at the endpoints of S
)
and hence [H
1}2+c
(R). Thus,

[L
1
(R) so r
is continuous. To show that r is analytic on D, we can dierentiate formally with respect to :; the
resulting integrals can be shown using Lemma 1 to converge uniformly on a neighborhood of any
point : D, thus justifying the dierentiation. Any antiderivative q of r is also continuous on DS
and analytic on D, and since q
: :
= r
:
= 0, Re(q) satises the Laplace equation on D.
Finally, to show that r matches the prescribed boundary values, we x 1, . . . , n and : S
,
multiply the global condition (8) by e
ik:
and integrate over

to obtain
_
)=1
e
ik:
j
)
(k) dk = 0. (40)
Since each of the integrals is nite, we can interchange the order of integration and summation to
obtain
n
)=1
_
e
ik:
j
)
(k) dk = 0. (41)
Dividing by 2 and subtracting from (22) yields
r(:) =
1
2
n
)=1
_
_
)
e
ik:
j
)
(k) dk
_
e
ik:
j
)
(k) dk
_
. (42)
For ) ,= the two integrals cancel by Lemma 2, leaving
r(:) =
1
2
_
_
e
ik:
j
(k) dk
_
e
ik:
j
(k) dk
_
= r
()
(:), : S
, (43)
where the last step follows from (39).
4. Numerical solution of the global condition
It follows from Proposition 2 that the key to solving the Laplace problem is solving the global
condition (8) for j
)
(k) () =1, . . . , n) in terms of the given boundary data. When this can be solved
analytically [2], the resulting solution is given by (13). In this section we give the details of the
numerical method outlined in Section 2.2 for solving the global condition.
4.1. Parameterization
To write the global condition in a form appropriate for numerical solution, we rst parameterize
: on the side S
)
by : =m
)
+sh
)
, s , where m
)
:= (:
)
+:
)+1
)}2 and h
)
:= (:
)+1
:
)
)}(2),
as in the proof of Proposition 2. Then using (17)and reinterpreting [
())
and q
())
as functions of
s rather than :we can write (9) in the form
j
)
(k) = h
)
e
io
)
e
ikm
)
_
q
( ))
(kh
)
) + i

[
( ))
(kh
)
)
_
, (44)
where the hat denotes the Fourier transform [cf. (37)], now evaluated at complex arguments. Thus,
the global condition (8) takes the form
n
)=1
h
)
e
io
)
e
ikm
)

[
( ))
(kh
)
) = i
n
)=1
h
)
e
io
)
e
ikm
)
q
( ))
(kh
)
), (45)
where the functions

[
( ))
are the unknowns, corresponding to the unknown integrals
)
in Section
2.1 [cf. (18)].
4.2. The choice of k
The global condition holds for all k C. Which values should we use? Eq. (13) indicates that
j
)
(k) is needed for k on the ray
)
dened by (14), where by construction j
)
(k) is oscillatory in k.
Indeed, the term e
ikh
)
s
in the integrand of j
)
(k) is oscillatory (and thus bounded) for all s (, )
if and only if kh
)
is real. Now while kh
)
is real for k on the ray
)
, on this ray the term e
ikm
)
multiplying the unknown

[
( ))
(k) in the global condition (45) is exponentially small as [k[ , so
these unknowns will be only weakly coupled. In contrast, kh
)
is also real for k on the ray

)
dened
by (20). On this ray the term e
ikm
)
is exponentially large as [k[ , so the unknowns

[
( ))
(k)
will be strongly coupled by the global condition (45).
3
Consequently, to derive a well-conditioned
system of equations we choose k on the ray

)
.
To obtain such a system, we choose a side index 1, 2, . . . , n and set k =!}h
in the global
condition (45); then positive and negative ! correspond to k

and k
, respectively. Scaling
the result by the coecient of the term ) = leads to
n
)=1
t
, )
o
!
, )

[
( ))
(!h
)
}h
) = i
n
)=1
t
, )
o
!
, )
q
( ))
(!h
)
}h
), = 1, . . . , n, (46)
where o
, )
:= e
i(m
)
m
)}h
and t
, )
:= (h
)
}h
)e
i(o
)
o
)
. Thus, (46) is a system of n equations for the
n unknown functions

[
( ))
. We note that o
,
= 1, and that for ) ,= , [o
, )
[ 1 (by convexity) so
the coecient of

[
( ))
is exponentially small as ! . Also, for )= and ! Z the numbers

[
( ))
(!)
are simply the coecients in the Fourier series for [
())
. Since this function is real, the coecients
for ! 60 [i.e., k
)
as needed for (13)] are related to those for ! 0 [i.e., k

)
as determined
by (46)] via

[
( ))
(!) =

[
( ))
(!).
3
This argument assumes the domain D is convex and contains the origin. It can be made independent of the location
of the origin by scaling by the coecient of

[
()
as in (46).
4.3. Continuity conditions
Up to this point the only restriction imposed on the solution is that q
:
must be integrable on each
side. Near a vertex :
)
the behavior of the solution depends on the interior angle c
)
:= (:
)

:
)1
) (0, ) and the boundary conditions as determined by [
)
and [
)1
[cf. (15)]. Without loss of
generality we can assume that 0 6[
)
[
)1
. It can be shown that:
If c
)
+ [
)
[
)1
then the problem is regular (q
:
is bounded) near :
)
.
If c
)
+ [
)
[
)1
then the problem is singular (q
:
may be unbounded) near :
)
.
In the borderline case c
)
+ [
)
[
)1
= we have o
)
= o
)1
, which says the same component of
the derivative (e.g., q
x
) is specied along sides S
)
and S
)1
. For simplicity, we exclude both the
borderline and singular cases here. Then the endpoint values of the unknowns can be determined
from the continuity of q
:
, and this information can be used to overcome the degeneracy present in
(46) when ! = 0.
Specically, at any vertex :
)
we have two representations of q
:
(:
)
), namely, q
( )1)
:
(:
)
) and q
( ))
:
(:
)
).
Requiring that these representations match and using (17) leads to the condition
e
io
)
[q
())
() + i[
())
()] = e
io
)1
_
q
()1)
() + i[
()1)
()
. (47)
The real and imaginary parts of this equation yield two equations for the unknown values [
()1)
()
and [
())
(). Solving these equations we nd
[
()1)
() =
cos(o
)
o
)1
)q
()1)
() q
())
()
sin(o
)
o
)1
)
(48)
and
[
())
() =
q
()1)
() cos(o
)
o
)1
)q
())
()
sin(o
)
o
)1
)
. (49)
Knowing these endpoint values, we can set
[
())
= [
( ))
+ F
())
, (50)
where the linear functions [
( ))
are chosen so that F

())
() =F
())
() =0 for each ) =1, . . . , n. Using
(50) and an analogous substitution for q
())
reduces (46) to
n
)=1
t
, )
o
!
, )

F
( ))
(!h
)
}h
) = i
n
)=1
t
, )
o
!
, )
_

G
( ))
(!h
)
}h
) +

H
( ))
(!h
)
}h
)
_
(51)
for = 1, . . . , n, where

H
( ))
(k) is the Fourier transform of the linear function
H
())
(s) = q
( ))
(s) + i[
( ))
(s). (52)
4.4. Numerical implementation
The global condition in form (51) consists of n equations for the n unknown functions F
())
,
) = 1, . . . , n (more specically, their Fourier transforms), which are coupled through the argument
k =!h
)
}h
. Note that each of the functions F

())
and G
())
is in C
0
[, ]=[C[, ] : [()=
[() = 0, and that up to this point the formulation is continuous in k (or !). To discretize the
problem, for each side S
)
we choose a basis
( ))
r
(s)
N
)
r=1
for a subspace S
)
of C
0
[ , ] (with
dimension N
)
even), and approximate F
())
(s) and G
())
(s) by
F
( ))
N
)
(s) =
N
)
r=1
c
( ))
r

( ))
r
(s) (53)
and
G
( ))
N
)
(s) =
N
)
r=1
d
( ))
r

( ))
r
(s), (54)
respectively. One example of such a basis is the hat functions dened on [, ] with mesh spacing
2}(N
)
+1); this family of piecewise linear functions is popular in nite element methods. Another
example is the sine basis
( ))
r
(s) = sin
_
r
_
s +
2
__
, (55)
which we will use for the numerical results presented below.
The Fourier transforms of (53) and (54) are
F
( ))
N
)
(k) =
N
)
r=1
c
( ))
r

( ))
r
(k) (56)
and
G
( ))
N
)
(k) =
N
)
r=1
d
( ))
r

( ))
r
(k), (57)
respectively. Since G
())
is a known function, we can determine the coecients d
( ))
r
by a collocation
projection of (57) in k, i.e., we require

G
( ))
(k) =

G
( ))
N
)
(k) for k = ! = 1, . . . , M
)
where M
)
= N
)
}2.
This leads to the linear system
N
)
r=1
d
( ))
r

( ))
r
(!) =

G
( ))
(!) =
1
2
_

e
i!s
G
())
(s) ds, (58)
the right-hand side of which can be computed by direct or numerical integration, or more eciently
by the FFT algorithm. Since the basis functions
( ))
r
are linearly independent, system (58) is uniquely
solvable.
Likewise, we determine the coecients c
( ))
r
by a collocation projection of the global condition
(51) using approximations (56) and (57). The resulting discrete equations are
n
)=1
t
, )
o
!
, )
N
)
r=1
c
( ))
r

( ))
r
(!h
)
}h
)
=i
n
)=1
t
, )
o
!
, )
_
N
)
r=1
d
( ))
r

( ))
r
(!h
)
}h
) +

H
( ))
(!h
)
}h
)
_
, (59)
which are applied for ! =1, . . . , M
)
and =1, . . . , n. Thus, the discrete global condition (59) consists
of M =
n
=1
M
complex equations for the N =
n
)=1
N
)
real coecients c
( ))
r
, where N =2M. The
matrix form of this system is straightforward, although tedious due to the need to work with the
real and imaginary parts separately; the details appear in Appendix A. The system could be solved
by a natural block-iterative method (e.g., Jacobi or Gauss-Seidel); here for simplicity we solve it
directly (using Matlab).
5. Numerical results
To illustrate the method, we apply it to the Laplace equation on a variety of convex polygonal
domains with dierent boundary conditions. For concreteness, in each case we take the analytical
solution to be
q(x, ,) = sinh(3x) sin(3,) (60)
and generate the corresponding boundary data analytically. To avoid unrepresentative results due
to alignment with the coordinate axes, in each case the specied domain is rotated by the angle
0.2. The performance of the method is quantied by comparing the analytical solution [, composed
piecewise of [
())
on sides S
)
[cf. (50)], and the corresponding numerical solution [
N
, composed
piecewise of [
( ))
+ F
N
)
[cf. (53)]. For the gures we measure the maximum relative error in the
norm
E
:=
|[ [
N
|
|[|
, |[|
:= max
16)6n
_
max
6s6
[[
())
(s)[
_
(61)
with the max over s taken over a large number of discrete points. In all cases we use the same
number of basis functions on each side of the domain.
First, we consider the solution for regular polygons. For n=3, 4, 5, 6, and 8 we construct a domain
as a regular n-gon centered at the origin with : =1 the midpoint of one side. Figs. 4 and 5 show the
10
1
10
2
10
0
10
1
10
-1
10
-2
10
-3
10
-4
10
-5
10
0
BASIS FUNCTIONS PER SIDE
R
E
L
A
T
I
V
E

E
R
R
O
R

(
M
A
X

N
O
R
M
)
Equilateral Triangle
Square
Regular Pentagon
Regular Hexagon
Regular Octagon
Fig. 4. Relative errors on regular polygons with Dirichlet boundary conditions.
R
E
L
A
T
I
V
E

E
R
R
O
R

(
M
A
X

N
O
R
M
)
Equilateral Triangle
Square
Regular Pentagon
Regular Hexagon
Regular Octagon
10
1
10
2
10
0
10
1
10
-1
10
-2
10
-3
10
-4
10
-5
10
0
Fig. 5. Relative errors on regular polygons with Neumann boundary conditions.
R
E
L
A
T
I
V
E

E
R
R
O
R

(
M
A
X

N
O
R
M
)
10
1
10
2
10
0
10
-1
10
-2
10
-3
10
-4
10
-5
10
0
Triangle
Trapezoid
Pentagon
Fig. 6. Relative errors on general polygons with mixed boundary conditions.
corresponding errors for Dirichlet ([
)
=0) and Neumann ([
)
=}2) boundary conditions, respectively.
The dotted lines in each gure show slopes for convergence of order 1 and 2. It is evident that
the method converges in each case, with order of convergence between 1 and 2. These results were
computed with the sine basis (55); with a basis of hat functions the corresponding errors (not shown)
are similar but larger by at least a factor of two.
Likewise, Fig. 6 shows corresponding results for three more general polygons with mixed boundary
conditions ([
)
= }3):
A triangle with vertices (0, 0), (0, 2), and (1, 0).
A trapezoid with vertices (0, 0), (0, 1), (1, 2), and (1, 0).
A pentagon with vertices (0, 0), (0, 1), (1, 2), (2, 1), and (1, 0).
Once again, the numerical solution converges in each case.
6. Conclusions and remarks
A new numerical method for solving linear elliptic PDEs with constant coecients has been in-
troduced. Proposition 2 of Section 3 reduces the solution of the Laplace equation on an arbitrary
convex bounded polygon to the solution of the global condition, which couples known and unknown
components of the derivative on the boundary. The numerical method introduced in Section 4 pro-
vides an approximate solution of the global condition. Results presented in Section 5 demonstrate
the convergence of the method; however, a rigorous proof of convergence remains open.
An advantage of this new numerical method is that it is formulated only in terms of the boundary
values. This should be particularly useful if one is interested only in the missing boundary values,
as opposed to q(x, ,) in the interior of the polygon. Examples of such cases include the analysis of
potential elds [5] and the determination of the drag force in creeping (Stokes) ow [7].
We note that the method introduced here can be extended in various ways:
(1) Problems with singularities at the vertices: Recall from Section 4.3 that if c
)
+[
)
[
)1
then the problem is singular near :

)
. In this case the corner values q
())
() cannot be computed
using the continuity conditions (47), but must be coupled with the unknown coecients c
( ))
r
.
(2) Other boundary conditions: We can replace the mixed boundary conditions (15) with the
more general Poincar e boundary conditions
cos([
)
)q
( ))
s
+ sin([
)
)q
( ))
n
+
)
q
())
= q
())
, : S
)
. (62)
If
)
,= 0 we can assume without loss of generality that sin([
)
) ,= 0; otherwise q
())
and hence q
( ))
s
is known, reducing to a special case of (15). Solving (62) for q
( ))
n
, substituting the result into (10)
to obtain q
:
, substituting that into (9) and integrating by parts yields
j
)
(k) =
ih
)
e
io
)
e
ikm
)
sin([
)
)
_
e
i[
)
q
( ))
(kh
)
) (
)
e
i[
)
+ ike
i:
)
) q
( ))
(kh
)
)
1
2[h
)
[
_
e
ikh
)
q
())
() e
ikh
)
q
())
()
_
, (63)
which replaces (44) in the development of the method. The global condition (8) can then be solved
in a manner similar to that outlined in Section 4, with two dierences. First, the discrete expansion
must now be for the variable q
())
, rather than a derivative of q
())
. Second, the corner values q
())
()
in (63) cannot be eliminated in advance, but must be coupled with the unknown coecients of the
solution as in point 1 above.
(3) Other linear elliptic PDEs: The results obtained here can be generalized to other linear elliptic
equations. For example, for the modied Helmholtz equation (1) with : = [
2
, j
)
(k) is given by
[cf. (5)]
j
)
(k) =
_
S
)
e
ik:+i[
2
:}k
_
q
:
d: +
i[
2
k
q d :
_
, k C, (64)
where q satises Eq. (1) with : =[
2
. In particular, the analysis of the global condition associated
with other elliptic PDEs is conceptually similar to the analysis of the global relation (8). In this sense
the application to the Laplace equation is the simplest but still a generic example of the application
of a new approach to linear elliptic PDEs.
(4) Forced problems: Consider for example the Poisson problem
q
: :
= F(:, :), : D. (65)
In this case the global condition (8) is replaced by
n
)=1
j
)
(k) = 2i
_ _
D
e
ik:
F(:, :) dx d,, k C. (66)
The right-hand side of this equation is known, thus its analysis is similar to the analysis of Eq. (8).
However, computing this right-hand side requires a double integration for each value of k, so this
approach may be practical only if these integrals can be computed analytically.
Finally, several aspects of the numerical method introduced here deserve further investigation.
In particular, in subsequent research we plan to address the convergence analysis, iterative solution
methods, and comparison with other methods in terms of accuracy and eciency.
Appendix A. Matrix form of the linear system
To write the linear system (59) in matrix form we treat the real and imaginary parts of the
equations separately. To that end, for any z C let
F
(, ))
z
:= [B
T
1
(

F
( ))
N
)
(z)), B
T
1
(

F
( ))
N
)
(2z)), . . . , B
T
1
(

F
( ))
N
)
(M
z))]
T
(A.1)
in R
2M
, where B
1
(:) := [Re(:), Im(:)]
T
. Then applying (56) at k = z, 2z, . . . , M
z we can write
the result in matrix form as
F
(, ))
z
:=

P
(, ))
z
c
( ))
, (A.2)
where c
( ))
:=
_
c
( ))
1
, . . . , c
( ))
N
)
_
T
and
P
(, ))
z
:=
_
_
B
1
_

( ))
1
(z)
_
B
1
_

( ))
N
)
(z)
_
.
.
.
.
.
.
B
1
_

( ))
1
(M
z)
_
B
1
_

( ))
N
)
(M
z)
_
_
_
. (A.3)
For any : C we dene the block-diagonal matrices
D
M
(:) := diag[B
2
(:), B
2
(:), . . . , B
2
(:)], (A.4)
E
M
(:) := diag[B
2
(:
1
), B
2
(:
2
), . . . , B
2
(:
M
)] (A.5)
in R
2M
2M
, where
B
2
(:) :=
_
Re(:) Im(:)
Im(:) Re(:)
_
. (A.6)
Then if we write vector forms of

G
( ))
and

H
( ))
analogous to (A.1), the global condition (59) takes
the matrix form
n
)=1
D
M
(t
, )
)E
M
(o
, )
)

P
(, ))
h
)
}h
c
( ))
=
n
)=1
D
M
(it
, )
)E
M
(o
, )
)
_
P
(, ))
h
)
}h
d
( ))
+

H
(, ))
h
)
}h
_
(A.7)
for = 1, . . . , n, where d
( ))
:=
_
d
( ))
1
, . . . , d
( ))
N
)
_
T
. We can write this linear system as
Ac = b, (A.8)
where AR
NN
has blocks
A
(, ))
= D
M
(t
, )
)E
M
(o
, )
)

P
(, ))
h
)
}h
, (A.9)
c = [(c
(1)
)
T
, . . . , (c
(n)
)
T
]
T
and similarly for b, with b
()
denoting the right-hand side of (A.7).
It should be noted that system (A.8) can also be written in the form
A

F = b, (A.10)
where the blocks comprising

AR
NN
are
A
(, ))
= A
(, ))
_
P
( ))
1
(A.11)
and

F =
_
(

F
(1)
)
T
, . . . , (

F
(n)
)
T
_
T
with

F
( ))
=

F
( ), ))
1
and

P
( ))
=

P
( ), ))
1
. Indeed, (A.8) may be referred to
as a physical space form of the global condition, since the unknown c gives the functions F
())
as functions of s via (53), and (A.10) may be referred to as a spectral space form, since the
unknown

F gives the Fourier coecients of those functions. The latter may appear preferable:
the unknowns are (essentially) the unknown functions j
)
(k) needed to compute the solution of the
Laplace equation [cf. (13)], and the diagonal blocks of

A are identity matrices. Nevertheless, the
former may be better in practice, since the condition number of the physical-space matrix A grows
more slowly with N.
References
[1] A.S. Fokas, A unied transform method for solving linear and certain nonlinear PDEs, Proc. Roy. Soc. London A
53 (1997) 14111443.
[2] A.S. Fokas, Two-dimensional linear PDEs in a convex polygon, Proc. Roy. Soc. London A 457 (2001) 371393.
[3] A.S. Fokas, A. Kapaev, On a transform method for the Laplace equation in a polygon, IMA J. Appl. Math. 68
(2003) 155.
[4] A. Greenbaum, L. Greengard, G.B. McFadden, Laplaces equation and the DirichletNeumann map in multiply
connected domains, J. Comput. Phys. 105 (1993) 267278.
[5] H. Igarashi, T. Honma, A boundary element method for potential elds with corner singularities, Appl. Math.
Modelling 20 (1996) 847852.
[6] C. Johnson, J.C. Nedelec, On the coupling of boundary integral and nite element methods, Math. Comput. 35
(1980) 10631079.
[7] T.C. Papanastasiou, G. Georgiou, A. Alexandrou, Viscous Fluid Flow, CRC Press, Boca Raton, FL, 1999.
[8] F. Pars, J. Ca nas, Boundary Element Method, Fundamentals and Applications, Oxford University Press, New York,
1997.
[9] T. Shigeta, T. Harayama, K. Shirota, Combination of the nite element and the boundary element methods for an
external boundary value problem of the Poisson equation, J. Chinese Inst. Engrs. 22 (1999) 777783.
[10] W.L. Wendland, Boundary element methods and their asymptotic convergence, in: P. Filippi (Ed.), Theoretical
Acoustics and Numerical Techniques, CISM, Courses and Lectures, Vol. 277, Springer, New York, 1983, pp. 135
216.
[11] J.P. Wolf, C. Song, The scaled boundary nite-element methoda fundamental solution-less boundary-element
method, Comp. Meth. Appl. Mech. Eng. 190 (2001) 55515569.

An Analytical Method For Linear Elliptic Pdes and Its Numerical Implementation

Uploaded by

Copyright:

Available Formats

An Analytical Method For Linear Elliptic Pdes and Its Numerical Implementation

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

An Analytical Method For Linear Elliptic Pdes and Its Numerical Implementation

Uploaded by

Copyright:

Available Formats

Available online at www.sciencedirect.

, where the linear function [

)] [k[p sin(c), where p := inf

are parallel) there is nothing to prove, since the

coincide. Thus, we consider the case 0 :

can be treated similarly. Note that if ,= ) + 1 then the integrand decays

and the proof is straightforward. Therefore,

are chosen so that F

. Note that each of the functions F

complex equations for the N =

then the problem is singular near :

You might also like