A sparse discretisation for integral equation formulations of high frequency scattering problems

Daan Huybrechs and Stefan Vandewalle Report TW 447, January 2006 (revised December 2006)

Katholieke Universiteit Leuven

A sparse discretisation for integral equation formulations of high frequency scattering problems
Daan Huybrechs and Stefan Vandewalle Report TW 447, January 2006 (revised December 2006)

Department of Computer Science, K.U.Leuven

Abstract We consider two-dimensional scattering problems, formulated as an integral equation dened on the boundary of the scattering obstacle. The oscillatory nature of high-frequency scattering problems necessitates a large number of unknowns in classical boundary element methods. In addition, the corresponding discretisation matrix of the integral equation is dense. We formulate a boundary element method with basis functions that incorporate the asymptotic behaviour of the solution at high frequencies. The method exhibits the eectiveness of asymptotic methods at high frequencies with only few unknowns, but retains accuracy for lower frequencies. New in our approach is that we combine this hybrid method with very eective quadrature rules for oscillatory integrals. As a result, we obtain a sparse discretisation matrix for the oscillatory problem. Moreover, numerical experiments indicate that the accuracy of the solution actually increases with increasing frequency. The sparse discretisation applies to problems where the phase of the solution can be predicted a priori, for example in the case of smooth and convex scatterers.

Keywords : oscillatory integral, steepest descent, numerical integration AMS(MOS) Classication : Primary : 45B05, Secondary : 65D30, 41A60.

A sparse discretisation for integral equation formulations of high frequency scattering problems
Daan Huybrechs and Stefan Vandewalle Katholieke Universiteit Leuven Department of Computer Science Celestijnenlaan 200A, B-3001 Leuven, Belgium. {Daan.Huybrechs,Stefan.Vandewalle}@cs.kuleuven.be

Abstract We consider two-dimensional scattering problems, formulated as an integral equation dened on the boundary of the scattering obstacle. The oscillatory nature of highfrequency scattering problems necessitates a large number of unknowns in classical boundary element methods. In addition, the corresponding discretisation matrix of the integral equation is dense. We formulate a boundary element method with basis functions that incorporate the asymptotic behaviour of the solution at high frequencies. The method exhibits the eectiveness of asymptotic methods at high frequencies with only few unknowns, but retains accuracy for lower frequencies. New in our approach is that we combine this hybrid method with very eective quadrature rules for oscillatory integrals. As a result, we obtain a sparse discretisation matrix for the oscillatory problem. Moreover, numerical experiments indicate that the accuracy of the solution actually increases with increasing frequency. The sparse discretisation applies to problems where the phase of the solution can be predicted a priori, for example in the case of smooth and convex scatterers.

Introduction

The accurate numerical modelling of physical problems involving strong oscillations is a challenging problem. Scattering problems in unbounded domains are often modelled by an integral equation dened on the boundary of the scattering obstacle. As such, the problem on an unbounded domain is reduced to a lower-dimensional problem on a bounded domain. Numerically, these are two important advantages. Still, the method has its drawbacks and diculties arise as the frequency of the problem increases. Contrary to the case for partial dierential equations, the discretisation matrix of an integral equation is dense. Furthermore, in order to represent an oscillatory solution, the number of unknowns in a boundary element approach has to be large. Typically, one chooses a xed number of unknowns per wavelength per dimension. This results in a linear system with a very large and dense discretisation matrix. Hence, classical solution methods for scattering problems rapidly become prohibitively expensive [4]. Substantial eorts have been made over the last two decades to overcome these diculties. One direction has been to improve on the solution time required to solve the dense linear system. The Fast Multipole Method achieves a fast matrix-vector product in O(N log N ) 1

operations, when the number of unknowns N increases at least linearly with the wavenumber [30]. The fast matrix-vector product can be combined with a preconditioned iterative technique for an ecient overall solution algorithm [10]. Another direction is given by a number of asymptotic methods, such as geometrical optics, physical optics and the geometrical theory of diraction [17, 23]. A common characteristic of asymptotic methods is that they have an error of the order O(kn ), where k is the wavenumber and where the exponent n is typically equal to 1 or 2. This means that the accuracy of such methods improves with increasing frequencies. Asymptotic methods break down for low to moderate frequencies however. A more recent trend is the combination of nite element methods with asymptotic methods. This is achieved by considering basis functions that are, e.g., piecewise polynomial, multiplied by the asymptotic form of the solution at large frequencies. The asymptotic behaviour of the solution to the problem of scattering by smooth convex obstacles was analysed in [27]. Motivated by these results, a hybrid scheme was considered in [1]. The authors report an overall solution method that requires O(k1/3 ) operations as a function of the wavenumber, a huge improvement over the linear dependence on k. The basis functions are piecewise polynomials, multiplied by plane waves in a number of directions. Similar hybrid methods with even better results are proposed in [2, 7, 25, 24, 16, 12]. A number of operations that is independent of the wavenumber, for a xed error, is achieved by Bruno, Geuzaine, Monro and Reitich in [7] for the scattering by smooth convex obstacles, and by Langdon and ChandlerWilde in [25] for scattering on a half-plane. In the present paper, we combine a similar approach with recent insights in the behaviour of oscillatory integrals [20, 19]. As a result, we obtain a small, and highly sparse discretisation matrix. In addition, the accuracy of the solution actually increases with increasing frequency. We start the paper in 2 with a brief review of a suitable quadrature rule for the evaluation of oscillatory integrals that have the general form
b

I[f ] :=
a

f (x)eikg(x) dx,

(1.1)

where both f and g are smooth functions, and where the wavenumber k determines the frequency of the oscillations. The review is based on results by Iserles and Nrsett in [20] and by the authors in [19]. We show how Filon-type quadrature rules using derivatives can be constructed that have the form
n dl

I[f ] Q[f ] :=
l=0 j=1

wl,j f (j) (xl ),

(1.2)

such that the accuracy of the rule improves with increasing frequency. This is quite contrary to the rapid deterioration of classical quadrature rules, based on polynomial interpolation, for increasingly oscillatory integrals. The asymptotic order of accuracy as a function of the frequency is O(ks1 ), where the value of s depends on the number of derivatives used. This is much like the behaviour of truncated asymptotic expansions. However, due to the nature of the constructed rules, the result is exact for an arbitrarily large family of functions regardless of the frequency. The main dierence with asymptotic expansions therefore is that there is no breakdown at low frequencies. In general, one can also expect a smaller error for moderate frequencies. 2

The results of [19] are generalised in 3 to an appropriately chosen model form

IH [f ] :=
a

(1) f (x)H (kg1 (x))eikg2 (x) dx, (1)

where f , g1 and g2 are assumed to be smooth functions. The function H (z) is the Hankel function of the rst kind and order ; it is an oscillatory function for large arguments. The model form is chosen to represent the integrals that will appear later in the solution of scattering problems. The resulting quadrature rule has the same form as (1.2). The scattering problem is introduced in 4. The quadrature rules are subsequently used in the discretisation of the oscillatory integral equation in 5. We identify a setting in which each row of the discretisation matrix corresponds to the discretisation of a specic one-dimensional oscillatory integral with a known phase. Owing to the small number of quadrature points required for the evaluation of such integrals, the discretisation matrix is sparse. The accuracy of the solution improves because the quadrature rules themselves improve for increasing frequencies. We illustrate the method with numerical results in 6. We consider the scattering of a plane wave and of a circular wave, by a circle and by an ellipse respectively. We end the paper with some concluding remarks in 7.

The ecient evaluation of oscillatory integrals

The classical approach for the evaluation of oscillatory integrals is to use a xed number of quadrature points per wavelength. This automatically leads to a number of operations that scales linearly with the frequency. However, the asymptotic expansion of an oscillatory integral for large frequencies reveals that the value of the integral is actually determined by the behaviour of the integrand near a small set of special points [31, 6]. These are the boundary points of the interval, and the so-called stationary points. Consider the oscillatory integral (1.1), where both f and g are smooth functions. We call f the amplitude, and g the oscillator of the integral. The stationary points of (1.1) are all solutions to the equation g () = 0, [a, b]. A stationary point is said to have order r if g(j) () = 0, j = 1, . . . , r, but g(r+1) () = 0. The importance of such points lies in the fact that, locally, the integrand does not oscillate near a stationary point. Away from all stationary points and the boundary points a and b, the oscillations of the integrand increasingly cancel out. Hence, such regions do not contribute much to the value of the integral. Mathematically, this property is reected in the asymptotic expansion of (1.1). It can be shown that integral (1.1) with one interior stationary point of order r has an asymptotic expansion of the form (see [31])

I[f ]
j=0

cj [f ] , k(j+1)/(r+1)

k .

(2.1)

It is proved in [20] that the rst few coecients cj [f ] of the asymptotic expansion depend only on the rst few derivatives of f and g, evaluated at the boundary points and at the stationary points. A number of recent methods exploit this behaviour, in order to obtain an approximation of I[f ] that improves with increasing k [20, 19, 26, 29]. We recall one particularly useful approach, that leads to a quadrature rule with a classical form. We refer the reader to [21] for a more general overview. 3

2.1

A numerical steepest descent method

A well-known technique for obtaining the asymptotic expansion (2.1) is the method of steepest descent [6, 33]. Justied by Cauchys integral theorem for analytic functions, the integration path is deformed into a complex integration path that has more desirable properties. Specifically, the new path is chosen such that the integrand is not oscillatory along the path, and has exponential decay. The asymptotic expansion can then be derived from the resulting line integrals in the complex plane. Asymptotic expansions have the disadvantage however that the error is not controllable for low values of k. Rather than forming the asymptotic expansion, the line integrals can also be evaluated numerically. This numerical approach leads to a method that improves with increasing k, but that retains accuracy for small values of k [19]. Assume that both f and g are analytic functions. Subdivide the integration interval [a, b] into subintervals [al , bl ], l = 1, . . . , L, such that g (x) = 0, x (al , bl ), i.e., the oscillator g is a monotonic function on [al , bl ]. Then the inverse of g exists uniquely on [al , bl ], and it is also analytic. We denote it by gl1 . The integral over [al , bl ] is evaluated by deforming the integration path into the complex plane. Dene the path hx,l (p) by hx,l (p) = gl1 (g(x) + ip), x [al , bl ]. (2.2)

This path is called the path of steepest descent. If follows immediately from the denition that eikg(hx,l (p)) = ekp eikg(x) . Hence, the line integral along the new integration path, originating in the point x and terminating in the point hx,l (P ), can be written as Sl [f ; x] = eikg(x)
0 P

f (hx,l (p))h (p)ekp dp, x,l

P > 0.

(2.3)

The integrand of (2.3) is not oscillatory, and decays exponentially fast. The integral over [al , bl ] can now be approximated by Sl [f ; al ] Sl [f ; bl ]. In general, the value of I[f ] can be approximated by a sum of non-oscillatory line integrals with exponentially small error,
L

I[f ] =
l=1

(Sl [f ; al ] Sl [f ; bl ]) + O(ekP ),

(2.4)

with P > 0. The larger P , the better the approximation, but for all nite P the error decays exponentially fast as the frequency parameter k increases. In many cases of practical interest, the limit case P is possible, and the error O(ekP ) vanishes. In that case, decomposition (2.4) is exact.

2.2

A quadrature rule for oscillatory integrals

The decomposition (2.4) can be used to obtain a quadrature rule for I[f ] with a classical form. Approximating f by its truncated Taylor series at each point al and bl , we nd a quadrature rule using derivatives of the form
L dl

Q[f ] :=
l=0 j=0

wl,j f (j) (xl ),

(2.5)

with x0 = a, xl = bl , l = 1, . . . , L, and with the weights given by w0,j = S1 wl,j wl,j (x a)j ;a , j! (x xl )j (x xl )j ; xl + Sl+1 ; xl , = Sl j! j! (x b)j = Sl ;b . j!

l = 1, . . . , L 1,

Note that this rule uses only local information of f : its values and derivatives at the endpoints a and b of the integration region, and at all the stationary points in between. Yet, the approximation I[f ] Q[f ] is very accurate, and becomes increasingly accurate with increasing values of the frequency parameter k. It can be shown that the rst few terms of the asymptotic expansion of the error I[f ] Q[f ] vanish. This is due to the fact that these terms depend only on the derivatives of f at exactly the points we consider in the quadrature rule. Hence, the approximation has a high asymptotic accuracy. The exact order depends on the number of derivatives dl that are used and on the order of the stationary points involved [19]. Another important characteristic of the quadrature rules (2.5) is that, by construction, it is exact for all polynomials up to degree p = min dl .
l

(2.6)

Similar rules are constructed with higher polynomial exactness, so-called Filon-type methods, in [20]. This property means that, contrary to asymptotic expansions, these quadrature rules will converge for a large class of functions even at low frequencies. The rules combine the desirable numerical properties of classical quadrature rules and asymptotic expansions.

Specialised quadrature rules

The quadrature rule of the previous section applies only to the model integral (1.1). Intuitively however, one sees that the ideas can be readily generalised to any oscillatory integral. The value of an oscillatory integral is determined by the behaviour of the integrand near the endpoints of the integration interval, and near the points where the integrand locally does not oscillate. In order to construct similar quadrature rules, one requires knowledge of the phase of the integral. In this section, we will construct such rules for a family of integrals that will arise in the scattering problem discussed later. In particular, the integrand involves an oscillatory Hankel function.

3.1

A generalised model form

Consider the oscillatory integral

IH [f ] =
a

(1) f (x)H (kg1 (x))eikg2 (x) dx, (1)

(3.1)

where f , g1 and g2 are smooth functions, and H (z) is the Hankel function of the rst kind (1) of order . The Hankel function of order zero H0 (z) has a logarithmic singularity at z = 0. Hankel functions of higher order have algebraic singularities of the form 1/z , z 0 [3]. 5

For large arguments, the Hankel functions behave like an oscillatory complex exponential with a decaying amplitude,
(1) H (z)

2 i(z 1 1/4) 2 e , z

< arg z < 2,

|z| .

(3.2)

Hence, the oscillator of the integrand of (3.1) is approximately given by g(x) = g1 (x) + g2 (x), (3.3)

up to the addition of a constant. The Hankel function decays exponentially fast for complex arguments with a positive imaginary part, as can be seen from the asymptotic behaviour (3.2). This means that the approach of 2.2 using the path of steepest descent is applicable. Hence, we may conjecture that a quadrature rule exists of the form
L dl H wl,j f (j) (xl ). l=0 j=0

IH [f ] QH [f ] :=

(3.4)

In the remainder of the section, we will prove this conjecture, determine the quadrature abscissae xl and show how the weights can be computed eciently.

3.2

Construction of the quadrature rule

We start by stating some assumptions on the functions f and g, that are needed to guarantee the integrability and analyticity of the integrand in (3.1). First, we assume that f is analytic in an open complex neighbourhood D of [a, b], so that [a, b] int D. Likewise, we assume that g1 and g2 are non-singular and analytic in D, except possibly along a branch cut that extends from a or b to the boundary of the region D, i.e., a and b may be branch points but not singular points. We assume furthermore that g(x), dened by (3.3), is strictly monotonic on the open interval (a, b) and hence invertible, but possibly g (a) = 0 or g (b) = 0. Also, we assume that g1 (x) = 0, x (a, b). Finally, if > 0 and g1 () = 0 we assume that f behaves like f (x) (x )1+ , x , with > 0. (3.5)

Condition (3.5) guarantees that the integrand of IH [f ] is integrable. Subject only to condition (3.5) and the analyticity requirements, the integration interval of (3.1) can always be split into a number of subintervals that satisfy the conditions. The assumptions guarantee that the integrand of IH [f ] is analytic on [a, b] except possibly in the points a and b. In particular, this will allow us to apply Cauchys integral theorem to select the integration path of (3.1). Theorem 3.1. Under the assumptions stated above, the integral IH [f ] can be approximated by a sum of contributions IH [f ] = S H [f ; a] S H [f ; b] + O(ekP ), with P > 0, and with the contributions given by the integrals S H [f ; x] =
0 P (1) f (hx (p))H (kg1 (hx (p)))eikg2 (hx (p)) h (p) dp, x

(3.6)

(3.7)

where hx (p) satises g(hx (p)) = g(x) + ip. (3.8)

The proof is almost identical to the proof of Lemma 4.1 in [18] and is omitted; it diers mainly in the special treatment of the Hankel function based on the asymptotic expression (3.2). We note from the asymptotic behaviour (3.2) that the integrand of the line integral S H [f ; x] in (3.7) is non-oscillatory and exponentially decaying in the integration variable p,
(1) H (kg1 (hx (p)))eikg2 (hx (p))
1 2 eikg(x) ei( 2 1/4) ekp , kg1 (hx (p))

k .

The size of the constant P is related to the size of the region of analyticity of f and g [18]. In the numerical examples of the scattering problem, given in 6, we can in fact consider the limit case P = . The error of decomposition (3.6) then vanishes even at low frequencies. We proceed in a similar way as in 2.2. Since f is analytic in D, it has an absolutely convergent Taylor series. By the linearity of S H , we may write

S [f ; x0 ] =
j=0

f (j)(x0 )S H

(x x0 )j ; x0 . j!

Now, consider a subdivision of [a, b] into subintervals [al , bl ], l = 1, . . . , L, such that on each subinterval the conditions of Theorem 3.1 are satised. Truncating the Taylor series of f at each special point al and bl after a nite number of terms, we arrive at a quadrature rule QH [f ] of the form (3.4), with weights given by
H H w0,j = S1 H wl,j H wl,j

(x a)j ;a , j! (x xl )j (x xl )j H ; xl + Sl+1 ; xl , = SlH j! j! (x b)j = SlH ;b . j!

(3.9) l = 1, . . . , L 1, (3.10) (3.11)

The weights can be explicitly computed very eciently, by using Gauss-Laguerre quadrature or similar techniques [11]. The accuracy of these methods improves rapidly as a function of k, due to the faster decay of the integrands as k increases. For the purposes of our application, this advantageous characteristic is not needed. It suces already that the number of operations for a xed accuracy is bounded with respect to k. We therefore choose to focus on the convergence properties of the quadrature rule itself, rather than on the convergence of methods to compute the weights.

3.3

Convergence properties of the quadrature rule

We discuss the properties of the quadrature rule QH [f ], with weights given by (3.9)-(3.11), as a function of k. The rule is exact by construction for polynomials of degree less than or equal to p = min dl .
l

(3.12) 7

For more general functions, the accuracy as a function of k is determined by the asymptotic size of the weights. We will show that the size of the weights decreases both with increasing frequency and with increasing order of the corresponding derivative. The order of accuracy of the quadrature rule is therefore equal to the asymptotic size of the rst weight that is discarded by truncation. In order to quantify this size, we require a few technical lemmas. Lemma 3.2. Assume x0 is a stationary point that has order r. The parameterisation of the path (3.8) behaves as hx0 (p) = x0 + O(p1/(r+1) ), h 0 (p) x = O(p
1/(r+1)1

p 0, p 0.

(3.13) (3.14)

),

Proof. Since g(j) (x0 ) = 0, j = 1, . . . , r, we can write the Taylor series of g as g(x) = g(x0 ) + g(r+1) (x0 ) (x x0 )r+1 + O((x x0 )r+2 ). (r + 1)!

The path hx0 (p) = g1 (g(x) + ip) solves g(hx0 (p)) = g(x) + ip, and hence hx0 (p) x0 +
r+1

ip(r + 1)! , g(r+1) (x0 )

p 0.

(3.15)

The second result follows by dierentiation. Note that the complex root is multi-valued: the 1 correct root is selected by using the analytic continuation of the inverse gi that satises 1 gi (g(x)) = x on [ai , bi ] in expression (3.8). The size of the weights follows from the size of the line integrals S H (xx0 ) ; x0 . Recall that j! the integral may be singular if g1 (x) = 0. We will assume for the sake of brevity that, in that case, g1 (x) = 0. This condition is always satised by the applications in 6. Lemma 3.3. Let S H [f ; x] be dened by (3.7) with P = , and g dened by (3.3). Assume that g (x0 ) = 0, i.e., x0 is not a stationary point. If g1 (x0 ) = 0, we have S H [(x x0 )j ; x0 ] = O(kj3/2 ), k .
j

If g1 (x0 ) = 0 and g1 (x0 ) = 0, the integral is singular and we have

S H [(x x0 )j ; x0 ] = O(kj1 ),

j ,

k .

Proof. We write the integral S H [(x x0 )j ; x0 ] as S H [(x x0 )j ; x0 ] =

u(p)ekp dp =

1 k

u(q/k)eq dq,

(3.16)

with
(1) u(p) = (hx0 (p) x0 )j h 0 (p)H (kg1 (hx0 (p)))eikg2 (hx0 (p)) ekp . x

(3.17)

It is a consequence of Watsons Lemma that the asymptotic expansion of the integral can 1 be obtained by integrating the asymptotic expansion of k u(q/k) as k term by term 8

in (3.16) [6, 32]. Generalising Watsons Lemma, this remains true for integrals of the form n s z 0 u(p)h(kp)dp where h(z) log(z) z e , n 0, s Z, z 0, if the integrand is integrable [5]. This means that the singularity of the Hankel function has no inuence on the asymptotic expansion. First, consider the case g1 (x0 ) = 0. Then, combining the asymptotic behaviour of the Hankel function for large arguments (3.2) with the results (3.13)-(3.14) of Lemma 3.2 for r = 0, we have u(q/k) kj1/2 . From (3.16) we can conclude S H [(x x0 )j ; x0 ] = O(kj3/2 ).
Next, consider the case g1 (x0 ) = 0. If g1 (x0 ) = 0, then we have g1 (hx0 (p)) p1/(r+1) = p. It (1) follows that H (kg1 (hx0 (q/k))) = O(1), k . Hence, by the generalisation of Watsons Lemma, we may conclude S H [(x x0 )j ; x0 ] = O(kj1 ), j .

The corresponding lemma for stationary points is very similar. The dierence is due to the dierent behaviour of the parameterisation as described by Lemma 3.2. Lemma 3.4. Let S H [f ; x] be dened by (3.7) with P = , and g dened by (3.3). Assume that x0 is a stationary point of order r. If g1 (x0 ) = 0, then we have S H [(x x0 )j ; x0 ] = O(k(j+1)/(r+1)1/2 ),
If g1 (x0 ) = 0 and g1 (x0 ) = 0 then we have

k .

S H [(x x0 )j ; x0 ] = O(k(j+1+r/2)/(r+1) ),

j ,

k .

Proof. Consider again the function u(p), given by (3.17). Assume rst that g1 (x0 ) = 0. We have (hx0 (q/k) x0 )j kj/(r+1) and h 0 (q/k) kr/(r+1) . Since kg1 (hx0 (q/k)) x kg1 (hx0 (0)) = kg1 (x0 ), we also have H (kg1 (hx0 (q/k))) k1/2 . Combined in (3.16), and by applying Watsons Lemma, this yields the rst result.
The case where g1 (x0 ) = 0 is slightly dierent. Since g1 (x0 ) = 0, we have g1 (hx0 (q/k)) 1/(r+1) and, hence, kg (h (q/k)) k r/(r+1) . The Hankel function therefore yields the k 1 x0 factor k(r/2)/(r+1) instead of k1/2 as in the rst case. (1)

The convergence of the quadrature rule (3.4) as a function of k can now be established. Note that the results of Lemma 3.3 agree with those of Lemma 3.4 if we take the order of a regular point to be r = 0. Hence, we need not distinguish between stationary points and regular (end)points. Lemma 3.5. The error of the approximation of SlH [f ; x0 ], x0 [al , bl ], is
dl

SlH [f ; x0 ] QS [f ; x0 ] = SlH [f ; x0 ] l
j=0

H wl,j f (j) (x0 ) = O(kl ),

k . (3.18)

If g1 (x0 ) = 0 then l := (dl + 2)/(r + 1) 1/2. If g1 (x0 ) = 0 and g1 (x0 ) = 0, then l := (dl + 2 + r/2)/(r + 1).

Proof. Since the weights decay as a function of k, and as a function of the order of derivative j, the error of the quadrature scheme is asymptotically determined by the size of the rst discarded weight. The result follows from Lemmas 3.3 and 3.4 by setting j = dl + 1. 9

The theorem that characterises the accuracy of the complete quadrature rule follows immediately. Theorem 3.6. Consider the approximation of IH [f ] by QH [f ]. The error has asymptotic order = minl l , where l is specied in Lemma 3.5. As an example, we consider the integral 0 cos(x 1)H0 (kx)eik(x +x x) dx. The total oscillator for this integral is g(x) = x2 + x3 . There are two quadrature points: there is a singularity and a stationary point of order 1 at x = 0, and a regular endpoint at x = 1. The H H weights w0,j and w1,j are given by (3.9) and (3.11) respectively. From Lemma 3.4 we have H H |w0,j | = O(k(j+1)/21/4 ) and from Lemma 3.3 we have |w1,j | = O(kj3/2 ). Using d0 and (d0 +2)/21/4 ), O(k (d1 +1)3/2 )} by Theorem 3.6. d1 derivatives, the error has order min{O(k We choose d1 = max{0, (2d0 5)/4} to match the errors. Table 1 shows the convergence of the quadrature rule QH [f ] as a function of k and d0 . Table 1: Absolute error of the approximation of IH [f ] by QH [f ], with f (x) = cos(x 1), g1 (x) = x and g2 (x) = x2 + x3 x. The last row shows the value of log2 (e400 /e800 ): this value should approximate d0 /2 + 5/4 (shown in parentheses). k \ d0 100 200 400 800 rate 0 1.2E 3 5.1E 4 2.2E 4 9.3E 5 1.23 (1.25) 1 2.8E 5 8.6E 6 2.6E 6 7.8.1E 7 1.73 (1.75) 2 1.3E 6 2.9E 7 6.4E 8 1.4E 8 2.20 (2.25) 3 2.6E 8 4.1E 9 6.2E 10 9.7E 11 2.68 (2.75)
1 (1)
2 3

4
4.1

High frequency scattering problems

We are interested in the solution of the two-dimensional Helmholtz equation u + k2 u = 0 (4.1)

in the domain + = R2 \ that is exterior to the open domain , subject to a Dirichlet boundary condition u = 0 on the boundary = . We assume that a periodic parameterisation for is given by ( ) : [0, 1] R2 , (4.2)

with |( )| > 0. The Helmholtz equation models the scattering of time-harmonic waves of the form u(x)eit . In the absence of damping we have k = /c, with c the speed of propagation. For the scattering of an incoming acoustic or electromagnetic wave ui (x) by the obstacle , the solution can be written as u = ui + us . The function us (x) represents the scattered wave; it satises the Helmholtz equation with the Dirichlet boundary condition us (x) = ui (x) on . 10

The scattered wave can be represented in terms of the single-layer potential, (Sq)(x) =

K(x, y)q(y) dsy ,

with

i (1) K(x, y) = H0 (k|x y|), 4

(4.3)

where q(x) is the density function dened on the boundary , and where K(x, y) is the Greens function of the two-dimensional Helmholtz equation. The scattered wave due to an incoming wave ui is given by us = Sq, if the density function q(x) is found as the solution to the combined potential integral equation q(x) + 2 K ui (x, y) + iK(x, y) q(y) dsy = (x) + iui (x), nx n x , (4.4)

where nx is used to denote the normal with respect to the variable x. Equation (4.4) is uniquely solvable for all values of the wavenumber k [8].
u An important observation is that q(x) = n (x), i.e., the density function is exactly the (exterior) normal derivative of the solution to the Helmholtz problem. This means that the solution to equation (4.4) is directly related to a physical property of the problem. For example, in electromagnetics, the normal derivative of the electric eld is proportional to the induced current on the surface of the conducting obstacle [28].

4.2

High frequency integral equation formulation

The density function q(x) is highly oscillatory for large values of the wavenumber k. The solution of equation (4.4) therefore generally requires a large number of unknowns. In some cases however, one has a priori information about the phase of the solution. For example, if the obstacle is convex, and if the incoming wave is a plane wave, then the phase of the solution q is approximately the same as the phase of the incoming wave. Assume the incoming wave i is given by ui (x) = ui (x)eikg (x) . Then we can write s q((t)) = qs (t)eikg
i ((t))

t [0, 1],

(4.5)

where qs (t) is a function that is less oscillatory than q((t)), in a sense that will be made more precise in 4.3. In physical terms, the oscillations of the induced current on a perfectly conducting surface tend to follow the oscillations of the incoming electromagnetic wave. This is the main reason why the problem is formulated such that the solution q(x) corresponds to a physical variable - only in that case is the phase known in the form of (4.5). This was noted in [7]; the integral equation formulation of this section follows the same pattern as [7]. Substituting the ansatz (4.5) into (4.4) leads to the integral equation of the second kind 1 i qs (t)eikg ((t)) + 2 with G(t, ) := K ((t), ( )) + iK(x, ( )) |( )|, nx
1 0

G(t, )qs (( ))eikg

i (( ))

d = v i ((t)),

(4.6)

and with the right hand side given by v i (x) := ui (x) + iui (x). n 11

shadow boundary reflection

diffraction

ui
illuminated region shadow region

1 0

boundary region

Figure 1: Reection and diraction eects in the scattering of an incoming wave ui by a smooth and convex obstacle. The unknown in (4.6) is the function qs (t). Since qs (t) is less oscillatory than the original density function q((t)), one can solve (4.6) using a coarser discretisation. In the next section, we discuss the remaining oscillatory behaviour of qs (t).

4.3

Asymptotic behaviour of the solution

In the past decades, a lot of eort has been invested in studying the asymptotic behaviour of the solution q(x) to (4.4) as a function of the wavenumber, concentrating mainly on the scattering of a plane wave (see, e.g., [23, 27] and references therein). For smooth and convex obstacles, there are three important regions with dierent properties, illustrated in Figure 1: the illuminated region, the shadow region, and the transitional shadow boundary region. In the illuminated region, the scattered wave is described asymptotically by geometrical optics: a wave is reected such that the angle of incidence and the angle of reection are identical. A wave tangential to the shadow boundary is diracted. The density function decays rapidly away from the shadow boundary into the shadow region, due to the continuous emission of diracted waves. In the shadow region the function q(x) approaches zero as k increases. The asymptotic expansion of qs (t) for large k reects these three regions. Assume an incoming plane wave in the direction of the form ui (x) = eikx , with || = 1. It was proved in [27] that qs (t) has a uniform asymptotic expansion of the form qs (t)
m,n0

k2/3n2m/3 bm,n (, t)(n) (k1/3 Z(, t)),

(4.7)

that remains valid as t crosses the shadow boundaries. The shadow boundary points can be characterised here by = 0, with the exterior normal to . For a thorough discussion and analysis of this expansion, we refer the reader to [12]. Here, we only recall the main characteristics of the expansion, as a means to motivate the choices that are made later on. The function Z C is innitely smooth, and has a simple root at the two shadow boundary points. The function Z is positive when < 0, i.e., in the illuminated region, and it is negative in the shadow region. The function (z) is smooth for positive arguments, with (z) z, z , (4.8)

and it is rapidly but oscillatory decaying for large negative arguments. It follows that the function qs (t) is smooth in the illuminated region, but may be oscillatory in the shadow 12

region, corresponding to the behaviour of (z) for large positive arguments and large negative arguments respectively. One can also deduce the following asymptotic properties from the leading order of (4.7), |qs (t)| = O(k), illuminated region, 2/3 ), shadow boundary. O(k (4.9)

There is a k-dependent transition region near the shadow-boundary. Motivated by (4.7), we introduce a transition region of size O(k1/3 ) around the shadow boundary points, and dene the shadow boundary regions as TB1 (k) = [tsb1 D1 k1/3 , tsb1 + C1 k1/3 ], TB2 (k) = [tsb2 C2 k1/3 , tsb2 + D2 k1/3 ], (4.10) (4.11)

with constants C1 , C2 , D1 , D2 > 0 independent of k, but small enough such that TB1 (k) and TB2 (k) are non overlapping, and with tsb1 and tsb2 the locations of the two shadow boundary points in the parameter domain [0, 1]. The illuminated region is dened as TI (k) = (tsb1 + C1 k1/3 , tsb2 C2 k1/3 ). The shadow region is the remaining part of the interval [0, 1]. The size of the transition region is related to the behaviour of the argument k1/3 Z(, t) of the function (z) in (4.7). Because Z(, t) has a simple zero at tsb1 , we have Z(, t) Z (, tsb1 )(t tsb1 ), t tsb1 . (4.12)

Hence, there exists a constant c such that for suciently large k we have |k1/3 Z(, t)| c, if t TB1 (k). (4.13)

A similar bound holds for the region TB2 (k). Since the oscillations of qs (t) originate in the oscillatory behaviour of (z) for negative arguments, the meaning of the bound (4.13) is that the number of oscillations of qs (t) in the shadow boundary region TB1 (k) is bounded in k. Remark 4.1. The remaining oscillatory behaviour of qc (t) in the shadow boundary region indicates that the oscillations of the incoming wave do not completely describe the oscillatory behaviour of q(x). Still, we will see that the ansatz (4.5) is suitable for further computations.

A hybrid high frequency boundary element method

The collocation of integral equation (4.6) in a point xn leads to a one-dimensional and oscillatory integral in the integration variable . In this section, we show how an ecient quadrature rule can be used for the discretisation of that collocation integral. First, we discuss the classical boundary element approach in 5.1. The application of specialised quadrature rules, leading to a highly sparse discretisation matrix, is described in 5.2. We comment on the location of the quadrature points 5.3 and on the limitations of the approach in 5.4. It turns out that the quadrature rule can not be applied everywhere. Finally, we arrive at a method combining a sparse discretisation with a dense part in 5.5. 13

5.1

Collocation approach for the discretisation

Consider a collocation scheme for integral equation (4.6), with a set of N distinct collocation points xn = (tn ), tn [0, 1], n = 1, . . . , N . The classical way to proceed is to look for an approximation qc to solution qs in the form
N

qc (t) =
m=1

cm m (t),

(5.1)

where the m functions are a set of linearly independent basis functions with support m := supp(m ). The number of basis functions may be small, since the exact solution qs is not very oscillatory. Collocating equation (4.6) in the points tn , with qs replaced by qc , leads to the equations
i 1 qc (tn )eikg (xn ) + 2

1 0

G(tn , )qc ( )eikg

i (( ))

d = v i (xn ),

n = 1, . . . , N.

(5.2)

The collocation approach therefore leads to a linear system Ac = b of size N N , where the elements of the discretisation matrix A are given by An,m =
m

G(tn , )eikg

i (( )

m ( ) d,

(5.3)

and the right hand side by bn = v i (xn ). The discretisation matrix A is dense, but small compared to the classical boundary element discretisation for the original equation. Hence, this is a big improvement over the direct discretisation of (4.4). Since the elements (5.3) are given by oscillatory integrals, they can be computed eciently using the numerical steepest descent technique described in 2.1. This would yield an ecient total solution method, that remains ecient when k increases, but that is also accurate for small values of k. However, there are still some issues associated with this approach. Since the matrix is dense, the method requires the evaluation of N 2 integrals. Although N may be rather small, the computational cost can still be high. Interestingly, it was observed in [15] that many of the elements are small, and can in fact be discarded, reducing the computation time. A second, more important issue is that the results of [15] indicate that the error of the scheme increases with increasing wavenumber. Here, we examine a dierent discretisation of (5.2) that aims to address these issues, based on the quadrature rule developed in 3 and motivated by the accuracy of this rule for high wavenumbers. Owing to the small number of required quadrature points, the resulting discretisation matrix will be highly sparse.

5.2

The application of specialised quadrature rules

Based on the collocation integral (5.2), we dene the following general oscillatory integral,
1

Ic [f ; tn ] :=
0

f ( )G(tn , )eikg

i (( )

d.

(5.4)

The oscillatory function G(t, ) in the integrand is given explicitly by i ( ) ((t) ( )) (1) |( )|H1 (k|(t) ( )|) G(t, ) = k 4 |(t) ( )| (1) |( )|H0 (k|(t) ( )|). 4 14

Choosing = k, this function can further be written as G(t, ) = a( )kH0 (k|(t) ( )|) + b(t, )kH1 (k|(t) ( )|),
(1) (1)

(5.5)

with functions a( ) and b(t, ) that are independent of k. Hence, integral (5.4) can be written as a sum of two model integrals of the form IH [f ], involving Hankel functions, that were introduced in 3. Both model integrals have the same oscillator, g( ; tn ) := |(tn ) ( )| + gi (( )). In particular, this means that one can nd a quadrature rule such that
Ln dn,l

(5.6)

Ic [f ; tn ] Qc [f ; tn ] :=
l=0 j=0

c wn,l,j f (j)(n,l ).

(5.7)

The rule Qc [f ; tn ] is computed as follows:

S c [f ; x] :=
0

f (hx (p))G(tn , hx (p))eikg

i ((h

x (p))

h (p) dp, x

(5.8)

where the integration path is the path of steepest descent corresponding to the oscillator (5.6). Compared to S H [f ; x], we have included the additional factors ka( ) and kb(t, ) from (5.5) into the weight function of the integral. We note that it implies the assumption that ( ) is analytic. The scaling by a factor k reduces the asymptotic order, based on the absolute error, by 1. the quadrature points n,l are found from the oscillator (5.6). They are the points where the integrand becomes singular (and hence non-analytic), and the stationary points of the oscillator g( ; tn ). These points are derived by a straightforward, but technical analysis of g( ; tn ). There are no contributing endpoints, as the integrand is periodic on the closed curve . We now describe how the quadrature rule (5.7) can be used in the discretisation. The derivatives of qc can be written in terms of the basis functions m ,
N (j) qc ( ) = m=1

cm (j) ( ). m

(5.9)

Hence, applying the quadrature rule to qc yields a matrix B with entries Bn,m =
l:n,l m dn,l c (j) j=0 wn,l,j m (n,l ),

l [0, Ln ] : n,l m , otherwise

(5.10)

The entry Bn,m is nonzero only if at least one quadrature point n,l exists that lies in the support of the basis function m . The number of nonzero points therefore depends on the size of the supports of the basis functions. If all basis functions are local, then the matrix B is highly sparse. Remark 5.1. The asymptotic order of quadrature rule (5.7) is limited by the number of derivatives dn,l that are used, as described by Theorem 3.6. In turn, note from (5.9) that dn,l is limited by the number of derivatives that exist for the chosen basis functions. 15

0.2

0.4

t
0.6 0.8 1 0

0.2

0.4

0.6

0.8

Figure 2: The location of the contributing points of the collocation integral for the scattering of a plane wave by a circular obstacle. Each row corresponds to a xed value t [0, 1], with the shadow boundary points at 0.25 and 0.75, and the illuminated region in between. The singular points are located along the diagonal. The o-diagonal points correspond to stationary points.

5.3

Location of the quadrature points

The location of the quadrature points is illustrated in Figure 2 for the scattering of a plane wave by a circular obstacle. There is one stationary point if tn lies in the illuminated region, and there are three stationary points if tn lies in the shadow region. Two of these points coalesce into one stationary point of order r = 2 exactly at the shadow boundary. Using local basis functions, the structure of the discretisation matrix B closely resembles that of Figure 2. The coalescence of two stationary points at the shadow boundary is a general property for convex obstacles [22]. The quadrature points on the diagonal reect the general principle of localisation of high frequency scattering, which states that the scattering of a ray is asymptotically determined by properties local to the point of incidence [13]. As such, Figure 2 is illustrative for more general convex shapes. Remark 5.2. Figure 2 shows the existence of a stationary point in the shadow region when t TI (k) is a point in the illuminated region. This seems to contradict the localisation principle. Indeed, this stationary point has no physical relevance for the scattering problem. Following Remark 4.1, it is an artefact of the incorrect ansatz in the shadow region.

5.4

Limitations of the sparse discretisation

The convergence of quadrature rule Qc [f ; tn ] as a function of k is described by the corresponding results for QH [f ] discussed in 3.3. In particular, the results imply that the accuracy of the approximation increases with increasing k, at a rate that depends on the number of derivatives that are used in the quadrature rule. For example, for t TI (k), we have |Ic [f ; tn ] Qc [f ; tn ]| = O k min{dn,0 +1,(dn,1 +2)/23/2} , (5.11)

where n,0 = tn and n,1 is the stationary point of order r = 1 in the shadow region. Hence, the sparse discretisation matrix B is very well suited for computing the application of the 16

1 0.5

0 0.5

1 0

0.2

0.4

0.6

0.8

Figure 3: The real part of the integrand for tn = tsb1 for a circular obstacle and k = 100. integral operator under consideration to a function of the form f (t)eikg
i ((t))

The convergence results in 3.3 were derived however with the assumption that the function f is independent of k. No statements can therefore be made regarding the computed solution of an integral equation using the sparse discretisation matrix B, because the solution generally i depends on k. Moreover, it was shown that the assumed oscillatory behaviour eikg ((t)) in the ansatz (4.5) does not fully describe the oscillatory behaviour of the exact solution q(x). Therefore, Qc [qs ; tn ] might not be a good approximation to Ic [qs ; tn ]. It is important that the quadrature rule is accurate when applied to the exact solution qs . Otherwise, one cannot expect that solving the system with the sparse representation matrix B can yield a good approximation of the exact solution. It can be veried that Qc [qs ; tn ] is accurate when tn TI (k), i.e., when tn lies in the illuminated region. The cases t Tsb1 (k) and t Tsb2 (k) are more problematic. Figure 3 shows the real part of the collocation integral Ic [1; tsb1 ], where tn = tsb1 coincides with the rst shadow boundary. This integral has a stationary point of order r = 2 at tsb1 , but only to the right of the singularity. The integral is highly oscillatory to the left of the shadow boundary, and not oscillatory to the right. It turns out that a quadrature rule similar to Qc [f ; tn ] can be used on the left interval [0, tsb1 ] and in the illuminated region TI (k). It is not suited for the intermediate interval [tsb1 , tsb1 + C1 k1/3 ]. Due to the stationary point however, the integrand is not oscillatory in that interval. In the following section, we will construct a sparse approximation matrix that is based on the use of specialised quadrature rules in the regions where they apply. They are combined with a classical dense discretisation in the remaining part of the integration domain. Remark 5.3. The failure of Qc [qs ; tsb1 ] for the exact solution qs can be seen as follows: the asymptotic decay of the quadrature weights wn,l,j for increasing k is cancelled exactly by the (j) growth of the derivatives qs (tsb1 ) for increasing k. Hence, the rule does not converge.

5.5

A sparse discretisation for the scattering problem

Basis functions and collocation points. The basis functions for the discretisation are chosen corresponding to the behaviour of the solution in the three dierent regions identied in 4.3. Recall that the solution is smooth in the illuminated region TI (k), and oscillatory but rapidly decaying in the shadow region. First, following [7], we approximate the solution by zero in the shadow region. We choose a xed number of basis functions in the illuminated region TI (k). Finally, we also choose a xed number of basis functions in the transitional shadow boundary regions TB1 (k) and TB2 (k), independently of k. This corresponds to using 17

0 50 100 150 200 0

50

100

150

200

Figure 4: Illustration of the sparse discretisation matrix using cubic B-splines for the scattering of a plane wave by a circular obstacle. The middle part of the matrix is tridiagonal. a xed number of basis functions per oscillation of the solution in these regions. The collocation points V are chosen equidistantly in the regions TB1 (k), TB2 (k) and TI (k). We have V = V1 V2 V3 , where V1 := {tsb1 D1 k1/3 + jh1 }N1 , j=0 V3 := {tsb2 C2 k1/3 + jh3 }N3 , j=0 with h1 := (C1 + D1 )k1/3 /N1 , with h3 := (C2 + D2 )k1/3 /N3 .

V2 := {tsb1 + C1 k1/3 + jh2 }N2 2 , with h2 := (tsb2 tsb1 (C1 + C2 )k1/3 )/(N2 1), j=1

These are N = N1 + N2 + N3 points. As basis functions, we choose B-splines of odd degree s with nodes U given by U := V {tsb1 D1 k1/3 jh1 }j=1
(s+1)/2

{tsb2 + D1 k1/3 + jh3 }j=1

(s+1)/2

(5.12)

These N + s + 1 points uniquely determine N distinct B-splines. The basis functions are s 1 times continuously dierentiable. The nodes in the shadow region are added in order to ensure a smooth transition into the shadow region where the solution is approximated by zero. Discretisation. For the discretisation, we propose the following scheme.

If tn TI (k), a collocation point in the illuminated region, then the quadrature rule Qc [f ; tn ] is applied for the singular point tn , the stationary points lie in the shadow region, and they are discarded. For splines of degree s, the singular point tn lies in the support of only s separate basis functions. The corresponding matrix entries are given by (5.10). The contributions of the stationary points are discarded because the solution is approximated by zero in the shadow region. Recall that in the context of the scattering problem, following Remark 5.2, these are only spurious stationary points. If tn TB1 (k), corresponding to the rst shadow boundary region, then 18

 a specialised quadrature rule for oscillatory integrals is applied on the interval [0, tn ], a classical dense discretisation is used on the interval [tn , tsb1 + C1 k1/3 ], a specialised quadrature rule for oscillatory integrals is again applied on the interval [tsb1 + C1 k1/3 , 1]. The quadrature rule on [0, tn ] reduces to the contribution of the singular point tn . The corresponding weights have the form of (3.11) with SlH replaced by Slc ,
c wn,ln ,j = Slcn

(x tn )j ; tn , j!

where ln is the index such that tn = n,ln . The quadrature rule on [tsb1 + C1 k1/3 , 1] consists of the contributions of the stationary points n,l outside the shadow boundary region, and of the endpoint tr := tsb1 + C1 k1/3 . The weights corresponding to that endpoint have the form of (3.9),
c wn,lr ,j = Slcr +1

(x tr )j ; tr , j!

where lr is the index such that tr [n,lr , n,lr +1 ]. Finally, the dense discretization in the interval [tn , tr ] leads to elements of the form n,m =
m [tn ,tr ]

G(tn , )eikg

i (( ))

m ( ) d,

(5.13)

The only dierence compared to (5.3) is that the integration domain may be cut at the boundaries of [tn , tr ]. Summarizing, the elements of the discretisation matrix for tn TB1 (k) can be written as if m [tn , tr ] = n,m (j) + dn,ln wc if tn m j=0 n,ln ,j m (tn ) dn,lr c (j) + if tr m j=0 wn,lr ,j m (tr ) Cn,m = c + l:tr < m dn,l wn,l,j (j) (n,l ), l [0, Ln ] : tr < n,l m , m j=0 n,l dn,l c (j) l [0, Ln ] : tr < n,l m , j=0 wn,l,j m (n,l ), l:tr <n,l m 0 otherwise.

The case tn TB2 (k) can be treated similarly. The structure of the sparse matrix C is illustrated in Figure 4. The two small dense parts correspond to the dense discretisation in the intervals [tn , tsb1 + C1 k1/3 ] and [tsb2 C2 k1/3 , tsb2 ]. For simplicity, we have chosen the constant C1 large enough such that, for tn TB1 , all stationary points n,l TB1 (k) also lie in the shadow boundary region. The constant C2 was chosen similarly. One can verify that the required integrals of the form (5.13) are not oscillatory. Due to 3 the stationary point of order r = 2, the integrand behaves as eikc( tsb1 ) near the shadow boundary. The argument of the exponential is bounded in k, since by construction we have | tsb1 | max{C1 , D1 }k1/3 . 19


R=0.5 (0,0) (2,0)
(a) Circle

(2,1) 0.5 0.3 (0,0)

(b) Ellipse

Figure 5: Illustration of two smooth convex scattering obstacles. The boundary conditions are plane waves, or circular waves originating from a point source. Hence, there is only a bounded number of oscillations in the integrals for increasing k. It is therefore reasonable to expect that the integrals can be evaluated with a number of operations that is independent of k. Since the weights of the quadrature rule can be evaluated eciently as well, and because the number of unknowns is xed, the matrix in Figure 4 can be computed with a total number of operations that is independent of k.

Numerical results

We consider the scattering by two convex obstacles, a circle and an ellipse, shown in Figure 5. We use two types of boundary conditions: a plane wave, modelled in the form ui (x) = eikx , (1) and a point source, modelled by ui (x) = H0 (|x x0 |), with x0 a point in the exterior + of the obstacle. The circle and ellipse are parameterised by (t) = respectively. In all examples, we have chosen the constants C1 and C2 large enough (w.r.t. D1 and D2 ) such that the sparse matrix has the form shown in Figure 4. The highest order of the derivative used in the specialised quadrature rule for oscillatory integrals is kept xed, dn,l = d, n, l, with d s 1 where s is the degree of the piecewise polynomial spline basis. The integrals (5.13) for the elements of the dense parts that are singular were evaluated using Cubpack [9]. R cos(2t) R sin(2t) and (t) = R1 cos(2t) R2 sin(2t)

6.1

Total solution time

The smooth function qc (t) is illustrated in Figure 6 as it was computed for the dierent scattering problems. The mild oscillatory behaviour of the function near the shadow boundary is illustrated in the left panel of Figure 7, showing only the real part of the solution. Two spikes are present near the shadow boundary, with a peak value that scales as O(k2/3 ) as predicted by the estimate (4.9). The dashed line shows the eect of doubling k. The O(k) behaviour in the illuminated region is clear from the imaginary part illustrated in the right panel of Figure 7.

20

500 400

15

10
300 200 100

0
0 100 0 0.2 0.4 0.6 0.8 1

5 0

0.2

0.4

0.6

0.8

(a) Circle, plane wave

(b) Circle, point source

100 0 100 0 0.2 0.4 0.6 0.8 1

2 0 2 0 0.2 0.4 0.6 0.8 1

(c) Ellipse, plane wave

(d) Ellipse, point source

Figure 6: The real part (dashed line) and imaginary part (full line) of the computed smooth function qc (t). The parameters correspond to the case k = 200 in Table 2.
10 0 10 400 20 200 30 40 0 0 0.2 0.4 0.6 0.8 1 200 0 0.2 0.4 0.6 0.8 1 1000 800 600

Figure 7: The real part (left) and imaginary part (right) of qc (t) for the scattering of a plane wave by a circle. The dense line corresponds to k = 200, the dashed line to k = 400. Table 2 shows the timings for an implementation of the algorithm of 5.5 in Matlab. In all examples considered, the time actually decreases with increasing wavenumber k. This is due to the fact that, at larger frequencies, the weights of the specialised quadrature rule Qc [f ; tn ] are easier to compute. Also, the integrals (5.13) that determine the elements of the dense parts do not become more oscillatory. In a classical boundary element method, and using 10 unknowns per wavelength, the case k = 105 would correspond to a dense matrix with approximately N = 0.5 106 unknowns.

6.2

Error and convergence

In this section, we present numerical results evaluating the error of the proposed scheme. First, we investigate the absolute and relative error for increasing values of k. Next, we show that the accuracy is improved by increasing the number of derivatives d that are used in the specialised quadrature rules. We then illustrate the inuence of the constants C1 , C2 , D1 and D2 that determine the size of the shadow boundary region. Finally, we illustrate the inuence 21

Table 2: Total solution time in seconds for the dierent scattering problems. All parameters are kept xed, except the wavenumber k. We used d = 2 derivatives in the quadrature rules. Circle Plane wave Point source (1.3, 1.3) (1, 0.1.0) (1.0, 1.0) (0.8, 0.8) (30, 30, 30) (30, 30, 30) 287s 312s 283s 308s 281s 306s 279s 302s 273s 294s 269s 289s Ellipse Plane wave Point source (1.5, 1.3) (1.3, 1.3) (1.3, 1.0) (1.0, 1.0) (60, 30, 60) (60, 30, 60) 463s 512s 450s 496s 443s 491s 438s 484s 422s 470s 416s 459s

(C1 , C2 ) (D1 , D2 ) (N1 , N2 , N3 ) k = 200 k = 400 k = 800 k = 1600 k = 10000 k = 100000

of the parameters N1 , N2 and N3 that determine the total degrees of freedom. In applications, one is usually interested in the quantity qs /k. For example in electromagnetics, this quantity is proportional to the induced current on the surface of the obstacle with a proportionality constant that is independent of k. One can see from (4.9) that qs /k corresponds to a normalisation of qs for dierent k. For this reason, in all following results we evaluate the absolute error in a point t as Ea (t) = |qs (t) qc (t)| , k (6.1)

and the relative error as Er (t) = |qs (t) qc (t)| . |qs (t)| (6.2)

The exact solution qs of the scattering problem is known analytically only for the case of scattering of a plane wave by a circle. For the other cases, the error is estimated by comparing computed solutions to each other. The relative and absolute error for the scattering of a plane wave by the circle is shown in Figure 8 for increasing values of k. We have chosen to use cubic spline basis functions, and derivatives up to order d = 1 in each specialised quadrature rule. The gures show that both the absolute error and the relative error decrease with increasing k. This is due to the higher accuracy of the quadrature rule Qc [f ; t] at larger frequencies. The relative error tends to 100% in the deep shadow region because we have approximated the solution by 0 in that region. One can verify from the gures that the absolute error in that region is still quite small compared to the average value of the function qs (t). At k = 1600, a relative error of order 105 is achieved in the illuminated region and near the shadow boundary. The accuracy shown in Figure 8 for the smaller value k = 200 is quite poor however. The accuracy is much improved by increasing the number of derivatives used in the quadrature rules. This is illustrated in Figure 9(a) for the same scattering example with k = 200. The number of derivatives d ranges from 0 to 4. We used cubic spline basis functions for d = 0, 1, 2 22

10

10

10

10 10
5

10

10

k=200 k=400 k=800 k=1600 0.2 0.4 0.6 0.8 1

10

10

15

10

10

0.2

0.4

0.6

0.8

(a) Absolute error Ea

(b) Relative error Er

Figure 8: Absolute and relative error for the scattering of a plane wave by a circle for dierent values of k. We have used d = 1 derivative in the specialised quadrature rules. The other parameters correspond to those in Table 2. and splines of degree 5 for d = 3, 4. A relative error of order 108 is achieved in the shadow boundary region, and a relative error of order 105 in the illuminated region. A similar experiment is illustrated in Figure 9(b), corresponding to an incident circular wave on an ellipse-shaped boundary. The error for d = 0 . . . 3 in this case was estimated by comparing to the results of using d = 4. Table 3: Maximal absolute error as a function of the size of the shadow boundary region, and pointwise absolute and relative error in a xed point T = 0.13 in the shadow boundary region. The constants Cj = C and Dj = D are chosen pairwise equal (k = 200, d = 2, Nj = 50). (C, D) Ea Ea (T ) Er (T ) (1.0, 0.7) 1.0E 3 4.4E 4 3.2E 1 (1.1, 0.8) 4.6E 4 7.9E 5 5.6E 2 (1.2, 0.9) 2.4E 4 3.6E 5 2.6E 2 (1.3, 1.0) 1.8E 4 5.1E 6 3.7E 3

The constants C1 , C2 , D1 and D2 determine the size of the transitional shadow boundary regions. Since the solution is approximated by zero in the shadow region outside these intervals, it can be expected that increasing the value of these parameters results in smaller errors. This is conrmed by the results in Table 3 that show the maximum of the absolute error Ea . The maximum is reached in the points tsb1 D1 k1/3 and tsb2 + D2 k1/3 , i.e., on the border of the shadow region. It is also shown that the accuracy of a point in the shadow boundary region increases if the size of that region is increased. The accuracy of the points in the illuminated region is unaected by this change. Finally, Table 4 shows the result of increasing the total number of degrees of freedom for the same scattering example. The accuracy initially increases rapidly, but remains stable as N increases further. In conclusion, the accuracy of the proposed method is inuenced by a number of factors: the value of k, the accuracy of the specialised quadrature rules, the size of the shadow boundary 23

10

10

10 10
0

10 10
5

10 d=0 0.2 d=1 0.4 d=2 0.6 d=3 0.8 d=4 1

d=0 10
8

d=1 0.4

d=2 0.6

d=3 0.8 1

10

10

0.2

(a) Relative error for the case of a circle

(b) Absolute error for the case of an ellipse

Figure 9: Comparison of the absolute and relative error at k = 200 for dierent values of d, the number of derivatives used in the specialised quadrature rules. The left panel corresponds to the case of column 1 in Table 2, the right panel corresponds to the case of column 4. Table 4: Maximal absolute error as a function of the number of degrees of freedom. N1 = N2 = N3 Ea 20 1.5E 2 30 2.4E 3 40 4.9E 4 50 2.2E 4 60 1.9E 4

region and the number of degrees of freedom. The results show that the error is quite small, and can be improved by changing the parameters. However, the experiments do not indicate that arbitrary accuracy can be achieved for a given value of k. For achieving high accuracy at very low values of k, the use of specialised quadrature rules is not recommended; one can instead employ the dense discretisation described in 5.1.

Concluding remarks

The method presented in this paper achieves a sparse discretisation matrix for an integral equation, which is made possible by the presence of strong oscillations. The method is based on a classical boundary element technique near the shadow boundary, and uses a new quadrature formula that is very eective for oscillatory integrals in the illuminated region. The a priori knowledge of the phase of the solution allows a discretisation with a xed, small number of unknowns. Since the new quadrature rule requires only few quadrature points, a sparse discretisation matrix is obtained. Moreover, numerical results indicate that the accuracy of the solution increases with increasing frequency. The method can still be improved in a number of ways. For example, we have only considered a localised Filon-type method, while numerical experiments indicate that regular Filon-type methods are typically more accurate for the same frequency. One could also add additional quadrature points besides the singular and stationary points to improve the accuracy. The small, densely discretised part near the shadow boundary may possibly be avoided by using

24

a more elaborate ansatz for the asymptotic behaviour of the solution, such as those described by the Geometric Theory of Diraction [17, 23, 15]. The method is limited to smooth convex obstacles. The approach of [7] for smooth convex obstacles can be extended to multiple scattering congurations using an iterative approach [14]. We expect the same will hold for our approach. The extension to three-dimensional problems is the subject of future research. This might be based on suitable cubature rules for multivariate highly oscillatory integrals which have recently been constructed [18].

References
[1] T. Abboud, J.-C. Ndlec, and B. Zhou. Mthode des quations intgrales pour les e e e e e hautes frquences. C. R. Acad. Sci. Paris, 318:165170, 1994. e [2] K. R. Aberegg and A. F. Peterson. Application of the integral equation-asymptotic phase method to two-dimensional scattering. IEEE Trans. Antennas Propagat., 43(5):534537, 1995. [3] M. Abramowitz and I. A. Stegun. Handbook of mathematical functions with formulas, graphs, and mathematical tables. Dover Publications, New York, 1965. [4] P. Bettess. Short wave scattering, problems and techniques. Phil. Trans. R. Soc. Lond. A, 362:421443, 2004. [5] N. Bleistein. Asymptotic expansions of integral transforms of functions with logarithmic singularities. SIAM J. Math. Anal., 8(4):655672, 1977. [6] N. Bleistein and R. Handelsman. Asymptotic expansions of integrals. Holt, Rinehart and Winston, New York, 1975. [7] O. P. Bruno, C. A. Geuzaine, J. A. Monro, and F. Reitich. Prescribed error tolerances within xed computational times for scattering problems of arbitrarily high frequency: the convex case. Phil. Trans. R. Soc. Lond. A, 362(1816):629645, 2004. [8] D. Colton and R. Kress. Integral equation methods in scattering theory. Wiley, New York, 1983. [9] R. Cools and A. Haegemans. Algorithm 824: CUBPACK: A package for automatic cubature; framework description. ACM Transactions on Mathematical Software, 29(3):287 296, 2003. [10] E. Darve. The fast multipole method: Numerical implementation. J. Comput. Phys., 160(1):195240, 2000. [11] P. J. Davis and P. Rabinowitz. Methods of numerical integration. Computer Science and Applied Mathematics. Academic Press, New York, 1984. [12] V. Dom nguez, I. G. Graham, and V. P. Smyshlyaev. A hybrid numerical-asymptotic boundary integral method for high-frequency acoustic scattering. Numerische Mathematik, (BICSP 1/06), 2006.

25

[13] V. A. Fock. Electromagnetic diraction and propagation problems. Pergamon Press, New York, 1965. [14] C. Geuzaine, O. Bruno, and F. Reitich. On the O(1) solution of multiple-scattering problems. IEEE Trans. Magn., 41(5):14881491, 2005. [15] E. Giladi. Asymptotically derived boundary elements for the Helmholtz equation in high frequencies. J. Comput. Appl. Math., 198(1):5274, 2007. [16] E. Giladi and J. B. Keller. A hybrid numerical asymptotic method for scattering problems. J. Comput. Phys., 174(1):226 247, 2001. [17] R. F. Harrington. Time-harmonic electromagnetic elds. McGraw-Hill, New York, 1961. [18] D. Huybrechs and S. Vandewalle. The construction of cubature rules for multivariate highly oscillatory integrals. Math. Comp., 2006. To appear. [19] D. Huybrechs and S. Vandewalle. On the evaluation of highly oscillatory integrals by analytic continuation. SIAM J. Numer. Anal., 44(3):10261048, 2006. [20] A. Iserles and S. P. Nrsett. Ecient quadrature of highly oscillatory integrals using derivatives. Proc. R. Soc. Lond. Ser. A Math. Phys. Eng. Sci., 461:13831399, 2005. [21] A. Iserles, S. P. Nrsett, and S. Olver. Highly oscillatory quadrature: the story so far. In Proceedings of ENuMath 2005, Santiago de Compostela, Berlin, 2006. Springer Verlag. [22] D. S. Jones. High-frequency refraction and diraction in general media. Phil. Trans. R. Soc. Lond. A, 255:341387, 1963. [23] J. B. Keller. Geometrical theory of diraction. J. Opt. Soc. Am., 52:116130, 1962. [24] S. Langdon and S. N. Chandler-Wilde. Implementation of a boundary element method for high frequency scattering by convex polygons. In K. Chen, editor, Proc. 5th U.K. Conf. on Boundary Integral Methods, pages 211, 2005. [25] S. Langdon and S. N. Chandler-Wilde. A wavenumber independent boundary element method for an acoustic scattering problem. SIAM J. Numer. Anal., 43(6):24502477, 2006. [26] D. Levin. Fast integration of rapidly oscillatory functions. J. Comput. Appl. Math., 67(1):95101, 1996. [27] R. B. Melrose and M. E. Taylor. Near peak scattering and the corrected Kirchho approximation for a convex obstacle. Adv. in Math., 55(3):242315, 1985. [28] J.-C. Ndlec. Acoustic and Electromagnetic Equations, volume 144 of Applied Mathee e matical Sciences. Springer, Berlin, 2001. [29] S. Olver. Moment-free numerical integration of highly oscillatory functions. IMA J. Num. Anal., 26(2):213227, 2006. [30] V. Rokhlin. Diagonal forms of translation operators for the Helmholtz equation in three dimensions. Appl. Comput. Harmon. Anal., 1(1):8293, 1993. 26

[31] E. M. Stein. Harmonic analysis: Real-variable methods, orthogonality and oscillatory integrals. Princeton University Press, Princeton, 1993. [32] G. N. Watson. Harmonic functions associated with the parabolic cylinder. Proc. London Math. Soc. Series 2, 17:116148, 1918. [33] R. Wong. Asymptotic approximation of integrals. SIAM, Philadelphia, 2001.

27