Sunteți pe pagina 1din 10

COMMUNICATIONS IN NUMERICAL METHODS IN ENGINEERING

Commun. Numer. Meth. Engng 2009; 25:237–246


Published online 28 March 2008 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/cnm.1123

Exact integration of polynomial–exponential products


with application to wave-based numerical methods

G. Gabard∗, †
Institute of Sound and Vibration Research, University of Southampton, Southampton SO17 1BJ, U.K.

SUMMARY
Wave-based numerical methods often require to integrate products of polynomials and exponentials. With
quadrature methods, this task can be particularly expensive at high frequencies as large numbers of
integration points are required. This paper presents a set of closed-form solutions for the integrals of
polynomial–exponential products in two and three dimensions. These results apply to arbitrary polygons
in two dimensions, and for arbitrary polygonal surfaces or polyhedral volumes in three dimensions.
Quadrature methods are therefore not required for this class of integrals that can be evaluated quickly and
exactly. Copyright q 2008 John Wiley & Sons, Ltd.

Received 29 July 2007; Revised 22 November 2007; Accepted 17 February 2008

KEY WORDS: wave-based numerical methods; quadrature; highly oscillatory integrals

1. INTRODUCTION

The present work is motivated by the current development of wave-based numerical methods for
acoustics, electromagnetism and other wave propagation problems. Standard numerical methods
(such as finite elements or finite differences) rely in one way or another on polynomial interpolation
to approximate the solution and its derivatives. For high frequencies, these methods become less
efficient, and alternatives have been actively developed to perform efficient numerical simulations
at high frequencies. One approach is to use local solutions of the problem at hand, such as plane
waves, to build the numerical approximations of the global solution. An overview of wave-based
methods is given by Bettess [1, Chapter 12]. Examples of wave-based numerical methods include
the partition of unity method originally proposed by Melenk and Babuška [2, 3] and which has
since received particular attention [4–6], the discontinuous enrichment method [7], the ultra-weak

∗ Correspondence to: G. Gabard, Institute of Sound and Vibration Research, University Road, Southampton SO17 1BJ,
U.K.

E-mail: gabard@soton.ac.uk

Copyright q 2008 John Wiley & Sons, Ltd.


238 G. GABARD

variational formulation [8] and the wave-based discontinuous Galerkin method [9]. Boundary
element methods using plane-wave interpolations are also under active development [10].
A significant issue when implementing some wave-based methods is the need for integrating
products of polynomials and exponentials. This task can be particularly demanding as a large
number of these integrals have to be evaluated for each element, and in some cases this task is
actually more expensive than solving the linear system of equations. An aggravating issue is that
wave-based methods are generally prone to ill-conditioning. As a consequence, the solution of the
numerical model is sensitive to any change in the entries of the element matrices. Hence, in order
not to introduce significant error in the solution, quadrature schemes must be used with a very
high level of accuracy (preferably up to the computer accuracy).
Numerical integration is obviously a very common task in computational engineering and a
vast literature is devoted to the subject. For univariate integrals with relatively smooth integrands,
standard and inexpensive methods are available such as Gauss–Legendre or Clenshaw–Curtis
quadrature techniques. However, these methods perform poorly for integrands oscillating rapidly.
In the literature, highly oscillatory integrands are generally expressed as f (x)eig(x) , where  is a
large constant. Various techniques have been developed specifically for this class of integrals. Evans
and Webster [11] provide a summary and a comparison of different extensions of the Clenshaw–
Curtis method to highly oscillatory integrals with linear oscillators, i.e. with g(x) = x. For irregular
oscillators (that is for a general phase function g(x)), Iserles and Nørsett have proposed several
techniques where the integral is approximated using the derivatives of f at the end points [12]. A
distinctive feature of the quadrature methods introduced in [12] is that they actually take advantage
of the oscillatory nature of the integrand as their accuracy increases as  → ∞. Following a
similar approach, Olver has devised a modified version of the Levin method featuring the same
property [13].
Comparatively, limited efforts have been put into the development of efficient quadrature methods
for multivariate integrals. When the integration domain allows for a separation of variables, it
is possible to rely on standard univariate integration schemes using the product rules. There are
no general methods for more complex integration domains. A notable exception is the extension
by Iserles and Nørsett of their quadrature schemes to general polytopes (e.g. polygons in two
dimensions and polyhedrons in three dimensions) [14]. It is also worth mentioning the analytical
result obtained by Gordon in the context of wave scattering through apertures [15]. This analytical
solution is, however, limited to the integral of an exponential over a polygonal surface in three
dimensions. The present work is concerned with the more general case of polynomial–exponential
products integrated over surfaces and volumes.
A few integration techniques have been proposed specifically for wave-based numerical methods.
Bettess et al. [16, 17] have proposed quadrature methods for rectangular and triangular elements. In
a paper on the partition of unity finite element method, Ortiz and Sanchez [18] outlined a method
where the domain of integration is rotated so that the exponential behaviour of the integrand is
limited to one direction. The integration along that direction is carried out using an extension of
the Clenshaw–Curtis integration scheme, while in the other direction one is left with a standard
integral of a polynomial. However, this method has only been applied to linear triangular elements
and is difficult to extend to three dimensions.
This paper addresses the issue of integrating polynomial–exponential products by deriving exact
expressions for these integrals in the case where the domain of integration is polygonal. This
includes arbitrary polygons in two dimensions and arbitrary polygonal surfaces or polyhedral
volumes in three dimensions. For these types of integrals, it is shown that there is no need to use

Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2009; 25:237–246
DOI: 10.1002/cnm
INTEGRATION OF POLYNOMIAL–EXPONENTIAL PRODUCTS 239

expensive quadrature schemes as exact solutions can be obtained in closed form. The remainder of
this paper is as follows. In the following section, the closed-form solutions for the two-dimensional
integrals of polynomial–exponential products are first presented. These results are then extended
in Sections 3 and 4 to surface and volume integrals in three dimensions. Some guidelines to
implement efficiently these solutions are given and some examples of timing are presented.

2. TWO-DIMENSIONAL INTEGRALS

We are interested in evaluating the following two-dimensional integral in closed form:




M  N
I = P(x)ek·x dx with P(x) = Amn x m y n (1)
S m=0 n=0

where the coefficients Amn of the polynomial are complex valued. The vector notation x = (x, y)
is used and the domain of integration S is an arbitrary polygonal region of R2 . Note that the vector
k = (, ) can also be complex valued. For instance, in the context of wave-based methods for
the Helmholtz equation, one uses k = ik̂ where k̂ is the real-valued wave number of a progressive
acoustic wave.

2.1. From surface integrals to boundary integrals


The first step is to rewrite the surface integral over the polygon into a line integral along its
boundary. To that end the integrand is substituted by the divergence of a vector field, i.e. we need
to choose f such that

M 
N
∇ ·f = Amn x m y n ex+ y (2)
m=0 n=0

Several choices for this vector field are possible and the following form will be considered:

M 
N
f= Fmn x m y n ex+ y a (3)
m=0 n=0

where the particular choice of vector a = (a, b) remains to be decided. By substituting this expres-
sion in Equation (2) and by equating with the coefficients of the polynomials, one finds that
Amn = (a·k)Fmn +a(m +1)Fm+1,n +b(n +1)Fm,n+1
The coefficients Fmn of the vector field are therefore given by a simple recurrence formula
Fmn = [Amn −a(m +1)Fm+1,n −b(n +1)Fm,n+1 ]/(a·k) (4)
where we have Fmn = 0 whenever m>M or n>N . This result holds for any a and k except when
a·k = 0. This special case is treated in Section 2.3. Now that the integrand has been expressed as
the divergence of a vector field, we can use Gauss’ theorem to obtain a contour integral


M  N
I= Fmn x m y n ex+ y a·n dC
m=0 n=0 C

Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2009; 25:237–246
DOI: 10.1002/cnm
240 G. GABARD

where C = *S is the boundary of the integration domain (note that C is a simple polygon) and n
is the outward unit normal on C. One is left with integrals of polynomial–exponential products
along the straight lines that form the boundary of S.

2.2. Evaluation of boundary integrals


We now focus on the contribution J of one of the straight lines along the polygon C. Let x0
and x0 +x1 denote the start and end vertices of the line, with x0 = (x0 , y0 ) and x1 = (x1 , y1 ). The
normal along this line being constant, one has to evaluate an integral of the form
 1
k·x0  
M N
J = a·nLe Fmn (x0 +sx1 )m (y0 +sy1 )n es  ds
m=0 n=0 0

where  = k·x1 and L is the length of the line. Using the binomial theorem, one finds that
  
k·x0   
M N m  n m n m−u n−v u v
J = a·nLe Fmn x y0 x1 y1 Q u+v () (5)
m=0 n=0 u=0 v=0 u v 0
where we have also defined that
 1
Q n () = s n es ds (6)
0

We need to evaluate Q n for all values of n from 0 to M + N . Evans and Webster studied this
particular problem and carried out a comparison of different methods [11]. They considered various
schemes based on extended Clenshaw–Curtis methods. For the partition of unity method, Ortiz
and Sanchez used an extension of the Clenshaw–Curtis method proposed by Alaylioglu, Evans
and Hyslop to evaluate a similar integral [18]. However, the results obtained by Evans and Webster
show that the use of recurrence formulae is consistently faster and more stable than other methods.
Therefore, the following recurrence formula is used in the present work:

Q n () = [e −n Q n−1 ()]/ and Q 0 () = (e −1)/ (7)


When  = 0 this recurrence relation does not apply, but instead we simply have Q n (0) = 1/(n +1).

2.3. Choice of the vector field direction


The choice of the vector field direction a leads to different integration formulae.
A first possible choice is to use one of the unit vector axes, for instance, a = (1, 0). In that case,
a·k =  and the recurrence formula (4) is significantly simplified as b = 0. Equation (5) can also be
simplified by noting that a·nL = y1 . This choice of vector a is, however, not possible when  = 0
as in that case the recurrence relation (4) is not valid. A simple way to fix this is to use a = (0, 1)
instead.
Another possibility is to define a = k̄/|k|2 where the overbar denotes the complex conjugate. We
then have a·k = 1 and the recurrence formula (4) is simplified. Equation (5) can also be simplified
¯ 1 )/|k|2 . Although a bit more complex, this choice is valid for any
by noting that a·nL = (¯ y1 − x
non-zero vector k.
Finally, when  =  = 0 the problem reduces to integrating a polynomial. For triangles and
quadrilaterals, standard quadrature methods can then be used.

Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2009; 25:237–246
DOI: 10.1002/cnm
INTEGRATION OF POLYNOMIAL–EXPONENTIAL PRODUCTS 241

3. THREE-DIMENSIONAL SURFACE INTEGRALS

We now extend our analysis to three-dimensional integrals of the form




M  N  R
I = P(x)ek·x dS(x) with P(x) = Am,n,r x m y n z r (8)
S m=0 n=0 r =0

where the wave-number vector k = (, , ) and the coefficients Amnr of the polynomial are complex
valued. The domain of integration S is an arbitrary polygonal planar surface in R3 . Using Stokes’
theorem, this surface integral can be rewritten as an integral along its contour. To that end we have
to write the integrand in the form n·(∇ ×g), where g is a vector field and n is the normal on S.
Again there are several possible choices for g. Here we use the following expression:

M 
N 
R
g = aek·x G m,n,r x m y n z r
m=0 n=0 r =0

where a = (a, b, c) is a constant. One finds that



M 
N 
R
n·(∇ ×g) = ek·x x m y n z r [a·(n×k)G m,n,r +(m +1)(bn z −cn y )G m+1,n,r
m=0 n=0 r =0

+(n +1)(cn x −an z )G m,n+1,r +(r +1)(an y −bn x )G m,n,r +1 ]


Then by identifying the coefficients of the polynomials in the integrand one obtains the following
recurrence formula that defines G m,n,r :
G m,n,r = [Am,n,r −(m +1)(bn z −cn y )G m+1,n,r
−(n +1)(cn x −an z )G m,n+1,r −(r +1)(an y −bn x )G m,n,r +1 ]/[a·(n×k)] (9)
where G m,n,r = 0 whenever m>M, n>N or r >R. The possible choices of vector a will be discussed
in Section 3.2. Using Stokes’ theorem, the integral over the polygonal surface can be expressed
as an integral on the contour


M  N R
I= G m,n,r x m y n z r ek·x a·ds
m=0 n=0 r =0 *S

where s is the unit tangential vector along the contour. Gordon has used a similar strategy to obtain
a closed-form solution for the integral of an exponential over a polygonal surface [15]. The present
result is a generalization to integrals of polynomial–exponential products and the key to obtain a
closed-form solution in that case is the recurrence relation (9).

3.1. Evaluation of contour integrals


We now consider the contribution J of one segment of the polygonal contour *S. Let x0 =
(x0 , y0 , z 0 ) and x1 = (x1 , y1 , z 1 ) such that x0 and x0 +x1 denote the start and end vertices of the
segment. Using the same approach as in Section 2.1, the integral along the segment can be evaluated
in the closed form
 1
k·x0   
M N R
J = Ls·ae G m,n,r (x0 +sx1 )m (y0 +sy1 )n (z 0 +sz 1 )r es  ds
m=0 n=0 r =0 0

Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2009; 25:237–246
DOI: 10.1002/cnm
242 G. GABARD

with  = k·x1 . Using the binomial formula, one finds that


M 
N 
R
J = x1 ·aek·x0 G m,n,r
m=0 n=0 r =0
   

m 
n 
r m n r m−u n−v r −w u v w
× x0 y0 z 0 x1 y1 z 1 Q u+v+w () (10)
u=0 v=0 w=0 u v w

where we have used Ls = x1 and Q is given by (6).

3.2. Choice of the vector field direction


Again several choices are possible for vector a.
A first possible choice is to use one of the unit vector axes, for instance a = (1, 0, 0). In that
case, a·k =  and the recurrence formula (4) is significantly simplified as b = c = 0. This choice
of vector a is, however, not possible when  = 0, as in that case the recurrence relation (4) is not
valid. A simple way to fix this is to use either a = (0, 1, 0) or a = (0, 0, 1) instead, provided that
  = 0 or   = 0.
Another possibility is to define a = n× k̄/|k|2 where the overbar denotes the complex conjugate.
We have then a·(n×k) = 1 and the recurrence formula (4) is simplified. The advantage of this
definition of a is that it is valid for any direction of the wave-number vector k as long as |k|  = 0.

4. THREE-DIMENSIONAL VOLUME INTEGRALS

Finally, the present analysis can also be extended to three-dimensional integrals of the form

M  N R
I = P(x)ek·x dx with P(x) = Am,n,r x m y n z r
V m=0 n=0 r =0

where the wave-number vector k = (, , ) and the coefficients Amnr of the polynomial are complex
valued. The domain of integration V ⊂ R3 is an arbitrary polyhedron.
We follow the same approach as in Section 2.1 and the integrand is expressed as the divergence
of a vector field h of the form

M 
N 
R
h = aek·x Hm,n,r x m y n z r (11)
m=0 n=0 r =0

with a = (a, b, c). For the divergence of this vector field to correspond to the integrand, the following
expression has to be satisfied:

Am,n,r = (a·k)Hm,n,r +a(m +1)Hm+1,n,r +b(n +1)Hm,n+1,r +c(r +1)Hm,n,r +1

As for two-dimensional integrals, it is found that the coefficients Hm,n,r of the vector field are
obtained by applying a recurrence relation

Hm,n,r = [Am,n,r −a(m +1)Hm+1,n,r −b(n +1)Hm,n+1,r −c(r +1)Hm,n,r +1 ]/(a·k) (12)

Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2009; 25:237–246
DOI: 10.1002/cnm
INTEGRATION OF POLYNOMIAL–EXPONENTIAL PRODUCTS 243

where we have Hmnr = 0 whenever m>M, n>N or r >R. The divergence theorem then yields

M N  R
I= Hm,n,r x m y n z r ek·x a·ndS (13)
m=0 n=0 r =0 *V

where the boundary *V of the polyhedron is composed of polygonal surfaces. The integral on
each of these surfaces can be evaluated in closed form using the results from Section 3.

5. IMPLEMENTATION

Expressions (5) for the two-dimensional integrals and expressions (10) and (13) for the three-
dimensional integrals appear rather complicated. However, a closer inspection reveals that several
quantities can be calculated in advance, stored and then used when necessary in the nested sums.
By taking the following remarks into account, one can easily reduce the cost of evaluating (5),
(10) and (13).
First note that the binomials used in (5) and (10) are known constants and hence do not require
additional calculations.
Second, it is important to minimize the use of time-consuming functions such as exponentials,
square roots, etc. A benefit of using the recurrence formula (7) to evaluate integrals (6) is that it
minimizes the evaluation of exponentials by requiring only a single evaluation of e (which can
also be stored for further use). This is in contrast with the AEH formulae used in Reference [18]
where several evaluations of complex sine and cosine functions are required.
Similarly, there is no need to use the function power to calculate x 1u , y1u , etc. As all the values
of x1u for 0<u<M are required, it is preferable to evaluate and store these values using the simple
recurrence formula x1u+1 = x1 x1u . The values of x1u can then be used efficiently in the nested sums.
Finally, it is also possible to minimize the use of the square root function. For instance, for
the two-dimensional integral (5), one has a·nL = y1 or a·nL = (¯ y1 − x ¯ 1 )/|k|2 depending on the
choice of vector a. There is therefore no need to evaluate the unit normal n on the boundary and
this avoids the calculation of a square root for normalizing the normal.
When these remarks are taken into account, the present algorithm requires only simple
operands (multiplications, additions, etc.) and a limited number of evaluations of the exponential
function.

6. EXAMPLES

When used for evaluating element matrices of finite element models, such as the partition of
unity method, for instance, the present integration scheme can be used several millions times. It
is therefore important to illustrate the cost of calculating integrals (1) and (8). The time needed to
evaluate these equations has been measured for different orders of polynomials and for different
numbers of edges on the boundary of the polygon, see Figures 1 and 2. The timings were carried
out by making one million runs for each case with a randomly chosen wave number for each run.
It appears that the cost of an evaluation varies linearly with the number of edges. The time also
increases with the polynomial order. However, overall the evaluation of the integrals is quite quick
and several millions of evaluations can be carried out in a few seconds.

Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2009; 25:237–246
DOI: 10.1002/cnm
244 G. GABARD

25 25

20 20

time (µs)
15 15
time (µs)

10 10

5 5

0 0
2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4 3 3.5 4 4.5 5 5.5 6
polynomial order number of edges

Figure 1. Time in microseconds per two-dimensional integration as a function of the polynomial


order M = N (left) and as a function of the number of edges of the polygon (right). Left: number
of edges 3 (solid line), 4 (dashed line), 5 (dot-dashed line), 6 (dotted line). Right: polynomial
order 2 (solid line), 3 (dashed line), 4 (dot-dashed line).

250 250

200 200

150 150
time (µs)

time (µs)

100 100

50 50

0 0
2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4 3 3.2 3.4 3.6 3.8 4 4.2 4.4 4.6 4.8 5
polynomial order number of edges

Figure 2. Time in microseconds per three-dimensional surface integration as a function of the


polynomial order M = N (left) and as a function of the number of edges of the polygonal
surface (right). Left: number of edges 3 (solid line), 4 (dashed line), 5 (dot-dashed line). Right:
polynomial order 2 (solid line), 3 (dashed line), 4 (dot-dashed line).

To put these results in context, a simple calculation with a wave-based finite element method
is now presented as an illustration. The two-dimensional Helmholtz equation is solved using the
partition of unity method [3]. The computational domain and the mesh are shown in Figure 3. A
source of sound is generated along the lower boundary y = 0, where the normal acoustic velocity
is given by exp(−x 2 /2 ) with  = 0.05. This induces a simple monopole source of sound radiating
from the origin. The Bayliss–Turkel non-reflecting boundary condition is used on the circular
boundary so as to represent the radiation of sound to the far field. The sound speed and the angular
frequency are taken to be c0 = 1 and  = 60.

Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2009; 25:237–246
DOI: 10.1002/cnm
INTEGRATION OF POLYNOMIAL–EXPONENTIAL PRODUCTS 245

–3
x 10
5

3
2 2

acoustic field
1
1.5
0

1 –1
y

–2
0.5
–3

0 –4
–2 –1.5 –1 –0.5 0 0.5 1 1.5 2 –2 –1.5 –1 – 0.5 0 0.5 1 1.5 2
x x

Figure 3. Left: finite element mesh. Right: solutions for the radiation problem with different numbers of
Gauss points. Dashed line: Ng =9. Dot-dashed line: Ng =10. Dotted line: Ng =11. Solid line: exact solution.

This represents a particularly stringent test case for integration schemes as each element contains
almost two wavelengths. The mesh contains 506 triangular elements and the finite element basis
is enriched with 24 plane waves at each node. For this particular test case, more than 2.6 millions
integrations of polynomial–exponential products are required. The time needed to compute and
assemble the element matrices was measured to be 14.1 s. To provide some elements of comparison,
the same calculation is performed using Gauss integration. The number of Gauss points in one
dimension was varied from Ng = 9 to Ng = 11, corresponding to a total of Ng (Ng +1)/2 Gauss
points in each element. With Ng = 9, 10, 11 the time needed to compute the element matrices is
29.3, 35.2 and 41.5 s, respectively. The solutions obtained are also shown in Figure 3. It can be
seen that Ng = 9 is not enough to obtain a correct solution. Furthermore, it was found that the
numerical solution is independent of the number of Gauss points for Ng = 11 and above. With
Ng = 11 (that is a total of 66 Gauss points per element), the numerical solution is the same as that
obtained with the exact integration formulae derived in the present paper. With these formulae,
the time used for the element matrices calculation can be reduced by a factor of 3 compared with
Gauss integration.

7. CONCLUDING REMARKS

Exact expressions have been derived for the integrals of products of polynomials and exponentials.
These results can be used for polygonal two- and three-dimensional surface integrals and for three-
dimensional integrals on polyhedral volumes. Using these closed-form solutions, there is no need
for quadrature methods that can be time consuming and can introduce some additional numerical
errors.
It should be noted that when used for wave-based numerical methods, there is no need to work
in the coordinate system of the reference element, the integration is directly carried out in the
physical coordinate system. When implemented with care, these exact solutions require a limited
number of operations. Furthermore, the computational cost for these calculations is independent
of the wave number used in the exponential.

Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2009; 25:237–246
DOI: 10.1002/cnm
246 G. GABARD

Extending the present results to non-polyhedral domains of integration should also be possible.
This will, however, involve irregular oscillators, that is, exponentials with arguments depending
non-linearly with the argument of the integrals.

ACKNOWLEDGEMENTS
The author is grateful to one of the referees for pointing out Reference [15].

REFERENCES
1. Zienkiewicz OC, Taylor RL, Nithiarasu P. The Finite Element Method for Fluid Dynamics. Elsevier: Amsterdam,
2005.
2. Melenk JM, Babuška I. The partition of unity finite element method: basic theory and applications. Computer
Methods in Applied Mechanics and Engineering 1996; 139:289–314.
3. Babuška I, Melenk JM. The partition of unity method. International Journal for Numerical Methods in Engineering
1997; 40:727–758.
4. Laghrouche O, Bettess P, Perrey-Debain E, Trevelyan J. Plane wave basis finite-elements for wave scattering in
three dimensions. Communications in Numerical Methods in Engineering 2003; 19:715–723.
5. Sugimoto R, Bettess P. Coupling of mapped wave infinite elements and plane wave basis finite elements for
the Helmholtz equation in exterior domains. Communications in Numerical Methods in Engineering 2003;
19:761–777.
6. Bettess P. Special wave basis finite elements for very short wave refraction and scattering problems.
Communications in Numerical Methods in Engineering 2004; 20:291–298.
7. Farhat C, Harari I, Franca LP. The discontinuous enrichment method. Computer Methods in Applied Mechanics
and Engineering 2001; 190:6455–6479.
8. Cessenat O, Després B. Application of an ultra weak variational formulation of elliptic PDEs to the two-
dimensional Helmholtz problem. SIAM Journal on Numerical Analysis 1998; 35:255–299.
9. Gabard G. Discontinuous Galerkin methods with plane waves for time-harmonic problems. Journal of
Computational Physics 2007; 225:1961–1984.
10. Perrey-Debain E, Trevelyan J, Bettess P. New special wave boundary elements for short wave problems.
Communications in Numerical Methods in Engineering 2002; 18:259–268.
11. Evans GA, Webster JR. A comparison of some methods for the evaluation of highly oscillatory integrals. Journal
of Computational and Applied Mathematics 1999; 112:55–69.
12. Iserles A, Nørsett SP. Efficient quadrature of highly oscillatory integrals using derivatives. Proceedings of the
Royal Society of London, Series A 2005; 461:1383–1399.
13. Olver S. Moment-free numerical integration of highly oscillatory functions. IMA Journal of Numerical Analysis
2006; 26:213–227.
14. Iserles A, Nørsett SP. Quadrature methods for multivariate highly oscillatory integrals using derivatives.
Mathematics of Computation 2006; 75:1233–1258.
15. Gordon W. Far-field approximation to the Kirchhoff–Helmholtz representations of scattered fields. IEEE
Transactions on Antennas and Propagation 1975; 23:590–592.
16. Bettess P, Shirron J, Laghrouche O, Peseux B, Sugimoto R, Trevelyan JA. A numerical integration scheme for
special finite elements for the Helmholtz equation. International Journal for Numerical Methods in Engineering
2003; 56:531–552.
17. Sugimoto R, Bettess P, Trevelyan J. A numerical integration scheme for special quadrilateral finite elements for
the Helmholtz equation. Communications in Numerical Methods in Engineering 2003; 19:233–245.
18. Ortiz P, Sanchez E. An improved partition of unity finite element model for diffraction problems. International
Journal for Numerical Methods in Engineering 2001; 50:2727–2740.

Copyright q 2008 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2009; 25:237–246
DOI: 10.1002/cnm

S-ar putea să vă placă și