hp-Finite Elements for Highly Indefinite Helmholtz
Problems
Lecture Notes of the Zürich Summerschool 2016.
S.A. Sauter∗
August 21, 2016
Abstract
These lecture notes comprise the talks of the author given at the Zürich Summerschool 2016. They are based on the papers [14], [17], [5].
1
Model Helmholtz Problems and their Discretization
1.1
Model Problems
The Helmholtz equation describes wave phenomena in the frequency domain which, e.g., arises
if electromagnetic or acoustic waves are scattered from or emitted by bounded physical objects.
In this light, the computational domain for such wave problems, typically, is the unbounded
complement of a bounded domain Ωin ⊂ Rd , d = 1, 2, 3, i.e., Ωout := Rd \Ωin . We assume that
Ωin has a Lipschitz boundary Γin := ∂Ωin . The classical Helmholtz problem depends on the
constant wavenumber k > 0.
1.1.1
Some Function Spaces
In this section, we will introduce some spaces and sets of functions.
Let Ω ⊂ Rd be a bounded Lipschitz domain. For s ≥ 0, 1 ≤ p ≤ ∞, let W s,p (Ω) denote
the classical (complex-valued) Sobolev spaces. As usual we write Lp (Ω) short for W 0,p (Ω)
and H s (Ω) for W s,2 (Ω). The scalar product and norm in L2 (Ω) and L2 (Γ) are denoted by
(u, v) := Ω uv̄ and
(u, v)Γ := Γ uv̄ and
u := (u, u)1/2 in L2 (Ω) ,
1/2
u Γ := (u, u)Γ in L2 (Γ) .
For parameters −∞ < a < b < ∞, we introduce the subset
L∞ (Ω, [a, b]) :=
w ∈ L∞ (Ω, R) : a ≤ ess inf w (x) ≤ ess supw (x) ≤ b .
x∈Ω
x∈Ω
∗
(stas@math.uzh.ch), Institut für Mathematik, Universität Zürich, Winterthurerstr 190, CH-8057 Zürich,
Switzerland
1
Let the boundary Γ = ∂Ω be partitioned into two disjoint, relatively open subsets ΓD , ΓN
such that the decomposition
Γ = ΓD ∪ ΓN ∪ Π with Π := ΓD ∩ ΓN
(1.1a)
forms a Lipschitz dissection of Γ (cf. [11]). For simplicity, we assume for the following that
both, ΓD and ΓN , are closed Lipschitz manifolds so that (see Fig. 1)
ΓD ∩ ΓN = ∅.
(1.1b)
For functions u ∈ C 0 Ω , the restriction to the Dirichlet boundary ΓD is well defined and
denoted by (γD,0 u) (x) := u (x) for all x ∈ ΓD . This mapping can be extended continuously to
γD,0 : H 1 (Ω) → L2 (ΓD ) and allows to define H 1/2 (ΓD ) := γD,0 (H 1 (Ω)) as the range of γD,0 .
We set
H := u ∈ H 1 (Ω) | γD,0 u = 0 .
(1.2)
The trace operator γN,0 : H 1 (Ω) → H 1/2 (ΓN ) is defined in an analogous way. The dual
′
′
spaces are denoted by H −1/2 (ΓN ) := H 1/2 (ΓN ) and H −1/2 (ΓD ) := H 1/2 (ΓD ) .
Since the boundary ∂Ω is Lipschitz, it follows by Rademacher’s theorem that an exterior
unit normal vector field n : Γ → Sd−1 exists almost everywhere and allows to define a normal
derivative γN,1 for function in C 1 Ω by (γN,1 u) (x) := n (x) , ∇u (x) for almost every x ∈ ΓN .
This mapping can be extended continuously to a mapping γN,1 : H 1 (Ω) → H −1/2 (ΓN ) — for
details we refer, e.g., to [11], [22].
1
We will also need the Sobolev space Hloc
(Ωc ) for the unbounded exterior Ωc := Rd \Ω. It
contains all functions u : Ωc → C with the property that ϕu ∈ H ℓ (Ωc ) for all
∞
ϕ ∈ Ccomp
(Ω) := u|Ω : u ∈ C0∞ Rd
1.1.2
.
Derivation of the Classical Helmholtz Equation with DtN and Robin Boundary Conditions
1
For a given right-hand side f ∈ L2 (Ωout ), the Helmholtz problem is to seek U ∈ Hloc
(Ωout )
such that
−∆ − k 2 U = f
in Ωout
(1.3a)
is satisfied. Towards infinity, Sommerfeld’s radiation condition is imposed
|∂r U − i kU | = o
x
1−d
2
for x → ∞,
(1.3b)
where ∂r denotes differentiation in radial direction and |·| the Euclidian vector norm. For
simplicity we restrict here to homogeneous Dirichlet boundary condition on Γin
U|Γin = 0.
(1.3c)
We assume that f is local in the sense that there exists some bounded, simply connected
Lipschitz domain1 Ω⋆ such that a) Ωin ⊂ Ω⋆ and b) supp(f) ⊂ Ω⋆ . The computational domain
(cf. Figure 1) will be
Ω := Ω⋆ \Ωin
(1.4)
2
G
W
in
G
W
o u t
in
W
o u t
W
*
Figure 1: Scatterer Ωin with boundary Γin and exterior domain Ωout . The support of f is
assumed to be contained in the bounded region Ω⋆ . The domain for the weak variational
formulation is Ω = Ω⋆ \Ωin .
and, next, we will derive appropriate boundary conditions at the outer boundary Γout := ∂Ω⋆ .
Problem (1.3) can be reformulated in an equivalent way as a transmission problem by seeking
1
functions u ∈ H 1 (Ω) and uout ∈ Hloc
(Rd \Ω⋆ ) such that
(−∆ − k 2 ) u = f
(−∆ − k 2 ) uout = 0
u=0
out
u=u
and ∂n u = ∂n uout
∂r uout − i kuout = o |x|
1−d
2
in Ω,
in Rd \Ω⋆ ,
on Γin ,
on Γout ,
(1.5)
for |x| → ∞.
Here, n denotes the normal vector pointing into the exterior domain Rd \Ω⋆ and ∂n denotes
differentiation in normal direction.
1
It can be shown that, for given g ∈ H 1/2 (Γout ), the problem: Find w ∈ Hloc
Rd \Ω⋆ such
that
(−∆ − k 2 ) w = 0
in Rd \Ω⋆ ,
w=g
on Γout ,
(1.6)
|∂r w − i kw| = o |x|
1−d
2
for |x| → ∞
has a unique solution. The mapping g → w is called the Steklov—Poincaré operator and
1
denoted by SP : H 1/2 (Γout ) → Hloc
Rd \Ω⋆ . The Dirichlet-to-Neumann (DtN) map is given
1/2
out
−1/2
by Tk := γ1 SP : H (Γ ) → H
(Γout ), where γ1 := ∂n is the normal derivative operator
at Γout . Hence, problem (1.5) can be reformulated as: Find u ∈ H 1 (Ω) such that
(−∆ − k 2 ) u = f
u=0
∂n u = Tk u
1
in Ω,
on Γin ,
on Γout .
(1.7)
Since Ωin is bounded, Ω⋆ always can be chosen as a ball. Other choices of Ω⋆ might be preferable in
certain situations.
3
The previous problems are posed in the weak formulation given by: Find u ∈ H (cf. (1.2))
such that
ADtN (u, v) :=
Ω
∇u, ∇v̄ − k 2 uv̄ −
(Tk u) v̄ =
Γout
for all v ∈ H.
fv
Ω
(1.8)
In this section we restrict to the case ΓD := Γin , while, in general, the definition (1.2) corresponds to the Lipschitz disection introduced in (1.1). Since the numerical realization of the
nonlocal DtN map Tk is costly, various approaches exist in the literature to approximate this
operator by a local operator. The most simple one is the use of Robin boundary conditions
leading to
(−∆ − k 2 ) u = f
in Ω,
u=0
on Γin ,
(1.9)
∂n u = i ku
on Γout .
The weak formulation of this equation is given by: Find u ∈ H such that
ARobin (u, v) :=
Ω
1.1.3
∇u, ∇v̄ − k 2 uv̄ −
i kuv̄ =
Γout
fv
Ω
for all v ∈ H.
(1.10)
Helmholtz-type Equations with Variable Coefficients
In the previous section, we have explained how the classical Helmholtz problem on an unbounded domain can be reformulated and approximated by a Helmholtz problem on a bounded
domain. In the following we will restrict to Helmholtz(-type) problems on bounded domains.
We define the precise problem class which depends on parameters βmax , cmin , cmax , ρmin , ρmax ,
k0 ∈ R satisfying
0 < cmin ≤ cmax < ∞, 0 < ρmin ≤ ρmax < ∞,
(1.11)
k0 > 0,
βmax > 0.
For parameters which satisfy (1.11) and coefficients , k ≥ k0 , ρ ∈ L∞ (Ω, [ρmin , ρmax ]), c ∈
L∞ (Ω, [cmin , cmax ]), β ∈ L∞ (Γ, [0, βmax ]), we introduce the sesquilinear form Aρ,c : H 1 (Ω) ×
H 1 (Ω) → C by
Aρ,c (u, v) :=
1
∇u, ∇v −
ρ
k
c
2
u, v
− i k (βu, v)ΓN .
(1.12)
We consider the Helmholtz problem with variable coefficients: For given F ∈ H′ , find u ∈ H
such that
∀v ∈ H.
(1.13)
Aρ,c (u, v) = F (v)
Remark 1.1 For sufficiently smooth coefficient function ρ, c, β and F (v) = (f, v) + (g, v)ΓN
for some f ∈ L2 (Ω) and g ∈ L2 (Γ) the strong formulation of (1.13) is given by
− div
2
− kc u = f in Ω,
i kβ = g
on ΓN ,
u=0
on ΓD .
1
∇u
ρ
1 ∂u
−
ρ ∂n
4
Definition 1.2 We equip the space H with the norm
u
H
:= u
H,ρ,c
2
∇u
√
ρ
:=
k
u
+
c
2
(1.14)
which is obviously equivalent to the H 1 (Ω)-norm.
Remark 1.3 The analysis of the Helmholtz equation with variable coefficients is a topic of
vivid research and many questions are still open. In our lectures we will concentrate on the
case of the classical Helmholtz equation with ρ = c = 1 and restrict to the case
k0 ≥ 1.
1.2
(1.15)
Abstract Variational Formulation
Since ΓN is a Lipschitz manifold , it is well known that the following trace estimates hold (see
[2, (1.6.6) Theorem]).
Lemma 1.4 Let ΓN be as in (1.1b). There exists a constant Ctr such that
∀u ∈ H 1 (Ω) :
u
H 1/2 (ΓN )
≤ Ctr u
and
∀u ∈ H 1 (Ω) :
u
Corollary 1.5 For u ∈ H 1 (Ω), we have
√
ku
L2 (ΓN )
L2 (ΓN )
≤ Ctr u
≤ Ctr u
H
1/2
u
(1.16a)
H
1/2
H 1 (Ω)
.
(1.16b)
.
Proof. It holds
k u
2
L2 (ΓN )
≤ Ctr2 k u
u
H 1 (Ω)
Ctr2
k 2 u 2 + u 2H 1 (Ω)
2
Ctr2
=
1 + k 2 u 2 + ∇u
2
≤
(1.15)
≤ Ctr2
ku
2
+ ∇u
2
2
(1.17)
.
Both sesquilinear forms ADtN (1.8) and ARobin (1.10) belong to the following class of forms
(see Proposition 1.8).
Assumption 1.6 (Variational formulation) Let Ω ⊂ Rd , for d ∈ {2, 3}, be a bounded
Lipschitz domain. Then H as in (1.2), equipped with the norm · H , is a closed subspace
of H 1 (Ω). We consider a sesquilinear form A : H × H → C that can be decomposed into
A = a − b, where
a (v, w) :=
Ω
∇v, ∇w̄ − k 2 v w̄
and the sesquilinear form b satisfies the following properties:
5
(a) b : H × H → C is a continuous sesquilinear form with
|b(v, w)| ≤ Cb v
H
w
for all v, w ∈ H,
H
for some positive constant Cb .
(b) There exist θ ≥ 0 and γell > 0 such that the following Gårding inequality holds:
Re A (v, v) + θ kv
2
2
H
≥ γell v
(c) The adjoint problem: Find z ∈ H such that
for all v ∈ H.
for all v ∈ H
A (v, z) = (v, f )L2 (Ω)
(1.18)
(1.19)
(1.20)
is uniquely solvable for every f ∈ L2 (Ω) with bounded solution operator Q⋆k : L2 (Ω) → H,
f → z, more precisely, the (k-dependent) constant
Ckadj := k
Q⋆k f
f
sup
f ∈L2 (Ω)\{0}
H
(1.21)
is finite.
Problem 1.7 Let A be a sesquilinear form as in Assumption 1.6. For given f ∈ L2 (Ω), we
seek u ∈ H such that
f v̄
for all v ∈ H.
(1.22)
A (u, v) =
Ω
Proposition 1.8 Both sesquilinear forms ARobin (1.10) and ADtN (1.8) (under the additional
condition that Γout is a sufficiently large sphere) satisfy Assumption 1.6.
Proof. See [19] and [15]. Here, we restrict to the sesquilinear form ARobin . Condition (a) for
ARobin follows from Corollary 1.5.
For condition (b) and Robin boundary conditions, we employ
Re (ARobin (v, v)) + 2 kv
2
= v
2
H
and (1.19) holds with θ = 2 and γell = 1.
For condition (c) we may apply Fredholm’s theory and, hence, it suffices to prove the
implication
(A (u, v) = 0 for all v ∈ H) =⇒ u = 0.
(1.23)
For Robin boundary conditions we consider the imaginary part of the sesquilinear form (1.10)
for the test function v = u
Im (ARobin (u, u)) = − u
2
L2 (ΓN )
to see that (1.23) implies u ∈ H01 (Ω). Hence, u solves
Ω
∇u, ∇v̄ − k 2 uv̄ = 0
for all v ∈ H.
(1.24)
Let Ω⋆⋆ be a bounded domain such that Ω ⊂ Ω⋆⋆ and ΓN ⊂ Ω⋆⋆ , ΓD ⊂ ∂Ω∗∗ . The extension of
u by zero to Ω⋆⋆ is denoted by u0 . It satisfies u ∈ H (Ω⋆⋆ ) := u ∈ H 1 (Ω⋆⋆ ) | u|ΓD = 0 and
Ω
∇u0 , ∇v̄ − k 2 u0 v̄ = 0
for all v ∈ H (Ω⋆⋆ ) .
Elliptic regularity theory implies that u0 ∈ H 2 (Q) for any compact subset Q ⊂ Ω⋆⋆ , in
particular, in an open Ω⋆⋆ neighborhood of ΓN . The unique continuation principle (cf. [10,
Ch. 4.3]) implies that u0 = 0 in Ω⋆⋆ so that u = 0 in Ω.
6
A
R
t
a ffin e ,
h -s c a lin g
t
d is to rtio n ,
h -in d e p e n d e n t
E le m e n t m a p : F t= R
t
A
t
Figure 2: hp-Finite element mesh a pullback to the reference element.
1.3
1.3.1
Discretization
Conforming Galerkin Discretization
A conforming Galerkin discretization of Problem 1.7 is based on the definition of a finite
dimensional subspace S ⊂ H and is given by: Find uS ∈ S such that
for all v ∈ S.
A (uS , v) = (f, v)
1.3.2
(1.25)
hp-Finite Elements
As an example for S as above, we will define hp-finite elements on a finite element mesh T
consisting of simplices with maximal mesh width h and local polynomial degree p. Before
formulating the conditions on the mesh in an abstract way, we give an example of a typical
construction (cf. Fig. 2).
Example 1.9 (Patchwise construction of FE mesh.) Let Ω denote a bounded domain.
(a) We assume that a polyhedral (polygonal in 2D) domain Ω along with a bi-Lipschitz
mapping χ : Ω → Ω is given. Let T macro = Kimacro : 1 ≤ i ≤ q denote a conforming
finite element mesh for Ω consisting of open simplices which are regular in the sense of
[3]. T macro is considered as a coarse partition of Ω, i.e., the diameters of the elements
in T macro are of order 1. We assume that the restrictions χi := χ|K macro are analytic for
i
all 1 ≤ i ≤ q.
(b) The finite element mesh with mesh width h is generated by refining the mesh T macro in
some standard (conforming) way and denoted by T = Ki : 1 ≤ i ≤ N . The corresponding finite element mesh for Ω then is defined by T = K = χ K : K ∈ T .
Note that, for any K = χ K ∈ T , there exists an affine bijection AK : K → K which
d
i=1
maps the reference element K := x ∈ Rd>0 :
xi < 1
to the simplex K. A parametriza-
tion FK : K → K can be chosen by FK := RK ◦ AK , where RK := χ|K is independent of the
mesh width h := max {hK : K ∈ T }, where hK := diam (K).
7
Concerning the polynomial degree distribution, it will be convenient (cf.[18]) to assume
that the polynomial degrees of neighboring elements are comparable: There exists a constant
cp > 0 such that
c−1
p (pK + 1) ≤ pK ′ + 1 ≤ cp (pK + 1)
for all K, K ′ ∈ T with K ∩ K ′ = ∅.
(1.26)
To formulate the smoothness and scaling assumptions on RK and AK in an abstract way we
have to introduce some notation first. For a function v : Ω → R, Ω ⊂ Rd , we write
|∇n v(x)|2 =
α∈Nd0 :|α|=n
n! α
|∂ v(x)|2 .
α!
(1.27)
Assumption 1.10 Each element map FK can be written as FK = RK ◦ AK , where AK is an
affine map and the maps RK and AK satisfy for constants Caffine , Cmetric , γ > 0 independent
of hK :
A′K
L∞ (K) ≤ Caffine hK ,
′ −1
(RK
) L∞ (K) ≤ Cmetric ,
(A′K )−1
∇n RK
L∞ (K)
≤ Caffine h−1
K
L∞ (K)
≤ Cmetric γ n n!
for all n ∈ N0 .
Here, K = AK (K).
Definition 1.11 (hp-finite element space) For meshes T with element maps FK as in
Assumption 1.10 the hp-finite element space of piecewise (mapped) polynomials is given by
S p,1 (T ) := {v ∈ H : v|K ◦ FK ∈ Pp for all K ∈ T },
(1.28)
where Pp denotes the space of polynomials of degree p. For chosen T and p, we may let
S = S p,1 (T ).
2
A Priori Analysis
This section is devoted to existence, uniqueness, stability, and regularity for the Helmholtz
problem (1.7).
2.1
Well-posedness
′
Proposition 2.1 Let Ω be a bounded Lipschitz domain. For all f ∈ (H 1 (Ω)) , a unique
solution u of problem (1.10) exists and depends continuously on the data.
Proof. The proof follows from Fredholm’s alternative and it suffices to prove uniqueness of
Problem (1.7). However, the same arguments as in the proof of Proposition 1.8 for the adjoint
problem can be applied to prove uniqueness for the forward problem.
8
2.2
Discrete Stability and Quasi-Optimality
An essential role for the stability and convergence of the Galerkin discretization is played by
the adjoint approximability which has been introduced in [17]; see also [21], [1].
Definition 2.2 (Adjoint approximability) For a finite dimensional subspace S ⊂ H, we
define the adjoint approximability of Problem 1.7 by
ηk⋆ (S)
inf v∈S Q⋆k f − v
:= k
sup
f
f ∈L2 (Ω)\{0}
H
,
(2.1)
where Q⋆k is as in (1.21).
Theorem 2.3 (Stability and convergence) Let γell , θ, Cb , Ckadj be as in Assumption 1.6
and S as in Section 1.3.1. Then the condition
γell
ηk⋆ (S) ≤
,
(2.2)
2θ (1 + Cb )
implies the following statements:
(a) The discrete inf-sup condition is satisfied:
inf
sup
v∈S\{0} w∈S\{0}
|A (v, w)|
γell
≥
> 0.
v H;Ω w H;Ω
2 + γell /(1 + Cb ) + 2θCkadj
(2.3)
(b) Let S satisfy (2.2). Then, the Galerkin method based on S is quasi-optimal, i.e., for
every u ∈ H there exists a unique uS ∈ S with A (uS , v) = (f, v) for all v ∈ S, and there
holds
2
(1 + Cb ) inf u − v H ,
v∈S
γell
2
k (u − uS ) ≤
(1 + Cb )2 ηk⋆ (S) inf u − v
v∈S
γell
u − uS
H
≤
(2.4)
H.
(2.5)
Proof. Let u ∈ S and set z := θk 2 Q⋆k u. Then,
A (u, u + z) = A (u, u) + θ kv
2
= A (u, u) + θ kv
2
+ A (u, z) − θ kv
2
.
This directly implies (cf. (1.19))
Re A (u, u + z) = γell u
2
H
.
Let zS ∈ S denote the best approximation of z with respect to the ·
|A (u, v)| ≤ Cc u
v
H
H
H -norm.
Note that
∀u, v ∈ H with Cc := 1 + Cb .
Then,
Re A (u, u + zS ) ≥ Re A (u, u + z) − |A (u, z − zS )| ≥ γell u
≥ u
H
(γell u
∗
H
2
H
− Cc u
∗
H
z − zS
− θkCc η (S) u ) ≥ (γell − θCc η (S)) u
9
2
H
.
H
The stability of the continuous adjoint problem (cf. (1.21)) implies
u + zS
H
≤ u
H
+ z − zS
H
+ z
≤ 1 + θη ∗ (S) + θCkadj
so that
Re A (u, u + zS ) ≥
H
u
≤ u
H
+ θkη ∗ (S) u + θkCkadj u
H
γell − θCc η ∗ (S)
u
1 + θη ∗ (S) + θCkadj
H
u + zS
H
.
Therefore, in view of the assumption (2.2), we have proved
inf sup
u∈S v∈S\{0}
|A (u, v)|
γell
≥
.
u H v H
2 + γell / (1 + Cb ) + 2θCkadj
The convergence of the finite element discretization is proved by applying the theory as
developed in [21] (see also [23, 15, 1], [2, Sec.5.7]).
In the first step, we will estimate the L2 -error by the H 1 -error and employ the AubinNitsche technique. The Galerkin error is denoted by e = u − uS . We set ψ := Q⋆k e (cf.
Assumption 1.6(c)) and denote by ψS ∈ S the best approximation of ψ with respect to the
H-norm.
The L2 -error can be estimated by using the Galerkin orthogonality
2
k e
= kA (e, ψ) ≤ kA (e, ψ − ψS ) ≤ Cc k e
≤ Cc η ∗ (S) e H e ,
H
ψ − ψS
H
(2.6)
i.e.,
k e ≤ Cc η ∗ (S) e
H
= Re (A (e, e)) + γell e
2
H
.
(2.7)
To estimate the H-norm of the error we proceed as follows. For any vS ∈ S it holds
γell e
2
H
≤ Re A (e, u − vS ) + θk 2 e
≤ Cc e
H
u − vS
H
− Re A (e, e)
2
+ θCc η∗ (S) e
2
H.
This leads to
(γell − θCc η ∗ (S)) e
H
≤ Cc u − vS
H
.
Noting that (2.2) implies γell − θCc η ∗ (S) ≥ γell /2, we arrive at the final estimate.
2.3
Regularity
In this section, we derive some explicit bounds for the solution operator Q∗k in (1.21) under
the assumption that the right-hand side is in L2 (Ω). To simplify the anylsis we restrict to the
full space case Ω = Rd , d = 1, 2, 3, with Sommerfeld’s radiation condition (1.3b) at infinity
(the case of bounded domains has been analysed in [17]). Under this assumption, the exact
solution of (1.3) can be written as the acoustic volume potential. Let Gk : Rd \ {0} → C
denote the fundamental solution to the operator Lk := −∆ − k 2 , i.e., Gk (z) = gk ( z ), where

ei kr

d = 1,
 − 2ik
(1)
i
gk (r) :=
H (kr) d = 2,
4 0

 ei kr
d = 3.
4πr
10
Then, the solution of (1.3) is given by
∀x ∈ Rd .
Gk (x − y) f (y) dy
U (x) := (Qk f ) (x) :=
Ω
(2.8)
The key ingredient of the analysis of the hp-FEM in Section 2.4 is the following decomposition result.
Lemma 2.4 (decomposition lemma) Let Ω be contained in a ball of radius R > 0. Then
there exists a constant C > 0 depending only on R and k0 such that for f ∈ L2 (Ω) the function
v given by
v(x) = Qk f(x) =
Ω
Gk (x − y)f (y) dy,
x ∈ Ω,
satisfies
k −1 v
H 2 (Ω)
+ v
H 1 (Ω)
+k v
≤C f
L2 (Ω)
L2 (Ω) .
Furthermore, for every λ > 1, there exists a λ- and k-dependent splitting v = vH 2 + vA with
∇p vH 2
L2 (Ω)
≤C 1+
∇p vA
L2 (Ω)
≤ Cλ
λ2
√
dλk
1
−1
(λk)p−2 f
p−1
f
∀p ∈ {0, 1, 2},
L2 (Ω)
∀p ∈ N0 .
L2 (Ω)
(2.9a)
(2.9b)
Here, ∇p vA stands for a sum over all derivatives of order p (see (1.27) for details).
Remark 2.5 For f ∈ L2 (Ω) the function v = Qk (f ) cannot be expected to have more Sobolev
regularity than H 2 . The decomposition v = vH 2 + vA of Lemma 2.4 splits v into an H 2 -regular
part vH 2 and an analytic part vA . The essential feature of this splitting is that the H 2 -part
vH 2 has a better H 2 -regularity constant in terms of k than v itself, namely, (2.9a), (2.9b), and
the triangle inequality ∇2 v L2 (Ω) ≤ ∇2 vH 2 L2 (Ω) + ∇2 vA L2 (Ω) imply
∇2 vH 2
L2 (Ω)
≤C f
versus
L2 (Ω)
∇2 v
L2 (Ω)
≤ Ck f
L2 (Ω) .
The fact that vH 2 H 2 ≤ C f L2 for a C > 0 independent of k is essential for the stability
and convergence analysis.
The decomposition lemma has been generalized to more smooth and less smooth right-hand
sides and domains in [16], [6], [13].
Proof of Lemma 2.4. The estimates for v follow directly from those for vH 2 and vA
by fixing a parameter λ > 1. In order to construct the splitting v = vH 2 + vA , we start by
recalling the definition of the Fourier transform for functions u with compact support
û (ξ) = (2π)−d/2
e− i ξ,x u (x) dx
Rd
∀ξ ∈ Rd
and the inversion formula
u (x) = (2π)−d/2
ei x,ξ û (ξ) dξ
Rd
11
∀x ∈ Rd .
Let BΩ ⊂ Rd be a ball of radius R containing Ω. Extend f by zero outside of Ω and denote
this extended function again by f . Let µ ∈ C ∞ (R≥0 ) be a cutoff function such that
supp µ ⊂ [0, 4R] ,
µ|[0,2R] = 1,
∀x ∈ R≥0 : 0 ≤ µ (x) ≤ 1, µ|[4R,∞[
|µ|W 1,∞ (R≥0 ) ≤
C
,
R
(2.10)
C
= 0, |µ|W 2,∞ (R≥0 ) ≤ 2 .
R
Define M (z) := µ ( z ) and
vµ (x) :=
BΩ
∀x ∈ Rd .
Gk (x − y) M (x − y) f (y) dy
The properties of µ guarantee vµ |BΩ = v|BΩ so that we may restrict our attention to the
function vµ . Since supp f ⊂ BΩ we may write
vµ = (Gk M ) ⋆ f,
(2.11)
where “⋆” denotes the convolution in Rd . We will define a decomposition of vµ (which will
determine the decomposition of v on BΩ ) by decomposing its Fourier transform, i.e.,
vµ = vH 2 + vA .
(2.12)
In order to define the two terms on the right-hand side of (2.12), we let Bλk (0) denote the
ball of radius λk centered at the origin where λ > 1 is the fixed constant (independent of k)
selected in the statement of the lemma. The characteristic function of Bλk (0) is denoted by
χλk . The Fourier transform of f is then decomposed as
f = f χλk + (1 − χλk )f =: fk + fkc .
By the inverse Fourier transformation, this decomposition of f entails a decomposition of f
into fk and fkc given by
ei x,ξ χλk (ξ) fˆ (ξ) dξ
fk (x) := (2π)−d/2
Rd
and fkc (x) := f − fk .
(2.13)
Accordingly, we define the decomposition of vµ by
vµ,H 2 := (Gk M ) ⋆ fkc
and vµ,A := (Gk M ) ⋆ fk .
(2.14)
The functions vH 2 and vA in (2.12) are then obtained by setting vH 2 := vµ,H 2 |Ω and vA :=
vµ,A |Ω . We will obtain the desired estimates by showing the following, stronger estimates:
vµ,H 2
H 2 (Rd )
Dα vµ,A
L2 (Rd )
≤C f
L2 (Rd ) ,
|α|−1
≤ Cλ (λk)
(2.15a)
f
L2 (Rd ) ,
∀α ∈ Nd0 .
(2.15b)
The estimates (2.15) are obtained by Fourier techniques. To that end, we compute the Fourier
transform of Gk M :
Gk M (ξ) = (2π)−d/2
= (2π)−d/2
e− i ξ,x Gk (x) M (x) dx
Rd
∞
gk (r) µ (r) rd−1
e− i r ξ,ζ dSζ
Sd−1
0
= (2π)−d/2 I (ξ) .
12
dr
The inner integral in I (ξ) can be evaluated analytically2 and I (ξ) = ι ( ξ ) with

∞


2
gk (r) µ (r) cos (sr) dr
d = 1,



 0 ∞
2π
gk (r) µ (r) rJ0 (rs) dr
d = 2,
ι (s) =

0∞



sin (rs)

 4π
dr d = 3.
gk (r) µ (r) r2
(rs)
0
(2.17)
Applying the Fourier transform to the convolutions (2.14) leads to
vµ,H 2 = (2π)d/2 Gk M fkc = (2π)d/2 Gk M f(1 − χλk ),
vµ,A = (2π)d/2 Gk M fk = (2π)d/2 Gk M f χλk .
To estimate higher order derivatives of vµ,H 2 and vµ,A we define for a multi-index α ∈ Nd0
the function Pα : Rd → Rd by Pα (ξ) := ξ α and obtain — by using standard properties of the
Fourier transformation and the support properties of χλk — for all |α| ≤ 2
∂ α vµ,H 2
= (2π)d/2 Pα Gk M (1 − χλk ) f
L2 (Rd )
≤ (2π)d/2
max
ξ∈Rd :|ξ|≥λk
|Pα I (ξ)|
≤ (2π)d/2 max s|α| ι (s)
f
s≥λk
(2.18)
L2 (Rd )
(1 − χλk ) f
L2 (Ω)
L2 (Rd )
.
Lemma 2.6, (iv) implies for |α| ∈ {0, 1, 2}
max s|α| ι (s) ≤ C (λk)|α|−2 1 +
s≥λk
λ2
1
−1
.
Thus,
∂ α vH 2
L2 (BΩ )
1
−1
f
L2 (Ω)
max s|α| ι (s)
f
L2 (Ω)
≤ C (λk)|α|−2 1 +
and (2.9a) follows.
Completely analogously, we derive for all α ∈ Nd0
∂ α vµ,A
L2 (Rd )
≤ (2π)d/2
0≤s≤λk
λ2
.
(2.19)
We can complete the proof of the lemma using the bounds on the function ι given in Lemma 2.6,
(v) below and using (1.27).
Lemma 2.6 For the function ι defined in (2.17) the quantity sm ι (s) can be estimated
2
This is trivial for d = 1 and follows for d = 2 from [7, (3.338)(4.)]. For d = 3, we use the formula
e−i
S2
x,x̂
x̂
x̂
Yℓm (x) dx = gℓ ( x̂ ) Yℓm
with gℓ (r) = (−i)ℓ 4πjℓ (r)
(which follows by a comparison of [20, Section 3.2.4, formula (3.2.44) and (3.2.54)]) for m = ℓ = 0, where
Y00 = const and g0 (r) = 4π sin (r) /r.
13
i. for m = 0 by
|ι (s)| ≤ C
R
,
k
ii. for m = 1 by

1 + (Rk)−1



|log kR|
|sι (s)| ≤ CR
1



1
d = 1,
d = 2 and 4Rk ≤ 1,
d = 2 and 4Rk > 1,
d = 3,
iii. and for m = 2 by

1


Rk +


Rk
|log(kR)|
s2 |ι (s)| ≤ C


Rk


1 + kR
d = 1,
d = 2 and 4Rk ≤ 1,
d = 2 and 4Rk > 1,
d = 3.
iv. For fixed R0 , R1 > 0 there exists C > 0 (depending only on R0 , R1 , k0 , d, and the
constant appearing in (2.10)) such that for any R ∈ [R0 , R1 ] and any λ > 1
sup s2 |ι (s)| ≤ C 1 +
|s|≥λk
λ2
1
−1
v. For any λ > 0 and all m ∈ N0 we have
sup |s|m |ι (s)| ≤ CλR (λk)m−1 .
|s|≤λk
The proof of this lemma is quite technical and requires some uniform estimates for some
Bessel and exponential functions. We refer to [19, Theorem 3.7] for the details.
2.4
Approximation Properties
On meshes Th satisfying Assumption 1.10 with element maps FK we consider the space of
piecewise (mapped) polynomials S p,1 (Th ) as in Definition 1.11. It is desirable to construct
an approximant Iu ∈ S p,1 (Th ) of a given (sufficiently smooth) function u in an elementwise
fashion and we employ the following concept:
Definition 2.7 (element-by-element construction) Let K be the reference simplex in
Rd , d ∈ {2, 3}. A polynomial π is said to permit an element-by-element construction of
polynomial degree p for u ∈ H s (K), s > d/2, if:
i. π(V ) = u(V ) for all d + 1 vertices V of K,
ii. for every edge e of K, the restriction π|e ∈ Pp is the unique minimizer of
π → p1/2 u − π
L2 (e)
+ u−π
1/2
H00 (e)
under the constraint that π satisfies the condition in Case i.
14
(2.20)
iii. (for d = 3) for every face f of K, the restriction π|f ∈ Pp is the unique minimizer of
π →p u−π
L2 (f )
+ u−π
(2.21)
H 1 (f )
under the constraint that π satisfies the conditions in Case i and ii for all vertices and
edges of the face f .
Remark 2.8 The conditions of Definition 2.7 are a variation of similar proposals in the
literature, e.g., [4] and [8]. For example, the effective difference between the projection-based
interpolation of [4] and the present construction lies in the choice of the norms employed in
the minimization process in Definition 2.7. Our motivation for formulating the conditions
in Definition 2.7 is that they permit us to construct approximation operators with optimal
simultaneous approximation properties in L2 and H 1 . Previously, the literature had focused
on H 1 -approximation alone.
Lemma 2.9 Let d ∈ {1, 2, 3}, and let K ⊂ Rd be the reference simplex. Let γ, C > 0 be given.
Then there exist constants C, σ > 0 that depend solely on γ and C such that the following is
true: For any function u that satisfies for some Cu , h, R > 0, κ ≥ 1 the conditions
∇n u
L2 (K)
≤ Cu (γh)n max{n/R, κ}n
∀n ∈ N,
n ≥ 2,
(2.22)
and for any polynomial degree p ∈ N that satisfies
h/R + κh/p ≤ C
(2.23)
there holds for m = 1, 2
inf u − π
π∈Pp
W m,∞ (K)
h/R
σ + h/R
≤ CCu
p+1
+
κh
σp
p+1
(2.24)
and additionally admits an element-by-element construction as defined in Definition 2.7.
The approximation of functions with finite regularity is considered in the following theorem.
Theorem 2.10 Let K ⊂ Rd be the reference triangle or the reference tetrahedron. Let s >
d/2. Then there exists C > 0 (depending only on s and d) and for every p a linear operator
π : H s (K) → Pp that permits an element-by-element construction in the sense of Definition 2.7
such that
p u − πu
2.5
L2 (K)
+ u − πu
H 1 (K)
≤ Cp−(s−1) |u|H s (K)
∀p ≥ s − 1.
(2.25)
Convergence Rates
We are now in position to show that the solution v = Nk⋆ f can be approximated well by the
FEM space S p,1 (Th ) provided that kh/p is sufficiently small and p ≥ c ln k.
Theorem 2.11 Let d ∈ {1, 2, 3} and Ω ⊂ Rd be a bounded domain. Then there exist constants
C, σ > 0 that depend solely on the constants appearing in Assumption 1.10 such that for every
f ∈ L2 (Ω) the function v := Nk⋆ f satisfies
inf
w∈S p,1 (Th )
k v−w
H
≤C f
L2 (Ω)
15
kh
1+
p
kh
+k
p
kh
σp
p
.
Proof. We will only prove the cases d ∈ {2, 3}. The case d = 1 follows by similar arguments
where the appeal to Theorem 2.11 and Lemma 2.9 is replaced with that to [24, Thm. 3.17].
We note v = Nk⋆ f = Nk f, fix λ > 1 in Lemma 2.4, and split with its aid v = vH 2 + vA
with vH 2 ∈ H 2 (Ω) and vA analytic; we have the following bounds
vH 2
H 2 (Ω)
≤C f
∇p vA
L2 (Ω) ,
L2 (Ω)
≤ C(λk)p−1 f
∀p ∈ N0 .
L2 (Ω)
We approximate vH 2 and vA separately. Theorem 2.10 and a scaling argument provides an
approximant wH 2 ∈ S p,1 (Th ) such that for every K ∈Th we have, for q = 0, 1,
vH 2 − wH 2
H q (K)
h
p
≤C
2−q
vH 2
∀K ∈ Th .
H 2 (K)
Hence, by summation over all elements, we arrive at
k vH 2 − wH 2
H
kh
+
p
≤C
2
kh
p
f
L2 (Ω) .
We now turn to the approximation of vA . Again, we construct the approximation wA ∈
S p,1 (Th ) in an element-by-element fashion. We start by defining for each element K ∈Th the
constant CK by
∇p vA 2L2 (K)
2
(2.26)
CK :=
2p
(2λk)
p∈N
0
and we note
∇p vA
L2 (K)
K∈Th
≤ (2λk)p CK
2
CK
≤
4
3
C
λk
∀p ∈ N0 ,
2
f
(2.27)
2
L2 (Ω) .
(2.28)
Let the element map for K be FK = RK ◦ AK . Then the function ṽ := vA |K ◦ RK satisfies, for
suitable constants C̃, C (which depend additionally on the constants describing the analyticity
of the element maps RK )3
∇p ṽ
L2 (K)
≤ C C̃ p max{p, k}p CK
∀p ∈ N0 .
Since AK is affine, the function v̂ := vA |K ◦ FK = ṽ ◦ AK therefore satisfies
∇p v̂
L2 (K)
≤ Ch−d/2 C̃ p hp max{p, k}p CK
∀p ∈ N0 .
Hence, the assumptions of Lemma 2.9 (with R = 1 there) are satisfied, and we get an approximation w on the element K by lifting an element-by-element construction on K to K via FK
which satisfies for q ∈ {0, 1}
vA − w
H q (K)
≤ Chd/2−q h−d/2 CK
3
h
h+σ
p+1
+
kh
σp
p+1
.
This is essentially proved in [12, Lemma 4.3.1]. Specifically, [12, Lemma 4.3.1] analyzes the case d = 2
and states that C, γ1 depends on the function g. Inspection of the proof shows that the case d = 3 can be
handled analogously and shows that the dependence on the function g can be reduced to a dependence on Cg ,
γg , and γf .
16
Summation over all elements K ∈Th gives
vA −w
2
H
≤
h
h+σ
2p
+k
2
h
h+σ
2p+2
k2
+ 2
p
2p
kh
σp
+k
2
kh
σp
2p+2
2
CK
. (2.29)
K∈Th
The combination of (2.29) and (2.28) yields
k vA − w
H
h
h+σ
≤C
p
hk
1+
h+σ
+k
kh
σp
p
1 kh
+
p σp
!
f
L2 (Ω) .
Furthermore, we estimate using h ≤ diamΩ and σ > 0 (independent of h)
h
h+σ
p
1+
kh
σ+h
≤ Ch(1 + kh)
p−1
h
σ+h
≤ Ch(1 + kh)p−2 ≤ C
h
p
1 kh
+
.
p
p
We therefore arrive at
k vA − w
H
≤C
1 kh
+
p
p
kh
+k
p
kh
σp
p!
f
L2 (Ω) ,
which completes the proof of the theorem.
Combining Theorems 2.11, 2.3 produces the condition
kh
+k
p
p
kh
σp
≤C
for sufficiently small C independent from k, h, and p required for quasi-optimality of the
hp-FEM. We extract from Theorem 2.11 that quasi-optimality of the h-version FEM can be
achieved under the side condition that p ≥ C log k:
Corollary 2.12 Let Ω = B1 (0) with the additional condition k0 ≥ 1 in the case d = 2. Let
Assumption 1.10 be valid. Then there exist constants c1 , c2 > 0 independent of k, h, and p
such that (2.2) is implied by the following condition:
kh
≤ c1
p
p ≥ c2 ln k.
together with
(2.30)
Alternatively, the discrete stability follows from
p = O (1) fixed independent of k
and kh + k (kh)p ≤ C
(2.31)
which is understood as a condition on the maximal step size h.
Proof. Theorem 2.11 implies
kh
kη(S) ≤ C 1 +
p
kh
+k
p
kh
σp
p
.
The right-hand side needs to be bounded by 1/Cc . It is now easy to see that we can select c1 ,
c2 such that this can be ensured.
An easy consequence of the stability result Corollary 2.12 is:
17
Corollary 2.13 Let the assumptions of Corollary 2.12 be satisfied and let (2.30) or (2.31)
hold. Then, the Galerkin solution uS exists and satisfies the error estimate
u − uS
H
≤ Cc
h
+
p
kh
σp
p
f
L2 (Ω)
.
Remark 2.14 To the best of the authors’ knowledge, discrete stability in 2D and 3D has only
been shown under much more restrictive conditions than (2.30), e.g., the condition k 2 h 1.
Even in one dimension, condition (2.30) improves the stability condition kh
1 that was
required in [9].
References
[1] L. Banjai and S. Sauter. A Refined Galerkin Error and Stability Analysis for Highly
Indefinite Variational Problems. SIAM J. Numer. Anal., 45(1):37—53, 2007.
[2] S. Brenner and L. Scott. The Mathematical Theory of Finite Element Methods. SpringerVerlag, New York, 1994.
[3] P. Ciarlet. The finite element method for elliptic problems. North-Holland, 1987.
[4] L. Demkowicz. Polynomial exact sequences and projection-based interpolation with applications to Maxwell’s equations. In D. Boffi, F. Brezzi, L. Demkowicz, L. Durán,
R. Falk, and M. Fortin, editors, Mixed Finite Elements, Compatibility Conditions, and
Applications, volume 1939 of Lectures Notes in Mathematics. Springer Verlag, 2008.
[5] W. Dörfler and S. Sauter. A posteriori error estimation for highly indefinite Helmholtz
problems. Comput. Methods Appl. Math., 13(3):333—347, 2013.
[6] S. Esterhazy and J. Melenk. On stability of discretizations of the Helmholtz equation. In
I. Graham, T. Hou, O. Lakkis, and R. Scheichl, editors, Numerical Analysis of Multiscale
Problems, volume 83 of Lect. Notes Comput. Sci. Eng., pages 285—324. Springer, Berlin,
2012.
[7] I. S. Gradshteyn and I. Ryzhik. Table of Integrals, Series, and Products. Academic Press,
New York, London, 1965.
[8] B. Guo and J. Zhang. Stable and compatible polynomial extensions in three dimensions
and applications to the p and h-p finite element method. 47(2):1195—1225, 2009.
[9] F. Ihlenburg. Finite Element Analysis of Acousting Scattering. Springer, New York,
1998.
[10] R. Leis. Initial Boundary Value Problems in Mathematical Physics. Teubner, Wiley Sons,
Stuttgart, Chichester, 1986.
[11] W. McLean. Strongly Elliptic Systems and Boundary Integral Equations. Cambridge,
Univ. Press, 2000.
[12] J. Melenk. hp-Finite Element Methods for Singular Perturbations. Springer, Berlin, 2002.
18
[13] J. Melenk, A. Parsania, and S. Sauter. Generalized DG Methods for Highly Indefinite
Helmholtz Problems Based on the Ultra-Weak Variational Formulation. J Sci Comput,
57:536—581, 2013.
[14] J. Melenk and S. Sauter. Convergence Analysis for Finite Element Discretizations of
the Helmholtz equation with Dirichlet-to-Neumann boundary condition. Math. Comp,
79:1871—1914, 2010.
[15] J. M. Melenk. On Generalized Finite Element Methods. PhD thesis, University of Maryland at College Park, 1995.
[16] J. M. Melenk. Mapping properties of combined field Helmholtz boundary integral operators. SIAM J. Math. Anal., 44(4):2599—2636, 2012.
[17] J. M. Melenk and S. A. Sauter. Wave-Number Explicit Convergence Analysis for Galerkin
Discretizations of the Helmholtz Equation. SIAM J. Numer. Anal., 49(3):1210—1243,
2011.
[18] J. M. Melenk and B. I. Wohlmuth. On residual-based a posteriori error estimation in
hp-FEM. Adv. Comput. Math., 15(1-4):311—331 (2002), 2001.
[19] M. Melenk and S. Sauter. Convergence Analysis for Finite Element Discretizations of the
Helmholtz Equation with Dirichlet-to-Neumann Boundary Conditions. Technical Report
09-2008, Universität Zürich, 2008. to appear in MathComp.
[20] J. C. Nédélec. Acoustic and Electromagnetic Equations. Springer, New York, 2001.
[21] S. Sauter. A Refined Finite Element Convergence Theory for Highly Indefinite Helmholtz
Problems. Computing, 78(2):101—115, 2006.
[22] S. Sauter and C. Schwab. Boundary Element Methods. Springer, Heidelberg, 2010.
[23] A. Schatz. An obeservation concerning Ritz-Galerkin methods with indefinite bilinear
forms. Math. Comp., 28:959—962, 1974.
[24] C. Schwab. p- and hp-finite element methods. The Clarendon Press Oxford University
Press, New York, 1998. Theory and applications in solid and fluid mechanics.
19