See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/265741784

Evaluating Local Approximations of the L-2-Orthogonal Projection Between Non-

Nested Finite Element Spaces

Article in Numerical Mathematics Theory Methods and Applications · August 2014

DOI: 10.4208/nmtma.2014.1218nm

CITATIONS READS

3 84

2 authors, including:

Rolf Krause
University of Lugano
170 PUBLICATIONS 1,309 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Time integration with SDC View project

Swiss Competence Center for Energy Research on the Electrical Infrastructure (SCCER-FURIES) View project

All content following this page was uploaded by Rolf Krause on 03 January 2016.

The user has requested enhancement of the downloaded file.

 Thomas Dickopf, Rolf Krause

Evaluating local approximations of the

L2-orthogonal projection between non-nested
finite element spaces

ICS Preprint No. 2012-01

8 March 2012

Via Giuseppe Buffi 13 ˆ CH-6904 Lugano ˆ phone: +41 58 666 4690 ˆ fax: +41 58 666 4536 ˆ
http://ics.unisi.ch
Evaluating local approximations of the L2 -orthogonal
projection between non-nested finite element spaces

Thomas Dickopf · Rolf Krause

Abstract We present quantitative studies of transfer operators between finite element spaces
associated with unrelated meshes. Several local approximations of the global L2 -orthogonal
projection are reviewed and evaluated computationally. The numerical studies in 3D pro-
vide the first estimates of the quantitative differences between a range of transfer operators
between non-nested finite element spaces. We consider the standard finite element inter-
polation, Clément’s quasi-interpolation with different local polynomial degrees, the global
L2 -orthogonal projection, a local L2 -quasi-projection via a discrete inner product, and a
pseudo-L2 -projection defined by a Petrov–Galerkin variational equation with a discontin-
uous test space. Understanding their qualitative and quantitative behaviors in this compu-
tational way is interesting per se; it could also be relevant in the context of discretization
and solution techniques which make use of different non-nested meshes. It turns out that
the pseudo-L2 -projection approximates the actual L2 -orthogonal projection best. The ob-
tained results seem to be largely independent of the underlying computational domain; this
is demonstrated by four examples (ball, cylinder, half torus and Stanford Bunny).
Keywords Finite elements · Unstructured meshes · Non-nested spaces · Transfer operators ·
Interpolation · Projection

1 Introduction

The question of how to interpolate functions in finite element spaces is as old as the fi-
nite element method itself. Approximation operators which map a given function to a finite
element space appear frequently in numerical analysis for a variety of reasons. For both
a priori and a posteriori discretization error estimates, one often needs to treat general el-
ements of infinite dimensional function spaces. However, such operators are usually not
evaluated computationally but merely used in the analysis. A famous example is the early
work by Clément [25] on finite element quasi-interpolation of discontinuous (non-smooth)
functions.
In this paper, we study different operators mapping to finite element spaces. Although
a couple of the reviewed properties hold true in case the input function is in an infinite

T. Dickopf · R. Krause
University of Lugano, Institute of Computational Science, Via G. Buffi 13, 6904 Lugano, Switzerland
Tel.: +41 58 666-4812, E-mail: [email protected], [email protected]
2 Thomas Dickopf, Rolf Krause

dimensional function space, our main focus is the information transfer between two finite
element spaces. We investigate this by a series of numerical studies. The respective spaces
are Lagrange conforming finite elements of first order associated with (either two or a whole
hierarchy of) non-nested meshes. This non-nested information transfer is important, for in-
stance, in non-conforming domain decomposition methods [8,9, 40,61,62] or in domain de-
composition or multigrid methods with non-nested coarse spaces [11, 15,18,19,23,29,46,
47,64] for the solution of partial differential equations. It appears both in the analysis and in
practical computations.
We present computational results on the behavior of local approximations of the L2 -
projection between non-nested finite element spaces. Our numerical studies are based on
a detailed overview of different transfer operators. Let us emphasize that there is neither
a conclusive characterization of the information transfer between finite element spaces as-
sociated with non-nested meshes nor a comprehensive classification of transfer operators
in this setting yet. In the present investigations, we show that, apart from basic similarities,
there are substantial conceptual and qualitative differences as well as substantial quantitative
differences between the studied operators.
Our research is in part motivated by the fundamental work on quasi-interpolation [25,
53]. We also learned about advanced techniques for the construction of transfer operators
from [40, 61,62] in the context of non-conforming domain decomposition methods. Other
interesting studies giving basic insights into the analysis of approximation operators in finite
element spaces, which influenced our work, can be found in [4,12,13,19,23,55–57, 63].
The evaluation criteria that are discussed in the theoretical part are the H 1 -stability,
an L2 -approximation property, the locality of the information transfer, and the projection
properties. In the computational part, we study the mutual relations between the diverse
transfer operators by numerical experiments. The operators are evaluated for four examples
of computational domains (ball, cylinder, half torus and Stanford Bunny), each time for
a series of independently generated meshes. All distances with respect to certain operator
norms are computed by solving the corresponding generalized eigenvalue problems. To our
knowledge, similar studies estimating distances between transfer operators in the present
context cannot be found elsewhere.
Let us briefly comment on an application for the non-nested information transfer studied
in this paper to multilevel preconditioners, which are among the most efficient algorithms
for the solution of discretized partial differential equations in many applications; see, e. g.,
[5,14,17,38,39,45,49,58,65] for some of the most influential achievements. For applica-
tions in computational engineering involving complicated geometries in three dimensions,
the construction of coarse spaces is often demanding and a flexible choice requires suitable
transfer operators. In case of elliptic problems discretized with first order elements on un-
structured meshes, it is known from the literature on domain decomposition methods [19,
23,29,57,64] that coarse spaces associated with non-nested meshes can be applied success-
fully to construct efficient preconditioners. This holds true both for two-level overlapping
Schwarz methods with global coarse space and for multigrid methods. These classes of
methods can be constructed either in a variational or in a non-variational setting. The first
one is characterized by a recursive Galerkin relation whereas the second one works with
virtually independent coarse level problems. In any case, transfer operators between finite
element spaces associated with non-nested meshes are necessary ingredients.

Outline This paper is organized as follows. In the remainder of Section 1, we introduce use-
ful notations concerning the information transfer in finite element spaces and comment on
characteristic properties of transfer operators. Section 2 is the first main part where different
Evaluating local approximations of the L2 -orthogonal projection 3

transfer operators are investigated comprehensively. Here, we also comment on a uniform

H 1 -stability estimate for the one-dimensional nodal interpolation. In a second main part,
Section 3, we present the quantitative studies of the diverse transfer operators and their mu-
tual relations for three geometric shapes in 3D. Section 4, is devoted to numerical studies
for a more complex geometry (Stanford Bunny).

1.1 Operators between finite element spaces associated with non-nested meshes

We recall several standard

R
notations from functional analysis; see, e. g., [1,34]. The Lebesgue
integral is denoted by · dxx. For a Lipschitz domain Ω ⊂ Rd , let L2 (Ω ) beR the Hilbert
space of square integrable functions in Ω with inner product (v, w)L2 (Ω ) := Ω vw dxx and
1/2
norm k · kL2 (Ω ) := (·, ·)L2 (Ω ) . The symbol L∞ (Ω ) represents the space of essentially bounded
functions with norm kvkL∞ (Ω ) := ess supx ∈Ω |v(xx)|. By H m (Ω ), as customary, we denote the
Sobolev space of functions with m ≥ 1 square integrable weak derivatives in Ω . Let α ∈ Nd
be a multi-index of order |α| := ∑1≤i≤d αi . Then, ∂ α denotes the weak differentiation and
the corresponding norm and semi-norm in H m (Ω ) are
 1/2  1/2
kvkH m (Ω ) := ∑|α|≤m k∂ α vk2L2 (Ω ) and |v|H m (Ω ) := ∑|α|=m k∂ α vk2L2 (Ω ) .

Moreover, the subspace of H 1 (Ω ) with vanishing image of the usual trace operator to the
boundary ∂ Ω is called H01 (Ω ); we have H01 (Ω ) = {v ∈ H 1 (Ω ) | v|∂ Ω = 0 in H 1/2 (∂ Ω )} =
{v ∈ H 1 (Ω ) | v = 0 a. e. on ∂ Ω }.
Let (Tℓ )ℓ∈N be a family of non-nested shape regular meshes (i. e., of non-overlapping
decompositions into finitely many open polytopes) of Lipschitz domains (Ωℓ ⊂ Rd )ℓ∈N
of dimension d ∈ {2, 3}, i. e., there is a constant c such that supℓ∈N maxT ∈Tℓ hT /rT ≤ c.
Here, for an element T ∈ Tℓ , let hT := diam(T ) be the diameter of T and rT the ra-
dius of the largest ball inscribed in T . For simplicity, assume Ωℓ ⊂ Ωℓ−1 for ℓ > 0. We
say (Tℓ )ℓ∈N is quasi-uniform if, in addition, there is another constant c, independent of ℓ,
such that maxT ∈Tℓ hT ≤ c minT ∈Tℓ hT for all ℓ ∈ N. As usual, local mesh size functions
hℓ ∈ L>∞ (Ω ) := {v ∈ L∞ (Ω ) | ∃ α > 0, such that v(x x) > α for a. e. x ∈ Ωℓ } are introduced
ℓ ℓ
for instance defined a. e. by hℓ (xx) := hT if x ∈ T . Assume the sequence (hℓ )ℓ∈N is non-
increasing where well-defined.
We denote the set of nodes of Tℓ by Nℓ and abbreviate nℓ := |Nℓ |. At each level ℓ,
we consider the space Xℓ of Lagrange conforming finite elements of first order and denote
its nodal basis as Λℓ = (λ pℓ ) p∈Nℓ with λ pℓ (q) = δ pq , p, q ∈ Nℓ . Let ω p := supp(λ pℓ ) be the
support of the basis function at node p ∈ Nℓ commonly called patch. We occasionally use
coordinate isomorphisms of the form Φℓ : Rnℓ → Xℓ , Φℓ (vv) := ∑ p∈Nℓ v p λ pℓ . See [24].
This paper is devoted to the numerics of the information transfer between non-nested
finite element spaces. A generic or unspecific transfer operator is denoted by Π . To every
concrete operator we will assign a different calligraphic symbol (I , P, Q, R), sometimes
varied by a tilde. If an operator maps between the two (non-nested) spaces Xℓ−1 and Xℓ , this
ℓ . Similarly, an operator mapping some other space, such as
will be indicated by, e. g., Πℓ−1
a Lebesgue or Sobolev space, to the finite element space Xℓ will be denoted by, e. g., Πℓ .
This shall suggest that a mesh Tℓ with a local mesh size function hℓ ∈ L> ∞ (Ω ) is always
ℓ
ℓ n ×n ℓ
involved. A matrix Π ℓ−1 ∈ R ℓ ℓ−1 represents an operator Πℓ−1 with respect to the chosen
ℓ v = Φ (Π ℓ −1
bases if Πℓ−1 ℓ Π ℓ−1 Φℓ−1 (v)) for all v ∈ Xℓ−1 .
4 Thomas Dickopf, Rolf Krause

1.2 On the properties of transfer operators

In this section, we briefly review characteristic properties of transfer operators which are
studied in this paper. Naturally, one does not only examine single entities but rather considers
entire types or families of transfer operators. For example, the terms “linear interpolation” or
“orthogonal projection” specify different instructions each providing an operator Π : X → Y
depending on certain data, namely the domains ΩY ⊂ ΩX ⊂ Rd and the spaces X ⊂ H 1 (ΩX )
and Y ⊂ H 1 (ΩY ). In this paper, the target space Y is always a finite element space.

Definition 1 Let ΩY ⊂ ΩX ⊂ Rd be domains. Given a subspace X ⊂ H 1 (ΩX ) and a (target)

∞ (Ω ), an operator
finite element space Y ⊂ H 1 (ΩY ) with discretization parameter hY ∈ L> Y
1
Π : X → Y is called H -stable in X if

|Π v|H 1 (ΩY ) . |v|H 1 (ΩX ) , ∀ v ∈ X. (1)

We say that the operator Π satisfies the L2 -approximation property if

khY−1 (v − Π v)kL2 (ΩY ) . |v|H 1 (ΩX ) , ∀ v ∈ X. (2)

Here and in the following, we write a . b if there is a constant c, which is independent

of the meshes, particularly of hY in (1) and (2), and the considered functions, such that
a ≤ cb. Note that no relation between X and Y has been specified other than the fact that
functions from X are also well-defined in the domain ΩY . Definition 1 constitutes a quite
general but common concept; the notions are used for both coarse-to-fine and fine-to-coarse
operators. The term H 1 -stability is slightly stronger than H 1 -continuity, especially in the
finite-dimensional case where every linear operator is continuous (with respect to every
equivalent norm), because the latter notion includes mappings with a continuity constant
dependent on the mesh.
A direct consequence of the L2 -approximation property (2) is that kv − Π vkL2 (ΩY ) → 0
for khY kL∞ (ΩY ) → 0, which holds for all v ∈ X. This observation, which has already been
considered by Clément in his work [25] on finite element interpolation of non-smooth func-
tions, further illustrates the nature of the operators addressed in this paper: An approximation
operator mapping to a finite element space converges (with respect to the L2 -norm) to the
natural embedding for increasing dimension of the target space.

Definition 2 Let Y be a finite element space associated with a mesh of a domain ΩY with
nodes NY , nY := |NY |, and coordinate isomorphism ΦY : RnY → Y . An operator ΠY :
H 1 (ΩY ) → Y is called local if



ΦY−1 (ΠY v) p = ΦY−1 (ΠY w) p , ∀ p ∈ NY , ∀ v, w ∈ H 1 (ΩY ), v|ω p = w|ω p .

In this paper, we discuss both non-local (global) and local operators in the sense of
Definition 2. Note that, indeed, the patch ω p is generally the most reasonable “domain of
influence” of the value at node p for the construction of a local approximation operator.
Finally, consider the case that an operator Πℓ : H 1 (Ωℓ ) → Xℓ acts as the identity map-
ping on the target space. This is true if, for instance, Πℓ is a surjective projection. Then, if
the meshes Tℓ−1 and Tℓ are nested, the fact that the corresponding spaces Xℓ−1 ⊂ Xℓ are
also nested implies immediately that the restricted mapping Πℓ−1 ℓ :X
ℓ−1 → Xℓ is the natural
embedding. This is a desirable property because one might argue that otherwise (if there
exists an element v ∈ Xℓ−1 such that Πℓ Πℓ v 6= Πℓ v) a considerable part of structure (namely
the appreciable fact that Xℓ−1 ⊂ Xℓ ) is unnecessarily disregarded.
Evaluating local approximations of the L2 -orthogonal projection 5

2 Operators between finite element spaces associated with non-nested meshes

In this section, we present various transfer operators and review their fundamental proper-
ties. Both intuitive and more elaborate mappings are examined to understand the information
transfer between finite element spaces associated with non-nested meshes. We discuss lo-
cally and globally defined operators including well-known quasi-interpolation concepts and
also focus on their algorithmic structure. Each transfer operator is at least “geometrically
inspired”, namely the algorithm to compute a concrete realization incorporates geometric
information in terms of finite element meshes. All operators presented here are studied ex-
perimentally in Section 3 and Section 4.

2.1 Standard finite element interpolation

First, we consider the most elementary operator. The standard finite element interpolation
or nodal interpolation in case of first order Lagrange elements is defined by

Iℓ : C 0 (Ω ) → Xℓ , u 7→ Iℓ u := ∑ p∈N u(p)λ pℓ .
ℓ

Here and in the following, assume that the generic domain Ω is sufficiently large. The op-
erator is surjective, namely Iℓ (C 0 (Ω )) = Xℓ , and a projection, i. e., for every v ∈ C 0 (Ω )
we have Iℓ Iℓ v = Iℓ v. The interpolation in Xℓ with the domain restricted to the finite ele-
ment space Xℓ−1 is called Iℓ−1 ℓ . Evidently, the operator is local according to Definition 2.

Moreover, when restricted to finite element spaces, it possesses the H 1 -stability and L2 -
approximation properties given in Definition 1 for shape regular meshes. This result can be
found in several papers; see, e. g., [19, 23,57]. For counterexamples for general H 1 -functions
and d = 2, see [3].
From a computational point of view, the standard nodal interpolation is very attractive.
Given an arbitrary function in C 0 (Ω ), the computation of the interpolant is very cheap with
one function evaluation per node in Nℓ , i. e., per basis function in Λℓ . It is without any doubt
the least expensive way to transfer information to a finite element space in a reasonable way.
For the computation of the matrix representation in Rnℓ ×nℓ−1 , this amounts to the evaluation
of λqℓ−1 (p) for all q ∈ Nℓ−1 and p ∈ Nℓ . Naturally, one may neglect the combinations with
p 6∈ ωq . This is straightforward if successive meshes are nested and parent–child relations
are known. In the non-nested setting, such neighborhood relations have to be computed; see
Section 2.6.

On the H 1 -stability of the nodal interpolation

In the literature, several different proofs have been brought forth for the stability and approx-
imation property of the nodal interpolation in case the domain is restricted to a (coarser)
finite element space; see [19, 23,57]. The stability estimates usually depend on the shape
regularity of the meshes which, in general, leads to constants greater than one. However,
we found an elementary proof in case d = 1, where no shape regularity assumption needs
to be considered, for the fact that “interpolation smoothes”. The following lemma states for
ℓ :X
the one-dimensional setting that the linear interpolation operator Iℓ−1 ℓ−1 → Xℓ has an
H 1 -stability constant less than or equal to one whether or not Dirichlet conditions come
into play. The uniform estimates hold without any assumptions on the mesh sizes or on the
relations between the meshes.
6 Thomas Dickopf, Rolf Krause

ℓ :X
Lemma 1 Let d = 1. Then, the nodal interpolation operator Iℓ−1 ℓ−1 → Xℓ satisfies the
1
following uniform H -stability estimates:
If Ωℓ ⊂ Ωℓ−1 ,
ℓ
|Iℓ−1 v|H 1 (Ωℓ ) ≤ |v|H 1 (Ωℓ ) ≤ |v|H 1 (Ωℓ−1 ) , ∀ v ∈ Xℓ−1 ,

otherwise,
ℓ
|Iℓ−1 E v|H 1 (Ωℓ ) ≤ |v|H 1 (Ωℓ ∩Ωℓ−1 ) , ∀ v ∈ Xℓ−1 , v|∂ Ωℓ−1 = 0,

where E : H01 (Ωℓ−1 ) → H 1 (Ωℓ−1 ∪ Ωℓ ) is the natural extension by zero. Moreover, the in-
ℓ,0
terpolation operator Iℓ−1 : Xℓ−1 → Xℓ ∩ H01 (Ωℓ ) enforcing zero function values at ∂ Ωℓ
satisfies
ℓ,0
|Iℓ−1 E v|H 1 (Ωℓ ) ≤ |v|H 1 (Ωℓ ∩Ωℓ−1 ) , ∀ v ∈ Xℓ−1 , v|∂ Ωℓ−1 = 0, v|∂ Ωℓ ∩Ω ℓ−1 = 0.

We emphasize that the symbols ⊂ and ⊃ always include the case of equality. The proof
and more details are elaborated in [30]. There, we also give counterexamples for the nodal
interpolation in higher order finite element spaces.

2.2 Clément-type quasi-interpolation

The following class of approximation operators has originally been introduced in [25] to
generalize the nodal interpolation in finite element spaces if the considered functions are
discontinuous. Quasi-interpolation is probably most famous for its frequent usage in proofs
of the reliability of a posteriori error estimators; see [22, 59] for a detailed review.
The Clément operator is defined by

Rℓ : L2 (Ω ) → Xℓ , u 7→ Rℓ u := ∑ p∈N (Q p u)(p)λ pℓ , (3)

ℓ

with the L2 -projections Q p onto the local polynomial spaces Pr (ω p ) of degree r ∈ N, i. e.,

u 7→ Q p u ∈ Pr (ω p ) : (Q p u, v)L2 (ω p ) = (u, v)L2 (ω p ) , ∀ v ∈ Pr (ω p ), p ∈ Nℓ . (4)

R
For instance, each projection Q p simply acts as local averaging by Q p u = |ω1p | ω p u dxx if
r = 0. By construction, the generated operators are local according to Definition 2.

Lemma 2 ([4,12,25]) The Clément operator Rℓ is H 1 -stable and has the L2 -approximation
property for r ∈ N.

Whereas the original results in [25] have been derived for triangular meshes in case
d = 2, the relevant properties can indeed be proved for finite element spaces associated with
general, not necessarily affine meshes and d ∈ {2, 3}. Note that the assertion holds true for
non-quasi-uniform meshes; we refer the reader to the discussion in [4]. The technical ideas
of the proof are perhaps most clearly elaborated in [12, Lemma 3.1], although in a slightly
different context.
By definition, the Clément interpolation acts as the nodal interpolation on polynomials
of degree r, namely Rℓ v = Iℓ v for all v ∈ Pr (Ω ). For the purpose of information transfer
between finite element spaces, regardless of whether nested or non-nested, which are built
from piecewise polynomials, this cannot be exploited, though.
Evaluating local approximations of the L2 -orthogonal projection 7

Restricting the attention to the discrete space Xℓ , one notes that the Clément interpolation
Rℓ : Xℓ → Xℓ does not keep invariant the basis functions; see (5) below. But this information
is not sufficient to determine the projection properties of the operator in the spaces L2 (Ω )
and Xℓ−1 , respectively. This is because, in general, one does not know whether the functions
λ pℓ ∈ Λℓ are contained in the range Rℓ (L2 (Ω )) or even Rℓ (Xℓ−1 ). However, considering the
size of the supports of the images of certain functions, we can prove the following

Proposition 1 Let the mesh Tℓ contain at least two interior nodes. Then, the quasi-interpo-
lation Rℓ : L2 (Ω ) → Xℓ is not a projection.

Proof Let v ∈ L2 (Ω ) be a non-negative, non-trivial function such that supp(Rℓ v) 6= Ω . It

is easy to see that such a “local” function v exists if the mesh Tℓ has at least two interior
nodes. Then, one can find an element T0 ∈ Tℓ with T0 6⊂ supp(Rℓ v) but T 0 ∩ supp(Rℓ v) 6= 0,
/
in other words an element adjacent to the support of Rℓ v. It is obvious that
[
supp(Rℓ λ pℓ ) = T | T ∈ Tℓ , T ∩ ω p 6= 0/ , ∀ p ∈ Nℓ . (5)

By definition, we have the linear combination Rℓ Rℓ v = ∑ p∈Nℓ (Q p v)(p)Rℓ λ pℓ with numbers

(Q p v)(p) ≥ 0. Because the functions Rℓ λ pℓ = ∑r∈Nℓ (Qr λ pℓ )(r)λrℓ , p ∈ Nℓ , are also non-
negative, the contributions coming from Rℓ λ pℓ and Rℓ λqℓ , p 6= q, do not cancel out each
other in the calculation of the effective coefficients of Rℓ Rℓ v with respect to the basis Λℓ .
Thus, it follows that T0 ⊂ supp(Rℓ Rℓ v) and, consequently, Rℓ Rℓ v 6= Rℓ v. This concludes
the proof of the proposition. ⊓
⊔

There are in fact subspaces U ⊂ L2 (Ω ) such that Rℓ Rℓ u = Rℓ u for all u ∈ U; for in-
stance, Pr (Ω ) has this property, as mentioned before. We now investigate to what extent
the above considerations also hold true for Rℓ−1ℓ : Xℓ−1 → Xℓ , namely if the domain of the
operator is restricted to the discrete subspace Xℓ−1 . For this purpose, suppose that there is a
node p ∈ Nℓ−1 and an element T1 ∈ Tℓ such that
[  [ 
int {T | T ∈ Tℓ , T ∩ supp(λ pℓ−1 ) 6= 0}
/ ∩ int {T | T ∈ Tℓ , T ∩ T 1 6= 0}
/ = 0.
/ (6)

Simply put, T1 needs to be sufficiently far away from the “reach of p”. This implies that
int(supp(Rℓ λ pℓ−1 )) ∩ T1 = 0.
/ Thus, supp(Rℓ λ pℓ−1 ) 6= Ω and one can find an element T0 ∈ Tℓ
which is adjacent to the support of Rℓ λ pℓ−1 . Concluding as before, we have the following

Proposition 2 Provided that (6) can be fulfilled, the Clément interpolation is not a projec-
tion even if its domain is restricted to the discrete subspace Xℓ−1 .

Note that the relatively weak assumption (6) is valid for virtually every pair of meshes
(Tℓ−1 , Tℓ ) one might handle. Therefore, we have shown that the Clément interpolation op-
erator is practically never a projection.
From Proposition 2 we observe the following: Neither does the Clément operator reduce
to the standard interpolation in case of nested meshes Tℓ−1 and Tℓ nor is it the identity map-
ping if the meshes and hence the associated spaces are identical. Evidently, this observation
is valid for any polynomial degree r ∈ N. In addition, this deficiency cannot be overcome by
changing the local domains (ω p ) p∈Nℓ in (4) by introducing another type of (overlapping or
non-overlapping) decomposition of local neighborhoods.
8 Thomas Dickopf, Rolf Krause

2.3 The L2 -projection

In this section, we comment on the use of an operator which appears naturally in the present
context. Let Qℓ : L2 (Ω ) → Xℓ be the L2 -projection onto Xℓ , i. e., the orthogonal projection
in the Hilbert space L2 (Ω ) to the subspace Xℓ characterized by the variational equation

u 7→ Qℓ u ∈ Xℓ : (Qℓ u, v)L2 (Ω ) = (u, v)L2 (Ω ) , ∀ v ∈ Xℓ .

The mapping Qℓ is global as opposed to Definition 2. This can be understood considering

ℓ
the algebraic representation of the fully discrete operator Qℓ−1 via a product
ℓ
Qℓ−1 M −1
v = Φℓ (M −1
ℓ B ℓ Φℓ−1 (v)), ∀ v ∈ Xℓ−1 , (7)

with the mass matrix M ℓ ∈ Rnℓ ×nℓ with respect to Λℓ , i. e., (M

M ℓ ) pq = (λ pℓ , λqℓ )L2 (Ω ) for p, q ∈
Nℓ , and a sparse coupling matrix B ℓ ∈ R n ℓ ×n ℓ−1 with the entries

Bℓ ) pq = (λ pℓ , λqℓ−1 )L2 (Ω ) ,
(B ∀ p ∈ Nℓ , q ∈ Nℓ−1 . (8)

Therefore, the L2 -projection “as is” cannot be expected to yield a computationally efficient
information transfer unless the evaluation of M −1 ℓ can be avoided. Note that the operator
attained by simply lumping the matrix M ℓ is considered in Section 2.4. We emphasize that
some of the (to a greater or lesser extent sophisticated) operators discussed in the literature
and in this paper are distinctly motivated by the idea to find an L2 -projection-like mapping
or a weighted interpolation which is more suitable for computations.
To obtain a stability estimate for the L2 -projection, the requirement of quasi-uniformity
of the mesh Tℓ has been considered inevitable for quite a long time. Meanwhile, weaker
criteria ensuring the H 1 -stability of Qℓ are available; see, e. g. [12, 21,27,55]. For estimates
with respect to other Lebesgue norms, see [33] and the references therein. Two different
proofs both using inverse estimates of Bernstein-type, which generally hold true only for
quasi-uniform meshes, can be found in [16, Theorem 3.4] and [10, Folgerung II.7.8]. The
L2 -approximation property in case of quasi-uniform meshes can be proved with elementary
techniques employing another suitable approximation operator such as the Clément quasi-
interpolation. In contrast, for a direct proof, see, e. g., [16, Theorem 3.2]. The latter employs
the fact that Qℓ yields the best approximation in Xℓ with respect to the norm k · kL2 (Ω ) . Fur-
ther ingredients are a standard finite element interpolation error estimate and an interpolation
technique between Sobolev spaces.

2.4 On L2 -quasi-projections

In this section, we consider a concept from the literature yielding local approximation op-
erators. The following quasi-projection operator has been employed in [13] to approximate
the L2 -projection from the space H 1 (Ω ) to the discrete spaces Xℓ . It is a mapping directly
defined via the formula
(λ pℓ , u)L2 (Ω )
Qeℓ : L2 (Ω ) → Xℓ , u 7→ Qeℓ u := ∑ p∈N λ pℓ , (9)
ℓ (λ pℓ , 1)L2 (Ω )

where 1 denotes the constant function with value 1; see also [20]. After all, we can obtain
a matrix representation of the fully discrete operator Qeℓ−1
ℓ : Xℓ−1 → Xℓ from the one of
2
the standard L -projection in a simple way by lumping the mass matrix M ℓ in (7). In the
Evaluating local approximations of the L2 -orthogonal projection 9

numerical practice, this seems a very natural thing to do. Moreover, for simplicial meshes, it
is easy to verify by integration over the reference element that (λ pℓ , 1)L2 (Ω ) = |ω p |/(d + 1)
for all p ∈ Nℓ ; thus, the operator Qeℓ may equivalently be defined by the variational equation

u 7→ Qeℓ u : (Qeℓ u, v)ℓ = (u, v)L2 (Ω ) , ∀ v ∈ Xℓ , (10)

with a specific discrete inner product (·, ·)ℓ in Xℓ , namely

1
d + 1 ∑T ∈Tℓ
(u, v)ℓ := |T | ∑ p∈N ∩T u(p)v(p), ∀ u, v ∈ Xℓ , (11)
ℓ

as proposed, e. g., in [63]. In other words, Qeℓ is the orthogonal projection to the space Xℓ
equipped with (·, ·)ℓ . Another discrete inner product also motivated by a quadrature rule (on
centroids of faces instead of nodes) can be found in [11]. There, it is used in the fashion
of (10) to define a prolongation operator between the non-nested spaces associated with a
discretization with Crouzeix–Raviart elements on nested meshes.
Note that Qeℓ is usually not a projection; this motivates the term quasi-projection. A proof
of this assertion can be achieved analogously to the ones of Proposition 1 and Proposition 2,
which treat the same issue for the Clément interpolation. In addition, one can easily see that
for p ∈ Nℓ we have (λ pℓ , λ pℓ )L2 (Ω ) < (λ pℓ , 1)L2 (Ω ) ; thus, Qeℓ λ pℓ 6= λ pℓ .
One needs to notice that there is virtually no experience with quasi-projections in practi-
cal computations. However, proofs of the H 1 -stability and the L2 -approximation property of
the operators Qeℓ for shape regular families of meshes are derived by well-known arguments
as described at the end of Section 2.5.
Finally, we notice that Qeℓ is self-adjoint with respect to the L2 -inner product, i. e.,

(λ pℓ , u)L2 (Ω ) (λ pℓ , v)L2 (Ω )
(Qeℓ u, v)L2 (Ω ) = ∑ p∈N = (u, Qeℓ v)L2 (Ω ) , ∀ u, v ∈ L2 (Ω ).
ℓ (λ pℓ , 1)L2 (Ω )

However, we do not know whether this property may be put to a good use in the analysis or
the practical computations at this point. This is because the two involved spaces are usually
not identical in applications. Note that the only operator that is self-adjoint and at the same
time a projection is the orthogonal projection Qℓ .

2.5 The pseudo-L2 -projection

In this section, a transfer operator is considered which is different in some respects. It will be
denoted by the symbol P with the appropriate indices. Generally speaking, we introduce a
Petrov–Galerkin scheme with a discontinuous test space built from a set of functions which
are biorthogonal to the standard nodal basis with respect to the L2 -inner product (·, ·)L2 (Ω ) .
By this means, the global variational formulation defines local mappings, which is similar
to (10). In the fully discrete setting, this yields a band matrix representation of the operator
as no mass matrix has to be inverted.
The mapping Pℓ is in fact a projection from L2 (Ω ) onto the finite element space Xℓ .
Additionally, in the authors’ view, the operator represents a way to get “as close as possible”
to the real L2 -projection while at the same time it guarantees an efficient evaluation. This is
clearly confirmed by the numerical experiments in Section 3 and Section 4. Therefore, we
suggest to call this oblique projection operator “pseudo-L2 -projection” in this context. This
term is also meant to contrast, e. g., the L2 -quasi-projection concepts of Section 2.4, which
10 Thomas Dickopf, Rolf Krause

yield in actual fact no projections. Moreover, the pseudo-L2 -projection seems to be the only,
reasonably straightforward operator in the fashion of the previous ones (9) and (10) which
is actually a projection.
For the definition of the operator, choose a set of functions Ψℓ = (ψ pℓ ) p∈Nℓ with ψ pℓ |ω p ∈
0
C (ω p ) for all p ∈ Nℓ , extended to Ω by zero, such that

(ψ pℓ , λqℓ )L2 (Ω ) = δ pq (λ pℓ , 1)L2 (Ω ) , ∀ p, q ∈ Nℓ , (12)

and set the discontinuous test space as Yℓ := span{ψ pℓ | p ∈ Nℓ } 6⊂ C 0 (Ω ). Note that such
a dual basis with respect to (·, ·)L2 (Ω ) of the nodal finite element basis Λℓ = (λ pℓ ) p∈Nℓ ex-
ists. This can also be seen, for example, in the various procedures in [36, 48] for an explicit
construction of the set Ψℓ . In particular, on its support ω p , each ψ pℓ can be represented by
a linear combination of the nodal basis functions associated with the adjacent elements re-
stricted to ω p . In case of affine elements, the coefficients do not depend on the actual node
p and element T but can be computed on the reference element in a one-time process; no
inverse element mass matrices are necessary. This is due to the scaling with (λ pℓ , 1)L2 (Ω )
on the right hand side of (12) which also implies the boundedness kψ pℓ kL∞ (Ω ) . 1 for all
p ∈ Nℓ . For a more detailed analysis of biorthogonal bases, carried out in the context of
the mortar finite element method [9], and the construction of such systems for higher order
finite element spaces, we refer to [40, 42,43,62].
Now, we define the pseudo-L2 -projection Pℓ : L2 (Ω ) → Xℓ by a global Petrov–Galerkin
variational formulation with trial space Xℓ and test space Yℓ , i. e.,

u 7→ Pℓ u : (Pℓ u, v)L2 (Ω ) = (u, v)L2 (Ω ) , ∀ v ∈ Yℓ . (13)

This variational problem has a unique solution because dim(Yℓ ) = dim(Xℓ ) < ∞ and for
u ∈ Xℓ it is (u, v)L2 (Ω ) = 0 for all v ∈ Yℓ if and only if u = 0. In particular, the definition
yields the obvious representation formula

(ψ pℓ , u)L2 (Ω )
Pℓ u = ∑ p∈N λ pℓ , ∀ u ∈ L2 (Ω ). (14)
ℓ (λ pℓ , 1)L2 (Ω )

The operator Pℓ is well-defined by Hölder’s inequality. The fully discrete representation in

Rnℓ ×nℓ−1 of the pseudo-L2 -projection Pℓ−1ℓ : Xℓ−1 → Xℓ is obtained analogously to the ones
2 2
of the L -projection and the L -quasi-projection from the previous sections.
The idea to use a Petrov–Galerkin scheme to define a generalized projection operator
can be found in [55] for d ∈ {1, 2}, too. There, the test space is constructed differently; the
local test functions are associated with a dual mesh. Undoubtedly, the root of the considered
operators lies in the research of quasi-interpolation concepts by Clément [25]. However, the
first appearance of a weighted interpolation operator using a system of biorthogonal test
functions was in [53]. Biorthogonal systems of some form or another are classic in linear
algebra and considerably more common in the context of wavelets; see, e. g., [26, 28,60].
Generalized projections using dual test functions have first been introduced to the area of
domain decomposition methods by [61, 62] and then [40]. In practical computations, oper-
ators of this type have been used to map trace functions between non-matching interfaces;
see also [31, 32] and the references therein.
As opposed to the earlier version proposed in [53], we do not aspire to preserve Dirichlet
boundary conditions on ∂ Ω . This is immediately reflected by the definition as supp(ψ pℓ ) =
ω p also for boundary nodes. We do not have to choose suitable (d − 1)-dimensional sub-
simplices on the boundary but rather work with the given finite element meshes, which
Evaluating local approximations of the L2 -orthogonal projection 11

makes the definition, in a sense, more symmetric. A further advantage, which we do not yet
exploit here, is the lower requirement for the regularity of the considered functions, namely
L1 (Ω ) instead of Wpm (Ω ) with m ≥ 1 if p = 1 and m > 1/p otherwise; see also [20].
Let us now examine the properties of the operator more closely. First of all, the map-
ping Pℓ is surjective, namely Pℓ (L2 (Ω )) = Xℓ , because (12) and (14) immediately imply
Pℓ λ pℓ = λ pℓ for all p ∈ Nℓ . Moreover, it is a projection onto Xℓ . This is a simple consequence
of the linearity of the operator and, again, the biorthogonality property (12). In addition, it
is important to note the following
Lemma 3 The pseudo-L2 -projection Pℓ is H 1 -stable and has the L2 -approximation prop-
erty for all shape regular families of meshes.
The assertion may proved by well-known arguments as Poincaré’s inequality and the fact
that constant functions are reproduced locally; see [29, Section 5.5]. It indeed holds true for
shape regular (not necessarily quasi-uniform) meshes with a mesh size function hℓ ∈ L> ∞ (Ω )

because all estimates are local; in an element T ∈ Tℓ they only involve the values in a small
neighborhood. Similar arguments yield the H 1 -stability and the L2 -approximation property
of the L2 -quasi-projection Qeℓ described in Section 2.4 as it also reproduces the constant
functions locally.

2.6 Implementation aspects

In this section, we focus on the realization of the specific transfer operators in practical finite
element codes. All described methods are implemented in a module nnmglib (developed in
[29]) in the package obslib++, which is maintained by the second author and his work
group. The software uses fundamental components of the finite element toolbox ug; see [7].
For the computation of a matrix representation of a linear operator from Xℓ−1 to Xℓ , one
needs to deal with quantities associated with different meshes without any usable a priori
relation. Therefore, we have incorporated the quadtree/octree implementation of [2] into
obslib++. Suitable advancing front techniques exploiting the connectivities of the single
meshes can be applied instead; see, e. g., [37] in a related context. In any case, for each node
p
p ∈ Nℓ , a set Nℓ−1 ⊂ Nℓ−1 containing a sufficiently small number of nodes is determined
such that
p
q ∈ Nℓ−1 , int(ωq ) ∩ int(ω p ) 6= 0/ =⇒ q ∈ Nℓ−1 .
Then, all terms which appear in the presented discrete operators may evidently be computed
only based on these local index subsets. This results in almost linear complexity of the
assembly procedure.

2.6.1 Numerical integration

For all transfer operators, with the exception of the nodal interpolation, L2 -inner products
of functions associated with different meshes need to be computed. This is obvious for the
ℓ
operators Qℓ−1 and Qeℓ−1
ℓ involving the sparse but global coupling matrix B ℓ ∈ Rnℓ ×nℓ−1
ℓ
defined in (8). The analogon for the pseudo-L2 -projection Pℓ−1 requires the entries

(ψ pℓ , λqℓ−1 )L2 (ω p ∩ωq ) , ∀ p ∈ Nℓ , q ∈ Nℓ−1 . (15)

We turn to the other mappings which employ local orthogonal projections below.
12 Thomas Dickopf, Rolf Krause

To evaluate (8) or (15) exactly, one has to compute the intersections of the elements in
the consecutive meshes. As we have previously done in [31] for the intersection of locally
projected non-matching interface meshes, we employ the quickhull algorithm in an imple-
mentation by [6] for this purpose. After a suitable remeshing of the computed intersection
polytopes, one achieves an exact integration, up to roundoff errors, by the application of low
order quadrature rules. We have implemented the methods concerning element intersections
in a module cutlib.
In practice, good results may be obtained by an approximate numerical integration via a
quadrature rule solely based on the finer mesh. The order of the employed quadrature rules
should be adequate such that they are exact at least in case of nested meshes. This requires
order two for the above operators and order r + 1 for the Clément quasi-interpolations. We
are aware of the fact that such an approach might fail to retain optimal (discretization) error
estimates, for instance, in the mortar finite element setting; see [35, 44]. However, let us refer
to the numerical studies in Section 3.3, where we demonstrate that the error in the operator
itself due to approximate integration is small if the quadrature rule is chosen adequately.

2.6.2 Computation of orthogonal projections

The evaluation of the operator Qℓ−1 ℓ , which is the orthogonal projection to the space X
ℓ
2
with respect to the L -inner product, is very expensive. For the experiments in Section 3 and
Section 4, we employ the direct sparse solver pardiso [51, 52] to decompose the appearing
mass matrices. This is more efficient than an iterative solver in this special case as the re-
spective inverse needs to be applied to a large number of vectors. The pseudo-L2 -projection
ℓ
Pℓ−1 is also defined via a global variational formulation but can be evaluated efficiently
by construction. Note that the quasi-projection Qeℓ−1 ℓ yields a simple formulation, too. The
same holds true for a quasi-interpolation operator in case the local trial and test spaces are
one-dimensional, e. g., for the Clément operator Rℓ−1 ℓ with r = 0.
In contrast, we have seen that general transfer operators may require the evaluation of
local orthogonal projections. In the following, we sketch the implementation of the operators
ℓ
Rℓ−1 with r > 0. To solve the corresponding local variational equations (4) for the right hand
sides given by the coarse level basis functions, one needs to compute coarse-to-fine coupling
matrices and mass matrices similar to the ones in (7) but associated with the local spaces.
Let (φip )1≤i≤n p be a basis of the considered trial space at p ∈ Nℓ . Then,
M p )i j = (φip , φ jp )L2 (ω p ) ,
(M ∀ 1 ≤ i, j ≤ n p ,
and
B p )iq = (φip , λqℓ−1 )L2 (ω p ) ,
(B ∀ 1 ≤ i ≤ n p , q ∈ Nℓ−1 , (16)
are the respective local matrices. We omit the level index ℓ as it is clear from the choice
of p. For the Clément quasi-interpolation operators, the trial and test spaces are obtained
by restrictions of global polynomial spaces to the patches. Therefore, one may choose a
universal basis for the implementation; for instance, (φi )1≤i≤d+1 with φi (xx) = x · e i for 1 ≤
i ≤ d and φd+1 ≡ 1 is a convenient choice in case r = 1. The issues concerning the numerical
integration of (16) are solved as before for the global coupling matrices. As usual in finite
element assembly algorithms, a single loop over all elements in Tℓ makes sure that no
redundant computations are carried out; each integral contribution is only computed once.
Finally, the entries of the global matrix representations in Rnℓ ×nℓ−1 of the Clément operators
ℓ
Rℓ−1 : Xℓ−1 → Xℓ read as
n n
p
∑i=1 p
∑ j=1 M −1
(M B p ) jq φi (p),
p )i j (B ∀ p ∈ Nℓ , q ∈ Nℓ−1 .
Evaluating local approximations of the L2 -orthogonal projection 13

Fig. 1 Examples for the unstructured meshes for the numerical studies: ball B, half torus H , cylinder C
(from left to right). The characteristics of the independently generated meshes are given in Table 1.

This formula is immediately derived by solving the variational equation (4) for the basis
functions (λqℓ−1 )q∈Nℓ and evaluating the result at the node p. The inversion of the nℓ local
mass matrices (MM p ) p∈Nℓ is required with dimension n p = dim(Pr (ω p )) = (d + r)!/(d! r!).

3 Numerical studies of the diverse transfer operators

Let us now focus on the practical properties of the described transfer operators. In this sec-
tion, we report on various numerical experiments which are performed to assess intercon-
nections between the single operators. Subjecting the discrete mappings to a close examina-
tion, we want to understand better what the fundamental characteristics of the information
transfer between non-nested finite element space are.
The examples of computational domains studied here are three geometric shapes (ball,
cylinder, half torus) each with a series of independently generated meshes; see Section 3.1.
To investigate the behaviors of the mappings, we look at suitable operator norms with re-
spect to the L2 -norm and the H 1 -semi-norm associated with the appropriate domains. The
desired quantities are computed by solving generalized eigenvalue problems as described
in Section 3.2. We obtain results on the accuracy of an approximate numerical integration
in Section 3.3. Finally, we examine the quantitative differences of the transfer operators by
measuring the distances between them in Section 3.4.
The assessment performed here is motivated by the desire to become more familiar
with the application of (to a greater or lesser extent sophisticated) (quasi-)interpolation and
(quasi-)projection operators in practical computations. It allows for “drawing a map” arrang-
ing the operators by their mutual relations. To our knowledge, the evaluation of operators for
the information transfer between finite element spaces associated with non-nested meshes
has never been studied in such a manner so far.

3.1 Setup of the experiments

For the experiments presented in this section, we consider a number of independently gen-
erated meshes of a ball, a cylinder and a half torus, respectively. These are appropriate
geometric settings as one can easily obtain completely independent unstructured volume
meshes for a large variety of different mesh sizes by standard tetrahedral mesh generation
tools, e. g., from CUBIT [50]. In addition, they yield very good reproducibility. Note that
the setting is also sufficiently general. On the one hand, this can be seen in an illustrative
example in Remark 1 at the beginning of Section 3.4. On the other hand, we study a more
complex geometry in Section 4.
14 Thomas Dickopf, Rolf Krause

Table 1 Characteristics of the independently generated meshes (Bi )1≤i≤9 of a ball, (Ci )1≤i≤9 of a cylinder
and (Hi )1≤i≤9 of a half torus. The meshes do not stem from a refinement routine; they cover a broad range
of sizes.
#elements #nodes #elements #nodes #elements #nodes
B1 292 88 C1 239 78 H1 649 191
B2 580 150 C2 751 196 H2 1,421 371
B3 1,708 392 C3 1,424 348 H3 1,964 500
B4 3,616 778 C4 4,407 947 H4 6,392 1,418
B5 10,711 2,168 C5 8,100 1,690 H5 12,329 2,613
B6 48,320 9,228 C6 16,591 3,313 H6 33,486 6,700
B7 64,773 12,294 C7 27,681 5,372 H7 56,959 11,126
B8 93,620 17,647 C8 50,195 9,570 H8 80,881 15,590
B9 123,946 23,259 C9 103,746 19,373 H9 111,439 21,235

We use a set of tetrahedral meshes (Bi )1≤i≤9 of a ball, (Ci )1≤i≤9 of a cylinder and
(Hi )1≤i≤9 of a half torus, respectively, with their characteristics given in Table 1 ordered
by the number of elements. For each of the geometric shapes, one mesh is illustrated in
Figure 1. Note that the situation between the single meshes is sufficiently general in the
sense that there are no mutual relations other than that they approximate the same domain.
In particular, none of the meshes stems from a refinement routine; they are all imported
separately.
For the three cases, we consider mappings between the different meshes and introduce
the notations, again, by using the generic operator symbol Π with i and j as indices and
exponents. For the purposes of the present section, we do not need to distinguish between
the different objects associated with the ball, the cylinder and the half torus by marking the
symbols for the domains, spaces and operators, respectively. This is because the geometric
shapes are treated one at a time. Consequently, let the corresponding domains be denoted
by (Ωi )1≤i≤9 ; accordingly, (Xi )1≤i≤9 are the standard finite element spaces associated either
with the meshes (Bi )1≤i≤9 or (Ci )1≤i≤9 or (Hi )1≤i≤9 without any boundary modifications.
Then, we denote the connecting operators, e. g., by Πij : Xi → X j for 1 ≤ i < j ≤ 9.

3.2 Computation of operator norms

In the following, operator norms play a central role; for Πij , Π

e j ∈ Lin(Xi , X j ), we study
i
terms of the form
kΠij vk j kΠij v − Π
e j vk j
i
sup or sup (17)
v∈Xi ∩H 1 (Ωi ), kvki 6=0 kvki
0
1
v∈Xi ∩H (Ωi ), kvki 6=0
0
kvki

where k · ki and k · k j are suitably chosen (semi-)norms in Xi and X j , respectively. All men-
tioned transfer operators may be employed to map an infinite-dimensional function space to
a finite element space which is normally a subspace. However, we emphasize that we do not
consider a general Hilbert space setting but restrict the attention to the case of two finite el-
ement spaces which is relevant for the outlined applications. Therefore, the suprema in (17)
are taken over finite element functions in Xi only. Finally, we require the test functions to be
in H01 (Ωi ) such that their extensions by zero to the possibly larger domain Ω j are continuous
and piecewise first order polynomials and, thus, weakly differentiable.
To compute quantities of the form (17) with respect to the L2 -norm and the H 1 -semi-
norm, respectively, the corresponding generalized eigenvalue problems are considered. For
Evaluating local approximations of the L2 -orthogonal projection 15

this purpose, we abbreviate Nℓ 0 := {p ∈ Nℓ | p 6∈ ∂ Ωℓ } and n0ℓ := |Nℓ 0 |. Let the matrix

R
Aℓ ∈
Aℓ ) pq = Ωℓ ∇λ pℓ ·
Rnℓ ×nℓ be the representation of the H 1 -semi-norm with respect to Λℓ , i. e., (A
0 0 R
∇λqℓ dxx for p, q ∈ Nℓ , and similarly A 0ℓ ∈ Rnℓ ×nℓ with entries (A A0ℓ ) pq = Ωℓ ∇λ pℓ · ∇λqℓ dxx
0
for p, q ∈ Nℓ 0 . Naturally, A ℓ is symmetric positive definite. We also introduce the mass
0 0
matrix in the interior M 0ℓ ∈ Rnℓ ×nℓ with (M
M 0ℓ ) pq = (λ pℓ , λqℓ )L2 (Ωℓ ) for p, q ∈ Nℓ 0 . Finally, let
0
Π ij ∈ Rn j ×ni be the matrix representation of an operator Πij : Xi ∩ H01 (Ωi ) → X j with respect
to the chosen bases. Then, we have the operator norms
n√ 0
o
kΠij kL2 = max Π ij )T M j Π ij v = ηM
η ∈ R | ∃ v ∈ Rni such that (Π M 0i v (18)

and n√ o
0
|Πij |H 1 = max Π ij )T A j Π ij v = ηA
η ∈ R | ∃ v ∈ Rni such that (Π A0i v . (19)

For brevity, in the notations of the operator norms, we omit the two different spaces with
the two different domains. Distances between operators may be measured likewise by, e. g.,
kΠij − Πe j k 2 for some Π j , Π
i L i
e j ∈ Lin(Xi , X j ). We also consider relative quantities, namely
i
j e | 1 /|Π j | 1 . Note that a sampling procedure that we used previ-
j
terms of the form |Πi − Π i H i H
ously to estimate operator norms yielded essentially the same results.
The generalized eigenvalue problems in (18) and (19) are solved iteratively by the
locally optimal block preconditioned conjugate gradient method (LOBPCG) [41] for the
largest eigenvalues. We proceed as outlined in Section 2.6 to obtain numerical representa-
tions of the transfer operators. Then, to compute the desired quantities by solving (18) and
(19), respectively, one step of the LOBPCG method requires several matrix-vector multi-
0 0
plications involving mass or stiffness matrices both in Rn j ×n j and in Rni ×ni as well as pro-
n ×n 0 2
longation matrices in R j i and their transposes. To evaluate the L -projection, additional
forward-backward substitutions are necessary.

3.3 Influence of numerical integration

In this section, we consider the inexact integration of the coupling terms between the basis
functions of Xi and X j by means of a quadrature rule solely associated with the target mesh,
as described in Section 2.6. We verify that this approximation is very accurate in case of suf-
ficiently many function evaluations per element. (Note that inner products of finite element
functions associated with the same mesh are always evaluated exactly except for roundoff
errors.) To quantify the effect not on the integrals as such but on the actual mappings, we
estimate the relative differences between the transfer operators Q, P, Q, e Rr=0 , Rr=1 and
Rr=2 on the one hand and approximate versions on the other hand.
For this purpose, composite quadrature formulas are employed. These rules are gener-
ated by regular decompositions of the tetrahedron into m3 tetrahedra of the same volume,
m ∈ N+ . Then, a rule of second order with four points is used on each of the sub-elements.
As the integrands are of low order but the integration domains may have relatively compli-
cated shapes, this is an appropriate choice; see Section 2.6.
We look at the decay of the quadrature error for the composite rules with m = 1, 2, 3
(that is 4, 32 and 108 points per element, respectively) using the results for m = 5 (i. e.,
500 points per element) as reference. Figure 2 shows the relative errors in the evaluation of
the transfer operators in logarithmic scale. The results are given for the combinations of the
16 Thomas Dickopf, Rolf Krause

−1 −1 −1 −1 −1 −1
10 10 10 10 10 10

−2 −2 −2 −2 −2 −2
10 10 10 10 10 10

−3 −3 −3 −3 −3 −3
10 10 10 10 10 10

4 32 108 4 32 108 4 32 108 4 32 108 4 32 108 4 32 108

e Rr=0 , Rr=1 and Rr=2 (from

Fig. 2 Estimated relative errors with respect to | · |H 1 of the operators Q, P, Q,
left to right) depending on the number of integration points per element. Each line represents the error decay
in one of the combinations (Bi , B j )1≤i< j≤5 .

Table 2 Symbols and colors of the operators in the charts

I Rr=0 Rr=1 Rr=2 Q Qe P

circle ◦ plus + square  crossing × diamond ♦ triangle △ dot •
(black) (orange) (green) (red) (petrol) (purple) (blue)
Section 2.1 Section 2.2 Section 2.2 Section 2.2 Section 2.3 Section 2.4 Section 2.5

first five meshes of the ball, (Bi )1≤i≤5 , namely we investigate Πij : Xi → X j for 1 ≤ i < j ≤ 5
for different types of Π .
As expected, the quality of the approximation improves considerably as the number of
integration points is increased. We also note that, for fixed coarse mesh Bi , the error be-
comes smaller with increasing index j. This is obvious but cannot be seen in the figure as
we do not intend to label all the single lines. Other than that, we do not experience any de-
pendence on the mesh size. In particular, for the most critical combinations (Bi , Bi+1 )1≤i≤4 ,
the errors depicted in Figure 2 do not grow with increasing i.
The error decay is slightly different for the six transfer operators because both the
coarse-to-fine integrand (e. g., λ pj λqi for Q versus ψ pj λqi for P) and the form of the error
transport vary (e. g., inverse mass matrix for Q versus diagonal scaling for Q). e Regardless,
we note that the composite quadrature rules produce very accurate approximations of the
operators. The relative errors in the operators with respect to | · |H 1 are of the order of 1% or
less; the errors with respect to k · kL2 (not shown here) are even smaller.

3.4 Quantitative analysis of the relations between the transfer operators

In this section, we present a quantitative study of the diverse transfer operators. This eventu-
ally allows for arranging them in a map-like sketch illustrating similarities and differences.
In the charts designed for this purpose, the operators are marked by the symbols and with
the colors specified in Table 2 where they appear.
Remark 1 To illustrate that the relations between the employed meshes are sufficiently gen-
eral, we consider rotations of the mesh B4 about the axis spanned by the sum of the standard
basis vectors (eei )1≤i≤3 by different, arbitrarily chosen angles. Table 3 states several relative
differences for operators between these rotated meshes (for the angles specified at the head)
and the meshes B5 and B6 . If there were distinguished relations between the unrotated
mesh and (some of) the other meshes, one would expect the computed quantities to vary
more significantly. This is not the case in this study and the other studies we performed.
Evaluating local approximations of the L2 -orthogonal projection 17

Table 3 The setting is sufficiently general. The computed estimates of the operator norms are independent
of rotations of the meshes. We show exemplarily the relative differences between P and Q and between Qe
and Q w. r. t. | · |H 1 (top) and k · kL2 (bottom), respectively, between meshes of different sizes.

0◦ 1◦ 2◦ 3◦ 4◦ 9.7◦ 17.1◦ 41.3◦

|P45 − Q45 |H 1 /|Q45 |H 1 0.29 0.29 0.30 0.30 0.30 0.26 0.31 0.32
|Qe5 − Q 5 | 1 /|Q 5 | 1
4 4 H 4 H 0.61 0.61 0.61 0.63 0.64 0.62 0.59 0.61
|P46 − Q46 |H 1 /|Q46 |H 1 0.29 0.29 0.28 0.27 0.26 0.25 0.28 0.28
|Qe46 − Q46 |H 1 /|Q46 |H 1 0.58 0.58 0.57 0.58 0.58 0.57 0.60 0.58
kP45 − Q45 kL2 /kQ45 kL2 0.25 0.26 0.25 0.25 0.24 0.24 0.24 0.24
kQe45 − Q45 kL2 /kQ45 kL2 0.53 0.53 0.53 0.54 0.55 0.55 0.53 0.53
kP46 − Q46 kL2 /kQ46 kL2 0.21 0.21 0.21 0.21 0.21 0.21 0.21 0.20
kQe6 − Q 6 k 2 /kQ 6 k 2
4 4 L 4 L 0.50 0.50 0.50 0.50 0.50 0.50 0.49 0.50

Now, for each geometric shape and then each of the following choices of two finite
element meshes, (Bi , B j ), (Ci , C j ) and (Hi , H j ), 1 ≤ i < j ≤ 9, we consider the distances
of the generated operators Ii j , (Rr=0 )ij , (Rr=1 )ij , (Rr=2 )ij , Qeij , and Pij to the L2 -projection
Qij between the spaces Xi and X j . Figure 3 and Figure 4 show the relative differences with
respect to | · |H 1 and k · kL2 , respectively. The diagrams are arranged such that a section
marked by either Bi or Ci or Hi below (for some index i) comprises the results for all the
situations (Bi , B j )i< j or (Ci , H j )i< j or (Ci , H j )i< j , each time ordered by increasing j from
left to right.
We point out two distinct facts established by the performed experiments and readily
understood by the figures. First, with decreasing ratio between fine and coarse mesh size, all
depicted operators approximate Q more accurately. This is because they have the common
property to preserve the constant functions, which has been mentioned before. In a certain
sense, a very fine mesh is “almost nested” in a very coarse mesh and the coarse function is
“almost constant” in the patches of the fine mesh; thus, the operators asymptotically become
more and more like the identity if the coarse mesh is fixed.
The second, even more important result is the following. We see that, consistently for all
experiments, the pseudo-L2 -projection is clearly the closest to the actual L2 -projection. In
fact, it is remarkable how much closer this operator is to the orthogonal projection compared
to all other approaches. The standard interpolation and the Clément-type interpolation with
local polynomial degree r = 2, although being only moderately close to each other as we
show shortly, have a very similar distance to Q. These two operators are the next closest
to the orthogonal projection; they are roughly twice as far away from Q as the pseudo-L2 -
projection is. The others are considerably further away.
Another important point is that the ratio between fine and coarse mesh size is most rel-
evant for the considered distances but not the mesh size itself. This is illustrated in Figure 5
and Figure 6. Here, we have collected several cases, each of four mesh combinations, each
with a roughly comparable ratio of the numbers of elements. The ranges of this ratio are
given in the description of Figure 5. The classification is somewhat arbitrary; however, the
diagrams show that the approximate differences to Q and Rr=2 , respectively, do not vary
significantly in the considered situations. Note that all charts have the same scale on the
vertical axes. As one has seen before in Figure 3, the behavior of I appears to be the least
predictable.
18 Thomas Dickopf, Rolf Krause

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
B1 B2 B3 B4 B5 B6 B7 B8
(a) Ball

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
C1 C2 C3 C4 C5 C6 C7 C8
(b) Cylinder

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
H1 H2 H3 H4 H5 H6 H7 H8
(c) Half torus

Fig. 3 Relative distances to the L2 -projection Q with respect to | · |H 1 for the meshes from Table 1. Each
marker represents one measurement (for example, the first blue dot marks |P12 − Q12 |H 1 /|Q12 |H 1 , the tenth
black circle is |I24 − Q24 |H 1 /|Q24 |H 1 and so forth). The labels (Bi )1≤i≤9 , (Ci )1≤i≤9 and (Hi )1≤i≤9 indicate
the sections where the particular space Xi is the same. In each of these sections, the results are given for
increasing j from left to right.

Finally, to highlight the interconnections, we state the complete data, namely the mutual
relative distances between the operators with respect to | · |H 1 , for one typical setting. The
results for the mappings generated between the spaces associated with B3 and B7 are given
in Table 4 ordered by their proximity to the L2 -orthogonal projection. In each cell, we state
the relative difference of the two specified operators with respect to the one in the current
row. Please be assured that this example is indeed representative.
Evaluating local approximations of the L2 -orthogonal projection 19

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
B1 B2 B3 B4 B5 B6 B7 B8
(a) Ball

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
C1 C2 C3 C4 C5 C6 C7 C8
(b) Cylinder

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
H1 H2 H3 H4 H5 H6 H7 H8
(c) Half torus

Fig. 4 Relative distances to the L2 -projection Q with respect to k · kL2 (cf. Figure 3)

Table 4 Relative distances w. r. t. | · |H 1 of the operators to each other in the situation (B3 , B7 ). The operators
are ordered by their proximity to Q. The value in a cell is relative to the operator specified by the row.

Q P Rr=2 I Qe Rr=1 Rr=0

Q 0.16 0.28 0.30 0.45 0.49 0.53
P 0.17 0.22 0.21 0.39 0.47 0.49
Rr=2 0.32 0.23 0.22 0.23 0.26 0.31
I 0.35 0.22 0.22 0.35 0.39 0.44
Qe 0.52 0.41 0.23 0.36 0.24 0.10
Rr=1 0.55 0.48 0.26 0.39 0.24 0.29
Rr=0 0.61 0.51 0.32 0.45 0.10 0.30
20 Thomas Dickopf, Rolf Krause

0.8 0.8 0.8 0.8 0.8 0.8

0.7 0.7 0.7 0.7 0.7 0.7

0.6 0.6 0.6 0.6 0.6 0.6

0.5 0.5 0.5 0.5 0.5 0.5

0.4 0.4 0.4 0.4 0.4 0.4

0.3 0.3 0.3 0.3 0.3 0.3

0.2 0.2 0.2 0.2 0.2 0.2

0.1 0.1 0.1 0.1 0.1 0.1

0 0 0 0 0 0

(a) Ball (b) Cylinder (c) Half torus

Fig. 5 Relative distances to Q with respect to | · |H 1 in selected situations. For the four cases within each dia-
gram, the ratio of the numbers of elements is roughly comparable. The ranges of this ratio are the following:
Ball: 1.9 to 2.1 (left); 5.8 to 6.3 (right). Cylinder: 3.0 to 3.4 (left); 5.9 to 6.3 (right). Half torus: 2.7 to 3.3
(left); 8.7 to 9.9 (right).

0.8 0.8 0.8 0.8 0.8 0.8

0.7 0.7 0.7 0.7 0.7 0.7

0.6 0.6 0.6 0.6 0.6 0.6

0.5 0.5 0.5 0.5 0.5 0.5

0.4 0.4 0.4 0.4 0.4 0.4

0.3 0.3 0.3 0.3 0.3 0.3

0.2 0.2 0.2 0.2 0.2 0.2

0.1 0.1 0.1 0.1 0.1 0.1

0 0 0 0 0 0

(a) Ball (b) Cylinder (c) Half torus

Fig. 6 Relative distances to Rr=2 with respect to | · |H 1 in selected situations (cf. Figure 5)

This section is concluded with a sketch summarizing the overall state. We visualize
the interconnections between the transfer operators by Figure 7. In this map-like graph, the
lengths of the lines represent the distances of the connected operators with respect to | · |H 1 .
We pick a typical situation; here, the operators generated from B3 to B7 are considered. The
studies throughout the paper confirm that other situations or some averages yield essentially
the same result as the sizes of the mutual distances are reasonably stable.

4 Numerical studies for a complex geometry: The Stanford Bunny

In this section, we present numerical studies for the widely used bunny model provided
by the Stanford 3D Scanning Repository [54]. This is done to further demonstrate that the
results seem to be largely independent of the underlying computational domain.
The original geometry data from [54] describes a surface with boundary by 69,451 tri-
angles. (There are five holes in the lower part of the geometry.) We want to consider “the
interior” of this surface as 3D computational domain. For this purpose, we fixed the holes to
obtain a closed surface which is the boundary of a simply connected domain with the shape
of the Stanford Bunny. This surface is in an intermediate step approximated sufficiently
accurately using NURBS, which is fairly standard in industrial applications.
Evaluating local approximations of the L2 -orthogonal projection 21

P
Q
Rr=0

Qe

Rr=2

Rr=1

Fig. 7 The mutual relations of the single operators visualized as a map-like graph. The length of each connect-
ing line represents the H 1 -distance between the respective operators. The lines from or to the L2 -orthogonal
projection Q are straight.

Table 5 Stanford Bunny: Characteristics of the independently generated meshes (Si )1≤i≤7

#elements #nodes
S1 6,447 1,479
S2 12,732 2,720
S3 26,355 5,380
S4 41,256 8,216
S5 83,213 16,095
S6 124,576 23,856
S7 184,783 34,860

Fig. 8 Stanford Bunny [54]: Examples for the meshes of Table 5. Here we show S2 (left) and S6 (right).
22 Thomas Dickopf, Rolf Krause

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
S1 S2 S3 S4 S5 S6
Fig. 9 Stanford Bunny: Relative distances to the L2 -projection Q with respect to | · |H 1 (cf. Figure 3)

Then, as done in Section 3.1 for the simpler geometric shapes, unstructured volume
meshes are generated by standard tools from CUBIT [50]. We use the seven tetrahedral
meshes (Si )1≤i≤7 with their characteristics given in Table 5 ordered by the number of el-
ements. They are again completely independent of each other and cover a broad range of
sizes. Two of the meshes are illustrated in Figure 8.
Proceeding as in Section 3, we study the mutual relations between the different transfer
operators when evaluated between the finite element spaces associated with pairs of meshes
(Si , S j )1≤i< j≤7 . Figure 9 shows the relative distances of the operators to the L2 -projection
with respect to | · |H 1 . The results are very similar to the ones obtained earlier for the simpler
geometric shapes; cf. Figure 3. The other findings of Section 3.4 hold true for the study
of the Stanford Bunny, too. In particular the other measured distances are very similar (not
shown here).

5 Conclusion

To investigate the information transfer between finite element spaces associated with non-
nested meshes, we examined a variety of transfer operators. The numerical studies provided
insight into their mutual relations. We considered the standard (nodal) finite element inter-
polation, Clément’s quasi-interpolation with different local polynomial degrees, the global
L2 -orthogonal projection, a local L2 -quasi-projection via a discrete inner product motivated
by a quadrature rule, and a pseudo-L2 -projection defined by a Petrov–Galerkin variational
equation with a discontinuous test space. We reviewed basic properties of the operators and
pointed out conceptual and implementational similarities and differences. The comprehen-
sive computational comparison of the transfer operators showed that the differences between
them are substantial. For the presented geometric test cases in 3D (unstructered meshes of
ball, cylinder, half torus and Stanford Bunny), the pseudo-L2 -projection turned out to be
clearly the closest to the actual L2 -projection compared to all other operators.

Acknowledgements This work was partly supported by the Bonn International Graduate School in Mathe-
matics and by the Iniziativa Ticino in Rete. We would like to thank the participants of the Söllerhaus Work-
shop on Domain Decomposition Methods in October 2011 and an anonymous referee for useful remarks on
the topic of this paper. We appreciate the help of Johannes Steiner in preparing the Stanford Bunny for the
experiments.
Evaluating local approximations of the L2 -orthogonal projection 23

References

1. Adams, R.A.: Sobolev Spaces. Academic Press, New York (1975)

2. Ainsworth, H.: Octree C++ General Component (2005)
3. Apel, T.: Anisotropic Finite Elements: Local Estimates and Applications. Advances in Numerical Math-
ematics. Teubner, Stuttgart (1999)
4. Apel, T.: Interpolation in h-version finite element spaces. In: E. Stein, R. de Borst, T.J.R. Hughes (eds.)
Encyclopedia of Computational Mechanics. Vol. 1. Fundamentals, pp. 55–72. Wiley, Chichester (2004)
5. Bank, R.E., Dupont, T.F., Yserentant, H.: The hierarchical basis multigrid method. Numer. Math. 52(4),
427–458 (1988)
6. Barber, C.M., Dobkin, D.P., Huhdanpaa, H.: The quickhull algorithm for convex hulls. ACM Trans.
Math. Softw. 22(4), 469–483 (1996)
7. Bastian, P., Birken, K., Johannsen, K., Lang, S., Neuß, N., Rentz-Reichert, H., Wieners, C.: UG – a
flexible software toolbox for solving partial differential equations. Comput. Vis. Sci. 1(1), 27–40 (1997)
8. Ben Belgacem, F.: The mortar finite element method with Lagrange multipliers. Numer. Math. 84(2),
173–197 (1999)
9. Bernardi, C., Maday, Y., Patera, A.T.: A new nonconforming approach to domain decomposition: the
mortar element method. In: H. Brezis, J.L. Lions (eds.) Nonlinear Partial Differential Equations and
Their Applications, Pitman Res. Notes Math. Ser., vol. 299, pp. 13–51. Harlow: Longman Scientific &
Technical, New York (1994)
10. Braess, D.: Finite Elemente. Springer, Berlin (2007)
11. Braess, D., Verfürth, R.: Multigrid methods for nonconforming finite element methods. SIAM J. Numer.
Anal. 27(4), 979–986 (1990)
12. Bramble, J.H., Pasciak, J.E., Steinbach, O.: On the stability of the L2 -projection in H 1 (Ω ). Math. Comp.
71(237), 147–156 (2002)
13. Bramble, J.H., Pasciak, J.E., Vassilevski, P.S.: Computational scales of Sobolev norms with applications
to preconditioning. Math. Comp. 69(230), 463–480 (2000)
14. Bramble, J.H., Pasciak, J.E., Wang, J., Xu, J.: Convergence estimates for multigrid algorithms without
regularity assumptions. Math. Comp. 57(195), 23–45 (1991)
15. Bramble, J.H., Pasciak, J.E., Xu, J.: The analysis of multigrid algorithms with nonnested spaces or non-
inherited quadratic forms. Math. Comp. 56(193), 1–34 (1991)
16. Bramble, J.H., Xu, J.: Some estimates for a weighted L2 projection. Math. Comp. 56(194), 463–476
(1991)
17. Brandt, A.: Multi-level adaptive solutions to boundary-value problems. Math. Comp. 31(138), 333–390
(1977)
18. Brenner, S.C.: Convergence of nonconforming V -cycle and F-cycle multigrid algorithms for second
order elliptic boundary value problems. Math. Comp. 73(247), 1041–1066 (2004)
19. X.-C. Cai: The use of pointwise interpolation in domain decomposition methods with non-nested meshes.
SIAM J. Sci. Comput. 16(1), 250–256 (1995)
20. Carstensen, C.: Quasi-interpolation and a posteriori error analysis in finite element methods. Math.
Model. Numer. Anal. 33(6), 1187–1202 (1999)
21. Carstensen, C.: Merging the Bramble-Pasciak-Steinbach and the Crouzeix-Thomée criterion for H 1 -
stability of the L2 -projection onto finite element spaces. Math. Comp. 71(237), 157–163 (2002)
22. Carstensen, C.: Clément interpolation and its role in adaptive finite element error control. In: E. Koelink,
J. van Neerven, B. de Pagter, G. Sweers (eds.) Partial Differential Equations and Functional Analysis –
The Philippe Clément Festschrift, Oper. Theory Adv. Appl., vol. 168, pp. 27–43. Birkhäuser, Basel (2006)
23. Chan, T.F., Smith, B.F., Zou, J.: Overlapping Schwarz methods on unstructured meshes using non-
matching coarse grids. Numer. Math. 73(2), 149–167 (1996)
24. Ciarlet, P.G.: The Finite Element Method for Elliptic Problems, Studies in Mathematics and its Applica-
tions, vol. 4. North-Holland, Amsterdam (1978)
25. Clément, P.: Approximation by finite element functions using local regularization. RAIRO Anal. Numér.
9(R-2), 77–84 (1975)
26. Cohen, A., Daubechies, I., Feauveau, J.C.: Biorthogonal bases of compactly supported wavelets. Comm.
Pure Appl. Math. 45(5), 485–560 (1992)
27. Crouzeix, M., Thomée, V.: The stability in L p and Wp1 of the L2 -projection onto finite element function
spaces. Math. Comp. 48(178), 521–532 (1987)
28. Dahmen, W., Kunoth, A., Urban, K.: Biorthogonal spline wavelets on the interval – stability and moment
conditions. Appl. Comput. Harmon. Anal. 6(2), 132–196 (1999)
29. Dickopf, T.: Multilevel methods based on non-nested meshes. Ph.D. thesis, University of Bonn (2010).
http://hss.ulb.uni-bonn.de/2010/2365
24 Thomas Dickopf, Rolf Krause

30. Dickopf, T.: Nodal interpolation between first-order finite element spaces in 1d is uniformly H 1 -stable.
Tech. Rep. 2011-09, Institute of Computational Science, University of Lugano (2011). Accepted by
Numerical Mathematics and Advanced Applications, Proceedings of ENUMATH 2011, Springer
31. Dickopf, T., Krause, R.: Efficient simulation of multi-body contact problems on complex geometries: a
flexible decomposition approach using constrained minimization. Int. J. Numer. Methods Engrg. 77(13),
1834–1862 (2009)
32. Dickopf, T., Krause, R.: Weak information transfer between non-matching warped interfaces. In:
M. Bercovier, M.J. Gander, R. Kornhuber, O.B. Widlund (eds.) Domain Decomposition Methods in
Science and Engineering XVIII, Lect. Notes Comput. Sci. Eng., vol. 70, pp. 283–290. Springer, Berlin
(2009)
33. Douglas jun., J., Dupont, T.F., Wahlbin, L.: The stability in Lq of the L2 -projection into finite element
function spaces. Numer. Math. 23(3), 193–197 (1975)
34. Evans, L.C.: Partial Differential Equations, Graduate Studies in Mathematics, vol. 19. AMS, Providence,
RI, USA (1998)
35. Falletta, S.: The approximate integration in the mortar method constraint. In: O.B. Widlund, D.E. Keyes
(eds.) Domain Decomposition Methods in Science and Engineering XVI, Lect. Notes Comput. Sci. Eng.,
vol. 55, pp. 555–563. Springer, Berlin (2007)
36. Flemisch, B., Wohlmuth, B.: Stable Lagrange multipliers for quadrilateral meshes of curved interfaces
in 3D. Comput. Methods Appl. Mech. Eng. 196(8), 1589–1602 (2007)
37. Gander, M.J., Japhet, C.: An algorithm for non-matching grid projections with linear complexity. In:
M. Bercovier, M.J. Gander, R. Kornhuber, O.B. Widlund (eds.) Domain Decomposition Methods in
Science and Engineering XVIII, Lect. Notes Comput. Sci. Eng., vol. 70, pp. 185–192. Springer, Berlin
(2009)
38. Griebel, M.: Multilevelmethoden als Iterationsverfahren über Erzeugendensystemen. Teubner Skripten
zur Numerik. Teubner, Stuttgart (1994)
39. Hackbusch, W.: Multi-Grid Methods and Applications, Springer Series in Computational Mathematics,
vol. 4. Springer, Berlin (1985)
40. Kim, C., Lazarov, R.D., Pasciak, J.E., Vassilevski, P.S.: Multiplier spaces for the mortar finite element
method in three dimensions. SIAM J. Numer. Anal. 39(2), 519–538 (2001)
41. Knyazev, A.V.: Toward the optimal preconditioned eigensolver: Locally optimal block preconditioned
conjugate gradient method. SIAM J. Sci. Comput. 23(2), 517–541 (2001)
42. Lamichhane, B.P.: Higher order mortar finite elements with dual lagrange multiplier spaces and applica-
tions. Ph.D. thesis, University of Stuttgart (2006)
43. Lamichhane, B.P., Wohlmuth, B.: Biorthogonal bases with local support and approximation properties.
Math. Comp. 76(257), 233–249 (2007)
44. Maday, Y., Rapetti, F., Wohlmuth, B.: The influence of quadrature formulas in 2d and 3d mortar element
methods. In: L.F. Pavarino, A. Toselli (eds.) Recent Developments in Domain Decomposition Methods,
Lect. Notes Comput. Sci. Eng., vol. 23, pp. 203–221. Springer, Berlin (2002)
45. Oswald, P.: Multilevel Finite Element Approximation. Theory and Applications. Teubner Skripten zur
Numerik. Teubner, Stuttgart (1994)
46. Oswald, P.: Intergrid transfer operators and multilevel preconditioners for nonconforming discretizations.
Appl. Numer. Math. 23(1), 139–158 (1997)
47. Oswald, P.: Optimality of multilevel preconditioning for nonconforming P1 finite elements. Numer.
Math. 111(2), 267–291 (2008)
48. Oswald, P., Wohlmuth, B.: On polynomial reproduction of dual FE bases. In: N. Debit, M. Garbey,
R. Hoppe, J. Périaux, D.E. Keyes, Y. Kuznetsov (eds.) Thirteenth International Conference on Domain
Decomposition Methods, pp. 85–96. CIMNE, Barcelona (2002)
49. Ruge, J.W., Stüben, K.: Algebraic multigrid. In: S.F. McCormick (ed.) Multigrid Methods, Frontiers in
Applied Mathematics, vol. 3, pp. 73–130. SIAM, Philadelphia, PA, USA (1987)
50. Sandia National Laboratories: CUBIT (2012). http://cubit.sandia.gov
51. Schenk, O., Gärtner, K.: Solving unsymmetric sparse systems of linear equations with PARDISO. Future
Generation Computer Systems 20(3), 475–487 (2004)
52. Schenk, O., Gärtner, K.: On fast factorization pivoting methods for sparse symmetric indefinite systems.
Electron. Trans. Numer. Anal. 23, 158–179 (2006)
53. Scott, L.R., Zhang, S.: Finite element interpolation of nonsmooth functions satisfying boundary condi-
tions. Math. Comp. 54(190), 483–493 (1990)
54. Stanford 3D Scanning Repository: Stanford Bunny (1994). http://graphics.stanford.edu/data/3Dscanrep
55. Steinbach, O.: On a generalized L2 -projection and some related stability estimates in Sobolev spaces.
Numer. Math. 90(4), 775–786 (2002)
56. Steinbach, O.: Stability Estimates for Hybrid Coupled Domain Decomposition Methods, Lecture Notes
in Mathematics, vol. 1809. Springer, Berlin (2003)
 Evaluating local approximations of the L2 -orthogonal projection 25

57. Toselli, A., Widlund, O.B.: Domain Decomposition Methods – Algorithms and Theory, Springer Ser.
Comput. Math., vol. 34. Springer, Berlin (2005)
58. Trottenberg, U., Oosterlee, C.W., Schüller, A.: Multigrid. Academic Press, Orlando (2001)
59. Verfürth, R.: Error estimates for some quasi-interpolation operators. Math. Model. Numer. Anal. 33(4),
695–713 (1999)
60. Vujičić, M.: Linear Algebra Thoroughly Explained. Springer, Berlin (2008)
61. Wohlmuth, B.: A mortar finite element method using dual spaces for the lagrange multiplier. SIAM J.
Numer. Anal. 38(3), 989–1012 (2000)
62. Wohlmuth, B.: Discretization Methods and Iterative Solvers Based on Domain Decomposition, Lect.
Notes Comput. Sci. Eng., vol. 17. Springer, Berlin (2001)
63. Xu, J.: Theory of multilevel methods. Ph.D. thesis, Cornell University (1989)
64. Xu, J.: The auxiliary space method and optimal multigrid preconditioning techniques for unstructured
grids. Computing 56(3), 215–235 (1996)
65. Yserentant, H.: Old and new convergence proofs for multigrid methods. Acta Numerica 2, 285–326
(1993)

View publication stats