Journal of Computational and Applied Mathematics 440 (2024) 115622

Contents lists available at ScienceDirect

Journal of Computational and Applied

Mathematics
journal homepage: www.elsevier.com/locate/cam

Nyström discretizations of boundary integral equations for the

solution of 2D elastic scattering problems
⇤
Víctor Domínguez a,b , , Catalin Turc c
a
Dep. Estadística, Informática y Matemáticas, Universidad Pública de Navarra, Campus de Tudela 31500 Tudela, Spain
b
Institute for Advanced Materials (INAMAT2 ), Campus de Arrosadia, 31500 Pamplona, Spain
c
Department of Mathematical Sciences, New Jersey Institute of Technology, University Heights. 323 Dr. M. L. King Jr.
Blvd, Newark, NJ 07102, USA

article info a b s t r a c t

Article history: We present three high-order Nyström discretization strategies of various boundary
Received 28 June 2022 integral equation formulations of the impenetrable time-harmonic Navier equations
Received in revised form 6 October 2023 in two dimensions. One class of such formulations is based on the four classical
MSC: Boundary Integral Operators (BIOs) associated with the Green’s function of the Navier
65N38 operator. We consider two types of Nyström discretizations of these operators, one that
35J05 relies on Kussmaul–Martensen logarithmic splittings (Chapko et al., 2000; Domínguez
65T40 and Turc, 2000), and the other on Alpert quadratures (Alpert, 1999). In addition, we
65F08 consider an alternative formulation of Navier scattering problems based on Helmholtz
decompositions of the elastic fields (Dong et al., 2021), which can be solved via a
Keywords:
Time-domain and time-harmonic Navier system of boundary integral equations that feature integral operators associated with
scattering problems the Helmholtz equation. Owing to the fact that some of the BIOs that are featured in
Boundary integral equations those formulations are non-standard, we use Quadrature by Expansion (QBX) methods
Nyström discretizations for their high order Nyström discretization. Alternatively, we use Maue integration by
Preconditioners parts techniques to recast those non-standard operators in terms of single and double
layer Helmholtz BIOs whose Nyström discretizations is amenable to the Kussmaul–
Martensen methodology. We present a variety of numerical results concerning the high
order accuracy that our Nyström discretization elastic scattering solvers achieve for both
smooth and Lipschitz boundaries. We also present extensive comparisons regarding the
iterative behavior of solvers based on different integral equations in the high frequency
regime. Finally, we illustrate how some of the Nyström discretizations we considered
can be incorporated seamlessly into the Convolution Quadrature (CQ) methodology to
deliver high-order solutions of the time domain elastic scattering problems.
© 2023 Elsevier B.V. All rights reserved.

1. Introduction

Boundary integral equation based numerical solutions of scattering problems have certain inherent advantages over
their volumetric counterparts, and thus they have attracted significant attention in the literature in the past four decades.
The dimensional reduction and implicit enforcement of radiation conditions that solvers based on BIE enjoy, however,
come with the challenge of resolving singular boundary integrals. In the arena of Nyström discretizations a powerful
methodology to deal with such singularities is the logarithmic splitting technique of Kussmaul–Martensen, which consists

⇤ Corresponding author.
E-mail addresses: [email protected] (V. Domínguez), [email protected] (C. Turc).

https://doi.org/10.1016/j.cam.2023.115622
0377-0427/© 2023 Elsevier B.V. All rights reserved.
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

of making explicit the singular parts of various boundary operator kernels and subsequently dealing with them on a case
by case basis. This technique has been successfully applied to first produce Nyström discretizations of the Helmholtz
Boundary Integral Operators (BIOs) [1,2] and subsequently of Navier BIOs [3,4]. In the case of elastic scattering problem
the fundamental solution of the Navier problem is significantly more complicated than its Helmholtz counterpart, which
brings a host of new challenges in performing the singularity splitting technique. As documented in [4], the Kussmaul–
Martensen singularity splitting technique, although effective, becomes extremely cumbersome in the case of Navier BIOs,
which may be a deterrent for its practical use. We explore in this paper alternative high-order Nyström discretization
strategies that, owing to their being simpler to implement, may be more attractive than the Kussmaul–Martensen
approach.
One such alternative strategy is the use of Alpert quadratures [5] which resolve to high order logarithmic singularities.
Basically, as shown in [4], the differences between the three elastodynamics BIO which are not themselves weakly singular
and their elastostatic counterparts are all weakly singular, that is their most singular component is logarithmic, and
thus the Alpert quadrature is directly applicable to the discretization of such difference operators. The most important
observation is that Alpert quadratures are applicable in a black box manner to the discretization of those difference
operators without the need whatsoever to explicitly account for the logarithmic part of those kernels. The elastostatic
BIOs (sometimes referred to as elasticity BIOs), on the other hand, have been studied in detail [6], and their Nyström
discretization is amenable to integration by parts and periodic Hilbert transform techniques [2]. Based on the strategy
outlined above we implement high-order Alpert quadrature Nyström discretizations of elastodynamics BIOs for both
smooth as well as Lipchitz boundaries, incorporating in the latter case sigmoid graded meshes [1] to resolve corner
singularities. Furthermore, following the roadmap in [7,8], we extend our frequency domain elastodynamics solvers to
high-order solutions of the time domain elasticity scattering problems via the Convolution Quadrature (CQ) methodology.
Indeed, using Laplace transforms, the CQ methodology [9–11] reduces the solution of retarded potential formulations
of wave equations to the solution of ensembles of Laplace domain elastodynamic problems. Alpert quadratures are
seamlessly applicable to the discretization of the elastodynamic BIOs featuring fundamental solutions of the Navier
equation with complex frequencies which are required in the CQ methods, and thus we derive high-order in time solutions
of time domain elastic scattering equations.
We also consider an entirely different BIE strategy for the solution of elastodynamics scattering problems that relies on
the Helmholtz decomposition of the elastic waves into compressional and shear waves. The scalar functions corresponding
to this Helmholtz decomposition, in turn, are radiative solutions of the Helmholtz equation with the pressure and
respectively shear wave number, whose normal and tangential traces are coupled on the boundary of the scatterer.
Seeking those aforementioned scalar functions in the form of Helmholtz layer potentials of corresponding wave numbers,
the Navier scattering problems is reduced to a system of BIE which feature Helmholtz BIOs. This approach has been
advocated in [12,13] for Dirichlet boundary conditions, and we extend it in this work to the case of Neumann boundary
conditions. Using Helmholtz layer potential representations, we arrive at a system of BIE that feature certain Helmholtz
BIOs which, in addition to being expressed in terms of Hadamard finite parts integrals as they involve the Hessian of the
Helmholtz Green function, are non-standard in the sense that they do not result from the application of the Dirichlet
and Neumann boundary traces to the usual single and double layer Helmholtz potentials. Furthermore, the derivation of
the ensuing system of BIE requires the use of certain jump conditions [14] which are not typically encountered in this
context. This procedure allows us to recast non-standard Helmholtz BIOs in terms of double and single layer Helmholtz
BIOs whose Nyström discretization is relatively straightforward within the Kussmaul–Martensen paradigm. Since the
considered Nystrom method uses product integration rules for trigonometric polynomials, it can be understood as a
combination of trigonometric interpolation and exact integration. It is easy to construct a reproducing kernel for the
evaluation of trigonometric interpolating polynomials of degree n. This kind of construction has been used for both,
evaluation and design of numerical methods. See for instance [15] for an introducing and this topic and [16–19] for
examples of application of these techniques.
We find that the Quadrature by Expansion (QBX) method is a more straightforward strategy [20,21] for the discretiza-
tion of the Helmholtz decomposition elasticity BIEs. QBX relies on smooth extensions of layer potentials evaluated in the
exterior/interior PDE domains onto the boundary which are achieved in practice via Fourier–Bessel expansions connected
to the addition theorem for Hankel functions. Thus, the application of boundary traces to layer potentials amounts to term
by term differentiation of the Fourier–Bessel expansions in the QBX framework. An attractive feature of QBX methods is
that the evaluation of the Fourier–Bessel expansion coefficients does not require resolution of kernel singularities. As such,
we derive relatively straightforward QBX Nyström discretizations of the Helmholtz decomposition BIE which converge
with high order for both smooth and Lipchitz scatterers through the use of panel Chebyshev meshes and Crenshaw–Curtis
quadratures for the evaluation of the QBX coefficients.
Arguably, the discretization of the Helmholtz decomposition BIE formulation of the Navier scattering problems is
simpler than that of the BIE counterparts that use the Navier Green functions. In order to gain more insight into the
properties of these two types of BIE formulations, we undergo a comparative study on their iterative behavior in the high
frequency regime. In the case of Navier Green function based BIE formulations for the solution of scattering problems, both
combined field and regularized combined field formulations are available in the literature [4,22,23]. These formulations
are proven to be robust for all frequencies, and they exhibit superior iterative behavior in the high frequency regime to
the Helmholtz decomposition BIE we consider in this paper, even after the combined field approach is applied to the
latter. The analysis of the robustness of the combined field Helmholtz decomposition BIE is currently under investigation.
2
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

The paper is organized as follows: in Section 2 we introduce the Navier equations that govern elastodynamics waves
in two dimensions and their fundamental solution, and we review their associated boundary layer potentials and integral
operators associated; in Section 3 we review several Combined Field BIE formulations of Navier scattering problems which
were discussed in detail in our previous contribution [4] as well as the Helmholtz decomposition BIE derived in [13] for
Dirichlet boundary conditions and extended in this work to the case of Neumann boundary conditions; several versions
of Nyström discretizations of the elastodynamics BIOs as well as the Helmholtz BIOs that are featured in the Helmholtz
decomposition BIE are presented in Section 4; finally, we present in Section 5 a variety of numerical results showcasing
the high-order convergence achieved by the Nyström discretizations of elastodynamics frequency as well as the iterative
behavior of various BIE formulations in the high frequency regime ; we conclude Section 5 with an application of the
discretizations discussed in the present work to the solution of the transient elastic wave equation by multistep and
multistage Convolution Quadrature methods.

2. Boundary elements methods for Navier equations

2.1. Navier equations

Let u(x1 , x2 ) = (u1 (x1 , x2 ), u2 (x1 , x2 )) : R2 ! R2 be a vector function. For a linear isotropic and homogeneous elastic
medium with Lamé constants and µ such that > 2µ, the strain and stress tensor are given by
 1
1 @x1 u1 @x1 u2 + @x2 u1
✏(u) := (r u + (r u)> ) = 1
2
2 2
@x1 u2 + @x2 u1 @x2 u2
(u) := 2µ✏(u) + (div u)I2 ,

where div u := @x1 u1 +@x2 u2 is the divergence operator and I2 , obviously, the identity matrix of order 2. The time-harmonic
elastic wave (Navier) equation is defined by

div (u) + !2 u = µ u + ( + µ)r (div u) + !2 u = 0

where the frequency ! 2 R+ and the divergence operator is applied to (u) row-wise. Considering a bounded domain
⌦ in R2 whose boundary is a closed Lipchitz curve, we are interested in solving the impenetrable elastic scattering
problem in ⌦+ , the exterior of ⌦ , that is look for solutions of the time-harmonic Navier equation

div (u) + !2 u = 0 in ⌦+ := R2 \ ⌦ (2.1)

that satisfy the Kupradze radiation condition at infinity (cf. [24], [25, Ch. 2]; see also (5.1)). On the boundary the solution
u of (2.1) satisfies either the Dirichlet boundary condition

u = u| =f
or the Neumann boundary condition

Tu = g

where f, g : ! C are sufficiently regular functions and T is the associated normal stress tensor (or traction operator)
on given by

T u := (u)n = (div u)n + 2µ(n · r )u µ(curl u)t .

Here, curl u = @x1 u2 @x2 u1 is the rotational or scalar curl of u, n the unit outward normal derivative and t = ( n2 , n1 )
is the unit positive orientated tangent vector.

2.2. Fundamental solution of Navier equation and the associated boundary integral operators

For x = (x1 , x2 ), y = (y1 , y2 ) 2 R2 , we will denote

r := x y, r := |r | = |x y |.

The fundamental solution of the time-harmonic elastic wave is given by

1 1 i (1)
(x, y) := (r) = 0 (ks r)I2 + 2
rx rx> ( 0 (ks r) 0 (kp r)), 0 (z) := H0 (z), (2.2)
µ ! 4
(1)
with H0 the Hankel function of first kind and order 0 so that 0 (kz) is the fundamental solution of the Helmholtz equation
2
0 +k 0 = 0 and

!2 !2
k2p := , k2s := (2.3)
+ 2µ µ
3
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

the (squared) pressure and shear wave-number. Boundary integral equation formulations of the Navier equations rely on
the Navier layer potentials and their associated BIOs which we will review in what follows.
For a given density (vector) function : ! C2 , the Navier single layer potential is defined as
Z
(S )(z) := (z , y) (y) dy , z 2 R2 \ . (2.4)

The single layer potential is continuous in R2 , and thus the single layer boundary integral operator can be defined as
Z
(V )(x) := lim (S )(x + " n(x)) = (x, y) (y) dy , x2 . (2.5)
"!0

For g : ! C2 , the double layer potential is, on the other hand, given by
Z
⇥ ⇤>
(D g )(z) := Ty (z , y) g (y) dy , z 2 R2 \ (2.6)

where Ty (z , y) is the normal stress tensor applied column-wise to (z , y) with respect to the y variable. The double
layer potential D undergoes a jump discontinuity across so that
1
lim (D g )(x ± " n(x)) = ± g + (K g )(x), x2
"!0+ 2
if is sufficiently smooth around x, where the Double Layer BIO is defined explicitly as
Z
⇥ ⇤>
(K g)(x) := p.v. K (x, y)g (y) dy , K (x, y) := Ty (r) .

Unlike the single layer BIO, the double layer BIO K is no longer weakly singular and the integral above exists only in the
sense of a Cauchy principal value, hence the ‘‘p.v.’’ notation used above.
The application of the traction operator to the single layer potential gives rise to jump discontinuities
lim (T S ')(x ± " n(x)) = ⌥'(x) + (K > ')(x), x2
"!0

where the adjoint double layer operator is given by

Z
>
(K ')(x) := K > (x, y)'(y) dy , K > (x, y) := Tx (x, y).

Finally, applying the traction operator to the double layer potential we obtain
lim Tx (D g )(x ± " n(x)) = (W g )(x), x2
"!0

where the BIO W is defined as

Z
⇥ ⇤
(W g )(x) := f.p. W (x, y)g (y) dy , W (x, y) := Ty Tx (x, y) .

The kernel W (x, y) is strongly singular (that is, it behaves like O(|x y | 2 ) as y ! x), and as such the integral in its
definition must be interpreted in a Hadamard finite part sense which is precisely what ‘‘f.p.’’ stands for.

2.2.1. Singularity subtraction and parameterized version for the Navier boundary integral operators
Let us consider the BIOs for elasticity V 0 , K 0 , K >
0 and W 0 , operators defined in the same manner from the fundamental
solution of the elasticity problem
✓ ◆
+ 3µ +µ 1
0 (x , y) = log r I2 + G(r) , G(r) = rr > ,
4⇡ µ( + 2µ) + 3µ r2
It can be shown that the difference between the corresponding operators K K 0 , K > K > 0 and W W 0 are logarithmic,
and so weakly singular, integral operators.
Indeed, if x : R ! is a smooth, regular, 2⇡ periodic, counterclockwise oriented parameterization of a smooth curve
, introducing the parameterized version of the densities
(t) = (x(t))|x0 (t)|, g (t) = g (x(t)),
and the parameterized normal stress tensor defined by
T u(t) = (T u)(x(t))|x0 (t)| (2.7)
we have the corresponding parameterized BIO:
Z 2⇡ Z 2⇡
(V )(t) := (x(t), x(⌧ )) (⌧ ) d⌧ , (K g )(t) := p.v. [T⌧ (x(t), x(⌧ ))]> g (⌧ ) d⌧
0 0
4
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

Z 2⇡ Z 2⇡
(K > )(t) := p.v. Tt (x(t), x(⌧ )) (⌧ ) d⌧ , (W g )(t) := f.p. Tt [T⌧ (x(t), x(⌧ ))]> g (⌧ ) d⌧ .
0 0

We then have that the kernels of the operators, denoted with a slight abuse of notation (we are confident that the context
will make it clear whenever one of these operators arises whether it is the BIOs of the curve or its parameterized version)
by the same symbols V , K , K > and W can be factorized as
t ⌧ t ⌧
V (t , ⌧ ) = V 0 (t , ⌧ ) + A(t , ⌧ ) sin2 log sin2 + B(t , ⌧ )
2 2
t ⌧ t ⌧
K (t , ⌧ ) = K 0 (t , ⌧ ) + C (t , ⌧ ) sin log sin2 + D(t , ⌧ ), K > (t , ⌧ ) = (K (⌧ , t))> , (2.8)
2 2
t ⌧
W (t , ⌧ ) = W 0 (t , ⌧ ) + E(t , ⌧ ) log sin2 + F (t , ⌧ )
2
where
+ 3µ 1 +µ 1
V 0 (t , ⌧ ) := log rI2 + G(r)
µ( + 2µ) 4⇡ µ( + 2µ) 4⇡
✓ ◆
1 0 µ +µ
K 0 (t , ⌧ ) := (x (t) · r) I 2 + 2 G(r)
2⇡ r 2 + 2µ + 2µ
µ( + µ) @ 2 1
W 0 (t , ⌧ ) := ( log rI2 + G(r))
+ 2µ @⌧ @ t ⇡
and r and r are given now by
1
r = r(t , ⌧ ) = x(t) x(⌧ ), r = |r |, G(r) = rr > .
|r |2
Let us notice that V 0 , K 0 , K >
0 and W 0 turn out to be the kernels of the corresponding BIO for elasticity operator, which
with the convention followed so far will be denoted, also, by the same symbols. Functions A, B, C , D and G are smooth
periodic matrix functions. Besides, we notice that the derivation with respect the parameters t , ⌧ is nothing but the
weighted tangential derivative. That is, if
(@s g)(x) := (rv (x)) · t(x), x2 , with v=g
we have
g 0 (t) = |x0 (t)|(@s g)(x(t)).
We will make extensive use of the tangential derivative operator in the next sections.
As an interesting byproduct, the principal symbol of the operators can be easily derived from these expressions see [4]
(we refer the reader to [6, Ch. 10] or [26, Ch. 7] for an introduction to the theory pseudodifferential operators in the
context of boundary integral equations in R2 ). Observe also that the strongly singular part of W can be rewritten as
Z 2⇡
µ( + µ) 1
(W 0 g )(t) = @⌧ ( log rI2 + G(r)) g 0 (⌧ ) d⌧
+ 2µ ⇡ 0

in the parameterized space, which is simply the parameterized version of the well-known Maue-type formula for
elasticity:
Z
µ( + µ) 1
(W 0 g )(x) = @sx ( log rI2 + G(r)) @sy g (y) dy . (2.9)
+ 2µ ⇡

3. Boundary integral formulations

We present in what follows various strategies to derive BIE formulations of elastic scattering problems. Besides the
classical combined field formulations CFIE we consider regularized formulations that rely on the use of approximations of
the Dirichlet-to-Neumann (DtN in what follows) operators whose analysis was given in our previous contribution [4].
The design of the regularized formulations for elastic scattering problems follows the blueprint from the Helmholtz
case [27,28].

3.1. Combined field integral equations

Just like in the Helmholtz case [29], the classical approach [30,31] in the case of Dirichlet boundary conditions is to
look for a scattered field in the form of a Combined Field representation
u(x) := (D g )(x) i⌘D (S g )(x), x 2 R2 \ ⌦
5
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

where the coupling parameter ⌘D 6 = 0, leading to the Combined Field Integral Equation (CFIE)
1
g + K g i⌘D V g = f. (3.1)
2
The question of selecting a value of the coupling parameter ⌘D that leads to formulations with superior spectral properties
(and thus faster convergence rates for iterative solver solutions) can be settled via DtN arguments [31]. We begin with
Somigliana’s identities
u=D u S T u, in ⌦+ = R2 \ ⌦
which we rewrite considering as u as the primary unknown boundary density and incorporating the DtN operator Y
in the form
u=D u S (Y u), in ⌦+ .
The main idea in constructing regularized formulations is to use easy to construct approximations RD of the DtN operator
Y and look for combined field representations in the form

u := D g S (R D g ) (3.2)
leading to the Combined Field Regularized Integral Equation (CFIER)
1
g + Kg V RD g = f. (3.3)
2
Clearly, the better the regularizing operator RD approximates the DtN operator Y , the closer the operator in the left
hand side of the CFIER equation (3.3) is to the identity operator, and, in conclusion, the ensuing CFIER formulations are
more suitable to iterative solver solutions (e.g. GMRES). In our previous effort [4] we proposed the regularizing operator
RD = PS (Y ) which is a Fourier multiplier operator whose symbol  [Y ](⇠ ) is defined by
✓  ◆
4µ( + 2µ) 1 iµ sign(⇠ )
 [Y ](⇠ ) = (|⇠ |2  2 )1/2 I2 + (3.4)
+ 3µ 2 2( + 2µ) sign(⇠ )
where = > 0.
However, the regularizing operator RD = PS (Y ) is a pseudodifferential operator of order 1 whose numerical evaluation
can be consequently more involved. Using the high-frequency approximation || ! 1 in Eq. (3.4) we can construct a
simple regularizing operator
2µ( + 2µ)
RD
1 = i,
+ 3µ
which, incidental, can be interpreted as delivering a quasi-optimal choice for the coupling parameter ⌘D in the CFIE
formulation
2µ( + 2µ)
⌘Dopt = ks (3.5)
+ 3µ
if we choose  = ks . We remark that a similar, easily implementable low-order approximation of the DtN operator was
proposed in [31] as a regularizing operator in the Dirichlet case.
In the case of Neumann boundary conditions, we can look for a scattered field in the form of a Combined Field
representation akin to the Burton–Miller formulation in the Helmholtz case [32]
u := S ' + i⌘N D ', in ⌦+
where the coupling parameter ⌘N 6 = 0, leading to the Combined Field Integral Equation (CFIE)
1
' K > ' + i⌘N W ' = g. (3.6)
2
Here again we start by recasting the Somigliana’s identities looking at T u as the primary unknown boundary density and
making use of the Neumann-to-Dirichlet (NtD) operator (which is the inverse of the DtN operator)
1
u = D (Y T u) S T u, in ⌦+ .
The construction of regularized formulations relies again on available approximations RN of the NtD operator Y 1
via
looking for combined field representations in the form

u := D (RN ') S ', in ⌦+ (3.7)

leading to the Combined Field Regularized Integral Equation (CFIER)
1
' K > ' + W RN ' = g. (3.8)
2
6
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

Again here, the choice RN = PS (Y 1 ) as the Fourier multiplier operator whose symbol is (  [Y ](⇠ )) 1 (the inverse must
be understood in matrix sense per formula (3.4)), leading again to well posed CFIER formulations, at least in the case
when the boundary is smooth [4].
Also, using high-frequency approximations we can construct a simple regularizing operator
+ 3µ
RN
1 = i  1
,
2µ( + 2µ)
which, delivers a quasi-optimal choice for the coupling parameter ⌘N in the CFIE formulation
+ 3µ
⌘Nopt = k 1. (3.9)
2µ( + 2µ) s
We remark that similar low-order approximations of NtD operators have been used in [23] to construct CFIE formulations
with superior spectral properties.
Similarly, we can construct direct regularized formulations in the case of Neumann boundary conditions following the
ideas in [33]. Assuming that a smooth incident field uinc (which is a solution of the Navier equation in the whole R2 )
impinges on the obstacle ⌦ , we will derive these BIEs in terms of unknown boundary quantity if utot = (u + uinc )| . We
obtain from Somigliana’s identities by taking into account the fact that T utot = 0 on

u = D( utot ), in ⌦+ . (3.10)

Applying the Dirichlet trace on to formula (3.10) we obtain

1
( utot ) K( utot ) = uinc , (3.11)
2
while applying the traction operator to formula (3.10) we get

W( utot ) = T uinc . (3.12)

We combine BIE (3.11) and a preconditioned (on the left) version of the BIE (3.12) to arrive at the DCFIER
1
utot K( utot ) + RN W ( utot ) = utot RN (T uinc ). (3.13)
2
We note that the operators on the left hand side of the DCFIER formulation is the real L2 ( ) ⇥ L2 ( ) adjoint of the operator
in the CFIER formulation, a situation that is, similar to that in the Helmholtz case [33]. We will use direct formulations in
the case when ⌦ is a Lipschitz domain in order to take advantage of the increased regularity of utot .

3.1.1. Open arcs

In the case when is an open arc in R2 , the combined field methodology is no longer available. Instead, first kind
formulations can be derived from Somigliana’s identities [3]. Assuming again smooth incident fields uinc , the scattering
problem off of an arc is solved in the case of Dirichlet boundary conditions via the following BIE of the first kind

V [T utot ] = uinc (3.14)

while in the Neumann case via the BIE (3.12). The first kind integral equations of elastodynamic scattering from arcs
can be preconditioned using the approximations of DtN operators introduced above. Thus, we will also consider the
preconditioned BIE

(PS (Y )V )[T utot ] = PS (Y )uinc on (3.15)

and
1
(PS (Y )W )[utot ] = PS (Y 1
)[T uinc ] on . (3.16)

We remark that a different strategy based on Calderón preconditioners have been proposed in [34,35] in order to produce
formulations of elastodynamics scattering problems from arcs that have more suitable spectral properties to iterative
solvers.

3.2. Helmholtz decomposition formulations

Another possibility to construct BIE formulations of Navier problems is via Helmholtz decompositions of the fields u.
Indeed, defining
1 1 !
up := r div u us := curl curl u (3.17)
k2p k2s
7
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

!
with curl ' = [r']? = [@x2 ' @x1 ']> the vector curl operator, we have that u = up + us . Hence, we can look for the
fields u in the form
!
u = r'p + curl 's (3.18)
where the scalar functions 'p and 's are radiative solutions of scalar Helmholtz equations in ⌦+ with wave-numbers kp
and ks respectively. In the case of Dirichlet boundary conditions, it is simply to see that 'p and 's satisfy the following
coupled boundary conditions
@n 'p + @s 's = uinc · n
@s 'p @n 's = uinc · t . (3.19)
Straightforward calculations yield
(r'p ) = 2µH'p k2p 'p I
 
! 1 1
(curl 's ) = 2µH's + µk2s 's
1 1
where H' denotes the Hessian of the scalar function ' . Using these calculations in the case of Neumann boundary
conditions we get in turn that 'p and 's are coupled via the following boundary conditions
⇢
2µ n> H'p n k2p 'p + 2µ n> H's t = T uinc · n
(3.20)
2µ t > H'p n 2µ n> H's n µk2s 's = T uinc · t .
In both cases we look for 'p and 's in the form of Helmholtz single layer potentials corresponding to wave-numbers kp
and respectively ks . That is, we look for unknown functional densities gp and gs defined on such that

'p = Skp gp 's = Sks gs in ⌦+ , (3.21)

where ( 0 is the fundamental solution of the Helmholtz equation cf. (2.2))
Z
Sk g(x) := 0 (k |x y |) g(y) dy, x 2 ⌦+

is the Single Layer BIO for Helmholtz equation with wave-number k and. In the case of Dirichlet boundary conditions, the
system of equations (3.19) is equivalent to the following system of BIE for the boundary densities gp and gs
   1
gp uinc · n I + Kk>p @s Vks
ASL = ASL
DH =
2
(3.22)
DH gs uinc · t @s Vkp 1
I Kk>s
2

where Vk and Kk> are the Helmholtz single layer and respectively the adjoint double layer BIOs associated with the Green’s
function 0 (k·) and @s denotes the tangential derivative operator on . The system of equations (3.22) was shown to
be uniquely solvable in the case when k2p and k2s are not eigenvalues of in the interior domain ⌦ with Dirichlet
boundary conditions [13]. However, the analysis of the invertibility of the operator ASL DH is quite involved on account of
the degeneracy of its principal symbol in the pseudodifferential sense. Indeed, the principal symbol of the operator ASL DH is
a nilpotent matrix operator (actually its square equals zero), and thus the integral formulation (3.22) is far from an optimal
formulation with regards to iterative solvers. Of course, it is possible to look for 'p and 's in the form of Helmholtz double
layer potentials corresponding to wave-numbers kp and respectively ks

'p := Dkp gp 's := Dks gs in ⌦+ , (3.23)

where
Z
Dk g(x) = @n(y) 0 (k |x y |)g(y) dy , x 2 ⌦+

is the Double Layer BIO for the Helmholtz Equation with wave-number k, gp and gs are unknown functional densities
defined on . In the case of Dirichlet boundary conditions, the system of equations (3.19) is equivalent to the following
system of BIE for the boundary densities gp and gs
   1
gp uinc · n Wkp @ + k2s t · Vks [n]
2 s
Kk>s @s
ADL = ADL := 1 (3.24)
DH gs uinc · t DH @ +
2 s
k2p t · Vkp [n] >
Kkp @s Wks

where we denoted by Wk the hyper singular Helmholtz BIO associated with Helmholtz equation with wave-number k.
We also took into account the Maue type formula [1] describing the behavior of the tangential derivative @s on applied
to the exterior double layer potential, which we recount next
1
t· (r + D k ' ) = @s ' + k2 t · Vk [' n] Kk> [@s '], (3.25)
2
8
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

The notation r+ in Eq. (3.25) refers to the application of the gradient in the exterior domain ⌦+ . It is not our intention
to analyze the operator ADL
DH in what follows. Rather, we use a combined field approach

'p := Dkp gp ikp Skp gp 's := Dks gs iks Sks gs in ⌦+ . (3.26)

leading to the CFIE formulation
  
gp uinc · n ikp
ACFIE = ACFIE = ADL ASL (3.27)
DH gs uinc · t DH DH iks DH

which is more suitable for iterative solutions of elastic scattering problems. We leave the analysis of the well possedness
of the CFIE formulation (3.27) for future work.
In order to derive the Neumann counterpart of the BIE system (3.22) we rely on the following trace relation:
11
[H+ Sk g ] = Hk g + ~ ( I + 2nn> )g (nt > + tn> )@s g . (3.28)
2 2
In the identity above, H+ means that the Hessian matrix operator is applied in ⌦+ , the exterior of , (to Sk g) ~ (x) denotes
the signed curvature of at x 2 , and the matrix BIO Hk [g ](x) (understood in the sense of Hadamard finite parts) is
defined as
Z
Hk [g ](x) := f.p. rx rx> 0 (x y)g(y)dy , x2 .

We note that the formula (3.28) also appears in [21] and its justification relies on results established in [14]. We obtain
the following system of BIE
 
gp T uinc · n
ASL =
NH gs T uinc · t
 
2µ n> H kp n µ@s µ I k2p Vkp 2µt > H ks n
ASL
NH = + . (3.29)
µ@s >
2µ n H ks n >
2µ t H kp n µ~ I µk2s Vks
Indeed, formulas (3.20) rely on the following jump relation for the Hessian of the single layer potential on which can be
viewed as Maue’s type formulas Hadamard finite parts integral operators into alternative expressions that involve Cauchy
Principal Value and weakly singular integral operators that recast
Z
lim HSk ['](x+ ) = k2 0 (x y)n(y)n> (y)' (y)dy
x+ !x
Z
⇥ ⇤
+ rx 0 (x y)t > (y) Q rx 0 (x y)n> (y) @s ' (y)dy
Z
⇥ ⇤
rx 0 (x y)n> (y) + Q rx 0 (x y)t > (y) ~ (y)' (y)dy

1⇥ ⇤ 1⇥ ⇤
I + 2n(x)n> (x) ~ (x)' (x)
+ t(x)n> (x) + n(x)t > (x) @s ' (x). (3.30)
2 2
We do not intend to provide neither a full derivation of formulas (3.30), nor an analysis of the integral formula-
tion (3.20) in this paper. It suffices to mention that it is straightforward to see that the principal symbol of the operators
ANH is again defective, and thus the analysis of the formulation (3.20) requires a pseudodifferential calculus beyond
the principal symbol. Also, the single layer formulation (3.20) is not uniquely solvable for all material parameters. This
situation can be remedied through the use of combined field formulations. However, the double layer analogue of the
jump conditions (3.28) are significantly more involved, and we will devote a separate effort to their derivation in which
we will present a complete analysis of robust integral formulations of Helmholtz decomposition based reformulations of
elastic scattering problems. For the sake of completeness we simply recount the following jump relations
Z
k2
lim HDk ['](x+ ) = k2 rx 0 (x y)n> (y)' (y)dy n(x)n> (x)' (x)
x+ !x 2
Z
+ k2 0 (x y)n(y)t > (y)@s ' (y)dy
Z
⇥ ⇤
+ rx 0 (x y)n> (y) + Q rx 0 (x y)t > (y) @s2 ' (y)dy
Z
⇥ ⇤
+ rx 0 (x y)t > (y) Q rx 0 (x y)n> (y) ~ (y)@s ' (y)dy

1⇥ > ⇤ 1⇥ ⇤
Q + 2n(x)t > (x) ~ (x)@s ' (x) I + 2n(x)n> (x) @s2 ' (x) (3.31)
2 2

I
where Q = .
I
9
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

We then demote by ADL NH the boundary integral matrix operator derived from (3.31) which, for the sake of brevity do
not detail here. Hence, the corresponding CFIER formulation, counterpart of is given by
  
gp T uinc · n ikp
ACFIE = ACFIE DL
NH = ANH ASL
NH . (3.32)
NH gs T uinc · t iks

4. Nyström discretizations

4.1. Kussmaul-Martensen based discretizations

We presented in a previous effort a Nyström discretization strategy based on singularity splitting technique and
a resolution of logarithmic singularities which via the Kussmaul–Martensen quadrature [36,37]. In a nutshell, this
discretization strategy relies on global trigonometric interpolation with 2n nodes
j⇡
tj = , j = 0, 1, . . . , 2n 1
n
onto the space of trigonometric polynomials
( n n 1
)
X X
Tn = ' (t) = am cos mt + bm sin mt : am , bm 2 C
m=0 m=1

and the use of the Kussmaul–Martensen quadrature method, a product quadrature method based on the identity
Z 2⇡
1 1
log(sin2 (⌧ /2)) cos m⌧ d⌧ = , m 2 Z \ {0}
2⇡ 0 |m|
to resolve the logarithmic singularity. We will refer to this method as K-M Nyström method and is applicable to the
kernels of the BIOs V , K K 0 , K > K >0 and W W 0 ). With some modifications can be adapted to cover Hilbert transform
quadratures [2], as those appearing in K 0 , K >
0 , and half grid size shifted quadrature methods [1] for the discretization of
the static counterparts W 0 . Given that the Nyström discretizations of the principal parts of the elastodynamic operators
are available in the literature, the main difficulty of the overall collocation schemes resides in the logarithmic splitting
of various kernels. In other words, a precise description of the kernels of the elastodynamic operator presented in (2.8)
is required which can be found in [4, Appendix]. Alternatively, we will present in next subsection a different collocation,
based on specialized quadrature rules for logarithmic singular functions, that acts in a black-box manner in the case of
weakly singular kernels and thus bypasses the need for complicated kernel splittings.
The Kussmaul–Martensen quadratures can be applied also for the Nyström discretization of the Helmholtz decompo-
sition formulations (3.19) and (3.20). While the details of these were provided in the literature [13] in the case of the
single layer formulation (3.19) Dirichlet boundary conditions, the Neumann boundary conditions counterpart (3.20) is
also amenable to such discretizations via the Maue type formulas (3.30) and respectively (3.31). The details concerning
the application of the Kussmaul–Martensen quadratures to the operators in the Eqs. (3.31) and (3.31) that feature the
weakly singular kernels 0 (r) and @n(x) 0 (r) (recall r = |x y |) can be found for instance in the classical Ref. [2], while
the operators that feature the singular kernel @sx 0 (r) can be dealt with using the half grid shifting technique, see for
instance [4]. Furthermore, the derivatives of the functional densities featured in Eqs. (3.30) and (3.31) can be performed
using Fourier differentiation.

4.2. Alpert quadrature

The guiding principle in the splitting calculations presented in the previous sections is the fact that the differences
between dynamic and static versions of the elasticity BIOs are all weakly singular. However, those calculations become
increasingly involved and cumbersome for the double layer and hypersingular elastodynamic BIO. We explore in what
follows an alternative strategy for evaluations of those weakly singular operators that does not require complex singularity
splittings. This alternative relies on Alpert quadratures [5,38] which can be applied seamlessly in our context. Specifically,
Alpert quadrature takes on the form
Z 2⇡ 2n 2a
X
k(ti , ⌧ ) (⌧ )d⌧ ⇡ h k(ti , ti + ah + ph) (ti + ah + ph)
0 p=0
m
X
+h wp k(ti , ti + p h) (ti + p h)
p=1
m
X 2⇡
+h wp k(ti , ti + 2⇡ p h) (ti + 2⇡ p h) , h :=
n
p=1

10
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

where the kernel k(ti , ⌧ ) has a logarithmic singularity at ⌧ = ti . Assuming that the density function is a regular enough
2⇡ -periodic density, the integer parameter a, the weights wp , and the nodes p can be selected in such a matter so that
the ensuing Alpert quadratures achieve prescribed high order convergence. The endpoint correction nodes p are typically
not integers, and as such the Alpert quadratures require evaluation of the density outside of the equispaced mesh {ti }.
This can be achieved by resorting to Lagrange interpolation of order m + 3 that shifts the grid points around the non-grid
points where the density function needs to be evaluated [5,38]. Specifically, we apply Alpert quadrature rules to the
parametrized versions of the BIOs V , K K 0 , K > K > 0 and W W 0 (cf. (2.8)) without resorting to singularity splitting
of their weakly singular kernels. The static BIOs (whose kernels correspond to the 0 subindex in the notations above), on
the other hand, are evaluated via the trigonometric interpolation Nyström methods described above.

4.3. Domains with corners

The extension of the Nyström discretizations of the elastodynamics BIOs to Lipschitz domains can be performed via
sigmoid transforms that accumulate with algebraic orders discretization points toward corners points [28]. We assume
that the domain ⌦ has corners at x1 , x2 , . . . , xP and that \ {x1 , x2 , . . . , xP } is piecewise smooth. We assume that
the boundary curve has a regular 2⇡ -periodic smooth parametrization so that each of the curved segments [xj , xj+1 ]
is parametrized by

xw (t) := (x1 (w (t)), x2 (w (t))), t 2 [Tj , Tj+1 ]

(so that xj = x(Tj )) where 0 = T1 < T2 < · · · < TP < TP +1 = 2⇡ . Here

w : [Tj , Tj+1 ] ! [Tj , Tj+1 ], 1  j  P

is the sigmoid transform of order p introduced by Kress [39]. The function w is a smooth and increasing bijection on each
of the intervals [Tj , Tj+1 ] for 1  j  P, with w (k) (Tj ) = w (k) (Tj+1 ) = 0 for 1  k  p 1 and all 1  j  P. With the aid of
the graded meshes, we can introduce the parameterized traction operator as defined in (2.7)

(T w u)(⌧ ) := |x0 (w (⌧ ))w 0 (⌧ )| (T u)(xw (⌧ ))

as well as the (weighted) parameterized adjoint double layer BIOs

Z 2⇡
K >,w [T w u](t) := Ttw [ (xw (t), xw (⌧ ))] [T w u](⌧ ) d⌧ .
0

Since |w 0 (⌧ )| has zeros of order p 1 at the corners, the singularity of the parameterized density and the kernel of the
adjoint double layer BIO at the corners is cancel out.
With regards to the double layer operator, we remark that while the kernel K (x, y) K0 (x, y) continues to be weakly
singular in the Lipschitz case, the kernel K0 (x, y) itself now becomes singular. We use a trick similar to that employed for
the evaluation of the Laplace double layer operator in the case of domains with corners, that is, we recast the evaluation
of the elastostatic double layer operator as
Z
1
g (x) + (K 0 g )(x) = K 0 (x, y)(g (y) g (x)) dy .
2
The previous identity is a direct consequence of the integration by parts formula (3.3.20) in [6] and well known results
for the Laplace double layer operator [1]. For the evaluation of the hypersingular operator W , we note that the operator
W W 0 is still weakly singular in the case when is Lipschitz. According to the discussion in Section 3, we prefer the
use of direct formulations of scattering and transmission elastodynamic problems in Lipschitz domains, whose unknowns
are u and/or T w u and which incorporate weighted BIOs. The Alpert discretizations can be immediately extended to
the case of Lipschitz domains through the aid sigmoid transforms, an insight which has been first proposed in [7] in the
case of Helmholtz BIO. It is indeed an advantage of Alpert quadratures the feature that they are readily amenable to any
situation which involves weakly singular kernels.
Furthermore, the Kussmaul–Martensen quadratures can be also extended in principle via sigmoid transforms for the
discretization of the Helmholtz decomposition formulations (3.19) and (3.20) in the case of domains with corners.

4.4. QBX

Given that the BIOs featured in the BIE systems (3.22) and (3.20) result from the application of various traces on to
Helmholtz single layer potentials, QBX methods [20,21] are particularly advantageous alternatives in delivering high-order
Nyström discretizations of such BIOs, especially for the nonstandard ones involving Hessians of single layer potentials.
Indeed, for a given wave-number k and functional density ' on , the main thrust of QBX methods is extending single
layer potentials Sk ['](x+ ), x+ 2 ⌦+ to ⌦+ using Fourier–Bessel series expansions. The application of boundary traces to
single layer potentials, therefore, is a matter of term by term differentiation of the Fourier–Bessel series expansions in the
QBX discretization paradigm, as we explain in what follows. We start with presenting the details of the QBX method for
11
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

the evaluation of single layer BIO. QBX relies on expansion centers, that is for x 2 we define x± = x ±" (x)n(x), " (x) > 0,
and on the addition theorem for Hankel functions
1
X
(1) (1) 0 )+ i`✓ +
H0 (k|x y |) = H` (k|x+ y |)ei`(✓ J` (k|x x+ |)e , y2 (4.1)
`= 1

where ✓ + and (✓ 0 )+ are the angular coordinates of x and respectively y in the polar coordinate system centered at
x+ . Considering the smooth extension of Sk ['](x+ ) onto as " (x) ! 0, we obtain the following Fourier–Bessel series
representation of the single layer BIO Vk ['](x)
1
X Z
i`✓ + i (1) 0 )+
Vk ['](x) = ↵`+ (x)J` (k|x x+ |)e , ↵`+ (x) := H` (k|x+ y |)ei`(✓ ' (y) dy . (4.2)
4
`= 1

Since for any function defined on a neighborhood of the tangential and normal derivatives of the expansions (4.2)
in polar coordinates centered at x+ are given by
1
@s (x) = r (x) · t(x) = @✓ (r , ✓ ), @n (x) = r (x) · n(x) = @r (r , ✓ ),
|x x+ |
(here (r , ✓ ) = (x+ + (r cos ✓ , r sin ✓ )), with r = |x x+ |) we have
1
X J` (k|x x+ |) i`✓ +
@s Vk ['](x) = i `↵`+ e (4.3)
|x x+ |
`= 1

and
1
X ✓ ◆
1 ` i`✓ +
' (x) + Kk> ['](x) = k ↵`+ (x) J`+1 (k|x x+ |) + J` (k|x x+ |) e , (4.4)
2 k|x x+ |
`= 1

that is, the Fourier–Bessel series representations of the BIOs featuring in Eqs. (3.22). Notice that in (4.4) we have used the
well-known identity for the derivative of the Bessel functions
`
J`0 (x) = J`+1 (x) + J` (x), ` 2 Z.
x
Similarly, applying the Hessian in polar coordinates on the expansions (4.2), taking into account appropriate jump relations
on as " (x) ! 0 and using instead
`
J`0 (x) = J` 1 (x) J` (x), `2Z
x
we obtain the following Fourier–Bessel series representations of the BIOs featuring in Eqs. (3.20), that is
 (x)
' (x) + n> (x)Hk ['](x)n(x) =
2 !
1
X 2
kJ`+1 (k|x x+ |) (`2 ` k2 |x x+ | )J` (k|x x+ |) i`✓ +
↵`+ (x) + e (4.5)
`= 1
|x x+ | |x x+ |2

and respectively
1
@s ' (x) + n> (x)Hk ['](x)t(x) =
2
X1 ✓ ◆
kJ`+1 (k|x x+ |) (1 `)J` (k|x x+ |) i`✓ +
i `↵`+ (x) + e . (4.6)
`= 1
|x x |+
|x x+ |2

It is also possible to average limits from ⌦+ and ⌦ to obtain alternative series expansions for traces on of single layer
potentials, especially in the case when those undergo jump discontinuities across the boundary. Specifically, we denote by
✓ and (✓ 0 ) the angular coordinates of x and respectively x0 in the polar coordinate system centered at x , and we derive
the alternative Fourier–Bessel series expansion of the single layer BIO via smooth extensions from the interior domain
⌦ onto :
1
X Z
i`✓ i (1) 0)
Vk ['](x) = ↵` (x)J` (k|x x |)e , ↵` (x) := H` (k|x0 x |)ei`(✓ ' (x0 ) dx0 . (4.7)
4
`= 1

12
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

Using the jump conditions of traces of single layer potentials, we derive the two-sided Fourier–Bessel series representa-
tions
1 ✓ ◆
1 k X ` i`✓ +
Kk> ['](x) = ' (x) ↵`+ (x) J`+1 (k|x x+ |) + J` (k|x x+ |) e
2 2 k|x x+ |
`= 1
1 ✓ ◆
k X ` i`✓
+ ↵` (x) J`+1 (k|x x |) + J` (k|x x |) e , (4.8)
2 k|x x |
`= 1

we easily derive

n> (x)Hk ['](x)n(x) =

1
!
2
1 X kJ`+1 (k|x x+ |) (` 2 ` k2 |x x+ | )J` (k|x x+ |) i`✓ +
+ ↵`+ (x) + e
2
`= 1
|x x+ | |x x+ |2
1
!
X 2
1 kJ`+1 (k|x x |) (` 2 ` k2 |x x | )J` (k|x x |) i`✓
+ ↵` (x) + e (4.9)
2
`= 1
|x x | |x x |2

and finally

n> (x)Hk ['](x)t(x) =

1 ✓ ◆
i X kJ`+1 (k|x x+ |) (1 `)J` (k|x x+ |) i`✓ +
+ `↵`+ (x) + e
2
`= 1
|x x+ | |x x+ |2
X1 ✓ ◆
i kJ`+1 (k|x x |) (1 `)J` (k|x x |) i`✓
+ `↵` (x) + e . (4.10)
2
`= 1
|x x | |x x |2

In the case when Double layer potentials are used in the Helmholtz decomposition formulations, their related BIOs
that feature in the systems of boundary conditions (3.19) and (3.20) can be evaluated using the same term by term
differentiation of Fourier–Bessel series strategy presented above simply replacing the coefficients ↵`+ by their double
layer counterparts defined as
Z ✓ ◆
i (1) ` (1) (x+ y) · n(y) 0 )+
↵`DL,+ (x) := H`+1 (k|x+ y |) H` (k|x+ y |) ei`(✓ ' (y)dy
4 k|x+ y| |x+ y|
Z
` (1) ( sin (✓ 0 )+ , cos (✓ 0 )+ ) · n(y) 0 )+
H` (k|x+ y |) ei`(✓ ' (y)dy . (4.11)
4 |x+ y|
The Fourier–Bessel expansions above constitute the basis of QBX Nyström discretizations of BIOs in Eqs. (3.22) and
(3.20). The full discretizations of those BIOs is achieved by (a) selecting a truncating parameter p in the Fourier–Bessel
series above, and (b) projecting the functional densities ' into appropriate discrete functional spaces and thus effecting
corresponding collocation quadratures for the evaluation of the expansion coefficients ↵`± for p  `  p. We note that
the integrands in the definition of coefficients ↵`± do not exhibit kernel singularities. In the case of smooth boundaries ,
we consider global trigonometric interpolation of the densities ' using the 2n equispaced nodes tm = mn⇡ , 0  m  2n 1
and trapezoidal quadratures for the evaluation of Fourier–Bessel expansion coefficients ↵`± (x(tm )), 0  m  2n 1.
In order to achieve uniform errors using the quadratures (4.12) for all indices ` (the integrands in Eqs. (4.12) get
increasingly oscillatory and the near-singularities of the Hankel functions more stringent as the indices ` get larger),
Fourier interpolation is used to oversample the density ' on a finer uniform mesh of size 2n0 = 2 n where 2 Z, > 1.
j⇡
We then define ⌧j = 2n0 and the coefficients ↵`± are evaluated using the trapezoidal rule on the finer mesh:
2n 1 0
i⇡ X (1) 0 )±
↵`± (x(tm )) ⇡ ±
↵`, n0
(x(tm )) := H` (k|x± (tm ) x(⌧j )|)ei`(✓ ' (x(⌧j ))|x0 (⌧j )| (4.12)
4n0
j=0

where x± (tm ) = x(tm ) ± "m n(x(tm )) with "m = min(|x(tm ) x(tm 1 )|, |x(tm ) x(tm+1 )|) (here we use cyclical indexing so
that t 1 := t2n 1 and t2n := t0 ). In short, we have approximations as
p
X
± i`✓
Vk ['](x(tm )) ⇡ ↵`, n0
(x(tm ))J` (k|x(t) x± (t)|)e (4.13)
`= p

We will also accurately evaluate the QBX expansion coefficients ↵`± using more general quadrature rules that are
applicable in the case when is piecewise smooth. Specifically, we consider a panel representation of the boundary
13
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

SM
curve in the form = r =1 r where the panels r are non overlapping. In the case when exhibits corner points
x1 , x2 , . . . , xP , each corner point is an end point of a panel. We thus have
M
X Z
i (1) 0 )±
↵`± (x) = ±
↵`, r (x),
±
↵`, r (x) := H` (k|x± y |)ei`(✓ ' (y) dy . (4.14)
4 r
r =1

Assuming that each panel r is parametrized in the form r = {zr (t) : t 2 [ 1, 1]} where zr : [ 1, 1] ! r is smooth,
we consider a Chebyshev mesh on the parameter space [ 1, 1]
(2m 1)⇡
tm := cos #m , #m := , m = 1, . . . , nr .
2nr
±
The coefficients ↵`, r are evaluated by a combination of oversampling and Fejér–Clenshaw–Curtis quadratures. Specifically,

nm 0
i X i`(✓ 0 )±
±
↵`, r (x(tm ) ⇡ ±
↵`, r ,n 0
(x(tm )) := !j H`(1) (k|x± (tm ) zr (⌧j )|)e j ' (zr (⌧j ))|zr (⌧j )| (4.15)
4
j=1

with, as before, n0m = nm and consequently

(2m 1)⇡
⌧` := cos #` , #` := , m = 1, . . . , n0r .
2n0 r

The Fejér quadrature weights !j , in turn, are given by

0 1
[n0r /2]
2 X 1
!j := @1 2 cos(2q#j )A , j = 1, . . . , n0r .
n0r 4q2 1
q=1

Again here, for a Chebyshev mesh point zr (tm ) on , we choose its corresponding centers along the exterior/interior
normal to at zr (tm ) located distance "j,r = min(|zr (tm ) zr (tm+1 )|, |zr (tm ) zr (tm 1 )|) from z(tm ). The evaluation of the
density at n0r quadrature points is also carried out by (Chebyshev) interpolation of ' from the coarse mesh with nr nodes.

5. Numerical results

We present in this section a variety of numerical results about the accuracy of Nyström discretizations of the
elastodynamic BIE solvers discussed in this text. Specifically, we show far field accuracy results of solvers based on
CFIE formulations as well as Helmholtz decomposition formulations (3.20) and (3.22). In addition, we study the iterative
behavior of solvers based on the aforementioned formulations using GMRES [40] iterative solvers for the solution of
the linear systems ensuing from Nyström discretizations. While the size of the linear systems we considered allows
for application of direct solvers, the iterative behavior of BIE formulations does shed light on the iterative properties
of their three dimensional counterparts. Finally, we present numerical results concerning BIE based CQ solutions of time
dependent elasticity scattering problems.
For a scattered elastic field u the associated longitudinal wave up and the transversal wave us [25] or [24, Ch. 2] are
defined as in Eqs. (3.17). The Kupradze radiation conditions [25] simply state that the functions 'p and 's defined in
Eqs. (3.18) are radiative solutions of the Helmholtz equation in the unbounded domain ⌦+ with wave-numbers kp and
respectively ks , which in vector form amounts to
✓ ✓ ◆◆ ✓ ✓ ◆◆
eikp |x| 1 eiks |x| 1
up (x) = p up,1 (x̂) + O us (x) = p us,1 (x̂) + O (5.1)
|x| |x| |x| |x|
as |x| ! 1 where x̂ = x/|x|. We assess the accuracy of our solvers using the metric of maximum far field errors "1 of
the quantities up,1 and us,1 evaluated at fine enough meshes on the unit circle |x̂| = 1. Per usual, we consider both types
of incident fields in our scattering experiments, that is, elastodynamic point sources and plane waves. The former type of
incident field is used in the context of manufactured solutions wherein the numerical errors are evaluated against an exact
solution. On the other hand, in the latter case of incident fields we computed far field errors with respect to reference
solutions produced through very fine discretizations of the underlying BIE. We also report the number of unknowns N
used in the discretization of each of the two unknowns of in the elastodynamic BIE systems considered in this text (in
the case of Nyström discretizations that use the trigonometric polynomial space Tn we have N = 2n). We start with
numerical examples related to the method of manufactured solutions.
The geometries considered in our numerical experiments are:

1. The starfish domain [38] whose 2⇡ periodic parametrization is given by

✓ ◆
1
x(t) = 1+ sin 5t (cos t , sin t). (5.2)
4
14
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

2. The cavity-like geometry whose parametrization is given by

1
x(t) = 4
(cos t + 2 cos 2t , A(t)/2 As (t)/48),
A(t) = sin t + sin 2t + 1/2 sin 3t , As (t) = 4 sin t + 7 sin 2t 6 sin 3t + 2 sin 4t . (5.3)

3. The kite domain cf. [41]

x(t) = (cos t + 0.65(cos 2t 1), 1.5 sin t),

4. The teardrop domain given by

x(t) = 2| sin 2t |, sin t , (5.4)

which presents a corner point at x(0) = (0, 0)

5. The boomerang domain,
2 3t
x(t) = 3
sin 2
, sin t ,

which present also a corner point at x(0) = (0, 0).

6. The flat line, an open arc given by

x(t) = (t , 0), t 2 [ 1, 1].

7. The V-shaped polygonal line,

⇣ ⌘
1 p1
x(t) = t , |t |
2 2
p , t2[
2
, p12 ]

and open curve with a corner point in t = 0 of length 2.

We depict in Fig. 1 such geometries.

5.1. The method of manufactured solutions

The method of manufactured solutions amounts to solving time-harmonic Navier equation with boundary value data
produced by point sources x0 placed inside the scatter ⌦

uinc (x) = (x, x0 ), x 2 , x0 2 ⌦

so that the solution of impenetrable scattering problems in the exterior domain ⌦+ are the point sources themselves, that
is, u(x) = (x, x0 ) for all x 2 ⌦+ . In the case of Dirichlet boundary conditions, looking for scattered fields in the form of
a single and double layer potentials corresponding to boundary functional densities ' and respectively g , we solve the
ensuing BIEs
1
(V ')(x) = (x, x0 ) and g (x) + (K g)(x) = (x, x0 ), x 2 .
2
In the case of Neumann boundary conditions, the same approach leads to solving the following BIEs
1
'(x) + (K > ')(x) = Tx (x, x0 ), and (W g )(x) = Tx (x, x0 ), x 2 .
2
In either type of boundary conditions we compare the numerical solutions in the far field against the exact point source
solution u(x) = (x, x0 ). The method of manufactured solutions in the case of Helmholtz decomposition BIE (3.20)
and (3.22) also amounts to consider the same incident field uinc (x) = (x, x0 ), x 2 whose decomposition (3.18) is
straightforward to effect by simply separating the kp and ks contributions via the appropriate Hankel functions in the
fundamental solution of the Navier equation.
In Table 1 we report the far field errors in the method of manufactured solutions in the case of a smooth starfish
boundary. Specifically, we present errors corresponding to (1) the single layer formulation with Dirichlet boundary
conditions (in the rubric V , given that the single layer BIO V is used to validate the method of manufactured solutions),
(2) the double layer formulation with Dirichlet boundary conditions (in the rubric K ), and (3) the double layer formulation
with Neumann boundary conditions (in the rubric W ). We mention that the BIO K > is the (real) L2 ⇥L2 adjoint of the BIO K ,
and at the discrete level the Nyström matrices corresponding to those two operators are the transpose of one another. We
considered two types of Nyström discretizations, one based on Kussmaul–Martensen (K-M) logarithmic splitting strategy
of all of the weakly singular kernels (which we labeled under the KM header) and another based on Alpert 10th order
quadrature (which does not require any splittings whatsoever of same kernels). The weights and the location of the off-
grid nodes required by the Alpert 10th order quadrature are tabulated in [5,38], which, for the sake of brevity, we chose
not to reproduce here. While the errors produced using the Alpert quadratures seem to saturate around 10 10 for both
the single and double layer formulations of the Dirichlet problems and to 10 6 in the case of the hyper singular operators
15
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

Fig. 1. Geometries for the experiments considered in this section, on the top, from left to right: the smooth domains, the starfish, the cavity, and
the boomerang curve; on the middle, the corner (Lipchitz) closed geometries: the teardrop and the boomerang curve; on the bottom, the open arcs:
the flat line and the V-shaped curve.

Table 1
Errors in the method of manufactured solution using the elastodynamics BIOs for the smooth starfish geometry for
different values of the frequency ! and parameter values = 1, µ = 1, at various levels of discretization using the
Kussmaul–Martensen logarithmic splitting Nyström discretizations as well as the 10th order Alpert discretization with
parameters a = 6 and m = 10.
! N V K W
KM "1 10 Alpert "1 KM "1 10 Alpert "1 KM "1 10 Alpert "1
2 2
16 64 3.0 ⇥ 10 1.5 ⇥ 10 9.9⇥ 10 2 1.3 ⇥ 10 1
1.3 ⇥ 10 1
1.7 ⇥ 10 1
7 6
16 128 7.1 ⇥ 10 1.7 ⇥ 10 1.1 ⇥ 10 3 3.1 ⇥ 10 5
1.6 ⇥ 10 6
7.4 ⇥ 10 3
14 10
16 256 5.5 ⇥ 10 1.7 ⇥ 10 4.2 ⇥ 10 13 3.4 ⇥ 10 9
1.6 ⇥ 10 9
3.2 ⇥ 10 5

2 2 1 1 1 2
32 128 3.0 ⇥ 10 1.3 ⇥ 10 2.5 ⇥ 10 2.4 ⇥ 10 3.4 ⇥ 10 6.5 ⇥ 10
7 6 4 6
32 256 2.2 ⇥ 10 1.4 ⇥ 10 2.4 ⇥ 10 7.0 ⇥ 10 5.0 ⇥ 10 8 4.6 ⇥ 10 3
15 10 13 10
32 512 1.0 ⇥ 10 1.5 ⇥ 10 8.3 ⇥ 10 9.0 ⇥ 10 5.4 ⇥10 12 2.8 ⇥ 10 5

(in contrast, the corresponding errors using Kussmaul–Martensen splittings can reach full double precision levels), the
application of Alpert quadratures to the elastodynamic 2D BIE solvers is significantly simpler.
We present in Table 2 errors in the method of manufactured solutions achieved by QBX Nyström discretizations of
the Helmholtz decomposition BIE formulations (3.22) of the elastodynamics scattering problem with Dirichlet boundary
conditions (K-M results are available in the literature [13] for the solution of the same equation (3.22)). Both, equispaced
16
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

Table 2
Errors in the method of manufactured solution using the Helmholtz decomposition BIE
(3.22) for the smooth starfish geometry with Dirichlet boundary conditions for different
values of the frequency ! and parameter values = 1, µ = 1, at various levels of
discretization using QBX discretizations with global equispaced and Chebyshev meshes
with M = 1 (only one panel) and with expansions parameters p = 16 and = 10.
! N Dirichlet b.c. Helmholtz decomposition BIE (3.22)
"1 Trapezoidal QBX (4.12) "1 Chebyshev QBX (4.15)
4 2
16 64 2.7 ⇥ 10 1.5 ⇥ 10
8 6
16 128 6.6 ⇥ 10 1.7 ⇥ 10
12 10
16 256 8.6 ⇥ 10 1.7 ⇥ 10
1 2
32 128 2.2 ⇥ 10 1.1 ⇥ 10
5 7
32 256 5.3 ⇥ 10 1.8 ⇥ 10
11 9
32 512 2.9 ⇥ 10 1.3 ⇥ 10

Table 3
Errors in the method of manufactured solution using the Helmholtz decomposition BIE
(3.22) for the smooth starfish geometry with Dirichlet boundary conditions for the
frequency ! = 16 and parameter values = 1, µ = 1, using different numbers (M) of
Chebyshev panels of different sizes to discretize the boundary, as well as various choices
of the QBX expansion parameters p and . The number of unknowns for each experiment
is N = nM.
Chebyshev QBX (4.15) Chebyshev QBX (4.15)
(M , n) (p, ) "1 (M , n) (p, ) "1
3 2
(8,16) (4,4) 2.4 ⇥ 10 (32,8) (2,2) 1.5 ⇥ 10
5 3
(8,32) (4,4) 9.7 ⇥ 10 (32,8) (4,2) 2.0 ⇥ 10
6 4
(8,32) (8,6) 4.0 ⇥ 10 (32,8) (6,4) 2.6 ⇥ 10
7 5
(8,32) (16,10) 1.9 ⇥ 10 (32,8) (12,8) 2.1 ⇥ 10

and Chebyshev meshes are used in this experiment, with p = 16, i.e. 33 terms in the expansion (see (4.13)), = 10 in
the oversampling and a global panel, i.e. M = 1 which means that only one panel is being used in the Chebyshev mesh
case. We point out that in this experiment the one sided (4.4) and two-sided (4.8) QBX expansions led to almost identical
levels of accuracy for the same size of discretizations.
We continue in Table 3 with the same setup in the method of manufactured solutions but using different numbers (M)
of Chebyshev panels on the starfish contour (the interval [0, 2⇡] was split into M equal parts and each panel corresponds
to the mapping of one such subinterval via the parametrization of ) and values of the QBX expansion parameters p
and . The density is therefore computed at nr points per panel, and so it amounts to nr M on the curve. In this and the
following experiments in this section, we will take nr = n, i.e., for Chebyshev mesh the same number of points per panel
will always be used.
Qualitatively similar results are obtained when the double layer formulation (3.24) (and hence the CFIE (3.27)) is used
for the solution of the Helmholtz decomposition approach.
Table 4 illustrates the accuracy levels achieved by the Nyström K-M and QBX discretizations of the Neumann Helmholtz
decomposition BIE (3.20) in the case of the starfish scatterer. The K-M Nyström discretization is applied to the recasting of
the Hessian operators that feature in Eqs. (3.20) via the Maue integration by parts techniques in formulas (3.30) and (3.31).
In the case of QBX discretizations we used a global equispaced mesh and corresponding trapezoidal quadratures for the
evaluation of the Fourier–Bessel coefficients (4.12). We discretized the BIOs that enter the BIE formulation (3.20) via one
sided QBX representations (4.5) and (4.6) and respectively two sided QBX representations (4.9) and (4.10). Interestingly, it
appears that the one sided QBX representations lead to more accurate solutions, which is relevant given that (a) they entail
half the computational cost of two-sided expansions, and, more importantly (b) the one sided expansions are oblivious of
the complicated jump relations (3.28). We mention that we observed a similar slight increase in accuracy when using one
sided QBX representations for other geometries and incident fields. We continue in Table 5 with numerical experiments
concerning QBX discretizations of the BIE formulations (3.20) using one sided expansions, and representations of as
a union of Chebyshev panels and various levels of discretizations and values of the QBX expansion parameters. Again,
high-order convergence is observed. Qualitatively similar results are obtained when the double layer approach is used for
the solution of the Neumann Helmholtz decomposition approach.
We will examine now the accuracy in the method of manufactured solutions in the case of Lipschitz geometries in
Table 6. Specifically, we used the teardrop geometry (5.4) in our numerical tests in conjunction with polynomially graded
meshes of order p = 4 in the sigmoid transform and Kussmaul–Martensen kernel splittings as well as Alpert 3rd order
method (using Lagrange interpolation with stencils of size 6 to evaluate the values of the densities at off grid locations).
We have found that amongst all possible Alpert quadratures, the use of 3rd order Alpert quadratures delivers in practice
nearly optimal accuracy given the lower order of regularity of BIOs functional densities in the Lipschitz case. However, we
17
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

Table 4
Errors in the method of Neumann manufactured solution for the smooth starfish geometry using the
Helmholtz decomposition formulation (3.20) for different values of the frequency ! and parameter values
= 1, µ = 1, at various levels of discretization using K-M and QBX discretizations with a global
equispaced mesh and expansion parameters p = 16 and = 10. The number of unknowns for each
experiment is N = nM.
! N Neumann b.c. Helmholtz decomposition BIE (3.20)
K-M One sided QBX "1 Two sided QBX "1
3 3 3
16 64 5.1 ⇥ 10 2.9 ⇥ 10 3.3 ⇥ 10
5 7 6
16 128 1.3 ⇥ 10 2.1 ⇥ 10 1.4 ⇥ 10
11 11 10
16 256 8.2 ⇥ 10 2.8 ⇥ 10 1.2 ⇥ 10
3 3 3
32 128 2.4 ⇥ 10 1.4 ⇥ 10 1.3 ⇥ 10
10 8 8
32 256 2.1 ⇥ 10 4.9 ⇥ 10 5.0 ⇥ 10
13 11 10
32 512 9.4 ⇥ 10 5.6 ⇥ 10 3.1 ⇥ 10

Table 5
Errors in the method of manufactured solution using the Helmholtz decomposition BIE
(3.20) for the smooth starfish geometry with Neumann boundary conditions for the
frequency ! = 16 and parameter values = 1, µ = 1, using different numbers (M) of
Chebyshev panels of different sizes to discretize the boundary, as well as various choices
of the QBX expansion parameters p and . The number of unknowns for each experiment
is N = nM.
Chebyshev QBX (4.15) Chebyshev QBX (4.15)
(M , n) (p, ) "1 (M , n) (p, ) "1
3 2
(8,16) (4,4) 2.3 ⇥ 10 (32,8) (2,2) 1.3 ⇥ 10
4 4
(8,32) (4,4) 3.3 ⇥ 10 (32,8) (4,4) 3.7 ⇥ 10
6 5
(8,32) (6,6) 6.6 ⇥ 10 (32,8) (8,6) 3.5 ⇥ 10
7 6
(8,32) (12,8) 2.4 ⇥ 10 (32,16) (8,8) 2.0 ⇥ 10

Table 6
Errors in the method of manufactured solution for the teardrop geometry for different values of the
frequency ! and parameter values = 1, µ = 1, at various levels of discretization using graded sigmoid
meshes with p = 3, the Kussmaul–Martensen logarithmic splitting Nyström discretizations as well as
the 3rd order Alpert discretization with parameters a = 2 and m = 3.
! n Vw K W
KM "1 3 Alpert KM "1 KM "1 3 Alpert "1 KM "1
2 3 1 2 1
16 32 8.1 ⇥ 10 3.8 ⇥ 10 1.6 ⇥ 10 1.0 ⇥ 10 1.7 ⇥ 10
5 5 4 3 3
16 64 8.4 ⇥ 10 6.4 ⇥ 10 3.4 ⇥ 10 2.3 ⇥ 10 2.2 ⇥ 10
12 6 9 4 4
16 128 4.1 ⇥ 10 2.0 ⇥ 10 2.3 ⇥ 10 2.8 ⇥ 10 4.2 ⇥ 10
2 3 1 2 1
32 64 4.7 ⇥ 10 6.4 ⇥ 10 2.0 ⇥ 10 1.1 ⇥ 10 1.3 ⇥ 10
6 4 4 3 3
32 128 8.1 ⇥ 10 1.7 ⇥ 10 1.7 ⇥ 10 1.7 ⇥ 10 1.4 ⇥ 10
13 6 10 4 4
32 256 2.6 ⇥ 10 6.4 ⇥ 10 2.7 ⇥ 10 2.7 ⇥ 10 3.5 ⇥ 10

do not observe high-order convergence when we applied Alpert quadratures to the Nyström discretization of the hyper
singular operator W . We note that similar levels of accuracy are attained in the case of plane wave incidence.
Tables 7 and 8 illustrate, again for Lipschitz scatterers, the accuracy of Helmholtz decomposition BIE formulations (3.20)
and (3.22) for the Navier equations with Dirichlet and respectively Neumann boundary conditions. Global Chebyshev
meshes in those experiments with Clenshaw–Curtis quadratures (4.15) are applied for the evaluations of the Fourier–
Bessel coefficients in the QBX method. Again here, the one sided expansions appear to lead to more accurate solutions. We
perform the same experiments in Tables 9 and 10 but using this time dyadic refinement of Chebyshev panels around the
corner, and various values of the QBX expansion parameters p and . Similar levels of accuracy are observed for scatterers
with re-entrant corners as well as multiple corners. We remark that the QBX discretizations of the BIE formulations (3.22)
and (3.20) are rather straightforward to implement even for Lipschitz boundaries, at least when one sided expansions are
used (since one needs not take into account more complicated jump properties across ), and significantly simpler than
discretizations of BIE based on the Navier fundamental solution.
In short, we have illustrated in this part the fact that QBX discretizations of the Helmholtz decomposition BIE
formulations of the Navier scattering equations lead to the same levels of accuracy as the other Nyström discretizations
of the elastodynamics BIEs that use the Navier fundamental solution, while being simpler to implement. We turn our
attention next to the iterative behavior of the BIE formulations considered in this text in the high frequency regime.
18
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

Table 7
Errors in the method of manufactured solution using the Helmholtz decomposition BIE (3.22) for the
teardrop geometry with Dirichlet boundary conditions for different values of the frequency ! and
parameter values = 1, µ = 1, at various levels of discretization using QBX discretizations with global
equispaced and Chebyshev meshes, with expansions parameters p = 16 and = 10.
! N Dirichlet b.c. Helmholtz decomposition BIE (3.22)
"1 Chebyshev QBX one sided (4.15) "1 Chebyshev QBX two-sided (4.15)
6 4
16 64 4.2 ⇥ 10 1.6 ⇥ 10
9 6
16 128 6.8 ⇥ 10 1.6 ⇥ 10
10 9
16 256 8.8 ⇥ 10 6.5 ⇥ 10
4 4
32 128 1.0 ⇥ 10 1.6 ⇥ 10
8 7
32 256 8.5 ⇥ 10 4.6 ⇥ 10
10 9
32 512 1.1 ⇥ 10 4.3 ⇥ 10

Table 8
Errors in the method of manufactured solution using the Helmholtz decomposition BIE (3.20) for the
teardrop geometry with Neumann boundary conditions for different values of the frequency ! and
parameter values = 1, µ = 1, at various levels of discretization using QBX discretizations with global
equispaced and Chebyshev meshes, with expansions parameters p = 12 and = 8.
! N Neumann b.c. Helmholtz decomposition BIE (3.20)
"1 Chebyshev QBX one sided (4.15) "1 Chebyshev QBX two-sided (4.15)
5 4
16 64 1.4 ⇥ 10 1.4 ⇥ 10
6 5
16 128 3.6 ⇥ 10 2.0 ⇥ 10
7 7
16 256 5.3 ⇥ 10 4.0 ⇥ 10
4 3
32 128 3.0 ⇥ 10 1.4 ⇥ 10
6 7
32 256 2.0 ⇥ 10 4.0 ⇥ 10
8 8
32 512 1.1 ⇥ 10 3.1 ⇥ 10

Table 9
Errors in the method of manufactured solution using the Helmholtz decomposition BIE (3.22) for the
teardrop geometry with Dirichlet boundary conditions for the frequency ! = 16 and parameter values
= 1, µ = 1, using different numbers (M) of Chebyshev panels with dyadic corner refinement of
different sizes to discretize the boundary, as well as various choices of the QBX expansion parameters
p and . The number of unknowns for each experiment is N = nM.
Dirichlet BIE (3.22) QBX (4.15) Dirichlet (3.22) QBX (4.15)
(M , n) (p, ) "1 (M , n) (p, ) "1
4 2
(8,16) (4,4) 7.8 ⇥ 10 (32,16) (2,2) 1.6 ⇥ 10
5 4
(8,32) (4,4) 6.1 ⇥ 10 (32,16) (4,4) 6.1 ⇥ 10
6 5
(8,32) (6,6) 8.1 ⇥ 10 (32,16) (8,6) 5.0 ⇥ 10
7 6
(8,32) (12,8) 2.2 ⇥ 10 (32,16) (8,8) 6.8 ⇥ 10

Table 10
Errors in the method of manufactured solution using the Helmholtz decomposition BIE (3.20) for the
teardrop geometry with Neumann boundary conditions for the frequency ! = 16 and parameter values
= 1, µ = 1, using different numbers (M) of Chebyshev panels with dyadic corner refinement of
different sizes to discretize the boundary, as well as various choices of the QBX expansion parameters
p and . The number of unknowns for each experiment is N = nM.
Neumann BIE (3.20) QBX (4.15) Neumann BIE (3.20) QBX (4.15)
(M , n) (p, ) "1 (M , n) (p, ) "1
3 2
(8,16) (4,4) 1.9 ⇥ 10 (32,16) (2,2) 1.0 ⇥ 10
4 3
(8,32) (4,4) 4.0 ⇥ 10 (32,16) (4,4) 1.6 ⇥ 10
6 4
(8,32) (6,6) 7.7 ⇥ 10 (32,16) (8,6) 3.7 ⇥ 10
7 5
(8,32) (12,8) 6.2 ⇥ 10 (32,16) (8,8) 4.5 ⇥ 10

5.2. Iterative behavior of BIE formulations

We devote this section to comparisons between the iterative behavior of the various BIE formulations for the solution
of high-frequency elastic impenetrable scattering problems considered in this text under plane wave incident fields of
the form
1 1
uinc (x) = eiks x·d (d ⇥ p) ⇥ d + eikp x·d (d · p)d (5.5)
µ + 2µ
19
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

Table 11
Numbers of GMRES iterations of various formulations to reach residuals of 10 5 for various BIE
formulations of Dirichlet elastic scattering problems at high frequencies in the case when is the kite
contour. The material parameters are = 2 and µ = 1, and the incidence was P-wave. The discretizations
used in these numerical experiments delivered results accurate at the level of 10 5 . The CFIE formulation
was discretized using Kussmaul–Martensen (K-M) logarithmic splittings Nyström method, while the BIE
(3.22) were discretized using QBX with a global equispaced mesh and expansion parameters p = 12 and
= 6.
! N # iter CFIE (3.5) ⌘D opt # iter Helmholtz decomposition
BIE (3.22) and (3.27)
10 64 25 205/100
20 128 27 302/124
40 256 31 437/145
80 512 34 727/166
160 1024 38 1401/183

Table 12
Numbers of GMRES iterations of various formulations to reach residuals of 10 5 for various BIE
formulations of Dirichlet elastic scattering problems at high frequencies in the case when is the
teardrop. The material parameters are = 2 and µ = 1, and the incidence was P-wave. The
discretizations used in these numerical experiments delivered results accurate at the level of 10 5 .
The BIE (3.22) and (3.27) were discretized using QBX with a global Chebyshev mesh and expansion
parameters p = 12 and = 6.
! N # iter CFIE (3.5) # iter one sided # iter two sided # iter one sided
⌘D opt QBX (3.22) QBX (3.22) QBX (3.27)
10 64 23 170 206 70
20 128 28 292 338 80
40 256 34 605 578 99
80 512 40 1207 1137 137
160 1024 50 1497 1405 186

where the direction d has unit length |d | = 1. If the vector p is chosen such as p = ±d, the incident plane is a pressure
wave or P-wave. In the case when p is orthogonal to the direction of propagation
⇥ p,
⇤>the incident plane wave is referred to
as a shear wave or S-wave. We considered plane waves of direction d = 0 1 in all of our numerical experiments;
⇥ ⇤>
in the case of S-wave incidence we selected p = 1 0 . We observed that other choices of the direction d and of the
vector p lead to qualitatively similar results.

5.2.1. Dirichlet boundary conditions

We investigated the iterative behavior of Dirichlet integral solvers based on two formulations: (1) the CFIE formulation
with the optimal coupling constant ⌘D given in Eq. (3.5)—which we refer to by the acronym ‘‘CFIE ⌘D opt’’; and (2) the
Helmholtz decomposition single layer BIE formulations (3.22) and their CFIE versions (3.27). We report in Tables 11– 13
the number of GMRES iterations required by each of these two BIE formulations to reach relative residuals of 10 5 . The
corresponding far field errors are also at the 10 5 level. in the case of smooth scatterers. Besides the starfish and the
teardrop geometry, we also considered the Lipschitz boomerang geometry given by the discretization. As it can be seen
from the results presented in Tables 11–13, the Navier CFIE formulation (3.3) with the optimal coupling parameter ⌘D
exhibits the best iterative behavior in the high-frequency regime. The iterations counts corresponding to the Navier CFIE
formulation (3.3) displayed in this section are almost identical for K-M and Alpert discretizations.
We note that given that the Navier double layer operators K and K > as well as the Helmholtz operators @s Vk
and Wk are not compact, neither of the Navier CFIE (3.3) nor the single layer and CFIE Helmholtz decomposition BIE
formulations (3.22) and (3.27) are of the second kind. Nevertheless, the Navier CFIE formulation (3.3) and the single
layer Helmholtz decomposition BIE formulations (3.22) behave in practice as second kind formulations in the sense that
the numbers of GMRES iterations required to reach a certain residual do not increase with more refined discretizations.
However, that is not the case for the CFIE Helmholtz decomposition BIE formulations (3.27), largely on the account of the
hyper singular operators Wk . Yet, the CFIE Helmholtz decomposition BIE formulations appear to be a superior alternative
to the single layer Helmholtz decomposition BIE with regards to iterative solvers in the high frequency regime. We used
QBX Nyström discretizations based on global equispaced (in the smooth boundaries case) and respectively Chebyshev
meshes for the discretizations of the Helmholtz decomposition BIE formulations (3.22). We note that a Kussmaul–
Martensen splitting Nyström discretizations of the BIE formulations (3.22) is relatively straightforward to implement,
and the ensuing numbers of GMRES iterations are slightly smaller than the ones resulting from QBX discretizations. In
the QBX discretizations the use of one sided expansions led to identical results with respect to GMRES iterations with the
versions using two sided expansions. Finally, we observed that the numbers of GMRES iterations grow with the number of
panels used to represent the boundary when QBX discretizations are used (with the same size of overall discretization).
20
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

Table 13
Numbers of GMRES iterations of various formulations to reach residuals of 10 5 for various BIE
formulations of Dirichlet elastic scattering problems at high frequencies in the case when is the
boomerang. The material parameters are = 2 and µ = 1, and the incidence was P-wave. The
discretizations used in these numerical experiments delivered results accurate at the level of 10 5 .
The BIE (3.22) and (3.27) were discretized using QBX with a global Chebyshev mesh and expansion
parameters p = 8 and = 4.
! N # iter CFIE (3.5) # iter one sided # iter two sided # iter one sided
⌘D opt QBX (3.22) QBX (3.22) QBX (3.27)
10 128 29 198 204 72
20 256 34 346 367 107
40 512 41 623 635 130
80 1024 49 1270 1229 165
160 2048 59 1468 1418 214

Table 14
Numbers of GMRES iterations of various formulations to reach residuals of 10 6 for various BIE formulations of Dirichlet
elastic scattering problems at high frequencies in the case when is a flat strip of length 2 (left panel) and a V-shaped
arc of length 2 (right panel). The material parameters are = 2 and µ = 1, and the incidence was an P-wave. The
discretizations used in these numerical experiments delivered results accurate at the level of 10 5 .
! N # iter V w # iter PS (Y+ )V w # iter V w # iter PS (Y+ )V w
10 32/64 15/18 21/21 33/39 26/27
20 64/128 20/23 27/27 46/52 38/39
40 128/256 25/29 34/34 67/75 56/55
80 256/512 36/42 42/42 86/98 80/80
160 512/1024 51/61 53/53 111/126 117/116

We present next in Table 14 results concerning scattering from arcs with Dirichlet boundary conditions. We considered
a flat strip of length 2 and the V-shaped arc. We display the number of iterations required by the classical single layer
formulation (referred to as V w in the column header) and the preconditioned on the left formulation that uses the
operator composition PS (Y+ )V w . We regularize the end-point square-root singularities of T utot [34,35] by resorting to its
weighted version T w utot which, on account that the derivatives of the sigmoid transform vanish at the end-point of the
2⇡ parametrization of the arc, vanishes itself at both endpoints 0 and 2⇡ and thus can be extended as a regular enough
2⇡ periodic function. Clearly, the use of a weighted traction calls for the use of the weighted single layer BIO V w . We
report numbers of GMRES iterations required to reach residuals of 10 6 and discretizations that deliver results at a 10 5
level of accuracy for the line segment and respectively 10 4 for the V-shaped arc. The V w formulation is a first kind BIE,
and hence the numbers of GMRES iterations required to reach a given tolerance grow upon refinement of discretizations,
albeit modestly so. On the other hand, the analytically preconditioned formulation PS (Y+ )V w appears to behave like an
integral equation of the second kind whose numbers of GMRES iterations are insensitive to the size of the discretization
for a given frequency. We remark that the computational overhead required by the application of the Fourier multiplier
PS (Y+ ) is insignificant as its application can be performed efficiently using FFTs. The findings reported in Table 14 appear
to be qualitatively similar to those in [34] where Calderón preconditioners are used.

5.2.2. Neumann boundary conditions

We present next in Tables 15–18 results related to the iterative behavior of the BIE formulations considered in
this paper for the solution of elastodynamic problems with Neumann boundary conditions in the case of smooth and
Lipschitz scatterers. Specifically, we consider (1) the Navier CFIE formulation (3.6) with the optimal coupling constant ⌘N ,
(2) the Navier CFIER formulation with the regularizing operator RN = (PS (Y+ )) 1 and (3) the single layer Helmholtz
decomposition BIE formulation (3.20) and its CFIE version. Both formulation (3.6) and (3.20) feature hyper singular
operators and therefore the number of GMRES iterations required for the iterative solution of their associated Nyström
linear systems grows with the size of the discretizations. The results reported in Tables 15–18 correspond to K-M
discretizations for smooth scatterers as well as one-sided QBX expansions for the discretization of the BIE (3.20); the use of
two sided expansions leads to very similar iteration counts. Several remarks and observations are in order. First, the K-M
and Alpert Nyström discretizations of the Navier CFIE and CFIER formulations lead to almost identical iteration counts.
Second, the Navier CFIE/CFIER formulations appear to enjoy superior iterative performance over the single layer and
combined field Helmholtz decomposition formulations of the Neumann elastic scattering problems in the high frequency
regime. Furthermore, the CFIE approach performs worse than the single layer approach in the case of the Helmholtz
decomposition approach for Neumann elastic scattering, a fact which can be explained by the fact that the double layer
formulation (3.31) involves higher order derivatives than its single layer counterpart (3.28). The construction of regularized
formulations of the Helmholtz decomposition approach is currently under investigation.
Finally, we conclude the numerical results with an illustration in Table 19 of scattering results in the case when
is an open arc with Neumann boundary conditions. The solution utot of the BIE (3.12) can be shown to vanish at the
21
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

Table 15
Numbers of GMRES iterations of various formulations to reach residuals of 10 5 for various BIE formulations of Neumann
elastic scattering problems at high frequencies in the case when is the kite contour. The material parameters are
= 2 and µ = 1, and the incidence was an P-wave. The discretizations used in these numerical experiments delivered
results accurate at the level of 10 4 . Global equispaced meshes were used for QBX discretizations with expansion
parameters p = 12 and = 6.
! N # iter CFIE opt / # Iter K-M / # Iter one sided QBX (3.32)
CFIER RN = (PS (Y+ )) 1
one sided QBX (3.20)
10 128 51/29 187/260 238
20 256 72/44 338/398 405
40 512 121/71 620/737 783
80 1024 195/123 1209/1332 1558
160 2048 285/156 2085/2501 3459

Table 16
Numbers of GMRES iterations of various formulations to reach residuals of 10 5 for various BIE formulations of Neumann
elastic scattering problems at high frequencies in the case when is the starfish contour. The material parameters are
= 2 and µ = 1, and the incidence was an P-wave. The discretizations used in these numerical experiments delivered
results accurate at the level of 10 4 . Global equispaced meshes were used for QBX discretizations with expansion
parameters p = 12 and = 6.
! N # iter CFIE opt/ # Iter K-M/ # Iter one sided QBX (3.32)
CFIER RN = (PS (Y+ )) 1
one sided QBX (3.20)
10 128 41/22 167/183 179
20 256 52/31 220/237 279
40 512 81/46 330/445 534
80 1024 161/84 627/818 1061
160 2048 348/167 1187/1419 2245

Table 17
Numbers of GMRES iterations of various formulations to reach residuals of 10 5 for various BIE formulations of Neumann
elastic scattering problems at high frequencies in the case when is a teardrop. The material parameters are = 2
and µ = 1, and the incidence was an P-wave. The discretizations used in these numerical experiments delivered results
4
accurate at the level of 10 . Global Chebyshev meshes were used for QBX discretizations with expansion parameters
p = 12 and = 6.
! N # iter CFIE opt / # Iter one sided # Iter one sided
DCFIER (3.13) QBX (3.20) QBX (3.32)
RN = (PS (Y+ )) 1
10 128 49/30 157 253
20 256 75/43 279 446
40 512 126/69 478 796
80 1024 211/116 910 1558
160 2048 369/210 1876 1708

Table 18
Numbers of GMRES iterations of various formulations to reach residuals of 10 5 for various BIE formulations of Neumann
elastic scattering problems at high frequencies in the case when is a boomerang. The material parameters are = 2
and µ = 1, and the incidence was a P-wave. The discretizations used in these numerical experiments delivered results
accurate at the level of 10 4 . Global Chebyshev meshes were used for QBX discretizations with expansion parameters
p = 12 and = 6.
! N # iter iter CFIE opt/ # Iter one sided # Iter one sided
DCFIER (3.13) QBX (3.20) QBX (3.32)
RN = (PS (Y+ )) 1
10 128 72/36 216 252
20 256 97/50 319 435
40 512 137/77 650 798
80 1024 229/135 1191 1479
160 2048 394/242 2267 2795

end points of the (open) arc [3,34], a salient feature that must be incorporated in the Nyström discretization of the
formulations (3.12). To do so, we assume the arc is smooth and follow the prescriptions in [3], that is, we use Chebyshev
meshes on whereby the density function is parametrized on the interval [0, ⇡], and we take into account the fact that
the density must vanish at the end points to extend it to the interval [0, 2⇡] via odd extension. Just like in [3], we use
the fact the operator W W 0 is weakly singular as the basis of our Nyström discretization of the hypersingular operator
W , and we project the discrete Nyström equations back onto the interval (0, ⇡ ). We present results for the first kind
formulation involving the hypersingular operator W (according to Eq. (3.12)), as well as for the preconditioned formulation
22
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

Table 19
Numbers of GMRES iterations of various formulations to reach residuals of
10 5 for various BIE formulations of Neumann elastic scattering problems
at high frequencies in the case when is a flat strip of length 2. The
material parameters are = 2 and µ = 1, and the incidence was an
P-wave. The discretizations used in these numerical experiments delivered
results accurate at the level of 10 3 .
1
! n # iter W (3.12) # iter (PS (Y+ )) W
10 128 29 16
20 256 49 22
40 512 103 38
80 1024 225 59
160 2048 490 76

(PS (Y+ )) 1 W . The discretization of the Fourier multiplier (PS (Y+ )) 1 is performed using trigonometric interpolation for
even functions, that is, only cosines functions are used. The extension of the Helmholtz decomposition approach to the
case of open arcs is currently underway.

5.3. Time domain simulations

We will show in this part how the numerical schemes introduced in this work can be applied to solve the transient
elastic wave equation
@t2 u(x; t) = div (u(x; t)) in ⌦+ ⇥ (0, 1) (5.6)

with either Dirichlet

u(x; 0) = 0, u(x; t) = f(x; t), x2 , t>0 (5.7a)
or Neumann conditions
u(x; 0) = 0, T u(x; t) = g(x; t), x2 , t>0 (5.7b)
by Convolution Quadrature (CQ) Methods. CQ constitutes a powerful framework for the numerical solutions of elastic
wave equation [42]. In particular, Runge–Kutta based CQ methods (RKCQ) can deliver higher order in time solvers for
the wave equation [11] than those based on multistep, as BDF, methods which are limited to order 2. We will present
next the details of the implementation of RKCQ methods. The case of linear multi-step solvers (e.g. BDF2) was discussed
in [42].

5.3.1. Runge Kutta convolution quadrature methods for transient elastic wave equation
We will follow closely the exposition in [43] and we restrict ourselves to the Dirichlet case (5.7a), since the Neumann
problem cf. (5.7b) is completely analogous. First, problem (5.6)–(5.7a) is rewritten as a first order system
8
> @ Y (x; t)
>
< = LY (x; t), (x, t) 2 (R2 \ ⌦ ) ⇥ (0, 1)
@t
> BY (x; t) = F (x; t), (x; t) 2 ⇥ (0, 1)
>
:Y (x; 0) = 0, x 2 ⌦+
where we have introduced the following notations
" # 
u(x; t)
0 I
Y (x; t) = @ u(x; t) L=
div 0
 @t 
I 0 g(x; t)
B= F (x; t) = .
0 0 0
Given a final time T , the CQ methods deliver a sequence of N + 1 approximations of the exact solution in [0, T ]
[u t (x; tn )]0nN , tn = n t .
Here N is a positive integer and t = NT . We apply an m stage Runge–Kutta (RK) scheme to the solution of the first order
system reformulation of the elastic wave equation
8 m
> X
>
> Vi (x; tn ) = Y t (x; tn ) + t aij LVj (x; tn ) i 2 1, . . . , m,
>
<
j=1
m (5.8)
>
> X
>
: Y t (x; tn+1 )
> = Y t (x; tn ) + t bj LVj (x; tn )
j=1

23
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

where the ensemble of matrices and vectors

A = [aij ]1i,jm b = [bj ]1jm c = [cj ]1jm ,

make up the Butcher Tableau associated with the RK scheme. The A-stability requirement on the underlying ODE solver
of CQ methods motivates the choice of stiffly accurate or L stable RK schemes (5.8). In these schemes,
1
⇥ ⇤
b> A = 0 0 ··· 0 1 , (5.9)

which trivially implies Y t (x; tn+1 ) = Vm (x; tn ). Hence we can focus on the computation of the internal stages of the RK
method.
Applying the ⇣ transform to the Eqs. (5.8), and taking into account (5.9), we obtain
8 m
> X
>
> V (x; ⇣ ) = Y t (x; ⇣ ) + t aij LVj (x; ⇣ ) i 2 1, . . . , m
>
< i
j=1
m (5.10)
>
> Y t (x;⇣ )
X
>
> = Y t (x; ⇣ ) + t bj LVj (x; ⇣ ).
: ⇣
j=1

where
X X
Vi (x; ⇣ ) := Vi (x; tn )⇣ n , Yd (x; ⇣ ) := Y t (x; tn )⇣ n .
n 0 n 0

We note that the second equation in (5.10) leads to

m
⇣ X
Y t (x; ⇣ ) = t bj LVj (x; ⇣ ). (5.11)
1 ⇣
j=1

Writing the vector quantities Vj in explicit form as [Rj (x; ⇣ ) Sj (x; ⇣ )]> and plugging it in the first equation in (5.10),
yields to the following system of equations
 m ✓
X ◆ 
Ri (x; ⇣ ) ⇣ Rj (x; ⇣ )
= t bj + aij L , i = 1, . . . , m. (5.12)
Si (x; ⇣ ) 1 ⇣ Sj (x; ⇣ )
j=1

Gathering the quantities above in vector form

R(x; ⇣ ) := [Ri (x; ⇣ )]i=1,...,m , S(x; ⇣ ) := [Si (x; ⇣ )]i=1,...,m ,

we have from the definition of the matrix differential operator L

✓ ◆ ✓ ◆
⇣ ⇣
R(x; ⇣ ) = t 1b> + A ⌦ S(x; ⇣ ), S(x; ⇣ ) = t 1b> + A ⌦ div (R(x; ⇣ ))
1 ⇣ 1 ⇣
(the differential operator div is obviously applied element-wise) where
> m
1 = [1, . . . , 1] 2 R

and A ⌦ B denoting the Kronecker product:

2 3
a11 B ··· a1m B
6 . .. .. 7 .
A ⌦ B = 4 .. . . 5
am1 B ··· amm B
Therefore, we can recast the linear system (5.12) in the equivalent form
✓ ◆2
(⇣ )
⌦R(x; ⇣ ) = div (R(x; ⇣ )) (5.13)
t
where (⇣ ) is a matrix operator defined by
✓ ◆ 1
⇣ >
(⇣ ) = A+ 1b .
1 ⇣
A natural idea is to attempt to decouple the linear system (5.13) via diagonalization. To this end, we begin by diagonalizing
the operator (⇣ ). Let P(⇣ ) = [Pij (⇣ )]1i,jm be the matrix consisting of the eigenvectors of (⇣ ) and define

D(⇣ ) = diag( 1( ⇣ ), . . . , m( ⇣ ))
24
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

the diagonal matrix with the corresponding eigenvalues of (⇣ ). Thus

1
(⇣ ) = P(⇣ )D(⇣ )P(⇣ ) .
Clearly, the system (5.13) is decoupled into m frequency domain Navier equations
✓ ◆2
j(⇣)
Wj (x; ⇣ ) = div (Wj (x; ⇣ )), 1  j  m (5.14)
t
where
m
X
1
Wj (x; ⇣ ) = (P (⇣ ))j` (x; ⇣ )R` (x; ⇣ ), or equivalently R(x; ⇣ ) = P(⇣ ) ⌦ W (x; ⇣ ).
`=1

Finally, boundary conditions ought to be provided for the frequency domain Navier Eqs. (5.14). For the intermediate RK
stages Vj these take on the form
2 3
g(x; t + c1 t)

g(x; tn + cj t) 6 g(x; t + c2 t) 7
BVj = F (x; tn + cj t) = , g(x; t + c t) = 6 .. 7.
0 4
.
5
g(x; t + cm t)
Applying the ⇣ transform to the boundary conditions above we get
X
R(x; ⇣ ) = G(x; ⇣ )= g(x; tn + c t)⇣ n (5.15)
n 0

and thus we derive the corresponding boundary conditions for the Navier solutions Wj :
3 2
w1 (x; ⇣ )
6 w2 (x; ⇣ ) 7 1
W (x; ⇣ ) = w(x; ⇣ ) = 6
4 .. 7 := P
5 (⇣ ) ⌦ G(x; ⇣ ). (5.16)
.
wm (x; ⇣ )
Consequently, the following modified Navier equations must be solved in the Laplace domain
( ⇣ ⌘2
j (⇣ )
div (Wj ( · ; ⇣ )) Wj ( · ; ⇣ ) = 0 in ⌦+
t (5.17)
Wj ( · ; ⇣ ) = wj ( · ; ⇣ ).
Once the Navier equations above are solved, the ⇣ -transform approximation of the solution of the wave equation is
retrieved via the formula (recall that Y t (x; tn+1 ) = Vm (x; tn ))
m
X
u t (x; ⇣ ) = ⇣ Rm (x; ⇣ ) = ⇣ Pmj (⇣ )Wj (x; ⇣ ). (5.18)
j=1

Finally, reverting to the physical domain from the Laplace domain is performed in the same manner as in the case of
linear multistep methods:
N
R n X
u t (x; n t) := u t (x; R` ⇣N` +1 )⇣N`n+1 ⇡ u(x, n t), ⇣N +1 = exp 2⇡ i
N
.
N +1
`=0
1
Here 0 < R < 1 with R = " 2N +2 with " being the unit round-off as suggested optimal value.

5.3.2. Numerical results for Convolution Quadrature methods

In our experiments the boundary conditions (5.7b) on ⇥ (0, 1) correspond to the incident field
uinc (x; t) = H(cL t x · d) sin(cL t x · d)d , cL
where d = (1, 0), = 1, µ = 1 and H is a smoothed version of the Heaviside function. We present numerical results
for two smooth scatterers, the starfish (5.2) and the cavity-like (5.3) geometry. both Dirichlet and Neumann boundary
conditions, and we present numerical results in the near field (the observation points are placed equispaced on a circle
situated at distance 1 from the scatterers) for a final time T = 3. The numerical approximations u t (· ; tn ) of time domain
simulations are produced at the time grid tn = n t , 0  n  N such that T = N t = 3 with BDF2 (multistep) or
RK3/RK5 (multistage) solvers. The ensemble of Laplace domain Navier Eqs. (5.17) are solved using the single layer potential
formulation (that is, we use the BIO V ) in the case of Dirichlet boundary conditions (e.g. Table 20) and the double layer
25
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

Table 20
CQ simulations of the solution of equations (5.6) with Dirichlet boundary conditions (5.7a) using Nyström discretizations
based on Alpert 10th order quadratures for the single layer BIE formulations that use the BIO V of the ensemble of
Laplace domain Navier equations (5.17).
N Starfish Cavity
BDF2 "1 e.o.c. RK3 "1 e.o.c. BDF2 "1 e.o.c. RK3 "1 e.o.c.
2 3 2 3
32 5.5 ⇥ 10 1.1 ⇥ 10 2.4 ⇥ 10 1.1 ⇥ 10
2 4 3 4
64 1.8 ⇥ 10 1.57 1.4 ⇥ 10 2.97 7.5 ⇥ 10 1.73 1.3 ⇥ 10 3.01
3 5 3 5
128 5.0 ⇥ 10 1.90 1.7 ⇥ 10 3.00 1.9 ⇥ 10 1.98 1.7 ⇥ 10 3.00
3 6 4 6
256 1.3 ⇥ 10 1.94 2.1 ⇥ 10 3.00 4.9 ⇥ 10 1.93 2.2 ⇥ 10 3.00
4 7 4 7
512 3.2 ⇥ 10 2.02 2.7 ⇥ 10 3.01 1.3 ⇥ 10 1.92 2.8 ⇥ 10 2.97
5 8 5 7
1024 8.0 ⇥ 10 1.99 3.3 ⇥ 10 3.01 4.1 ⇥ 10 1.65 2.7 ⇥ 10 3.01
5 8 5 7
2048 2.0 ⇥ 10 2.00 2.3 ⇥ 10 1.2 ⇥ 10 1.77 2.9 ⇥ 10

Table 21
CQ simulations of the solution of equations (5.6) with Neumann boundary conditions (5.7b) using Nyström discretiza-
tions based on Alpert 10th order quadratures for the double layer BIE formulations that use the BIO W of the ensemble
of Laplace domain Navier equations (5.17).
N Starfish Cavity
BDF2 "1 e.o.c. RK3 "1 e.o.c. BDF2 "1 e.o.c. RK3 "1 e.o.c.
2 3 3 4
32 1.9 ⇥ 10 1.4 ⇥ 10 2.6 ⇥ 10 3.9 ⇥ 10
3 4 3 5
64 6.5 ⇥ 10 1.58 1.7 ⇥ 10 3.01 1.1 ⇥ 10 1.24 4.8 ⇥ 10 3.01
3 4 4 6
128 1.7 ⇥ 10 1.93 2.2 ⇥ 10 2.86 3.6 ⇥ 10 1.61 5.9 ⇥ 10 3.03
4 5 4 7
256 4.0 ⇥ 10 2.06 2.1 ⇥ 10 3.07 1.0 ⇥ 10 1.81 8.9 ⇥ 10 2.73
5 6 5 7
512 9.8 ⇥ 10 2.03 2.4 ⇥ 10 3.06 2.7 ⇥ 10 1.92 2.6 ⇥ 10 1.77
5 7 6 7
1024 2.4 ⇥ 10 2.01 2.9 ⇥ 10 3.08 7.0 ⇥ 10 1.93 2.6 ⇥ 10
6 7 6 7
2048 6.1 ⇥ 10 2.00 3.2 ⇥ 10 1.9 ⇥ 10 1.87 2.5 ⇥ 10

formulation (that is, we use the BIO W ) in the case of Neumann boundary conditions (e.g. Table 21). These BIE of the first
kind are the ones that are most frequently used in the CQ literature [42], and therefore we chose to present numerical
results based on those in order to illustrate the levels of accuracy than can be achieved by our Nyström discretizations.
The discretization of these first kind BIE, in turn, is effected through Nyström discretizations based on Alpert quadratures
of order 10 (i.e. a = 6 and m = 10) using 2n = 512 discretization points on the boundaries for each of the Laplace
domain frequencies that feature in Eqs. (5.17). We present in Tables 20 and 21 the orders of convergence in time (under
the headings estimated order of convergence ‘‘e.o.c’’.) achieved by the CQ BDF2 and CQ RK3 when the errors "1 are
computed in the near field at final time T = 3 with respect to reference solutions (in the near field) obtained using
CQ RK5 for N = 2048 time steps. We note that the CQ errors in this case appear to saturate at the level of 10 7 /10 8
when RK3 solvers are used, which is the highest level of accuracy the CQ methods can actually achieve [10]. We end
this section with a comment on the iterative behavior of BIE formulations of the Laplace domain Navier problems (5.17).
Based on our experience, it is the BIE formulations of the second kind that perform best in this regard: in the case of the
starfish/cavity geometry, the double layer formulation in the case of Dirichlet boundary conditions and the single layer
formulation in the case of Neumann boundary conditions require at most 29/32 iterations to reach GMRES residuals of
10 7 for all the frequencies corresponding to all values of time steps N considered, with similar levels of accuracy to those
reported in Tables 20 and 21. In contrast, both formulations considered in Tables 20 and 21, being first kind formulations,
require larger numbers of iterations for convergence—up to one order of magnitude more, as the number of frequencies
is increased, a feature that is, shared by CFIER formulations as well. Qualitatively similar results are obtained when the
Helmholtz decomposition integral formulations (3.22) and (3.20) are used to solve the ensemble of CQ Laplace domain
problems.

6. Conclusions

We presented two high-order Nyström methods for the discretization of the four BIOs associated with time-harmonic
Navier equations in two dimensions for both smooth as well as Lipschitz boundaries. These discretizations were
used for the solution of elastic scattering problems in frequency and time domain based on BIE formulations. We
presented high-order Nyström discretizations of the Helmholtz BIOs that feature in the Helmholtz decomposition BIE
alternative formulation of elastodynamics scattering problems. Comparisons between the iterative behavior of various
BIE formulations of elastodynamics scattering problems were carried out in the high-frequency regime. Extensions to
three dimensional configurations are currently underway.

Data availability

No data was used for the research described in the article.

26
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

Acknowledgments

Catalin Turc gratefully acknowledges support from NSF through contract DMS-1908602. Víctor Domínguez is partially
supported by projects ‘‘Adquisición de conocimiento y minería de datos, funciones especiales y métodos numéricos
avanzados’’ from Universidad Pública de Navarra, Spain and ‘‘Técnicas innovadoras para la resolución de problemas
evolutivos’’, ref. PID2022-136441NB-I00 from Ministerio de Ciencia e Innovación, Gobierno de España, Spain.

References

[1] R. Kress, Linear Integral Equations, second ed., in: Applied Mathematical Sciences., vol. 82, Springer-Verlag, New York, 1999.
[2] R. Kress, On the numerical solution of a hypersingular integral equation in scattering theory, J. Comput. Appl. Math. 61 (3) (1995) 345–360.
[3] R. Chapko, R. Kress, L. Monch, On the numerical solution of a hypersingular integral equation for elastic scattering from a planar crack, IMA J.
Numer. Anal. 20 (4) (2000) 601–619.
[4] V. Domínguez, C. Turc, Boundary integral equation methods for the solution of scattering and transmission 2D elastodynamic problems, IMA
J. Appl. Math. 87 (4) (2022) 647–706.
[5] B.K. Alpert, Hybrid Gauss-trapezoidal quadrature rules, SIAM J. Sci. Comput. 20 (5) (1999) 1551–1584.
[6] G.C. Hsiao, W.L. Wendland, Boundary Integral Equations, Springer, 2008.
[7] I. Labarca, L.M.. Faria, C. Pérez-Arancibia, Convolution quadrature methods for time-domain scattering from unbounded penetrable interfaces,
Proc. R. Soc. Lond. Ser. A Math. Phys. Eng. Sci. 475 (2227) (2019) 20190029.
[8] P.G. Petropoulos, C. Turc, E. Wind-andersen, Nyström methods for high-order CQ solutions of the wave equation in two dimensions, 2021,
arXiv preprint arXiv:2111.06829.
[9] L. Banjai, S. Sauter, Rapid solution of the wave equation in unbounded domains, SIAM J. Numer. Anal. 47 (1) (2009) 227–249.
[10] L. Banjai, Multistep and multistage convolution quadrature for the wave equation: algorithms and experiments, SIAM J. Sci. Comput. 32 (5)
(2010) 2964–2994.
[11] L. Banjai, C. Lubich, J.M. Melenk, Runge–Kutta convolution quadrature for operators arising in wave propagation, Numer. Math. 119 (1) (2011)
1–20.
[12] J. Lai, P. Li, A framework for simulation of multiple elastic scattering in two dimensions, SIAM J. Sci. Comput. 41 (5) (2019) A3276–A3299.
[13] H. Dong, J. Lai, P. Li, A highly accurate boundary integral method for the elastic obstacle scattering problem, Math. Comp. (2021).
[14] P. Kolm, S. Jiang, V. Rokhlin, Quadruple and octuple layer potentials in two dimensions i: Analytical apparatus, Appl. Comput. Harmon. Anal.
14 (1) (2003) 47–74.
[15] Y. Sawano, Theory of Reproducing Kernels and Applications, Springer, 2017.
[16] O.A. Arqub, Numerical solutions for the robin time-fractional partial differential equations of heat and fluid flows based on the reproducing
kernel algorithm, Internat. J. Numer. Methods Heat Fluid Flow 28 (2018) 828–856.
[17] O.A. Arqub, Numerical simulation of time-fractional partial differential equations arising in fluid flows via reproducing kernel method, Internat.
J. Numer. Methods Heat Fluid Flow 30 (2019) 4711–4733.
[18] O.A. Arqub, M. Al-Smadi, Numerical solutions of Riesz fractional diffusion and advection–dispersion equations in porous media using iterative
reproducing kernel algorithm, J. Porous Media 23 (8) (2020) 783–804.
[19] O.A. Arqub, S. Nabil, Application of reproducing kernel algorithm for solving dirichlet time-fractional diffusion-gordon types equations in porous
media, J. Porous Media 22 (4) (2019) 411–434.
[20] C.L. Epstein, L. Greengard, A. Klöckner, On the convergence of local expansions of layer potentials, SIAM J. Numer. Anal. 51 (5) (2013) 2660–2679.
[21] A. Klöckner, A. Barnett, M. Greengard, Land O’Neil, Quadrature by expansion: A new method for the evaluation of layer potentials, J. Comput.
Phys. 252 (2013) 332–349.
[22] S. Chaillat, M. Darbas, F. Le Louër, Approximate local Dirichlet-to-Neumann map for three-dimensional time-harmonic elastic waves, Comput.
Methods Appl. Mech. Engrg. 297 (2015) 62–83.
[23] S. Chaillat, M. Darbas, F. Le Louër, Analytical preconditioners for Neumann elastodynamic boundary element methods, Partial Differ. Equ. Appl.
2 (2) (2021) 26, Paper (22).
[24] H. Ammari, H. Kang, H. Lee, Layer Potential Techniques in Spectral Analysis, in: Mathematical Surveys and Monographs, vol. 153, American
Mathematical Society, Providence, RI, 2009.
[25] V.D. Kupradze, T.G. Gegelia, M.O. Basheleı̆shvili, T.V. Burchuladze, Three-dimensional problems of the mathematical theory of elasticity and
thermoelasticity, in: V.D. Kupradze (Ed.), Russian ed., in: North-Holland Series in Applied Mathematics and Mechanics, vol. 25, North-Holland
Publishing Co. Amsterdam-New York, 1979.
[26] J. Saranen, G. Vainikko, Periodic integral and pseudodifferential equations with numerical approximation, in: Springer Monographs in
Mathematics, Springer-Verlag, Berlin, 2002.
[27] Y. Boubendir, C. Turc, Wave-number estimates for regularized combined field boundary integral operators in acoustic scattering problems with
Neumann boundary conditions, IMA J. Numer. Anal. 33 (4) (2013) 1176–1225.
[28] V. Domínguez, M. Lyon, C. Turc, Well-posed boundary integral equation formulations and nyström discretizations for the solution of Helmholtz
transmission problems in two-dimensional lipschitz domains, J. Integral Equations Appl. 28 (3) (2016) 395–440.
[29] H. Brakhage, P. Werner, Über das Dirichletsche A.ussenraumproblem für die Helmholtzsche schwingungsgleichung, Arch. Math. 16 (1965)
325–329.
[30] S. Chaillat, Marc Bonnet, Jean-François Semblat, A fast multipole accelerated bem for 3-D elastic wave computation, Eur. J. Comput. Mech.
/Revue Européenne de Mécanique Numérique 17 (5–7) (2008) 701–712.
[31] S. Chaillat, M. Darbas, F. Le Louër, Fast iterative boundary element methods for high-frequency scattering problems in 3d elastodynamics, J.
Comput. Phys. 341 (2017) 429–446.
[32] A.J. Burton, G.F. Miller, The application of integral equation methods to the numerical solution of some exterior boundary-value problems, Proc.
Roy. Soc. London. Ser. A 323 (1971) 201–210, A discussion on numerical analysis of partial differential equations (1970).
[33] A. Anand, J.S. Ovall, C. Turc, Well-conditioned boundary integral equations for two-dimensional sound-hard scattering problems in domains
with corners, J. Integral Equations Appl. 24 (3) (2012) 321–358.
[34] O.P. Bruno, L. Xu, T. Yin, Weighted integral solvers for elastic scattering by open arcs in two dimensions, Internat. J. Numer. Methods Engrg.
122 (11) (2021) 2733–2750.
[35] O.P. Bruno, T. Yin, Regularized integral equation methods for elastic scattering problems in three dimensions, J. Comput. Phys. (2020) 109350.
[36] R. Kussmaul, Ein numerisches Verfahren zur Lösung des Neumannschen A.ussenraumproblems für die Helmholtzsche Schwingungsgleichung,
Comput. (Arch. Elektron. Rechnen) 4 (1969) 246–273.

27
V. Domínguez and C. Turc Journal of Computational and Applied Mathematics 440 (2024) 115622

[37] E. Martensen, Über eine Methode zum räumlichen Neumannschen Problem mit einer A.nwendung für torusartige Berandungen, Acta Math.
109 (1963) 75–135.
[38] S. Hao, A.H. Barnett, P.-G. Martinsson, P. Young, High-order accurate methods for Nyström discretization of integral equations on smooth curves
in the plane, Adv. Comput. Math. 40 (1) (2014) 245–272.
[39] R. Kress, A Nyström method for boundary integral equations in domains with corners, Numer. Math. 58 (2) (1990) 145–161.
[40] Y. Saad, M.H. Schultz, GMRES: a generalized minimal residual algorithm for solving nonsymmetric linear systems, SIAM J. Sci. Stat. Comput. 7
(3) (1986) 856–869.
[41] D. Colton, R. Kress, Inverse Acoustic and Electromagnetic Scattering Theory, second ed., in: Applied Mathematical Sciences, vol. 93,
Springer-Verlag, Berlin, 1998.
[42] V. Domínguez, T. Sánchez-Vizuet, F.-J. Sayas, A fully discrete calderón calculus for the two-dimensional elastic wave equation, Comput. Math.
Appl. 69 (7) (2015) 620–635.
[43] T. Betcke, N. Salles, Wojciech Smigaj, Overresolving in the Laplace domain for convolution quadrature methods, SIAM J. Sci. Comput. 39 (1)
(2017) A188–A213.

28