3.1. Width of Exterior Boundary Layers
In the DNFEM and the MFS, the source nodes are located not only outside
S but also outside the exterior boundary layers
, which are similar to the interior boundary layers for the boundary layer computation in
Section 4.4. One may ask, what is the width of the boundary layers
of
? It is shown below that
. Since the maximal
N in numerical computations is still finite, the source nodes
must be located at a larger distance of
to
. In
Section 4.4, the solutions in the interior boundary layers can only be computed after the
in Equation (
4) with a given
N are obtained. Since the widths of the interior and exterior boundary layers are similar, the width study in this subsection is more important for the DNFEM.
To explore the width of the exterior boundary layers, we need the error analysis in
Section 2.2. For simplicity, we consider a disk domain with radius
a and choose an outside circle
with radius
. The radius
is given to bypass the degenerate scales (see Equations (
23) and (
25)). Suppose that the solution
is highly smooth with large
p. In the error bounds of Equation (
48), the second term is dominant:
where we choose
and denote
. For large
N, we have
From Equations (
50) and (
51), we find large errors, as
.
Next, suppose that small errors of
, where
, are required for precision. Then, we may choose a slightly larger circle
with radius
. We have
where
. Hence, we have
In contrast, from Equations (
52) and (
53), we obtain small errors, as
.
We summarize the above results as a proposition.
Proposition 3. Consider disk domains with radius a and use the circular pseudo-boundaries with radius in the DNFEM. Suppose that the solution is highly smooth with large p. When choosing the circular pseudo-boundaries with radius , large errors of may occur. However, when choosing circular pseudo-boundaries with a slightly larger radius , small errors of , where , can be achieved. Hence, the width of the exterior boundary layers is determined as .
Consider the MFS, where the solution
in Equation (
10) is given from Equation (
11). From [
8] (Chapter 2), we have the error bounds:
When
, the second dominant term leads to
Based on Equation (
55), we can similarly prove the following corollary for the MFS.
Corollary 2. Consider disk domains with radius a and use the circular pseudo-boundaries with radius in the MFS. When radius , we have . When radius , we have . The same width of the exterior boundary layers is also found for the MFS.
Define the boundary and domain errors for the DNFEM and the MFS by
where
,
, and
and
are obtained from the DNFEM and the MFS, respectively.
To evaluate the width of the exterior boundary layers, we choose the solution
. The results are given in
Table 1 and
Table 2. The curves based on the results in
Table 1 and
Table 2 are given in
Figure 1 and
Figure 2. For
, we have
. In
Table 1 and
Table 2, the large errors of
are given for
in agreement with Proposition 3 and Corollary 2. Next, for the errors
, from Proposition 3, we use
, where the theoretical width
. Hence, when
and
, we have
Since
on
, we find
from Equation (
56). In
Table 1, small errors of
are shown for
, which also supports the convergence results in Proposition 3. To achieve higher accuracy, the
rule was proposed by Barnett [
3]. Later, more computations were carried out for the source nodes outside the exterior boundary layers (i.e., at least at a distance larger than
).
Next, for the DNFEM, when the source nodes are as far as
, the errors become larger (see
Table 1 and
Figure 1) due to ill-conditioning. Hence, for the DNFEM, the source nodes should be chosen not too far from
(i.e.,
). However, for the MFS, source nodes far from
for
are beneficial (see
Table 2 and
Figure 2). The locations of source nodes relative to
in the DNFEM are more sensitive than those in the MFS.
The term “boundary layers" originated from fluid mechanics, and the thickness (i.e., width) has been studied. In this paper, we also use the “boundary layers" resulting from the singularity of the FS:
as
. The exterior boundary layers occur for both the DNFEM and the MFS, but the interior boundary layers occur only for the DNFEM. In [
3], the width of the interior boundary layers was studied for the boundary integral equation method (BIEM). In this paper, we explore the width of the exterior boundary layers for the DNFEM and the MFS, which is consistent with [
3].
3.2. Pseudo-Boundaries from Conformal Mapping
For Jordan domains, conformal mapping was employed by Katsurada and Okamoto [
9] to transform both the domain boundary and the pseudo-boundary into concentric circles. Then, exponential convergence rates were obtained for analytic solutions using the MFS. We use conformal mapping for the DNFEM (as well as the MFS) to find suitable pseudo-boundaries. We can express the coordinate functions using the Fourier series, as shown in [
4]:
Let
and denote the uniform nodes on the unit circle
with
. Also, denote the boundary nodes
. The conformal mapping
T from the unit circle to
in Equation (
57) at
can be defined, where the coefficients
were obtained from [
4]:
Let the boundary
be defined by
, where
is a function in polar coordinates. In Equations (
58) and (
59), we choose
and
with
. The pseudo-boundary
of the source nodes can be determined using Equation (
57) with small
. For
, we choose the Fourier series:
The conformal mapping
T from
to
with
, is given by Equations (
60) and (
61) for
, where the Fourier coefficients are
and
where
and
with
.
For disk domains
S with radius
a, the conformal mapping from a unit circle to a circle with radius
a is given in Equations (
57), (
60) and (
61) for
, where
, while all other coefficients
are zero. In the general case, where
is defined by
, nonzero coefficients
arise. The conformal mappings in Equations (
57), (
60) and (
61) are valid only for small values of
and not for large values of
L.
Proposition 4. Let denote the conformal mapping, and suppose that the -periodic function with . For Fourier expansions in Equation (57), small values of and not large values of n (i.e., not large L) are required in applications. One necessary condition iswhere is a bounded constant independent of ρ and n. For the given n, the condition in Equation (66) leads to Proof. To find the conditions for the conformal mapping, we consider
as an example. The function
can be expressed using a Fourier series as follows:
where
and
are the true expansion coefficients. Then, we have
where the coefficients in Equation (
58) are determined by
Under the assumption
with
, we have the bounds for the coefficients in Equations (
68) and (
70):
From Equation (
71), we have
Since
is bounded, the last term,
, in the series
from Equation (
72) must also be bounded, which leads to the necessary condition in Equation (
66). Hence, small vallues of
and not large values of
n are required in applications. For a given
n, the condition in Equation (
67) follows directly from Equation (
66). This completes the proof of Proposition 4. □
When
, the Fourier expansions in Equations (
60) and (
61) are used. We take the peanut-like domain in Equation (
85) below for testing purposes. Equations (
60) and (
61) work well when
and
, as shown in
Figure 3, where only
source nodes on
are shown. Other
source nodes can also be obtained from the same pseudo-boundary
defined by Equations (
60) and (
61). When
, the source nodes at
are illustrated in
Figure 4. The source nodes at
are useless, and some of them at
are unacceptable. Several source nodes are too close to each other at
and even located inside
at
.
Figure 4 displays a failure of the conformal mapping for
. To test the failure of the conformal mapping for large
, we choose the same
as shown in
Figure 3. In
Figure 5, failure occurs for
, thus verifying the necessary condition in Proposition 4.
In [
9], the conformal mapping was carried out using complex functions. Let
, and choose the complex transformation
where the complex function
with
and
. The complex coefficients
with
. Choose the complex variable
. From Equation (
73), we have
Then, from Equation (
74), we have
where the coefficients are given by
When
, the conformal mapping in Equation (
75) from [
9] is equivalent to Equation (
57) using the Fourier series (see [
14], p.178). However, for
, the two algorithms are slightly different. Hence, the pseudo-boundaries from [
9] are also slightly different. Under some stronger analytic assumptions for the boundary
and the function
f in Equation (
1) (i.e., the analytic Jordan curve
and the analytic function
f), the distance
must be small but
N may be large. Then, the fast Fourier transformation (FFT) can be used for large
to save CPU time. The algorithms are given in [
9] (p. 128).
Remark 1. In this paper, the number L is different from N, which is used in the DNFEM and the MFS. The L is confined to not be large (e.g., for or for for the peanut-like domain in Equation (85)), which may be smaller than N (even ) for some applications. For large N, we can use the known continuous curve in Figure 3 with and . The source nodes can be obtained directly from Equations (60) and (61) by (e.g., ) (see Section 5.2). Such algorithms are simple since the FFT in [9] can be bypassed. Hence, the small is only one key necessary condition of Proposition 4. Let us check the condition in Equation (66) again. If the boundary Γ is highly smooth with , where q is large, both ρ and n in Equation (66) could be larger. When Γ is polygons or piecewise smooth curves, we have only. The conformal mapping fails because only is allowed from Equation (66) at . Remark 2. In Section 3.1, the width of the exterior boundary layers is studied for disk domains with circular pseudo-boundaries. In this remark, we seek the width of the exterior boundary layers for smooth domains when the pseudo-boundaries are obtained from the conformal mapping. For the conformal mapping, the width of the exterior boundary layers can also be determined using the error estimates of the DNFEM. Consider both Γ and to be highly smooth Jordan curves. By a one-by-one conformal mapping , Γ and are, respectively, converted to a real axis and a parallel line in the complex plane (see Barnett ([3], Figure 2.1):For the analytic solutions from the DNFEM, using the trapezoidal rule, we have the following error bounds:When , from Equation (79). Thus, the DNFEM diverges. To achieve small , a slightly larger may be required to satisfy . This yieldsThe width is again found to coincide with Proposition 3 and Corollary 2 at . The conformal mapping in Equation (77) is based on the Schwarz function in Davis [15]. In the complex plane, let and , and we have and . A closed highly smooth contour can be denoted by the analytic function , where g can be also regarded as a -periodical function in . Thus, the boundary Γ of S is denoted by with and (see [15], Chapters 2 and 5). We can relate the above mapping to the conformal mapping used here. Denote the other one-to-one conformal mapping by . From Equation (77), we haveEquation (81) also denotes circles with radii . When , we have . By the composite conformal transformation , the boundary Γ and the pseudo-boundary are converted to a unit circle and a concentric circle with radius , as shown in this section. Hence, the width, , of the exterior boundary layer is again confirmed. Hence, by using , Equation (57) is rewritten as Remark 3. The study in this paper can be extended to n-dimensional (nD) Laplace’s equation in Equation (1). First, consider the 3D case. Since the FS is , the Green’s representation formula in Equation (2) is modified as follows (see Atkinson [16], p.430):where the node . The NFE from the third equation of Equation (82) is given byFor D, since the FS is given by (see Chen and Zhou [17], p.214), the NFE is given byThe algorithms of the DNFEM can be formulated similarly, where the source nodes are located outside the boundary Γ. Since the conformal mapping in Section 3.2 can be extended to nD, the techniques for locating in this paper can be developed further. Let us take the 3D case as an example. The conformal mapping in 3D is given in [8] (Section 13.2). The smooth surface Γ is converted using the conformal mapping T into a unit sphere with radius . Then, the pseudo-surface is obtained from the larger sphere with radius by applying its inverse transformation . For spheres, a distribution of good nodes is illustrated in ([8], p. 326). The collocation nodes on Γ and the source nodes on can be found using . For the conformal mapping in Section 3.2, the techniques are simple and straightforward without solving linear algebraic equations. Such a remarkable advantage remains for the conformal mapping in 3D. For the DNFEM in 3D, the stability and error analysis in Section 2 needs to explored, and the width of the boundary layers in Section 3.1 can then be discussed similarly.