ESAIM: M2AN 46 (2012) 647660

DOI: 10.1051/m2an/2011063

ESAIM: Mathematical Modelling and Numerical Analysis

www.esaim-m2an.org

ITERATIVE SCHEMES FOR HIGH ORDER COMPACT DISCRETIZATIONS

TO THE EXTERIOR HELMHOLTZ EQUATION

Yogi Erlangga1,2 and Eli Turkel3

Abstract. We consider high order nite dierence approximations to the Helmholtz equation in an
exterior domain. We include a simplied absorbing boundary condition to approximate the Sommerfeld
radiation condition. This yields a large, but sparse, complex system, which is not self-adjoint and not
positive denite. We discretize the equation with a compact fourth or sixth order accurate scheme. We
solve this large system of linear equations with a Krylov subspace iterative method. Since the method
converges slowly, a preconditioner is introduced, which is a Helmholtz equation but with a modied
complex wavenumber. This is discretized by a second or fourth order compact scheme. The system is
solved by BICGSTAB with multigrid used for the preconditioner. We study, both by Fourier analysis
and computations this preconditioned system especially for the eects of high order discretizations.
Mathematics Subject Classification. 35J47, 65M06.
Received August 26, 2009. Revised January 3, 2011.
Published online January 11, 2012.

1. Introduction
Many physical phenomena such as acoustics, elasticity and electromagnetic waves are governed in the frequency domain by the Fourier transform of the wave equation, the Helmholtz equation, with appropriate
boundary conditions. Examples include acoustic scattering about bodies and geophysical problems. In these
problems, the boundaries of the computational domain are typically treated as absorbing ones to guarantee
that the outgoing waves are not reected back into the domain. The main challenge comes from the fact that
the Helmholtz equation has to be solved for a wide range of frequencies, from low to high, especially if one is
interested in the associated time-domain solution of the problem, obtained by an inverse Fourier-transform of
the frequency-domain solutions.
In order to compute the numerical solution, a mesh needs to be dened, and an appropriate discretization is
applied to the Helmholtz equation. It is well known [1, 3] that the mesh length, h, required for a xed accuracy
increases faster than linear in the frequency. Babuska and coworkers named this the pollution eect [1]. In
Keywords and phrases. Helmholtz equation, high order compact schemes.

Dedicated to the memory of a dear friend, David Gottlieb.

Department of Earth and Ocean Sciences, The University of British Columbia, 2329 West Mall Vancouver, BC V6T 124,
Canada.
2 Currently at Mathematics Division, College of Sciences, Alfaisal University, P.O. Box 50927, Riyadh, 11533 Kingdom of Saudi
Arabia.
3 Department of Mathematics, Tel Aviv University, Tel Aviv, Israel. [email protected]
Article published by EDP Sciences

c EDP Sciences, SMAI 2012



648

Y. ERLANGGA AND E. TURKEL

particular, the number of grid points N follows the relation

N=

p+1
1
k p ,
h

where p is the order of accuracy of the scheme. Thus, for a second order accurate scheme the grid grows like
5
k 3/2 , in order to maintain a constant accuracy. For a fourth order algorithm N k 4 , while for a sixth order
7
accurate scheme N k 6 . Hence, for high wave numbers (frequencies) there is a need for very ne grids. This
requirement is, however, less demanding if a higher order discretization is used. For nite dierences, examples
of high order accurate nite-dierence discretizations on uniform meshes are introduced and analyzed by Harari
and Turkel [12], and later by Singer and Turkel [13, 14].
Given a discrete approximation to the Helmholtz equation, we need to solve the associated linear system.
As the problem size increases, the eciency, in terms of computational work and storage, of a direct method
deteriorates rapidly, especially in 3D. For Helmholtz problems, iterative methods, however, are not an alternative to direct methods unless a special preconditioner is incorporated. In [9], Erlangga et al. introduce a
preconditioner based on a discrete approximation to the damped wave equation. The idea of using an operator
as a preconditioner for the Helmholtz equation is not new. A prototype of this can be traced back to the work
by Bayliss et al. [2] in the early 1980s. In [9] it is shown that a Krylov method preconditioned by a damped wave
equation can attain convergence, which depends linearly on the frequency (with a small proportional constant).
It is well known that classical multigrid methods cannot be used to solve the Helmholtz equation with the
main diculty being intermediate frequencies. Instead the multigrid can be used only for very ne grids that
still maintain some accuracy. Elman and OLeary [6] suggested a combination of multigrid and Krylov strategies
on various grids. Alternatively, Brandt and Livshitz [5] developed a new multigrid strategy based on waves.
In this study we solve the Helmholtz equation with a preconditioned Krylov method. The preconditioner is
based on a modied Helmholtz equation which is then solved by classical multigrid techniques utilizing all coarser
levels of grids. In contrast to [9], which focused only on second-order accurate schemes, we shall concentrate on
the eect of high-order nite-dierence schemes on the convergence.
The paper is organized as follows. In Section 2, we shall briey discussed the Helmholtz equation and the
construction of some high-order nite dierence schemes. Discretization of boundary conditions is described in
Section 3. In Section 4, the iterative method for solving the resultant linear system is discussed, followed by
Section 5, which presents numerical results. Some concluding remarks are drawn in Section 6.

2. Discretization
A dissipative term, for the wave equation, translates to either a plus or minus sign in the exponent in the
frequency domain and similarly for the radiation condition. However, the two have to be consistent. Hence,
depending on how the Sommerfeld radiation condition is implemented it will aect the sign of an additional
imaginary dissipative term to the Helmholtz equation. Note that, with zero dissipation, we have
vxx + vyy + k 2 v = 0,

(2.1)

the standard Helmholtz equation. In general, the right-hand side may not be equal to zero. In this paper we
only consider a zero right hand side. In [13] the formulae for including the right hand is given.
Let i,j be a numerical approximation to v (xi , yj ). We wish to use symmetric stencil in both directions x
and y. A nine point scheme having these properties has the form
A0 i,j + As s + Ac c = 0,
where
s = i,j+1 + i+1,j + i,j1 + i1,j

(2.2)

ITERATIVE SCHEMES FOR HELMHOLTZ EQUATION

649

is the sum of the values of the mid-side points, while

c = i+1,j+1 + i+1,j1 + i1,j1 + i1,j+1
is the sum of the values at the corner points.
The standard second order accurate approximation is given by
2

A0 = 4 + (kh) ,

As = 1,

Ac = 0.

A fourth order accurate scheme is given by the relations



2

10
2
A0 = + (kh)
+
,
3
3 36


1

2
1
2
2

,
, Ac = + (kh)
As = + (kh)
3
12 72
6
144

(2.3)

(2.4)

with an arbitrary constant. For variable k the scheme needs to be slightly modied. Instead of (kh)2 appearing
in formula (2.4) multiplying the corresponding A in (2.2) now for each node of the stencil the (kh)2 of that
2

. This is then
node is used. So, for example, i+1,j+1 is multiplied by the modied Ac = 16 + (kh)i+1,j+1 144
4
14
fourth-order accurate also for variable k(x, y) as shown in [12]. By choosing = 5 and adding O((kh) ) terms
to the coecients, one can achieve sixth order accuracy for constant k [14]. In this case, for arbitrary we have
10 67
3
2
4
+
(kh) +
(kh) ,
3
90
180
2
2
3 2
2
4
(kh) +
(kh) ,
As = +
3 45
720
7
1

2
4
(kh) +
(kh) .
Ac = +
6 360
720

A0 =

(2.5)

3. Boundary condition
We next wish to construct approximations to the boundary conditions that maintain the high order accuracy.
Suppose that along the boundary x = 1 we have a simplied radiation condition of the form
u
+ iu = 0.
x

(3.1)

We approximate the derivative by (dropping the index in the y direction for simplicity)
u h2 3 u
uN +1 uN 1
h4 5 u
=
+
+
+ O(h6 ),
2h
x
6 x3
120 x5
where the N + 1-th grid point lies outside the computational domain, and hence is a virtual point.
Dierentiating (3.1) twice we obtain a relation for the third-order and fth-order derivatives in (3.2):
3u
= (i)3 u = i 3 u
x3

5u
= i 5 u
x5

Substitution of the above relation into (3.2) yields

u i 3 h2
uN +1 uN 1
=
+
2h
x
6



2 h2
1
uN + O(h6 ),
20

(3.2)

650

Y. ERLANGGA AND E. TURKEL

or
uN +1
u
=
x

i 3 h3
3

2 h2
20

uN uN 1

2h

+ O(h6 )

and hence from (3.1),

0=

u
+ iu =
x


uN +1 + 2ih 1

2 h2
6

4 h4
120

uN uN 1

2h

+ O(h6 ).

(3.3)

This is combined with the discretization of the Helmholtz equation (2.2) at the N -th grid point to eliminate
the uN +1 term.

4. Iterative methods
The resultant linear system of equations has a matrix of coecients, which is sparse, has complex quantities,
is not self adjoint and is not positive denite. Furthermore, the resultant systems are too large, especially in
three dimensions to directly solve. In order to solve this linear system, we shall concentrate on iterative methods.
In particular we shall consider Krylov space methods.
Numerous authors have found that BICGSTAB [18] is an eective iterative method to solve the linear system
arising from the Helmholtz equation. However, this method, along with other Krylov space methods, converges
reasonably well only if the equations are preconditioned. Hence, we replace
Av = b
by the right preconditioning
AP 1 w = b

P v = w.

(4.1)

For the preconditioner to be ecient we require two conditions. The rst is that P is an approximation to A
and second that P v = w is much easier to solve than Av = b. Over the years two types of preconditioners have
been developed [16,17]. In the rst kind, the preconditioner utilizes algebraic properties of the matrix. A typical
example is ILU , where lower and upper factors are derived which are sparse and LU agrees with A for some
set of elements of A.
A dierent class of preconditioners is based on properties of the dierential system, i.e. an operator based
preconditioner. An example of this is basing the preconditioner on the solution of the Laplace equation [2].
Erlangga et al. [710] extended this idea by considering a Helmholtz equation with a complex wave number as
the preconditioner, namely
2
2
v + kprec
v = 0,
kprec
= (1 + i)k 2 ,
(4.2)
with the sign of depending on formulation 1 or formulation 2 as noted in Section 2, and similarly for the sign
in the radiation boundary condition. After discretization, the linear system associated with the preconditioner
is then approximately inverted by one multigrid iteration, with point damped Jacobi as a smoother [9]. This
will make the preconditioning step cheap, i.e., it is an O(N ) work.
In this paper, in addition to the use of damped Jacobi smoother, we also consider a (red-black (RB)) point
Gauss-Seidel (GS) smoother with a damping factor 0 < R 1. This should be referred to as successive
underrelaxation. Nevertheless, we shall use the more standard notation of either SOR or GS. In both case, one
pre- and post- smoothing step is employed. The coarsest grid goes to a 2 2 grid independent of k. We compare
V, F and W cycles, with bilinear interpolation and full weighting as the restriction and prolongation operator,
respectively.
Since the original Helmholtz equation (without a dissipative term) is discretized by a high order method we
study what is the appropriate discretization for (4.2).

ITERATIVE SCHEMES FOR HELMHOLTZ EQUATION

651

Figure 1. Spectrum of the unpreconditioned system (top left) and preconditioned systems with
second-order (top right), fourth order (bottom left) and sixth order (bottom right) discretization
used in A and P . k = 20, grid = 322 , and = 0.5.

4.1. Spectra of the preconditioned system

For Krylov subspace iterative methods, spectrum of the matrix in the linear system to be solved often
provides information about convergence. In general, a clustered spectrum is favorable for fast convergence, and
this clustering should be suciently far from zero. While spectral information is important, this information
is dicult to obtain theoretically for general problems. In this section, we present some spectral information,
numerically computed to show the eectiveness of the preconditioner (4.2).
Figure 1 shows spectra of AP 1 (4.1) for matrices A and P obtained with the same order of discretization,
with k = 20, 322 mesh (kh = 0.625), and = 0.5 in (4.2). To obtain these spectra, we assume that P is
inverted exactly. Compared with the spectrum of unpreconditioned system, the matrix P eectively reduces
the condition number and also clusters the eigenvalues around the circle with a center C = 0.5 and a radius
R = 0.5. This result conrms the algebraic results discussed in [19] and holds independent of the discretization
and kh.
Since the preconditioner is dened based on a partial dierential equation, we examine the eect of dierent
discretizations of the Helmholtz and preconditioning operators. Figure 2 shows spectra of AP 1 , with P based
on second order accurate dierences and A based either on fourth order or sixth order accurate scheme. Again,
similar clustering to Figure 1 is observed, except that some dierences in the detail, especially in the half right
of the spectra. Both possible combinations, however, lead to very similar spectra. Of course the spectra are not
identical but the dierences do not seem to impact the convergence rate as demonstrated by the computational
results. In general, it is dicult to deduce which will lead to a better convergence. Hence, we nd that the
order of accuracy of both the scheme and the preconditioner has little eect on the spectrum and this is borne

652

Y. ERLANGGA AND E. TURKEL

Figure 2. Spectrum of the preconditioned system, with a combination of second order discretization for P and fourth order (left) and sixth order (right) for A.
Noncompact, 2nd order
1

0.5

0.5
Imag

Imag

Noncompact, Noncompact
1

0.5

1
0.5

0.5

0.5
Real

1.5

1
0.5

0.5
Real

1.5

Figure 3. Spectrum of the preconditioned system, with a noncompact fourth order scheme
preconditioned by noncompact fourth order scheme (left) and a second order scheme (right).
out by the computational results. In Figure 3 we plot the spectrum for a noncompact fourth order scheme
preconditioned by the same fourth order noncompact scheme (left gure) and by the standard second order ve
point scheme (right gure). For the rst case, the spectrum is very similar to that in Figure 1. In the latter
case, however, we see a tail extending from the spectrum. It is not completely clear what eect this will have
on convergence. In any case a noncompact scheme has diculties near boundaries and has no clear advantage
and so no computational results will be presented.
We note that since we solve (4.2) using only one multigrid iteration, the exact inverse of the preconditioning
matrix is never available. Hence, the actual spectrum will deviate from the ideal situation.

4.2. Multigrid as the preconditioner solver

Multigrid is typically very ecient for Poisson-like equations. For this class of problems, a smoothing method
(typically based on a basic iterative method) eectively reduces high frequency components of error, while a
coarse-grid correction takes care the low frequency components. This is however not the case for the Helmholtz
equation, unless the wavenumber is suciently small, i.e., k 2 less than the smallest eigenvalue of the Laplacian.
For high wavenumbers, some eigenvalues will have a negative sign (indenite). For indenite systems, Jacobi

653

ITERATIVE SCHEMES FOR HELMHOLTZ EQUATION

Table 1. Fourier analysis asymptotic convergence for the Helmholtz equation. is the smoothing factor, 2 and 3 are the two- and three-grid convergence, respectively. Bilinear interpolation
and full weighting are used, with coarse-grid approximation based on the Galerkin operator and
one pre- and post-smoothing.
k
5
20
50

Grid
322
322
1282

0.56
0.60
0.58

2
0.56
1.82
1.05

3
0.56
2.04
4.04

iteration, for instance, does not converge. In multigrid, the smoothing factor of a basic iterative method is more
important than its convergence, and for the Helmholtz equation, Jacobi iteration still gives satisfactory smoothing properties. Due to the increase of indeniteness, its smoothing properties will deteriorate. (A convergent
Jacobi iteration can be based on two-step Jacobi iteration with complex relaxation, as discussed in [11]. But
experiments show no advantage of using this method as a smoother for multigrid as compared to a one-step
Jacobi iteration). Another source of diculty comes from the coarse-grid correction. Due to indeniteness, a
sign change may occur to the eigenvalues close to zero after coarse-grid discretization. This will eventually lead
to a convergence degradation.
A powerful tool to evaluate properties of a multigrid method is based on Fourier analysis [4,20? ]. An example
of results from Fourier analysis is shown in Table 1 for a W-cycle multigrid method with one pre- and post
point-wise Jacobi smoothing, with underrelaxation R = 0.5. (For three-grid methods, F-cycle is identical to
W-cycle). The coarse-grid correction is based on a Galerkin approximation, with bilinear interpolation and
full weighting as the prolongation and the restriction operator, respectively. For k = 5 (small wavenumber), a
convergent multigrid method can still be obtained, indicated by a satisfactory smoothing factor , and two- and
three-grid convergence, denoted by 2 and 3 , respectively. For higher wavenumbers, even though the smoothing
properties are still satisfactory ( < 1), results from Fourier analysis indicate problematic coarse-grid correction,
especially with three grids.
Having a complex wavenumber in the Helmholtz equation makes the situation quite dierent. We rst analyze
the use of damped Jacobi for the complex equation (4.2). To simplify the analysis we consider the discretization
of the one dimensional equation with central dierences. So we consider
vj+1 2vj + vj1 + (1 + i)(kh)2 vj = 0.

(4.3)

The damped Jacobi iteration for this is given by

vj+1 + vj1
2 (1 + i)(kh)2
vj = R vj + (1 R )vj

vj =

(4.4)

The amplication matrix is given by

2 2
1 + 1+i
cos()
2 k h
2
2
2 4 4
1+i 2 2 = 1 R + R cos()
1 2 k h
(1 + k 2h )2 + k4 h


2 2
R cos() 1 + k 2h
R cos()
= 1 R +
+ i
2
4 h4
(2 +1)k4 h4
2
2
2
1+k h +
1 + k h2 + ( +1)k
4
4

G() = 1 R + R

So

|G()| = 1 R +
2


R cos() 1 +
1 + k 2 h2 +

k 2 h2
2

 2

+
(2 +1)k4 h4
4

k 2 h2 R cos()
1 + k 2 h2 +

(2 +1)k4 h4
4

654

Y. ERLANGGA AND E. TURKEL

Table 2. Fourier analysis asymptotic convergence for the Helmholtz equation with complex
wavenumber. is the smoothing factor, 2 and 3 are the two- and three-grid convergence,
respectively. Bilinear interpolation and full weighting are used, with coarse-grid approximation based on the Galerkin operator and one pre- and post-smoothing with point Jacobi and
R = 0.5.

2th order
4th order
6th order
Noncompact

k = 20, 322

2
0.60 0.60
0.53 0.53
0.52 0.52
0.62 0.62

grid
3
0.59
0.73
0.52
0.61

k = 50, 1282

2
0.58 0.58
0.50 0.50
0.50 0.50
0.60 0.60

grid
3
0.58
0.50
0.50
0.60

k = 100, 5122

2
0.57 0.57
0.49 0.49
0.49 0.49
0.59 0.59

grid
3
0.57
0.49
0.49
0.59

We thus, see that only multiplies terms k 4 h4 . Hence, if kh is small enough the eect of on the convergence
is minimal. In other words changing k 2 to a complex coecient in the Helmholtz equation has a minimal eect on
the smoother on ne grids. Only on coarser grids where kh is not small does it have an eect. It is straightforward
to check that changes to the discretization of the Helmholtz equation has no eect on these conclusions.
To quantify multigrid convergence properties, we perform Fourier analysis on (4.2), with = 0.5. The
associated matrix is obtained via second, fourth, or sixth order nite dierences. Results from this analysis are
shown in Table 2. In addition to the satisfactory smoothing factor, these results now indicate no problematic
coarse-grid corrections. Fourier analysis, including for based on noncompact scheme, also shows that these
properties hold almost independently of discretizations used and the wave number k.

5. Results
We consider the following problem on a uniform Cartesian grid with mesh width h. We have Dirichlet
boundary conditions on three sides and a simplied absorbing boundary condition (ABC) on the fourth side.
The results are fairly independent of the details of the discretization of the original problem. Hence, replacing
this poor absorbing boundary condition by an improved one should have little eect on the convergence rates.
To verify the high order accuracy of the scheme we consider a system with an explicit solution:


1 1
2
2
u + k u = 0 in = [0, 1] ,
,
(5.1)
2 2




1
1
= u x,
= 0,
u x,
2
2
u(0, y) = cos(y),


u
+ iu
= 0.
x
x=1
The exact solution is u(x, y) = cos(y)eix with 2 + 2 = k 2 .
The preconditioned system is similar with k replaced by kprec in the Helmholtz equation but not the ABC.
Using this replacement also in the Robin boundary condition has a minor eect on the convergence.
We consider several discretizations of both the original and preconditioned Helmholtz equation. Using a
second (2.3) or fourth (2.4) or sixth order (2.5) (with = 0) accurate approximation for both equations will be
denoted as scheme 2, HO4, HO6 respectively. If we use a fourth order or sixth high order accurate approximation
for the original equation but second order for the preconditioned equation it will be denoted as scheme HO42,
HO62 respectively. Scheme HO64 is similarly dened.
In Table 3 we check the accuracy of the code. We reduced the residual by 10 orders of magnitude. This
was necessary to get the full accuracy on the nest mesh 512512. On coarser meshes we give the number of

655

ITERATIVE SCHEMES FOR HELMHOLTZ EQUATION

Table 3. k = 30, BICGSTAB with F cycle; damped Jacobi = .5, R = 0.4.

Scheme
HO42
HO42
HO42
HO42
HO62
HO62
HO62
HO62

Mesh
65 65
129 129
257 257
513 513
65 65
129 129
257 257
513 513

Tol
10
10
10
10
10
10
10
12

# iter
78
61
54
53
78
64
54
53

Sec
0.39
1.27
5.03
22.22
0.38
1.33
5.16
21.56

Tol
6
7
8
10
6
7
8
12

# iter
53
47
45
53
53
49
45
53

L2 error
1.08 103
7.10 105
4.57 106
2.83 107
4.21 105
1.39 106
4.54 108
5.09 109

Conv rate
3.93
3.96
3.98
4.92
4.93
4.94

(kh)2 = 1.0

(kh)2 = 0.1
1

0.9

0.9

0.8
0.8
amplification factor

amplification factor

0.7
0.6
0.5
0.4

0.7
0.6
0.5

0.3
0.4
0.2
0.3

0.1
0

0.5

1.5

2.5

3.5

0.2

0.5

1.5

theta

2.5

3.5

theta

Figure 4. Amplication factor of damped Jacobi.

iterations required to reduce the residual by 10 orders of magnitude and also to reduce the residual by enough
not to aect the nal accuracy. The number of orders of magnitude of reduction is given by tol. We clearly
see the global fourth order accuracy of HO42 including the simplied absorbing boundary condition while HO62
seems to give only fth order accuracy. In this test the Helmholtz equation was approximated by a fourth or
sixth order approximation while the preconditioner was approximated by a standard ve point second order
nite dierence (HO42/HO62). BICGSTAB was used for the solver with point Jacobi used as the smoother for
the MG of the preconditioner. To graphically see the amplication factor as a function of we plot in Figure 4
the amplication factor for the 512 512 grid.
For the Laplacian the best smoother with SOR is r = 1, i.e. Gauss-Seidel. However, for the modied
Helmholtz equation we have found that under-relaxation is required. In addition using a red-black ordering
increases the convergence rate and allows for parallelization. Hence, in Table 4 we repeat the computations but
using RB-SOR with R = 0.5 as the smoother. We see that this converges slower than with the damped Jacobi.
Lexicographic GS with the same damping factor was not as robust.
In Table 5 we consider the two smoothers for a range of wavenumbers k. All the numbers in the table are
still based on a reduction of 10 orders. This uses a fourth (or sixth) order nine point discretization of the
original equation but a ve point second order discretization of the modied (preconditioner) equation. The
preconditioned equation is solved by a F-cycle multigrid down to a 2 2 grid with one iteration of a damped
point Jacobi preconditioner with a relaxation factor, R = .5 or R = 23 . The dependence on R is very slight.

656

Y. ERLANGGA AND E. TURKEL

Table 4. k = 30, HO42, BICGSTAB with F cycle; SOR; = .7, R = 0.5.

Mesh
65 65
129 129
257 257
513 513

Tol
10
10
10
10

# iter
81
64
61
57

Sec
0.43
1.62
4.56
29.20

Table 5. BICGSTAB F cycle MG and damped Jacobi, 513 513 mesh, HO42.
k
5
10
20
30
40
50
60

.2
.4
.5
.5
.5
.5
.5

Jacobi R = .4
12
25
33
57
63
98
153

.5
.5
.5
.6
.7
.7
.7

GS R = .5
12
19
45
70
96
118
208

L2 error
1.73 1011
1.28 109
3.03 108
2.89 107
1.33 106
3.96 106
9.17 106

For the highest k the number of iterations is large but this is mainly due to a slowdown after the residual has
2
been reduced 7 orders of magnitude. We also present the optimal for kprec
= (1 + i)k 2 . We again see that
the damped Jacobi is more ecient and more robust compared with RB-SOR. In Table 6 we solve (5.1) with
BICGSTAB and the preconditioned system (4.2) with multigrid. A F cycle is used with 1 iteration of damped
Jacobi with R = 0.4 as the smoother, before and after each cycle. However, for higher values of k we found
that lower values of R is necessary. We reduce the residual by 10 orders of magnitude for all k.
We see from Table 6 that the convergence for xed k is independent of the mesh. Similarly, when we use a
standard second order ve point stencil to approximate the preconditioned equation we get the same convergence
independent of whether a second order or high order discretization is used for the original equation. We stress
that the accuracy of the total scheme is independent of the treatment of the preconditioner. As k increases, (the
imaginary part of k prec , (4.2)), the convergence slows down. For k 20 the number of iterations grows linear
in k. However, for larger k it grows faster than linear. If instead of damped Jacobi we use GS underrelaxation
then the results are very dependent on the R chosen. For large k we need to choose R = 0.5 and the number
of iterations for convergence is larger than with the damped Jacobi. In Figure 5 we display the convergence
history for the case k = 50, HO42 and a 512 512 grid. We see that the convergence slows down for smaller
residual levels. Hence, if we reduce the residual by less than 10 orders of magnitude the time required is much
less than indicated in the tables. This is typical of all the cases described here.
A main point is that the convergence rate is insensitive to the details of the discretization both of the
equation and the preconditioner. Higher and lower order accuracy schemes for both the Helmholtz equation and
the preconditioner converge at similar rates. The damping factor R has a greater inuence on the convergence
than as long as is suciently large which depends on k.
We also note that the convergence in Tables 6 and 5 depends on k. Similar behavior is also observed for lower
order scheme, as reported in [9]. This convergence degradation when k increases, is relatively mild, compared
to the unpreconditioned Bi-CGSTAB. The latter typically shows extremely slow convergence for even small
k [8, 10] and an exponential increase of the number of iterations as k increases.
In Table 7 we compare, on a 512 512 mesh, Red-Black SOR, SOR or damped Jacobi as the smoother for the
multigrid for the preconditioner. We again reduce the residual by 10 orders of magnitude. We see that RB-SOR
requires fewer iterations but more computer time. The extra CPU is caused by needing 2 loops to go through
the mesh and moving data to and from a temporary storage. However, this is very dependent on the computer
architecture.

ITERATIVE SCHEMES FOR HELMHOLTZ EQUATION

657

Table 6. BICGSTAB with F cycle, damped Jacobi.

Mesh
129 129
129 129
513 513
513 513
513 513
513 513
129 129
129 129
129 129
129 129
129 129
513 513
513 513
513 513
513 513
513 513
513 513
513 513
513 513
513 513
513 513
513 513
513 513
513 513
513 513
513 513
513 513
513 513
513 513
513 513
513 513
513 513

k
5
5
5
5
5
5
10
10
10
10
10
10
10
10
10
10
10
10
10
10
30
30
30
30
30
30
50
50
50
50
50
50

R
.4
.3
.4
.3
.3
.3
.2
.2
.4
.4
.4
.4
.2
.4
.2
.2
.2
.4
.2
.2
.4
.4
.4
.4
.4
.4
.4
.4
.4
.4
.4
.4

0
0
0
0
0
0
.2
.2
.2
.2
.2
.2
.2
.2
.4
.2
.4
.2
.2
.4
.5
.5
.5
.5
.5
.7
.5
.5
.5
.5
.5
.5

Scheme
2
HO42
2
2
HO4
HO6
2
HO4
HO4
HO42
HO6
2
2
HO4
HO4
HO4
HO42
HO42
HO42
HO6
2
HO4
HO42
HO6
HO62
HO64
2
HO4
HO42
HO6
HO62
HO64

0
0
0

0
0
0
0
0
0

0
0

0
0

# iterations
18
11
14
11
11
11
32
27
37
37
35
58
26
64
25
24
26
54
27
26
54
51
56
50
53
56
101
95
98
101
96
101

In Table 8 we compare the use of V-, F- or W-cycles within the multigrid used to solve the preconditioned
equation. We now only consider the fourth order discretization for the original equation and the ve point second
order stencil for the preconditioned equation (HO42) on a 513513 mesh. As usual we reduce the residual by
10 orders of magnitude. The V cycle is less ecient than either F and W cycles especially for larger k.
In Table 9 we examine the eect of k and the mesh on the optimal . To keep all other parameters the same
we choose the HO42 scheme with a F multigrid strategy and a damped Jacobi smoother, = 0.5. To account
for the pollution eect whenever we double k we increase the number of mesh points in each direction, N , by
5
approximately k 4 2 which should keep the error approximately constant. In this table we present various k
and the appropriate mesh and the experimentally found best (i.e. the one that minimized the number of
iterations). In many cases changing slightly had little eect on the number of iterations and so the optimal
is actually an interval around that listed in the table. The optimal seems to grow as approximately k 1/2 . We
also present the L2 and maximum errors, to indeed verify the fourth order accuracy and the pollution eect. In
Figure 5 we plot a typical convergence history for the above case with a 512 512 mesh. We see that there is

658

Y. ERLANGGA AND E. TURKEL

Table 7. BICGSTAB with F cycle and SOR or damped Jacobi smoother, mesh 513 513.
k
5
5
5
5
5
5
30
30
30
30
30
30
50
50
50
50
50
50

.2
.2
.2
.2
.2
.2
.5
.5
.5
.5
.5
.5
.5
.5
.5
.5
.5
.5

Smoother
RB-SOR
SOR
JAC
RB-SOR
SOR
JAC
RB-SOR
SOR
JAC
RB-SOR
SOR
JAC
RB-SOR
SOR
JAC
RB-SOR
SOR
JAC

R
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.8
0.4
0.4
0.8
0.4
0.5
0.5
0.5
0.5
0.5
0.5

Scheme
2
2
2
HO42
HO42
HO42
2
2
2
HO42
HO42
HO42
2
2
2
HO42
HO42
HO42

# iterations
12
12
12
12
12
12
55
106
54
57
113
57
84
122
100
84
120
88

Time (s)
6.70
5.35
5.41
6.74
5.36
5.37
28.45
41.75
21.78
29.68
44.97
20.94
42.52
47.86
39.20
42.64
46.58
34.73

Table 8. HO42, BICGSTAB with V,F,W MG cycles; RB-SOR or Jacobi, mesh 513 513.
k
5
5
5
5
5
5
30
30
30
30
30
30
50
50
50
50
50
50

0
.2
.2
0
.2
0
.5
.5
.5
.5
.5
.5
.5
.5
.5
.5
.5
.5

Smoother
RB-SOR
RB-SOR
RB-SOR
JAC
JAC
JAC
RB-SOR
RB-SOR
RB-SOR
JAC
JAC
JAC
RB-SOR
RB-SOR
RB-SOR
JAC
JAC
JAC

R
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.6
0.5
0.5
0.5
0.5
0.5

Cycle
V
F
W
V
F
W
V
F
W
V
F
W
V
F
W
V
F
W

# iterations
11
12
12
12
14
16
123
57
62
108
57
65
151
84
79
160
88
81

Time (s)
5.84
7.23
7.53
4.95
5.60
5.03
56.31
30.94
35.35
37.56
23.51
28.11
99.37
45.47
63.34
77.40
50.14
48.29

initially a slow convergence. After settling down we have a rapid convergence until the residual is reduced by
about 6 orders of magnitude. Then the convergence slows down to achieve a greater reduction in the residual.
We nally did the computations with a variable k. Thus, we now solve (5.1) with k(x, y) = k0 g(x, y) and

g(x, y) =

1
1 + f sin(px) sin(qy)

near boundary
in the interior f = 0.05

659

ITERATIVE SCHEMES FOR HELMHOLTZ EQUATION

Table 9. HO42, BICGSTAB with F,cycle and damped Jacobi. Optimal and error for various
k with N = k 5/4 . mesh 513 513.
k
4
8
16
32
64
128
256

Mesh N N
88
16 16
38 38
88 88
210 210
512 512
1210 1210

0.0
0.15
0.35
0.35
0.55
0.85
1.08

# iterations
8
13
32
70
119
104
177

L2 error
4.06104
2.52104
4.23104
4.09104
4.30104
4.31104
4.33104

Max error
6.73 104
9.00 104
1.04 103
1.06 103
1.09 103
1.06 103
1.07 103

Table 10. Variable k same parameters as Table 9.

k
4
8
16
32
64
128
256

Mesh N N
88
16 16
38 38
88 88
210 210
512 512
1210 1210

Optimal
0.00
0.08
0.30
0.45
0.55
0.70
1.10

# iterations
7
13
37
77
177
1590
263

HO62 513x513 mesh, F cycle, damped Jacobi

2

log residual

10
0

10

20

30

40

50
60
# iterations

70

80

90

100

Figure 5. typical convergence, HO42, F cycle, damped Jacobi, 512 512 mesh.
p and q are chosen so that there is a complete sine wave perturbation in the interior in a centered subdomain with
half the width and length of the entire domain. We ran the exact parameters used in Table 9. The results are
presented in Table 10. We see that approximately the same optimal is needed but that the number of iterations
required increases. With greater perturbations than f = 0.05 the residuals only converged about 5 orders of
magnitude, for large k, rather than the previous 7 orders of magnitude reduction. For k up to about 16 we could
considerably increase the amount of variability (f ) without any adverse eects. Note, that the perturbation is
multiplicative and so smaller in absolute numbers for smaller k. Since only kh appears in the discretization a
nonuniform mesh has a similar eect to a variable wavenumber.

660

Y. ERLANGGA AND E. TURKEL

6. Conclusion
We have considered second, fourth and sixth order compact discretizations of the Helmholtz equation coupled
with an absorbing boundary condition. The resultant linear system has complex coecients, is not self-adjoint
and not positive denite. The system is solved using BICGSTAB. To accelerate the convergence we employ a
right preconditioner using a modied Helmholtz equation with a complex wavenumber. This modied equation
is then solved by multigrid going down to a 2 2 coarse grid.
The convergence rate of the resultant scheme is relatively independent of the grid. Furthermore, if the
preconditioned is discretized by a 5-point second-order stencil then the convergence rate is independent of the
discretization of the original equation which determines the nal accuracy. The multigrid scheme uses only
1 iteration of a pre and post smoother. Using a SOR scheme for the preconditioner uses fewer iterations but
more CPU time than a point damped Jacobi scheme. Both the SOR and damped Jacobi work best with a
relaxation parameter R = .5. Because of the complex wavenumber this relaxation factor is dierent than that
appropriate for the Laplace equation. A Fourier analysis is presented to justify the computational results.
Acknowledgements. E. Turkel was supported by a grant from the Ministry of Science, Culture & Sport, Israel & the
Ministry of Research, France.

References
[1] I.M. Babuska and S.A. Sauter, Is the pollution eect of the FEM avoidable for the Helmholtz equation considering high wave
numbers? SIAM Rev. 42 (2000) 451484.
[2] A. Bayliss, C.I. Goldstein and E. Turkel, An iterative method for the Helmholtz equation. J. Comput. Phys. 49 (1983) 443457.
[3] A. Bayliss, C.I. Goldstein and E. Turkel, On accuracy conditions for the numerical computation of waves. J. Comput. Phys.
59 (1985) 396404.
[4] A. Brandt, Multi-level adaptive solution to the boundary- value problems. Math. Comp. 31 (1977) 333-390.
[5] A. Brandt and I. Livshits, Remarks on the wave-ray Multigrid Solvers for Helmholtz Equations, Computational Fluid and
Solid Mechanics, edited by K.J. Bathe. Elsevier (2003) 18711871.
[6] H.C. Elman and D.P. OLeary, Ecient iterative solution of the three dimensional Helmholtz equation. J. Comput. Phys. 142
(1998) 163181.
[7] Y.A. Erlangga, Advances in iterative methods and preconditioners for the Helmholtz equation. Arch. Comput. Methods Eng.
15 (2008) 3766.
[8] Y.A. Erlangga, C. Vuik and C.W. Oosterlee, On a class of preconditioners for the Helmholtz equation. Appl. Numer. Math.
50 (2004) 409425.
[9] Y.A. Erlangga, C.W. Oosterlee and C. Vuik, A novel multigrid based preconditioner for heterogeneous Helmholtz problems.
SIAM J. Sci. Comput. 27 (2006) 14711492.
[10] Y.A. Erlangga, C. Vuik and C.W. Oosterlee, Comparison of multigrid and incomplete LU shifted-Laplace preconditioners for
the inhomogeneous Helmholtz equation. Appl. Numer. Math. 56 (2006) 648666.
[11] G.R. Hadley, A complex Jacobi iterative method for the indenite Helmholtz equation. J. Comput. Phys. 203 (2005) 358370.
[12] I. Harari and E. Turkel, Accurate nite dierence methods for time-harmonic wave propagation. J. Comput. Phys. 119 (1995)
252270.
[13] I. Singer and E. Turkel, High order nite dierence methods for the Helmholtz equation. Comput. Meth. Appl. Mech. Eng.
163 (1998) 343358.
[14] I. Singer and E. Turkel, Sixth order accurate nite dierence schemes for the Helmholtz equation. J. Comp. Acous. 14 (2006)
339351.
[15] H. Tal-Ezer and E. Turkel, Iterative Solver for the Exterior Helmholtz Problem. SIAM J. Sci. Comput. 32 (2010) 463475.
[16] E. Turkel, Numerical methods and nature. J. Sci. Comput. 28 (2006) 549570.
[17] E. Turkel, Boundary Conditions and Iterative Schemes for the Helmholtz Equation in Unbounded Regions, Computational
Methods for Acoustics Problems, edited by F. Magoules. Saxe-Coburg Publ. UK (2008).
[18] H.A. van der Vorst, Bi-CGSTAB: A fast and smoothly converging variant of BI-CG for the solution of nonsymmetric linear
systems. SIAM J. Sci. Stat. Comput. 13 (1992) 631644.
[19] M.B. van Gijzen, Y.A. Erlangga and C. Vuik, Spectral analysis of the discrete Helmholtz operator preconditioned with a shifted
Laplace precondtioner. SIAM J. Sci. Comput. 29 (2006) 19421958.
[20] R. Wienands, C.W. Oosterlee, On three-grid Fourier analysis for multigrid. SIAM J. Sci. Comput. 22 (2001) 651671.